xfs: ensure that sync updates the log tail correctly
[linux-2.6.git] / fs / xfs / xfs_log.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
42 #include "xfs_rw.h"
43 #include "xfs_trace.h"
44
45 kmem_zone_t     *xfs_log_ticket_zone;
46
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
48         { (ptr) += (bytes); \
49           (len) -= (bytes); \
50           (off) += (bytes);}
51
52 /* Local miscellaneous function prototypes */
53 STATIC int       xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
54                                     xlog_in_core_t **, xfs_lsn_t *);
55 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
56                                 xfs_buftarg_t   *log_target,
57                                 xfs_daddr_t     blk_offset,
58                                 int             num_bblks);
59 STATIC int       xlog_space_left(xlog_t *log, int cycle, int bytes);
60 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
61 STATIC void      xlog_dealloc_log(xlog_t *log);
62 STATIC int       xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
63                             int nentries, struct xlog_ticket *tic,
64                             xfs_lsn_t *start_lsn,
65                             xlog_in_core_t **commit_iclog,
66                             uint flags);
67
68 /* local state machine functions */
69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
70 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
71 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
72                                        int              len,
73                                        xlog_in_core_t   **iclog,
74                                        xlog_ticket_t    *ticket,
75                                        int              *continued_write,
76                                        int              *logoffsetp);
77 STATIC int  xlog_state_release_iclog(xlog_t             *log,
78                                      xlog_in_core_t     *iclog);
79 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
80                                      xlog_in_core_t *iclog,
81                                      int                eventual_size);
82 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
83
84 /* local functions to manipulate grant head */
85 STATIC int  xlog_grant_log_space(xlog_t         *log,
86                                  xlog_ticket_t  *xtic);
87 STATIC void xlog_grant_push_ail(xfs_mount_t     *mp,
88                                 int             need_bytes);
89 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
90                                            xlog_ticket_t *ticket);
91 STATIC int xlog_regrant_write_log_space(xlog_t          *log,
92                                          xlog_ticket_t  *ticket);
93 STATIC void xlog_ungrant_log_space(xlog_t        *log,
94                                    xlog_ticket_t *ticket);
95
96
97 /* local ticket functions */
98 STATIC xlog_ticket_t    *xlog_ticket_alloc(xlog_t *log,
99                                          int    unit_bytes,
100                                          int    count,
101                                          char   clientid,
102                                          uint   flags);
103
104 #if defined(DEBUG)
105 STATIC void     xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
106 STATIC void     xlog_verify_grant_head(xlog_t *log, int equals);
107 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
108                                   int count, boolean_t syncing);
109 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
110                                      xfs_lsn_t tail_lsn);
111 #else
112 #define xlog_verify_dest_ptr(a,b)
113 #define xlog_verify_grant_head(a,b)
114 #define xlog_verify_iclog(a,b,c,d)
115 #define xlog_verify_tail_lsn(a,b,c)
116 #endif
117
118 STATIC int      xlog_iclogs_empty(xlog_t *log);
119
120
121 static void
122 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
123 {
124         if (*qp) {
125                 tic->t_next         = (*qp);
126                 tic->t_prev         = (*qp)->t_prev;
127                 (*qp)->t_prev->t_next = tic;
128                 (*qp)->t_prev       = tic;
129         } else {
130                 tic->t_prev = tic->t_next = tic;
131                 *qp = tic;
132         }
133
134         tic->t_flags |= XLOG_TIC_IN_Q;
135 }
136
137 static void
138 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
139 {
140         if (tic == tic->t_next) {
141                 *qp = NULL;
142         } else {
143                 *qp = tic->t_next;
144                 tic->t_next->t_prev = tic->t_prev;
145                 tic->t_prev->t_next = tic->t_next;
146         }
147
148         tic->t_next = tic->t_prev = NULL;
149         tic->t_flags &= ~XLOG_TIC_IN_Q;
150 }
151
152 static void
153 xlog_grant_sub_space(struct log *log, int bytes)
154 {
155         log->l_grant_write_bytes -= bytes;
156         if (log->l_grant_write_bytes < 0) {
157                 log->l_grant_write_bytes += log->l_logsize;
158                 log->l_grant_write_cycle--;
159         }
160
161         log->l_grant_reserve_bytes -= bytes;
162         if ((log)->l_grant_reserve_bytes < 0) {
163                 log->l_grant_reserve_bytes += log->l_logsize;
164                 log->l_grant_reserve_cycle--;
165         }
166
167 }
168
169 static void
170 xlog_grant_add_space_write(struct log *log, int bytes)
171 {
172         int tmp = log->l_logsize - log->l_grant_write_bytes;
173         if (tmp > bytes)
174                 log->l_grant_write_bytes += bytes;
175         else {
176                 log->l_grant_write_cycle++;
177                 log->l_grant_write_bytes = bytes - tmp;
178         }
179 }
180
181 static void
182 xlog_grant_add_space_reserve(struct log *log, int bytes)
183 {
184         int tmp = log->l_logsize - log->l_grant_reserve_bytes;
185         if (tmp > bytes)
186                 log->l_grant_reserve_bytes += bytes;
187         else {
188                 log->l_grant_reserve_cycle++;
189                 log->l_grant_reserve_bytes = bytes - tmp;
190         }
191 }
192
193 static inline void
194 xlog_grant_add_space(struct log *log, int bytes)
195 {
196         xlog_grant_add_space_write(log, bytes);
197         xlog_grant_add_space_reserve(log, bytes);
198 }
199
200 static void
201 xlog_tic_reset_res(xlog_ticket_t *tic)
202 {
203         tic->t_res_num = 0;
204         tic->t_res_arr_sum = 0;
205         tic->t_res_num_ophdrs = 0;
206 }
207
208 static void
209 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
210 {
211         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
212                 /* add to overflow and start again */
213                 tic->t_res_o_flow += tic->t_res_arr_sum;
214                 tic->t_res_num = 0;
215                 tic->t_res_arr_sum = 0;
216         }
217
218         tic->t_res_arr[tic->t_res_num].r_len = len;
219         tic->t_res_arr[tic->t_res_num].r_type = type;
220         tic->t_res_arr_sum += len;
221         tic->t_res_num++;
222 }
223
224 /*
225  * NOTES:
226  *
227  *      1. currblock field gets updated at startup and after in-core logs
228  *              marked as with WANT_SYNC.
229  */
230
231 /*
232  * This routine is called when a user of a log manager ticket is done with
233  * the reservation.  If the ticket was ever used, then a commit record for
234  * the associated transaction is written out as a log operation header with
235  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
236  * a given ticket.  If the ticket was one with a permanent reservation, then
237  * a few operations are done differently.  Permanent reservation tickets by
238  * default don't release the reservation.  They just commit the current
239  * transaction with the belief that the reservation is still needed.  A flag
240  * must be passed in before permanent reservations are actually released.
241  * When these type of tickets are not released, they need to be set into
242  * the inited state again.  By doing this, a start record will be written
243  * out when the next write occurs.
244  */
245 xfs_lsn_t
246 xfs_log_done(
247         struct xfs_mount        *mp,
248         struct xlog_ticket      *ticket,
249         struct xlog_in_core     **iclog,
250         uint                    flags)
251 {
252         struct log              *log = mp->m_log;
253         xfs_lsn_t               lsn = 0;
254
255         if (XLOG_FORCED_SHUTDOWN(log) ||
256             /*
257              * If nothing was ever written, don't write out commit record.
258              * If we get an error, just continue and give back the log ticket.
259              */
260             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
261              (xlog_commit_record(mp, ticket, iclog, &lsn)))) {
262                 lsn = (xfs_lsn_t) -1;
263                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
264                         flags |= XFS_LOG_REL_PERM_RESERV;
265                 }
266         }
267
268
269         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
270             (flags & XFS_LOG_REL_PERM_RESERV)) {
271                 trace_xfs_log_done_nonperm(log, ticket);
272
273                 /*
274                  * Release ticket if not permanent reservation or a specific
275                  * request has been made to release a permanent reservation.
276                  */
277                 xlog_ungrant_log_space(log, ticket);
278                 xfs_log_ticket_put(ticket);
279         } else {
280                 trace_xfs_log_done_perm(log, ticket);
281
282                 xlog_regrant_reserve_log_space(log, ticket);
283                 /* If this ticket was a permanent reservation and we aren't
284                  * trying to release it, reset the inited flags; so next time
285                  * we write, a start record will be written out.
286                  */
287                 ticket->t_flags |= XLOG_TIC_INITED;
288         }
289
290         return lsn;
291 }
292
293 /*
294  * Attaches a new iclog I/O completion callback routine during
295  * transaction commit.  If the log is in error state, a non-zero
296  * return code is handed back and the caller is responsible for
297  * executing the callback at an appropriate time.
298  */
299 int
300 xfs_log_notify(
301         struct xfs_mount        *mp,
302         struct xlog_in_core     *iclog,
303         xfs_log_callback_t      *cb)
304 {
305         int     abortflg;
306
307         spin_lock(&iclog->ic_callback_lock);
308         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
309         if (!abortflg) {
310                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
311                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
312                 cb->cb_next = NULL;
313                 *(iclog->ic_callback_tail) = cb;
314                 iclog->ic_callback_tail = &(cb->cb_next);
315         }
316         spin_unlock(&iclog->ic_callback_lock);
317         return abortflg;
318 }
319
320 int
321 xfs_log_release_iclog(
322         struct xfs_mount        *mp,
323         struct xlog_in_core     *iclog)
324 {
325         if (xlog_state_release_iclog(mp->m_log, iclog)) {
326                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
327                 return EIO;
328         }
329
330         return 0;
331 }
332
333 /*
334  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
335  *      to the reservation.
336  *  2. Potentially, push buffers at tail of log to disk.
337  *
338  * Each reservation is going to reserve extra space for a log record header.
339  * When writes happen to the on-disk log, we don't subtract the length of the
340  * log record header from any reservation.  By wasting space in each
341  * reservation, we prevent over allocation problems.
342  */
343 int
344 xfs_log_reserve(
345         struct xfs_mount        *mp,
346         int                     unit_bytes,
347         int                     cnt,
348         struct xlog_ticket      **ticket,
349         __uint8_t               client,
350         uint                    flags,
351         uint                    t_type)
352 {
353         struct log              *log = mp->m_log;
354         struct xlog_ticket      *internal_ticket;
355         int                     retval = 0;
356
357         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
358         ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
359
360         if (XLOG_FORCED_SHUTDOWN(log))
361                 return XFS_ERROR(EIO);
362
363         XFS_STATS_INC(xs_try_logspace);
364
365
366         if (*ticket != NULL) {
367                 ASSERT(flags & XFS_LOG_PERM_RESERV);
368                 internal_ticket = *ticket;
369
370                 trace_xfs_log_reserve(log, internal_ticket);
371
372                 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
373                 retval = xlog_regrant_write_log_space(log, internal_ticket);
374         } else {
375                 /* may sleep if need to allocate more tickets */
376                 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
377                                                   client, flags);
378                 if (!internal_ticket)
379                         return XFS_ERROR(ENOMEM);
380                 internal_ticket->t_trans_type = t_type;
381                 *ticket = internal_ticket;
382
383                 trace_xfs_log_reserve(log, internal_ticket);
384
385                 xlog_grant_push_ail(mp,
386                                     (internal_ticket->t_unit_res *
387                                      internal_ticket->t_cnt));
388                 retval = xlog_grant_log_space(log, internal_ticket);
389         }
390
391         return retval;
392 }       /* xfs_log_reserve */
393
394
395 /*
396  * Mount a log filesystem
397  *
398  * mp           - ubiquitous xfs mount point structure
399  * log_target   - buftarg of on-disk log device
400  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
401  * num_bblocks  - Number of BBSIZE blocks in on-disk log
402  *
403  * Return error or zero.
404  */
405 int
406 xfs_log_mount(
407         xfs_mount_t     *mp,
408         xfs_buftarg_t   *log_target,
409         xfs_daddr_t     blk_offset,
410         int             num_bblks)
411 {
412         int             error;
413
414         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
415                 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
416         else {
417                 cmn_err(CE_NOTE,
418                         "!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
419                         mp->m_fsname);
420                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
421         }
422
423         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
424         if (IS_ERR(mp->m_log)) {
425                 error = -PTR_ERR(mp->m_log);
426                 goto out;
427         }
428
429         /*
430          * Initialize the AIL now we have a log.
431          */
432         error = xfs_trans_ail_init(mp);
433         if (error) {
434                 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
435                 goto out_free_log;
436         }
437         mp->m_log->l_ailp = mp->m_ail;
438
439         /*
440          * skip log recovery on a norecovery mount.  pretend it all
441          * just worked.
442          */
443         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
444                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
445
446                 if (readonly)
447                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
448
449                 error = xlog_recover(mp->m_log);
450
451                 if (readonly)
452                         mp->m_flags |= XFS_MOUNT_RDONLY;
453                 if (error) {
454                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
455                         goto out_destroy_ail;
456                 }
457         }
458
459         /* Normal transactions can now occur */
460         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
461
462         return 0;
463
464 out_destroy_ail:
465         xfs_trans_ail_destroy(mp);
466 out_free_log:
467         xlog_dealloc_log(mp->m_log);
468 out:
469         return error;
470 }
471
472 /*
473  * Finish the recovery of the file system.  This is separate from
474  * the xfs_log_mount() call, because it depends on the code in
475  * xfs_mountfs() to read in the root and real-time bitmap inodes
476  * between calling xfs_log_mount() and here.
477  *
478  * mp           - ubiquitous xfs mount point structure
479  */
480 int
481 xfs_log_mount_finish(xfs_mount_t *mp)
482 {
483         int     error;
484
485         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
486                 error = xlog_recover_finish(mp->m_log);
487         else {
488                 error = 0;
489                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
490         }
491
492         return error;
493 }
494
495 /*
496  * Final log writes as part of unmount.
497  *
498  * Mark the filesystem clean as unmount happens.  Note that during relocation
499  * this routine needs to be executed as part of source-bag while the
500  * deallocation must not be done until source-end.
501  */
502
503 /*
504  * Unmount record used to have a string "Unmount filesystem--" in the
505  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
506  * We just write the magic number now since that particular field isn't
507  * currently architecture converted and "nUmount" is a bit foo.
508  * As far as I know, there weren't any dependencies on the old behaviour.
