Merge git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[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_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
42 #include "xfs_rw.h"
43
44
45 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
46         { (ptr) += (bytes); \
47           (len) -= (bytes); \
48           (off) += (bytes);}
49
50 /* Local miscellaneous function prototypes */
51 STATIC int       xlog_bdstrat_cb(struct xfs_buf *);
52 STATIC int       xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
53                                     xlog_in_core_t **, xfs_lsn_t *);
54 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
55                                 xfs_buftarg_t   *log_target,
56                                 xfs_daddr_t     blk_offset,
57                                 int             num_bblks);
58 STATIC int       xlog_space_left(xlog_t *log, int cycle, int bytes);
59 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
60 STATIC void      xlog_dealloc_log(xlog_t *log);
61 STATIC int       xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
62                             int nentries, xfs_log_ticket_t tic,
63                             xfs_lsn_t *start_lsn,
64                             xlog_in_core_t **commit_iclog,
65                             uint flags);
66
67 /* local state machine functions */
68 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
70 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
71                                        int              len,
72                                        xlog_in_core_t   **iclog,
73                                        xlog_ticket_t    *ticket,
74                                        int              *continued_write,
75                                        int              *logoffsetp);
76 STATIC void xlog_state_put_ticket(xlog_t        *log,
77                                   xlog_ticket_t *tic);
78 STATIC int  xlog_state_release_iclog(xlog_t             *log,
79                                      xlog_in_core_t     *iclog);
80 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
81                                      xlog_in_core_t *iclog,
82                                      int                eventual_size);
83 STATIC int  xlog_state_sync(xlog_t                      *log,
84                             xfs_lsn_t                   lsn,
85                             uint                        flags,
86                             int                         *log_flushed);
87 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
88 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
89
90 /* local functions to manipulate grant head */
91 STATIC int  xlog_grant_log_space(xlog_t         *log,
92                                  xlog_ticket_t  *xtic);
93 STATIC void xlog_grant_push_ail(xfs_mount_t     *mp,
94                                 int             need_bytes);
95 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
96                                            xlog_ticket_t *ticket);
97 STATIC int xlog_regrant_write_log_space(xlog_t          *log,
98                                          xlog_ticket_t  *ticket);
99 STATIC void xlog_ungrant_log_space(xlog_t        *log,
100                                    xlog_ticket_t *ticket);
101
102
103 /* local ticket functions */
104 STATIC void             xlog_state_ticket_alloc(xlog_t *log);
105 STATIC xlog_ticket_t    *xlog_ticket_get(xlog_t *log,
106                                          int    unit_bytes,
107                                          int    count,
108                                          char   clientid,
109                                          uint   flags);
110 STATIC void             xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
111
112 #if defined(DEBUG)
113 STATIC void     xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
114 STATIC void     xlog_verify_grant_head(xlog_t *log, int equals);
115 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
116                                   int count, boolean_t syncing);
117 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
118                                      xfs_lsn_t tail_lsn);
119 #else
120 #define xlog_verify_dest_ptr(a,b)
121 #define xlog_verify_grant_head(a,b)
122 #define xlog_verify_iclog(a,b,c,d)
123 #define xlog_verify_tail_lsn(a,b,c)
124 #endif
125
126 STATIC int      xlog_iclogs_empty(xlog_t *log);
127
128 #if defined(XFS_LOG_TRACE)
129 void
130 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
131 {
132         unsigned long cnts;
133
134         if (!log->l_grant_trace) {
135                 log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
136                 if (!log->l_grant_trace)
137                         return;
138         }
139         /* ticket counts are 1 byte each */
140         cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
141
142         ktrace_enter(log->l_grant_trace,
143                      (void *)tic,
144                      (void *)log->l_reserve_headq,
145                      (void *)log->l_write_headq,
146                      (void *)((unsigned long)log->l_grant_reserve_cycle),
147                      (void *)((unsigned long)log->l_grant_reserve_bytes),
148                      (void *)((unsigned long)log->l_grant_write_cycle),
149                      (void *)((unsigned long)log->l_grant_write_bytes),
150                      (void *)((unsigned long)log->l_curr_cycle),
151                      (void *)((unsigned long)log->l_curr_block),
152                      (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
153                      (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
154                      (void *)string,
155                      (void *)((unsigned long)tic->t_trans_type),
156                      (void *)cnts,
157                      (void *)((unsigned long)tic->t_curr_res),
158                      (void *)((unsigned long)tic->t_unit_res));
159 }
160
161 void
162 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
163 {
164         if (!iclog->ic_trace)
165                 iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
166         ktrace_enter(iclog->ic_trace,
167                      (void *)((unsigned long)state),
168                      (void *)((unsigned long)current_pid()),
169                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
170                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
171                      (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172                      (void *)NULL, (void *)NULL);
173 }
174 #else
175 #define xlog_trace_loggrant(log,tic,string)
176 #define xlog_trace_iclog(iclog,state)
177 #endif /* XFS_LOG_TRACE */
178
179
180 static void
181 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
182 {
183         if (*qp) {
184                 tic->t_next         = (*qp);
185                 tic->t_prev         = (*qp)->t_prev;
186                 (*qp)->t_prev->t_next = tic;
187                 (*qp)->t_prev       = tic;
188         } else {
189                 tic->t_prev = tic->t_next = tic;
190                 *qp = tic;
191         }
192
193         tic->t_flags |= XLOG_TIC_IN_Q;
194 }
195
196 static void
197 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
198 {
199         if (tic == tic->t_next) {
200                 *qp = NULL;
201         } else {
202                 *qp = tic->t_next;
203                 tic->t_next->t_prev = tic->t_prev;
204                 tic->t_prev->t_next = tic->t_next;
205         }
206
207         tic->t_next = tic->t_prev = NULL;
208         tic->t_flags &= ~XLOG_TIC_IN_Q;
209 }
210
211 static void
212 xlog_grant_sub_space(struct log *log, int bytes)
213 {
214         log->l_grant_write_bytes -= bytes;
215         if (log->l_grant_write_bytes < 0) {
216                 log->l_grant_write_bytes += log->l_logsize;
217                 log->l_grant_write_cycle--;
218         }
219
220         log->l_grant_reserve_bytes -= bytes;
221         if ((log)->l_grant_reserve_bytes < 0) {
222                 log->l_grant_reserve_bytes += log->l_logsize;
223                 log->l_grant_reserve_cycle--;
224         }
225
226 }
227
228 static void
229 xlog_grant_add_space_write(struct log *log, int bytes)
230 {
231         log->l_grant_write_bytes += bytes;
232         if (log->l_grant_write_bytes > log->l_logsize) {
233                 log->l_grant_write_bytes -= log->l_logsize;
234                 log->l_grant_write_cycle++;
235         }
236 }
237
238 static void
239 xlog_grant_add_space_reserve(struct log *log, int bytes)
240 {
241         log->l_grant_reserve_bytes += bytes;
242         if (log->l_grant_reserve_bytes > log->l_logsize) {
243                 log->l_grant_reserve_bytes -= log->l_logsize;
244                 log->l_grant_reserve_cycle++;
245         }
246 }
247
248 static inline void
249 xlog_grant_add_space(struct log *log, int bytes)
250 {
251         xlog_grant_add_space_write(log, bytes);
252         xlog_grant_add_space_reserve(log, bytes);
253 }
254
255
256 /*
257  * NOTES:
258  *
259  *      1. currblock field gets updated at startup and after in-core logs
260  *              marked as with WANT_SYNC.
261  */
262
263 /*
264  * This routine is called when a user of a log manager ticket is done with
265  * the reservation.  If the ticket was ever used, then a commit record for
266  * the associated transaction is written out as a log operation header with
267  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
268  * a given ticket.  If the ticket was one with a permanent reservation, then
269  * a few operations are done differently.  Permanent reservation tickets by
270  * default don't release the reservation.  They just commit the current
271  * transaction with the belief that the reservation is still needed.  A flag
272  * must be passed in before permanent reservations are actually released.
273  * When these type of tickets are not released, they need to be set into
274  * the inited state again.  By doing this, a start record will be written
275  * out when the next write occurs.
276  */
277 xfs_lsn_t
278 xfs_log_done(xfs_mount_t        *mp,
279              xfs_log_ticket_t   xtic,
280              void               **iclog,
281              uint               flags)
282 {
283         xlog_t          *log    = mp->m_log;
284         xlog_ticket_t   *ticket = (xfs_log_ticket_t) xtic;
285         xfs_lsn_t       lsn     = 0;
286
287         if (XLOG_FORCED_SHUTDOWN(log) ||
288             /*
289              * If nothing was ever written, don't write out commit record.
290              * If we get an error, just continue and give back the log ticket.
291              */
292             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
293              (xlog_commit_record(mp, ticket,
294                                  (xlog_in_core_t **)iclog, &lsn)))) {
295                 lsn = (xfs_lsn_t) -1;
296                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
297                         flags |= XFS_LOG_REL_PERM_RESERV;
298                 }
299         }
300
301
302         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
303             (flags & XFS_LOG_REL_PERM_RESERV)) {
304                 /*
305                  * Release ticket if not permanent reservation or a specific
306                  * request has been made to release a permanent reservation.
307                  */
308                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
309                 xlog_ungrant_log_space(log, ticket);
310                 xlog_state_put_ticket(log, ticket);
311         } else {
312                 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
313                 xlog_regrant_reserve_log_space(log, ticket);
314         }
315
316         /* If this ticket was a permanent reservation and we aren't
317          * trying to release it, reset the inited flags; so next time
318          * we write, a start record will be written out.
319          */
320         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
321             (flags & XFS_LOG_REL_PERM_RESERV) == 0)
322                 ticket->t_flags |= XLOG_TIC_INITED;
323
324         return lsn;
325 }       /* xfs_log_done */
326
327
328 /*
329  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
330  *      the force is done synchronously.
331  *
332  * Asynchronous forces are implemented by setting the WANT_SYNC
333  * bit in the appropriate in-core log and then returning.
334  *
335  * Synchronous forces are implemented with a semaphore.  All callers
336  * to force a given lsn to disk will wait on a semaphore attached to the
337  * specific in-core log.  When given in-core log finally completes its
338  * write to disk, that thread will wake up all threads waiting on the
339  * semaphore.
340  */
341 int
342 _xfs_log_force(
343         xfs_mount_t     *mp,
344         xfs_lsn_t       lsn,
345         uint            flags,
346         int             *log_flushed)
347 {
348         xlog_t          *log = mp->m_log;
349         int             dummy;
350
351         if (!log_flushed)
352                 log_flushed = &dummy;
353
354         ASSERT(flags & XFS_LOG_FORCE);
355
356         XFS_STATS_INC(xs_log_force);
357
358         if (log->l_flags & XLOG_IO_ERROR)
359                 return XFS_ERROR(EIO);
360         if (lsn == 0)
361                 return xlog_state_sync_all(log, flags, log_flushed);
362         else
363                 return xlog_state_sync(log, lsn, flags, log_flushed);
364 }       /* xfs_log_force */
365
366 /*
367  * Attaches a new iclog I/O completion callback routine during
368  * transaction commit.  If the log is in error state, a non-zero
369  * return code is handed back and the caller is responsible for
370  * executing the callback at an appropriate time.
371  */
372 int
373 xfs_log_notify(xfs_mount_t        *mp,          /* mount of partition */
374                void               *iclog_hndl,  /* iclog to hang callback off */
375                xfs_log_callback_t *cb)
376 {
377         xlog_t *log = mp->m_log;
378         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
379         int     abortflg, spl;
380
381         cb->cb_next = NULL;
382         spl = LOG_LOCK(log);
383         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
384         if (!abortflg) {
385                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
386                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
387                 cb->cb_next = NULL;
388                 *(iclog->ic_callback_tail) = cb;
389                 iclog->ic_callback_tail = &(cb->cb_next);
390         }
391         LOG_UNLOCK(log, spl);
392         return abortflg;
393 }       /* xfs_log_notify */
394
395 int
396 xfs_log_release_iclog(xfs_mount_t *mp,
397                       void        *iclog_hndl)
398 {
399         xlog_t *log = mp->m_log;
400         xlog_in_core_t    *iclog = (xlog_in_core_t *)iclog_hndl;
401
402         if (xlog_state_release_iclog(log, iclog)) {
403                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
404                 return EIO;
405         }
406
407         return 0;
408 }
409
410 /*
411  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
412  *      to the reservation.
413  *  2. Potentially, push buffers at tail of log to disk.
414  *
415  * Each reservation is going to reserve extra space for a log record header.
416  * When writes happen to the on-disk log, we don't subtract the length of the
417  * log record header from any reservation.  By wasting space in each
418  * reservation, we prevent over allocation problems.
419  */
420 int
421 xfs_log_reserve(xfs_mount_t      *mp,
422                 int              unit_bytes,
423                 int              cnt,
424                 xfs_log_ticket_t *ticket,
425                 __uint8_t        client,
426                 uint             flags,
427                 uint             t_type)
428 {
429         xlog_t          *log = mp->m_log;
430         xlog_ticket_t   *internal_ticket;
431         int             retval = 0;
432
433         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
434         ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
435
436         if (XLOG_FORCED_SHUTDOWN(log))
437                 return XFS_ERROR(EIO);
438
439         XFS_STATS_INC(xs_try_logspace);
440
441         if (*ticket != NULL) {
442                 ASSERT(flags & XFS_LOG_PERM_RESERV);
443                 internal_ticket = (xlog_ticket_t *)*ticket;
444                 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
445                 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
446                 retval = xlog_regrant_write_log_space(log, internal_ticket);
447         } else {
448                 /* may sleep if need to allocate more tickets */
449                 internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
450                                                   client, flags);
451                 internal_ticket->t_trans_type = t_type;
452                 *ticket = internal_ticket;
453                 xlog_trace_loggrant(log, internal_ticket, 
454                         (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
455                         "xfs_log_reserve: create new ticket (permanent trans)" :
456                         "xfs_log_reserve: create new ticket");
457                 xlog_grant_push_ail(mp,
458                                     (internal_ticket->t_unit_res *
459                                      internal_ticket->t_cnt));
460                 retval = xlog_grant_log_space(log, internal_ticket);
461         }
462
463         return retval;
464 }       /* xfs_log_reserve */
465
466
467 /*
468  * Mount a log filesystem
469  *
470  * mp           - ubiquitous xfs mount point structure
471  * log_target   - buftarg of on-disk log device
472  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
473  * num_bblocks  - Number of BBSIZE blocks in on-disk log
474  *
475  * Return error or zero.
476  */
477 int
478 xfs_log_mount(xfs_mount_t       *mp,
479               xfs_buftarg_t     *log_target,
480               xfs_daddr_t       blk_offset,
481               int               num_bblks)
482 {
483         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
484                 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
485         else {
486                 cmn_err(CE_NOTE,
487                         "!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
488                         mp->m_fsname);
489                 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
490         }
491
492         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
493
494         /*
495          * skip log recovery on a norecovery mount.  pretend it all
496          * just worked.
