Merge rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / 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_dir.h"
28 #include "xfs_dir2.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_error.h"
32 #include "xfs_log_priv.h"
33 #include "xfs_buf_item.h"
34 #include "xfs_bmap_btree.h"
35 #include "xfs_alloc_btree.h"
36 #include "xfs_ialloc_btree.h"
37 #include "xfs_log_recover.h"
38 #include "xfs_trans_priv.h"
39 #include "xfs_dir_sf.h"
40 #include "xfs_dir2_sf.h"
41 #include "xfs_attr_sf.h"
42 #include "xfs_dinode.h"
43 #include "xfs_inode.h"
44 #include "xfs_rw.h"
45
46
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
48         { (ptr) += (bytes); \
49           (len) -= (bytes); \
50           (off) += (bytes);}
51
52 /* Local miscellaneous function prototypes */
53 STATIC int       xlog_bdstrat_cb(struct xfs_buf *);
54 STATIC int       xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
55                                     xlog_in_core_t **, xfs_lsn_t *);
56 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
57                                 xfs_buftarg_t   *log_target,
58                                 xfs_daddr_t     blk_offset,
59                                 int             num_bblks);
60 STATIC int       xlog_space_left(xlog_t *log, int cycle, int bytes);
61 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
62 STATIC void      xlog_unalloc_log(xlog_t *log);
63 STATIC int       xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
64                             int nentries, xfs_log_ticket_t tic,
65                             xfs_lsn_t *start_lsn,
66                             xlog_in_core_t **commit_iclog,
67                             uint flags);
68
69 /* local state machine functions */
70 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
71 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
72 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
73                                        int              len,
74                                        xlog_in_core_t   **iclog,
75                                        xlog_ticket_t    *ticket,
76                                        int              *continued_write,
77                                        int              *logoffsetp);
78 STATIC void xlog_state_put_ticket(xlog_t        *log,
79                                   xlog_ticket_t *tic);
80 STATIC int  xlog_state_release_iclog(xlog_t             *log,
81                                      xlog_in_core_t     *iclog);
82 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
83                                      xlog_in_core_t *iclog,
84                                      int                eventual_size);
85 STATIC int  xlog_state_sync(xlog_t                      *log,
86                             xfs_lsn_t                   lsn,
87                             uint                        flags,
88                             int                         *log_flushed);
89 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
90 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
91
92 /* local functions to manipulate grant head */
93 STATIC int  xlog_grant_log_space(xlog_t         *log,
94                                  xlog_ticket_t  *xtic);
95 STATIC void xlog_grant_push_ail(xfs_mount_t     *mp,
96                                 int             need_bytes);
97 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
98                                            xlog_ticket_t *ticket);
99 STATIC int xlog_regrant_write_log_space(xlog_t          *log,
100                                          xlog_ticket_t  *ticket);
101 STATIC void xlog_ungrant_log_space(xlog_t        *log,
102                                    xlog_ticket_t *ticket);
103
104
105 /* local ticket functions */
106 STATIC void             xlog_state_ticket_alloc(xlog_t *log);
107 STATIC xlog_ticket_t    *xlog_ticket_get(xlog_t *log,
108                                          int    unit_bytes,
109                                          int    count,
110                                          char   clientid,
111                                          uint   flags);
112 STATIC void             xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
113
114 #if defined(DEBUG)
115 STATIC void     xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
116 STATIC void     xlog_verify_grant_head(xlog_t *log, int equals);
117 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
118                                   int count, boolean_t syncing);
119 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
120                                      xfs_lsn_t tail_lsn);
121 #else
122 #define xlog_verify_dest_ptr(a,b)
123 #define xlog_verify_grant_head(a,b)
124 #define xlog_verify_iclog(a,b,c,d)
125 #define xlog_verify_tail_lsn(a,b,c)
126 #endif
127
128 STATIC int      xlog_iclogs_empty(xlog_t *log);
129
130 #if defined(XFS_LOG_TRACE)
131 void
132 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
133 {
134         unsigned long cnts;
135
136         if (!log->l_grant_trace) {
137                 log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
138                 if (!log->l_grant_trace)
139                         return;
140         }
141         /* ticket counts are 1 byte each */
142         cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
143
144         ktrace_enter(log->l_grant_trace,
145                      (void *)tic,
146                      (void *)log->l_reserve_headq,
147                      (void *)log->l_write_headq,
148                      (void *)((unsigned long)log->l_grant_reserve_cycle),
149                      (void *)((unsigned long)log->l_grant_reserve_bytes),
150                      (void *)((unsigned long)log->l_grant_write_cycle),
151                      (void *)((unsigned long)log->l_grant_write_bytes),
152                      (void *)((unsigned long)log->l_curr_cycle),
153                      (void *)((unsigned long)log->l_curr_block),
154                      (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
155                      (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
156                      (void *)string,
157                      (void *)((unsigned long)tic->t_trans_type),
158                      (void *)cnts,
159                      (void *)((unsigned long)tic->t_curr_res),
160                      (void *)((unsigned long)tic->t_unit_res));
161 }
162
163 void
164 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
165 {
166         if (!iclog->ic_trace)
167                 iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
168         ktrace_enter(iclog->ic_trace,
169                      (void *)((unsigned long)state),
170                      (void *)((unsigned long)current_pid()),
171                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
173                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
174                      (void *)NULL, (void *)NULL);
175 }
176 #else
177 #define xlog_trace_loggrant(log,tic,string)
178 #define xlog_trace_iclog(iclog,state)
179 #endif /* XFS_LOG_TRACE */
180
181
182 static void
183 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
184 {
185         if (*qp) {
186                 tic->t_next         = (*qp);
187                 tic->t_prev         = (*qp)->t_prev;
188                 (*qp)->t_prev->t_next = tic;
189                 (*qp)->t_prev       = tic;
190         } else {
191                 tic->t_prev = tic->t_next = tic;
192                 *qp = tic;
193         }
194
195         tic->t_flags |= XLOG_TIC_IN_Q;
196 }
197
198 static void
199 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
200 {
201         if (tic == tic->t_next) {
202                 *qp = NULL;
203         } else {
204                 *qp = tic->t_next;
205                 tic->t_next->t_prev = tic->t_prev;
206                 tic->t_prev->t_next = tic->t_next;
207         }
208
209         tic->t_next = tic->t_prev = NULL;
210         tic->t_flags &= ~XLOG_TIC_IN_Q;
211 }
212
213 static void
214 xlog_grant_sub_space(struct log *log, int bytes)
215 {
216         log->l_grant_write_bytes -= bytes;
217         if (log->l_grant_write_bytes < 0) {
218                 log->l_grant_write_bytes += log->l_logsize;
219                 log->l_grant_write_cycle--;
220         }
221
222         log->l_grant_reserve_bytes -= bytes;
223         if ((log)->l_grant_reserve_bytes < 0) {
224                 log->l_grant_reserve_bytes += log->l_logsize;
225                 log->l_grant_reserve_cycle--;
226         }
227
228 }
229
230 static void
231 xlog_grant_add_space_write(struct log *log, int bytes)
232 {
233         log->l_grant_write_bytes += bytes;
234         if (log->l_grant_write_bytes > log->l_logsize) {
235                 log->l_grant_write_bytes -= log->l_logsize;
236                 log->l_grant_write_cycle++;
237         }
238 }
239
240 static void
241 xlog_grant_add_space_reserve(struct log *log, int bytes)
242 {
243         log->l_grant_reserve_bytes += bytes;
244         if (log->l_grant_reserve_bytes > log->l_logsize) {
245                 log->l_grant_reserve_bytes -= log->l_logsize;
246                 log->l_grant_reserve_cycle++;
247         }
248 }
249
250 static inline void
251 xlog_grant_add_space(struct log *log, int bytes)
252 {
253         xlog_grant_add_space_write(log, bytes);
254         xlog_grant_add_space_reserve(log, bytes);
255 }
256
257
258 /*
259  * NOTES:
260  *
261  *      1. currblock field gets updated at startup and after in-core logs
262  *              marked as with WANT_SYNC.
263  */
264
265 /*
266  * This routine is called when a user of a log manager ticket is done with
267  * the reservation.  If the ticket was ever used, then a commit record for
268  * the associated transaction is written out as a log operation header with
269  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
270  * a given ticket.  If the ticket was one with a permanent reservation, then
271  * a few operations are done differently.  Permanent reservation tickets by
272  * default don't release the reservation.  They just commit the current
273  * transaction with the belief that the reservation is still needed.  A flag
274  * must be passed in before permanent reservations are actually released.
275  * When these type of tickets are not released, they need to be set into
276  * the inited state again.  By doing this, a start record will be written
277  * out when the next write occurs.
278  */
279 xfs_lsn_t
280 xfs_log_done(xfs_mount_t        *mp,
281              xfs_log_ticket_t   xtic,
282              void               **iclog,
283              uint               flags)
284 {
285         xlog_t          *log    = mp->m_log;
286         xlog_ticket_t   *ticket = (xfs_log_ticket_t) xtic;
287         xfs_lsn_t       lsn     = 0;
288
289         if (XLOG_FORCED_SHUTDOWN(log) ||
290             /*
291              * If nothing was ever written, don't write out commit record.
292              * If we get an error, just continue and give back the log ticket.
293              */
294             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
295              (xlog_commit_record(mp, ticket,
296                                  (xlog_in_core_t **)iclog, &lsn)))) {
297                 lsn = (xfs_lsn_t) -1;
298                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
299                         flags |= XFS_LOG_REL_PERM_RESERV;
300                 }
301         }
302
303
304         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
305             (flags & XFS_LOG_REL_PERM_RESERV)) {
306                 /*
307                  * Release ticket if not permanent reservation or a specifc
308                  * request has been made to release a permanent reservation.
309                  */
310                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
311                 xlog_ungrant_log_space(log, ticket);
312                 xlog_state_put_ticket(log, ticket);
313         } else {
314                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
315                 xlog_regrant_reserve_log_space(log, ticket);
316         }
317
318         /* If this ticket was a permanent reservation and we aren't
319          * trying to release it, reset the inited flags; so next time
320          * we write, a start record will be written out.
321          */
322         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
323             (flags & XFS_LOG_REL_PERM_RESERV) == 0)
324                 ticket->t_flags |= XLOG_TIC_INITED;
325
326         return lsn;
327 }       /* xfs_log_done */
328
329
330 /*
331  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
332  *      the force is done synchronously.
333  *
334  * Asynchronous forces are implemented by setting the WANT_SYNC
335  * bit in the appropriate in-core log and then returning.
336  *
337  * Synchronous forces are implemented with a semaphore.  All callers
338  * to force a given lsn to disk will wait on a semaphore attached to the
339  * specific in-core log.  When given in-core log finally completes its
340  * write to disk, that thread will wake up all threads waiting on the
341  * semaphore.
342  */
343 int
344 _xfs_log_force(
345         xfs_mount_t     *mp,
346         xfs_lsn_t       lsn,
347         uint            flags,
348         int             *log_flushed)
349 {
350         xlog_t          *log = mp->m_log;
351         int             dummy;
352
353         if (!log_flushed)
354                 log_flushed = &dummy;
355
356         ASSERT(flags & XFS_LOG_FORCE);
357
358         XFS_STATS_INC(xs_log_force);
359
360         if (log->l_flags & XLOG_IO_ERROR)
361                 return XFS_ERROR(EIO);
362         if (lsn == 0)
363                 return xlog_state_sync_all(log, flags, log_flushed);
364         else
365                 return xlog_state_sync(log, lsn, flags, log_flushed);
366 }       /* xfs_log_force */
367
368 /*
369  * Attaches a new iclog I/O completion callback routine during
370  * transaction commit.  If the log is in error state, a non-zero
371  * return code is handed back and the caller is responsible for
372  * executing the callback at an appropriate time.
373  */
374 int
375 xfs_log_notify(xfs_mount_t        *mp,          /* mount of partition */
376                void               *iclog_hndl,  /* iclog to hang callback off */
377                xfs_log_callback_t *cb)
378 {
379         xlog_t *log = mp->m_log;
380         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
381         int     abortflg, spl;
382
383         cb->cb_next = NULL;
384         spl = LOG_LOCK(log);
385         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
386         if (!abortflg) {
387                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
388                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
389                 cb->cb_next = NULL;
390                 *(iclog->ic_callback_tail) = cb;
391                 iclog->ic_callback_tail = &(cb->cb_next);
392         }
393         LOG_UNLOCK(log, spl);
394         return abortflg;
395 }       /* xfs_log_notify */
396
397 int
398 xfs_log_release_iclog(xfs_mount_t *mp,
399                       void        *iclog_hndl)
400 {
401         xlog_t *log = mp->m_log;
402         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
403
404         if (xlog_state_release_iclog(log, iclog)) {
405                 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
406                 return EIO;
407         }
408
409         return 0;
410 }
411
412 /*
413  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
414  *      to the reservation.
415  *  2. Potentially, push buffers at tail of log to disk.
416  *
417  * Each reservation is going to reserve extra space for a log record header.
418  * When writes happen to the on-disk log, we don't subtract the length of the
419  * log record header from any reservation.  By wasting space in each
420  * reservation, we prevent over allocation problems.
421  */
422 int
423 xfs_log_reserve(xfs_mount_t      *mp,
424                 int              unit_bytes,
425                 int              cnt,
426                 xfs_log_ticket_t *ticket,
427                 __uint8_t        client,
428                 uint             flags,
429                 uint             t_type)
430 {
431         xlog_t          *log = mp->m_log;
432         xlog_ticket_t   *internal_ticket;
433         int             retval = 0;
434
435         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
436         ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
437
438         if (XLOG_FORCED_SHUTDOWN(log))
439                 return XFS_ERROR(EIO);
440
441         XFS_STATS_INC(xs_try_logspace);
442
443         if (*ticket != NULL) {
444                 ASSERT(flags & XFS_LOG_PERM_RESERV);
445                 internal_ticket = (xlog_ticket_t *)*ticket;
446                 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
447                 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
448                 retval = xlog_regrant_write_log_space(log, internal_ticket);
449         } else {
450                 /* may sleep if need to allocate more tickets */
451                 internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
452                                                   client, flags);
453                 internal_ticket->t_trans_type = t_type;
454                 *ticket = internal_ticket;
455                 xlog_trace_loggrant(log, internal_ticket, 
456                         (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
457                         "xfs_log_reserve: create new ticket (permanent trans)" :
458                         "xfs_log_reserve: create new ticket");
459                 xlog_grant_push_ail(mp,
460                                     (internal_ticket->t_unit_res *
461                                      internal_ticket->t_cnt));
462                 retval = xlog_grant_log_space(log, internal_ticket);
463         }
464
465         return retval;
466 }       /* xfs_log_reserve */
467
468
469 /*
470  * Mount a log filesystem
471  *
472  * mp           - ubiquitous xfs mount point structure
473  * log_target   - buftarg of on-disk log device
474  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
475  * num_bblocks  - Number of BBSIZE blocks in on-disk log
476  *
477  * Return error or zero.
