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(struct log *log, atomic64_t *head);
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(struct log *log,
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_tail(struct log *log);
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_tail(a)
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);
104 int64_t head_val = atomic64_read(head);
110 xlog_crack_grant_head_val(head_val, &cycle, &space);
114 space += log->l_logsize;
119 new = xlog_assign_grant_head_val(cycle, space);
120 head_val = atomic64_cmpxchg(head, old, new);
121 } while (head_val != old);
125 xlog_grant_add_space(
130 int64_t head_val = atomic64_read(head);
137 xlog_crack_grant_head_val(head_val, &cycle, &space);
139 tmp = log->l_logsize - space;
148 new = xlog_assign_grant_head_val(cycle, space);
149 head_val = atomic64_cmpxchg(head, old, new);
150 } while (head_val != old);
158 struct xlog_ticket *tic;
161 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
162 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
163 need_bytes = tic->t_unit_res * tic->t_cnt;
165 need_bytes = tic->t_unit_res;
167 if (*free_bytes < need_bytes)
169 *free_bytes -= need_bytes;
171 trace_xfs_log_grant_wake_up(log, tic);
172 wake_up(&tic->t_wait);
183 struct xlog_ticket *tic;
186 list_for_each_entry(tic, &log->l_writeq, t_queue) {
187 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
189 need_bytes = tic->t_unit_res;
191 if (*free_bytes < need_bytes)
193 *free_bytes -= need_bytes;
195 trace_xfs_log_regrant_write_wake_up(log, tic);
196 wake_up(&tic->t_wait);
205 struct xlog_ticket *tic,
208 list_add_tail(&tic->t_queue, &log->l_reserveq);
211 if (XLOG_FORCED_SHUTDOWN(log))
213 xlog_grant_push_ail(log, need_bytes);
215 XFS_STATS_INC(xs_sleep_logspace);
216 trace_xfs_log_grant_sleep(log, tic);
218 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
219 trace_xfs_log_grant_wake(log, tic);
221 spin_lock(&log->l_grant_reserve_lock);
222 if (XLOG_FORCED_SHUTDOWN(log))
224 } while (xlog_space_left(log, &log->l_grant_reserve_head) < need_bytes);
226 list_del_init(&tic->t_queue);
229 list_del_init(&tic->t_queue);
230 return XFS_ERROR(EIO);
236 struct xlog_ticket *tic,
239 list_add_tail(&tic->t_queue, &log->l_writeq);
242 if (XLOG_FORCED_SHUTDOWN(log))
244 xlog_grant_push_ail(log, need_bytes);
246 XFS_STATS_INC(xs_sleep_logspace);
247 trace_xfs_log_regrant_write_sleep(log, tic);
249 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
250 trace_xfs_log_regrant_write_wake(log, tic);
252 spin_lock(&log->l_grant_write_lock);
253 if (XLOG_FORCED_SHUTDOWN(log))
255 } while (xlog_space_left(log, &log->l_grant_write_head) < need_bytes);
257 list_del_init(&tic->t_queue);
260 list_del_init(&tic->t_queue);
261 return XFS_ERROR(EIO);
265 xlog_tic_reset_res(xlog_ticket_t *tic)
268 tic->t_res_arr_sum = 0;
269 tic->t_res_num_ophdrs = 0;
273 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
275 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
276 /* add to overflow and start again */
277 tic->t_res_o_flow += tic->t_res_arr_sum;
279 tic->t_res_arr_sum = 0;
282 tic->t_res_arr[tic->t_res_num].r_len = len;
283 tic->t_res_arr[tic->t_res_num].r_type = type;
284 tic->t_res_arr_sum += len;
291 * 1. currblock field gets updated at startup and after in-core logs
292 * marked as with WANT_SYNC.
296 * This routine is called when a user of a log manager ticket is done with
297 * the reservation. If the ticket was ever used, then a commit record for
298 * the associated transaction is written out as a log operation header with
299 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
300 * a given ticket. If the ticket was one with a permanent reservation, then
301 * a few operations are done differently. Permanent reservation tickets by
302 * default don't release the reservation. They just commit the current
303 * transaction with the belief that the reservation is still needed. A flag
304 * must be passed in before permanent reservations are actually released.
305 * When these type of tickets are not released, they need to be set into
306 * the inited state again. By doing this, a start record will be written
307 * out when the next write occurs.
311 struct xfs_mount *mp,
312 struct xlog_ticket *ticket,
313 struct xlog_in_core **iclog,
316 struct log *log = mp->m_log;
319 if (XLOG_FORCED_SHUTDOWN(log) ||
321 * If nothing was ever written, don't write out commit record.
322 * If we get an error, just continue and give back the log ticket.
324 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
325 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
326 lsn = (xfs_lsn_t) -1;
327 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
328 flags |= XFS_LOG_REL_PERM_RESERV;
333 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
334 (flags & XFS_LOG_REL_PERM_RESERV)) {
335 trace_xfs_log_done_nonperm(log, ticket);
338 * Release ticket if not permanent reservation or a specific
339 * request has been made to release a permanent reservation.
341 xlog_ungrant_log_space(log, ticket);
342 xfs_log_ticket_put(ticket);
344 trace_xfs_log_done_perm(log, ticket);
346 xlog_regrant_reserve_log_space(log, ticket);
347 /* If this ticket was a permanent reservation and we aren't
348 * trying to release it, reset the inited flags; so next time
349 * we write, a start record will be written out.
351 ticket->t_flags |= XLOG_TIC_INITED;
358 * Attaches a new iclog I/O completion callback routine during
359 * transaction commit. If the log is in error state, a non-zero
360 * return code is handed back and the caller is responsible for
361 * executing the callback at an appropriate time.
365 struct xfs_mount *mp,
366 struct xlog_in_core *iclog,
367 xfs_log_callback_t *cb)
371 spin_lock(&iclog->ic_callback_lock);
372 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
374 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
375 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
377 *(iclog->ic_callback_tail) = cb;
378 iclog->ic_callback_tail = &(cb->cb_next);
380 spin_unlock(&iclog->ic_callback_lock);
385 xfs_log_release_iclog(
386 struct xfs_mount *mp,
387 struct xlog_in_core *iclog)
389 if (xlog_state_release_iclog(mp->m_log, iclog)) {
390 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
398 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
399 * to the reservation.
400 * 2. Potentially, push buffers at tail of log to disk.
402 * Each reservation is going to reserve extra space for a log record header.
403 * When writes happen to the on-disk log, we don't subtract the length of the
404 * log record header from any reservation. By wasting space in each
405 * reservation, we prevent over allocation problems.
409 struct xfs_mount *mp,
412 struct xlog_ticket **ticket,
417 struct log *log = mp->m_log;
418 struct xlog_ticket *internal_ticket;
421 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
423 if (XLOG_FORCED_SHUTDOWN(log))
424 return XFS_ERROR(EIO);
426 XFS_STATS_INC(xs_try_logspace);
429 if (*ticket != NULL) {
430 ASSERT(flags & XFS_LOG_PERM_RESERV);
431 internal_ticket = *ticket;
434 * this is a new transaction on the ticket, so we need to
435 * change the transaction ID so that the next transaction has a
436 * different TID in the log. Just add one to the existing tid
437 * so that we can see chains of rolling transactions in the log
440 internal_ticket->t_tid++;
442 trace_xfs_log_reserve(log, internal_ticket);
444 xlog_grant_push_ail(log, internal_ticket->t_unit_res);
445 retval = xlog_regrant_write_log_space(log, internal_ticket);
447 /* may sleep if need to allocate more tickets */
448 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
450 KM_SLEEP|KM_MAYFAIL);
451 if (!internal_ticket)
452 return XFS_ERROR(ENOMEM);
453 internal_ticket->t_trans_type = t_type;
454 *ticket = internal_ticket;
456 trace_xfs_log_reserve(log, internal_ticket);
458 xlog_grant_push_ail(log,
459 (internal_ticket->t_unit_res *
460 internal_ticket->t_cnt));
461 retval = xlog_grant_log_space(log, internal_ticket);
464 if (unlikely(retval)) {
466 * If we are failing, make sure the ticket doesn't have any
467 * current reservations. We don't want to add this back
468 * when the ticket/ transaction gets cancelled.
470 internal_ticket->t_curr_res = 0;
471 /* ungrant will give back unit_res * t_cnt. */
472 internal_ticket->t_cnt = 0;
480 * Mount a log filesystem
482 * mp - ubiquitous xfs mount point structure
483 * log_target - buftarg of on-disk log device
484 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
485 * num_bblocks - Number of BBSIZE blocks in on-disk log
487 * Return error or zero.
492 xfs_buftarg_t *log_target,
493 xfs_daddr_t blk_offset,
498 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
499 xfs_notice(mp, "Mounting Filesystem");
502 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
503 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
506 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
507 if (IS_ERR(mp->m_log)) {
508 error = -PTR_ERR(mp->m_log);
513 * Initialize the AIL now we have a log.
515 error = xfs_trans_ail_init(mp);
517 xfs_warn(mp, "AIL initialisation failed: error %d", error);
520 mp->m_log->l_ailp = mp->m_ail;
523 * skip log recovery on a norecovery mount. pretend it all
526 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
527 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
530 mp->m_flags &= ~XFS_MOUNT_RDONLY;
532 error = xlog_recover(mp->m_log);
535 mp->m_flags |= XFS_MOUNT_RDONLY;
537 xfs_warn(mp, "log mount/recovery failed: error %d",
539 goto out_destroy_ail;
543 /* Normal transactions can now occur */
544 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
547 * Now the log has been fully initialised and we know were our
548 * space grant counters are, we can initialise the permanent ticket
549 * needed for delayed logging to work.
551 xlog_cil_init_post_recovery(mp->m_log);
556 xfs_trans_ail_destroy(mp);
558 xlog_dealloc_log(mp->m_log);
564 * Finish the recovery of the file system. This is separate from
565 * the xfs_log_mount() call, because it depends on the code in
566 * xfs_mountfs() to read in the root and real-time bitmap inodes
567 * between calling xfs_log_mount() and here.
569 * mp - ubiquitous xfs mount point structure
572 xfs_log_mount_finish(xfs_mount_t *mp)
576 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
577 error = xlog_recover_finish(mp->m_log);
580 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
587 * Final log writes as part of unmount.
589 * Mark the filesystem clean as unmount happens. Note that during relocation
590 * this routine needs to be executed as part of source-bag while the
591 * deallocation must not be done until source-end.
595 * Unmount record used to have a string "Unmount filesystem--" in the
596 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
597 * We just write the magic number now since that particular field isn't
598 * currently architecture converted and "nUmount" is a bit foo.
599 * As far as I know, there weren't any dependencies on the old behaviour.
603 xfs_log_unmount_write(xfs_mount_t *mp)
605 xlog_t *log = mp->m_log;
606 xlog_in_core_t *iclog;
608 xlog_in_core_t *first_iclog;
610 xlog_ticket_t *tic = NULL;
615 * Don't write out unmount record on read-only mounts.
