2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
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.
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.
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
20 #include "xfs_types.h"
24 #include "xfs_trans.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"
44 kmem_zone_t *xfs_log_ticket_zone;
46 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
51 /* Local miscellaneous function prototypes */
52 STATIC int xlog_bdstrat_cb(struct xfs_buf *);
53 STATIC int xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
54 xlog_in_core_t **, xfs_lsn_t *);
55 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
56 xfs_buftarg_t *log_target,
57 xfs_daddr_t blk_offset,
59 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
60 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
61 STATIC void xlog_dealloc_log(xlog_t *log);
62 STATIC int xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
63 int nentries, xfs_log_ticket_t tic,
65 xlog_in_core_t **commit_iclog,
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,
73 xlog_in_core_t **iclog,
74 xlog_ticket_t *ticket,
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,
82 STATIC int xlog_state_sync(xlog_t *log,
86 STATIC int xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
87 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
89 /* local functions to manipulate grant head */
90 STATIC int xlog_grant_log_space(xlog_t *log,
92 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
94 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
95 xlog_ticket_t *ticket);
96 STATIC int xlog_regrant_write_log_space(xlog_t *log,
97 xlog_ticket_t *ticket);
98 STATIC void xlog_ungrant_log_space(xlog_t *log,
99 xlog_ticket_t *ticket);
102 /* local ticket functions */
103 STATIC xlog_ticket_t *xlog_ticket_get(xlog_t *log,
108 STATIC void xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
111 STATIC void xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
112 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
113 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
114 int count, boolean_t syncing);
115 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
118 #define xlog_verify_dest_ptr(a,b)
119 #define xlog_verify_grant_head(a,b)
120 #define xlog_verify_iclog(a,b,c,d)
121 #define xlog_verify_tail_lsn(a,b,c)
124 STATIC int xlog_iclogs_empty(xlog_t *log);
126 #if defined(XFS_LOG_TRACE)
128 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
132 if (!log->l_grant_trace) {
133 log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
134 if (!log->l_grant_trace)
137 /* ticket counts are 1 byte each */
138 cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
140 ktrace_enter(log->l_grant_trace,
142 (void *)log->l_reserve_headq,
143 (void *)log->l_write_headq,
144 (void *)((unsigned long)log->l_grant_reserve_cycle),
145 (void *)((unsigned long)log->l_grant_reserve_bytes),
146 (void *)((unsigned long)log->l_grant_write_cycle),
147 (void *)((unsigned long)log->l_grant_write_bytes),
148 (void *)((unsigned long)log->l_curr_cycle),
149 (void *)((unsigned long)log->l_curr_block),
150 (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
151 (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
153 (void *)((unsigned long)tic->t_trans_type),
155 (void *)((unsigned long)tic->t_curr_res),
156 (void *)((unsigned long)tic->t_unit_res));
160 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
162 if (!iclog->ic_trace)
163 iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
164 ktrace_enter(iclog->ic_trace,
165 (void *)((unsigned long)state),
166 (void *)((unsigned long)current_pid()),
167 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
168 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
169 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
170 (void *)NULL, (void *)NULL);
173 #define xlog_trace_loggrant(log,tic,string)
174 #define xlog_trace_iclog(iclog,state)
175 #endif /* XFS_LOG_TRACE */
179 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
183 tic->t_prev = (*qp)->t_prev;
184 (*qp)->t_prev->t_next = tic;
187 tic->t_prev = tic->t_next = tic;
191 tic->t_flags |= XLOG_TIC_IN_Q;
195 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
197 if (tic == tic->t_next) {
201 tic->t_next->t_prev = tic->t_prev;
202 tic->t_prev->t_next = tic->t_next;
205 tic->t_next = tic->t_prev = NULL;
206 tic->t_flags &= ~XLOG_TIC_IN_Q;
210 xlog_grant_sub_space(struct log *log, int bytes)
212 log->l_grant_write_bytes -= bytes;
213 if (log->l_grant_write_bytes < 0) {
214 log->l_grant_write_bytes += log->l_logsize;
215 log->l_grant_write_cycle--;
218 log->l_grant_reserve_bytes -= bytes;
219 if ((log)->l_grant_reserve_bytes < 0) {
220 log->l_grant_reserve_bytes += log->l_logsize;
221 log->l_grant_reserve_cycle--;
227 xlog_grant_add_space_write(struct log *log, int bytes)
229 log->l_grant_write_bytes += bytes;
230 if (log->l_grant_write_bytes > log->l_logsize) {
231 log->l_grant_write_bytes -= log->l_logsize;
232 log->l_grant_write_cycle++;
237 xlog_grant_add_space_reserve(struct log *log, int bytes)
239 log->l_grant_reserve_bytes += bytes;
240 if (log->l_grant_reserve_bytes > log->l_logsize) {
241 log->l_grant_reserve_bytes -= log->l_logsize;
242 log->l_grant_reserve_cycle++;
247 xlog_grant_add_space(struct log *log, int bytes)
249 xlog_grant_add_space_write(log, bytes);
250 xlog_grant_add_space_reserve(log, bytes);
254 xlog_tic_reset_res(xlog_ticket_t *tic)
257 tic->t_res_arr_sum = 0;
258 tic->t_res_num_ophdrs = 0;
262 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
264 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
265 /* add to overflow and start again */
266 tic->t_res_o_flow += tic->t_res_arr_sum;
268 tic->t_res_arr_sum = 0;
271 tic->t_res_arr[tic->t_res_num].r_len = len;
272 tic->t_res_arr[tic->t_res_num].r_type = type;
273 tic->t_res_arr_sum += len;
280 * 1. currblock field gets updated at startup and after in-core logs
281 * marked as with WANT_SYNC.
285 * This routine is called when a user of a log manager ticket is done with
286 * the reservation. If the ticket was ever used, then a commit record for
287 * the associated transaction is written out as a log operation header with
288 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
289 * a given ticket. If the ticket was one with a permanent reservation, then
290 * a few operations are done differently. Permanent reservation tickets by
291 * default don't release the reservation. They just commit the current
292 * transaction with the belief that the reservation is still needed. A flag
293 * must be passed in before permanent reservations are actually released.
294 * When these type of tickets are not released, they need to be set into
295 * the inited state again. By doing this, a start record will be written
296 * out when the next write occurs.
299 xfs_log_done(xfs_mount_t *mp,
300 xfs_log_ticket_t xtic,
304 xlog_t *log = mp->m_log;
305 xlog_ticket_t *ticket = (xfs_log_ticket_t) xtic;
308 if (XLOG_FORCED_SHUTDOWN(log) ||
310 * If nothing was ever written, don't write out commit record.
311 * If we get an error, just continue and give back the log ticket.
313 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
314 (xlog_commit_record(mp, ticket,
315 (xlog_in_core_t **)iclog, &lsn)))) {
316 lsn = (xfs_lsn_t) -1;
317 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
318 flags |= XFS_LOG_REL_PERM_RESERV;
323 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
324 (flags & XFS_LOG_REL_PERM_RESERV)) {
326 * Release ticket if not permanent reservation or a specific
327 * request has been made to release a permanent reservation.
329 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
330 xlog_ungrant_log_space(log, ticket);
331 xlog_ticket_put(log, ticket);
333 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
334 xlog_regrant_reserve_log_space(log, ticket);
337 /* If this ticket was a permanent reservation and we aren't
338 * trying to release it, reset the inited flags; so next time
339 * we write, a start record will be written out.
341 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
342 (flags & XFS_LOG_REL_PERM_RESERV) == 0)
343 ticket->t_flags |= XLOG_TIC_INITED;
350 * Force the in-core log to disk. If flags == XFS_LOG_SYNC,
351 * the force is done synchronously.
353 * Asynchronous forces are implemented by setting the WANT_SYNC
354 * bit in the appropriate in-core log and then returning.
356 * Synchronous forces are implemented with a semaphore. All callers
357 * to force a given lsn to disk will wait on a semaphore attached to the
358 * specific in-core log. When given in-core log finally completes its
359 * write to disk, that thread will wake up all threads waiting on the
369 xlog_t *log = mp->m_log;
373 log_flushed = &dummy;
375 ASSERT(flags & XFS_LOG_FORCE);
377 XFS_STATS_INC(xs_log_force);
379 if (log->l_flags & XLOG_IO_ERROR)
380 return XFS_ERROR(EIO);
382 return xlog_state_sync_all(log, flags, log_flushed);
384 return xlog_state_sync(log, lsn, flags, log_flushed);
385 } /* xfs_log_force */
388 * Attaches a new iclog I/O completion callback routine during
389 * transaction commit. If the log is in error state, a non-zero
390 * return code is handed back and the caller is responsible for
391 * executing the callback at an appropriate time.
394 xfs_log_notify(xfs_mount_t *mp, /* mount of partition */
395 void *iclog_hndl, /* iclog to hang callback off */
396 xfs_log_callback_t *cb)
398 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
401 spin_lock(&iclog->ic_callback_lock);
402 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
404 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
405 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
407 *(iclog->ic_callback_tail) = cb;
408 iclog->ic_callback_tail = &(cb->cb_next);
410 spin_unlock(&iclog->ic_callback_lock);
412 } /* xfs_log_notify */
415 xfs_log_release_iclog(xfs_mount_t *mp,
418 xlog_t *log = mp->m_log;
419 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
421 if (xlog_state_release_iclog(log, iclog)) {
422 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
430 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
431 * to the reservation.
432 * 2. Potentially, push buffers at tail of log to disk.
434 * Each reservation is going to reserve extra space for a log record header.
435 * When writes happen to the on-disk log, we don't subtract the length of the
436 * log record header from any reservation. By wasting space in each
437 * reservation, we prevent over allocation problems.
440 xfs_log_reserve(xfs_mount_t *mp,
443 xfs_log_ticket_t *ticket,
448 xlog_t *log = mp->m_log;
449 xlog_ticket_t *internal_ticket;
452 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
453 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
455 if (XLOG_FORCED_SHUTDOWN(log))
456 return XFS_ERROR(EIO);
458 XFS_STATS_INC(xs_try_logspace);
460 if (*ticket != NULL) {
461 ASSERT(flags & XFS_LOG_PERM_RESERV);
462 internal_ticket = (xlog_ticket_t *)*ticket;
463 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
464 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
465 retval = xlog_regrant_write_log_space(log, internal_ticket);
467 /* may sleep if need to allocate more tickets */
468 internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
470 if (!internal_ticket)
471 return XFS_ERROR(ENOMEM);
472 internal_ticket->t_trans_type = t_type;
473 *ticket = internal_ticket;
474 xlog_trace_loggrant(log, internal_ticket,
475 (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
476 "xfs_log_reserve: create new ticket (permanent trans)" :
477 "xfs_log_reserve: create new ticket");
478 xlog_grant_push_ail(mp,
479 (internal_ticket->t_unit_res *
480 internal_ticket->t_cnt));
481 retval = xlog_grant_log_space(log, internal_ticket);
485 } /* xfs_log_reserve */
489 * Mount a log filesystem
491 * mp - ubiquitous xfs mount point structure
492 * log_target - buftarg of on-disk log device
493 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
494 * num_bblocks - Number of BBSIZE blocks in on-disk log
496 * Return error or zero.
