2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_buf_item.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_trans_priv.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
39 #include "xfs_trace.h"
41 kmem_zone_t *xfs_log_ticket_zone;
43 /* Local miscellaneous function prototypes */
44 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
50 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
51 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52 STATIC void xlog_dealloc_log(xlog_t *log);
54 /* local state machine functions */
55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57 STATIC int xlog_state_get_iclog_space(xlog_t *log,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
63 STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65 STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
70 /* local functions to manipulate grant head */
71 STATIC int xlog_grant_log_space(xlog_t *log,
73 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
75 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
76 xlog_ticket_t *ticket);
77 STATIC int xlog_regrant_write_log_space(xlog_t *log,
78 xlog_ticket_t *ticket);
79 STATIC void xlog_ungrant_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
83 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
85 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
86 int count, boolean_t syncing);
87 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
90 #define xlog_verify_dest_ptr(a,b)
91 #define xlog_verify_grant_head(a,b)
92 #define xlog_verify_iclog(a,b,c,d)
93 #define xlog_verify_tail_lsn(a,b,c)
96 STATIC int xlog_iclogs_empty(xlog_t *log);
100 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
104 tic->t_prev = (*qp)->t_prev;
105 (*qp)->t_prev->t_next = tic;
108 tic->t_prev = tic->t_next = tic;
112 tic->t_flags |= XLOG_TIC_IN_Q;
116 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
118 if (tic == tic->t_next) {
122 tic->t_next->t_prev = tic->t_prev;
123 tic->t_prev->t_next = tic->t_next;
126 tic->t_next = tic->t_prev = NULL;
127 tic->t_flags &= ~XLOG_TIC_IN_Q;
131 xlog_grant_sub_space(struct log *log, int bytes)
133 log->l_grant_write_bytes -= bytes;
134 if (log->l_grant_write_bytes < 0) {
135 log->l_grant_write_bytes += log->l_logsize;
136 log->l_grant_write_cycle--;
139 log->l_grant_reserve_bytes -= bytes;
140 if ((log)->l_grant_reserve_bytes < 0) {
141 log->l_grant_reserve_bytes += log->l_logsize;
142 log->l_grant_reserve_cycle--;
148 xlog_grant_add_space_write(struct log *log, int bytes)
150 int tmp = log->l_logsize - log->l_grant_write_bytes;
152 log->l_grant_write_bytes += bytes;
154 log->l_grant_write_cycle++;
155 log->l_grant_write_bytes = bytes - tmp;
160 xlog_grant_add_space_reserve(struct log *log, int bytes)
162 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
164 log->l_grant_reserve_bytes += bytes;
166 log->l_grant_reserve_cycle++;
167 log->l_grant_reserve_bytes = bytes - tmp;
172 xlog_grant_add_space(struct log *log, int bytes)
174 xlog_grant_add_space_write(log, bytes);
175 xlog_grant_add_space_reserve(log, bytes);
179 xlog_tic_reset_res(xlog_ticket_t *tic)
182 tic->t_res_arr_sum = 0;
183 tic->t_res_num_ophdrs = 0;
187 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
189 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
190 /* add to overflow and start again */
191 tic->t_res_o_flow += tic->t_res_arr_sum;
193 tic->t_res_arr_sum = 0;
196 tic->t_res_arr[tic->t_res_num].r_len = len;
197 tic->t_res_arr[tic->t_res_num].r_type = type;
198 tic->t_res_arr_sum += len;
205 * 1. currblock field gets updated at startup and after in-core logs
206 * marked as with WANT_SYNC.
210 * This routine is called when a user of a log manager ticket is done with
211 * the reservation. If the ticket was ever used, then a commit record for
212 * the associated transaction is written out as a log operation header with
213 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
214 * a given ticket. If the ticket was one with a permanent reservation, then
215 * a few operations are done differently. Permanent reservation tickets by
216 * default don't release the reservation. They just commit the current
217 * transaction with the belief that the reservation is still needed. A flag
218 * must be passed in before permanent reservations are actually released.
219 * When these type of tickets are not released, they need to be set into
220 * the inited state again. By doing this, a start record will be written
221 * out when the next write occurs.
225 struct xfs_mount *mp,
226 struct xlog_ticket *ticket,
227 struct xlog_in_core **iclog,
230 struct log *log = mp->m_log;
233 if (XLOG_FORCED_SHUTDOWN(log) ||
235 * If nothing was ever written, don't write out commit record.
236 * If we get an error, just continue and give back the log ticket.
238 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
239 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
240 lsn = (xfs_lsn_t) -1;
241 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
242 flags |= XFS_LOG_REL_PERM_RESERV;
247 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
248 (flags & XFS_LOG_REL_PERM_RESERV)) {
249 trace_xfs_log_done_nonperm(log, ticket);
252 * Release ticket if not permanent reservation or a specific
253 * request has been made to release a permanent reservation.
255 xlog_ungrant_log_space(log, ticket);
256 xfs_log_ticket_put(ticket);
258 trace_xfs_log_done_perm(log, ticket);
260 xlog_regrant_reserve_log_space(log, ticket);
261 /* If this ticket was a permanent reservation and we aren't
262 * trying to release it, reset the inited flags; so next time
263 * we write, a start record will be written out.
265 ticket->t_flags |= XLOG_TIC_INITED;
272 * Attaches a new iclog I/O completion callback routine during
273 * transaction commit. If the log is in error state, a non-zero
274 * return code is handed back and the caller is responsible for
275 * executing the callback at an appropriate time.
279 struct xfs_mount *mp,
280 struct xlog_in_core *iclog,
281 xfs_log_callback_t *cb)
285 spin_lock(&iclog->ic_callback_lock);
286 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
288 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
289 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
291 *(iclog->ic_callback_tail) = cb;
292 iclog->ic_callback_tail = &(cb->cb_next);
294 spin_unlock(&iclog->ic_callback_lock);
299 xfs_log_release_iclog(
300 struct xfs_mount *mp,
301 struct xlog_in_core *iclog)
303 if (xlog_state_release_iclog(mp->m_log, iclog)) {
304 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
312 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
313 * to the reservation.
314 * 2. Potentially, push buffers at tail of log to disk.
316 * Each reservation is going to reserve extra space for a log record header.
317 * When writes happen to the on-disk log, we don't subtract the length of the
318 * log record header from any reservation. By wasting space in each
319 * reservation, we prevent over allocation problems.
323 struct xfs_mount *mp,
326 struct xlog_ticket **ticket,
331 struct log *log = mp->m_log;
332 struct xlog_ticket *internal_ticket;
335 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
337 if (XLOG_FORCED_SHUTDOWN(log))
338 return XFS_ERROR(EIO);
340 XFS_STATS_INC(xs_try_logspace);
343 if (*ticket != NULL) {
344 ASSERT(flags & XFS_LOG_PERM_RESERV);
345 internal_ticket = *ticket;
348 * this is a new transaction on the ticket, so we need to
349 * change the transaction ID so that the next transaction has a
350 * different TID in the log. Just add one to the existing tid
351 * so that we can see chains of rolling transactions in the log
354 internal_ticket->t_tid++;
356 trace_xfs_log_reserve(log, internal_ticket);
358 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
359 retval = xlog_regrant_write_log_space(log, internal_ticket);
361 /* may sleep if need to allocate more tickets */
362 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
364 KM_SLEEP|KM_MAYFAIL);
365 if (!internal_ticket)
366 return XFS_ERROR(ENOMEM);
367 internal_ticket->t_trans_type = t_type;
368 *ticket = internal_ticket;
370 trace_xfs_log_reserve(log, internal_ticket);
372 xlog_grant_push_ail(mp,
373 (internal_ticket->t_unit_res *
374 internal_ticket->t_cnt));
375 retval = xlog_grant_log_space(log, internal_ticket);
379 } /* xfs_log_reserve */
383 * Mount a log filesystem
385 * mp - ubiquitous xfs mount point structure
386 * log_target - buftarg of on-disk log device
387 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
388 * num_bblocks - Number of BBSIZE blocks in on-disk log
390 * Return error or zero.
395 xfs_buftarg_t *log_target,
396 xfs_daddr_t blk_offset,
401 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
402 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
405 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
407 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
410 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
411 if (IS_ERR(mp->m_log)) {
412 error = -PTR_ERR(mp->m_log);
417 * Initialize the AIL now we have a log.
