Merge branch 'for-linus' of git://oss.sgi.com/xfs/xfs
[sfrench/cifs-2.6.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 (IS_ERR(mp->m_log)) {
566                 error = -PTR_ERR(mp->m_log);
567                 goto out;
568         }
569
570         /*
571          * Initialize the AIL now we have a log.
572          */
573         error = xfs_trans_ail_init(mp);
574         if (error) {
575                 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
576                 goto out_free_log;
577         }
578         mp->m_log->l_ailp = mp->m_ail;
579
580         /*
581          * skip log recovery on a norecovery mount.  pretend it all
582          * just worked.
583          */
584         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
585                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
586
587                 if (readonly)
588                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
589
590                 error = xlog_recover(mp->m_log);
591
592                 if (readonly)
593                         mp->m_flags |= XFS_MOUNT_RDONLY;
594                 if (error) {
595                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
596                         goto out_destroy_ail;
597                 }
598         }
599
600         /* Normal transactions can now occur */
601         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
602
603         return 0;
604
605 out_destroy_ail:
606         xfs_trans_ail_destroy(mp);
607 out_free_log:
608         xlog_dealloc_log(mp->m_log);
609 out:
610         return error;
611 }
612
613 /*
614  * Finish the recovery of the file system.  This is separate from
615  * the xfs_log_mount() call, because it depends on the code in
616  * xfs_mountfs() to read in the root and real-time bitmap inodes
617  * between calling xfs_log_mount() and here.
618  *
619  * mp           - ubiquitous xfs mount point structure
620  */
621 int
622 xfs_log_mount_finish(xfs_mount_t *mp)
623 {
624         int     error;
625
626         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
627                 error = xlog_recover_finish(mp->m_log);
628         else {
629                 error = 0;
630                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
631         }
632
633         return error;
634 }
635
636 /*
637  * Final log writes as part of unmount.
638  *
639  * Mark the filesystem clean as unmount happens.  Note that during relocation
640  * this routine needs to be executed as part of source-bag while the
641  * deallocation must not be done until source-end.
642  */
643
644 /*
645  * Unmount record used to have a string "Unmount filesystem--" in the
646  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
647  * We just write the magic number now since that particular field isn't
648  * currently architecture converted and "nUmount" is a bit foo.
649  * As far as I know, there weren't any dependencies on the old behaviour.
650  */
651
652 int
653 xfs_log_unmount_write(xfs_mount_t *mp)
654 {
655         xlog_t           *log = mp->m_log;
656         xlog_in_core_t   *iclog;
657 #ifdef DEBUG
658         xlog_in_core_t   *first_iclog;
659 #endif
660         xfs_log_iovec_t  reg[1];
661         xfs_log_ticket_t tic = NULL;
662         xfs_lsn_t        lsn;
663         int              error;
664
665         /* the data section must be 32 bit size aligned */
666         struct {
667             __uint16_t magic;
668             __uint16_t pad1;
669             __uint32_t pad2; /* may as well make it 64 bits */
670         } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
671
672         /*
673          * Don't write out unmount record on read-only mounts.
674          * Or, if we are doing a forced umount (typically because of IO errors).
675          */
676         if (mp->m_flags & XFS_MOUNT_RDONLY)
677                 return 0;
678
679         error = _xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC, NULL);
680         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
681
682 #ifdef DEBUG
683         first_iclog = iclog = log->l_iclog;
684         do {
685                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
686                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
687                         ASSERT(iclog->ic_offset == 0);
688                 }
689                 iclog = iclog->ic_next;
690         } while (iclog != first_iclog);
691 #endif
692         if (! (XLOG_FORCED_SHUTDOWN(log))) {
693                 reg[0].i_addr = (void*)&magic;
694                 reg[0].i_len  = sizeof(magic);
695                 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
696
697                 error = xfs_log_reserve(mp, 600, 1, &tic,
698                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
699                 if (!error) {
700                         /* remove inited flag */
701                         ((xlog_ticket_t *)tic)->t_flags = 0;
702                         error = xlog_write(mp, reg, 1, tic, &lsn,
703                                            NULL, XLOG_UNMOUNT_TRANS);
704                         /*
705                          * At this point, we're umounting anyway,
706                          * so there's no point in transitioning log state
707                          * to IOERROR. Just continue...
708                          */
709                 }
710
711                 if (error) {
712                         xfs_fs_cmn_err(CE_ALERT, mp,
713                                 "xfs_log_unmount: unmount record failed");
714                 }
715
716
717                 spin_lock(&log->l_icloglock);
718                 iclog = log->l_iclog;
719                 atomic_inc(&iclog->ic_refcnt);
720                 xlog_state_want_sync(log, iclog);
721                 spin_unlock(&log->l_icloglock);
722                 error = xlog_state_release_iclog(log, iclog);
723
724                 spin_lock(&log->l_icloglock);
725                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
726                       iclog->ic_state == XLOG_STATE_DIRTY)) {
727                         if (!XLOG_FORCED_SHUTDOWN(log)) {
728                                 sv_wait(&iclog->ic_force_wait, PMEM,
729                                         &log->l_icloglock, s);
730                         } else {
731                                 spin_unlock(&log->l_icloglock);
732                         }
733                 } else {
734                         spin_unlock(&log->l_icloglock);
735                 }
736                 if (tic) {
737                         xlog_trace_loggrant(log, tic, "unmount rec");
738                         xlog_ungrant_log_space(log, tic);
739                         xfs_log_ticket_put(tic);
740                 }
741         } else {
742                 /*
743                  * We're already in forced_shutdown mode, couldn't
744                  * even attempt to write out the unmount transaction.
745                  *
746                  * Go through the motions of sync'ing and releasing
747                  * the iclog, even though no I/O will actually happen,
748                  * we need to wait for other log I/Os that may already
749                  * be in progress.  Do this as a separate section of
750                  * code so we'll know if we ever get stuck here that
751                  * we're in this odd situation of trying to unmount
752                  * a file system that went into forced_shutdown as
753                  * the result of an unmount..
754                  */
755                 spin_lock(&log->l_icloglock);
756                 iclog = log->l_iclog;
757                 atomic_inc(&iclog->ic_refcnt);
758
759                 xlog_state_want_sync(log, iclog);
760                 spin_unlock(&log->l_icloglock);
761                 error =  xlog_state_release_iclog(log, iclog);
762
763                 spin_lock(&log->l_icloglock);
764
765                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
766                         || iclog->ic_state == XLOG_STATE_DIRTY
767                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
768
769                                 sv_wait(&iclog->ic_force_wait, PMEM,
770                                         &log->l_icloglock, s);
771                 } else {
772                         spin_unlock(&log->l_icloglock);
773                 }
774         }
775
776         return error;
777 }       /* xfs_log_unmount_write */
778
779 /*
780  * Deallocate log structures for unmount/relocation.
781  *
782  * We need to stop the aild from running before we destroy
783  * and deallocate the log as the aild references the log.
784  */
785 void
786 xfs_log_unmount(xfs_mount_t *mp)
787 {
788         xfs_trans_ail_destroy(mp);
789         xlog_dealloc_log(mp->m_log);
790 }
791
792 /*
793  * Write region vectors to log.  The write happens using the space reservation
794  * of the ticket (tic).  It is not a requirement that all writes for a given
795  * transaction occur with one call to xfs_log_write().
796  */
797 int
798 xfs_log_write(xfs_mount_t *     mp,
799               xfs_log_iovec_t   reg[],
800               int               nentries,
801               xfs_log_ticket_t  tic,
802               xfs_lsn_t         *start_lsn)
803 {
804         int     error;
805         xlog_t *log = mp->m_log;
806
807         if (XLOG_FORCED_SHUTDOWN(log))
808                 return XFS_ERROR(EIO);
809
810         if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
811                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
812         }
813         return error;
814 }       /* xfs_log_write */
815
816
817 void
818 xfs_log_move_tail(xfs_mount_t   *mp,
819                   xfs_lsn_t     tail_lsn)
820 {
821         xlog_ticket_t   *tic;
822         xlog_t          *log = mp->m_log;
823         int             need_bytes, free_bytes, cycle, bytes;
824
825         if (XLOG_FORCED_SHUTDOWN(log))
826                 return;
827
828         if (tail_lsn == 0) {
829                 /* needed since sync_lsn is 64 bits */
830                 spin_lock(&log->l_icloglock);
831                 tail_lsn = log->l_last_sync_lsn;
832                 spin_unlock(&log->l_icloglock);
833         }
834
835         spin_lock(&log->l_grant_lock);
836
837         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
838          * tail_lsn.
839          */
840         if (tail_lsn != 1) {
841                 log->l_tail_lsn = tail_lsn;
842         }
843
844         if ((tic = log->l_write_headq)) {
845 #ifdef DEBUG
846                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
847                         panic("Recovery problem");
848 #endif
849                 cycle = log->l_grant_write_cycle;
850                 bytes = log->l_grant_write_bytes;
851                 free_bytes = xlog_space_left(log, cycle, bytes);
852                 do {
853                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
854
855                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
856                                 break;
857                         tail_lsn = 0;
858                         free_bytes -= tic->t_unit_res;
859                         sv_signal(&tic->t_wait);
860                         tic = tic->t_next;
861                 } while (tic != log->l_write_headq);
862         }
863         if ((tic = log->l_reserve_headq)) {
864 #ifdef DEBUG
865                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
866                         panic("Recovery problem");
867 #endif
868                 cycle = log->l_grant_reserve_cycle;
869                 bytes = log->l_grant_reserve_bytes;
870                 free_bytes = xlog_space_left(log, cycle, bytes);
871                 do {
872                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
873                                 need_bytes = tic->t_unit_res*tic->t_cnt;
874                         else
875                                 need_bytes = tic->t_unit_res;
876                         if (free_bytes < need_bytes && tail_lsn != 1)
877                                 break;
878                         tail_lsn = 0;
879                         free_bytes -= need_bytes;
880                         sv_signal(&tic->t_wait);
881                         tic = tic->t_next;
882                 } while (tic != log->l_reserve_headq);
883         }
884         spin_unlock(&log->l_grant_lock);
885 }       /* xfs_log_move_tail */
886
887 /*
888  * Determine if we have a transaction that has gone to disk
889  * that needs to be covered. Log activity needs to be idle (no AIL and
890  * nothing in the iclogs). And, we need to be in the right state indicating
891  * something has gone out.