509  */
510
511 int
512 xfs_log_unmount_write(xfs_mount_t *mp)
513 {
514         xlog_t           *log = mp->m_log;
515         xlog_in_core_t   *iclog;
516 #ifdef DEBUG
517         xlog_in_core_t   *first_iclog;
518 #endif
519         xfs_log_iovec_t  reg[1];
520         xlog_ticket_t   *tic = NULL;
521         xfs_lsn_t        lsn;
522         int              error;
523
524         /* the data section must be 32 bit size aligned */
525         struct {
526             __uint16_t magic;
527             __uint16_t pad1;
528             __uint32_t pad2; /* may as well make it 64 bits */
529         } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
530
531         /*
532          * Don't write out unmount record on read-only mounts.
533          * Or, if we are doing a forced umount (typically because of IO errors).
534          */
535         if (mp->m_flags & XFS_MOUNT_RDONLY)
536                 return 0;
537
538         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
539         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
540
541 #ifdef DEBUG
542         first_iclog = iclog = log->l_iclog;
543         do {
544                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
545                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
546                         ASSERT(iclog->ic_offset == 0);
547                 }
548                 iclog = iclog->ic_next;
549         } while (iclog != first_iclog);
550 #endif
551         if (! (XLOG_FORCED_SHUTDOWN(log))) {
552                 reg[0].i_addr = (void*)&magic;
553                 reg[0].i_len  = sizeof(magic);
554                 reg[0].i_type = XLOG_REG_TYPE_UNMOUNT;
555
556                 error = xfs_log_reserve(mp, 600, 1, &tic,
557                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
558                 if (!error) {
559                         /* remove inited flag */
560                         ((xlog_ticket_t *)tic)->t_flags = 0;
561                         error = xlog_write(mp, reg, 1, tic, &lsn,
562                                            NULL, XLOG_UNMOUNT_TRANS);
563                         /*
564                          * At this point, we're umounting anyway,
565                          * so there's no point in transitioning log state
566                          * to IOERROR. Just continue...
567                          */
568                 }
569
570                 if (error) {
571                         xfs_fs_cmn_err(CE_ALERT, mp,
572                                 "xfs_log_unmount: unmount record failed");
573                 }
574
575
576                 spin_lock(&log->l_icloglock);
577                 iclog = log->l_iclog;
578                 atomic_inc(&iclog->ic_refcnt);
579                 xlog_state_want_sync(log, iclog);
580                 spin_unlock(&log->l_icloglock);
581                 error = xlog_state_release_iclog(log, iclog);
582
583                 spin_lock(&log->l_icloglock);
584                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
585                       iclog->ic_state == XLOG_STATE_DIRTY)) {
586                         if (!XLOG_FORCED_SHUTDOWN(log)) {
587                                 sv_wait(&iclog->ic_force_wait, PMEM,
588                                         &log->l_icloglock, s);
589                         } else {
590                                 spin_unlock(&log->l_icloglock);
591                         }
592                 } else {
593                         spin_unlock(&log->l_icloglock);
594                 }
595                 if (tic) {
596                         trace_xfs_log_umount_write(log, tic);
597                         xlog_ungrant_log_space(log, tic);
598                         xfs_log_ticket_put(tic);
599                 }
600         } else {
601                 /*
602                  * We're already in forced_shutdown mode, couldn't
603                  * even attempt to write out the unmount transaction.
604                  *
605                  * Go through the motions of sync'ing and releasing
606                  * the iclog, even though no I/O will actually happen,
607                  * we need to wait for other log I/Os that may already
608                  * be in progress.  Do this as a separate section of
609                  * code so we'll know if we ever get stuck here that
610                  * we're in this odd situation of trying to unmount
611                  * a file system that went into forced_shutdown as
612                  * the result of an unmount..
613                  */
614                 spin_lock(&log->l_icloglock);
615                 iclog = log->l_iclog;
616                 atomic_inc(&iclog->ic_refcnt);
617
618                 xlog_state_want_sync(log, iclog);
619                 spin_unlock(&log->l_icloglock);
620                 error =  xlog_state_release_iclog(log, iclog);
621
622                 spin_lock(&log->l_icloglock);
623
624                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
625                         || iclog->ic_state == XLOG_STATE_DIRTY
626                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
627
628                                 sv_wait(&iclog->ic_force_wait, PMEM,
629                                         &log->l_icloglock, s);
630                 } else {
631                         spin_unlock(&log->l_icloglock);
632                 }
633         }
634
635         return error;
636 }       /* xfs_log_unmount_write */
637
638 /*
639  * Deallocate log structures for unmount/relocation.
640  *
641  * We need to stop the aild from running before we destroy
642  * and deallocate the log as the aild references the log.
643  */
644 void
645 xfs_log_unmount(xfs_mount_t *mp)
646 {
647         xfs_trans_ail_destroy(mp);
648         xlog_dealloc_log(mp->m_log);
649 }
650
651 /*
652  * Write region vectors to log.  The write happens using the space reservation
653  * of the ticket (tic).  It is not a requirement that all writes for a given
654  * transaction occur with one call to xfs_log_write().
655  */
656 int
657 xfs_log_write(
658         struct xfs_mount        *mp,
659         struct xfs_log_iovec    reg[],
660         int                     nentries,
661         struct xlog_ticket      *tic,
662         xfs_lsn_t               *start_lsn)
663 {
664         struct log              *log = mp->m_log;
665         int                     error;
666
667         if (XLOG_FORCED_SHUTDOWN(log))
668                 return XFS_ERROR(EIO);
669
670         error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0);
671         if (error)
672                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
673         return error;
674 }
675
676 void
677 xfs_log_move_tail(xfs_mount_t   *mp,
678                   xfs_lsn_t     tail_lsn)
679 {
680         xlog_ticket_t   *tic;
681         xlog_t          *log = mp->m_log;
682         int             need_bytes, free_bytes, cycle, bytes;
683
684         if (XLOG_FORCED_SHUTDOWN(log))
685                 return;
686
687         if (tail_lsn == 0) {
688                 /* needed since sync_lsn is 64 bits */
689                 spin_lock(&log->l_icloglock);
690                 tail_lsn = log->l_last_sync_lsn;
691                 spin_unlock(&log->l_icloglock);
692         }
693
694         spin_lock(&log->l_grant_lock);
695
696         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
697          * tail_lsn.
698          */
699         if (tail_lsn != 1) {
700                 log->l_tail_lsn = tail_lsn;
701         }
702
703         if ((tic = log->l_write_headq)) {
704 #ifdef DEBUG
705                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
706                         panic("Recovery problem");
707 #endif
708                 cycle = log->l_grant_write_cycle;
709                 bytes = log->l_grant_write_bytes;
710                 free_bytes = xlog_space_left(log, cycle, bytes);
711                 do {
712                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
713
714                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
715                                 break;
716                         tail_lsn = 0;
717                         free_bytes -= tic->t_unit_res;
718                         sv_signal(&tic->t_wait);
719                         tic = tic->t_next;
720                 } while (tic != log->l_write_headq);
721         }
722         if ((tic = log->l_reserve_headq)) {
723 #ifdef DEBUG
724                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
725                         panic("Recovery problem");
726 #endif
727                 cycle = log->l_grant_reserve_cycle;
728                 bytes = log->l_grant_reserve_bytes;
729                 free_bytes = xlog_space_left(log, cycle, bytes);
730                 do {
731                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
732                                 need_bytes = tic->t_unit_res*tic->t_cnt;
733                         else
734                                 need_bytes = tic->t_unit_res;
735                         if (free_bytes < need_bytes && tail_lsn != 1)
736                                 break;
737                         tail_lsn = 0;
738                         free_bytes -= need_bytes;
739                         sv_signal(&tic->t_wait);
740                         tic = tic->t_next;
741                 } while (tic != log->l_reserve_headq);
742         }
743         spin_unlock(&log->l_grant_lock);
744 }       /* xfs_log_move_tail */
745
746 /*
747  * Determine if we have a transaction that has gone to disk
748  * that needs to be covered. To begin the transition to the idle state
749  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
750  * If we are then in a state where covering is needed, the caller is informed
751  * that dummy transactions are required to move the log into the idle state.
752  *
753  * Because this is called as part of the sync process, we should also indicate
754  * that dummy transactions should be issued in anything but the covered or
755  * idle states. This ensures that the log tail is accurately reflected in
756  * the log at the end of the sync, hence if a crash occurrs avoids replay
757  * of transactions where the metadata is already on disk.
758  */
759 int
760 xfs_log_need_covered(xfs_mount_t *mp)
761 {
762         int             needed = 0;
763         xlog_t          *log = mp->m_log;
764
765         if (!xfs_fs_writable(mp))
766                 return 0;
767
768         spin_lock(&log->l_icloglock);
769         switch (log->l_covered_state) {
770         case XLOG_STATE_COVER_DONE:
771         case XLOG_STATE_COVER_DONE2:
772         case XLOG_STATE_COVER_IDLE:
773                 break;
774         case XLOG_STATE_COVER_NEED:
775         case XLOG_STATE_COVER_NEED2:
776                 if (!xfs_trans_ail_tail(log->l_ailp) &&
777                     xlog_iclogs_empty(log)) {
778                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
779                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
780                         else
781                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
782                 }
783                 /* FALLTHRU */
784         default:
785                 needed = 1;
786                 break;
787         }
788         spin_unlock(&log->l_icloglock);
789         return needed;
790 }
791
792 /******************************************************************************
793  *
794  *      local routines
795  *
796  ******************************************************************************
797  */
798
799 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
800  * The log manager must keep track of the last LR which was committed
801  * to disk.  The lsn of this LR will become the new tail_lsn whenever
802  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
803  * the situation where stuff could be written into the log but nothing
804  * was ever in the AIL when asked.  Eventually, we panic since the
805  * tail hits the head.
806  *
807  * We may be holding the log iclog lock upon entering this routine.
808  */
809 xfs_lsn_t
810 xlog_assign_tail_lsn(xfs_mount_t *mp)
811 {
812         xfs_lsn_t tail_lsn;
813         xlog_t    *log = mp->m_log;
814
815         tail_lsn = xfs_trans_ail_tail(mp->m_ail);
816         spin_lock(&log->l_grant_lock);
817         if (tail_lsn != 0) {
818                 log->l_tail_lsn = tail_lsn;
819         } else {
820                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
821         }
822         spin_unlock(&log->l_grant_lock);
823
824         return tail_lsn;
825 }       /* xlog_assign_tail_lsn */
826
827
828 /*
829  * Return the space in the log between the tail and the head.  The head
830  * is passed in the cycle/bytes formal parms.  In the special case where
831  * the reserve head has wrapped passed the tail, this calculation is no
832  * longer valid.  In this case, just return 0 which means there is no space
833  * in the log.  This works for all places where this function is called
834  * with the reserve head.  Of course, if the write head were to ever
835  * wrap the tail, we should blow up.  Rather than catch this case here,
836  * we depend on other ASSERTions in other parts of the code.   XXXmiken
837  *
838  * This code also handles the case where the reservation head is behind
839  * the tail.  The details of this case are described below, but the end
840  * result is that we return the size of the log as the amount of space left.
841  */
842 STATIC int
843 xlog_space_left(xlog_t *log, int cycle, int bytes)
844 {
845         int free_bytes;
846         int tail_bytes;
847         int tail_cycle;
848
849         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
850         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
851         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
852                 free_bytes = log->l_logsize - (bytes - tail_bytes);
853         } else if ((tail_cycle + 1) < cycle) {
854                 return 0;
855         } else if (tail_cycle < cycle) {
856                 ASSERT(tail_cycle == (cycle - 1));
857                 free_bytes = tail_bytes - bytes;
858         } else {
859                 /*
860                  * The reservation head is behind the tail.
861                  * In this case we just want to return the size of the
862                  * log as the amount of space left.
863                  */
864                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
865                         "xlog_space_left: head behind tail\n"
866                         "  tail_cycle = %d, tail_bytes = %d\n"
867                         "  GH   cycle = %d, GH   bytes = %d",
868                         tail_cycle, tail_bytes, cycle, bytes);
869                 ASSERT(0);
870                 free_bytes = log->l_logsize;
871         }
872         return free_bytes;
873 }       /* xlog_space_left */
874
875
876 /*
877  * Log function which is called when an io completes.
878  *
879  * The log manager needs its own routine, in order to control what
880  * happens with the buffer after the write completes.
881  */
882 void
883 xlog_iodone(xfs_buf_t *bp)
884 {
885         xlog_in_core_t  *iclog;
886         xlog_t          *l;
887         int             aborted;
888
889         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
890         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
891         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
892         aborted = 0;
893         l = iclog->ic_log;
894
895         /*
896          * If the _XFS_BARRIER_FAILED flag was set by a lower
897          * layer, it means the underlying device no longer supports
898          * barrier I/O. Warn loudly and turn off barriers.
899          */
900         if (bp->b_flags & _XFS_BARRIER_FAILED) {
901                 bp->b_flags &= ~_XFS_BARRIER_FAILED;
902                 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
903                 xfs_fs_cmn_err(CE_WARN, l->l_mp,
904                                 "xlog_iodone: Barriers are no longer supported"
905                                 " by device. Disabling barriers\n");
906         }
907
908         /*
909          * Race to shutdown the filesystem if we see an error.
910          */
911         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
912                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
913                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
914                 XFS_BUF_STALE(bp);
915                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
916                 /*
917                  * This flag will be propagated to the trans-committed
918                  * callback routines to let them know that the log-commit
919                  * didn't succeed.
920                  */
921                 aborted = XFS_LI_ABORTED;
922         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
923                 aborted = XFS_LI_ABORTED;
924         }
925
926         /* log I/O is always issued ASYNC */
927         ASSERT(XFS_BUF_ISASYNC(bp));
928         xlog_state_done_syncing(iclog, aborted);
929         /*
930          * do not reference the buffer (bp) here as we could race
931          * with it being freed after writing the unmount record to the
932          * log.
933          */
934
935 }       /* xlog_iodone */
936
937 /*
938  * Return size of each in-core log record buffer.
939  *
940  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
941  *
942  * If the filesystem blocksize is too large, we may need to choose a
943  * larger size since the directory code currently logs entire blocks.
944  */
945
946 STATIC void
947 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
948                            xlog_t       *log)
949 {
950         int size;
951         int xhdrs;
952
953         if (mp->m_logbufs <= 0)
954                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
955         else
956                 log->l_iclog_bufs = mp->m_logbufs;
957
958         /*
959          * Buffer size passed in from mount system call.