419 error = xfs_trans_ail_init(mp);
421 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
424 mp->m_log->l_ailp = mp->m_ail;
427 * skip log recovery on a norecovery mount. pretend it all
430 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
431 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
434 mp->m_flags &= ~XFS_MOUNT_RDONLY;
436 error = xlog_recover(mp->m_log);
439 mp->m_flags |= XFS_MOUNT_RDONLY;
441 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
442 goto out_destroy_ail;
446 /* Normal transactions can now occur */
447 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
450 * Now the log has been fully initialised and we know were our
451 * space grant counters are, we can initialise the permanent ticket
452 * needed for delayed logging to work.
454 xlog_cil_init_post_recovery(mp->m_log);
459 xfs_trans_ail_destroy(mp);
461 xlog_dealloc_log(mp->m_log);
467 * Finish the recovery of the file system. This is separate from
468 * the xfs_log_mount() call, because it depends on the code in
469 * xfs_mountfs() to read in the root and real-time bitmap inodes
470 * between calling xfs_log_mount() and here.
472 * mp - ubiquitous xfs mount point structure
475 xfs_log_mount_finish(xfs_mount_t *mp)
479 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
480 error = xlog_recover_finish(mp->m_log);
483 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
490 * Final log writes as part of unmount.
492 * Mark the filesystem clean as unmount happens. Note that during relocation
493 * this routine needs to be executed as part of source-bag while the
494 * deallocation must not be done until source-end.
498 * Unmount record used to have a string "Unmount filesystem--" in the
499 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
500 * We just write the magic number now since that particular field isn't
501 * currently architecture converted and "nUmount" is a bit foo.
502 * As far as I know, there weren't any dependencies on the old behaviour.
506 xfs_log_unmount_write(xfs_mount_t *mp)
508 xlog_t *log = mp->m_log;
509 xlog_in_core_t *iclog;
511 xlog_in_core_t *first_iclog;
513 xlog_ticket_t *tic = NULL;
518 * Don't write out unmount record on read-only mounts.
519 * Or, if we are doing a forced umount (typically because of IO errors).
521 if (mp->m_flags & XFS_MOUNT_RDONLY)
524 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
525 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
528 first_iclog = iclog = log->l_iclog;
530 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
531 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
532 ASSERT(iclog->ic_offset == 0);
534 iclog = iclog->ic_next;
535 } while (iclog != first_iclog);
537 if (! (XLOG_FORCED_SHUTDOWN(log))) {
538 error = xfs_log_reserve(mp, 600, 1, &tic,
539 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
541 /* the data section must be 32 bit size aligned */
545 __uint32_t pad2; /* may as well make it 64 bits */
547 .magic = XLOG_UNMOUNT_TYPE,
549 struct xfs_log_iovec reg = {
551 .i_len = sizeof(magic),
552 .i_type = XLOG_REG_TYPE_UNMOUNT,
554 struct xfs_log_vec vec = {
559 /* remove inited flag */
561 error = xlog_write(log, &vec, tic, &lsn,
562 NULL, XLOG_UNMOUNT_TRANS);
564 * At this point, we're umounting anyway,
565 * so there's no point in transitioning log state
566 * to IOERROR. Just continue...
571 xfs_fs_cmn_err(CE_ALERT, mp,
572 "xfs_log_unmount: unmount record failed");
576 spin_lock(&log->l_icloglock);
577 iclog = log->l_iclog;
578 atomic_inc(&iclog->ic_refcnt);
579 xlog_state_want_sync(log, iclog);
580 spin_unlock(&log->l_icloglock);
581 error = xlog_state_release_iclog(log, iclog);
583 spin_lock(&log->l_icloglock);
584 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
585 iclog->ic_state == XLOG_STATE_DIRTY)) {
586 if (!XLOG_FORCED_SHUTDOWN(log)) {
587 sv_wait(&iclog->ic_force_wait, PMEM,
588 &log->l_icloglock, s);
590 spin_unlock(&log->l_icloglock);
593 spin_unlock(&log->l_icloglock);
596 trace_xfs_log_umount_write(log, tic);
597 xlog_ungrant_log_space(log, tic);
598 xfs_log_ticket_put(tic);
602 * We're already in forced_shutdown mode, couldn't
603 * even attempt to write out the unmount transaction.
605 * Go through the motions of sync'ing and releasing
606 * the iclog, even though no I/O will actually happen,
607 * we need to wait for other log I/Os that may already
608 * be in progress. Do this as a separate section of
609 * code so we'll know if we ever get stuck here that
610 * we're in this odd situation of trying to unmount
611 * a file system that went into forced_shutdown as
612 * the result of an unmount..
614 spin_lock(&log->l_icloglock);
615 iclog = log->l_iclog;
616 atomic_inc(&iclog->ic_refcnt);
618 xlog_state_want_sync(log, iclog);
619 spin_unlock(&log->l_icloglock);
620 error = xlog_state_release_iclog(log, iclog);
622 spin_lock(&log->l_icloglock);
624 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
625 || iclog->ic_state == XLOG_STATE_DIRTY
626 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
628 sv_wait(&iclog->ic_force_wait, PMEM,
629 &log->l_icloglock, s);
631 spin_unlock(&log->l_icloglock);
636 } /* xfs_log_unmount_write */
639 * Deallocate log structures for unmount/relocation.
641 * We need to stop the aild from running before we destroy
642 * and deallocate the log as the aild references the log.
645 xfs_log_unmount(xfs_mount_t *mp)
647 xfs_trans_ail_destroy(mp);
648 xlog_dealloc_log(mp->m_log);
653 struct xfs_mount *mp,
654 struct xfs_log_item *item,
656 struct xfs_item_ops *ops)
658 item->li_mountp = mp;
659 item->li_ailp = mp->m_ail;
660 item->li_type = type;
664 INIT_LIST_HEAD(&item->li_ail);
665 INIT_LIST_HEAD(&item->li_cil);
669 * Write region vectors to log. The write happens using the space reservation
670 * of the ticket (tic). It is not a requirement that all writes for a given
671 * transaction occur with one call to xfs_log_write(). However, it is important
672 * to note that the transaction reservation code makes an assumption about the
673 * number of log headers a transaction requires that may be violated if you
674 * don't pass all the transaction vectors in one call....
678 struct xfs_mount *mp,
679 struct xfs_log_iovec reg[],
681 struct xlog_ticket *tic,
682 xfs_lsn_t *start_lsn)
684 struct log *log = mp->m_log;
686 struct xfs_log_vec vec = {
687 .lv_niovecs = nentries,
691 if (XLOG_FORCED_SHUTDOWN(log))
692 return XFS_ERROR(EIO);
694 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
696 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
701 xfs_log_move_tail(xfs_mount_t *mp,
705 xlog_t *log = mp->m_log;
706 int need_bytes, free_bytes, cycle, bytes;
708 if (XLOG_FORCED_SHUTDOWN(log))
712 /* needed since sync_lsn is 64 bits */
713 spin_lock(&log->l_icloglock);
714 tail_lsn = log->l_last_sync_lsn;
715 spin_unlock(&log->l_icloglock);
718 spin_lock(&log->l_grant_lock);
720 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
724 log->l_tail_lsn = tail_lsn;
727 if ((tic = log->l_write_headq)) {
729 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
730 panic("Recovery problem");
732 cycle = log->l_grant_write_cycle;
733 bytes = log->l_grant_write_bytes;
734 free_bytes = xlog_space_left(log, cycle, bytes);
736 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
738 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
741 free_bytes -= tic->t_unit_res;
742 sv_signal(&tic->t_wait);
744 } while (tic != log->l_write_headq);
746 if ((tic = log->l_reserve_headq)) {
748 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
749 panic("Recovery problem");
751 cycle = log->l_grant_reserve_cycle;
752 bytes = log->l_grant_reserve_bytes;
753 free_bytes = xlog_space_left(log, cycle, bytes);
755 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
756 need_bytes = tic->t_unit_res*tic->t_cnt;
758 need_bytes = tic->t_unit_res;
759 if (free_bytes < need_bytes && tail_lsn != 1)
762 free_bytes -= need_bytes;
763 sv_signal(&tic->t_wait);
765 } while (tic != log->l_reserve_headq);
767 spin_unlock(&log->l_grant_lock);
768 } /* xfs_log_move_tail */
771 * Determine if we have a transaction that has gone to disk
772 * that needs to be covered. To begin the transition to the idle state
773 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
774 * If we are then in a state where covering is needed, the caller is informed
775 * that dummy transactions are required to move the log into the idle state.
777 * Because this is called as part of the sync process, we should also indicate
778 * that dummy transactions should be issued in anything but the covered or
779 * idle states. This ensures that the log tail is accurately reflected in
780 * the log at the end of the sync, hence if a crash occurrs avoids replay
781 * of transactions where the metadata is already on disk.
