Merge branch 'master' of git://git.kernel.org/pub/scm/fs/xfs/xfs
[pandora-kernel.git] / fs / xfs / xfs_log.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
42 #include "xfs_rw.h"
43
44 kmem_zone_t     *xfs_log_ticket_zone;
45
46 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
47         { (ptr) += (bytes); \
48           (len) -= (bytes); \
49           (off) += (bytes);}
50
51 /* Local miscellaneous function prototypes */
52 STATIC int       xlog_bdstrat_cb(struct xfs_buf *);
53 STATIC int       xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
54                                     xlog_in_core_t **, xfs_lsn_t *);
55 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
56                                 xfs_buftarg_t   *log_target,
57                                 xfs_daddr_t     blk_offset,
58                                 int             num_bblks);
59 STATIC int       xlog_space_left(xlog_t *log, int cycle, int bytes);
60 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
61 STATIC void      xlog_dealloc_log(xlog_t *log);
62 STATIC int       xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
63                             int nentries, xfs_log_ticket_t tic,
64                             xfs_lsn_t *start_lsn,
65                             xlog_in_core_t **commit_iclog,
66                             uint flags);
67
68 /* local state machine functions */
69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
70 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
71 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
72                                        int              len,
73                                        xlog_in_core_t   **iclog,
74                                        xlog_ticket_t    *ticket,
75                                        int              *continued_write,
76                                        int              *logoffsetp);
77 STATIC int  xlog_state_release_iclog(xlog_t             *log,
78                                      xlog_in_core_t     *iclog);
79 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
80                                      xlog_in_core_t *iclog,
81                                      int                eventual_size);
82 STATIC int  xlog_state_sync(xlog_t                      *log,
83                             xfs_lsn_t                   lsn,
84                             uint                        flags,
85                             int                         *log_flushed);
86 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
87 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
88
89 /* local functions to manipulate grant head */
90 STATIC int  xlog_grant_log_space(xlog_t         *log,
91                                  xlog_ticket_t  *xtic);
92 STATIC void xlog_grant_push_ail(xfs_mount_t     *mp,
93                                 int             need_bytes);
94 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
95                                            xlog_ticket_t *ticket);
96 STATIC int xlog_regrant_write_log_space(xlog_t          *log,
97                                          xlog_ticket_t  *ticket);
98 STATIC void xlog_ungrant_log_space(xlog_t        *log,
99                                    xlog_ticket_t *ticket);
100
101
102 /* local ticket functions */
103 STATIC xlog_ticket_t    *xlog_ticket_alloc(xlog_t *log,
104                                          int    unit_bytes,
105                                          int    count,
106                                          char   clientid,
107                                          uint   flags);
108
109 #if defined(DEBUG)
110 STATIC void     xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
111 STATIC void     xlog_verify_grant_head(xlog_t *log, int equals);
112 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
113                                   int count, boolean_t syncing);
114 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
115                                      xfs_lsn_t tail_lsn);
116 #else
117 #define xlog_verify_dest_ptr(a,b)
118 #define xlog_verify_grant_head(a,b)
119 #define xlog_verify_iclog(a,b,c,d)
120 #define xlog_verify_tail_lsn(a,b,c)
121 #endif
122
123 STATIC int      xlog_iclogs_empty(xlog_t *log);
124
125 #if defined(XFS_LOG_TRACE)
126
127 #define XLOG_TRACE_LOGGRANT_SIZE        2048
128 #define XLOG_TRACE_ICLOG_SIZE           256
129
130 void
131 xlog_trace_loggrant_alloc(xlog_t *log)
132 {
133         log->l_grant_trace = ktrace_alloc(XLOG_TRACE_LOGGRANT_SIZE, KM_NOFS);
134 }
135
136 void
137 xlog_trace_loggrant_dealloc(xlog_t *log)
138 {
139         ktrace_free(log->l_grant_trace);
140 }
141
142 void
143 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
144 {
145         unsigned long cnts;
146
147         /* ticket counts are 1 byte each */
148         cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
149
150         ktrace_enter(log->l_grant_trace,
151                      (void *)tic,
152                      (void *)log->l_reserve_headq,
153                      (void *)log->l_write_headq,
154                      (void *)((unsigned long)log->l_grant_reserve_cycle),
155                      (void *)((unsigned long)log->l_grant_reserve_bytes),
156                      (void *)((unsigned long)log->l_grant_write_cycle),
157                      (void *)((unsigned long)log->l_grant_write_bytes),
158                      (void *)((unsigned long)log->l_curr_cycle),
159                      (void *)((unsigned long)log->l_curr_block),
160                      (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
161                      (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
162                      (void *)string,
163                      (void *)((unsigned long)tic->t_trans_type),
164                      (void *)cnts,
165                      (void *)((unsigned long)tic->t_curr_res),
166                      (void *)((unsigned long)tic->t_unit_res));
167 }
168
169 void
170 xlog_trace_iclog_alloc(xlog_in_core_t *iclog)
171 {
172         iclog->ic_trace = ktrace_alloc(XLOG_TRACE_ICLOG_SIZE, KM_NOFS);
173 }
174
175 void
176 xlog_trace_iclog_dealloc(xlog_in_core_t *iclog)
177 {
178         ktrace_free(iclog->ic_trace);
179 }
180
181 void
182 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
183 {
184         ktrace_enter(iclog->ic_trace,
185                      (void *)((unsigned long)state),
186                      (void *)((unsigned long)current_pid()),
187                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
188                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
189                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
190                      (void *)NULL, (void *)NULL);
191 }
192 #else
193
194 #define xlog_trace_loggrant_alloc(log)
195 #define xlog_trace_loggrant_dealloc(log)
196 #define xlog_trace_loggrant(log,tic,string)
197
198 #define xlog_trace_iclog_alloc(iclog)
199 #define xlog_trace_iclog_dealloc(iclog)
200 #define xlog_trace_iclog(iclog,state)
201
202 #endif /* XFS_LOG_TRACE */
203
204
205 static void
206 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
207 {
208         if (*qp) {
209                 tic->t_next         = (*qp);
210                 tic->t_prev         = (*qp)->t_prev;
211                 (*qp)->t_prev->t_next = tic;
212                 (*qp)->t_prev       = tic;
213         } else {
214                 tic->t_prev = tic->t_next = tic;
215                 *qp = tic;
216         }
217
218         tic->t_flags |= XLOG_TIC_IN_Q;
219 }
220
221 static void
222 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
223 {
224         if (tic == tic->t_next) {
225                 *qp = NULL;
226         } else {
227                 *qp = tic->t_next;
228                 tic->t_next->t_prev = tic->t_prev;
229                 tic->t_prev->t_next = tic->t_next;
230         }
231
232         tic->t_next = tic->t_prev = NULL;
233         tic->t_flags &= ~XLOG_TIC_IN_Q;
234 }
235
236 static void
237 xlog_grant_sub_space(struct log *log, int bytes)
238 {
239         log->l_grant_write_bytes -= bytes;
240         if (log->l_grant_write_bytes < 0) {
241                 log->l_grant_write_bytes += log->l_logsize;
242                 log->l_grant_write_cycle--;
243         }
244
245         log->l_grant_reserve_bytes -= bytes;
246         if ((log)->l_grant_reserve_bytes < 0) {
247                 log->l_grant_reserve_bytes += log->l_logsize;
248                 log->l_grant_reserve_cycle--;
249         }
250
251 }
252
253 static void
254 xlog_grant_add_space_write(struct log *log, int bytes)
255 {
256         int tmp = log->l_logsize - log->l_grant_write_bytes;
257         if (tmp > bytes)
258                 log->l_grant_write_bytes += bytes;
259         else {
260                 log->l_grant_write_cycle++;
261                 log->l_grant_write_bytes = bytes - tmp;
262         }
263 }
264
265 static void
266 xlog_grant_add_space_reserve(struct log *log, int bytes)
267 {
268         int tmp = log->l_logsize - log->l_grant_reserve_bytes;
269         if (tmp > bytes)
270                 log->l_grant_reserve_bytes += bytes;
271         else {
272                 log->l_grant_reserve_cycle++;
273                 log->l_grant_reserve_bytes = bytes - tmp;
274         }
275 }
276
277 static inline void
278 xlog_grant_add_space(struct log *log, int bytes)
279 {
280         xlog_grant_add_space_write(log, bytes);
281         xlog_grant_add_space_reserve(log, bytes);
282 }
283
284 static void
285 xlog_tic_reset_res(xlog_ticket_t *tic)
286 {
287         tic->t_res_num = 0;
288         tic->t_res_arr_sum = 0;
289         tic->t_res_num_ophdrs = 0;
290 }
291
292 static void
293 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
294 {
295         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
296                 /* add to overflow and start again */
297                 tic->t_res_o_flow += tic->t_res_arr_sum;
298                 tic->t_res_num = 0;
299                 tic->t_res_arr_sum = 0;
300         }
301
302         tic->t_res_arr[tic->t_res_num].r_len = len;
303         tic->t_res_arr[tic->t_res_num].r_type = type;
304         tic->t_res_arr_sum += len;
305         tic->t_res_num++;
306 }
307
308 /*
309  * NOTES:
310  *
311  *      1. currblock field gets updated at startup and after in-core logs
312  *              marked as with WANT_SYNC.
313  */
314
315 /*
316  * This routine is called when a user of a log manager ticket is done with
317  * the reservation.  If the ticket was ever used, then a commit record for
318  * the associated transaction is written out as a log operation header with
319  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
320  * a given ticket.  If the ticket was one with a permanent reservation, then
321  * a few operations are done differently.  Permanent reservation tickets by
322  * default don't release the reservation.  They just commit the current
323  * transaction with the belief that the reservation is still needed.  A flag
324  * must be passed in before permanent reservations are actually released.
325  * When these type of tickets are not released, they need to be set into
326  * the inited state again.  By doing this, a start record will be written
327  * out when the next write occurs.
328  */
329 xfs_lsn_t
330 xfs_log_done(xfs_mount_t        *mp,
331              xfs_log_ticket_t   xtic,
332              void               **iclog,
333              uint               flags)
334 {
335         xlog_t          *log    = mp->m_log;
336         xlog_ticket_t   *ticket = (xfs_log_ticket_t) xtic;
337         xfs_lsn_t       lsn     = 0;
338
339         if (XLOG_FORCED_SHUTDOWN(log) ||
340             /*
341              * If nothing was ever written, don't write out commit record.
342              * If we get an error, just continue and give back the log ticket.
343              */
344             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
345              (xlog_commit_record(mp, ticket,
346                                  (xlog_in_core_t **)iclog, &lsn)))) {
347                 lsn = (xfs_lsn_t) -1;
348                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
349                         flags |= XFS_LOG_REL_PERM_RESERV;
350                 }
351         }
352
353
354         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
355             (flags & XFS_LOG_REL_PERM_RESERV)) {
356                 /*
357                  * Release ticket if not permanent reservation or a specific
358                  * request has been made to release a permanent reservation.
359                  */
360                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
361                 xlog_ungrant_log_space(log, ticket);
362                 xfs_log_ticket_put(ticket);
363         } else {
364                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
365                 xlog_regrant_reserve_log_space(log, ticket);
366                 /* If this ticket was a permanent reservation and we aren't
367                  * trying to release it, reset the inited flags; so next time
368                  * we write, a start record will be written out.
369                  */
370                 ticket->t_flags |= XLOG_TIC_INITED;
371         }
372
373         return lsn;
374 }       /* xfs_log_done */
375
376
377 /*
378  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
379  *      the force is done synchronously.
380  *
381  * Asynchronous forces are implemented by setting the WANT_SYNC
382  * bit in the appropriate in-core log and then returning.
383  *
384  * Synchronous forces are implemented with a signal variable. All callers
385  * to force a given lsn to disk will wait on a the sv attached to the
386  * specific in-core log.  When given in-core log finally completes its
387  * write to disk, that thread will wake up all threads waiting on the
388  * sv.
