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);
154 xlog_tic_reset_res(xlog_ticket_t *tic)
157 tic->t_res_arr_sum = 0;
158 tic->t_res_num_ophdrs = 0;
162 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
164 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
165 /* add to overflow and start again */
166 tic->t_res_o_flow += tic->t_res_arr_sum;
168 tic->t_res_arr_sum = 0;
171 tic->t_res_arr[tic->t_res_num].r_len = len;
172 tic->t_res_arr[tic->t_res_num].r_type = type;
173 tic->t_res_arr_sum += len;
180 * 1. currblock field gets updated at startup and after in-core logs
181 * marked as with WANT_SYNC.
185 * This routine is called when a user of a log manager ticket is done with
186 * the reservation. If the ticket was ever used, then a commit record for
187 * the associated transaction is written out as a log operation header with
188 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
189 * a given ticket. If the ticket was one with a permanent reservation, then
190 * a few operations are done differently. Permanent reservation tickets by
191 * default don't release the reservation. They just commit the current
192 * transaction with the belief that the reservation is still needed. A flag
193 * must be passed in before permanent reservations are actually released.
194 * When these type of tickets are not released, they need to be set into
195 * the inited state again. By doing this, a start record will be written
196 * out when the next write occurs.
200 struct xfs_mount *mp,
201 struct xlog_ticket *ticket,
202 struct xlog_in_core **iclog,
205 struct log *log = mp->m_log;
208 if (XLOG_FORCED_SHUTDOWN(log) ||
210 * If nothing was ever written, don't write out commit record.
211 * If we get an error, just continue and give back the log ticket.
213 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
214 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
215 lsn = (xfs_lsn_t) -1;
216 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
217 flags |= XFS_LOG_REL_PERM_RESERV;
222 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
223 (flags & XFS_LOG_REL_PERM_RESERV)) {
224 trace_xfs_log_done_nonperm(log, ticket);
227 * Release ticket if not permanent reservation or a specific
228 * request has been made to release a permanent reservation.
230 xlog_ungrant_log_space(log, ticket);
231 xfs_log_ticket_put(ticket);
233 trace_xfs_log_done_perm(log, ticket);
235 xlog_regrant_reserve_log_space(log, ticket);
236 /* If this ticket was a permanent reservation and we aren't
237 * trying to release it, reset the inited flags; so next time
238 * we write, a start record will be written out.
240 ticket->t_flags |= XLOG_TIC_INITED;
247 * Attaches a new iclog I/O completion callback routine during
248 * transaction commit. If the log is in error state, a non-zero
249 * return code is handed back and the caller is responsible for
250 * executing the callback at an appropriate time.
254 struct xfs_mount *mp,
255 struct xlog_in_core *iclog,
256 xfs_log_callback_t *cb)
260 spin_lock(&iclog->ic_callback_lock);
261 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
263 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
264 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
266 *(iclog->ic_callback_tail) = cb;
267 iclog->ic_callback_tail = &(cb->cb_next);
269 spin_unlock(&iclog->ic_callback_lock);
274 xfs_log_release_iclog(
275 struct xfs_mount *mp,
276 struct xlog_in_core *iclog)
278 if (xlog_state_release_iclog(mp->m_log, iclog)) {
279 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
287 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
288 * to the reservation.
289 * 2. Potentially, push buffers at tail of log to disk.
291 * Each reservation is going to reserve extra space for a log record header.
292 * When writes happen to the on-disk log, we don't subtract the length of the
293 * log record header from any reservation. By wasting space in each
294 * reservation, we prevent over allocation problems.
298 struct xfs_mount *mp,
301 struct xlog_ticket **ticket,
306 struct log *log = mp->m_log;
307 struct xlog_ticket *internal_ticket;
310 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
312 if (XLOG_FORCED_SHUTDOWN(log))
313 return XFS_ERROR(EIO);
315 XFS_STATS_INC(xs_try_logspace);
318 if (*ticket != NULL) {
319 ASSERT(flags & XFS_LOG_PERM_RESERV);
320 internal_ticket = *ticket;
323 * this is a new transaction on the ticket, so we need to
324 * change the transaction ID so that the next transaction has a
325 * different TID in the log. Just add one to the existing tid
326 * so that we can see chains of rolling transactions in the log
329 internal_ticket->t_tid++;
331 trace_xfs_log_reserve(log, internal_ticket);
333 xlog_grant_push_ail(log, internal_ticket->t_unit_res);
334 retval = xlog_regrant_write_log_space(log, internal_ticket);
336 /* may sleep if need to allocate more tickets */
337 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
339 KM_SLEEP|KM_MAYFAIL);
340 if (!internal_ticket)
341 return XFS_ERROR(ENOMEM);
342 internal_ticket->t_trans_type = t_type;
343 *ticket = internal_ticket;
345 trace_xfs_log_reserve(log, internal_ticket);
347 xlog_grant_push_ail(log,
348 (internal_ticket->t_unit_res *
349 internal_ticket->t_cnt));
350 retval = xlog_grant_log_space(log, internal_ticket);
354 } /* xfs_log_reserve */
358 * Mount a log filesystem
360 * mp - ubiquitous xfs mount point structure
361 * log_target - buftarg of on-disk log device
362 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
363 * num_bblocks - Number of BBSIZE blocks in on-disk log
365 * Return error or zero.
370 xfs_buftarg_t *log_target,
371 xfs_daddr_t blk_offset,
376 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
377 xfs_notice(mp, "Mounting Filesystem");
380 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
381 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
384 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
385 if (IS_ERR(mp->m_log)) {
386 error = -PTR_ERR(mp->m_log);
391 * Initialize the AIL now we have a log.
393 error = xfs_trans_ail_init(mp);
395 xfs_warn(mp, "AIL initialisation failed: error %d", error);
398 mp->m_log->l_ailp = mp->m_ail;
401 * skip log recovery on a norecovery mount. pretend it all
404 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
405 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
408 mp->m_flags &= ~XFS_MOUNT_RDONLY;
410 error = xlog_recover(mp->m_log);
413 mp->m_flags |= XFS_MOUNT_RDONLY;
415 xfs_warn(mp, "log mount/recovery failed: error %d",
417 goto out_destroy_ail;
421 /* Normal transactions can now occur */
422 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
425 * Now the log has been fully initialised and we know were our
426 * space grant counters are, we can initialise the permanent ticket
427 * needed for delayed logging to work.
429 xlog_cil_init_post_recovery(mp->m_log);
434 xfs_trans_ail_destroy(mp);
436 xlog_dealloc_log(mp->m_log);
442 * Finish the recovery of the file system. This is separate from
443 * the xfs_log_mount() call, because it depends on the code in
444 * xfs_mountfs() to read in the root and real-time bitmap inodes
445 * between calling xfs_log_mount() and here.
447 * mp - ubiquitous xfs mount point structure
450 xfs_log_mount_finish(xfs_mount_t *mp)
454 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
455 error = xlog_recover_finish(mp->m_log);
458 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
465 * Final log writes as part of unmount.
467 * Mark the filesystem clean as unmount happens. Note that during relocation
468 * this routine needs to be executed as part of source-bag while the
469 * deallocation must not be done until source-end.
473 * Unmount record used to have a string "Unmount filesystem--" in the
474 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
475 * We just write the magic number now since that particular field isn't
476 * currently architecture converted and "nUmount" is a bit foo.
477 * As far as I know, there weren't any dependencies on the old behaviour.
481 xfs_log_unmount_write(xfs_mount_t *mp)
483 xlog_t *log = mp->m_log;
484 xlog_in_core_t *iclog;
486 xlog_in_core_t *first_iclog;
488 xlog_ticket_t *tic = NULL;
493 * Don't write out unmount record on read-only mounts.
494 * Or, if we are doing a forced umount (typically because of IO errors).
496 if (mp->m_flags & XFS_MOUNT_RDONLY)
499 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
500 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
503 first_iclog = iclog = log->l_iclog;
505 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
506 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
507 ASSERT(iclog->ic_offset == 0);
509 iclog = iclog->ic_next;
510 } while (iclog != first_iclog);
512 if (! (XLOG_FORCED_SHUTDOWN(log))) {
513 error = xfs_log_reserve(mp, 600, 1, &tic,
514 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
516 /* the data section must be 32 bit size aligned */
520 __uint32_t pad2; /* may as well make it 64 bits */
522 .magic = XLOG_UNMOUNT_TYPE,
524 struct xfs_log_iovec reg = {
526 .i_len = sizeof(magic),
527 .i_type = XLOG_REG_TYPE_UNMOUNT,
529 struct xfs_log_vec vec = {
534 /* remove inited flag */
536 error = xlog_write(log, &vec, tic, &lsn,
537 NULL, XLOG_UNMOUNT_TRANS);
539 * At this point, we're umounting anyway,
540 * so there's no point in transitioning log state
541 * to IOERROR. Just continue...
546 xfs_alert(mp, "%s: unmount record failed", __func__);
549 spin_lock(&log->l_icloglock);
550 iclog = log->l_iclog;
551 atomic_inc(&iclog->ic_refcnt);
552 xlog_state_want_sync(log, iclog);
553 spin_unlock(&log->l_icloglock);
554 error = xlog_state_release_iclog(log, iclog);
556 spin_lock(&log->l_icloglock);
557 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
558 iclog->ic_state == XLOG_STATE_DIRTY)) {
559 if (!XLOG_FORCED_SHUTDOWN(log)) {
560 xlog_wait(&iclog->ic_force_wait,
563 spin_unlock(&log->l_icloglock);
566 spin_unlock(&log->l_icloglock);
569 trace_xfs_log_umount_write(log, tic);
570 xlog_ungrant_log_space(log, tic);
571 xfs_log_ticket_put(tic);
575 * We're already in forced_shutdown mode, couldn't
576 * even attempt to write out the unmount transaction.
578 * Go through the motions of sync'ing and releasing
579 * the iclog, even though no I/O will actually happen,
580 * we need to wait for other log I/Os that may already
581 * be in progress. Do this as a separate section of
582 * code so we'll know if we ever get stuck here that
583 * we're in this odd situation of trying to unmount
584 * a file system that went into forced_shutdown as
585 * the result of an unmount..
587 spin_lock(&log->l_icloglock);
588 iclog = log->l_iclog;
589 atomic_inc(&iclog->ic_refcnt);
591 xlog_state_want_sync(log, iclog);
592 spin_unlock(&log->l_icloglock);
593 error = xlog_state_release_iclog(log, iclog);
595 spin_lock(&log->l_icloglock);
597 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
598 || iclog->ic_state == XLOG_STATE_DIRTY
599 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
601 xlog_wait(&iclog->ic_force_wait,
604 spin_unlock(&log->l_icloglock);
609 } /* xfs_log_unmount_write */
612 * Deallocate log structures for unmount/relocation.
614 * We need to stop the aild from running before we destroy
615 * and deallocate the log as the aild references the log.
