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"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_buf_item.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_trans_priv.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
39 #include "xfs_trace.h"
41 kmem_zone_t *xfs_log_ticket_zone;
43 /* Local miscellaneous function prototypes */
44 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
50 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
51 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52 STATIC void xlog_dealloc_log(xlog_t *log);
54 /* local state machine functions */
55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57 STATIC int xlog_state_get_iclog_space(xlog_t *log,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
63 STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65 STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
70 /* local functions to manipulate grant head */
71 STATIC int xlog_grant_log_space(xlog_t *log,
73 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
75 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
76 xlog_ticket_t *ticket);
77 STATIC int xlog_regrant_write_log_space(xlog_t *log,
78 xlog_ticket_t *ticket);
79 STATIC void xlog_ungrant_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
83 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
85 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
86 int count, boolean_t syncing);
87 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
90 #define xlog_verify_dest_ptr(a,b)
91 #define xlog_verify_grant_head(a,b)
92 #define xlog_verify_iclog(a,b,c,d)
93 #define xlog_verify_tail_lsn(a,b,c)
96 STATIC int xlog_iclogs_empty(xlog_t *log);
99 xlog_grant_sub_space(struct log *log, int bytes)
101 log->l_grant_write_bytes -= bytes;
102 if (log->l_grant_write_bytes < 0) {
103 log->l_grant_write_bytes += log->l_logsize;
104 log->l_grant_write_cycle--;
107 log->l_grant_reserve_bytes -= bytes;
108 if ((log)->l_grant_reserve_bytes < 0) {
109 log->l_grant_reserve_bytes += log->l_logsize;
110 log->l_grant_reserve_cycle--;
116 xlog_grant_add_space_write(struct log *log, int bytes)
118 int tmp = log->l_logsize - log->l_grant_write_bytes;
120 log->l_grant_write_bytes += bytes;
122 log->l_grant_write_cycle++;
123 log->l_grant_write_bytes = bytes - tmp;
128 xlog_grant_add_space_reserve(struct log *log, int bytes)
130 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
132 log->l_grant_reserve_bytes += bytes;
134 log->l_grant_reserve_cycle++;
135 log->l_grant_reserve_bytes = bytes - tmp;
140 xlog_grant_add_space(struct log *log, int bytes)
142 xlog_grant_add_space_write(log, bytes);
143 xlog_grant_add_space_reserve(log, bytes);
147 xlog_tic_reset_res(xlog_ticket_t *tic)
150 tic->t_res_arr_sum = 0;
151 tic->t_res_num_ophdrs = 0;
155 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
157 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
158 /* add to overflow and start again */
159 tic->t_res_o_flow += tic->t_res_arr_sum;
161 tic->t_res_arr_sum = 0;
164 tic->t_res_arr[tic->t_res_num].r_len = len;
165 tic->t_res_arr[tic->t_res_num].r_type = type;
166 tic->t_res_arr_sum += len;
173 * 1. currblock field gets updated at startup and after in-core logs
174 * marked as with WANT_SYNC.
178 * This routine is called when a user of a log manager ticket is done with
179 * the reservation. If the ticket was ever used, then a commit record for
180 * the associated transaction is written out as a log operation header with
181 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
182 * a given ticket. If the ticket was one with a permanent reservation, then
183 * a few operations are done differently. Permanent reservation tickets by
184 * default don't release the reservation. They just commit the current
185 * transaction with the belief that the reservation is still needed. A flag
186 * must be passed in before permanent reservations are actually released.
187 * When these type of tickets are not released, they need to be set into
188 * the inited state again. By doing this, a start record will be written
189 * out when the next write occurs.
193 struct xfs_mount *mp,
194 struct xlog_ticket *ticket,
195 struct xlog_in_core **iclog,
198 struct log *log = mp->m_log;
201 if (XLOG_FORCED_SHUTDOWN(log) ||
203 * If nothing was ever written, don't write out commit record.
204 * If we get an error, just continue and give back the log ticket.
206 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
207 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
208 lsn = (xfs_lsn_t) -1;
209 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
210 flags |= XFS_LOG_REL_PERM_RESERV;
215 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
216 (flags & XFS_LOG_REL_PERM_RESERV)) {
217 trace_xfs_log_done_nonperm(log, ticket);
220 * Release ticket if not permanent reservation or a specific
221 * request has been made to release a permanent reservation.
223 xlog_ungrant_log_space(log, ticket);
224 xfs_log_ticket_put(ticket);
226 trace_xfs_log_done_perm(log, ticket);
228 xlog_regrant_reserve_log_space(log, ticket);
229 /* If this ticket was a permanent reservation and we aren't
230 * trying to release it, reset the inited flags; so next time
231 * we write, a start record will be written out.
233 ticket->t_flags |= XLOG_TIC_INITED;
240 * Attaches a new iclog I/O completion callback routine during
241 * transaction commit. If the log is in error state, a non-zero
242 * return code is handed back and the caller is responsible for
243 * executing the callback at an appropriate time.
247 struct xfs_mount *mp,
248 struct xlog_in_core *iclog,
249 xfs_log_callback_t *cb)
253 spin_lock(&iclog->ic_callback_lock);
254 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
256 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
257 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
259 *(iclog->ic_callback_tail) = cb;
260 iclog->ic_callback_tail = &(cb->cb_next);
262 spin_unlock(&iclog->ic_callback_lock);
267 xfs_log_release_iclog(
268 struct xfs_mount *mp,
269 struct xlog_in_core *iclog)
271 if (xlog_state_release_iclog(mp->m_log, iclog)) {
272 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
280 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
281 * to the reservation.
282 * 2. Potentially, push buffers at tail of log to disk.
284 * Each reservation is going to reserve extra space for a log record header.
285 * When writes happen to the on-disk log, we don't subtract the length of the
286 * log record header from any reservation. By wasting space in each
287 * reservation, we prevent over allocation problems.
291 struct xfs_mount *mp,
294 struct xlog_ticket **ticket,
299 struct log *log = mp->m_log;
300 struct xlog_ticket *internal_ticket;
303 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
305 if (XLOG_FORCED_SHUTDOWN(log))
306 return XFS_ERROR(EIO);
308 XFS_STATS_INC(xs_try_logspace);
311 if (*ticket != NULL) {
312 ASSERT(flags & XFS_LOG_PERM_RESERV);
313 internal_ticket = *ticket;
316 * this is a new transaction on the ticket, so we need to
317 * change the transaction ID so that the next transaction has a
318 * different TID in the log. Just add one to the existing tid
319 * so that we can see chains of rolling transactions in the log
322 internal_ticket->t_tid++;
324 trace_xfs_log_reserve(log, internal_ticket);
326 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
327 retval = xlog_regrant_write_log_space(log, internal_ticket);
329 /* may sleep if need to allocate more tickets */
330 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
332 KM_SLEEP|KM_MAYFAIL);
333 if (!internal_ticket)
334 return XFS_ERROR(ENOMEM);
335 internal_ticket->t_trans_type = t_type;
336 *ticket = internal_ticket;
338 trace_xfs_log_reserve(log, internal_ticket);
340 xlog_grant_push_ail(mp,
341 (internal_ticket->t_unit_res *
342 internal_ticket->t_cnt));
343 retval = xlog_grant_log_space(log, internal_ticket);
347 } /* xfs_log_reserve */
351 * Mount a log filesystem
353 * mp - ubiquitous xfs mount point structure
354 * log_target - buftarg of on-disk log device
355 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
356 * num_bblocks - Number of BBSIZE blocks in on-disk log
358 * Return error or zero.
363 xfs_buftarg_t *log_target,
364 xfs_daddr_t blk_offset,
369 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
370 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
373 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
375 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
378 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
379 if (IS_ERR(mp->m_log)) {
380 error = -PTR_ERR(mp->m_log);
385 * Initialize the AIL now we have a log.
387 error = xfs_trans_ail_init(mp);
389 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
392 mp->m_log->l_ailp = mp->m_ail;
395 * skip log recovery on a norecovery mount. pretend it all
398 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
399 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
402 mp->m_flags &= ~XFS_MOUNT_RDONLY;
404 error = xlog_recover(mp->m_log);
407 mp->m_flags |= XFS_MOUNT_RDONLY;
409 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
410 goto out_destroy_ail;
414 /* Normal transactions can now occur */
415 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
418 * Now the log has been fully initialised and we know were our
419 * space grant counters are, we can initialise the permanent ticket
420 * needed for delayed logging to work.
422 xlog_cil_init_post_recovery(mp->m_log);
427 xfs_trans_ail_destroy(mp);
429 xlog_dealloc_log(mp->m_log);
435 * Finish the recovery of the file system. This is separate from
436 * the xfs_log_mount() call, because it depends on the code in
437 * xfs_mountfs() to read in the root and real-time bitmap inodes
438 * between calling xfs_log_mount() and here.
440 * mp - ubiquitous xfs mount point structure
443 xfs_log_mount_finish(xfs_mount_t *mp)
447 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
448 error = xlog_recover_finish(mp->m_log);
451 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
458 * Final log writes as part of unmount.
460 * Mark the filesystem clean as unmount happens. Note that during relocation
461 * this routine needs to be executed as part of source-bag while the
462 * deallocation must not be done until source-end.
466 * Unmount record used to have a string "Unmount filesystem--" in the
467 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
468 * We just write the magic number now since that particular field isn't
469 * currently architecture converted and "nUmount" is a bit foo.
470 * As far as I know, there weren't any dependencies on the old behaviour.
474 xfs_log_unmount_write(xfs_mount_t *mp)
476 xlog_t *log = mp->m_log;
477 xlog_in_core_t *iclog;
479 xlog_in_core_t *first_iclog;
481 xlog_ticket_t *tic = NULL;
486 * Don't write out unmount record on read-only mounts.
487 * Or, if we are doing a forced umount (typically because of IO errors).
489 if (mp->m_flags & XFS_MOUNT_RDONLY)
492 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
493 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
496 first_iclog = iclog = log->l_iclog;
498 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
499 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
500 ASSERT(iclog->ic_offset == 0);
502 iclog = iclog->ic_next;
503 } while (iclog != first_iclog);
505 if (! (XLOG_FORCED_SHUTDOWN(log))) {
506 error = xfs_log_reserve(mp, 600, 1, &tic,
507 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
509 /* the data section must be 32 bit size aligned */
513 __uint32_t pad2; /* may as well make it 64 bits */
515 .magic = XLOG_UNMOUNT_TYPE,
517 struct xfs_log_iovec reg = {
519 .i_len = sizeof(magic),
520 .i_type = XLOG_REG_TYPE_UNMOUNT,
522 struct xfs_log_vec vec = {
527 /* remove inited flag */
529 error = xlog_write(log, &vec, tic, &lsn,
530 NULL, XLOG_UNMOUNT_TRANS);
532 * At this point, we're umounting anyway,
533 * so there's no point in transitioning log state
534 * to IOERROR. Just continue...