389  */
390 int
391 _xfs_log_force(
392         xfs_mount_t     *mp,
393         xfs_lsn_t       lsn,
394         uint            flags,
395         int             *log_flushed)
396 {
397         xlog_t          *log = mp->m_log;
398         int             dummy;
399
400         if (!log_flushed)
401                 log_flushed = &dummy;
402
403         ASSERT(flags & XFS_LOG_FORCE);
404
405         XFS_STATS_INC(xs_log_force);
406
407         if (log->l_flags & XLOG_IO_ERROR)
408                 return XFS_ERROR(EIO);
409         if (lsn == 0)
410                 return xlog_state_sync_all(log, flags, log_flushed);
411         else
412                 return xlog_state_sync(log, lsn, flags, log_flushed);
413 }       /* _xfs_log_force */
414
415 /*
416  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
417  * about errors or whether the log was flushed or not. This is the normal
418  * interface to use when trying to unpin items or move the log forward.
419  */
420 void
421 xfs_log_force(
422         xfs_mount_t     *mp,
423         xfs_lsn_t       lsn,
424         uint            flags)
425 {
426         int     error;
427         error = _xfs_log_force(mp, lsn, flags, NULL);
428         if (error) {
429                 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
430                         "error %d returned.", error);
431         }
432 }
433
434
435 /*
436  * Attaches a new iclog I/O completion callback routine during
437  * transaction commit.  If the log is in error state, a non-zero
438  * return code is handed back and the caller is responsible for
439  * executing the callback at an appropriate time.
440  */
441 int
442 xfs_log_notify(xfs_mount_t        *mp,          /* mount of partition */
443                void               *iclog_hndl,  /* iclog to hang callback off */
444                xfs_log_callback_t *cb)
445 {
446         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
447         int     abortflg;
448
449         spin_lock(&iclog->ic_callback_lock);
450         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
451         if (!abortflg) {
452                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
453                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
454                 cb->cb_next = NULL;
455                 *(iclog->ic_callback_tail) = cb;
456                 iclog->ic_callback_tail = &(cb->cb_next);
457         }
458         spin_unlock(&iclog->ic_callback_lock);
459         return abortflg;
460 }       /* xfs_log_notify */
461
462 int
463 xfs_log_release_iclog(xfs_mount_t *mp,
464                       void        *iclog_hndl)
465 {
466         xlog_t *log = mp->m_log;
467         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
468
469         if (xlog_state_release_iclog(log, iclog)) {
470                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
471                 return EIO;
472         }
473
474         return 0;
475 }
476
477 /*
478  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
479  *      to the reservation.
480  *  2. Potentially, push buffers at tail of log to disk.
481  *
482  * Each reservation is going to reserve extra space for a log record header.
483  * When writes happen to the on-disk log, we don't subtract the length of the
484  * log record header from any reservation.  By wasting space in each
485  * reservation, we prevent over allocation problems.
486  */
487 int
488 xfs_log_reserve(xfs_mount_t      *mp,
489                 int              unit_bytes,
490                 int              cnt,
491                 xfs_log_ticket_t *ticket,
492                 __uint8_t        client,
493                 uint             flags,
494                 uint             t_type)
495 {
496         xlog_t          *log = mp->m_log;
497         xlog_ticket_t   *internal_ticket;
498         int             retval = 0;
499
500         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
501         ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
502
503         if (XLOG_FORCED_SHUTDOWN(log))
504                 return XFS_ERROR(EIO);
505
506         XFS_STATS_INC(xs_try_logspace);
507
508         if (*ticket != NULL) {
509                 ASSERT(flags & XFS_LOG_PERM_RESERV);
510                 internal_ticket = (xlog_ticket_t *)*ticket;
511                 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
512                 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
513                 retval = xlog_regrant_write_log_space(log, internal_ticket);
514         } else {
515                 /* may sleep if need to allocate more tickets */
516                 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
517                                                   client, flags);
518                 if (!internal_ticket)
519                         return XFS_ERROR(ENOMEM);
520                 internal_ticket->t_trans_type = t_type;
521                 *ticket = internal_ticket;
522                 xlog_trace_loggrant(log, internal_ticket, 
523                         (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
524                         "xfs_log_reserve: create new ticket (permanent trans)" :
525                         "xfs_log_reserve: create new ticket");
526                 xlog_grant_push_ail(mp,
527                                     (internal_ticket->t_unit_res *
528                                      internal_ticket->t_cnt));
529                 retval = xlog_grant_log_space(log, internal_ticket);
530         }
531
532         return retval;
533 }       /* xfs_log_reserve */
534
535
536 /*
537  * Mount a log filesystem
538  *
539  * mp           - ubiquitous xfs mount point structure
540  * log_target   - buftarg of on-disk log device
541  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
542  * num_bblocks  - Number of BBSIZE blocks in on-disk log
543  *
544  * Return error or zero.
545  */
546 int
547 xfs_log_mount(
548         xfs_mount_t     *mp,
549         xfs_buftarg_t   *log_target,
550         xfs_daddr_t     blk_offset,
551         int             num_bblks)
552 {
553         int             error;
554
555         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
556                 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
557         else {
558                 cmn_err(CE_NOTE,
559                         "!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
560                         mp->m_fsname);
561                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
562         }
563
564         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
565         if (!mp->m_log) {
566                 cmn_err(CE_WARN, "XFS: Log allocation failed: No memory!");
567                 error = ENOMEM;
568                 goto out;
569         }
570
571         /*
572          * Initialize the AIL now we have a log.
573          */
574         error = xfs_trans_ail_init(mp);
575         if (error) {
576                 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
577                 goto out_free_log;
578         }
579         mp->m_log->l_ailp = mp->m_ail;
580
581         /*
582          * skip log recovery on a norecovery mount.  pretend it all
583          * just worked.
584          */
585         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
586                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
587
588                 if (readonly)
589                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
590
591                 error = xlog_recover(mp->m_log);
592
593                 if (readonly)
594                         mp->m_flags |= XFS_MOUNT_RDONLY;
595                 if (error) {
596                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
597                         goto out_destroy_ail;
598                 }
599         }
600
601         /* Normal transactions can now occur */
602         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
603
604         return 0;
605
606 out_destroy_ail:
607         xfs_trans_ail_destroy(mp);
608 out_free_log:
609         xlog_dealloc_log(mp->m_log);
610 out:
611         return error;
612 }
613
614 /*
615  * Finish the recovery of the file system.  This is separate from
616  * the xfs_log_mount() call, because it depends on the code in
617  * xfs_mountfs() to read in the root and real-time bitmap inodes
618  * between calling xfs_log_mount() and here.
619  *
620  * mp           - ubiquitous xfs mount point structure
621  */
622 int
623 xfs_log_mount_finish(xfs_mount_t *mp)
624 {
625         int     error;
626
627         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
628                 error = xlog_recover_finish(mp->m_log);
629         else {
630                 error = 0;
631                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
632         }
633
634         return error;
635 }
636
637 /*
638  * Final log writes as part of unmount.
639  *
640  * Mark the filesystem clean as unmount happens.  Note that during relocation
641  * this routine needs to be executed as part of source-bag while the
642  * deallocation must not be done until source-end.
643  */
644
645 /*
646  * Unmount record used to have a string "Unmount filesystem--" in the
647  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
648  * We just write the magic number now since that particular field isn't
649  * currently architecture converted and "nUmount" is a bit foo.
650  * As far as I know, there weren't any dependencies on the old behaviour.
651  */
652
653 int
654 xfs_log_unmount_write(xfs_mount_t *mp)
655 {
656         xlog_t           *log = mp->m_log;
657         xlog_in_core_t   *iclog;
658 #ifdef DEBUG
659         xlog_in_core_t   *first_iclog;
660 #endif
661         xfs_log_iovec_t  reg[1];
662         xfs_log_ticket_t tic = NULL;
663         xfs_lsn_t        lsn;
664         int              error;
665
666         /* the data section must be 32 bit size aligned */
667         struct {
668             __uint16_t magic;
669             __uint16_t pad1;
670             __uint32_t pad2; /* may as well make it 64 bits */
671         } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
672
673         /*
674          * Don't write out unmount record on read-only mounts.
675          * Or, if we are doing a forced umount (typically because of IO errors).
676          */
677         if (mp->m_flags & XFS_MOUNT_RDONLY)
678                 return 0;
679
680         error = _xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC, NULL);
681         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
682
683 #ifdef DEBUG
684         first_iclog = iclog = log->l_iclog;
685         do {
686                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
687                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
688                         ASSERT(iclog->ic_offset == 0);
689                 }
690                 iclog = iclog->ic_next;
691         } while (iclog != first_iclog);
692 #endif
693         if (! (XLOG_FORCED_SHUTDOWN(log))) {
694                 reg[0].i_addr = (void*)&magic;
695                 reg[0].i_len  = sizeof(magic);
696                 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
697
698                 error = xfs_log_reserve(mp, 600, 1, &tic,
699                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
700                 if (!error) {
701                         /* remove inited flag */
702                         ((xlog_ticket_t *)tic)->t_flags = 0;
703                         error = xlog_write(mp, reg, 1, tic, &lsn,
704                                            NULL, XLOG_UNMOUNT_TRANS);
705                         /*
706                          * At this point, we're umounting anyway,
707                          * so there's no point in transitioning log state
708                          * to IOERROR. Just continue...
709                          */
710                 }
711
712                 if (error) {
713                         xfs_fs_cmn_err(CE_ALERT, mp,
714                                 "xfs_log_unmount: unmount record failed");
715                 }
716
717
718                 spin_lock(&log->l_icloglock);
719                 iclog = log->l_iclog;
720                 atomic_inc(&iclog->ic_refcnt);
721                 xlog_state_want_sync(log, iclog);
722                 spin_unlock(&log->l_icloglock);
723                 error = xlog_state_release_iclog(log, iclog);
724
725                 spin_lock(&log->l_icloglock);
726                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
727                       iclog->ic_state == XLOG_STATE_DIRTY)) {
728                         if (!XLOG_FORCED_SHUTDOWN(log)) {
729                                 sv_wait(&iclog->ic_force_wait, PMEM,
730                                         &log->l_icloglock, s);
731                         } else {
732                                 spin_unlock(&log->l_icloglock);
733                         }
734                 } else {
735                         spin_unlock(&log->l_icloglock);
736                 }
737                 if (tic) {
738                         xlog_trace_loggrant(log, tic, "unmount rec");
739                         xlog_ungrant_log_space(log, tic);
740                         xfs_log_ticket_put(tic);
741                 }
742         } else {
743                 /*
744                  * We're already in forced_shutdown mode, couldn't
745                  * even attempt to write out the unmount transaction.
746                  *
747                  * Go through the motions of sync'ing and releasing
748                  * the iclog, even though no I/O will actually happen,
749                  * we need to wait for other log I/Os that may already
750                  * be in progress.  Do this as a separate section of
751                  * code so we'll know if we ever get stuck here that
752                  * we're in this odd situation of trying to unmount
753                  * a file system that went into forced_shutdown as
754                  * the result of an unmount..
755                  */
756                 spin_lock(&log->l_icloglock);
757                 iclog = log->l_iclog;
758                 atomic_inc(&iclog->ic_refcnt);
759
760                 xlog_state_want_sync(log, iclog);
761                 spin_unlock(&log->l_icloglock);
762                 error =  xlog_state_release_iclog(log, iclog);
763
764                 spin_lock(&log->l_icloglock);
765
766                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
767                         || iclog->ic_state == XLOG_STATE_DIRTY
768                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
769
770                                 sv_wait(&iclog->ic_force_wait, PMEM,
771                                         &log->l_icloglock, s);
772                 } else {
773                         spin_unlock(&log->l_icloglock);
774                 }
775         }
776
777         return error;
778 }       /* xfs_log_unmount_write */
779
780 /*
781  * Deallocate log structures for unmount/relocation.
782  *
783  * We need to stop the aild from running before we destroy
784  * and deallocate the log as the aild references the log.
785  */
786 void
787 xfs_log_unmount(xfs_mount_t *mp)
788 {
789         xfs_trans_ail_destroy(mp);
790         xlog_dealloc_log(mp->m_log);
791 }
792
793 /*
794  * Write region vectors to log.  The write happens using the space reservation
795  * of the ticket (tic).  It is not a requirement that all writes for a given
796  * transaction occur with one call to xfs_log_write().