892  */
893 int
894 xfs_log_need_covered(xfs_mount_t *mp)
895 {
896         int             needed = 0;
897         xlog_t          *log = mp->m_log;
898
899         if (!xfs_fs_writable(mp))
900                 return 0;
901
902         spin_lock(&log->l_icloglock);
903         if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
904                 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
905                         && !xfs_trans_ail_tail(log->l_ailp)
906                         && xlog_iclogs_empty(log)) {
907                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
908                         log->l_covered_state = XLOG_STATE_COVER_DONE;
909                 else {
910                         ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
911                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
912                 }
913                 needed = 1;
914         }
915         spin_unlock(&log->l_icloglock);
916         return needed;
917 }
918
919 /******************************************************************************
920  *
921  *      local routines
922  *
923  ******************************************************************************
924  */
925
926 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
927  * The log manager must keep track of the last LR which was committed
928  * to disk.  The lsn of this LR will become the new tail_lsn whenever
929  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
930  * the situation where stuff could be written into the log but nothing
931  * was ever in the AIL when asked.  Eventually, we panic since the
932  * tail hits the head.
933  *
934  * We may be holding the log iclog lock upon entering this routine.
935  */
936 xfs_lsn_t
937 xlog_assign_tail_lsn(xfs_mount_t *mp)
938 {
939         xfs_lsn_t tail_lsn;
940         xlog_t    *log = mp->m_log;
941
942         tail_lsn = xfs_trans_ail_tail(mp->m_ail);
943         spin_lock(&log->l_grant_lock);
944         if (tail_lsn != 0) {
945                 log->l_tail_lsn = tail_lsn;
946         } else {
947                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
948         }
949         spin_unlock(&log->l_grant_lock);
950
951         return tail_lsn;
952 }       /* xlog_assign_tail_lsn */
953
954
955 /*
956  * Return the space in the log between the tail and the head.  The head
957  * is passed in the cycle/bytes formal parms.  In the special case where
958  * the reserve head has wrapped passed the tail, this calculation is no
959  * longer valid.  In this case, just return 0 which means there is no space
960  * in the log.  This works for all places where this function is called
961  * with the reserve head.  Of course, if the write head were to ever
962  * wrap the tail, we should blow up.  Rather than catch this case here,
963  * we depend on other ASSERTions in other parts of the code.   XXXmiken
964  *
965  * This code also handles the case where the reservation head is behind
966  * the tail.  The details of this case are described below, but the end
967  * result is that we return the size of the log as the amount of space left.
968  */
969 STATIC int
970 xlog_space_left(xlog_t *log, int cycle, int bytes)
971 {
972         int free_bytes;
973         int tail_bytes;
974         int tail_cycle;
975
976         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
977         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
978         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
979                 free_bytes = log->l_logsize - (bytes - tail_bytes);
980         } else if ((tail_cycle + 1) < cycle) {
981                 return 0;
982         } else if (tail_cycle < cycle) {
983                 ASSERT(tail_cycle == (cycle - 1));
984                 free_bytes = tail_bytes - bytes;
985         } else {
986                 /*
987                  * The reservation head is behind the tail.
988                  * In this case we just want to return the size of the
989                  * log as the amount of space left.
990                  */
991                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
992                         "xlog_space_left: head behind tail\n"
993                         "  tail_cycle = %d, tail_bytes = %d\n"
994                         "  GH   cycle = %d, GH   bytes = %d",
995                         tail_cycle, tail_bytes, cycle, bytes);
996                 ASSERT(0);
997                 free_bytes = log->l_logsize;
998         }
999         return free_bytes;
1000 }       /* xlog_space_left */
1001
1002
1003 /*
1004  * Log function which is called when an io completes.
1005  *
1006  * The log manager needs its own routine, in order to control what
1007  * happens with the buffer after the write completes.
1008  */
1009 void
1010 xlog_iodone(xfs_buf_t *bp)
1011 {
1012         xlog_in_core_t  *iclog;
1013         xlog_t          *l;
1014         int             aborted;
1015
1016         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1017         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
1018         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1019         aborted = 0;
1020         l = iclog->ic_log;
1021
1022         /*
1023          * If the _XFS_BARRIER_FAILED flag was set by a lower
1024          * layer, it means the underlying device no longer supports
1025          * barrier I/O. Warn loudly and turn off barriers.
1026          */
1027         if (bp->b_flags & _XFS_BARRIER_FAILED) {
1028                 bp->b_flags &= ~_XFS_BARRIER_FAILED;
1029                 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
1030                 xfs_fs_cmn_err(CE_WARN, l->l_mp,
1031                                 "xlog_iodone: Barriers are no longer supported"
1032                                 " by device. Disabling barriers\n");
1033                 xfs_buftrace("XLOG_IODONE BARRIERS OFF", bp);
1034         }
1035
1036         /*
1037          * Race to shutdown the filesystem if we see an error.
1038          */
1039         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
1040                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1041                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
1042                 XFS_BUF_STALE(bp);
1043                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1044                 /*
1045                  * This flag will be propagated to the trans-committed
1046                  * callback routines to let them know that the log-commit
1047                  * didn't succeed.
1048                  */
1049                 aborted = XFS_LI_ABORTED;
1050         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1051                 aborted = XFS_LI_ABORTED;
1052         }
1053
1054         /* log I/O is always issued ASYNC */
1055         ASSERT(XFS_BUF_ISASYNC(bp));
1056         xlog_state_done_syncing(iclog, aborted);
1057         /*
1058          * do not reference the buffer (bp) here as we could race
1059          * with it being freed after writing the unmount record to the
1060          * log.
1061          */
1062
1063 }       /* xlog_iodone */
1064
1065 /*
1066  * The bdstrat callback function for log bufs. This gives us a central
1067  * place to trap bufs in case we get hit by a log I/O error and need to
1068  * shutdown. Actually, in practice, even when we didn't get a log error,
1069  * we transition the iclogs to IOERROR state *after* flushing all existing
1070  * iclogs to disk. This is because we don't want anymore new transactions to be
1071  * started or completed afterwards.
1072  */
1073 STATIC int
1074 xlog_bdstrat_cb(struct xfs_buf *bp)
1075 {
1076         xlog_in_core_t *iclog;
1077
1078         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1079
1080         if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
1081           /* note for irix bstrat will need  struct bdevsw passed
1082            * Fix the following macro if the code ever is merged
1083            */
1084             XFS_bdstrat(bp);
1085                 return 0;
1086         }
1087
1088         xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1089         XFS_BUF_ERROR(bp, EIO);
1090         XFS_BUF_STALE(bp);
1091         xfs_biodone(bp);
1092         return XFS_ERROR(EIO);
1093
1094
1095 }
1096
1097 /*
1098  * Return size of each in-core log record buffer.
1099  *
1100  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1101  *
1102  * If the filesystem blocksize is too large, we may need to choose a
1103  * larger size since the directory code currently logs entire blocks.
1104  */
1105
1106 STATIC void
1107 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1108                            xlog_t       *log)
1109 {
1110         int size;
1111         int xhdrs;
1112
1113         if (mp->m_logbufs <= 0)
1114                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1115         else
1116                 log->l_iclog_bufs = mp->m_logbufs;
1117
1118         /*
1119          * Buffer size passed in from mount system call.
1120          */
1121         if (mp->m_logbsize > 0) {
1122                 size = log->l_iclog_size = mp->m_logbsize;
1123                 log->l_iclog_size_log = 0;
1124                 while (size != 1) {
1125                         log->l_iclog_size_log++;
1126                         size >>= 1;
1127                 }
1128
1129                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1130                         /* # headers = size / 32k
1131                          * one header holds cycles from 32k of data
1132                          */
1133
1134                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1135                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1136                                 xhdrs++;
1137                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1138                         log->l_iclog_heads = xhdrs;
1139                 } else {
1140                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1141                         log->l_iclog_hsize = BBSIZE;
1142                         log->l_iclog_heads = 1;
1143                 }
1144                 goto done;
1145         }
1146
1147         /* All machines use 32kB buffers by default. */
1148         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1149         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1150
1151         /* the default log size is 16k or 32k which is one header sector */
1152         log->l_iclog_hsize = BBSIZE;
1153         log->l_iclog_heads = 1;
1154
1155 done:
1156         /* are we being asked to make the sizes selected above visible? */
1157         if (mp->m_logbufs == 0)
1158                 mp->m_logbufs = log->l_iclog_bufs;
1159         if (mp->m_logbsize == 0)
1160                 mp->m_logbsize = log->l_iclog_size;
1161 }       /* xlog_get_iclog_buffer_size */
1162
1163
1164 /*
1165  * This routine initializes some of the log structure for a given mount point.
1166  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1167  * some other stuff may be filled in too.
1168  */
1169 STATIC xlog_t *
1170 xlog_alloc_log(xfs_mount_t      *mp,
1171                xfs_buftarg_t    *log_target,
1172                xfs_daddr_t      blk_offset,
1173                int              num_bblks)
1174 {
1175         xlog_t                  *log;
1176         xlog_rec_header_t       *head;
1177         xlog_in_core_t          **iclogp;
1178         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1179         xfs_buf_t               *bp;
1180         int                     i;
1181         int                     iclogsize;
1182         int                     error = ENOMEM;
1183
1184         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1185         if (!log) {
1186                 xlog_warn("XFS: Log allocation failed: No memory!");
1187                 goto out;
1188         }
1189
1190         log->l_mp          = mp;
1191         log->l_targ        = log_target;
1192         log->l_logsize     = BBTOB(num_bblks);
1193         log->l_logBBstart  = blk_offset;
1194         log->l_logBBsize   = num_bblks;
1195         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1196         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1197
1198         log->l_prev_block  = -1;
1199         log->l_tail_lsn    = xlog_assign_lsn(1, 0);
1200         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1201         log->l_last_sync_lsn = log->l_tail_lsn;
1202         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1203         log->l_grant_reserve_cycle = 1;
1204         log->l_grant_write_cycle = 1;
1205
1206         error = EFSCORRUPTED;
1207         if (xfs_sb_version_hassector(&mp->m_sb)) {
1208                 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1209                 if (log->l_sectbb_log < 0 ||
1210                     log->l_sectbb_log > mp->m_sectbb_log) {
1211                         xlog_warn("XFS: Log sector size (0x%x) out of range.",
1212                                                 log->l_sectbb_log);
1213                         goto out_free_log;
1214                 }
1215
1216                 /* for larger sector sizes, must have v2 or external log */
1217                 if (log->l_sectbb_log != 0 &&
1218                     (log->l_logBBstart != 0 &&
1219                      !xfs_sb_version_haslogv2(&mp->m_sb))) {
1220                         xlog_warn("XFS: log sector size (0x%x) invalid "
1221                                   "for configuration.", log->l_sectbb_log);
1222                         goto out_free_log;
1223                 }
1224                 if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1225                         xlog_warn("XFS: Log sector log (0x%x) too small.",
1226                                                 mp->m_sb.sb_logsectlog);
1227                         goto out_free_log;
1228                 }
1229         }
1230         log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1231
1232         xlog_get_iclog_buffer_size(mp, log);
1233
1234         error = ENOMEM;
1235         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1236         if (!bp)
1237                 goto out_free_log;
1238         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1239         XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1240         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1241         ASSERT(XFS_BUF_ISBUSY(bp));
1242         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1243         log->l_xbuf = bp;
1244
1245         spin_lock_init(&log->l_icloglock);
1246         spin_lock_init(&log->l_grant_lock);
1247         sv_init(&log->l_flush_wait, 0, "flush_wait");
1248
1249         xlog_trace_loggrant_alloc(log);
1250         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1251         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1252
1253         iclogp = &log->l_iclog;
1254         /*
1255          * The amount of memory to allocate for the iclog structure is
1256          * rather funky due to the way the structure is defined.  It is
1257          * done this way so that we can use different sizes for machines
1258          * with different amounts of memory.  See the definition of
1259          * xlog_in_core_t in xfs_log_priv.h for details.