618 xfs_log_unmount(xfs_mount_t *mp)
620 xfs_trans_ail_destroy(mp);
621 xlog_dealloc_log(mp->m_log);
626 struct xfs_mount *mp,
627 struct xfs_log_item *item,
629 struct xfs_item_ops *ops)
631 item->li_mountp = mp;
632 item->li_ailp = mp->m_ail;
633 item->li_type = type;
637 INIT_LIST_HEAD(&item->li_ail);
638 INIT_LIST_HEAD(&item->li_cil);
642 * Write region vectors to log. The write happens using the space reservation
643 * of the ticket (tic). It is not a requirement that all writes for a given
644 * transaction occur with one call to xfs_log_write(). However, it is important
645 * to note that the transaction reservation code makes an assumption about the
646 * number of log headers a transaction requires that may be violated if you
647 * don't pass all the transaction vectors in one call....
651 struct xfs_mount *mp,
652 struct xfs_log_iovec reg[],
654 struct xlog_ticket *tic,
655 xfs_lsn_t *start_lsn)
657 struct log *log = mp->m_log;
659 struct xfs_log_vec vec = {
660 .lv_niovecs = nentries,
664 if (XLOG_FORCED_SHUTDOWN(log))
665 return XFS_ERROR(EIO);
667 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
669 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
674 xfs_log_move_tail(xfs_mount_t *mp,
678 xlog_t *log = mp->m_log;
679 int need_bytes, free_bytes;
681 if (XLOG_FORCED_SHUTDOWN(log))
685 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
687 /* tail_lsn == 1 implies that we weren't passed a valid value. */
689 atomic64_set(&log->l_tail_lsn, tail_lsn);
691 if (!list_empty_careful(&log->l_writeq)) {
693 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
694 panic("Recovery problem");
696 spin_lock(&log->l_grant_write_lock);
697 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
698 list_for_each_entry(tic, &log->l_writeq, t_queue) {
699 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
701 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
704 free_bytes -= tic->t_unit_res;
705 trace_xfs_log_regrant_write_wake_up(log, tic);
706 wake_up(&tic->t_wait);
708 spin_unlock(&log->l_grant_write_lock);
711 if (!list_empty_careful(&log->l_reserveq)) {
713 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
714 panic("Recovery problem");
716 spin_lock(&log->l_grant_reserve_lock);
717 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
718 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
719 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
720 need_bytes = tic->t_unit_res*tic->t_cnt;
722 need_bytes = tic->t_unit_res;
723 if (free_bytes < need_bytes && tail_lsn != 1)
726 free_bytes -= need_bytes;
727 trace_xfs_log_grant_wake_up(log, tic);
728 wake_up(&tic->t_wait);
730 spin_unlock(&log->l_grant_reserve_lock);
735 * Determine if we have a transaction that has gone to disk
736 * that needs to be covered. To begin the transition to the idle state
737 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
738 * If we are then in a state where covering is needed, the caller is informed
739 * that dummy transactions are required to move the log into the idle state.
741 * Because this is called as part of the sync process, we should also indicate
742 * that dummy transactions should be issued in anything but the covered or
743 * idle states. This ensures that the log tail is accurately reflected in
744 * the log at the end of the sync, hence if a crash occurrs avoids replay
745 * of transactions where the metadata is already on disk.
748 xfs_log_need_covered(xfs_mount_t *mp)
751 xlog_t *log = mp->m_log;
753 if (!xfs_fs_writable(mp))
756 spin_lock(&log->l_icloglock);
757 switch (log->l_covered_state) {
758 case XLOG_STATE_COVER_DONE:
759 case XLOG_STATE_COVER_DONE2:
760 case XLOG_STATE_COVER_IDLE:
762 case XLOG_STATE_COVER_NEED:
763 case XLOG_STATE_COVER_NEED2:
764 if (!xfs_ail_min_lsn(log->l_ailp) &&
765 xlog_iclogs_empty(log)) {
766 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
767 log->l_covered_state = XLOG_STATE_COVER_DONE;
769 log->l_covered_state = XLOG_STATE_COVER_DONE2;
776 spin_unlock(&log->l_icloglock);
780 /******************************************************************************
784 ******************************************************************************
787 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
788 * The log manager must keep track of the last LR which was committed
789 * to disk. The lsn of this LR will become the new tail_lsn whenever
790 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
791 * the situation where stuff could be written into the log but nothing
792 * was ever in the AIL when asked. Eventually, we panic since the
793 * tail hits the head.
795 * We may be holding the log iclog lock upon entering this routine.
798 xlog_assign_tail_lsn(
799 struct xfs_mount *mp)
802 struct log *log = mp->m_log;
804 tail_lsn = xfs_ail_min_lsn(mp->m_ail);
806 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
808 atomic64_set(&log->l_tail_lsn, tail_lsn);
813 * Return the space in the log between the tail and the head. The head
814 * is passed in the cycle/bytes formal parms. In the special case where
815 * the reserve head has wrapped passed the tail, this calculation is no
816 * longer valid. In this case, just return 0 which means there is no space
817 * in the log. This works for all places where this function is called
818 * with the reserve head. Of course, if the write head were to ever
819 * wrap the tail, we should blow up. Rather than catch this case here,
820 * we depend on other ASSERTions in other parts of the code. XXXmiken
822 * This code also handles the case where the reservation head is behind
823 * the tail. The details of this case are described below, but the end
824 * result is that we return the size of the log as the amount of space left.
837 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
838 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
839 tail_bytes = BBTOB(tail_bytes);
840 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
841 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
842 else if (tail_cycle + 1 < head_cycle)
844 else if (tail_cycle < head_cycle) {
845 ASSERT(tail_cycle == (head_cycle - 1));
846 free_bytes = tail_bytes - head_bytes;
849 * The reservation head is behind the tail.
850 * In this case we just want to return the size of the
851 * log as the amount of space left.
854 "xlog_space_left: head behind tail\n"
855 " tail_cycle = %d, tail_bytes = %d\n"
856 " GH cycle = %d, GH bytes = %d",
857 tail_cycle, tail_bytes, head_cycle, head_bytes);
859 free_bytes = log->l_logsize;
866 * Log function which is called when an io completes.
868 * The log manager needs its own routine, in order to control what
869 * happens with the buffer after the write completes.
872 xlog_iodone(xfs_buf_t *bp)
874 xlog_in_core_t *iclog;
878 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
879 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
880 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
885 * Race to shutdown the filesystem if we see an error.
887 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
888 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
889 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
891 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
893 * This flag will be propagated to the trans-committed
894 * callback routines to let them know that the log-commit
897 aborted = XFS_LI_ABORTED;
898 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
899 aborted = XFS_LI_ABORTED;
902 /* log I/O is always issued ASYNC */
903 ASSERT(XFS_BUF_ISASYNC(bp));
904 xlog_state_done_syncing(iclog, aborted);
906 * do not reference the buffer (bp) here as we could race
907 * with it being freed after writing the unmount record to the
914 * Return size of each in-core log record buffer.
916 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
918 * If the filesystem blocksize is too large, we may need to choose a
919 * larger size since the directory code currently logs entire blocks.
923 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
929 if (mp->m_logbufs <= 0)
930 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
932 log->l_iclog_bufs = mp->m_logbufs;
935 * Buffer size passed in from mount system call.
937 if (mp->m_logbsize > 0) {
938 size = log->l_iclog_size = mp->m_logbsize;
939 log->l_iclog_size_log = 0;
941 log->l_iclog_size_log++;
945 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
946 /* # headers = size / 32k
947 * one header holds cycles from 32k of data
950 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
951 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
953 log->l_iclog_hsize = xhdrs << BBSHIFT;
954 log->l_iclog_heads = xhdrs;
956 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
957 log->l_iclog_hsize = BBSIZE;
958 log->l_iclog_heads = 1;
963 /* All machines use 32kB buffers by default. */
964 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
965 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
967 /* the default log size is 16k or 32k which is one header sector */
968 log->l_iclog_hsize = BBSIZE;
969 log->l_iclog_heads = 1;
972 /* are we being asked to make the sizes selected above visible? */
973 if (mp->m_logbufs == 0)
974 mp->m_logbufs = log->l_iclog_bufs;
975 if (mp->m_logbsize == 0)
976 mp->m_logbsize = log->l_iclog_size;
977 } /* xlog_get_iclog_buffer_size */
981 * This routine initializes some of the log structure for a given mount point.
982 * Its primary purpose is to fill in enough, so recovery can occur. However,
983 * some other stuff may be filled in too.
986 xlog_alloc_log(xfs_mount_t *mp,
987 xfs_buftarg_t *log_target,
988 xfs_daddr_t blk_offset,
992 xlog_rec_header_t *head;
993 xlog_in_core_t **iclogp;
994 xlog_in_core_t *iclog, *prev_iclog=NULL;
1000 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1002 xfs_warn(mp, "Log allocation failed: No memory!");
1007 log->l_targ = log_target;
1008 log->l_logsize = BBTOB(num_bblks);
1009 log->l_logBBstart = blk_offset;
1010 log->l_logBBsize = num_bblks;
1011 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1012 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1014 log->l_prev_block = -1;
1015 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1016 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1017 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1018 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1019 xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0);
1020 xlog_assign_grant_head(&log->l_grant_write_head, 1, 0);
1021 INIT_LIST_HEAD(&log->l_reserveq);
1022 INIT_LIST_HEAD(&log->l_writeq);
1023 spin_lock_init(&log->l_grant_reserve_lock);
1024 spin_lock_init(&log->l_grant_write_lock);
1026 error = EFSCORRUPTED;
1027 if (xfs_sb_version_hassector(&mp->m_sb)) {
1028 log2_size = mp->m_sb.sb_logsectlog;
1029 if (log2_size < BBSHIFT) {
1030 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1031 log2_size, BBSHIFT);
1035 log2_size -= BBSHIFT;
1036 if (log2_size > mp->m_sectbb_log) {
1037 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1038 log2_size, mp->m_sectbb_log);
1042 /* for larger sector sizes, must have v2 or external log */
1043 if (log2_size && log->l_logBBstart > 0 &&
1044 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1046 "log sector size (0x%x) invalid for configuration.",
1051 log->l_sectBBsize = 1 << log2_size;
1053 xlog_get_iclog_buffer_size(mp, log);
1056 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1059 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1060 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1061 ASSERT(XFS_BUF_ISBUSY(bp));
1062 ASSERT(xfs_buf_islocked(bp));
1065 spin_lock_init(&log->l_icloglock);
1066 init_waitqueue_head(&log->l_flush_wait);
1068 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1069 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1071 iclogp = &log->l_iclog;
1073 * The amount of memory to allocate for the iclog structure is
1074 * rather funky due to the way the structure is defined. It is
1075 * done this way so that we can use different sizes for machines
1076 * with different amounts of memory. See the definition of
1077 * xlog_in_core_t in xfs_log_priv.h for details.