501 xfs_buftarg_t *log_target,
502 xfs_daddr_t blk_offset,
507 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
508 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
511 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
513 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
516 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
519 * Initialize the AIL now we have a log.
521 spin_lock_init(&mp->m_ail_lock);
522 error = xfs_trans_ail_init(mp);
524 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
529 * skip log recovery on a norecovery mount. pretend it all
532 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
533 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
536 mp->m_flags &= ~XFS_MOUNT_RDONLY;
538 error = xlog_recover(mp->m_log);
541 mp->m_flags |= XFS_MOUNT_RDONLY;
543 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
548 /* Normal transactions can now occur */
549 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
551 /* End mounting message in xfs_log_mount_finish */
554 xfs_log_unmount_dealloc(mp);
556 } /* xfs_log_mount */
559 * Finish the recovery of the file system. This is separate from
560 * the xfs_log_mount() call, because it depends on the code in
561 * xfs_mountfs() to read in the root and real-time bitmap inodes
562 * between calling xfs_log_mount() and here.
564 * mp - ubiquitous xfs mount point structure
567 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
571 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
572 error = xlog_recover_finish(mp->m_log, mfsi_flags);
575 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
582 * Unmount processing for the log.
585 xfs_log_unmount(xfs_mount_t *mp)
589 error = xfs_log_unmount_write(mp);
590 xfs_log_unmount_dealloc(mp);
595 * Final log writes as part of unmount.
597 * Mark the filesystem clean as unmount happens. Note that during relocation
598 * this routine needs to be executed as part of source-bag while the
599 * deallocation must not be done until source-end.
603 * Unmount record used to have a string "Unmount filesystem--" in the
604 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
605 * We just write the magic number now since that particular field isn't
606 * currently architecture converted and "nUmount" is a bit foo.
607 * As far as I know, there weren't any dependencies on the old behaviour.
611 xfs_log_unmount_write(xfs_mount_t *mp)
613 xlog_t *log = mp->m_log;
614 xlog_in_core_t *iclog;
616 xlog_in_core_t *first_iclog;
618 xfs_log_iovec_t reg[1];
619 xfs_log_ticket_t tic = NULL;
623 /* the data section must be 32 bit size aligned */
627 __uint32_t pad2; /* may as well make it 64 bits */
628 } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
631 * Don't write out unmount record on read-only mounts.
632 * Or, if we are doing a forced umount (typically because of IO errors).
634 if (mp->m_flags & XFS_MOUNT_RDONLY)
637 xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
640 first_iclog = iclog = log->l_iclog;
642 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
643 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
644 ASSERT(iclog->ic_offset == 0);
646 iclog = iclog->ic_next;
647 } while (iclog != first_iclog);
649 if (! (XLOG_FORCED_SHUTDOWN(log))) {
650 reg[0].i_addr = (void*)&magic;
651 reg[0].i_len = sizeof(magic);
652 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_UNMOUNT);
654 error = xfs_log_reserve(mp, 600, 1, &tic,
655 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
657 /* remove inited flag */
658 ((xlog_ticket_t *)tic)->t_flags = 0;
659 error = xlog_write(mp, reg, 1, tic, &lsn,
660 NULL, XLOG_UNMOUNT_TRANS);
662 * At this point, we're umounting anyway,
663 * so there's no point in transitioning log state
664 * to IOERROR. Just continue...
669 xfs_fs_cmn_err(CE_ALERT, mp,
670 "xfs_log_unmount: unmount record failed");
674 spin_lock(&log->l_icloglock);
675 iclog = log->l_iclog;
676 atomic_inc(&iclog->ic_refcnt);
677 spin_unlock(&log->l_icloglock);
678 xlog_state_want_sync(log, iclog);
679 (void) xlog_state_release_iclog(log, iclog);
681 spin_lock(&log->l_icloglock);
682 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
683 iclog->ic_state == XLOG_STATE_DIRTY)) {
684 if (!XLOG_FORCED_SHUTDOWN(log)) {
685 sv_wait(&iclog->ic_forcesema, PMEM,
686 &log->l_icloglock, s);
688 spin_unlock(&log->l_icloglock);
691 spin_unlock(&log->l_icloglock);
694 xlog_trace_loggrant(log, tic, "unmount rec");
695 xlog_ungrant_log_space(log, tic);
696 xlog_ticket_put(log, tic);
700 * We're already in forced_shutdown mode, couldn't
701 * even attempt to write out the unmount transaction.
703 * Go through the motions of sync'ing and releasing
704 * the iclog, even though no I/O will actually happen,
705 * we need to wait for other log I/Os that may already
706 * be in progress. Do this as a separate section of
707 * code so we'll know if we ever get stuck here that
708 * we're in this odd situation of trying to unmount
709 * a file system that went into forced_shutdown as
710 * the result of an unmount..
712 spin_lock(&log->l_icloglock);
713 iclog = log->l_iclog;
714 atomic_inc(&iclog->ic_refcnt);
715 spin_unlock(&log->l_icloglock);
717 xlog_state_want_sync(log, iclog);
718 (void) xlog_state_release_iclog(log, iclog);
720 spin_lock(&log->l_icloglock);
722 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
723 || iclog->ic_state == XLOG_STATE_DIRTY
724 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
726 sv_wait(&iclog->ic_forcesema, PMEM,
727 &log->l_icloglock, s);
729 spin_unlock(&log->l_icloglock);
734 } /* xfs_log_unmount_write */
737 * Deallocate log structures for unmount/relocation.
739 * We need to stop the aild from running before we destroy
740 * and deallocate the log as the aild references the log.
743 xfs_log_unmount_dealloc(xfs_mount_t *mp)
745 xfs_trans_ail_destroy(mp);
746 xlog_dealloc_log(mp->m_log);
750 * Write region vectors to log. The write happens using the space reservation
751 * of the ticket (tic). It is not a requirement that all writes for a given
752 * transaction occur with one call to xfs_log_write().
755 xfs_log_write(xfs_mount_t * mp,
756 xfs_log_iovec_t reg[],
758 xfs_log_ticket_t tic,
759 xfs_lsn_t *start_lsn)
762 xlog_t *log = mp->m_log;
764 if (XLOG_FORCED_SHUTDOWN(log))
765 return XFS_ERROR(EIO);
767 if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
768 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
771 } /* xfs_log_write */
775 xfs_log_move_tail(xfs_mount_t *mp,
779 xlog_t *log = mp->m_log;
780 int need_bytes, free_bytes, cycle, bytes;
782 if (XLOG_FORCED_SHUTDOWN(log))
786 /* needed since sync_lsn is 64 bits */
787 spin_lock(&log->l_icloglock);
788 tail_lsn = log->l_last_sync_lsn;
789 spin_unlock(&log->l_icloglock);
792 spin_lock(&log->l_grant_lock);
794 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
798 log->l_tail_lsn = tail_lsn;
801 if ((tic = log->l_write_headq)) {
803 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
804 panic("Recovery problem");
806 cycle = log->l_grant_write_cycle;
807 bytes = log->l_grant_write_bytes;
808 free_bytes = xlog_space_left(log, cycle, bytes);
810 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
812 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
815 free_bytes -= tic->t_unit_res;
816 sv_signal(&tic->t_sema);
818 } while (tic != log->l_write_headq);
820 if ((tic = log->l_reserve_headq)) {
822 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
823 panic("Recovery problem");
825 cycle = log->l_grant_reserve_cycle;
826 bytes = log->l_grant_reserve_bytes;
827 free_bytes = xlog_space_left(log, cycle, bytes);
829 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
830 need_bytes = tic->t_unit_res*tic->t_cnt;
832 need_bytes = tic->t_unit_res;
833 if (free_bytes < need_bytes && tail_lsn != 1)
836 free_bytes -= need_bytes;
837 sv_signal(&tic->t_sema);
839 } while (tic != log->l_reserve_headq);
841 spin_unlock(&log->l_grant_lock);
842 } /* xfs_log_move_tail */
845 * Determine if we have a transaction that has gone to disk
846 * that needs to be covered. Log activity needs to be idle (no AIL and
847 * nothing in the iclogs). And, we need to be in the right state indicating
848 * something has gone out.
851 xfs_log_need_covered(xfs_mount_t *mp)
854 xlog_t *log = mp->m_log;
856 if (!xfs_fs_writable(mp))
859 spin_lock(&log->l_icloglock);
860 if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
861 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
862 && !xfs_trans_first_ail(mp, &gen)
863 && xlog_iclogs_empty(log)) {
864 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
865 log->l_covered_state = XLOG_STATE_COVER_DONE;
867 ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
868 log->l_covered_state = XLOG_STATE_COVER_DONE2;
872 spin_unlock(&log->l_icloglock);
876 /******************************************************************************
880 ******************************************************************************
883 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
884 * The log manager must keep track of the last LR which was committed
885 * to disk. The lsn of this LR will become the new tail_lsn whenever
886 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
887 * the situation where stuff could be written into the log but nothing
888 * was ever in the AIL when asked. Eventually, we panic since the
889 * tail hits the head.
891 * We may be holding the log iclog lock upon entering this routine.
894 xlog_assign_tail_lsn(xfs_mount_t *mp)
897 xlog_t *log = mp->m_log;
899 tail_lsn = xfs_trans_tail_ail(mp);
900 spin_lock(&log->l_grant_lock);
902 log->l_tail_lsn = tail_lsn;
904 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
906 spin_unlock(&log->l_grant_lock);
909 } /* xlog_assign_tail_lsn */
913 * Return the space in the log between the tail and the head. The head
914 * is passed in the cycle/bytes formal parms. In the special case where
915 * the reserve head has wrapped passed the tail, this calculation is no
916 * longer valid. In this case, just return 0 which means there is no space
917 * in the log. This works for all places where this function is called
918 * with the reserve head. Of course, if the write head were to ever
919 * wrap the tail, we should blow up. Rather than catch this case here,
920 * we depend on other ASSERTions in other parts of the code. XXXmiken
922 * This code also handles the case where the reservation head is behind
923 * the tail. The details of this case are described below, but the end
924 * result is that we return the size of the log as the amount of space left.
927 xlog_space_left(xlog_t *log, int cycle, int bytes)
933 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
934 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
935 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
936 free_bytes = log->l_logsize - (bytes - tail_bytes);
937 } else if ((tail_cycle + 1) < cycle) {
939 } else if (tail_cycle < cycle) {
940 ASSERT(tail_cycle == (cycle - 1));
941 free_bytes = tail_bytes - bytes;
944 * The reservation head is behind the tail.
945 * In this case we just want to return the size of the
946 * log as the amount of space left.
948 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
949 "xlog_space_left: head behind tail\n"
950 " tail_cycle = %d, tail_bytes = %d\n"
951 " GH cycle = %d, GH bytes = %d",
952 tail_cycle, tail_bytes, cycle, bytes);
954 free_bytes = log->l_logsize;
957 } /* xlog_space_left */
961 * Log function which is called when an io completes.
963 * The log manager needs its own routine, in order to control what
964 * happens with the buffer after the write completes.
967 xlog_iodone(xfs_buf_t *bp)
969 xlog_in_core_t *iclog;
973 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
974 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
975 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
979 * Some versions of cpp barf on the recursive definition of
980 * ic_log -> hic_fields.ic_log and expand ic_log twice when
981 * it is passed through two macros. Workaround broken cpp.