478  */
479 int
480 xfs_log_mount(xfs_mount_t       *mp,
481               xfs_buftarg_t     *log_target,
482               xfs_daddr_t       blk_offset,
483               int               num_bblks)
484 {
485         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
486                 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
487         else {
488                 cmn_err(CE_NOTE,
489                         "!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
490                         mp->m_fsname);
491                 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
492         }
493
494         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
495
496         /*
497          * skip log recovery on a norecovery mount.  pretend it all
498          * just worked.
499          */
500         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
501                 int     error;
502                 vfs_t   *vfsp = XFS_MTOVFS(mp);
503                 int     readonly = (vfsp->vfs_flag & VFS_RDONLY);
504
505                 if (readonly)
506                         vfsp->vfs_flag &= ~VFS_RDONLY;
507
508                 error = xlog_recover(mp->m_log);
509
510                 if (readonly)
511                         vfsp->vfs_flag |= VFS_RDONLY;
512                 if (error) {
513                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
514                         xlog_unalloc_log(mp->m_log);
515                         return error;
516                 }
517         }
518
519         /* Normal transactions can now occur */
520         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
521
522         /* End mounting message in xfs_log_mount_finish */
523         return 0;
524 }       /* xfs_log_mount */
525
526 /*
527  * Finish the recovery of the file system.  This is separate from
528  * the xfs_log_mount() call, because it depends on the code in
529  * xfs_mountfs() to read in the root and real-time bitmap inodes
530  * between calling xfs_log_mount() and here.
531  *
532  * mp           - ubiquitous xfs mount point structure
533  */
534 int
535 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
536 {
537         int     error;
538
539         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
540                 error = xlog_recover_finish(mp->m_log, mfsi_flags);
541         else {
542                 error = 0;
543                 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
544         }
545
546         return error;
547 }
548
549 /*
550  * Unmount processing for the log.
551  */
552 int
553 xfs_log_unmount(xfs_mount_t *mp)
554 {
555         int             error;
556
557         error = xfs_log_unmount_write(mp);
558         xfs_log_unmount_dealloc(mp);
559         return error;
560 }
561
562 /*
563  * Final log writes as part of unmount.
564  *
565  * Mark the filesystem clean as unmount happens.  Note that during relocation
566  * this routine needs to be executed as part of source-bag while the
567  * deallocation must not be done until source-end.
568  */
569
570 /*
571  * Unmount record used to have a string "Unmount filesystem--" in the
572  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
573  * We just write the magic number now since that particular field isn't
574  * currently architecture converted and "nUmount" is a bit foo.
575  * As far as I know, there weren't any dependencies on the old behaviour.
576  */
577
578 int
579 xfs_log_unmount_write(xfs_mount_t *mp)
580 {
581         xlog_t           *log = mp->m_log;
582         xlog_in_core_t   *iclog;
583 #ifdef DEBUG
584         xlog_in_core_t   *first_iclog;
585 #endif
586         xfs_log_iovec_t  reg[1];
587         xfs_log_ticket_t tic = NULL;
588         xfs_lsn_t        lsn;
589         int              error;
590         SPLDECL(s);
591
592         /* the data section must be 32 bit size aligned */
593         struct {
594             __uint16_t magic;
595             __uint16_t pad1;
596             __uint32_t pad2; /* may as well make it 64 bits */
597         } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
598
599         /*
600          * Don't write out unmount record on read-only mounts.
601          * Or, if we are doing a forced umount (typically because of IO errors).
602          */
603         if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
604                 return 0;
605
606         xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
607
608 #ifdef DEBUG
609         first_iclog = iclog = log->l_iclog;
610         do {
611                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
612                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
613                         ASSERT(iclog->ic_offset == 0);
614                 }
615                 iclog = iclog->ic_next;
616         } while (iclog != first_iclog);
617 #endif
618         if (! (XLOG_FORCED_SHUTDOWN(log))) {
619                 reg[0].i_addr = (void*)&magic;
620                 reg[0].i_len  = sizeof(magic);
621                 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
622
623                 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
624                 if (!error) {
625                         /* remove inited flag */
626                         ((xlog_ticket_t *)tic)->t_flags = 0;
627                         error = xlog_write(mp, reg, 1, tic, &lsn,
628                                            NULL, XLOG_UNMOUNT_TRANS);
629                         /*
630                          * At this point, we're umounting anyway,
631                          * so there's no point in transitioning log state
632                          * to IOERROR. Just continue...
633                          */
634                 }
635
636                 if (error) {
637                         xfs_fs_cmn_err(CE_ALERT, mp,
638                                 "xfs_log_unmount: unmount record failed");
639                 }
640
641
642                 s = LOG_LOCK(log);
643                 iclog = log->l_iclog;
644                 iclog->ic_refcnt++;
645                 LOG_UNLOCK(log, s);
646                 xlog_state_want_sync(log, iclog);
647                 (void) xlog_state_release_iclog(log, iclog);
648
649                 s = LOG_LOCK(log);
650                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
651                       iclog->ic_state == XLOG_STATE_DIRTY)) {
652                         if (!XLOG_FORCED_SHUTDOWN(log)) {
653                                 sv_wait(&iclog->ic_forcesema, PMEM,
654                                         &log->l_icloglock, s);
655                         } else {
656                                 LOG_UNLOCK(log, s);
657                         }
658                 } else {
659                         LOG_UNLOCK(log, s);
660                 }
661                 if (tic)
662                         xlog_state_put_ticket(log, tic);
663         } else {
664                 /*
665                  * We're already in forced_shutdown mode, couldn't
666                  * even attempt to write out the unmount transaction.
667                  *
668                  * Go through the motions of sync'ing and releasing
669                  * the iclog, even though no I/O will actually happen,
670                  * we need to wait for other log I/O's that may already
671                  * be in progress.  Do this as a separate section of
672                  * code so we'll know if we ever get stuck here that
673                  * we're in this odd situation of trying to unmount
674                  * a file system that went into forced_shutdown as
675                  * the result of an unmount..
676                  */
677                 s = LOG_LOCK(log);
678                 iclog = log->l_iclog;
679                 iclog->ic_refcnt++;
680                 LOG_UNLOCK(log, s);
681
682                 xlog_state_want_sync(log, iclog);
683                 (void) xlog_state_release_iclog(log, iclog);
684
685                 s = LOG_LOCK(log);
686
687                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
688                         || iclog->ic_state == XLOG_STATE_DIRTY
689                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
690
691                                 sv_wait(&iclog->ic_forcesema, PMEM,
692                                         &log->l_icloglock, s);
693                 } else {
694                         LOG_UNLOCK(log, s);
695                 }
696         }
697
698         return 0;
699 }       /* xfs_log_unmount_write */
700
701 /*
702  * Deallocate log structures for unmount/relocation.
703  */
704 void
705 xfs_log_unmount_dealloc(xfs_mount_t *mp)
706 {
707         xlog_unalloc_log(mp->m_log);
708 }
709
710 /*
711  * Write region vectors to log.  The write happens using the space reservation
712  * of the ticket (tic).  It is not a requirement that all writes for a given
713  * transaction occur with one call to xfs_log_write().
714  */
715 int
716 xfs_log_write(xfs_mount_t *     mp,
717               xfs_log_iovec_t   reg[],
718               int               nentries,
719               xfs_log_ticket_t  tic,
720               xfs_lsn_t         *start_lsn)
721 {
722         int     error;
723         xlog_t *log = mp->m_log;
724
725         if (XLOG_FORCED_SHUTDOWN(log))
726                 return XFS_ERROR(EIO);
727
728         if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
729                 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
730         }
731         return error;
732 }       /* xfs_log_write */
733
734
735 void
736 xfs_log_move_tail(xfs_mount_t   *mp,
737                   xfs_lsn_t     tail_lsn)
738 {
739         xlog_ticket_t   *tic;
740         xlog_t          *log = mp->m_log;
741         int             need_bytes, free_bytes, cycle, bytes;
742         SPLDECL(s);
743
744         if (XLOG_FORCED_SHUTDOWN(log))
745                 return;
746         ASSERT(!XFS_FORCED_SHUTDOWN(mp));
747
748         if (tail_lsn == 0) {
749                 /* needed since sync_lsn is 64 bits */
750                 s = LOG_LOCK(log);
751                 tail_lsn = log->l_last_sync_lsn;
752                 LOG_UNLOCK(log, s);
753         }
754
755         s = GRANT_LOCK(log);
756
757         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
758          * tail_lsn.
759          */
760         if (tail_lsn != 1) {
761                 log->l_tail_lsn = tail_lsn;
762         }
763
764         if ((tic = log->l_write_headq)) {
765 #ifdef DEBUG
766                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
767                         panic("Recovery problem");
768 #endif
769                 cycle = log->l_grant_write_cycle;
770                 bytes = log->l_grant_write_bytes;
771                 free_bytes = xlog_space_left(log, cycle, bytes);
772                 do {
773                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
774
775                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
776                                 break;
777                         tail_lsn = 0;
778                         free_bytes -= tic->t_unit_res;
779                         sv_signal(&tic->t_sema);
780                         tic = tic->t_next;
781                 } while (tic != log->l_write_headq);
782         }
783         if ((tic = log->l_reserve_headq)) {
784 #ifdef DEBUG
785                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
786                         panic("Recovery problem");
787 #endif
788                 cycle = log->l_grant_reserve_cycle;
789                 bytes = log->l_grant_reserve_bytes;
790                 free_bytes = xlog_space_left(log, cycle, bytes);
791                 do {
792                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
793                                 need_bytes = tic->t_unit_res*tic->t_cnt;
794                         else
795                                 need_bytes = tic->t_unit_res;
796                         if (free_bytes < need_bytes && tail_lsn != 1)
797                                 break;
798                         tail_lsn = 0;
799                         free_bytes -= need_bytes;
800                         sv_signal(&tic->t_sema);
801                         tic = tic->t_next;
802                 } while (tic != log->l_reserve_headq);
803         }
804         GRANT_UNLOCK(log, s);
805 }       /* xfs_log_move_tail */
806
807 /*
808  * Determine if we have a transaction that has gone to disk
809  * that needs to be covered. Log activity needs to be idle (no AIL and
810  * nothing in the iclogs). And, we need to be in the right state indicating
811  * something has gone out.
812  */
813 int
814 xfs_log_need_covered(xfs_mount_t *mp)
815 {
816         SPLDECL(s);
817         int             needed = 0, gen;
818         xlog_t          *log = mp->m_log;
819         vfs_t           *vfsp = XFS_MTOVFS(mp);
820
821         if (fs_frozen(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
822             (vfsp->vfs_flag & VFS_RDONLY))
823                 return 0;
824
825         s = LOG_LOCK(log);
826         if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
827                 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
828                         && !xfs_trans_first_ail(mp, &gen)
829                         && xlog_iclogs_empty(log)) {
830                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
831                         log->l_covered_state = XLOG_STATE_COVER_DONE;
832                 else {
833                         ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
834                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
835                 }
836                 needed = 1;
837         }
838         LOG_UNLOCK(log, s);
839         return needed;
840 }
841
842 /******************************************************************************
843  *
844  *      local routines
845  *
846  ******************************************************************************
847  */
848
849 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
850  * The log manager must keep track of the last LR which was committed
851  * to disk.  The lsn of this LR will become the new tail_lsn whenever
852  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
853  * the situation where stuff could be written into the log but nothing
854  * was ever in the AIL when asked.  Eventually, we panic since the
855  * tail hits the head.
856  *
857  * We may be holding the log iclog lock upon entering this routine.
858  */
859 xfs_lsn_t
860 xlog_assign_tail_lsn(xfs_mount_t *mp)
861 {
862         xfs_lsn_t tail_lsn;
863         SPLDECL(s);
864         xlog_t    *log = mp->m_log;
865
866         tail_lsn = xfs_trans_tail_ail(mp);
867         s = GRANT_LOCK(log);
868         if (tail_lsn != 0) {
869                 log->l_tail_lsn = tail_lsn;
870         } else {
871                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
872         }
873         GRANT_UNLOCK(log, s);
874
875         return tail_lsn;
876 }       /* xlog_assign_tail_lsn */
877
878
879 /*
880  * Return the space in the log between the tail and the head.  The head
881  * is passed in the cycle/bytes formal parms.  In the special case where
882  * the reserve head has wrapped passed the tail, this calculation is no
883  * longer valid.  In this case, just return 0 which means there is no space
884  * in the log.  This works for all places where this function is called
885  * with the reserve head.  Of course, if the write head were to ever
886  * wrap the tail, we should blow up.  Rather than catch this case here,
887  * we depend on other ASSERTions in other parts of the code.   XXXmiken
888  *
889  * This code also handles the case where the reservation head is behind
890  * the tail.  The details of this case are described below, but the end
891  * result is that we return the size of the log as the amount of space left.
892  */
893 int
894 xlog_space_left(xlog_t *log, int cycle, int bytes)
895 {
896         int free_bytes;
897         int tail_bytes;
898         int tail_cycle;
899
900         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
901         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
902         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
903                 free_bytes = log->l_logsize - (bytes - tail_bytes);
904         } else if ((tail_cycle + 1) < cycle) {
905                 return 0;
906         } else if (tail_cycle < cycle) {
907                 ASSERT(tail_cycle == (cycle - 1));
908                 free_bytes = tail_bytes - bytes;
909         } else {
910                 /*
911                  * The reservation head is behind the tail.
912                  * In this case we just want to return the size of the
913                  * log as the amount of space left.
914                  */
915                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
916                         "xlog_space_left: head behind tail\n"
917                         "  tail_cycle = %d, tail_bytes = %d\n"
918                         "  GH   cycle = %d, GH   bytes = %d",
919                         tail_cycle, tail_bytes, cycle, bytes);
920                 ASSERT(0);
921                 free_bytes = log->l_logsize;
922         }
923         return free_bytes;
924 }       /* xlog_space_left */
925
926
927 /*
928  * Log function which is called when an io completes.
929  *
930  * The log manager needs its own routine, in order to control what
931  * happens with the buffer after the write completes.
932  */
933 void
934 xlog_iodone(xfs_buf_t *bp)
935 {
936         xlog_in_core_t  *iclog;
937         xlog_t          *l;
938         int             aborted;
939
940         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
941         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
942         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
943         aborted = 0;
944
945         /*
946          * Some versions of cpp barf on the recursive definition of
947          * ic_log -> hic_fields.ic_log and expand ic_log twice when
948          * it is passed through two macros.  Workaround broken cpp.
949          */
950         l = iclog->ic_log;
951
952         /*
953          * Race to shutdown the filesystem if we see an error.
954          */
955         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
956                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
957                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
958                 XFS_BUF_STALE(bp);
959                 xfs_force_shutdown(l->l_mp, XFS_LOG_IO_ERROR);
960                 /*
961                  * This flag will be propagated to the trans-committed
962                  * callback routines to let them know that the log-commit
963                  * didn't succeed.
964                  */
965                 aborted = XFS_LI_ABORTED;
966         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
967                 aborted = XFS_LI_ABORTED;
968         }
969         xlog_state_done_syncing(iclog, aborted);
970         if (!(XFS_BUF_ISASYNC(bp))) {
971                 /*
972                  * Corresponding psema() will be done in bwrite().  If we don't
973                  * vsema() here, panic.