497          */
498         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
499                 bhv_vfs_t       *vfsp = XFS_MTOVFS(mp);
500                 int             error, readonly = (vfsp->vfs_flag & VFS_RDONLY);
501
502                 if (readonly)
503                         vfsp->vfs_flag &= ~VFS_RDONLY;
504
505                 error = xlog_recover(mp->m_log);
506
507                 if (readonly)
508                         vfsp->vfs_flag |= VFS_RDONLY;
509                 if (error) {
510                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
511                         xlog_dealloc_log(mp->m_log);
512                         return error;
513                 }
514         }
515
516         /* Normal transactions can now occur */
517         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
518
519         /* End mounting message in xfs_log_mount_finish */
520         return 0;
521 }       /* xfs_log_mount */
522
523 /*
524  * Finish the recovery of the file system.  This is separate from
525  * the xfs_log_mount() call, because it depends on the code in
526  * xfs_mountfs() to read in the root and real-time bitmap inodes
527  * between calling xfs_log_mount() and here.
528  *
529  * mp           - ubiquitous xfs mount point structure
530  */
531 int
532 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
533 {
534         int     error;
535
536         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
537                 error = xlog_recover_finish(mp->m_log, mfsi_flags);
538         else {
539                 error = 0;
540                 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
541         }
542
543         return error;
544 }
545
546 /*
547  * Unmount processing for the log.
548  */
549 int
550 xfs_log_unmount(xfs_mount_t *mp)
551 {
552         int             error;
553
554         error = xfs_log_unmount_write(mp);
555         xfs_log_unmount_dealloc(mp);
556         return error;
557 }
558
559 /*
560  * Final log writes as part of unmount.
561  *
562  * Mark the filesystem clean as unmount happens.  Note that during relocation
563  * this routine needs to be executed as part of source-bag while the
564  * deallocation must not be done until source-end.
565  */
566
567 /*
568  * Unmount record used to have a string "Unmount filesystem--" in the
569  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
570  * We just write the magic number now since that particular field isn't
571  * currently architecture converted and "nUmount" is a bit foo.
572  * As far as I know, there weren't any dependencies on the old behaviour.
573  */
574
575 int
576 xfs_log_unmount_write(xfs_mount_t *mp)
577 {
578         xlog_t           *log = mp->m_log;
579         xlog_in_core_t   *iclog;
580 #ifdef DEBUG
581         xlog_in_core_t   *first_iclog;
582 #endif
583         xfs_log_iovec_t  reg[1];
584         xfs_log_ticket_t tic = NULL;
585         xfs_lsn_t        lsn;
586         int              error;
587         SPLDECL(s);
588
589         /* the data section must be 32 bit size aligned */
590         struct {
591             __uint16_t magic;
592             __uint16_t pad1;
593             __uint32_t pad2; /* may as well make it 64 bits */
594         } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
595
596         /*
597          * Don't write out unmount record on read-only mounts.
598          * Or, if we are doing a forced umount (typically because of IO errors).
599          */
600         if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
601                 return 0;
602
603         xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
604
605 #ifdef DEBUG
606         first_iclog = iclog = log->l_iclog;
607         do {
608                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
609                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
610                         ASSERT(iclog->ic_offset == 0);
611                 }
612                 iclog = iclog->ic_next;
613         } while (iclog != first_iclog);
614 #endif
615         if (! (XLOG_FORCED_SHUTDOWN(log))) {
616                 reg[0].i_addr = (void*)&magic;
617                 reg[0].i_len  = sizeof(magic);
618                 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
619
620                 error = xfs_log_reserve(mp, 600, 1, &tic,
621                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
622                 if (!error) {
623                         /* remove inited flag */
624                         ((xlog_ticket_t *)tic)->t_flags = 0;
625                         error = xlog_write(mp, reg, 1, tic, &lsn,
626                                            NULL, XLOG_UNMOUNT_TRANS);
627                         /*
628                          * At this point, we're umounting anyway,
629                          * so there's no point in transitioning log state
630                          * to IOERROR. Just continue...
631                          */
632                 }
633
634                 if (error) {
635                         xfs_fs_cmn_err(CE_ALERT, mp,
636                                 "xfs_log_unmount: unmount record failed");
637                 }
638
639
640                 s = LOG_LOCK(log);
641                 iclog = log->l_iclog;
642                 iclog->ic_refcnt++;
643                 LOG_UNLOCK(log, s);
644                 xlog_state_want_sync(log, iclog);
645                 (void) xlog_state_release_iclog(log, iclog);
646
647                 s = LOG_LOCK(log);
648                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
649                       iclog->ic_state == XLOG_STATE_DIRTY)) {
650                         if (!XLOG_FORCED_SHUTDOWN(log)) {
651                                 sv_wait(&iclog->ic_forcesema, PMEM,
652                                         &log->l_icloglock, s);
653                         } else {
654                                 LOG_UNLOCK(log, s);
655                         }
656                 } else {
657                         LOG_UNLOCK(log, s);
658                 }
659                 if (tic) {
660                         xlog_trace_loggrant(log, tic, "unmount rec");
661                         xlog_ungrant_log_space(log, tic);
662                         xlog_state_put_ticket(log, tic);
663                 }
664         } else {
665                 /*
666                  * We're already in forced_shutdown mode, couldn't
667                  * even attempt to write out the unmount transaction.
668                  *
669                  * Go through the motions of sync'ing and releasing
670                  * the iclog, even though no I/O will actually happen,
671                  * we need to wait for other log I/Os that may already
672                  * be in progress.  Do this as a separate section of
673                  * code so we'll know if we ever get stuck here that
674                  * we're in this odd situation of trying to unmount
675                  * a file system that went into forced_shutdown as
676                  * the result of an unmount..
677                  */
678                 s = LOG_LOCK(log);
679                 iclog = log->l_iclog;
680                 iclog->ic_refcnt++;
681                 LOG_UNLOCK(log, s);
682
683                 xlog_state_want_sync(log, iclog);
684                 (void) xlog_state_release_iclog(log, iclog);
685
686                 s = LOG_LOCK(log);
687
688                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
689                         || iclog->ic_state == XLOG_STATE_DIRTY
690                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
691
692                                 sv_wait(&iclog->ic_forcesema, PMEM,
693                                         &log->l_icloglock, s);
694                 } else {
695                         LOG_UNLOCK(log, s);
696                 }
697         }
698
699         return 0;
700 }       /* xfs_log_unmount_write */
701
702 /*
703  * Deallocate log structures for unmount/relocation.
704  */
705 void
706 xfs_log_unmount_dealloc(xfs_mount_t *mp)
707 {
708         xlog_dealloc_log(mp->m_log);
709 }
710
711 /*
712  * Write region vectors to log.  The write happens using the space reservation
713  * of the ticket (tic).  It is not a requirement that all writes for a given
714  * transaction occur with one call to xfs_log_write().
715  */
716 int
717 xfs_log_write(xfs_mount_t *     mp,
718               xfs_log_iovec_t   reg[],
719               int               nentries,
720               xfs_log_ticket_t  tic,
721               xfs_lsn_t         *start_lsn)
722 {
723         int     error;
724         xlog_t *log = mp->m_log;
725
726         if (XLOG_FORCED_SHUTDOWN(log))
727                 return XFS_ERROR(EIO);
728
729         if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
730                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
731         }
732         return error;
733 }       /* xfs_log_write */
734
735
736 void
737 xfs_log_move_tail(xfs_mount_t   *mp,
738                   xfs_lsn_t     tail_lsn)
739 {
740         xlog_ticket_t   *tic;
741         xlog_t          *log = mp->m_log;
742         int             need_bytes, free_bytes, cycle, bytes;
743         SPLDECL(s);
744
745         if (XLOG_FORCED_SHUTDOWN(log))
746                 return;
747         ASSERT(!XFS_FORCED_SHUTDOWN(mp));
748
749         if (tail_lsn == 0) {
750                 /* needed since sync_lsn is 64 bits */
751                 s = LOG_LOCK(log);
752                 tail_lsn = log->l_last_sync_lsn;
753                 LOG_UNLOCK(log, s);
754         }
755
756         s = GRANT_LOCK(log);
757
758         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
759          * tail_lsn.
760          */
761         if (tail_lsn != 1) {
762                 log->l_tail_lsn = tail_lsn;
763         }
764
765         if ((tic = log->l_write_headq)) {
766 #ifdef DEBUG
767                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
768                         panic("Recovery problem");
769 #endif
770                 cycle = log->l_grant_write_cycle;
771                 bytes = log->l_grant_write_bytes;
772                 free_bytes = xlog_space_left(log, cycle, bytes);
773                 do {
774                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
775
776                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
777                                 break;
778                         tail_lsn = 0;
779                         free_bytes -= tic->t_unit_res;
780                         sv_signal(&tic->t_sema);
781                         tic = tic->t_next;
782                 } while (tic != log->l_write_headq);
783         }
784         if ((tic = log->l_reserve_headq)) {
785 #ifdef DEBUG
786                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
787                         panic("Recovery problem");
788 #endif
789                 cycle = log->l_grant_reserve_cycle;
790                 bytes = log->l_grant_reserve_bytes;
791                 free_bytes = xlog_space_left(log, cycle, bytes);
792                 do {
793                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
794                                 need_bytes = tic->t_unit_res*tic->t_cnt;
795                         else
796                                 need_bytes = tic->t_unit_res;
797                         if (free_bytes < need_bytes && tail_lsn != 1)
798                                 break;
799                         tail_lsn = 0;
800                         free_bytes -= need_bytes;
801                         sv_signal(&tic->t_sema);
802                         tic = tic->t_next;
803                 } while (tic != log->l_reserve_headq);
804         }
805         GRANT_UNLOCK(log, s);
806 }       /* xfs_log_move_tail */
807
808 /*
809  * Determine if we have a transaction that has gone to disk
810  * that needs to be covered. Log activity needs to be idle (no AIL and
811  * nothing in the iclogs). And, we need to be in the right state indicating
812  * something has gone out.
813  */
814 int
815 xfs_log_need_covered(xfs_mount_t *mp)
816 {
817         SPLDECL(s);
818         int             needed = 0, gen;
819         xlog_t          *log = mp->m_log;
820         bhv_vfs_t       *vfsp = XFS_MTOVFS(mp);
821
822         if (vfs_test_for_freeze(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
823             (vfsp->vfs_flag & VFS_RDONLY))
824                 return 0;
825
826         s = LOG_LOCK(log);
827         if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
828                 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
829                         && !xfs_trans_first_ail(mp, &gen)
830                         && xlog_iclogs_empty(log)) {
831                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
832                         log->l_covered_state = XLOG_STATE_COVER_DONE;
833                 else {
834                         ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
835                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
836                 }
837                 needed = 1;
838         }
839         LOG_UNLOCK(log, s);
840         return needed;
841 }
842
843 /******************************************************************************
844  *
845  *      local routines
846  *
847  ******************************************************************************
848  */
849
850 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
851  * The log manager must keep track of the last LR which was committed
852  * to disk.  The lsn of this LR will become the new tail_lsn whenever
853  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
854  * the situation where stuff could be written into the log but nothing
855  * was ever in the AIL when asked.  Eventually, we panic since the
856  * tail hits the head.
857  *
858  * We may be holding the log iclog lock upon entering this routine.
859  */
860 xfs_lsn_t
861 xlog_assign_tail_lsn(xfs_mount_t *mp)
862 {
863         xfs_lsn_t tail_lsn;
864         SPLDECL(s);
865         xlog_t    *log = mp->m_log;
866
867         tail_lsn = xfs_trans_tail_ail(mp);
868         s = GRANT_LOCK(log);
869         if (tail_lsn != 0) {
870                 log->l_tail_lsn = tail_lsn;
871         } else {
872                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
873         }
874         GRANT_UNLOCK(log, s);
875
876         return tail_lsn;
877 }       /* xlog_assign_tail_lsn */
878
879
880 /*
881  * Return the space in the log between the tail and the head.  The head
882  * is passed in the cycle/bytes formal parms.  In the special case where
883  * the reserve head has wrapped passed the tail, this calculation is no
884  * longer valid.  In this case, just return 0 which means there is no space
885  * in the log.  This works for all places where this function is called
886  * with the reserve head.  Of course, if the write head were to ever
887  * wrap the tail, we should blow up.  Rather than catch this case here,
888  * we depend on other ASSERTions in other parts of the code.   XXXmiken
889  *
890  * This code also handles the case where the reservation head is behind
891  * the tail.  The details of this case are described below, but the end
892  * result is that we return the size of the log as the amount of space left.
893  */
894 int
895 xlog_space_left(xlog_t *log, int cycle, int bytes)
896 {
897         int free_bytes;
898         int tail_bytes;
899         int tail_cycle;
900
901         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
902         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
903         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
904                 free_bytes = log->l_logsize - (bytes - tail_bytes);
905         } else if ((tail_cycle + 1) < cycle) {
906                 return 0;
907         } else if (tail_cycle < cycle) {
908                 ASSERT(tail_cycle == (cycle - 1));
909                 free_bytes = tail_bytes - bytes;
910         } else {
911                 /*
912                  * The reservation head is behind the tail.
913                  * In this case we just want to return the size of the
914                  * log as the amount of space left.
915                  */
916                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
917                         "xlog_space_left: head behind tail\n"
918                         "  tail_cycle = %d, tail_bytes = %d\n"
919                         "  GH   cycle = %d, GH   bytes = %d",
920                         tail_cycle, tail_bytes, cycle, bytes);
921                 ASSERT(0);
922                 free_bytes = log->l_logsize;
923         }
924         return free_bytes;
925 }       /* xlog_space_left */
926
927
928 /*
929  * Log function which is called when an io completes.
930  *
931  * The log manager needs its own routine, in order to control what
932  * happens with the buffer after the write completes.
933  */
934 void
935 xlog_iodone(xfs_buf_t *bp)
936 {
937         xlog_in_core_t  *iclog;
938         xlog_t          *l;
939         int             aborted;
940
941         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
942         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
943         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
944         aborted = 0;
945
946         /*
947          * Some versions of cpp barf on the recursive definition of
948          * ic_log -> hic_fields.ic_log and expand ic_log twice when
949          * it is passed through two macros.  Workaround broken cpp.
950          */
951         l = iclog->ic_log;
952
953         /*
954          * Race to shutdown the filesystem if we see an error.
955          */
956         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
957                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
958                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
959                 XFS_BUF_STALE(bp);
960                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
961                 /*
962                  * This flag will be propagated to the trans-committed
963                  * callback routines to let them know that the log-commit
964                  * didn't succeed.
965                  */
966                 aborted = XFS_LI_ABORTED;
967         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
968                 aborted = XFS_LI_ABORTED;
969         }
970         xlog_state_done_syncing(iclog, aborted);
971         if (!(XFS_BUF_ISASYNC(bp))) {
972                 /*
973                  * Corresponding psema() will be done in bwrite().  If we don't
974                  * vsema() here, panic.