986 * If the ordered flag has been removed by a lower
987 * layer, it means the underlyin device no longer supports
988 * barrier I/O. Warn loudly and turn off barriers.
990 if ((l->l_mp->m_flags & XFS_MOUNT_BARRIER) && !XFS_BUF_ORDERED(bp)) {
991 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
992 xfs_fs_cmn_err(CE_WARN, l->l_mp,
993 "xlog_iodone: Barriers are no longer supported"
994 " by device. Disabling barriers\n");
995 xfs_buftrace("XLOG_IODONE BARRIERS OFF", bp);
999 * Race to shutdown the filesystem if we see an error.
1001 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
1002 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1003 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
1005 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1007 * This flag will be propagated to the trans-committed
1008 * callback routines to let them know that the log-commit
1011 aborted = XFS_LI_ABORTED;
1012 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1013 aborted = XFS_LI_ABORTED;
1016 /* log I/O is always issued ASYNC */
1017 ASSERT(XFS_BUF_ISASYNC(bp));
1018 xlog_state_done_syncing(iclog, aborted);
1020 * do not reference the buffer (bp) here as we could race
1021 * with it being freed after writing the unmount record to the
1028 * The bdstrat callback function for log bufs. This gives us a central
1029 * place to trap bufs in case we get hit by a log I/O error and need to
1030 * shutdown. Actually, in practice, even when we didn't get a log error,
1031 * we transition the iclogs to IOERROR state *after* flushing all existing
1032 * iclogs to disk. This is because we don't want anymore new transactions to be
1033 * started or completed afterwards.
1036 xlog_bdstrat_cb(struct xfs_buf *bp)
1038 xlog_in_core_t *iclog;
1040 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1042 if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
1043 /* note for irix bstrat will need struct bdevsw passed
1044 * Fix the following macro if the code ever is merged
1050 xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1051 XFS_BUF_ERROR(bp, EIO);
1054 return XFS_ERROR(EIO);
1060 * Return size of each in-core log record buffer.
1062 * All machines get 8 x 32KB buffers by default, unless tuned otherwise.
1064 * If the filesystem blocksize is too large, we may need to choose a
1065 * larger size since the directory code currently logs entire blocks.
1069 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1075 if (mp->m_logbufs <= 0)
1076 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1078 log->l_iclog_bufs = mp->m_logbufs;
1081 * Buffer size passed in from mount system call.
1083 if (mp->m_logbsize > 0) {
1084 size = log->l_iclog_size = mp->m_logbsize;
1085 log->l_iclog_size_log = 0;
1087 log->l_iclog_size_log++;
1091 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1092 /* # headers = size / 32K
1093 * one header holds cycles from 32K of data
1096 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1097 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1099 log->l_iclog_hsize = xhdrs << BBSHIFT;
1100 log->l_iclog_heads = xhdrs;
1102 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1103 log->l_iclog_hsize = BBSIZE;
1104 log->l_iclog_heads = 1;
1109 /* All machines use 32KB buffers by default. */
1110 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1111 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1113 /* the default log size is 16k or 32k which is one header sector */
1114 log->l_iclog_hsize = BBSIZE;
1115 log->l_iclog_heads = 1;
1118 * For 16KB, we use 3 32KB buffers. For 32KB block sizes, we use
1119 * 4 32KB buffers. For 64KB block sizes, we use 8 32KB buffers.
1121 if (mp->m_sb.sb_blocksize >= 16*1024) {
1122 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1123 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1124 if (mp->m_logbufs <= 0) {
1125 switch (mp->m_sb.sb_blocksize) {
1126 case 16*1024: /* 16 KB */
1127 log->l_iclog_bufs = 3;
1129 case 32*1024: /* 32 KB */
1130 log->l_iclog_bufs = 4;
1132 case 64*1024: /* 64 KB */
1133 log->l_iclog_bufs = 8;
1136 xlog_panic("XFS: Invalid blocksize");
1142 done: /* are we being asked to make the sizes selected above visible? */
1143 if (mp->m_logbufs == 0)
1144 mp->m_logbufs = log->l_iclog_bufs;
1145 if (mp->m_logbsize == 0)
1146 mp->m_logbsize = log->l_iclog_size;
1147 } /* xlog_get_iclog_buffer_size */
1151 * This routine initializes some of the log structure for a given mount point.
1152 * Its primary purpose is to fill in enough, so recovery can occur. However,
1153 * some other stuff may be filled in too.
1156 xlog_alloc_log(xfs_mount_t *mp,
1157 xfs_buftarg_t *log_target,
1158 xfs_daddr_t blk_offset,
1162 xlog_rec_header_t *head;
1163 xlog_in_core_t **iclogp;
1164 xlog_in_core_t *iclog, *prev_iclog=NULL;
1169 log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1172 log->l_targ = log_target;
1173 log->l_logsize = BBTOB(num_bblks);
1174 log->l_logBBstart = blk_offset;
1175 log->l_logBBsize = num_bblks;
1176 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1177 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1179 log->l_prev_block = -1;
1180 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1181 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1182 log->l_last_sync_lsn = log->l_tail_lsn;
1183 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1184 log->l_grant_reserve_cycle = 1;
1185 log->l_grant_write_cycle = 1;
1187 if (xfs_sb_version_hassector(&mp->m_sb)) {
1188 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1189 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1190 /* for larger sector sizes, must have v2 or external log */
1191 ASSERT(log->l_sectbb_log == 0 ||
1192 log->l_logBBstart == 0 ||
1193 xfs_sb_version_haslogv2(&mp->m_sb));
1194 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1196 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1198 xlog_get_iclog_buffer_size(mp, log);
1200 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1201 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1202 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1203 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1204 ASSERT(XFS_BUF_ISBUSY(bp));
1205 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1208 spin_lock_init(&log->l_icloglock);
1209 spin_lock_init(&log->l_grant_lock);
1210 initnsema(&log->l_flushsema, 0, "ic-flush");
1212 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1213 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1215 iclogp = &log->l_iclog;
1217 * The amount of memory to allocate for the iclog structure is
1218 * rather funky due to the way the structure is defined. It is
1219 * done this way so that we can use different sizes for machines
1220 * with different amounts of memory. See the definition of
1221 * xlog_in_core_t in xfs_log_priv.h for details.
1223 iclogsize = log->l_iclog_size;
1224 ASSERT(log->l_iclog_size >= 4096);
1225 for (i=0; i < log->l_iclog_bufs; i++) {
1226 *iclogp = (xlog_in_core_t *)
1227 kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1229 iclog->ic_prev = prev_iclog;
1232 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1233 if (!XFS_BUF_CPSEMA(bp))
1235 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1236 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1237 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1239 iclog->hic_data = bp->b_addr;
1241 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1243 head = &iclog->ic_header;
1244 memset(head, 0, sizeof(xlog_rec_header_t));
1245 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1246 head->h_version = cpu_to_be32(
1247 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1248 head->h_size = cpu_to_be32(log->l_iclog_size);
1250 head->h_fmt = cpu_to_be32(XLOG_FMT);
1251 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1253 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1254 iclog->ic_state = XLOG_STATE_ACTIVE;
1255 iclog->ic_log = log;
1256 atomic_set(&iclog->ic_refcnt, 0);
1257 spin_lock_init(&iclog->ic_callback_lock);
1258 iclog->ic_callback_tail = &(iclog->ic_callback);
1259 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1261 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1262 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1263 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1264 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1266 iclogp = &iclog->ic_next;
1268 *iclogp = log->l_iclog; /* complete ring */
1269 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1272 } /* xlog_alloc_log */
1276 * Write out the commit record of a transaction associated with the given
1277 * ticket. Return the lsn of the commit record.
1280 xlog_commit_record(xfs_mount_t *mp,
1281 xlog_ticket_t *ticket,
1282 xlog_in_core_t **iclog,
1283 xfs_lsn_t *commitlsnp)
1286 xfs_log_iovec_t reg[1];
1288 reg[0].i_addr = NULL;
1290 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_COMMIT);
1292 ASSERT_ALWAYS(iclog);
1293 if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1294 iclog, XLOG_COMMIT_TRANS))) {
1295 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1298 } /* xlog_commit_record */
1302 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1303 * log space. This code pushes on the lsn which would supposedly free up
1304 * the 25% which we want to leave free. We may need to adopt a policy which
1305 * pushes on an lsn which is further along in the log once we reach the high
1306 * water mark. In this manner, we would be creating a low water mark.
1309 xlog_grant_push_ail(xfs_mount_t *mp,
1312 xlog_t *log = mp->m_log; /* pointer to the log */
1313 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1314 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1315 int free_blocks; /* free blocks left to write to */
1316 int free_bytes; /* free bytes left to write to */
1317 int threshold_block; /* block in lsn we'd like to be at */
1318 int threshold_cycle; /* lsn cycle we'd like to be at */
1321 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1323 spin_lock(&log->l_grant_lock);
1324 free_bytes = xlog_space_left(log,
1325 log->l_grant_reserve_cycle,
1326 log->l_grant_reserve_bytes);
1327 tail_lsn = log->l_tail_lsn;
1328 free_blocks = BTOBBT(free_bytes);
1331 * Set the threshold for the minimum number of free blocks in the
1332 * log to the maximum of what the caller needs, one quarter of the
1333 * log, and 256 blocks.
1335 free_threshold = BTOBB(need_bytes);
1336 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1337 free_threshold = MAX(free_threshold, 256);
1338 if (free_blocks < free_threshold) {
1339 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1340 threshold_cycle = CYCLE_LSN(tail_lsn);
1341 if (threshold_block >= log->l_logBBsize) {
1342 threshold_block -= log->l_logBBsize;
1343 threshold_cycle += 1;
1345 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1347 /* Don't pass in an lsn greater than the lsn of the last
1348 * log record known to be on disk.
1350 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1351 threshold_lsn = log->l_last_sync_lsn;
1353 spin_unlock(&log->l_grant_lock);
1356 * Get the transaction layer to kick the dirty buffers out to
1357 * disk asynchronously. No point in trying to do this if
1358 * the filesystem is shutting down.
1360 if (threshold_lsn &&
1361 !XLOG_FORCED_SHUTDOWN(log))
1362 xfs_trans_push_ail(mp, threshold_lsn);
1363 } /* xlog_grant_push_ail */
1367 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1368 * fashion. Previously, we should have moved the current iclog
1369 * ptr in the log to point to the next available iclog. This allows further
1370 * write to continue while this code syncs out an iclog ready to go.
1371 * Before an in-core log can be written out, the data section must be scanned
1372 * to save away the 1st word of each BBSIZE block into the header. We replace
1373 * it with the current cycle count. Each BBSIZE block is tagged with the
1374 * cycle count because there in an implicit assumption that drives will
1375 * guarantee that entire 512 byte blocks get written at once. In other words,
1376 * we can't have part of a 512 byte block written and part not written. By
1377 * tagging each block, we will know which blocks are valid when recovering
1378 * after an unclean shutdown.
1380 * This routine is single threaded on the iclog. No other thread can be in
1381 * this routine with the same iclog. Changing contents of iclog can there-
1382 * fore be done without grabbing the state machine lock. Updating the global
1383 * log will require grabbing the lock though.