784 xfs_log_need_covered(xfs_mount_t *mp)
787 xlog_t *log = mp->m_log;
789 if (!xfs_fs_writable(mp))
792 spin_lock(&log->l_icloglock);
793 switch (log->l_covered_state) {
794 case XLOG_STATE_COVER_DONE:
795 case XLOG_STATE_COVER_DONE2:
796 case XLOG_STATE_COVER_IDLE:
798 case XLOG_STATE_COVER_NEED:
799 case XLOG_STATE_COVER_NEED2:
800 if (!xfs_trans_ail_tail(log->l_ailp) &&
801 xlog_iclogs_empty(log)) {
802 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
803 log->l_covered_state = XLOG_STATE_COVER_DONE;
805 log->l_covered_state = XLOG_STATE_COVER_DONE2;
812 spin_unlock(&log->l_icloglock);
816 /******************************************************************************
820 ******************************************************************************
823 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
824 * The log manager must keep track of the last LR which was committed
825 * to disk. The lsn of this LR will become the new tail_lsn whenever
826 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
827 * the situation where stuff could be written into the log but nothing
828 * was ever in the AIL when asked. Eventually, we panic since the
829 * tail hits the head.
831 * We may be holding the log iclog lock upon entering this routine.
834 xlog_assign_tail_lsn(xfs_mount_t *mp)
837 xlog_t *log = mp->m_log;
839 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
840 spin_lock(&log->l_grant_lock);
842 log->l_tail_lsn = tail_lsn;
844 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
846 spin_unlock(&log->l_grant_lock);
849 } /* xlog_assign_tail_lsn */
853 * Return the space in the log between the tail and the head. The head
854 * is passed in the cycle/bytes formal parms. In the special case where
855 * the reserve head has wrapped passed the tail, this calculation is no
856 * longer valid. In this case, just return 0 which means there is no space
857 * in the log. This works for all places where this function is called
858 * with the reserve head. Of course, if the write head were to ever
859 * wrap the tail, we should blow up. Rather than catch this case here,
860 * we depend on other ASSERTions in other parts of the code. XXXmiken
862 * This code also handles the case where the reservation head is behind
863 * the tail. The details of this case are described below, but the end
864 * result is that we return the size of the log as the amount of space left.
867 xlog_space_left(xlog_t *log, int cycle, int bytes)
873 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
874 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
875 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
876 free_bytes = log->l_logsize - (bytes - tail_bytes);
877 } else if ((tail_cycle + 1) < cycle) {
879 } else if (tail_cycle < cycle) {
880 ASSERT(tail_cycle == (cycle - 1));
881 free_bytes = tail_bytes - bytes;
884 * The reservation head is behind the tail.
885 * In this case we just want to return the size of the
886 * log as the amount of space left.
888 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
889 "xlog_space_left: head behind tail\n"
890 " tail_cycle = %d, tail_bytes = %d\n"
891 " GH cycle = %d, GH bytes = %d",
892 tail_cycle, tail_bytes, cycle, bytes);
894 free_bytes = log->l_logsize;
897 } /* xlog_space_left */
901 * Log function which is called when an io completes.
903 * The log manager needs its own routine, in order to control what
904 * happens with the buffer after the write completes.
907 xlog_iodone(xfs_buf_t *bp)
909 xlog_in_core_t *iclog;
913 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
914 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
915 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
920 * Race to shutdown the filesystem if we see an error.
922 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
923 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
924 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
926 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
928 * This flag will be propagated to the trans-committed
929 * callback routines to let them know that the log-commit
932 aborted = XFS_LI_ABORTED;
933 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
934 aborted = XFS_LI_ABORTED;
937 /* log I/O is always issued ASYNC */
938 ASSERT(XFS_BUF_ISASYNC(bp));
939 xlog_state_done_syncing(iclog, aborted);
941 * do not reference the buffer (bp) here as we could race
942 * with it being freed after writing the unmount record to the
949 * Return size of each in-core log record buffer.
951 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
953 * If the filesystem blocksize is too large, we may need to choose a
954 * larger size since the directory code currently logs entire blocks.
958 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
964 if (mp->m_logbufs <= 0)
965 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
967 log->l_iclog_bufs = mp->m_logbufs;
970 * Buffer size passed in from mount system call.
972 if (mp->m_logbsize > 0) {
973 size = log->l_iclog_size = mp->m_logbsize;
974 log->l_iclog_size_log = 0;
976 log->l_iclog_size_log++;
980 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
981 /* # headers = size / 32k
982 * one header holds cycles from 32k of data
985 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
986 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
988 log->l_iclog_hsize = xhdrs << BBSHIFT;
989 log->l_iclog_heads = xhdrs;
991 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
992 log->l_iclog_hsize = BBSIZE;
993 log->l_iclog_heads = 1;
998 /* All machines use 32kB buffers by default. */
999 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1000 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1002 /* the default log size is 16k or 32k which is one header sector */
1003 log->l_iclog_hsize = BBSIZE;
1004 log->l_iclog_heads = 1;
1007 /* are we being asked to make the sizes selected above visible? */
1008 if (mp->m_logbufs == 0)
1009 mp->m_logbufs = log->l_iclog_bufs;
1010 if (mp->m_logbsize == 0)
1011 mp->m_logbsize = log->l_iclog_size;
1012 } /* xlog_get_iclog_buffer_size */
1016 * This routine initializes some of the log structure for a given mount point.
1017 * Its primary purpose is to fill in enough, so recovery can occur. However,
1018 * some other stuff may be filled in too.
1021 xlog_alloc_log(xfs_mount_t *mp,
1022 xfs_buftarg_t *log_target,
1023 xfs_daddr_t blk_offset,
1027 xlog_rec_header_t *head;
1028 xlog_in_core_t **iclogp;
1029 xlog_in_core_t *iclog, *prev_iclog=NULL;
1035 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1037 xlog_warn("XFS: Log allocation failed: No memory!");
1042 log->l_targ = log_target;
1043 log->l_logsize = BBTOB(num_bblks);
1044 log->l_logBBstart = blk_offset;
1045 log->l_logBBsize = num_bblks;
1046 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1047 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1049 log->l_prev_block = -1;
1050 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1051 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1052 log->l_last_sync_lsn = log->l_tail_lsn;
1053 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1054 log->l_grant_reserve_cycle = 1;
1055 log->l_grant_write_cycle = 1;
1057 error = EFSCORRUPTED;
1058 if (xfs_sb_version_hassector(&mp->m_sb)) {
1059 log2_size = mp->m_sb.sb_logsectlog;
1060 if (log2_size < BBSHIFT) {
1061 xlog_warn("XFS: Log sector size too small "
1062 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1066 log2_size -= BBSHIFT;
1067 if (log2_size > mp->m_sectbb_log) {
1068 xlog_warn("XFS: Log sector size too large "
1069 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1073 /* for larger sector sizes, must have v2 or external log */
1074 if (log2_size && log->l_logBBstart > 0 &&
1075 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1077 xlog_warn("XFS: log sector size (0x%x) invalid "
1078 "for configuration.", log2_size);
1082 log->l_sectBBsize = 1 << log2_size;
1084 xlog_get_iclog_buffer_size(mp, log);
1087 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1090 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1091 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1092 ASSERT(XFS_BUF_ISBUSY(bp));
1093 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1096 spin_lock_init(&log->l_icloglock);
1097 spin_lock_init(&log->l_grant_lock);
1098 sv_init(&log->l_flush_wait, 0, "flush_wait");
1100 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1101 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1103 iclogp = &log->l_iclog;
1105 * The amount of memory to allocate for the iclog structure is
1106 * rather funky due to the way the structure is defined. It is
1107 * done this way so that we can use different sizes for machines
1108 * with different amounts of memory. See the definition of
1109 * xlog_in_core_t in xfs_log_priv.h for details.