1079 ASSERT(log->l_iclog_size >= 4096);
1080 for (i=0; i < log->l_iclog_bufs; i++) {
1081 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1083 goto out_free_iclog;
1086 iclog->ic_prev = prev_iclog;
1089 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1090 log->l_iclog_size, 0);
1092 goto out_free_iclog;
1094 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1095 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1097 iclog->ic_data = bp->b_addr;
1099 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1101 head = &iclog->ic_header;
1102 memset(head, 0, sizeof(xlog_rec_header_t));
1103 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1104 head->h_version = cpu_to_be32(
1105 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1106 head->h_size = cpu_to_be32(log->l_iclog_size);
1108 head->h_fmt = cpu_to_be32(XLOG_FMT);
1109 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1111 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1112 iclog->ic_state = XLOG_STATE_ACTIVE;
1113 iclog->ic_log = log;
1114 atomic_set(&iclog->ic_refcnt, 0);
1115 spin_lock_init(&iclog->ic_callback_lock);
1116 iclog->ic_callback_tail = &(iclog->ic_callback);
1117 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1119 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1120 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1121 init_waitqueue_head(&iclog->ic_force_wait);
1122 init_waitqueue_head(&iclog->ic_write_wait);
1124 iclogp = &iclog->ic_next;
1126 *iclogp = log->l_iclog; /* complete ring */
1127 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1129 error = xlog_cil_init(log);
1131 goto out_free_iclog;
1135 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1136 prev_iclog = iclog->ic_next;
1138 xfs_buf_free(iclog->ic_bp);
1141 spinlock_destroy(&log->l_icloglock);
1142 xfs_buf_free(log->l_xbuf);
1146 return ERR_PTR(-error);
1147 } /* xlog_alloc_log */
1151 * Write out the commit record of a transaction associated with the given
1152 * ticket. Return the lsn of the commit record.
1157 struct xlog_ticket *ticket,
1158 struct xlog_in_core **iclog,
1159 xfs_lsn_t *commitlsnp)
1161 struct xfs_mount *mp = log->l_mp;
1163 struct xfs_log_iovec reg = {
1166 .i_type = XLOG_REG_TYPE_COMMIT,
1168 struct xfs_log_vec vec = {
1173 ASSERT_ALWAYS(iclog);
1174 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1177 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1182 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1183 * log space. This code pushes on the lsn which would supposedly free up
1184 * the 25% which we want to leave free. We may need to adopt a policy which
1185 * pushes on an lsn which is further along in the log once we reach the high
1186 * water mark. In this manner, we would be creating a low water mark.
1189 xlog_grant_push_ail(
1193 xfs_lsn_t threshold_lsn = 0;
1194 xfs_lsn_t last_sync_lsn;
1197 int threshold_block;
1198 int threshold_cycle;
1201 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1203 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
1204 free_blocks = BTOBBT(free_bytes);
1207 * Set the threshold for the minimum number of free blocks in the
1208 * log to the maximum of what the caller needs, one quarter of the
1209 * log, and 256 blocks.
1211 free_threshold = BTOBB(need_bytes);
1212 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1213 free_threshold = MAX(free_threshold, 256);
1214 if (free_blocks >= free_threshold)
1217 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1219 threshold_block += free_threshold;
1220 if (threshold_block >= log->l_logBBsize) {
1221 threshold_block -= log->l_logBBsize;
1222 threshold_cycle += 1;
1224 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1227 * Don't pass in an lsn greater than the lsn of the last
1228 * log record known to be on disk. Use a snapshot of the last sync lsn
1229 * so that it doesn't change between the compare and the set.
1231 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1232 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1233 threshold_lsn = last_sync_lsn;
1236 * Get the transaction layer to kick the dirty buffers out to
1237 * disk asynchronously. No point in trying to do this if
1238 * the filesystem is shutting down.
1240 if (!XLOG_FORCED_SHUTDOWN(log))
1241 xfs_ail_push(log->l_ailp, threshold_lsn);
1245 * The bdstrat callback function for log bufs. This gives us a central
1246 * place to trap bufs in case we get hit by a log I/O error and need to
1247 * shutdown. Actually, in practice, even when we didn't get a log error,
1248 * we transition the iclogs to IOERROR state *after* flushing all existing
1249 * iclogs to disk. This is because we don't want anymore new transactions to be
1250 * started or completed afterwards.
1256 struct xlog_in_core *iclog;
1258 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1259 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1260 XFS_BUF_ERROR(bp, EIO);
1262 xfs_buf_ioend(bp, 0);
1264 * It would seem logical to return EIO here, but we rely on
1265 * the log state machine to propagate I/O errors instead of
1271 bp->b_flags |= _XBF_RUN_QUEUES;
1272 xfs_buf_iorequest(bp);
1277 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1278 * fashion. Previously, we should have moved the current iclog
1279 * ptr in the log to point to the next available iclog. This allows further
1280 * write to continue while this code syncs out an iclog ready to go.
1281 * Before an in-core log can be written out, the data section must be scanned
1282 * to save away the 1st word of each BBSIZE block into the header. We replace
1283 * it with the current cycle count. Each BBSIZE block is tagged with the
1284 * cycle count because there in an implicit assumption that drives will
1285 * guarantee that entire 512 byte blocks get written at once. In other words,
1286 * we can't have part of a 512 byte block written and part not written. By
1287 * tagging each block, we will know which blocks are valid when recovering
1288 * after an unclean shutdown.
1290 * This routine is single threaded on the iclog. No other thread can be in
1291 * this routine with the same iclog. Changing contents of iclog can there-
1292 * fore be done without grabbing the state machine lock. Updating the global
1293 * log will require grabbing the lock though.
1295 * The entire log manager uses a logical block numbering scheme. Only
1296 * log_sync (and then only bwrite()) know about the fact that the log may
1297 * not start with block zero on a given device. The log block start offset
1298 * is added immediately before calling bwrite().
1302 xlog_sync(xlog_t *log,
1303 xlog_in_core_t *iclog)
1305 xfs_caddr_t dptr; /* pointer to byte sized element */
1308 uint count; /* byte count of bwrite */
1309 uint count_init; /* initial count before roundup */
1310 int roundoff; /* roundoff to BB or stripe */
1311 int split = 0; /* split write into two regions */
1313 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1315 XFS_STATS_INC(xs_log_writes);
1316 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1318 /* Add for LR header */
1319 count_init = log->l_iclog_hsize + iclog->ic_offset;
1321 /* Round out the log write size */
1322 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1323 /* we have a v2 stripe unit to use */
1324 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1326 count = BBTOB(BTOBB(count_init));
1328 roundoff = count - count_init;
1329 ASSERT(roundoff >= 0);
1330 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1331 roundoff < log->l_mp->m_sb.sb_logsunit)
1333 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1334 roundoff < BBTOB(1)));
1336 /* move grant heads by roundoff in sync */
1337 xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff);
1338 xlog_grant_add_space(log, &log->l_grant_write_head, roundoff);
1340 /* put cycle number in every block */
1341 xlog_pack_data(log, iclog, roundoff);
1343 /* real byte length */
1345 iclog->ic_header.h_len =
1346 cpu_to_be32(iclog->ic_offset + roundoff);
1348 iclog->ic_header.h_len =
1349 cpu_to_be32(iclog->ic_offset);
1353 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1354 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1355 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1357 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1359 /* Do we need to split this write into 2 parts? */
1360 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1361 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1362 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1363 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1365 iclog->ic_bwritecnt = 1;
1367 XFS_BUF_SET_COUNT(bp, count);
1368 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1369 XFS_BUF_ZEROFLAGS(bp);
1372 bp->b_flags |= XBF_LOG_BUFFER;
1374 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1376 * If we have an external log device, flush the data device
1377 * before flushing the log to make sure all meta data
1378 * written back from the AIL actually made it to disk
1379 * before writing out the new log tail LSN in the log buffer.
1381 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1382 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1383 XFS_BUF_ORDERED(bp);
1386 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1387 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1389 xlog_verify_iclog(log, iclog, count, B_TRUE);
1391 /* account for log which doesn't start at block #0 */
1392 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1394 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1399 if ((error = xlog_bdstrat(bp))) {
1400 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1405 bp = iclog->ic_log->l_xbuf;
1406 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1408 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1409 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1410 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1411 (__psint_t)count), split);
1412 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1413 XFS_BUF_ZEROFLAGS(bp);
1416 bp->b_flags |= XBF_LOG_BUFFER;
1417 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1418 XFS_BUF_ORDERED(bp);
1419 dptr = XFS_BUF_PTR(bp);
1421 * Bump the cycle numbers at the start of each block
1422 * since this part of the buffer is at the start of
1423 * a new cycle. Watch out for the header magic number
1426 for (i = 0; i < split; i += BBSIZE) {
1427 be32_add_cpu((__be32 *)dptr, 1);
1428 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1429 be32_add_cpu((__be32 *)dptr, 1);
1433 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1434 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1436 /* account for internal log which doesn't start at block #0 */
1437 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1439 if ((error = xlog_bdstrat(bp))) {
1440 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1441 bp, XFS_BUF_ADDR(bp));
1450 * Deallocate a log structure
1453 xlog_dealloc_log(xlog_t *log)
1455 xlog_in_core_t *iclog, *next_iclog;
1458 xlog_cil_destroy(log);
1461 * always need to ensure that the extra buffer does not point to memory
1462 * owned by another log buffer before we free it.
1464 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1465 xfs_buf_free(log->l_xbuf);
1467 iclog = log->l_iclog;
1468 for (i=0; i<log->l_iclog_bufs; i++) {
1469 xfs_buf_free(iclog->ic_bp);
1470 next_iclog = iclog->ic_next;
1474 spinlock_destroy(&log->l_icloglock);
1476 log->l_mp->m_log = NULL;
1478 } /* xlog_dealloc_log */
1481 * Update counters atomically now that memcpy is done.
1485 xlog_state_finish_copy(xlog_t *log,
1486 xlog_in_core_t *iclog,
1490 spin_lock(&log->l_icloglock);
1492 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1493 iclog->ic_offset += copy_bytes;
1495 spin_unlock(&log->l_icloglock);
1496 } /* xlog_state_finish_copy */
1502 * print out info relating to regions written which consume
1507 struct xfs_mount *mp,
1508 struct xlog_ticket *ticket)
1511 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1513 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1514 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1535 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1579 "xfs_log_write: reservation summary:\n"
1580 " trans type = %s (%u)\n"
1581 " unit res = %d bytes\n"
1582 " current res = %d bytes\n"
1583 " total reg = %u bytes (o/flow = %u bytes)\n"
1584 " ophdrs = %u (ophdr space = %u bytes)\n"
1585 " ophdr + reg = %u bytes\n"
1586 " num regions = %u\n",
1587 ((ticket->t_trans_type <= 0 ||
1588 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1589 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1590 ticket->t_trans_type,
1593 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1594 ticket->t_res_num_ophdrs, ophdr_spc,
1595 ticket->t_res_arr_sum +
1596 ticket->t_res_o_flow + ophdr_spc,
1599 for (i = 0; i < ticket->t_res_num; i++) {
1600 uint r_type = ticket->t_res_arr[i].r_type;
1601 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1602 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1603 "bad-rtype" : res_type_str[r_type-1]),
1604 ticket->t_res_arr[i].r_len);
1607 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1608 "xfs_log_write: reservation ran out. Need to up reservation");
1609 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1613 * Calculate the potential space needed by the log vector. Each region gets
1614 * its own xlog_op_header_t and may need to be double word aligned.