960          */
961         if (mp->m_logbsize > 0) {
962                 size = log->l_iclog_size = mp->m_logbsize;
963                 log->l_iclog_size_log = 0;
964                 while (size != 1) {
965                         log->l_iclog_size_log++;
966                         size >>= 1;
967                 }
968
969                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
970                         /* # headers = size / 32k
971                          * one header holds cycles from 32k of data
972                          */
973
974                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
975                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
976                                 xhdrs++;
977                         log->l_iclog_hsize = xhdrs << BBSHIFT;
978                         log->l_iclog_heads = xhdrs;
979                 } else {
980                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
981                         log->l_iclog_hsize = BBSIZE;
982                         log->l_iclog_heads = 1;
983                 }
984                 goto done;
985         }
986
987         /* All machines use 32kB buffers by default. */
988         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
989         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
990
991         /* the default log size is 16k or 32k which is one header sector */
992         log->l_iclog_hsize = BBSIZE;
993         log->l_iclog_heads = 1;
994
995 done:
996         /* are we being asked to make the sizes selected above visible? */
997         if (mp->m_logbufs == 0)
998                 mp->m_logbufs = log->l_iclog_bufs;
999         if (mp->m_logbsize == 0)
1000                 mp->m_logbsize = log->l_iclog_size;
1001 }       /* xlog_get_iclog_buffer_size */
1002
1003
1004 /*
1005  * This routine initializes some of the log structure for a given mount point.
1006  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1007  * some other stuff may be filled in too.
1008  */
1009 STATIC xlog_t *
1010 xlog_alloc_log(xfs_mount_t      *mp,
1011                xfs_buftarg_t    *log_target,
1012                xfs_daddr_t      blk_offset,
1013                int              num_bblks)
1014 {
1015         xlog_t                  *log;
1016         xlog_rec_header_t       *head;
1017         xlog_in_core_t          **iclogp;
1018         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1019         xfs_buf_t               *bp;
1020         int                     i;
1021         int                     iclogsize;
1022         int                     error = ENOMEM;
1023
1024         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1025         if (!log) {
1026                 xlog_warn("XFS: Log allocation failed: No memory!");
1027                 goto out;
1028         }
1029
1030         log->l_mp          = mp;
1031         log->l_targ        = log_target;
1032         log->l_logsize     = BBTOB(num_bblks);
1033         log->l_logBBstart  = blk_offset;
1034         log->l_logBBsize   = num_bblks;
1035         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1036         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1037
1038         log->l_prev_block  = -1;
1039         log->l_tail_lsn    = xlog_assign_lsn(1, 0);
1040         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1041         log->l_last_sync_lsn = log->l_tail_lsn;
1042         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1043         log->l_grant_reserve_cycle = 1;
1044         log->l_grant_write_cycle = 1;
1045
1046         error = EFSCORRUPTED;
1047         if (xfs_sb_version_hassector(&mp->m_sb)) {
1048                 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1049                 if (log->l_sectbb_log < 0 ||
1050                     log->l_sectbb_log > mp->m_sectbb_log) {
1051                         xlog_warn("XFS: Log sector size (0x%x) out of range.",
1052                                                 log->l_sectbb_log);
1053                         goto out_free_log;
1054                 }
1055
1056                 /* for larger sector sizes, must have v2 or external log */
1057                 if (log->l_sectbb_log != 0 &&
1058                     (log->l_logBBstart != 0 &&
1059                      !xfs_sb_version_haslogv2(&mp->m_sb))) {
1060                         xlog_warn("XFS: log sector size (0x%x) invalid "
1061                                   "for configuration.", log->l_sectbb_log);
1062                         goto out_free_log;
1063                 }
1064                 if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1065                         xlog_warn("XFS: Log sector log (0x%x) too small.",
1066                                                 mp->m_sb.sb_logsectlog);
1067                         goto out_free_log;
1068                 }
1069         }
1070         log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1071
1072         xlog_get_iclog_buffer_size(mp, log);
1073
1074         error = ENOMEM;
1075         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1076         if (!bp)
1077                 goto out_free_log;
1078         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1079         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1080         ASSERT(XFS_BUF_ISBUSY(bp));
1081         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1082         log->l_xbuf = bp;
1083
1084         spin_lock_init(&log->l_icloglock);
1085         spin_lock_init(&log->l_grant_lock);
1086         sv_init(&log->l_flush_wait, 0, "flush_wait");
1087
1088         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1089         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1090
1091         iclogp = &log->l_iclog;
1092         /*
1093          * The amount of memory to allocate for the iclog structure is
1094          * rather funky due to the way the structure is defined.  It is
1095          * done this way so that we can use different sizes for machines
1096          * with different amounts of memory.  See the definition of
1097          * xlog_in_core_t in xfs_log_priv.h for details.
1098          */
1099         iclogsize = log->l_iclog_size;
1100         ASSERT(log->l_iclog_size >= 4096);
1101         for (i=0; i < log->l_iclog_bufs; i++) {
1102                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1103                 if (!*iclogp)
1104                         goto out_free_iclog;
1105
1106                 iclog = *iclogp;
1107                 iclog->ic_prev = prev_iclog;
1108                 prev_iclog = iclog;
1109
1110                 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1111                 if (!bp)
1112                         goto out_free_iclog;
1113                 if (!XFS_BUF_CPSEMA(bp))
1114                         ASSERT(0);
1115                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1116                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1117                 iclog->ic_bp = bp;
1118                 iclog->ic_data = bp->b_addr;
1119 #ifdef DEBUG
1120                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1121 #endif
1122                 head = &iclog->ic_header;
1123                 memset(head, 0, sizeof(xlog_rec_header_t));
1124                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1125                 head->h_version = cpu_to_be32(
1126                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1127                 head->h_size = cpu_to_be32(log->l_iclog_size);
1128                 /* new fields */
1129                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1130                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1131
1132                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1133                 iclog->ic_state = XLOG_STATE_ACTIVE;
1134                 iclog->ic_log = log;
1135                 atomic_set(&iclog->ic_refcnt, 0);
1136                 spin_lock_init(&iclog->ic_callback_lock);
1137                 iclog->ic_callback_tail = &(iclog->ic_callback);
1138                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1139
1140                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1141                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1142                 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1143                 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1144
1145                 iclogp = &iclog->ic_next;
1146         }
1147         *iclogp = log->l_iclog;                 /* complete ring */
1148         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1149
1150         return log;
1151
1152 out_free_iclog:
1153         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1154                 prev_iclog = iclog->ic_next;
1155                 if (iclog->ic_bp) {
1156                         sv_destroy(&iclog->ic_force_wait);
1157                         sv_destroy(&iclog->ic_write_wait);
1158                         xfs_buf_free(iclog->ic_bp);
1159                 }
1160                 kmem_free(iclog);
1161         }
1162         spinlock_destroy(&log->l_icloglock);
1163         spinlock_destroy(&log->l_grant_lock);
1164         xfs_buf_free(log->l_xbuf);
1165 out_free_log:
1166         kmem_free(log);
1167 out:
1168         return ERR_PTR(-error);
1169 }       /* xlog_alloc_log */
1170
1171
1172 /*
1173  * Write out the commit record of a transaction associated with the given
1174  * ticket.  Return the lsn of the commit record.
1175  */
1176 STATIC int
1177 xlog_commit_record(xfs_mount_t  *mp,
1178                    xlog_ticket_t *ticket,
1179                    xlog_in_core_t **iclog,
1180                    xfs_lsn_t    *commitlsnp)
1181 {
1182         int             error;
1183         xfs_log_iovec_t reg[1];
1184
1185         reg[0].i_addr = NULL;
1186         reg[0].i_len = 0;
1187         reg[0].i_type = XLOG_REG_TYPE_COMMIT;
1188
1189         ASSERT_ALWAYS(iclog);
1190         if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1191                                iclog, XLOG_COMMIT_TRANS))) {
1192                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1193         }
1194         return error;
1195 }       /* xlog_commit_record */
1196
1197
1198 /*
1199  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1200  * log space.  This code pushes on the lsn which would supposedly free up
1201  * the 25% which we want to leave free.  We may need to adopt a policy which
1202  * pushes on an lsn which is further along in the log once we reach the high
1203  * water mark.  In this manner, we would be creating a low water mark.
1204  */
1205 STATIC void
1206 xlog_grant_push_ail(xfs_mount_t *mp,
1207                     int         need_bytes)
1208 {
1209     xlog_t      *log = mp->m_log;       /* pointer to the log */
1210     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1211     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1212     int         free_blocks;            /* free blocks left to write to */
1213     int         free_bytes;             /* free bytes left to write to */
1214     int         threshold_block;        /* block in lsn we'd like to be at */
1215     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1216     int         free_threshold;
1217
1218     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1219
1220     spin_lock(&log->l_grant_lock);
1221     free_bytes = xlog_space_left(log,
1222                                  log->l_grant_reserve_cycle,
1223                                  log->l_grant_reserve_bytes);
1224     tail_lsn = log->l_tail_lsn;
1225     free_blocks = BTOBBT(free_bytes);
1226
1227     /*
1228      * Set the threshold for the minimum number of free blocks in the
1229      * log to the maximum of what the caller needs, one quarter of the
1230      * log, and 256 blocks.
1231      */
1232     free_threshold = BTOBB(need_bytes);
1233     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1234     free_threshold = MAX(free_threshold, 256);
1235     if (free_blocks < free_threshold) {
1236         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1237         threshold_cycle = CYCLE_LSN(tail_lsn);
1238         if (threshold_block >= log->l_logBBsize) {
1239             threshold_block -= log->l_logBBsize;
1240             threshold_cycle += 1;
1241         }
1242         threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1243
1244         /* Don't pass in an lsn greater than the lsn of the last
1245          * log record known to be on disk.
1246          */
1247         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1248             threshold_lsn = log->l_last_sync_lsn;
1249     }
1250     spin_unlock(&log->l_grant_lock);
1251
1252     /*
1253      * Get the transaction layer to kick the dirty buffers out to
1254      * disk asynchronously. No point in trying to do this if
1255      * the filesystem is shutting down.
1256      */
1257     if (threshold_lsn &&
1258         !XLOG_FORCED_SHUTDOWN(log))
1259             xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1260 }       /* xlog_grant_push_ail */
1261
1262 /*
1263  * The bdstrat callback function for log bufs. This gives us a central
1264  * place to trap bufs in case we get hit by a log I/O error and need to
1265  * shutdown. Actually, in practice, even when we didn't get a log error,
1266  * we transition the iclogs to IOERROR state *after* flushing all existing
1267  * iclogs to disk. This is because we don't want anymore new transactions to be
1268  * started or completed afterwards.
1269  */
1270 STATIC int
1271 xlog_bdstrat(
1272         struct xfs_buf          *bp)
1273 {
1274         struct xlog_in_core     *iclog;
1275
1276         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1277         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1278                 XFS_BUF_ERROR(bp, EIO);
1279                 XFS_BUF_STALE(bp);
1280                 xfs_biodone(bp);
1281                 /*
1282                  * It would seem logical to return EIO here, but we rely on
1283                  * the log state machine to propagate I/O errors instead of
1284                  * doing it here.
1285                  */
1286                 return 0;
1287         }
1288
1289         bp->b_flags |= _XBF_RUN_QUEUES;
1290         xfs_buf_iorequest(bp);
1291         return 0;
1292 }
1293
1294 /*
1295  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1296  * fashion.  Previously, we should have moved the current iclog
1297  * ptr in the log to point to the next available iclog.  This allows further
1298  * write to continue while this code syncs out an iclog ready to go.
1299  * Before an in-core log can be written out, the data section must be scanned
1300  * to save away the 1st word of each BBSIZE block into the header.  We replace
1301  * it with the current cycle count.  Each BBSIZE block is tagged with the
1302  * cycle count because there in an implicit assumption that drives will
1303  * guarantee that entire 512 byte blocks get written at once.  In other words,
1304  * we can't have part of a 512 byte block written and part not written.  By
1305  * tagging each block, we will know which blocks are valid when recovering
1306  * after an unclean shutdown.
1307  *
1308  * This routine is single threaded on the iclog.  No other thread can be in
1309  * this routine with the same iclog.  Changing contents of iclog can there-
1310  * fore be done without grabbing the state machine lock.  Updating the global
1311  * log will require grabbing the lock though.
1312  *
1313  * The entire log manager uses a logical block numbering scheme.  Only
1314  * log_sync (and then only bwrite()) know about the fact that the log may
1315  * not start with block zero on a given device.  The log block start offset
1316  * is added immediately before calling bwrite().
1317  */
1318
1319 STATIC int
1320 xlog_sync(xlog_t                *log,
1321           xlog_in_core_t        *iclog)
1322 {
1323         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1324         xfs_buf_t       *bp;
1325         int             i;
1326         uint            count;          /* byte count of bwrite */
1327         uint            count_init;     /* initial count before roundup */
1328         int             roundoff;       /* roundoff to BB or stripe */
1329         int             split = 0;      /* split write into two regions */
1330         int             error;
1331         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1332
1333         XFS_STATS_INC(xs_log_writes);
1334         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1335
1336         /* Add for LR header */
1337         count_init = log->l_iclog_hsize + iclog->ic_offset;
1338
1339         /* Round out the log write size */
1340         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1341                 /* we have a v2 stripe unit to use */
1342                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1343         } else {
1344                 count = BBTOB(BTOBB(count_init));
1345         }
1346         roundoff = count - count_init;
1347         ASSERT(roundoff >= 0);
1348         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1349                 roundoff < log->l_mp->m_sb.sb_logsunit)
1350                 || 
1351                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1352                  roundoff < BBTOB(1)));
1353
1354         /* move grant heads by roundoff in sync */
1355         spin_lock(&log->l_grant_lock);
1356         xlog_grant_add_space(log, roundoff);
1357         spin_unlock(&log->l_grant_lock);
1358
1359         /* put cycle number in every block */
1360         xlog_pack_data(log, iclog, roundoff); 
1361
1362         /* real byte length */
1363         if (v2) {
1364                 iclog->ic_header.h_len =
1365                         cpu_to_be32(iclog->ic_offset + roundoff);
1366         } else {
1367                 iclog->ic_header.h_len =
1368                         cpu_to_be32(iclog->ic_offset);
1369         }
1370
1371         bp = iclog->ic_bp;
1372         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1373         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1374         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1375
1376         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1377
1378         /* Do we need to split this write into 2 parts? */
1379         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1380                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1381                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1382                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1383         } else {
1384                 iclog->ic_bwritecnt = 1;
1385         }
1386         XFS_BUF_SET_COUNT(bp, count);
1387         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1388         XFS_BUF_ZEROFLAGS(bp);
1389         XFS_BUF_BUSY(bp);
1390         XFS_BUF_ASYNC(bp);
1391         bp->b_flags |= XBF_LOG_BUFFER;
1392         /*
1393          * Do an ordered write for the log block.
1394          * Its unnecessary to flush the first split block in the log wrap case.