1111 ASSERT(log->l_iclog_size >= 4096);
1112 for (i=0; i < log->l_iclog_bufs; i++) {
1113 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1115 goto out_free_iclog;
1118 iclog->ic_prev = prev_iclog;
1121 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1123 goto out_free_iclog;
1124 if (!XFS_BUF_CPSEMA(bp))
1126 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1127 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1129 iclog->ic_data = bp->b_addr;
1131 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1133 head = &iclog->ic_header;
1134 memset(head, 0, sizeof(xlog_rec_header_t));
1135 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1136 head->h_version = cpu_to_be32(
1137 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1138 head->h_size = cpu_to_be32(log->l_iclog_size);
1140 head->h_fmt = cpu_to_be32(XLOG_FMT);
1141 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1143 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1144 iclog->ic_state = XLOG_STATE_ACTIVE;
1145 iclog->ic_log = log;
1146 atomic_set(&iclog->ic_refcnt, 0);
1147 spin_lock_init(&iclog->ic_callback_lock);
1148 iclog->ic_callback_tail = &(iclog->ic_callback);
1149 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1151 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1152 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1153 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1154 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1156 iclogp = &iclog->ic_next;
1158 *iclogp = log->l_iclog; /* complete ring */
1159 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1161 error = xlog_cil_init(log);
1163 goto out_free_iclog;
1167 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1168 prev_iclog = iclog->ic_next;
1170 sv_destroy(&iclog->ic_force_wait);
1171 sv_destroy(&iclog->ic_write_wait);
1172 xfs_buf_free(iclog->ic_bp);
1176 spinlock_destroy(&log->l_icloglock);
1177 spinlock_destroy(&log->l_grant_lock);
1178 xfs_buf_free(log->l_xbuf);
1182 return ERR_PTR(-error);
1183 } /* xlog_alloc_log */
1187 * Write out the commit record of a transaction associated with the given
1188 * ticket. Return the lsn of the commit record.
1193 struct xlog_ticket *ticket,
1194 struct xlog_in_core **iclog,
1195 xfs_lsn_t *commitlsnp)
1197 struct xfs_mount *mp = log->l_mp;
1199 struct xfs_log_iovec reg = {
1202 .i_type = XLOG_REG_TYPE_COMMIT,
1204 struct xfs_log_vec vec = {
1209 ASSERT_ALWAYS(iclog);
1210 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1213 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1218 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1219 * log space. This code pushes on the lsn which would supposedly free up
1220 * the 25% which we want to leave free. We may need to adopt a policy which
1221 * pushes on an lsn which is further along in the log once we reach the high
1222 * water mark. In this manner, we would be creating a low water mark.
1225 xlog_grant_push_ail(xfs_mount_t *mp,
1228 xlog_t *log = mp->m_log; /* pointer to the log */
1229 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1230 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1231 int free_blocks; /* free blocks left to write to */
1232 int free_bytes; /* free bytes left to write to */
1233 int threshold_block; /* block in lsn we'd like to be at */
1234 int threshold_cycle; /* lsn cycle we'd like to be at */
1237 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1239 spin_lock(&log->l_grant_lock);
1240 free_bytes = xlog_space_left(log,
1241 log->l_grant_reserve_cycle,
1242 log->l_grant_reserve_bytes);
1243 tail_lsn = log->l_tail_lsn;
1244 free_blocks = BTOBBT(free_bytes);
1247 * Set the threshold for the minimum number of free blocks in the
1248 * log to the maximum of what the caller needs, one quarter of the
1249 * log, and 256 blocks.
1251 free_threshold = BTOBB(need_bytes);
1252 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1253 free_threshold = MAX(free_threshold, 256);
1254 if (free_blocks < free_threshold) {
1255 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1256 threshold_cycle = CYCLE_LSN(tail_lsn);
1257 if (threshold_block >= log->l_logBBsize) {
1258 threshold_block -= log->l_logBBsize;
1259 threshold_cycle += 1;
1261 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1263 /* Don't pass in an lsn greater than the lsn of the last
1264 * log record known to be on disk.
1266 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1267 threshold_lsn = log->l_last_sync_lsn;
1269 spin_unlock(&log->l_grant_lock);
1272 * Get the transaction layer to kick the dirty buffers out to
1273 * disk asynchronously. No point in trying to do this if
1274 * the filesystem is shutting down.
1276 if (threshold_lsn &&
1277 !XLOG_FORCED_SHUTDOWN(log))
1278 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1279 } /* xlog_grant_push_ail */
1282 * The bdstrat callback function for log bufs. This gives us a central
1283 * place to trap bufs in case we get hit by a log I/O error and need to
1284 * shutdown. Actually, in practice, even when we didn't get a log error,
1285 * we transition the iclogs to IOERROR state *after* flushing all existing
1286 * iclogs to disk. This is because we don't want anymore new transactions to be
1287 * started or completed afterwards.
1293 struct xlog_in_core *iclog;
1295 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1296 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1297 XFS_BUF_ERROR(bp, EIO);
1301 * It would seem logical to return EIO here, but we rely on
1302 * the log state machine to propagate I/O errors instead of
1308 bp->b_flags |= _XBF_RUN_QUEUES;
1309 xfs_buf_iorequest(bp);
1314 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1315 * fashion. Previously, we should have moved the current iclog
1316 * ptr in the log to point to the next available iclog. This allows further
1317 * write to continue while this code syncs out an iclog ready to go.
1318 * Before an in-core log can be written out, the data section must be scanned
1319 * to save away the 1st word of each BBSIZE block into the header. We replace
1320 * it with the current cycle count. Each BBSIZE block is tagged with the
1321 * cycle count because there in an implicit assumption that drives will
1322 * guarantee that entire 512 byte blocks get written at once. In other words,
1323 * we can't have part of a 512 byte block written and part not written. By
1324 * tagging each block, we will know which blocks are valid when recovering
1325 * after an unclean shutdown.
1327 * This routine is single threaded on the iclog. No other thread can be in
1328 * this routine with the same iclog. Changing contents of iclog can there-
1329 * fore be done without grabbing the state machine lock. Updating the global
1330 * log will require grabbing the lock though.
1332 * The entire log manager uses a logical block numbering scheme. Only
1333 * log_sync (and then only bwrite()) know about the fact that the log may
1334 * not start with block zero on a given device. The log block start offset
1335 * is added immediately before calling bwrite().
1339 xlog_sync(xlog_t *log,
1340 xlog_in_core_t *iclog)
1342 xfs_caddr_t dptr; /* pointer to byte sized element */
1345 uint count; /* byte count of bwrite */
1346 uint count_init; /* initial count before roundup */
1347 int roundoff; /* roundoff to BB or stripe */
1348 int split = 0; /* split write into two regions */
1350 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1352 XFS_STATS_INC(xs_log_writes);
1353 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1355 /* Add for LR header */
1356 count_init = log->l_iclog_hsize + iclog->ic_offset;
1358 /* Round out the log write size */
1359 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1360 /* we have a v2 stripe unit to use */
1361 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1363 count = BBTOB(BTOBB(count_init));
1365 roundoff = count - count_init;
1366 ASSERT(roundoff >= 0);
1367 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1368 roundoff < log->l_mp->m_sb.sb_logsunit)
1370 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1371 roundoff < BBTOB(1)));
1373 /* move grant heads by roundoff in sync */
1374 spin_lock(&log->l_grant_lock);
1375 xlog_grant_add_space(log, roundoff);
1376 spin_unlock(&log->l_grant_lock);
1378 /* put cycle number in every block */
1379 xlog_pack_data(log, iclog, roundoff);
1381 /* real byte length */
1383 iclog->ic_header.h_len =
1384 cpu_to_be32(iclog->ic_offset + roundoff);
1386 iclog->ic_header.h_len =
1387 cpu_to_be32(iclog->ic_offset);
1391 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1392 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1393 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1395 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1397 /* Do we need to split this write into 2 parts? */
1398 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1399 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1400 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1401 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1403 iclog->ic_bwritecnt = 1;
1405 XFS_BUF_SET_COUNT(bp, count);
1406 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1407 XFS_BUF_ZEROFLAGS(bp);
1410 bp->b_flags |= XBF_LOG_BUFFER;
1412 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1413 XFS_BUF_ORDERED(bp);
1415 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1416 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1418 xlog_verify_iclog(log, iclog, count, B_TRUE);
1420 /* account for log which doesn't start at block #0 */
1421 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1423 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1428 if ((error = xlog_bdstrat(bp))) {
1429 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1434 bp = iclog->ic_log->l_xbuf;
1435 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1437 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1438 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1439 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1440 (__psint_t)count), split);
1441 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1442 XFS_BUF_ZEROFLAGS(bp);
1445 bp->b_flags |= XBF_LOG_BUFFER;
1446 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1447 XFS_BUF_ORDERED(bp);
1448 dptr = XFS_BUF_PTR(bp);
1450 * Bump the cycle numbers at the start of each block
1451 * since this part of the buffer is at the start of
1452 * a new cycle. Watch out for the header magic number
1455 for (i = 0; i < split; i += BBSIZE) {
1456 be32_add_cpu((__be32 *)dptr, 1);
1457 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1458 be32_add_cpu((__be32 *)dptr, 1);
1462 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1463 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1465 /* account for internal log which doesn't start at block #0 */
1466 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1468 if ((error = xlog_bdstrat(bp))) {
1469 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1470 bp, XFS_BUF_ADDR(bp));
1479 * Deallocate a log structure
1482 xlog_dealloc_log(xlog_t *log)
1484 xlog_in_core_t *iclog, *next_iclog;
1487 xlog_cil_destroy(log);
1489 iclog = log->l_iclog;
1490 for (i=0; i<log->l_iclog_bufs; i++) {
1491 sv_destroy(&iclog->ic_force_wait);
1492 sv_destroy(&iclog->ic_write_wait);
1493 xfs_buf_free(iclog->ic_bp);
1494 next_iclog = iclog->ic_next;
1498 spinlock_destroy(&log->l_icloglock);
1499 spinlock_destroy(&log->l_grant_lock);
1501 xfs_buf_free(log->l_xbuf);
1502 log->l_mp->m_log = NULL;
1504 } /* xlog_dealloc_log */
1507 * Update counters atomically now that memcpy is done.