975                  */
976                 XFS_BUF_V_IODONESEMA(bp);
977         }
978 }       /* xlog_iodone */
979
980 /*
981  * The bdstrat callback function for log bufs. This gives us a central
982  * place to trap bufs in case we get hit by a log I/O error and need to
983  * shutdown. Actually, in practice, even when we didn't get a log error,
984  * we transition the iclogs to IOERROR state *after* flushing all existing
985  * iclogs to disk. This is because we don't want anymore new transactions to be
986  * started or completed afterwards.
987  */
988 STATIC int
989 xlog_bdstrat_cb(struct xfs_buf *bp)
990 {
991         xlog_in_core_t *iclog;
992
993         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
994
995         if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
996           /* note for irix bstrat will need  struct bdevsw passed
997            * Fix the following macro if the code ever is merged
998            */
999             XFS_bdstrat(bp);
1000                 return 0;
1001         }
1002
1003         xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1004         XFS_BUF_ERROR(bp, EIO);
1005         XFS_BUF_STALE(bp);
1006         xfs_biodone(bp);
1007         return XFS_ERROR(EIO);
1008
1009
1010 }
1011
1012 /*
1013  * Return size of each in-core log record buffer.
1014  *
1015  * Low memory machines only get 2 16KB buffers.  We don't want to waste
1016  * memory here.  However, all other machines get at least 2 32KB buffers.
1017  * The number is hard coded because we don't care about the minimum
1018  * memory size, just 32MB systems.
1019  *
1020  * If the filesystem blocksize is too large, we may need to choose a
1021  * larger size since the directory code currently logs entire blocks.
1022  */
1023
1024 STATIC void
1025 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1026                            xlog_t       *log)
1027 {
1028         int size;
1029         int xhdrs;
1030
1031         if (mp->m_logbufs <= 0) {
1032                 if (xfs_physmem <= btoc(128*1024*1024)) {
1033                         log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1034                 } else if (xfs_physmem <= btoc(400*1024*1024)) {
1035                         log->l_iclog_bufs = XLOG_MED_ICLOGS;
1036                 } else {        /* 256K with 32K bufs */
1037                         log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1038                 }
1039         } else {
1040                 log->l_iclog_bufs = mp->m_logbufs;
1041         }
1042
1043         /*
1044          * Buffer size passed in from mount system call.
1045          */
1046         if (mp->m_logbsize > 0) {
1047                 size = log->l_iclog_size = mp->m_logbsize;
1048                 log->l_iclog_size_log = 0;
1049                 while (size != 1) {
1050                         log->l_iclog_size_log++;
1051                         size >>= 1;
1052                 }
1053
1054                 if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1055                         /* # headers = size / 32K
1056                          * one header holds cycles from 32K of data
1057                          */
1058
1059                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1060                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1061                                 xhdrs++;
1062                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1063                         log->l_iclog_heads = xhdrs;
1064                 } else {
1065                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1066                         log->l_iclog_hsize = BBSIZE;
1067                         log->l_iclog_heads = 1;
1068                 }
1069                 goto done;
1070         }
1071
1072         /*
1073          * Special case machines that have less than 32MB of memory.
1074          * All machines with more memory use 32KB buffers.
1075          */
1076         if (xfs_physmem <= btoc(32*1024*1024)) {
1077                 /* Don't change; min configuration */
1078                 log->l_iclog_size = XLOG_RECORD_BSIZE;          /* 16k */
1079                 log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1080         } else {
1081                 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;      /* 32k */
1082                 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1083         }
1084
1085         /* the default log size is 16k or 32k which is one header sector */
1086         log->l_iclog_hsize = BBSIZE;
1087         log->l_iclog_heads = 1;
1088
1089         /*
1090          * For 16KB, we use 3 32KB buffers.  For 32KB block sizes, we use
1091          * 4 32KB buffers.  For 64KB block sizes, we use 8 32KB buffers.
1092          */
1093         if (mp->m_sb.sb_blocksize >= 16*1024) {
1094                 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1095                 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1096                 if (mp->m_logbufs <= 0) {
1097                         switch (mp->m_sb.sb_blocksize) {
1098                             case 16*1024:                       /* 16 KB */
1099                                 log->l_iclog_bufs = 3;
1100                                 break;
1101                             case 32*1024:                       /* 32 KB */
1102                                 log->l_iclog_bufs = 4;
1103                                 break;
1104                             case 64*1024:                       /* 64 KB */
1105                                 log->l_iclog_bufs = 8;
1106                                 break;
1107                             default:
1108                                 xlog_panic("XFS: Invalid blocksize");
1109                                 break;
1110                         }
1111                 }
1112         }
1113
1114 done:   /* are we being asked to make the sizes selected above visible? */
1115         if (mp->m_logbufs == 0)
1116                 mp->m_logbufs = log->l_iclog_bufs;
1117         if (mp->m_logbsize == 0)
1118                 mp->m_logbsize = log->l_iclog_size;
1119 }       /* xlog_get_iclog_buffer_size */
1120
1121
1122 /*
1123  * This routine initializes some of the log structure for a given mount point.
1124  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1125  * some other stuff may be filled in too.
1126  */
1127 STATIC xlog_t *
1128 xlog_alloc_log(xfs_mount_t      *mp,
1129                xfs_buftarg_t    *log_target,
1130                xfs_daddr_t      blk_offset,
1131                int              num_bblks)
1132 {
1133         xlog_t                  *log;
1134         xlog_rec_header_t       *head;
1135         xlog_in_core_t          **iclogp;
1136         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1137         xfs_buf_t               *bp;
1138         int                     i;
1139         int                     iclogsize;
1140
1141         log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1142
1143         log->l_mp          = mp;
1144         log->l_targ        = log_target;
1145         log->l_logsize     = BBTOB(num_bblks);
1146         log->l_logBBstart  = blk_offset;
1147         log->l_logBBsize   = num_bblks;
1148         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1149         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1150
1151         log->l_prev_block  = -1;
1152         ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1153         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1154         log->l_last_sync_lsn = log->l_tail_lsn;
1155         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1156         log->l_grant_reserve_cycle = 1;
1157         log->l_grant_write_cycle = 1;
1158
1159         if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1160                 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1161                 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1162                 /* for larger sector sizes, must have v2 or external log */
1163                 ASSERT(log->l_sectbb_log == 0 ||
1164                         log->l_logBBstart == 0 ||
1165                         XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1166                 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1167         }
1168         log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1169
1170         xlog_get_iclog_buffer_size(mp, log);
1171
1172         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1173         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1174         XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1175         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1176         ASSERT(XFS_BUF_ISBUSY(bp));
1177         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1178         log->l_xbuf = bp;
1179
1180         spinlock_init(&log->l_icloglock, "iclog");
1181         spinlock_init(&log->l_grant_lock, "grhead_iclog");
1182         initnsema(&log->l_flushsema, 0, "ic-flush");
1183         xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
1184
1185         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1186         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1187
1188         iclogp = &log->l_iclog;
1189         /*
1190          * The amount of memory to allocate for the iclog structure is
1191          * rather funky due to the way the structure is defined.  It is
1192          * done this way so that we can use different sizes for machines
1193          * with different amounts of memory.  See the definition of
1194          * xlog_in_core_t in xfs_log_priv.h for details.
1195          */
1196         iclogsize = log->l_iclog_size;
1197         ASSERT(log->l_iclog_size >= 4096);
1198         for (i=0; i < log->l_iclog_bufs; i++) {
1199                 *iclogp = (xlog_in_core_t *)
1200                           kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1201                 iclog = *iclogp;
1202                 iclog->hic_data = (xlog_in_core_2_t *)
1203                           kmem_zalloc(iclogsize, KM_SLEEP | KM_LARGE);
1204
1205                 iclog->ic_prev = prev_iclog;
1206                 prev_iclog = iclog;
1207                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1208
1209                 head = &iclog->ic_header;
1210                 memset(head, 0, sizeof(xlog_rec_header_t));
1211                 INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1212                 INT_SET(head->h_version, ARCH_CONVERT,
1213                         XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1214                 INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1215                 /* new fields */
1216                 INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1217                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1218
1219                 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1220                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1221                 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1222                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1223                 iclog->ic_bp = bp;
1224
1225                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1226                 iclog->ic_state = XLOG_STATE_ACTIVE;
1227                 iclog->ic_log = log;
1228                 iclog->ic_callback_tail = &(iclog->ic_callback);
1229                 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1230
1231                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1232                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1233                 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1234                 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1235
1236                 iclogp = &iclog->ic_next;
1237         }
1238         *iclogp = log->l_iclog;                 /* complete ring */
1239         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1240
1241         return log;
1242 }       /* xlog_alloc_log */
1243
1244
1245 /*
1246  * Write out the commit record of a transaction associated with the given
1247  * ticket.  Return the lsn of the commit record.
1248  */
1249 STATIC int
1250 xlog_commit_record(xfs_mount_t  *mp,
1251                    xlog_ticket_t *ticket,
1252                    xlog_in_core_t **iclog,
1253                    xfs_lsn_t    *commitlsnp)
1254 {
1255         int             error;
1256         xfs_log_iovec_t reg[1];
1257
1258         reg[0].i_addr = NULL;
1259         reg[0].i_len = 0;
1260         XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1261
1262         ASSERT_ALWAYS(iclog);
1263         if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1264                                iclog, XLOG_COMMIT_TRANS))) {
1265                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1266         }
1267         return error;
1268 }       /* xlog_commit_record */
1269
1270
1271 /*
1272  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1273  * log space.  This code pushes on the lsn which would supposedly free up
1274  * the 25% which we want to leave free.  We may need to adopt a policy which
1275  * pushes on an lsn which is further along in the log once we reach the high
1276  * water mark.  In this manner, we would be creating a low water mark.
1277  */
1278 void
1279 xlog_grant_push_ail(xfs_mount_t *mp,
1280                     int         need_bytes)
1281 {
1282     xlog_t      *log = mp->m_log;       /* pointer to the log */
1283     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1284     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1285     int         free_blocks;            /* free blocks left to write to */
1286     int         free_bytes;             /* free bytes left to write to */
1287     int         threshold_block;        /* block in lsn we'd like to be at */
1288     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1289     int         free_threshold;
1290     SPLDECL(s);
1291
1292     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1293
1294     s = GRANT_LOCK(log);
1295     free_bytes = xlog_space_left(log,
1296                                  log->l_grant_reserve_cycle,
1297                                  log->l_grant_reserve_bytes);
1298     tail_lsn = log->l_tail_lsn;
1299     free_blocks = BTOBBT(free_bytes);
1300
1301     /*
1302      * Set the threshold for the minimum number of free blocks in the
1303      * log to the maximum of what the caller needs, one quarter of the
1304      * log, and 256 blocks.
1305      */
1306     free_threshold = BTOBB(need_bytes);
1307     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1308     free_threshold = MAX(free_threshold, 256);
1309     if (free_blocks < free_threshold) {
1310         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1311         threshold_cycle = CYCLE_LSN(tail_lsn);
1312         if (threshold_block >= log->l_logBBsize) {
1313             threshold_block -= log->l_logBBsize;
1314             threshold_cycle += 1;
1315         }
1316         ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1317                        threshold_block);
1318
1319         /* Don't pass in an lsn greater than the lsn of the last
1320          * log record known to be on disk.
1321          */
1322         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1323             threshold_lsn = log->l_last_sync_lsn;
1324     }
1325     GRANT_UNLOCK(log, s);
1326
1327     /*
1328      * Get the transaction layer to kick the dirty buffers out to
1329      * disk asynchronously. No point in trying to do this if
1330      * the filesystem is shutting down.
1331      */
1332     if (threshold_lsn &&
1333         !XLOG_FORCED_SHUTDOWN(log))
1334             xfs_trans_push_ail(mp, threshold_lsn);
1335 }       /* xlog_grant_push_ail */
1336
1337
1338 /*
1339  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1340  * fashion.  Previously, we should have moved the current iclog
1341  * ptr in the log to point to the next available iclog.  This allows further
1342  * write to continue while this code syncs out an iclog ready to go.
1343  * Before an in-core log can be written out, the data section must be scanned
1344  * to save away the 1st word of each BBSIZE block into the header.  We replace
1345  * it with the current cycle count.  Each BBSIZE block is tagged with the
1346  * cycle count because there in an implicit assumption that drives will
1347  * guarantee that entire 512 byte blocks get written at once.  In other words,
1348  * we can't have part of a 512 byte block written and part not written.  By
1349  * tagging each block, we will know which blocks are valid when recovering
1350  * after an unclean shutdown.
1351  *
1352  * This routine is single threaded on the iclog.  No other thread can be in
1353  * this routine with the same iclog.  Changing contents of iclog can there-
1354  * fore be done without grabbing the state machine lock.  Updating the global
1355  * log will require grabbing the lock though.
1356  *
1357  * The entire log manager uses a logical block numbering scheme.  Only
1358  * log_sync (and then only bwrite()) know about the fact that the log may
1359  * not start with block zero on a given device.  The log block start offset
1360  * is added immediately before calling bwrite().
1361  */
1362
1363 int
1364 xlog_sync(xlog_t                *log,
1365           xlog_in_core_t        *iclog)
1366 {
1367         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1368         xfs_buf_t       *bp;
1369         int             i, ops;
1370         uint            count;          /* byte count of bwrite */
1371         uint            count_init;     /* initial count before roundup */
1372         int             roundoff;       /* roundoff to BB or stripe */
1373         int             split = 0;      /* split write into two regions */
1374         int             error;
1375         SPLDECL(s);
1376         int             v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1377
1378         XFS_STATS_INC(xs_log_writes);
1379         ASSERT(iclog->ic_refcnt == 0);
1380
1381         /* Add for LR header */
1382         count_init = log->l_iclog_hsize + iclog->ic_offset;
1383
1384         /* Round out the log write size */
1385         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1386                 /* we have a v2 stripe unit to use */
1387                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1388         } else {
1389                 count = BBTOB(BTOBB(count_init));
1390         }
1391         roundoff = count - count_init;
1392         ASSERT(roundoff >= 0);
1393         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1394                 roundoff < log->l_mp->m_sb.sb_logsunit)
1395                 || 
1396                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1397                  roundoff < BBTOB(1)));
1398
1399         /* move grant heads by roundoff in sync */
1400         s = GRANT_LOCK(log);
1401         xlog_grant_add_space(log, roundoff);
1402         GRANT_UNLOCK(log, s);
1403
1404         /* put cycle number in every block */
1405         xlog_pack_data(log, iclog, roundoff); 
1406
1407         /* real byte length */
1408         if (v2) {
1409                 INT_SET(iclog->ic_header.h_len, 
1410                         ARCH_CONVERT,
1411                         iclog->ic_offset + roundoff);
1412         } else {
1413                 INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1414         }
1415
1416         /* put ops count in correct order */
1417         ops = iclog->ic_header.h_num_logops;
1418         INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1419
1420         bp = iclog->ic_bp;
1421         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1422         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1423         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1424
1425         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1426
1427         /* Do we need to split this write into 2 parts? */
1428         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1429                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1430                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1431                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1432         } else {
1433                 iclog->ic_bwritecnt = 1;
1434         }
1435         XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1436         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1437         XFS_BUF_ZEROFLAGS(bp);
1438         XFS_BUF_BUSY(bp);
1439         XFS_BUF_ASYNC(bp);
1440         /*
1441          * Do an ordered write for the log block.