1617 xlog_write_calc_vec_length(
1618 struct xlog_ticket *ticket,
1619 struct xfs_log_vec *log_vector)
1621 struct xfs_log_vec *lv;
1626 /* acct for start rec of xact */
1627 if (ticket->t_flags & XLOG_TIC_INITED)
1630 for (lv = log_vector; lv; lv = lv->lv_next) {
1631 headers += lv->lv_niovecs;
1633 for (i = 0; i < lv->lv_niovecs; i++) {
1634 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1637 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1641 ticket->t_res_num_ophdrs += headers;
1642 len += headers * sizeof(struct xlog_op_header);
1648 * If first write for transaction, insert start record We can't be trying to
1649 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1652 xlog_write_start_rec(
1653 struct xlog_op_header *ophdr,
1654 struct xlog_ticket *ticket)
1656 if (!(ticket->t_flags & XLOG_TIC_INITED))
1659 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1660 ophdr->oh_clientid = ticket->t_clientid;
1662 ophdr->oh_flags = XLOG_START_TRANS;
1665 ticket->t_flags &= ~XLOG_TIC_INITED;
1667 return sizeof(struct xlog_op_header);
1670 static xlog_op_header_t *
1671 xlog_write_setup_ophdr(
1673 struct xlog_op_header *ophdr,
1674 struct xlog_ticket *ticket,
1677 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1678 ophdr->oh_clientid = ticket->t_clientid;
1681 /* are we copying a commit or unmount record? */
1682 ophdr->oh_flags = flags;
1685 * We've seen logs corrupted with bad transaction client ids. This
1686 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1687 * and shut down the filesystem.
1689 switch (ophdr->oh_clientid) {
1690 case XFS_TRANSACTION:
1696 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1697 ophdr->oh_clientid, ticket);
1705 * Set up the parameters of the region copy into the log. This has
1706 * to handle region write split across multiple log buffers - this
1707 * state is kept external to this function so that this code can
1708 * can be written in an obvious, self documenting manner.
1711 xlog_write_setup_copy(
1712 struct xlog_ticket *ticket,
1713 struct xlog_op_header *ophdr,
1714 int space_available,
1718 int *last_was_partial_copy,
1719 int *bytes_consumed)
1723 still_to_copy = space_required - *bytes_consumed;
1724 *copy_off = *bytes_consumed;
1726 if (still_to_copy <= space_available) {
1727 /* write of region completes here */
1728 *copy_len = still_to_copy;
1729 ophdr->oh_len = cpu_to_be32(*copy_len);
1730 if (*last_was_partial_copy)
1731 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1732 *last_was_partial_copy = 0;
1733 *bytes_consumed = 0;
1737 /* partial write of region, needs extra log op header reservation */
1738 *copy_len = space_available;
1739 ophdr->oh_len = cpu_to_be32(*copy_len);
1740 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1741 if (*last_was_partial_copy)
1742 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1743 *bytes_consumed += *copy_len;
1744 (*last_was_partial_copy)++;
1746 /* account for new log op header */
1747 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1748 ticket->t_res_num_ophdrs++;
1750 return sizeof(struct xlog_op_header);
1754 xlog_write_copy_finish(
1756 struct xlog_in_core *iclog,
1761 int *partial_copy_len,
1763 struct xlog_in_core **commit_iclog)
1765 if (*partial_copy) {
1767 * This iclog has already been marked WANT_SYNC by
1768 * xlog_state_get_iclog_space.
1770 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1773 return xlog_state_release_iclog(log, iclog);
1777 *partial_copy_len = 0;
1779 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1780 /* no more space in this iclog - push it. */
1781 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1785 spin_lock(&log->l_icloglock);
1786 xlog_state_want_sync(log, iclog);
1787 spin_unlock(&log->l_icloglock);
1790 return xlog_state_release_iclog(log, iclog);
1791 ASSERT(flags & XLOG_COMMIT_TRANS);
1792 *commit_iclog = iclog;
1799 * Write some region out to in-core log
1801 * This will be called when writing externally provided regions or when
1802 * writing out a commit record for a given transaction.
1804 * General algorithm:
1805 * 1. Find total length of this write. This may include adding to the
1806 * lengths passed in.
1807 * 2. Check whether we violate the tickets reservation.
1808 * 3. While writing to this iclog
1809 * A. Reserve as much space in this iclog as can get
1810 * B. If this is first write, save away start lsn
1811 * C. While writing this region:
1812 * 1. If first write of transaction, write start record
1813 * 2. Write log operation header (header per region)
1814 * 3. Find out if we can fit entire region into this iclog
1815 * 4. Potentially, verify destination memcpy ptr
1816 * 5. Memcpy (partial) region
1817 * 6. If partial copy, release iclog; otherwise, continue
1818 * copying more regions into current iclog
1819 * 4. Mark want sync bit (in simulation mode)
1820 * 5. Release iclog for potential flush to on-disk log.
1823 * 1. Panic if reservation is overrun. This should never happen since
1824 * reservation amounts are generated internal to the filesystem.
1826 * 1. Tickets are single threaded data structures.
1827 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1828 * syncing routine. When a single log_write region needs to span
1829 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1830 * on all log operation writes which don't contain the end of the
1831 * region. The XLOG_END_TRANS bit is used for the in-core log
1832 * operation which contains the end of the continued log_write region.
1833 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1834 * we don't really know exactly how much space will be used. As a result,
1835 * we don't update ic_offset until the end when we know exactly how many
1836 * bytes have been written out.
1841 struct xfs_log_vec *log_vector,
1842 struct xlog_ticket *ticket,
1843 xfs_lsn_t *start_lsn,
1844 struct xlog_in_core **commit_iclog,
1847 struct xlog_in_core *iclog = NULL;
1848 struct xfs_log_iovec *vecp;
1849 struct xfs_log_vec *lv;
1852 int partial_copy = 0;
1853 int partial_copy_len = 0;
1861 len = xlog_write_calc_vec_length(ticket, log_vector);
1864 * Region headers and bytes are already accounted for.
1865 * We only need to take into account start records and
1866 * split regions in this function.
1868 if (ticket->t_flags & XLOG_TIC_INITED)
1869 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1872 * Commit record headers need to be accounted for. These
1873 * come in as separate writes so are easy to detect.
1875 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1876 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1878 ticket->t_curr_res -= len;
1880 if (ticket->t_curr_res < 0)
1881 xlog_print_tic_res(log->l_mp, ticket);
1885 vecp = lv->lv_iovecp;
1886 while (lv && index < lv->lv_niovecs) {
1890 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1891 &contwr, &log_offset);
1895 ASSERT(log_offset <= iclog->ic_size - 1);
1896 ptr = iclog->ic_datap + log_offset;
1898 /* start_lsn is the first lsn written to. That's all we need. */
1900 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1903 * This loop writes out as many regions as can fit in the amount
1904 * of space which was allocated by xlog_state_get_iclog_space().
1906 while (lv && index < lv->lv_niovecs) {
1907 struct xfs_log_iovec *reg = &vecp[index];
1908 struct xlog_op_header *ophdr;
1913 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1914 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1916 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1917 if (start_rec_copy) {
1919 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1923 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1925 return XFS_ERROR(EIO);
1927 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1928 sizeof(struct xlog_op_header));
1930 len += xlog_write_setup_copy(ticket, ophdr,
1931 iclog->ic_size-log_offset,
1933 ©_off, ©_len,
1936 xlog_verify_dest_ptr(log, ptr);
1939 ASSERT(copy_len >= 0);
1940 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1941 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1943 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1945 data_cnt += contwr ? copy_len : 0;
1947 error = xlog_write_copy_finish(log, iclog, flags,
1948 &record_cnt, &data_cnt,
1957 * if we had a partial copy, we need to get more iclog
1958 * space but we don't want to increment the region
1959 * index because there is still more is this region to
1962 * If we completed writing this region, and we flushed
1963 * the iclog (indicated by resetting of the record
1964 * count), then we also need to get more log space. If
1965 * this was the last record, though, we are done and
1971 if (++index == lv->lv_niovecs) {
1975 vecp = lv->lv_iovecp;
1977 if (record_cnt == 0) {
1987 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1989 return xlog_state_release_iclog(log, iclog);
1991 ASSERT(flags & XLOG_COMMIT_TRANS);
1992 *commit_iclog = iclog;
1997 /*****************************************************************************
1999 * State Machine functions
2001 *****************************************************************************
2004 /* Clean iclogs starting from the head. This ordering must be
2005 * maintained, so an iclog doesn't become ACTIVE beyond one that
2006 * is SYNCING. This is also required to maintain the notion that we use
2007 * a ordered wait queue to hold off would be writers to the log when every
2008 * iclog is trying to sync to disk.
2010 * State Change: DIRTY -> ACTIVE
2013 xlog_state_clean_log(xlog_t *log)
2015 xlog_in_core_t *iclog;
2018 iclog = log->l_iclog;
2020 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2021 iclog->ic_state = XLOG_STATE_ACTIVE;
2022 iclog->ic_offset = 0;
2023 ASSERT(iclog->ic_callback == NULL);
2025 * If the number of ops in this iclog indicate it just
2026 * contains the dummy transaction, we can
2027 * change state into IDLE (the second time around).
2028 * Otherwise we should change the state into
2030 * We don't need to cover the dummy.
2033 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2038 * We have two dirty iclogs so start over
2039 * This could also be num of ops indicates
2040 * this is not the dummy going out.
2044 iclog->ic_header.h_num_logops = 0;
2045 memset(iclog->ic_header.h_cycle_data, 0,
2046 sizeof(iclog->ic_header.h_cycle_data));
2047 iclog->ic_header.h_lsn = 0;
2048 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2051 break; /* stop cleaning */
2052 iclog = iclog->ic_next;
2053 } while (iclog != log->l_iclog);
2055 /* log is locked when we are called */
2057 * Change state for the dummy log recording.
2058 * We usually go to NEED. But we go to NEED2 if the changed indicates
2059 * we are done writing the dummy record.