1511 xlog_state_finish_copy(xlog_t *log,
1512 xlog_in_core_t *iclog,
1516 spin_lock(&log->l_icloglock);
1518 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1519 iclog->ic_offset += copy_bytes;
1521 spin_unlock(&log->l_icloglock);
1522 } /* xlog_state_finish_copy */
1528 * print out info relating to regions written which consume
1533 struct xfs_mount *mp,
1534 struct xlog_ticket *ticket)
1537 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1539 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1540 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1561 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1604 xfs_fs_cmn_err(CE_WARN, mp,
1605 "xfs_log_write: reservation summary:\n"
1606 " trans type = %s (%u)\n"
1607 " unit res = %d bytes\n"
1608 " current res = %d bytes\n"
1609 " total reg = %u bytes (o/flow = %u bytes)\n"
1610 " ophdrs = %u (ophdr space = %u bytes)\n"
1611 " ophdr + reg = %u bytes\n"
1612 " num regions = %u\n",
1613 ((ticket->t_trans_type <= 0 ||
1614 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1615 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1616 ticket->t_trans_type,
1619 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1620 ticket->t_res_num_ophdrs, ophdr_spc,
1621 ticket->t_res_arr_sum +
1622 ticket->t_res_o_flow + ophdr_spc,
1625 for (i = 0; i < ticket->t_res_num; i++) {
1626 uint r_type = ticket->t_res_arr[i].r_type;
1628 "region[%u]: %s - %u bytes\n",
1630 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1631 "bad-rtype" : res_type_str[r_type-1]),
1632 ticket->t_res_arr[i].r_len);
1635 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1636 "xfs_log_write: reservation ran out. Need to up reservation");
1637 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1641 * Calculate the potential space needed by the log vector. Each region gets
1642 * its own xlog_op_header_t and may need to be double word aligned.
1645 xlog_write_calc_vec_length(
1646 struct xlog_ticket *ticket,
1647 struct xfs_log_vec *log_vector)
1649 struct xfs_log_vec *lv;
1654 /* acct for start rec of xact */
1655 if (ticket->t_flags & XLOG_TIC_INITED)
1658 for (lv = log_vector; lv; lv = lv->lv_next) {
1659 headers += lv->lv_niovecs;
1661 for (i = 0; i < lv->lv_niovecs; i++) {
1662 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1665 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1669 ticket->t_res_num_ophdrs += headers;
1670 len += headers * sizeof(struct xlog_op_header);
1676 * If first write for transaction, insert start record We can't be trying to
1677 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1680 xlog_write_start_rec(
1681 struct xlog_op_header *ophdr,
1682 struct xlog_ticket *ticket)
1684 if (!(ticket->t_flags & XLOG_TIC_INITED))
1687 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1688 ophdr->oh_clientid = ticket->t_clientid;
1690 ophdr->oh_flags = XLOG_START_TRANS;
1693 ticket->t_flags &= ~XLOG_TIC_INITED;
1695 return sizeof(struct xlog_op_header);
1698 static xlog_op_header_t *
1699 xlog_write_setup_ophdr(
1701 struct xlog_op_header *ophdr,
1702 struct xlog_ticket *ticket,
1705 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1706 ophdr->oh_clientid = ticket->t_clientid;
1709 /* are we copying a commit or unmount record? */
1710 ophdr->oh_flags = flags;
1713 * We've seen logs corrupted with bad transaction client ids. This
1714 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1715 * and shut down the filesystem.
1717 switch (ophdr->oh_clientid) {
1718 case XFS_TRANSACTION:
1723 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1724 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1725 ophdr->oh_clientid, ticket);
1733 * Set up the parameters of the region copy into the log. This has
1734 * to handle region write split across multiple log buffers - this
1735 * state is kept external to this function so that this code can
1736 * can be written in an obvious, self documenting manner.
1739 xlog_write_setup_copy(
1740 struct xlog_ticket *ticket,
1741 struct xlog_op_header *ophdr,
1742 int space_available,
1746 int *last_was_partial_copy,
1747 int *bytes_consumed)
1751 still_to_copy = space_required - *bytes_consumed;
1752 *copy_off = *bytes_consumed;
1754 if (still_to_copy <= space_available) {
1755 /* write of region completes here */
1756 *copy_len = still_to_copy;
1757 ophdr->oh_len = cpu_to_be32(*copy_len);
1758 if (*last_was_partial_copy)
1759 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1760 *last_was_partial_copy = 0;
1761 *bytes_consumed = 0;
1765 /* partial write of region, needs extra log op header reservation */
1766 *copy_len = space_available;
1767 ophdr->oh_len = cpu_to_be32(*copy_len);
1768 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1769 if (*last_was_partial_copy)
1770 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1771 *bytes_consumed += *copy_len;
1772 (*last_was_partial_copy)++;
1774 /* account for new log op header */
1775 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1776 ticket->t_res_num_ophdrs++;
1778 return sizeof(struct xlog_op_header);
1782 xlog_write_copy_finish(
1784 struct xlog_in_core *iclog,
1789 int *partial_copy_len,
1791 struct xlog_in_core **commit_iclog)
1793 if (*partial_copy) {
1795 * This iclog has already been marked WANT_SYNC by
1796 * xlog_state_get_iclog_space.
1798 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1801 return xlog_state_release_iclog(log, iclog);
1805 *partial_copy_len = 0;
1807 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1808 /* no more space in this iclog - push it. */
1809 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1813 spin_lock(&log->l_icloglock);
1814 xlog_state_want_sync(log, iclog);
1815 spin_unlock(&log->l_icloglock);
1818 return xlog_state_release_iclog(log, iclog);
1819 ASSERT(flags & XLOG_COMMIT_TRANS);
1820 *commit_iclog = iclog;
1827 * Write some region out to in-core log
1829 * This will be called when writing externally provided regions or when
1830 * writing out a commit record for a given transaction.
1832 * General algorithm:
1833 * 1. Find total length of this write. This may include adding to the
1834 * lengths passed in.
1835 * 2. Check whether we violate the tickets reservation.
1836 * 3. While writing to this iclog
1837 * A. Reserve as much space in this iclog as can get
1838 * B. If this is first write, save away start lsn
1839 * C. While writing this region:
1840 * 1. If first write of transaction, write start record
1841 * 2. Write log operation header (header per region)
1842 * 3. Find out if we can fit entire region into this iclog
1843 * 4. Potentially, verify destination memcpy ptr
1844 * 5. Memcpy (partial) region
1845 * 6. If partial copy, release iclog; otherwise, continue
1846 * copying more regions into current iclog
1847 * 4. Mark want sync bit (in simulation mode)
1848 * 5. Release iclog for potential flush to on-disk log.
1851 * 1. Panic if reservation is overrun. This should never happen since
1852 * reservation amounts are generated internal to the filesystem.
1854 * 1. Tickets are single threaded data structures.
1855 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1856 * syncing routine. When a single log_write region needs to span
1857 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1858 * on all log operation writes which don't contain the end of the
1859 * region. The XLOG_END_TRANS bit is used for the in-core log
1860 * operation which contains the end of the continued log_write region.
1861 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1862 * we don't really know exactly how much space will be used. As a result,
1863 * we don't update ic_offset until the end when we know exactly how many
1864 * bytes have been written out.