539 xfs_fs_cmn_err(CE_ALERT, mp,
540 "xfs_log_unmount: unmount record failed");
544 spin_lock(&log->l_icloglock);
545 iclog = log->l_iclog;
546 atomic_inc(&iclog->ic_refcnt);
547 xlog_state_want_sync(log, iclog);
548 spin_unlock(&log->l_icloglock);
549 error = xlog_state_release_iclog(log, iclog);
551 spin_lock(&log->l_icloglock);
552 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
553 iclog->ic_state == XLOG_STATE_DIRTY)) {
554 if (!XLOG_FORCED_SHUTDOWN(log)) {
555 sv_wait(&iclog->ic_force_wait, PMEM,
556 &log->l_icloglock, s);
558 spin_unlock(&log->l_icloglock);
561 spin_unlock(&log->l_icloglock);
564 trace_xfs_log_umount_write(log, tic);
565 xlog_ungrant_log_space(log, tic);
566 xfs_log_ticket_put(tic);
570 * We're already in forced_shutdown mode, couldn't
571 * even attempt to write out the unmount transaction.
573 * Go through the motions of sync'ing and releasing
574 * the iclog, even though no I/O will actually happen,
575 * we need to wait for other log I/Os that may already
576 * be in progress. Do this as a separate section of
577 * code so we'll know if we ever get stuck here that
578 * we're in this odd situation of trying to unmount
579 * a file system that went into forced_shutdown as
580 * the result of an unmount..
582 spin_lock(&log->l_icloglock);
583 iclog = log->l_iclog;
584 atomic_inc(&iclog->ic_refcnt);
586 xlog_state_want_sync(log, iclog);
587 spin_unlock(&log->l_icloglock);
588 error = xlog_state_release_iclog(log, iclog);
590 spin_lock(&log->l_icloglock);
592 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
593 || iclog->ic_state == XLOG_STATE_DIRTY
594 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
596 sv_wait(&iclog->ic_force_wait, PMEM,
597 &log->l_icloglock, s);
599 spin_unlock(&log->l_icloglock);
604 } /* xfs_log_unmount_write */
607 * Deallocate log structures for unmount/relocation.
609 * We need to stop the aild from running before we destroy
610 * and deallocate the log as the aild references the log.
613 xfs_log_unmount(xfs_mount_t *mp)
615 xfs_trans_ail_destroy(mp);
616 xlog_dealloc_log(mp->m_log);
621 struct xfs_mount *mp,
622 struct xfs_log_item *item,
624 struct xfs_item_ops *ops)
626 item->li_mountp = mp;
627 item->li_ailp = mp->m_ail;
628 item->li_type = type;
632 INIT_LIST_HEAD(&item->li_ail);
633 INIT_LIST_HEAD(&item->li_cil);
637 * Write region vectors to log. The write happens using the space reservation
638 * of the ticket (tic). It is not a requirement that all writes for a given
639 * transaction occur with one call to xfs_log_write(). However, it is important
640 * to note that the transaction reservation code makes an assumption about the
641 * number of log headers a transaction requires that may be violated if you
642 * don't pass all the transaction vectors in one call....
646 struct xfs_mount *mp,
647 struct xfs_log_iovec reg[],
649 struct xlog_ticket *tic,
650 xfs_lsn_t *start_lsn)
652 struct log *log = mp->m_log;
654 struct xfs_log_vec vec = {
655 .lv_niovecs = nentries,
659 if (XLOG_FORCED_SHUTDOWN(log))
660 return XFS_ERROR(EIO);
662 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
664 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
669 xfs_log_move_tail(xfs_mount_t *mp,
673 xlog_t *log = mp->m_log;
674 int need_bytes, free_bytes, cycle, bytes;
676 if (XLOG_FORCED_SHUTDOWN(log))
680 /* needed since sync_lsn is 64 bits */
681 spin_lock(&log->l_icloglock);
682 tail_lsn = log->l_last_sync_lsn;
683 spin_unlock(&log->l_icloglock);
686 spin_lock(&log->l_grant_lock);
688 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
692 log->l_tail_lsn = tail_lsn;
695 if (!list_empty(&log->l_writeq)) {
697 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
698 panic("Recovery problem");
700 cycle = log->l_grant_write_cycle;
701 bytes = log->l_grant_write_bytes;
702 free_bytes = xlog_space_left(log, cycle, bytes);
703 list_for_each_entry(tic, &log->l_writeq, t_queue) {
704 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
706 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
709 free_bytes -= tic->t_unit_res;
710 sv_signal(&tic->t_wait);
714 if (!list_empty(&log->l_reserveq)) {
716 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
717 panic("Recovery problem");
719 cycle = log->l_grant_reserve_cycle;
720 bytes = log->l_grant_reserve_bytes;
721 free_bytes = xlog_space_left(log, cycle, bytes);
722 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
723 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
724 need_bytes = tic->t_unit_res*tic->t_cnt;
726 need_bytes = tic->t_unit_res;
727 if (free_bytes < need_bytes && tail_lsn != 1)
730 free_bytes -= need_bytes;
731 sv_signal(&tic->t_wait);
734 spin_unlock(&log->l_grant_lock);
735 } /* xfs_log_move_tail */
738 * Determine if we have a transaction that has gone to disk
739 * that needs to be covered. To begin the transition to the idle state
740 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
741 * If we are then in a state where covering is needed, the caller is informed
742 * that dummy transactions are required to move the log into the idle state.
744 * Because this is called as part of the sync process, we should also indicate
745 * that dummy transactions should be issued in anything but the covered or
746 * idle states. This ensures that the log tail is accurately reflected in
747 * the log at the end of the sync, hence if a crash occurrs avoids replay
748 * of transactions where the metadata is already on disk.
751 xfs_log_need_covered(xfs_mount_t *mp)
754 xlog_t *log = mp->m_log;
756 if (!xfs_fs_writable(mp))
759 spin_lock(&log->l_icloglock);
760 switch (log->l_covered_state) {
761 case XLOG_STATE_COVER_DONE:
762 case XLOG_STATE_COVER_DONE2:
763 case XLOG_STATE_COVER_IDLE:
765 case XLOG_STATE_COVER_NEED:
766 case XLOG_STATE_COVER_NEED2:
767 if (!xfs_trans_ail_tail(log->l_ailp) &&
768 xlog_iclogs_empty(log)) {
769 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
770 log->l_covered_state = XLOG_STATE_COVER_DONE;
772 log->l_covered_state = XLOG_STATE_COVER_DONE2;
779 spin_unlock(&log->l_icloglock);
783 /******************************************************************************
787 ******************************************************************************
790 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
791 * The log manager must keep track of the last LR which was committed
792 * to disk. The lsn of this LR will become the new tail_lsn whenever
793 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
794 * the situation where stuff could be written into the log but nothing
795 * was ever in the AIL when asked. Eventually, we panic since the
796 * tail hits the head.
798 * We may be holding the log iclog lock upon entering this routine.
801 xlog_assign_tail_lsn(xfs_mount_t *mp)
804 xlog_t *log = mp->m_log;
806 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
807 spin_lock(&log->l_grant_lock);
809 log->l_tail_lsn = tail_lsn;
811 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
813 spin_unlock(&log->l_grant_lock);
816 } /* xlog_assign_tail_lsn */
820 * Return the space in the log between the tail and the head. The head
821 * is passed in the cycle/bytes formal parms. In the special case where
822 * the reserve head has wrapped passed the tail, this calculation is no
823 * longer valid. In this case, just return 0 which means there is no space
824 * in the log. This works for all places where this function is called
825 * with the reserve head. Of course, if the write head were to ever
826 * wrap the tail, we should blow up. Rather than catch this case here,
827 * we depend on other ASSERTions in other parts of the code. XXXmiken
829 * This code also handles the case where the reservation head is behind
830 * the tail. The details of this case are described below, but the end
831 * result is that we return the size of the log as the amount of space left.
834 xlog_space_left(xlog_t *log, int cycle, int bytes)
840 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
841 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
842 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
843 free_bytes = log->l_logsize - (bytes - tail_bytes);
844 } else if ((tail_cycle + 1) < cycle) {
846 } else if (tail_cycle < cycle) {
847 ASSERT(tail_cycle == (cycle - 1));
848 free_bytes = tail_bytes - bytes;
851 * The reservation head is behind the tail.
852 * In this case we just want to return the size of the
853 * log as the amount of space left.
855 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
856 "xlog_space_left: head behind tail\n"
857 " tail_cycle = %d, tail_bytes = %d\n"
858 " GH cycle = %d, GH bytes = %d",
859 tail_cycle, tail_bytes, cycle, bytes);
861 free_bytes = log->l_logsize;
864 } /* xlog_space_left */
868 * Log function which is called when an io completes.
870 * The log manager needs its own routine, in order to control what
871 * happens with the buffer after the write completes.
874 xlog_iodone(xfs_buf_t *bp)
876 xlog_in_core_t *iclog;
880 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
881 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
882 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
887 * Race to shutdown the filesystem if we see an error.
889 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
890 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
891 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
893 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
895 * This flag will be propagated to the trans-committed
896 * callback routines to let them know that the log-commit
899 aborted = XFS_LI_ABORTED;
900 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
901 aborted = XFS_LI_ABORTED;
904 /* log I/O is always issued ASYNC */
905 ASSERT(XFS_BUF_ISASYNC(bp));
906 xlog_state_done_syncing(iclog, aborted);
908 * do not reference the buffer (bp) here as we could race
909 * with it being freed after writing the unmount record to the
916 * Return size of each in-core log record buffer.
918 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
920 * If the filesystem blocksize is too large, we may need to choose a
921 * larger size since the directory code currently logs entire blocks.
925 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
931 if (mp->m_logbufs <= 0)
932 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
934 log->l_iclog_bufs = mp->m_logbufs;
937 * Buffer size passed in from mount system call.
939 if (mp->m_logbsize > 0) {
940 size = log->l_iclog_size = mp->m_logbsize;
941 log->l_iclog_size_log = 0;
943 log->l_iclog_size_log++;
947 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
948 /* # headers = size / 32k
949 * one header holds cycles from 32k of data
952 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
953 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
955 log->l_iclog_hsize = xhdrs << BBSHIFT;
956 log->l_iclog_heads = xhdrs;
958 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
959 log->l_iclog_hsize = BBSIZE;
960 log->l_iclog_heads = 1;
965 /* All machines use 32kB buffers by default. */
966 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
967 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
969 /* the default log size is 16k or 32k which is one header sector */
970 log->l_iclog_hsize = BBSIZE;
971 log->l_iclog_heads = 1;
974 /* are we being asked to make the sizes selected above visible? */
975 if (mp->m_logbufs == 0)
976 mp->m_logbufs = log->l_iclog_bufs;
977 if (mp->m_logbsize == 0)
978 mp->m_logbsize = log->l_iclog_size;
979 } /* xlog_get_iclog_buffer_size */
983 * This routine initializes some of the log structure for a given mount point.
984 * Its primary purpose is to fill in enough, so recovery can occur. However,
985 * some other stuff may be filled in too.