1442          * Its unnecessary to flush the first split block in the log wrap case.
1443          */
1444         if (!split && (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_ZEROFLAGS(bp);
1475                 XFS_BUF_BUSY(bp);
1476                 XFS_BUF_ASYNC(bp);
1477                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1478                         XFS_BUF_ORDERED(bp);
1479                 dptr = XFS_BUF_PTR(bp);
1480                 /*
1481                  * Bump the cycle numbers at the start of each block
1482                  * since this part of the buffer is at the start of
1483                  * a new cycle.  Watch out for the header magic number
1484                  * case, though.
1485                  */
1486                 for (i=0; i<split; i += BBSIZE) {
1487                         INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1488                         if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1489                                 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1490                         dptr += BBSIZE;
1491                 }
1492
1493                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1494                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1495
1496                 /* account for internal log which doesn't start at block #0 */
1497                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1498                 XFS_BUF_WRITE(bp);
1499                 if ((error = XFS_bwrite(bp))) {
1500                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1501                                           bp, XFS_BUF_ADDR(bp));
1502                         return error;
1503                 }
1504         }
1505         return 0;
1506 }       /* xlog_sync */
1507
1508
1509 /*
1510  * Deallocate a log structure
1511  */
1512 void
1513 xlog_dealloc_log(xlog_t *log)
1514 {
1515         xlog_in_core_t  *iclog, *next_iclog;
1516         xlog_ticket_t   *tic, *next_tic;
1517         int             i;
1518
1519
1520         iclog = log->l_iclog;
1521         for (i=0; i<log->l_iclog_bufs; i++) {
1522                 sv_destroy(&iclog->ic_forcesema);
1523                 sv_destroy(&iclog->ic_writesema);
1524                 xfs_buf_free(iclog->ic_bp);
1525 #ifdef XFS_LOG_TRACE
1526                 if (iclog->ic_trace != NULL) {
1527                         ktrace_free(iclog->ic_trace);
1528                 }
1529 #endif
1530                 next_iclog = iclog->ic_next;
1531                 kmem_free(iclog->hic_data, log->l_iclog_size);
1532                 kmem_free(iclog, sizeof(xlog_in_core_t));
1533                 iclog = next_iclog;
1534         }
1535         freesema(&log->l_flushsema);
1536         spinlock_destroy(&log->l_icloglock);
1537         spinlock_destroy(&log->l_grant_lock);
1538
1539         /* XXXsup take a look at this again. */
1540         if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
1541             !XLOG_FORCED_SHUTDOWN(log)) {
1542                 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1543                         "xlog_dealloc_log: (cnt: %d, total: %d)",
1544                         log->l_ticket_cnt, log->l_ticket_tcnt);
1545                 /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1546
1547         } else {
1548                 tic = log->l_unmount_free;
1549                 while (tic) {
1550                         next_tic = tic->t_next;
1551                         kmem_free(tic, NBPP);
1552                         tic = next_tic;
1553                 }
1554         }
1555         xfs_buf_free(log->l_xbuf);
1556 #ifdef XFS_LOG_TRACE
1557         if (log->l_trace != NULL) {
1558                 ktrace_free(log->l_trace);
1559         }
1560         if (log->l_grant_trace != NULL) {
1561                 ktrace_free(log->l_grant_trace);
1562         }
1563 #endif
1564         log->l_mp->m_log = NULL;
1565         kmem_free(log, sizeof(xlog_t));
1566 }       /* xlog_dealloc_log */
1567
1568 /*
1569  * Update counters atomically now that memcpy is done.
1570  */
1571 /* ARGSUSED */
1572 static inline void
1573 xlog_state_finish_copy(xlog_t           *log,
1574                        xlog_in_core_t   *iclog,
1575                        int              record_cnt,
1576                        int              copy_bytes)
1577 {
1578         SPLDECL(s);
1579
1580         s = LOG_LOCK(log);
1581
1582         iclog->ic_header.h_num_logops += record_cnt;
1583         iclog->ic_offset += copy_bytes;
1584
1585         LOG_UNLOCK(log, s);
1586 }       /* xlog_state_finish_copy */
1587
1588
1589
1590
1591 /*
1592  * print out info relating to regions written which consume
1593  * the reservation
1594  */
1595 STATIC void
1596 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1597 {
1598         uint i;
1599         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1600
1601         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1602         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1603             "bformat",
1604             "bchunk",
1605             "efi_format",
1606             "efd_format",
1607             "iformat",
1608             "icore",
1609             "iext",
1610             "ibroot",
1611             "ilocal",
1612             "iattr_ext",
1613             "iattr_broot",
1614             "iattr_local",
1615             "qformat",
1616             "dquot",
1617             "quotaoff",
1618             "LR header",
1619             "unmount",
1620             "commit",
1621             "trans header"
1622         };
1623         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1624             "SETATTR_NOT_SIZE",
1625             "SETATTR_SIZE",
1626             "INACTIVE",
1627             "CREATE",
1628             "CREATE_TRUNC",
1629             "TRUNCATE_FILE",
1630             "REMOVE",
1631             "LINK",
1632             "RENAME",
1633             "MKDIR",
1634             "RMDIR",
1635             "SYMLINK",
1636             "SET_DMATTRS",
1637             "GROWFS",
1638             "STRAT_WRITE",
1639             "DIOSTRAT",
1640             "WRITE_SYNC",
1641             "WRITEID",
1642             "ADDAFORK",
1643             "ATTRINVAL",
1644             "ATRUNCATE",
1645             "ATTR_SET",
1646             "ATTR_RM",
1647             "ATTR_FLAG",
1648             "CLEAR_AGI_BUCKET",
1649             "QM_SBCHANGE",
1650             "DUMMY1",
1651             "DUMMY2",
1652             "QM_QUOTAOFF",
1653             "QM_DQALLOC",
1654             "QM_SETQLIM",
1655             "QM_DQCLUSTER",
1656             "QM_QINOCREATE",
1657             "QM_QUOTAOFF_END",
1658             "SB_UNIT",
1659             "FSYNC_TS",
1660             "GROWFSRT_ALLOC",
1661             "GROWFSRT_ZERO",
1662             "GROWFSRT_FREE",
1663             "SWAPEXT"
1664         };
1665
1666         xfs_fs_cmn_err(CE_WARN, mp,
1667                         "xfs_log_write: reservation summary:\n"
1668                         "  trans type  = %s (%u)\n"
1669                         "  unit res    = %d bytes\n"
1670                         "  current res = %d bytes\n"
1671                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1672                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1673                         "  ophdr + reg = %u bytes\n"
1674                         "  num regions = %u\n",
1675                         ((ticket->t_trans_type <= 0 ||
1676                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1677                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1678                         ticket->t_trans_type,
1679                         ticket->t_unit_res,
1680                         ticket->t_curr_res,
1681                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1682                         ticket->t_res_num_ophdrs, ophdr_spc,
1683                         ticket->t_res_arr_sum + 
1684                         ticket->t_res_o_flow + ophdr_spc,
1685                         ticket->t_res_num);
1686
1687         for (i = 0; i < ticket->t_res_num; i++) {
1688                 uint r_type = ticket->t_res_arr[i].r_type; 
1689                 cmn_err(CE_WARN,
1690                             "region[%u]: %s - %u bytes\n",
1691                             i, 
1692                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1693                             "bad-rtype" : res_type_str[r_type-1]),
1694                             ticket->t_res_arr[i].r_len);
1695         }
1696 }
1697
1698 /*
1699  * Write some region out to in-core log
1700  *
1701  * This will be called when writing externally provided regions or when
1702  * writing out a commit record for a given transaction.
1703  *
1704  * General algorithm:
1705  *      1. Find total length of this write.  This may include adding to the
1706  *              lengths passed in.
1707  *      2. Check whether we violate the tickets reservation.
1708  *      3. While writing to this iclog
1709  *          A. Reserve as much space in this iclog as can get
1710  *          B. If this is first write, save away start lsn
1711  *          C. While writing this region:
1712  *              1. If first write of transaction, write start record
1713  *              2. Write log operation header (header per region)
1714  *              3. Find out if we can fit entire region into this iclog
1715  *              4. Potentially, verify destination memcpy ptr
1716  *              5. Memcpy (partial) region
1717  *              6. If partial copy, release iclog; otherwise, continue
1718  *                      copying more regions into current iclog
1719  *      4. Mark want sync bit (in simulation mode)
1720  *      5. Release iclog for potential flush to on-disk log.
1721  *
1722  * ERRORS:
1723  * 1.   Panic if reservation is overrun.  This should never happen since
1724  *      reservation amounts are generated internal to the filesystem.
1725  * NOTES:
1726  * 1. Tickets are single threaded data structures.
1727  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1728  *      syncing routine.  When a single log_write region needs to span
1729  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1730  *      on all log operation writes which don't contain the end of the
1731  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1732  *      operation which contains the end of the continued log_write region.
1733  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1734  *      we don't really know exactly how much space will be used.  As a result,
1735  *      we don't update ic_offset until the end when we know exactly how many
1736  *      bytes have been written out.
1737  */
1738 int
1739 xlog_write(xfs_mount_t *        mp,
1740            xfs_log_iovec_t      reg[],
1741            int                  nentries,
1742            xfs_log_ticket_t     tic,
1743            xfs_lsn_t            *start_lsn,
1744            xlog_in_core_t       **commit_iclog,
1745            uint                 flags)
1746 {
1747     xlog_t           *log = mp->m_log;
1748     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1749     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1750     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1751     __psint_t        ptr;            /* copy address into data region */
1752     int              len;            /* # xlog_write() bytes 2 still copy */
1753     int              index;          /* region index currently copying */
1754     int              log_offset;     /* offset (from 0) into data region */
1755     int              start_rec_copy; /* # bytes to copy for start record */
1756     int              partial_copy;   /* did we split a region? */
1757     int              partial_copy_len;/* # bytes copied if split region */
1758     int              need_copy;      /* # bytes need to memcpy this region */
1759     int              copy_len;       /* # bytes actually memcpy'ing */
1760     int              copy_off;       /* # bytes from entry start */
1761     int              contwr;         /* continued write of in-core log? */
1762     int              error;
1763     int              record_cnt = 0, data_cnt = 0;
1764
1765     partial_copy_len = partial_copy = 0;
1766
1767     /* Calculate potential maximum space.  Each region gets its own
1768      * xlog_op_header_t and may need to be double word aligned.
1769      */
1770     len = 0;
1771     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1772         len += sizeof(xlog_op_header_t);
1773         XLOG_TIC_ADD_OPHDR(ticket);
1774     }
1775
1776     for (index = 0; index < nentries; index++) {
1777         len += sizeof(xlog_op_header_t);            /* each region gets >= 1 */
1778         XLOG_TIC_ADD_OPHDR(ticket);
1779         len += reg[index].i_len;
1780         XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1781     }
1782     contwr = *start_lsn = 0;
1783
1784     if (ticket->t_curr_res < len) {
1785         xlog_print_tic_res(mp, ticket);
1786 #ifdef DEBUG
1787         xlog_panic(
1788                 "xfs_log_write: reservation ran out. Need to up reservation");
1789 #else
1790         /* Customer configurable panic */
1791         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1792                 "xfs_log_write: reservation ran out. Need to up reservation");
1793         /* If we did not panic, shutdown the filesystem */
1794         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1795 #endif
1796     } else
1797         ticket->t_curr_res -= len;
1798
1799     for (index = 0; index < nentries; ) {
1800         if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1801                                                &contwr, &log_offset)))
1802                 return error;
1803
1804         ASSERT(log_offset <= iclog->ic_size - 1);
1805         ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1806
1807         /* start_lsn is the first lsn written to. That's all we need. */
1808         if (! *start_lsn)
1809             *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1810
1811         /* This loop writes out as many regions as can fit in the amount
1812          * of space which was allocated by xlog_state_get_iclog_space().
1813          */
1814         while (index < nentries) {
1815             ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1816             ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1817             start_rec_copy = 0;
1818
1819             /* If first write for transaction, insert start record.
1820              * We can't be trying to commit if we are inited.  We can't
1821              * have any "partial_copy" if we are inited.
1822              */
1823             if (ticket->t_flags & XLOG_TIC_INITED) {
1824                 logop_head              = (xlog_op_header_t *)ptr;
1825                 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1826                 logop_head->oh_clientid = ticket->t_clientid;
1827                 logop_head->oh_len      = 0;
1828                 logop_head->oh_flags    = XLOG_START_TRANS;
1829                 logop_head->oh_res2     = 0;
1830                 ticket->t_flags         &= ~XLOG_TIC_INITED;    /* clear bit */
1831                 record_cnt++;
1832
1833                 start_rec_copy = sizeof(xlog_op_header_t);
1834                 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1835             }
1836
1837             /* Copy log operation header directly into data section */
1838             logop_head                  = (xlog_op_header_t *)ptr;
1839             INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1840             logop_head->oh_clientid     = ticket->t_clientid;
1841             logop_head->oh_res2         = 0;
1842
1843             /* header copied directly */
1844             xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1845
1846             /* are we copying a commit or unmount record? */
1847             logop_head->oh_flags = flags;
1848
1849             /*
1850              * We've seen logs corrupted with bad transaction client
1851              * ids.  This makes sure that XFS doesn't generate them on.
1852              * Turn this into an EIO and shut down the filesystem.
1853              */
1854             switch (logop_head->oh_clientid)  {
1855             case XFS_TRANSACTION:
1856             case XFS_VOLUME:
1857             case XFS_LOG:
1858                 break;
1859             default:
1860                 xfs_fs_cmn_err(CE_WARN, mp,
1861                     "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1862                     logop_head->oh_clientid, tic);
1863                 return XFS_ERROR(EIO);
1864             }
1865
1866             /* Partial write last time? => (partial_copy != 0)
1867              * need_copy is the amount we'd like to copy if everything could
1868              * fit in the current memcpy.