797  */
798 int
799 xfs_log_write(xfs_mount_t *     mp,
800               xfs_log_iovec_t   reg[],
801               int               nentries,
802               xfs_log_ticket_t  tic,
803               xfs_lsn_t         *start_lsn)
804 {
805         int     error;
806         xlog_t *log = mp->m_log;
807
808         if (XLOG_FORCED_SHUTDOWN(log))
809                 return XFS_ERROR(EIO);
810
811         if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
812                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
813         }
814         return error;
815 }       /* xfs_log_write */
816
817
818 void
819 xfs_log_move_tail(xfs_mount_t   *mp,
820                   xfs_lsn_t     tail_lsn)
821 {
822         xlog_ticket_t   *tic;
823         xlog_t          *log = mp->m_log;
824         int             need_bytes, free_bytes, cycle, bytes;
825
826         if (XLOG_FORCED_SHUTDOWN(log))
827                 return;
828
829         if (tail_lsn == 0) {
830                 /* needed since sync_lsn is 64 bits */
831                 spin_lock(&log->l_icloglock);
832                 tail_lsn = log->l_last_sync_lsn;
833                 spin_unlock(&log->l_icloglock);
834         }
835
836         spin_lock(&log->l_grant_lock);
837
838         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
839          * tail_lsn.
840          */
841         if (tail_lsn != 1) {
842                 log->l_tail_lsn = tail_lsn;
843         }
844
845         if ((tic = log->l_write_headq)) {
846 #ifdef DEBUG
847                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
848                         panic("Recovery problem");
849 #endif
850                 cycle = log->l_grant_write_cycle;
851                 bytes = log->l_grant_write_bytes;
852                 free_bytes = xlog_space_left(log, cycle, bytes);
853                 do {
854                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
855
856                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
857                                 break;
858                         tail_lsn = 0;
859                         free_bytes -= tic->t_unit_res;
860                         sv_signal(&tic->t_wait);
861                         tic = tic->t_next;
862                 } while (tic != log->l_write_headq);
863         }
864         if ((tic = log->l_reserve_headq)) {
865 #ifdef DEBUG
866                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
867                         panic("Recovery problem");
868 #endif
869                 cycle = log->l_grant_reserve_cycle;
870                 bytes = log->l_grant_reserve_bytes;
871                 free_bytes = xlog_space_left(log, cycle, bytes);
872                 do {
873                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
874                                 need_bytes = tic->t_unit_res*tic->t_cnt;
875                         else
876                                 need_bytes = tic->t_unit_res;
877                         if (free_bytes < need_bytes && tail_lsn != 1)
878                                 break;
879                         tail_lsn = 0;
880                         free_bytes -= need_bytes;
881                         sv_signal(&tic->t_wait);
882                         tic = tic->t_next;
883                 } while (tic != log->l_reserve_headq);
884         }
885         spin_unlock(&log->l_grant_lock);
886 }       /* xfs_log_move_tail */
887
888 /*
889  * Determine if we have a transaction that has gone to disk
890  * that needs to be covered. Log activity needs to be idle (no AIL and
891  * nothing in the iclogs). And, we need to be in the right state indicating
892  * something has gone out.
893  */
894 int
895 xfs_log_need_covered(xfs_mount_t *mp)
896 {
897         int             needed = 0;
898         xlog_t          *log = mp->m_log;
899
900         if (!xfs_fs_writable(mp))
901                 return 0;
902
903         spin_lock(&log->l_icloglock);
904         if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
905                 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
906                         && !xfs_trans_ail_tail(log->l_ailp)
907                         && xlog_iclogs_empty(log)) {
908                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
909                         log->l_covered_state = XLOG_STATE_COVER_DONE;
910                 else {
911                         ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
912                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
913                 }
914                 needed = 1;
915         }
916         spin_unlock(&log->l_icloglock);
917         return needed;
918 }
919
920 /******************************************************************************
921  *
922  *      local routines
923  *
924  ******************************************************************************
925  */
926
927 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
928  * The log manager must keep track of the last LR which was committed
929  * to disk.  The lsn of this LR will become the new tail_lsn whenever
930  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
931  * the situation where stuff could be written into the log but nothing
932  * was ever in the AIL when asked.  Eventually, we panic since the
933  * tail hits the head.
934  *
935  * We may be holding the log iclog lock upon entering this routine.
936  */
937 xfs_lsn_t
938 xlog_assign_tail_lsn(xfs_mount_t *mp)
939 {
940         xfs_lsn_t tail_lsn;
941         xlog_t    *log = mp->m_log;
942
943         tail_lsn = xfs_trans_ail_tail(mp->m_ail);
944         spin_lock(&log->l_grant_lock);
945         if (tail_lsn != 0) {
946                 log->l_tail_lsn = tail_lsn;
947         } else {
948                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
949         }
950         spin_unlock(&log->l_grant_lock);
951
952         return tail_lsn;
953 }       /* xlog_assign_tail_lsn */
954
955
956 /*
957  * Return the space in the log between the tail and the head.  The head
958  * is passed in the cycle/bytes formal parms.  In the special case where
959  * the reserve head has wrapped passed the tail, this calculation is no
960  * longer valid.  In this case, just return 0 which means there is no space
961  * in the log.  This works for all places where this function is called
962  * with the reserve head.  Of course, if the write head were to ever
963  * wrap the tail, we should blow up.  Rather than catch this case here,
964  * we depend on other ASSERTions in other parts of the code.   XXXmiken
965  *
966  * This code also handles the case where the reservation head is behind
967  * the tail.  The details of this case are described below, but the end
968  * result is that we return the size of the log as the amount of space left.
969  */
970 STATIC int
971 xlog_space_left(xlog_t *log, int cycle, int bytes)
972 {
973         int free_bytes;
974         int tail_bytes;
975         int tail_cycle;
976
977         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
978         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
979         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
980                 free_bytes = log->l_logsize - (bytes - tail_bytes);
981         } else if ((tail_cycle + 1) < cycle) {
982                 return 0;
983         } else if (tail_cycle < cycle) {
984                 ASSERT(tail_cycle == (cycle - 1));
985                 free_bytes = tail_bytes - bytes;
986         } else {
987                 /*
988                  * The reservation head is behind the tail.
989                  * In this case we just want to return the size of the
990                  * log as the amount of space left.
991                  */
992                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
993                         "xlog_space_left: head behind tail\n"
994                         "  tail_cycle = %d, tail_bytes = %d\n"
995                         "  GH   cycle = %d, GH   bytes = %d",
996                         tail_cycle, tail_bytes, cycle, bytes);
997                 ASSERT(0);
998                 free_bytes = log->l_logsize;
999         }
1000         return free_bytes;
1001 }       /* xlog_space_left */
1002
1003
1004 /*
1005  * Log function which is called when an io completes.
1006  *
1007  * The log manager needs its own routine, in order to control what
1008  * happens with the buffer after the write completes.
1009  */
1010 void
1011 xlog_iodone(xfs_buf_t *bp)
1012 {
1013         xlog_in_core_t  *iclog;
1014         xlog_t          *l;
1015         int             aborted;
1016
1017         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1018         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
1019         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1020         aborted = 0;
1021         l = iclog->ic_log;
1022
1023         /*
1024          * If the _XFS_BARRIER_FAILED flag was set by a lower
1025          * layer, it means the underlying device no longer supports
1026          * barrier I/O. Warn loudly and turn off barriers.
1027          */
1028         if (bp->b_flags & _XFS_BARRIER_FAILED) {
1029                 bp->b_flags &= ~_XFS_BARRIER_FAILED;
1030                 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
1031                 xfs_fs_cmn_err(CE_WARN, l->l_mp,
1032                                 "xlog_iodone: Barriers are no longer supported"
1033                                 " by device. Disabling barriers\n");
1034                 xfs_buftrace("XLOG_IODONE BARRIERS OFF", bp);
1035         }
1036
1037         /*
1038          * Race to shutdown the filesystem if we see an error.
1039          */
1040         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
1041                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1042                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
1043                 XFS_BUF_STALE(bp);
1044                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1045                 /*
1046                  * This flag will be propagated to the trans-committed
1047                  * callback routines to let them know that the log-commit
1048                  * didn't succeed.
1049                  */
1050                 aborted = XFS_LI_ABORTED;
1051         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1052                 aborted = XFS_LI_ABORTED;
1053         }
1054
1055         /* log I/O is always issued ASYNC */
1056         ASSERT(XFS_BUF_ISASYNC(bp));
1057         xlog_state_done_syncing(iclog, aborted);
1058         /*
1059          * do not reference the buffer (bp) here as we could race
1060          * with it being freed after writing the unmount record to the
1061          * log.
1062          */
1063
1064 }       /* xlog_iodone */
1065
1066 /*
1067  * The bdstrat callback function for log bufs. This gives us a central
1068  * place to trap bufs in case we get hit by a log I/O error and need to
1069  * shutdown. Actually, in practice, even when we didn't get a log error,
1070  * we transition the iclogs to IOERROR state *after* flushing all existing
1071  * iclogs to disk. This is because we don't want anymore new transactions to be
1072  * started or completed afterwards.
1073  */
1074 STATIC int
1075 xlog_bdstrat_cb(struct xfs_buf *bp)
1076 {
1077         xlog_in_core_t *iclog;
1078
1079         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1080
1081         if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
1082           /* note for irix bstrat will need  struct bdevsw passed
1083            * Fix the following macro if the code ever is merged
1084            */
1085             XFS_bdstrat(bp);
1086                 return 0;
1087         }
1088
1089         xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1090         XFS_BUF_ERROR(bp, EIO);
1091         XFS_BUF_STALE(bp);
1092         xfs_biodone(bp);
1093         return XFS_ERROR(EIO);
1094
1095
1096 }
1097
1098 /*
1099  * Return size of each in-core log record buffer.
1100  *
1101  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1102  *
1103  * If the filesystem blocksize is too large, we may need to choose a
1104  * larger size since the directory code currently logs entire blocks.
1105  */
1106
1107 STATIC void
1108 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1109                            xlog_t       *log)
1110 {
1111         int size;
1112         int xhdrs;
1113
1114         if (mp->m_logbufs <= 0)
1115                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1116         else
1117                 log->l_iclog_bufs = mp->m_logbufs;
1118
1119         /*
1120          * Buffer size passed in from mount system call.
1121          */
1122         if (mp->m_logbsize > 0) {
1123                 size = log->l_iclog_size = mp->m_logbsize;
1124                 log->l_iclog_size_log = 0;
1125                 while (size != 1) {
1126                         log->l_iclog_size_log++;
1127                         size >>= 1;
1128                 }
1129
1130                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1131                         /* # headers = size / 32k
1132                          * one header holds cycles from 32k of data
1133                          */
1134
1135                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1136                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1137                                 xhdrs++;
1138                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1139                         log->l_iclog_heads = xhdrs;
1140                 } else {
1141                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1142                         log->l_iclog_hsize = BBSIZE;
1143                         log->l_iclog_heads = 1;
1144                 }
1145                 goto done;
1146         }
1147
1148         /* All machines use 32kB buffers by default. */
1149         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1150         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1151
1152         /* the default log size is 16k or 32k which is one header sector */
1153         log->l_iclog_hsize = BBSIZE;
1154         log->l_iclog_heads = 1;
1155
1156 done:
1157         /* are we being asked to make the sizes selected above visible? */
1158         if (mp->m_logbufs == 0)
1159                 mp->m_logbufs = log->l_iclog_bufs;
1160         if (mp->m_logbsize == 0)
1161                 mp->m_logbsize = log->l_iclog_size;
1162 }       /* xlog_get_iclog_buffer_size */
1163
1164
1165 /*
1166  * This routine initializes some of the log structure for a given mount point.
1167  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1168  * some other stuff may be filled in too.