988 xlog_alloc_log(xfs_mount_t *mp,
989 xfs_buftarg_t *log_target,
990 xfs_daddr_t blk_offset,
994 xlog_rec_header_t *head;
995 xlog_in_core_t **iclogp;
996 xlog_in_core_t *iclog, *prev_iclog=NULL;
1002 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1004 xlog_warn("XFS: Log allocation failed: No memory!");
1009 log->l_targ = log_target;
1010 log->l_logsize = BBTOB(num_bblks);
1011 log->l_logBBstart = blk_offset;
1012 log->l_logBBsize = num_bblks;
1013 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1014 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1016 log->l_prev_block = -1;
1017 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1018 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1019 log->l_last_sync_lsn = log->l_tail_lsn;
1020 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1021 log->l_grant_reserve_cycle = 1;
1022 log->l_grant_write_cycle = 1;
1023 INIT_LIST_HEAD(&log->l_reserveq);
1024 INIT_LIST_HEAD(&log->l_writeq);
1026 error = EFSCORRUPTED;
1027 if (xfs_sb_version_hassector(&mp->m_sb)) {
1028 log2_size = mp->m_sb.sb_logsectlog;
1029 if (log2_size < BBSHIFT) {
1030 xlog_warn("XFS: Log sector size too small "
1031 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1035 log2_size -= BBSHIFT;
1036 if (log2_size > mp->m_sectbb_log) {
1037 xlog_warn("XFS: Log sector size too large "
1038 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1042 /* for larger sector sizes, must have v2 or external log */
1043 if (log2_size && log->l_logBBstart > 0 &&
1044 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1046 xlog_warn("XFS: log sector size (0x%x) invalid "
1047 "for configuration.", log2_size);
1051 log->l_sectBBsize = 1 << log2_size;
1053 xlog_get_iclog_buffer_size(mp, log);
1056 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1059 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1060 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1061 ASSERT(XFS_BUF_ISBUSY(bp));
1062 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1065 spin_lock_init(&log->l_icloglock);
1066 spin_lock_init(&log->l_grant_lock);
1067 sv_init(&log->l_flush_wait, 0, "flush_wait");
1069 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1070 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1072 iclogp = &log->l_iclog;
1074 * The amount of memory to allocate for the iclog structure is
1075 * rather funky due to the way the structure is defined. It is
1076 * done this way so that we can use different sizes for machines
1077 * with different amounts of memory. See the definition of
1078 * xlog_in_core_t in xfs_log_priv.h for details.
1080 ASSERT(log->l_iclog_size >= 4096);
1081 for (i=0; i < log->l_iclog_bufs; i++) {
1082 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1084 goto out_free_iclog;
1087 iclog->ic_prev = prev_iclog;
1090 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1091 log->l_iclog_size, 0);
1093 goto out_free_iclog;
1094 if (!XFS_BUF_CPSEMA(bp))
1096 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1097 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1099 iclog->ic_data = bp->b_addr;
1101 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1103 head = &iclog->ic_header;
1104 memset(head, 0, sizeof(xlog_rec_header_t));
1105 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1106 head->h_version = cpu_to_be32(
1107 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1108 head->h_size = cpu_to_be32(log->l_iclog_size);
1110 head->h_fmt = cpu_to_be32(XLOG_FMT);
1111 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1113 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1114 iclog->ic_state = XLOG_STATE_ACTIVE;
1115 iclog->ic_log = log;
1116 atomic_set(&iclog->ic_refcnt, 0);
1117 spin_lock_init(&iclog->ic_callback_lock);
1118 iclog->ic_callback_tail = &(iclog->ic_callback);
1119 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1121 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1122 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1123 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1124 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1126 iclogp = &iclog->ic_next;
1128 *iclogp = log->l_iclog; /* complete ring */
1129 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1131 error = xlog_cil_init(log);
1133 goto out_free_iclog;
1137 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1138 prev_iclog = iclog->ic_next;
1140 sv_destroy(&iclog->ic_force_wait);
1141 sv_destroy(&iclog->ic_write_wait);
1142 xfs_buf_free(iclog->ic_bp);
1146 spinlock_destroy(&log->l_icloglock);
1147 spinlock_destroy(&log->l_grant_lock);
1148 xfs_buf_free(log->l_xbuf);
1152 return ERR_PTR(-error);
1153 } /* xlog_alloc_log */
1157 * Write out the commit record of a transaction associated with the given
1158 * ticket. Return the lsn of the commit record.
1163 struct xlog_ticket *ticket,
1164 struct xlog_in_core **iclog,
1165 xfs_lsn_t *commitlsnp)
1167 struct xfs_mount *mp = log->l_mp;
1169 struct xfs_log_iovec reg = {
1172 .i_type = XLOG_REG_TYPE_COMMIT,
1174 struct xfs_log_vec vec = {
1179 ASSERT_ALWAYS(iclog);
1180 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1183 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1188 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1189 * log space. This code pushes on the lsn which would supposedly free up
1190 * the 25% which we want to leave free. We may need to adopt a policy which
1191 * pushes on an lsn which is further along in the log once we reach the high
1192 * water mark. In this manner, we would be creating a low water mark.
1195 xlog_grant_push_ail(xfs_mount_t *mp,
1198 xlog_t *log = mp->m_log; /* pointer to the log */
1199 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1200 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1201 int free_blocks; /* free blocks left to write to */
1202 int free_bytes; /* free bytes left to write to */
1203 int threshold_block; /* block in lsn we'd like to be at */
1204 int threshold_cycle; /* lsn cycle we'd like to be at */
1207 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1209 spin_lock(&log->l_grant_lock);
1210 free_bytes = xlog_space_left(log,
1211 log->l_grant_reserve_cycle,
1212 log->l_grant_reserve_bytes);
1213 tail_lsn = log->l_tail_lsn;
1214 free_blocks = BTOBBT(free_bytes);
1217 * Set the threshold for the minimum number of free blocks in the
1218 * log to the maximum of what the caller needs, one quarter of the
1219 * log, and 256 blocks.
1221 free_threshold = BTOBB(need_bytes);
1222 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1223 free_threshold = MAX(free_threshold, 256);
1224 if (free_blocks < free_threshold) {
1225 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1226 threshold_cycle = CYCLE_LSN(tail_lsn);
1227 if (threshold_block >= log->l_logBBsize) {
1228 threshold_block -= log->l_logBBsize;
1229 threshold_cycle += 1;
1231 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1233 /* Don't pass in an lsn greater than the lsn of the last
1234 * log record known to be on disk.
1236 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1237 threshold_lsn = log->l_last_sync_lsn;
1239 spin_unlock(&log->l_grant_lock);
1242 * Get the transaction layer to kick the dirty buffers out to
1243 * disk asynchronously. No point in trying to do this if
1244 * the filesystem is shutting down.
1246 if (threshold_lsn &&
1247 !XLOG_FORCED_SHUTDOWN(log))
1248 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1249 } /* xlog_grant_push_ail */
1252 * The bdstrat callback function for log bufs. This gives us a central
1253 * place to trap bufs in case we get hit by a log I/O error and need to
1254 * shutdown. Actually, in practice, even when we didn't get a log error,
1255 * we transition the iclogs to IOERROR state *after* flushing all existing
1256 * iclogs to disk. This is because we don't want anymore new transactions to be
1257 * started or completed afterwards.
1263 struct xlog_in_core *iclog;
1265 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1266 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1267 XFS_BUF_ERROR(bp, EIO);
1269 xfs_buf_ioend(bp, 0);
1271 * It would seem logical to return EIO here, but we rely on
1272 * the log state machine to propagate I/O errors instead of
1278 bp->b_flags |= _XBF_RUN_QUEUES;
1279 xfs_buf_iorequest(bp);
1284 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1285 * fashion. Previously, we should have moved the current iclog
1286 * ptr in the log to point to the next available iclog. This allows further
1287 * write to continue while this code syncs out an iclog ready to go.
1288 * Before an in-core log can be written out, the data section must be scanned
1289 * to save away the 1st word of each BBSIZE block into the header. We replace
1290 * it with the current cycle count. Each BBSIZE block is tagged with the
1291 * cycle count because there in an implicit assumption that drives will
1292 * guarantee that entire 512 byte blocks get written at once. In other words,
1293 * we can't have part of a 512 byte block written and part not written. By
1294 * tagging each block, we will know which blocks are valid when recovering
1295 * after an unclean shutdown.
1297 * This routine is single threaded on the iclog. No other thread can be in
1298 * this routine with the same iclog. Changing contents of iclog can there-
1299 * fore be done without grabbing the state machine lock. Updating the global
1300 * log will require grabbing the lock though.
1302 * The entire log manager uses a logical block numbering scheme. Only
1303 * log_sync (and then only bwrite()) know about the fact that the log may
1304 * not start with block zero on a given device. The log block start offset
1305 * is added immediately before calling bwrite().
1309 xlog_sync(xlog_t *log,
1310 xlog_in_core_t *iclog)
1312 xfs_caddr_t dptr; /* pointer to byte sized element */
1315 uint count; /* byte count of bwrite */
1316 uint count_init; /* initial count before roundup */
1317 int roundoff; /* roundoff to BB or stripe */
1318 int split = 0; /* split write into two regions */
1320 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1322 XFS_STATS_INC(xs_log_writes);
1323 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1325 /* Add for LR header */
1326 count_init = log->l_iclog_hsize + iclog->ic_offset;
1328 /* Round out the log write size */
1329 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1330 /* we have a v2 stripe unit to use */
1331 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1333 count = BBTOB(BTOBB(count_init));
1335 roundoff = count - count_init;
1336 ASSERT(roundoff >= 0);
1337 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1338 roundoff < log->l_mp->m_sb.sb_logsunit)
1340 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1341 roundoff < BBTOB(1)));
1343 /* move grant heads by roundoff in sync */
1344 spin_lock(&log->l_grant_lock);
1345 xlog_grant_add_space(log, roundoff);
1346 spin_unlock(&log->l_grant_lock);
1348 /* put cycle number in every block */
1349 xlog_pack_data(log, iclog, roundoff);
1351 /* real byte length */
1353 iclog->ic_header.h_len =
1354 cpu_to_be32(iclog->ic_offset + roundoff);
1356 iclog->ic_header.h_len =
1357 cpu_to_be32(iclog->ic_offset);
1361 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1362 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1363 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1365 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1367 /* Do we need to split this write into 2 parts? */
1368 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1369 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1370 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1371 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1373 iclog->ic_bwritecnt = 1;
1375 XFS_BUF_SET_COUNT(bp, count);
1376 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1377 XFS_BUF_ZEROFLAGS(bp);
1380 bp->b_flags |= XBF_LOG_BUFFER;
1382 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1383 XFS_BUF_ORDERED(bp);
1385 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1386 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1388 xlog_verify_iclog(log, iclog, count, B_TRUE);
1390 /* account for log which doesn't start at block #0 */
1391 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1393 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1398 if ((error = xlog_bdstrat(bp))) {
1399 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1404 bp = iclog->ic_log->l_xbuf;
1405 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1407 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1408 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1409 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1410 (__psint_t)count), split);
1411 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1412 XFS_BUF_ZEROFLAGS(bp);
1415 bp->b_flags |= XBF_LOG_BUFFER;
1416 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1417 XFS_BUF_ORDERED(bp);
1418 dptr = XFS_BUF_PTR(bp);
1420 * Bump the cycle numbers at the start of each block
1421 * since this part of the buffer is at the start of
1422 * a new cycle. Watch out for the header magic number
1425 for (i = 0; i < split; i += BBSIZE) {
1426 be32_add_cpu((__be32 *)dptr, 1);
1427 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1428 be32_add_cpu((__be32 *)dptr, 1);
1432 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1433 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1435 /* account for internal log which doesn't start at block #0 */
1436 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1438 if ((error = xlog_bdstrat(bp))) {
1439 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1440 bp, XFS_BUF_ADDR(bp));
1449 * Deallocate a log structure
1452 xlog_dealloc_log(xlog_t *log)
1454 xlog_in_core_t *iclog, *next_iclog;
1457 xlog_cil_destroy(log);
1459 iclog = log->l_iclog;
1460 for (i=0; i<log->l_iclog_bufs; i++) {
1461 sv_destroy(&iclog->ic_force_wait);
1462 sv_destroy(&iclog->ic_write_wait);
1463 xfs_buf_free(iclog->ic_bp);
1464 next_iclog = iclog->ic_next;
1468 spinlock_destroy(&log->l_icloglock);
1469 spinlock_destroy(&log->l_grant_lock);
1471 xfs_buf_free(log->l_xbuf);
1472 log->l_mp->m_log = NULL;
1474 } /* xlog_dealloc_log */
1477 * Update counters atomically now that memcpy is done.