1869              */
1870             need_copy = reg[index].i_len - partial_copy_len;
1871
1872             copy_off = partial_copy_len;
1873             if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1874                 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1875                 if (partial_copy)
1876                     logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1877                 partial_copy_len = partial_copy = 0;
1878             } else {                                        /* partial write */
1879                 copy_len = iclog->ic_size - log_offset;
1880                 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1881                 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1882                 if (partial_copy)
1883                         logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1884                 partial_copy_len += copy_len;
1885                 partial_copy++;
1886                 len += sizeof(xlog_op_header_t); /* from splitting of region */
1887                 /* account for new log op header */
1888                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1889                 XLOG_TIC_ADD_OPHDR(ticket);
1890             }
1891             xlog_verify_dest_ptr(log, ptr);
1892
1893             /* copy region */
1894             ASSERT(copy_len >= 0);
1895             memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1896             xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1897
1898             /* make copy_len total bytes copied, including headers */
1899             copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1900             record_cnt++;
1901             data_cnt += contwr ? copy_len : 0;
1902             if (partial_copy) {                 /* copied partial region */
1903                     /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1904                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1905                     record_cnt = data_cnt = 0;
1906                     if ((error = xlog_state_release_iclog(log, iclog)))
1907                             return error;
1908                     break;                      /* don't increment index */
1909             } else {                            /* copied entire region */
1910                 index++;
1911                 partial_copy_len = partial_copy = 0;
1912
1913                 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1914                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1915                     record_cnt = data_cnt = 0;
1916                     xlog_state_want_sync(log, iclog);
1917                     if (commit_iclog) {
1918                         ASSERT(flags & XLOG_COMMIT_TRANS);
1919                         *commit_iclog = iclog;
1920                     } else if ((error = xlog_state_release_iclog(log, iclog)))
1921                            return error;
1922                     if (index == nentries)
1923                             return 0;           /* we are done */
1924                     else
1925                             break;
1926                 }
1927             } /* if (partial_copy) */
1928         } /* while (index < nentries) */
1929     } /* for (index = 0; index < nentries; ) */
1930     ASSERT(len == 0);
1931
1932     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1933     if (commit_iclog) {
1934         ASSERT(flags & XLOG_COMMIT_TRANS);
1935         *commit_iclog = iclog;
1936         return 0;
1937     }
1938     return xlog_state_release_iclog(log, iclog);
1939 }       /* xlog_write */
1940
1941
1942 /*****************************************************************************
1943  *
1944  *              State Machine functions
1945  *
1946  *****************************************************************************
1947  */
1948
1949 /* Clean iclogs starting from the head.  This ordering must be
1950  * maintained, so an iclog doesn't become ACTIVE beyond one that
1951  * is SYNCING.  This is also required to maintain the notion that we use
1952  * a counting semaphore to hold off would be writers to the log when every
1953  * iclog is trying to sync to disk.
1954  *
1955  * State Change: DIRTY -> ACTIVE
1956  */
1957 STATIC void
1958 xlog_state_clean_log(xlog_t *log)
1959 {
1960         xlog_in_core_t  *iclog;
1961         int changed = 0;
1962
1963         iclog = log->l_iclog;
1964         do {
1965                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1966                         iclog->ic_state = XLOG_STATE_ACTIVE;
1967                         iclog->ic_offset       = 0;
1968                         iclog->ic_callback      = NULL;   /* don't need to free */
1969                         /*
1970                          * If the number of ops in this iclog indicate it just
1971                          * contains the dummy transaction, we can
1972                          * change state into IDLE (the second time around).
1973                          * Otherwise we should change the state into
1974                          * NEED a dummy.
1975                          * We don't need to cover the dummy.
1976                          */
1977                         if (!changed &&
1978                            (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1979                                 changed = 1;
1980                         } else {
1981                                 /*
1982                                  * We have two dirty iclogs so start over
1983                                  * This could also be num of ops indicates
1984                                  * this is not the dummy going out.
1985                                  */
1986                                 changed = 2;
1987                         }
1988                         iclog->ic_header.h_num_logops = 0;
1989                         memset(iclog->ic_header.h_cycle_data, 0,
1990                               sizeof(iclog->ic_header.h_cycle_data));
1991                         iclog->ic_header.h_lsn = 0;
1992                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1993                         /* do nothing */;
1994                 else
1995                         break;  /* stop cleaning */
1996                 iclog = iclog->ic_next;
1997         } while (iclog != log->l_iclog);
1998
1999         /* log is locked when we are called */
2000         /*
2001          * Change state for the dummy log recording.
2002          * We usually go to NEED. But we go to NEED2 if the changed indicates
2003          * we are done writing the dummy record.
2004          * If we are done with the second dummy recored (DONE2), then
2005          * we go to IDLE.
2006          */
2007         if (changed) {
2008                 switch (log->l_covered_state) {
2009                 case XLOG_STATE_COVER_IDLE:
2010                 case XLOG_STATE_COVER_NEED:
2011                 case XLOG_STATE_COVER_NEED2:
2012                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2013                         break;
2014
2015                 case XLOG_STATE_COVER_DONE:
2016                         if (changed == 1)
2017                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2018                         else
2019                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2020                         break;
2021
2022                 case XLOG_STATE_COVER_DONE2:
2023                         if (changed == 1)
2024                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2025                         else
2026                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2027                         break;
2028
2029                 default:
2030                         ASSERT(0);
2031                 }
2032         }
2033 }       /* xlog_state_clean_log */
2034
2035 STATIC xfs_lsn_t
2036 xlog_get_lowest_lsn(
2037         xlog_t          *log)
2038 {
2039         xlog_in_core_t  *lsn_log;
2040         xfs_lsn_t       lowest_lsn, lsn;
2041
2042         lsn_log = log->l_iclog;
2043         lowest_lsn = 0;
2044         do {
2045             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2046                 lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2047                 if ((lsn && !lowest_lsn) ||
2048                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2049                         lowest_lsn = lsn;
2050                 }
2051             }
2052             lsn_log = lsn_log->ic_next;
2053         } while (lsn_log != log->l_iclog);
2054         return lowest_lsn;
2055 }
2056
2057
2058 STATIC void
2059 xlog_state_do_callback(
2060         xlog_t          *log,
2061         int             aborted,
2062         xlog_in_core_t  *ciclog)
2063 {
2064         xlog_in_core_t     *iclog;
2065         xlog_in_core_t     *first_iclog;        /* used to know when we've
2066                                                  * processed all iclogs once */
2067         xfs_log_callback_t *cb, *cb_next;
2068         int                flushcnt = 0;
2069         xfs_lsn_t          lowest_lsn;
2070         int                ioerrors;    /* counter: iclogs with errors */
2071         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2072         int                funcdidcallbacks; /* flag: function did callbacks */
2073         int                repeats;     /* for issuing console warnings if
2074                                          * looping too many times */
2075         SPLDECL(s);
2076
2077         s = LOG_LOCK(log);
2078         first_iclog = iclog = log->l_iclog;
2079         ioerrors = 0;
2080         funcdidcallbacks = 0;
2081         repeats = 0;
2082
2083         do {
2084                 /*
2085                  * Scan all iclogs starting with the one pointed to by the
2086                  * log.  Reset this starting point each time the log is
2087                  * unlocked (during callbacks).
2088                  *
2089                  * Keep looping through iclogs until one full pass is made
2090                  * without running any callbacks.
2091                  */
2092                 first_iclog = log->l_iclog;
2093                 iclog = log->l_iclog;
2094                 loopdidcallbacks = 0;
2095                 repeats++;
2096
2097                 do {
2098
2099                         /* skip all iclogs in the ACTIVE & DIRTY states */
2100                         if (iclog->ic_state &
2101                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2102                                 iclog = iclog->ic_next;
2103                                 continue;
2104                         }
2105
2106                         /*
2107                          * Between marking a filesystem SHUTDOWN and stopping
2108                          * the log, we do flush all iclogs to disk (if there
2109                          * wasn't a log I/O error). So, we do want things to
2110                          * go smoothly in case of just a SHUTDOWN  w/o a
2111                          * LOG_IO_ERROR.
2112                          */
2113                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2114                                 /*
2115                                  * Can only perform callbacks in order.  Since
2116                                  * this iclog is not in the DONE_SYNC/
2117                                  * DO_CALLBACK state, we skip the rest and
2118                                  * just try to clean up.  If we set our iclog
2119                                  * to DO_CALLBACK, we will not process it when
2120                                  * we retry since a previous iclog is in the
2121                                  * CALLBACK and the state cannot change since
2122                                  * we are holding the LOG_LOCK.
2123                                  */
2124                                 if (!(iclog->ic_state &
2125                                         (XLOG_STATE_DONE_SYNC |
2126                                                  XLOG_STATE_DO_CALLBACK))) {
2127                                         if (ciclog && (ciclog->ic_state ==
2128                                                         XLOG_STATE_DONE_SYNC)) {
2129                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2130                                         }
2131                                         break;
2132                                 }
2133                                 /*
2134                                  * We now have an iclog that is in either the
2135                                  * DO_CALLBACK or DONE_SYNC states. The other
2136                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2137                                  * caught by the above if and are going to
2138                                  * clean (i.e. we aren't doing their callbacks)
2139                                  * see the above if.
2140                                  */
2141
2142                                 /*
2143                                  * We will do one more check here to see if we
2144                                  * have chased our tail around.
2145                                  */
2146
2147                                 lowest_lsn = xlog_get_lowest_lsn(log);
2148                                 if (lowest_lsn && (
2149                                         XFS_LSN_CMP(
2150                                                 lowest_lsn,
2151                                                 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2152                                         )<0)) {
2153                                         iclog = iclog->ic_next;
2154                                         continue; /* Leave this iclog for
2155                                                    * another thread */
2156                                 }
2157
2158                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2159
2160                                 LOG_UNLOCK(log, s);
2161
2162                                 /* l_last_sync_lsn field protected by
2163                                  * GRANT_LOCK. Don't worry about iclog's lsn.
2164                                  * No one else can be here except us.
2165                                  */
2166                                 s = GRANT_LOCK(log);
2167                                 ASSERT(XFS_LSN_CMP(
2168                                                 log->l_last_sync_lsn,
2169                                                 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2170                                         )<=0);
2171                                 log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2172                                 GRANT_UNLOCK(log, s);
2173
2174                                 /*
2175                                  * Keep processing entries in the callback list
2176                                  * until we come around and it is empty.  We
2177                                  * need to atomically see that the list is
2178                                  * empty and change the state to DIRTY so that
2179                                  * we don't miss any more callbacks being added.
2180                                  */
2181                                 s = LOG_LOCK(log);
2182                         } else {
2183                                 ioerrors++;
2184                         }
2185                         cb = iclog->ic_callback;
2186
2187                         while (cb != 0) {
2188                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2189                                 iclog->ic_callback = NULL;
2190                                 LOG_UNLOCK(log, s);
2191
2192                                 /* perform callbacks in the order given */
2193                                 for (; cb != 0; cb = cb_next) {
2194                                         cb_next = cb->cb_next;
2195                                         cb->cb_func(cb->cb_arg, aborted);
2196                                 }
2197                                 s = LOG_LOCK(log);
2198                                 cb = iclog->ic_callback;
2199                         }
2200
2201                         loopdidcallbacks++;
2202                         funcdidcallbacks++;
2203
2204                         ASSERT(iclog->ic_callback == 0);
2205                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2206                                 iclog->ic_state = XLOG_STATE_DIRTY;
2207
2208                         /*
2209                          * Transition from DIRTY to ACTIVE if applicable.
2210                          * NOP if STATE_IOERROR.
2211                          */
2212                         xlog_state_clean_log(log);
2213
2214                         /* wake up threads waiting in xfs_log_force() */
2215                         sv_broadcast(&iclog->ic_forcesema);
2216
2217                         iclog = iclog->ic_next;
2218                 } while (first_iclog != iclog);
2219
2220                 if (repeats > 5000) {
2221                         flushcnt += repeats;
2222                         repeats = 0;
2223                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2224                                 "%s: possible infinite loop (%d iterations)",
2225                                 __FUNCTION__, flushcnt);
2226                 }
2227         } while (!ioerrors && loopdidcallbacks);
2228
2229         /*
2230          * make one last gasp attempt to see if iclogs are being left in
2231          * limbo..
2232          */
2233 #ifdef DEBUG
2234         if (funcdidcallbacks) {
2235                 first_iclog = iclog = log->l_iclog;
2236                 do {
2237                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2238                         /*
2239                          * Terminate the loop if iclogs are found in states
2240                          * which will cause other threads to clean up iclogs.
2241                          *
2242                          * SYNCING - i/o completion will go through logs
2243                          * DONE_SYNC - interrupt thread should be waiting for
2244                          *              LOG_LOCK
2245                          * IOERROR - give up hope all ye who enter here
2246                          */
2247                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2248                             iclog->ic_state == XLOG_STATE_SYNCING ||
2249                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2250                             iclog->ic_state == XLOG_STATE_IOERROR )
2251                                 break;
2252                         iclog = iclog->ic_next;
2253                 } while (first_iclog != iclog);
2254         }
2255 #endif
2256
2257         flushcnt = 0;
2258         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2259                 flushcnt = log->l_flushcnt;
2260                 log->l_flushcnt = 0;
2261         }
2262         LOG_UNLOCK(log, s);
2263         while (flushcnt--)
2264                 vsema(&log->l_flushsema);
2265 }       /* xlog_state_do_callback */
2266
2267
2268 /*
2269  * Finish transitioning this iclog to the dirty state.
2270  *
2271  * Make sure that we completely execute this routine only when this is
2272  * the last call to the iclog.  There is a good chance that iclog flushes,
2273  * when we reach the end of the physical log, get turned into 2 separate
2274  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2275  * routine.  By using the reference count bwritecnt, we guarantee that only
2276  * the second completion goes through.
2277  *
2278  * Callbacks could take time, so they are done outside the scope of the
2279  * global state machine log lock.  Assume that the calls to cvsema won't
2280  * take a long time.  At least we know it won't sleep.
2281  */
2282 void
2283 xlog_state_done_syncing(
2284         xlog_in_core_t  *iclog,
2285         int             aborted)
2286 {
2287         xlog_t             *log = iclog->ic_log;
2288         SPLDECL(s);
2289
2290         s = LOG_LOCK(log);
2291
2292         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2293                iclog->ic_state == XLOG_STATE_IOERROR);
2294         ASSERT(iclog->ic_refcnt == 0);
2295         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2296
2297
2298         /*
2299          * If we got an error, either on the first buffer, or in the case of
2300          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2301          * and none should ever be attempted to be written to disk
2302          * again.
2303          */
2304         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2305                 if (--iclog->ic_bwritecnt == 1) {
2306                         LOG_UNLOCK(log, s);
2307                         return;
2308                 }
2309                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2310         }
2311
2312         /*
2313          * Someone could be sleeping prior to writing out the next
2314          * iclog buffer, we wake them all, one will get to do the
2315          * I/O, the others get to wait for the result.
2316          */
2317         sv_broadcast(&iclog->ic_writesema);
2318         LOG_UNLOCK(log, s);
2319         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2320 }       /* xlog_state_done_syncing */
2321
2322
2323 /*
2324  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2325  * sleep.  The flush semaphore is set to the number of in-core buffers and
2326  * decremented around disk syncing.  Therefore, if all buffers are syncing,
2327  * this semaphore will cause new writes to sleep until a sync completes.
2328  * Otherwise, this code just does p() followed by v().  This approximates
2329  * a sleep/wakeup except we can't race.
