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);
100 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
104 tic->t_prev = (*qp)->t_prev;
105 (*qp)->t_prev->t_next = tic;
108 tic->t_prev = tic->t_next = tic;
112 tic->t_flags |= XLOG_TIC_IN_Q;
116 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
118 if (tic == tic->t_next) {
122 tic->t_next->t_prev = tic->t_prev;
123 tic->t_prev->t_next = tic->t_next;
126 tic->t_next = tic->t_prev = NULL;
127 tic->t_flags &= ~XLOG_TIC_IN_Q;
131 xlog_grant_sub_space(struct log *log, int bytes)
133 log->l_grant_write_bytes -= bytes;
134 if (log->l_grant_write_bytes < 0) {
135 log->l_grant_write_bytes += log->l_logsize;
136 log->l_grant_write_cycle--;
139 log->l_grant_reserve_bytes -= bytes;
140 if ((log)->l_grant_reserve_bytes < 0) {
141 log->l_grant_reserve_bytes += log->l_logsize;
142 log->l_grant_reserve_cycle--;
148 xlog_grant_add_space_write(struct log *log, int bytes)
150 int tmp = log->l_logsize - log->l_grant_write_bytes;
152 log->l_grant_write_bytes += bytes;
154 log->l_grant_write_cycle++;
155 log->l_grant_write_bytes = bytes - tmp;
160 xlog_grant_add_space_reserve(struct log *log, int bytes)
162 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
164 log->l_grant_reserve_bytes += bytes;
166 log->l_grant_reserve_cycle++;
167 log->l_grant_reserve_bytes = bytes - tmp;
172 xlog_grant_add_space(struct log *log, int bytes)
174 xlog_grant_add_space_write(log, bytes);
175 xlog_grant_add_space_reserve(log, bytes);
179 xlog_tic_reset_res(xlog_ticket_t *tic)
182 tic->t_res_arr_sum = 0;
183 tic->t_res_num_ophdrs = 0;
187 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
189 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
190 /* add to overflow and start again */
191 tic->t_res_o_flow += tic->t_res_arr_sum;
193 tic->t_res_arr_sum = 0;
196 tic->t_res_arr[tic->t_res_num].r_len = len;
197 tic->t_res_arr[tic->t_res_num].r_type = type;
198 tic->t_res_arr_sum += len;
205 * 1. currblock field gets updated at startup and after in-core logs
206 * marked as with WANT_SYNC.
210 * This routine is called when a user of a log manager ticket is done with
211 * the reservation. If the ticket was ever used, then a commit record for
212 * the associated transaction is written out as a log operation header with
213 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
214 * a given ticket. If the ticket was one with a permanent reservation, then
215 * a few operations are done differently. Permanent reservation tickets by
216 * default don't release the reservation. They just commit the current
217 * transaction with the belief that the reservation is still needed. A flag
218 * must be passed in before permanent reservations are actually released.
219 * When these type of tickets are not released, they need to be set into
220 * the inited state again. By doing this, a start record will be written
221 * out when the next write occurs.
225 struct xfs_mount *mp,
226 struct xlog_ticket *ticket,
227 struct xlog_in_core **iclog,
230 struct log *log = mp->m_log;
233 if (XLOG_FORCED_SHUTDOWN(log) ||
235 * If nothing was ever written, don't write out commit record.
236 * If we get an error, just continue and give back the log ticket.
238 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
239 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
240 lsn = (xfs_lsn_t) -1;
241 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
242 flags |= XFS_LOG_REL_PERM_RESERV;
247 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
248 (flags & XFS_LOG_REL_PERM_RESERV)) {
249 trace_xfs_log_done_nonperm(log, ticket);
252 * Release ticket if not permanent reservation or a specific
253 * request has been made to release a permanent reservation.
255 xlog_ungrant_log_space(log, ticket);
256 xfs_log_ticket_put(ticket);
258 trace_xfs_log_done_perm(log, ticket);
260 xlog_regrant_reserve_log_space(log, ticket);
261 /* If this ticket was a permanent reservation and we aren't
262 * trying to release it, reset the inited flags; so next time
263 * we write, a start record will be written out.
265 ticket->t_flags |= XLOG_TIC_INITED;
272 * Attaches a new iclog I/O completion callback routine during
273 * transaction commit. If the log is in error state, a non-zero
274 * return code is handed back and the caller is responsible for
275 * executing the callback at an appropriate time.
279 struct xfs_mount *mp,
280 struct xlog_in_core *iclog,
281 xfs_log_callback_t *cb)
285 spin_lock(&iclog->ic_callback_lock);
286 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
288 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
289 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
291 *(iclog->ic_callback_tail) = cb;
292 iclog->ic_callback_tail = &(cb->cb_next);
294 spin_unlock(&iclog->ic_callback_lock);
299 xfs_log_release_iclog(
300 struct xfs_mount *mp,
301 struct xlog_in_core *iclog)
303 if (xlog_state_release_iclog(mp->m_log, iclog)) {
304 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
312 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
313 * to the reservation.
314 * 2. Potentially, push buffers at tail of log to disk.
316 * Each reservation is going to reserve extra space for a log record header.
317 * When writes happen to the on-disk log, we don't subtract the length of the
318 * log record header from any reservation. By wasting space in each
319 * reservation, we prevent over allocation problems.
323 struct xfs_mount *mp,
326 struct xlog_ticket **ticket,
331 struct log *log = mp->m_log;
332 struct xlog_ticket *internal_ticket;
335 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
337 if (XLOG_FORCED_SHUTDOWN(log))
338 return XFS_ERROR(EIO);
340 XFS_STATS_INC(xs_try_logspace);
343 if (*ticket != NULL) {
344 ASSERT(flags & XFS_LOG_PERM_RESERV);
345 internal_ticket = *ticket;
348 * this is a new transaction on the ticket, so we need to
349 * change the transaction ID so that the next transaction has a
350 * different TID in the log. Just add one to the existing tid
351 * so that we can see chains of rolling transactions in the log
354 internal_ticket->t_tid++;
356 trace_xfs_log_reserve(log, internal_ticket);
358 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
359 retval = xlog_regrant_write_log_space(log, internal_ticket);
361 /* may sleep if need to allocate more tickets */
362 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
364 KM_SLEEP|KM_MAYFAIL);
365 if (!internal_ticket)
366 return XFS_ERROR(ENOMEM);
367 internal_ticket->t_trans_type = t_type;
368 *ticket = internal_ticket;
370 trace_xfs_log_reserve(log, internal_ticket);
372 xlog_grant_push_ail(mp,
373 (internal_ticket->t_unit_res *
374 internal_ticket->t_cnt));
375 retval = xlog_grant_log_space(log, internal_ticket);
379 } /* xfs_log_reserve */
383 * Mount a log filesystem
385 * mp - ubiquitous xfs mount point structure
386 * log_target - buftarg of on-disk log device
387 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
388 * num_bblocks - Number of BBSIZE blocks in on-disk log
390 * Return error or zero.
395 xfs_buftarg_t *log_target,
396 xfs_daddr_t blk_offset,
401 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
402 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
405 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
407 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
410 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
411 if (IS_ERR(mp->m_log)) {
412 error = -PTR_ERR(mp->m_log);
417 * Initialize the AIL now we have a log.
419 error = xfs_trans_ail_init(mp);
421 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
424 mp->m_log->l_ailp = mp->m_ail;
427 * skip log recovery on a norecovery mount. pretend it all
430 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
431 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
434 mp->m_flags &= ~XFS_MOUNT_RDONLY;
436 error = xlog_recover(mp->m_log);
439 mp->m_flags |= XFS_MOUNT_RDONLY;
441 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
442 goto out_destroy_ail;
446 /* Normal transactions can now occur */
447 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
450 * Now the log has been fully initialised and we know were our
451 * space grant counters are, we can initialise the permanent ticket
452 * needed for delayed logging to work.
454 xlog_cil_init_post_recovery(mp->m_log);
459 xfs_trans_ail_destroy(mp);
461 xlog_dealloc_log(mp->m_log);
467 * Finish the recovery of the file system. This is separate from
468 * the xfs_log_mount() call, because it depends on the code in
469 * xfs_mountfs() to read in the root and real-time bitmap inodes
470 * between calling xfs_log_mount() and here.
472 * mp - ubiquitous xfs mount point structure
475 xfs_log_mount_finish(xfs_mount_t *mp)
479 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
480 error = xlog_recover_finish(mp->m_log);
483 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
490 * Final log writes as part of unmount.
492 * Mark the filesystem clean as unmount happens. Note that during relocation
493 * this routine needs to be executed as part of source-bag while the
494 * deallocation must not be done until source-end.
498 * Unmount record used to have a string "Unmount filesystem--" in the
499 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
500 * We just write the magic number now since that particular field isn't
501 * currently architecture converted and "nUmount" is a bit foo.
502 * As far as I know, there weren't any dependencies on the old behaviour.
506 xfs_log_unmount_write(xfs_mount_t *mp)
508 xlog_t *log = mp->m_log;
509 xlog_in_core_t *iclog;
511 xlog_in_core_t *first_iclog;
513 xlog_ticket_t *tic = NULL;
518 * Don't write out unmount record on read-only mounts.
519 * Or, if we are doing a forced umount (typically because of IO errors).
521 if (mp->m_flags & XFS_MOUNT_RDONLY)
524 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
525 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
528 first_iclog = iclog = log->l_iclog;
530 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
531 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
532 ASSERT(iclog->ic_offset == 0);
534 iclog = iclog->ic_next;
535 } while (iclog != first_iclog);
537 if (! (XLOG_FORCED_SHUTDOWN(log))) {
538 error = xfs_log_reserve(mp, 600, 1, &tic,
539 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
541 /* the data section must be 32 bit size aligned */
545 __uint32_t pad2; /* may as well make it 64 bits */
547 .magic = XLOG_UNMOUNT_TYPE,
549 struct xfs_log_iovec reg = {
551 .i_len = sizeof(magic),
552 .i_type = XLOG_REG_TYPE_UNMOUNT,
554 struct xfs_log_vec vec = {
559 /* remove inited flag */
561 error = xlog_write(log, &vec, tic, &lsn,
562 NULL, XLOG_UNMOUNT_TRANS);
564 * At this point, we're umounting anyway,
565 * so there's no point in transitioning log state
566 * to IOERROR. Just continue...
571 xfs_fs_cmn_err(CE_ALERT, mp,
572 "xfs_log_unmount: unmount record failed");
576 spin_lock(&log->l_icloglock);
577 iclog = log->l_iclog;
578 atomic_inc(&iclog->ic_refcnt);
579 xlog_state_want_sync(log, iclog);
580 spin_unlock(&log->l_icloglock);
581 error = xlog_state_release_iclog(log, iclog);
583 spin_lock(&log->l_icloglock);
584 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
585 iclog->ic_state == XLOG_STATE_DIRTY)) {
586 if (!XLOG_FORCED_SHUTDOWN(log)) {
587 sv_wait(&iclog->ic_force_wait, PMEM,
588 &log->l_icloglock, s);
590 spin_unlock(&log->l_icloglock);
593 spin_unlock(&log->l_icloglock);
596 trace_xfs_log_umount_write(log, tic);
597 xlog_ungrant_log_space(log, tic);
598 xfs_log_ticket_put(tic);
602 * We're already in forced_shutdown mode, couldn't
603 * even attempt to write out the unmount transaction.