1481 xlog_state_finish_copy(xlog_t *log,
1482 xlog_in_core_t *iclog,
1486 spin_lock(&log->l_icloglock);
1488 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1489 iclog->ic_offset += copy_bytes;
1491 spin_unlock(&log->l_icloglock);
1492 } /* xlog_state_finish_copy */
1498 * print out info relating to regions written which consume
1503 struct xfs_mount *mp,
1504 struct xlog_ticket *ticket)
1507 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1509 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1510 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1531 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1574 xfs_fs_cmn_err(CE_WARN, mp,
1575 "xfs_log_write: reservation summary:\n"
1576 " trans type = %s (%u)\n"
1577 " unit res = %d bytes\n"
1578 " current res = %d bytes\n"
1579 " total reg = %u bytes (o/flow = %u bytes)\n"
1580 " ophdrs = %u (ophdr space = %u bytes)\n"
1581 " ophdr + reg = %u bytes\n"
1582 " num regions = %u\n",
1583 ((ticket->t_trans_type <= 0 ||
1584 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1585 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1586 ticket->t_trans_type,
1589 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1590 ticket->t_res_num_ophdrs, ophdr_spc,
1591 ticket->t_res_arr_sum +
1592 ticket->t_res_o_flow + ophdr_spc,
1595 for (i = 0; i < ticket->t_res_num; i++) {
1596 uint r_type = ticket->t_res_arr[i].r_type;
1598 "region[%u]: %s - %u bytes\n",
1600 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1601 "bad-rtype" : res_type_str[r_type-1]),
1602 ticket->t_res_arr[i].r_len);
1605 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1606 "xfs_log_write: reservation ran out. Need to up reservation");
1607 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1611 * Calculate the potential space needed by the log vector. Each region gets
1612 * its own xlog_op_header_t and may need to be double word aligned.
1615 xlog_write_calc_vec_length(
1616 struct xlog_ticket *ticket,
1617 struct xfs_log_vec *log_vector)
1619 struct xfs_log_vec *lv;
1624 /* acct for start rec of xact */
1625 if (ticket->t_flags & XLOG_TIC_INITED)
1628 for (lv = log_vector; lv; lv = lv->lv_next) {
1629 headers += lv->lv_niovecs;
1631 for (i = 0; i < lv->lv_niovecs; i++) {
1632 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1635 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1639 ticket->t_res_num_ophdrs += headers;
1640 len += headers * sizeof(struct xlog_op_header);
1646 * If first write for transaction, insert start record We can't be trying to
1647 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1650 xlog_write_start_rec(
1651 struct xlog_op_header *ophdr,
1652 struct xlog_ticket *ticket)
1654 if (!(ticket->t_flags & XLOG_TIC_INITED))
1657 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1658 ophdr->oh_clientid = ticket->t_clientid;
1660 ophdr->oh_flags = XLOG_START_TRANS;
1663 ticket->t_flags &= ~XLOG_TIC_INITED;
1665 return sizeof(struct xlog_op_header);
1668 static xlog_op_header_t *
1669 xlog_write_setup_ophdr(
1671 struct xlog_op_header *ophdr,
1672 struct xlog_ticket *ticket,
1675 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1676 ophdr->oh_clientid = ticket->t_clientid;
1679 /* are we copying a commit or unmount record? */
1680 ophdr->oh_flags = flags;
1683 * We've seen logs corrupted with bad transaction client ids. This
1684 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1685 * and shut down the filesystem.
1687 switch (ophdr->oh_clientid) {
1688 case XFS_TRANSACTION:
1693 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1694 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1695 ophdr->oh_clientid, ticket);
1703 * Set up the parameters of the region copy into the log. This has
1704 * to handle region write split across multiple log buffers - this
1705 * state is kept external to this function so that this code can
1706 * can be written in an obvious, self documenting manner.
1709 xlog_write_setup_copy(
1710 struct xlog_ticket *ticket,
1711 struct xlog_op_header *ophdr,
1712 int space_available,
1716 int *last_was_partial_copy,
1717 int *bytes_consumed)
1721 still_to_copy = space_required - *bytes_consumed;
1722 *copy_off = *bytes_consumed;
1724 if (still_to_copy <= space_available) {
1725 /* write of region completes here */
1726 *copy_len = still_to_copy;
1727 ophdr->oh_len = cpu_to_be32(*copy_len);
1728 if (*last_was_partial_copy)
1729 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1730 *last_was_partial_copy = 0;
1731 *bytes_consumed = 0;
1735 /* partial write of region, needs extra log op header reservation */
1736 *copy_len = space_available;
1737 ophdr->oh_len = cpu_to_be32(*copy_len);
1738 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1739 if (*last_was_partial_copy)
1740 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1741 *bytes_consumed += *copy_len;
1742 (*last_was_partial_copy)++;
1744 /* account for new log op header */
1745 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1746 ticket->t_res_num_ophdrs++;
1748 return sizeof(struct xlog_op_header);
1752 xlog_write_copy_finish(
1754 struct xlog_in_core *iclog,
1759 int *partial_copy_len,
1761 struct xlog_in_core **commit_iclog)
1763 if (*partial_copy) {
1765 * This iclog has already been marked WANT_SYNC by
1766 * xlog_state_get_iclog_space.
1768 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1771 return xlog_state_release_iclog(log, iclog);
1775 *partial_copy_len = 0;
1777 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1778 /* no more space in this iclog - push it. */
1779 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1783 spin_lock(&log->l_icloglock);
1784 xlog_state_want_sync(log, iclog);
1785 spin_unlock(&log->l_icloglock);
1788 return xlog_state_release_iclog(log, iclog);
1789 ASSERT(flags & XLOG_COMMIT_TRANS);
1790 *commit_iclog = iclog;
1797 * Write some region out to in-core log
1799 * This will be called when writing externally provided regions or when
1800 * writing out a commit record for a given transaction.
1802 * General algorithm:
1803 * 1. Find total length of this write. This may include adding to the
1804 * lengths passed in.
1805 * 2. Check whether we violate the tickets reservation.
1806 * 3. While writing to this iclog
1807 * A. Reserve as much space in this iclog as can get
1808 * B. If this is first write, save away start lsn
1809 * C. While writing this region:
1810 * 1. If first write of transaction, write start record
1811 * 2. Write log operation header (header per region)
1812 * 3. Find out if we can fit entire region into this iclog
1813 * 4. Potentially, verify destination memcpy ptr
1814 * 5. Memcpy (partial) region
1815 * 6. If partial copy, release iclog; otherwise, continue
1816 * copying more regions into current iclog
1817 * 4. Mark want sync bit (in simulation mode)
1818 * 5. Release iclog for potential flush to on-disk log.
1821 * 1. Panic if reservation is overrun. This should never happen since
1822 * reservation amounts are generated internal to the filesystem.
1824 * 1. Tickets are single threaded data structures.
1825 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1826 * syncing routine. When a single log_write region needs to span
1827 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1828 * on all log operation writes which don't contain the end of the
1829 * region. The XLOG_END_TRANS bit is used for the in-core log
1830 * operation which contains the end of the continued log_write region.
1831 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1832 * we don't really know exactly how much space will be used. As a result,
1833 * we don't update ic_offset until the end when we know exactly how many
1834 * bytes have been written out.
1839 struct xfs_log_vec *log_vector,
1840 struct xlog_ticket *ticket,
1841 xfs_lsn_t *start_lsn,
1842 struct xlog_in_core **commit_iclog,
1845 struct xlog_in_core *iclog = NULL;
1846 struct xfs_log_iovec *vecp;
1847 struct xfs_log_vec *lv;
1850 int partial_copy = 0;
1851 int partial_copy_len = 0;
1859 len = xlog_write_calc_vec_length(ticket, log_vector);
1862 * Region headers and bytes are already accounted for.
1863 * We only need to take into account start records and
1864 * split regions in this function.
1866 if (ticket->t_flags & XLOG_TIC_INITED)
1867 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1870 * Commit record headers need to be accounted for. These
1871 * come in as separate writes so are easy to detect.
1873 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1874 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1876 ticket->t_curr_res -= len;
1878 if (ticket->t_curr_res < 0)
1879 xlog_print_tic_res(log->l_mp, ticket);
1883 vecp = lv->lv_iovecp;
1884 while (lv && index < lv->lv_niovecs) {
1888 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1889 &contwr, &log_offset);
1893 ASSERT(log_offset <= iclog->ic_size - 1);
1894 ptr = iclog->ic_datap + log_offset;
1896 /* start_lsn is the first lsn written to. That's all we need. */
1898 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1901 * This loop writes out as many regions as can fit in the amount
1902 * of space which was allocated by xlog_state_get_iclog_space().
1904 while (lv && index < lv->lv_niovecs) {
1905 struct xfs_log_iovec *reg = &vecp[index];
1906 struct xlog_op_header *ophdr;
1911 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1912 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1914 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1915 if (start_rec_copy) {
1917 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1921 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1923 return XFS_ERROR(EIO);
1925 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1926 sizeof(struct xlog_op_header));
1928 len += xlog_write_setup_copy(ticket, ophdr,
1929 iclog->ic_size-log_offset,
1931 ©_off, ©_len,
1934 xlog_verify_dest_ptr(log, ptr);
1937 ASSERT(copy_len >= 0);
1938 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1939 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1941 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1943 data_cnt += contwr ? copy_len : 0;
1945 error = xlog_write_copy_finish(log, iclog, flags,
1946 &record_cnt, &data_cnt,
1955 * if we had a partial copy, we need to get more iclog
1956 * space but we don't want to increment the region
1957 * index because there is still more is this region to
1960 * If we completed writing this region, and we flushed
1961 * the iclog (indicated by resetting of the record
1962 * count), then we also need to get more log space. If
1963 * this was the last record, though, we are done and
1969 if (++index == lv->lv_niovecs) {
1973 vecp = lv->lv_iovecp;
1975 if (record_cnt == 0) {
1985 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1987 return xlog_state_release_iclog(log, iclog);
1989 ASSERT(flags & XLOG_COMMIT_TRANS);
1990 *commit_iclog = iclog;
1995 /*****************************************************************************
1997 * State Machine functions
1999 *****************************************************************************
2002 /* Clean iclogs starting from the head. This ordering must be
2003 * maintained, so an iclog doesn't become ACTIVE beyond one that
2004 * is SYNCING. This is also required to maintain the notion that we use
2005 * a ordered wait queue to hold off would be writers to the log when every
2006 * iclog is trying to sync to disk.