2330  *
2331  * The in-core logs are used in a circular fashion. They are not used
2332  * out-of-order even when an iclog past the head is free.
2333  *
2334  * return:
2335  *      * log_offset where xlog_write() can start writing into the in-core
2336  *              log's data space.
2337  *      * in-core log pointer to which xlog_write() should write.
2338  *      * boolean indicating this is a continued write to an in-core log.
2339  *              If this is the last write, then the in-core log's offset field
2340  *              needs to be incremented, depending on the amount of data which
2341  *              is copied.
2342  */
2343 int
2344 xlog_state_get_iclog_space(xlog_t         *log,
2345                            int            len,
2346                            xlog_in_core_t **iclogp,
2347                            xlog_ticket_t  *ticket,
2348                            int            *continued_write,
2349                            int            *logoffsetp)
2350 {
2351         SPLDECL(s);
2352         int               log_offset;
2353         xlog_rec_header_t *head;
2354         xlog_in_core_t    *iclog;
2355         int               error;
2356
2357 restart:
2358         s = LOG_LOCK(log);
2359         if (XLOG_FORCED_SHUTDOWN(log)) {
2360                 LOG_UNLOCK(log, s);
2361                 return XFS_ERROR(EIO);
2362         }
2363
2364         iclog = log->l_iclog;
2365         if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2366                 log->l_flushcnt++;
2367                 LOG_UNLOCK(log, s);
2368                 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2369                 XFS_STATS_INC(xs_log_noiclogs);
2370                 /* Ensure that log writes happen */
2371                 psema(&log->l_flushsema, PINOD);
2372                 goto restart;
2373         }
2374         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2375         head = &iclog->ic_header;
2376
2377         iclog->ic_refcnt++;                     /* prevents sync */
2378         log_offset = iclog->ic_offset;
2379
2380         /* On the 1st write to an iclog, figure out lsn.  This works
2381          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2382          * committing to.  If the offset is set, that's how many blocks
2383          * must be written.
2384          */
2385         if (log_offset == 0) {
2386                 ticket->t_curr_res -= log->l_iclog_hsize;
2387                 XLOG_TIC_ADD_REGION(ticket,
2388                                     log->l_iclog_hsize,
2389                                     XLOG_REG_TYPE_LRHEADER);
2390                 INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2391                 ASSIGN_LSN(head->h_lsn, log);
2392                 ASSERT(log->l_curr_block >= 0);
2393         }
2394
2395         /* If there is enough room to write everything, then do it.  Otherwise,
2396          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2397          * bit is on, so this will get flushed out.  Don't update ic_offset
2398          * until you know exactly how many bytes get copied.  Therefore, wait
2399          * until later to update ic_offset.
2400          *
2401          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2402          * can fit into remaining data section.
2403          */
2404         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2405                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2406
2407                 /* If I'm the only one writing to this iclog, sync it to disk */
2408                 if (iclog->ic_refcnt == 1) {
2409                         LOG_UNLOCK(log, s);
2410                         if ((error = xlog_state_release_iclog(log, iclog)))
2411                                 return error;
2412                 } else {
2413                         iclog->ic_refcnt--;
2414                         LOG_UNLOCK(log, s);
2415                 }
2416                 goto restart;
2417         }
2418
2419         /* Do we have enough room to write the full amount in the remainder
2420          * of this iclog?  Or must we continue a write on the next iclog and
2421          * mark this iclog as completely taken?  In the case where we switch
2422          * iclogs (to mark it taken), this particular iclog will release/sync
2423          * to disk in xlog_write().
2424          */
2425         if (len <= iclog->ic_size - iclog->ic_offset) {
2426                 *continued_write = 0;
2427                 iclog->ic_offset += len;
2428         } else {
2429                 *continued_write = 1;
2430                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2431         }
2432         *iclogp = iclog;
2433
2434         ASSERT(iclog->ic_offset <= iclog->ic_size);
2435         LOG_UNLOCK(log, s);
2436
2437         *logoffsetp = log_offset;
2438         return 0;
2439 }       /* xlog_state_get_iclog_space */
2440
2441 /*
2442  * Atomically get the log space required for a log ticket.
2443  *
2444  * Once a ticket gets put onto the reserveq, it will only return after
2445  * the needed reservation is satisfied.
2446  */
2447 STATIC int
2448 xlog_grant_log_space(xlog_t        *log,
2449                      xlog_ticket_t *tic)
2450 {
2451         int              free_bytes;
2452         int              need_bytes;
2453         SPLDECL(s);
2454 #ifdef DEBUG
2455         xfs_lsn_t        tail_lsn;
2456 #endif
2457
2458
2459 #ifdef DEBUG
2460         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2461                 panic("grant Recovery problem");
2462 #endif
2463
2464         /* Is there space or do we need to sleep? */
2465         s = GRANT_LOCK(log);
2466         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2467
2468         /* something is already sleeping; insert new transaction at end */
2469         if (log->l_reserve_headq) {
2470                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2471                 xlog_trace_loggrant(log, tic,
2472                                     "xlog_grant_log_space: sleep 1");
2473                 /*
2474                  * Gotta check this before going to sleep, while we're
2475                  * holding the grant lock.
2476                  */
2477                 if (XLOG_FORCED_SHUTDOWN(log))
2478                         goto error_return;
2479
2480                 XFS_STATS_INC(xs_sleep_logspace);
2481                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2482                 /*
2483                  * If we got an error, and the filesystem is shutting down,
2484                  * we'll catch it down below. So just continue...
2485                  */
2486                 xlog_trace_loggrant(log, tic,
2487                                     "xlog_grant_log_space: wake 1");
2488                 s = GRANT_LOCK(log);
2489         }
2490         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2491                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2492         else
2493                 need_bytes = tic->t_unit_res;
2494
2495 redo:
2496         if (XLOG_FORCED_SHUTDOWN(log))
2497                 goto error_return;
2498
2499         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2500                                      log->l_grant_reserve_bytes);
2501         if (free_bytes < need_bytes) {
2502                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2503                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2504                 xlog_trace_loggrant(log, tic,
2505                                     "xlog_grant_log_space: sleep 2");
2506                 XFS_STATS_INC(xs_sleep_logspace);
2507                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2508
2509                 if (XLOG_FORCED_SHUTDOWN(log)) {
2510                         s = GRANT_LOCK(log);
2511                         goto error_return;
2512                 }
2513
2514                 xlog_trace_loggrant(log, tic,
2515                                     "xlog_grant_log_space: wake 2");
2516                 xlog_grant_push_ail(log->l_mp, need_bytes);
2517                 s = GRANT_LOCK(log);
2518                 goto redo;
2519         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2520                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2521
2522         /* we've got enough space */
2523         xlog_grant_add_space(log, need_bytes);
2524 #ifdef DEBUG
2525         tail_lsn = log->l_tail_lsn;
2526         /*
2527          * Check to make sure the grant write head didn't just over lap the
2528          * tail.  If the cycles are the same, we can't be overlapping.
2529          * Otherwise, make sure that the cycles differ by exactly one and
2530          * check the byte count.
2531          */
2532         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2533                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2534                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2535         }
2536 #endif
2537         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2538         xlog_verify_grant_head(log, 1);
2539         GRANT_UNLOCK(log, s);
2540         return 0;
2541
2542  error_return:
2543         if (tic->t_flags & XLOG_TIC_IN_Q)
2544                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2545         xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2546         /*
2547          * If we are failing, make sure the ticket doesn't have any
2548          * current reservations. We don't want to add this back when
2549          * the ticket/transaction gets cancelled.
2550          */
2551         tic->t_curr_res = 0;
2552         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2553         GRANT_UNLOCK(log, s);
2554         return XFS_ERROR(EIO);
2555 }       /* xlog_grant_log_space */
2556
2557
2558 /*
2559  * Replenish the byte reservation required by moving the grant write head.
2560  *
2561  *
2562  */
2563 STATIC int
2564 xlog_regrant_write_log_space(xlog_t        *log,
2565                              xlog_ticket_t *tic)
2566 {
2567         SPLDECL(s);
2568         int             free_bytes, need_bytes;
2569         xlog_ticket_t   *ntic;
2570 #ifdef DEBUG
2571         xfs_lsn_t       tail_lsn;
2572 #endif
2573
2574         tic->t_curr_res = tic->t_unit_res;
2575         XLOG_TIC_RESET_RES(tic);
2576
2577         if (tic->t_cnt > 0)
2578                 return 0;
2579
2580 #ifdef DEBUG
2581         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2582                 panic("regrant Recovery problem");
2583 #endif
2584
2585         s = GRANT_LOCK(log);
2586         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2587
2588         if (XLOG_FORCED_SHUTDOWN(log))
2589                 goto error_return;
2590
2591         /* If there are other waiters on the queue then give them a
2592          * chance at logspace before us. Wake up the first waiters,
2593          * if we do not wake up all the waiters then go to sleep waiting
2594          * for more free space, otherwise try to get some space for
2595          * this transaction.
2596          */
2597
2598         if ((ntic = log->l_write_headq)) {
2599                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2600                                              log->l_grant_write_bytes);
2601                 do {
2602                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2603
2604                         if (free_bytes < ntic->t_unit_res)
2605                                 break;
2606                         free_bytes -= ntic->t_unit_res;
2607                         sv_signal(&ntic->t_sema);
2608                         ntic = ntic->t_next;
2609                 } while (ntic != log->l_write_headq);
2610
2611                 if (ntic != log->l_write_headq) {
2612                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2613                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2614
2615                         xlog_trace_loggrant(log, tic,
2616                                     "xlog_regrant_write_log_space: sleep 1");
2617                         XFS_STATS_INC(xs_sleep_logspace);
2618                         sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2619                                 &log->l_grant_lock, s);
2620
2621                         /* If we're shutting down, this tic is already
2622                          * off the queue */
2623                         if (XLOG_FORCED_SHUTDOWN(log)) {
2624                                 s = GRANT_LOCK(log);
2625                                 goto error_return;
2626                         }
2627
2628                         xlog_trace_loggrant(log, tic,
2629                                     "xlog_regrant_write_log_space: wake 1");
2630                         xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2631                         s = GRANT_LOCK(log);
2632                 }
2633         }
2634
2635         need_bytes = tic->t_unit_res;
2636
2637 redo:
2638         if (XLOG_FORCED_SHUTDOWN(log))
2639                 goto error_return;
2640
2641         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2642                                      log->l_grant_write_bytes);
2643         if (free_bytes < need_bytes) {
2644                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2645                         xlog_ins_ticketq(&log->l_write_headq, tic);
2646                 XFS_STATS_INC(xs_sleep_logspace);
2647                 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2648
2649                 /* If we're shutting down, this tic is already off the queue */
2650                 if (XLOG_FORCED_SHUTDOWN(log)) {
2651                         s = GRANT_LOCK(log);
2652                         goto error_return;
2653                 }
2654
2655                 xlog_trace_loggrant(log, tic,
2656                                     "xlog_regrant_write_log_space: wake 2");
2657                 xlog_grant_push_ail(log->l_mp, need_bytes);
2658                 s = GRANT_LOCK(log);
2659                 goto redo;
2660         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2661                 xlog_del_ticketq(&log->l_write_headq, tic);
2662
2663         /* we've got enough space */
2664         xlog_grant_add_space_write(log, need_bytes);
2665 #ifdef DEBUG
2666         tail_lsn = log->l_tail_lsn;
2667         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2668                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2669                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2670         }
2671 #endif
2672
2673         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2674         xlog_verify_grant_head(log, 1);
2675         GRANT_UNLOCK(log, s);
2676         return 0;
2677
2678
2679  error_return:
2680         if (tic->t_flags & XLOG_TIC_IN_Q)
2681                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2682         xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2683         /*
2684          * If we are failing, make sure the ticket doesn't have any
2685          * current reservations. We don't want to add this back when
2686          * the ticket/transaction gets cancelled.
2687          */
2688         tic->t_curr_res = 0;
2689         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2690         GRANT_UNLOCK(log, s);
2691         return XFS_ERROR(EIO);
2692 }       /* xlog_regrant_write_log_space */
2693
2694
2695 /* The first cnt-1 times through here we don't need to
2696  * move the grant write head because the permanent
2697  * reservation has reserved cnt times the unit amount.
2698  * Release part of current permanent unit reservation and
2699  * reset current reservation to be one units worth.  Also
2700  * move grant reservation head forward.
2701  */
2702 STATIC void
2703 xlog_regrant_reserve_log_space(xlog_t        *log,
2704                                xlog_ticket_t *ticket)
2705 {
2706         SPLDECL(s);
2707
2708         xlog_trace_loggrant(log, ticket,
2709                             "xlog_regrant_reserve_log_space: enter");
2710         if (ticket->t_cnt > 0)
2711                 ticket->t_cnt--;
2712
2713         s = GRANT_LOCK(log);
2714         xlog_grant_sub_space(log, ticket->t_curr_res);
2715         ticket->t_curr_res = ticket->t_unit_res;
2716         XLOG_TIC_RESET_RES(ticket);
2717         xlog_trace_loggrant(log, ticket,
2718                             "xlog_regrant_reserve_log_space: sub current res");
2719         xlog_verify_grant_head(log, 1);
2720
2721         /* just return if we still have some of the pre-reserved space */
2722         if (ticket->t_cnt > 0) {
2723                 GRANT_UNLOCK(log, s);
2724                 return;
2725         }
2726
2727         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2728         xlog_trace_loggrant(log, ticket,
2729                             "xlog_regrant_reserve_log_space: exit");
2730         xlog_verify_grant_head(log, 0);
2731         GRANT_UNLOCK(log, s);
2732         ticket->t_curr_res = ticket->t_unit_res;
2733         XLOG_TIC_RESET_RES(ticket);
2734 }       /* xlog_regrant_reserve_log_space */
2735
2736
2737 /*
2738  * Give back the space left from a reservation.
2739  *
2740  * All the information we need to make a correct determination of space left
2741  * is present.  For non-permanent reservations, things are quite easy.  The
2742  * count should have been decremented to zero.  We only need to deal with the
2743  * space remaining in the current reservation part of the ticket.  If the
2744  * ticket contains a permanent reservation, there may be left over space which
2745  * needs to be released.  A count of N means that N-1 refills of the current
2746  * reservation can be done before we need to ask for more space.  The first
2747  * one goes to fill up the first current reservation.  Once we run out of
2748  * space, the count will stay at zero and the only space remaining will be
2749  * in the current reservation field.
2750  */
2751 STATIC void
2752 xlog_ungrant_log_space(xlog_t        *log,
2753                        xlog_ticket_t *ticket)
2754 {
2755         SPLDECL(s);
2756
2757         if (ticket->t_cnt > 0)
2758                 ticket->t_cnt--;
2759
2760         s = GRANT_LOCK(log);
2761         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2762
2763         xlog_grant_sub_space(log, ticket->t_curr_res);
2764
2765         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2766
2767         /* If this is a permanent reservation ticket, we may be able to free
2768          * up more space based on the remaining count.