616 * Or, if we are doing a forced umount (typically because of IO errors).
618 if (mp->m_flags & XFS_MOUNT_RDONLY)
621 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
622 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
625 first_iclog = iclog = log->l_iclog;
627 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
628 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
629 ASSERT(iclog->ic_offset == 0);
631 iclog = iclog->ic_next;
632 } while (iclog != first_iclog);
634 if (! (XLOG_FORCED_SHUTDOWN(log))) {
635 error = xfs_log_reserve(mp, 600, 1, &tic,
636 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
638 /* the data section must be 32 bit size aligned */
642 __uint32_t pad2; /* may as well make it 64 bits */
644 .magic = XLOG_UNMOUNT_TYPE,
646 struct xfs_log_iovec reg = {
648 .i_len = sizeof(magic),
649 .i_type = XLOG_REG_TYPE_UNMOUNT,
651 struct xfs_log_vec vec = {
656 /* remove inited flag, and account for space used */
658 tic->t_curr_res -= sizeof(magic);
659 error = xlog_write(log, &vec, tic, &lsn,
660 NULL, XLOG_UNMOUNT_TRANS);
662 * At this point, we're umounting anyway,
663 * so there's no point in transitioning log state
664 * to IOERROR. Just continue...
669 xfs_alert(mp, "%s: unmount record failed", __func__);
672 spin_lock(&log->l_icloglock);
673 iclog = log->l_iclog;
674 atomic_inc(&iclog->ic_refcnt);
675 xlog_state_want_sync(log, iclog);
676 spin_unlock(&log->l_icloglock);
677 error = xlog_state_release_iclog(log, iclog);
679 spin_lock(&log->l_icloglock);
680 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
681 iclog->ic_state == XLOG_STATE_DIRTY)) {
682 if (!XLOG_FORCED_SHUTDOWN(log)) {
683 xlog_wait(&iclog->ic_force_wait,
686 spin_unlock(&log->l_icloglock);
689 spin_unlock(&log->l_icloglock);
692 trace_xfs_log_umount_write(log, tic);
693 xlog_ungrant_log_space(log, tic);
694 xfs_log_ticket_put(tic);
698 * We're already in forced_shutdown mode, couldn't
699 * even attempt to write out the unmount transaction.
701 * Go through the motions of sync'ing and releasing
702 * the iclog, even though no I/O will actually happen,
703 * we need to wait for other log I/Os that may already
704 * be in progress. Do this as a separate section of
705 * code so we'll know if we ever get stuck here that
706 * we're in this odd situation of trying to unmount
707 * a file system that went into forced_shutdown as
708 * the result of an unmount..
710 spin_lock(&log->l_icloglock);
711 iclog = log->l_iclog;
712 atomic_inc(&iclog->ic_refcnt);
714 xlog_state_want_sync(log, iclog);
715 spin_unlock(&log->l_icloglock);
716 error = xlog_state_release_iclog(log, iclog);
718 spin_lock(&log->l_icloglock);
720 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
721 || iclog->ic_state == XLOG_STATE_DIRTY
722 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
724 xlog_wait(&iclog->ic_force_wait,
727 spin_unlock(&log->l_icloglock);
732 } /* xfs_log_unmount_write */
735 * Deallocate log structures for unmount/relocation.
737 * We need to stop the aild from running before we destroy
738 * and deallocate the log as the aild references the log.
741 xfs_log_unmount(xfs_mount_t *mp)
743 xfs_trans_ail_destroy(mp);
744 xlog_dealloc_log(mp->m_log);
749 struct xfs_mount *mp,
750 struct xfs_log_item *item,
752 const struct xfs_item_ops *ops)
754 item->li_mountp = mp;
755 item->li_ailp = mp->m_ail;
756 item->li_type = type;
760 INIT_LIST_HEAD(&item->li_ail);
761 INIT_LIST_HEAD(&item->li_cil);
765 * Write region vectors to log. The write happens using the space reservation
766 * of the ticket (tic). It is not a requirement that all writes for a given
767 * transaction occur with one call to xfs_log_write(). However, it is important
768 * to note that the transaction reservation code makes an assumption about the
769 * number of log headers a transaction requires that may be violated if you
770 * don't pass all the transaction vectors in one call....
774 struct xfs_mount *mp,
775 struct xfs_log_iovec reg[],
777 struct xlog_ticket *tic,
778 xfs_lsn_t *start_lsn)
780 struct log *log = mp->m_log;
782 struct xfs_log_vec vec = {
783 .lv_niovecs = nentries,
787 if (XLOG_FORCED_SHUTDOWN(log))
788 return XFS_ERROR(EIO);
790 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
792 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
797 xfs_log_move_tail(xfs_mount_t *mp,
801 xlog_t *log = mp->m_log;
802 int need_bytes, free_bytes;
804 if (XLOG_FORCED_SHUTDOWN(log))
808 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
810 /* tail_lsn == 1 implies that we weren't passed a valid value. */
812 atomic64_set(&log->l_tail_lsn, tail_lsn);
814 if (!list_empty_careful(&log->l_writeq)) {
816 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
817 panic("Recovery problem");
819 spin_lock(&log->l_grant_write_lock);
820 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
821 list_for_each_entry(tic, &log->l_writeq, t_queue) {
822 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
824 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
827 free_bytes -= tic->t_unit_res;
828 trace_xfs_log_regrant_write_wake_up(log, tic);
829 wake_up(&tic->t_wait);
831 spin_unlock(&log->l_grant_write_lock);
834 if (!list_empty_careful(&log->l_reserveq)) {
836 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
837 panic("Recovery problem");
839 spin_lock(&log->l_grant_reserve_lock);
840 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
841 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
842 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
843 need_bytes = tic->t_unit_res*tic->t_cnt;
845 need_bytes = tic->t_unit_res;
846 if (free_bytes < need_bytes && tail_lsn != 1)
849 free_bytes -= need_bytes;
850 trace_xfs_log_grant_wake_up(log, tic);
851 wake_up(&tic->t_wait);
853 spin_unlock(&log->l_grant_reserve_lock);
858 * Determine if we have a transaction that has gone to disk
859 * that needs to be covered. To begin the transition to the idle state
860 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
861 * If we are then in a state where covering is needed, the caller is informed
862 * that dummy transactions are required to move the log into the idle state.
864 * Because this is called as part of the sync process, we should also indicate
865 * that dummy transactions should be issued in anything but the covered or
866 * idle states. This ensures that the log tail is accurately reflected in
867 * the log at the end of the sync, hence if a crash occurrs avoids replay
868 * of transactions where the metadata is already on disk.
871 xfs_log_need_covered(xfs_mount_t *mp)
874 xlog_t *log = mp->m_log;
876 if (!xfs_fs_writable(mp))
879 spin_lock(&log->l_icloglock);
880 switch (log->l_covered_state) {
881 case XLOG_STATE_COVER_DONE:
882 case XLOG_STATE_COVER_DONE2:
883 case XLOG_STATE_COVER_IDLE:
885 case XLOG_STATE_COVER_NEED:
886 case XLOG_STATE_COVER_NEED2:
887 if (!xfs_ail_min_lsn(log->l_ailp) &&
888 xlog_iclogs_empty(log)) {
889 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
890 log->l_covered_state = XLOG_STATE_COVER_DONE;
892 log->l_covered_state = XLOG_STATE_COVER_DONE2;
899 spin_unlock(&log->l_icloglock);
903 /******************************************************************************
907 ******************************************************************************
910 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
911 * The log manager must keep track of the last LR which was committed
912 * to disk. The lsn of this LR will become the new tail_lsn whenever
913 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
914 * the situation where stuff could be written into the log but nothing
915 * was ever in the AIL when asked. Eventually, we panic since the
916 * tail hits the head.
918 * We may be holding the log iclog lock upon entering this routine.
921 xlog_assign_tail_lsn(
922 struct xfs_mount *mp)
925 struct log *log = mp->m_log;
927 tail_lsn = xfs_ail_min_lsn(mp->m_ail);
929 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
931 atomic64_set(&log->l_tail_lsn, tail_lsn);
936 * Return the space in the log between the tail and the head. The head
937 * is passed in the cycle/bytes formal parms. In the special case where
938 * the reserve head has wrapped passed the tail, this calculation is no
939 * longer valid. In this case, just return 0 which means there is no space
940 * in the log. This works for all places where this function is called
941 * with the reserve head. Of course, if the write head were to ever
942 * wrap the tail, we should blow up. Rather than catch this case here,
943 * we depend on other ASSERTions in other parts of the code. XXXmiken
945 * This code also handles the case where the reservation head is behind
946 * the tail. The details of this case are described below, but the end
947 * result is that we return the size of the log as the amount of space left.
960 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
961 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
962 tail_bytes = BBTOB(tail_bytes);
963 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
964 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
965 else if (tail_cycle + 1 < head_cycle)
967 else if (tail_cycle < head_cycle) {
968 ASSERT(tail_cycle == (head_cycle - 1));
969 free_bytes = tail_bytes - head_bytes;
972 * The reservation head is behind the tail.
973 * In this case we just want to return the size of the
974 * log as the amount of space left.
977 "xlog_space_left: head behind tail\n"
978 " tail_cycle = %d, tail_bytes = %d\n"
979 " GH cycle = %d, GH bytes = %d",
980 tail_cycle, tail_bytes, head_cycle, head_bytes);
982 free_bytes = log->l_logsize;
989 * Log function which is called when an io completes.
991 * The log manager needs its own routine, in order to control what
992 * happens with the buffer after the write completes.
995 xlog_iodone(xfs_buf_t *bp)
997 xlog_in_core_t *iclog = bp->b_fspriv;
998 xlog_t *l = iclog->ic_log;
1002 * Race to shutdown the filesystem if we see an error.
1004 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1005 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1006 xfs_buf_ioerror_alert(bp, __func__);
1008 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1010 * This flag will be propagated to the trans-committed
1011 * callback routines to let them know that the log-commit
1014 aborted = XFS_LI_ABORTED;
1015 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1016 aborted = XFS_LI_ABORTED;
1019 /* log I/O is always issued ASYNC */
1020 ASSERT(XFS_BUF_ISASYNC(bp));
1021 xlog_state_done_syncing(iclog, aborted);
1023 * do not reference the buffer (bp) here as we could race
1024 * with it being freed after writing the unmount record to the
1031 * Return size of each in-core log record buffer.
1033 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1035 * If the filesystem blocksize is too large, we may need to choose a
1036 * larger size since the directory code currently logs entire blocks.
1040 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1046 if (mp->m_logbufs <= 0)
1047 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1049 log->l_iclog_bufs = mp->m_logbufs;
1052 * Buffer size passed in from mount system call.