974                  */
975                 XFS_BUF_V_IODONESEMA(bp);
976         }
977 }       /* xlog_iodone */
978
979 /*
980  * The bdstrat callback function for log bufs. This gives us a central
981  * place to trap bufs in case we get hit by a log I/O error and need to
982  * shutdown. Actually, in practice, even when we didn't get a log error,
983  * we transition the iclogs to IOERROR state *after* flushing all existing
984  * iclogs to disk. This is because we don't want anymore new transactions to be
985  * started or completed afterwards.
986  */
987 STATIC int
988 xlog_bdstrat_cb(struct xfs_buf *bp)
989 {
990         xlog_in_core_t *iclog;
991
992         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
993
994         if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
995           /* note for irix bstrat will need  struct bdevsw passed
996            * Fix the following macro if the code ever is merged
997            */
998             XFS_bdstrat(bp);
999                 return 0;
1000         }
1001
1002         xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1003         XFS_BUF_ERROR(bp, EIO);
1004         XFS_BUF_STALE(bp);
1005         xfs_biodone(bp);
1006         return XFS_ERROR(EIO);
1007
1008
1009 }
1010
1011 /*
1012  * Return size of each in-core log record buffer.
1013  *
1014  * Low memory machines only get 2 16KB buffers.  We don't want to waste
1015  * memory here.  However, all other machines get at least 2 32KB buffers.
1016  * The number is hard coded because we don't care about the minimum
1017  * memory size, just 32MB systems.
1018  *
1019  * If the filesystem blocksize is too large, we may need to choose a
1020  * larger size since the directory code currently logs entire blocks.
1021  */
1022
1023 STATIC void
1024 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1025                            xlog_t       *log)
1026 {
1027         int size;
1028         int xhdrs;
1029
1030         if (mp->m_logbufs <= 0) {
1031                 if (xfs_physmem <= btoc(128*1024*1024)) {
1032                         log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1033                 } else if (xfs_physmem <= btoc(400*1024*1024)) {
1034                         log->l_iclog_bufs = XLOG_MED_ICLOGS;
1035                 } else {        /* 256K with 32K bufs */
1036                         log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1037                 }
1038         } else {
1039                 log->l_iclog_bufs = mp->m_logbufs;
1040         }
1041
1042         /*
1043          * Buffer size passed in from mount system call.
1044          */
1045         if (mp->m_logbsize > 0) {
1046                 size = log->l_iclog_size = mp->m_logbsize;
1047                 log->l_iclog_size_log = 0;
1048                 while (size != 1) {
1049                         log->l_iclog_size_log++;
1050                         size >>= 1;
1051                 }
1052
1053                 if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1054                         /* # headers = size / 32K
1055                          * one header holds cycles from 32K of data
1056                          */
1057
1058                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1059                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1060                                 xhdrs++;
1061                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1062                         log->l_iclog_heads = xhdrs;
1063                 } else {
1064                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1065                         log->l_iclog_hsize = BBSIZE;
1066                         log->l_iclog_heads = 1;
1067                 }
1068                 goto done;
1069         }
1070
1071         /*
1072          * Special case machines that have less than 32MB of memory.
1073          * All machines with more memory use 32KB buffers.
1074          */
1075         if (xfs_physmem <= btoc(32*1024*1024)) {
1076                 /* Don't change; min configuration */
1077                 log->l_iclog_size = XLOG_RECORD_BSIZE;          /* 16k */
1078                 log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1079         } else {
1080                 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;      /* 32k */
1081                 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1082         }
1083
1084         /* the default log size is 16k or 32k which is one header sector */
1085         log->l_iclog_hsize = BBSIZE;
1086         log->l_iclog_heads = 1;
1087
1088         /*
1089          * For 16KB, we use 3 32KB buffers.  For 32KB block sizes, we use
1090          * 4 32KB buffers.  For 64KB block sizes, we use 8 32KB buffers.
1091          */
1092         if (mp->m_sb.sb_blocksize >= 16*1024) {
1093                 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1094                 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1095                 if (mp->m_logbufs <= 0) {
1096                         switch (mp->m_sb.sb_blocksize) {
1097                             case 16*1024:                       /* 16 KB */
1098                                 log->l_iclog_bufs = 3;
1099                                 break;
1100                             case 32*1024:                       /* 32 KB */
1101                                 log->l_iclog_bufs = 4;
1102                                 break;
1103                             case 64*1024:                       /* 64 KB */
1104                                 log->l_iclog_bufs = 8;
1105                                 break;
1106                             default:
1107                                 xlog_panic("XFS: Invalid blocksize");
1108                                 break;
1109                         }
1110                 }
1111         }
1112
1113 done:   /* are we being asked to make the sizes selected above visible? */
1114         if (mp->m_logbufs == 0)
1115                 mp->m_logbufs = log->l_iclog_bufs;
1116         if (mp->m_logbsize == 0)
1117                 mp->m_logbsize = log->l_iclog_size;
1118 }       /* xlog_get_iclog_buffer_size */
1119
1120
1121 /*
1122  * This routine initializes some of the log structure for a given mount point.
1123  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1124  * some other stuff may be filled in too.
1125  */
1126 STATIC xlog_t *
1127 xlog_alloc_log(xfs_mount_t      *mp,
1128                xfs_buftarg_t    *log_target,
1129                xfs_daddr_t      blk_offset,
1130                int              num_bblks)
1131 {
1132         xlog_t                  *log;
1133         xlog_rec_header_t       *head;
1134         xlog_in_core_t          **iclogp;
1135         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1136         xfs_buf_t               *bp;
1137         int                     i;
1138         int                     iclogsize;
1139
1140         log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1141
1142         log->l_mp          = mp;
1143         log->l_targ        = log_target;
1144         log->l_logsize     = BBTOB(num_bblks);
1145         log->l_logBBstart  = blk_offset;
1146         log->l_logBBsize   = num_bblks;
1147         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1148         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1149
1150         log->l_prev_block  = -1;
1151         ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1152         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1153         log->l_last_sync_lsn = log->l_tail_lsn;
1154         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1155         log->l_grant_reserve_cycle = 1;
1156         log->l_grant_write_cycle = 1;
1157
1158         if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1159                 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1160                 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1161                 /* for larger sector sizes, must have v2 or external log */
1162                 ASSERT(log->l_sectbb_log == 0 ||
1163                         log->l_logBBstart == 0 ||
1164                         XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1165                 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1166         }
1167         log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1168
1169         xlog_get_iclog_buffer_size(mp, log);
1170
1171         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1172         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1173         XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1174         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1175         ASSERT(XFS_BUF_ISBUSY(bp));
1176         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1177         log->l_xbuf = bp;
1178
1179         spinlock_init(&log->l_icloglock, "iclog");
1180         spinlock_init(&log->l_grant_lock, "grhead_iclog");
1181         initnsema(&log->l_flushsema, 0, "ic-flush");
1182         xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
1183
1184         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1185         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1186
1187         iclogp = &log->l_iclog;
1188         /*
1189          * The amount of memory to allocate for the iclog structure is
1190          * rather funky due to the way the structure is defined.  It is
1191          * done this way so that we can use different sizes for machines
1192          * with different amounts of memory.  See the definition of
1193          * xlog_in_core_t in xfs_log_priv.h for details.
1194          */
1195         iclogsize = log->l_iclog_size;
1196         ASSERT(log->l_iclog_size >= 4096);
1197         for (i=0; i < log->l_iclog_bufs; i++) {
1198                 *iclogp = (xlog_in_core_t *)
1199                           kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1200                 iclog = *iclogp;
1201                 iclog->hic_data = (xlog_in_core_2_t *)
1202                           kmem_zalloc(iclogsize, KM_SLEEP);
1203
1204                 iclog->ic_prev = prev_iclog;
1205                 prev_iclog = iclog;
1206                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1207
1208                 head = &iclog->ic_header;
1209                 memset(head, 0, sizeof(xlog_rec_header_t));
1210                 INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1211                 INT_SET(head->h_version, ARCH_CONVERT,
1212                         XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1213                 INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1214                 /* new fields */
1215                 INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1216                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1217
1218                 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1219                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1220                 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1221                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1222                 iclog->ic_bp = bp;
1223
1224                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1225                 iclog->ic_state = XLOG_STATE_ACTIVE;
1226                 iclog->ic_log = log;
1227                 iclog->ic_callback_tail = &(iclog->ic_callback);
1228                 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1229
1230                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1231                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1232                 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1233                 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1234
1235                 iclogp = &iclog->ic_next;
1236         }
1237         *iclogp = log->l_iclog;                 /* complete ring */
1238         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1239
1240         return log;
1241 }       /* xlog_alloc_log */
1242
1243
1244 /*
1245  * Write out the commit record of a transaction associated with the given
1246  * ticket.  Return the lsn of the commit record.
1247  */
1248 STATIC int
1249 xlog_commit_record(xfs_mount_t  *mp,
1250                    xlog_ticket_t *ticket,
1251                    xlog_in_core_t **iclog,
1252                    xfs_lsn_t    *commitlsnp)
1253 {
1254         int             error;
1255         xfs_log_iovec_t reg[1];
1256
1257         reg[0].i_addr = NULL;
1258         reg[0].i_len = 0;
1259         XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1260
1261         ASSERT_ALWAYS(iclog);
1262         if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1263                                iclog, XLOG_COMMIT_TRANS))) {
1264                 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
1265         }
1266         return error;
1267 }       /* xlog_commit_record */
1268
1269
1270 /*
1271  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1272  * log space.  This code pushes on the lsn which would supposedly free up
1273  * the 25% which we want to leave free.  We may need to adopt a policy which
1274  * pushes on an lsn which is further along in the log once we reach the high
1275  * water mark.  In this manner, we would be creating a low water mark.
1276  */
1277 void
1278 xlog_grant_push_ail(xfs_mount_t *mp,
1279                     int         need_bytes)
1280 {
1281     xlog_t      *log = mp->m_log;       /* pointer to the log */
1282     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1283     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1284     int         free_blocks;            /* free blocks left to write to */
1285     int         free_bytes;             /* free bytes left to write to */
1286     int         threshold_block;        /* block in lsn we'd like to be at */
1287     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1288     int         free_threshold;
1289     SPLDECL(s);
1290
1291     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1292
1293     s = GRANT_LOCK(log);
1294     free_bytes = xlog_space_left(log,
1295                                  log->l_grant_reserve_cycle,
1296                                  log->l_grant_reserve_bytes);
1297     tail_lsn = log->l_tail_lsn;
1298     free_blocks = BTOBBT(free_bytes);
1299
1300     /*
1301      * Set the threshold for the minimum number of free blocks in the
1302      * log to the maximum of what the caller needs, one quarter of the
1303      * log, and 256 blocks.
1304      */
1305     free_threshold = BTOBB(need_bytes);
1306     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1307     free_threshold = MAX(free_threshold, 256);
1308     if (free_blocks < free_threshold) {
1309         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1310         threshold_cycle = CYCLE_LSN(tail_lsn);
1311         if (threshold_block >= log->l_logBBsize) {
1312             threshold_block -= log->l_logBBsize;
1313             threshold_cycle += 1;
1314         }
1315         ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1316                        threshold_block);
1317
1318         /* Don't pass in an lsn greater than the lsn of the last
1319          * log record known to be on disk.
1320          */
1321         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1322             threshold_lsn = log->l_last_sync_lsn;
1323     }
1324     GRANT_UNLOCK(log, s);
1325
1326     /*
1327      * Get the transaction layer to kick the dirty buffers out to
1328      * disk asynchronously. No point in trying to do this if
1329      * the filesystem is shutting down.
1330      */
1331     if (threshold_lsn &&
1332         !XLOG_FORCED_SHUTDOWN(log))
1333             xfs_trans_push_ail(mp, threshold_lsn);
1334 }       /* xlog_grant_push_ail */
1335
1336
1337 /*
1338  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1339  * fashion.  Previously, we should have moved the current iclog
1340  * ptr in the log to point to the next available iclog.  This allows further
1341  * write to continue while this code syncs out an iclog ready to go.
1342  * Before an in-core log can be written out, the data section must be scanned
1343  * to save away the 1st word of each BBSIZE block into the header.  We replace
1344  * it with the current cycle count.  Each BBSIZE block is tagged with the
1345  * cycle count because there in an implicit assumption that drives will
1346  * guarantee that entire 512 byte blocks get written at once.  In other words,
1347  * we can't have part of a 512 byte block written and part not written.  By
1348  * tagging each block, we will know which blocks are valid when recovering
1349  * after an unclean shutdown.
1350  *
1351  * This routine is single threaded on the iclog.  No other thread can be in
1352  * this routine with the same iclog.  Changing contents of iclog can there-
1353  * fore be done without grabbing the state machine lock.  Updating the global
1354  * log will require grabbing the lock though.
1355  *
1356  * The entire log manager uses a logical block numbering scheme.  Only
1357  * log_sync (and then only bwrite()) know about the fact that the log may
1358  * not start with block zero on a given device.  The log block start offset
1359  * is added immediately before calling bwrite().
1360  */
1361
1362 int
1363 xlog_sync(xlog_t                *log,
1364           xlog_in_core_t        *iclog)
1365 {
1366         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1367         xfs_buf_t       *bp;
1368         int             i, ops;
1369         uint            count;          /* byte count of bwrite */
1370         uint            count_init;     /* initial count before roundup */
1371         int             roundoff;       /* roundoff to BB or stripe */
1372         int             split = 0;      /* split write into two regions */
1373         int             error;
1374         SPLDECL(s);
1375         int             v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1376
1377         XFS_STATS_INC(xs_log_writes);
1378         ASSERT(iclog->ic_refcnt == 0);
1379
1380         /* Add for LR header */
1381         count_init = log->l_iclog_hsize + iclog->ic_offset;
1382
1383         /* Round out the log write size */
1384         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1385                 /* we have a v2 stripe unit to use */
1386                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1387         } else {
1388                 count = BBTOB(BTOBB(count_init));
1389         }
1390         roundoff = count - count_init;
1391         ASSERT(roundoff >= 0);
1392         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1393                 roundoff < log->l_mp->m_sb.sb_logsunit)
1394                 || 
1395                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1396                  roundoff < BBTOB(1)));
1397
1398         /* move grant heads by roundoff in sync */
1399         s = GRANT_LOCK(log);
1400         xlog_grant_add_space(log, roundoff);
1401         GRANT_UNLOCK(log, s);
1402
1403         /* put cycle number in every block */
1404         xlog_pack_data(log, iclog, roundoff); 
1405
1406         /* real byte length */
1407         if (v2) {
1408                 INT_SET(iclog->ic_header.h_len, 
1409                         ARCH_CONVERT,
1410                         iclog->ic_offset + roundoff);
1411         } else {
1412                 INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1413         }
1414
1415         /* put ops count in correct order */
1416         ops = iclog->ic_header.h_num_logops;
1417         INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1418
1419         bp          = iclog->ic_bp;
1420         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1421         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1422         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1423
1424         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1425
1426         /* Do we need to split this write into 2 parts? */
1427         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1428                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1429                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1430                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1431         } else {
1432                 iclog->ic_bwritecnt = 1;
1433         }
1434         XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1435         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1436         XFS_BUF_BUSY(bp);
1437         XFS_BUF_ASYNC(bp);
1438         /*
1439          * Do an ordered write for the log block.