1395          */
1396         if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1397                 XFS_BUF_ORDERED(bp);
1398
1399         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1400         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1401
1402         xlog_verify_iclog(log, iclog, count, B_TRUE);
1403
1404         /* account for log which doesn't start at block #0 */
1405         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1406         /*
1407          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1408          * is shutting down.
1409          */
1410         XFS_BUF_WRITE(bp);
1411
1412         if ((error = xlog_bdstrat(bp))) {
1413                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1414                                   XFS_BUF_ADDR(bp));
1415                 return error;
1416         }
1417         if (split) {
1418                 bp = iclog->ic_log->l_xbuf;
1419                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1420                                                         (unsigned long)1);
1421                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1422                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1423                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1424                                             (__psint_t)count), split);
1425                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1426                 XFS_BUF_ZEROFLAGS(bp);
1427                 XFS_BUF_BUSY(bp);
1428                 XFS_BUF_ASYNC(bp);
1429                 bp->b_flags |= XBF_LOG_BUFFER;
1430                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1431                         XFS_BUF_ORDERED(bp);
1432                 dptr = XFS_BUF_PTR(bp);
1433                 /*
1434                  * Bump the cycle numbers at the start of each block
1435                  * since this part of the buffer is at the start of
1436                  * a new cycle.  Watch out for the header magic number
1437                  * case, though.
1438                  */
1439                 for (i = 0; i < split; i += BBSIZE) {
1440                         be32_add_cpu((__be32 *)dptr, 1);
1441                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1442                                 be32_add_cpu((__be32 *)dptr, 1);
1443                         dptr += BBSIZE;
1444                 }
1445
1446                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1447                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1448
1449                 /* account for internal log which doesn't start at block #0 */
1450                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1451                 XFS_BUF_WRITE(bp);
1452                 if ((error = xlog_bdstrat(bp))) {
1453                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1454                                           bp, XFS_BUF_ADDR(bp));
1455                         return error;
1456                 }
1457         }
1458         return 0;
1459 }       /* xlog_sync */
1460
1461
1462 /*
1463  * Deallocate a log structure
1464  */
1465 STATIC void
1466 xlog_dealloc_log(xlog_t *log)
1467 {
1468         xlog_in_core_t  *iclog, *next_iclog;
1469         int             i;
1470
1471         iclog = log->l_iclog;
1472         for (i=0; i<log->l_iclog_bufs; i++) {
1473                 sv_destroy(&iclog->ic_force_wait);
1474                 sv_destroy(&iclog->ic_write_wait);
1475                 xfs_buf_free(iclog->ic_bp);
1476                 next_iclog = iclog->ic_next;
1477                 kmem_free(iclog);
1478                 iclog = next_iclog;
1479         }
1480         spinlock_destroy(&log->l_icloglock);
1481         spinlock_destroy(&log->l_grant_lock);
1482
1483         xfs_buf_free(log->l_xbuf);
1484         log->l_mp->m_log = NULL;
1485         kmem_free(log);
1486 }       /* xlog_dealloc_log */
1487
1488 /*
1489  * Update counters atomically now that memcpy is done.
1490  */
1491 /* ARGSUSED */
1492 static inline void
1493 xlog_state_finish_copy(xlog_t           *log,
1494                        xlog_in_core_t   *iclog,
1495                        int              record_cnt,
1496                        int              copy_bytes)
1497 {
1498         spin_lock(&log->l_icloglock);
1499
1500         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1501         iclog->ic_offset += copy_bytes;
1502
1503         spin_unlock(&log->l_icloglock);
1504 }       /* xlog_state_finish_copy */
1505
1506
1507
1508
1509 /*
1510  * print out info relating to regions written which consume
1511  * the reservation
1512  */
1513 STATIC void
1514 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1515 {
1516         uint i;
1517         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1518
1519         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1520         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1521             "bformat",
1522             "bchunk",
1523             "efi_format",
1524             "efd_format",
1525             "iformat",
1526             "icore",
1527             "iext",
1528             "ibroot",
1529             "ilocal",
1530             "iattr_ext",
1531             "iattr_broot",
1532             "iattr_local",
1533             "qformat",
1534             "dquot",
1535             "quotaoff",
1536             "LR header",
1537             "unmount",
1538             "commit",
1539             "trans header"
1540         };
1541         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1542             "SETATTR_NOT_SIZE",
1543             "SETATTR_SIZE",
1544             "INACTIVE",
1545             "CREATE",
1546             "CREATE_TRUNC",
1547             "TRUNCATE_FILE",
1548             "REMOVE",
1549             "LINK",
1550             "RENAME",
1551             "MKDIR",
1552             "RMDIR",
1553             "SYMLINK",
1554             "SET_DMATTRS",
1555             "GROWFS",
1556             "STRAT_WRITE",
1557             "DIOSTRAT",
1558             "WRITE_SYNC",
1559             "WRITEID",
1560             "ADDAFORK",
1561             "ATTRINVAL",
1562             "ATRUNCATE",
1563             "ATTR_SET",
1564             "ATTR_RM",
1565             "ATTR_FLAG",
1566             "CLEAR_AGI_BUCKET",
1567             "QM_SBCHANGE",
1568             "DUMMY1",
1569             "DUMMY2",
1570             "QM_QUOTAOFF",
1571             "QM_DQALLOC",
1572             "QM_SETQLIM",
1573             "QM_DQCLUSTER",
1574             "QM_QINOCREATE",
1575             "QM_QUOTAOFF_END",
1576             "SB_UNIT",
1577             "FSYNC_TS",
1578             "GROWFSRT_ALLOC",
1579             "GROWFSRT_ZERO",
1580             "GROWFSRT_FREE",
1581             "SWAPEXT"
1582         };
1583
1584         xfs_fs_cmn_err(CE_WARN, mp,
1585                         "xfs_log_write: reservation summary:\n"
1586                         "  trans type  = %s (%u)\n"
1587                         "  unit res    = %d bytes\n"
1588                         "  current res = %d bytes\n"
1589                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1590                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1591                         "  ophdr + reg = %u bytes\n"
1592                         "  num regions = %u\n",
1593                         ((ticket->t_trans_type <= 0 ||
1594                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1595                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1596                         ticket->t_trans_type,
1597                         ticket->t_unit_res,
1598                         ticket->t_curr_res,
1599                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1600                         ticket->t_res_num_ophdrs, ophdr_spc,
1601                         ticket->t_res_arr_sum + 
1602                         ticket->t_res_o_flow + ophdr_spc,
1603                         ticket->t_res_num);
1604
1605         for (i = 0; i < ticket->t_res_num; i++) {
1606                 uint r_type = ticket->t_res_arr[i].r_type; 
1607                 cmn_err(CE_WARN,
1608                             "region[%u]: %s - %u bytes\n",
1609                             i, 
1610                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1611                             "bad-rtype" : res_type_str[r_type-1]),
1612                             ticket->t_res_arr[i].r_len);
1613         }
1614 }
1615
1616 /*
1617  * Write some region out to in-core log
1618  *
1619  * This will be called when writing externally provided regions or when
1620  * writing out a commit record for a given transaction.
1621  *
1622  * General algorithm:
1623  *      1. Find total length of this write.  This may include adding to the
1624  *              lengths passed in.
1625  *      2. Check whether we violate the tickets reservation.
1626  *      3. While writing to this iclog
1627  *          A. Reserve as much space in this iclog as can get
1628  *          B. If this is first write, save away start lsn
1629  *          C. While writing this region:
1630  *              1. If first write of transaction, write start record
1631  *              2. Write log operation header (header per region)
1632  *              3. Find out if we can fit entire region into this iclog
1633  *              4. Potentially, verify destination memcpy ptr
1634  *              5. Memcpy (partial) region
1635  *              6. If partial copy, release iclog; otherwise, continue
1636  *                      copying more regions into current iclog
1637  *      4. Mark want sync bit (in simulation mode)
1638  *      5. Release iclog for potential flush to on-disk log.
1639  *
1640  * ERRORS:
1641  * 1.   Panic if reservation is overrun.  This should never happen since
1642  *      reservation amounts are generated internal to the filesystem.
1643  * NOTES:
1644  * 1. Tickets are single threaded data structures.
1645  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1646  *      syncing routine.  When a single log_write region needs to span
1647  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1648  *      on all log operation writes which don't contain the end of the
1649  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1650  *      operation which contains the end of the continued log_write region.
1651  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1652  *      we don't really know exactly how much space will be used.  As a result,
1653  *      we don't update ic_offset until the end when we know exactly how many
1654  *      bytes have been written out.
1655  */
1656 STATIC int
1657 xlog_write(
1658         struct xfs_mount        *mp,
1659         struct xfs_log_iovec    reg[],
1660         int                     nentries,
1661         struct xlog_ticket      *ticket,
1662         xfs_lsn_t               *start_lsn,
1663         struct xlog_in_core     **commit_iclog,
1664         uint                    flags)
1665 {
1666     xlog_t           *log = mp->m_log;
1667     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1668     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1669     __psint_t        ptr;            /* copy address into data region */
1670     int              len;            /* # xlog_write() bytes 2 still copy */
1671     int              index;          /* region index currently copying */
1672     int              log_offset;     /* offset (from 0) into data region */
1673     int              start_rec_copy; /* # bytes to copy for start record */
1674     int              partial_copy;   /* did we split a region? */
1675     int              partial_copy_len;/* # bytes copied if split region */
1676     int              need_copy;      /* # bytes need to memcpy this region */
1677     int              copy_len;       /* # bytes actually memcpy'ing */
1678     int              copy_off;       /* # bytes from entry start */
1679     int              contwr;         /* continued write of in-core log? */
1680     int              error;
1681     int              record_cnt = 0, data_cnt = 0;
1682
1683     partial_copy_len = partial_copy = 0;
1684
1685     /* Calculate potential maximum space.  Each region gets its own
1686      * xlog_op_header_t and may need to be double word aligned.
1687      */
1688     len = 0;
1689     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1690         len += sizeof(xlog_op_header_t);
1691         ticket->t_res_num_ophdrs++;
1692     }
1693
1694     for (index = 0; index < nentries; index++) {
1695         len += sizeof(xlog_op_header_t);            /* each region gets >= 1 */
1696         ticket->t_res_num_ophdrs++;
1697         len += reg[index].i_len;
1698         xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1699     }
1700     contwr = *start_lsn = 0;
1701
1702     if (ticket->t_curr_res < len) {
1703         xlog_print_tic_res(mp, ticket);
1704 #ifdef DEBUG
1705         xlog_panic(
1706                 "xfs_log_write: reservation ran out. Need to up reservation");
1707 #else
1708         /* Customer configurable panic */
1709         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1710                 "xfs_log_write: reservation ran out. Need to up reservation");
1711         /* If we did not panic, shutdown the filesystem */
1712         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1713 #endif
1714     } else
1715         ticket->t_curr_res -= len;
1716
1717     for (index = 0; index < nentries; ) {
1718         if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1719                                                &contwr, &log_offset)))
1720                 return error;
1721
1722         ASSERT(log_offset <= iclog->ic_size - 1);
1723         ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1724
1725         /* start_lsn is the first lsn written to. That's all we need. */
1726         if (! *start_lsn)
1727             *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1728
1729         /* This loop writes out as many regions as can fit in the amount
1730          * of space which was allocated by xlog_state_get_iclog_space().
1731          */
1732         while (index < nentries) {
1733             ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1734             ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1735             start_rec_copy = 0;
1736
1737             /* If first write for transaction, insert start record.
1738              * We can't be trying to commit if we are inited.  We can't
1739              * have any "partial_copy" if we are inited.
1740              */
1741             if (ticket->t_flags & XLOG_TIC_INITED) {
1742                 logop_head              = (xlog_op_header_t *)ptr;
1743                 logop_head->oh_tid      = cpu_to_be32(ticket->t_tid);
1744                 logop_head->oh_clientid = ticket->t_clientid;
1745                 logop_head->oh_len      = 0;
1746                 logop_head->oh_flags    = XLOG_START_TRANS;
1747                 logop_head->oh_res2     = 0;
1748                 ticket->t_flags         &= ~XLOG_TIC_INITED;    /* clear bit */
1749                 record_cnt++;
1750
1751                 start_rec_copy = sizeof(xlog_op_header_t);
1752                 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1753             }
1754
1755             /* Copy log operation header directly into data section */
1756             logop_head                  = (xlog_op_header_t *)ptr;
1757             logop_head->oh_tid          = cpu_to_be32(ticket->t_tid);
1758             logop_head->oh_clientid     = ticket->t_clientid;
1759             logop_head->oh_res2         = 0;
1760
1761             /* header copied directly */
1762             xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1763
1764             /* are we copying a commit or unmount record? */
1765             logop_head->oh_flags = flags;
1766
1767             /*
1768              * We've seen logs corrupted with bad transaction client
1769              * ids.  This makes sure that XFS doesn't generate them on.
1770              * Turn this into an EIO and shut down the filesystem.
1771              */
1772             switch (logop_head->oh_clientid)  {
1773             case XFS_TRANSACTION:
1774             case XFS_VOLUME:
1775             case XFS_LOG:
1776                 break;
1777             default:
1778                 xfs_fs_cmn_err(CE_WARN, mp,
1779                     "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1780                     logop_head->oh_clientid, ticket);
1781                 return XFS_ERROR(EIO);
1782             }
1783
1784             /* Partial write last time? => (partial_copy != 0)
1785              * need_copy is the amount we'd like to copy if everything could
1786              * fit in the current memcpy.