1054 if (mp->m_logbsize > 0) {
1055 size = log->l_iclog_size = mp->m_logbsize;
1056 log->l_iclog_size_log = 0;
1058 log->l_iclog_size_log++;
1062 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1063 /* # headers = size / 32k
1064 * one header holds cycles from 32k of data
1067 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1068 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1070 log->l_iclog_hsize = xhdrs << BBSHIFT;
1071 log->l_iclog_heads = xhdrs;
1073 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1074 log->l_iclog_hsize = BBSIZE;
1075 log->l_iclog_heads = 1;
1080 /* All machines use 32kB buffers by default. */
1081 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1082 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1084 /* the default log size is 16k or 32k which is one header sector */
1085 log->l_iclog_hsize = BBSIZE;
1086 log->l_iclog_heads = 1;
1089 /* are we being asked to make the sizes selected above visible? */
1090 if (mp->m_logbufs == 0)
1091 mp->m_logbufs = log->l_iclog_bufs;
1092 if (mp->m_logbsize == 0)
1093 mp->m_logbsize = log->l_iclog_size;
1094 } /* xlog_get_iclog_buffer_size */
1098 * This routine initializes some of the log structure for a given mount point.
1099 * Its primary purpose is to fill in enough, so recovery can occur. However,
1100 * some other stuff may be filled in too.
1103 xlog_alloc_log(xfs_mount_t *mp,
1104 xfs_buftarg_t *log_target,
1105 xfs_daddr_t blk_offset,
1109 xlog_rec_header_t *head;
1110 xlog_in_core_t **iclogp;
1111 xlog_in_core_t *iclog, *prev_iclog=NULL;
1117 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1119 xfs_warn(mp, "Log allocation failed: No memory!");
1124 log->l_targ = log_target;
1125 log->l_logsize = BBTOB(num_bblks);
1126 log->l_logBBstart = blk_offset;
1127 log->l_logBBsize = num_bblks;
1128 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1129 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1131 log->l_prev_block = -1;
1132 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1133 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1134 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1135 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1136 xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0);
1137 xlog_assign_grant_head(&log->l_grant_write_head, 1, 0);
1138 INIT_LIST_HEAD(&log->l_reserveq);
1139 INIT_LIST_HEAD(&log->l_writeq);
1140 spin_lock_init(&log->l_grant_reserve_lock);
1141 spin_lock_init(&log->l_grant_write_lock);
1143 error = EFSCORRUPTED;
1144 if (xfs_sb_version_hassector(&mp->m_sb)) {
1145 log2_size = mp->m_sb.sb_logsectlog;
1146 if (log2_size < BBSHIFT) {
1147 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1148 log2_size, BBSHIFT);
1152 log2_size -= BBSHIFT;
1153 if (log2_size > mp->m_sectbb_log) {
1154 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1155 log2_size, mp->m_sectbb_log);
1159 /* for larger sector sizes, must have v2 or external log */
1160 if (log2_size && log->l_logBBstart > 0 &&
1161 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1163 "log sector size (0x%x) invalid for configuration.",
1168 log->l_sectBBsize = 1 << log2_size;
1170 xlog_get_iclog_buffer_size(mp, log);
1173 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, log->l_iclog_size, 0);
1176 bp->b_iodone = xlog_iodone;
1177 ASSERT(xfs_buf_islocked(bp));
1180 spin_lock_init(&log->l_icloglock);
1181 init_waitqueue_head(&log->l_flush_wait);
1183 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1184 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1186 iclogp = &log->l_iclog;
1188 * The amount of memory to allocate for the iclog structure is
1189 * rather funky due to the way the structure is defined. It is
1190 * done this way so that we can use different sizes for machines
1191 * with different amounts of memory. See the definition of
1192 * xlog_in_core_t in xfs_log_priv.h for details.
1194 ASSERT(log->l_iclog_size >= 4096);
1195 for (i=0; i < log->l_iclog_bufs; i++) {
1196 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1198 goto out_free_iclog;
1201 iclog->ic_prev = prev_iclog;
1204 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1205 log->l_iclog_size, 0);
1207 goto out_free_iclog;
1209 bp->b_iodone = xlog_iodone;
1211 iclog->ic_data = bp->b_addr;
1213 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1215 head = &iclog->ic_header;
1216 memset(head, 0, sizeof(xlog_rec_header_t));
1217 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1218 head->h_version = cpu_to_be32(
1219 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1220 head->h_size = cpu_to_be32(log->l_iclog_size);
1222 head->h_fmt = cpu_to_be32(XLOG_FMT);
1223 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1225 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1226 iclog->ic_state = XLOG_STATE_ACTIVE;
1227 iclog->ic_log = log;
1228 atomic_set(&iclog->ic_refcnt, 0);
1229 spin_lock_init(&iclog->ic_callback_lock);
1230 iclog->ic_callback_tail = &(iclog->ic_callback);
1231 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1233 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1234 init_waitqueue_head(&iclog->ic_force_wait);
1235 init_waitqueue_head(&iclog->ic_write_wait);
1237 iclogp = &iclog->ic_next;
1239 *iclogp = log->l_iclog; /* complete ring */
1240 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1242 error = xlog_cil_init(log);
1244 goto out_free_iclog;
1248 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1249 prev_iclog = iclog->ic_next;
1251 xfs_buf_free(iclog->ic_bp);
1254 spinlock_destroy(&log->l_icloglock);
1255 xfs_buf_free(log->l_xbuf);
1259 return ERR_PTR(-error);
1260 } /* xlog_alloc_log */
1264 * Write out the commit record of a transaction associated with the given
1265 * ticket. Return the lsn of the commit record.
1270 struct xlog_ticket *ticket,
1271 struct xlog_in_core **iclog,
1272 xfs_lsn_t *commitlsnp)
1274 struct xfs_mount *mp = log->l_mp;
1276 struct xfs_log_iovec reg = {
1279 .i_type = XLOG_REG_TYPE_COMMIT,
1281 struct xfs_log_vec vec = {
1286 ASSERT_ALWAYS(iclog);
1287 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1290 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1295 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1296 * log space. This code pushes on the lsn which would supposedly free up
1297 * the 25% which we want to leave free. We may need to adopt a policy which
1298 * pushes on an lsn which is further along in the log once we reach the high
1299 * water mark. In this manner, we would be creating a low water mark.
1302 xlog_grant_push_ail(
1306 xfs_lsn_t threshold_lsn = 0;
1307 xfs_lsn_t last_sync_lsn;
1310 int threshold_block;
1311 int threshold_cycle;
1314 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1316 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
1317 free_blocks = BTOBBT(free_bytes);
1320 * Set the threshold for the minimum number of free blocks in the
1321 * log to the maximum of what the caller needs, one quarter of the
1322 * log, and 256 blocks.
1324 free_threshold = BTOBB(need_bytes);
1325 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1326 free_threshold = MAX(free_threshold, 256);
1327 if (free_blocks >= free_threshold)
1330 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1332 threshold_block += free_threshold;
1333 if (threshold_block >= log->l_logBBsize) {
1334 threshold_block -= log->l_logBBsize;
1335 threshold_cycle += 1;
1337 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1340 * Don't pass in an lsn greater than the lsn of the last
1341 * log record known to be on disk. Use a snapshot of the last sync lsn
1342 * so that it doesn't change between the compare and the set.
1344 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1345 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1346 threshold_lsn = last_sync_lsn;
1349 * Get the transaction layer to kick the dirty buffers out to
1350 * disk asynchronously. No point in trying to do this if
1351 * the filesystem is shutting down.
1353 if (!XLOG_FORCED_SHUTDOWN(log))
1354 xfs_ail_push(log->l_ailp, threshold_lsn);
1358 * The bdstrat callback function for log bufs. This gives us a central
1359 * place to trap bufs in case we get hit by a log I/O error and need to
1360 * shutdown. Actually, in practice, even when we didn't get a log error,
1361 * we transition the iclogs to IOERROR state *after* flushing all existing
1362 * iclogs to disk. This is because we don't want anymore new transactions to be
1363 * started or completed afterwards.
1369 struct xlog_in_core *iclog = bp->b_fspriv;
1371 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1372 xfs_buf_ioerror(bp, EIO);
1374 xfs_buf_ioend(bp, 0);
1376 * It would seem logical to return EIO here, but we rely on
1377 * the log state machine to propagate I/O errors instead of
1383 xfs_buf_iorequest(bp);
1388 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1389 * fashion. Previously, we should have moved the current iclog
1390 * ptr in the log to point to the next available iclog. This allows further
1391 * write to continue while this code syncs out an iclog ready to go.
1392 * Before an in-core log can be written out, the data section must be scanned
1393 * to save away the 1st word of each BBSIZE block into the header. We replace
1394 * it with the current cycle count. Each BBSIZE block is tagged with the
1395 * cycle count because there in an implicit assumption that drives will
1396 * guarantee that entire 512 byte blocks get written at once. In other words,
1397 * we can't have part of a 512 byte block written and part not written. By
1398 * tagging each block, we will know which blocks are valid when recovering
1399 * after an unclean shutdown.
1401 * This routine is single threaded on the iclog. No other thread can be in
1402 * this routine with the same iclog. Changing contents of iclog can there-
1403 * fore be done without grabbing the state machine lock. Updating the global
1404 * log will require grabbing the lock though.
1406 * The entire log manager uses a logical block numbering scheme. Only
1407 * log_sync (and then only bwrite()) know about the fact that the log may
1408 * not start with block zero on a given device. The log block start offset
1409 * is added immediately before calling bwrite().
1413 xlog_sync(xlog_t *log,
1414 xlog_in_core_t *iclog)
1416 xfs_caddr_t dptr; /* pointer to byte sized element */
1419 uint count; /* byte count of bwrite */
1420 uint count_init; /* initial count before roundup */
1421 int roundoff; /* roundoff to BB or stripe */
1422 int split = 0; /* split write into two regions */
1424 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1426 XFS_STATS_INC(xs_log_writes);
1427 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1429 /* Add for LR header */
1430 count_init = log->l_iclog_hsize + iclog->ic_offset;
1432 /* Round out the log write size */
1433 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1434 /* we have a v2 stripe unit to use */
1435 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1437 count = BBTOB(BTOBB(count_init));
1439 roundoff = count - count_init;
1440 ASSERT(roundoff >= 0);
1441 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1442 roundoff < log->l_mp->m_sb.sb_logsunit)
1444 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1445 roundoff < BBTOB(1)));
1447 /* move grant heads by roundoff in sync */
1448 xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff);
1449 xlog_grant_add_space(log, &log->l_grant_write_head, roundoff);
1451 /* put cycle number in every block */
1452 xlog_pack_data(log, iclog, roundoff);
1454 /* real byte length */
1456 iclog->ic_header.h_len =
1457 cpu_to_be32(iclog->ic_offset + roundoff);
1459 iclog->ic_header.h_len =
1460 cpu_to_be32(iclog->ic_offset);
1464 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1466 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1468 /* Do we need to split this write into 2 parts? */
1469 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1470 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1471 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1472 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1474 iclog->ic_bwritecnt = 1;
1476 XFS_BUF_SET_COUNT(bp, count);
1477 bp->b_fspriv = iclog;
1478 XFS_BUF_ZEROFLAGS(bp);
1480 bp->b_flags |= XBF_SYNCIO;
1482 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1483 bp->b_flags |= XBF_FUA;
1486 * Flush the data device before flushing the log to make
1487 * sure all meta data written back from the AIL actually made
1488 * it to disk before stamping the new log tail LSN into the
1489 * log buffer. For an external log we need to issue the
1490 * flush explicitly, and unfortunately synchronously here;
1491 * for an internal log we can simply use the block layer
1492 * state machine for preflushes.