2008 * State Change: DIRTY -> ACTIVE
2011 xlog_state_clean_log(xlog_t *log)
2013 xlog_in_core_t *iclog;
2016 iclog = log->l_iclog;
2018 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2019 iclog->ic_state = XLOG_STATE_ACTIVE;
2020 iclog->ic_offset = 0;
2021 ASSERT(iclog->ic_callback == NULL);
2023 * If the number of ops in this iclog indicate it just
2024 * contains the dummy transaction, we can
2025 * change state into IDLE (the second time around).
2026 * Otherwise we should change the state into
2028 * We don't need to cover the dummy.
2031 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2036 * We have two dirty iclogs so start over
2037 * This could also be num of ops indicates
2038 * this is not the dummy going out.
2042 iclog->ic_header.h_num_logops = 0;
2043 memset(iclog->ic_header.h_cycle_data, 0,
2044 sizeof(iclog->ic_header.h_cycle_data));
2045 iclog->ic_header.h_lsn = 0;
2046 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2049 break; /* stop cleaning */
2050 iclog = iclog->ic_next;
2051 } while (iclog != log->l_iclog);
2053 /* log is locked when we are called */
2055 * Change state for the dummy log recording.
2056 * We usually go to NEED. But we go to NEED2 if the changed indicates
2057 * we are done writing the dummy record.
2058 * If we are done with the second dummy recored (DONE2), then
2062 switch (log->l_covered_state) {
2063 case XLOG_STATE_COVER_IDLE:
2064 case XLOG_STATE_COVER_NEED:
2065 case XLOG_STATE_COVER_NEED2:
2066 log->l_covered_state = XLOG_STATE_COVER_NEED;
2069 case XLOG_STATE_COVER_DONE:
2071 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2073 log->l_covered_state = XLOG_STATE_COVER_NEED;
2076 case XLOG_STATE_COVER_DONE2:
2078 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2080 log->l_covered_state = XLOG_STATE_COVER_NEED;
2087 } /* xlog_state_clean_log */
2090 xlog_get_lowest_lsn(
2093 xlog_in_core_t *lsn_log;
2094 xfs_lsn_t lowest_lsn, lsn;
2096 lsn_log = log->l_iclog;
2099 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2100 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2101 if ((lsn && !lowest_lsn) ||
2102 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2106 lsn_log = lsn_log->ic_next;
2107 } while (lsn_log != log->l_iclog);
2113 xlog_state_do_callback(
2116 xlog_in_core_t *ciclog)
2118 xlog_in_core_t *iclog;
2119 xlog_in_core_t *first_iclog; /* used to know when we've
2120 * processed all iclogs once */
2121 xfs_log_callback_t *cb, *cb_next;
2123 xfs_lsn_t lowest_lsn;
2124 int ioerrors; /* counter: iclogs with errors */
2125 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2126 int funcdidcallbacks; /* flag: function did callbacks */
2127 int repeats; /* for issuing console warnings if
2128 * looping too many times */
2131 spin_lock(&log->l_icloglock);
2132 first_iclog = iclog = log->l_iclog;
2134 funcdidcallbacks = 0;
2139 * Scan all iclogs starting with the one pointed to by the
2140 * log. Reset this starting point each time the log is
2141 * unlocked (during callbacks).
2143 * Keep looping through iclogs until one full pass is made
2144 * without running any callbacks.
2146 first_iclog = log->l_iclog;
2147 iclog = log->l_iclog;
2148 loopdidcallbacks = 0;
2153 /* skip all iclogs in the ACTIVE & DIRTY states */
2154 if (iclog->ic_state &
2155 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2156 iclog = iclog->ic_next;
2161 * Between marking a filesystem SHUTDOWN and stopping
2162 * the log, we do flush all iclogs to disk (if there
2163 * wasn't a log I/O error). So, we do want things to
2164 * go smoothly in case of just a SHUTDOWN w/o a
2167 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2169 * Can only perform callbacks in order. Since
2170 * this iclog is not in the DONE_SYNC/
2171 * DO_CALLBACK state, we skip the rest and
2172 * just try to clean up. If we set our iclog
2173 * to DO_CALLBACK, we will not process it when
2174 * we retry since a previous iclog is in the
2175 * CALLBACK and the state cannot change since
2176 * we are holding the l_icloglock.
2178 if (!(iclog->ic_state &
2179 (XLOG_STATE_DONE_SYNC |
2180 XLOG_STATE_DO_CALLBACK))) {
2181 if (ciclog && (ciclog->ic_state ==
2182 XLOG_STATE_DONE_SYNC)) {
2183 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2188 * We now have an iclog that is in either the
2189 * DO_CALLBACK or DONE_SYNC states. The other
2190 * states (WANT_SYNC, SYNCING, or CALLBACK were
2191 * caught by the above if and are going to
2192 * clean (i.e. we aren't doing their callbacks)
2197 * We will do one more check here to see if we
2198 * have chased our tail around.
2201 lowest_lsn = xlog_get_lowest_lsn(log);
2203 XFS_LSN_CMP(lowest_lsn,
2204 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2205 iclog = iclog->ic_next;
2206 continue; /* Leave this iclog for
2210 iclog->ic_state = XLOG_STATE_CALLBACK;
2212 spin_unlock(&log->l_icloglock);
2214 /* l_last_sync_lsn field protected by
2215 * l_grant_lock. Don't worry about iclog's lsn.
2216 * No one else can be here except us.
2218 spin_lock(&log->l_grant_lock);
2219 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2220 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2221 log->l_last_sync_lsn =
2222 be64_to_cpu(iclog->ic_header.h_lsn);
2223 spin_unlock(&log->l_grant_lock);
2226 spin_unlock(&log->l_icloglock);
2231 * Keep processing entries in the callback list until
2232 * we come around and it is empty. We need to
2233 * atomically see that the list is empty and change the
2234 * state to DIRTY so that we don't miss any more
2235 * callbacks being added.
2237 spin_lock(&iclog->ic_callback_lock);
2238 cb = iclog->ic_callback;
2240 iclog->ic_callback_tail = &(iclog->ic_callback);
2241 iclog->ic_callback = NULL;
2242 spin_unlock(&iclog->ic_callback_lock);
2244 /* perform callbacks in the order given */
2245 for (; cb; cb = cb_next) {
2246 cb_next = cb->cb_next;
2247 cb->cb_func(cb->cb_arg, aborted);
2249 spin_lock(&iclog->ic_callback_lock);
2250 cb = iclog->ic_callback;
2256 spin_lock(&log->l_icloglock);
2257 ASSERT(iclog->ic_callback == NULL);
2258 spin_unlock(&iclog->ic_callback_lock);
2259 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2260 iclog->ic_state = XLOG_STATE_DIRTY;
2263 * Transition from DIRTY to ACTIVE if applicable.
2264 * NOP if STATE_IOERROR.
2266 xlog_state_clean_log(log);
2268 /* wake up threads waiting in xfs_log_force() */
2269 sv_broadcast(&iclog->ic_force_wait);
2271 iclog = iclog->ic_next;
2272 } while (first_iclog != iclog);
2274 if (repeats > 5000) {
2275 flushcnt += repeats;
2277 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2278 "%s: possible infinite loop (%d iterations)",
2279 __func__, flushcnt);
2281 } while (!ioerrors && loopdidcallbacks);
2284 * make one last gasp attempt to see if iclogs are being left in
2288 if (funcdidcallbacks) {
2289 first_iclog = iclog = log->l_iclog;
2291 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2293 * Terminate the loop if iclogs are found in states
2294 * which will cause other threads to clean up iclogs.
2296 * SYNCING - i/o completion will go through logs
2297 * DONE_SYNC - interrupt thread should be waiting for
2299 * IOERROR - give up hope all ye who enter here
2301 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2302 iclog->ic_state == XLOG_STATE_SYNCING ||
2303 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2304 iclog->ic_state == XLOG_STATE_IOERROR )
2306 iclog = iclog->ic_next;
2307 } while (first_iclog != iclog);
2311 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2313 spin_unlock(&log->l_icloglock);
2316 sv_broadcast(&log->l_flush_wait);
2321 * Finish transitioning this iclog to the dirty state.
2323 * Make sure that we completely execute this routine only when this is
2324 * the last call to the iclog. There is a good chance that iclog flushes,
2325 * when we reach the end of the physical log, get turned into 2 separate
2326 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2327 * routine. By using the reference count bwritecnt, we guarantee that only
2328 * the second completion goes through.
2330 * Callbacks could take time, so they are done outside the scope of the
2331 * global state machine log lock.
2334 xlog_state_done_syncing(
2335 xlog_in_core_t *iclog,
2338 xlog_t *log = iclog->ic_log;
2340 spin_lock(&log->l_icloglock);
2342 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2343 iclog->ic_state == XLOG_STATE_IOERROR);
2344 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2345 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2349 * If we got an error, either on the first buffer, or in the case of
2350 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2351 * and none should ever be attempted to be written to disk
2354 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2355 if (--iclog->ic_bwritecnt == 1) {
2356 spin_unlock(&log->l_icloglock);
2359 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2363 * Someone could be sleeping prior to writing out the next
2364 * iclog buffer, we wake them all, one will get to do the
2365 * I/O, the others get to wait for the result.
2367 sv_broadcast(&iclog->ic_write_wait);
2368 spin_unlock(&log->l_icloglock);
2369 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2370 } /* xlog_state_done_syncing */
2374 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2375 * sleep. We wait on the flush queue on the head iclog as that should be
2376 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2377 * we will wait here and all new writes will sleep until a sync completes.
2379 * The in-core logs are used in a circular fashion. They are not used
2380 * out-of-order even when an iclog past the head is free.
2383 * * log_offset where xlog_write() can start writing into the in-core
2385 * * in-core log pointer to which xlog_write() should write.
2386 * * boolean indicating this is a continued write to an in-core log.
2387 * If this is the last write, then the in-core log's offset field
2388 * needs to be incremented, depending on the amount of data which
2392 xlog_state_get_iclog_space(xlog_t *log,
2394 xlog_in_core_t **iclogp,
2395 xlog_ticket_t *ticket,
2396 int *continued_write,
2400 xlog_rec_header_t *head;
2401 xlog_in_core_t *iclog;
2405 spin_lock(&log->l_icloglock);
2406 if (XLOG_FORCED_SHUTDOWN(log)) {
2407 spin_unlock(&log->l_icloglock);
2408 return XFS_ERROR(EIO);
2411 iclog = log->l_iclog;
2412 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2413 XFS_STATS_INC(xs_log_noiclogs);
2415 /* Wait for log writes to have flushed */
2416 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2420 head = &iclog->ic_header;
2422 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2423 log_offset = iclog->ic_offset;
2425 /* On the 1st write to an iclog, figure out lsn. This works
2426 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2427 * committing to. If the offset is set, that's how many blocks
2430 if (log_offset == 0) {
2431 ticket->t_curr_res -= log->l_iclog_hsize;
2432 xlog_tic_add_region(ticket,
2434 XLOG_REG_TYPE_LRHEADER);
2435 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2436 head->h_lsn = cpu_to_be64(
2437 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2438 ASSERT(log->l_curr_block >= 0);
2441 /* If there is enough room to write everything, then do it. Otherwise,
2442 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2443 * bit is on, so this will get flushed out. Don't update ic_offset
2444 * until you know exactly how many bytes get copied. Therefore, wait
2445 * until later to update ic_offset.