1385 * The entire log manager uses a logical block numbering scheme. Only
1386 * log_sync (and then only bwrite()) know about the fact that the log may
1387 * not start with block zero on a given device. The log block start offset
1388 * is added immediately before calling bwrite().
1392 xlog_sync(xlog_t *log,
1393 xlog_in_core_t *iclog)
1395 xfs_caddr_t dptr; /* pointer to byte sized element */
1398 uint count; /* byte count of bwrite */
1399 uint count_init; /* initial count before roundup */
1400 int roundoff; /* roundoff to BB or stripe */
1401 int split = 0; /* split write into two regions */
1403 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1405 XFS_STATS_INC(xs_log_writes);
1406 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1408 /* Add for LR header */
1409 count_init = log->l_iclog_hsize + iclog->ic_offset;
1411 /* Round out the log write size */
1412 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1413 /* we have a v2 stripe unit to use */
1414 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1416 count = BBTOB(BTOBB(count_init));
1418 roundoff = count - count_init;
1419 ASSERT(roundoff >= 0);
1420 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1421 roundoff < log->l_mp->m_sb.sb_logsunit)
1423 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1424 roundoff < BBTOB(1)));
1426 /* move grant heads by roundoff in sync */
1427 spin_lock(&log->l_grant_lock);
1428 xlog_grant_add_space(log, roundoff);
1429 spin_unlock(&log->l_grant_lock);
1431 /* put cycle number in every block */
1432 xlog_pack_data(log, iclog, roundoff);
1434 /* real byte length */
1436 iclog->ic_header.h_len =
1437 cpu_to_be32(iclog->ic_offset + roundoff);
1439 iclog->ic_header.h_len =
1440 cpu_to_be32(iclog->ic_offset);
1444 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1445 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1446 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1448 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1450 /* Do we need to split this write into 2 parts? */
1451 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1452 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1453 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1454 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1456 iclog->ic_bwritecnt = 1;
1458 XFS_BUF_SET_COUNT(bp, count);
1459 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1460 XFS_BUF_ZEROFLAGS(bp);
1464 * Do an ordered write for the log block.
1465 * Its unnecessary to flush the first split block in the log wrap case.
1467 if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1468 XFS_BUF_ORDERED(bp);
1470 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1471 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1473 xlog_verify_iclog(log, iclog, count, B_TRUE);
1475 /* account for log which doesn't start at block #0 */
1476 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1478 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1483 if ((error = XFS_bwrite(bp))) {
1484 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1489 bp = iclog->ic_log->l_xbuf;
1490 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1492 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1493 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1494 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1495 (__psint_t)count), split);
1496 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1497 XFS_BUF_ZEROFLAGS(bp);
1500 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1501 XFS_BUF_ORDERED(bp);
1502 dptr = XFS_BUF_PTR(bp);
1504 * Bump the cycle numbers at the start of each block
1505 * since this part of the buffer is at the start of
1506 * a new cycle. Watch out for the header magic number
1509 for (i = 0; i < split; i += BBSIZE) {
1510 be32_add_cpu((__be32 *)dptr, 1);
1511 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1512 be32_add_cpu((__be32 *)dptr, 1);
1516 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1517 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1519 /* account for internal log which doesn't start at block #0 */
1520 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1522 if ((error = XFS_bwrite(bp))) {
1523 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1524 bp, XFS_BUF_ADDR(bp));
1533 * Deallocate a log structure
1536 xlog_dealloc_log(xlog_t *log)
1538 xlog_in_core_t *iclog, *next_iclog;
1541 iclog = log->l_iclog;
1542 for (i=0; i<log->l_iclog_bufs; i++) {
1543 sv_destroy(&iclog->ic_forcesema);
1544 sv_destroy(&iclog->ic_writesema);
1545 xfs_buf_free(iclog->ic_bp);
1546 #ifdef XFS_LOG_TRACE
1547 if (iclog->ic_trace != NULL) {
1548 ktrace_free(iclog->ic_trace);
1551 next_iclog = iclog->ic_next;
1552 kmem_free(iclog, sizeof(xlog_in_core_t));
1555 freesema(&log->l_flushsema);
1556 spinlock_destroy(&log->l_icloglock);
1557 spinlock_destroy(&log->l_grant_lock);
1559 xfs_buf_free(log->l_xbuf);
1560 #ifdef XFS_LOG_TRACE
1561 if (log->l_trace != NULL) {
1562 ktrace_free(log->l_trace);
1564 if (log->l_grant_trace != NULL) {
1565 ktrace_free(log->l_grant_trace);
1568 log->l_mp->m_log = NULL;
1569 kmem_free(log, sizeof(xlog_t));
1570 } /* xlog_dealloc_log */
1573 * Update counters atomically now that memcpy is done.
1577 xlog_state_finish_copy(xlog_t *log,
1578 xlog_in_core_t *iclog,
1582 spin_lock(&log->l_icloglock);
1584 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1585 iclog->ic_offset += copy_bytes;
1587 spin_unlock(&log->l_icloglock);
1588 } /* xlog_state_finish_copy */
1594 * print out info relating to regions written which consume
1598 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1601 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1603 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1604 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1625 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1668 xfs_fs_cmn_err(CE_WARN, mp,
1669 "xfs_log_write: reservation summary:\n"
1670 " trans type = %s (%u)\n"
1671 " unit res = %d bytes\n"
1672 " current res = %d bytes\n"
1673 " total reg = %u bytes (o/flow = %u bytes)\n"
1674 " ophdrs = %u (ophdr space = %u bytes)\n"
1675 " ophdr + reg = %u bytes\n"
1676 " num regions = %u\n",
1677 ((ticket->t_trans_type <= 0 ||
1678 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1679 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1680 ticket->t_trans_type,
1683 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1684 ticket->t_res_num_ophdrs, ophdr_spc,
1685 ticket->t_res_arr_sum +
1686 ticket->t_res_o_flow + ophdr_spc,
1689 for (i = 0; i < ticket->t_res_num; i++) {
1690 uint r_type = ticket->t_res_arr[i].r_type;
1692 "region[%u]: %s - %u bytes\n",
1694 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1695 "bad-rtype" : res_type_str[r_type-1]),
1696 ticket->t_res_arr[i].r_len);
1701 * Write some region out to in-core log
1703 * This will be called when writing externally provided regions or when
1704 * writing out a commit record for a given transaction.
1706 * General algorithm:
1707 * 1. Find total length of this write. This may include adding to the
1708 * lengths passed in.
1709 * 2. Check whether we violate the tickets reservation.
1710 * 3. While writing to this iclog
1711 * A. Reserve as much space in this iclog as can get
1712 * B. If this is first write, save away start lsn
1713 * C. While writing this region:
1714 * 1. If first write of transaction, write start record
1715 * 2. Write log operation header (header per region)
1716 * 3. Find out if we can fit entire region into this iclog
1717 * 4. Potentially, verify destination memcpy ptr
1718 * 5. Memcpy (partial) region
1719 * 6. If partial copy, release iclog; otherwise, continue
1720 * copying more regions into current iclog
1721 * 4. Mark want sync bit (in simulation mode)
1722 * 5. Release iclog for potential flush to on-disk log.
1725 * 1. Panic if reservation is overrun. This should never happen since
1726 * reservation amounts are generated internal to the filesystem.
1728 * 1. Tickets are single threaded data structures.
1729 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1730 * syncing routine. When a single log_write region needs to span
1731 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1732 * on all log operation writes which don't contain the end of the
1733 * region. The XLOG_END_TRANS bit is used for the in-core log
1734 * operation which contains the end of the continued log_write region.
1735 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1736 * we don't really know exactly how much space will be used. As a result,
1737 * we don't update ic_offset until the end when we know exactly how many
1738 * bytes have been written out.
1741 xlog_write(xfs_mount_t * mp,
1742 xfs_log_iovec_t reg[],
1744 xfs_log_ticket_t tic,
1745 xfs_lsn_t *start_lsn,
1746 xlog_in_core_t **commit_iclog,
1749 xlog_t *log = mp->m_log;
1750 xlog_ticket_t *ticket = (xlog_ticket_t *)tic;
1751 xlog_in_core_t *iclog = NULL; /* ptr to current in-core log */
1752 xlog_op_header_t *logop_head; /* ptr to log operation header */
1753 __psint_t ptr; /* copy address into data region */
1754 int len; /* # xlog_write() bytes 2 still copy */
1755 int index; /* region index currently copying */
1756 int log_offset; /* offset (from 0) into data region */
1757 int start_rec_copy; /* # bytes to copy for start record */
1758 int partial_copy; /* did we split a region? */
1759 int partial_copy_len;/* # bytes copied if split region */
1760 int need_copy; /* # bytes need to memcpy this region */
1761 int copy_len; /* # bytes actually memcpy'ing */
1762 int copy_off; /* # bytes from entry start */
1763 int contwr; /* continued write of in-core log? */
1765 int record_cnt = 0, data_cnt = 0;
1767 partial_copy_len = partial_copy = 0;
1769 /* Calculate potential maximum space. Each region gets its own
1770 * xlog_op_header_t and may need to be double word aligned.
1773 if (ticket->t_flags & XLOG_TIC_INITED) { /* acct for start rec of xact */
1774 len += sizeof(xlog_op_header_t);
1775 ticket->t_res_num_ophdrs++;
1778 for (index = 0; index < nentries; index++) {
1779 len += sizeof(xlog_op_header_t); /* each region gets >= 1 */
1780 ticket->t_res_num_ophdrs++;
1781 len += reg[index].i_len;
1782 xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1784 contwr = *start_lsn = 0;
1786 if (ticket->t_curr_res < len) {
1787 xlog_print_tic_res(mp, ticket);
1790 "xfs_log_write: reservation ran out. Need to up reservation");
1792 /* Customer configurable panic */
1793 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1794 "xfs_log_write: reservation ran out. Need to up reservation");
1795 /* If we did not panic, shutdown the filesystem */
1796 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1799 ticket->t_curr_res -= len;
1801 for (index = 0; index < nentries; ) {
1802 if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1803 &contwr, &log_offset)))
1806 ASSERT(log_offset <= iclog->ic_size - 1);
1807 ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1809 /* start_lsn is the first lsn written to. That's all we need. */
1811 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1813 /* This loop writes out as many regions as can fit in the amount
1814 * of space which was allocated by xlog_state_get_iclog_space().
1816 while (index < nentries) {
1817 ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1818 ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1821 /* If first write for transaction, insert start record.
1822 * We can't be trying to commit if we are inited. We can't
1823 * have any "partial_copy" if we are inited.
1825 if (ticket->t_flags & XLOG_TIC_INITED) {
1826 logop_head = (xlog_op_header_t *)ptr;
1827 logop_head->oh_tid = cpu_to_be32(ticket->t_tid);
1828 logop_head->oh_clientid = ticket->t_clientid;
1829 logop_head->oh_len = 0;
1830 logop_head->oh_flags = XLOG_START_TRANS;
1831 logop_head->oh_res2 = 0;
1832 ticket->t_flags &= ~XLOG_TIC_INITED; /* clear bit */
1835 start_rec_copy = sizeof(xlog_op_header_t);
1836 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1839 /* Copy log operation header directly into data section */
1840 logop_head = (xlog_op_header_t *)ptr;
1841 logop_head->oh_tid = cpu_to_be32(ticket->t_tid);
1842 logop_head->oh_clientid = ticket->t_clientid;
1843 logop_head->oh_res2 = 0;
1845 /* header copied directly */
1846 xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1848 /* are we copying a commit or unmount record? */
1849 logop_head->oh_flags = flags;
1852 * We've seen logs corrupted with bad transaction client
1853 * ids. This makes sure that XFS doesn't generate them on.