1494 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1495 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1497 bp->b_flags |= XBF_FLUSH;
1500 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1501 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1503 xlog_verify_iclog(log, iclog, count, B_TRUE);
1505 /* account for log which doesn't start at block #0 */
1506 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1508 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1513 error = xlog_bdstrat(bp);
1515 xfs_buf_ioerror_alert(bp, "xlog_sync");
1519 bp = iclog->ic_log->l_xbuf;
1520 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1521 xfs_buf_associate_memory(bp,
1522 (char *)&iclog->ic_header + count, split);
1523 bp->b_fspriv = iclog;
1524 XFS_BUF_ZEROFLAGS(bp);
1526 bp->b_flags |= XBF_SYNCIO;
1527 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1528 bp->b_flags |= XBF_FUA;
1531 * Bump the cycle numbers at the start of each block
1532 * since this part of the buffer is at the start of
1533 * a new cycle. Watch out for the header magic number
1536 for (i = 0; i < split; i += BBSIZE) {
1537 be32_add_cpu((__be32 *)dptr, 1);
1538 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1539 be32_add_cpu((__be32 *)dptr, 1);
1543 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1544 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1546 /* account for internal log which doesn't start at block #0 */
1547 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1549 error = xlog_bdstrat(bp);
1551 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1560 * Deallocate a log structure
1563 xlog_dealloc_log(xlog_t *log)
1565 xlog_in_core_t *iclog, *next_iclog;
1568 xlog_cil_destroy(log);
1571 * always need to ensure that the extra buffer does not point to memory
1572 * owned by another log buffer before we free it.
1574 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1575 xfs_buf_free(log->l_xbuf);
1577 iclog = log->l_iclog;
1578 for (i=0; i<log->l_iclog_bufs; i++) {
1579 xfs_buf_free(iclog->ic_bp);
1580 next_iclog = iclog->ic_next;
1584 spinlock_destroy(&log->l_icloglock);
1586 log->l_mp->m_log = NULL;
1588 } /* xlog_dealloc_log */
1591 * Update counters atomically now that memcpy is done.
1595 xlog_state_finish_copy(xlog_t *log,
1596 xlog_in_core_t *iclog,
1600 spin_lock(&log->l_icloglock);
1602 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1603 iclog->ic_offset += copy_bytes;
1605 spin_unlock(&log->l_icloglock);
1606 } /* xlog_state_finish_copy */
1612 * print out info relating to regions written which consume
1617 struct xfs_mount *mp,
1618 struct xlog_ticket *ticket)
1621 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1623 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1624 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1645 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1689 "xfs_log_write: reservation summary:\n"
1690 " trans type = %s (%u)\n"
1691 " unit res = %d bytes\n"
1692 " current res = %d bytes\n"
1693 " total reg = %u bytes (o/flow = %u bytes)\n"
1694 " ophdrs = %u (ophdr space = %u bytes)\n"
1695 " ophdr + reg = %u bytes\n"
1696 " num regions = %u\n",
1697 ((ticket->t_trans_type <= 0 ||
1698 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1699 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1700 ticket->t_trans_type,
1703 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1704 ticket->t_res_num_ophdrs, ophdr_spc,
1705 ticket->t_res_arr_sum +
1706 ticket->t_res_o_flow + ophdr_spc,
1709 for (i = 0; i < ticket->t_res_num; i++) {
1710 uint r_type = ticket->t_res_arr[i].r_type;
1711 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1712 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1713 "bad-rtype" : res_type_str[r_type-1]),
1714 ticket->t_res_arr[i].r_len);
1717 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1718 "xfs_log_write: reservation ran out. Need to up reservation");
1719 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1723 * Calculate the potential space needed by the log vector. Each region gets
1724 * its own xlog_op_header_t and may need to be double word aligned.
1727 xlog_write_calc_vec_length(
1728 struct xlog_ticket *ticket,
1729 struct xfs_log_vec *log_vector)
1731 struct xfs_log_vec *lv;
1736 /* acct for start rec of xact */
1737 if (ticket->t_flags & XLOG_TIC_INITED)
1740 for (lv = log_vector; lv; lv = lv->lv_next) {
1741 headers += lv->lv_niovecs;
1743 for (i = 0; i < lv->lv_niovecs; i++) {
1744 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1747 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1751 ticket->t_res_num_ophdrs += headers;
1752 len += headers * sizeof(struct xlog_op_header);
1758 * If first write for transaction, insert start record We can't be trying to
1759 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1762 xlog_write_start_rec(
1763 struct xlog_op_header *ophdr,
1764 struct xlog_ticket *ticket)
1766 if (!(ticket->t_flags & XLOG_TIC_INITED))
1769 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1770 ophdr->oh_clientid = ticket->t_clientid;
1772 ophdr->oh_flags = XLOG_START_TRANS;
1775 ticket->t_flags &= ~XLOG_TIC_INITED;
1777 return sizeof(struct xlog_op_header);
1780 static xlog_op_header_t *
1781 xlog_write_setup_ophdr(
1783 struct xlog_op_header *ophdr,
1784 struct xlog_ticket *ticket,
1787 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1788 ophdr->oh_clientid = ticket->t_clientid;
1791 /* are we copying a commit or unmount record? */
1792 ophdr->oh_flags = flags;
1795 * We've seen logs corrupted with bad transaction client ids. This
1796 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1797 * and shut down the filesystem.
1799 switch (ophdr->oh_clientid) {
1800 case XFS_TRANSACTION:
1806 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1807 ophdr->oh_clientid, ticket);
1815 * Set up the parameters of the region copy into the log. This has
1816 * to handle region write split across multiple log buffers - this
1817 * state is kept external to this function so that this code can
1818 * can be written in an obvious, self documenting manner.
1821 xlog_write_setup_copy(
1822 struct xlog_ticket *ticket,
1823 struct xlog_op_header *ophdr,
1824 int space_available,
1828 int *last_was_partial_copy,
1829 int *bytes_consumed)
1833 still_to_copy = space_required - *bytes_consumed;
1834 *copy_off = *bytes_consumed;
1836 if (still_to_copy <= space_available) {
1837 /* write of region completes here */
1838 *copy_len = still_to_copy;
1839 ophdr->oh_len = cpu_to_be32(*copy_len);
1840 if (*last_was_partial_copy)
1841 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1842 *last_was_partial_copy = 0;
1843 *bytes_consumed = 0;
1847 /* partial write of region, needs extra log op header reservation */
1848 *copy_len = space_available;
1849 ophdr->oh_len = cpu_to_be32(*copy_len);
1850 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1851 if (*last_was_partial_copy)
1852 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1853 *bytes_consumed += *copy_len;
1854 (*last_was_partial_copy)++;
1856 /* account for new log op header */
1857 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1858 ticket->t_res_num_ophdrs++;
1860 return sizeof(struct xlog_op_header);
1864 xlog_write_copy_finish(
1866 struct xlog_in_core *iclog,
1871 int *partial_copy_len,
1873 struct xlog_in_core **commit_iclog)
1875 if (*partial_copy) {
1877 * This iclog has already been marked WANT_SYNC by
1878 * xlog_state_get_iclog_space.
1880 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1883 return xlog_state_release_iclog(log, iclog);
1887 *partial_copy_len = 0;
1889 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1890 /* no more space in this iclog - push it. */
1891 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1895 spin_lock(&log->l_icloglock);
1896 xlog_state_want_sync(log, iclog);
1897 spin_unlock(&log->l_icloglock);
1900 return xlog_state_release_iclog(log, iclog);
1901 ASSERT(flags & XLOG_COMMIT_TRANS);
1902 *commit_iclog = iclog;
1909 * Write some region out to in-core log
1911 * This will be called when writing externally provided regions or when
1912 * writing out a commit record for a given transaction.
1914 * General algorithm:
1915 * 1. Find total length of this write. This may include adding to the
1916 * lengths passed in.
1917 * 2. Check whether we violate the tickets reservation.
1918 * 3. While writing to this iclog
1919 * A. Reserve as much space in this iclog as can get
1920 * B. If this is first write, save away start lsn
1921 * C. While writing this region:
1922 * 1. If first write of transaction, write start record
1923 * 2. Write log operation header (header per region)
1924 * 3. Find out if we can fit entire region into this iclog
1925 * 4. Potentially, verify destination memcpy ptr
1926 * 5. Memcpy (partial) region
1927 * 6. If partial copy, release iclog; otherwise, continue
1928 * copying more regions into current iclog
1929 * 4. Mark want sync bit (in simulation mode)
1930 * 5. Release iclog for potential flush to on-disk log.
1933 * 1. Panic if reservation is overrun. This should never happen since
1934 * reservation amounts are generated internal to the filesystem.
1936 * 1. Tickets are single threaded data structures.
1937 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1938 * syncing routine. When a single log_write region needs to span
1939 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1940 * on all log operation writes which don't contain the end of the
1941 * region. The XLOG_END_TRANS bit is used for the in-core log
1942 * operation which contains the end of the continued log_write region.
1943 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1944 * we don't really know exactly how much space will be used. As a result,
1945 * we don't update ic_offset until the end when we know exactly how many
1946 * bytes have been written out.
1951 struct xfs_log_vec *log_vector,
1952 struct xlog_ticket *ticket,
1953 xfs_lsn_t *start_lsn,
1954 struct xlog_in_core **commit_iclog,
1957 struct xlog_in_core *iclog = NULL;
1958 struct xfs_log_iovec *vecp;
1959 struct xfs_log_vec *lv;
1962 int partial_copy = 0;
1963 int partial_copy_len = 0;
1971 len = xlog_write_calc_vec_length(ticket, log_vector);
1974 * Region headers and bytes are already accounted for.
1975 * We only need to take into account start records and
1976 * split regions in this function.
1978 if (ticket->t_flags & XLOG_TIC_INITED)
1979 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1982 * Commit record headers need to be accounted for. These
1983 * come in as separate writes so are easy to detect.
1985 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1986 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1988 ticket->t_curr_res -= len;
1990 if (ticket->t_curr_res < 0)
1991 xlog_print_tic_res(log->l_mp, ticket);
1995 vecp = lv->lv_iovecp;
1996 while (lv && index < lv->lv_niovecs) {
2000 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2001 &contwr, &log_offset);
2005 ASSERT(log_offset <= iclog->ic_size - 1);
2006 ptr = iclog->ic_datap + log_offset;
2008 /* start_lsn is the first lsn written to. That's all we need. */
2010 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2013 * This loop writes out as many regions as can fit in the amount
2014 * of space which was allocated by xlog_state_get_iclog_space().