1440          *
1441          * It may not be needed to flush the first split block in the log wrap
1442          * case, but do it anyways to be safe -AK
1443          */
1444         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1445                 XFS_BUF_ORDERED(bp);
1446
1447         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1448         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1449
1450         xlog_verify_iclog(log, iclog, count, B_TRUE);
1451
1452         /* account for log which doesn't start at block #0 */
1453         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1454         /*
1455          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1456          * is shutting down.
1457          */
1458         XFS_BUF_WRITE(bp);
1459
1460         if ((error = XFS_bwrite(bp))) {
1461                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1462                                   XFS_BUF_ADDR(bp));
1463                 return error;
1464         }
1465         if (split) {
1466                 bp              = iclog->ic_log->l_xbuf;
1467                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1468                                                         (unsigned long)1);
1469                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1470                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1471                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1472                                             (__psint_t)count), split);
1473                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1474                 XFS_BUF_BUSY(bp);
1475                 XFS_BUF_ASYNC(bp);
1476                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1477                         XFS_BUF_ORDERED(bp);
1478                 dptr = XFS_BUF_PTR(bp);
1479                 /*
1480                  * Bump the cycle numbers at the start of each block
1481                  * since this part of the buffer is at the start of
1482                  * a new cycle.  Watch out for the header magic number
1483                  * case, though.
1484                  */
1485                 for (i=0; i<split; i += BBSIZE) {
1486                         INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1487                         if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1488                                 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1489                         dptr += BBSIZE;
1490                 }
1491
1492                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1493                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1494
1495                 /* account for internal log which does't start at block #0 */
1496                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1497                 XFS_BUF_WRITE(bp);
1498                 if ((error = XFS_bwrite(bp))) {
1499                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1500                                           bp, XFS_BUF_ADDR(bp));
1501                         return error;
1502                 }
1503         }
1504         return 0;
1505 }       /* xlog_sync */
1506
1507
1508 /*
1509  * Unallocate a log structure
1510  */
1511 void
1512 xlog_unalloc_log(xlog_t *log)
1513 {
1514         xlog_in_core_t  *iclog, *next_iclog;
1515         xlog_ticket_t   *tic, *next_tic;
1516         int             i;
1517
1518
1519         iclog = log->l_iclog;
1520         for (i=0; i<log->l_iclog_bufs; i++) {
1521                 sv_destroy(&iclog->ic_forcesema);
1522                 sv_destroy(&iclog->ic_writesema);
1523                 xfs_buf_free(iclog->ic_bp);
1524 #ifdef XFS_LOG_TRACE
1525                 if (iclog->ic_trace != NULL) {
1526                         ktrace_free(iclog->ic_trace);
1527                 }
1528 #endif
1529                 next_iclog = iclog->ic_next;
1530                 kmem_free(iclog->hic_data, log->l_iclog_size);
1531                 kmem_free(iclog, sizeof(xlog_in_core_t));
1532                 iclog = next_iclog;
1533         }
1534         freesema(&log->l_flushsema);
1535         spinlock_destroy(&log->l_icloglock);
1536         spinlock_destroy(&log->l_grant_lock);
1537
1538         /* XXXsup take a look at this again. */
1539         if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
1540             !XLOG_FORCED_SHUTDOWN(log)) {
1541                 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1542                         "xlog_unalloc_log: (cnt: %d, total: %d)",
1543                         log->l_ticket_cnt, log->l_ticket_tcnt);
1544                 /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1545
1546         } else {
1547                 tic = log->l_unmount_free;
1548                 while (tic) {
1549                         next_tic = tic->t_next;
1550                         kmem_free(tic, NBPP);
1551                         tic = next_tic;
1552                 }
1553         }
1554         xfs_buf_free(log->l_xbuf);
1555 #ifdef XFS_LOG_TRACE
1556         if (log->l_trace != NULL) {
1557                 ktrace_free(log->l_trace);
1558         }
1559         if (log->l_grant_trace != NULL) {
1560                 ktrace_free(log->l_grant_trace);
1561         }
1562 #endif
1563         log->l_mp->m_log = NULL;
1564         kmem_free(log, sizeof(xlog_t));
1565 }       /* xlog_unalloc_log */
1566
1567 /*
1568  * Update counters atomically now that memcpy is done.
1569  */
1570 /* ARGSUSED */
1571 static inline void
1572 xlog_state_finish_copy(xlog_t           *log,
1573                        xlog_in_core_t   *iclog,
1574                        int              record_cnt,
1575                        int              copy_bytes)
1576 {
1577         SPLDECL(s);
1578
1579         s = LOG_LOCK(log);
1580
1581         iclog->ic_header.h_num_logops += record_cnt;
1582         iclog->ic_offset += copy_bytes;
1583
1584         LOG_UNLOCK(log, s);
1585 }       /* xlog_state_finish_copy */
1586
1587
1588
1589
1590 /*
1591  * print out info relating to regions written which consume
1592  * the reservation
1593  */
1594 STATIC void
1595 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1596 {
1597         uint i;
1598         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1599
1600         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1601         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1602             "bformat",
1603             "bchunk",
1604             "efi_format",
1605             "efd_format",
1606             "iformat",
1607             "icore",
1608             "iext",
1609             "ibroot",
1610             "ilocal",
1611             "iattr_ext",
1612             "iattr_broot",
1613             "iattr_local",
1614             "qformat",
1615             "dquot",
1616             "quotaoff",
1617             "LR header",
1618             "unmount",
1619             "commit",
1620             "trans header"
1621         };
1622         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1623             "SETATTR_NOT_SIZE",
1624             "SETATTR_SIZE",
1625             "INACTIVE",
1626             "CREATE",
1627             "CREATE_TRUNC",
1628             "TRUNCATE_FILE",
1629             "REMOVE",
1630             "LINK",
1631             "RENAME",
1632             "MKDIR",
1633             "RMDIR",
1634             "SYMLINK",
1635             "SET_DMATTRS",
1636             "GROWFS",
1637             "STRAT_WRITE",
1638             "DIOSTRAT",
1639             "WRITE_SYNC",
1640             "WRITEID",
1641             "ADDAFORK",
1642             "ATTRINVAL",
1643             "ATRUNCATE",
1644             "ATTR_SET",
1645             "ATTR_RM",
1646             "ATTR_FLAG",
1647             "CLEAR_AGI_BUCKET",
1648             "QM_SBCHANGE",
1649             "DUMMY1",
1650             "DUMMY2",
1651             "QM_QUOTAOFF",
1652             "QM_DQALLOC",
1653             "QM_SETQLIM",
1654             "QM_DQCLUSTER",
1655             "QM_QINOCREATE",
1656             "QM_QUOTAOFF_END",
1657             "SB_UNIT",
1658             "FSYNC_TS",
1659             "GROWFSRT_ALLOC",
1660             "GROWFSRT_ZERO",
1661             "GROWFSRT_FREE",
1662             "SWAPEXT"
1663         };
1664
1665         xfs_fs_cmn_err(CE_WARN, mp,
1666                         "xfs_log_write: reservation summary:\n"
1667                         "  trans type  = %s (%u)\n"
1668                         "  unit res    = %d bytes\n"
1669                         "  current res = %d bytes\n"
1670                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1671                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1672                         "  ophdr + reg = %u bytes\n"
1673                         "  num regions = %u\n",
1674                         ((ticket->t_trans_type <= 0 ||
1675                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1676                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1677                         ticket->t_trans_type,
1678                         ticket->t_unit_res,
1679                         ticket->t_curr_res,
1680                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1681                         ticket->t_res_num_ophdrs, ophdr_spc,
1682                         ticket->t_res_arr_sum + 
1683                         ticket->t_res_o_flow + ophdr_spc,
1684                         ticket->t_res_num);
1685
1686         for (i = 0; i < ticket->t_res_num; i++) {
1687                 uint r_type = ticket->t_res_arr[i].r_type; 
1688                 cmn_err(CE_WARN,
1689                             "region[%u]: %s - %u bytes\n",
1690                             i, 
1691                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1692                             "bad-rtype" : res_type_str[r_type-1]),
1693                             ticket->t_res_arr[i].r_len);
1694         }
1695 }
1696
1697 /*
1698  * Write some region out to in-core log
1699  *
1700  * This will be called when writing externally provided regions or when
1701  * writing out a commit record for a given transaction.
1702  *
1703  * General algorithm:
1704  *      1. Find total length of this write.  This may include adding to the
1705  *              lengths passed in.
1706  *      2. Check whether we violate the tickets reservation.
1707  *      3. While writing to this iclog
1708  *          A. Reserve as much space in this iclog as can get
1709  *          B. If this is first write, save away start lsn
1710  *          C. While writing this region:
1711  *              1. If first write of transaction, write start record
1712  *              2. Write log operation header (header per region)
1713  *              3. Find out if we can fit entire region into this iclog
1714  *              4. Potentially, verify destination memcpy ptr
1715  *              5. Memcpy (partial) region
1716  *              6. If partial copy, release iclog; otherwise, continue
1717  *                      copying more regions into current iclog
1718  *      4. Mark want sync bit (in simulation mode)
1719  *      5. Release iclog for potential flush to on-disk log.
1720  *
1721  * ERRORS:
1722  * 1.   Panic if reservation is overrun.  This should never happen since
1723  *      reservation amounts are generated internal to the filesystem.
1724  * NOTES:
1725  * 1. Tickets are single threaded data structures.
1726  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1727  *      syncing routine.  When a single log_write region needs to span
1728  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1729  *      on all log operation writes which don't contain the end of the
1730  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1731  *      operation which contains the end of the continued log_write region.
1732  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1733  *      we don't really know exactly how much space will be used.  As a result,
1734  *      we don't update ic_offset until the end when we know exactly how many
1735  *      bytes have been written out.
1736  */
1737 int
1738 xlog_write(xfs_mount_t *        mp,
1739            xfs_log_iovec_t      reg[],
1740            int                  nentries,
1741            xfs_log_ticket_t     tic,
1742            xfs_lsn_t            *start_lsn,
1743            xlog_in_core_t       **commit_iclog,
1744            uint                 flags)
1745 {
1746     xlog_t           *log    = mp->m_log;
1747     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1748     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1749     xlog_in_core_t   *iclog;         /* ptr to current in-core log */
1750     __psint_t        ptr;            /* copy address into data region */
1751     int              len;            /* # xlog_write() bytes 2 still copy */
1752     int              index;          /* region index currently copying */
1753     int              log_offset;     /* offset (from 0) into data region */
1754     int              start_rec_copy; /* # bytes to copy for start record */
1755     int              partial_copy;   /* did we split a region? */
1756     int              partial_copy_len;/* # bytes copied if split region */
1757     int              need_copy;      /* # bytes need to memcpy this region */
1758     int              copy_len;       /* # bytes actually memcpy'ing */
1759     int              copy_off;       /* # bytes from entry start */
1760     int              contwr;         /* continued write of in-core log? */
1761     int              error;
1762     int              record_cnt = 0, data_cnt = 0;
1763
1764     partial_copy_len = partial_copy = 0;
1765
1766     /* Calculate potential maximum space.  Each region gets its own
1767      * xlog_op_header_t and may need to be double word aligned.
1768      */
1769     len = 0;
1770     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1771         len += sizeof(xlog_op_header_t);
1772         XLOG_TIC_ADD_OPHDR(ticket);
1773     }
1774
1775     for (index = 0; index < nentries; index++) {
1776         len += sizeof(xlog_op_header_t);            /* each region gets >= 1 */
1777         XLOG_TIC_ADD_OPHDR(ticket);
1778         len += reg[index].i_len;
1779         XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1780     }
1781     contwr = *start_lsn = 0;
1782
1783     if (ticket->t_curr_res < len) {
1784         xlog_print_tic_res(mp, ticket);
1785 #ifdef DEBUG
1786         xlog_panic(
1787                 "xfs_log_write: reservation ran out. Need to up reservation");
1788 #else
1789         /* Customer configurable panic */
1790         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1791                 "xfs_log_write: reservation ran out. Need to up reservation");
1792         /* If we did not panic, shutdown the filesystem */
1793         xfs_force_shutdown(mp, XFS_CORRUPT_INCORE);
1794 #endif
1795     } else
1796         ticket->t_curr_res -= len;
1797
1798     for (index = 0; index < nentries; ) {
1799         if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1800                                                &contwr, &log_offset)))
1801                 return error;
1802
1803         ASSERT(log_offset <= iclog->ic_size - 1);
1804         ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1805
1806         /* start_lsn is the first lsn written to. That's all we need. */
1807         if (! *start_lsn)
1808             *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1809
1810         /* This loop writes out as many regions as can fit in the amount
1811          * of space which was allocated by xlog_state_get_iclog_space().
1812          */
1813         while (index < nentries) {
1814             ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1815             ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1816             start_rec_copy = 0;
1817
1818             /* If first write for transaction, insert start record.
1819              * We can't be trying to commit if we are inited.  We can't
1820              * have any "partial_copy" if we are inited.
1821              */
1822             if (ticket->t_flags & XLOG_TIC_INITED) {
1823                 logop_head              = (xlog_op_header_t *)ptr;
1824                 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1825                 logop_head->oh_clientid = ticket->t_clientid;
1826                 logop_head->oh_len      = 0;
1827                 logop_head->oh_flags    = XLOG_START_TRANS;
1828                 logop_head->oh_res2     = 0;
1829                 ticket->t_flags         &= ~XLOG_TIC_INITED;    /* clear bit */
1830                 record_cnt++;
1831
1832                 start_rec_copy = sizeof(xlog_op_header_t);
1833                 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1834             }
1835
1836             /* Copy log operation header directly into data section */
1837             logop_head                  = (xlog_op_header_t *)ptr;
1838             INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1839             logop_head->oh_clientid     = ticket->t_clientid;
1840             logop_head->oh_res2         = 0;
1841
1842             /* header copied directly */
1843             xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1844
1845             /* are we copying a commit or unmount record? */
1846             logop_head->oh_flags = flags;
1847
1848             /*
1849              * We've seen logs corrupted with bad transaction client
1850              * ids.  This makes sure that XFS doesn't generate them on.
1851              * Turn this into an EIO and shut down the filesystem.
1852              */
1853             switch (logop_head->oh_clientid)  {
1854             case XFS_TRANSACTION:
1855             case XFS_VOLUME:
1856             case XFS_LOG:
1857                 break;
1858             default:
1859                 xfs_fs_cmn_err(CE_WARN, mp,
1860                     "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1861                     logop_head->oh_clientid, tic);
1862                 return XFS_ERROR(EIO);
1863             }
1864
1865             /* Partial write last time? => (partial_copy != 0)
1866              * need_copy is the amount we'd like to copy if everything could
1867              * fit in the current memcpy.