605 * Go through the motions of sync'ing and releasing
606 * the iclog, even though no I/O will actually happen,
607 * we need to wait for other log I/Os that may already
608 * be in progress. Do this as a separate section of
609 * code so we'll know if we ever get stuck here that
610 * we're in this odd situation of trying to unmount
611 * a file system that went into forced_shutdown as
612 * the result of an unmount..
614 spin_lock(&log->l_icloglock);
615 iclog = log->l_iclog;
616 atomic_inc(&iclog->ic_refcnt);
618 xlog_state_want_sync(log, iclog);
619 spin_unlock(&log->l_icloglock);
620 error = xlog_state_release_iclog(log, iclog);
622 spin_lock(&log->l_icloglock);
624 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
625 || iclog->ic_state == XLOG_STATE_DIRTY
626 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
628 sv_wait(&iclog->ic_force_wait, PMEM,
629 &log->l_icloglock, s);
631 spin_unlock(&log->l_icloglock);
636 } /* xfs_log_unmount_write */
639 * Deallocate log structures for unmount/relocation.
641 * We need to stop the aild from running before we destroy
642 * and deallocate the log as the aild references the log.
645 xfs_log_unmount(xfs_mount_t *mp)
647 xfs_trans_ail_destroy(mp);
648 xlog_dealloc_log(mp->m_log);
653 struct xfs_mount *mp,
654 struct xfs_log_item *item,
656 struct xfs_item_ops *ops)
658 item->li_mountp = mp;
659 item->li_ailp = mp->m_ail;
660 item->li_type = type;
664 INIT_LIST_HEAD(&item->li_ail);
665 INIT_LIST_HEAD(&item->li_cil);
669 * Write region vectors to log. The write happens using the space reservation
670 * of the ticket (tic). It is not a requirement that all writes for a given
671 * transaction occur with one call to xfs_log_write(). However, it is important
672 * to note that the transaction reservation code makes an assumption about the
673 * number of log headers a transaction requires that may be violated if you
674 * don't pass all the transaction vectors in one call....
678 struct xfs_mount *mp,
679 struct xfs_log_iovec reg[],
681 struct xlog_ticket *tic,
682 xfs_lsn_t *start_lsn)
684 struct log *log = mp->m_log;
686 struct xfs_log_vec vec = {
687 .lv_niovecs = nentries,
691 if (XLOG_FORCED_SHUTDOWN(log))
692 return XFS_ERROR(EIO);
694 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
696 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
701 xfs_log_move_tail(xfs_mount_t *mp,
705 xlog_t *log = mp->m_log;
706 int need_bytes, free_bytes, cycle, bytes;
708 if (XLOG_FORCED_SHUTDOWN(log))
712 /* needed since sync_lsn is 64 bits */
713 spin_lock(&log->l_icloglock);
714 tail_lsn = log->l_last_sync_lsn;
715 spin_unlock(&log->l_icloglock);
718 spin_lock(&log->l_grant_lock);
720 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
724 log->l_tail_lsn = tail_lsn;
727 if ((tic = log->l_write_headq)) {
729 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
730 panic("Recovery problem");
732 cycle = log->l_grant_write_cycle;
733 bytes = log->l_grant_write_bytes;
734 free_bytes = xlog_space_left(log, cycle, bytes);
736 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
738 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
741 free_bytes -= tic->t_unit_res;
742 sv_signal(&tic->t_wait);
744 } while (tic != log->l_write_headq);
746 if ((tic = log->l_reserve_headq)) {
748 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
749 panic("Recovery problem");
751 cycle = log->l_grant_reserve_cycle;
752 bytes = log->l_grant_reserve_bytes;
753 free_bytes = xlog_space_left(log, cycle, bytes);
755 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
756 need_bytes = tic->t_unit_res*tic->t_cnt;
758 need_bytes = tic->t_unit_res;
759 if (free_bytes < need_bytes && tail_lsn != 1)
762 free_bytes -= need_bytes;
763 sv_signal(&tic->t_wait);
765 } while (tic != log->l_reserve_headq);
767 spin_unlock(&log->l_grant_lock);
768 } /* xfs_log_move_tail */
771 * Determine if we have a transaction that has gone to disk
772 * that needs to be covered. To begin the transition to the idle state
773 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
774 * If we are then in a state where covering is needed, the caller is informed
775 * that dummy transactions are required to move the log into the idle state.
777 * Because this is called as part of the sync process, we should also indicate
778 * that dummy transactions should be issued in anything but the covered or
779 * idle states. This ensures that the log tail is accurately reflected in
780 * the log at the end of the sync, hence if a crash occurrs avoids replay
781 * of transactions where the metadata is already on disk.
784 xfs_log_need_covered(xfs_mount_t *mp)
787 xlog_t *log = mp->m_log;
789 if (!xfs_fs_writable(mp))
792 spin_lock(&log->l_icloglock);
793 switch (log->l_covered_state) {
794 case XLOG_STATE_COVER_DONE:
795 case XLOG_STATE_COVER_DONE2:
796 case XLOG_STATE_COVER_IDLE:
798 case XLOG_STATE_COVER_NEED:
799 case XLOG_STATE_COVER_NEED2:
800 if (!xfs_trans_ail_tail(log->l_ailp) &&
801 xlog_iclogs_empty(log)) {
802 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
803 log->l_covered_state = XLOG_STATE_COVER_DONE;
805 log->l_covered_state = XLOG_STATE_COVER_DONE2;
812 spin_unlock(&log->l_icloglock);
816 /******************************************************************************
820 ******************************************************************************
823 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
824 * The log manager must keep track of the last LR which was committed
825 * to disk. The lsn of this LR will become the new tail_lsn whenever
826 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
827 * the situation where stuff could be written into the log but nothing
828 * was ever in the AIL when asked. Eventually, we panic since the
829 * tail hits the head.
831 * We may be holding the log iclog lock upon entering this routine.
834 xlog_assign_tail_lsn(xfs_mount_t *mp)
837 xlog_t *log = mp->m_log;
839 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
840 spin_lock(&log->l_grant_lock);
842 log->l_tail_lsn = tail_lsn;
844 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
846 spin_unlock(&log->l_grant_lock);
849 } /* xlog_assign_tail_lsn */
853 * Return the space in the log between the tail and the head. The head
854 * is passed in the cycle/bytes formal parms. In the special case where
855 * the reserve head has wrapped passed the tail, this calculation is no
856 * longer valid. In this case, just return 0 which means there is no space
857 * in the log. This works for all places where this function is called
858 * with the reserve head. Of course, if the write head were to ever
859 * wrap the tail, we should blow up. Rather than catch this case here,
860 * we depend on other ASSERTions in other parts of the code. XXXmiken
862 * This code also handles the case where the reservation head is behind
863 * the tail. The details of this case are described below, but the end
864 * result is that we return the size of the log as the amount of space left.
867 xlog_space_left(xlog_t *log, int cycle, int bytes)
873 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
874 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
875 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
876 free_bytes = log->l_logsize - (bytes - tail_bytes);
877 } else if ((tail_cycle + 1) < cycle) {
879 } else if (tail_cycle < cycle) {
880 ASSERT(tail_cycle == (cycle - 1));
881 free_bytes = tail_bytes - bytes;
884 * The reservation head is behind the tail.
885 * In this case we just want to return the size of the
886 * log as the amount of space left.
888 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
889 "xlog_space_left: head behind tail\n"
890 " tail_cycle = %d, tail_bytes = %d\n"
891 " GH cycle = %d, GH bytes = %d",
892 tail_cycle, tail_bytes, cycle, bytes);
894 free_bytes = log->l_logsize;
897 } /* xlog_space_left */
901 * Log function which is called when an io completes.
903 * The log manager needs its own routine, in order to control what
904 * happens with the buffer after the write completes.
907 xlog_iodone(xfs_buf_t *bp)
909 xlog_in_core_t *iclog;
913 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
914 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
915 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
920 * If the _XFS_BARRIER_FAILED flag was set by a lower
921 * layer, it means the underlying device no longer supports
922 * barrier I/O. Warn loudly and turn off barriers.
924 if (bp->b_flags & _XFS_BARRIER_FAILED) {
925 bp->b_flags &= ~_XFS_BARRIER_FAILED;
926 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
927 xfs_fs_cmn_err(CE_WARN, l->l_mp,
928 "xlog_iodone: Barriers are no longer supported"
929 " by device. Disabling barriers\n");
933 * Race to shutdown the filesystem if we see an error.
935 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
936 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
937 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
939 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
941 * This flag will be propagated to the trans-committed
942 * callback routines to let them know that the log-commit
945 aborted = XFS_LI_ABORTED;
946 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
947 aborted = XFS_LI_ABORTED;
950 /* log I/O is always issued ASYNC */
951 ASSERT(XFS_BUF_ISASYNC(bp));
952 xlog_state_done_syncing(iclog, aborted);
954 * do not reference the buffer (bp) here as we could race
955 * with it being freed after writing the unmount record to the
962 * Return size of each in-core log record buffer.
964 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
966 * If the filesystem blocksize is too large, we may need to choose a
967 * larger size since the directory code currently logs entire blocks.
971 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
977 if (mp->m_logbufs <= 0)
978 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
980 log->l_iclog_bufs = mp->m_logbufs;
983 * Buffer size passed in from mount system call.
985 if (mp->m_logbsize > 0) {
986 size = log->l_iclog_size = mp->m_logbsize;
987 log->l_iclog_size_log = 0;
989 log->l_iclog_size_log++;
993 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
994 /* # headers = size / 32k
995 * one header holds cycles from 32k of data
998 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
999 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1001 log->l_iclog_hsize = xhdrs << BBSHIFT;
1002 log->l_iclog_heads = xhdrs;
1004 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1005 log->l_iclog_hsize = BBSIZE;
1006 log->l_iclog_heads = 1;
1011 /* All machines use 32kB buffers by default. */
1012 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1013 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1015 /* the default log size is 16k or 32k which is one header sector */
1016 log->l_iclog_hsize = BBSIZE;
1017 log->l_iclog_heads = 1;
1020 /* are we being asked to make the sizes selected above visible? */
1021 if (mp->m_logbufs == 0)
1022 mp->m_logbufs = log->l_iclog_bufs;
1023 if (mp->m_logbsize == 0)
1024 mp->m_logbsize = log->l_iclog_size;
1025 } /* xlog_get_iclog_buffer_size */
1029 * This routine initializes some of the log structure for a given mount point.
1030 * Its primary purpose is to fill in enough, so recovery can occur. However,
1031 * some other stuff may be filled in too.