2447 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2448 * can fit into remaining data section.
2450 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2451 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2454 * If I'm the only one writing to this iclog, sync it to disk.
2455 * We need to do an atomic compare and decrement here to avoid
2456 * racing with concurrent atomic_dec_and_lock() calls in
2457 * xlog_state_release_iclog() when there is more than one
2458 * reference to the iclog.
2460 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2461 /* we are the only one */
2462 spin_unlock(&log->l_icloglock);
2463 error = xlog_state_release_iclog(log, iclog);
2467 spin_unlock(&log->l_icloglock);
2472 /* Do we have enough room to write the full amount in the remainder
2473 * of this iclog? Or must we continue a write on the next iclog and
2474 * mark this iclog as completely taken? In the case where we switch
2475 * iclogs (to mark it taken), this particular iclog will release/sync
2476 * to disk in xlog_write().
2478 if (len <= iclog->ic_size - iclog->ic_offset) {
2479 *continued_write = 0;
2480 iclog->ic_offset += len;
2482 *continued_write = 1;
2483 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2487 ASSERT(iclog->ic_offset <= iclog->ic_size);
2488 spin_unlock(&log->l_icloglock);
2490 *logoffsetp = log_offset;
2492 } /* xlog_state_get_iclog_space */
2495 * Atomically get the log space required for a log ticket.
2497 * Once a ticket gets put onto the reserveq, it will only return after
2498 * the needed reservation is satisfied.
2501 xlog_grant_log_space(xlog_t *log,
2512 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2513 panic("grant Recovery problem");
2516 /* Is there space or do we need to sleep? */
2517 spin_lock(&log->l_grant_lock);
2519 trace_xfs_log_grant_enter(log, tic);
2521 /* something is already sleeping; insert new transaction at end */
2522 if (!list_empty(&log->l_reserveq)) {
2523 list_add_tail(&tic->t_queue, &log->l_reserveq);
2525 trace_xfs_log_grant_sleep1(log, tic);
2528 * Gotta check this before going to sleep, while we're
2529 * holding the grant lock.
2531 if (XLOG_FORCED_SHUTDOWN(log))
2534 XFS_STATS_INC(xs_sleep_logspace);
2535 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2537 * If we got an error, and the filesystem is shutting down,
2538 * we'll catch it down below. So just continue...
2540 trace_xfs_log_grant_wake1(log, tic);
2541 spin_lock(&log->l_grant_lock);
2543 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2544 need_bytes = tic->t_unit_res*tic->t_ocnt;
2546 need_bytes = tic->t_unit_res;
2549 if (XLOG_FORCED_SHUTDOWN(log))
2552 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2553 log->l_grant_reserve_bytes);
2554 if (free_bytes < need_bytes) {
2555 if (list_empty(&tic->t_queue))
2556 list_add_tail(&tic->t_queue, &log->l_reserveq);
2558 trace_xfs_log_grant_sleep2(log, tic);
2560 spin_unlock(&log->l_grant_lock);
2561 xlog_grant_push_ail(log->l_mp, need_bytes);
2562 spin_lock(&log->l_grant_lock);
2564 XFS_STATS_INC(xs_sleep_logspace);
2565 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2567 spin_lock(&log->l_grant_lock);
2568 if (XLOG_FORCED_SHUTDOWN(log))
2571 trace_xfs_log_grant_wake2(log, tic);
2576 list_del_init(&tic->t_queue);
2578 /* we've got enough space */
2579 xlog_grant_add_space(log, need_bytes);
2581 tail_lsn = log->l_tail_lsn;
2583 * Check to make sure the grant write head didn't just over lap the
2584 * tail. If the cycles are the same, we can't be overlapping.
2585 * Otherwise, make sure that the cycles differ by exactly one and
2586 * check the byte count.
2588 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2589 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2590 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2593 trace_xfs_log_grant_exit(log, tic);
2594 xlog_verify_grant_head(log, 1);
2595 spin_unlock(&log->l_grant_lock);
2599 list_del_init(&tic->t_queue);
2600 trace_xfs_log_grant_error(log, tic);
2603 * If we are failing, make sure the ticket doesn't have any
2604 * current reservations. We don't want to add this back when
2605 * the ticket/transaction gets cancelled.
2607 tic->t_curr_res = 0;
2608 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2609 spin_unlock(&log->l_grant_lock);
2610 return XFS_ERROR(EIO);
2611 } /* xlog_grant_log_space */
2615 * Replenish the byte reservation required by moving the grant write head.
2620 xlog_regrant_write_log_space(xlog_t *log,
2623 int free_bytes, need_bytes;
2628 tic->t_curr_res = tic->t_unit_res;
2629 xlog_tic_reset_res(tic);
2635 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2636 panic("regrant Recovery problem");
2639 spin_lock(&log->l_grant_lock);
2641 trace_xfs_log_regrant_write_enter(log, tic);
2643 if (XLOG_FORCED_SHUTDOWN(log))
2646 /* If there are other waiters on the queue then give them a
2647 * chance at logspace before us. Wake up the first waiters,
2648 * if we do not wake up all the waiters then go to sleep waiting
2649 * for more free space, otherwise try to get some space for
2652 need_bytes = tic->t_unit_res;
2653 if (!list_empty(&log->l_writeq)) {
2654 struct xlog_ticket *ntic;
2655 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2656 log->l_grant_write_bytes);
2657 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2658 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2660 if (free_bytes < ntic->t_unit_res)
2662 free_bytes -= ntic->t_unit_res;
2663 sv_signal(&ntic->t_wait);
2666 if (ntic != list_first_entry(&log->l_writeq,
2667 struct xlog_ticket, t_queue)) {
2668 if (list_empty(&tic->t_queue))
2669 list_add_tail(&tic->t_queue, &log->l_writeq);
2671 trace_xfs_log_regrant_write_sleep1(log, tic);
2673 spin_unlock(&log->l_grant_lock);
2674 xlog_grant_push_ail(log->l_mp, need_bytes);
2675 spin_lock(&log->l_grant_lock);
2677 XFS_STATS_INC(xs_sleep_logspace);
2678 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2679 &log->l_grant_lock, s);
2681 /* If we're shutting down, this tic is already
2683 spin_lock(&log->l_grant_lock);
2684 if (XLOG_FORCED_SHUTDOWN(log))
2687 trace_xfs_log_regrant_write_wake1(log, tic);
2692 if (XLOG_FORCED_SHUTDOWN(log))
2695 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2696 log->l_grant_write_bytes);
2697 if (free_bytes < need_bytes) {
2698 if (list_empty(&tic->t_queue))
2699 list_add_tail(&tic->t_queue, &log->l_writeq);
2700 spin_unlock(&log->l_grant_lock);
2701 xlog_grant_push_ail(log->l_mp, need_bytes);
2702 spin_lock(&log->l_grant_lock);
2704 XFS_STATS_INC(xs_sleep_logspace);
2705 trace_xfs_log_regrant_write_sleep2(log, tic);
2707 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2709 /* If we're shutting down, this tic is already off the queue */
2710 spin_lock(&log->l_grant_lock);
2711 if (XLOG_FORCED_SHUTDOWN(log))
2714 trace_xfs_log_regrant_write_wake2(log, tic);
2718 list_del_init(&tic->t_queue);
2720 /* we've got enough space */
2721 xlog_grant_add_space_write(log, need_bytes);
2723 tail_lsn = log->l_tail_lsn;
2724 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2725 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2726 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2730 trace_xfs_log_regrant_write_exit(log, tic);
2732 xlog_verify_grant_head(log, 1);
2733 spin_unlock(&log->l_grant_lock);
2738 list_del_init(&tic->t_queue);
2739 trace_xfs_log_regrant_write_error(log, tic);
2742 * If we are failing, make sure the ticket doesn't have any
2743 * current reservations. We don't want to add this back when
2744 * the ticket/transaction gets cancelled.
2746 tic->t_curr_res = 0;
2747 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2748 spin_unlock(&log->l_grant_lock);
2749 return XFS_ERROR(EIO);
2750 } /* xlog_regrant_write_log_space */
2753 /* The first cnt-1 times through here we don't need to
2754 * move the grant write head because the permanent
2755 * reservation has reserved cnt times the unit amount.
2756 * Release part of current permanent unit reservation and
2757 * reset current reservation to be one units worth. Also
2758 * move grant reservation head forward.
2761 xlog_regrant_reserve_log_space(xlog_t *log,
2762 xlog_ticket_t *ticket)
2764 trace_xfs_log_regrant_reserve_enter(log, ticket);
2766 if (ticket->t_cnt > 0)
2769 spin_lock(&log->l_grant_lock);
2770 xlog_grant_sub_space(log, ticket->t_curr_res);
2771 ticket->t_curr_res = ticket->t_unit_res;
2772 xlog_tic_reset_res(ticket);
2774 trace_xfs_log_regrant_reserve_sub(log, ticket);
2776 xlog_verify_grant_head(log, 1);
2778 /* just return if we still have some of the pre-reserved space */
2779 if (ticket->t_cnt > 0) {
2780 spin_unlock(&log->l_grant_lock);
2784 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2786 trace_xfs_log_regrant_reserve_exit(log, ticket);
2788 xlog_verify_grant_head(log, 0);
2789 spin_unlock(&log->l_grant_lock);
2790 ticket->t_curr_res = ticket->t_unit_res;
2791 xlog_tic_reset_res(ticket);
2792 } /* xlog_regrant_reserve_log_space */
2796 * Give back the space left from a reservation.
2798 * All the information we need to make a correct determination of space left
2799 * is present. For non-permanent reservations, things are quite easy. The
2800 * count should have been decremented to zero. We only need to deal with the
2801 * space remaining in the current reservation part of the ticket. If the
2802 * ticket contains a permanent reservation, there may be left over space which
2803 * needs to be released. A count of N means that N-1 refills of the current
2804 * reservation can be done before we need to ask for more space. The first
2805 * one goes to fill up the first current reservation. Once we run out of
2806 * space, the count will stay at zero and the only space remaining will be
2807 * in the current reservation field.
2810 xlog_ungrant_log_space(xlog_t *log,
2811 xlog_ticket_t *ticket)
2813 if (ticket->t_cnt > 0)
2816 spin_lock(&log->l_grant_lock);
2817 trace_xfs_log_ungrant_enter(log, ticket);
2819 xlog_grant_sub_space(log, ticket->t_curr_res);
2821 trace_xfs_log_ungrant_sub(log, ticket);
2823 /* If this is a permanent reservation ticket, we may be able to free
2824 * up more space based on the remaining count.