1869 struct xfs_log_vec *log_vector,
1870 struct xlog_ticket *ticket,
1871 xfs_lsn_t *start_lsn,
1872 struct xlog_in_core **commit_iclog,
1875 struct xlog_in_core *iclog = NULL;
1876 struct xfs_log_iovec *vecp;
1877 struct xfs_log_vec *lv;
1880 int partial_copy = 0;
1881 int partial_copy_len = 0;
1889 len = xlog_write_calc_vec_length(ticket, log_vector);
1892 * Region headers and bytes are already accounted for.
1893 * We only need to take into account start records and
1894 * split regions in this function.
1896 if (ticket->t_flags & XLOG_TIC_INITED)
1897 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1900 * Commit record headers need to be accounted for. These
1901 * come in as separate writes so are easy to detect.
1903 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1904 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1906 ticket->t_curr_res -= len;
1908 if (ticket->t_curr_res < 0)
1909 xlog_print_tic_res(log->l_mp, ticket);
1913 vecp = lv->lv_iovecp;
1914 while (lv && index < lv->lv_niovecs) {
1918 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1919 &contwr, &log_offset);
1923 ASSERT(log_offset <= iclog->ic_size - 1);
1924 ptr = iclog->ic_datap + log_offset;
1926 /* start_lsn is the first lsn written to. That's all we need. */
1928 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1931 * This loop writes out as many regions as can fit in the amount
1932 * of space which was allocated by xlog_state_get_iclog_space().
1934 while (lv && index < lv->lv_niovecs) {
1935 struct xfs_log_iovec *reg = &vecp[index];
1936 struct xlog_op_header *ophdr;
1941 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1942 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1944 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1945 if (start_rec_copy) {
1947 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1951 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1953 return XFS_ERROR(EIO);
1955 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1956 sizeof(struct xlog_op_header));
1958 len += xlog_write_setup_copy(ticket, ophdr,
1959 iclog->ic_size-log_offset,
1961 ©_off, ©_len,
1964 xlog_verify_dest_ptr(log, ptr);
1967 ASSERT(copy_len >= 0);
1968 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1969 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1971 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1973 data_cnt += contwr ? copy_len : 0;
1975 error = xlog_write_copy_finish(log, iclog, flags,
1976 &record_cnt, &data_cnt,
1985 * if we had a partial copy, we need to get more iclog
1986 * space but we don't want to increment the region
1987 * index because there is still more is this region to
1990 * If we completed writing this region, and we flushed
1991 * the iclog (indicated by resetting of the record
1992 * count), then we also need to get more log space. If
1993 * this was the last record, though, we are done and
1999 if (++index == lv->lv_niovecs) {
2003 vecp = lv->lv_iovecp;
2005 if (record_cnt == 0) {
2015 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2017 return xlog_state_release_iclog(log, iclog);
2019 ASSERT(flags & XLOG_COMMIT_TRANS);
2020 *commit_iclog = iclog;
2025 /*****************************************************************************
2027 * State Machine functions
2029 *****************************************************************************
2032 /* Clean iclogs starting from the head. This ordering must be
2033 * maintained, so an iclog doesn't become ACTIVE beyond one that
2034 * is SYNCING. This is also required to maintain the notion that we use
2035 * a ordered wait queue to hold off would be writers to the log when every
2036 * iclog is trying to sync to disk.
2038 * State Change: DIRTY -> ACTIVE
2041 xlog_state_clean_log(xlog_t *log)
2043 xlog_in_core_t *iclog;
2046 iclog = log->l_iclog;
2048 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2049 iclog->ic_state = XLOG_STATE_ACTIVE;
2050 iclog->ic_offset = 0;
2051 ASSERT(iclog->ic_callback == NULL);
2053 * If the number of ops in this iclog indicate it just
2054 * contains the dummy transaction, we can
2055 * change state into IDLE (the second time around).
2056 * Otherwise we should change the state into
2058 * We don't need to cover the dummy.
2061 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2066 * We have two dirty iclogs so start over
2067 * This could also be num of ops indicates
2068 * this is not the dummy going out.
2072 iclog->ic_header.h_num_logops = 0;
2073 memset(iclog->ic_header.h_cycle_data, 0,
2074 sizeof(iclog->ic_header.h_cycle_data));
2075 iclog->ic_header.h_lsn = 0;
2076 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2079 break; /* stop cleaning */
2080 iclog = iclog->ic_next;
2081 } while (iclog != log->l_iclog);
2083 /* log is locked when we are called */
2085 * Change state for the dummy log recording.
2086 * We usually go to NEED. But we go to NEED2 if the changed indicates
2087 * we are done writing the dummy record.
2088 * If we are done with the second dummy recored (DONE2), then
2092 switch (log->l_covered_state) {
2093 case XLOG_STATE_COVER_IDLE:
2094 case XLOG_STATE_COVER_NEED:
2095 case XLOG_STATE_COVER_NEED2:
2096 log->l_covered_state = XLOG_STATE_COVER_NEED;
2099 case XLOG_STATE_COVER_DONE:
2101 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2103 log->l_covered_state = XLOG_STATE_COVER_NEED;
2106 case XLOG_STATE_COVER_DONE2:
2108 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2110 log->l_covered_state = XLOG_STATE_COVER_NEED;
2117 } /* xlog_state_clean_log */
2120 xlog_get_lowest_lsn(
2123 xlog_in_core_t *lsn_log;
2124 xfs_lsn_t lowest_lsn, lsn;
2126 lsn_log = log->l_iclog;
2129 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2130 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2131 if ((lsn && !lowest_lsn) ||
2132 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2136 lsn_log = lsn_log->ic_next;
2137 } while (lsn_log != log->l_iclog);
2143 xlog_state_do_callback(
2146 xlog_in_core_t *ciclog)
2148 xlog_in_core_t *iclog;
2149 xlog_in_core_t *first_iclog; /* used to know when we've
2150 * processed all iclogs once */
2151 xfs_log_callback_t *cb, *cb_next;
2153 xfs_lsn_t lowest_lsn;
2154 int ioerrors; /* counter: iclogs with errors */
2155 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2156 int funcdidcallbacks; /* flag: function did callbacks */
2157 int repeats; /* for issuing console warnings if
2158 * looping too many times */
2161 spin_lock(&log->l_icloglock);
2162 first_iclog = iclog = log->l_iclog;
2164 funcdidcallbacks = 0;
2169 * Scan all iclogs starting with the one pointed to by the
2170 * log. Reset this starting point each time the log is
2171 * unlocked (during callbacks).
2173 * Keep looping through iclogs until one full pass is made
2174 * without running any callbacks.
2176 first_iclog = log->l_iclog;
2177 iclog = log->l_iclog;
2178 loopdidcallbacks = 0;
2183 /* skip all iclogs in the ACTIVE & DIRTY states */
2184 if (iclog->ic_state &
2185 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2186 iclog = iclog->ic_next;
2191 * Between marking a filesystem SHUTDOWN and stopping
2192 * the log, we do flush all iclogs to disk (if there
2193 * wasn't a log I/O error). So, we do want things to
2194 * go smoothly in case of just a SHUTDOWN w/o a
2197 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2199 * Can only perform callbacks in order. Since
2200 * this iclog is not in the DONE_SYNC/
2201 * DO_CALLBACK state, we skip the rest and
2202 * just try to clean up. If we set our iclog
2203 * to DO_CALLBACK, we will not process it when
2204 * we retry since a previous iclog is in the
2205 * CALLBACK and the state cannot change since
2206 * we are holding the l_icloglock.
2208 if (!(iclog->ic_state &
2209 (XLOG_STATE_DONE_SYNC |
2210 XLOG_STATE_DO_CALLBACK))) {
2211 if (ciclog && (ciclog->ic_state ==
2212 XLOG_STATE_DONE_SYNC)) {
2213 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2218 * We now have an iclog that is in either the
2219 * DO_CALLBACK or DONE_SYNC states. The other
2220 * states (WANT_SYNC, SYNCING, or CALLBACK were
2221 * caught by the above if and are going to
2222 * clean (i.e. we aren't doing their callbacks)
2227 * We will do one more check here to see if we
2228 * have chased our tail around.
2231 lowest_lsn = xlog_get_lowest_lsn(log);
2233 XFS_LSN_CMP(lowest_lsn,
2234 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2235 iclog = iclog->ic_next;
2236 continue; /* Leave this iclog for
2240 iclog->ic_state = XLOG_STATE_CALLBACK;
2242 spin_unlock(&log->l_icloglock);
2244 /* l_last_sync_lsn field protected by
2245 * l_grant_lock. Don't worry about iclog's lsn.
2246 * No one else can be here except us.