1868              */
1869             need_copy = reg[index].i_len - partial_copy_len;
1870
1871             copy_off = partial_copy_len;
1872             if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1873                 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1874                 if (partial_copy)
1875                     logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1876                 partial_copy_len = partial_copy = 0;
1877             } else {                                        /* partial write */
1878                 copy_len = iclog->ic_size - log_offset;
1879                 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1880                 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1881                 if (partial_copy)
1882                         logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1883                 partial_copy_len += copy_len;
1884                 partial_copy++;
1885                 len += sizeof(xlog_op_header_t); /* from splitting of region */
1886                 /* account for new log op header */
1887                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1888                 XLOG_TIC_ADD_OPHDR(ticket);
1889             }
1890             xlog_verify_dest_ptr(log, ptr);
1891
1892             /* copy region */
1893             ASSERT(copy_len >= 0);
1894             memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1895             xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1896
1897             /* make copy_len total bytes copied, including headers */
1898             copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1899             record_cnt++;
1900             data_cnt += contwr ? copy_len : 0;
1901             if (partial_copy) {                 /* copied partial region */
1902                     /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1903                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1904                     record_cnt = data_cnt = 0;
1905                     if ((error = xlog_state_release_iclog(log, iclog)))
1906                             return error;
1907                     break;                      /* don't increment index */
1908             } else {                            /* copied entire region */
1909                 index++;
1910                 partial_copy_len = partial_copy = 0;
1911
1912                 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1913                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1914                     record_cnt = data_cnt = 0;
1915                     xlog_state_want_sync(log, iclog);
1916                     if (commit_iclog) {
1917                         ASSERT(flags & XLOG_COMMIT_TRANS);
1918                         *commit_iclog = iclog;
1919                     } else if ((error = xlog_state_release_iclog(log, iclog)))
1920                            return error;
1921                     if (index == nentries)
1922                             return 0;           /* we are done */
1923                     else
1924                             break;
1925                 }
1926             } /* if (partial_copy) */
1927         } /* while (index < nentries) */
1928     } /* for (index = 0; index < nentries; ) */
1929     ASSERT(len == 0);
1930
1931     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1932     if (commit_iclog) {
1933         ASSERT(flags & XLOG_COMMIT_TRANS);
1934         *commit_iclog = iclog;
1935         return 0;
1936     }
1937     return xlog_state_release_iclog(log, iclog);
1938 }       /* xlog_write */
1939
1940
1941 /*****************************************************************************
1942  *
1943  *              State Machine functions
1944  *
1945  *****************************************************************************
1946  */
1947
1948 /* Clean iclogs starting from the head.  This ordering must be
1949  * maintained, so an iclog doesn't become ACTIVE beyond one that
1950  * is SYNCING.  This is also required to maintain the notion that we use
1951  * a counting semaphore to hold off would be writers to the log when every
1952  * iclog is trying to sync to disk.
1953  *
1954  * State Change: DIRTY -> ACTIVE
1955  */
1956 STATIC void
1957 xlog_state_clean_log(xlog_t *log)
1958 {
1959         xlog_in_core_t  *iclog;
1960         int changed = 0;
1961
1962         iclog = log->l_iclog;
1963         do {
1964                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1965                         iclog->ic_state = XLOG_STATE_ACTIVE;
1966                         iclog->ic_offset       = 0;
1967                         iclog->ic_callback      = NULL;   /* don't need to free */
1968                         /*
1969                          * If the number of ops in this iclog indicate it just
1970                          * contains the dummy transaction, we can
1971                          * change state into IDLE (the second time around).
1972                          * Otherwise we should change the state into
1973                          * NEED a dummy.
1974                          * We don't need to cover the dummy.
1975                          */
1976                         if (!changed &&
1977                            (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1978                                 changed = 1;
1979                         } else {
1980                                 /*
1981                                  * We have two dirty iclogs so start over
1982                                  * This could also be num of ops indicates
1983                                  * this is not the dummy going out.
1984                                  */
1985                                 changed = 2;
1986                         }
1987                         iclog->ic_header.h_num_logops = 0;
1988                         memset(iclog->ic_header.h_cycle_data, 0,
1989                               sizeof(iclog->ic_header.h_cycle_data));
1990                         iclog->ic_header.h_lsn = 0;
1991                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1992                         /* do nothing */;
1993                 else
1994                         break;  /* stop cleaning */
1995                 iclog = iclog->ic_next;
1996         } while (iclog != log->l_iclog);
1997
1998         /* log is locked when we are called */
1999         /*
2000          * Change state for the dummy log recording.
2001          * We usually go to NEED. But we go to NEED2 if the changed indicates
2002          * we are done writing the dummy record.
2003          * If we are done with the second dummy recored (DONE2), then
2004          * we go to IDLE.
2005          */
2006         if (changed) {
2007                 switch (log->l_covered_state) {
2008                 case XLOG_STATE_COVER_IDLE:
2009                 case XLOG_STATE_COVER_NEED:
2010                 case XLOG_STATE_COVER_NEED2:
2011                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2012                         break;
2013
2014                 case XLOG_STATE_COVER_DONE:
2015                         if (changed == 1)
2016                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2017                         else
2018                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2019                         break;
2020
2021                 case XLOG_STATE_COVER_DONE2:
2022                         if (changed == 1)
2023                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2024                         else
2025                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2026                         break;
2027
2028                 default:
2029                         ASSERT(0);
2030                 }
2031         }
2032 }       /* xlog_state_clean_log */
2033
2034 STATIC xfs_lsn_t
2035 xlog_get_lowest_lsn(
2036         xlog_t          *log)
2037 {
2038         xlog_in_core_t  *lsn_log;
2039         xfs_lsn_t       lowest_lsn, lsn;
2040
2041         lsn_log = log->l_iclog;
2042         lowest_lsn = 0;
2043         do {
2044             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2045                 lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2046                 if ((lsn && !lowest_lsn) ||
2047                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2048                         lowest_lsn = lsn;
2049                 }
2050             }
2051             lsn_log = lsn_log->ic_next;
2052         } while (lsn_log != log->l_iclog);
2053         return lowest_lsn;
2054 }
2055
2056
2057 STATIC void
2058 xlog_state_do_callback(
2059         xlog_t          *log,
2060         int             aborted,
2061         xlog_in_core_t  *ciclog)
2062 {
2063         xlog_in_core_t     *iclog;
2064         xlog_in_core_t     *first_iclog;        /* used to know when we've
2065                                                  * processed all iclogs once */
2066         xfs_log_callback_t *cb, *cb_next;
2067         int                flushcnt = 0;
2068         xfs_lsn_t          lowest_lsn;
2069         int                ioerrors;    /* counter: iclogs with errors */
2070         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2071         int                funcdidcallbacks; /* flag: function did callbacks */
2072         int                repeats;     /* for issuing console warnings if
2073                                          * looping too many times */
2074         SPLDECL(s);
2075
2076         s = LOG_LOCK(log);
2077         first_iclog = iclog = log->l_iclog;
2078         ioerrors = 0;
2079         funcdidcallbacks = 0;
2080         repeats = 0;
2081
2082         do {
2083                 /*
2084                  * Scan all iclogs starting with the one pointed to by the
2085                  * log.  Reset this starting point each time the log is
2086                  * unlocked (during callbacks).
2087                  *
2088                  * Keep looping through iclogs until one full pass is made
2089                  * without running any callbacks.
2090                  */
2091                 first_iclog = log->l_iclog;
2092                 iclog = log->l_iclog;
2093                 loopdidcallbacks = 0;
2094                 repeats++;
2095
2096                 do {
2097
2098                         /* skip all iclogs in the ACTIVE & DIRTY states */
2099                         if (iclog->ic_state &
2100                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2101                                 iclog = iclog->ic_next;
2102                                 continue;
2103                         }
2104
2105                         /*
2106                          * Between marking a filesystem SHUTDOWN and stopping
2107                          * the log, we do flush all iclogs to disk (if there
2108                          * wasn't a log I/O error). So, we do want things to
2109                          * go smoothly in case of just a SHUTDOWN  w/o a
2110                          * LOG_IO_ERROR.
2111                          */
2112                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2113                                 /*
2114                                  * Can only perform callbacks in order.  Since
2115                                  * this iclog is not in the DONE_SYNC/
2116                                  * DO_CALLBACK state, we skip the rest and
2117                                  * just try to clean up.  If we set our iclog
2118                                  * to DO_CALLBACK, we will not process it when
2119                                  * we retry since a previous iclog is in the
2120                                  * CALLBACK and the state cannot change since
2121                                  * we are holding the LOG_LOCK.
2122                                  */
2123                                 if (!(iclog->ic_state &
2124                                         (XLOG_STATE_DONE_SYNC |
2125                                                  XLOG_STATE_DO_CALLBACK))) {
2126                                         if (ciclog && (ciclog->ic_state ==
2127                                                         XLOG_STATE_DONE_SYNC)) {
2128                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2129                                         }
2130                                         break;
2131                                 }
2132                                 /*
2133                                  * We now have an iclog that is in either the
2134                                  * DO_CALLBACK or DONE_SYNC states. The other
2135                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2136                                  * caught by the above if and are going to
2137                                  * clean (i.e. we aren't doing their callbacks)
2138                                  * see the above if.
2139                                  */
2140
2141                                 /*
2142                                  * We will do one more check here to see if we
2143                                  * have chased our tail around.
2144                                  */
2145
2146                                 lowest_lsn = xlog_get_lowest_lsn(log);
2147                                 if (lowest_lsn && (
2148                                         XFS_LSN_CMP(
2149                                                 lowest_lsn,
2150                                                 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2151                                         )<0)) {
2152                                         iclog = iclog->ic_next;
2153                                         continue; /* Leave this iclog for
2154                                                    * another thread */
2155                                 }
2156
2157                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2158
2159                                 LOG_UNLOCK(log, s);
2160
2161                                 /* l_last_sync_lsn field protected by
2162                                  * GRANT_LOCK. Don't worry about iclog's lsn.
2163                                  * No one else can be here except us.
2164                                  */
2165                                 s = GRANT_LOCK(log);
2166                                 ASSERT(XFS_LSN_CMP(
2167                                                 log->l_last_sync_lsn,
2168                                                 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2169                                         )<=0);
2170                                 log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2171                                 GRANT_UNLOCK(log, s);
2172
2173                                 /*
2174                                  * Keep processing entries in the callback list
2175                                  * until we come around and it is empty.  We
2176                                  * need to atomically see that the list is
2177                                  * empty and change the state to DIRTY so that
2178                                  * we don't miss any more callbacks being added.
2179                                  */
2180                                 s = LOG_LOCK(log);
2181                         } else {
2182                                 ioerrors++;
2183                         }
2184                         cb = iclog->ic_callback;
2185
2186                         while (cb != 0) {
2187                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2188                                 iclog->ic_callback = NULL;
2189                                 LOG_UNLOCK(log, s);
2190
2191                                 /* perform callbacks in the order given */
2192                                 for (; cb != 0; cb = cb_next) {
2193                                         cb_next = cb->cb_next;
2194                                         cb->cb_func(cb->cb_arg, aborted);
2195                                 }
2196                                 s = LOG_LOCK(log);
2197                                 cb = iclog->ic_callback;
2198                         }
2199
2200                         loopdidcallbacks++;
2201                         funcdidcallbacks++;
2202
2203                         ASSERT(iclog->ic_callback == 0);
2204                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2205                                 iclog->ic_state = XLOG_STATE_DIRTY;
2206
2207                         /*
2208                          * Transition from DIRTY to ACTIVE if applicable.
2209                          * NOP if STATE_IOERROR.
2210                          */
2211                         xlog_state_clean_log(log);
2212
2213                         /* wake up threads waiting in xfs_log_force() */
2214                         sv_broadcast(&iclog->ic_forcesema);
2215
2216                         iclog = iclog->ic_next;
2217                 } while (first_iclog != iclog);
2218                 if (repeats && (repeats % 10) == 0) {
2219                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2220                                 "xlog_state_do_callback: looping %d", repeats);
2221                 }
2222         } while (!ioerrors && loopdidcallbacks);
2223
2224         /*
2225          * make one last gasp attempt to see if iclogs are being left in
2226          * limbo..
2227          */
2228 #ifdef DEBUG
2229         if (funcdidcallbacks) {
2230                 first_iclog = iclog = log->l_iclog;
2231                 do {
2232                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2233                         /*
2234                          * Terminate the loop if iclogs are found in states
2235                          * which will cause other threads to clean up iclogs.
2236                          *
2237                          * SYNCING - i/o completion will go through logs
2238                          * DONE_SYNC - interrupt thread should be waiting for
2239                          *              LOG_LOCK
2240                          * IOERROR - give up hope all ye who enter here
2241                          */
2242                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2243                             iclog->ic_state == XLOG_STATE_SYNCING ||
2244                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2245                             iclog->ic_state == XLOG_STATE_IOERROR )
2246                                 break;
2247                         iclog = iclog->ic_next;
2248                 } while (first_iclog != iclog);
2249         }
2250 #endif
2251
2252         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2253                 flushcnt = log->l_flushcnt;
2254                 log->l_flushcnt = 0;
2255         }
2256         LOG_UNLOCK(log, s);
2257         while (flushcnt--)
2258                 vsema(&log->l_flushsema);
2259 }       /* xlog_state_do_callback */
2260
2261
2262 /*
2263  * Finish transitioning this iclog to the dirty state.
2264  *
2265  * Make sure that we completely execute this routine only when this is
2266  * the last call to the iclog.  There is a good chance that iclog flushes,
2267  * when we reach the end of the physical log, get turned into 2 separate
2268  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2269  * routine.  By using the reference count bwritecnt, we guarantee that only
2270  * the second completion goes through.
2271  *
2272  * Callbacks could take time, so they are done outside the scope of the
2273  * global state machine log lock.  Assume that the calls to cvsema won't
2274  * take a long time.  At least we know it won't sleep.
2275  */
2276 void
2277 xlog_state_done_syncing(
2278         xlog_in_core_t  *iclog,
2279         int             aborted)
2280 {
2281         xlog_t             *log = iclog->ic_log;
2282         SPLDECL(s);
2283
2284         s = LOG_LOCK(log);
2285
2286         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2287                iclog->ic_state == XLOG_STATE_IOERROR);
2288         ASSERT(iclog->ic_refcnt == 0);
2289         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2290
2291
2292         /*
2293          * If we got an error, either on the first buffer, or in the case of
2294          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2295          * and none should ever be attempted to be written to disk
2296          * again.
2297          */
2298         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2299                 if (--iclog->ic_bwritecnt == 1) {
2300                         LOG_UNLOCK(log, s);
2301                         return;
2302                 }
2303                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2304         }
2305
2306         /*
2307          * Someone could be sleeping prior to writing out the next
2308          * iclog buffer, we wake them all, one will get to do the
2309          * I/O, the others get to wait for the result.