1034 xlog_alloc_log(xfs_mount_t *mp,
1035 xfs_buftarg_t *log_target,
1036 xfs_daddr_t blk_offset,
1040 xlog_rec_header_t *head;
1041 xlog_in_core_t **iclogp;
1042 xlog_in_core_t *iclog, *prev_iclog=NULL;
1049 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1051 xlog_warn("XFS: Log allocation failed: No memory!");
1056 log->l_targ = log_target;
1057 log->l_logsize = BBTOB(num_bblks);
1058 log->l_logBBstart = blk_offset;
1059 log->l_logBBsize = num_bblks;
1060 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1061 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1063 log->l_prev_block = -1;
1064 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1065 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1066 log->l_last_sync_lsn = log->l_tail_lsn;
1067 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1068 log->l_grant_reserve_cycle = 1;
1069 log->l_grant_write_cycle = 1;
1071 error = EFSCORRUPTED;
1072 if (xfs_sb_version_hassector(&mp->m_sb)) {
1073 log2_size = mp->m_sb.sb_logsectlog;
1074 if (log2_size < BBSHIFT) {
1075 xlog_warn("XFS: Log sector size too small "
1076 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1080 log2_size -= BBSHIFT;
1081 if (log2_size > mp->m_sectbb_log) {
1082 xlog_warn("XFS: Log sector size too large "
1083 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1087 /* for larger sector sizes, must have v2 or external log */
1088 if (log2_size && log->l_logBBstart > 0 &&
1089 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1091 xlog_warn("XFS: log sector size (0x%x) invalid "
1092 "for configuration.", log2_size);
1096 log->l_sectBBsize = 1 << log2_size;
1098 xlog_get_iclog_buffer_size(mp, log);
1101 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1104 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1105 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1106 ASSERT(XFS_BUF_ISBUSY(bp));
1107 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1110 spin_lock_init(&log->l_icloglock);
1111 spin_lock_init(&log->l_grant_lock);
1112 sv_init(&log->l_flush_wait, 0, "flush_wait");
1114 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1115 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1117 iclogp = &log->l_iclog;
1119 * The amount of memory to allocate for the iclog structure is
1120 * rather funky due to the way the structure is defined. It is
1121 * done this way so that we can use different sizes for machines
1122 * with different amounts of memory. See the definition of
1123 * xlog_in_core_t in xfs_log_priv.h for details.
1125 iclogsize = log->l_iclog_size;
1126 ASSERT(log->l_iclog_size >= 4096);
1127 for (i=0; i < log->l_iclog_bufs; i++) {
1128 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1130 goto out_free_iclog;
1133 iclog->ic_prev = prev_iclog;
1136 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1138 goto out_free_iclog;
1139 if (!XFS_BUF_CPSEMA(bp))
1141 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1142 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1144 iclog->ic_data = bp->b_addr;
1146 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1148 head = &iclog->ic_header;
1149 memset(head, 0, sizeof(xlog_rec_header_t));
1150 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1151 head->h_version = cpu_to_be32(
1152 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1153 head->h_size = cpu_to_be32(log->l_iclog_size);
1155 head->h_fmt = cpu_to_be32(XLOG_FMT);
1156 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1158 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1159 iclog->ic_state = XLOG_STATE_ACTIVE;
1160 iclog->ic_log = log;
1161 atomic_set(&iclog->ic_refcnt, 0);
1162 spin_lock_init(&iclog->ic_callback_lock);
1163 iclog->ic_callback_tail = &(iclog->ic_callback);
1164 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1166 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1167 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1168 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1169 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1171 iclogp = &iclog->ic_next;
1173 *iclogp = log->l_iclog; /* complete ring */
1174 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1176 error = xlog_cil_init(log);
1178 goto out_free_iclog;
1182 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1183 prev_iclog = iclog->ic_next;
1185 sv_destroy(&iclog->ic_force_wait);
1186 sv_destroy(&iclog->ic_write_wait);
1187 xfs_buf_free(iclog->ic_bp);
1191 spinlock_destroy(&log->l_icloglock);
1192 spinlock_destroy(&log->l_grant_lock);
1193 xfs_buf_free(log->l_xbuf);
1197 return ERR_PTR(-error);
1198 } /* xlog_alloc_log */
1202 * Write out the commit record of a transaction associated with the given
1203 * ticket. Return the lsn of the commit record.
1208 struct xlog_ticket *ticket,
1209 struct xlog_in_core **iclog,
1210 xfs_lsn_t *commitlsnp)
1212 struct xfs_mount *mp = log->l_mp;
1214 struct xfs_log_iovec reg = {
1217 .i_type = XLOG_REG_TYPE_COMMIT,
1219 struct xfs_log_vec vec = {
1224 ASSERT_ALWAYS(iclog);
1225 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1228 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1233 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1234 * log space. This code pushes on the lsn which would supposedly free up
1235 * the 25% which we want to leave free. We may need to adopt a policy which
1236 * pushes on an lsn which is further along in the log once we reach the high
1237 * water mark. In this manner, we would be creating a low water mark.
1240 xlog_grant_push_ail(xfs_mount_t *mp,
1243 xlog_t *log = mp->m_log; /* pointer to the log */
1244 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1245 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1246 int free_blocks; /* free blocks left to write to */
1247 int free_bytes; /* free bytes left to write to */
1248 int threshold_block; /* block in lsn we'd like to be at */
1249 int threshold_cycle; /* lsn cycle we'd like to be at */
1252 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1254 spin_lock(&log->l_grant_lock);
1255 free_bytes = xlog_space_left(log,
1256 log->l_grant_reserve_cycle,
1257 log->l_grant_reserve_bytes);
1258 tail_lsn = log->l_tail_lsn;
1259 free_blocks = BTOBBT(free_bytes);
1262 * Set the threshold for the minimum number of free blocks in the
1263 * log to the maximum of what the caller needs, one quarter of the
1264 * log, and 256 blocks.
1266 free_threshold = BTOBB(need_bytes);
1267 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1268 free_threshold = MAX(free_threshold, 256);
1269 if (free_blocks < free_threshold) {
1270 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1271 threshold_cycle = CYCLE_LSN(tail_lsn);
1272 if (threshold_block >= log->l_logBBsize) {
1273 threshold_block -= log->l_logBBsize;
1274 threshold_cycle += 1;
1276 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1278 /* Don't pass in an lsn greater than the lsn of the last
1279 * log record known to be on disk.
1281 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1282 threshold_lsn = log->l_last_sync_lsn;
1284 spin_unlock(&log->l_grant_lock);
1287 * Get the transaction layer to kick the dirty buffers out to
1288 * disk asynchronously. No point in trying to do this if
1289 * the filesystem is shutting down.
1291 if (threshold_lsn &&
1292 !XLOG_FORCED_SHUTDOWN(log))
1293 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1294 } /* xlog_grant_push_ail */
1297 * The bdstrat callback function for log bufs. This gives us a central
1298 * place to trap bufs in case we get hit by a log I/O error and need to
1299 * shutdown. Actually, in practice, even when we didn't get a log error,
1300 * we transition the iclogs to IOERROR state *after* flushing all existing
1301 * iclogs to disk. This is because we don't want anymore new transactions to be
1302 * started or completed afterwards.
1308 struct xlog_in_core *iclog;
1310 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1311 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1312 XFS_BUF_ERROR(bp, EIO);
1316 * It would seem logical to return EIO here, but we rely on
1317 * the log state machine to propagate I/O errors instead of
1323 bp->b_flags |= _XBF_RUN_QUEUES;
1324 xfs_buf_iorequest(bp);
1329 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1330 * fashion. Previously, we should have moved the current iclog
1331 * ptr in the log to point to the next available iclog. This allows further
1332 * write to continue while this code syncs out an iclog ready to go.
1333 * Before an in-core log can be written out, the data section must be scanned
1334 * to save away the 1st word of each BBSIZE block into the header. We replace
1335 * it with the current cycle count. Each BBSIZE block is tagged with the
1336 * cycle count because there in an implicit assumption that drives will
1337 * guarantee that entire 512 byte blocks get written at once. In other words,
1338 * we can't have part of a 512 byte block written and part not written. By
1339 * tagging each block, we will know which blocks are valid when recovering
1340 * after an unclean shutdown.
1342 * This routine is single threaded on the iclog. No other thread can be in
1343 * this routine with the same iclog. Changing contents of iclog can there-
1344 * fore be done without grabbing the state machine lock. Updating the global
1345 * log will require grabbing the lock though.
1347 * The entire log manager uses a logical block numbering scheme. Only
1348 * log_sync (and then only bwrite()) know about the fact that the log may
1349 * not start with block zero on a given device. The log block start offset
1350 * is added immediately before calling bwrite().
1354 xlog_sync(xlog_t *log,
1355 xlog_in_core_t *iclog)
1357 xfs_caddr_t dptr; /* pointer to byte sized element */
1360 uint count; /* byte count of bwrite */
1361 uint count_init; /* initial count before roundup */
1362 int roundoff; /* roundoff to BB or stripe */
1363 int split = 0; /* split write into two regions */
1365 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1367 XFS_STATS_INC(xs_log_writes);
1368 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1370 /* Add for LR header */
1371 count_init = log->l_iclog_hsize + iclog->ic_offset;
1373 /* Round out the log write size */
1374 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1375 /* we have a v2 stripe unit to use */
1376 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1378 count = BBTOB(BTOBB(count_init));
1380 roundoff = count - count_init;
1381 ASSERT(roundoff >= 0);
1382 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1383 roundoff < log->l_mp->m_sb.sb_logsunit)
1385 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1386 roundoff < BBTOB(1)));
1388 /* move grant heads by roundoff in sync */
1389 spin_lock(&log->l_grant_lock);
1390 xlog_grant_add_space(log, roundoff);
1391 spin_unlock(&log->l_grant_lock);
1393 /* put cycle number in every block */
1394 xlog_pack_data(log, iclog, roundoff);
1396 /* real byte length */
1398 iclog->ic_header.h_len =
1399 cpu_to_be32(iclog->ic_offset + roundoff);
1401 iclog->ic_header.h_len =
1402 cpu_to_be32(iclog->ic_offset);
1406 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1407 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1408 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1410 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1412 /* Do we need to split this write into 2 parts? */
1413 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1414 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1415 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1416 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1418 iclog->ic_bwritecnt = 1;
1420 XFS_BUF_SET_COUNT(bp, count);
1421 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1422 XFS_BUF_ZEROFLAGS(bp);
1425 bp->b_flags |= XBF_LOG_BUFFER;
1427 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1428 XFS_BUF_ORDERED(bp);
1430 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1431 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1433 xlog_verify_iclog(log, iclog, count, B_TRUE);
1435 /* account for log which doesn't start at block #0 */
1436 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1438 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1443 if ((error = xlog_bdstrat(bp))) {
1444 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1449 bp = iclog->ic_log->l_xbuf;
1450 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1452 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1453 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1454 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1455 (__psint_t)count), split);
1456 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1457 XFS_BUF_ZEROFLAGS(bp);
1460 bp->b_flags |= XBF_LOG_BUFFER;
1461 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1462 XFS_BUF_ORDERED(bp);
1463 dptr = XFS_BUF_PTR(bp);
1465 * Bump the cycle numbers at the start of each block
1466 * since this part of the buffer is at the start of
1467 * a new cycle. Watch out for the header magic number
1470 for (i = 0; i < split; i += BBSIZE) {
1471 be32_add_cpu((__be32 *)dptr, 1);
1472 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1473 be32_add_cpu((__be32 *)dptr, 1);
1477 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1478 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1480 /* account for internal log which doesn't start at block #0 */
1481 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1483 if ((error = xlog_bdstrat(bp))) {
1484 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1485 bp, XFS_BUF_ADDR(bp));
1494 * Deallocate a log structure
1497 xlog_dealloc_log(xlog_t *log)
1499 xlog_in_core_t *iclog, *next_iclog;
1502 xlog_cil_destroy(log);
1504 iclog = log->l_iclog;
1505 for (i=0; i<log->l_iclog_bufs; i++) {
1506 sv_destroy(&iclog->ic_force_wait);
1507 sv_destroy(&iclog->ic_write_wait);
1508 xfs_buf_free(iclog->ic_bp);
1509 next_iclog = iclog->ic_next;
1513 spinlock_destroy(&log->l_icloglock);
1514 spinlock_destroy(&log->l_grant_lock);
1516 xfs_buf_free(log->l_xbuf);
1517 log->l_mp->m_log = NULL;
1519 } /* xlog_dealloc_log */
1522 * Update counters atomically now that memcpy is done.