1787              */
1788             need_copy = reg[index].i_len - partial_copy_len;
1789
1790             copy_off = partial_copy_len;
1791             if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1792                 copy_len = need_copy;
1793                 logop_head->oh_len = cpu_to_be32(copy_len);
1794                 if (partial_copy)
1795                     logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1796                 partial_copy_len = partial_copy = 0;
1797             } else {                                        /* partial write */
1798                 copy_len = iclog->ic_size - log_offset;
1799                 logop_head->oh_len = cpu_to_be32(copy_len);
1800                 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1801                 if (partial_copy)
1802                         logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1803                 partial_copy_len += copy_len;
1804                 partial_copy++;
1805                 len += sizeof(xlog_op_header_t); /* from splitting of region */
1806                 /* account for new log op header */
1807                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1808                 ticket->t_res_num_ophdrs++;
1809             }
1810             xlog_verify_dest_ptr(log, ptr);
1811
1812             /* copy region */
1813             ASSERT(copy_len >= 0);
1814             memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1815             xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1816
1817             /* make copy_len total bytes copied, including headers */
1818             copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1819             record_cnt++;
1820             data_cnt += contwr ? copy_len : 0;
1821             if (partial_copy) {                 /* copied partial region */
1822                     /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1823                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1824                     record_cnt = data_cnt = 0;
1825                     if ((error = xlog_state_release_iclog(log, iclog)))
1826                             return error;
1827                     break;                      /* don't increment index */
1828             } else {                            /* copied entire region */
1829                 index++;
1830                 partial_copy_len = partial_copy = 0;
1831
1832                 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1833                     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1834                     record_cnt = data_cnt = 0;
1835                     spin_lock(&log->l_icloglock);
1836                     xlog_state_want_sync(log, iclog);
1837                     spin_unlock(&log->l_icloglock);
1838                     if (commit_iclog) {
1839                         ASSERT(flags & XLOG_COMMIT_TRANS);
1840                         *commit_iclog = iclog;
1841                     } else if ((error = xlog_state_release_iclog(log, iclog)))
1842                            return error;
1843                     if (index == nentries)
1844                             return 0;           /* we are done */
1845                     else
1846                             break;
1847                 }
1848             } /* if (partial_copy) */
1849         } /* while (index < nentries) */
1850     } /* for (index = 0; index < nentries; ) */
1851     ASSERT(len == 0);
1852
1853     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1854     if (commit_iclog) {
1855         ASSERT(flags & XLOG_COMMIT_TRANS);
1856         *commit_iclog = iclog;
1857         return 0;
1858     }
1859     return xlog_state_release_iclog(log, iclog);
1860 }       /* xlog_write */
1861
1862
1863 /*****************************************************************************
1864  *
1865  *              State Machine functions
1866  *
1867  *****************************************************************************
1868  */
1869
1870 /* Clean iclogs starting from the head.  This ordering must be
1871  * maintained, so an iclog doesn't become ACTIVE beyond one that
1872  * is SYNCING.  This is also required to maintain the notion that we use
1873  * a ordered wait queue to hold off would be writers to the log when every
1874  * iclog is trying to sync to disk.
1875  *
1876  * State Change: DIRTY -> ACTIVE
1877  */
1878 STATIC void
1879 xlog_state_clean_log(xlog_t *log)
1880 {
1881         xlog_in_core_t  *iclog;
1882         int changed = 0;
1883
1884         iclog = log->l_iclog;
1885         do {
1886                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1887                         iclog->ic_state = XLOG_STATE_ACTIVE;
1888                         iclog->ic_offset       = 0;
1889                         ASSERT(iclog->ic_callback == NULL);
1890                         /*
1891                          * If the number of ops in this iclog indicate it just
1892                          * contains the dummy transaction, we can
1893                          * change state into IDLE (the second time around).
1894                          * Otherwise we should change the state into
1895                          * NEED a dummy.
1896                          * We don't need to cover the dummy.
1897                          */
1898                         if (!changed &&
1899                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
1900                                         XLOG_COVER_OPS)) {
1901                                 changed = 1;
1902                         } else {
1903                                 /*
1904                                  * We have two dirty iclogs so start over
1905                                  * This could also be num of ops indicates
1906                                  * this is not the dummy going out.
1907                                  */
1908                                 changed = 2;
1909                         }
1910                         iclog->ic_header.h_num_logops = 0;
1911                         memset(iclog->ic_header.h_cycle_data, 0,
1912                               sizeof(iclog->ic_header.h_cycle_data));
1913                         iclog->ic_header.h_lsn = 0;
1914                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1915                         /* do nothing */;
1916                 else
1917                         break;  /* stop cleaning */
1918                 iclog = iclog->ic_next;
1919         } while (iclog != log->l_iclog);
1920
1921         /* log is locked when we are called */
1922         /*
1923          * Change state for the dummy log recording.
1924          * We usually go to NEED. But we go to NEED2 if the changed indicates
1925          * we are done writing the dummy record.
1926          * If we are done with the second dummy recored (DONE2), then
1927          * we go to IDLE.
1928          */
1929         if (changed) {
1930                 switch (log->l_covered_state) {
1931                 case XLOG_STATE_COVER_IDLE:
1932                 case XLOG_STATE_COVER_NEED:
1933                 case XLOG_STATE_COVER_NEED2:
1934                         log->l_covered_state = XLOG_STATE_COVER_NEED;
1935                         break;
1936
1937                 case XLOG_STATE_COVER_DONE:
1938                         if (changed == 1)
1939                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
1940                         else
1941                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
1942                         break;
1943
1944                 case XLOG_STATE_COVER_DONE2:
1945                         if (changed == 1)
1946                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1947                         else
1948                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
1949                         break;
1950
1951                 default:
1952                         ASSERT(0);
1953                 }
1954         }
1955 }       /* xlog_state_clean_log */
1956
1957 STATIC xfs_lsn_t
1958 xlog_get_lowest_lsn(
1959         xlog_t          *log)
1960 {
1961         xlog_in_core_t  *lsn_log;
1962         xfs_lsn_t       lowest_lsn, lsn;
1963
1964         lsn_log = log->l_iclog;
1965         lowest_lsn = 0;
1966         do {
1967             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
1968                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
1969                 if ((lsn && !lowest_lsn) ||
1970                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
1971                         lowest_lsn = lsn;
1972                 }
1973             }
1974             lsn_log = lsn_log->ic_next;
1975         } while (lsn_log != log->l_iclog);
1976         return lowest_lsn;
1977 }
1978
1979
1980 STATIC void
1981 xlog_state_do_callback(
1982         xlog_t          *log,
1983         int             aborted,
1984         xlog_in_core_t  *ciclog)
1985 {
1986         xlog_in_core_t     *iclog;
1987         xlog_in_core_t     *first_iclog;        /* used to know when we've
1988                                                  * processed all iclogs once */
1989         xfs_log_callback_t *cb, *cb_next;
1990         int                flushcnt = 0;
1991         xfs_lsn_t          lowest_lsn;
1992         int                ioerrors;    /* counter: iclogs with errors */
1993         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
1994         int                funcdidcallbacks; /* flag: function did callbacks */
1995         int                repeats;     /* for issuing console warnings if
1996                                          * looping too many times */
1997         int                wake = 0;
1998
1999         spin_lock(&log->l_icloglock);
2000         first_iclog = iclog = log->l_iclog;
2001         ioerrors = 0;
2002         funcdidcallbacks = 0;
2003         repeats = 0;
2004
2005         do {
2006                 /*
2007                  * Scan all iclogs starting with the one pointed to by the
2008                  * log.  Reset this starting point each time the log is
2009                  * unlocked (during callbacks).
2010                  *
2011                  * Keep looping through iclogs until one full pass is made
2012                  * without running any callbacks.
2013                  */
2014                 first_iclog = log->l_iclog;
2015                 iclog = log->l_iclog;
2016                 loopdidcallbacks = 0;
2017                 repeats++;
2018
2019                 do {
2020
2021                         /* skip all iclogs in the ACTIVE & DIRTY states */
2022                         if (iclog->ic_state &
2023                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2024                                 iclog = iclog->ic_next;
2025                                 continue;
2026                         }
2027
2028                         /*
2029                          * Between marking a filesystem SHUTDOWN and stopping
2030                          * the log, we do flush all iclogs to disk (if there
2031                          * wasn't a log I/O error). So, we do want things to
2032                          * go smoothly in case of just a SHUTDOWN  w/o a
2033                          * LOG_IO_ERROR.
2034                          */
2035                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2036                                 /*
2037                                  * Can only perform callbacks in order.  Since
2038                                  * this iclog is not in the DONE_SYNC/
2039                                  * DO_CALLBACK state, we skip the rest and
2040                                  * just try to clean up.  If we set our iclog
2041                                  * to DO_CALLBACK, we will not process it when
2042                                  * we retry since a previous iclog is in the
2043                                  * CALLBACK and the state cannot change since
2044                                  * we are holding the l_icloglock.
2045                                  */
2046                                 if (!(iclog->ic_state &
2047                                         (XLOG_STATE_DONE_SYNC |
2048                                                  XLOG_STATE_DO_CALLBACK))) {
2049                                         if (ciclog && (ciclog->ic_state ==
2050                                                         XLOG_STATE_DONE_SYNC)) {
2051                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2052                                         }
2053                                         break;
2054                                 }
2055                                 /*
2056                                  * We now have an iclog that is in either the
2057                                  * DO_CALLBACK or DONE_SYNC states. The other
2058                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2059                                  * caught by the above if and are going to
2060                                  * clean (i.e. we aren't doing their callbacks)
2061                                  * see the above if.
2062                                  */
2063
2064                                 /*
2065                                  * We will do one more check here to see if we
2066                                  * have chased our tail around.
2067                                  */
2068
2069                                 lowest_lsn = xlog_get_lowest_lsn(log);
2070                                 if (lowest_lsn &&
2071                                     XFS_LSN_CMP(lowest_lsn,
2072                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2073                                         iclog = iclog->ic_next;
2074                                         continue; /* Leave this iclog for
2075                                                    * another thread */
2076                                 }
2077
2078                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2079
2080                                 spin_unlock(&log->l_icloglock);
2081
2082                                 /* l_last_sync_lsn field protected by
2083                                  * l_grant_lock. Don't worry about iclog's lsn.
2084                                  * No one else can be here except us.
2085                                  */
2086                                 spin_lock(&log->l_grant_lock);
2087                                 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2088                                        be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2089                                 log->l_last_sync_lsn =
2090                                         be64_to_cpu(iclog->ic_header.h_lsn);
2091                                 spin_unlock(&log->l_grant_lock);
2092
2093                         } else {
2094                                 spin_unlock(&log->l_icloglock);
2095                                 ioerrors++;
2096                         }
2097
2098                         /*
2099                          * Keep processing entries in the callback list until
2100                          * we come around and it is empty.  We need to
2101                          * atomically see that the list is empty and change the
2102                          * state to DIRTY so that we don't miss any more
2103                          * callbacks being added.
2104                          */
2105                         spin_lock(&iclog->ic_callback_lock);
2106                         cb = iclog->ic_callback;
2107                         while (cb) {
2108                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2109                                 iclog->ic_callback = NULL;
2110                                 spin_unlock(&iclog->ic_callback_lock);
2111
2112                                 /* perform callbacks in the order given */
2113                                 for (; cb; cb = cb_next) {
2114                                         cb_next = cb->cb_next;
2115                                         cb->cb_func(cb->cb_arg, aborted);
2116                                 }
2117                                 spin_lock(&iclog->ic_callback_lock);
2118                                 cb = iclog->ic_callback;
2119                         }
2120
2121                         loopdidcallbacks++;
2122                         funcdidcallbacks++;
2123
2124                         spin_lock(&log->l_icloglock);
2125                         ASSERT(iclog->ic_callback == NULL);
2126                         spin_unlock(&iclog->ic_callback_lock);
2127                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2128                                 iclog->ic_state = XLOG_STATE_DIRTY;
2129
2130                         /*
2131                          * Transition from DIRTY to ACTIVE if applicable.
2132                          * NOP if STATE_IOERROR.
2133                          */
2134                         xlog_state_clean_log(log);
2135
2136                         /* wake up threads waiting in xfs_log_force() */
2137                         sv_broadcast(&iclog->ic_force_wait);
2138
2139                         iclog = iclog->ic_next;
2140                 } while (first_iclog != iclog);
2141
2142                 if (repeats > 5000) {
2143                         flushcnt += repeats;
2144                         repeats = 0;
2145                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2146                                 "%s: possible infinite loop (%d iterations)",
2147                                 __func__, flushcnt);
2148                 }
2149         } while (!ioerrors && loopdidcallbacks);
2150
2151         /*
2152          * make one last gasp attempt to see if iclogs are being left in
2153          * limbo..
2154          */
2155 #ifdef DEBUG
2156         if (funcdidcallbacks) {
2157                 first_iclog = iclog = log->l_iclog;
2158                 do {
2159                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2160                         /*
2161                          * Terminate the loop if iclogs are found in states
2162                          * which will cause other threads to clean up iclogs.
2163                          *
2164                          * SYNCING - i/o completion will go through logs
2165                          * DONE_SYNC - interrupt thread should be waiting for
2166                          *              l_icloglock
2167                          * IOERROR - give up hope all ye who enter here
2168                          */
2169                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2170                             iclog->ic_state == XLOG_STATE_SYNCING ||
2171                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2172                             iclog->ic_state == XLOG_STATE_IOERROR )
2173                                 break;
2174                         iclog = iclog->ic_next;
2175                 } while (first_iclog != iclog);
2176         }
2177 #endif
2178
2179         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2180                 wake = 1;
2181         spin_unlock(&log->l_icloglock);
2182
2183         if (wake)
2184                 sv_broadcast(&log->l_flush_wait);
2185 }
2186
2187
2188 /*
2189  * Finish transitioning this iclog to the dirty state.
2190  *
2191  * Make sure that we completely execute this routine only when this is
2192  * the last call to the iclog.  There is a good chance that iclog flushes,
2193  * when we reach the end of the physical log, get turned into 2 separate
2194  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2195  * routine.  By using the reference count bwritecnt, we guarantee that only
2196  * the second completion goes through.
2197  *
2198  * Callbacks could take time, so they are done outside the scope of the
2199  * global state machine log lock.
2200  */
2201 STATIC void
2202 xlog_state_done_syncing(
2203         xlog_in_core_t  *iclog,
2204         int             aborted)
2205 {
2206         xlog_t             *log = iclog->ic_log;
2207
2208         spin_lock(&log->l_icloglock);
2209
2210         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2211                iclog->ic_state == XLOG_STATE_IOERROR);
2212         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2213         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2214
2215
2216         /*
2217          * If we got an error, either on the first buffer, or in the case of
2218          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2219          * and none should ever be attempted to be written to disk
2220          * again.
2221          */
2222         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2223                 if (--iclog->ic_bwritecnt == 1) {
2224                         spin_unlock(&log->l_icloglock);
2225                         return;
2226                 }
2227                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2228         }
2229
2230         /*
2231          * Someone could be sleeping prior to writing out the next
2232          * iclog buffer, we wake them all, one will get to do the
2233          * I/O, the others get to wait for the result.
2234          */
2235         sv_broadcast(&iclog->ic_write_wait);
2236         spin_unlock(&log->l_icloglock);
2237         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2238 }       /* xlog_state_done_syncing */
2239
2240
2241 /*
2242  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2243  * sleep.  We wait on the flush queue on the head iclog as that should be
2244  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2245  * we will wait here and all new writes will sleep until a sync completes.
2246  *
2247  * The in-core logs are used in a circular fashion. They are not used
2248  * out-of-order even when an iclog past the head is free.