1169  */
1170 STATIC xlog_t *
1171 xlog_alloc_log(xfs_mount_t      *mp,
1172                xfs_buftarg_t    *log_target,
1173                xfs_daddr_t      blk_offset,
1174                int              num_bblks)
1175 {
1176         xlog_t                  *log;
1177         xlog_rec_header_t       *head;
1178         xlog_in_core_t          **iclogp;
1179         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1180         xfs_buf_t               *bp;
1181         int                     i;
1182         int                     iclogsize;
1183
1184         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1185         if (!log)
1186                 return NULL;
1187
1188         log->l_mp          = mp;
1189         log->l_targ        = log_target;
1190         log->l_logsize     = BBTOB(num_bblks);
1191         log->l_logBBstart  = blk_offset;
1192         log->l_logBBsize   = num_bblks;
1193         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1194         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1195
1196         log->l_prev_block  = -1;
1197         log->l_tail_lsn    = xlog_assign_lsn(1, 0);
1198         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1199         log->l_last_sync_lsn = log->l_tail_lsn;
1200         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1201         log->l_grant_reserve_cycle = 1;
1202         log->l_grant_write_cycle = 1;
1203
1204         if (xfs_sb_version_hassector(&mp->m_sb)) {
1205                 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1206                 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1207                 /* for larger sector sizes, must have v2 or external log */
1208                 ASSERT(log->l_sectbb_log == 0 ||
1209                         log->l_logBBstart == 0 ||
1210                         xfs_sb_version_haslogv2(&mp->m_sb));
1211                 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1212         }
1213         log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1214
1215         xlog_get_iclog_buffer_size(mp, log);
1216
1217         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1218         if (!bp)
1219                 goto out_free_log;
1220         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1221         XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1222         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1223         ASSERT(XFS_BUF_ISBUSY(bp));
1224         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1225         log->l_xbuf = bp;
1226
1227         spin_lock_init(&log->l_icloglock);
1228         spin_lock_init(&log->l_grant_lock);
1229         sv_init(&log->l_flush_wait, 0, "flush_wait");
1230
1231         xlog_trace_loggrant_alloc(log);
1232         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1233         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1234
1235         iclogp = &log->l_iclog;
1236         /*
1237          * The amount of memory to allocate for the iclog structure is
1238          * rather funky due to the way the structure is defined.  It is
1239          * done this way so that we can use different sizes for machines
1240          * with different amounts of memory.  See the definition of
1241          * xlog_in_core_t in xfs_log_priv.h for details.
1242          */
1243         iclogsize = log->l_iclog_size;
1244         ASSERT(log->l_iclog_size >= 4096);
1245         for (i=0; i < log->l_iclog_bufs; i++) {
1246                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1247                 if (!*iclogp)
1248                         goto out_free_iclog;
1249
1250                 iclog = *iclogp;
1251                 iclog->ic_prev = prev_iclog;
1252                 prev_iclog = iclog;
1253
1254                 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1255                 if (!bp)
1256                         goto out_free_iclog;
1257                 if (!XFS_BUF_CPSEMA(bp))
1258                         ASSERT(0);
1259                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1260                 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1261                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1262                 iclog->ic_bp = bp;
1263                 iclog->ic_data = bp->b_addr;
1264 #ifdef DEBUG
1265                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1266 #endif
1267                 head = &iclog->ic_header;
1268                 memset(head, 0, sizeof(xlog_rec_header_t));
1269                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1270                 head->h_version = cpu_to_be32(
1271                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1272                 head->h_size = cpu_to_be32(log->l_iclog_size);
1273                 /* new fields */
1274                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1275                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1276
1277                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1278                 iclog->ic_state = XLOG_STATE_ACTIVE;
1279                 iclog->ic_log = log;
1280                 atomic_set(&iclog->ic_refcnt, 0);
1281                 spin_lock_init(&iclog->ic_callback_lock);
1282                 iclog->ic_callback_tail = &(iclog->ic_callback);
1283                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1284
1285                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1286                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1287                 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1288                 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1289
1290                 xlog_trace_iclog_alloc(iclog);
1291
1292                 iclogp = &iclog->ic_next;
1293         }
1294         *iclogp = log->l_iclog;                 /* complete ring */
1295         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1296
1297         return log;
1298
1299 out_free_iclog:
1300         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1301                 prev_iclog = iclog->ic_next;
1302                 if (iclog->ic_bp) {
1303                         sv_destroy(&iclog->ic_force_wait);
1304                         sv_destroy(&iclog->ic_write_wait);
1305                         xfs_buf_free(iclog->ic_bp);
1306                         xlog_trace_iclog_dealloc(iclog);
1307                 }
1308                 kmem_free(iclog);
1309         }
1310         spinlock_destroy(&log->l_icloglock);
1311         spinlock_destroy(&log->l_grant_lock);
1312         xlog_trace_loggrant_dealloc(log);
1313         xfs_buf_free(log->l_xbuf);
1314 out_free_log:
1315         kmem_free(log);
1316         return NULL;
1317 }       /* xlog_alloc_log */
1318
1319
1320 /*
1321  * Write out the commit record of a transaction associated with the given
1322  * ticket.  Return the lsn of the commit record.
1323  */
1324 STATIC int
1325 xlog_commit_record(xfs_mount_t  *mp,
1326                    xlog_ticket_t *ticket,
1327                    xlog_in_core_t **iclog,
1328                    xfs_lsn_t    *commitlsnp)
1329 {
1330         int             error;
1331         xfs_log_iovec_t reg[1];
1332
1333         reg[0].i_addr = NULL;
1334         reg[0].i_len = 0;
1335         XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1336
1337         ASSERT_ALWAYS(iclog);
1338         if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1339                                iclog, XLOG_COMMIT_TRANS))) {
1340                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1341         }
1342         return error;
1343 }       /* xlog_commit_record */
1344
1345
1346 /*
1347  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1348  * log space.  This code pushes on the lsn which would supposedly free up
1349  * the 25% which we want to leave free.  We may need to adopt a policy which
1350  * pushes on an lsn which is further along in the log once we reach the high
1351  * water mark.  In this manner, we would be creating a low water mark.
1352  */
1353 STATIC void
1354 xlog_grant_push_ail(xfs_mount_t *mp,
1355                     int         need_bytes)
1356 {
1357     xlog_t      *log = mp->m_log;       /* pointer to the log */
1358     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1359     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1360     int         free_blocks;            /* free blocks left to write to */
1361     int         free_bytes;             /* free bytes left to write to */
1362     int         threshold_block;        /* block in lsn we'd like to be at */
1363     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1364     int         free_threshold;
1365
1366     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1367
1368     spin_lock(&log->l_grant_lock);
1369     free_bytes = xlog_space_left(log,
1370                                  log->l_grant_reserve_cycle,
1371                                  log->l_grant_reserve_bytes);
1372     tail_lsn = log->l_tail_lsn;
1373     free_blocks = BTOBBT(free_bytes);
1374
1375     /*
1376      * Set the threshold for the minimum number of free blocks in the
1377      * log to the maximum of what the caller needs, one quarter of the
1378      * log, and 256 blocks.
1379      */
1380     free_threshold = BTOBB(need_bytes);
1381     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1382     free_threshold = MAX(free_threshold, 256);
1383     if (free_blocks < free_threshold) {
1384         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1385         threshold_cycle = CYCLE_LSN(tail_lsn);
1386         if (threshold_block >= log->l_logBBsize) {
1387             threshold_block -= log->l_logBBsize;
1388             threshold_cycle += 1;
1389         }
1390         threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1391
1392         /* Don't pass in an lsn greater than the lsn of the last
1393          * log record known to be on disk.
1394          */
1395         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1396             threshold_lsn = log->l_last_sync_lsn;
1397     }
1398     spin_unlock(&log->l_grant_lock);
1399
1400     /*
1401      * Get the transaction layer to kick the dirty buffers out to
1402      * disk asynchronously. No point in trying to do this if
1403      * the filesystem is shutting down.
1404      */
1405     if (threshold_lsn &&
1406         !XLOG_FORCED_SHUTDOWN(log))
1407             xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1408 }       /* xlog_grant_push_ail */
1409
1410
1411 /*
1412  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1413  * fashion.  Previously, we should have moved the current iclog
1414  * ptr in the log to point to the next available iclog.  This allows further
1415  * write to continue while this code syncs out an iclog ready to go.
1416  * Before an in-core log can be written out, the data section must be scanned
1417  * to save away the 1st word of each BBSIZE block into the header.  We replace
1418  * it with the current cycle count.  Each BBSIZE block is tagged with the
1419  * cycle count because there in an implicit assumption that drives will
1420  * guarantee that entire 512 byte blocks get written at once.  In other words,
1421  * we can't have part of a 512 byte block written and part not written.  By
1422  * tagging each block, we will know which blocks are valid when recovering
1423  * after an unclean shutdown.
1424  *
1425  * This routine is single threaded on the iclog.  No other thread can be in
1426  * this routine with the same iclog.  Changing contents of iclog can there-
1427  * fore be done without grabbing the state machine lock.  Updating the global
1428  * log will require grabbing the lock though.
1429  *
1430  * The entire log manager uses a logical block numbering scheme.  Only
1431  * log_sync (and then only bwrite()) know about the fact that the log may
1432  * not start with block zero on a given device.  The log block start offset
1433  * is added immediately before calling bwrite().
1434  */
1435
1436 STATIC int
1437 xlog_sync(xlog_t                *log,
1438           xlog_in_core_t        *iclog)
1439 {
1440         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1441         xfs_buf_t       *bp;
1442         int             i;
1443         uint            count;          /* byte count of bwrite */
1444         uint            count_init;     /* initial count before roundup */
1445         int             roundoff;       /* roundoff to BB or stripe */
1446         int             split = 0;      /* split write into two regions */
1447         int             error;
1448         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1449
1450         XFS_STATS_INC(xs_log_writes);
1451         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1452
1453         /* Add for LR header */
1454         count_init = log->l_iclog_hsize + iclog->ic_offset;
1455
1456         /* Round out the log write size */
1457         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1458                 /* we have a v2 stripe unit to use */
1459                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1460         } else {
1461                 count = BBTOB(BTOBB(count_init));
1462         }
1463         roundoff = count - count_init;
1464         ASSERT(roundoff >= 0);
1465         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1466                 roundoff < log->l_mp->m_sb.sb_logsunit)
1467                 || 
1468                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1469                  roundoff < BBTOB(1)));
1470
1471         /* move grant heads by roundoff in sync */
1472         spin_lock(&log->l_grant_lock);
1473         xlog_grant_add_space(log, roundoff);
1474         spin_unlock(&log->l_grant_lock);
1475
1476         /* put cycle number in every block */
1477         xlog_pack_data(log, iclog, roundoff); 
1478
1479         /* real byte length */
1480         if (v2) {
1481                 iclog->ic_header.h_len =
1482                         cpu_to_be32(iclog->ic_offset + roundoff);
1483         } else {
1484                 iclog->ic_header.h_len =
1485                         cpu_to_be32(iclog->ic_offset);
1486         }
1487
1488         bp = iclog->ic_bp;
1489         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1490         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1491         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1492
1493         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1494
1495         /* Do we need to split this write into 2 parts? */
1496         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1497                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1498                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1499                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1500         } else {
1501                 iclog->ic_bwritecnt = 1;
1502         }
1503         XFS_BUF_SET_COUNT(bp, count);
1504         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1505         XFS_BUF_ZEROFLAGS(bp);
1506         XFS_BUF_BUSY(bp);
1507         XFS_BUF_ASYNC(bp);
1508         /*
1509          * Do an ordered write for the log block.
1510          * Its unnecessary to flush the first split block in the log wrap case.
1511          */
1512         if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1513                 XFS_BUF_ORDERED(bp);
1514
1515         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1516         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1517
1518         xlog_verify_iclog(log, iclog, count, B_TRUE);
1519
1520         /* account for log which doesn't start at block #0 */
1521         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1522         /*
1523          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1524          * is shutting down.
1525          */
1526         XFS_BUF_WRITE(bp);
1527
1528         if ((error = XFS_bwrite(bp))) {
1529                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1530                                   XFS_BUF_ADDR(bp));
1531                 return error;
1532         }
1533         if (split) {
1534                 bp = iclog->ic_log->l_xbuf;
1535                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1536                                                         (unsigned long)1);
1537                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1538                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1539                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1540                                             (__psint_t)count), split);
1541                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1542                 XFS_BUF_ZEROFLAGS(bp);
1543                 XFS_BUF_BUSY(bp);
1544                 XFS_BUF_ASYNC(bp);
1545                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1546                         XFS_BUF_ORDERED(bp);
1547                 dptr = XFS_BUF_PTR(bp);
1548                 /*
1549                  * Bump the cycle numbers at the start of each block
1550                  * since this part of the buffer is at the start of
1551                  * a new cycle.  Watch out for the header magic number
1552                  * case, though.