1526 xlog_state_finish_copy(xlog_t *log,
1527 xlog_in_core_t *iclog,
1531 spin_lock(&log->l_icloglock);
1533 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1534 iclog->ic_offset += copy_bytes;
1536 spin_unlock(&log->l_icloglock);
1537 } /* xlog_state_finish_copy */
1543 * print out info relating to regions written which consume
1548 struct xfs_mount *mp,
1549 struct xlog_ticket *ticket)
1552 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1554 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1555 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1576 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1619 xfs_fs_cmn_err(CE_WARN, mp,
1620 "xfs_log_write: reservation summary:\n"
1621 " trans type = %s (%u)\n"
1622 " unit res = %d bytes\n"
1623 " current res = %d bytes\n"
1624 " total reg = %u bytes (o/flow = %u bytes)\n"
1625 " ophdrs = %u (ophdr space = %u bytes)\n"
1626 " ophdr + reg = %u bytes\n"
1627 " num regions = %u\n",
1628 ((ticket->t_trans_type <= 0 ||
1629 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1630 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1631 ticket->t_trans_type,
1634 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1635 ticket->t_res_num_ophdrs, ophdr_spc,
1636 ticket->t_res_arr_sum +
1637 ticket->t_res_o_flow + ophdr_spc,
1640 for (i = 0; i < ticket->t_res_num; i++) {
1641 uint r_type = ticket->t_res_arr[i].r_type;
1643 "region[%u]: %s - %u bytes\n",
1645 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1646 "bad-rtype" : res_type_str[r_type-1]),
1647 ticket->t_res_arr[i].r_len);
1650 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1651 "xfs_log_write: reservation ran out. Need to up reservation");
1652 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1656 * Calculate the potential space needed by the log vector. Each region gets
1657 * its own xlog_op_header_t and may need to be double word aligned.
1660 xlog_write_calc_vec_length(
1661 struct xlog_ticket *ticket,
1662 struct xfs_log_vec *log_vector)
1664 struct xfs_log_vec *lv;
1669 /* acct for start rec of xact */
1670 if (ticket->t_flags & XLOG_TIC_INITED)
1673 for (lv = log_vector; lv; lv = lv->lv_next) {
1674 headers += lv->lv_niovecs;
1676 for (i = 0; i < lv->lv_niovecs; i++) {
1677 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1680 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1684 ticket->t_res_num_ophdrs += headers;
1685 len += headers * sizeof(struct xlog_op_header);
1691 * If first write for transaction, insert start record We can't be trying to
1692 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1695 xlog_write_start_rec(
1696 struct xlog_op_header *ophdr,
1697 struct xlog_ticket *ticket)
1699 if (!(ticket->t_flags & XLOG_TIC_INITED))
1702 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1703 ophdr->oh_clientid = ticket->t_clientid;
1705 ophdr->oh_flags = XLOG_START_TRANS;
1708 ticket->t_flags &= ~XLOG_TIC_INITED;
1710 return sizeof(struct xlog_op_header);
1713 static xlog_op_header_t *
1714 xlog_write_setup_ophdr(
1716 struct xlog_op_header *ophdr,
1717 struct xlog_ticket *ticket,
1720 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1721 ophdr->oh_clientid = ticket->t_clientid;
1724 /* are we copying a commit or unmount record? */
1725 ophdr->oh_flags = flags;
1728 * We've seen logs corrupted with bad transaction client ids. This
1729 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1730 * and shut down the filesystem.
1732 switch (ophdr->oh_clientid) {
1733 case XFS_TRANSACTION:
1738 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1739 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1740 ophdr->oh_clientid, ticket);
1748 * Set up the parameters of the region copy into the log. This has
1749 * to handle region write split across multiple log buffers - this
1750 * state is kept external to this function so that this code can
1751 * can be written in an obvious, self documenting manner.
1754 xlog_write_setup_copy(
1755 struct xlog_ticket *ticket,
1756 struct xlog_op_header *ophdr,
1757 int space_available,
1761 int *last_was_partial_copy,
1762 int *bytes_consumed)
1766 still_to_copy = space_required - *bytes_consumed;
1767 *copy_off = *bytes_consumed;
1769 if (still_to_copy <= space_available) {
1770 /* write of region completes here */
1771 *copy_len = still_to_copy;
1772 ophdr->oh_len = cpu_to_be32(*copy_len);
1773 if (*last_was_partial_copy)
1774 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1775 *last_was_partial_copy = 0;
1776 *bytes_consumed = 0;
1780 /* partial write of region, needs extra log op header reservation */
1781 *copy_len = space_available;
1782 ophdr->oh_len = cpu_to_be32(*copy_len);
1783 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1784 if (*last_was_partial_copy)
1785 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1786 *bytes_consumed += *copy_len;
1787 (*last_was_partial_copy)++;
1789 /* account for new log op header */
1790 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1791 ticket->t_res_num_ophdrs++;
1793 return sizeof(struct xlog_op_header);
1797 xlog_write_copy_finish(
1799 struct xlog_in_core *iclog,
1804 int *partial_copy_len,
1806 struct xlog_in_core **commit_iclog)
1808 if (*partial_copy) {
1810 * This iclog has already been marked WANT_SYNC by
1811 * xlog_state_get_iclog_space.
1813 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1816 return xlog_state_release_iclog(log, iclog);
1820 *partial_copy_len = 0;
1822 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1823 /* no more space in this iclog - push it. */
1824 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1828 spin_lock(&log->l_icloglock);
1829 xlog_state_want_sync(log, iclog);
1830 spin_unlock(&log->l_icloglock);
1833 return xlog_state_release_iclog(log, iclog);
1834 ASSERT(flags & XLOG_COMMIT_TRANS);
1835 *commit_iclog = iclog;
1842 * Write some region out to in-core log
1844 * This will be called when writing externally provided regions or when
1845 * writing out a commit record for a given transaction.
1847 * General algorithm:
1848 * 1. Find total length of this write. This may include adding to the
1849 * lengths passed in.
1850 * 2. Check whether we violate the tickets reservation.
1851 * 3. While writing to this iclog
1852 * A. Reserve as much space in this iclog as can get
1853 * B. If this is first write, save away start lsn
1854 * C. While writing this region:
1855 * 1. If first write of transaction, write start record
1856 * 2. Write log operation header (header per region)
1857 * 3. Find out if we can fit entire region into this iclog
1858 * 4. Potentially, verify destination memcpy ptr
1859 * 5. Memcpy (partial) region
1860 * 6. If partial copy, release iclog; otherwise, continue
1861 * copying more regions into current iclog
1862 * 4. Mark want sync bit (in simulation mode)
1863 * 5. Release iclog for potential flush to on-disk log.
1866 * 1. Panic if reservation is overrun. This should never happen since
1867 * reservation amounts are generated internal to the filesystem.
1869 * 1. Tickets are single threaded data structures.
1870 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1871 * syncing routine. When a single log_write region needs to span
1872 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1873 * on all log operation writes which don't contain the end of the
1874 * region. The XLOG_END_TRANS bit is used for the in-core log
1875 * operation which contains the end of the continued log_write region.
1876 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1877 * we don't really know exactly how much space will be used. As a result,
1878 * we don't update ic_offset until the end when we know exactly how many
1879 * bytes have been written out.
1884 struct xfs_log_vec *log_vector,
1885 struct xlog_ticket *ticket,
1886 xfs_lsn_t *start_lsn,
1887 struct xlog_in_core **commit_iclog,
1890 struct xlog_in_core *iclog = NULL;
1891 struct xfs_log_iovec *vecp;
1892 struct xfs_log_vec *lv;
1895 int partial_copy = 0;
1896 int partial_copy_len = 0;
1904 len = xlog_write_calc_vec_length(ticket, log_vector);
1907 * Region headers and bytes are already accounted for.
1908 * We only need to take into account start records and
1909 * split regions in this function.
1911 if (ticket->t_flags & XLOG_TIC_INITED)
1912 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1915 * Commit record headers need to be accounted for. These
1916 * come in as separate writes so are easy to detect.
1918 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1919 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1921 ticket->t_curr_res -= len;
1923 if (ticket->t_curr_res < 0)
1924 xlog_print_tic_res(log->l_mp, ticket);
1928 vecp = lv->lv_iovecp;
1929 while (lv && index < lv->lv_niovecs) {
1933 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1934 &contwr, &log_offset);
1938 ASSERT(log_offset <= iclog->ic_size - 1);
1939 ptr = iclog->ic_datap + log_offset;
1941 /* start_lsn is the first lsn written to. That's all we need. */
1943 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1946 * This loop writes out as many regions as can fit in the amount
1947 * of space which was allocated by xlog_state_get_iclog_space().
1949 while (lv && index < lv->lv_niovecs) {
1950 struct xfs_log_iovec *reg = &vecp[index];
1951 struct xlog_op_header *ophdr;
1956 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1957 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1959 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1960 if (start_rec_copy) {
1962 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1966 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1968 return XFS_ERROR(EIO);
1970 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1971 sizeof(struct xlog_op_header));
1973 len += xlog_write_setup_copy(ticket, ophdr,
1974 iclog->ic_size-log_offset,
1976 ©_off, ©_len,
1979 xlog_verify_dest_ptr(log, ptr);
1982 ASSERT(copy_len >= 0);
1983 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1984 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1986 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1988 data_cnt += contwr ? copy_len : 0;
1990 error = xlog_write_copy_finish(log, iclog, flags,
1991 &record_cnt, &data_cnt,
2000 * if we had a partial copy, we need to get more iclog
2001 * space but we don't want to increment the region
2002 * index because there is still more is this region to
2005 * If we completed writing this region, and we flushed
2006 * the iclog (indicated by resetting of the record
2007 * count), then we also need to get more log space. If
2008 * this was the last record, though, we are done and
2014 if (++index == lv->lv_niovecs) {
2018 vecp = lv->lv_iovecp;
2020 if (record_cnt == 0) {
2030 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2032 return xlog_state_release_iclog(log, iclog);
2034 ASSERT(flags & XLOG_COMMIT_TRANS);
2035 *commit_iclog = iclog;
2040 /*****************************************************************************
2042 * State Machine functions
2044 *****************************************************************************
2047 /* Clean iclogs starting from the head. This ordering must be
2048 * maintained, so an iclog doesn't become ACTIVE beyond one that
2049 * is SYNCING. This is also required to maintain the notion that we use
2050 * a ordered wait queue to hold off would be writers to the log when every
2051 * iclog is trying to sync to disk.