2249  *
2250  * return:
2251  *      * log_offset where xlog_write() can start writing into the in-core
2252  *              log's data space.
2253  *      * in-core log pointer to which xlog_write() should write.
2254  *      * boolean indicating this is a continued write to an in-core log.
2255  *              If this is the last write, then the in-core log's offset field
2256  *              needs to be incremented, depending on the amount of data which
2257  *              is copied.
2258  */
2259 STATIC int
2260 xlog_state_get_iclog_space(xlog_t         *log,
2261                            int            len,
2262                            xlog_in_core_t **iclogp,
2263                            xlog_ticket_t  *ticket,
2264                            int            *continued_write,
2265                            int            *logoffsetp)
2266 {
2267         int               log_offset;
2268         xlog_rec_header_t *head;
2269         xlog_in_core_t    *iclog;
2270         int               error;
2271
2272 restart:
2273         spin_lock(&log->l_icloglock);
2274         if (XLOG_FORCED_SHUTDOWN(log)) {
2275                 spin_unlock(&log->l_icloglock);
2276                 return XFS_ERROR(EIO);
2277         }
2278
2279         iclog = log->l_iclog;
2280         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2281                 XFS_STATS_INC(xs_log_noiclogs);
2282
2283                 /* Wait for log writes to have flushed */
2284                 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2285                 goto restart;
2286         }
2287
2288         head = &iclog->ic_header;
2289
2290         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2291         log_offset = iclog->ic_offset;
2292
2293         /* On the 1st write to an iclog, figure out lsn.  This works
2294          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2295          * committing to.  If the offset is set, that's how many blocks
2296          * must be written.
2297          */
2298         if (log_offset == 0) {
2299                 ticket->t_curr_res -= log->l_iclog_hsize;
2300                 xlog_tic_add_region(ticket,
2301                                     log->l_iclog_hsize,
2302                                     XLOG_REG_TYPE_LRHEADER);
2303                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2304                 head->h_lsn = cpu_to_be64(
2305                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2306                 ASSERT(log->l_curr_block >= 0);
2307         }
2308
2309         /* If there is enough room to write everything, then do it.  Otherwise,
2310          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2311          * bit is on, so this will get flushed out.  Don't update ic_offset
2312          * until you know exactly how many bytes get copied.  Therefore, wait
2313          * until later to update ic_offset.
2314          *
2315          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2316          * can fit into remaining data section.
2317          */
2318         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2319                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2320
2321                 /*
2322                  * If I'm the only one writing to this iclog, sync it to disk.
2323                  * We need to do an atomic compare and decrement here to avoid
2324                  * racing with concurrent atomic_dec_and_lock() calls in
2325                  * xlog_state_release_iclog() when there is more than one
2326                  * reference to the iclog.
2327                  */
2328                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2329                         /* we are the only one */
2330                         spin_unlock(&log->l_icloglock);
2331                         error = xlog_state_release_iclog(log, iclog);
2332                         if (error)
2333                                 return error;
2334                 } else {
2335                         spin_unlock(&log->l_icloglock);
2336                 }
2337                 goto restart;
2338         }
2339
2340         /* Do we have enough room to write the full amount in the remainder
2341          * of this iclog?  Or must we continue a write on the next iclog and
2342          * mark this iclog as completely taken?  In the case where we switch
2343          * iclogs (to mark it taken), this particular iclog will release/sync
2344          * to disk in xlog_write().
2345          */
2346         if (len <= iclog->ic_size - iclog->ic_offset) {
2347                 *continued_write = 0;
2348                 iclog->ic_offset += len;
2349         } else {
2350                 *continued_write = 1;
2351                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2352         }
2353         *iclogp = iclog;
2354
2355         ASSERT(iclog->ic_offset <= iclog->ic_size);
2356         spin_unlock(&log->l_icloglock);
2357
2358         *logoffsetp = log_offset;
2359         return 0;
2360 }       /* xlog_state_get_iclog_space */
2361
2362 /*
2363  * Atomically get the log space required for a log ticket.
2364  *
2365  * Once a ticket gets put onto the reserveq, it will only return after
2366  * the needed reservation is satisfied.
2367  */
2368 STATIC int
2369 xlog_grant_log_space(xlog_t        *log,
2370                      xlog_ticket_t *tic)
2371 {
2372         int              free_bytes;
2373         int              need_bytes;
2374 #ifdef DEBUG
2375         xfs_lsn_t        tail_lsn;
2376 #endif
2377
2378
2379 #ifdef DEBUG
2380         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2381                 panic("grant Recovery problem");
2382 #endif
2383
2384         /* Is there space or do we need to sleep? */
2385         spin_lock(&log->l_grant_lock);
2386
2387         trace_xfs_log_grant_enter(log, tic);
2388
2389         /* something is already sleeping; insert new transaction at end */
2390         if (log->l_reserve_headq) {
2391                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2392
2393                 trace_xfs_log_grant_sleep1(log, tic);
2394
2395                 /*
2396                  * Gotta check this before going to sleep, while we're
2397                  * holding the grant lock.
2398                  */
2399                 if (XLOG_FORCED_SHUTDOWN(log))
2400                         goto error_return;
2401
2402                 XFS_STATS_INC(xs_sleep_logspace);
2403                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2404                 /*
2405                  * If we got an error, and the filesystem is shutting down,
2406                  * we'll catch it down below. So just continue...
2407                  */
2408                 trace_xfs_log_grant_wake1(log, tic);
2409                 spin_lock(&log->l_grant_lock);
2410         }
2411         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2412                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2413         else
2414                 need_bytes = tic->t_unit_res;
2415
2416 redo:
2417         if (XLOG_FORCED_SHUTDOWN(log))
2418                 goto error_return;
2419
2420         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2421                                      log->l_grant_reserve_bytes);
2422         if (free_bytes < need_bytes) {
2423                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2424                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2425
2426                 trace_xfs_log_grant_sleep2(log, tic);
2427
2428                 spin_unlock(&log->l_grant_lock);
2429                 xlog_grant_push_ail(log->l_mp, need_bytes);
2430                 spin_lock(&log->l_grant_lock);
2431
2432                 XFS_STATS_INC(xs_sleep_logspace);
2433                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2434
2435                 spin_lock(&log->l_grant_lock);
2436                 if (XLOG_FORCED_SHUTDOWN(log))
2437                         goto error_return;
2438
2439                 trace_xfs_log_grant_wake2(log, tic);
2440
2441                 goto redo;
2442         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2443                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2444
2445         /* we've got enough space */
2446         xlog_grant_add_space(log, need_bytes);
2447 #ifdef DEBUG
2448         tail_lsn = log->l_tail_lsn;
2449         /*
2450          * Check to make sure the grant write head didn't just over lap the
2451          * tail.  If the cycles are the same, we can't be overlapping.
2452          * Otherwise, make sure that the cycles differ by exactly one and
2453          * check the byte count.
2454          */
2455         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2456                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2457                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2458         }
2459 #endif
2460         trace_xfs_log_grant_exit(log, tic);
2461         xlog_verify_grant_head(log, 1);
2462         spin_unlock(&log->l_grant_lock);
2463         return 0;
2464
2465  error_return:
2466         if (tic->t_flags & XLOG_TIC_IN_Q)
2467                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2468
2469         trace_xfs_log_grant_error(log, tic);
2470
2471         /*
2472          * If we are failing, make sure the ticket doesn't have any
2473          * current reservations. We don't want to add this back when
2474          * the ticket/transaction gets cancelled.
2475          */
2476         tic->t_curr_res = 0;
2477         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2478         spin_unlock(&log->l_grant_lock);
2479         return XFS_ERROR(EIO);
2480 }       /* xlog_grant_log_space */
2481
2482
2483 /*
2484  * Replenish the byte reservation required by moving the grant write head.
2485  *
2486  *
2487  */
2488 STATIC int
2489 xlog_regrant_write_log_space(xlog_t        *log,
2490                              xlog_ticket_t *tic)
2491 {
2492         int             free_bytes, need_bytes;
2493         xlog_ticket_t   *ntic;
2494 #ifdef DEBUG
2495         xfs_lsn_t       tail_lsn;
2496 #endif
2497
2498         tic->t_curr_res = tic->t_unit_res;
2499         xlog_tic_reset_res(tic);
2500
2501         if (tic->t_cnt > 0)
2502                 return 0;
2503
2504 #ifdef DEBUG
2505         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2506                 panic("regrant Recovery problem");
2507 #endif
2508
2509         spin_lock(&log->l_grant_lock);
2510
2511         trace_xfs_log_regrant_write_enter(log, tic);
2512
2513         if (XLOG_FORCED_SHUTDOWN(log))
2514                 goto error_return;
2515
2516         /* If there are other waiters on the queue then give them a
2517          * chance at logspace before us. Wake up the first waiters,
2518          * if we do not wake up all the waiters then go to sleep waiting
2519          * for more free space, otherwise try to get some space for
2520          * this transaction.
2521          */
2522         need_bytes = tic->t_unit_res;
2523         if ((ntic = log->l_write_headq)) {
2524                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2525                                              log->l_grant_write_bytes);
2526                 do {
2527                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2528
2529                         if (free_bytes < ntic->t_unit_res)
2530                                 break;
2531                         free_bytes -= ntic->t_unit_res;
2532                         sv_signal(&ntic->t_wait);
2533                         ntic = ntic->t_next;
2534                 } while (ntic != log->l_write_headq);
2535
2536                 if (ntic != log->l_write_headq) {
2537                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2538                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2539
2540                         trace_xfs_log_regrant_write_sleep1(log, tic);
2541
2542                         spin_unlock(&log->l_grant_lock);
2543                         xlog_grant_push_ail(log->l_mp, need_bytes);
2544                         spin_lock(&log->l_grant_lock);
2545
2546                         XFS_STATS_INC(xs_sleep_logspace);
2547                         sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2548                                 &log->l_grant_lock, s);
2549
2550                         /* If we're shutting down, this tic is already
2551                          * off the queue */
2552                         spin_lock(&log->l_grant_lock);
2553                         if (XLOG_FORCED_SHUTDOWN(log))
2554                                 goto error_return;
2555
2556                         trace_xfs_log_regrant_write_wake1(log, tic);
2557                 }
2558         }
2559
2560 redo:
2561         if (XLOG_FORCED_SHUTDOWN(log))
2562                 goto error_return;
2563
2564         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2565                                      log->l_grant_write_bytes);
2566         if (free_bytes < need_bytes) {
2567                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2568                         xlog_ins_ticketq(&log->l_write_headq, tic);
2569                 spin_unlock(&log->l_grant_lock);
2570                 xlog_grant_push_ail(log->l_mp, need_bytes);
2571                 spin_lock(&log->l_grant_lock);
2572
2573                 XFS_STATS_INC(xs_sleep_logspace);
2574                 trace_xfs_log_regrant_write_sleep2(log, tic);
2575
2576                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2577
2578                 /* If we're shutting down, this tic is already off the queue */
2579                 spin_lock(&log->l_grant_lock);
2580                 if (XLOG_FORCED_SHUTDOWN(log))
2581                         goto error_return;
2582
2583                 trace_xfs_log_regrant_write_wake2(log, tic);
2584                 goto redo;
2585         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2586                 xlog_del_ticketq(&log->l_write_headq, tic);
2587
2588         /* we've got enough space */
2589         xlog_grant_add_space_write(log, need_bytes);
2590 #ifdef DEBUG
2591         tail_lsn = log->l_tail_lsn;
2592         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2593                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2594                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2595         }
2596 #endif
2597
2598         trace_xfs_log_regrant_write_exit(log, tic);
2599
2600         xlog_verify_grant_head(log, 1);
2601         spin_unlock(&log->l_grant_lock);
2602         return 0;
2603
2604
2605  error_return:
2606         if (tic->t_flags & XLOG_TIC_IN_Q)
2607                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2608
2609         trace_xfs_log_regrant_write_error(log, tic);
2610
2611         /*
2612          * If we are failing, make sure the ticket doesn't have any
2613          * current reservations. We don't want to add this back when
2614          * the ticket/transaction gets cancelled.
2615          */
2616         tic->t_curr_res = 0;
2617         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2618         spin_unlock(&log->l_grant_lock);
2619         return XFS_ERROR(EIO);
2620 }       /* xlog_regrant_write_log_space */
2621
2622
2623 /* The first cnt-1 times through here we don't need to
2624  * move the grant write head because the permanent
2625  * reservation has reserved cnt times the unit amount.
2626  * Release part of current permanent unit reservation and
2627  * reset current reservation to be one units worth.  Also
2628  * move grant reservation head forward.
2629  */
2630 STATIC void
2631 xlog_regrant_reserve_log_space(xlog_t        *log,
2632                                xlog_ticket_t *ticket)
2633 {
2634         trace_xfs_log_regrant_reserve_enter(log, ticket);
2635
2636         if (ticket->t_cnt > 0)
2637                 ticket->t_cnt--;
2638
2639         spin_lock(&log->l_grant_lock);
2640         xlog_grant_sub_space(log, ticket->t_curr_res);
2641         ticket->t_curr_res = ticket->t_unit_res;
2642         xlog_tic_reset_res(ticket);
2643
2644         trace_xfs_log_regrant_reserve_sub(log, ticket);
2645
2646         xlog_verify_grant_head(log, 1);
2647
2648         /* just return if we still have some of the pre-reserved space */
2649         if (ticket->t_cnt > 0) {
2650                 spin_unlock(&log->l_grant_lock);
2651                 return;
2652         }
2653
2654         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2655
2656         trace_xfs_log_regrant_reserve_exit(log, ticket);
2657
2658         xlog_verify_grant_head(log, 0);
2659         spin_unlock(&log->l_grant_lock);
2660         ticket->t_curr_res = ticket->t_unit_res;
2661         xlog_tic_reset_res(ticket);
2662 }       /* xlog_regrant_reserve_log_space */
2663
2664
2665 /*
2666  * Give back the space left from a reservation.
2667  *
2668  * All the information we need to make a correct determination of space left
2669  * is present.  For non-permanent reservations, things are quite easy.  The
2670  * count should have been decremented to zero.  We only need to deal with the
2671  * space remaining in the current reservation part of the ticket.  If the
2672  * ticket contains a permanent reservation, there may be left over space which
2673  * needs to be released.  A count of N means that N-1 refills of the current
2674  * reservation can be done before we need to ask for more space.  The first
2675  * one goes to fill up the first current reservation.  Once we run out of
2676  * space, the count will stay at zero and the only space remaining will be
2677  * in the current reservation field.