1553                  */
1554                 for (i = 0; i < split; i += BBSIZE) {
1555                         be32_add_cpu((__be32 *)dptr, 1);
1556                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1557                                 be32_add_cpu((__be32 *)dptr, 1);
1558                         dptr += BBSIZE;
1559                 }
1560
1561                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1562                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1563
1564                 /* account for internal log which doesn't start at block #0 */
1565                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1566                 XFS_BUF_WRITE(bp);
1567                 if ((error = XFS_bwrite(bp))) {
1568                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1569                                           bp, XFS_BUF_ADDR(bp));
1570                         return error;
1571                 }
1572         }
1573         return 0;
1574 }       /* xlog_sync */
1575
1576
1577 /*
1578  * Deallocate a log structure
1579  */
1580 STATIC void
1581 xlog_dealloc_log(xlog_t *log)
1582 {
1583         xlog_in_core_t  *iclog, *next_iclog;
1584         int             i;
1585
1586         iclog = log->l_iclog;
1587         for (i=0; i<log->l_iclog_bufs; i++) {
1588                 sv_destroy(&iclog->ic_force_wait);
1589                 sv_destroy(&iclog->ic_write_wait);
1590                 xfs_buf_free(iclog->ic_bp);
1591                 xlog_trace_iclog_dealloc(iclog);
1592                 next_iclog = iclog->ic_next;
1593                 kmem_free(iclog);
1594                 iclog = next_iclog;
1595         }
1596         spinlock_destroy(&log->l_icloglock);
1597         spinlock_destroy(&log->l_grant_lock);
1598
1599         xfs_buf_free(log->l_xbuf);
1600         xlog_trace_loggrant_dealloc(log);
1601         log->l_mp->m_log = NULL;
1602         kmem_free(log);
1603 }       /* xlog_dealloc_log */
1604
1605 /*
1606  * Update counters atomically now that memcpy is done.
1607  */
1608 /* ARGSUSED */
1609 static inline void
1610 xlog_state_finish_copy(xlog_t           *log,
1611                        xlog_in_core_t   *iclog,
1612                        int              record_cnt,
1613                        int              copy_bytes)
1614 {
1615         spin_lock(&log->l_icloglock);
1616
1617         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1618         iclog->ic_offset += copy_bytes;
1619
1620         spin_unlock(&log->l_icloglock);
1621 }       /* xlog_state_finish_copy */
1622
1623
1624
1625
1626 /*
1627  * print out info relating to regions written which consume
1628  * the reservation
1629  */
1630 STATIC void
1631 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1632 {
1633         uint i;
1634         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1635
1636         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1637         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1638             "bformat",
1639             "bchunk",
1640             "efi_format",
1641             "efd_format",
1642             "iformat",
1643             "icore",
1644             "iext",
1645             "ibroot",
1646             "ilocal",
1647             "iattr_ext",
1648             "iattr_broot",
1649             "iattr_local",
1650             "qformat",
1651             "dquot",
1652             "quotaoff",
1653             "LR header",
1654             "unmount",
1655             "commit",
1656             "trans header"
1657         };
1658         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1659             "SETATTR_NOT_SIZE",
1660             "SETATTR_SIZE",
1661             "INACTIVE",
1662             "CREATE",
1663             "CREATE_TRUNC",
1664             "TRUNCATE_FILE",
1665             "REMOVE",
1666             "LINK",
1667             "RENAME",
1668             "MKDIR",
1669             "RMDIR",
1670             "SYMLINK",
1671             "SET_DMATTRS",
1672             "GROWFS",
1673             "STRAT_WRITE",
1674             "DIOSTRAT",
1675             "WRITE_SYNC",
1676             "WRITEID",
1677             "ADDAFORK",
1678             "ATTRINVAL",
1679             "ATRUNCATE",
1680             "ATTR_SET",
1681             "ATTR_RM",
1682             "ATTR_FLAG",
1683             "CLEAR_AGI_BUCKET",
1684             "QM_SBCHANGE",
1685             "DUMMY1",
1686             "DUMMY2",
1687             "QM_QUOTAOFF",
1688             "QM_DQALLOC",
1689             "QM_SETQLIM",
1690             "QM_DQCLUSTER",
1691             "QM_QINOCREATE",
1692             "QM_QUOTAOFF_END",
1693             "SB_UNIT",
1694             "FSYNC_TS",
1695             "GROWFSRT_ALLOC",
1696             "GROWFSRT_ZERO",
1697             "GROWFSRT_FREE",
1698             "SWAPEXT"
1699         };
1700
1701         xfs_fs_cmn_err(CE_WARN, mp,
1702                         "xfs_log_write: reservation summary:\n"
1703                         "  trans type  = %s (%u)\n"
1704                         "  unit res    = %d bytes\n"
1705                         "  current res = %d bytes\n"
1706                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1707                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1708                         "  ophdr + reg = %u bytes\n"
1709                         "  num regions = %u\n",
1710                         ((ticket->t_trans_type <= 0 ||
1711                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1712                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1713                         ticket->t_trans_type,
1714                         ticket->t_unit_res,
1715                         ticket->t_curr_res,
1716                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1717                         ticket->t_res_num_ophdrs, ophdr_spc,
1718                         ticket->t_res_arr_sum + 
1719                         ticket->t_res_o_flow + ophdr_spc,
1720                         ticket->t_res_num);
1721
1722         for (i = 0; i < ticket->t_res_num; i++) {
1723                 uint r_type = ticket->t_res_arr[i].r_type; 
1724                 cmn_err(CE_WARN,
1725                             "region[%u]: %s - %u bytes\n",
1726                             i, 
1727                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1728                             "bad-rtype" : res_type_str[r_type-1]),
1729                             ticket->t_res_arr[i].r_len);
1730         }
1731 }
1732
1733 /*
1734  * Write some region out to in-core log
1735  *
1736  * This will be called when writing externally provided regions or when
1737  * writing out a commit record for a given transaction.
1738  *
1739  * General algorithm:
1740  *      1. Find total length of this write.  This may include adding to the
1741  *              lengths passed in.
1742  *      2. Check whether we violate the tickets reservation.
1743  *      3. While writing to this iclog
1744  *          A. Reserve as much space in this iclog as can get
1745  *          B. If this is first write, save away start lsn
1746  *          C. While writing this region:
1747  *              1. If first write of transaction, write start record
1748  *              2. Write log operation header (header per region)
1749  *              3. Find out if we can fit entire region into this iclog
1750  *              4. Potentially, verify destination memcpy ptr
1751  *              5. Memcpy (partial) region
1752  *              6. If partial copy, release iclog; otherwise, continue
1753  *                      copying more regions into current iclog
1754  *      4. Mark want sync bit (in simulation mode)
1755  *      5. Release iclog for potential flush to on-disk log.
1756  *
1757  * ERRORS:
1758  * 1.   Panic if reservation is overrun.  This should never happen since
1759  *      reservation amounts are generated internal to the filesystem.
1760  * NOTES:
1761  * 1. Tickets are single threaded data structures.
1762  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1763  *      syncing routine.  When a single log_write region needs to span
1764  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1765  *      on all log operation writes which don't contain the end of the
1766  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1767  *      operation which contains the end of the continued log_write region.
1768  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1769  *      we don't really know exactly how much space will be used.  As a result,
1770  *      we don't update ic_offset until the end when we know exactly how many
1771  *      bytes have been written out.
1772  */
1773 STATIC int
1774 xlog_write(xfs_mount_t *        mp,
1775            xfs_log_iovec_t      reg[],
1776            int                  nentries,
1777            xfs_log_ticket_t     tic,
1778            xfs_lsn_t            *start_lsn,
1779            xlog_in_core_t       **commit_iclog,
1780            uint                 flags)
1781 {
1782     xlog_t           *log = mp->m_log;
1783     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1784     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1785     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1786     __psint_t        ptr;            /* copy address into data region */
1787     int              len;            /* # xlog_write() bytes 2 still copy */
1788     int              index;          /* region index currently copying */
1789     int              log_offset;     /* offset (from 0) into data region */
1790     int              start_rec_copy; /* # bytes to copy for start record */
1791     int              partial_copy;   /* did we split a region? */
1792     int              partial_copy_len;/* # bytes copied if split region */
1793     int              need_copy;      /* # bytes need to memcpy this region */
1794     int              copy_len;       /* # bytes actually memcpy'ing */
1795     int              copy_off;       /* # bytes from entry start */
1796     int              contwr;         /* continued write of in-core log? */
1797     int              error;
1798     int              record_cnt = 0, data_cnt = 0;
1799
1800     partial_copy_len = partial_copy = 0;
1801
1802     /* Calculate potential maximum space.  Each region gets its own
1803      * xlog_op_header_t and may need to be double word aligned.
1804      */
1805     len = 0;
1806     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1807         len += sizeof(xlog_op_header_t);
1808         ticket->t_res_num_ophdrs++;
1809     }
1810
1811     for (index = 0; index < nentries; index++) {
1812         len += sizeof(xlog_op_header_t);            /* each region gets >= 1 */
1813         ticket->t_res_num_ophdrs++;
1814         len += reg[index].i_len;
1815         xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1816     }
1817     contwr = *start_lsn = 0;
1818
1819     if (ticket->t_curr_res < len) {
1820         xlog_print_tic_res(mp, ticket);
1821 #ifdef DEBUG
1822         xlog_panic(
1823                 "xfs_log_write: reservation ran out. Need to up reservation");
1824 #else
1825         /* Customer configurable panic */
1826         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1827                 "xfs_log_write: reservation ran out. Need to up reservation");
1828         /* If we did not panic, shutdown the filesystem */
1829         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1830 #endif
1831     } else
1832         ticket->t_curr_res -= len;
1833
1834     for (index = 0; index < nentries; ) {
1835         if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1836                                                &contwr, &log_offset)))
1837                 return error;
1838
1839         ASSERT(log_offset <= iclog->ic_size - 1);
1840         ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1841
1842         /* start_lsn is the first lsn written to. That's all we need. */
1843         if (! *start_lsn)
1844             *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1845
1846         /* This loop writes out as many regions as can fit in the amount
1847          * of space which was allocated by xlog_state_get_iclog_space().
1848          */
1849         while (index < nentries) {
1850             ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1851             ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1852             start_rec_copy = 0;
1853
1854             /* If first write for transaction, insert start record.
1855              * We can't be trying to commit if we are inited.  We can't
1856              * have any "partial_copy" if we are inited.
1857              */
1858             if (ticket->t_flags & XLOG_TIC_INITED) {
1859                 logop_head              = (xlog_op_header_t *)ptr;
1860                 logop_head->oh_tid      = cpu_to_be32(ticket->t_tid);
1861                 logop_head->oh_clientid = ticket->t_clientid;
1862                 logop_head->oh_len      = 0;
1863                 logop_head->oh_flags    = XLOG_START_TRANS;
1864                 logop_head->oh_res2     = 0;
1865                 ticket->t_flags         &= ~XLOG_TIC_INITED;    /* clear bit */
1866                 record_cnt++;
1867
1868                 start_rec_copy = sizeof(xlog_op_header_t);
1869                 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1870             }
1871
1872             /* Copy log operation header directly into data section */
1873             logop_head                  = (xlog_op_header_t *)ptr;
1874             logop_head->oh_tid          = cpu_to_be32(ticket->t_tid);
1875             logop_head->oh_clientid     = ticket->t_clientid;
1876             logop_head->oh_res2         = 0;
1877
1878             /* header copied directly */
1879             xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1880
1881             /* are we copying a commit or unmount record? */
1882             logop_head->oh_flags = flags;
1883
1884             /*
1885              * We've seen logs corrupted with bad transaction client
1886              * ids.  This makes sure that XFS doesn't generate them on.
1887              * Turn this into an EIO and shut down the filesystem.
1888              */
1889             switch (logop_head->oh_clientid)  {
1890             case XFS_TRANSACTION:
1891             case XFS_VOLUME:
1892             case XFS_LOG:
1893                 break;
1894             default:
1895                 xfs_fs_cmn_err(CE_WARN, mp,
1896                     "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1897                     logop_head->oh_clientid, tic);
1898                 return XFS_ERROR(EIO);
1899             }
1900
1901             /* Partial write last time? => (partial_copy != 0)
1902              * need_copy is the amount we'd like to copy if everything could
1903              * fit in the current memcpy.