2060 * If we are done with the second dummy recored (DONE2), then
2064 switch (log->l_covered_state) {
2065 case XLOG_STATE_COVER_IDLE:
2066 case XLOG_STATE_COVER_NEED:
2067 case XLOG_STATE_COVER_NEED2:
2068 log->l_covered_state = XLOG_STATE_COVER_NEED;
2071 case XLOG_STATE_COVER_DONE:
2073 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2075 log->l_covered_state = XLOG_STATE_COVER_NEED;
2078 case XLOG_STATE_COVER_DONE2:
2080 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2082 log->l_covered_state = XLOG_STATE_COVER_NEED;
2089 } /* xlog_state_clean_log */
2092 xlog_get_lowest_lsn(
2095 xlog_in_core_t *lsn_log;
2096 xfs_lsn_t lowest_lsn, lsn;
2098 lsn_log = log->l_iclog;
2101 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2102 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2103 if ((lsn && !lowest_lsn) ||
2104 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2108 lsn_log = lsn_log->ic_next;
2109 } while (lsn_log != log->l_iclog);
2115 xlog_state_do_callback(
2118 xlog_in_core_t *ciclog)
2120 xlog_in_core_t *iclog;
2121 xlog_in_core_t *first_iclog; /* used to know when we've
2122 * processed all iclogs once */
2123 xfs_log_callback_t *cb, *cb_next;
2125 xfs_lsn_t lowest_lsn;
2126 int ioerrors; /* counter: iclogs with errors */
2127 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2128 int funcdidcallbacks; /* flag: function did callbacks */
2129 int repeats; /* for issuing console warnings if
2130 * looping too many times */
2133 spin_lock(&log->l_icloglock);
2134 first_iclog = iclog = log->l_iclog;
2136 funcdidcallbacks = 0;
2141 * Scan all iclogs starting with the one pointed to by the
2142 * log. Reset this starting point each time the log is
2143 * unlocked (during callbacks).
2145 * Keep looping through iclogs until one full pass is made
2146 * without running any callbacks.
2148 first_iclog = log->l_iclog;
2149 iclog = log->l_iclog;
2150 loopdidcallbacks = 0;
2155 /* skip all iclogs in the ACTIVE & DIRTY states */
2156 if (iclog->ic_state &
2157 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2158 iclog = iclog->ic_next;
2163 * Between marking a filesystem SHUTDOWN and stopping
2164 * the log, we do flush all iclogs to disk (if there
2165 * wasn't a log I/O error). So, we do want things to
2166 * go smoothly in case of just a SHUTDOWN w/o a
2169 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2171 * Can only perform callbacks in order. Since
2172 * this iclog is not in the DONE_SYNC/
2173 * DO_CALLBACK state, we skip the rest and
2174 * just try to clean up. If we set our iclog
2175 * to DO_CALLBACK, we will not process it when
2176 * we retry since a previous iclog is in the
2177 * CALLBACK and the state cannot change since
2178 * we are holding the l_icloglock.
2180 if (!(iclog->ic_state &
2181 (XLOG_STATE_DONE_SYNC |
2182 XLOG_STATE_DO_CALLBACK))) {
2183 if (ciclog && (ciclog->ic_state ==
2184 XLOG_STATE_DONE_SYNC)) {
2185 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2190 * We now have an iclog that is in either the
2191 * DO_CALLBACK or DONE_SYNC states. The other
2192 * states (WANT_SYNC, SYNCING, or CALLBACK were
2193 * caught by the above if and are going to
2194 * clean (i.e. we aren't doing their callbacks)
2199 * We will do one more check here to see if we
2200 * have chased our tail around.
2203 lowest_lsn = xlog_get_lowest_lsn(log);
2205 XFS_LSN_CMP(lowest_lsn,
2206 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2207 iclog = iclog->ic_next;
2208 continue; /* Leave this iclog for
2212 iclog->ic_state = XLOG_STATE_CALLBACK;
2216 * update the last_sync_lsn before we drop the
2217 * icloglock to ensure we are the only one that
2220 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2221 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2222 atomic64_set(&log->l_last_sync_lsn,
2223 be64_to_cpu(iclog->ic_header.h_lsn));
2228 spin_unlock(&log->l_icloglock);
2231 * Keep processing entries in the callback list until
2232 * we come around and it is empty. We need to
2233 * atomically see that the list is empty and change the
2234 * state to DIRTY so that we don't miss any more
2235 * callbacks being added.
2237 spin_lock(&iclog->ic_callback_lock);
2238 cb = iclog->ic_callback;
2240 iclog->ic_callback_tail = &(iclog->ic_callback);
2241 iclog->ic_callback = NULL;
2242 spin_unlock(&iclog->ic_callback_lock);
2244 /* perform callbacks in the order given */
2245 for (; cb; cb = cb_next) {
2246 cb_next = cb->cb_next;
2247 cb->cb_func(cb->cb_arg, aborted);
2249 spin_lock(&iclog->ic_callback_lock);
2250 cb = iclog->ic_callback;
2256 spin_lock(&log->l_icloglock);
2257 ASSERT(iclog->ic_callback == NULL);
2258 spin_unlock(&iclog->ic_callback_lock);
2259 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2260 iclog->ic_state = XLOG_STATE_DIRTY;
2263 * Transition from DIRTY to ACTIVE if applicable.
2264 * NOP if STATE_IOERROR.
2266 xlog_state_clean_log(log);
2268 /* wake up threads waiting in xfs_log_force() */
2269 wake_up_all(&iclog->ic_force_wait);
2271 iclog = iclog->ic_next;
2272 } while (first_iclog != iclog);
2274 if (repeats > 5000) {
2275 flushcnt += repeats;
2278 "%s: possible infinite loop (%d iterations)",
2279 __func__, flushcnt);
2281 } while (!ioerrors && loopdidcallbacks);
2284 * make one last gasp attempt to see if iclogs are being left in
2288 if (funcdidcallbacks) {
2289 first_iclog = iclog = log->l_iclog;
2291 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2293 * Terminate the loop if iclogs are found in states
2294 * which will cause other threads to clean up iclogs.
2296 * SYNCING - i/o completion will go through logs
2297 * DONE_SYNC - interrupt thread should be waiting for
2299 * IOERROR - give up hope all ye who enter here
2301 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2302 iclog->ic_state == XLOG_STATE_SYNCING ||
2303 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2304 iclog->ic_state == XLOG_STATE_IOERROR )
2306 iclog = iclog->ic_next;
2307 } while (first_iclog != iclog);
2311 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2313 spin_unlock(&log->l_icloglock);
2316 wake_up_all(&log->l_flush_wait);
2321 * Finish transitioning this iclog to the dirty state.
2323 * Make sure that we completely execute this routine only when this is
2324 * the last call to the iclog. There is a good chance that iclog flushes,
2325 * when we reach the end of the physical log, get turned into 2 separate
2326 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2327 * routine. By using the reference count bwritecnt, we guarantee that only
2328 * the second completion goes through.
2330 * Callbacks could take time, so they are done outside the scope of the
2331 * global state machine log lock.
2334 xlog_state_done_syncing(
2335 xlog_in_core_t *iclog,
2338 xlog_t *log = iclog->ic_log;
2340 spin_lock(&log->l_icloglock);
2342 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2343 iclog->ic_state == XLOG_STATE_IOERROR);
2344 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2345 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2349 * If we got an error, either on the first buffer, or in the case of
2350 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2351 * and none should ever be attempted to be written to disk
2354 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2355 if (--iclog->ic_bwritecnt == 1) {
2356 spin_unlock(&log->l_icloglock);
2359 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2363 * Someone could be sleeping prior to writing out the next
2364 * iclog buffer, we wake them all, one will get to do the
2365 * I/O, the others get to wait for the result.
2367 wake_up_all(&iclog->ic_write_wait);
2368 spin_unlock(&log->l_icloglock);
2369 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2370 } /* xlog_state_done_syncing */
2374 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2375 * sleep. We wait on the flush queue on the head iclog as that should be
2376 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2377 * we will wait here and all new writes will sleep until a sync completes.
2379 * The in-core logs are used in a circular fashion. They are not used
2380 * out-of-order even when an iclog past the head is free.
2383 * * log_offset where xlog_write() can start writing into the in-core
2385 * * in-core log pointer to which xlog_write() should write.
2386 * * boolean indicating this is a continued write to an in-core log.
2387 * If this is the last write, then the in-core log's offset field
2388 * needs to be incremented, depending on the amount of data which
2392 xlog_state_get_iclog_space(xlog_t *log,
2394 xlog_in_core_t **iclogp,
2395 xlog_ticket_t *ticket,
2396 int *continued_write,
2400 xlog_rec_header_t *head;
2401 xlog_in_core_t *iclog;
2405 spin_lock(&log->l_icloglock);
2406 if (XLOG_FORCED_SHUTDOWN(log)) {
2407 spin_unlock(&log->l_icloglock);
2408 return XFS_ERROR(EIO);
2411 iclog = log->l_iclog;
2412 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2413 XFS_STATS_INC(xs_log_noiclogs);
2415 /* Wait for log writes to have flushed */
2416 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2420 head = &iclog->ic_header;
2422 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2423 log_offset = iclog->ic_offset;
2425 /* On the 1st write to an iclog, figure out lsn. This works
2426 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2427 * committing to. If the offset is set, that's how many blocks
2430 if (log_offset == 0) {
2431 ticket->t_curr_res -= log->l_iclog_hsize;
2432 xlog_tic_add_region(ticket,
2434 XLOG_REG_TYPE_LRHEADER);
2435 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2436 head->h_lsn = cpu_to_be64(
2437 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2438 ASSERT(log->l_curr_block >= 0);
2441 /* If there is enough room to write everything, then do it. Otherwise,
2442 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2443 * bit is on, so this will get flushed out. Don't update ic_offset
2444 * until you know exactly how many bytes get copied. Therefore, wait
2445 * until later to update ic_offset.
2447 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2448 * can fit into remaining data section.
2450 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2451 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2454 * If I'm the only one writing to this iclog, sync it to disk.
2455 * We need to do an atomic compare and decrement here to avoid
2456 * racing with concurrent atomic_dec_and_lock() calls in
2457 * xlog_state_release_iclog() when there is more than one
2458 * reference to the iclog.
2460 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2461 /* we are the only one */
2462 spin_unlock(&log->l_icloglock);
2463 error = xlog_state_release_iclog(log, iclog);
2467 spin_unlock(&log->l_icloglock);
2472 /* Do we have enough room to write the full amount in the remainder
2473 * of this iclog? Or must we continue a write on the next iclog and
2474 * mark this iclog as completely taken? In the case where we switch
2475 * iclogs (to mark it taken), this particular iclog will release/sync
2476 * to disk in xlog_write().
2478 if (len <= iclog->ic_size - iclog->ic_offset) {
2479 *continued_write = 0;
2480 iclog->ic_offset += len;
2482 *continued_write = 1;
2483 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2487 ASSERT(iclog->ic_offset <= iclog->ic_size);
2488 spin_unlock(&log->l_icloglock);
2490 *logoffsetp = log_offset;
2492 } /* xlog_state_get_iclog_space */
2495 * Atomically get the log space required for a log ticket.