2769          */
2770         if (ticket->t_cnt > 0) {
2771                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2772                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2773         }
2774
2775         xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2776         xlog_verify_grant_head(log, 1);
2777         GRANT_UNLOCK(log, s);
2778         xfs_log_move_tail(log->l_mp, 1);
2779 }       /* xlog_ungrant_log_space */
2780
2781
2782 /*
2783  * Atomically put back used ticket.
2784  */
2785 void
2786 xlog_state_put_ticket(xlog_t        *log,
2787                       xlog_ticket_t *tic)
2788 {
2789         unsigned long s;
2790
2791         s = LOG_LOCK(log);
2792         xlog_ticket_put(log, tic);
2793         LOG_UNLOCK(log, s);
2794 }       /* xlog_state_put_ticket */
2795
2796 /*
2797  * Flush iclog to disk if this is the last reference to the given iclog and
2798  * the WANT_SYNC bit is set.
2799  *
2800  * When this function is entered, the iclog is not necessarily in the
2801  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2802  *
2803  *
2804  */
2805 int
2806 xlog_state_release_iclog(xlog_t         *log,
2807                          xlog_in_core_t *iclog)
2808 {
2809         SPLDECL(s);
2810         int             sync = 0;       /* do we sync? */
2811
2812         xlog_assign_tail_lsn(log->l_mp);
2813
2814         s = LOG_LOCK(log);
2815
2816         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2817                 LOG_UNLOCK(log, s);
2818                 return XFS_ERROR(EIO);
2819         }
2820
2821         ASSERT(iclog->ic_refcnt > 0);
2822         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2823                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2824
2825         if (--iclog->ic_refcnt == 0 &&
2826             iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2827                 sync++;
2828                 iclog->ic_state = XLOG_STATE_SYNCING;
2829                 INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2830                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2831                 /* cycle incremented when incrementing curr_block */
2832         }
2833
2834         LOG_UNLOCK(log, s);
2835
2836         /*
2837          * We let the log lock go, so it's possible that we hit a log I/O
2838          * error or some other SHUTDOWN condition that marks the iclog
2839          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2840          * this iclog has consistent data, so we ignore IOERROR
2841          * flags after this point.
2842          */
2843         if (sync) {
2844                 return xlog_sync(log, iclog);
2845         }
2846         return 0;
2847
2848 }       /* xlog_state_release_iclog */
2849
2850
2851 /*
2852  * This routine will mark the current iclog in the ring as WANT_SYNC
2853  * and move the current iclog pointer to the next iclog in the ring.
2854  * When this routine is called from xlog_state_get_iclog_space(), the
2855  * exact size of the iclog has not yet been determined.  All we know is
2856  * that every data block.  We have run out of space in this log record.
2857  */
2858 STATIC void
2859 xlog_state_switch_iclogs(xlog_t         *log,
2860                          xlog_in_core_t *iclog,
2861                          int            eventual_size)
2862 {
2863         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2864         if (!eventual_size)
2865                 eventual_size = iclog->ic_offset;
2866         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2867         INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2868         log->l_prev_block = log->l_curr_block;
2869         log->l_prev_cycle = log->l_curr_cycle;
2870
2871         /* roll log?: ic_offset changed later */
2872         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2873
2874         /* Round up to next log-sunit */
2875         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2876             log->l_mp->m_sb.sb_logsunit > 1) {
2877                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2878                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2879         }
2880
2881         if (log->l_curr_block >= log->l_logBBsize) {
2882                 log->l_curr_cycle++;
2883                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2884                         log->l_curr_cycle++;
2885                 log->l_curr_block -= log->l_logBBsize;
2886                 ASSERT(log->l_curr_block >= 0);
2887         }
2888         ASSERT(iclog == log->l_iclog);
2889         log->l_iclog = iclog->ic_next;
2890 }       /* xlog_state_switch_iclogs */
2891
2892
2893 /*
2894  * Write out all data in the in-core log as of this exact moment in time.
2895  *
2896  * Data may be written to the in-core log during this call.  However,
2897  * we don't guarantee this data will be written out.  A change from past
2898  * implementation means this routine will *not* write out zero length LRs.
2899  *
2900  * Basically, we try and perform an intelligent scan of the in-core logs.
2901  * If we determine there is no flushable data, we just return.  There is no
2902  * flushable data if:
2903  *
2904  *      1. the current iclog is active and has no data; the previous iclog
2905  *              is in the active or dirty state.
2906  *      2. the current iclog is drity, and the previous iclog is in the
2907  *              active or dirty state.
2908  *
2909  * We may sleep (call psema) if:
2910  *
2911  *      1. the current iclog is not in the active nor dirty state.
2912  *      2. the current iclog dirty, and the previous iclog is not in the
2913  *              active nor dirty state.
2914  *      3. the current iclog is active, and there is another thread writing
2915  *              to this particular iclog.
2916  *      4. a) the current iclog is active and has no other writers
2917  *         b) when we return from flushing out this iclog, it is still
2918  *              not in the active nor dirty state.
2919  */
2920 STATIC int
2921 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2922 {
2923         xlog_in_core_t  *iclog;
2924         xfs_lsn_t       lsn;
2925         SPLDECL(s);
2926
2927         s = LOG_LOCK(log);
2928
2929         iclog = log->l_iclog;
2930         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2931                 LOG_UNLOCK(log, s);
2932                 return XFS_ERROR(EIO);
2933         }
2934
2935         /* If the head iclog is not active nor dirty, we just attach
2936          * ourselves to the head and go to sleep.
2937          */
2938         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2939             iclog->ic_state == XLOG_STATE_DIRTY) {
2940                 /*
2941                  * If the head is dirty or (active and empty), then
2942                  * we need to look at the previous iclog.  If the previous
2943                  * iclog is active or dirty we are done.  There is nothing
2944                  * to sync out.  Otherwise, we attach ourselves to the
2945                  * previous iclog and go to sleep.
2946                  */
2947                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2948                     (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2949                         iclog = iclog->ic_prev;
2950                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2951                             iclog->ic_state == XLOG_STATE_DIRTY)
2952                                 goto no_sleep;
2953                         else
2954                                 goto maybe_sleep;
2955                 } else {
2956                         if (iclog->ic_refcnt == 0) {
2957                                 /* We are the only one with access to this
2958                                  * iclog.  Flush it out now.  There should
2959                                  * be a roundoff of zero to show that someone
2960                                  * has already taken care of the roundoff from
2961                                  * the previous sync.
2962                                  */
2963                                 iclog->ic_refcnt++;
2964                                 lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2965                                 xlog_state_switch_iclogs(log, iclog, 0);
2966                                 LOG_UNLOCK(log, s);
2967
2968                                 if (xlog_state_release_iclog(log, iclog))
2969                                         return XFS_ERROR(EIO);
2970                                 *log_flushed = 1;
2971                                 s = LOG_LOCK(log);
2972                                 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2973                                     iclog->ic_state != XLOG_STATE_DIRTY)
2974                                         goto maybe_sleep;
2975                                 else
2976                                         goto no_sleep;
2977                         } else {
2978                                 /* Someone else is writing to this iclog.
2979                                  * Use its call to flush out the data.  However,
2980                                  * the other thread may not force out this LR,
2981                                  * so we mark it WANT_SYNC.
2982                                  */
2983                                 xlog_state_switch_iclogs(log, iclog, 0);
2984                                 goto maybe_sleep;
2985                         }
2986                 }
2987         }
2988
2989         /* By the time we come around again, the iclog could've been filled
2990          * which would give it another lsn.  If we have a new lsn, just
2991          * return because the relevant data has been flushed.
2992          */
2993 maybe_sleep:
2994         if (flags & XFS_LOG_SYNC) {
2995                 /*
2996                  * We must check if we're shutting down here, before
2997                  * we wait, while we're holding the LOG_LOCK.
2998                  * Then we check again after waking up, in case our
2999                  * sleep was disturbed by a bad news.
3000                  */
3001                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3002                         LOG_UNLOCK(log, s);
3003                         return XFS_ERROR(EIO);
3004                 }
3005                 XFS_STATS_INC(xs_log_force_sleep);
3006                 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3007                 /*
3008                  * No need to grab the log lock here since we're
3009                  * only deciding whether or not to return EIO
3010                  * and the memory read should be atomic.
3011                  */
3012                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3013                         return XFS_ERROR(EIO);
3014                 *log_flushed = 1;
3015
3016         } else {
3017
3018 no_sleep:
3019                 LOG_UNLOCK(log, s);
3020         }
3021         return 0;
3022 }       /* xlog_state_sync_all */
3023
3024
3025 /*
3026  * Used by code which implements synchronous log forces.
3027  *
3028  * Find in-core log with lsn.
3029  *      If it is in the DIRTY state, just return.
3030  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3031  *              state and go to sleep or return.
3032  *      If it is in any other state, go to sleep or return.
3033  *
3034  * If filesystem activity goes to zero, the iclog will get flushed only by
3035  * bdflush().
3036  */
3037 int
3038 xlog_state_sync(xlog_t    *log,
3039                 xfs_lsn_t lsn,
3040                 uint      flags,
3041                 int       *log_flushed)
3042 {
3043     xlog_in_core_t      *iclog;
3044     int                 already_slept = 0;
3045     SPLDECL(s);
3046
3047
3048 try_again:
3049     s = LOG_LOCK(log);
3050     iclog = log->l_iclog;
3051
3052     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3053             LOG_UNLOCK(log, s);
3054             return XFS_ERROR(EIO);
3055     }
3056
3057     do {
3058         if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3059             iclog = iclog->ic_next;
3060             continue;
3061         }
3062
3063         if (iclog->ic_state == XLOG_STATE_DIRTY) {
3064                 LOG_UNLOCK(log, s);
3065                 return 0;
3066         }
3067
3068         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3069                 /*
3070                  * We sleep here if we haven't already slept (e.g.
3071                  * this is the first time we've looked at the correct
3072                  * iclog buf) and the buffer before us is going to
3073                  * be sync'ed. The reason for this is that if we
3074                  * are doing sync transactions here, by waiting for
3075                  * the previous I/O to complete, we can allow a few
3076                  * more transactions into this iclog before we close
3077                  * it down.
3078                  *
3079                  * Otherwise, we mark the buffer WANT_SYNC, and bump
3080                  * up the refcnt so we can release the log (which drops
3081                  * the ref count).  The state switch keeps new transaction
3082                  * commits from using this buffer.  When the current commits
3083                  * finish writing into the buffer, the refcount will drop to
3084                  * zero and the buffer will go out then.
3085                  */
3086                 if (!already_slept &&
3087                     (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3088                                                  XLOG_STATE_SYNCING))) {
3089                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3090                         XFS_STATS_INC(xs_log_force_sleep);
3091                         sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3092                                 &log->l_icloglock, s);
3093                         *log_flushed = 1;
3094                         already_slept = 1;
3095                         goto try_again;
3096                 } else {
3097                         iclog->ic_refcnt++;
3098                         xlog_state_switch_iclogs(log, iclog, 0);
3099                         LOG_UNLOCK(log, s);
3100                         if (xlog_state_release_iclog(log, iclog))
3101                                 return XFS_ERROR(EIO);
3102                         *log_flushed = 1;
3103                         s = LOG_LOCK(log);
3104                 }
3105         }
3106
3107         if ((flags & XFS_LOG_SYNC) && /* sleep */
3108             !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3109
3110                 /*
3111                  * Don't wait on the forcesema if we know that we've
3112                  * gotten a log write error.
3113                  */
3114                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3115                         LOG_UNLOCK(log, s);
3116                         return XFS_ERROR(EIO);
3117                 }
3118                 XFS_STATS_INC(xs_log_force_sleep);
3119                 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3120                 /*
3121                  * No need to grab the log lock here since we're
3122                  * only deciding whether or not to return EIO
3123                  * and the memory read should be atomic.
3124                  */
3125                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3126                         return XFS_ERROR(EIO);
3127                 *log_flushed = 1;
3128         } else {                /* just return */
3129                 LOG_UNLOCK(log, s);
3130         }
3131         return 0;
3132
3133     } while (iclog != log->l_iclog);
3134
3135     LOG_UNLOCK(log, s);
3136     return 0;
3137 }       /* xlog_state_sync */
3138
3139
3140 /*
3141  * Called when we want to mark the current iclog as being ready to sync to
3142  * disk.
3143  */
3144 void
3145 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3146 {
3147         SPLDECL(s);
3148
3149         s = LOG_LOCK(log);
3150
3151         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3152                 xlog_state_switch_iclogs(log, iclog, 0);
3153         } else {
3154                 ASSERT(iclog->ic_state &
3155                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3156         }
3157
3158         LOG_UNLOCK(log, s);
3159 }       /* xlog_state_want_sync */
3160
3161
3162
3163 /*****************************************************************************
3164  *
3165  *              TICKET functions
3166  *
3167  *****************************************************************************
3168  */
3169
3170 /*
3171  *      Algorithm doesn't take into account page size. ;-(
3172  */
3173 STATIC void
3174 xlog_state_ticket_alloc(xlog_t *log)
3175 {
3176         xlog_ticket_t   *t_list;
3177         xlog_ticket_t   *next;
3178         xfs_caddr_t     buf;
3179         uint            i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3180         SPLDECL(s);
3181
3182         /*
3183          * The kmem_zalloc may sleep, so we shouldn't be holding the
3184          * global lock.  XXXmiken: may want to use zone allocator.
3185          */
3186         buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3187
3188         s = LOG_LOCK(log);
3189
3190         /* Attach 1st ticket to Q, so we can keep track of allocated memory */
3191         t_list = (xlog_ticket_t *)buf;
3192         t_list->t_next = log->l_unmount_free;
3193         log->l_unmount_free = t_list++;
3194         log->l_ticket_cnt++;
3195         log->l_ticket_tcnt++;
3196
3197         /* Next ticket becomes first ticket attached to ticket free list */
3198         if (log->l_freelist != NULL) {
3199                 ASSERT(log->l_tail != NULL);
3200                 log->l_tail->t_next = t_list;
3201         } else {
3202                 log->l_freelist = t_list;
3203         }
3204         log->l_ticket_cnt++;
3205         log->l_ticket_tcnt++;
3206
3207         /* Cycle through rest of alloc'ed memory, building up free Q */
3208         for ( ; i > 0; i--) {
3209                 next = t_list + 1;
3210                 t_list->t_next = next;
3211                 t_list = next;
3212                 log->l_ticket_cnt++;
3213                 log->l_ticket_tcnt++;
3214         }
3215         t_list->t_next = NULL;
3216         log->l_tail = t_list;
3217         LOG_UNLOCK(log, s);
3218 }       /* xlog_state_ticket_alloc */
3219
3220
3221 /*
3222  * Put ticket into free list
3223  *
3224  * Assumption: log lock is held around this call.