2826 if (ticket->t_cnt > 0) {
2827 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2828 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2831 trace_xfs_log_ungrant_exit(log, ticket);
2833 xlog_verify_grant_head(log, 1);
2834 spin_unlock(&log->l_grant_lock);
2835 xfs_log_move_tail(log->l_mp, 1);
2836 } /* xlog_ungrant_log_space */
2840 * Flush iclog to disk if this is the last reference to the given iclog and
2841 * the WANT_SYNC bit is set.
2843 * When this function is entered, the iclog is not necessarily in the
2844 * WANT_SYNC state. It may be sitting around waiting to get filled.
2849 xlog_state_release_iclog(
2851 xlog_in_core_t *iclog)
2853 int sync = 0; /* do we sync? */
2855 if (iclog->ic_state & XLOG_STATE_IOERROR)
2856 return XFS_ERROR(EIO);
2858 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2859 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2862 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2863 spin_unlock(&log->l_icloglock);
2864 return XFS_ERROR(EIO);
2866 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2867 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2869 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2870 /* update tail before writing to iclog */
2871 xlog_assign_tail_lsn(log->l_mp);
2873 iclog->ic_state = XLOG_STATE_SYNCING;
2874 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2875 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2876 /* cycle incremented when incrementing curr_block */
2878 spin_unlock(&log->l_icloglock);
2881 * We let the log lock go, so it's possible that we hit a log I/O
2882 * error or some other SHUTDOWN condition that marks the iclog
2883 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2884 * this iclog has consistent data, so we ignore IOERROR
2885 * flags after this point.
2888 return xlog_sync(log, iclog);
2890 } /* xlog_state_release_iclog */
2894 * This routine will mark the current iclog in the ring as WANT_SYNC
2895 * and move the current iclog pointer to the next iclog in the ring.
2896 * When this routine is called from xlog_state_get_iclog_space(), the
2897 * exact size of the iclog has not yet been determined. All we know is
2898 * that every data block. We have run out of space in this log record.
2901 xlog_state_switch_iclogs(xlog_t *log,
2902 xlog_in_core_t *iclog,
2905 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2907 eventual_size = iclog->ic_offset;
2908 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2909 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2910 log->l_prev_block = log->l_curr_block;
2911 log->l_prev_cycle = log->l_curr_cycle;
2913 /* roll log?: ic_offset changed later */
2914 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2916 /* Round up to next log-sunit */
2917 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2918 log->l_mp->m_sb.sb_logsunit > 1) {
2919 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2920 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2923 if (log->l_curr_block >= log->l_logBBsize) {
2924 log->l_curr_cycle++;
2925 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2926 log->l_curr_cycle++;
2927 log->l_curr_block -= log->l_logBBsize;
2928 ASSERT(log->l_curr_block >= 0);
2930 ASSERT(iclog == log->l_iclog);
2931 log->l_iclog = iclog->ic_next;
2932 } /* xlog_state_switch_iclogs */
2935 * Write out all data in the in-core log as of this exact moment in time.
2937 * Data may be written to the in-core log during this call. However,
2938 * we don't guarantee this data will be written out. A change from past
2939 * implementation means this routine will *not* write out zero length LRs.
2941 * Basically, we try and perform an intelligent scan of the in-core logs.
2942 * If we determine there is no flushable data, we just return. There is no
2943 * flushable data if:
2945 * 1. the current iclog is active and has no data; the previous iclog
2946 * is in the active or dirty state.
2947 * 2. the current iclog is drity, and the previous iclog is in the
2948 * active or dirty state.
2952 * 1. the current iclog is not in the active nor dirty state.
2953 * 2. the current iclog dirty, and the previous iclog is not in the
2954 * active nor dirty state.
2955 * 3. the current iclog is active, and there is another thread writing
2956 * to this particular iclog.
2957 * 4. a) the current iclog is active and has no other writers
2958 * b) when we return from flushing out this iclog, it is still
2959 * not in the active nor dirty state.
2963 struct xfs_mount *mp,
2967 struct log *log = mp->m_log;
2968 struct xlog_in_core *iclog;
2971 XFS_STATS_INC(xs_log_force);
2974 xlog_cil_force(log);
2976 spin_lock(&log->l_icloglock);
2978 iclog = log->l_iclog;
2979 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2980 spin_unlock(&log->l_icloglock);
2981 return XFS_ERROR(EIO);
2984 /* If the head iclog is not active nor dirty, we just attach
2985 * ourselves to the head and go to sleep.
2987 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2988 iclog->ic_state == XLOG_STATE_DIRTY) {
2990 * If the head is dirty or (active and empty), then
2991 * we need to look at the previous iclog. If the previous
2992 * iclog is active or dirty we are done. There is nothing
2993 * to sync out. Otherwise, we attach ourselves to the
2994 * previous iclog and go to sleep.
2996 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2997 (atomic_read(&iclog->ic_refcnt) == 0
2998 && iclog->ic_offset == 0)) {
2999 iclog = iclog->ic_prev;
3000 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3001 iclog->ic_state == XLOG_STATE_DIRTY)
3006 if (atomic_read(&iclog->ic_refcnt) == 0) {
3007 /* We are the only one with access to this
3008 * iclog. Flush it out now. There should
3009 * be a roundoff of zero to show that someone
3010 * has already taken care of the roundoff from
3011 * the previous sync.
3013 atomic_inc(&iclog->ic_refcnt);
3014 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3015 xlog_state_switch_iclogs(log, iclog, 0);
3016 spin_unlock(&log->l_icloglock);
3018 if (xlog_state_release_iclog(log, iclog))
3019 return XFS_ERROR(EIO);
3023 spin_lock(&log->l_icloglock);
3024 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3025 iclog->ic_state != XLOG_STATE_DIRTY)
3030 /* Someone else is writing to this iclog.
3031 * Use its call to flush out the data. However,
3032 * the other thread may not force out this LR,
3033 * so we mark it WANT_SYNC.
3035 xlog_state_switch_iclogs(log, iclog, 0);
3041 /* By the time we come around again, the iclog could've been filled
3042 * which would give it another lsn. If we have a new lsn, just
3043 * return because the relevant data has been flushed.
3046 if (flags & XFS_LOG_SYNC) {
3048 * We must check if we're shutting down here, before
3049 * we wait, while we're holding the l_icloglock.
3050 * Then we check again after waking up, in case our
3051 * sleep was disturbed by a bad news.
3053 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3054 spin_unlock(&log->l_icloglock);
3055 return XFS_ERROR(EIO);
3057 XFS_STATS_INC(xs_log_force_sleep);
3058 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3060 * No need to grab the log lock here since we're
3061 * only deciding whether or not to return EIO
3062 * and the memory read should be atomic.
3064 if (iclog->ic_state & XLOG_STATE_IOERROR)
3065 return XFS_ERROR(EIO);
3071 spin_unlock(&log->l_icloglock);
3077 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3078 * about errors or whether the log was flushed or not. This is the normal
3079 * interface to use when trying to unpin items or move the log forward.
3088 error = _xfs_log_force(mp, flags, NULL);
3090 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3091 "error %d returned.", error);
3096 * Force the in-core log to disk for a specific LSN.
3098 * Find in-core log with lsn.
3099 * If it is in the DIRTY state, just return.
3100 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3101 * state and go to sleep or return.
3102 * If it is in any other state, go to sleep or return.
3104 * Synchronous forces are implemented with a signal variable. All callers
3105 * to force a given lsn to disk will wait on a the sv attached to the
3106 * specific in-core log. When given in-core log finally completes its
3107 * write to disk, that thread will wake up all threads waiting on the
3112 struct xfs_mount *mp,
3117 struct log *log = mp->m_log;
3118 struct xlog_in_core *iclog;
3119 int already_slept = 0;
3123 XFS_STATS_INC(xs_log_force);
3126 lsn = xlog_cil_force_lsn(log, lsn);
3127 if (lsn == NULLCOMMITLSN)
3132 spin_lock(&log->l_icloglock);
3133 iclog = log->l_iclog;
3134 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3135 spin_unlock(&log->l_icloglock);
3136 return XFS_ERROR(EIO);
3140 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3141 iclog = iclog->ic_next;
3145 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3146 spin_unlock(&log->l_icloglock);
3150 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3152 * We sleep here if we haven't already slept (e.g.
3153 * this is the first time we've looked at the correct
3154 * iclog buf) and the buffer before us is going to
3155 * be sync'ed. The reason for this is that if we
3156 * are doing sync transactions here, by waiting for
3157 * the previous I/O to complete, we can allow a few
3158 * more transactions into this iclog before we close
3161 * Otherwise, we mark the buffer WANT_SYNC, and bump
3162 * up the refcnt so we can release the log (which
3163 * drops the ref count). The state switch keeps new
3164 * transaction commits from using this buffer. When
3165 * the current commits finish writing into the buffer,
3166 * the refcount will drop to zero and the buffer will
3169 if (!already_slept &&
3170 (iclog->ic_prev->ic_state &
3171 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3172 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3174 XFS_STATS_INC(xs_log_force_sleep);
3176 sv_wait(&iclog->ic_prev->ic_write_wait,
3177 PSWP, &log->l_icloglock, s);
3183 atomic_inc(&iclog->ic_refcnt);
3184 xlog_state_switch_iclogs(log, iclog, 0);
3185 spin_unlock(&log->l_icloglock);
3186 if (xlog_state_release_iclog(log, iclog))
3187 return XFS_ERROR(EIO);
3190 spin_lock(&log->l_icloglock);
3193 if ((flags & XFS_LOG_SYNC) && /* sleep */
3195 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3197 * Don't wait on completion if we know that we've
3198 * gotten a log write error.
3200 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3201 spin_unlock(&log->l_icloglock);
3202 return XFS_ERROR(EIO);
3204 XFS_STATS_INC(xs_log_force_sleep);
3205 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3207 * No need to grab the log lock here since we're
3208 * only deciding whether or not to return EIO
3209 * and the memory read should be atomic.
3211 if (iclog->ic_state & XLOG_STATE_IOERROR)
3212 return XFS_ERROR(EIO);
3216 } else { /* just return */
3217 spin_unlock(&log->l_icloglock);
3221 } while (iclog != log->l_iclog);
3223 spin_unlock(&log->l_icloglock);
3228 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3229 * about errors or whether the log was flushed or not. This is the normal
3230 * interface to use when trying to unpin items or move the log forward.
3240 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3242 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3243 "error %d returned.", error);
3248 * Called when we want to mark the current iclog as being ready to sync to
3252 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3254 assert_spin_locked(&log->l_icloglock);
3256 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3257 xlog_state_switch_iclogs(log, iclog, 0);
3259 ASSERT(iclog->ic_state &
3260 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3265 /*****************************************************************************
3269 *****************************************************************************
3273 * Free a used ticket when its refcount falls to zero.
3277 xlog_ticket_t *ticket)
3279 ASSERT(atomic_read(&ticket->t_ref) > 0);
3280 if (atomic_dec_and_test(&ticket->t_ref)) {
3281 sv_destroy(&ticket->t_wait);
3282 kmem_zone_free(xfs_log_ticket_zone, ticket);
3288 xlog_ticket_t *ticket)
3290 ASSERT(atomic_read(&ticket->t_ref) > 0);
3291 atomic_inc(&ticket->t_ref);
3296 xfs_log_get_trans_ident(
3297 struct xfs_trans *tp)
3299 return tp->t_ticket->t_tid;
3303 * Allocate and initialise a new log ticket.