2016 while (lv && index < lv->lv_niovecs) {
2017 struct xfs_log_iovec *reg = &vecp[index];
2018 struct xlog_op_header *ophdr;
2023 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2024 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2026 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2027 if (start_rec_copy) {
2029 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2033 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2035 return XFS_ERROR(EIO);
2037 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2038 sizeof(struct xlog_op_header));
2040 len += xlog_write_setup_copy(ticket, ophdr,
2041 iclog->ic_size-log_offset,
2043 ©_off, ©_len,
2046 xlog_verify_dest_ptr(log, ptr);
2049 ASSERT(copy_len >= 0);
2050 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2051 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2053 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2055 data_cnt += contwr ? copy_len : 0;
2057 error = xlog_write_copy_finish(log, iclog, flags,
2058 &record_cnt, &data_cnt,
2067 * if we had a partial copy, we need to get more iclog
2068 * space but we don't want to increment the region
2069 * index because there is still more is this region to
2072 * If we completed writing this region, and we flushed
2073 * the iclog (indicated by resetting of the record
2074 * count), then we also need to get more log space. If
2075 * this was the last record, though, we are done and
2081 if (++index == lv->lv_niovecs) {
2085 vecp = lv->lv_iovecp;
2087 if (record_cnt == 0) {
2097 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2099 return xlog_state_release_iclog(log, iclog);
2101 ASSERT(flags & XLOG_COMMIT_TRANS);
2102 *commit_iclog = iclog;
2107 /*****************************************************************************
2109 * State Machine functions
2111 *****************************************************************************
2114 /* Clean iclogs starting from the head. This ordering must be
2115 * maintained, so an iclog doesn't become ACTIVE beyond one that
2116 * is SYNCING. This is also required to maintain the notion that we use
2117 * a ordered wait queue to hold off would be writers to the log when every
2118 * iclog is trying to sync to disk.
2120 * State Change: DIRTY -> ACTIVE
2123 xlog_state_clean_log(xlog_t *log)
2125 xlog_in_core_t *iclog;
2128 iclog = log->l_iclog;
2130 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2131 iclog->ic_state = XLOG_STATE_ACTIVE;
2132 iclog->ic_offset = 0;
2133 ASSERT(iclog->ic_callback == NULL);
2135 * If the number of ops in this iclog indicate it just
2136 * contains the dummy transaction, we can
2137 * change state into IDLE (the second time around).
2138 * Otherwise we should change the state into
2140 * We don't need to cover the dummy.
2143 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2148 * We have two dirty iclogs so start over
2149 * This could also be num of ops indicates
2150 * this is not the dummy going out.
2154 iclog->ic_header.h_num_logops = 0;
2155 memset(iclog->ic_header.h_cycle_data, 0,
2156 sizeof(iclog->ic_header.h_cycle_data));
2157 iclog->ic_header.h_lsn = 0;
2158 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2161 break; /* stop cleaning */
2162 iclog = iclog->ic_next;
2163 } while (iclog != log->l_iclog);
2165 /* log is locked when we are called */
2167 * Change state for the dummy log recording.
2168 * We usually go to NEED. But we go to NEED2 if the changed indicates
2169 * we are done writing the dummy record.
2170 * If we are done with the second dummy recored (DONE2), then
2174 switch (log->l_covered_state) {
2175 case XLOG_STATE_COVER_IDLE:
2176 case XLOG_STATE_COVER_NEED:
2177 case XLOG_STATE_COVER_NEED2:
2178 log->l_covered_state = XLOG_STATE_COVER_NEED;
2181 case XLOG_STATE_COVER_DONE:
2183 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2185 log->l_covered_state = XLOG_STATE_COVER_NEED;
2188 case XLOG_STATE_COVER_DONE2:
2190 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2192 log->l_covered_state = XLOG_STATE_COVER_NEED;
2199 } /* xlog_state_clean_log */
2202 xlog_get_lowest_lsn(
2205 xlog_in_core_t *lsn_log;
2206 xfs_lsn_t lowest_lsn, lsn;
2208 lsn_log = log->l_iclog;
2211 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2212 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2213 if ((lsn && !lowest_lsn) ||
2214 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2218 lsn_log = lsn_log->ic_next;
2219 } while (lsn_log != log->l_iclog);
2225 xlog_state_do_callback(
2228 xlog_in_core_t *ciclog)
2230 xlog_in_core_t *iclog;
2231 xlog_in_core_t *first_iclog; /* used to know when we've
2232 * processed all iclogs once */
2233 xfs_log_callback_t *cb, *cb_next;
2235 xfs_lsn_t lowest_lsn;
2236 int ioerrors; /* counter: iclogs with errors */
2237 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2238 int funcdidcallbacks; /* flag: function did callbacks */
2239 int repeats; /* for issuing console warnings if
2240 * looping too many times */
2243 spin_lock(&log->l_icloglock);
2244 first_iclog = iclog = log->l_iclog;
2246 funcdidcallbacks = 0;
2251 * Scan all iclogs starting with the one pointed to by the
2252 * log. Reset this starting point each time the log is
2253 * unlocked (during callbacks).
2255 * Keep looping through iclogs until one full pass is made
2256 * without running any callbacks.
2258 first_iclog = log->l_iclog;
2259 iclog = log->l_iclog;
2260 loopdidcallbacks = 0;
2265 /* skip all iclogs in the ACTIVE & DIRTY states */
2266 if (iclog->ic_state &
2267 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2268 iclog = iclog->ic_next;
2273 * Between marking a filesystem SHUTDOWN and stopping
2274 * the log, we do flush all iclogs to disk (if there
2275 * wasn't a log I/O error). So, we do want things to
2276 * go smoothly in case of just a SHUTDOWN w/o a
2279 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2281 * Can only perform callbacks in order. Since
2282 * this iclog is not in the DONE_SYNC/
2283 * DO_CALLBACK state, we skip the rest and
2284 * just try to clean up. If we set our iclog
2285 * to DO_CALLBACK, we will not process it when
2286 * we retry since a previous iclog is in the
2287 * CALLBACK and the state cannot change since
2288 * we are holding the l_icloglock.
2290 if (!(iclog->ic_state &
2291 (XLOG_STATE_DONE_SYNC |
2292 XLOG_STATE_DO_CALLBACK))) {
2293 if (ciclog && (ciclog->ic_state ==
2294 XLOG_STATE_DONE_SYNC)) {
2295 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2300 * We now have an iclog that is in either the
2301 * DO_CALLBACK or DONE_SYNC states. The other
2302 * states (WANT_SYNC, SYNCING, or CALLBACK were
2303 * caught by the above if and are going to
2304 * clean (i.e. we aren't doing their callbacks)
2309 * We will do one more check here to see if we
2310 * have chased our tail around.
2313 lowest_lsn = xlog_get_lowest_lsn(log);
2315 XFS_LSN_CMP(lowest_lsn,
2316 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2317 iclog = iclog->ic_next;
2318 continue; /* Leave this iclog for
2322 iclog->ic_state = XLOG_STATE_CALLBACK;
2326 * update the last_sync_lsn before we drop the
2327 * icloglock to ensure we are the only one that
2330 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2331 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2332 atomic64_set(&log->l_last_sync_lsn,
2333 be64_to_cpu(iclog->ic_header.h_lsn));
2338 spin_unlock(&log->l_icloglock);
2341 * Keep processing entries in the callback list until
2342 * we come around and it is empty. We need to
2343 * atomically see that the list is empty and change the
2344 * state to DIRTY so that we don't miss any more
2345 * callbacks being added.
2347 spin_lock(&iclog->ic_callback_lock);
2348 cb = iclog->ic_callback;
2350 iclog->ic_callback_tail = &(iclog->ic_callback);
2351 iclog->ic_callback = NULL;
2352 spin_unlock(&iclog->ic_callback_lock);
2354 /* perform callbacks in the order given */
2355 for (; cb; cb = cb_next) {
2356 cb_next = cb->cb_next;
2357 cb->cb_func(cb->cb_arg, aborted);
2359 spin_lock(&iclog->ic_callback_lock);
2360 cb = iclog->ic_callback;
2366 spin_lock(&log->l_icloglock);
2367 ASSERT(iclog->ic_callback == NULL);
2368 spin_unlock(&iclog->ic_callback_lock);
2369 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2370 iclog->ic_state = XLOG_STATE_DIRTY;
2373 * Transition from DIRTY to ACTIVE if applicable.
2374 * NOP if STATE_IOERROR.
2376 xlog_state_clean_log(log);
2378 /* wake up threads waiting in xfs_log_force() */
2379 wake_up_all(&iclog->ic_force_wait);
2381 iclog = iclog->ic_next;
2382 } while (first_iclog != iclog);
2384 if (repeats > 5000) {
2385 flushcnt += repeats;
2388 "%s: possible infinite loop (%d iterations)",
2389 __func__, flushcnt);
2391 } while (!ioerrors && loopdidcallbacks);
2394 * make one last gasp attempt to see if iclogs are being left in
2398 if (funcdidcallbacks) {
2399 first_iclog = iclog = log->l_iclog;
2401 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2403 * Terminate the loop if iclogs are found in states
2404 * which will cause other threads to clean up iclogs.
2406 * SYNCING - i/o completion will go through logs
2407 * DONE_SYNC - interrupt thread should be waiting for
2409 * IOERROR - give up hope all ye who enter here
2411 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2412 iclog->ic_state == XLOG_STATE_SYNCING ||
2413 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2414 iclog->ic_state == XLOG_STATE_IOERROR )
2416 iclog = iclog->ic_next;
2417 } while (first_iclog != iclog);
2421 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2423 spin_unlock(&log->l_icloglock);
2426 wake_up_all(&log->l_flush_wait);
2431 * Finish transitioning this iclog to the dirty state.
2433 * Make sure that we completely execute this routine only when this is
2434 * the last call to the iclog. There is a good chance that iclog flushes,
2435 * when we reach the end of the physical log, get turned into 2 separate
2436 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2437 * routine. By using the reference count bwritecnt, we guarantee that only
2438 * the second completion goes through.
2440 * Callbacks could take time, so they are done outside the scope of the
2441 * global state machine log lock.
2444 xlog_state_done_syncing(
2445 xlog_in_core_t *iclog,
2448 xlog_t *log = iclog->ic_log;
2450 spin_lock(&log->l_icloglock);
2452 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2453 iclog->ic_state == XLOG_STATE_IOERROR);
2454 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2455 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2459 * If we got an error, either on the first buffer, or in the case of
2460 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2461 * and none should ever be attempted to be written to disk
2464 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2465 if (--iclog->ic_bwritecnt == 1) {
2466 spin_unlock(&log->l_icloglock);
2469 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2473 * Someone could be sleeping prior to writing out the next
2474 * iclog buffer, we wake them all, one will get to do the
2475 * I/O, the others get to wait for the result.