1854 * Turn this into an EIO and shut down the filesystem.
1856 switch (logop_head->oh_clientid) {
1857 case XFS_TRANSACTION:
1862 xfs_fs_cmn_err(CE_WARN, mp,
1863 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1864 logop_head->oh_clientid, tic);
1865 return XFS_ERROR(EIO);
1868 /* Partial write last time? => (partial_copy != 0)
1869 * need_copy is the amount we'd like to copy if everything could
1870 * fit in the current memcpy.
1872 need_copy = reg[index].i_len - partial_copy_len;
1874 copy_off = partial_copy_len;
1875 if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1876 copy_len = need_copy;
1877 logop_head->oh_len = cpu_to_be32(copy_len);
1879 logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1880 partial_copy_len = partial_copy = 0;
1881 } else { /* partial write */
1882 copy_len = iclog->ic_size - log_offset;
1883 logop_head->oh_len = cpu_to_be32(copy_len);
1884 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1886 logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1887 partial_copy_len += copy_len;
1889 len += sizeof(xlog_op_header_t); /* from splitting of region */
1890 /* account for new log op header */
1891 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1892 ticket->t_res_num_ophdrs++;
1894 xlog_verify_dest_ptr(log, ptr);
1897 ASSERT(copy_len >= 0);
1898 memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1899 xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1901 /* make copy_len total bytes copied, including headers */
1902 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1904 data_cnt += contwr ? copy_len : 0;
1905 if (partial_copy) { /* copied partial region */
1906 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1907 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1908 record_cnt = data_cnt = 0;
1909 if ((error = xlog_state_release_iclog(log, iclog)))
1911 break; /* don't increment index */
1912 } else { /* copied entire region */
1914 partial_copy_len = partial_copy = 0;
1916 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1917 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1918 record_cnt = data_cnt = 0;
1919 xlog_state_want_sync(log, iclog);
1921 ASSERT(flags & XLOG_COMMIT_TRANS);
1922 *commit_iclog = iclog;
1923 } else if ((error = xlog_state_release_iclog(log, iclog)))
1925 if (index == nentries)
1926 return 0; /* we are done */
1930 } /* if (partial_copy) */
1931 } /* while (index < nentries) */
1932 } /* for (index = 0; index < nentries; ) */
1935 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1937 ASSERT(flags & XLOG_COMMIT_TRANS);
1938 *commit_iclog = iclog;
1941 return xlog_state_release_iclog(log, iclog);
1945 /*****************************************************************************
1947 * State Machine functions
1949 *****************************************************************************
1952 /* Clean iclogs starting from the head. This ordering must be
1953 * maintained, so an iclog doesn't become ACTIVE beyond one that
1954 * is SYNCING. This is also required to maintain the notion that we use
1955 * a counting semaphore to hold off would be writers to the log when every
1956 * iclog is trying to sync to disk.
1958 * State Change: DIRTY -> ACTIVE
1961 xlog_state_clean_log(xlog_t *log)
1963 xlog_in_core_t *iclog;
1966 iclog = log->l_iclog;
1968 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1969 iclog->ic_state = XLOG_STATE_ACTIVE;
1970 iclog->ic_offset = 0;
1971 ASSERT(iclog->ic_callback == NULL);
1973 * If the number of ops in this iclog indicate it just
1974 * contains the dummy transaction, we can
1975 * change state into IDLE (the second time around).
1976 * Otherwise we should change the state into
1978 * We don't need to cover the dummy.
1981 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
1986 * We have two dirty iclogs so start over
1987 * This could also be num of ops indicates
1988 * this is not the dummy going out.
1992 iclog->ic_header.h_num_logops = 0;
1993 memset(iclog->ic_header.h_cycle_data, 0,
1994 sizeof(iclog->ic_header.h_cycle_data));
1995 iclog->ic_header.h_lsn = 0;
1996 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1999 break; /* stop cleaning */
2000 iclog = iclog->ic_next;
2001 } while (iclog != log->l_iclog);
2003 /* log is locked when we are called */
2005 * Change state for the dummy log recording.
2006 * We usually go to NEED. But we go to NEED2 if the changed indicates
2007 * we are done writing the dummy record.
2008 * If we are done with the second dummy recored (DONE2), then
2012 switch (log->l_covered_state) {
2013 case XLOG_STATE_COVER_IDLE:
2014 case XLOG_STATE_COVER_NEED:
2015 case XLOG_STATE_COVER_NEED2:
2016 log->l_covered_state = XLOG_STATE_COVER_NEED;
2019 case XLOG_STATE_COVER_DONE:
2021 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2023 log->l_covered_state = XLOG_STATE_COVER_NEED;
2026 case XLOG_STATE_COVER_DONE2:
2028 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2030 log->l_covered_state = XLOG_STATE_COVER_NEED;
2037 } /* xlog_state_clean_log */
2040 xlog_get_lowest_lsn(
2043 xlog_in_core_t *lsn_log;
2044 xfs_lsn_t lowest_lsn, lsn;
2046 lsn_log = log->l_iclog;
2049 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2050 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2051 if ((lsn && !lowest_lsn) ||
2052 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2056 lsn_log = lsn_log->ic_next;
2057 } while (lsn_log != log->l_iclog);
2063 xlog_state_do_callback(
2066 xlog_in_core_t *ciclog)
2068 xlog_in_core_t *iclog;
2069 xlog_in_core_t *first_iclog; /* used to know when we've
2070 * processed all iclogs once */
2071 xfs_log_callback_t *cb, *cb_next;
2073 xfs_lsn_t lowest_lsn;
2074 int ioerrors; /* counter: iclogs with errors */
2075 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2076 int funcdidcallbacks; /* flag: function did callbacks */
2077 int repeats; /* for issuing console warnings if
2078 * looping too many times */
2080 spin_lock(&log->l_icloglock);
2081 first_iclog = iclog = log->l_iclog;
2083 funcdidcallbacks = 0;
2088 * Scan all iclogs starting with the one pointed to by the
2089 * log. Reset this starting point each time the log is
2090 * unlocked (during callbacks).
2092 * Keep looping through iclogs until one full pass is made
2093 * without running any callbacks.
2095 first_iclog = log->l_iclog;
2096 iclog = log->l_iclog;
2097 loopdidcallbacks = 0;
2102 /* skip all iclogs in the ACTIVE & DIRTY states */
2103 if (iclog->ic_state &
2104 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2105 iclog = iclog->ic_next;
2110 * Between marking a filesystem SHUTDOWN and stopping
2111 * the log, we do flush all iclogs to disk (if there
2112 * wasn't a log I/O error). So, we do want things to
2113 * go smoothly in case of just a SHUTDOWN w/o a
2116 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2118 * Can only perform callbacks in order. Since
2119 * this iclog is not in the DONE_SYNC/
2120 * DO_CALLBACK state, we skip the rest and
2121 * just try to clean up. If we set our iclog
2122 * to DO_CALLBACK, we will not process it when
2123 * we retry since a previous iclog is in the
2124 * CALLBACK and the state cannot change since
2125 * we are holding the l_icloglock.
2127 if (!(iclog->ic_state &
2128 (XLOG_STATE_DONE_SYNC |
2129 XLOG_STATE_DO_CALLBACK))) {
2130 if (ciclog && (ciclog->ic_state ==
2131 XLOG_STATE_DONE_SYNC)) {
2132 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2137 * We now have an iclog that is in either the
2138 * DO_CALLBACK or DONE_SYNC states. The other
2139 * states (WANT_SYNC, SYNCING, or CALLBACK were
2140 * caught by the above if and are going to
2141 * clean (i.e. we aren't doing their callbacks)
2146 * We will do one more check here to see if we
2147 * have chased our tail around.
2150 lowest_lsn = xlog_get_lowest_lsn(log);
2152 XFS_LSN_CMP(lowest_lsn,
2153 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2154 iclog = iclog->ic_next;
2155 continue; /* Leave this iclog for
2159 iclog->ic_state = XLOG_STATE_CALLBACK;
2161 spin_unlock(&log->l_icloglock);
2163 /* l_last_sync_lsn field protected by
2164 * l_grant_lock. Don't worry about iclog's lsn.
2165 * No one else can be here except us.
2167 spin_lock(&log->l_grant_lock);
2168 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2169 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2170 log->l_last_sync_lsn =
2171 be64_to_cpu(iclog->ic_header.h_lsn);
2172 spin_unlock(&log->l_grant_lock);
2175 spin_unlock(&log->l_icloglock);
2180 * Keep processing entries in the callback list until
2181 * we come around and it is empty. We need to
2182 * atomically see that the list is empty and change the
2183 * state to DIRTY so that we don't miss any more
2184 * callbacks being added.
2186 spin_lock(&iclog->ic_callback_lock);
2187 cb = iclog->ic_callback;
2189 iclog->ic_callback_tail = &(iclog->ic_callback);
2190 iclog->ic_callback = NULL;
2191 spin_unlock(&iclog->ic_callback_lock);
2193 /* perform callbacks in the order given */
2194 for (; cb; cb = cb_next) {
2195 cb_next = cb->cb_next;
2196 cb->cb_func(cb->cb_arg, aborted);
2198 spin_lock(&iclog->ic_callback_lock);
2199 cb = iclog->ic_callback;
2205 spin_lock(&log->l_icloglock);
2206 ASSERT(iclog->ic_callback == NULL);
2207 spin_unlock(&iclog->ic_callback_lock);
2208 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2209 iclog->ic_state = XLOG_STATE_DIRTY;
2212 * Transition from DIRTY to ACTIVE if applicable.
2213 * NOP if STATE_IOERROR.
2215 xlog_state_clean_log(log);
2217 /* wake up threads waiting in xfs_log_force() */
2218 sv_broadcast(&iclog->ic_forcesema);
2220 iclog = iclog->ic_next;
2221 } while (first_iclog != iclog);
2223 if (repeats > 5000) {
2224 flushcnt += repeats;
2226 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2227 "%s: possible infinite loop (%d iterations)",
2228 __func__, flushcnt);
2230 } while (!ioerrors && loopdidcallbacks);
2233 * make one last gasp attempt to see if iclogs are being left in
2237 if (funcdidcallbacks) {
2238 first_iclog = iclog = log->l_iclog;
2240 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2242 * Terminate the loop if iclogs are found in states
2243 * which will cause other threads to clean up iclogs.
2245 * SYNCING - i/o completion will go through logs
2246 * DONE_SYNC - interrupt thread should be waiting for
2248 * IOERROR - give up hope all ye who enter here
2250 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2251 iclog->ic_state == XLOG_STATE_SYNCING ||
2252 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2253 iclog->ic_state == XLOG_STATE_IOERROR )
2255 iclog = iclog->ic_next;
2256 } while (first_iclog != iclog);
2261 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2262 flushcnt = log->l_flushcnt;
2263 log->l_flushcnt = 0;
2265 spin_unlock(&log->l_icloglock);
2267 vsema(&log->l_flushsema);
2268 } /* xlog_state_do_callback */
2272 * Finish transitioning this iclog to the dirty state.