1904              */
1905             need_copy = reg[index].i_len - partial_copy_len;
1906
1907             copy_off = partial_copy_len;
1908             if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1909                 copy_len = need_copy;
1910                 logop_head->oh_len = cpu_to_be32(copy_len);
1911                 if (partial_copy)
1912                     logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1913                 partial_copy_len = partial_copy = 0;
1914             } else {                                        /* partial write */
1915                 copy_len = iclog->ic_size - log_offset;
1916                 logop_head->oh_len = cpu_to_be32(copy_len);
1917                 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1918                 if (partial_copy)
1919                         logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1920                 partial_copy_len += copy_len;
1921                 partial_copy++;
1922                 len += sizeof(xlog_op_header_t); /* from splitting of region */
1923                 /* account for new log op header */
1924                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1925                 ticket->t_res_num_ophdrs++;
1926             }
1927             xlog_verify_dest_ptr(log, ptr);
1928
1929             /* copy region */
1930             ASSERT(copy_len >= 0);
1931             memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1932             xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1933
1934             /* make copy_len total bytes copied, including headers */
1935             copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1936             record_cnt++;
1937             data_cnt += contwr ? copy_len : 0;
1938             if (partial_copy) {                 /* copied partial region */
1939                     /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1940                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1941                     record_cnt = data_cnt = 0;
1942                     if ((error = xlog_state_release_iclog(log, iclog)))
1943                             return error;
1944                     break;                      /* don't increment index */
1945             } else {                            /* copied entire region */
1946                 index++;
1947                 partial_copy_len = partial_copy = 0;
1948
1949                 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1950                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1951                     record_cnt = data_cnt = 0;
1952                     spin_lock(&log->l_icloglock);
1953                     xlog_state_want_sync(log, iclog);
1954                     spin_unlock(&log->l_icloglock);
1955                     if (commit_iclog) {
1956                         ASSERT(flags & XLOG_COMMIT_TRANS);
1957                         *commit_iclog = iclog;
1958                     } else if ((error = xlog_state_release_iclog(log, iclog)))
1959                            return error;
1960                     if (index == nentries)
1961                             return 0;           /* we are done */
1962                     else
1963                             break;
1964                 }
1965             } /* if (partial_copy) */
1966         } /* while (index < nentries) */
1967     } /* for (index = 0; index < nentries; ) */
1968     ASSERT(len == 0);
1969
1970     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1971     if (commit_iclog) {
1972         ASSERT(flags & XLOG_COMMIT_TRANS);
1973         *commit_iclog = iclog;
1974         return 0;
1975     }
1976     return xlog_state_release_iclog(log, iclog);
1977 }       /* xlog_write */
1978
1979
1980 /*****************************************************************************
1981  *
1982  *              State Machine functions
1983  *
1984  *****************************************************************************
1985  */
1986
1987 /* Clean iclogs starting from the head.  This ordering must be
1988  * maintained, so an iclog doesn't become ACTIVE beyond one that
1989  * is SYNCING.  This is also required to maintain the notion that we use
1990  * a ordered wait queue to hold off would be writers to the log when every
1991  * iclog is trying to sync to disk.
1992  *
1993  * State Change: DIRTY -> ACTIVE
1994  */
1995 STATIC void
1996 xlog_state_clean_log(xlog_t *log)
1997 {
1998         xlog_in_core_t  *iclog;
1999         int changed = 0;
2000
2001         iclog = log->l_iclog;
2002         do {
2003                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2004                         iclog->ic_state = XLOG_STATE_ACTIVE;
2005                         iclog->ic_offset       = 0;
2006                         ASSERT(iclog->ic_callback == NULL);
2007                         /*
2008                          * If the number of ops in this iclog indicate it just
2009                          * contains the dummy transaction, we can
2010                          * change state into IDLE (the second time around).
2011                          * Otherwise we should change the state into
2012                          * NEED a dummy.
2013                          * We don't need to cover the dummy.
2014                          */
2015                         if (!changed &&
2016                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2017                                         XLOG_COVER_OPS)) {
2018                                 changed = 1;
2019                         } else {
2020                                 /*
2021                                  * We have two dirty iclogs so start over
2022                                  * This could also be num of ops indicates
2023                                  * this is not the dummy going out.
2024                                  */
2025                                 changed = 2;
2026                         }
2027                         iclog->ic_header.h_num_logops = 0;
2028                         memset(iclog->ic_header.h_cycle_data, 0,
2029                               sizeof(iclog->ic_header.h_cycle_data));
2030                         iclog->ic_header.h_lsn = 0;
2031                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2032                         /* do nothing */;
2033                 else
2034                         break;  /* stop cleaning */
2035                 iclog = iclog->ic_next;
2036         } while (iclog != log->l_iclog);
2037
2038         /* log is locked when we are called */
2039         /*
2040          * Change state for the dummy log recording.
2041          * We usually go to NEED. But we go to NEED2 if the changed indicates
2042          * we are done writing the dummy record.
2043          * If we are done with the second dummy recored (DONE2), then
2044          * we go to IDLE.
2045          */
2046         if (changed) {
2047                 switch (log->l_covered_state) {
2048                 case XLOG_STATE_COVER_IDLE:
2049                 case XLOG_STATE_COVER_NEED:
2050                 case XLOG_STATE_COVER_NEED2:
2051                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2052                         break;
2053
2054                 case XLOG_STATE_COVER_DONE:
2055                         if (changed == 1)
2056                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2057                         else
2058                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2059                         break;
2060
2061                 case XLOG_STATE_COVER_DONE2:
2062                         if (changed == 1)
2063                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2064                         else
2065                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2066                         break;
2067
2068                 default:
2069                         ASSERT(0);
2070                 }
2071         }
2072 }       /* xlog_state_clean_log */
2073
2074 STATIC xfs_lsn_t
2075 xlog_get_lowest_lsn(
2076         xlog_t          *log)
2077 {
2078         xlog_in_core_t  *lsn_log;
2079         xfs_lsn_t       lowest_lsn, lsn;
2080
2081         lsn_log = log->l_iclog;
2082         lowest_lsn = 0;
2083         do {
2084             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2085                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2086                 if ((lsn && !lowest_lsn) ||
2087                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2088                         lowest_lsn = lsn;
2089                 }
2090             }
2091             lsn_log = lsn_log->ic_next;
2092         } while (lsn_log != log->l_iclog);
2093         return lowest_lsn;
2094 }
2095
2096
2097 STATIC void
2098 xlog_state_do_callback(
2099         xlog_t          *log,
2100         int             aborted,
2101         xlog_in_core_t  *ciclog)
2102 {
2103         xlog_in_core_t     *iclog;
2104         xlog_in_core_t     *first_iclog;        /* used to know when we've
2105                                                  * processed all iclogs once */
2106         xfs_log_callback_t *cb, *cb_next;
2107         int                flushcnt = 0;
2108         xfs_lsn_t          lowest_lsn;
2109         int                ioerrors;    /* counter: iclogs with errors */
2110         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2111         int                funcdidcallbacks; /* flag: function did callbacks */
2112         int                repeats;     /* for issuing console warnings if
2113                                          * looping too many times */
2114         int                wake = 0;
2115
2116         spin_lock(&log->l_icloglock);
2117         first_iclog = iclog = log->l_iclog;
2118         ioerrors = 0;
2119         funcdidcallbacks = 0;
2120         repeats = 0;
2121
2122         do {
2123                 /*
2124                  * Scan all iclogs starting with the one pointed to by the
2125                  * log.  Reset this starting point each time the log is
2126                  * unlocked (during callbacks).
2127                  *
2128                  * Keep looping through iclogs until one full pass is made
2129                  * without running any callbacks.
2130                  */
2131                 first_iclog = log->l_iclog;
2132                 iclog = log->l_iclog;
2133                 loopdidcallbacks = 0;
2134                 repeats++;
2135
2136                 do {
2137
2138                         /* skip all iclogs in the ACTIVE & DIRTY states */
2139                         if (iclog->ic_state &
2140                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2141                                 iclog = iclog->ic_next;
2142                                 continue;
2143                         }
2144
2145                         /*
2146                          * Between marking a filesystem SHUTDOWN and stopping
2147                          * the log, we do flush all iclogs to disk (if there
2148                          * wasn't a log I/O error). So, we do want things to
2149                          * go smoothly in case of just a SHUTDOWN  w/o a
2150                          * LOG_IO_ERROR.
2151                          */
2152                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2153                                 /*
2154                                  * Can only perform callbacks in order.  Since
2155                                  * this iclog is not in the DONE_SYNC/
2156                                  * DO_CALLBACK state, we skip the rest and
2157                                  * just try to clean up.  If we set our iclog
2158                                  * to DO_CALLBACK, we will not process it when
2159                                  * we retry since a previous iclog is in the
2160                                  * CALLBACK and the state cannot change since
2161                                  * we are holding the l_icloglock.
2162                                  */
2163                                 if (!(iclog->ic_state &
2164                                         (XLOG_STATE_DONE_SYNC |
2165                                                  XLOG_STATE_DO_CALLBACK))) {
2166                                         if (ciclog && (ciclog->ic_state ==
2167                                                         XLOG_STATE_DONE_SYNC)) {
2168                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2169                                         }
2170                                         break;
2171                                 }
2172                                 /*
2173                                  * We now have an iclog that is in either the
2174                                  * DO_CALLBACK or DONE_SYNC states. The other
2175                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2176                                  * caught by the above if and are going to
2177                                  * clean (i.e. we aren't doing their callbacks)
2178                                  * see the above if.
2179                                  */
2180
2181                                 /*
2182                                  * We will do one more check here to see if we
2183                                  * have chased our tail around.
2184                                  */
2185
2186                                 lowest_lsn = xlog_get_lowest_lsn(log);
2187                                 if (lowest_lsn &&
2188                                     XFS_LSN_CMP(lowest_lsn,
2189                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2190                                         iclog = iclog->ic_next;
2191                                         continue; /* Leave this iclog for
2192                                                    * another thread */
2193                                 }
2194
2195                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2196
2197                                 spin_unlock(&log->l_icloglock);
2198
2199                                 /* l_last_sync_lsn field protected by
2200                                  * l_grant_lock. Don't worry about iclog's lsn.
2201                                  * No one else can be here except us.
2202                                  */
2203                                 spin_lock(&log->l_grant_lock);
2204                                 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2205                                        be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2206                                 log->l_last_sync_lsn =
2207                                         be64_to_cpu(iclog->ic_header.h_lsn);
2208                                 spin_unlock(&log->l_grant_lock);
2209
2210                         } else {
2211                                 spin_unlock(&log->l_icloglock);
2212                                 ioerrors++;
2213                         }
2214
2215                         /*
2216                          * Keep processing entries in the callback list until
2217                          * we come around and it is empty.  We need to
2218                          * atomically see that the list is empty and change the
2219                          * state to DIRTY so that we don't miss any more
2220                          * callbacks being added.
2221                          */
2222                         spin_lock(&iclog->ic_callback_lock);
2223                         cb = iclog->ic_callback;
2224                         while (cb) {
2225                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2226                                 iclog->ic_callback = NULL;
2227                                 spin_unlock(&iclog->ic_callback_lock);
2228
2229                                 /* perform callbacks in the order given */
2230                                 for (; cb; cb = cb_next) {
2231                                         cb_next = cb->cb_next;
2232                                         cb->cb_func(cb->cb_arg, aborted);
2233                                 }
2234                                 spin_lock(&iclog->ic_callback_lock);
2235                                 cb = iclog->ic_callback;
2236                         }
2237
2238                         loopdidcallbacks++;
2239                         funcdidcallbacks++;
2240
2241                         spin_lock(&log->l_icloglock);
2242                         ASSERT(iclog->ic_callback == NULL);
2243                         spin_unlock(&iclog->ic_callback_lock);
2244                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2245                                 iclog->ic_state = XLOG_STATE_DIRTY;
2246
2247                         /*
2248                          * Transition from DIRTY to ACTIVE if applicable.
2249                          * NOP if STATE_IOERROR.
2250                          */
2251                         xlog_state_clean_log(log);
2252
2253                         /* wake up threads waiting in xfs_log_force() */
2254                         sv_broadcast(&iclog->ic_force_wait);
2255
2256                         iclog = iclog->ic_next;
2257                 } while (first_iclog != iclog);
2258
2259                 if (repeats > 5000) {
2260                         flushcnt += repeats;
2261                         repeats = 0;
2262                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2263                                 "%s: possible infinite loop (%d iterations)",
2264                                 __func__, flushcnt);
2265                 }
2266         } while (!ioerrors && loopdidcallbacks);
2267
2268         /*
2269          * make one last gasp attempt to see if iclogs are being left in
2270          * limbo..