2678  */
2679 STATIC void
2680 xlog_ungrant_log_space(xlog_t        *log,
2681                        xlog_ticket_t *ticket)
2682 {
2683         if (ticket->t_cnt > 0)
2684                 ticket->t_cnt--;
2685
2686         spin_lock(&log->l_grant_lock);
2687         trace_xfs_log_ungrant_enter(log, ticket);
2688
2689         xlog_grant_sub_space(log, ticket->t_curr_res);
2690
2691         trace_xfs_log_ungrant_sub(log, ticket);
2692
2693         /* If this is a permanent reservation ticket, we may be able to free
2694          * up more space based on the remaining count.
2695          */
2696         if (ticket->t_cnt > 0) {
2697                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2698                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2699         }
2700
2701         trace_xfs_log_ungrant_exit(log, ticket);
2702
2703         xlog_verify_grant_head(log, 1);
2704         spin_unlock(&log->l_grant_lock);
2705         xfs_log_move_tail(log->l_mp, 1);
2706 }       /* xlog_ungrant_log_space */
2707
2708
2709 /*
2710  * Flush iclog to disk if this is the last reference to the given iclog and
2711  * the WANT_SYNC bit is set.
2712  *
2713  * When this function is entered, the iclog is not necessarily in the
2714  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2715  *
2716  *
2717  */
2718 STATIC int
2719 xlog_state_release_iclog(
2720         xlog_t          *log,
2721         xlog_in_core_t  *iclog)
2722 {
2723         int             sync = 0;       /* do we sync? */
2724
2725         if (iclog->ic_state & XLOG_STATE_IOERROR)
2726                 return XFS_ERROR(EIO);
2727
2728         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2729         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2730                 return 0;
2731
2732         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2733                 spin_unlock(&log->l_icloglock);
2734                 return XFS_ERROR(EIO);
2735         }
2736         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2737                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2738
2739         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2740                 /* update tail before writing to iclog */
2741                 xlog_assign_tail_lsn(log->l_mp);
2742                 sync++;
2743                 iclog->ic_state = XLOG_STATE_SYNCING;
2744                 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2745                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2746                 /* cycle incremented when incrementing curr_block */
2747         }
2748         spin_unlock(&log->l_icloglock);
2749
2750         /*
2751          * We let the log lock go, so it's possible that we hit a log I/O
2752          * error or some other SHUTDOWN condition that marks the iclog
2753          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2754          * this iclog has consistent data, so we ignore IOERROR
2755          * flags after this point.
2756          */
2757         if (sync)
2758                 return xlog_sync(log, iclog);
2759         return 0;
2760 }       /* xlog_state_release_iclog */
2761
2762
2763 /*
2764  * This routine will mark the current iclog in the ring as WANT_SYNC
2765  * and move the current iclog pointer to the next iclog in the ring.
2766  * When this routine is called from xlog_state_get_iclog_space(), the
2767  * exact size of the iclog has not yet been determined.  All we know is
2768  * that every data block.  We have run out of space in this log record.
2769  */
2770 STATIC void
2771 xlog_state_switch_iclogs(xlog_t         *log,
2772                          xlog_in_core_t *iclog,
2773                          int            eventual_size)
2774 {
2775         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2776         if (!eventual_size)
2777                 eventual_size = iclog->ic_offset;
2778         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2779         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2780         log->l_prev_block = log->l_curr_block;
2781         log->l_prev_cycle = log->l_curr_cycle;
2782
2783         /* roll log?: ic_offset changed later */
2784         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2785
2786         /* Round up to next log-sunit */
2787         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2788             log->l_mp->m_sb.sb_logsunit > 1) {
2789                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2790                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2791         }
2792
2793         if (log->l_curr_block >= log->l_logBBsize) {
2794                 log->l_curr_cycle++;
2795                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2796                         log->l_curr_cycle++;
2797                 log->l_curr_block -= log->l_logBBsize;
2798                 ASSERT(log->l_curr_block >= 0);
2799         }
2800         ASSERT(iclog == log->l_iclog);
2801         log->l_iclog = iclog->ic_next;
2802 }       /* xlog_state_switch_iclogs */
2803
2804 /*
2805  * Write out all data in the in-core log as of this exact moment in time.
2806  *
2807  * Data may be written to the in-core log during this call.  However,
2808  * we don't guarantee this data will be written out.  A change from past
2809  * implementation means this routine will *not* write out zero length LRs.
2810  *
2811  * Basically, we try and perform an intelligent scan of the in-core logs.
2812  * If we determine there is no flushable data, we just return.  There is no
2813  * flushable data if:
2814  *
2815  *      1. the current iclog is active and has no data; the previous iclog
2816  *              is in the active or dirty state.
2817  *      2. the current iclog is drity, and the previous iclog is in the
2818  *              active or dirty state.
2819  *
2820  * We may sleep if:
2821  *
2822  *      1. the current iclog is not in the active nor dirty state.
2823  *      2. the current iclog dirty, and the previous iclog is not in the
2824  *              active nor dirty state.
2825  *      3. the current iclog is active, and there is another thread writing
2826  *              to this particular iclog.
2827  *      4. a) the current iclog is active and has no other writers
2828  *         b) when we return from flushing out this iclog, it is still
2829  *              not in the active nor dirty state.
2830  */
2831 int
2832 _xfs_log_force(
2833         struct xfs_mount        *mp,
2834         uint                    flags,
2835         int                     *log_flushed)
2836 {
2837         struct log              *log = mp->m_log;
2838         struct xlog_in_core     *iclog;
2839         xfs_lsn_t               lsn;
2840
2841         XFS_STATS_INC(xs_log_force);
2842
2843         spin_lock(&log->l_icloglock);
2844
2845         iclog = log->l_iclog;
2846         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2847                 spin_unlock(&log->l_icloglock);
2848                 return XFS_ERROR(EIO);
2849         }
2850
2851         /* If the head iclog is not active nor dirty, we just attach
2852          * ourselves to the head and go to sleep.
2853          */
2854         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2855             iclog->ic_state == XLOG_STATE_DIRTY) {
2856                 /*
2857                  * If the head is dirty or (active and empty), then
2858                  * we need to look at the previous iclog.  If the previous
2859                  * iclog is active or dirty we are done.  There is nothing
2860                  * to sync out.  Otherwise, we attach ourselves to the
2861                  * previous iclog and go to sleep.
2862                  */
2863                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2864                     (atomic_read(&iclog->ic_refcnt) == 0
2865                      && iclog->ic_offset == 0)) {
2866                         iclog = iclog->ic_prev;
2867                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2868                             iclog->ic_state == XLOG_STATE_DIRTY)
2869                                 goto no_sleep;
2870                         else
2871                                 goto maybe_sleep;
2872                 } else {
2873                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2874                                 /* We are the only one with access to this
2875                                  * iclog.  Flush it out now.  There should
2876                                  * be a roundoff of zero to show that someone
2877                                  * has already taken care of the roundoff from
2878                                  * the previous sync.
2879                                  */
2880                                 atomic_inc(&iclog->ic_refcnt);
2881                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2882                                 xlog_state_switch_iclogs(log, iclog, 0);
2883                                 spin_unlock(&log->l_icloglock);
2884
2885                                 if (xlog_state_release_iclog(log, iclog))
2886                                         return XFS_ERROR(EIO);
2887
2888                                 if (log_flushed)
2889                                         *log_flushed = 1;
2890                                 spin_lock(&log->l_icloglock);
2891                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2892                                     iclog->ic_state != XLOG_STATE_DIRTY)
2893                                         goto maybe_sleep;
2894                                 else
2895                                         goto no_sleep;
2896                         } else {
2897                                 /* Someone else is writing to this iclog.
2898                                  * Use its call to flush out the data.  However,
2899                                  * the other thread may not force out this LR,
2900                                  * so we mark it WANT_SYNC.
2901                                  */
2902                                 xlog_state_switch_iclogs(log, iclog, 0);
2903                                 goto maybe_sleep;
2904                         }
2905                 }
2906         }
2907
2908         /* By the time we come around again, the iclog could've been filled
2909          * which would give it another lsn.  If we have a new lsn, just
2910          * return because the relevant data has been flushed.
2911          */
2912 maybe_sleep:
2913         if (flags & XFS_LOG_SYNC) {
2914                 /*
2915                  * We must check if we're shutting down here, before
2916                  * we wait, while we're holding the l_icloglock.
2917                  * Then we check again after waking up, in case our
2918                  * sleep was disturbed by a bad news.
2919                  */
2920                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2921                         spin_unlock(&log->l_icloglock);
2922                         return XFS_ERROR(EIO);
2923                 }
2924                 XFS_STATS_INC(xs_log_force_sleep);
2925                 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
2926                 /*
2927                  * No need to grab the log lock here since we're
2928                  * only deciding whether or not to return EIO
2929                  * and the memory read should be atomic.
2930                  */
2931                 if (iclog->ic_state & XLOG_STATE_IOERROR)
2932                         return XFS_ERROR(EIO);
2933                 if (log_flushed)
2934                         *log_flushed = 1;
2935         } else {
2936
2937 no_sleep:
2938                 spin_unlock(&log->l_icloglock);
2939         }
2940         return 0;
2941 }
2942
2943 /*
2944  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2945  * about errors or whether the log was flushed or not. This is the normal
2946  * interface to use when trying to unpin items or move the log forward.
2947  */
2948 void
2949 xfs_log_force(
2950         xfs_mount_t     *mp,
2951         uint            flags)
2952 {
2953         int     error;
2954
2955         error = _xfs_log_force(mp, flags, NULL);
2956         if (error) {
2957                 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
2958                         "error %d returned.", error);
2959         }
2960 }
2961
2962 /*
2963  * Force the in-core log to disk for a specific LSN.
2964  *
2965  * Find in-core log with lsn.
2966  *      If it is in the DIRTY state, just return.
2967  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2968  *              state and go to sleep or return.
2969  *      If it is in any other state, go to sleep or return.
2970  *
2971  * Synchronous forces are implemented with a signal variable. All callers
2972  * to force a given lsn to disk will wait on a the sv attached to the
2973  * specific in-core log.  When given in-core log finally completes its
2974  * write to disk, that thread will wake up all threads waiting on the
2975  * sv.
2976  */
2977 int
2978 _xfs_log_force_lsn(
2979         struct xfs_mount        *mp,
2980         xfs_lsn_t               lsn,
2981         uint                    flags,
2982         int                     *log_flushed)
2983 {
2984         struct log              *log = mp->m_log;
2985         struct xlog_in_core     *iclog;
2986         int                     already_slept = 0;
2987
2988         ASSERT(lsn != 0);
2989
2990         XFS_STATS_INC(xs_log_force);
2991
2992 try_again:
2993         spin_lock(&log->l_icloglock);
2994         iclog = log->l_iclog;
2995         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2996                 spin_unlock(&log->l_icloglock);
2997                 return XFS_ERROR(EIO);
2998         }
2999
3000         do {
3001                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3002                         iclog = iclog->ic_next;
3003                         continue;
3004                 }
3005
3006                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3007                         spin_unlock(&log->l_icloglock);
3008                         return 0;
3009                 }
3010
3011                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3012                         /*
3013                          * We sleep here if we haven't already slept (e.g.
3014                          * this is the first time we've looked at the correct
3015                          * iclog buf) and the buffer before us is going to
3016                          * be sync'ed. The reason for this is that if we
3017                          * are doing sync transactions here, by waiting for
3018                          * the previous I/O to complete, we can allow a few
3019                          * more transactions into this iclog before we close
3020                          * it down.
3021                          *
3022                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3023                          * up the refcnt so we can release the log (which
3024                          * drops the ref count).  The state switch keeps new
3025                          * transaction commits from using this buffer.  When
3026                          * the current commits finish writing into the buffer,
3027                          * the refcount will drop to zero and the buffer will
3028                          * go out then.
3029                          */
3030                         if (!already_slept &&
3031                             (iclog->ic_prev->ic_state &
3032                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3033                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3034
3035                                 XFS_STATS_INC(xs_log_force_sleep);
3036
3037                                 sv_wait(&iclog->ic_prev->ic_write_wait,
3038                                         PSWP, &log->l_icloglock, s);
3039                                 if (log_flushed)
3040                                         *log_flushed = 1;
3041                                 already_slept = 1;
3042                                 goto try_again;
3043                         }
3044                         atomic_inc(&iclog->ic_refcnt);
3045                         xlog_state_switch_iclogs(log, iclog, 0);
3046                         spin_unlock(&log->l_icloglock);
3047                         if (xlog_state_release_iclog(log, iclog))
3048                                 return XFS_ERROR(EIO);
3049                         if (log_flushed)
3050                                 *log_flushed = 1;
3051                         spin_lock(&log->l_icloglock);
3052                 }
3053
3054                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3055                     !(iclog->ic_state &
3056                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3057                         /*
3058                          * Don't wait on completion if we know that we've
3059                          * gotten a log write error.
3060                          */
3061                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3062                                 spin_unlock(&log->l_icloglock);
3063                                 return XFS_ERROR(EIO);
3064                         }
3065                         XFS_STATS_INC(xs_log_force_sleep);
3066                         sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3067                         /*
3068                          * No need to grab the log lock here since we're
3069                          * only deciding whether or not to return EIO
3070                          * and the memory read should be atomic.
3071                          */
3072                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3073                                 return XFS_ERROR(EIO);
3074
3075                         if (log_flushed)
3076                                 *log_flushed = 1;
3077                 } else {                /* just return */
3078                         spin_unlock(&log->l_icloglock);
3079                 }
3080
3081                 return 0;
3082         } while (iclog != log->l_iclog);
3083
3084         spin_unlock(&log->l_icloglock);
3085         return 0;
3086 }
3087
3088 /*
3089  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3090  * about errors or whether the log was flushed or not. This is the normal
3091  * interface to use when trying to unpin items or move the log forward.
3092  */
3093 void
3094 xfs_log_force_lsn(
3095         xfs_mount_t     *mp,
3096         xfs_lsn_t       lsn,
3097         uint            flags)
3098 {
3099         int     error;
3100
3101         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3102         if (error) {
3103                 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3104                         "error %d returned.", error);
3105         }
3106 }
3107
3108 /*
3109  * Called when we want to mark the current iclog as being ready to sync to
3110  * disk.
3111  */
3112 STATIC void
3113 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3114 {
3115         assert_spin_locked(&log->l_icloglock);
3116
3117         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3118                 xlog_state_switch_iclogs(log, iclog, 0);
3119         } else {
3120                 ASSERT(iclog->ic_state &
3121                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3122         }
3123 }
3124
3125
3126 /*****************************************************************************
3127  *
3128  *              TICKET functions
3129  *
3130  *****************************************************************************
3131  */
3132
3133 /*
3134  * Free a used ticket when its refcount falls to zero.