2310          */
2311         sv_broadcast(&iclog->ic_writesema);
2312         LOG_UNLOCK(log, s);
2313         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2314 }       /* xlog_state_done_syncing */
2315
2316
2317 /*
2318  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2319  * sleep.  The flush semaphore is set to the number of in-core buffers and
2320  * decremented around disk syncing.  Therefore, if all buffers are syncing,
2321  * this semaphore will cause new writes to sleep until a sync completes.
2322  * Otherwise, this code just does p() followed by v().  This approximates
2323  * a sleep/wakeup except we can't race.
2324  *
2325  * The in-core logs are used in a circular fashion. They are not used
2326  * out-of-order even when an iclog past the head is free.
2327  *
2328  * return:
2329  *      * log_offset where xlog_write() can start writing into the in-core
2330  *              log's data space.
2331  *      * in-core log pointer to which xlog_write() should write.
2332  *      * boolean indicating this is a continued write to an in-core log.
2333  *              If this is the last write, then the in-core log's offset field
2334  *              needs to be incremented, depending on the amount of data which
2335  *              is copied.
2336  */
2337 int
2338 xlog_state_get_iclog_space(xlog_t         *log,
2339                            int            len,
2340                            xlog_in_core_t **iclogp,
2341                            xlog_ticket_t  *ticket,
2342                            int            *continued_write,
2343                            int            *logoffsetp)
2344 {
2345         SPLDECL(s);
2346         int               log_offset;
2347         xlog_rec_header_t *head;
2348         xlog_in_core_t    *iclog;
2349         int               error;
2350
2351 restart:
2352         s = LOG_LOCK(log);
2353         if (XLOG_FORCED_SHUTDOWN(log)) {
2354                 LOG_UNLOCK(log, s);
2355                 return XFS_ERROR(EIO);
2356         }
2357
2358         iclog = log->l_iclog;
2359         if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2360                 log->l_flushcnt++;
2361                 LOG_UNLOCK(log, s);
2362                 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2363                 XFS_STATS_INC(xs_log_noiclogs);
2364                 /* Ensure that log writes happen */
2365                 psema(&log->l_flushsema, PINOD);
2366                 goto restart;
2367         }
2368         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2369         head = &iclog->ic_header;
2370
2371         iclog->ic_refcnt++;                     /* prevents sync */
2372         log_offset = iclog->ic_offset;
2373
2374         /* On the 1st write to an iclog, figure out lsn.  This works
2375          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2376          * committing to.  If the offset is set, that's how many blocks
2377          * must be written.
2378          */
2379         if (log_offset == 0) {
2380                 ticket->t_curr_res -= log->l_iclog_hsize;
2381                 XLOG_TIC_ADD_REGION(ticket,
2382                                     log->l_iclog_hsize,
2383                                     XLOG_REG_TYPE_LRHEADER);
2384                 INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2385                 ASSIGN_LSN(head->h_lsn, log);
2386                 ASSERT(log->l_curr_block >= 0);
2387         }
2388
2389         /* If there is enough room to write everything, then do it.  Otherwise,
2390          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2391          * bit is on, so this will get flushed out.  Don't update ic_offset
2392          * until you know exactly how many bytes get copied.  Therefore, wait
2393          * until later to update ic_offset.
2394          *
2395          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2396          * can fit into remaining data section.
2397          */
2398         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2399                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2400
2401                 /* If I'm the only one writing to this iclog, sync it to disk */
2402                 if (iclog->ic_refcnt == 1) {
2403                         LOG_UNLOCK(log, s);
2404                         if ((error = xlog_state_release_iclog(log, iclog)))
2405                                 return error;
2406                 } else {
2407                         iclog->ic_refcnt--;
2408                         LOG_UNLOCK(log, s);
2409                 }
2410                 goto restart;
2411         }
2412
2413         /* Do we have enough room to write the full amount in the remainder
2414          * of this iclog?  Or must we continue a write on the next iclog and
2415          * mark this iclog as completely taken?  In the case where we switch
2416          * iclogs (to mark it taken), this particular iclog will release/sync
2417          * to disk in xlog_write().
2418          */
2419         if (len <= iclog->ic_size - iclog->ic_offset) {
2420                 *continued_write = 0;
2421                 iclog->ic_offset += len;
2422         } else {
2423                 *continued_write = 1;
2424                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2425         }
2426         *iclogp = iclog;
2427
2428         ASSERT(iclog->ic_offset <= iclog->ic_size);
2429         LOG_UNLOCK(log, s);
2430
2431         *logoffsetp = log_offset;
2432         return 0;
2433 }       /* xlog_state_get_iclog_space */
2434
2435 /*
2436  * Atomically get the log space required for a log ticket.
2437  *
2438  * Once a ticket gets put onto the reserveq, it will only return after
2439  * the needed reservation is satisfied.
2440  */
2441 STATIC int
2442 xlog_grant_log_space(xlog_t        *log,
2443                      xlog_ticket_t *tic)
2444 {
2445         int              free_bytes;
2446         int              need_bytes;
2447         SPLDECL(s);
2448 #ifdef DEBUG
2449         xfs_lsn_t        tail_lsn;
2450 #endif
2451
2452
2453 #ifdef DEBUG
2454         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2455                 panic("grant Recovery problem");
2456 #endif
2457
2458         /* Is there space or do we need to sleep? */
2459         s = GRANT_LOCK(log);
2460         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2461
2462         /* something is already sleeping; insert new transaction at end */
2463         if (log->l_reserve_headq) {
2464                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2465                 xlog_trace_loggrant(log, tic,
2466                                     "xlog_grant_log_space: sleep 1");
2467                 /*
2468                  * Gotta check this before going to sleep, while we're
2469                  * holding the grant lock.
2470                  */
2471                 if (XLOG_FORCED_SHUTDOWN(log))
2472                         goto error_return;
2473
2474                 XFS_STATS_INC(xs_sleep_logspace);
2475                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2476                 /*
2477                  * If we got an error, and the filesystem is shutting down,
2478                  * we'll catch it down below. So just continue...
2479                  */
2480                 xlog_trace_loggrant(log, tic,
2481                                     "xlog_grant_log_space: wake 1");
2482                 s = GRANT_LOCK(log);
2483         }
2484         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2485                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2486         else
2487                 need_bytes = tic->t_unit_res;
2488
2489 redo:
2490         if (XLOG_FORCED_SHUTDOWN(log))
2491                 goto error_return;
2492
2493         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2494                                      log->l_grant_reserve_bytes);
2495         if (free_bytes < need_bytes) {
2496                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2497                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2498                 xlog_trace_loggrant(log, tic,
2499                                     "xlog_grant_log_space: sleep 2");
2500                 XFS_STATS_INC(xs_sleep_logspace);
2501                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2502
2503                 if (XLOG_FORCED_SHUTDOWN(log)) {
2504                         s = GRANT_LOCK(log);
2505                         goto error_return;
2506                 }
2507
2508                 xlog_trace_loggrant(log, tic,
2509                                     "xlog_grant_log_space: wake 2");
2510                 xlog_grant_push_ail(log->l_mp, need_bytes);
2511                 s = GRANT_LOCK(log);
2512                 goto redo;
2513         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2514                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2515
2516         /* we've got enough space */
2517         xlog_grant_add_space(log, need_bytes);
2518 #ifdef DEBUG
2519         tail_lsn = log->l_tail_lsn;
2520         /*
2521          * Check to make sure the grant write head didn't just over lap the
2522          * tail.  If the cycles are the same, we can't be overlapping.
2523          * Otherwise, make sure that the cycles differ by exactly one and
2524          * check the byte count.
2525          */
2526         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2527                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2528                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2529         }
2530 #endif
2531         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2532         xlog_verify_grant_head(log, 1);
2533         GRANT_UNLOCK(log, s);
2534         return 0;
2535
2536  error_return:
2537         if (tic->t_flags & XLOG_TIC_IN_Q)
2538                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2539         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2540         /*
2541          * If we are failing, make sure the ticket doesn't have any
2542          * current reservations. We don't want to add this back when
2543          * the ticket/transaction gets cancelled.
2544          */
2545         tic->t_curr_res = 0;
2546         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2547         GRANT_UNLOCK(log, s);
2548         return XFS_ERROR(EIO);
2549 }       /* xlog_grant_log_space */
2550
2551
2552 /*
2553  * Replenish the byte reservation required by moving the grant write head.
2554  *
2555  *
2556  */
2557 STATIC int
2558 xlog_regrant_write_log_space(xlog_t        *log,
2559                              xlog_ticket_t *tic)
2560 {
2561         SPLDECL(s);
2562         int             free_bytes, need_bytes;
2563         xlog_ticket_t   *ntic;
2564 #ifdef DEBUG
2565         xfs_lsn_t       tail_lsn;
2566 #endif
2567
2568         tic->t_curr_res = tic->t_unit_res;
2569         XLOG_TIC_RESET_RES(tic);
2570
2571         if (tic->t_cnt > 0)
2572                 return 0;
2573
2574 #ifdef DEBUG
2575         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2576                 panic("regrant Recovery problem");
2577 #endif
2578
2579         s = GRANT_LOCK(log);
2580         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2581
2582         if (XLOG_FORCED_SHUTDOWN(log))
2583                 goto error_return;
2584
2585         /* If there are other waiters on the queue then give them a
2586          * chance at logspace before us. Wake up the first waiters,
2587          * if we do not wake up all the waiters then go to sleep waiting
2588          * for more free space, otherwise try to get some space for
2589          * this transaction.
2590          */
2591
2592         if ((ntic = log->l_write_headq)) {
2593                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2594                                              log->l_grant_write_bytes);
2595                 do {
2596                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2597
2598                         if (free_bytes < ntic->t_unit_res)
2599                                 break;
2600                         free_bytes -= ntic->t_unit_res;
2601                         sv_signal(&ntic->t_sema);
2602                         ntic = ntic->t_next;
2603                 } while (ntic != log->l_write_headq);
2604
2605                 if (ntic != log->l_write_headq) {
2606                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2607                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2608
2609                         xlog_trace_loggrant(log, tic,
2610                                     "xlog_regrant_write_log_space: sleep 1");
2611                         XFS_STATS_INC(xs_sleep_logspace);
2612                         sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2613                                 &log->l_grant_lock, s);
2614
2615                         /* If we're shutting down, this tic is already
2616                          * off the queue */
2617                         if (XLOG_FORCED_SHUTDOWN(log)) {
2618                                 s = GRANT_LOCK(log);
2619                                 goto error_return;
2620                         }
2621
2622                         xlog_trace_loggrant(log, tic,
2623                                     "xlog_regrant_write_log_space: wake 1");
2624                         xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2625                         s = GRANT_LOCK(log);
2626                 }
2627         }
2628
2629         need_bytes = tic->t_unit_res;
2630
2631 redo:
2632         if (XLOG_FORCED_SHUTDOWN(log))
2633                 goto error_return;
2634
2635         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2636                                      log->l_grant_write_bytes);
2637         if (free_bytes < need_bytes) {
2638                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2639                         xlog_ins_ticketq(&log->l_write_headq, tic);
2640                 XFS_STATS_INC(xs_sleep_logspace);
2641                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2642
2643                 /* If we're shutting down, this tic is already off the queue */
2644                 if (XLOG_FORCED_SHUTDOWN(log)) {
2645                         s = GRANT_LOCK(log);
2646                         goto error_return;
2647                 }
2648
2649                 xlog_trace_loggrant(log, tic,
2650                                     "xlog_regrant_write_log_space: wake 2");
2651                 xlog_grant_push_ail(log->l_mp, need_bytes);
2652                 s = GRANT_LOCK(log);
2653                 goto redo;
2654         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2655                 xlog_del_ticketq(&log->l_write_headq, tic);
2656
2657         /* we've got enough space */
2658         xlog_grant_add_space_write(log, need_bytes);
2659 #ifdef DEBUG
2660         tail_lsn = log->l_tail_lsn;
2661         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2662                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2663                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2664         }
2665 #endif
2666
2667         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2668         xlog_verify_grant_head(log, 1);
2669         GRANT_UNLOCK(log, s);
2670         return 0;
2671
2672
2673  error_return:
2674         if (tic->t_flags & XLOG_TIC_IN_Q)
2675                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2676         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2677         /*
2678          * If we are failing, make sure the ticket doesn't have any
2679          * current reservations. We don't want to add this back when
2680          * the ticket/transaction gets cancelled.
2681          */
2682         tic->t_curr_res = 0;
2683         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2684         GRANT_UNLOCK(log, s);
2685         return XFS_ERROR(EIO);
2686 }       /* xlog_regrant_write_log_space */
2687
2688
2689 /* The first cnt-1 times through here we don't need to
2690  * move the grant write head because the permanent
2691  * reservation has reserved cnt times the unit amount.
2692  * Release part of current permanent unit reservation and
2693  * reset current reservation to be one units worth.  Also
2694  * move grant reservation head forward.
2695  */
2696 STATIC void
2697 xlog_regrant_reserve_log_space(xlog_t        *log,
2698                                xlog_ticket_t *ticket)
2699 {
2700         SPLDECL(s);
2701
2702         xlog_trace_loggrant(log, ticket,
2703                             "xlog_regrant_reserve_log_space: enter");
2704         if (ticket->t_cnt > 0)
2705                 ticket->t_cnt--;
2706
2707         s = GRANT_LOCK(log);
2708         xlog_grant_sub_space(log, ticket->t_curr_res);
2709         ticket->t_curr_res = ticket->t_unit_res;
2710         XLOG_TIC_RESET_RES(ticket);
2711         xlog_trace_loggrant(log, ticket,
2712                             "xlog_regrant_reserve_log_space: sub current res");
2713         xlog_verify_grant_head(log, 1);
2714
2715         /* just return if we still have some of the pre-reserved space */
2716         if (ticket->t_cnt > 0) {
2717                 GRANT_UNLOCK(log, s);
2718                 return;
2719         }
2720
2721         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2722         xlog_trace_loggrant(log, ticket,
2723                             "xlog_regrant_reserve_log_space: exit");
2724         xlog_verify_grant_head(log, 0);
2725         GRANT_UNLOCK(log, s);
2726         ticket->t_curr_res = ticket->t_unit_res;
2727         XLOG_TIC_RESET_RES(ticket);
2728 }       /* xlog_regrant_reserve_log_space */
2729
2730
2731 /*
2732  * Give back the space left from a reservation.
2733  *
2734  * All the information we need to make a correct determination of space left
2735  * is present.  For non-permanent reservations, things are quite easy.  The
2736  * count should have been decremented to zero.  We only need to deal with the
2737  * space remaining in the current reservation part of the ticket.  If the
2738  * ticket contains a permanent reservation, there may be left over space which
2739  * needs to be released.  A count of N means that N-1 refills of the current
2740  * reservation can be done before we need to ask for more space.  The first
2741  * one goes to fill up the first current reservation.  Once we run out of
2742  * space, the count will stay at zero and the only space remaining will be
2743  * in the current reservation field.