2271          */
2272 #ifdef DEBUG
2273         if (funcdidcallbacks) {
2274                 first_iclog = iclog = log->l_iclog;
2275                 do {
2276                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2277                         /*
2278                          * Terminate the loop if iclogs are found in states
2279                          * which will cause other threads to clean up iclogs.
2280                          *
2281                          * SYNCING - i/o completion will go through logs
2282                          * DONE_SYNC - interrupt thread should be waiting for
2283                          *              l_icloglock
2284                          * IOERROR - give up hope all ye who enter here
2285                          */
2286                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2287                             iclog->ic_state == XLOG_STATE_SYNCING ||
2288                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2289                             iclog->ic_state == XLOG_STATE_IOERROR )
2290                                 break;
2291                         iclog = iclog->ic_next;
2292                 } while (first_iclog != iclog);
2293         }
2294 #endif
2295
2296         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2297                 wake = 1;
2298         spin_unlock(&log->l_icloglock);
2299
2300         if (wake)
2301                 sv_broadcast(&log->l_flush_wait);
2302 }
2303
2304
2305 /*
2306  * Finish transitioning this iclog to the dirty state.
2307  *
2308  * Make sure that we completely execute this routine only when this is
2309  * the last call to the iclog.  There is a good chance that iclog flushes,
2310  * when we reach the end of the physical log, get turned into 2 separate
2311  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2312  * routine.  By using the reference count bwritecnt, we guarantee that only
2313  * the second completion goes through.
2314  *
2315  * Callbacks could take time, so they are done outside the scope of the
2316  * global state machine log lock.
2317  */
2318 STATIC void
2319 xlog_state_done_syncing(
2320         xlog_in_core_t  *iclog,
2321         int             aborted)
2322 {
2323         xlog_t             *log = iclog->ic_log;
2324
2325         spin_lock(&log->l_icloglock);
2326
2327         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2328                iclog->ic_state == XLOG_STATE_IOERROR);
2329         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2330         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2331
2332
2333         /*
2334          * If we got an error, either on the first buffer, or in the case of
2335          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2336          * and none should ever be attempted to be written to disk
2337          * again.
2338          */
2339         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2340                 if (--iclog->ic_bwritecnt == 1) {
2341                         spin_unlock(&log->l_icloglock);
2342                         return;
2343                 }
2344                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2345         }
2346
2347         /*
2348          * Someone could be sleeping prior to writing out the next
2349          * iclog buffer, we wake them all, one will get to do the
2350          * I/O, the others get to wait for the result.
2351          */
2352         sv_broadcast(&iclog->ic_write_wait);
2353         spin_unlock(&log->l_icloglock);
2354         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2355 }       /* xlog_state_done_syncing */
2356
2357
2358 /*
2359  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2360  * sleep.  We wait on the flush queue on the head iclog as that should be
2361  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2362  * we will wait here and all new writes will sleep until a sync completes.
2363  *
2364  * The in-core logs are used in a circular fashion. They are not used
2365  * out-of-order even when an iclog past the head is free.
2366  *
2367  * return:
2368  *      * log_offset where xlog_write() can start writing into the in-core
2369  *              log's data space.
2370  *      * in-core log pointer to which xlog_write() should write.
2371  *      * boolean indicating this is a continued write to an in-core log.
2372  *              If this is the last write, then the in-core log's offset field
2373  *              needs to be incremented, depending on the amount of data which
2374  *              is copied.
2375  */
2376 STATIC int
2377 xlog_state_get_iclog_space(xlog_t         *log,
2378                            int            len,
2379                            xlog_in_core_t **iclogp,
2380                            xlog_ticket_t  *ticket,
2381                            int            *continued_write,
2382                            int            *logoffsetp)
2383 {
2384         int               log_offset;
2385         xlog_rec_header_t *head;
2386         xlog_in_core_t    *iclog;
2387         int               error;
2388
2389 restart:
2390         spin_lock(&log->l_icloglock);
2391         if (XLOG_FORCED_SHUTDOWN(log)) {
2392                 spin_unlock(&log->l_icloglock);
2393                 return XFS_ERROR(EIO);
2394         }
2395
2396         iclog = log->l_iclog;
2397         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2398                 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2399                 XFS_STATS_INC(xs_log_noiclogs);
2400
2401                 /* Wait for log writes to have flushed */
2402                 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
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 STATIC int
2487 xlog_grant_log_space(xlog_t        *log,
2488                      xlog_ticket_t *tic)
2489 {
2490         int              free_bytes;
2491         int              need_bytes;
2492 #ifdef DEBUG
2493         xfs_lsn_t        tail_lsn;
2494 #endif
2495
2496
2497 #ifdef DEBUG
2498         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2499                 panic("grant Recovery problem");
2500 #endif
2501
2502         /* Is there space or do we need to sleep? */
2503         spin_lock(&log->l_grant_lock);
2504         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2505
2506         /* something is already sleeping; insert new transaction at end */
2507         if (log->l_reserve_headq) {
2508                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2509                 xlog_trace_loggrant(log, tic,
2510                                     "xlog_grant_log_space: sleep 1");
2511                 /*
2512                  * Gotta check this before going to sleep, while we're
2513                  * holding the grant lock.
2514                  */
2515                 if (XLOG_FORCED_SHUTDOWN(log))
2516                         goto error_return;
2517
2518                 XFS_STATS_INC(xs_sleep_logspace);
2519                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2520                 /*
2521                  * If we got an error, and the filesystem is shutting down,
2522                  * we'll catch it down below. So just continue...
2523                  */
2524                 xlog_trace_loggrant(log, tic,
2525                                     "xlog_grant_log_space: wake 1");
2526                 spin_lock(&log->l_grant_lock);
2527         }
2528         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2529                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2530         else
2531                 need_bytes = tic->t_unit_res;
2532
2533 redo:
2534         if (XLOG_FORCED_SHUTDOWN(log))
2535                 goto error_return;
2536
2537         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2538                                      log->l_grant_reserve_bytes);
2539         if (free_bytes < need_bytes) {
2540                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2541                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2542                 xlog_trace_loggrant(log, tic,
2543                                     "xlog_grant_log_space: sleep 2");
2544                 XFS_STATS_INC(xs_sleep_logspace);
2545                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2546
2547                 if (XLOG_FORCED_SHUTDOWN(log)) {
2548                         spin_lock(&log->l_grant_lock);
2549                         goto error_return;
2550                 }
2551
2552                 xlog_trace_loggrant(log, tic,
2553                                     "xlog_grant_log_space: wake 2");
2554                 xlog_grant_push_ail(log->l_mp, need_bytes);
2555                 spin_lock(&log->l_grant_lock);
2556                 goto redo;
2557         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2558                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2559
2560         /* we've got enough space */
2561         xlog_grant_add_space(log, need_bytes);
2562 #ifdef DEBUG
2563         tail_lsn = log->l_tail_lsn;
2564         /*
2565          * Check to make sure the grant write head didn't just over lap the
2566          * tail.  If the cycles are the same, we can't be overlapping.
2567          * Otherwise, make sure that the cycles differ by exactly one and
2568          * check the byte count.
2569          */
2570         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2571                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2572                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2573         }
2574 #endif
2575         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2576         xlog_verify_grant_head(log, 1);
2577         spin_unlock(&log->l_grant_lock);
2578         return 0;
2579
2580  error_return:
2581         if (tic->t_flags & XLOG_TIC_IN_Q)
2582                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2583         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2584         /*
2585          * If we are failing, make sure the ticket doesn't have any
2586          * current reservations. We don't want to add this back when
2587          * the ticket/transaction gets cancelled.
2588          */
2589         tic->t_curr_res = 0;
2590         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2591         spin_unlock(&log->l_grant_lock);
2592         return XFS_ERROR(EIO);
2593 }       /* xlog_grant_log_space */
2594
2595
2596 /*
2597  * Replenish the byte reservation required by moving the grant write head.
2598  *
2599  *
2600  */
2601 STATIC int
2602 xlog_regrant_write_log_space(xlog_t        *log,
2603                              xlog_ticket_t *tic)
2604 {
2605         int             free_bytes, need_bytes;
2606         xlog_ticket_t   *ntic;
2607 #ifdef DEBUG
2608         xfs_lsn_t       tail_lsn;
2609 #endif
2610
2611         tic->t_curr_res = tic->t_unit_res;
2612         xlog_tic_reset_res(tic);
2613
2614         if (tic->t_cnt > 0)
2615                 return 0;
2616
2617 #ifdef DEBUG
2618         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2619                 panic("regrant Recovery problem");
2620 #endif
2621
2622         spin_lock(&log->l_grant_lock);
2623         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2624
2625         if (XLOG_FORCED_SHUTDOWN(log))
2626                 goto error_return;
2627
2628         /* If there are other waiters on the queue then give them a
2629          * chance at logspace before us. Wake up the first waiters,
2630          * if we do not wake up all the waiters then go to sleep waiting
2631          * for more free space, otherwise try to get some space for
2632          * this transaction.
2633          */
2634
2635         if ((ntic = log->l_write_headq)) {
2636                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2637                                              log->l_grant_write_bytes);
2638                 do {
2639                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2640
2641                         if (free_bytes < ntic->t_unit_res)
2642                                 break;
2643                         free_bytes -= ntic->t_unit_res;
2644                         sv_signal(&ntic->t_wait);
2645                         ntic = ntic->t_next;
2646                 } while (ntic != log->l_write_headq);
2647
2648                 if (ntic != log->l_write_headq) {
2649                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2650                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2651
2652                         xlog_trace_loggrant(log, tic,
2653                                     "xlog_regrant_write_log_space: sleep 1");
2654                         XFS_STATS_INC(xs_sleep_logspace);
2655                         sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2656                                 &log->l_grant_lock, s);
2657
2658                         /* If we're shutting down, this tic is already
2659                          * off the queue */
2660                         if (XLOG_FORCED_SHUTDOWN(log)) {
2661                                 spin_lock(&log->l_grant_lock);
2662                                 goto error_return;
2663                         }
2664
2665                         xlog_trace_loggrant(log, tic,
2666                                     "xlog_regrant_write_log_space: wake 1");
2667                         xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2668                         spin_lock(&log->l_grant_lock);
2669                 }
2670         }
2671
2672         need_bytes = tic->t_unit_res;
2673
2674 redo:
2675         if (XLOG_FORCED_SHUTDOWN(log))
2676                 goto error_return;
2677
2678         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2679                                      log->l_grant_write_bytes);
2680         if (free_bytes < need_bytes) {
2681                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2682                         xlog_ins_ticketq(&log->l_write_headq, tic);
2683                 XFS_STATS_INC(xs_sleep_logspace);
2684                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2685
2686                 /* If we're shutting down, this tic is already off the queue */
2687                 if (XLOG_FORCED_SHUTDOWN(log)) {
2688                         spin_lock(&log->l_grant_lock);
2689                         goto error_return;
2690                 }
2691
2692                 xlog_trace_loggrant(log, tic,
2693                                     "xlog_regrant_write_log_space: wake 2");
2694                 xlog_grant_push_ail(log->l_mp, need_bytes);
2695                 spin_lock(&log->l_grant_lock);
2696                 goto redo;
2697         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2698                 xlog_del_ticketq(&log->l_write_headq, tic);
2699
2700         /* we've got enough space */
2701         xlog_grant_add_space_write(log, need_bytes);
2702 #ifdef DEBUG
2703         tail_lsn = log->l_tail_lsn;
2704         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2705                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2706                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2707         }
2708 #endif
2709
2710         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2711         xlog_verify_grant_head(log, 1);
2712         spin_unlock(&log->l_grant_lock);
2713         return 0;
2714
2715
2716  error_return:
2717         if (tic->t_flags & XLOG_TIC_IN_Q)
2718                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2719         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2720         /*
2721          * If we are failing, make sure the ticket doesn't have any
2722          * current reservations. We don't want to add this back when
2723          * the ticket/transaction gets cancelled.
2724          */
2725         tic->t_curr_res = 0;
2726         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2727         spin_unlock(&log->l_grant_lock);
2728         return XFS_ERROR(EIO);
2729 }       /* xlog_regrant_write_log_space */
2730
2731
2732 /* The first cnt-1 times through here we don't need to
2733  * move the grant write head because the permanent
2734  * reservation has reserved cnt times the unit amount.
2735  * Release part of current permanent unit reservation and
2736  * reset current reservation to be one units worth.  Also
2737  * move grant reservation head forward.