2497 * Once a ticket gets put onto the reserveq, it will only return after
2498 * the needed reservation is satisfied.
2500 * This function is structured so that it has a lock free fast path. This is
2501 * necessary because every new transaction reservation will come through this
2502 * path. Hence any lock will be globally hot if we take it unconditionally on
2505 * As tickets are only ever moved on and off the reserveq under the
2506 * l_grant_reserve_lock, we only need to take that lock if we are going
2507 * to add the ticket to the queue and sleep. We can avoid taking the lock if the
2508 * ticket was never added to the reserveq because the t_queue list head will be
2509 * empty and we hold the only reference to it so it can safely be checked
2513 xlog_grant_log_space(xlog_t *log,
2520 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2521 panic("grant Recovery problem");
2524 trace_xfs_log_grant_enter(log, tic);
2526 need_bytes = tic->t_unit_res;
2527 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2528 need_bytes *= tic->t_ocnt;
2530 /* something is already sleeping; insert new transaction at end */
2531 if (!list_empty_careful(&log->l_reserveq)) {
2532 spin_lock(&log->l_grant_reserve_lock);
2533 /* recheck the queue now we are locked */
2534 if (list_empty(&log->l_reserveq)) {
2535 spin_unlock(&log->l_grant_reserve_lock);
2538 list_add_tail(&tic->t_queue, &log->l_reserveq);
2540 trace_xfs_log_grant_sleep1(log, tic);
2543 * Gotta check this before going to sleep, while we're
2544 * holding the grant lock.
2546 if (XLOG_FORCED_SHUTDOWN(log))
2549 XFS_STATS_INC(xs_sleep_logspace);
2550 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2553 * If we got an error, and the filesystem is shutting down,
2554 * we'll catch it down below. So just continue...
2556 trace_xfs_log_grant_wake1(log, tic);
2560 if (XLOG_FORCED_SHUTDOWN(log))
2561 goto error_return_unlocked;
2563 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
2564 if (free_bytes < need_bytes) {
2565 spin_lock(&log->l_grant_reserve_lock);
2566 if (list_empty(&tic->t_queue))
2567 list_add_tail(&tic->t_queue, &log->l_reserveq);
2569 trace_xfs_log_grant_sleep2(log, tic);
2571 if (XLOG_FORCED_SHUTDOWN(log))
2574 xlog_grant_push_ail(log, need_bytes);
2576 XFS_STATS_INC(xs_sleep_logspace);
2577 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2579 trace_xfs_log_grant_wake2(log, tic);
2583 if (!list_empty(&tic->t_queue)) {
2584 spin_lock(&log->l_grant_reserve_lock);
2585 list_del_init(&tic->t_queue);
2586 spin_unlock(&log->l_grant_reserve_lock);
2589 /* we've got enough space */
2590 xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes);
2591 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2592 trace_xfs_log_grant_exit(log, tic);
2593 xlog_verify_grant_tail(log);
2596 error_return_unlocked:
2597 spin_lock(&log->l_grant_reserve_lock);
2599 list_del_init(&tic->t_queue);
2600 spin_unlock(&log->l_grant_reserve_lock);
2601 trace_xfs_log_grant_error(log, tic);
2604 * If we are failing, make sure the ticket doesn't have any
2605 * current reservations. We don't want to add this back when
2606 * the ticket/transaction gets cancelled.
2608 tic->t_curr_res = 0;
2609 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2610 return XFS_ERROR(EIO);
2611 } /* xlog_grant_log_space */
2615 * Replenish the byte reservation required by moving the grant write head.
2617 * Similar to xlog_grant_log_space, the function is structured to have a lock
2621 xlog_regrant_write_log_space(xlog_t *log,
2624 int free_bytes, need_bytes;
2626 tic->t_curr_res = tic->t_unit_res;
2627 xlog_tic_reset_res(tic);
2633 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2634 panic("regrant Recovery problem");
2637 trace_xfs_log_regrant_write_enter(log, tic);
2638 if (XLOG_FORCED_SHUTDOWN(log))
2639 goto error_return_unlocked;
2641 /* If there are other waiters on the queue then give them a
2642 * chance at logspace before us. Wake up the first waiters,
2643 * if we do not wake up all the waiters then go to sleep waiting
2644 * for more free space, otherwise try to get some space for
2647 need_bytes = tic->t_unit_res;
2648 if (!list_empty_careful(&log->l_writeq)) {
2649 struct xlog_ticket *ntic;
2651 spin_lock(&log->l_grant_write_lock);
2652 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2653 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2654 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2656 if (free_bytes < ntic->t_unit_res)
2658 free_bytes -= ntic->t_unit_res;
2659 wake_up(&ntic->t_wait);
2662 if (ntic != list_first_entry(&log->l_writeq,
2663 struct xlog_ticket, t_queue)) {
2664 if (list_empty(&tic->t_queue))
2665 list_add_tail(&tic->t_queue, &log->l_writeq);
2666 trace_xfs_log_regrant_write_sleep1(log, tic);
2668 xlog_grant_push_ail(log, need_bytes);
2670 XFS_STATS_INC(xs_sleep_logspace);
2671 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2672 trace_xfs_log_regrant_write_wake1(log, tic);
2674 spin_unlock(&log->l_grant_write_lock);
2678 if (XLOG_FORCED_SHUTDOWN(log))
2679 goto error_return_unlocked;
2681 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2682 if (free_bytes < need_bytes) {
2683 spin_lock(&log->l_grant_write_lock);
2684 if (list_empty(&tic->t_queue))
2685 list_add_tail(&tic->t_queue, &log->l_writeq);
2687 if (XLOG_FORCED_SHUTDOWN(log))
2690 xlog_grant_push_ail(log, need_bytes);
2692 XFS_STATS_INC(xs_sleep_logspace);
2693 trace_xfs_log_regrant_write_sleep2(log, tic);
2694 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2696 trace_xfs_log_regrant_write_wake2(log, tic);
2700 if (!list_empty(&tic->t_queue)) {
2701 spin_lock(&log->l_grant_write_lock);
2702 list_del_init(&tic->t_queue);
2703 spin_unlock(&log->l_grant_write_lock);
2706 /* we've got enough space */
2707 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2708 trace_xfs_log_regrant_write_exit(log, tic);
2709 xlog_verify_grant_tail(log);
2713 error_return_unlocked:
2714 spin_lock(&log->l_grant_write_lock);
2716 list_del_init(&tic->t_queue);
2717 spin_unlock(&log->l_grant_write_lock);
2718 trace_xfs_log_regrant_write_error(log, tic);
2721 * If we are failing, make sure the ticket doesn't have any
2722 * current reservations. We don't want to add this back when
2723 * the ticket/transaction gets cancelled.
2725 tic->t_curr_res = 0;
2726 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2727 return XFS_ERROR(EIO);
2728 } /* xlog_regrant_write_log_space */
2731 /* The first cnt-1 times through here we don't need to
2732 * move the grant write head because the permanent
2733 * reservation has reserved cnt times the unit amount.
2734 * Release part of current permanent unit reservation and
2735 * reset current reservation to be one units worth. Also
2736 * move grant reservation head forward.
2739 xlog_regrant_reserve_log_space(xlog_t *log,
2740 xlog_ticket_t *ticket)
2742 trace_xfs_log_regrant_reserve_enter(log, ticket);
2744 if (ticket->t_cnt > 0)
2747 xlog_grant_sub_space(log, &log->l_grant_reserve_head,
2748 ticket->t_curr_res);
2749 xlog_grant_sub_space(log, &log->l_grant_write_head,
2750 ticket->t_curr_res);
2751 ticket->t_curr_res = ticket->t_unit_res;
2752 xlog_tic_reset_res(ticket);
2754 trace_xfs_log_regrant_reserve_sub(log, ticket);
2756 /* just return if we still have some of the pre-reserved space */
2757 if (ticket->t_cnt > 0)
2760 xlog_grant_add_space(log, &log->l_grant_reserve_head,
2761 ticket->t_unit_res);
2763 trace_xfs_log_regrant_reserve_exit(log, ticket);
2765 ticket->t_curr_res = ticket->t_unit_res;
2766 xlog_tic_reset_res(ticket);
2767 } /* xlog_regrant_reserve_log_space */
2771 * Give back the space left from a reservation.
2773 * All the information we need to make a correct determination of space left
2774 * is present. For non-permanent reservations, things are quite easy. The
2775 * count should have been decremented to zero. We only need to deal with the
2776 * space remaining in the current reservation part of the ticket. If the
2777 * ticket contains a permanent reservation, there may be left over space which
2778 * needs to be released. A count of N means that N-1 refills of the current
2779 * reservation can be done before we need to ask for more space. The first
2780 * one goes to fill up the first current reservation. Once we run out of
2781 * space, the count will stay at zero and the only space remaining will be
2782 * in the current reservation field.
2785 xlog_ungrant_log_space(xlog_t *log,
2786 xlog_ticket_t *ticket)
2790 if (ticket->t_cnt > 0)
2793 trace_xfs_log_ungrant_enter(log, ticket);
2794 trace_xfs_log_ungrant_sub(log, ticket);
2797 * If this is a permanent reservation ticket, we may be able to free
2798 * up more space based on the remaining count.
2800 bytes = ticket->t_curr_res;
2801 if (ticket->t_cnt > 0) {
2802 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2803 bytes += ticket->t_unit_res*ticket->t_cnt;
2806 xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes);
2807 xlog_grant_sub_space(log, &log->l_grant_write_head, bytes);
2809 trace_xfs_log_ungrant_exit(log, ticket);
2811 xfs_log_move_tail(log->l_mp, 1);
2812 } /* xlog_ungrant_log_space */
2816 * Flush iclog to disk if this is the last reference to the given iclog and
2817 * the WANT_SYNC bit is set.
2819 * When this function is entered, the iclog is not necessarily in the
2820 * WANT_SYNC state. It may be sitting around waiting to get filled.
2825 xlog_state_release_iclog(
2827 xlog_in_core_t *iclog)
2829 int sync = 0; /* do we sync? */
2831 if (iclog->ic_state & XLOG_STATE_IOERROR)
2832 return XFS_ERROR(EIO);
2834 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2835 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2838 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2839 spin_unlock(&log->l_icloglock);
2840 return XFS_ERROR(EIO);
2842 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2843 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2845 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2846 /* update tail before writing to iclog */
2847 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2849 iclog->ic_state = XLOG_STATE_SYNCING;
2850 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2851 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2852 /* cycle incremented when incrementing curr_block */
2854 spin_unlock(&log->l_icloglock);
2857 * We let the log lock go, so it's possible that we hit a log I/O
2858 * error or some other SHUTDOWN condition that marks the iclog
2859 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2860 * this iclog has consistent data, so we ignore IOERROR
2861 * flags after this point.