2744  */
2745 STATIC void
2746 xlog_ungrant_log_space(xlog_t        *log,
2747                        xlog_ticket_t *ticket)
2748 {
2749         SPLDECL(s);
2750
2751         if (ticket->t_cnt > 0)
2752                 ticket->t_cnt--;
2753
2754         s = GRANT_LOCK(log);
2755         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2756
2757         xlog_grant_sub_space(log, ticket->t_curr_res);
2758
2759         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2760
2761         /* If this is a permanent reservation ticket, we may be able to free
2762          * up more space based on the remaining count.
2763          */
2764         if (ticket->t_cnt > 0) {
2765                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2766                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2767         }
2768
2769         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2770         xlog_verify_grant_head(log, 1);
2771         GRANT_UNLOCK(log, s);
2772         xfs_log_move_tail(log->l_mp, 1);
2773 }       /* xlog_ungrant_log_space */
2774
2775
2776 /*
2777  * Atomically put back used ticket.
2778  */
2779 void
2780 xlog_state_put_ticket(xlog_t        *log,
2781                       xlog_ticket_t *tic)
2782 {
2783         unsigned long s;
2784
2785         s = LOG_LOCK(log);
2786         xlog_ticket_put(log, tic);
2787         LOG_UNLOCK(log, s);
2788 }       /* xlog_state_put_ticket */
2789
2790 /*
2791  * Flush iclog to disk if this is the last reference to the given iclog and
2792  * the WANT_SYNC bit is set.
2793  *
2794  * When this function is entered, the iclog is not necessarily in the
2795  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2796  *
2797  *
2798  */
2799 int
2800 xlog_state_release_iclog(xlog_t         *log,
2801                          xlog_in_core_t *iclog)
2802 {
2803         SPLDECL(s);
2804         int             sync = 0;       /* do we sync? */
2805
2806         xlog_assign_tail_lsn(log->l_mp);
2807
2808         s = LOG_LOCK(log);
2809
2810         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2811                 LOG_UNLOCK(log, s);
2812                 return XFS_ERROR(EIO);
2813         }
2814
2815         ASSERT(iclog->ic_refcnt > 0);
2816         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2817                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2818
2819         if (--iclog->ic_refcnt == 0 &&
2820             iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2821                 sync++;
2822                 iclog->ic_state = XLOG_STATE_SYNCING;
2823                 INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2824                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2825                 /* cycle incremented when incrementing curr_block */
2826         }
2827
2828         LOG_UNLOCK(log, s);
2829
2830         /*
2831          * We let the log lock go, so it's possible that we hit a log I/O
2832          * error or someother SHUTDOWN condition that marks the iclog
2833          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2834          * this iclog has consistent data, so we ignore IOERROR
2835          * flags after this point.
2836          */
2837         if (sync) {
2838                 return xlog_sync(log, iclog);
2839         }
2840         return 0;
2841
2842 }       /* xlog_state_release_iclog */
2843
2844
2845 /*
2846  * This routine will mark the current iclog in the ring as WANT_SYNC
2847  * and move the current iclog pointer to the next iclog in the ring.
2848  * When this routine is called from xlog_state_get_iclog_space(), the
2849  * exact size of the iclog has not yet been determined.  All we know is
2850  * that every data block.  We have run out of space in this log record.
2851  */
2852 STATIC void
2853 xlog_state_switch_iclogs(xlog_t         *log,
2854                          xlog_in_core_t *iclog,
2855                          int            eventual_size)
2856 {
2857         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2858         if (!eventual_size)
2859                 eventual_size = iclog->ic_offset;
2860         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2861         INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2862         log->l_prev_block = log->l_curr_block;
2863         log->l_prev_cycle = log->l_curr_cycle;
2864
2865         /* roll log?: ic_offset changed later */
2866         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2867
2868         /* Round up to next log-sunit */
2869         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2870             log->l_mp->m_sb.sb_logsunit > 1) {
2871                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2872                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2873         }
2874
2875         if (log->l_curr_block >= log->l_logBBsize) {
2876                 log->l_curr_cycle++;
2877                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2878                         log->l_curr_cycle++;
2879                 log->l_curr_block -= log->l_logBBsize;
2880                 ASSERT(log->l_curr_block >= 0);
2881         }
2882         ASSERT(iclog == log->l_iclog);
2883         log->l_iclog = iclog->ic_next;
2884 }       /* xlog_state_switch_iclogs */
2885
2886
2887 /*
2888  * Write out all data in the in-core log as of this exact moment in time.
2889  *
2890  * Data may be written to the in-core log during this call.  However,
2891  * we don't guarantee this data will be written out.  A change from past
2892  * implementation means this routine will *not* write out zero length LRs.
2893  *
2894  * Basically, we try and perform an intelligent scan of the in-core logs.
2895  * If we determine there is no flushable data, we just return.  There is no
2896  * flushable data if:
2897  *
2898  *      1. the current iclog is active and has no data; the previous iclog
2899  *              is in the active or dirty state.
2900  *      2. the current iclog is drity, and the previous iclog is in the
2901  *              active or dirty state.
2902  *
2903  * We may sleep (call psema) if:
2904  *
2905  *      1. the current iclog is not in the active nor dirty state.
2906  *      2. the current iclog dirty, and the previous iclog is not in the
2907  *              active nor dirty state.
2908  *      3. the current iclog is active, and there is another thread writing
2909  *              to this particular iclog.
2910  *      4. a) the current iclog is active and has no other writers
2911  *         b) when we return from flushing out this iclog, it is still
2912  *              not in the active nor dirty state.
2913  */
2914 STATIC int
2915 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2916 {
2917         xlog_in_core_t  *iclog;
2918         xfs_lsn_t       lsn;
2919         SPLDECL(s);
2920
2921         s = LOG_LOCK(log);
2922
2923         iclog = log->l_iclog;
2924         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2925                 LOG_UNLOCK(log, s);
2926                 return XFS_ERROR(EIO);
2927         }
2928
2929         /* If the head iclog is not active nor dirty, we just attach
2930          * ourselves to the head and go to sleep.
2931          */
2932         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2933             iclog->ic_state == XLOG_STATE_DIRTY) {
2934                 /*
2935                  * If the head is dirty or (active and empty), then
2936                  * we need to look at the previous iclog.  If the previous
2937                  * iclog is active or dirty we are done.  There is nothing
2938                  * to sync out.  Otherwise, we attach ourselves to the
2939                  * previous iclog and go to sleep.
2940                  */
2941                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2942                     (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2943                         iclog = iclog->ic_prev;
2944                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2945                             iclog->ic_state == XLOG_STATE_DIRTY)
2946                                 goto no_sleep;
2947                         else
2948                                 goto maybe_sleep;
2949                 } else {
2950                         if (iclog->ic_refcnt == 0) {
2951                                 /* We are the only one with access to this
2952                                  * iclog.  Flush it out now.  There should
2953                                  * be a roundoff of zero to show that someone
2954                                  * has already taken care of the roundoff from
2955                                  * the previous sync.
2956                                  */
2957                                 iclog->ic_refcnt++;
2958                                 lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2959                                 xlog_state_switch_iclogs(log, iclog, 0);
2960                                 LOG_UNLOCK(log, s);
2961
2962                                 if (xlog_state_release_iclog(log, iclog))
2963                                         return XFS_ERROR(EIO);
2964                                 *log_flushed = 1;
2965                                 s = LOG_LOCK(log);
2966                                 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2967                                     iclog->ic_state != XLOG_STATE_DIRTY)
2968                                         goto maybe_sleep;
2969                                 else
2970                                         goto no_sleep;
2971                         } else {
2972                                 /* Someone else is writing to this iclog.
2973                                  * Use its call to flush out the data.  However,
2974                                  * the other thread may not force out this LR,
2975                                  * so we mark it WANT_SYNC.
2976                                  */
2977                                 xlog_state_switch_iclogs(log, iclog, 0);
2978                                 goto maybe_sleep;
2979                         }
2980                 }
2981         }
2982
2983         /* By the time we come around again, the iclog could've been filled
2984          * which would give it another lsn.  If we have a new lsn, just
2985          * return because the relevant data has been flushed.
2986          */
2987 maybe_sleep:
2988         if (flags & XFS_LOG_SYNC) {
2989                 /*
2990                  * We must check if we're shutting down here, before
2991                  * we wait, while we're holding the LOG_LOCK.
2992                  * Then we check again after waking up, in case our
2993                  * sleep was disturbed by a bad news.
2994                  */
2995                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2996                         LOG_UNLOCK(log, s);
2997                         return XFS_ERROR(EIO);
2998                 }
2999                 XFS_STATS_INC(xs_log_force_sleep);
3000                 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3001                 /*
3002                  * No need to grab the log lock here since we're
3003                  * only deciding whether or not to return EIO
3004                  * and the memory read should be atomic.
3005                  */
3006                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3007                         return XFS_ERROR(EIO);
3008                 *log_flushed = 1;
3009
3010         } else {
3011
3012 no_sleep:
3013                 LOG_UNLOCK(log, s);
3014         }
3015         return 0;
3016 }       /* xlog_state_sync_all */
3017
3018
3019 /*
3020  * Used by code which implements synchronous log forces.
3021  *
3022  * Find in-core log with lsn.
3023  *      If it is in the DIRTY state, just return.
3024  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3025  *              state and go to sleep or return.
3026  *      If it is in any other state, go to sleep or return.
3027  *
3028  * If filesystem activity goes to zero, the iclog will get flushed only by
3029  * bdflush().
3030  */
3031 int
3032 xlog_state_sync(xlog_t    *log,
3033                 xfs_lsn_t lsn,
3034                 uint      flags,
3035                 int       *log_flushed)
3036 {
3037     xlog_in_core_t      *iclog;
3038     int                 already_slept = 0;
3039     SPLDECL(s);
3040
3041
3042 try_again:
3043     s = LOG_LOCK(log);
3044     iclog = log->l_iclog;
3045
3046     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3047             LOG_UNLOCK(log, s);
3048             return XFS_ERROR(EIO);
3049     }
3050
3051     do {
3052         if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3053             iclog = iclog->ic_next;
3054             continue;
3055         }
3056
3057         if (iclog->ic_state == XLOG_STATE_DIRTY) {
3058                 LOG_UNLOCK(log, s);
3059                 return 0;
3060         }
3061
3062         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3063                 /*
3064                  * We sleep here if we haven't already slept (e.g.
3065                  * this is the first time we've looked at the correct
3066                  * iclog buf) and the buffer before us is going to
3067                  * be sync'ed. The reason for this is that if we
3068                  * are doing sync transactions here, by waiting for
3069                  * the previous I/O to complete, we can allow a few
3070                  * more transactions into this iclog before we close
3071                  * it down.
3072                  *
3073                  * Otherwise, we mark the buffer WANT_SYNC, and bump
3074                  * up the refcnt so we can release the log (which drops
3075                  * the ref count).  The state switch keeps new transaction
3076                  * commits from using this buffer.  When the current commits
3077                  * finish writing into the buffer, the refcount will drop to
3078                  * zero and the buffer will go out then.
3079                  */
3080                 if (!already_slept &&
3081                     (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3082                                                  XLOG_STATE_SYNCING))) {
3083                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3084                         XFS_STATS_INC(xs_log_force_sleep);
3085                         sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3086                                 &log->l_icloglock, s);
3087                         *log_flushed = 1;
3088                         already_slept = 1;
3089                         goto try_again;
3090                 } else {
3091                         iclog->ic_refcnt++;
3092                         xlog_state_switch_iclogs(log, iclog, 0);
3093                         LOG_UNLOCK(log, s);
3094                         if (xlog_state_release_iclog(log, iclog))
3095                                 return XFS_ERROR(EIO);
3096                         *log_flushed = 1;
3097                         s = LOG_LOCK(log);
3098                 }
3099         }
3100
3101         if ((flags & XFS_LOG_SYNC) && /* sleep */
3102             !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3103
3104                 /*
3105                  * Don't wait on the forcesema if we know that we've
3106                  * gotten a log write error.
3107                  */
3108                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3109                         LOG_UNLOCK(log, s);
3110                         return XFS_ERROR(EIO);
3111                 }
3112                 XFS_STATS_INC(xs_log_force_sleep);
3113                 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3114                 /*
3115                  * No need to grab the log lock here since we're
3116                  * only deciding whether or not to return EIO
3117                  * and the memory read should be atomic.
3118                  */
3119                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3120                         return XFS_ERROR(EIO);
3121                 *log_flushed = 1;
3122         } else {                /* just return */
3123                 LOG_UNLOCK(log, s);
3124         }
3125         return 0;
3126
3127     } while (iclog != log->l_iclog);
3128
3129     LOG_UNLOCK(log, s);
3130     return 0;
3131 }       /* xlog_state_sync */
3132
3133
3134 /*
3135  * Called when we want to mark the current iclog as being ready to sync to
3136  * disk.
3137  */
3138 void
3139 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3140 {
3141         SPLDECL(s);
3142
3143         s = LOG_LOCK(log);
3144
3145         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3146                 xlog_state_switch_iclogs(log, iclog, 0);
3147         } else {
3148                 ASSERT(iclog->ic_state &
3149                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3150         }
3151
3152         LOG_UNLOCK(log, s);
3153 }       /* xlog_state_want_sync */
3154
3155
3156
3157 /*****************************************************************************
3158  *
3159  *              TICKET functions
3160  *
3161  *****************************************************************************
3162  */
3163
3164 /*
3165  *      Algorithm doesn't take into account page size. ;-(
3166  */
3167 STATIC void
3168 xlog_state_ticket_alloc(xlog_t *log)
3169 {
3170         xlog_ticket_t   *t_list;
3171         xlog_ticket_t   *next;
3172         xfs_caddr_t     buf;
3173         uint            i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3174         SPLDECL(s);
3175
3176         /*
3177          * The kmem_zalloc may sleep, so we shouldn't be holding the
3178          * global lock.  XXXmiken: may want to use zone allocator.
3179          */
3180         buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3181
3182         s = LOG_LOCK(log);
3183
3184         /* Attach 1st ticket to Q, so we can keep track of allocated memory */
3185         t_list = (xlog_ticket_t *)buf;
3186         t_list->t_next = log->l_unmount_free;
3187         log->l_unmount_free = t_list++;
3188         log->l_ticket_cnt++;
3189         log->l_ticket_tcnt++;
3190
3191         /* Next ticket becomes first ticket attached to ticket free list */
3192         if (log->l_freelist != NULL) {
3193                 ASSERT(log->l_tail != NULL);
3194                 log->l_tail->t_next = t_list;
3195         } else {
3196                 log->l_freelist = t_list;
3197         }
3198         log->l_ticket_cnt++;
3199         log->l_ticket_tcnt++;
3200
3201         /* Cycle through rest of alloc'ed memory, building up free Q */
3202         for ( ; i > 0; i--) {
3203                 next = t_list + 1;
3204                 t_list->t_next = next;
3205                 t_list = next;
3206                 log->l_ticket_cnt++;
3207                 log->l_ticket_tcnt++;
3208         }
3209         t_list->t_next = NULL;
3210         log->l_tail = t_list;
3211         LOG_UNLOCK(log, s);
3212 }       /* xlog_state_ticket_alloc */
3213
3214
3215 /*
3216  * Put ticket into free list
3217  *
3218  * Assumption: log lock is held around this call.