2053 * State Change: DIRTY -> ACTIVE
2056 xlog_state_clean_log(xlog_t *log)
2058 xlog_in_core_t *iclog;
2061 iclog = log->l_iclog;
2063 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2064 iclog->ic_state = XLOG_STATE_ACTIVE;
2065 iclog->ic_offset = 0;
2066 ASSERT(iclog->ic_callback == NULL);
2068 * If the number of ops in this iclog indicate it just
2069 * contains the dummy transaction, we can
2070 * change state into IDLE (the second time around).
2071 * Otherwise we should change the state into
2073 * We don't need to cover the dummy.
2076 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2081 * We have two dirty iclogs so start over
2082 * This could also be num of ops indicates
2083 * this is not the dummy going out.
2087 iclog->ic_header.h_num_logops = 0;
2088 memset(iclog->ic_header.h_cycle_data, 0,
2089 sizeof(iclog->ic_header.h_cycle_data));
2090 iclog->ic_header.h_lsn = 0;
2091 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2094 break; /* stop cleaning */
2095 iclog = iclog->ic_next;
2096 } while (iclog != log->l_iclog);
2098 /* log is locked when we are called */
2100 * Change state for the dummy log recording.
2101 * We usually go to NEED. But we go to NEED2 if the changed indicates
2102 * we are done writing the dummy record.
2103 * If we are done with the second dummy recored (DONE2), then
2107 switch (log->l_covered_state) {
2108 case XLOG_STATE_COVER_IDLE:
2109 case XLOG_STATE_COVER_NEED:
2110 case XLOG_STATE_COVER_NEED2:
2111 log->l_covered_state = XLOG_STATE_COVER_NEED;
2114 case XLOG_STATE_COVER_DONE:
2116 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2118 log->l_covered_state = XLOG_STATE_COVER_NEED;
2121 case XLOG_STATE_COVER_DONE2:
2123 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2125 log->l_covered_state = XLOG_STATE_COVER_NEED;
2132 } /* xlog_state_clean_log */
2135 xlog_get_lowest_lsn(
2138 xlog_in_core_t *lsn_log;
2139 xfs_lsn_t lowest_lsn, lsn;
2141 lsn_log = log->l_iclog;
2144 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2145 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2146 if ((lsn && !lowest_lsn) ||
2147 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2151 lsn_log = lsn_log->ic_next;
2152 } while (lsn_log != log->l_iclog);
2158 xlog_state_do_callback(
2161 xlog_in_core_t *ciclog)
2163 xlog_in_core_t *iclog;
2164 xlog_in_core_t *first_iclog; /* used to know when we've
2165 * processed all iclogs once */
2166 xfs_log_callback_t *cb, *cb_next;
2168 xfs_lsn_t lowest_lsn;
2169 int ioerrors; /* counter: iclogs with errors */
2170 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2171 int funcdidcallbacks; /* flag: function did callbacks */
2172 int repeats; /* for issuing console warnings if
2173 * looping too many times */
2176 spin_lock(&log->l_icloglock);
2177 first_iclog = iclog = log->l_iclog;
2179 funcdidcallbacks = 0;
2184 * Scan all iclogs starting with the one pointed to by the
2185 * log. Reset this starting point each time the log is
2186 * unlocked (during callbacks).
2188 * Keep looping through iclogs until one full pass is made
2189 * without running any callbacks.
2191 first_iclog = log->l_iclog;
2192 iclog = log->l_iclog;
2193 loopdidcallbacks = 0;
2198 /* skip all iclogs in the ACTIVE & DIRTY states */
2199 if (iclog->ic_state &
2200 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2201 iclog = iclog->ic_next;
2206 * Between marking a filesystem SHUTDOWN and stopping
2207 * the log, we do flush all iclogs to disk (if there
2208 * wasn't a log I/O error). So, we do want things to
2209 * go smoothly in case of just a SHUTDOWN w/o a
2212 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2214 * Can only perform callbacks in order. Since
2215 * this iclog is not in the DONE_SYNC/
2216 * DO_CALLBACK state, we skip the rest and
2217 * just try to clean up. If we set our iclog
2218 * to DO_CALLBACK, we will not process it when
2219 * we retry since a previous iclog is in the
2220 * CALLBACK and the state cannot change since
2221 * we are holding the l_icloglock.
2223 if (!(iclog->ic_state &
2224 (XLOG_STATE_DONE_SYNC |
2225 XLOG_STATE_DO_CALLBACK))) {
2226 if (ciclog && (ciclog->ic_state ==
2227 XLOG_STATE_DONE_SYNC)) {
2228 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2233 * We now have an iclog that is in either the
2234 * DO_CALLBACK or DONE_SYNC states. The other
2235 * states (WANT_SYNC, SYNCING, or CALLBACK were
2236 * caught by the above if and are going to
2237 * clean (i.e. we aren't doing their callbacks)
2242 * We will do one more check here to see if we
2243 * have chased our tail around.
2246 lowest_lsn = xlog_get_lowest_lsn(log);
2248 XFS_LSN_CMP(lowest_lsn,
2249 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2250 iclog = iclog->ic_next;
2251 continue; /* Leave this iclog for
2255 iclog->ic_state = XLOG_STATE_CALLBACK;
2257 spin_unlock(&log->l_icloglock);
2259 /* l_last_sync_lsn field protected by
2260 * l_grant_lock. Don't worry about iclog's lsn.
2261 * No one else can be here except us.
2263 spin_lock(&log->l_grant_lock);
2264 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2265 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2266 log->l_last_sync_lsn =
2267 be64_to_cpu(iclog->ic_header.h_lsn);
2268 spin_unlock(&log->l_grant_lock);
2271 spin_unlock(&log->l_icloglock);
2276 * Keep processing entries in the callback list until
2277 * we come around and it is empty. We need to
2278 * atomically see that the list is empty and change the
2279 * state to DIRTY so that we don't miss any more
2280 * callbacks being added.
2282 spin_lock(&iclog->ic_callback_lock);
2283 cb = iclog->ic_callback;
2285 iclog->ic_callback_tail = &(iclog->ic_callback);
2286 iclog->ic_callback = NULL;
2287 spin_unlock(&iclog->ic_callback_lock);
2289 /* perform callbacks in the order given */
2290 for (; cb; cb = cb_next) {
2291 cb_next = cb->cb_next;
2292 cb->cb_func(cb->cb_arg, aborted);
2294 spin_lock(&iclog->ic_callback_lock);
2295 cb = iclog->ic_callback;
2301 spin_lock(&log->l_icloglock);
2302 ASSERT(iclog->ic_callback == NULL);
2303 spin_unlock(&iclog->ic_callback_lock);
2304 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2305 iclog->ic_state = XLOG_STATE_DIRTY;
2308 * Transition from DIRTY to ACTIVE if applicable.
2309 * NOP if STATE_IOERROR.
2311 xlog_state_clean_log(log);
2313 /* wake up threads waiting in xfs_log_force() */
2314 sv_broadcast(&iclog->ic_force_wait);
2316 iclog = iclog->ic_next;
2317 } while (first_iclog != iclog);
2319 if (repeats > 5000) {
2320 flushcnt += repeats;
2322 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2323 "%s: possible infinite loop (%d iterations)",
2324 __func__, flushcnt);
2326 } while (!ioerrors && loopdidcallbacks);
2329 * make one last gasp attempt to see if iclogs are being left in
2333 if (funcdidcallbacks) {
2334 first_iclog = iclog = log->l_iclog;
2336 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2338 * Terminate the loop if iclogs are found in states
2339 * which will cause other threads to clean up iclogs.
2341 * SYNCING - i/o completion will go through logs
2342 * DONE_SYNC - interrupt thread should be waiting for
2344 * IOERROR - give up hope all ye who enter here
2346 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2347 iclog->ic_state == XLOG_STATE_SYNCING ||
2348 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2349 iclog->ic_state == XLOG_STATE_IOERROR )
2351 iclog = iclog->ic_next;
2352 } while (first_iclog != iclog);
2356 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2358 spin_unlock(&log->l_icloglock);
2361 sv_broadcast(&log->l_flush_wait);
2366 * Finish transitioning this iclog to the dirty state.
2368 * Make sure that we completely execute this routine only when this is
2369 * the last call to the iclog. There is a good chance that iclog flushes,
2370 * when we reach the end of the physical log, get turned into 2 separate
2371 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2372 * routine. By using the reference count bwritecnt, we guarantee that only
2373 * the second completion goes through.
2375 * Callbacks could take time, so they are done outside the scope of the
2376 * global state machine log lock.
2379 xlog_state_done_syncing(
2380 xlog_in_core_t *iclog,
2383 xlog_t *log = iclog->ic_log;
2385 spin_lock(&log->l_icloglock);
2387 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2388 iclog->ic_state == XLOG_STATE_IOERROR);
2389 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2390 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2394 * If we got an error, either on the first buffer, or in the case of
2395 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2396 * and none should ever be attempted to be written to disk
2399 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2400 if (--iclog->ic_bwritecnt == 1) {
2401 spin_unlock(&log->l_icloglock);
2404 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2408 * Someone could be sleeping prior to writing out the next
2409 * iclog buffer, we wake them all, one will get to do the
2410 * I/O, the others get to wait for the result.
2412 sv_broadcast(&iclog->ic_write_wait);
2413 spin_unlock(&log->l_icloglock);
2414 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2415 } /* xlog_state_done_syncing */
2419 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2420 * sleep. We wait on the flush queue on the head iclog as that should be
2421 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2422 * we will wait here and all new writes will sleep until a sync completes.
2424 * The in-core logs are used in a circular fashion. They are not used
2425 * out-of-order even when an iclog past the head is free.
2428 * * log_offset where xlog_write() can start writing into the in-core
2430 * * in-core log pointer to which xlog_write() should write.
2431 * * boolean indicating this is a continued write to an in-core log.
2432 * If this is the last write, then the in-core log's offset field
2433 * needs to be incremented, depending on the amount of data which
2437 xlog_state_get_iclog_space(xlog_t *log,
2439 xlog_in_core_t **iclogp,
2440 xlog_ticket_t *ticket,
2441 int *continued_write,
2445 xlog_rec_header_t *head;
2446 xlog_in_core_t *iclog;
2450 spin_lock(&log->l_icloglock);
2451 if (XLOG_FORCED_SHUTDOWN(log)) {
2452 spin_unlock(&log->l_icloglock);
2453 return XFS_ERROR(EIO);
2456 iclog = log->l_iclog;
2457 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2458 XFS_STATS_INC(xs_log_noiclogs);
2460 /* Wait for log writes to have flushed */
2461 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2465 head = &iclog->ic_header;
2467 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2468 log_offset = iclog->ic_offset;
2470 /* On the 1st write to an iclog, figure out lsn. This works
2471 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2472 * committing to. If the offset is set, that's how many blocks
2475 if (log_offset == 0) {
2476 ticket->t_curr_res -= log->l_iclog_hsize;
2477 xlog_tic_add_region(ticket,
2479 XLOG_REG_TYPE_LRHEADER);
2480 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2481 head->h_lsn = cpu_to_be64(
2482 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2483 ASSERT(log->l_curr_block >= 0);
2486 /* If there is enough room to write everything, then do it. Otherwise,
2487 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2488 * bit is on, so this will get flushed out. Don't update ic_offset
2489 * until you know exactly how many bytes get copied. Therefore, wait
2490 * until later to update ic_offset.