1260          */
1261         iclogsize = log->l_iclog_size;
1262         ASSERT(log->l_iclog_size >= 4096);
1263         for (i=0; i < log->l_iclog_bufs; i++) {
1264                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1265                 if (!*iclogp)
1266                         goto out_free_iclog;
1267
1268                 iclog = *iclogp;
1269                 iclog->ic_prev = prev_iclog;
1270                 prev_iclog = iclog;
1271
1272                 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1273                 if (!bp)
1274                         goto out_free_iclog;
1275                 if (!XFS_BUF_CPSEMA(bp))
1276                         ASSERT(0);
1277                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1278                 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1279                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1280                 iclog->ic_bp = bp;
1281                 iclog->ic_data = bp->b_addr;
1282 #ifdef DEBUG
1283                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1284 #endif
1285                 head = &iclog->ic_header;
1286                 memset(head, 0, sizeof(xlog_rec_header_t));
1287                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1288                 head->h_version = cpu_to_be32(
1289                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1290                 head->h_size = cpu_to_be32(log->l_iclog_size);
1291                 /* new fields */
1292                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1293                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1294
1295                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1296                 iclog->ic_state = XLOG_STATE_ACTIVE;
1297                 iclog->ic_log = log;
1298                 atomic_set(&iclog->ic_refcnt, 0);
1299                 spin_lock_init(&iclog->ic_callback_lock);
1300                 iclog->ic_callback_tail = &(iclog->ic_callback);
1301                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1302
1303                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1304                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1305                 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1306                 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1307
1308                 xlog_trace_iclog_alloc(iclog);
1309
1310                 iclogp = &iclog->ic_next;
1311         }
1312         *iclogp = log->l_iclog;                 /* complete ring */
1313         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1314
1315         return log;
1316
1317 out_free_iclog:
1318         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1319                 prev_iclog = iclog->ic_next;
1320                 if (iclog->ic_bp) {
1321                         sv_destroy(&iclog->ic_force_wait);
1322                         sv_destroy(&iclog->ic_write_wait);
1323                         xfs_buf_free(iclog->ic_bp);
1324                         xlog_trace_iclog_dealloc(iclog);
1325                 }
1326                 kmem_free(iclog);
1327         }
1328         spinlock_destroy(&log->l_icloglock);
1329         spinlock_destroy(&log->l_grant_lock);
1330         xlog_trace_loggrant_dealloc(log);
1331         xfs_buf_free(log->l_xbuf);
1332 out_free_log:
1333         kmem_free(log);
1334 out:
1335         return ERR_PTR(-error);
1336 }       /* xlog_alloc_log */
1337
1338
1339 /*
1340  * Write out the commit record of a transaction associated with the given
1341  * ticket.  Return the lsn of the commit record.
1342  */
1343 STATIC int
1344 xlog_commit_record(xfs_mount_t  *mp,
1345                    xlog_ticket_t *ticket,
1346                    xlog_in_core_t **iclog,
1347                    xfs_lsn_t    *commitlsnp)
1348 {
1349         int             error;
1350         xfs_log_iovec_t reg[1];
1351
1352         reg[0].i_addr = NULL;
1353         reg[0].i_len = 0;
1354         XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1355
1356         ASSERT_ALWAYS(iclog);
1357         if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1358                                iclog, XLOG_COMMIT_TRANS))) {
1359                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1360         }
1361         return error;
1362 }       /* xlog_commit_record */
1363
1364
1365 /*
1366  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1367  * log space.  This code pushes on the lsn which would supposedly free up
1368  * the 25% which we want to leave free.  We may need to adopt a policy which
1369  * pushes on an lsn which is further along in the log once we reach the high
1370  * water mark.  In this manner, we would be creating a low water mark.
1371  */
1372 STATIC void
1373 xlog_grant_push_ail(xfs_mount_t *mp,
1374                     int         need_bytes)
1375 {
1376     xlog_t      *log = mp->m_log;       /* pointer to the log */
1377     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1378     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1379     int         free_blocks;            /* free blocks left to write to */
1380     int         free_bytes;             /* free bytes left to write to */
1381     int         threshold_block;        /* block in lsn we'd like to be at */
1382     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1383     int         free_threshold;
1384
1385     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1386
1387     spin_lock(&log->l_grant_lock);
1388     free_bytes = xlog_space_left(log,
1389                                  log->l_grant_reserve_cycle,
1390                                  log->l_grant_reserve_bytes);
1391     tail_lsn = log->l_tail_lsn;
1392     free_blocks = BTOBBT(free_bytes);
1393
1394     /*
1395      * Set the threshold for the minimum number of free blocks in the
1396      * log to the maximum of what the caller needs, one quarter of the
1397      * log, and 256 blocks.
1398      */
1399     free_threshold = BTOBB(need_bytes);
1400     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1401     free_threshold = MAX(free_threshold, 256);
1402     if (free_blocks < free_threshold) {
1403         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1404         threshold_cycle = CYCLE_LSN(tail_lsn);
1405         if (threshold_block >= log->l_logBBsize) {
1406             threshold_block -= log->l_logBBsize;
1407             threshold_cycle += 1;
1408         }
1409         threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1410
1411         /* Don't pass in an lsn greater than the lsn of the last
1412          * log record known to be on disk.
1413          */
1414         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1415             threshold_lsn = log->l_last_sync_lsn;
1416     }
1417     spin_unlock(&log->l_grant_lock);
1418
1419     /*
1420      * Get the transaction layer to kick the dirty buffers out to
1421      * disk asynchronously. No point in trying to do this if
1422      * the filesystem is shutting down.
1423      */
1424     if (threshold_lsn &&
1425         !XLOG_FORCED_SHUTDOWN(log))
1426             xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1427 }       /* xlog_grant_push_ail */
1428
1429
1430 /*
1431  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1432  * fashion.  Previously, we should have moved the current iclog
1433  * ptr in the log to point to the next available iclog.  This allows further
1434  * write to continue while this code syncs out an iclog ready to go.
1435  * Before an in-core log can be written out, the data section must be scanned
1436  * to save away the 1st word of each BBSIZE block into the header.  We replace
1437  * it with the current cycle count.  Each BBSIZE block is tagged with the
1438  * cycle count because there in an implicit assumption that drives will
1439  * guarantee that entire 512 byte blocks get written at once.  In other words,
1440  * we can't have part of a 512 byte block written and part not written.  By
1441  * tagging each block, we will know which blocks are valid when recovering
1442  * after an unclean shutdown.
1443  *
1444  * This routine is single threaded on the iclog.  No other thread can be in
1445  * this routine with the same iclog.  Changing contents of iclog can there-
1446  * fore be done without grabbing the state machine lock.  Updating the global
1447  * log will require grabbing the lock though.
1448  *
1449  * The entire log manager uses a logical block numbering scheme.  Only
1450  * log_sync (and then only bwrite()) know about the fact that the log may
1451  * not start with block zero on a given device.  The log block start offset
1452  * is added immediately before calling bwrite().
1453  */
1454
1455 STATIC int
1456 xlog_sync(xlog_t                *log,
1457           xlog_in_core_t        *iclog)
1458 {
1459         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1460         xfs_buf_t       *bp;
1461         int             i;
1462         uint            count;          /* byte count of bwrite */
1463         uint            count_init;     /* initial count before roundup */
1464         int             roundoff;       /* roundoff to BB or stripe */
1465         int             split = 0;      /* split write into two regions */
1466         int             error;
1467         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1468
1469         XFS_STATS_INC(xs_log_writes);
1470         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1471
1472         /* Add for LR header */
1473         count_init = log->l_iclog_hsize + iclog->ic_offset;
1474
1475         /* Round out the log write size */
1476         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1477                 /* we have a v2 stripe unit to use */
1478                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1479         } else {
1480                 count = BBTOB(BTOBB(count_init));
1481         }
1482         roundoff = count - count_init;
1483         ASSERT(roundoff >= 0);
1484         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1485                 roundoff < log->l_mp->m_sb.sb_logsunit)
1486                 || 
1487                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1488                  roundoff < BBTOB(1)));
1489
1490         /* move grant heads by roundoff in sync */
1491         spin_lock(&log->l_grant_lock);
1492         xlog_grant_add_space(log, roundoff);
1493         spin_unlock(&log->l_grant_lock);
1494
1495         /* put cycle number in every block */
1496         xlog_pack_data(log, iclog, roundoff); 
1497
1498         /* real byte length */
1499         if (v2) {
1500                 iclog->ic_header.h_len =
1501                         cpu_to_be32(iclog->ic_offset + roundoff);
1502         } else {
1503                 iclog->ic_header.h_len =
1504                         cpu_to_be32(iclog->ic_offset);
1505         }
1506
1507         bp = iclog->ic_bp;
1508         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1509         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1510         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1511
1512         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1513
1514         /* Do we need to split this write into 2 parts? */
1515         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1516                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1517                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1518                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1519         } else {
1520                 iclog->ic_bwritecnt = 1;
1521         }
1522         XFS_BUF_SET_COUNT(bp, count);
1523         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1524         XFS_BUF_ZEROFLAGS(bp);
1525         XFS_BUF_BUSY(bp);
1526         XFS_BUF_ASYNC(bp);
1527         /*
1528          * Do an ordered write for the log block.
1529          * Its unnecessary to flush the first split block in the log wrap case.
1530          */
1531         if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1532                 XFS_BUF_ORDERED(bp);
1533
1534         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1535         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1536
1537         xlog_verify_iclog(log, iclog, count, B_TRUE);
1538
1539         /* account for log which doesn't start at block #0 */
1540         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1541         /*
1542          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1543          * is shutting down.
1544          */
1545         XFS_BUF_WRITE(bp);
1546
1547         if ((error = XFS_bwrite(bp))) {
1548                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1549                                   XFS_BUF_ADDR(bp));
1550                 return error;
1551         }
1552         if (split) {
1553                 bp = iclog->ic_log->l_xbuf;
1554                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1555                                                         (unsigned long)1);
1556                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1557                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1558                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1559                                             (__psint_t)count), split);
1560                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1561                 XFS_BUF_ZEROFLAGS(bp);
1562                 XFS_BUF_BUSY(bp);
1563                 XFS_BUF_ASYNC(bp);
1564                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1565                         XFS_BUF_ORDERED(bp);
1566                 dptr = XFS_BUF_PTR(bp);
1567                 /*
1568                  * Bump the cycle numbers at the start of each block
1569                  * since this part of the buffer is at the start of
1570                  * a new cycle.  Watch out for the header magic number
1571                  * case, though.