2248 spin_lock(&log->l_grant_lock);
2249 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2250 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2251 log->l_last_sync_lsn =
2252 be64_to_cpu(iclog->ic_header.h_lsn);
2253 spin_unlock(&log->l_grant_lock);
2256 spin_unlock(&log->l_icloglock);
2261 * Keep processing entries in the callback list until
2262 * we come around and it is empty. We need to
2263 * atomically see that the list is empty and change the
2264 * state to DIRTY so that we don't miss any more
2265 * callbacks being added.
2267 spin_lock(&iclog->ic_callback_lock);
2268 cb = iclog->ic_callback;
2270 iclog->ic_callback_tail = &(iclog->ic_callback);
2271 iclog->ic_callback = NULL;
2272 spin_unlock(&iclog->ic_callback_lock);
2274 /* perform callbacks in the order given */
2275 for (; cb; cb = cb_next) {
2276 cb_next = cb->cb_next;
2277 cb->cb_func(cb->cb_arg, aborted);
2279 spin_lock(&iclog->ic_callback_lock);
2280 cb = iclog->ic_callback;
2286 spin_lock(&log->l_icloglock);
2287 ASSERT(iclog->ic_callback == NULL);
2288 spin_unlock(&iclog->ic_callback_lock);
2289 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2290 iclog->ic_state = XLOG_STATE_DIRTY;
2293 * Transition from DIRTY to ACTIVE if applicable.
2294 * NOP if STATE_IOERROR.
2296 xlog_state_clean_log(log);
2298 /* wake up threads waiting in xfs_log_force() */
2299 sv_broadcast(&iclog->ic_force_wait);
2301 iclog = iclog->ic_next;
2302 } while (first_iclog != iclog);
2304 if (repeats > 5000) {
2305 flushcnt += repeats;
2307 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2308 "%s: possible infinite loop (%d iterations)",
2309 __func__, flushcnt);
2311 } while (!ioerrors && loopdidcallbacks);
2314 * make one last gasp attempt to see if iclogs are being left in
2318 if (funcdidcallbacks) {
2319 first_iclog = iclog = log->l_iclog;
2321 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2323 * Terminate the loop if iclogs are found in states
2324 * which will cause other threads to clean up iclogs.
2326 * SYNCING - i/o completion will go through logs
2327 * DONE_SYNC - interrupt thread should be waiting for
2329 * IOERROR - give up hope all ye who enter here
2331 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2332 iclog->ic_state == XLOG_STATE_SYNCING ||
2333 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2334 iclog->ic_state == XLOG_STATE_IOERROR )
2336 iclog = iclog->ic_next;
2337 } while (first_iclog != iclog);
2341 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2343 spin_unlock(&log->l_icloglock);
2346 sv_broadcast(&log->l_flush_wait);
2351 * Finish transitioning this iclog to the dirty state.
2353 * Make sure that we completely execute this routine only when this is
2354 * the last call to the iclog. There is a good chance that iclog flushes,
2355 * when we reach the end of the physical log, get turned into 2 separate
2356 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2357 * routine. By using the reference count bwritecnt, we guarantee that only
2358 * the second completion goes through.
2360 * Callbacks could take time, so they are done outside the scope of the
2361 * global state machine log lock.
2364 xlog_state_done_syncing(
2365 xlog_in_core_t *iclog,
2368 xlog_t *log = iclog->ic_log;
2370 spin_lock(&log->l_icloglock);
2372 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2373 iclog->ic_state == XLOG_STATE_IOERROR);
2374 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2375 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2379 * If we got an error, either on the first buffer, or in the case of
2380 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2381 * and none should ever be attempted to be written to disk
2384 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2385 if (--iclog->ic_bwritecnt == 1) {
2386 spin_unlock(&log->l_icloglock);
2389 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2393 * Someone could be sleeping prior to writing out the next
2394 * iclog buffer, we wake them all, one will get to do the
2395 * I/O, the others get to wait for the result.
2397 sv_broadcast(&iclog->ic_write_wait);
2398 spin_unlock(&log->l_icloglock);
2399 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2400 } /* xlog_state_done_syncing */
2404 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2405 * sleep. We wait on the flush queue on the head iclog as that should be
2406 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2407 * we will wait here and all new writes will sleep until a sync completes.
2409 * The in-core logs are used in a circular fashion. They are not used
2410 * out-of-order even when an iclog past the head is free.
2413 * * log_offset where xlog_write() can start writing into the in-core
2415 * * in-core log pointer to which xlog_write() should write.
2416 * * boolean indicating this is a continued write to an in-core log.
2417 * If this is the last write, then the in-core log's offset field
2418 * needs to be incremented, depending on the amount of data which
2422 xlog_state_get_iclog_space(xlog_t *log,
2424 xlog_in_core_t **iclogp,
2425 xlog_ticket_t *ticket,
2426 int *continued_write,
2430 xlog_rec_header_t *head;
2431 xlog_in_core_t *iclog;
2435 spin_lock(&log->l_icloglock);
2436 if (XLOG_FORCED_SHUTDOWN(log)) {
2437 spin_unlock(&log->l_icloglock);
2438 return XFS_ERROR(EIO);
2441 iclog = log->l_iclog;
2442 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2443 XFS_STATS_INC(xs_log_noiclogs);
2445 /* Wait for log writes to have flushed */
2446 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2450 head = &iclog->ic_header;
2452 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2453 log_offset = iclog->ic_offset;
2455 /* On the 1st write to an iclog, figure out lsn. This works
2456 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2457 * committing to. If the offset is set, that's how many blocks
2460 if (log_offset == 0) {
2461 ticket->t_curr_res -= log->l_iclog_hsize;
2462 xlog_tic_add_region(ticket,
2464 XLOG_REG_TYPE_LRHEADER);
2465 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2466 head->h_lsn = cpu_to_be64(
2467 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2468 ASSERT(log->l_curr_block >= 0);
2471 /* If there is enough room to write everything, then do it. Otherwise,
2472 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2473 * bit is on, so this will get flushed out. Don't update ic_offset
2474 * until you know exactly how many bytes get copied. Therefore, wait
2475 * until later to update ic_offset.
2477 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2478 * can fit into remaining data section.
2480 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2481 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2484 * If I'm the only one writing to this iclog, sync it to disk.
2485 * We need to do an atomic compare and decrement here to avoid
2486 * racing with concurrent atomic_dec_and_lock() calls in
2487 * xlog_state_release_iclog() when there is more than one
2488 * reference to the iclog.
2490 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2491 /* we are the only one */
2492 spin_unlock(&log->l_icloglock);
2493 error = xlog_state_release_iclog(log, iclog);
2497 spin_unlock(&log->l_icloglock);
2502 /* Do we have enough room to write the full amount in the remainder
2503 * of this iclog? Or must we continue a write on the next iclog and
2504 * mark this iclog as completely taken? In the case where we switch
2505 * iclogs (to mark it taken), this particular iclog will release/sync
2506 * to disk in xlog_write().
2508 if (len <= iclog->ic_size - iclog->ic_offset) {
2509 *continued_write = 0;
2510 iclog->ic_offset += len;
2512 *continued_write = 1;
2513 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2517 ASSERT(iclog->ic_offset <= iclog->ic_size);
2518 spin_unlock(&log->l_icloglock);
2520 *logoffsetp = log_offset;
2522 } /* xlog_state_get_iclog_space */
2525 * Atomically get the log space required for a log ticket.
2527 * Once a ticket gets put onto the reserveq, it will only return after
2528 * the needed reservation is satisfied.
2531 xlog_grant_log_space(xlog_t *log,
2542 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2543 panic("grant Recovery problem");
2546 /* Is there space or do we need to sleep? */
2547 spin_lock(&log->l_grant_lock);
2549 trace_xfs_log_grant_enter(log, tic);
2551 /* something is already sleeping; insert new transaction at end */
2552 if (log->l_reserve_headq) {
2553 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2555 trace_xfs_log_grant_sleep1(log, tic);
2558 * Gotta check this before going to sleep, while we're
2559 * holding the grant lock.
2561 if (XLOG_FORCED_SHUTDOWN(log))
2564 XFS_STATS_INC(xs_sleep_logspace);
2565 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2567 * If we got an error, and the filesystem is shutting down,
2568 * we'll catch it down below. So just continue...