2492 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2493 * can fit into remaining data section.
2495 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2496 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2499 * If I'm the only one writing to this iclog, sync it to disk.
2500 * We need to do an atomic compare and decrement here to avoid
2501 * racing with concurrent atomic_dec_and_lock() calls in
2502 * xlog_state_release_iclog() when there is more than one
2503 * reference to the iclog.
2505 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2506 /* we are the only one */
2507 spin_unlock(&log->l_icloglock);
2508 error = xlog_state_release_iclog(log, iclog);
2512 spin_unlock(&log->l_icloglock);
2517 /* Do we have enough room to write the full amount in the remainder
2518 * of this iclog? Or must we continue a write on the next iclog and
2519 * mark this iclog as completely taken? In the case where we switch
2520 * iclogs (to mark it taken), this particular iclog will release/sync
2521 * to disk in xlog_write().
2523 if (len <= iclog->ic_size - iclog->ic_offset) {
2524 *continued_write = 0;
2525 iclog->ic_offset += len;
2527 *continued_write = 1;
2528 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2532 ASSERT(iclog->ic_offset <= iclog->ic_size);
2533 spin_unlock(&log->l_icloglock);
2535 *logoffsetp = log_offset;
2537 } /* xlog_state_get_iclog_space */
2540 * Atomically get the log space required for a log ticket.
2542 * Once a ticket gets put onto the reserveq, it will only return after
2543 * the needed reservation is satisfied.
2546 xlog_grant_log_space(xlog_t *log,
2557 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2558 panic("grant Recovery problem");
2561 /* Is there space or do we need to sleep? */
2562 spin_lock(&log->l_grant_lock);
2564 trace_xfs_log_grant_enter(log, tic);
2566 /* something is already sleeping; insert new transaction at end */
2567 if (log->l_reserve_headq) {
2568 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2570 trace_xfs_log_grant_sleep1(log, tic);
2573 * Gotta check this before going to sleep, while we're
2574 * holding the grant lock.
2576 if (XLOG_FORCED_SHUTDOWN(log))
2579 XFS_STATS_INC(xs_sleep_logspace);
2580 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2582 * If we got an error, and the filesystem is shutting down,
2583 * we'll catch it down below. So just continue...
2585 trace_xfs_log_grant_wake1(log, tic);
2586 spin_lock(&log->l_grant_lock);
2588 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2589 need_bytes = tic->t_unit_res*tic->t_ocnt;
2591 need_bytes = tic->t_unit_res;
2594 if (XLOG_FORCED_SHUTDOWN(log))
2597 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2598 log->l_grant_reserve_bytes);
2599 if (free_bytes < need_bytes) {
2600 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2601 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2603 trace_xfs_log_grant_sleep2(log, tic);
2605 spin_unlock(&log->l_grant_lock);
2606 xlog_grant_push_ail(log->l_mp, need_bytes);
2607 spin_lock(&log->l_grant_lock);
2609 XFS_STATS_INC(xs_sleep_logspace);
2610 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2612 spin_lock(&log->l_grant_lock);
2613 if (XLOG_FORCED_SHUTDOWN(log))
2616 trace_xfs_log_grant_wake2(log, tic);
2619 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2620 xlog_del_ticketq(&log->l_reserve_headq, tic);
2622 /* we've got enough space */
2623 xlog_grant_add_space(log, need_bytes);
2625 tail_lsn = log->l_tail_lsn;
2627 * Check to make sure the grant write head didn't just over lap the
2628 * tail. If the cycles are the same, we can't be overlapping.
2629 * Otherwise, make sure that the cycles differ by exactly one and
2630 * check the byte count.
2632 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2633 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2634 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2637 trace_xfs_log_grant_exit(log, tic);
2638 xlog_verify_grant_head(log, 1);
2639 spin_unlock(&log->l_grant_lock);
2643 if (tic->t_flags & XLOG_TIC_IN_Q)
2644 xlog_del_ticketq(&log->l_reserve_headq, tic);
2646 trace_xfs_log_grant_error(log, tic);
2649 * If we are failing, make sure the ticket doesn't have any
2650 * current reservations. We don't want to add this back when
2651 * the ticket/transaction gets cancelled.
2653 tic->t_curr_res = 0;
2654 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2655 spin_unlock(&log->l_grant_lock);
2656 return XFS_ERROR(EIO);
2657 } /* xlog_grant_log_space */
2661 * Replenish the byte reservation required by moving the grant write head.
2666 xlog_regrant_write_log_space(xlog_t *log,
2669 int free_bytes, need_bytes;
2670 xlog_ticket_t *ntic;
2675 tic->t_curr_res = tic->t_unit_res;
2676 xlog_tic_reset_res(tic);
2682 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2683 panic("regrant Recovery problem");
2686 spin_lock(&log->l_grant_lock);
2688 trace_xfs_log_regrant_write_enter(log, tic);
2690 if (XLOG_FORCED_SHUTDOWN(log))
2693 /* If there are other waiters on the queue then give them a
2694 * chance at logspace before us. Wake up the first waiters,
2695 * if we do not wake up all the waiters then go to sleep waiting
2696 * for more free space, otherwise try to get some space for
2699 need_bytes = tic->t_unit_res;
2700 if ((ntic = log->l_write_headq)) {
2701 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2702 log->l_grant_write_bytes);
2704 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2706 if (free_bytes < ntic->t_unit_res)
2708 free_bytes -= ntic->t_unit_res;
2709 sv_signal(&ntic->t_wait);
2710 ntic = ntic->t_next;
2711 } while (ntic != log->l_write_headq);
2713 if (ntic != log->l_write_headq) {
2714 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2715 xlog_ins_ticketq(&log->l_write_headq, tic);
2717 trace_xfs_log_regrant_write_sleep1(log, tic);
2719 spin_unlock(&log->l_grant_lock);
2720 xlog_grant_push_ail(log->l_mp, need_bytes);
2721 spin_lock(&log->l_grant_lock);
2723 XFS_STATS_INC(xs_sleep_logspace);
2724 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2725 &log->l_grant_lock, s);
2727 /* If we're shutting down, this tic is already
2729 spin_lock(&log->l_grant_lock);
2730 if (XLOG_FORCED_SHUTDOWN(log))
2733 trace_xfs_log_regrant_write_wake1(log, tic);
2738 if (XLOG_FORCED_SHUTDOWN(log))
2741 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2742 log->l_grant_write_bytes);
2743 if (free_bytes < need_bytes) {
2744 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2745 xlog_ins_ticketq(&log->l_write_headq, tic);
2746 spin_unlock(&log->l_grant_lock);
2747 xlog_grant_push_ail(log->l_mp, need_bytes);
2748 spin_lock(&log->l_grant_lock);
2750 XFS_STATS_INC(xs_sleep_logspace);
2751 trace_xfs_log_regrant_write_sleep2(log, tic);
2753 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2755 /* If we're shutting down, this tic is already off the queue */
2756 spin_lock(&log->l_grant_lock);
2757 if (XLOG_FORCED_SHUTDOWN(log))
2760 trace_xfs_log_regrant_write_wake2(log, tic);
2762 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2763 xlog_del_ticketq(&log->l_write_headq, tic);
2765 /* we've got enough space */
2766 xlog_grant_add_space_write(log, need_bytes);
2768 tail_lsn = log->l_tail_lsn;
2769 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2770 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2771 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2775 trace_xfs_log_regrant_write_exit(log, tic);
2777 xlog_verify_grant_head(log, 1);
2778 spin_unlock(&log->l_grant_lock);
2783 if (tic->t_flags & XLOG_TIC_IN_Q)
2784 xlog_del_ticketq(&log->l_reserve_headq, tic);
2786 trace_xfs_log_regrant_write_error(log, tic);
2789 * If we are failing, make sure the ticket doesn't have any
2790 * current reservations. We don't want to add this back when
2791 * the ticket/transaction gets cancelled.
2793 tic->t_curr_res = 0;
2794 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2795 spin_unlock(&log->l_grant_lock);
2796 return XFS_ERROR(EIO);
2797 } /* xlog_regrant_write_log_space */
2800 /* The first cnt-1 times through here we don't need to
2801 * move the grant write head because the permanent
2802 * reservation has reserved cnt times the unit amount.
2803 * Release part of current permanent unit reservation and
2804 * reset current reservation to be one units worth. Also
2805 * move grant reservation head forward.
2808 xlog_regrant_reserve_log_space(xlog_t *log,
2809 xlog_ticket_t *ticket)
2811 trace_xfs_log_regrant_reserve_enter(log, ticket);
2813 if (ticket->t_cnt > 0)
2816 spin_lock(&log->l_grant_lock);
2817 xlog_grant_sub_space(log, ticket->t_curr_res);
2818 ticket->t_curr_res = ticket->t_unit_res;
2819 xlog_tic_reset_res(ticket);
2821 trace_xfs_log_regrant_reserve_sub(log, ticket);
2823 xlog_verify_grant_head(log, 1);
2825 /* just return if we still have some of the pre-reserved space */
2826 if (ticket->t_cnt > 0) {
2827 spin_unlock(&log->l_grant_lock);
2831 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2833 trace_xfs_log_regrant_reserve_exit(log, ticket);
2835 xlog_verify_grant_head(log, 0);
2836 spin_unlock(&log->l_grant_lock);
2837 ticket->t_curr_res = ticket->t_unit_res;
2838 xlog_tic_reset_res(ticket);
2839 } /* xlog_regrant_reserve_log_space */
2843 * Give back the space left from a reservation.
2845 * All the information we need to make a correct determination of space left
2846 * is present. For non-permanent reservations, things are quite easy. The
2847 * count should have been decremented to zero. We only need to deal with the
2848 * space remaining in the current reservation part of the ticket. If the
2849 * ticket contains a permanent reservation, there may be left over space which
2850 * needs to be released. A count of N means that N-1 refills of the current
2851 * reservation can be done before we need to ask for more space. The first
2852 * one goes to fill up the first current reservation. Once we run out of
2853 * space, the count will stay at zero and the only space remaining will be
2854 * in the current reservation field.
2857 xlog_ungrant_log_space(xlog_t *log,
2858 xlog_ticket_t *ticket)
2860 if (ticket->t_cnt > 0)
2863 spin_lock(&log->l_grant_lock);
2864 trace_xfs_log_ungrant_enter(log, ticket);
2866 xlog_grant_sub_space(log, ticket->t_curr_res);
2868 trace_xfs_log_ungrant_sub(log, ticket);
2870 /* If this is a permanent reservation ticket, we may be able to free
2871 * up more space based on the remaining count.
2873 if (ticket->t_cnt > 0) {
2874 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2875 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2878 trace_xfs_log_ungrant_exit(log, ticket);
2880 xlog_verify_grant_head(log, 1);
2881 spin_unlock(&log->l_grant_lock);
2882 xfs_log_move_tail(log->l_mp, 1);
2883 } /* xlog_ungrant_log_space */
2887 * Flush iclog to disk if this is the last reference to the given iclog and
2888 * the WANT_SYNC bit is set.