1572                  */
1573                 for (i = 0; i < split; i += BBSIZE) {
1574                         be32_add_cpu((__be32 *)dptr, 1);
1575                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1576                                 be32_add_cpu((__be32 *)dptr, 1);
1577                         dptr += BBSIZE;
1578                 }
1579
1580                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1581                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1582
1583                 /* account for internal log which doesn't start at block #0 */
1584                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1585                 XFS_BUF_WRITE(bp);
1586                 if ((error = XFS_bwrite(bp))) {
1587                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1588                                           bp, XFS_BUF_ADDR(bp));
1589                         return error;
1590                 }
1591         }
1592         return 0;
1593 }       /* xlog_sync */
1594
1595
1596 /*
1597  * Deallocate a log structure
1598  */
1599 STATIC void
1600 xlog_dealloc_log(xlog_t *log)
1601 {
1602         xlog_in_core_t  *iclog, *next_iclog;
1603         int             i;
1604
1605         iclog = log->l_iclog;
1606         for (i=0; i<log->l_iclog_bufs; i++) {
1607                 sv_destroy(&iclog->ic_force_wait);
1608                 sv_destroy(&iclog->ic_write_wait);
1609                 xfs_buf_free(iclog->ic_bp);
1610                 xlog_trace_iclog_dealloc(iclog);
1611                 next_iclog = iclog->ic_next;
1612                 kmem_free(iclog);
1613                 iclog = next_iclog;
1614         }
1615         spinlock_destroy(&log->l_icloglock);
1616         spinlock_destroy(&log->l_grant_lock);
1617
1618         xfs_buf_free(log->l_xbuf);
1619         xlog_trace_loggrant_dealloc(log);
1620         log->l_mp->m_log = NULL;
1621         kmem_free(log);
1622 }       /* xlog_dealloc_log */
1623
1624 /*
1625  * Update counters atomically now that memcpy is done.
1626  */
1627 /* ARGSUSED */
1628 static inline void
1629 xlog_state_finish_copy(xlog_t           *log,
1630                        xlog_in_core_t   *iclog,
1631                        int              record_cnt,
1632                        int              copy_bytes)
1633 {
1634         spin_lock(&log->l_icloglock);
1635
1636         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1637         iclog->ic_offset += copy_bytes;
1638
1639         spin_unlock(&log->l_icloglock);
1640 }       /* xlog_state_finish_copy */
1641
1642
1643
1644
1645 /*
1646  * print out info relating to regions written which consume
1647  * the reservation
1648  */
1649 STATIC void
1650 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1651 {
1652         uint i;
1653         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1654
1655         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1656         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1657             "bformat",
1658             "bchunk",
1659             "efi_format",
1660             "efd_format",
1661             "iformat",
1662             "icore",
1663             "iext",
1664             "ibroot",
1665             "ilocal",
1666             "iattr_ext",
1667             "iattr_broot",
1668             "iattr_local",
1669             "qformat",
1670             "dquot",
1671             "quotaoff",
1672             "LR header",
1673             "unmount",
1674             "commit",
1675             "trans header"
1676         };
1677         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1678             "SETATTR_NOT_SIZE",
1679             "SETATTR_SIZE",
1680             "INACTIVE",
1681             "CREATE",
1682             "CREATE_TRUNC",
1683             "TRUNCATE_FILE",
1684             "REMOVE",
1685             "LINK",
1686             "RENAME",
1687             "MKDIR",
1688             "RMDIR",
1689             "SYMLINK",
1690             "SET_DMATTRS",
1691             "GROWFS",
1692             "STRAT_WRITE",
1693             "DIOSTRAT",
1694             "WRITE_SYNC",
1695             "WRITEID",
1696             "ADDAFORK",
1697             "ATTRINVAL",
1698             "ATRUNCATE",
1699             "ATTR_SET",
1700             "ATTR_RM",
1701             "ATTR_FLAG",
1702             "CLEAR_AGI_BUCKET",
1703             "QM_SBCHANGE",
1704             "DUMMY1",
1705             "DUMMY2",
1706             "QM_QUOTAOFF",
1707             "QM_DQALLOC",
1708             "QM_SETQLIM",
1709             "QM_DQCLUSTER",
1710             "QM_QINOCREATE",
1711             "QM_QUOTAOFF_END",
1712             "SB_UNIT",
1713             "FSYNC_TS",
1714             "GROWFSRT_ALLOC",
1715             "GROWFSRT_ZERO",
1716             "GROWFSRT_FREE",
1717             "SWAPEXT"
1718         };
1719
1720         xfs_fs_cmn_err(CE_WARN, mp,
1721                         "xfs_log_write: reservation summary:\n"
1722                         "  trans type  = %s (%u)\n"
1723                         "  unit res    = %d bytes\n"
1724                         "  current res = %d bytes\n"
1725                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1726                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1727                         "  ophdr + reg = %u bytes\n"
1728                         "  num regions = %u\n",
1729                         ((ticket->t_trans_type <= 0 ||
1730                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1731                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1732                         ticket->t_trans_type,
1733                         ticket->t_unit_res,
1734                         ticket->t_curr_res,
1735                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1736                         ticket->t_res_num_ophdrs, ophdr_spc,
1737                         ticket->t_res_arr_sum + 
1738                         ticket->t_res_o_flow + ophdr_spc,
1739                         ticket->t_res_num);
1740
1741         for (i = 0; i < ticket->t_res_num; i++) {
1742                 uint r_type = ticket->t_res_arr[i].r_type; 
1743                 cmn_err(CE_WARN,
1744                             "region[%u]: %s - %u bytes\n",
1745                             i, 
1746                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1747                             "bad-rtype" : res_type_str[r_type-1]),
1748                             ticket->t_res_arr[i].r_len);
1749         }
1750 }
1751
1752 /*
1753  * Write some region out to in-core log
1754  *
1755  * This will be called when writing externally provided regions or when
1756  * writing out a commit record for a given transaction.
1757  *
1758  * General algorithm:
1759  *      1. Find total length of this write.  This may include adding to the
1760  *              lengths passed in.
1761  *      2. Check whether we violate the tickets reservation.
1762  *      3. While writing to this iclog
1763  *          A. Reserve as much space in this iclog as can get
1764  *          B. If this is first write, save away start lsn
1765  *          C. While writing this region:
1766  *              1. If first write of transaction, write start record
1767  *              2. Write log operation header (header per region)
1768  *              3. Find out if we can fit entire region into this iclog
1769  *              4. Potentially, verify destination memcpy ptr
1770  *              5. Memcpy (partial) region
1771  *              6. If partial copy, release iclog; otherwise, continue
1772  *                      copying more regions into current iclog
1773  *      4. Mark want sync bit (in simulation mode)
1774  *      5. Release iclog for potential flush to on-disk log.
1775  *
1776  * ERRORS:
1777  * 1.   Panic if reservation is overrun.  This should never happen since
1778  *      reservation amounts are generated internal to the filesystem.
1779  * NOTES:
1780  * 1. Tickets are single threaded data structures.
1781  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1782  *      syncing routine.  When a single log_write region needs to span
1783  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1784  *      on all log operation writes which don't contain the end of the
1785  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1786  *      operation which contains the end of the continued log_write region.
1787  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1788  *      we don't really know exactly how much space will be used.  As a result,
1789  *      we don't update ic_offset until the end when we know exactly how many
1790  *      bytes have been written out.
1791  */
1792 STATIC int
1793 xlog_write(xfs_mount_t *        mp,
1794            xfs_log_iovec_t      reg[],
1795            int                  nentries,
1796            xfs_log_ticket_t     tic,
1797            xfs_lsn_t            *start_lsn,
1798            xlog_in_core_t       **commit_iclog,
1799            uint                 flags)
1800 {
1801     xlog_t           *log = mp->m_log;
1802     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1803     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1804     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1805     __psint_t        ptr;            /* copy address into data region */
1806     int              len;            /* # xlog_write() bytes 2 still copy */
1807     int              index;          /* region index currently copying */
1808     int              log_offset;     /* offset (from 0) into data region */
1809     int              start_rec_copy; /* # bytes to copy for start record */
1810     int              partial_copy;   /* did we split a region? */
1811     int              partial_copy_len;/* # bytes copied if split region */
1812     int              need_copy;      /* # bytes need to memcpy this region */
1813     int              copy_len;       /* # bytes actually memcpy'ing */
1814     int              copy_off;       /* # bytes from entry start */
1815     int              contwr;         /* continued write of in-core log? */
1816     int              error;
1817     int              record_cnt = 0, data_cnt = 0;
1818
1819     partial_copy_len = partial_copy = 0;
1820
1821     /* Calculate potential maximum space.  Each region gets its own
1822      * xlog_op_header_t and may need to be double word aligned.
1823      */
1824     len = 0;
1825     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1826         len += sizeof(xlog_op_header_t);
1827         ticket->t_res_num_ophdrs++;
1828     }
1829
1830     for (index = 0; index < nentries; index++) {
1831         len += sizeof(xlog_op_header_t);            /* each region gets >= 1 */
1832         ticket->t_res_num_ophdrs++;
1833         len += reg[index].i_len;
1834         xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1835     }
1836     contwr = *start_lsn = 0;
1837
1838     if (ticket->t_curr_res < len) {
1839         xlog_print_tic_res(mp, ticket);
1840 #ifdef DEBUG
1841         xlog_panic(
1842                 "xfs_log_write: reservation ran out. Need to up reservation");
1843 #else
1844         /* Customer configurable panic */
1845         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1846                 "xfs_log_write: reservation ran out. Need to up reservation");
1847         /* If we did not panic, shutdown the filesystem */
1848         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1849 #endif
1850     } else
1851         ticket->t_curr_res -= len;
1852
1853     for (index = 0; index < nentries; ) {
1854         if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1855                                                &contwr, &log_offset)))
1856                 return error;
1857
1858         ASSERT(log_offset <= iclog->ic_size - 1);
1859         ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1860
1861         /* start_lsn is the first lsn written to. That's all we need. */
1862         if (! *start_lsn)
1863             *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1864
1865         /* This loop writes out as many regions as can fit in the amount
1866          * of space which was allocated by xlog_state_get_iclog_space().
1867          */
1868         while (index < nentries) {
1869             ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1870             ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1871             start_rec_copy = 0;
1872
1873             /* If first write for transaction, insert start record.
1874              * We can't be trying to commit if we are inited.  We can't
1875              * have any "partial_copy" if we are inited.