2864 return xlog_sync(log, iclog);
2866 } /* xlog_state_release_iclog */
2870 * This routine will mark the current iclog in the ring as WANT_SYNC
2871 * and move the current iclog pointer to the next iclog in the ring.
2872 * When this routine is called from xlog_state_get_iclog_space(), the
2873 * exact size of the iclog has not yet been determined. All we know is
2874 * that every data block. We have run out of space in this log record.
2877 xlog_state_switch_iclogs(xlog_t *log,
2878 xlog_in_core_t *iclog,
2881 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2883 eventual_size = iclog->ic_offset;
2884 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2885 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2886 log->l_prev_block = log->l_curr_block;
2887 log->l_prev_cycle = log->l_curr_cycle;
2889 /* roll log?: ic_offset changed later */
2890 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2892 /* Round up to next log-sunit */
2893 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2894 log->l_mp->m_sb.sb_logsunit > 1) {
2895 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2896 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2899 if (log->l_curr_block >= log->l_logBBsize) {
2900 log->l_curr_cycle++;
2901 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2902 log->l_curr_cycle++;
2903 log->l_curr_block -= log->l_logBBsize;
2904 ASSERT(log->l_curr_block >= 0);
2906 ASSERT(iclog == log->l_iclog);
2907 log->l_iclog = iclog->ic_next;
2908 } /* xlog_state_switch_iclogs */
2911 * Write out all data in the in-core log as of this exact moment in time.
2913 * Data may be written to the in-core log during this call. However,
2914 * we don't guarantee this data will be written out. A change from past
2915 * implementation means this routine will *not* write out zero length LRs.
2917 * Basically, we try and perform an intelligent scan of the in-core logs.
2918 * If we determine there is no flushable data, we just return. There is no
2919 * flushable data if:
2921 * 1. the current iclog is active and has no data; the previous iclog
2922 * is in the active or dirty state.
2923 * 2. the current iclog is drity, and the previous iclog is in the
2924 * active or dirty state.
2928 * 1. the current iclog is not in the active nor dirty state.
2929 * 2. the current iclog dirty, and the previous iclog is not in the
2930 * active nor dirty state.
2931 * 3. the current iclog is active, and there is another thread writing
2932 * to this particular iclog.
2933 * 4. a) the current iclog is active and has no other writers
2934 * b) when we return from flushing out this iclog, it is still
2935 * not in the active nor dirty state.
2939 struct xfs_mount *mp,
2943 struct log *log = mp->m_log;
2944 struct xlog_in_core *iclog;
2947 XFS_STATS_INC(xs_log_force);
2950 xlog_cil_force(log);
2952 spin_lock(&log->l_icloglock);
2954 iclog = log->l_iclog;
2955 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2956 spin_unlock(&log->l_icloglock);
2957 return XFS_ERROR(EIO);
2960 /* If the head iclog is not active nor dirty, we just attach
2961 * ourselves to the head and go to sleep.
2963 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2964 iclog->ic_state == XLOG_STATE_DIRTY) {
2966 * If the head is dirty or (active and empty), then
2967 * we need to look at the previous iclog. If the previous
2968 * iclog is active or dirty we are done. There is nothing
2969 * to sync out. Otherwise, we attach ourselves to the
2970 * previous iclog and go to sleep.
2972 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2973 (atomic_read(&iclog->ic_refcnt) == 0
2974 && iclog->ic_offset == 0)) {
2975 iclog = iclog->ic_prev;
2976 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2977 iclog->ic_state == XLOG_STATE_DIRTY)
2982 if (atomic_read(&iclog->ic_refcnt) == 0) {
2983 /* We are the only one with access to this
2984 * iclog. Flush it out now. There should
2985 * be a roundoff of zero to show that someone
2986 * has already taken care of the roundoff from
2987 * the previous sync.
2989 atomic_inc(&iclog->ic_refcnt);
2990 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2991 xlog_state_switch_iclogs(log, iclog, 0);
2992 spin_unlock(&log->l_icloglock);
2994 if (xlog_state_release_iclog(log, iclog))
2995 return XFS_ERROR(EIO);
2999 spin_lock(&log->l_icloglock);
3000 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3001 iclog->ic_state != XLOG_STATE_DIRTY)
3006 /* Someone else is writing to this iclog.
3007 * Use its call to flush out the data. However,
3008 * the other thread may not force out this LR,
3009 * so we mark it WANT_SYNC.
3011 xlog_state_switch_iclogs(log, iclog, 0);
3017 /* By the time we come around again, the iclog could've been filled
3018 * which would give it another lsn. If we have a new lsn, just
3019 * return because the relevant data has been flushed.
3022 if (flags & XFS_LOG_SYNC) {
3024 * We must check if we're shutting down here, before
3025 * we wait, while we're holding the l_icloglock.
3026 * Then we check again after waking up, in case our
3027 * sleep was disturbed by a bad news.
3029 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3030 spin_unlock(&log->l_icloglock);
3031 return XFS_ERROR(EIO);
3033 XFS_STATS_INC(xs_log_force_sleep);
3034 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3036 * No need to grab the log lock here since we're
3037 * only deciding whether or not to return EIO
3038 * and the memory read should be atomic.
3040 if (iclog->ic_state & XLOG_STATE_IOERROR)
3041 return XFS_ERROR(EIO);
3047 spin_unlock(&log->l_icloglock);
3053 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3054 * about errors or whether the log was flushed or not. This is the normal
3055 * interface to use when trying to unpin items or move the log forward.
3064 error = _xfs_log_force(mp, flags, NULL);
3066 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3070 * Force the in-core log to disk for a specific LSN.
3072 * Find in-core log with lsn.
3073 * If it is in the DIRTY state, just return.
3074 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3075 * state and go to sleep or return.
3076 * If it is in any other state, go to sleep or return.
3078 * Synchronous forces are implemented with a signal variable. All callers
3079 * to force a given lsn to disk will wait on a the sv attached to the
3080 * specific in-core log. When given in-core log finally completes its
3081 * write to disk, that thread will wake up all threads waiting on the
3086 struct xfs_mount *mp,
3091 struct log *log = mp->m_log;
3092 struct xlog_in_core *iclog;
3093 int already_slept = 0;
3097 XFS_STATS_INC(xs_log_force);
3100 lsn = xlog_cil_force_lsn(log, lsn);
3101 if (lsn == NULLCOMMITLSN)
3106 spin_lock(&log->l_icloglock);
3107 iclog = log->l_iclog;
3108 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3109 spin_unlock(&log->l_icloglock);
3110 return XFS_ERROR(EIO);
3114 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3115 iclog = iclog->ic_next;
3119 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3120 spin_unlock(&log->l_icloglock);
3124 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3126 * We sleep here if we haven't already slept (e.g.
3127 * this is the first time we've looked at the correct
3128 * iclog buf) and the buffer before us is going to
3129 * be sync'ed. The reason for this is that if we
3130 * are doing sync transactions here, by waiting for
3131 * the previous I/O to complete, we can allow a few
3132 * more transactions into this iclog before we close
3135 * Otherwise, we mark the buffer WANT_SYNC, and bump
3136 * up the refcnt so we can release the log (which
3137 * drops the ref count). The state switch keeps new
3138 * transaction commits from using this buffer. When
3139 * the current commits finish writing into the buffer,
3140 * the refcount will drop to zero and the buffer will
3143 if (!already_slept &&
3144 (iclog->ic_prev->ic_state &
3145 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3146 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3148 XFS_STATS_INC(xs_log_force_sleep);
3150 xlog_wait(&iclog->ic_prev->ic_write_wait,
3157 atomic_inc(&iclog->ic_refcnt);
3158 xlog_state_switch_iclogs(log, iclog, 0);
3159 spin_unlock(&log->l_icloglock);
3160 if (xlog_state_release_iclog(log, iclog))
3161 return XFS_ERROR(EIO);
3164 spin_lock(&log->l_icloglock);
3167 if ((flags & XFS_LOG_SYNC) && /* sleep */
3169 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3171 * Don't wait on completion if we know that we've
3172 * gotten a log write error.
3174 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3175 spin_unlock(&log->l_icloglock);
3176 return XFS_ERROR(EIO);
3178 XFS_STATS_INC(xs_log_force_sleep);
3179 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3181 * No need to grab the log lock here since we're
3182 * only deciding whether or not to return EIO
3183 * and the memory read should be atomic.
3185 if (iclog->ic_state & XLOG_STATE_IOERROR)
3186 return XFS_ERROR(EIO);
3190 } else { /* just return */
3191 spin_unlock(&log->l_icloglock);
3195 } while (iclog != log->l_iclog);
3197 spin_unlock(&log->l_icloglock);
3202 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3203 * about errors or whether the log was flushed or not. This is the normal
3204 * interface to use when trying to unpin items or move the log forward.
3214 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3216 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3220 * Called when we want to mark the current iclog as being ready to sync to
3224 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3226 assert_spin_locked(&log->l_icloglock);
3228 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3229 xlog_state_switch_iclogs(log, iclog, 0);
3231 ASSERT(iclog->ic_state &
3232 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3237 /*****************************************************************************
3241 *****************************************************************************
3245 * Free a used ticket when its refcount falls to zero.
3249 xlog_ticket_t *ticket)
3251 ASSERT(atomic_read(&ticket->t_ref) > 0);
3252 if (atomic_dec_and_test(&ticket->t_ref))
3253 kmem_zone_free(xfs_log_ticket_zone, ticket);
3258 xlog_ticket_t *ticket)
3260 ASSERT(atomic_read(&ticket->t_ref) > 0);
3261 atomic_inc(&ticket->t_ref);
3266 * Allocate and initialise a new log ticket.
3277 struct xlog_ticket *tic;
3281 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3286 * Permanent reservations have up to 'cnt'-1 active log operations
3287 * in the log. A unit in this case is the amount of space for one
3288 * of these log operations. Normal reservations have a cnt of 1
3289 * and their unit amount is the total amount of space required.
3291 * The following lines of code account for non-transaction data
3292 * which occupy space in the on-disk log.
3294 * Normal form of a transaction is:
3295 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3296 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3298 * We need to account for all the leadup data and trailer data
3299 * around the transaction data.
3300 * And then we need to account for the worst case in terms of using
3302 * The worst case will happen if:
3303 * - the placement of the transaction happens to be such that the
3304 * roundoff is at its maximum
3305 * - the transaction data is synced before the commit record is synced
3306 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3307 * Therefore the commit record is in its own Log Record.
3308 * This can happen as the commit record is called with its
3309 * own region to xlog_write().
3310 * This then means that in the worst case, roundoff can happen for
3311 * the commit-rec as well.
3312 * The commit-rec is smaller than padding in this scenario and so it is
3313 * not added separately.