2477 wake_up_all(&iclog->ic_write_wait);
2478 spin_unlock(&log->l_icloglock);
2479 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2480 } /* xlog_state_done_syncing */
2484 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2485 * sleep. We wait on the flush queue on the head iclog as that should be
2486 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2487 * we will wait here and all new writes will sleep until a sync completes.
2489 * The in-core logs are used in a circular fashion. They are not used
2490 * out-of-order even when an iclog past the head is free.
2493 * * log_offset where xlog_write() can start writing into the in-core
2495 * * in-core log pointer to which xlog_write() should write.
2496 * * boolean indicating this is a continued write to an in-core log.
2497 * If this is the last write, then the in-core log's offset field
2498 * needs to be incremented, depending on the amount of data which
2502 xlog_state_get_iclog_space(xlog_t *log,
2504 xlog_in_core_t **iclogp,
2505 xlog_ticket_t *ticket,
2506 int *continued_write,
2510 xlog_rec_header_t *head;
2511 xlog_in_core_t *iclog;
2515 spin_lock(&log->l_icloglock);
2516 if (XLOG_FORCED_SHUTDOWN(log)) {
2517 spin_unlock(&log->l_icloglock);
2518 return XFS_ERROR(EIO);
2521 iclog = log->l_iclog;
2522 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2523 XFS_STATS_INC(xs_log_noiclogs);
2525 /* Wait for log writes to have flushed */
2526 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2530 head = &iclog->ic_header;
2532 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2533 log_offset = iclog->ic_offset;
2535 /* On the 1st write to an iclog, figure out lsn. This works
2536 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2537 * committing to. If the offset is set, that's how many blocks
2540 if (log_offset == 0) {
2541 ticket->t_curr_res -= log->l_iclog_hsize;
2542 xlog_tic_add_region(ticket,
2544 XLOG_REG_TYPE_LRHEADER);
2545 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2546 head->h_lsn = cpu_to_be64(
2547 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2548 ASSERT(log->l_curr_block >= 0);
2551 /* If there is enough room to write everything, then do it. Otherwise,
2552 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2553 * bit is on, so this will get flushed out. Don't update ic_offset
2554 * until you know exactly how many bytes get copied. Therefore, wait
2555 * until later to update ic_offset.
2557 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2558 * can fit into remaining data section.
2560 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2561 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2564 * If I'm the only one writing to this iclog, sync it to disk.
2565 * We need to do an atomic compare and decrement here to avoid
2566 * racing with concurrent atomic_dec_and_lock() calls in
2567 * xlog_state_release_iclog() when there is more than one
2568 * reference to the iclog.
2570 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2571 /* we are the only one */
2572 spin_unlock(&log->l_icloglock);
2573 error = xlog_state_release_iclog(log, iclog);
2577 spin_unlock(&log->l_icloglock);
2582 /* Do we have enough room to write the full amount in the remainder
2583 * of this iclog? Or must we continue a write on the next iclog and
2584 * mark this iclog as completely taken? In the case where we switch
2585 * iclogs (to mark it taken), this particular iclog will release/sync
2586 * to disk in xlog_write().
2588 if (len <= iclog->ic_size - iclog->ic_offset) {
2589 *continued_write = 0;
2590 iclog->ic_offset += len;
2592 *continued_write = 1;
2593 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2597 ASSERT(iclog->ic_offset <= iclog->ic_size);
2598 spin_unlock(&log->l_icloglock);
2600 *logoffsetp = log_offset;
2602 } /* xlog_state_get_iclog_space */
2605 * Atomically get the log space required for a log ticket.
2607 * Once a ticket gets put onto the reserveq, it will only return after the
2608 * needed reservation is satisfied.
2610 * This function is structured so that it has a lock free fast path. This is
2611 * necessary because every new transaction reservation will come through this
2612 * path. Hence any lock will be globally hot if we take it unconditionally on
2615 * As tickets are only ever moved on and off the reserveq under the
2616 * l_grant_reserve_lock, we only need to take that lock if we are going to add
2617 * the ticket to the queue and sleep. We can avoid taking the lock if the ticket
2618 * was never added to the reserveq because the t_queue list head will be empty
2619 * and we hold the only reference to it so it can safely be checked unlocked.
2622 xlog_grant_log_space(
2624 struct xlog_ticket *tic)
2626 int free_bytes, need_bytes;
2629 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
2631 trace_xfs_log_grant_enter(log, tic);
2634 * If there are other waiters on the queue then give them a chance at
2635 * logspace before us. Wake up the first waiters, if we do not wake
2636 * up all the waiters then go to sleep waiting for more free space,
2637 * otherwise try to get some space for this transaction.
2639 need_bytes = tic->t_unit_res;
2640 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2641 need_bytes *= tic->t_ocnt;
2642 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
2643 if (!list_empty_careful(&log->l_reserveq)) {
2644 spin_lock(&log->l_grant_reserve_lock);
2645 if (!xlog_reserveq_wake(log, &free_bytes) ||
2646 free_bytes < need_bytes)
2647 error = xlog_reserveq_wait(log, tic, need_bytes);
2648 spin_unlock(&log->l_grant_reserve_lock);
2649 } else if (free_bytes < need_bytes) {
2650 spin_lock(&log->l_grant_reserve_lock);
2651 error = xlog_reserveq_wait(log, tic, need_bytes);
2652 spin_unlock(&log->l_grant_reserve_lock);
2657 xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes);
2658 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2659 trace_xfs_log_grant_exit(log, tic);
2660 xlog_verify_grant_tail(log);
2665 * Replenish the byte reservation required by moving the grant write head.
2667 * Similar to xlog_grant_log_space, the function is structured to have a lock
2671 xlog_regrant_write_log_space(
2673 struct xlog_ticket *tic)
2675 int free_bytes, need_bytes;
2678 tic->t_curr_res = tic->t_unit_res;
2679 xlog_tic_reset_res(tic);
2684 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
2686 trace_xfs_log_regrant_write_enter(log, tic);
2689 * If there are other waiters on the queue then give them a chance at
2690 * logspace before us. Wake up the first waiters, if we do not wake
2691 * up all the waiters then go to sleep waiting for more free space,
2692 * otherwise try to get some space for this transaction.
2694 need_bytes = tic->t_unit_res;
2695 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2696 if (!list_empty_careful(&log->l_writeq)) {
2697 spin_lock(&log->l_grant_write_lock);
2698 if (!xlog_writeq_wake(log, &free_bytes) ||
2699 free_bytes < need_bytes)
2700 error = xlog_writeq_wait(log, tic, need_bytes);
2701 spin_unlock(&log->l_grant_write_lock);
2702 } else if (free_bytes < need_bytes) {
2703 spin_lock(&log->l_grant_write_lock);
2704 error = xlog_writeq_wait(log, tic, need_bytes);
2705 spin_unlock(&log->l_grant_write_lock);
2711 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2712 trace_xfs_log_regrant_write_exit(log, tic);
2713 xlog_verify_grant_tail(log);
2717 /* The first cnt-1 times through here we don't need to
2718 * move the grant write head because the permanent
2719 * reservation has reserved cnt times the unit amount.
2720 * Release part of current permanent unit reservation and
2721 * reset current reservation to be one units worth. Also
2722 * move grant reservation head forward.
2725 xlog_regrant_reserve_log_space(xlog_t *log,
2726 xlog_ticket_t *ticket)
2728 trace_xfs_log_regrant_reserve_enter(log, ticket);
2730 if (ticket->t_cnt > 0)
2733 xlog_grant_sub_space(log, &log->l_grant_reserve_head,
2734 ticket->t_curr_res);
2735 xlog_grant_sub_space(log, &log->l_grant_write_head,
2736 ticket->t_curr_res);
2737 ticket->t_curr_res = ticket->t_unit_res;
2738 xlog_tic_reset_res(ticket);
2740 trace_xfs_log_regrant_reserve_sub(log, ticket);
2742 /* just return if we still have some of the pre-reserved space */
2743 if (ticket->t_cnt > 0)
2746 xlog_grant_add_space(log, &log->l_grant_reserve_head,
2747 ticket->t_unit_res);
2749 trace_xfs_log_regrant_reserve_exit(log, ticket);
2751 ticket->t_curr_res = ticket->t_unit_res;
2752 xlog_tic_reset_res(ticket);
2753 } /* xlog_regrant_reserve_log_space */
2757 * Give back the space left from a reservation.
2759 * All the information we need to make a correct determination of space left
2760 * is present. For non-permanent reservations, things are quite easy. The
2761 * count should have been decremented to zero. We only need to deal with the
2762 * space remaining in the current reservation part of the ticket. If the
2763 * ticket contains a permanent reservation, there may be left over space which
2764 * needs to be released. A count of N means that N-1 refills of the current
2765 * reservation can be done before we need to ask for more space. The first
2766 * one goes to fill up the first current reservation. Once we run out of
2767 * space, the count will stay at zero and the only space remaining will be
2768 * in the current reservation field.
2771 xlog_ungrant_log_space(xlog_t *log,
2772 xlog_ticket_t *ticket)
2776 if (ticket->t_cnt > 0)
2779 trace_xfs_log_ungrant_enter(log, ticket);
2780 trace_xfs_log_ungrant_sub(log, ticket);
2783 * If this is a permanent reservation ticket, we may be able to free
2784 * up more space based on the remaining count.
2786 bytes = ticket->t_curr_res;
2787 if (ticket->t_cnt > 0) {
2788 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2789 bytes += ticket->t_unit_res*ticket->t_cnt;
2792 xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes);
2793 xlog_grant_sub_space(log, &log->l_grant_write_head, bytes);
2795 trace_xfs_log_ungrant_exit(log, ticket);
2797 xfs_log_move_tail(log->l_mp, 1);
2798 } /* xlog_ungrant_log_space */
2802 * Flush iclog to disk if this is the last reference to the given iclog and
2803 * the WANT_SYNC bit is set.
2805 * When this function is entered, the iclog is not necessarily in the
2806 * WANT_SYNC state. It may be sitting around waiting to get filled.
2811 xlog_state_release_iclog(
2813 xlog_in_core_t *iclog)
2815 int sync = 0; /* do we sync? */
2817 if (iclog->ic_state & XLOG_STATE_IOERROR)
2818 return XFS_ERROR(EIO);
2820 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2821 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2824 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2825 spin_unlock(&log->l_icloglock);
2826 return XFS_ERROR(EIO);
2828 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2829 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2831 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2832 /* update tail before writing to iclog */
2833 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2835 iclog->ic_state = XLOG_STATE_SYNCING;
2836 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2837 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2838 /* cycle incremented when incrementing curr_block */
2840 spin_unlock(&log->l_icloglock);
2843 * We let the log lock go, so it's possible that we hit a log I/O
2844 * error or some other SHUTDOWN condition that marks the iclog
2845 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2846 * this iclog has consistent data, so we ignore IOERROR
2847 * flags after this point.