1876              */
1877             if (ticket->t_flags & XLOG_TIC_INITED) {
1878                 logop_head              = (xlog_op_header_t *)ptr;
1879                 logop_head->oh_tid      = cpu_to_be32(ticket->t_tid);
1880                 logop_head->oh_clientid = ticket->t_clientid;
1881                 logop_head->oh_len      = 0;
1882                 logop_head->oh_flags    = XLOG_START_TRANS;
1883                 logop_head->oh_res2     = 0;
1884                 ticket->t_flags         &= ~XLOG_TIC_INITED;    /* clear bit */
1885                 record_cnt++;
1886
1887                 start_rec_copy = sizeof(xlog_op_header_t);
1888                 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1889             }
1890
1891             /* Copy log operation header directly into data section */
1892             logop_head                  = (xlog_op_header_t *)ptr;
1893             logop_head->oh_tid          = cpu_to_be32(ticket->t_tid);
1894             logop_head->oh_clientid     = ticket->t_clientid;
1895             logop_head->oh_res2         = 0;
1896
1897             /* header copied directly */
1898             xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1899
1900             /* are we copying a commit or unmount record? */
1901             logop_head->oh_flags = flags;
1902
1903             /*
1904              * We've seen logs corrupted with bad transaction client
1905              * ids.  This makes sure that XFS doesn't generate them on.
1906              * Turn this into an EIO and shut down the filesystem.
1907              */
1908             switch (logop_head->oh_clientid)  {
1909             case XFS_TRANSACTION:
1910             case XFS_VOLUME:
1911             case XFS_LOG:
1912                 break;
1913             default:
1914                 xfs_fs_cmn_err(CE_WARN, mp,
1915                     "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1916                     logop_head->oh_clientid, tic);
1917                 return XFS_ERROR(EIO);
1918             }
1919
1920             /* Partial write last time? => (partial_copy != 0)
1921              * need_copy is the amount we'd like to copy if everything could
1922              * fit in the current memcpy.
1923              */
1924             need_copy = reg[index].i_len - partial_copy_len;
1925
1926             copy_off = partial_copy_len;
1927             if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1928                 copy_len = need_copy;
1929                 logop_head->oh_len = cpu_to_be32(copy_len);
1930                 if (partial_copy)
1931                     logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1932                 partial_copy_len = partial_copy = 0;
1933             } else {                                        /* partial write */
1934                 copy_len = iclog->ic_size - log_offset;
1935                 logop_head->oh_len = cpu_to_be32(copy_len);
1936                 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1937                 if (partial_copy)
1938                         logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1939                 partial_copy_len += copy_len;
1940                 partial_copy++;
1941                 len += sizeof(xlog_op_header_t); /* from splitting of region */
1942                 /* account for new log op header */
1943                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1944                 ticket->t_res_num_ophdrs++;
1945             }
1946             xlog_verify_dest_ptr(log, ptr);
1947
1948             /* copy region */
1949             ASSERT(copy_len >= 0);
1950             memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1951             xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1952
1953             /* make copy_len total bytes copied, including headers */
1954             copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1955             record_cnt++;
1956             data_cnt += contwr ? copy_len : 0;
1957             if (partial_copy) {                 /* copied partial region */
1958                     /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1959                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1960                     record_cnt = data_cnt = 0;
1961                     if ((error = xlog_state_release_iclog(log, iclog)))
1962                             return error;
1963                     break;                      /* don't increment index */
1964             } else {                            /* copied entire region */
1965                 index++;
1966                 partial_copy_len = partial_copy = 0;
1967
1968                 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1969                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1970                     record_cnt = data_cnt = 0;
1971                     spin_lock(&log->l_icloglock);
1972                     xlog_state_want_sync(log, iclog);
1973                     spin_unlock(&log->l_icloglock);
1974                     if (commit_iclog) {
1975                         ASSERT(flags & XLOG_COMMIT_TRANS);
1976                         *commit_iclog = iclog;
1977                     } else if ((error = xlog_state_release_iclog(log, iclog)))
1978                            return error;
1979                     if (index == nentries)
1980                             return 0;           /* we are done */
1981                     else
1982                             break;
1983                 }
1984             } /* if (partial_copy) */
1985         } /* while (index < nentries) */
1986     } /* for (index = 0; index < nentries; ) */
1987     ASSERT(len == 0);
1988
1989     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1990     if (commit_iclog) {
1991         ASSERT(flags & XLOG_COMMIT_TRANS);
1992         *commit_iclog = iclog;
1993         return 0;
1994     }
1995     return xlog_state_release_iclog(log, iclog);
1996 }       /* xlog_write */
1997
1998
1999 /*****************************************************************************
2000  *
2001  *              State Machine functions
2002  *
2003  *****************************************************************************
2004  */
2005
2006 /* Clean iclogs starting from the head.  This ordering must be
2007  * maintained, so an iclog doesn't become ACTIVE beyond one that
2008  * is SYNCING.  This is also required to maintain the notion that we use
2009  * a ordered wait queue to hold off would be writers to the log when every
2010  * iclog is trying to sync to disk.
2011  *
2012  * State Change: DIRTY -> ACTIVE
2013  */
2014 STATIC void
2015 xlog_state_clean_log(xlog_t *log)
2016 {
2017         xlog_in_core_t  *iclog;
2018         int changed = 0;
2019
2020         iclog = log->l_iclog;
2021         do {
2022                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2023                         iclog->ic_state = XLOG_STATE_ACTIVE;
2024                         iclog->ic_offset       = 0;
2025                         ASSERT(iclog->ic_callback == NULL);
2026                         /*
2027                          * If the number of ops in this iclog indicate it just
2028                          * contains the dummy transaction, we can
2029                          * change state into IDLE (the second time around).
2030                          * Otherwise we should change the state into
2031                          * NEED a dummy.
2032                          * We don't need to cover the dummy.
2033                          */
2034                         if (!changed &&
2035                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2036                                         XLOG_COVER_OPS)) {
2037                                 changed = 1;
2038                         } else {
2039                                 /*
2040                                  * We have two dirty iclogs so start over
2041                                  * This could also be num of ops indicates
2042                                  * this is not the dummy going out.
2043                                  */
2044                                 changed = 2;
2045                         }
2046                         iclog->ic_header.h_num_logops = 0;
2047                         memset(iclog->ic_header.h_cycle_data, 0,
2048                               sizeof(iclog->ic_header.h_cycle_data));
2049                         iclog->ic_header.h_lsn = 0;
2050                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2051                         /* do nothing */;
2052                 else
2053                         break;  /* stop cleaning */
2054                 iclog = iclog->ic_next;
2055         } while (iclog != log->l_iclog);
2056
2057         /* log is locked when we are called */
2058         /*
2059          * Change state for the dummy log recording.
2060          * We usually go to NEED. But we go to NEED2 if the changed indicates
2061          * we are done writing the dummy record.
2062          * If we are done with the second dummy recored (DONE2), then
2063          * we go to IDLE.
2064          */
2065         if (changed) {
2066                 switch (log->l_covered_state) {
2067                 case XLOG_STATE_COVER_IDLE:
2068                 case XLOG_STATE_COVER_NEED:
2069                 case XLOG_STATE_COVER_NEED2:
2070                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2071                         break;
2072
2073                 case XLOG_STATE_COVER_DONE:
2074                         if (changed == 1)
2075                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2076                         else
2077                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2078                         break;
2079
2080                 case XLOG_STATE_COVER_DONE2:
2081                         if (changed == 1)
2082                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2083                         else
2084                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2085                         break;
2086
2087                 default:
2088                         ASSERT(0);
2089                 }
2090         }
2091 }       /* xlog_state_clean_log */
2092
2093 STATIC xfs_lsn_t
2094 xlog_get_lowest_lsn(
2095         xlog_t          *log)
2096 {
2097         xlog_in_core_t  *lsn_log;
2098         xfs_lsn_t       lowest_lsn, lsn;
2099
2100         lsn_log = log->l_iclog;
2101         lowest_lsn = 0;
2102         do {
2103             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2104                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2105                 if ((lsn && !lowest_lsn) ||
2106                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2107                         lowest_lsn = lsn;
2108                 }
2109             }
2110             lsn_log = lsn_log->ic_next;
2111         } while (lsn_log != log->l_iclog);
2112         return lowest_lsn;
2113 }
2114
2115
2116 STATIC void
2117 xlog_state_do_callback(
2118         xlog_t          *log,
2119         int             aborted,
2120         xlog_in_core_t  *ciclog)
2121 {
2122         xlog_in_core_t     *iclog;
2123         xlog_in_core_t     *first_iclog;        /* used to know when we've
2124                                                  * processed all iclogs once */
2125         xfs_log_callback_t *cb, *cb_next;
2126         int                flushcnt = 0;
2127         xfs_lsn_t          lowest_lsn;
2128         int                ioerrors;    /* counter: iclogs with errors */
2129         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2130         int                funcdidcallbacks; /* flag: function did callbacks */
2131         int                repeats;     /* for issuing console warnings if
2132                                          * looping too many times */
2133         int                wake = 0;
2134
2135         spin_lock(&log->l_icloglock);
2136         first_iclog = iclog = log->l_iclog;
2137         ioerrors = 0;
2138         funcdidcallbacks = 0;
2139         repeats = 0;
2140
2141         do {
2142                 /*
2143                  * Scan all iclogs starting with the one pointed to by the
2144                  * log.  Reset this starting point each time the log is
2145                  * unlocked (during callbacks).
2146                  *
2147                  * Keep looping through iclogs until one full pass is made
2148                  * without running any callbacks.
2149                  */
2150                 first_iclog = log->l_iclog;
2151                 iclog = log->l_iclog;
2152                 loopdidcallbacks = 0;
2153                 repeats++;
2154
2155                 do {
2156
2157                         /* skip all iclogs in the ACTIVE & DIRTY states */
2158                         if (iclog->ic_state &
2159                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2160                                 iclog = iclog->ic_next;
2161                                 continue;
2162                         }
2163
2164                         /*
2165                          * Between marking a filesystem SHUTDOWN and stopping
2166                          * the log, we do flush all iclogs to disk (if there
2167                          * wasn't a log I/O error). So, we do want things to
2168                          * go smoothly in case of just a SHUTDOWN  w/o a
2169                          * LOG_IO_ERROR.
2170                          */
2171                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2172                                 /*
2173                                  * Can only perform callbacks in order.  Since
2174                                  * this iclog is not in the DONE_SYNC/
2175                                  * DO_CALLBACK state, we skip the rest and
2176                                  * just try to clean up.  If we set our iclog
2177                                  * to DO_CALLBACK, we will not process it when
2178                                  * we retry since a previous iclog is in the
2179                                  * CALLBACK and the state cannot change since
2180                                  * we are holding the l_icloglock.