3225  */
3226 STATIC void
3227 xlog_ticket_put(xlog_t          *log,
3228                 xlog_ticket_t   *ticket)
3229 {
3230         sv_destroy(&ticket->t_sema);
3231
3232         /*
3233          * Don't think caching will make that much difference.  It's
3234          * more important to make debug easier.
3235          */
3236 #if 0
3237         /* real code will want to use LIFO for caching */
3238         ticket->t_next = log->l_freelist;
3239         log->l_freelist = ticket;
3240         /* no need to clear fields */
3241 #else
3242         /* When we debug, it is easier if tickets are cycled */
3243         ticket->t_next     = NULL;
3244         if (log->l_tail != 0) {
3245                 log->l_tail->t_next = ticket;
3246         } else {
3247                 ASSERT(log->l_freelist == 0);
3248                 log->l_freelist = ticket;
3249         }
3250         log->l_tail         = ticket;
3251 #endif /* DEBUG */
3252         log->l_ticket_cnt++;
3253 }       /* xlog_ticket_put */
3254
3255
3256 /*
3257  * Grab ticket off freelist or allocation some more
3258  */
3259 xlog_ticket_t *
3260 xlog_ticket_get(xlog_t          *log,
3261                 int             unit_bytes,
3262                 int             cnt,
3263                 char            client,
3264                 uint            xflags)
3265 {
3266         xlog_ticket_t   *tic;
3267         uint            num_headers;
3268         SPLDECL(s);
3269
3270  alloc:
3271         if (log->l_freelist == NULL)
3272                 xlog_state_ticket_alloc(log);           /* potentially sleep */
3273
3274         s = LOG_LOCK(log);
3275         if (log->l_freelist == NULL) {
3276                 LOG_UNLOCK(log, s);
3277                 goto alloc;
3278         }
3279         tic             = log->l_freelist;
3280         log->l_freelist = tic->t_next;
3281         if (log->l_freelist == NULL)
3282                 log->l_tail = NULL;
3283         log->l_ticket_cnt--;
3284         LOG_UNLOCK(log, s);
3285
3286         /*
3287          * Permanent reservations have up to 'cnt'-1 active log operations
3288          * in the log.  A unit in this case is the amount of space for one
3289          * of these log operations.  Normal reservations have a cnt of 1
3290          * and their unit amount is the total amount of space required.
3291          *
3292          * The following lines of code account for non-transaction data
3293          * which occupy space in the on-disk log.
3294          *
3295          * Normal form of a transaction is:
3296          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3297          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3298          *
3299          * We need to account for all the leadup data and trailer data
3300          * around the transaction data.
3301          * And then we need to account for the worst case in terms of using
3302          * more space.
3303          * The worst case will happen if:
3304          * - the placement of the transaction happens to be such that the
3305          *   roundoff is at its maximum
3306          * - the transaction data is synced before the commit record is synced
3307          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3308          *   Therefore the commit record is in its own Log Record.
3309          *   This can happen as the commit record is called with its
3310          *   own region to xlog_write().
3311          *   This then means that in the worst case, roundoff can happen for
3312          *   the commit-rec as well.
3313          *   The commit-rec is smaller than padding in this scenario and so it is
3314          *   not added separately.
3315          */
3316
3317         /* for trans header */
3318         unit_bytes += sizeof(xlog_op_header_t);
3319         unit_bytes += sizeof(xfs_trans_header_t);
3320
3321         /* for start-rec */
3322         unit_bytes += sizeof(xlog_op_header_t);
3323
3324         /* for LR headers */
3325         num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3326         unit_bytes += log->l_iclog_hsize * num_headers;
3327
3328         /* for commit-rec LR header - note: padding will subsume the ophdr */
3329         unit_bytes += log->l_iclog_hsize;
3330
3331         /* for split-recs - ophdrs added when data split over LRs */
3332         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3333
3334         /* for roundoff padding for transaction data and one for commit record */
3335         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3336             log->l_mp->m_sb.sb_logsunit > 1) {
3337                 /* log su roundoff */
3338                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3339         } else {
3340                 /* BB roundoff */
3341                 unit_bytes += 2*BBSIZE;
3342         }
3343
3344         tic->t_unit_res         = unit_bytes;
3345         tic->t_curr_res         = unit_bytes;
3346         tic->t_cnt              = cnt;
3347         tic->t_ocnt             = cnt;
3348         tic->t_tid              = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3349         tic->t_clientid         = client;
3350         tic->t_flags            = XLOG_TIC_INITED;
3351         tic->t_trans_type       = 0;
3352         if (xflags & XFS_LOG_PERM_RESERV)
3353                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3354         sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3355
3356         XLOG_TIC_RESET_RES(tic);
3357
3358         return tic;
3359 }       /* xlog_ticket_get */
3360
3361
3362 /******************************************************************************
3363  *
3364  *              Log debug routines
3365  *
3366  ******************************************************************************
3367  */
3368 #if defined(DEBUG)
3369 /*
3370  * Make sure that the destination ptr is within the valid data region of
3371  * one of the iclogs.  This uses backup pointers stored in a different
3372  * part of the log in case we trash the log structure.
3373  */
3374 void
3375 xlog_verify_dest_ptr(xlog_t     *log,
3376                      __psint_t  ptr)
3377 {
3378         int i;
3379         int good_ptr = 0;
3380
3381         for (i=0; i < log->l_iclog_bufs; i++) {
3382                 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3383                     ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3384                         good_ptr++;
3385         }
3386         if (! good_ptr)
3387                 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3388 }       /* xlog_verify_dest_ptr */
3389
3390 STATIC void
3391 xlog_verify_grant_head(xlog_t *log, int equals)
3392 {
3393     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3394         if (equals)
3395             ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3396         else
3397             ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3398     } else {
3399         ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3400         ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3401     }
3402 }       /* xlog_verify_grant_head */
3403
3404 /* check if it will fit */
3405 STATIC void
3406 xlog_verify_tail_lsn(xlog_t         *log,
3407                      xlog_in_core_t *iclog,
3408                      xfs_lsn_t      tail_lsn)
3409 {
3410     int blocks;
3411
3412     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3413         blocks =
3414             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3415         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3416             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3417     } else {
3418         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3419
3420         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3421             xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3422
3423         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3424         if (blocks < BTOBB(iclog->ic_offset) + 1)
3425             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3426     }
3427 }       /* xlog_verify_tail_lsn */
3428
3429 /*
3430  * Perform a number of checks on the iclog before writing to disk.
3431  *
3432  * 1. Make sure the iclogs are still circular
3433  * 2. Make sure we have a good magic number
3434  * 3. Make sure we don't have magic numbers in the data
3435  * 4. Check fields of each log operation header for:
3436  *      A. Valid client identifier
3437  *      B. tid ptr value falls in valid ptr space (user space code)
3438  *      C. Length in log record header is correct according to the
3439  *              individual operation headers within record.
3440  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3441  *      log, check the preceding blocks of the physical log to make sure all
3442  *      the cycle numbers agree with the current cycle number.
3443  */
3444 STATIC void
3445 xlog_verify_iclog(xlog_t         *log,
3446                   xlog_in_core_t *iclog,
3447                   int            count,
3448                   boolean_t      syncing)
3449 {
3450         xlog_op_header_t        *ophead;
3451         xlog_in_core_t          *icptr;
3452         xlog_in_core_2_t        *xhdr;
3453         xfs_caddr_t             ptr;
3454         xfs_caddr_t             base_ptr;
3455         __psint_t               field_offset;
3456         __uint8_t               clientid;
3457         int                     len, i, j, k, op_len;
3458         int                     idx;
3459         SPLDECL(s);
3460
3461         /* check validity of iclog pointers */
3462         s = LOG_LOCK(log);
3463         icptr = log->l_iclog;
3464         for (i=0; i < log->l_iclog_bufs; i++) {
3465                 if (icptr == 0)
3466                         xlog_panic("xlog_verify_iclog: invalid ptr");
3467                 icptr = icptr->ic_next;
3468         }
3469         if (icptr != log->l_iclog)
3470                 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3471         LOG_UNLOCK(log, s);
3472
3473         /* check log magic numbers */
3474         ptr = (xfs_caddr_t) &(iclog->ic_header);
3475         if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3476                 xlog_panic("xlog_verify_iclog: invalid magic num");
3477
3478         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3479              ptr += BBSIZE) {
3480                 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3481                         xlog_panic("xlog_verify_iclog: unexpected magic num");
3482         }
3483
3484         /* check fields */
3485         len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3486         ptr = iclog->ic_datap;
3487         base_ptr = ptr;
3488         ophead = (xlog_op_header_t *)ptr;
3489         xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3490         for (i = 0; i < len; i++) {
3491                 ophead = (xlog_op_header_t *)ptr;
3492
3493                 /* clientid is only 1 byte */
3494                 field_offset = (__psint_t)
3495                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3496                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3497                         clientid = ophead->oh_clientid;
3498                 } else {
3499                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3500                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3501                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3502                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3503                                 clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3504                         } else {
3505                                 clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3506                         }
3507                 }
3508                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3509                         cmn_err(CE_WARN, "xlog_verify_iclog: "
3510                                 "invalid clientid %d op 0x%p offset 0x%lx",
3511                                 clientid, ophead, (unsigned long)field_offset);
3512
3513                 /* check length */
3514                 field_offset = (__psint_t)
3515                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3516                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3517                         op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3518                 } else {
3519                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3520                                     (__psint_t)iclog->ic_datap);
3521                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3522                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3523                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3524                                 op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3525                         } else {
3526                                 op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3527                         }
3528                 }
3529                 ptr += sizeof(xlog_op_header_t) + op_len;
3530         }
3531 }       /* xlog_verify_iclog */
3532 #endif
3533
3534 /*
3535  * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3536  */
3537 STATIC int
3538 xlog_state_ioerror(
3539         xlog_t  *log)
3540 {
3541         xlog_in_core_t  *iclog, *ic;
3542
3543         iclog = log->l_iclog;
3544         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3545                 /*
3546                  * Mark all the incore logs IOERROR.
3547                  * From now on, no log flushes will result.
3548                  */
3549                 ic = iclog;
3550                 do {
3551                         ic->ic_state = XLOG_STATE_IOERROR;
3552                         ic = ic->ic_next;
3553                 } while (ic != iclog);
3554                 return 0;
3555         }
3556         /*
3557          * Return non-zero, if state transition has already happened.
3558          */
3559         return 1;
3560 }
3561
3562 /*
3563  * This is called from xfs_force_shutdown, when we're forcibly
3564  * shutting down the filesystem, typically because of an IO error.
3565  * Our main objectives here are to make sure that:
3566  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3567  *         parties to find out, 'atomically'.
3568  *      b. those who're sleeping on log reservations, pinned objects and
3569  *          other resources get woken up, and be told the bad news.
3570  *      c. nothing new gets queued up after (a) and (b) are done.
3571  *      d. if !logerror, flush the iclogs to disk, then seal them off
3572  *         for business.
3573  */
3574 int
3575 xfs_log_force_umount(
3576         struct xfs_mount        *mp,
3577         int                     logerror)
3578 {
3579         xlog_ticket_t   *tic;
3580         xlog_t          *log;
3581         int             retval;
3582         int             dummy;
3583         SPLDECL(s);
3584         SPLDECL(s2);
3585
3586         log = mp->m_log;
3587
3588         /*
3589          * If this happens during log recovery, don't worry about
3590          * locking; the log isn't open for business yet.
3591          */
3592         if (!log ||
3593             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3594                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3595                 XFS_BUF_DONE(mp->m_sb_bp);
3596                 return 0;
3597         }
3598
3599         /*
3600          * Somebody could've already done the hard work for us.
3601          * No need to get locks for this.
3602          */
3603         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3604                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3605                 return 1;
3606         }
3607         retval = 0;
3608         /*
3609          * We must hold both the GRANT lock and the LOG lock,
3610          * before we mark the filesystem SHUTDOWN and wake
3611          * everybody up to tell the bad news.
3612          */
3613         s = GRANT_LOCK(log);
3614         s2 = LOG_LOCK(log);
3615         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3616         XFS_BUF_DONE(mp->m_sb_bp);
3617         /*
3618          * This flag is sort of redundant because of the mount flag, but
3619          * it's good to maintain the separation between the log and the rest
3620          * of XFS.
3621          */
3622         log->l_flags |= XLOG_IO_ERROR;
3623
3624         /*
3625          * If we hit a log error, we want to mark all the iclogs IOERROR
3626          * while we're still holding the loglock.
3627          */
3628         if (logerror)
3629                 retval = xlog_state_ioerror(log);
3630         LOG_UNLOCK(log, s2);
3631
3632         /*
3633          * We don't want anybody waiting for log reservations
3634          * after this. That means we have to wake up everybody
3635          * queued up on reserve_headq as well as write_headq.
3636          * In addition, we make sure in xlog_{re}grant_log_space
3637          * that we don't enqueue anything once the SHUTDOWN flag
3638          * is set, and this action is protected by the GRANTLOCK.
3639          */
3640         if ((tic = log->l_reserve_headq)) {
3641                 do {
3642                         sv_signal(&tic->t_sema);
3643                         tic = tic->t_next;
3644                 } while (tic != log->l_reserve_headq);
3645         }
3646
3647         if ((tic = log->l_write_headq)) {
3648                 do {
3649                         sv_signal(&tic->t_sema);
3650                         tic = tic->t_next;
3651                 } while (tic != log->l_write_headq);
3652         }
3653         GRANT_UNLOCK(log, s);
3654
3655         if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3656                 ASSERT(!logerror);
3657                 /*
3658                  * Force the incore logs to disk before shutting the
3659                  * log down completely.
3660                  */
3661                 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3662                 s2 = LOG_LOCK(log);
3663                 retval = xlog_state_ioerror(log);
3664                 LOG_UNLOCK(log, s2);
3665         }
3666         /*
3667          * Wake up everybody waiting on xfs_log_force.
3668          * Callback all log item committed functions as if the
3669          * log writes were completed.
3670          */
3671         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3672
3673 #ifdef XFSERRORDEBUG
3674         {
3675                 xlog_in_core_t  *iclog;
3676
3677                 s = LOG_LOCK(log);
3678                 iclog = log->l_iclog;
3679                 do {
3680                         ASSERT(iclog->ic_callback == 0);
3681                         iclog = iclog->ic_next;
3682                 } while (iclog != log->l_iclog);
3683                 LOG_UNLOCK(log, s);
3684         }
3685 #endif
3686         /* return non-zero if log IOERROR transition had already happened */
3687         return retval;
3688 }
3689
3690 STATIC int
3691 xlog_iclogs_empty(xlog_t *log)
3692 {
3693         xlog_in_core_t  *iclog;
3694
3695         iclog = log->l_iclog;
3696         do {
3697                 /* endianness does not matter here, zero is zero in
3698                  * any language.
3699                  */
3700                 if (iclog->ic_header.h_num_logops)
3701                         return 0;
3702                 iclog = iclog->ic_next;
3703         } while (iclog != log->l_iclog);
3704         return 1;
3705 }