3219  */
3220 STATIC void
3221 xlog_ticket_put(xlog_t          *log,
3222                 xlog_ticket_t   *ticket)
3223 {
3224         sv_destroy(&ticket->t_sema);
3225
3226         /*
3227          * Don't think caching will make that much difference.  It's
3228          * more important to make debug easier.
3229          */
3230 #if 0
3231         /* real code will want to use LIFO for caching */
3232         ticket->t_next = log->l_freelist;
3233         log->l_freelist = ticket;
3234         /* no need to clear fields */
3235 #else
3236         /* When we debug, it is easier if tickets are cycled */
3237         ticket->t_next     = NULL;
3238         if (log->l_tail != 0) {
3239                 log->l_tail->t_next = ticket;
3240         } else {
3241                 ASSERT(log->l_freelist == 0);
3242                 log->l_freelist = ticket;
3243         }
3244         log->l_tail         = ticket;
3245 #endif /* DEBUG */
3246         log->l_ticket_cnt++;
3247 }       /* xlog_ticket_put */
3248
3249
3250 /*
3251  * Grab ticket off freelist or allocation some more
3252  */
3253 xlog_ticket_t *
3254 xlog_ticket_get(xlog_t          *log,
3255                 int             unit_bytes,
3256                 int             cnt,
3257                 char            client,
3258                 uint            xflags)
3259 {
3260         xlog_ticket_t   *tic;
3261         uint            num_headers;
3262         SPLDECL(s);
3263
3264  alloc:
3265         if (log->l_freelist == NULL)
3266                 xlog_state_ticket_alloc(log);           /* potentially sleep */
3267
3268         s = LOG_LOCK(log);
3269         if (log->l_freelist == NULL) {
3270                 LOG_UNLOCK(log, s);
3271                 goto alloc;
3272         }
3273         tic             = log->l_freelist;
3274         log->l_freelist = tic->t_next;
3275         if (log->l_freelist == NULL)
3276                 log->l_tail = NULL;
3277         log->l_ticket_cnt--;
3278         LOG_UNLOCK(log, s);
3279
3280         /*
3281          * Permanent reservations have up to 'cnt'-1 active log operations
3282          * in the log.  A unit in this case is the amount of space for one
3283          * of these log operations.  Normal reservations have a cnt of 1
3284          * and their unit amount is the total amount of space required.
3285          *
3286          * The following lines of code account for non-transaction data
3287          * which occupy space in the on-disk log.
3288          *
3289          * Normal form of a transaction is:
3290          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3291          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3292          *
3293          * We need to account for all the leadup data and trailer data
3294          * around the transaction data.
3295          * And then we need to account for the worst case in terms of using
3296          * more space.
3297          * The worst case will happen if:
3298          * - the placement of the transaction happens to be such that the
3299          *   roundoff is at its maximum
3300          * - the transaction data is synced before the commit record is synced
3301          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3302          *   Therefore the commit record is in its own Log Record.
3303          *   This can happen as the commit record is called with its
3304          *   own region to xlog_write().
3305          *   This then means that in the worst case, roundoff can happen for
3306          *   the commit-rec as well.
3307          *   The commit-rec is smaller than padding in this scenario and so it is
3308          *   not added separately.
3309          */
3310
3311         /* for trans header */
3312         unit_bytes += sizeof(xlog_op_header_t);
3313         unit_bytes += sizeof(xfs_trans_header_t);
3314
3315         /* for start-rec */
3316         unit_bytes += sizeof(xlog_op_header_t);
3317
3318         /* for LR headers */
3319         num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3320         unit_bytes += log->l_iclog_hsize * num_headers;
3321
3322         /* for commit-rec LR header - note: padding will subsume the ophdr */
3323         unit_bytes += log->l_iclog_hsize;
3324
3325         /* for split-recs - ophdrs added when data split over LRs */
3326         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3327
3328         /* for roundoff padding for transaction data and one for commit record */
3329         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3330             log->l_mp->m_sb.sb_logsunit > 1) {
3331                 /* log su roundoff */
3332                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3333         } else {
3334                 /* BB roundoff */
3335                 unit_bytes += 2*BBSIZE;
3336         }
3337
3338         tic->t_unit_res         = unit_bytes;
3339         tic->t_curr_res         = unit_bytes;
3340         tic->t_cnt              = cnt;
3341         tic->t_ocnt             = cnt;
3342         tic->t_tid              = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3343         tic->t_clientid         = client;
3344         tic->t_flags            = XLOG_TIC_INITED;
3345         tic->t_trans_type       = 0;
3346         if (xflags & XFS_LOG_PERM_RESERV)
3347                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3348         sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3349
3350         XLOG_TIC_RESET_RES(tic);
3351
3352         return tic;
3353 }       /* xlog_ticket_get */
3354
3355
3356 /******************************************************************************
3357  *
3358  *              Log debug routines
3359  *
3360  ******************************************************************************
3361  */
3362 #if defined(DEBUG)
3363 /*
3364  * Make sure that the destination ptr is within the valid data region of
3365  * one of the iclogs.  This uses backup pointers stored in a different
3366  * part of the log in case we trash the log structure.
3367  */
3368 void
3369 xlog_verify_dest_ptr(xlog_t     *log,
3370                      __psint_t  ptr)
3371 {
3372         int i;
3373         int good_ptr = 0;
3374
3375         for (i=0; i < log->l_iclog_bufs; i++) {
3376                 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3377                     ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3378                         good_ptr++;
3379         }
3380         if (! good_ptr)
3381                 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3382 }       /* xlog_verify_dest_ptr */
3383
3384 STATIC void
3385 xlog_verify_grant_head(xlog_t *log, int equals)
3386 {
3387     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3388         if (equals)
3389             ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3390         else
3391             ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3392     } else {
3393         ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3394         ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3395     }
3396 }       /* xlog_verify_grant_head */
3397
3398 /* check if it will fit */
3399 STATIC void
3400 xlog_verify_tail_lsn(xlog_t         *log,
3401                      xlog_in_core_t *iclog,
3402                      xfs_lsn_t      tail_lsn)
3403 {
3404     int blocks;
3405
3406     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3407         blocks =
3408             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3409         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3410             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3411     } else {
3412         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3413
3414         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3415             xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3416
3417         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3418         if (blocks < BTOBB(iclog->ic_offset) + 1)
3419             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3420     }
3421 }       /* xlog_verify_tail_lsn */
3422
3423 /*
3424  * Perform a number of checks on the iclog before writing to disk.
3425  *
3426  * 1. Make sure the iclogs are still circular
3427  * 2. Make sure we have a good magic number
3428  * 3. Make sure we don't have magic numbers in the data
3429  * 4. Check fields of each log operation header for:
3430  *      A. Valid client identifier
3431  *      B. tid ptr value falls in valid ptr space (user space code)
3432  *      C. Length in log record header is correct according to the
3433  *              individual operation headers within record.
3434  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3435  *      log, check the preceding blocks of the physical log to make sure all
3436  *      the cycle numbers agree with the current cycle number.
3437  */
3438 STATIC void
3439 xlog_verify_iclog(xlog_t         *log,
3440                   xlog_in_core_t *iclog,
3441                   int            count,
3442                   boolean_t      syncing)
3443 {
3444         xlog_op_header_t        *ophead;
3445         xlog_in_core_t          *icptr;
3446         xlog_in_core_2_t        *xhdr;
3447         xfs_caddr_t             ptr;
3448         xfs_caddr_t             base_ptr;
3449         __psint_t               field_offset;
3450         __uint8_t               clientid;
3451         int                     len, i, j, k, op_len;
3452         int                     idx;
3453         SPLDECL(s);
3454
3455         /* check validity of iclog pointers */
3456         s = LOG_LOCK(log);
3457         icptr = log->l_iclog;
3458         for (i=0; i < log->l_iclog_bufs; i++) {
3459                 if (icptr == 0)
3460                         xlog_panic("xlog_verify_iclog: invalid ptr");
3461                 icptr = icptr->ic_next;
3462         }
3463         if (icptr != log->l_iclog)
3464                 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3465         LOG_UNLOCK(log, s);
3466
3467         /* check log magic numbers */
3468         ptr = (xfs_caddr_t) &(iclog->ic_header);
3469         if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3470                 xlog_panic("xlog_verify_iclog: invalid magic num");
3471
3472         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3473              ptr += BBSIZE) {
3474                 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3475                         xlog_panic("xlog_verify_iclog: unexpected magic num");
3476         }
3477
3478         /* check fields */
3479         len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3480         ptr = iclog->ic_datap;
3481         base_ptr = ptr;
3482         ophead = (xlog_op_header_t *)ptr;
3483         xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3484         for (i = 0; i < len; i++) {
3485                 ophead = (xlog_op_header_t *)ptr;
3486
3487                 /* clientid is only 1 byte */
3488                 field_offset = (__psint_t)
3489                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3490                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3491                         clientid = ophead->oh_clientid;
3492                 } else {
3493                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3494                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3495                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3496                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3497                                 clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3498                         } else {
3499                                 clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3500                         }
3501                 }
3502                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3503                         cmn_err(CE_WARN, "xlog_verify_iclog: "
3504                                 "invalid clientid %d op 0x%p offset 0x%lx",
3505                                 clientid, ophead, (unsigned long)field_offset);
3506
3507                 /* check length */
3508                 field_offset = (__psint_t)
3509                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3510                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3511                         op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3512                 } else {
3513                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3514                                     (__psint_t)iclog->ic_datap);
3515                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3516                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3517                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3518                                 op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3519                         } else {
3520                                 op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3521                         }
3522                 }
3523                 ptr += sizeof(xlog_op_header_t) + op_len;
3524         }
3525 }       /* xlog_verify_iclog */
3526 #endif
3527
3528 /*
3529  * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3530  */
3531 STATIC int
3532 xlog_state_ioerror(
3533         xlog_t  *log)
3534 {
3535         xlog_in_core_t  *iclog, *ic;
3536
3537         iclog = log->l_iclog;
3538         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3539                 /*
3540                  * Mark all the incore logs IOERROR.
3541                  * From now on, no log flushes will result.
3542                  */
3543                 ic = iclog;
3544                 do {
3545                         ic->ic_state = XLOG_STATE_IOERROR;
3546                         ic = ic->ic_next;
3547                 } while (ic != iclog);
3548                 return 0;
3549         }
3550         /*
3551          * Return non-zero, if state transition has already happened.
3552          */
3553         return 1;
3554 }
3555
3556 /*
3557  * This is called from xfs_force_shutdown, when we're forcibly
3558  * shutting down the filesystem, typically because of an IO error.
3559  * Our main objectives here are to make sure that:
3560  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3561  *         parties to find out, 'atomically'.
3562  *      b. those who're sleeping on log reservations, pinned objects and
3563  *          other resources get woken up, and be told the bad news.
3564  *      c. nothing new gets queued up after (a) and (b) are done.
3565  *      d. if !logerror, flush the iclogs to disk, then seal them off
3566  *         for business.
3567  */
3568 int
3569 xfs_log_force_umount(
3570         struct xfs_mount        *mp,
3571         int                     logerror)
3572 {
3573         xlog_ticket_t   *tic;
3574         xlog_t          *log;
3575         int             retval;
3576         int             dummy;
3577         SPLDECL(s);
3578         SPLDECL(s2);
3579
3580         log = mp->m_log;
3581
3582         /*
3583          * If this happens during log recovery, don't worry about
3584          * locking; the log isn't open for business yet.
3585          */
3586         if (!log ||
3587             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3588                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3589                 XFS_BUF_DONE(mp->m_sb_bp);
3590                 return 0;
3591         }
3592
3593         /*
3594          * Somebody could've already done the hard work for us.
3595          * No need to get locks for this.
3596          */
3597         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3598                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3599                 return 1;
3600         }
3601         retval = 0;
3602         /*
3603          * We must hold both the GRANT lock and the LOG lock,
3604          * before we mark the filesystem SHUTDOWN and wake
3605          * everybody up to tell the bad news.
3606          */
3607         s = GRANT_LOCK(log);
3608         s2 = LOG_LOCK(log);
3609         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3610         XFS_BUF_DONE(mp->m_sb_bp);
3611         /*
3612          * This flag is sort of redundant because of the mount flag, but
3613          * it's good to maintain the separation between the log and the rest
3614          * of XFS.
3615          */
3616         log->l_flags |= XLOG_IO_ERROR;
3617
3618         /*
3619          * If we hit a log error, we want to mark all the iclogs IOERROR
3620          * while we're still holding the loglock.
3621          */
3622         if (logerror)
3623                 retval = xlog_state_ioerror(log);
3624         LOG_UNLOCK(log, s2);
3625
3626         /*
3627          * We don't want anybody waiting for log reservations
3628          * after this. That means we have to wake up everybody
3629          * queued up on reserve_headq as well as write_headq.
3630          * In addition, we make sure in xlog_{re}grant_log_space
3631          * that we don't enqueue anything once the SHUTDOWN flag
3632          * is set, and this action is protected by the GRANTLOCK.
3633          */
3634         if ((tic = log->l_reserve_headq)) {
3635                 do {
3636                         sv_signal(&tic->t_sema);
3637                         tic = tic->t_next;
3638                 } while (tic != log->l_reserve_headq);
3639         }
3640
3641         if ((tic = log->l_write_headq)) {
3642                 do {
3643                         sv_signal(&tic->t_sema);
3644                         tic = tic->t_next;
3645                 } while (tic != log->l_write_headq);
3646         }
3647         GRANT_UNLOCK(log, s);
3648
3649         if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3650                 ASSERT(!logerror);
3651                 /*
3652                  * Force the incore logs to disk before shutting the
3653                  * log down completely.
3654                  */
3655                 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3656                 s2 = LOG_LOCK(log);
3657                 retval = xlog_state_ioerror(log);
3658                 LOG_UNLOCK(log, s2);
3659         }
3660         /*
3661          * Wake up everybody waiting on xfs_log_force.
3662          * Callback all log item committed functions as if the
3663          * log writes were completed.
3664          */
3665         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3666
3667 #ifdef XFSERRORDEBUG
3668         {
3669                 xlog_in_core_t  *iclog;
3670
3671                 s = LOG_LOCK(log);
3672                 iclog = log->l_iclog;
3673                 do {
3674                         ASSERT(iclog->ic_callback == 0);
3675                         iclog = iclog->ic_next;
3676                 } while (iclog != log->l_iclog);
3677                 LOG_UNLOCK(log, s);
3678         }
3679 #endif
3680         /* return non-zero if log IOERROR transition had already happened */
3681         return retval;
3682 }
3683
3684 STATIC int
3685 xlog_iclogs_empty(xlog_t *log)
3686 {
3687         xlog_in_core_t  *iclog;
3688
3689         iclog = log->l_iclog;
3690         do {
3691                 /* endianness does not matter here, zero is zero in
3692                  * any language.
3693                  */
3694                 if (iclog->ic_header.h_num_logops)
3695                         return 0;
3696                 iclog = iclog->ic_next;
3697         } while (iclog != log->l_iclog);
3698         return 1;
3699 }