2181                                  */
2182                                 if (!(iclog->ic_state &
2183                                         (XLOG_STATE_DONE_SYNC |
2184                                                  XLOG_STATE_DO_CALLBACK))) {
2185                                         if (ciclog && (ciclog->ic_state ==
2186                                                         XLOG_STATE_DONE_SYNC)) {
2187                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2188                                         }
2189                                         break;
2190                                 }
2191                                 /*
2192                                  * We now have an iclog that is in either the
2193                                  * DO_CALLBACK or DONE_SYNC states. The other
2194                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2195                                  * caught by the above if and are going to
2196                                  * clean (i.e. we aren't doing their callbacks)
2197                                  * see the above if.
2198                                  */
2199
2200                                 /*
2201                                  * We will do one more check here to see if we
2202                                  * have chased our tail around.
2203                                  */
2204
2205                                 lowest_lsn = xlog_get_lowest_lsn(log);
2206                                 if (lowest_lsn &&
2207                                     XFS_LSN_CMP(lowest_lsn,
2208                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2209                                         iclog = iclog->ic_next;
2210                                         continue; /* Leave this iclog for
2211                                                    * another thread */
2212                                 }
2213
2214                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2215
2216                                 spin_unlock(&log->l_icloglock);
2217
2218                                 /* l_last_sync_lsn field protected by
2219                                  * l_grant_lock. Don't worry about iclog's lsn.
2220                                  * No one else can be here except us.
2221                                  */
2222                                 spin_lock(&log->l_grant_lock);
2223                                 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2224                                        be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2225                                 log->l_last_sync_lsn =
2226                                         be64_to_cpu(iclog->ic_header.h_lsn);
2227                                 spin_unlock(&log->l_grant_lock);
2228
2229                         } else {
2230                                 spin_unlock(&log->l_icloglock);
2231                                 ioerrors++;
2232                         }
2233
2234                         /*
2235                          * Keep processing entries in the callback list until
2236                          * we come around and it is empty.  We need to
2237                          * atomically see that the list is empty and change the
2238                          * state to DIRTY so that we don't miss any more
2239                          * callbacks being added.
2240                          */
2241                         spin_lock(&iclog->ic_callback_lock);
2242                         cb = iclog->ic_callback;
2243                         while (cb) {
2244                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2245                                 iclog->ic_callback = NULL;
2246                                 spin_unlock(&iclog->ic_callback_lock);
2247
2248                                 /* perform callbacks in the order given */
2249                                 for (; cb; cb = cb_next) {
2250                                         cb_next = cb->cb_next;
2251                                         cb->cb_func(cb->cb_arg, aborted);
2252                                 }
2253                                 spin_lock(&iclog->ic_callback_lock);
2254                                 cb = iclog->ic_callback;
2255                         }
2256
2257                         loopdidcallbacks++;
2258                         funcdidcallbacks++;
2259
2260                         spin_lock(&log->l_icloglock);
2261                         ASSERT(iclog->ic_callback == NULL);
2262                         spin_unlock(&iclog->ic_callback_lock);
2263                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2264                                 iclog->ic_state = XLOG_STATE_DIRTY;
2265
2266                         /*
2267                          * Transition from DIRTY to ACTIVE if applicable.
2268                          * NOP if STATE_IOERROR.
2269                          */
2270                         xlog_state_clean_log(log);
2271
2272                         /* wake up threads waiting in xfs_log_force() */
2273                         sv_broadcast(&iclog->ic_force_wait);
2274
2275                         iclog = iclog->ic_next;
2276                 } while (first_iclog != iclog);
2277
2278                 if (repeats > 5000) {
2279                         flushcnt += repeats;
2280                         repeats = 0;
2281                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2282                                 "%s: possible infinite loop (%d iterations)",
2283                                 __func__, flushcnt);
2284                 }
2285         } while (!ioerrors && loopdidcallbacks);
2286
2287         /*
2288          * make one last gasp attempt to see if iclogs are being left in
2289          * limbo..
2290          */
2291 #ifdef DEBUG
2292         if (funcdidcallbacks) {
2293                 first_iclog = iclog = log->l_iclog;
2294                 do {
2295                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2296                         /*
2297                          * Terminate the loop if iclogs are found in states
2298                          * which will cause other threads to clean up iclogs.
2299                          *
2300                          * SYNCING - i/o completion will go through logs
2301                          * DONE_SYNC - interrupt thread should be waiting for
2302                          *              l_icloglock
2303                          * IOERROR - give up hope all ye who enter here
2304                          */
2305                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2306                             iclog->ic_state == XLOG_STATE_SYNCING ||
2307                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2308                             iclog->ic_state == XLOG_STATE_IOERROR )
2309                                 break;
2310                         iclog = iclog->ic_next;
2311                 } while (first_iclog != iclog);
2312         }
2313 #endif
2314
2315         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2316                 wake = 1;
2317         spin_unlock(&log->l_icloglock);
2318
2319         if (wake)
2320                 sv_broadcast(&log->l_flush_wait);
2321 }
2322
2323
2324 /*
2325  * Finish transitioning this iclog to the dirty state.
2326  *
2327  * Make sure that we completely execute this routine only when this is
2328  * the last call to the iclog.  There is a good chance that iclog flushes,
2329  * when we reach the end of the physical log, get turned into 2 separate
2330  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2331  * routine.  By using the reference count bwritecnt, we guarantee that only
2332  * the second completion goes through.
2333  *
2334  * Callbacks could take time, so they are done outside the scope of the
2335  * global state machine log lock.
2336  */
2337 STATIC void
2338 xlog_state_done_syncing(
2339         xlog_in_core_t  *iclog,
2340         int             aborted)
2341 {
2342         xlog_t             *log = iclog->ic_log;
2343
2344         spin_lock(&log->l_icloglock);
2345
2346         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2347                iclog->ic_state == XLOG_STATE_IOERROR);
2348         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2349         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2350
2351
2352         /*
2353          * If we got an error, either on the first buffer, or in the case of
2354          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2355          * and none should ever be attempted to be written to disk
2356          * again.
2357          */
2358         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2359                 if (--iclog->ic_bwritecnt == 1) {
2360                         spin_unlock(&log->l_icloglock);
2361                         return;
2362                 }
2363                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2364         }
2365
2366         /*
2367          * Someone could be sleeping prior to writing out the next
2368          * iclog buffer, we wake them all, one will get to do the
2369          * I/O, the others get to wait for the result.
2370          */
2371         sv_broadcast(&iclog->ic_write_wait);
2372         spin_unlock(&log->l_icloglock);
2373         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2374 }       /* xlog_state_done_syncing */
2375
2376
2377 /*
2378  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2379  * sleep.  We wait on the flush queue on the head iclog as that should be
2380  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2381  * we will wait here and all new writes will sleep until a sync completes.
2382  *
2383  * The in-core logs are used in a circular fashion. They are not used
2384  * out-of-order even when an iclog past the head is free.
2385  *
2386  * return:
2387  *      * log_offset where xlog_write() can start writing into the in-core
2388  *              log's data space.
2389  *      * in-core log pointer to which xlog_write() should write.
2390  *      * boolean indicating this is a continued write to an in-core log.
2391  *              If this is the last write, then the in-core log's offset field
2392  *              needs to be incremented, depending on the amount of data which
2393  *              is copied.
2394  */
2395 STATIC int
2396 xlog_state_get_iclog_space(xlog_t         *log,
2397                            int            len,
2398                            xlog_in_core_t **iclogp,
2399                            xlog_ticket_t  *ticket,
2400                            int            *continued_write,
2401                            int            *logoffsetp)
2402 {
2403         int               log_offset;
2404         xlog_rec_header_t *head;
2405         xlog_in_core_t    *iclog;
2406         int               error;
2407
2408 restart:
2409         spin_lock(&log->l_icloglock);
2410         if (XLOG_FORCED_SHUTDOWN(log)) {
2411                 spin_unlock(&log->l_icloglock);
2412                 return XFS_ERROR(EIO);
2413         }
2414
2415         iclog = log->l_iclog;
2416         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2417                 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2418                 XFS_STATS_INC(xs_log_noiclogs);
2419
2420                 /* Wait for log writes to have flushed */
2421                 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2422                 goto restart;
2423         }
2424
2425         head = &iclog->ic_header;
2426
2427         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2428         log_offset = iclog->ic_offset;
2429
2430         /* On the 1st write to an iclog, figure out lsn.  This works
2431          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2432          * committing to.  If the offset is set, that's how many blocks
2433          * must be written.
2434          */
2435         if (log_offset == 0) {
2436                 ticket->t_curr_res -= log->l_iclog_hsize;
2437                 xlog_tic_add_region(ticket,
2438                                     log->l_iclog_hsize,
2439                                     XLOG_REG_TYPE_LRHEADER);
2440                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2441                 head->h_lsn = cpu_to_be64(
2442                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2443                 ASSERT(log->l_curr_block >= 0);
2444         }
2445
2446         /* If there is enough room to write everything, then do it.  Otherwise,
2447          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2448          * bit is on, so this will get flushed out.  Don't update ic_offset
2449          * until you know exactly how many bytes get copied.  Therefore, wait
2450          * until later to update ic_offset.
2451          *
2452          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2453          * can fit into remaining data section.
2454          */
2455         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2456                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2457
2458                 /*
2459                  * If I'm the only one writing to this iclog, sync it to disk.
2460                  * We need to do an atomic compare and decrement here to avoid
2461                  * racing with concurrent atomic_dec_and_lock() calls in
2462                  * xlog_state_release_iclog() when there is more than one
2463                  * reference to the iclog.
2464                  */
2465                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2466                         /* we are the only one */
2467                         spin_unlock(&log->l_icloglock);
2468                         error = xlog_state_release_iclog(log, iclog);
2469                         if (error)
2470                                 return error;
2471                 } else {
2472                         spin_unlock(&log->l_icloglock);
2473                 }
2474                 goto restart;
2475         }
2476
2477         /* Do we have enough room to write the full amount in the remainder
2478          * of this iclog?  Or must we continue a write on the next iclog and
2479          * mark this iclog as completely taken?  In the case where we switch
2480          * iclogs (to mark it taken), this particular iclog will release/sync
2481          * to disk in xlog_write().
2482          */
2483         if (len <= iclog->ic_size - iclog->ic_offset) {
2484                 *continued_write = 0;
2485                 iclog->ic_offset += len;
2486         } else {
2487                 *continued_write = 1;
2488                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2489         }
2490         *iclogp = iclog;
2491
2492         ASSERT(iclog->ic_offset <= iclog->ic_size);
2493         spin_unlock(&log->l_icloglock);
2494
2495         *logoffsetp = log_offset;
2496         return 0;
2497 }       /* xlog_state_get_iclog_space */
2498
2499 /*
2500  * Atomically get the log space required for a log ticket.
2501  *
2502  * Once a ticket gets put onto the reserveq, it will only return after
2503  * the needed reservation is satisfied.