2850 return xlog_sync(log, iclog);
2852 } /* xlog_state_release_iclog */
2856 * This routine will mark the current iclog in the ring as WANT_SYNC
2857 * and move the current iclog pointer to the next iclog in the ring.
2858 * When this routine is called from xlog_state_get_iclog_space(), the
2859 * exact size of the iclog has not yet been determined. All we know is
2860 * that every data block. We have run out of space in this log record.
2863 xlog_state_switch_iclogs(xlog_t *log,
2864 xlog_in_core_t *iclog,
2867 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2869 eventual_size = iclog->ic_offset;
2870 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2871 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2872 log->l_prev_block = log->l_curr_block;
2873 log->l_prev_cycle = log->l_curr_cycle;
2875 /* roll log?: ic_offset changed later */
2876 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2878 /* Round up to next log-sunit */
2879 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2880 log->l_mp->m_sb.sb_logsunit > 1) {
2881 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2882 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2885 if (log->l_curr_block >= log->l_logBBsize) {
2886 log->l_curr_cycle++;
2887 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2888 log->l_curr_cycle++;
2889 log->l_curr_block -= log->l_logBBsize;
2890 ASSERT(log->l_curr_block >= 0);
2892 ASSERT(iclog == log->l_iclog);
2893 log->l_iclog = iclog->ic_next;
2894 } /* xlog_state_switch_iclogs */
2897 * Write out all data in the in-core log as of this exact moment in time.
2899 * Data may be written to the in-core log during this call. However,
2900 * we don't guarantee this data will be written out. A change from past
2901 * implementation means this routine will *not* write out zero length LRs.
2903 * Basically, we try and perform an intelligent scan of the in-core logs.
2904 * If we determine there is no flushable data, we just return. There is no
2905 * flushable data if:
2907 * 1. the current iclog is active and has no data; the previous iclog
2908 * is in the active or dirty state.
2909 * 2. the current iclog is drity, and the previous iclog is in the
2910 * active or dirty state.
2914 * 1. the current iclog is not in the active nor dirty state.
2915 * 2. the current iclog dirty, and the previous iclog is not in the
2916 * active nor dirty state.
2917 * 3. the current iclog is active, and there is another thread writing
2918 * to this particular iclog.
2919 * 4. a) the current iclog is active and has no other writers
2920 * b) when we return from flushing out this iclog, it is still
2921 * not in the active nor dirty state.
2925 struct xfs_mount *mp,
2929 struct log *log = mp->m_log;
2930 struct xlog_in_core *iclog;
2933 XFS_STATS_INC(xs_log_force);
2936 xlog_cil_force(log);
2938 spin_lock(&log->l_icloglock);
2940 iclog = log->l_iclog;
2941 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2942 spin_unlock(&log->l_icloglock);
2943 return XFS_ERROR(EIO);
2946 /* If the head iclog is not active nor dirty, we just attach
2947 * ourselves to the head and go to sleep.
2949 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2950 iclog->ic_state == XLOG_STATE_DIRTY) {
2952 * If the head is dirty or (active and empty), then
2953 * we need to look at the previous iclog. If the previous
2954 * iclog is active or dirty we are done. There is nothing
2955 * to sync out. Otherwise, we attach ourselves to the
2956 * previous iclog and go to sleep.
2958 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2959 (atomic_read(&iclog->ic_refcnt) == 0
2960 && iclog->ic_offset == 0)) {
2961 iclog = iclog->ic_prev;
2962 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2963 iclog->ic_state == XLOG_STATE_DIRTY)
2968 if (atomic_read(&iclog->ic_refcnt) == 0) {
2969 /* We are the only one with access to this
2970 * iclog. Flush it out now. There should
2971 * be a roundoff of zero to show that someone
2972 * has already taken care of the roundoff from
2973 * the previous sync.
2975 atomic_inc(&iclog->ic_refcnt);
2976 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2977 xlog_state_switch_iclogs(log, iclog, 0);
2978 spin_unlock(&log->l_icloglock);
2980 if (xlog_state_release_iclog(log, iclog))
2981 return XFS_ERROR(EIO);
2985 spin_lock(&log->l_icloglock);
2986 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2987 iclog->ic_state != XLOG_STATE_DIRTY)
2992 /* Someone else is writing to this iclog.
2993 * Use its call to flush out the data. However,
2994 * the other thread may not force out this LR,
2995 * so we mark it WANT_SYNC.
2997 xlog_state_switch_iclogs(log, iclog, 0);
3003 /* By the time we come around again, the iclog could've been filled
3004 * which would give it another lsn. If we have a new lsn, just
3005 * return because the relevant data has been flushed.
3008 if (flags & XFS_LOG_SYNC) {
3010 * We must check if we're shutting down here, before
3011 * we wait, while we're holding the l_icloglock.
3012 * Then we check again after waking up, in case our
3013 * sleep was disturbed by a bad news.
3015 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3016 spin_unlock(&log->l_icloglock);
3017 return XFS_ERROR(EIO);
3019 XFS_STATS_INC(xs_log_force_sleep);
3020 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3022 * No need to grab the log lock here since we're
3023 * only deciding whether or not to return EIO
3024 * and the memory read should be atomic.
3026 if (iclog->ic_state & XLOG_STATE_IOERROR)
3027 return XFS_ERROR(EIO);
3033 spin_unlock(&log->l_icloglock);
3039 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3040 * about errors or whether the log was flushed or not. This is the normal
3041 * interface to use when trying to unpin items or move the log forward.
3050 error = _xfs_log_force(mp, flags, NULL);
3052 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3056 * Force the in-core log to disk for a specific LSN.
3058 * Find in-core log with lsn.
3059 * If it is in the DIRTY state, just return.
3060 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3061 * state and go to sleep or return.
3062 * If it is in any other state, go to sleep or return.
3064 * Synchronous forces are implemented with a signal variable. All callers
3065 * to force a given lsn to disk will wait on a the sv attached to the
3066 * specific in-core log. When given in-core log finally completes its
3067 * write to disk, that thread will wake up all threads waiting on the
3072 struct xfs_mount *mp,
3077 struct log *log = mp->m_log;
3078 struct xlog_in_core *iclog;
3079 int already_slept = 0;
3083 XFS_STATS_INC(xs_log_force);
3086 lsn = xlog_cil_force_lsn(log, lsn);
3087 if (lsn == NULLCOMMITLSN)
3092 spin_lock(&log->l_icloglock);
3093 iclog = log->l_iclog;
3094 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3095 spin_unlock(&log->l_icloglock);
3096 return XFS_ERROR(EIO);
3100 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3101 iclog = iclog->ic_next;
3105 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3106 spin_unlock(&log->l_icloglock);
3110 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3112 * We sleep here if we haven't already slept (e.g.
3113 * this is the first time we've looked at the correct
3114 * iclog buf) and the buffer before us is going to
3115 * be sync'ed. The reason for this is that if we
3116 * are doing sync transactions here, by waiting for
3117 * the previous I/O to complete, we can allow a few
3118 * more transactions into this iclog before we close
3121 * Otherwise, we mark the buffer WANT_SYNC, and bump
3122 * up the refcnt so we can release the log (which
3123 * drops the ref count). The state switch keeps new
3124 * transaction commits from using this buffer. When
3125 * the current commits finish writing into the buffer,
3126 * the refcount will drop to zero and the buffer will
3129 if (!already_slept &&
3130 (iclog->ic_prev->ic_state &
3131 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3132 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3134 XFS_STATS_INC(xs_log_force_sleep);
3136 xlog_wait(&iclog->ic_prev->ic_write_wait,
3143 atomic_inc(&iclog->ic_refcnt);
3144 xlog_state_switch_iclogs(log, iclog, 0);
3145 spin_unlock(&log->l_icloglock);
3146 if (xlog_state_release_iclog(log, iclog))
3147 return XFS_ERROR(EIO);
3150 spin_lock(&log->l_icloglock);
3153 if ((flags & XFS_LOG_SYNC) && /* sleep */
3155 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3157 * Don't wait on completion if we know that we've
3158 * gotten a log write error.
3160 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3161 spin_unlock(&log->l_icloglock);
3162 return XFS_ERROR(EIO);
3164 XFS_STATS_INC(xs_log_force_sleep);
3165 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3167 * No need to grab the log lock here since we're
3168 * only deciding whether or not to return EIO
3169 * and the memory read should be atomic.
3171 if (iclog->ic_state & XLOG_STATE_IOERROR)
3172 return XFS_ERROR(EIO);
3176 } else { /* just return */
3177 spin_unlock(&log->l_icloglock);
3181 } while (iclog != log->l_iclog);
3183 spin_unlock(&log->l_icloglock);
3188 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3189 * about errors or whether the log was flushed or not. This is the normal
3190 * interface to use when trying to unpin items or move the log forward.
3200 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3202 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3206 * Called when we want to mark the current iclog as being ready to sync to
3210 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3212 assert_spin_locked(&log->l_icloglock);
3214 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3215 xlog_state_switch_iclogs(log, iclog, 0);
3217 ASSERT(iclog->ic_state &
3218 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3223 /*****************************************************************************
3227 *****************************************************************************
3231 * Free a used ticket when its refcount falls to zero.
3235 xlog_ticket_t *ticket)
3237 ASSERT(atomic_read(&ticket->t_ref) > 0);
3238 if (atomic_dec_and_test(&ticket->t_ref))
3239 kmem_zone_free(xfs_log_ticket_zone, ticket);
3244 xlog_ticket_t *ticket)
3246 ASSERT(atomic_read(&ticket->t_ref) > 0);
3247 atomic_inc(&ticket->t_ref);
3252 * Allocate and initialise a new log ticket.
3263 struct xlog_ticket *tic;
3267 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3272 * Permanent reservations have up to 'cnt'-1 active log operations
3273 * in the log. A unit in this case is the amount of space for one
3274 * of these log operations. Normal reservations have a cnt of 1
3275 * and their unit amount is the total amount of space required.
3277 * The following lines of code account for non-transaction data
3278 * which occupy space in the on-disk log.
3280 * Normal form of a transaction is:
3281 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3282 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3284 * We need to account for all the leadup data and trailer data
3285 * around the transaction data.
3286 * And then we need to account for the worst case in terms of using
3288 * The worst case will happen if:
3289 * - the placement of the transaction happens to be such that the
3290 * roundoff is at its maximum
3291 * - the transaction data is synced before the commit record is synced
3292 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3293 * Therefore the commit record is in its own Log Record.
3294 * This can happen as the commit record is called with its
3295 * own region to xlog_write().
3296 * This then means that in the worst case, roundoff can happen for
3297 * the commit-rec as well.
3298 * The commit-rec is smaller than padding in this scenario and so it is
3299 * not added separately.