2274 * Make sure that we completely execute this routine only when this is
2275 * the last call to the iclog. There is a good chance that iclog flushes,
2276 * when we reach the end of the physical log, get turned into 2 separate
2277 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2278 * routine. By using the reference count bwritecnt, we guarantee that only
2279 * the second completion goes through.
2281 * Callbacks could take time, so they are done outside the scope of the
2282 * global state machine log lock. Assume that the calls to cvsema won't
2283 * take a long time. At least we know it won't sleep.
2286 xlog_state_done_syncing(
2287 xlog_in_core_t *iclog,
2290 xlog_t *log = iclog->ic_log;
2292 spin_lock(&log->l_icloglock);
2294 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2295 iclog->ic_state == XLOG_STATE_IOERROR);
2296 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2297 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2301 * If we got an error, either on the first buffer, or in the case of
2302 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2303 * and none should ever be attempted to be written to disk
2306 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2307 if (--iclog->ic_bwritecnt == 1) {
2308 spin_unlock(&log->l_icloglock);
2311 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2315 * Someone could be sleeping prior to writing out the next
2316 * iclog buffer, we wake them all, one will get to do the
2317 * I/O, the others get to wait for the result.
2319 sv_broadcast(&iclog->ic_writesema);
2320 spin_unlock(&log->l_icloglock);
2321 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2322 } /* xlog_state_done_syncing */
2326 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2327 * sleep. The flush semaphore is set to the number of in-core buffers and
2328 * decremented around disk syncing. Therefore, if all buffers are syncing,
2329 * this semaphore will cause new writes to sleep until a sync completes.
2330 * Otherwise, this code just does p() followed by v(). This approximates
2331 * a sleep/wakeup except we can't race.
2333 * The in-core logs are used in a circular fashion. They are not used
2334 * out-of-order even when an iclog past the head is free.
2337 * * log_offset where xlog_write() can start writing into the in-core
2339 * * in-core log pointer to which xlog_write() should write.
2340 * * boolean indicating this is a continued write to an in-core log.
2341 * If this is the last write, then the in-core log's offset field
2342 * needs to be incremented, depending on the amount of data which
2346 xlog_state_get_iclog_space(xlog_t *log,
2348 xlog_in_core_t **iclogp,
2349 xlog_ticket_t *ticket,
2350 int *continued_write,
2354 xlog_rec_header_t *head;
2355 xlog_in_core_t *iclog;
2359 spin_lock(&log->l_icloglock);
2360 if (XLOG_FORCED_SHUTDOWN(log)) {
2361 spin_unlock(&log->l_icloglock);
2362 return XFS_ERROR(EIO);
2365 iclog = log->l_iclog;
2366 if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2368 spin_unlock(&log->l_icloglock);
2369 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2370 XFS_STATS_INC(xs_log_noiclogs);
2371 /* Ensure that log writes happen */
2372 psema(&log->l_flushsema, PINOD);
2375 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2376 head = &iclog->ic_header;
2378 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2379 log_offset = iclog->ic_offset;
2381 /* On the 1st write to an iclog, figure out lsn. This works
2382 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2383 * committing to. If the offset is set, that's how many blocks
2386 if (log_offset == 0) {
2387 ticket->t_curr_res -= log->l_iclog_hsize;
2388 xlog_tic_add_region(ticket,
2390 XLOG_REG_TYPE_LRHEADER);
2391 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2392 head->h_lsn = cpu_to_be64(
2393 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2394 ASSERT(log->l_curr_block >= 0);
2397 /* If there is enough room to write everything, then do it. Otherwise,
2398 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2399 * bit is on, so this will get flushed out. Don't update ic_offset
2400 * until you know exactly how many bytes get copied. Therefore, wait
2401 * until later to update ic_offset.
2403 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2404 * can fit into remaining data section.
2406 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2407 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2409 /* If I'm the only one writing to this iclog, sync it to disk */
2410 if (atomic_read(&iclog->ic_refcnt) == 1) {
2411 spin_unlock(&log->l_icloglock);
2412 if ((error = xlog_state_release_iclog(log, iclog)))
2415 atomic_dec(&iclog->ic_refcnt);
2416 spin_unlock(&log->l_icloglock);
2421 /* Do we have enough room to write the full amount in the remainder
2422 * of this iclog? Or must we continue a write on the next iclog and
2423 * mark this iclog as completely taken? In the case where we switch
2424 * iclogs (to mark it taken), this particular iclog will release/sync
2425 * to disk in xlog_write().
2427 if (len <= iclog->ic_size - iclog->ic_offset) {
2428 *continued_write = 0;
2429 iclog->ic_offset += len;
2431 *continued_write = 1;
2432 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2436 ASSERT(iclog->ic_offset <= iclog->ic_size);
2437 spin_unlock(&log->l_icloglock);
2439 *logoffsetp = log_offset;
2441 } /* xlog_state_get_iclog_space */
2444 * Atomically get the log space required for a log ticket.
2446 * Once a ticket gets put onto the reserveq, it will only return after
2447 * the needed reservation is satisfied.
2450 xlog_grant_log_space(xlog_t *log,
2461 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2462 panic("grant Recovery problem");
2465 /* Is there space or do we need to sleep? */
2466 spin_lock(&log->l_grant_lock);
2467 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2469 /* something is already sleeping; insert new transaction at end */
2470 if (log->l_reserve_headq) {
2471 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2472 xlog_trace_loggrant(log, tic,
2473 "xlog_grant_log_space: sleep 1");
2475 * Gotta check this before going to sleep, while we're
2476 * holding the grant lock.
2478 if (XLOG_FORCED_SHUTDOWN(log))
2481 XFS_STATS_INC(xs_sleep_logspace);
2482 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2484 * If we got an error, and the filesystem is shutting down,
2485 * we'll catch it down below. So just continue...
2487 xlog_trace_loggrant(log, tic,
2488 "xlog_grant_log_space: wake 1");
2489 spin_lock(&log->l_grant_lock);
2491 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2492 need_bytes = tic->t_unit_res*tic->t_ocnt;
2494 need_bytes = tic->t_unit_res;
2497 if (XLOG_FORCED_SHUTDOWN(log))
2500 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2501 log->l_grant_reserve_bytes);
2502 if (free_bytes < need_bytes) {
2503 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2504 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2505 xlog_trace_loggrant(log, tic,
2506 "xlog_grant_log_space: sleep 2");
2507 XFS_STATS_INC(xs_sleep_logspace);
2508 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2510 if (XLOG_FORCED_SHUTDOWN(log)) {
2511 spin_lock(&log->l_grant_lock);
2515 xlog_trace_loggrant(log, tic,
2516 "xlog_grant_log_space: wake 2");
2517 xlog_grant_push_ail(log->l_mp, need_bytes);
2518 spin_lock(&log->l_grant_lock);
2520 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2521 xlog_del_ticketq(&log->l_reserve_headq, tic);
2523 /* we've got enough space */
2524 xlog_grant_add_space(log, need_bytes);
2526 tail_lsn = log->l_tail_lsn;
2528 * Check to make sure the grant write head didn't just over lap the
2529 * tail. If the cycles are the same, we can't be overlapping.
2530 * Otherwise, make sure that the cycles differ by exactly one and
2531 * check the byte count.
2533 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2534 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2535 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2538 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2539 xlog_verify_grant_head(log, 1);
2540 spin_unlock(&log->l_grant_lock);
2544 if (tic->t_flags & XLOG_TIC_IN_Q)
2545 xlog_del_ticketq(&log->l_reserve_headq, tic);
2546 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2548 * If we are failing, make sure the ticket doesn't have any
2549 * current reservations. We don't want to add this back when
2550 * the ticket/transaction gets cancelled.
2552 tic->t_curr_res = 0;
2553 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2554 spin_unlock(&log->l_grant_lock);
2555 return XFS_ERROR(EIO);
2556 } /* xlog_grant_log_space */
2560 * Replenish the byte reservation required by moving the grant write head.
2565 xlog_regrant_write_log_space(xlog_t *log,
2568 int free_bytes, need_bytes;
2569 xlog_ticket_t *ntic;
2574 tic->t_curr_res = tic->t_unit_res;
2575 xlog_tic_reset_res(tic);
2581 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2582 panic("regrant Recovery problem");
2585 spin_lock(&log->l_grant_lock);
2586 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2588 if (XLOG_FORCED_SHUTDOWN(log))
2591 /* If there are other waiters on the queue then give them a
2592 * chance at logspace before us. Wake up the first waiters,
2593 * if we do not wake up all the waiters then go to sleep waiting
2594 * for more free space, otherwise try to get some space for
2598 if ((ntic = log->l_write_headq)) {
2599 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2600 log->l_grant_write_bytes);
2602 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2604 if (free_bytes < ntic->t_unit_res)
2606 free_bytes -= ntic->t_unit_res;
2607 sv_signal(&ntic->t_sema);
2608 ntic = ntic->t_next;
2609 } while (ntic != log->l_write_headq);
2611 if (ntic != log->l_write_headq) {
2612 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2613 xlog_ins_ticketq(&log->l_write_headq, tic);
2615 xlog_trace_loggrant(log, tic,
2616 "xlog_regrant_write_log_space: sleep 1");
2617 XFS_STATS_INC(xs_sleep_logspace);
2618 sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2619 &log->l_grant_lock, s);
2621 /* If we're shutting down, this tic is already
2623 if (XLOG_FORCED_SHUTDOWN(log)) {
2624 spin_lock(&log->l_grant_lock);
2628 xlog_trace_loggrant(log, tic,
2629 "xlog_regrant_write_log_space: wake 1");
2630 xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2631 spin_lock(&log->l_grant_lock);
2635 need_bytes = tic->t_unit_res;
2638 if (XLOG_FORCED_SHUTDOWN(log))
2641 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2642 log->l_grant_write_bytes);
2643 if (free_bytes < need_bytes) {
2644 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2645 xlog_ins_ticketq(&log->l_write_headq, tic);
2646 XFS_STATS_INC(xs_sleep_logspace);
2647 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2649 /* If we're shutting down, this tic is already off the queue */
2650 if (XLOG_FORCED_SHUTDOWN(log)) {
2651 spin_lock(&log->l_grant_lock);
2655 xlog_trace_loggrant(log, tic,
2656 "xlog_regrant_write_log_space: wake 2");
2657 xlog_grant_push_ail(log->l_mp, need_bytes);
2658 spin_lock(&log->l_grant_lock);
2660 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2661 xlog_del_ticketq(&log->l_write_headq, tic);
2663 /* we've got enough space */
2664 xlog_grant_add_space_write(log, need_bytes);
2666 tail_lsn = log->l_tail_lsn;
2667 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2668 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2669 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2673 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2674 xlog_verify_grant_head(log, 1);
2675 spin_unlock(&log->l_grant_lock);
2680 if (tic->t_flags & XLOG_TIC_IN_Q)
2681 xlog_del_ticketq(&log->l_reserve_headq, tic);
2682 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2684 * If we are failing, make sure the ticket doesn't have any
2685 * current reservations. We don't want to add this back when
2686 * the ticket/transaction gets cancelled.