2504  */
2505 STATIC int
2506 xlog_grant_log_space(xlog_t        *log,
2507                      xlog_ticket_t *tic)
2508 {
2509         int              free_bytes;
2510         int              need_bytes;
2511 #ifdef DEBUG
2512         xfs_lsn_t        tail_lsn;
2513 #endif
2514
2515
2516 #ifdef DEBUG
2517         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2518                 panic("grant Recovery problem");
2519 #endif
2520
2521         /* Is there space or do we need to sleep? */
2522         spin_lock(&log->l_grant_lock);
2523         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2524
2525         /* something is already sleeping; insert new transaction at end */
2526         if (log->l_reserve_headq) {
2527                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2528                 xlog_trace_loggrant(log, tic,
2529                                     "xlog_grant_log_space: sleep 1");
2530                 /*
2531                  * Gotta check this before going to sleep, while we're
2532                  * holding the grant lock.
2533                  */
2534                 if (XLOG_FORCED_SHUTDOWN(log))
2535                         goto error_return;
2536
2537                 XFS_STATS_INC(xs_sleep_logspace);
2538                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2539                 /*
2540                  * If we got an error, and the filesystem is shutting down,
2541                  * we'll catch it down below. So just continue...
2542                  */
2543                 xlog_trace_loggrant(log, tic,
2544                                     "xlog_grant_log_space: wake 1");
2545                 spin_lock(&log->l_grant_lock);
2546         }
2547         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2548                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2549         else
2550                 need_bytes = tic->t_unit_res;
2551
2552 redo:
2553         if (XLOG_FORCED_SHUTDOWN(log))
2554                 goto error_return;
2555
2556         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2557                                      log->l_grant_reserve_bytes);
2558         if (free_bytes < need_bytes) {
2559                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2560                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2561                 xlog_trace_loggrant(log, tic,
2562                                     "xlog_grant_log_space: sleep 2");
2563                 spin_unlock(&log->l_grant_lock);
2564                 xlog_grant_push_ail(log->l_mp, need_bytes);
2565                 spin_lock(&log->l_grant_lock);
2566
2567                 XFS_STATS_INC(xs_sleep_logspace);
2568                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2569
2570                 spin_lock(&log->l_grant_lock);
2571                 if (XLOG_FORCED_SHUTDOWN(log))
2572                         goto error_return;
2573
2574                 xlog_trace_loggrant(log, tic,
2575                                     "xlog_grant_log_space: wake 2");
2576                 goto redo;
2577         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2578                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2579
2580         /* we've got enough space */
2581         xlog_grant_add_space(log, need_bytes);
2582 #ifdef DEBUG
2583         tail_lsn = log->l_tail_lsn;
2584         /*
2585          * Check to make sure the grant write head didn't just over lap the
2586          * tail.  If the cycles are the same, we can't be overlapping.
2587          * Otherwise, make sure that the cycles differ by exactly one and
2588          * check the byte count.
2589          */
2590         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2591                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2592                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2593         }
2594 #endif
2595         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2596         xlog_verify_grant_head(log, 1);
2597         spin_unlock(&log->l_grant_lock);
2598         return 0;
2599
2600  error_return:
2601         if (tic->t_flags & XLOG_TIC_IN_Q)
2602                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2603         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2604         /*
2605          * If we are failing, make sure the ticket doesn't have any
2606          * current reservations. We don't want to add this back when
2607          * the ticket/transaction gets cancelled.
2608          */
2609         tic->t_curr_res = 0;
2610         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2611         spin_unlock(&log->l_grant_lock);
2612         return XFS_ERROR(EIO);
2613 }       /* xlog_grant_log_space */
2614
2615
2616 /*
2617  * Replenish the byte reservation required by moving the grant write head.
2618  *
2619  *
2620  */
2621 STATIC int
2622 xlog_regrant_write_log_space(xlog_t        *log,
2623                              xlog_ticket_t *tic)
2624 {
2625         int             free_bytes, need_bytes;
2626         xlog_ticket_t   *ntic;
2627 #ifdef DEBUG
2628         xfs_lsn_t       tail_lsn;
2629 #endif
2630
2631         tic->t_curr_res = tic->t_unit_res;
2632         xlog_tic_reset_res(tic);
2633
2634         if (tic->t_cnt > 0)
2635                 return 0;
2636
2637 #ifdef DEBUG
2638         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2639                 panic("regrant Recovery problem");
2640 #endif
2641
2642         spin_lock(&log->l_grant_lock);
2643         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2644
2645         if (XLOG_FORCED_SHUTDOWN(log))
2646                 goto error_return;
2647
2648         /* If there are other waiters on the queue then give them a
2649          * chance at logspace before us. Wake up the first waiters,
2650          * if we do not wake up all the waiters then go to sleep waiting
2651          * for more free space, otherwise try to get some space for
2652          * this transaction.
2653          */
2654         need_bytes = tic->t_unit_res;
2655         if ((ntic = log->l_write_headq)) {
2656                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2657                                              log->l_grant_write_bytes);
2658                 do {
2659                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2660
2661                         if (free_bytes < ntic->t_unit_res)
2662                                 break;
2663                         free_bytes -= ntic->t_unit_res;
2664                         sv_signal(&ntic->t_wait);
2665                         ntic = ntic->t_next;
2666                 } while (ntic != log->l_write_headq);
2667
2668                 if (ntic != log->l_write_headq) {
2669                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2670                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2671
2672                         xlog_trace_loggrant(log, tic,
2673                                     "xlog_regrant_write_log_space: sleep 1");
2674                         spin_unlock(&log->l_grant_lock);
2675                         xlog_grant_push_ail(log->l_mp, need_bytes);
2676                         spin_lock(&log->l_grant_lock);
2677
2678                         XFS_STATS_INC(xs_sleep_logspace);
2679                         sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2680                                 &log->l_grant_lock, s);
2681
2682                         /* If we're shutting down, this tic is already
2683                          * off the queue */
2684                         spin_lock(&log->l_grant_lock);
2685                         if (XLOG_FORCED_SHUTDOWN(log))
2686                                 goto error_return;
2687
2688                         xlog_trace_loggrant(log, tic,
2689                                     "xlog_regrant_write_log_space: wake 1");
2690                 }
2691         }
2692
2693 redo:
2694         if (XLOG_FORCED_SHUTDOWN(log))
2695                 goto error_return;
2696
2697         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2698                                      log->l_grant_write_bytes);
2699         if (free_bytes < need_bytes) {
2700                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2701                         xlog_ins_ticketq(&log->l_write_headq, tic);
2702                 spin_unlock(&log->l_grant_lock);
2703                 xlog_grant_push_ail(log->l_mp, need_bytes);
2704                 spin_lock(&log->l_grant_lock);
2705
2706                 XFS_STATS_INC(xs_sleep_logspace);
2707                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2708
2709                 /* If we're shutting down, this tic is already off the queue */
2710                 spin_lock(&log->l_grant_lock);
2711                 if (XLOG_FORCED_SHUTDOWN(log))
2712                         goto error_return;
2713
2714                 xlog_trace_loggrant(log, tic,
2715                                     "xlog_regrant_write_log_space: wake 2");
2716                 goto redo;
2717         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2718                 xlog_del_ticketq(&log->l_write_headq, tic);
2719
2720         /* we've got enough space */
2721         xlog_grant_add_space_write(log, need_bytes);
2722 #ifdef DEBUG
2723         tail_lsn = log->l_tail_lsn;
2724         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2725                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2726                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2727         }
2728 #endif
2729
2730         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2731         xlog_verify_grant_head(log, 1);
2732         spin_unlock(&log->l_grant_lock);
2733         return 0;
2734
2735
2736  error_return:
2737         if (tic->t_flags & XLOG_TIC_IN_Q)
2738                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2739         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2740         /*
2741          * If we are failing, make sure the ticket doesn't have any
2742          * current reservations. We don't want to add this back when
2743          * the ticket/transaction gets cancelled.
2744          */
2745         tic->t_curr_res = 0;
2746         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2747         spin_unlock(&log->l_grant_lock);
2748         return XFS_ERROR(EIO);
2749 }       /* xlog_regrant_write_log_space */
2750
2751
2752 /* The first cnt-1 times through here we don't need to
2753  * move the grant write head because the permanent
2754  * reservation has reserved cnt times the unit amount.
2755  * Release part of current permanent unit reservation and
2756  * reset current reservation to be one units worth.  Also
2757  * move grant reservation head forward.
2758  */
2759 STATIC void
2760 xlog_regrant_reserve_log_space(xlog_t        *log,
2761                                xlog_ticket_t *ticket)
2762 {
2763         xlog_trace_loggrant(log, ticket,
2764                             "xlog_regrant_reserve_log_space: enter");
2765         if (ticket->t_cnt > 0)
2766                 ticket->t_cnt--;
2767
2768         spin_lock(&log->l_grant_lock);
2769         xlog_grant_sub_space(log, ticket->t_curr_res);
2770         ticket->t_curr_res = ticket->t_unit_res;
2771         xlog_tic_reset_res(ticket);
2772         xlog_trace_loggrant(log, ticket,
2773                             "xlog_regrant_reserve_log_space: sub current res");
2774         xlog_verify_grant_head(log, 1);
2775
2776         /* just return if we still have some of the pre-reserved space */
2777         if (ticket->t_cnt > 0) {
2778                 spin_unlock(&log->l_grant_lock);
2779                 return;
2780         }
2781
2782         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2783         xlog_trace_loggrant(log, ticket,
2784                             "xlog_regrant_reserve_log_space: exit");
2785         xlog_verify_grant_head(log, 0);
2786         spin_unlock(&log->l_grant_lock);
2787         ticket->t_curr_res = ticket->t_unit_res;
2788         xlog_tic_reset_res(ticket);
2789 }       /* xlog_regrant_reserve_log_space */
2790
2791
2792 /*
2793  * Give back the space left from a reservation.
2794  *
2795  * All the information we need to make a correct determination of space left
2796  * is present.  For non-permanent reservations, things are quite easy.  The
2797  * count should have been decremented to zero.  We only need to deal with the
2798  * space remaining in the current reservation part of the ticket.  If the
2799  * ticket contains a permanent reservation, there may be left over space which
2800  * needs to be released.  A count of N means that N-1 refills of the current
2801  * reservation can be done before we need to ask for more space.  The first
2802  * one goes to fill up the first current reservation.  Once we run out of
2803  * space, the count will stay at zero and the only space remaining will be
2804  * in the current reservation field.
2805  */
2806 STATIC void
2807 xlog_ungrant_log_space(xlog_t        *log,
2808                        xlog_ticket_t *ticket)
2809 {
2810         if (ticket->t_cnt > 0)
2811                 ticket->t_cnt--;
2812
2813         spin_lock(&log->l_grant_lock);
2814         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2815
2816         xlog_grant_sub_space(log, ticket->t_curr_res);
2817
2818         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2819
2820         /* If this is a permanent reservation ticket, we may be able to free
2821          * up more space based on the remaining count.