3316 /* for trans header */
3317 unit_bytes += sizeof(xlog_op_header_t);
3318 unit_bytes += sizeof(xfs_trans_header_t);
3321 unit_bytes += sizeof(xlog_op_header_t);
3324 * for LR headers - the space for data in an iclog is the size minus
3325 * the space used for the headers. If we use the iclog size, then we
3326 * undercalculate the number of headers required.
3328 * Furthermore - the addition of op headers for split-recs might
3329 * increase the space required enough to require more log and op
3330 * headers, so take that into account too.
3332 * IMPORTANT: This reservation makes the assumption that if this
3333 * transaction is the first in an iclog and hence has the LR headers
3334 * accounted to it, then the remaining space in the iclog is
3335 * exclusively for this transaction. i.e. if the transaction is larger
3336 * than the iclog, it will be the only thing in that iclog.
3337 * Fundamentally, this means we must pass the entire log vector to
3338 * xlog_write to guarantee this.
3340 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3341 num_headers = howmany(unit_bytes, iclog_space);
3343 /* for split-recs - ophdrs added when data split over LRs */
3344 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3346 /* add extra header reservations if we overrun */
3347 while (!num_headers ||
3348 howmany(unit_bytes, iclog_space) > num_headers) {
3349 unit_bytes += sizeof(xlog_op_header_t);
3352 unit_bytes += log->l_iclog_hsize * num_headers;
3354 /* for commit-rec LR header - note: padding will subsume the ophdr */
3355 unit_bytes += log->l_iclog_hsize;
3357 /* for roundoff padding for transaction data and one for commit record */
3358 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3359 log->l_mp->m_sb.sb_logsunit > 1) {
3360 /* log su roundoff */
3361 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3364 unit_bytes += 2*BBSIZE;
3367 atomic_set(&tic->t_ref, 1);
3368 INIT_LIST_HEAD(&tic->t_queue);
3369 tic->t_unit_res = unit_bytes;
3370 tic->t_curr_res = unit_bytes;
3373 tic->t_tid = random32();
3374 tic->t_clientid = client;
3375 tic->t_flags = XLOG_TIC_INITED;
3376 tic->t_trans_type = 0;
3377 if (xflags & XFS_LOG_PERM_RESERV)
3378 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3379 init_waitqueue_head(&tic->t_wait);
3381 xlog_tic_reset_res(tic);
3387 /******************************************************************************
3389 * Log debug routines
3391 ******************************************************************************
3395 * Make sure that the destination ptr is within the valid data region of
3396 * one of the iclogs. This uses backup pointers stored in a different
3397 * part of the log in case we trash the log structure.
3400 xlog_verify_dest_ptr(
3407 for (i = 0; i < log->l_iclog_bufs; i++) {
3408 if (ptr >= log->l_iclog_bak[i] &&
3409 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3414 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3418 * Check to make sure the grant write head didn't just over lap the tail. If
3419 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3420 * the cycles differ by exactly one and check the byte count.
3422 * This check is run unlocked, so can give false positives. Rather than assert
3423 * on failures, use a warn-once flag and a panic tag to allow the admin to
3424 * determine if they want to panic the machine when such an error occurs. For
3425 * debug kernels this will have the same effect as using an assert but, unlinke
3426 * an assert, it can be turned off at runtime.
3429 xlog_verify_grant_tail(
3432 int tail_cycle, tail_blocks;
3435 xlog_crack_grant_head(&log->l_grant_write_head, &cycle, &space);
3436 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3437 if (tail_cycle != cycle) {
3438 if (cycle - 1 != tail_cycle &&
3439 !(log->l_flags & XLOG_TAIL_WARN)) {
3440 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3441 "%s: cycle - 1 != tail_cycle", __func__);
3442 log->l_flags |= XLOG_TAIL_WARN;
3445 if (space > BBTOB(tail_blocks) &&
3446 !(log->l_flags & XLOG_TAIL_WARN)) {
3447 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3448 "%s: space > BBTOB(tail_blocks)", __func__);
3449 log->l_flags |= XLOG_TAIL_WARN;
3454 /* check if it will fit */
3456 xlog_verify_tail_lsn(xlog_t *log,
3457 xlog_in_core_t *iclog,
3462 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3464 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3465 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3466 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3468 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3470 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3471 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3473 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3474 if (blocks < BTOBB(iclog->ic_offset) + 1)
3475 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3477 } /* xlog_verify_tail_lsn */
3480 * Perform a number of checks on the iclog before writing to disk.
3482 * 1. Make sure the iclogs are still circular
3483 * 2. Make sure we have a good magic number
3484 * 3. Make sure we don't have magic numbers in the data
3485 * 4. Check fields of each log operation header for:
3486 * A. Valid client identifier
3487 * B. tid ptr value falls in valid ptr space (user space code)
3488 * C. Length in log record header is correct according to the
3489 * individual operation headers within record.
3490 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3491 * log, check the preceding blocks of the physical log to make sure all
3492 * the cycle numbers agree with the current cycle number.
3495 xlog_verify_iclog(xlog_t *log,
3496 xlog_in_core_t *iclog,
3500 xlog_op_header_t *ophead;
3501 xlog_in_core_t *icptr;
3502 xlog_in_core_2_t *xhdr;
3504 xfs_caddr_t base_ptr;
3505 __psint_t field_offset;
3507 int len, i, j, k, op_len;
3510 /* check validity of iclog pointers */
3511 spin_lock(&log->l_icloglock);
3512 icptr = log->l_iclog;
3513 for (i=0; i < log->l_iclog_bufs; i++) {
3515 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3516 icptr = icptr->ic_next;
3518 if (icptr != log->l_iclog)
3519 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3520 spin_unlock(&log->l_icloglock);
3522 /* check log magic numbers */
3523 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3524 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3526 ptr = (xfs_caddr_t) &iclog->ic_header;
3527 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3529 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3530 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3535 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3536 ptr = iclog->ic_datap;
3538 ophead = (xlog_op_header_t *)ptr;
3539 xhdr = iclog->ic_data;
3540 for (i = 0; i < len; i++) {
3541 ophead = (xlog_op_header_t *)ptr;
3543 /* clientid is only 1 byte */
3544 field_offset = (__psint_t)
3545 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3546 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3547 clientid = ophead->oh_clientid;
3549 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3550 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3551 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3552 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3553 clientid = xlog_get_client_id(
3554 xhdr[j].hic_xheader.xh_cycle_data[k]);
3556 clientid = xlog_get_client_id(
3557 iclog->ic_header.h_cycle_data[idx]);
3560 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3562 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3563 __func__, clientid, ophead,
3564 (unsigned long)field_offset);
3567 field_offset = (__psint_t)
3568 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3569 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3570 op_len = be32_to_cpu(ophead->oh_len);
3572 idx = BTOBBT((__psint_t)&ophead->oh_len -
3573 (__psint_t)iclog->ic_datap);
3574 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3575 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3576 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3577 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3579 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3582 ptr += sizeof(xlog_op_header_t) + op_len;
3584 } /* xlog_verify_iclog */
3588 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3594 xlog_in_core_t *iclog, *ic;
3596 iclog = log->l_iclog;
3597 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3599 * Mark all the incore logs IOERROR.
3600 * From now on, no log flushes will result.
3604 ic->ic_state = XLOG_STATE_IOERROR;
3606 } while (ic != iclog);
3610 * Return non-zero, if state transition has already happened.
3616 * This is called from xfs_force_shutdown, when we're forcibly
3617 * shutting down the filesystem, typically because of an IO error.
3618 * Our main objectives here are to make sure that:
3619 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3620 * parties to find out, 'atomically'.
3621 * b. those who're sleeping on log reservations, pinned objects and
3622 * other resources get woken up, and be told the bad news.
3623 * c. nothing new gets queued up after (a) and (b) are done.
3624 * d. if !logerror, flush the iclogs to disk, then seal them off
3627 * Note: for delayed logging the !logerror case needs to flush the regions
3628 * held in memory out to the iclogs before flushing them to disk. This needs
3629 * to be done before the log is marked as shutdown, otherwise the flush to the
3633 xfs_log_force_umount(
3634 struct xfs_mount *mp,
3644 * If this happens during log recovery, don't worry about
3645 * locking; the log isn't open for business yet.
3648 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3649 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3651 XFS_BUF_DONE(mp->m_sb_bp);
3656 * Somebody could've already done the hard work for us.
3657 * No need to get locks for this.
3659 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3660 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3666 * Flush the in memory commit item list before marking the log as
3667 * being shut down. We need to do it in this order to ensure all the
3668 * completed transactions are flushed to disk with the xfs_log_force()
3671 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3672 xlog_cil_force(log);
3675 * mark the filesystem and the as in a shutdown state and wake
3676 * everybody up to tell them the bad news.
3678 spin_lock(&log->l_icloglock);
3679 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3681 XFS_BUF_DONE(mp->m_sb_bp);
3684 * This flag is sort of redundant because of the mount flag, but
3685 * it's good to maintain the separation between the log and the rest
3688 log->l_flags |= XLOG_IO_ERROR;
3691 * If we hit a log error, we want to mark all the iclogs IOERROR
3692 * while we're still holding the loglock.
3695 retval = xlog_state_ioerror(log);
3696 spin_unlock(&log->l_icloglock);
3699 * We don't want anybody waiting for log reservations after this. That
3700 * means we have to wake up everybody queued up on reserveq as well as
3701 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3702 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3703 * action is protected by the grant locks.
3705 spin_lock(&log->l_grant_reserve_lock);
3706 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3707 wake_up(&tic->t_wait);
3708 spin_unlock(&log->l_grant_reserve_lock);
3710 spin_lock(&log->l_grant_write_lock);
3711 list_for_each_entry(tic, &log->l_writeq, t_queue)
3712 wake_up(&tic->t_wait);
3713 spin_unlock(&log->l_grant_write_lock);
3715 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3718 * Force the incore logs to disk before shutting the
3719 * log down completely.
3721 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3723 spin_lock(&log->l_icloglock);
3724 retval = xlog_state_ioerror(log);
3725 spin_unlock(&log->l_icloglock);
3728 * Wake up everybody waiting on xfs_log_force.
3729 * Callback all log item committed functions as if the
3730 * log writes were completed.
3732 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3734 #ifdef XFSERRORDEBUG
3736 xlog_in_core_t *iclog;
3738 spin_lock(&log->l_icloglock);
3739 iclog = log->l_iclog;
3741 ASSERT(iclog->ic_callback == 0);
3742 iclog = iclog->ic_next;
3743 } while (iclog != log->l_iclog);
3744 spin_unlock(&log->l_icloglock);
3747 /* return non-zero if log IOERROR transition had already happened */
3752 xlog_iclogs_empty(xlog_t *log)
3754 xlog_in_core_t *iclog;
3756 iclog = log->l_iclog;
3758 /* endianness does not matter here, zero is zero in
3761 if (iclog->ic_header.h_num_logops)
3763 iclog = iclog->ic_next;
3764 } while (iclog != log->l_iclog);