3302 /* for trans header */
3303 unit_bytes += sizeof(xlog_op_header_t);
3304 unit_bytes += sizeof(xfs_trans_header_t);
3307 unit_bytes += sizeof(xlog_op_header_t);
3310 * for LR headers - the space for data in an iclog is the size minus
3311 * the space used for the headers. If we use the iclog size, then we
3312 * undercalculate the number of headers required.
3314 * Furthermore - the addition of op headers for split-recs might
3315 * increase the space required enough to require more log and op
3316 * headers, so take that into account too.
3318 * IMPORTANT: This reservation makes the assumption that if this
3319 * transaction is the first in an iclog and hence has the LR headers
3320 * accounted to it, then the remaining space in the iclog is
3321 * exclusively for this transaction. i.e. if the transaction is larger
3322 * than the iclog, it will be the only thing in that iclog.
3323 * Fundamentally, this means we must pass the entire log vector to
3324 * xlog_write to guarantee this.
3326 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3327 num_headers = howmany(unit_bytes, iclog_space);
3329 /* for split-recs - ophdrs added when data split over LRs */
3330 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3332 /* add extra header reservations if we overrun */
3333 while (!num_headers ||
3334 howmany(unit_bytes, iclog_space) > num_headers) {
3335 unit_bytes += sizeof(xlog_op_header_t);
3338 unit_bytes += log->l_iclog_hsize * num_headers;
3340 /* for commit-rec LR header - note: padding will subsume the ophdr */
3341 unit_bytes += log->l_iclog_hsize;
3343 /* for roundoff padding for transaction data and one for commit record */
3344 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3345 log->l_mp->m_sb.sb_logsunit > 1) {
3346 /* log su roundoff */
3347 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3350 unit_bytes += 2*BBSIZE;
3353 atomic_set(&tic->t_ref, 1);
3354 INIT_LIST_HEAD(&tic->t_queue);
3355 tic->t_unit_res = unit_bytes;
3356 tic->t_curr_res = unit_bytes;
3359 tic->t_tid = random32();
3360 tic->t_clientid = client;
3361 tic->t_flags = XLOG_TIC_INITED;
3362 tic->t_trans_type = 0;
3363 if (xflags & XFS_LOG_PERM_RESERV)
3364 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3365 init_waitqueue_head(&tic->t_wait);
3367 xlog_tic_reset_res(tic);
3373 /******************************************************************************
3375 * Log debug routines
3377 ******************************************************************************
3381 * Make sure that the destination ptr is within the valid data region of
3382 * one of the iclogs. This uses backup pointers stored in a different
3383 * part of the log in case we trash the log structure.
3386 xlog_verify_dest_ptr(
3393 for (i = 0; i < log->l_iclog_bufs; i++) {
3394 if (ptr >= log->l_iclog_bak[i] &&
3395 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3400 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3404 * Check to make sure the grant write head didn't just over lap the tail. If
3405 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3406 * the cycles differ by exactly one and check the byte count.
3408 * This check is run unlocked, so can give false positives. Rather than assert
3409 * on failures, use a warn-once flag and a panic tag to allow the admin to
3410 * determine if they want to panic the machine when such an error occurs. For
3411 * debug kernels this will have the same effect as using an assert but, unlinke
3412 * an assert, it can be turned off at runtime.
3415 xlog_verify_grant_tail(
3418 int tail_cycle, tail_blocks;
3421 xlog_crack_grant_head(&log->l_grant_write_head, &cycle, &space);
3422 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3423 if (tail_cycle != cycle) {
3424 if (cycle - 1 != tail_cycle &&
3425 !(log->l_flags & XLOG_TAIL_WARN)) {
3426 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3427 "%s: cycle - 1 != tail_cycle", __func__);
3428 log->l_flags |= XLOG_TAIL_WARN;
3431 if (space > BBTOB(tail_blocks) &&
3432 !(log->l_flags & XLOG_TAIL_WARN)) {
3433 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3434 "%s: space > BBTOB(tail_blocks)", __func__);
3435 log->l_flags |= XLOG_TAIL_WARN;
3440 /* check if it will fit */
3442 xlog_verify_tail_lsn(xlog_t *log,
3443 xlog_in_core_t *iclog,
3448 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3450 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3451 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3452 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3454 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3456 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3457 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3459 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3460 if (blocks < BTOBB(iclog->ic_offset) + 1)
3461 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3463 } /* xlog_verify_tail_lsn */
3466 * Perform a number of checks on the iclog before writing to disk.
3468 * 1. Make sure the iclogs are still circular
3469 * 2. Make sure we have a good magic number
3470 * 3. Make sure we don't have magic numbers in the data
3471 * 4. Check fields of each log operation header for:
3472 * A. Valid client identifier
3473 * B. tid ptr value falls in valid ptr space (user space code)
3474 * C. Length in log record header is correct according to the
3475 * individual operation headers within record.
3476 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3477 * log, check the preceding blocks of the physical log to make sure all
3478 * the cycle numbers agree with the current cycle number.
3481 xlog_verify_iclog(xlog_t *log,
3482 xlog_in_core_t *iclog,
3486 xlog_op_header_t *ophead;
3487 xlog_in_core_t *icptr;
3488 xlog_in_core_2_t *xhdr;
3490 xfs_caddr_t base_ptr;
3491 __psint_t field_offset;
3493 int len, i, j, k, op_len;
3496 /* check validity of iclog pointers */
3497 spin_lock(&log->l_icloglock);
3498 icptr = log->l_iclog;
3499 for (i=0; i < log->l_iclog_bufs; i++) {
3501 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3502 icptr = icptr->ic_next;
3504 if (icptr != log->l_iclog)
3505 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3506 spin_unlock(&log->l_icloglock);
3508 /* check log magic numbers */
3509 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3510 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3512 ptr = (xfs_caddr_t) &iclog->ic_header;
3513 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3515 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3516 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3521 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3522 ptr = iclog->ic_datap;
3524 ophead = (xlog_op_header_t *)ptr;
3525 xhdr = iclog->ic_data;
3526 for (i = 0; i < len; i++) {
3527 ophead = (xlog_op_header_t *)ptr;
3529 /* clientid is only 1 byte */
3530 field_offset = (__psint_t)
3531 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3532 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3533 clientid = ophead->oh_clientid;
3535 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3536 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3537 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3538 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3539 clientid = xlog_get_client_id(
3540 xhdr[j].hic_xheader.xh_cycle_data[k]);
3542 clientid = xlog_get_client_id(
3543 iclog->ic_header.h_cycle_data[idx]);
3546 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3548 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3549 __func__, clientid, ophead,
3550 (unsigned long)field_offset);
3553 field_offset = (__psint_t)
3554 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3555 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3556 op_len = be32_to_cpu(ophead->oh_len);
3558 idx = BTOBBT((__psint_t)&ophead->oh_len -
3559 (__psint_t)iclog->ic_datap);
3560 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3561 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3562 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3563 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3565 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3568 ptr += sizeof(xlog_op_header_t) + op_len;
3570 } /* xlog_verify_iclog */
3574 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3580 xlog_in_core_t *iclog, *ic;
3582 iclog = log->l_iclog;
3583 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3585 * Mark all the incore logs IOERROR.
3586 * From now on, no log flushes will result.
3590 ic->ic_state = XLOG_STATE_IOERROR;
3592 } while (ic != iclog);
3596 * Return non-zero, if state transition has already happened.
3602 * This is called from xfs_force_shutdown, when we're forcibly
3603 * shutting down the filesystem, typically because of an IO error.
3604 * Our main objectives here are to make sure that:
3605 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3606 * parties to find out, 'atomically'.
3607 * b. those who're sleeping on log reservations, pinned objects and
3608 * other resources get woken up, and be told the bad news.
3609 * c. nothing new gets queued up after (a) and (b) are done.
3610 * d. if !logerror, flush the iclogs to disk, then seal them off
3613 * Note: for delayed logging the !logerror case needs to flush the regions
3614 * held in memory out to the iclogs before flushing them to disk. This needs
3615 * to be done before the log is marked as shutdown, otherwise the flush to the
3619 xfs_log_force_umount(
3620 struct xfs_mount *mp,
3630 * If this happens during log recovery, don't worry about
3631 * locking; the log isn't open for business yet.
3634 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3635 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3637 XFS_BUF_DONE(mp->m_sb_bp);
3642 * Somebody could've already done the hard work for us.
3643 * No need to get locks for this.
3645 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3646 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3652 * Flush the in memory commit item list before marking the log as
3653 * being shut down. We need to do it in this order to ensure all the
3654 * completed transactions are flushed to disk with the xfs_log_force()
3657 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3658 xlog_cil_force(log);
3661 * mark the filesystem and the as in a shutdown state and wake
3662 * everybody up to tell them the bad news.
3664 spin_lock(&log->l_icloglock);
3665 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3667 XFS_BUF_DONE(mp->m_sb_bp);
3670 * This flag is sort of redundant because of the mount flag, but
3671 * it's good to maintain the separation between the log and the rest
3674 log->l_flags |= XLOG_IO_ERROR;
3677 * If we hit a log error, we want to mark all the iclogs IOERROR
3678 * while we're still holding the loglock.
3681 retval = xlog_state_ioerror(log);
3682 spin_unlock(&log->l_icloglock);
3685 * We don't want anybody waiting for log reservations after this. That
3686 * means we have to wake up everybody queued up on reserveq as well as
3687 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3688 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3689 * action is protected by the grant locks.
3691 spin_lock(&log->l_grant_reserve_lock);
3692 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3693 wake_up(&tic->t_wait);
3694 spin_unlock(&log->l_grant_reserve_lock);
3696 spin_lock(&log->l_grant_write_lock);
3697 list_for_each_entry(tic, &log->l_writeq, t_queue)
3698 wake_up(&tic->t_wait);
3699 spin_unlock(&log->l_grant_write_lock);
3701 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3704 * Force the incore logs to disk before shutting the
3705 * log down completely.
3707 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3709 spin_lock(&log->l_icloglock);
3710 retval = xlog_state_ioerror(log);
3711 spin_unlock(&log->l_icloglock);
3714 * Wake up everybody waiting on xfs_log_force.
3715 * Callback all log item committed functions as if the
3716 * log writes were completed.
3718 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3720 #ifdef XFSERRORDEBUG
3722 xlog_in_core_t *iclog;
3724 spin_lock(&log->l_icloglock);
3725 iclog = log->l_iclog;
3727 ASSERT(iclog->ic_callback == 0);
3728 iclog = iclog->ic_next;
3729 } while (iclog != log->l_iclog);
3730 spin_unlock(&log->l_icloglock);
3733 /* return non-zero if log IOERROR transition had already happened */
3738 xlog_iclogs_empty(xlog_t *log)
3740 xlog_in_core_t *iclog;
3742 iclog = log->l_iclog;
3744 /* endianness does not matter here, zero is zero in
3747 if (iclog->ic_header.h_num_logops)
3749 iclog = iclog->ic_next;
3750 } while (iclog != log->l_iclog);