2688 tic->t_curr_res = 0;
2689 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2690 spin_unlock(&log->l_grant_lock);
2691 return XFS_ERROR(EIO);
2692 } /* xlog_regrant_write_log_space */
2695 /* The first cnt-1 times through here we don't need to
2696 * move the grant write head because the permanent
2697 * reservation has reserved cnt times the unit amount.
2698 * Release part of current permanent unit reservation and
2699 * reset current reservation to be one units worth. Also
2700 * move grant reservation head forward.
2703 xlog_regrant_reserve_log_space(xlog_t *log,
2704 xlog_ticket_t *ticket)
2706 xlog_trace_loggrant(log, ticket,
2707 "xlog_regrant_reserve_log_space: enter");
2708 if (ticket->t_cnt > 0)
2711 spin_lock(&log->l_grant_lock);
2712 xlog_grant_sub_space(log, ticket->t_curr_res);
2713 ticket->t_curr_res = ticket->t_unit_res;
2714 xlog_tic_reset_res(ticket);
2715 xlog_trace_loggrant(log, ticket,
2716 "xlog_regrant_reserve_log_space: sub current res");
2717 xlog_verify_grant_head(log, 1);
2719 /* just return if we still have some of the pre-reserved space */
2720 if (ticket->t_cnt > 0) {
2721 spin_unlock(&log->l_grant_lock);
2725 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2726 xlog_trace_loggrant(log, ticket,
2727 "xlog_regrant_reserve_log_space: exit");
2728 xlog_verify_grant_head(log, 0);
2729 spin_unlock(&log->l_grant_lock);
2730 ticket->t_curr_res = ticket->t_unit_res;
2731 xlog_tic_reset_res(ticket);
2732 } /* xlog_regrant_reserve_log_space */
2736 * Give back the space left from a reservation.
2738 * All the information we need to make a correct determination of space left
2739 * is present. For non-permanent reservations, things are quite easy. The
2740 * count should have been decremented to zero. We only need to deal with the
2741 * space remaining in the current reservation part of the ticket. If the
2742 * ticket contains a permanent reservation, there may be left over space which
2743 * needs to be released. A count of N means that N-1 refills of the current
2744 * reservation can be done before we need to ask for more space. The first
2745 * one goes to fill up the first current reservation. Once we run out of
2746 * space, the count will stay at zero and the only space remaining will be
2747 * in the current reservation field.
2750 xlog_ungrant_log_space(xlog_t *log,
2751 xlog_ticket_t *ticket)
2753 if (ticket->t_cnt > 0)
2756 spin_lock(&log->l_grant_lock);
2757 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2759 xlog_grant_sub_space(log, ticket->t_curr_res);
2761 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2763 /* If this is a permanent reservation ticket, we may be able to free
2764 * up more space based on the remaining count.
2766 if (ticket->t_cnt > 0) {
2767 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2768 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2771 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2772 xlog_verify_grant_head(log, 1);
2773 spin_unlock(&log->l_grant_lock);
2774 xfs_log_move_tail(log->l_mp, 1);
2775 } /* xlog_ungrant_log_space */
2779 * Flush iclog to disk if this is the last reference to the given iclog and
2780 * the WANT_SYNC bit is set.
2782 * When this function is entered, the iclog is not necessarily in the
2783 * WANT_SYNC state. It may be sitting around waiting to get filled.
2788 xlog_state_release_iclog(
2790 xlog_in_core_t *iclog)
2792 int sync = 0; /* do we sync? */
2794 if (iclog->ic_state & XLOG_STATE_IOERROR)
2795 return XFS_ERROR(EIO);
2797 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2798 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2801 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2802 spin_unlock(&log->l_icloglock);
2803 return XFS_ERROR(EIO);
2805 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2806 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2808 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2809 /* update tail before writing to iclog */
2810 xlog_assign_tail_lsn(log->l_mp);
2812 iclog->ic_state = XLOG_STATE_SYNCING;
2813 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2814 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2815 /* cycle incremented when incrementing curr_block */
2817 spin_unlock(&log->l_icloglock);
2820 * We let the log lock go, so it's possible that we hit a log I/O
2821 * error or some other SHUTDOWN condition that marks the iclog
2822 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2823 * this iclog has consistent data, so we ignore IOERROR
2824 * flags after this point.
2827 return xlog_sync(log, iclog);
2829 } /* xlog_state_release_iclog */
2833 * This routine will mark the current iclog in the ring as WANT_SYNC
2834 * and move the current iclog pointer to the next iclog in the ring.
2835 * When this routine is called from xlog_state_get_iclog_space(), the
2836 * exact size of the iclog has not yet been determined. All we know is
2837 * that every data block. We have run out of space in this log record.
2840 xlog_state_switch_iclogs(xlog_t *log,
2841 xlog_in_core_t *iclog,
2844 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2846 eventual_size = iclog->ic_offset;
2847 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2848 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2849 log->l_prev_block = log->l_curr_block;
2850 log->l_prev_cycle = log->l_curr_cycle;
2852 /* roll log?: ic_offset changed later */
2853 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2855 /* Round up to next log-sunit */
2856 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2857 log->l_mp->m_sb.sb_logsunit > 1) {
2858 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2859 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2862 if (log->l_curr_block >= log->l_logBBsize) {
2863 log->l_curr_cycle++;
2864 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2865 log->l_curr_cycle++;
2866 log->l_curr_block -= log->l_logBBsize;
2867 ASSERT(log->l_curr_block >= 0);
2869 ASSERT(iclog == log->l_iclog);
2870 log->l_iclog = iclog->ic_next;
2871 } /* xlog_state_switch_iclogs */
2875 * Write out all data in the in-core log as of this exact moment in time.
2877 * Data may be written to the in-core log during this call. However,
2878 * we don't guarantee this data will be written out. A change from past
2879 * implementation means this routine will *not* write out zero length LRs.
2881 * Basically, we try and perform an intelligent scan of the in-core logs.
2882 * If we determine there is no flushable data, we just return. There is no
2883 * flushable data if:
2885 * 1. the current iclog is active and has no data; the previous iclog
2886 * is in the active or dirty state.
2887 * 2. the current iclog is drity, and the previous iclog is in the
2888 * active or dirty state.
2890 * We may sleep (call psema) if:
2892 * 1. the current iclog is not in the active nor dirty state.
2893 * 2. the current iclog dirty, and the previous iclog is not in the
2894 * active nor dirty state.
2895 * 3. the current iclog is active, and there is another thread writing
2896 * to this particular iclog.
2897 * 4. a) the current iclog is active and has no other writers
2898 * b) when we return from flushing out this iclog, it is still
2899 * not in the active nor dirty state.
2902 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2904 xlog_in_core_t *iclog;
2907 spin_lock(&log->l_icloglock);
2909 iclog = log->l_iclog;
2910 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2911 spin_unlock(&log->l_icloglock);
2912 return XFS_ERROR(EIO);
2915 /* If the head iclog is not active nor dirty, we just attach
2916 * ourselves to the head and go to sleep.
2918 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2919 iclog->ic_state == XLOG_STATE_DIRTY) {
2921 * If the head is dirty or (active and empty), then
2922 * we need to look at the previous iclog. If the previous
2923 * iclog is active or dirty we are done. There is nothing
2924 * to sync out. Otherwise, we attach ourselves to the
2925 * previous iclog and go to sleep.
2927 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2928 (atomic_read(&iclog->ic_refcnt) == 0
2929 && iclog->ic_offset == 0)) {
2930 iclog = iclog->ic_prev;
2931 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2932 iclog->ic_state == XLOG_STATE_DIRTY)
2937 if (atomic_read(&iclog->ic_refcnt) == 0) {
2938 /* We are the only one with access to this
2939 * iclog. Flush it out now. There should
2940 * be a roundoff of zero to show that someone
2941 * has already taken care of the roundoff from
2942 * the previous sync.
2944 atomic_inc(&iclog->ic_refcnt);
2945 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2946 xlog_state_switch_iclogs(log, iclog, 0);
2947 spin_unlock(&log->l_icloglock);
2949 if (xlog_state_release_iclog(log, iclog))
2950 return XFS_ERROR(EIO);
2952 spin_lock(&log->l_icloglock);
2953 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2954 iclog->ic_state != XLOG_STATE_DIRTY)
2959 /* Someone else is writing to this iclog.
2960 * Use its call to flush out the data. However,
2961 * the other thread may not force out this LR,
2962 * so we mark it WANT_SYNC.
2964 xlog_state_switch_iclogs(log, iclog, 0);
2970 /* By the time we come around again, the iclog could've been filled
2971 * which would give it another lsn. If we have a new lsn, just
2972 * return because the relevant data has been flushed.
2975 if (flags & XFS_LOG_SYNC) {
2977 * We must check if we're shutting down here, before
2978 * we wait, while we're holding the l_icloglock.
2979 * Then we check again after waking up, in case our
2980 * sleep was disturbed by a bad news.
2982 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2983 spin_unlock(&log->l_icloglock);
2984 return XFS_ERROR(EIO);
2986 XFS_STATS_INC(xs_log_force_sleep);
2987 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
2989 * No need to grab the log lock here since we're
2990 * only deciding whether or not to return EIO
2991 * and the memory read should be atomic.
2993 if (iclog->ic_state & XLOG_STATE_IOERROR)
2994 return XFS_ERROR(EIO);
3000 spin_unlock(&log->l_icloglock);
3003 } /* xlog_state_sync_all */
3007 * Used by code which implements synchronous log forces.
3009 * Find in-core log with lsn.
3010 * If it is in the DIRTY state, just return.
3011 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3012 * state and go to sleep or return.
3013 * If it is in any other state, go to sleep or return.
3015 * If filesystem activity goes to zero, the iclog will get flushed only by
3019 xlog_state_sync(xlog_t *log,
3024 xlog_in_core_t *iclog;
3025 int already_slept = 0;
3028 spin_lock(&log->l_icloglock);
3029 iclog = log->l_iclog;
3031 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3032 spin_unlock(&log->l_icloglock);
3033 return XFS_ERROR(EIO);
3037 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3038 iclog = iclog->ic_next;
3042 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3043 spin_unlock(&log->l_icloglock);
3047 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3049 * We sleep here if we haven't already slept (e.g.
3050 * this is the first time we've looked at the correct
3051 * iclog buf) and the buffer before us is going to
3052 * be sync'ed. The reason for this is that if we
3053 * are doing sync transactions here, by waiting for
3054 * the previous I/O to complete, we can allow a few
3055 * more transactions into this iclog before we close
3058 * Otherwise, we mark the buffer WANT_SYNC, and bump
3059 * up the refcnt so we can release the log (which drops
3060 * the ref count). The state switch keeps new transaction
3061 * commits from using this buffer. When the current commits
3062 * finish writing into the buffer, the refcount will drop to
3063 * zero and the buffer will go out then.
3065 if (!already_slept &&
3066 (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3067 XLOG_STATE_SYNCING))) {
3068 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3069 XFS_STATS_INC(xs_log_force_sleep);
3070 sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3071 &log->l_icloglock, s);
3076 atomic_inc(&iclog->ic_refcnt);
3077 xlog_state_switch_iclogs(log, iclog, 0);
3078 spin_unlock(&log->l_icloglock);
3079 if (xlog_state_release_iclog(log, iclog))
3080 return XFS_ERROR(EIO);
3082 spin_lock(&log->l_icloglock);
3086 if ((flags & XFS_LOG_SYNC) && /* sleep */
3087 !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3090 * Don't wait on the forcesema if we know that we've
3091 * gotten a log write error.