2890 * When this function is entered, the iclog is not necessarily in the
2891 * WANT_SYNC state. It may be sitting around waiting to get filled.
2896 xlog_state_release_iclog(
2898 xlog_in_core_t *iclog)
2900 int sync = 0; /* do we sync? */
2902 if (iclog->ic_state & XLOG_STATE_IOERROR)
2903 return XFS_ERROR(EIO);
2905 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2906 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2909 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2910 spin_unlock(&log->l_icloglock);
2911 return XFS_ERROR(EIO);
2913 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2914 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2916 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2917 /* update tail before writing to iclog */
2918 xlog_assign_tail_lsn(log->l_mp);
2920 iclog->ic_state = XLOG_STATE_SYNCING;
2921 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2922 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2923 /* cycle incremented when incrementing curr_block */
2925 spin_unlock(&log->l_icloglock);
2928 * We let the log lock go, so it's possible that we hit a log I/O
2929 * error or some other SHUTDOWN condition that marks the iclog
2930 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2931 * this iclog has consistent data, so we ignore IOERROR
2932 * flags after this point.
2935 return xlog_sync(log, iclog);
2937 } /* xlog_state_release_iclog */
2941 * This routine will mark the current iclog in the ring as WANT_SYNC
2942 * and move the current iclog pointer to the next iclog in the ring.
2943 * When this routine is called from xlog_state_get_iclog_space(), the
2944 * exact size of the iclog has not yet been determined. All we know is
2945 * that every data block. We have run out of space in this log record.
2948 xlog_state_switch_iclogs(xlog_t *log,
2949 xlog_in_core_t *iclog,
2952 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2954 eventual_size = iclog->ic_offset;
2955 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2956 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2957 log->l_prev_block = log->l_curr_block;
2958 log->l_prev_cycle = log->l_curr_cycle;
2960 /* roll log?: ic_offset changed later */
2961 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2963 /* Round up to next log-sunit */
2964 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2965 log->l_mp->m_sb.sb_logsunit > 1) {
2966 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2967 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2970 if (log->l_curr_block >= log->l_logBBsize) {
2971 log->l_curr_cycle++;
2972 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2973 log->l_curr_cycle++;
2974 log->l_curr_block -= log->l_logBBsize;
2975 ASSERT(log->l_curr_block >= 0);
2977 ASSERT(iclog == log->l_iclog);
2978 log->l_iclog = iclog->ic_next;
2979 } /* xlog_state_switch_iclogs */
2982 * Write out all data in the in-core log as of this exact moment in time.
2984 * Data may be written to the in-core log during this call. However,
2985 * we don't guarantee this data will be written out. A change from past
2986 * implementation means this routine will *not* write out zero length LRs.
2988 * Basically, we try and perform an intelligent scan of the in-core logs.
2989 * If we determine there is no flushable data, we just return. There is no
2990 * flushable data if:
2992 * 1. the current iclog is active and has no data; the previous iclog
2993 * is in the active or dirty state.
2994 * 2. the current iclog is drity, and the previous iclog is in the
2995 * active or dirty state.
2999 * 1. the current iclog is not in the active nor dirty state.
3000 * 2. the current iclog dirty, and the previous iclog is not in the
3001 * active nor dirty state.
3002 * 3. the current iclog is active, and there is another thread writing
3003 * to this particular iclog.
3004 * 4. a) the current iclog is active and has no other writers
3005 * b) when we return from flushing out this iclog, it is still
3006 * not in the active nor dirty state.
3010 struct xfs_mount *mp,
3014 struct log *log = mp->m_log;
3015 struct xlog_in_core *iclog;
3018 XFS_STATS_INC(xs_log_force);
3020 xlog_cil_push(log, 1);
3022 spin_lock(&log->l_icloglock);
3024 iclog = log->l_iclog;
3025 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3026 spin_unlock(&log->l_icloglock);
3027 return XFS_ERROR(EIO);
3030 /* If the head iclog is not active nor dirty, we just attach
3031 * ourselves to the head and go to sleep.
3033 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3034 iclog->ic_state == XLOG_STATE_DIRTY) {
3036 * If the head is dirty or (active and empty), then
3037 * we need to look at the previous iclog. If the previous
3038 * iclog is active or dirty we are done. There is nothing
3039 * to sync out. Otherwise, we attach ourselves to the
3040 * previous iclog and go to sleep.
3042 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3043 (atomic_read(&iclog->ic_refcnt) == 0
3044 && iclog->ic_offset == 0)) {
3045 iclog = iclog->ic_prev;
3046 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3047 iclog->ic_state == XLOG_STATE_DIRTY)
3052 if (atomic_read(&iclog->ic_refcnt) == 0) {
3053 /* We are the only one with access to this
3054 * iclog. Flush it out now. There should
3055 * be a roundoff of zero to show that someone
3056 * has already taken care of the roundoff from
3057 * the previous sync.
3059 atomic_inc(&iclog->ic_refcnt);
3060 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3061 xlog_state_switch_iclogs(log, iclog, 0);
3062 spin_unlock(&log->l_icloglock);
3064 if (xlog_state_release_iclog(log, iclog))
3065 return XFS_ERROR(EIO);
3069 spin_lock(&log->l_icloglock);
3070 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3071 iclog->ic_state != XLOG_STATE_DIRTY)
3076 /* Someone else is writing to this iclog.
3077 * Use its call to flush out the data. However,
3078 * the other thread may not force out this LR,
3079 * so we mark it WANT_SYNC.
3081 xlog_state_switch_iclogs(log, iclog, 0);
3087 /* By the time we come around again, the iclog could've been filled
3088 * which would give it another lsn. If we have a new lsn, just
3089 * return because the relevant data has been flushed.
3092 if (flags & XFS_LOG_SYNC) {
3094 * We must check if we're shutting down here, before
3095 * we wait, while we're holding the l_icloglock.
3096 * Then we check again after waking up, in case our
3097 * sleep was disturbed by a bad news.
3099 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3100 spin_unlock(&log->l_icloglock);
3101 return XFS_ERROR(EIO);
3103 XFS_STATS_INC(xs_log_force_sleep);
3104 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3106 * No need to grab the log lock here since we're
3107 * only deciding whether or not to return EIO
3108 * and the memory read should be atomic.
3110 if (iclog->ic_state & XLOG_STATE_IOERROR)
3111 return XFS_ERROR(EIO);
3117 spin_unlock(&log->l_icloglock);
3123 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3124 * about errors or whether the log was flushed or not. This is the normal
3125 * interface to use when trying to unpin items or move the log forward.
3134 error = _xfs_log_force(mp, flags, NULL);
3136 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3137 "error %d returned.", error);
3142 * Force the in-core log to disk for a specific LSN.
3144 * Find in-core log with lsn.
3145 * If it is in the DIRTY state, just return.
3146 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3147 * state and go to sleep or return.
3148 * If it is in any other state, go to sleep or return.
3150 * Synchronous forces are implemented with a signal variable. All callers
3151 * to force a given lsn to disk will wait on a the sv attached to the
3152 * specific in-core log. When given in-core log finally completes its
3153 * write to disk, that thread will wake up all threads waiting on the
3158 struct xfs_mount *mp,
3163 struct log *log = mp->m_log;
3164 struct xlog_in_core *iclog;
3165 int already_slept = 0;
3169 XFS_STATS_INC(xs_log_force);
3172 lsn = xlog_cil_push_lsn(log, lsn);
3173 if (lsn == NULLCOMMITLSN)
3178 spin_lock(&log->l_icloglock);
3179 iclog = log->l_iclog;
3180 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3181 spin_unlock(&log->l_icloglock);
3182 return XFS_ERROR(EIO);
3186 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3187 iclog = iclog->ic_next;
3191 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3192 spin_unlock(&log->l_icloglock);
3196 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3198 * We sleep here if we haven't already slept (e.g.
3199 * this is the first time we've looked at the correct
3200 * iclog buf) and the buffer before us is going to
3201 * be sync'ed. The reason for this is that if we
3202 * are doing sync transactions here, by waiting for
3203 * the previous I/O to complete, we can allow a few
3204 * more transactions into this iclog before we close
3207 * Otherwise, we mark the buffer WANT_SYNC, and bump
3208 * up the refcnt so we can release the log (which
3209 * drops the ref count). The state switch keeps new
3210 * transaction commits from using this buffer. When
3211 * the current commits finish writing into the buffer,
3212 * the refcount will drop to zero and the buffer will
3215 if (!already_slept &&
3216 (iclog->ic_prev->ic_state &
3217 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3218 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3220 XFS_STATS_INC(xs_log_force_sleep);
3222 sv_wait(&iclog->ic_prev->ic_write_wait,
3223 PSWP, &log->l_icloglock, s);
3229 atomic_inc(&iclog->ic_refcnt);
3230 xlog_state_switch_iclogs(log, iclog, 0);
3231 spin_unlock(&log->l_icloglock);
3232 if (xlog_state_release_iclog(log, iclog))
3233 return XFS_ERROR(EIO);
3236 spin_lock(&log->l_icloglock);
3239 if ((flags & XFS_LOG_SYNC) && /* sleep */
3241 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3243 * Don't wait on completion if we know that we've
3244 * gotten a log write error.
3246 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3247 spin_unlock(&log->l_icloglock);
3248 return XFS_ERROR(EIO);
3250 XFS_STATS_INC(xs_log_force_sleep);
3251 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3253 * No need to grab the log lock here since we're
3254 * only deciding whether or not to return EIO
3255 * and the memory read should be atomic.
3257 if (iclog->ic_state & XLOG_STATE_IOERROR)
3258 return XFS_ERROR(EIO);
3262 } else { /* just return */
3263 spin_unlock(&log->l_icloglock);
3267 } while (iclog != log->l_iclog);
3269 spin_unlock(&log->l_icloglock);
3274 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3275 * about errors or whether the log was flushed or not. This is the normal
3276 * interface to use when trying to unpin items or move the log forward.
3286 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3288 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3289 "error %d returned.", error);
3294 * Called when we want to mark the current iclog as being ready to sync to
3298 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3300 assert_spin_locked(&log->l_icloglock);
3302 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3303 xlog_state_switch_iclogs(log, iclog, 0);
3305 ASSERT(iclog->ic_state &
3306 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3311 /*****************************************************************************
3315 *****************************************************************************
3319 * Free a used ticket when its refcount falls to zero.
3323 xlog_ticket_t *ticket)
3325 ASSERT(atomic_read(&ticket->t_ref) > 0);
3326 if (atomic_dec_and_test(&ticket->t_ref)) {
3327 sv_destroy(&ticket->t_wait);
3328 kmem_zone_free(xfs_log_ticket_zone, ticket);
3334 xlog_ticket_t *ticket)
3336 ASSERT(atomic_read(&ticket->t_ref) > 0);
3337 atomic_inc(&ticket->t_ref);
3342 xfs_log_get_trans_ident(
3343 struct xfs_trans *tp)
3345 return tp->t_ticket->t_tid;
3349 * Allocate and initialise a new log ticket.
3360 struct xlog_ticket *tic;
3364 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3369 * Permanent reservations have up to 'cnt'-1 active log operations
3370 * in the log. A unit in this case is the amount of space for one
3371 * of these log operations. Normal reservations have a cnt of 1
3372 * and their unit amount is the total amount of space required.