2822          */
2823         if (ticket->t_cnt > 0) {
2824                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2825                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2826         }
2827
2828         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2829         xlog_verify_grant_head(log, 1);
2830         spin_unlock(&log->l_grant_lock);
2831         xfs_log_move_tail(log->l_mp, 1);
2832 }       /* xlog_ungrant_log_space */
2833
2834
2835 /*
2836  * Flush iclog to disk if this is the last reference to the given iclog and
2837  * the WANT_SYNC bit is set.
2838  *
2839  * When this function is entered, the iclog is not necessarily in the
2840  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2841  *
2842  *
2843  */
2844 STATIC int
2845 xlog_state_release_iclog(
2846         xlog_t          *log,
2847         xlog_in_core_t  *iclog)
2848 {
2849         int             sync = 0;       /* do we sync? */
2850
2851         if (iclog->ic_state & XLOG_STATE_IOERROR)
2852                 return XFS_ERROR(EIO);
2853
2854         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2855         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2856                 return 0;
2857
2858         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2859                 spin_unlock(&log->l_icloglock);
2860                 return XFS_ERROR(EIO);
2861         }
2862         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2863                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2864
2865         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2866                 /* update tail before writing to iclog */
2867                 xlog_assign_tail_lsn(log->l_mp);
2868                 sync++;
2869                 iclog->ic_state = XLOG_STATE_SYNCING;
2870                 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2871                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2872                 /* cycle incremented when incrementing curr_block */
2873         }
2874         spin_unlock(&log->l_icloglock);
2875
2876         /*
2877          * We let the log lock go, so it's possible that we hit a log I/O
2878          * error or some other SHUTDOWN condition that marks the iclog
2879          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2880          * this iclog has consistent data, so we ignore IOERROR
2881          * flags after this point.
2882          */
2883         if (sync)
2884                 return xlog_sync(log, iclog);
2885         return 0;
2886 }       /* xlog_state_release_iclog */
2887
2888
2889 /*
2890  * This routine will mark the current iclog in the ring as WANT_SYNC
2891  * and move the current iclog pointer to the next iclog in the ring.
2892  * When this routine is called from xlog_state_get_iclog_space(), the
2893  * exact size of the iclog has not yet been determined.  All we know is
2894  * that every data block.  We have run out of space in this log record.
2895  */
2896 STATIC void
2897 xlog_state_switch_iclogs(xlog_t         *log,
2898                          xlog_in_core_t *iclog,
2899                          int            eventual_size)
2900 {
2901         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2902         if (!eventual_size)
2903                 eventual_size = iclog->ic_offset;
2904         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2905         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2906         log->l_prev_block = log->l_curr_block;
2907         log->l_prev_cycle = log->l_curr_cycle;
2908
2909         /* roll log?: ic_offset changed later */
2910         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2911
2912         /* Round up to next log-sunit */
2913         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2914             log->l_mp->m_sb.sb_logsunit > 1) {
2915                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2916                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2917         }
2918
2919         if (log->l_curr_block >= log->l_logBBsize) {
2920                 log->l_curr_cycle++;
2921                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2922                         log->l_curr_cycle++;
2923                 log->l_curr_block -= log->l_logBBsize;
2924                 ASSERT(log->l_curr_block >= 0);
2925         }
2926         ASSERT(iclog == log->l_iclog);
2927         log->l_iclog = iclog->ic_next;
2928 }       /* xlog_state_switch_iclogs */
2929
2930
2931 /*
2932  * Write out all data in the in-core log as of this exact moment in time.
2933  *
2934  * Data may be written to the in-core log during this call.  However,
2935  * we don't guarantee this data will be written out.  A change from past
2936  * implementation means this routine will *not* write out zero length LRs.
2937  *
2938  * Basically, we try and perform an intelligent scan of the in-core logs.
2939  * If we determine there is no flushable data, we just return.  There is no
2940  * flushable data if:
2941  *
2942  *      1. the current iclog is active and has no data; the previous iclog
2943  *              is in the active or dirty state.
2944  *      2. the current iclog is drity, and the previous iclog is in the
2945  *              active or dirty state.
2946  *
2947  * We may sleep if:
2948  *
2949  *      1. the current iclog is not in the active nor dirty state.
2950  *      2. the current iclog dirty, and the previous iclog is not in the
2951  *              active nor dirty state.
2952  *      3. the current iclog is active, and there is another thread writing
2953  *              to this particular iclog.
2954  *      4. a) the current iclog is active and has no other writers
2955  *         b) when we return from flushing out this iclog, it is still
2956  *              not in the active nor dirty state.
2957  */
2958 STATIC int
2959 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2960 {
2961         xlog_in_core_t  *iclog;
2962         xfs_lsn_t       lsn;
2963
2964         spin_lock(&log->l_icloglock);
2965
2966         iclog = log->l_iclog;
2967         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2968                 spin_unlock(&log->l_icloglock);
2969                 return XFS_ERROR(EIO);
2970         }
2971
2972         /* If the head iclog is not active nor dirty, we just attach
2973          * ourselves to the head and go to sleep.
2974          */
2975         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2976             iclog->ic_state == XLOG_STATE_DIRTY) {
2977                 /*
2978                  * If the head is dirty or (active and empty), then
2979                  * we need to look at the previous iclog.  If the previous
2980                  * iclog is active or dirty we are done.  There is nothing
2981                  * to sync out.  Otherwise, we attach ourselves to the
2982                  * previous iclog and go to sleep.
2983                  */
2984                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2985                     (atomic_read(&iclog->ic_refcnt) == 0
2986                      && iclog->ic_offset == 0)) {
2987                         iclog = iclog->ic_prev;
2988                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2989                             iclog->ic_state == XLOG_STATE_DIRTY)
2990                                 goto no_sleep;
2991                         else
2992                                 goto maybe_sleep;
2993                 } else {
2994                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2995                                 /* We are the only one with access to this
2996                                  * iclog.  Flush it out now.  There should
2997                                  * be a roundoff of zero to show that someone
2998                                  * has already taken care of the roundoff from
2999                                  * the previous sync.
3000                                  */
3001                                 atomic_inc(&iclog->ic_refcnt);
3002                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3003                                 xlog_state_switch_iclogs(log, iclog, 0);
3004                                 spin_unlock(&log->l_icloglock);
3005
3006                                 if (xlog_state_release_iclog(log, iclog))
3007                                         return XFS_ERROR(EIO);
3008                                 *log_flushed = 1;
3009                                 spin_lock(&log->l_icloglock);
3010                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3011                                     iclog->ic_state != XLOG_STATE_DIRTY)
3012                                         goto maybe_sleep;
3013                                 else
3014                                         goto no_sleep;
3015                         } else {
3016                                 /* Someone else is writing to this iclog.
3017                                  * Use its call to flush out the data.  However,
3018                                  * the other thread may not force out this LR,
3019                                  * so we mark it WANT_SYNC.
3020                                  */
3021                                 xlog_state_switch_iclogs(log, iclog, 0);
3022                                 goto maybe_sleep;
3023                         }
3024                 }
3025         }
3026
3027         /* By the time we come around again, the iclog could've been filled
3028          * which would give it another lsn.  If we have a new lsn, just
3029          * return because the relevant data has been flushed.
3030          */
3031 maybe_sleep:
3032         if (flags & XFS_LOG_SYNC) {
3033                 /*
3034                  * We must check if we're shutting down here, before
3035                  * we wait, while we're holding the l_icloglock.
3036                  * Then we check again after waking up, in case our
3037                  * sleep was disturbed by a bad news.
3038                  */
3039                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3040                         spin_unlock(&log->l_icloglock);
3041                         return XFS_ERROR(EIO);
3042                 }
3043                 XFS_STATS_INC(xs_log_force_sleep);
3044                 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3045                 /*
3046                  * No need to grab the log lock here since we're
3047                  * only deciding whether or not to return EIO
3048                  * and the memory read should be atomic.
3049                  */
3050                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3051                         return XFS_ERROR(EIO);
3052                 *log_flushed = 1;
3053
3054         } else {
3055
3056 no_sleep:
3057                 spin_unlock(&log->l_icloglock);
3058         }
3059         return 0;
3060 }       /* xlog_state_sync_all */
3061
3062
3063 /*
3064  * Used by code which implements synchronous log forces.
3065  *
3066  * Find in-core log with lsn.
3067  *      If it is in the DIRTY state, just return.
3068  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3069  *              state and go to sleep or return.
3070  *      If it is in any other state, go to sleep or return.
3071  *
3072  * If filesystem activity goes to zero, the iclog will get flushed only by
3073  * bdflush().
3074  */
3075 STATIC int
3076 xlog_state_sync(xlog_t    *log,
3077                 xfs_lsn_t lsn,
3078                 uint      flags,
3079                 int       *log_flushed)
3080 {
3081     xlog_in_core_t      *iclog;
3082     int                 already_slept = 0;
3083
3084 try_again:
3085     spin_lock(&log->l_icloglock);
3086     iclog = log->l_iclog;
3087
3088     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3089             spin_unlock(&log->l_icloglock);
3090             return XFS_ERROR(EIO);
3091     }
3092
3093     do {
3094         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3095                 iclog = iclog->ic_next;
3096                 continue;
3097         }
3098
3099         if (iclog->ic_state == XLOG_STATE_DIRTY) {
3100                 spin_unlock(&log->l_icloglock);
3101                 return 0;
3102         }
3103
3104         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3105                 /*
3106                  * We sleep here if we haven't already slept (e.g.
3107                  * this is the first time we've looked at the correct
3108                  * iclog buf) and the buffer before us is going to
3109                  * be sync'ed. The reason for this is that if we
3110                  * are doing sync transactions here, by waiting for
3111                  * the previous I/O to complete, we can allow a few
3112                  * more transactions into this iclog before we close
3113                  * it down.
3114                  *
3115                  * Otherwise, we mark the buffer WANT_SYNC, and bump
3116                  * up the refcnt so we can release the log (which drops
3117                  * the ref count).  The state switch keeps new transaction
3118                  * commits from using this buffer.  When the current commits
3119                  * finish writing into the buffer, the refcount will drop to
3120                  * zero and the buffer will go out then.
3121                  */
3122                 if (!already_slept &&
3123                     (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3124                                                  XLOG_STATE_SYNCING))) {
3125                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3126                         XFS_STATS_INC(xs_log_force_sleep);
3127                         sv_wait(&iclog->ic_prev->ic_write_wait, PSWP,
3128                                 &log->l_icloglock, s);
3129                         *log_flushed = 1;
3130                         already_slept = 1;
3131                         goto try_again;
3132                 } else {
3133                         atomic_inc(&iclog->ic_refcnt);
3134                         xlog_state_switch_iclogs(log, iclog, 0);
3135                         spin_unlock(&log->l_icloglock);
3136                         if (xlog_state_release_iclog(log, iclog))
3137                                 return XFS_ERROR(EIO);
3138                         *log_flushed = 1;
3139                         spin_lock(&log->l_icloglock);
3140                 }