3093 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3094 spin_unlock(&log->l_icloglock);
3095 return XFS_ERROR(EIO);
3097 XFS_STATS_INC(xs_log_force_sleep);
3098 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3100 * No need to grab the log lock here since we're
3101 * only deciding whether or not to return EIO
3102 * and the memory read should be atomic.
3104 if (iclog->ic_state & XLOG_STATE_IOERROR)
3105 return XFS_ERROR(EIO);
3107 } else { /* just return */
3108 spin_unlock(&log->l_icloglock);
3112 } while (iclog != log->l_iclog);
3114 spin_unlock(&log->l_icloglock);
3116 } /* xlog_state_sync */
3120 * Called when we want to mark the current iclog as being ready to sync to
3124 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3126 spin_lock(&log->l_icloglock);
3128 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3129 xlog_state_switch_iclogs(log, iclog, 0);
3131 ASSERT(iclog->ic_state &
3132 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3135 spin_unlock(&log->l_icloglock);
3136 } /* xlog_state_want_sync */
3140 /*****************************************************************************
3144 *****************************************************************************
3148 * Free a used ticket.
3151 xlog_ticket_put(xlog_t *log,
3152 xlog_ticket_t *ticket)
3154 sv_destroy(&ticket->t_sema);
3155 kmem_zone_free(xfs_log_ticket_zone, ticket);
3156 } /* xlog_ticket_put */
3160 * Allocate and initialise a new log ticket.
3162 STATIC xlog_ticket_t *
3163 xlog_ticket_get(xlog_t *log,
3172 tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3177 * Permanent reservations have up to 'cnt'-1 active log operations
3178 * in the log. A unit in this case is the amount of space for one
3179 * of these log operations. Normal reservations have a cnt of 1
3180 * and their unit amount is the total amount of space required.
3182 * The following lines of code account for non-transaction data
3183 * which occupy space in the on-disk log.
3185 * Normal form of a transaction is:
3186 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3187 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3189 * We need to account for all the leadup data and trailer data
3190 * around the transaction data.
3191 * And then we need to account for the worst case in terms of using
3193 * The worst case will happen if:
3194 * - the placement of the transaction happens to be such that the
3195 * roundoff is at its maximum
3196 * - the transaction data is synced before the commit record is synced
3197 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3198 * Therefore the commit record is in its own Log Record.
3199 * This can happen as the commit record is called with its
3200 * own region to xlog_write().
3201 * This then means that in the worst case, roundoff can happen for
3202 * the commit-rec as well.
3203 * The commit-rec is smaller than padding in this scenario and so it is
3204 * not added separately.
3207 /* for trans header */
3208 unit_bytes += sizeof(xlog_op_header_t);
3209 unit_bytes += sizeof(xfs_trans_header_t);
3212 unit_bytes += sizeof(xlog_op_header_t);
3214 /* for LR headers */
3215 num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3216 unit_bytes += log->l_iclog_hsize * num_headers;
3218 /* for commit-rec LR header - note: padding will subsume the ophdr */
3219 unit_bytes += log->l_iclog_hsize;
3221 /* for split-recs - ophdrs added when data split over LRs */
3222 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3224 /* for roundoff padding for transaction data and one for commit record */
3225 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3226 log->l_mp->m_sb.sb_logsunit > 1) {
3227 /* log su roundoff */
3228 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3231 unit_bytes += 2*BBSIZE;
3234 tic->t_unit_res = unit_bytes;
3235 tic->t_curr_res = unit_bytes;
3238 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3239 tic->t_clientid = client;
3240 tic->t_flags = XLOG_TIC_INITED;
3241 tic->t_trans_type = 0;
3242 if (xflags & XFS_LOG_PERM_RESERV)
3243 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3244 sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3246 xlog_tic_reset_res(tic);
3249 } /* xlog_ticket_get */
3252 /******************************************************************************
3254 * Log debug routines
3256 ******************************************************************************
3260 * Make sure that the destination ptr is within the valid data region of
3261 * one of the iclogs. This uses backup pointers stored in a different
3262 * part of the log in case we trash the log structure.
3265 xlog_verify_dest_ptr(xlog_t *log,
3271 for (i=0; i < log->l_iclog_bufs; i++) {
3272 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3273 ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3277 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3278 } /* xlog_verify_dest_ptr */
3281 xlog_verify_grant_head(xlog_t *log, int equals)
3283 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3285 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3287 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3289 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3290 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3292 } /* xlog_verify_grant_head */
3294 /* check if it will fit */
3296 xlog_verify_tail_lsn(xlog_t *log,
3297 xlog_in_core_t *iclog,
3302 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3304 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3305 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3306 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3308 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3310 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3311 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3313 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3314 if (blocks < BTOBB(iclog->ic_offset) + 1)
3315 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3317 } /* xlog_verify_tail_lsn */
3320 * Perform a number of checks on the iclog before writing to disk.
3322 * 1. Make sure the iclogs are still circular
3323 * 2. Make sure we have a good magic number
3324 * 3. Make sure we don't have magic numbers in the data
3325 * 4. Check fields of each log operation header for:
3326 * A. Valid client identifier
3327 * B. tid ptr value falls in valid ptr space (user space code)
3328 * C. Length in log record header is correct according to the
3329 * individual operation headers within record.
3330 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3331 * log, check the preceding blocks of the physical log to make sure all
3332 * the cycle numbers agree with the current cycle number.
3335 xlog_verify_iclog(xlog_t *log,
3336 xlog_in_core_t *iclog,
3340 xlog_op_header_t *ophead;
3341 xlog_in_core_t *icptr;
3342 xlog_in_core_2_t *xhdr;
3344 xfs_caddr_t base_ptr;
3345 __psint_t field_offset;
3347 int len, i, j, k, op_len;
3350 /* check validity of iclog pointers */
3351 spin_lock(&log->l_icloglock);
3352 icptr = log->l_iclog;
3353 for (i=0; i < log->l_iclog_bufs; i++) {
3355 xlog_panic("xlog_verify_iclog: invalid ptr");
3356 icptr = icptr->ic_next;
3358 if (icptr != log->l_iclog)
3359 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3360 spin_unlock(&log->l_icloglock);
3362 /* check log magic numbers */
3363 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3364 xlog_panic("xlog_verify_iclog: invalid magic num");
3366 ptr = (xfs_caddr_t) &iclog->ic_header;
3367 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3369 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3370 xlog_panic("xlog_verify_iclog: unexpected magic num");
3374 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3375 ptr = iclog->ic_datap;
3377 ophead = (xlog_op_header_t *)ptr;
3378 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3379 for (i = 0; i < len; i++) {
3380 ophead = (xlog_op_header_t *)ptr;
3382 /* clientid is only 1 byte */
3383 field_offset = (__psint_t)
3384 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3385 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3386 clientid = ophead->oh_clientid;
3388 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3389 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3390 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3391 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3392 clientid = xlog_get_client_id(
3393 xhdr[j].hic_xheader.xh_cycle_data[k]);
3395 clientid = xlog_get_client_id(
3396 iclog->ic_header.h_cycle_data[idx]);
3399 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3400 cmn_err(CE_WARN, "xlog_verify_iclog: "
3401 "invalid clientid %d op 0x%p offset 0x%lx",
3402 clientid, ophead, (unsigned long)field_offset);
3405 field_offset = (__psint_t)
3406 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3407 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3408 op_len = be32_to_cpu(ophead->oh_len);
3410 idx = BTOBBT((__psint_t)&ophead->oh_len -
3411 (__psint_t)iclog->ic_datap);
3412 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3413 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3414 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3415 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3417 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3420 ptr += sizeof(xlog_op_header_t) + op_len;
3422 } /* xlog_verify_iclog */
3426 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3432 xlog_in_core_t *iclog, *ic;
3434 iclog = log->l_iclog;
3435 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3437 * Mark all the incore logs IOERROR.
3438 * From now on, no log flushes will result.
3442 ic->ic_state = XLOG_STATE_IOERROR;
3444 } while (ic != iclog);
3448 * Return non-zero, if state transition has already happened.
3454 * This is called from xfs_force_shutdown, when we're forcibly
3455 * shutting down the filesystem, typically because of an IO error.
3456 * Our main objectives here are to make sure that:
3457 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3458 * parties to find out, 'atomically'.
3459 * b. those who're sleeping on log reservations, pinned objects and
3460 * other resources get woken up, and be told the bad news.
3461 * c. nothing new gets queued up after (a) and (b) are done.
3462 * d. if !logerror, flush the iclogs to disk, then seal them off
3466 xfs_log_force_umount(
3467 struct xfs_mount *mp,
3478 * If this happens during log recovery, don't worry about
3479 * locking; the log isn't open for business yet.
3482 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3483 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3484 XFS_BUF_DONE(mp->m_sb_bp);
3489 * Somebody could've already done the hard work for us.
3490 * No need to get locks for this.
3492 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3493 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3498 * We must hold both the GRANT lock and the LOG lock,
3499 * before we mark the filesystem SHUTDOWN and wake
3500 * everybody up to tell the bad news.
3502 spin_lock(&log->l_icloglock);
3503 spin_lock(&log->l_grant_lock);
3504 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3505 XFS_BUF_DONE(mp->m_sb_bp);
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
3511 log->l_flags |= XLOG_IO_ERROR;
3514 * If we hit a log error, we want to mark all the iclogs IOERROR
3515 * while we're still holding the loglock.
3518 retval = xlog_state_ioerror(log);
3519 spin_unlock(&log->l_icloglock);
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.
3529 if ((tic = log->l_reserve_headq)) {
3531 sv_signal(&tic->t_sema);
3533 } while (tic != log->l_reserve_headq);
3536 if ((tic = log->l_write_headq)) {
3538 sv_signal(&tic->t_sema);
3540 } while (tic != log->l_write_headq);
3542 spin_unlock(&log->l_grant_lock);
3544 if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3547 * Force the incore logs to disk before shutting the
3548 * log down completely.
3550 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3551 spin_lock(&log->l_icloglock);
3552 retval = xlog_state_ioerror(log);
3553 spin_unlock(&log->l_icloglock);
3556 * Wake up everybody waiting on xfs_log_force.
3557 * Callback all log item committed functions as if the
3558 * log writes were completed.
3560 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3562 #ifdef XFSERRORDEBUG
3564 xlog_in_core_t *iclog;
3566 spin_lock(&log->l_icloglock);
3567 iclog = log->l_iclog;
3569 ASSERT(iclog->ic_callback == 0);
3570 iclog = iclog->ic_next;
3571 } while (iclog != log->l_iclog);
3572 spin_unlock(&log->l_icloglock);
3575 /* return non-zero if log IOERROR transition had already happened */
3580 xlog_iclogs_empty(xlog_t *log)
3582 xlog_in_core_t *iclog;
3584 iclog = log->l_iclog;
3586 /* endianness does not matter here, zero is zero in
3589 if (iclog->ic_header.h_num_logops)
3591 iclog = iclog->ic_next;
3592 } while (iclog != log->l_iclog);