3135  */
3136 void
3137 xfs_log_ticket_put(
3138         xlog_ticket_t   *ticket)
3139 {
3140         ASSERT(atomic_read(&ticket->t_ref) > 0);
3141         if (atomic_dec_and_test(&ticket->t_ref)) {
3142                 sv_destroy(&ticket->t_wait);
3143                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3144         }
3145 }
3146
3147 xlog_ticket_t *
3148 xfs_log_ticket_get(
3149         xlog_ticket_t   *ticket)
3150 {
3151         ASSERT(atomic_read(&ticket->t_ref) > 0);
3152         atomic_inc(&ticket->t_ref);
3153         return ticket;
3154 }
3155
3156 /*
3157  * Allocate and initialise a new log ticket.
3158  */
3159 STATIC xlog_ticket_t *
3160 xlog_ticket_alloc(xlog_t                *log,
3161                 int             unit_bytes,
3162                 int             cnt,
3163                 char            client,
3164                 uint            xflags)
3165 {
3166         xlog_ticket_t   *tic;
3167         uint            num_headers;
3168
3169         tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3170         if (!tic)
3171                 return NULL;
3172
3173         /*
3174          * Permanent reservations have up to 'cnt'-1 active log operations
3175          * in the log.  A unit in this case is the amount of space for one
3176          * of these log operations.  Normal reservations have a cnt of 1
3177          * and their unit amount is the total amount of space required.
3178          *
3179          * The following lines of code account for non-transaction data
3180          * which occupy space in the on-disk log.
3181          *
3182          * Normal form of a transaction is:
3183          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3184          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3185          *
3186          * We need to account for all the leadup data and trailer data
3187          * around the transaction data.
3188          * And then we need to account for the worst case in terms of using
3189          * more space.
3190          * The worst case will happen if:
3191          * - the placement of the transaction happens to be such that the
3192          *   roundoff is at its maximum
3193          * - the transaction data is synced before the commit record is synced
3194          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3195          *   Therefore the commit record is in its own Log Record.
3196          *   This can happen as the commit record is called with its
3197          *   own region to xlog_write().
3198          *   This then means that in the worst case, roundoff can happen for
3199          *   the commit-rec as well.
3200          *   The commit-rec is smaller than padding in this scenario and so it is
3201          *   not added separately.
3202          */
3203
3204         /* for trans header */
3205         unit_bytes += sizeof(xlog_op_header_t);
3206         unit_bytes += sizeof(xfs_trans_header_t);
3207
3208         /* for start-rec */
3209         unit_bytes += sizeof(xlog_op_header_t);
3210
3211         /* for LR headers */
3212         num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3213         unit_bytes += log->l_iclog_hsize * num_headers;
3214
3215         /* for commit-rec LR header - note: padding will subsume the ophdr */
3216         unit_bytes += log->l_iclog_hsize;
3217
3218         /* for split-recs - ophdrs added when data split over LRs */
3219         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3220
3221         /* for roundoff padding for transaction data and one for commit record */
3222         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3223             log->l_mp->m_sb.sb_logsunit > 1) {
3224                 /* log su roundoff */
3225                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3226         } else {
3227                 /* BB roundoff */
3228                 unit_bytes += 2*BBSIZE;
3229         }
3230
3231         atomic_set(&tic->t_ref, 1);
3232         tic->t_unit_res         = unit_bytes;
3233         tic->t_curr_res         = unit_bytes;
3234         tic->t_cnt              = cnt;
3235         tic->t_ocnt             = cnt;
3236         tic->t_tid              = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3237         tic->t_clientid         = client;
3238         tic->t_flags            = XLOG_TIC_INITED;
3239         tic->t_trans_type       = 0;
3240         if (xflags & XFS_LOG_PERM_RESERV)
3241                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3242         sv_init(&(tic->t_wait), SV_DEFAULT, "logtick");
3243
3244         xlog_tic_reset_res(tic);
3245
3246         return tic;
3247 }
3248
3249
3250 /******************************************************************************
3251  *
3252  *              Log debug routines
3253  *
3254  ******************************************************************************
3255  */
3256 #if defined(DEBUG)
3257 /*
3258  * Make sure that the destination ptr is within the valid data region of
3259  * one of the iclogs.  This uses backup pointers stored in a different
3260  * part of the log in case we trash the log structure.
3261  */
3262 void
3263 xlog_verify_dest_ptr(xlog_t     *log,
3264                      __psint_t  ptr)
3265 {
3266         int i;
3267         int good_ptr = 0;
3268
3269         for (i=0; i < log->l_iclog_bufs; i++) {
3270                 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3271                     ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3272                         good_ptr++;
3273         }
3274         if (! good_ptr)
3275                 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3276 }       /* xlog_verify_dest_ptr */
3277
3278 STATIC void
3279 xlog_verify_grant_head(xlog_t *log, int equals)
3280 {
3281     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3282         if (equals)
3283             ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3284         else
3285             ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3286     } else {
3287         ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3288         ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3289     }
3290 }       /* xlog_verify_grant_head */
3291
3292 /* check if it will fit */
3293 STATIC void
3294 xlog_verify_tail_lsn(xlog_t         *log,
3295                      xlog_in_core_t *iclog,
3296                      xfs_lsn_t      tail_lsn)
3297 {
3298     int blocks;
3299
3300     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3301         blocks =
3302             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3303         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3304             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3305     } else {
3306         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3307
3308         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3309             xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3310
3311         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3312         if (blocks < BTOBB(iclog->ic_offset) + 1)
3313             xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3314     }
3315 }       /* xlog_verify_tail_lsn */
3316
3317 /*
3318  * Perform a number of checks on the iclog before writing to disk.
3319  *
3320  * 1. Make sure the iclogs are still circular
3321  * 2. Make sure we have a good magic number
3322  * 3. Make sure we don't have magic numbers in the data
3323  * 4. Check fields of each log operation header for:
3324  *      A. Valid client identifier
3325  *      B. tid ptr value falls in valid ptr space (user space code)
3326  *      C. Length in log record header is correct according to the
3327  *              individual operation headers within record.
3328  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3329  *      log, check the preceding blocks of the physical log to make sure all
3330  *      the cycle numbers agree with the current cycle number.
3331  */
3332 STATIC void
3333 xlog_verify_iclog(xlog_t         *log,
3334                   xlog_in_core_t *iclog,
3335                   int            count,
3336                   boolean_t      syncing)
3337 {
3338         xlog_op_header_t        *ophead;
3339         xlog_in_core_t          *icptr;
3340         xlog_in_core_2_t        *xhdr;
3341         xfs_caddr_t             ptr;
3342         xfs_caddr_t             base_ptr;
3343         __psint_t               field_offset;
3344         __uint8_t               clientid;
3345         int                     len, i, j, k, op_len;
3346         int                     idx;
3347
3348         /* check validity of iclog pointers */
3349         spin_lock(&log->l_icloglock);
3350         icptr = log->l_iclog;
3351         for (i=0; i < log->l_iclog_bufs; i++) {
3352                 if (icptr == NULL)
3353                         xlog_panic("xlog_verify_iclog: invalid ptr");
3354                 icptr = icptr->ic_next;
3355         }
3356         if (icptr != log->l_iclog)
3357                 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3358         spin_unlock(&log->l_icloglock);
3359
3360         /* check log magic numbers */
3361         if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3362                 xlog_panic("xlog_verify_iclog: invalid magic num");
3363
3364         ptr = (xfs_caddr_t) &iclog->ic_header;
3365         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3366              ptr += BBSIZE) {
3367                 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3368                         xlog_panic("xlog_verify_iclog: unexpected magic num");
3369         }
3370
3371         /* check fields */
3372         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3373         ptr = iclog->ic_datap;
3374         base_ptr = ptr;
3375         ophead = (xlog_op_header_t *)ptr;
3376         xhdr = iclog->ic_data;
3377         for (i = 0; i < len; i++) {
3378                 ophead = (xlog_op_header_t *)ptr;
3379
3380                 /* clientid is only 1 byte */
3381                 field_offset = (__psint_t)
3382                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3383                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3384                         clientid = ophead->oh_clientid;
3385                 } else {
3386                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3387                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3388                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3389                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3390                                 clientid = xlog_get_client_id(
3391                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3392                         } else {
3393                                 clientid = xlog_get_client_id(
3394                                         iclog->ic_header.h_cycle_data[idx]);
3395                         }
3396                 }
3397                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3398                         cmn_err(CE_WARN, "xlog_verify_iclog: "
3399                                 "invalid clientid %d op 0x%p offset 0x%lx",
3400                                 clientid, ophead, (unsigned long)field_offset);
3401
3402                 /* check length */
3403                 field_offset = (__psint_t)
3404                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3405                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3406                         op_len = be32_to_cpu(ophead->oh_len);
3407                 } else {
3408                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3409                                     (__psint_t)iclog->ic_datap);
3410                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3411                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3412                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3413                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3414                         } else {
3415                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3416                         }
3417                 }
3418                 ptr += sizeof(xlog_op_header_t) + op_len;
3419         }
3420 }       /* xlog_verify_iclog */
3421 #endif
3422
3423 /*
3424  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3425  */
3426 STATIC int
3427 xlog_state_ioerror(
3428         xlog_t  *log)
3429 {
3430         xlog_in_core_t  *iclog, *ic;
3431
3432         iclog = log->l_iclog;
3433         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3434                 /*
3435                  * Mark all the incore logs IOERROR.
3436                  * From now on, no log flushes will result.
3437                  */
3438                 ic = iclog;
3439                 do {
3440                         ic->ic_state = XLOG_STATE_IOERROR;
3441                         ic = ic->ic_next;
3442                 } while (ic != iclog);
3443                 return 0;
3444         }
3445         /*
3446          * Return non-zero, if state transition has already happened.
3447          */
3448         return 1;
3449 }
3450
3451 /*
3452  * This is called from xfs_force_shutdown, when we're forcibly
3453  * shutting down the filesystem, typically because of an IO error.
3454  * Our main objectives here are to make sure that:
3455  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3456  *         parties to find out, 'atomically'.
3457  *      b. those who're sleeping on log reservations, pinned objects and
3458  *          other resources get woken up, and be told the bad news.
3459  *      c. nothing new gets queued up after (a) and (b) are done.
3460  *      d. if !logerror, flush the iclogs to disk, then seal them off
3461  *         for business.
3462  */
3463 int
3464 xfs_log_force_umount(
3465         struct xfs_mount        *mp,
3466         int                     logerror)
3467 {
3468         xlog_ticket_t   *tic;
3469         xlog_t          *log;
3470         int             retval;
3471
3472         log = mp->m_log;
3473
3474         /*
3475          * If this happens during log recovery, don't worry about
3476          * locking; the log isn't open for business yet.
3477          */
3478         if (!log ||
3479             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3480                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3481                 if (mp->m_sb_bp)
3482                         XFS_BUF_DONE(mp->m_sb_bp);
3483                 return 0;
3484         }
3485
3486         /*
3487          * Somebody could've already done the hard work for us.
3488          * No need to get locks for this.
3489          */
3490         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3491                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3492                 return 1;
3493         }
3494         retval = 0;
3495         /*
3496          * We must hold both the GRANT lock and the LOG lock,
3497          * before we mark the filesystem SHUTDOWN and wake
3498          * everybody up to tell the bad news.
3499          */
3500         spin_lock(&log->l_icloglock);
3501         spin_lock(&log->l_grant_lock);
3502         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3503         if (mp->m_sb_bp)
3504                 XFS_BUF_DONE(mp->m_sb_bp);
3505
3506         /*
3507          * This flag is sort of redundant because of the mount flag, but
3508          * it's good to maintain the separation between the log and the rest
3509          * of XFS.
3510          */
3511         log->l_flags |= XLOG_IO_ERROR;
3512
3513         /*
3514          * If we hit a log error, we want to mark all the iclogs IOERROR
3515          * while we're still holding the loglock.
3516          */
3517         if (logerror)
3518                 retval = xlog_state_ioerror(log);
3519         spin_unlock(&log->l_icloglock);
3520
3521         /*
3522          * We don't want anybody waiting for log reservations
3523          * after this. That means we have to wake up everybody
3524          * queued up on reserve_headq as well as write_headq.
3525          * In addition, we make sure in xlog_{re}grant_log_space
3526          * that we don't enqueue anything once the SHUTDOWN flag
3527          * is set, and this action is protected by the GRANTLOCK.
3528          */
3529         if ((tic = log->l_reserve_headq)) {
3530                 do {
3531                         sv_signal(&tic->t_wait);
3532                         tic = tic->t_next;
3533                 } while (tic != log->l_reserve_headq);
3534         }
3535
3536         if ((tic = log->l_write_headq)) {
3537                 do {
3538                         sv_signal(&tic->t_wait);
3539                         tic = tic->t_next;
3540                 } while (tic != log->l_write_headq);
3541         }
3542         spin_unlock(&log->l_grant_lock);
3543
3544         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3545                 ASSERT(!logerror);
3546                 /*
3547                  * Force the incore logs to disk before shutting the
3548                  * log down completely.
3549                  */
3550                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3551
3552                 spin_lock(&log->l_icloglock);
3553                 retval = xlog_state_ioerror(log);
3554                 spin_unlock(&log->l_icloglock);
3555         }
3556         /*
3557          * Wake up everybody waiting on xfs_log_force.
3558          * Callback all log item committed functions as if the
3559          * log writes were completed.
3560          */
3561         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3562
3563 #ifdef XFSERRORDEBUG
3564         {
3565                 xlog_in_core_t  *iclog;
3566
3567                 spin_lock(&log->l_icloglock);
3568                 iclog = log->l_iclog;
3569                 do {
3570                         ASSERT(iclog->ic_callback == 0);
3571                         iclog = iclog->ic_next;
3572                 } while (iclog != log->l_iclog);
3573                 spin_unlock(&log->l_icloglock);
3574         }
3575 #endif
3576         /* return non-zero if log IOERROR transition had already happened */
3577         return retval;
3578 }
3579
3580 STATIC int
3581 xlog_iclogs_empty(xlog_t *log)
3582 {
3583         xlog_in_core_t  *iclog;
3584
3585         iclog = log->l_iclog;
3586         do {
3587                 /* endianness does not matter here, zero is zero in
3588                  * any language.
3589                  */
3590                 if (iclog->ic_header.h_num_logops)
3591                         return 0;
3592                 iclog = iclog->ic_next;
3593         } while (iclog != log->l_iclog);
3594         return 1;
3595 }