2738  */
2739 STATIC void
2740 xlog_regrant_reserve_log_space(xlog_t        *log,
2741                                xlog_ticket_t *ticket)
2742 {
2743         xlog_trace_loggrant(log, ticket,
2744                             "xlog_regrant_reserve_log_space: enter");
2745         if (ticket->t_cnt > 0)
2746                 ticket->t_cnt--;
2747
2748         spin_lock(&log->l_grant_lock);
2749         xlog_grant_sub_space(log, ticket->t_curr_res);
2750         ticket->t_curr_res = ticket->t_unit_res;
2751         xlog_tic_reset_res(ticket);
2752         xlog_trace_loggrant(log, ticket,
2753                             "xlog_regrant_reserve_log_space: sub current res");
2754         xlog_verify_grant_head(log, 1);
2755
2756         /* just return if we still have some of the pre-reserved space */
2757         if (ticket->t_cnt > 0) {
2758                 spin_unlock(&log->l_grant_lock);
2759                 return;
2760         }
2761
2762         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2763         xlog_trace_loggrant(log, ticket,
2764                             "xlog_regrant_reserve_log_space: exit");
2765         xlog_verify_grant_head(log, 0);
2766         spin_unlock(&log->l_grant_lock);
2767         ticket->t_curr_res = ticket->t_unit_res;
2768         xlog_tic_reset_res(ticket);
2769 }       /* xlog_regrant_reserve_log_space */
2770
2771
2772 /*
2773  * Give back the space left from a reservation.
2774  *
2775  * All the information we need to make a correct determination of space left
2776  * is present.  For non-permanent reservations, things are quite easy.  The
2777  * count should have been decremented to zero.  We only need to deal with the
2778  * space remaining in the current reservation part of the ticket.  If the
2779  * ticket contains a permanent reservation, there may be left over space which
2780  * needs to be released.  A count of N means that N-1 refills of the current
2781  * reservation can be done before we need to ask for more space.  The first
2782  * one goes to fill up the first current reservation.  Once we run out of
2783  * space, the count will stay at zero and the only space remaining will be
2784  * in the current reservation field.
2785  */
2786 STATIC void
2787 xlog_ungrant_log_space(xlog_t        *log,
2788                        xlog_ticket_t *ticket)
2789 {
2790         if (ticket->t_cnt > 0)
2791                 ticket->t_cnt--;
2792
2793         spin_lock(&log->l_grant_lock);
2794         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2795
2796         xlog_grant_sub_space(log, ticket->t_curr_res);
2797
2798         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2799
2800         /* If this is a permanent reservation ticket, we may be able to free
2801          * up more space based on the remaining count.
2802          */
2803         if (ticket->t_cnt > 0) {
2804                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2805                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2806         }
2807
2808         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2809         xlog_verify_grant_head(log, 1);
2810         spin_unlock(&log->l_grant_lock);
2811         xfs_log_move_tail(log->l_mp, 1);
2812 }       /* xlog_ungrant_log_space */
2813
2814
2815 /*
2816  * Flush iclog to disk if this is the last reference to the given iclog and
2817  * the WANT_SYNC bit is set.
2818  *
2819  * When this function is entered, the iclog is not necessarily in the
2820  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2821  *
2822  *
2823  */
2824 STATIC int
2825 xlog_state_release_iclog(
2826         xlog_t          *log,
2827         xlog_in_core_t  *iclog)
2828 {
2829         int             sync = 0;       /* do we sync? */
2830
2831         if (iclog->ic_state & XLOG_STATE_IOERROR)
2832                 return XFS_ERROR(EIO);
2833
2834         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2835         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2836                 return 0;
2837
2838         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2839                 spin_unlock(&log->l_icloglock);
2840                 return XFS_ERROR(EIO);
2841         }
2842         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2843                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2844
2845         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2846                 /* update tail before writing to iclog */
2847                 xlog_assign_tail_lsn(log->l_mp);
2848                 sync++;
2849                 iclog->ic_state = XLOG_STATE_SYNCING;
2850                 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2851                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2852                 /* cycle incremented when incrementing curr_block */
2853         }
2854         spin_unlock(&log->l_icloglock);
2855
2856         /*
2857          * We let the log lock go, so it's possible that we hit a log I/O
2858          * error or some other SHUTDOWN condition that marks the iclog
2859          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2860          * this iclog has consistent data, so we ignore IOERROR
2861          * flags after this point.
2862          */
2863         if (sync)
2864                 return xlog_sync(log, iclog);
2865         return 0;
2866 }       /* xlog_state_release_iclog */
2867
2868
2869 /*
2870  * This routine will mark the current iclog in the ring as WANT_SYNC
2871  * and move the current iclog pointer to the next iclog in the ring.
2872  * When this routine is called from xlog_state_get_iclog_space(), the
2873  * exact size of the iclog has not yet been determined.  All we know is
2874  * that every data block.  We have run out of space in this log record.
2875  */
2876 STATIC void
2877 xlog_state_switch_iclogs(xlog_t         *log,
2878                          xlog_in_core_t *iclog,
2879                          int            eventual_size)
2880 {
2881         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2882         if (!eventual_size)
2883                 eventual_size = iclog->ic_offset;
2884         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2885         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2886         log->l_prev_block = log->l_curr_block;
2887         log->l_prev_cycle = log->l_curr_cycle;
2888
2889         /* roll log?: ic_offset changed later */
2890         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2891
2892         /* Round up to next log-sunit */
2893         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2894             log->l_mp->m_sb.sb_logsunit > 1) {
2895                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2896                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2897         }
2898
2899         if (log->l_curr_block >= log->l_logBBsize) {
2900                 log->l_curr_cycle++;
2901                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2902                         log->l_curr_cycle++;
2903                 log->l_curr_block -= log->l_logBBsize;
2904                 ASSERT(log->l_curr_block >= 0);
2905         }
2906         ASSERT(iclog == log->l_iclog);
2907         log->l_iclog = iclog->ic_next;
2908 }       /* xlog_state_switch_iclogs */
2909
2910
2911 /*
2912  * Write out all data in the in-core log as of this exact moment in time.
2913  *
2914  * Data may be written to the in-core log during this call.  However,
2915  * we don't guarantee this data will be written out.  A change from past
2916  * implementation means this routine will *not* write out zero length LRs.
2917  *
2918  * Basically, we try and perform an intelligent scan of the in-core logs.
2919  * If we determine there is no flushable data, we just return.  There is no
2920  * flushable data if:
2921  *
2922  *      1. the current iclog is active and has no data; the previous iclog
2923  *              is in the active or dirty state.
2924  *      2. the current iclog is drity, and the previous iclog is in the
2925  *              active or dirty state.
2926  *
2927  * We may sleep if:
2928  *
2929  *      1. the current iclog is not in the active nor dirty state.
2930  *      2. the current iclog dirty, and the previous iclog is not in the
2931  *              active nor dirty state.
2932  *      3. the current iclog is active, and there is another thread writing
2933  *              to this particular iclog.
2934  *      4. a) the current iclog is active and has no other writers
2935  *         b) when we return from flushing out this iclog, it is still
2936  *              not in the active nor dirty state.
2937  */
2938 STATIC int
2939 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2940 {
2941         xlog_in_core_t  *iclog;
2942         xfs_lsn_t       lsn;
2943
2944         spin_lock(&log->l_icloglock);
2945
2946         iclog = log->l_iclog;
2947         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2948                 spin_unlock(&log->l_icloglock);
2949                 return XFS_ERROR(EIO);
2950         }
2951
2952         /* If the head iclog is not active nor dirty, we just attach
2953          * ourselves to the head and go to sleep.
2954          */
2955         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2956             iclog->ic_state == XLOG_STATE_DIRTY) {
2957                 /*
2958                  * If the head is dirty or (active and empty), then
2959                  * we need to look at the previous iclog.  If the previous
2960                  * iclog is active or dirty we are done.  There is nothing
2961                  * to sync out.  Otherwise, we attach ourselves to the
2962                  * previous iclog and go to sleep.
2963                  */
2964                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2965                     (atomic_read(&iclog->ic_refcnt) == 0
2966                      && iclog->ic_offset == 0)) {
2967                         iclog = iclog->ic_prev;
2968                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2969                             iclog->ic_state == XLOG_STATE_DIRTY)
2970                                 goto no_sleep;
2971                         else
2972                                 goto maybe_sleep;
2973                 } else {
2974                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2975                                 /* We are the only one with access to this
2976                                  * iclog.  Flush it out now.  There should
2977                                  * be a roundoff of zero to show that someone
2978                                  * has already taken care of the roundoff from
2979                                  * the previous sync.
2980                                  */
2981                                 atomic_inc(&iclog->ic_refcnt);
2982                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2983                                 xlog_state_switch_iclogs(log, iclog, 0);
2984                                 spin_unlock(&log->l_icloglock);
2985
2986                                 if (xlog_state_release_iclog(log, iclog))
2987                                         return XFS_ERROR(EIO);
2988                                 *log_flushed = 1;
2989                                 spin_lock(&log->l_icloglock);
2990                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2991                                     iclog->ic_state != XLOG_STATE_DIRTY)
2992                                         goto maybe_sleep;
2993                                 else
2994                                         goto no_sleep;
2995                         } else {
2996                                 /* Someone else is writing to this iclog.
2997                                  * Use its call to flush out the data.  However,
2998                                  * the other thread may not force out this LR,
2999                                  * so we mark it WANT_SYNC.
3000                                  */
3001                                 xlog_state_switch_iclogs(log, iclog, 0);
3002                                 goto maybe_sleep;
3003                         }
3004                 }
3005         }
3006
3007         /* By the time we come around again, the iclog could've been filled
3008          * which would give it another lsn.  If we have a new lsn, just
3009          * return because the relevant data has been flushed.
3010          */
3011 maybe_sleep:
3012         if (flags & XFS_LOG_SYNC) {
3013                 /*
3014                  * We must check if we're shutting down here, before
3015                  * we wait, while we're holding the l_icloglock.
3016                  * Then we check again after waking up, in case our
3017                  * sleep was disturbed by a bad news.
3018                  */
3019                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3020                         spin_unlock(&log->l_icloglock);
3021                         return XFS_ERROR(EIO);
3022                 }
3023                 XFS_STATS_INC(xs_log_force_sleep);
3024                 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3025                 /*
3026                  * No need to grab the log lock here since we're
3027                  * only deciding whether or not to return EIO
3028                  * and the memory read should be atomic.
3029                  */
3030                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3031                         return XFS_ERROR(EIO);
3032                 *log_flushed = 1;
3033
3034         } else {
3035
3036 no_sleep:
3037                 spin_unlock(&log->l_icloglock);
3038         }
3039         return 0;
3040 }       /* xlog_state_sync_all */
3041
3042
3043 /*
3044  * Used by code which implements synchronous log forces.
3045  *
3046  * Find in-core log with lsn.
3047  *      If it is in the DIRTY state, just return.
3048  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3049  *              state and go to sleep or return.
3050  *      If it is in any other state, go to sleep or return.
3051  *
3052  * If filesystem activity goes to zero, the iclog will get flushed only by
3053  * bdflush().
3054  */
3055 STATIC int
3056 xlog_state_sync(xlog_t    *log,
3057                 xfs_lsn_t lsn,
3058                 uint      flags,
3059                 int       *log_flushed)
3060 {
3061     xlog_in_core_t      *iclog;
3062     int                 already_slept = 0;
3063
3064 try_again:
3065     spin_lock(&log->l_icloglock);
3066     iclog = log->l_iclog;
3067
3068     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3069             spin_unlock(&log->l_icloglock);
3070             return XFS_ERROR(EIO);
3071     }
3072
3073     do {
3074         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3075                 iclog = iclog->ic_next;
3076                 continue;
3077         }
3078
3079         if (iclog->ic_state == XLOG_STATE_DIRTY) {
3080                 spin_unlock(&log->l_icloglock);
3081                 return 0;
3082         }
3083
3084         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3085                 /*
3086                  * We sleep here if we haven't already slept (e.g.
3087                  * this is the first time we've looked at the correct
3088                  * iclog buf) and the buffer before us is going to
3089                  * be sync'ed. The reason for this is that if we
3090                  * are doing sync transactions here, by waiting for
3091                  * the previous I/O to complete, we can allow a few
3092                  * more transactions into this iclog before we close