3374 * The following lines of code account for non-transaction data
3375 * which occupy space in the on-disk log.
3377 * Normal form of a transaction is:
3378 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3379 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3381 * We need to account for all the leadup data and trailer data
3382 * around the transaction data.
3383 * And then we need to account for the worst case in terms of using
3385 * The worst case will happen if:
3386 * - the placement of the transaction happens to be such that the
3387 * roundoff is at its maximum
3388 * - the transaction data is synced before the commit record is synced
3389 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3390 * Therefore the commit record is in its own Log Record.
3391 * This can happen as the commit record is called with its
3392 * own region to xlog_write().
3393 * This then means that in the worst case, roundoff can happen for
3394 * the commit-rec as well.
3395 * The commit-rec is smaller than padding in this scenario and so it is
3396 * not added separately.
3399 /* for trans header */
3400 unit_bytes += sizeof(xlog_op_header_t);
3401 unit_bytes += sizeof(xfs_trans_header_t);
3404 unit_bytes += sizeof(xlog_op_header_t);
3407 * for LR headers - the space for data in an iclog is the size minus
3408 * the space used for the headers. If we use the iclog size, then we
3409 * undercalculate the number of headers required.
3411 * Furthermore - the addition of op headers for split-recs might
3412 * increase the space required enough to require more log and op
3413 * headers, so take that into account too.
3415 * IMPORTANT: This reservation makes the assumption that if this
3416 * transaction is the first in an iclog and hence has the LR headers
3417 * accounted to it, then the remaining space in the iclog is
3418 * exclusively for this transaction. i.e. if the transaction is larger
3419 * than the iclog, it will be the only thing in that iclog.
3420 * Fundamentally, this means we must pass the entire log vector to
3421 * xlog_write to guarantee this.
3423 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3424 num_headers = howmany(unit_bytes, iclog_space);
3426 /* for split-recs - ophdrs added when data split over LRs */
3427 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3429 /* add extra header reservations if we overrun */
3430 while (!num_headers ||
3431 howmany(unit_bytes, iclog_space) > num_headers) {
3432 unit_bytes += sizeof(xlog_op_header_t);
3435 unit_bytes += log->l_iclog_hsize * num_headers;
3437 /* for commit-rec LR header - note: padding will subsume the ophdr */
3438 unit_bytes += log->l_iclog_hsize;
3440 /* for roundoff padding for transaction data and one for commit record */
3441 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3442 log->l_mp->m_sb.sb_logsunit > 1) {
3443 /* log su roundoff */
3444 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3447 unit_bytes += 2*BBSIZE;
3450 atomic_set(&tic->t_ref, 1);
3451 tic->t_unit_res = unit_bytes;
3452 tic->t_curr_res = unit_bytes;
3455 tic->t_tid = random32();
3456 tic->t_clientid = client;
3457 tic->t_flags = XLOG_TIC_INITED;
3458 tic->t_trans_type = 0;
3459 if (xflags & XFS_LOG_PERM_RESERV)
3460 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3461 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3463 xlog_tic_reset_res(tic);
3469 /******************************************************************************
3471 * Log debug routines
3473 ******************************************************************************
3477 * Make sure that the destination ptr is within the valid data region of
3478 * one of the iclogs. This uses backup pointers stored in a different
3479 * part of the log in case we trash the log structure.
3482 xlog_verify_dest_ptr(
3489 for (i = 0; i < log->l_iclog_bufs; i++) {
3490 if (ptr >= log->l_iclog_bak[i] &&
3491 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3496 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3500 xlog_verify_grant_head(xlog_t *log, int equals)
3502 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3504 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3506 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3508 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3509 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3511 } /* xlog_verify_grant_head */
3513 /* check if it will fit */
3515 xlog_verify_tail_lsn(xlog_t *log,
3516 xlog_in_core_t *iclog,
3521 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3523 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3524 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3525 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3527 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3529 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3530 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3532 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3533 if (blocks < BTOBB(iclog->ic_offset) + 1)
3534 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3536 } /* xlog_verify_tail_lsn */
3539 * Perform a number of checks on the iclog before writing to disk.
3541 * 1. Make sure the iclogs are still circular
3542 * 2. Make sure we have a good magic number
3543 * 3. Make sure we don't have magic numbers in the data
3544 * 4. Check fields of each log operation header for:
3545 * A. Valid client identifier
3546 * B. tid ptr value falls in valid ptr space (user space code)
3547 * C. Length in log record header is correct according to the
3548 * individual operation headers within record.
3549 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3550 * log, check the preceding blocks of the physical log to make sure all
3551 * the cycle numbers agree with the current cycle number.
3554 xlog_verify_iclog(xlog_t *log,
3555 xlog_in_core_t *iclog,
3559 xlog_op_header_t *ophead;
3560 xlog_in_core_t *icptr;
3561 xlog_in_core_2_t *xhdr;
3563 xfs_caddr_t base_ptr;
3564 __psint_t field_offset;
3566 int len, i, j, k, op_len;
3569 /* check validity of iclog pointers */
3570 spin_lock(&log->l_icloglock);
3571 icptr = log->l_iclog;
3572 for (i=0; i < log->l_iclog_bufs; i++) {
3574 xlog_panic("xlog_verify_iclog: invalid ptr");
3575 icptr = icptr->ic_next;
3577 if (icptr != log->l_iclog)
3578 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3579 spin_unlock(&log->l_icloglock);
3581 /* check log magic numbers */
3582 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3583 xlog_panic("xlog_verify_iclog: invalid magic num");
3585 ptr = (xfs_caddr_t) &iclog->ic_header;
3586 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3588 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3589 xlog_panic("xlog_verify_iclog: unexpected magic num");
3593 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3594 ptr = iclog->ic_datap;
3596 ophead = (xlog_op_header_t *)ptr;
3597 xhdr = iclog->ic_data;
3598 for (i = 0; i < len; i++) {
3599 ophead = (xlog_op_header_t *)ptr;
3601 /* clientid is only 1 byte */
3602 field_offset = (__psint_t)
3603 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3604 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3605 clientid = ophead->oh_clientid;
3607 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3608 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3609 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3610 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3611 clientid = xlog_get_client_id(
3612 xhdr[j].hic_xheader.xh_cycle_data[k]);
3614 clientid = xlog_get_client_id(
3615 iclog->ic_header.h_cycle_data[idx]);
3618 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3619 cmn_err(CE_WARN, "xlog_verify_iclog: "
3620 "invalid clientid %d op 0x%p offset 0x%lx",
3621 clientid, ophead, (unsigned long)field_offset);
3624 field_offset = (__psint_t)
3625 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3626 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3627 op_len = be32_to_cpu(ophead->oh_len);
3629 idx = BTOBBT((__psint_t)&ophead->oh_len -
3630 (__psint_t)iclog->ic_datap);
3631 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3632 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3633 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3634 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3636 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3639 ptr += sizeof(xlog_op_header_t) + op_len;
3641 } /* xlog_verify_iclog */
3645 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3651 xlog_in_core_t *iclog, *ic;
3653 iclog = log->l_iclog;
3654 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3656 * Mark all the incore logs IOERROR.
3657 * From now on, no log flushes will result.
3661 ic->ic_state = XLOG_STATE_IOERROR;
3663 } while (ic != iclog);
3667 * Return non-zero, if state transition has already happened.
3673 * This is called from xfs_force_shutdown, when we're forcibly
3674 * shutting down the filesystem, typically because of an IO error.
3675 * Our main objectives here are to make sure that:
3676 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3677 * parties to find out, 'atomically'.
3678 * b. those who're sleeping on log reservations, pinned objects and
3679 * other resources get woken up, and be told the bad news.
3680 * c. nothing new gets queued up after (a) and (b) are done.
3681 * d. if !logerror, flush the iclogs to disk, then seal them off
3684 * Note: for delayed logging the !logerror case needs to flush the regions
3685 * held in memory out to the iclogs before flushing them to disk. This needs
3686 * to be done before the log is marked as shutdown, otherwise the flush to the
3690 xfs_log_force_umount(
3691 struct xfs_mount *mp,
3701 * If this happens during log recovery, don't worry about
3702 * locking; the log isn't open for business yet.
3705 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3706 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3708 XFS_BUF_DONE(mp->m_sb_bp);
3713 * Somebody could've already done the hard work for us.
3714 * No need to get locks for this.
3716 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3717 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3723 * Flush the in memory commit item list before marking the log as
3724 * being shut down. We need to do it in this order to ensure all the
3725 * completed transactions are flushed to disk with the xfs_log_force()
3728 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3729 xlog_cil_push(log, 1);
3732 * We must hold both the GRANT lock and the LOG lock,
3733 * before we mark the filesystem SHUTDOWN and wake
3734 * everybody up to tell the bad news.
3736 spin_lock(&log->l_icloglock);
3737 spin_lock(&log->l_grant_lock);
3738 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3740 XFS_BUF_DONE(mp->m_sb_bp);
3743 * This flag is sort of redundant because of the mount flag, but
3744 * it's good to maintain the separation between the log and the rest
3747 log->l_flags |= XLOG_IO_ERROR;
3750 * If we hit a log error, we want to mark all the iclogs IOERROR
3751 * while we're still holding the loglock.
3754 retval = xlog_state_ioerror(log);
3755 spin_unlock(&log->l_icloglock);
3758 * We don't want anybody waiting for log reservations
3759 * after this. That means we have to wake up everybody
3760 * queued up on reserve_headq as well as write_headq.
3761 * In addition, we make sure in xlog_{re}grant_log_space
3762 * that we don't enqueue anything once the SHUTDOWN flag
3763 * is set, and this action is protected by the GRANTLOCK.
3765 if ((tic = log->l_reserve_headq)) {
3767 sv_signal(&tic->t_wait);
3769 } while (tic != log->l_reserve_headq);
3772 if ((tic = log->l_write_headq)) {
3774 sv_signal(&tic->t_wait);
3776 } while (tic != log->l_write_headq);
3778 spin_unlock(&log->l_grant_lock);
3780 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3783 * Force the incore logs to disk before shutting the
3784 * log down completely.
3786 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3788 spin_lock(&log->l_icloglock);
3789 retval = xlog_state_ioerror(log);
3790 spin_unlock(&log->l_icloglock);
3793 * Wake up everybody waiting on xfs_log_force.
3794 * Callback all log item committed functions as if the
3795 * log writes were completed.
3797 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3799 #ifdef XFSERRORDEBUG
3801 xlog_in_core_t *iclog;
3803 spin_lock(&log->l_icloglock);
3804 iclog = log->l_iclog;
3806 ASSERT(iclog->ic_callback == 0);
3807 iclog = iclog->ic_next;
3808 } while (iclog != log->l_iclog);
3809 spin_unlock(&log->l_icloglock);
3812 /* return non-zero if log IOERROR transition had already happened */
3817 xlog_iclogs_empty(xlog_t *log)
3819 xlog_in_core_t *iclog;
3821 iclog = log->l_iclog;
3823 /* endianness does not matter here, zero is zero in
3826 if (iclog->ic_header.h_num_logops)
3828 iclog = iclog->ic_next;
3829 } while (iclog != log->l_iclog);