3314 struct xlog_ticket *tic;
3318 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3323 * Permanent reservations have up to 'cnt'-1 active log operations
3324 * in the log. A unit in this case is the amount of space for one
3325 * of these log operations. Normal reservations have a cnt of 1
3326 * and their unit amount is the total amount of space required.
3328 * The following lines of code account for non-transaction data
3329 * which occupy space in the on-disk log.
3331 * Normal form of a transaction is:
3332 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3333 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3335 * We need to account for all the leadup data and trailer data
3336 * around the transaction data.
3337 * And then we need to account for the worst case in terms of using
3339 * The worst case will happen if:
3340 * - the placement of the transaction happens to be such that the
3341 * roundoff is at its maximum
3342 * - the transaction data is synced before the commit record is synced
3343 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3344 * Therefore the commit record is in its own Log Record.
3345 * This can happen as the commit record is called with its
3346 * own region to xlog_write().
3347 * This then means that in the worst case, roundoff can happen for
3348 * the commit-rec as well.
3349 * The commit-rec is smaller than padding in this scenario and so it is
3350 * not added separately.
3353 /* for trans header */
3354 unit_bytes += sizeof(xlog_op_header_t);
3355 unit_bytes += sizeof(xfs_trans_header_t);
3358 unit_bytes += sizeof(xlog_op_header_t);
3361 * for LR headers - the space for data in an iclog is the size minus
3362 * the space used for the headers. If we use the iclog size, then we
3363 * undercalculate the number of headers required.
3365 * Furthermore - the addition of op headers for split-recs might
3366 * increase the space required enough to require more log and op
3367 * headers, so take that into account too.
3369 * IMPORTANT: This reservation makes the assumption that if this
3370 * transaction is the first in an iclog and hence has the LR headers
3371 * accounted to it, then the remaining space in the iclog is
3372 * exclusively for this transaction. i.e. if the transaction is larger
3373 * than the iclog, it will be the only thing in that iclog.
3374 * Fundamentally, this means we must pass the entire log vector to
3375 * xlog_write to guarantee this.
3377 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3378 num_headers = howmany(unit_bytes, iclog_space);
3380 /* for split-recs - ophdrs added when data split over LRs */
3381 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3383 /* add extra header reservations if we overrun */
3384 while (!num_headers ||
3385 howmany(unit_bytes, iclog_space) > num_headers) {
3386 unit_bytes += sizeof(xlog_op_header_t);
3389 unit_bytes += log->l_iclog_hsize * num_headers;
3391 /* for commit-rec LR header - note: padding will subsume the ophdr */
3392 unit_bytes += log->l_iclog_hsize;
3394 /* for roundoff padding for transaction data and one for commit record */
3395 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3396 log->l_mp->m_sb.sb_logsunit > 1) {
3397 /* log su roundoff */
3398 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3401 unit_bytes += 2*BBSIZE;
3404 atomic_set(&tic->t_ref, 1);
3405 INIT_LIST_HEAD(&tic->t_queue);
3406 tic->t_unit_res = unit_bytes;
3407 tic->t_curr_res = unit_bytes;
3410 tic->t_tid = random32();
3411 tic->t_clientid = client;
3412 tic->t_flags = XLOG_TIC_INITED;
3413 tic->t_trans_type = 0;
3414 if (xflags & XFS_LOG_PERM_RESERV)
3415 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3416 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3418 xlog_tic_reset_res(tic);
3424 /******************************************************************************
3426 * Log debug routines
3428 ******************************************************************************
3432 * Make sure that the destination ptr is within the valid data region of
3433 * one of the iclogs. This uses backup pointers stored in a different
3434 * part of the log in case we trash the log structure.
3437 xlog_verify_dest_ptr(
3444 for (i = 0; i < log->l_iclog_bufs; i++) {
3445 if (ptr >= log->l_iclog_bak[i] &&
3446 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3451 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3455 xlog_verify_grant_head(xlog_t *log, int equals)
3457 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3459 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3461 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3463 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3464 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3466 } /* xlog_verify_grant_head */
3468 /* check if it will fit */
3470 xlog_verify_tail_lsn(xlog_t *log,
3471 xlog_in_core_t *iclog,
3476 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3478 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3479 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3480 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3482 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3484 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3485 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3487 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3488 if (blocks < BTOBB(iclog->ic_offset) + 1)
3489 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3491 } /* xlog_verify_tail_lsn */
3494 * Perform a number of checks on the iclog before writing to disk.
3496 * 1. Make sure the iclogs are still circular
3497 * 2. Make sure we have a good magic number
3498 * 3. Make sure we don't have magic numbers in the data
3499 * 4. Check fields of each log operation header for:
3500 * A. Valid client identifier
3501 * B. tid ptr value falls in valid ptr space (user space code)
3502 * C. Length in log record header is correct according to the
3503 * individual operation headers within record.
3504 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3505 * log, check the preceding blocks of the physical log to make sure all
3506 * the cycle numbers agree with the current cycle number.
3509 xlog_verify_iclog(xlog_t *log,
3510 xlog_in_core_t *iclog,
3514 xlog_op_header_t *ophead;
3515 xlog_in_core_t *icptr;
3516 xlog_in_core_2_t *xhdr;
3518 xfs_caddr_t base_ptr;
3519 __psint_t field_offset;
3521 int len, i, j, k, op_len;
3524 /* check validity of iclog pointers */
3525 spin_lock(&log->l_icloglock);
3526 icptr = log->l_iclog;
3527 for (i=0; i < log->l_iclog_bufs; i++) {
3529 xlog_panic("xlog_verify_iclog: invalid ptr");
3530 icptr = icptr->ic_next;
3532 if (icptr != log->l_iclog)
3533 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3534 spin_unlock(&log->l_icloglock);
3536 /* check log magic numbers */
3537 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3538 xlog_panic("xlog_verify_iclog: invalid magic num");
3540 ptr = (xfs_caddr_t) &iclog->ic_header;
3541 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3543 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3544 xlog_panic("xlog_verify_iclog: unexpected magic num");
3548 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3549 ptr = iclog->ic_datap;
3551 ophead = (xlog_op_header_t *)ptr;
3552 xhdr = iclog->ic_data;
3553 for (i = 0; i < len; i++) {
3554 ophead = (xlog_op_header_t *)ptr;
3556 /* clientid is only 1 byte */
3557 field_offset = (__psint_t)
3558 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3559 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3560 clientid = ophead->oh_clientid;
3562 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3563 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3564 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3565 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3566 clientid = xlog_get_client_id(
3567 xhdr[j].hic_xheader.xh_cycle_data[k]);
3569 clientid = xlog_get_client_id(
3570 iclog->ic_header.h_cycle_data[idx]);
3573 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3574 cmn_err(CE_WARN, "xlog_verify_iclog: "
3575 "invalid clientid %d op 0x%p offset 0x%lx",
3576 clientid, ophead, (unsigned long)field_offset);
3579 field_offset = (__psint_t)
3580 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3581 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3582 op_len = be32_to_cpu(ophead->oh_len);
3584 idx = BTOBBT((__psint_t)&ophead->oh_len -
3585 (__psint_t)iclog->ic_datap);
3586 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3587 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3588 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3589 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3591 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3594 ptr += sizeof(xlog_op_header_t) + op_len;
3596 } /* xlog_verify_iclog */
3600 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3606 xlog_in_core_t *iclog, *ic;
3608 iclog = log->l_iclog;
3609 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3611 * Mark all the incore logs IOERROR.
3612 * From now on, no log flushes will result.
3616 ic->ic_state = XLOG_STATE_IOERROR;
3618 } while (ic != iclog);
3622 * Return non-zero, if state transition has already happened.
3628 * This is called from xfs_force_shutdown, when we're forcibly
3629 * shutting down the filesystem, typically because of an IO error.
3630 * Our main objectives here are to make sure that:
3631 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3632 * parties to find out, 'atomically'.
3633 * b. those who're sleeping on log reservations, pinned objects and
3634 * other resources get woken up, and be told the bad news.
3635 * c. nothing new gets queued up after (a) and (b) are done.
3636 * d. if !logerror, flush the iclogs to disk, then seal them off
3639 * Note: for delayed logging the !logerror case needs to flush the regions
3640 * held in memory out to the iclogs before flushing them to disk. This needs
3641 * to be done before the log is marked as shutdown, otherwise the flush to the
3645 xfs_log_force_umount(
3646 struct xfs_mount *mp,
3656 * If this happens during log recovery, don't worry about
3657 * locking; the log isn't open for business yet.
3660 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3661 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3663 XFS_BUF_DONE(mp->m_sb_bp);
3668 * Somebody could've already done the hard work for us.
3669 * No need to get locks for this.
3671 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3672 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3678 * Flush the in memory commit item list before marking the log as
3679 * being shut down. We need to do it in this order to ensure all the
3680 * completed transactions are flushed to disk with the xfs_log_force()
3683 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3684 xlog_cil_force(log);
3687 * We must hold both the GRANT lock and the LOG lock,
3688 * before we mark the filesystem SHUTDOWN and wake
3689 * everybody up to tell the bad news.
3691 spin_lock(&log->l_icloglock);
3692 spin_lock(&log->l_grant_lock);
3693 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3695 XFS_BUF_DONE(mp->m_sb_bp);
3698 * This flag is sort of redundant because of the mount flag, but
3699 * it's good to maintain the separation between the log and the rest
3702 log->l_flags |= XLOG_IO_ERROR;
3705 * If we hit a log error, we want to mark all the iclogs IOERROR
3706 * while we're still holding the loglock.
3709 retval = xlog_state_ioerror(log);
3710 spin_unlock(&log->l_icloglock);
3713 * We don't want anybody waiting for log reservations after this. That
3714 * means we have to wake up everybody queued up on reserveq as well as
3715 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3716 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3717 * action is protected by the GRANTLOCK.
3719 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3720 sv_signal(&tic->t_wait);
3722 list_for_each_entry(tic, &log->l_writeq, t_queue)
3723 sv_signal(&tic->t_wait);
3724 spin_unlock(&log->l_grant_lock);
3726 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3729 * Force the incore logs to disk before shutting the
3730 * log down completely.
3732 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3734 spin_lock(&log->l_icloglock);
3735 retval = xlog_state_ioerror(log);
3736 spin_unlock(&log->l_icloglock);
3739 * Wake up everybody waiting on xfs_log_force.
3740 * Callback all log item committed functions as if the
3741 * log writes were completed.
3743 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3745 #ifdef XFSERRORDEBUG
3747 xlog_in_core_t *iclog;
3749 spin_lock(&log->l_icloglock);
3750 iclog = log->l_iclog;
3752 ASSERT(iclog->ic_callback == 0);
3753 iclog = iclog->ic_next;
3754 } while (iclog != log->l_iclog);
3755 spin_unlock(&log->l_icloglock);
3758 /* return non-zero if log IOERROR transition had already happened */
3763 xlog_iclogs_empty(xlog_t *log)
3765 xlog_in_core_t *iclog;
3767 iclog = log->l_iclog;
3769 /* endianness does not matter here, zero is zero in
3772 if (iclog->ic_header.h_num_logops)
3774 iclog = iclog->ic_next;
3775 } while (iclog != log->l_iclog);