2570 trace_xfs_log_grant_wake1(log, tic);
2571 spin_lock(&log->l_grant_lock);
2573 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2574 need_bytes = tic->t_unit_res*tic->t_ocnt;
2576 need_bytes = tic->t_unit_res;
2579 if (XLOG_FORCED_SHUTDOWN(log))
2582 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2583 log->l_grant_reserve_bytes);
2584 if (free_bytes < need_bytes) {
2585 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2586 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2588 trace_xfs_log_grant_sleep2(log, tic);
2590 spin_unlock(&log->l_grant_lock);
2591 xlog_grant_push_ail(log->l_mp, need_bytes);
2592 spin_lock(&log->l_grant_lock);
2594 XFS_STATS_INC(xs_sleep_logspace);
2595 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2597 spin_lock(&log->l_grant_lock);
2598 if (XLOG_FORCED_SHUTDOWN(log))
2601 trace_xfs_log_grant_wake2(log, tic);
2604 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2605 xlog_del_ticketq(&log->l_reserve_headq, tic);
2607 /* we've got enough space */
2608 xlog_grant_add_space(log, need_bytes);
2610 tail_lsn = log->l_tail_lsn;
2612 * Check to make sure the grant write head didn't just over lap the
2613 * tail. If the cycles are the same, we can't be overlapping.
2614 * Otherwise, make sure that the cycles differ by exactly one and
2615 * check the byte count.
2617 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2618 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2619 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2622 trace_xfs_log_grant_exit(log, tic);
2623 xlog_verify_grant_head(log, 1);
2624 spin_unlock(&log->l_grant_lock);
2628 if (tic->t_flags & XLOG_TIC_IN_Q)
2629 xlog_del_ticketq(&log->l_reserve_headq, tic);
2631 trace_xfs_log_grant_error(log, tic);
2634 * If we are failing, make sure the ticket doesn't have any
2635 * current reservations. We don't want to add this back when
2636 * the ticket/transaction gets cancelled.
2638 tic->t_curr_res = 0;
2639 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2640 spin_unlock(&log->l_grant_lock);
2641 return XFS_ERROR(EIO);
2642 } /* xlog_grant_log_space */
2646 * Replenish the byte reservation required by moving the grant write head.
2651 xlog_regrant_write_log_space(xlog_t *log,
2654 int free_bytes, need_bytes;
2655 xlog_ticket_t *ntic;
2660 tic->t_curr_res = tic->t_unit_res;
2661 xlog_tic_reset_res(tic);
2667 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2668 panic("regrant Recovery problem");
2671 spin_lock(&log->l_grant_lock);
2673 trace_xfs_log_regrant_write_enter(log, tic);
2675 if (XLOG_FORCED_SHUTDOWN(log))
2678 /* If there are other waiters on the queue then give them a
2679 * chance at logspace before us. Wake up the first waiters,
2680 * if we do not wake up all the waiters then go to sleep waiting
2681 * for more free space, otherwise try to get some space for
2684 need_bytes = tic->t_unit_res;
2685 if ((ntic = log->l_write_headq)) {
2686 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2687 log->l_grant_write_bytes);
2689 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2691 if (free_bytes < ntic->t_unit_res)
2693 free_bytes -= ntic->t_unit_res;
2694 sv_signal(&ntic->t_wait);
2695 ntic = ntic->t_next;
2696 } while (ntic != log->l_write_headq);
2698 if (ntic != log->l_write_headq) {
2699 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2700 xlog_ins_ticketq(&log->l_write_headq, tic);
2702 trace_xfs_log_regrant_write_sleep1(log, tic);
2704 spin_unlock(&log->l_grant_lock);
2705 xlog_grant_push_ail(log->l_mp, need_bytes);
2706 spin_lock(&log->l_grant_lock);
2708 XFS_STATS_INC(xs_sleep_logspace);
2709 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2710 &log->l_grant_lock, s);
2712 /* If we're shutting down, this tic is already
2714 spin_lock(&log->l_grant_lock);
2715 if (XLOG_FORCED_SHUTDOWN(log))
2718 trace_xfs_log_regrant_write_wake1(log, tic);
2723 if (XLOG_FORCED_SHUTDOWN(log))
2726 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2727 log->l_grant_write_bytes);
2728 if (free_bytes < need_bytes) {
2729 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2730 xlog_ins_ticketq(&log->l_write_headq, tic);
2731 spin_unlock(&log->l_grant_lock);
2732 xlog_grant_push_ail(log->l_mp, need_bytes);
2733 spin_lock(&log->l_grant_lock);
2735 XFS_STATS_INC(xs_sleep_logspace);
2736 trace_xfs_log_regrant_write_sleep2(log, tic);
2738 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2740 /* If we're shutting down, this tic is already off the queue */
2741 spin_lock(&log->l_grant_lock);
2742 if (XLOG_FORCED_SHUTDOWN(log))
2745 trace_xfs_log_regrant_write_wake2(log, tic);
2747 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2748 xlog_del_ticketq(&log->l_write_headq, tic);
2750 /* we've got enough space */
2751 xlog_grant_add_space_write(log, need_bytes);
2753 tail_lsn = log->l_tail_lsn;
2754 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2755 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2756 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2760 trace_xfs_log_regrant_write_exit(log, tic);
2762 xlog_verify_grant_head(log, 1);
2763 spin_unlock(&log->l_grant_lock);
2768 if (tic->t_flags & XLOG_TIC_IN_Q)
2769 xlog_del_ticketq(&log->l_reserve_headq, tic);
2771 trace_xfs_log_regrant_write_error(log, tic);
2774 * If we are failing, make sure the ticket doesn't have any
2775 * current reservations. We don't want to add this back when
2776 * the ticket/transaction gets cancelled.
2778 tic->t_curr_res = 0;
2779 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2780 spin_unlock(&log->l_grant_lock);
2781 return XFS_ERROR(EIO);
2782 } /* xlog_regrant_write_log_space */
2785 /* The first cnt-1 times through here we don't need to
2786 * move the grant write head because the permanent
2787 * reservation has reserved cnt times the unit amount.
2788 * Release part of current permanent unit reservation and
2789 * reset current reservation to be one units worth. Also
2790 * move grant reservation head forward.
2793 xlog_regrant_reserve_log_space(xlog_t *log,
2794 xlog_ticket_t *ticket)
2796 trace_xfs_log_regrant_reserve_enter(log, ticket);
2798 if (ticket->t_cnt > 0)
2801 spin_lock(&log->l_grant_lock);
2802 xlog_grant_sub_space(log, ticket->t_curr_res);
2803 ticket->t_curr_res = ticket->t_unit_res;
2804 xlog_tic_reset_res(ticket);
2806 trace_xfs_log_regrant_reserve_sub(log, ticket);
2808 xlog_verify_grant_head(log, 1);
2810 /* just return if we still have some of the pre-reserved space */
2811 if (ticket->t_cnt > 0) {
2812 spin_unlock(&log->l_grant_lock);
2816 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2818 trace_xfs_log_regrant_reserve_exit(log, ticket);
2820 xlog_verify_grant_head(log, 0);
2821 spin_unlock(&log->l_grant_lock);
2822 ticket->t_curr_res = ticket->t_unit_res;
2823 xlog_tic_reset_res(ticket);
2824 } /* xlog_regrant_reserve_log_space */
2828 * Give back the space left from a reservation.
2830 * All the information we need to make a correct determination of space left
2831 * is present. For non-permanent reservations, things are quite easy. The
2832 * count should have been decremented to zero. We only need to deal with the
2833 * space remaining in the current reservation part of the ticket. If the
2834 * ticket contains a permanent reservation, there may be left over space which
2835 * needs to be released. A count of N means that N-1 refills of the current
2836 * reservation can be done before we need to ask for more space. The first
2837 * one goes to fill up the first current reservation. Once we run out of
2838 * space, the count will stay at zero and the only space remaining will be
2839 * in the current reservation field.
2842 xlog_ungrant_log_space(xlog_t *log,
2843 xlog_ticket_t *ticket)
2845 if (ticket->t_cnt > 0)
2848 spin_lock(&log->l_grant_lock);
2849 trace_xfs_log_ungrant_enter(log, ticket);
2851 xlog_grant_sub_space(log, ticket->t_curr_res);
2853 trace_xfs_log_ungrant_sub(log, ticket);
2855 /* If this is a permanent reservation ticket, we may be able to free
2856 * up more space based on the remaining count.
2858 if (ticket->t_cnt > 0) {
2859 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2860 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2863 trace_xfs_log_ungrant_exit(log, ticket);
2865 xlog_verify_grant_head(log, 1);
2866 spin_unlock(&log->l_grant_lock);
2867 xfs_log_move_tail(log->l_mp, 1);
2868 } /* xlog_ungrant_log_space */
2872 * Flush iclog to disk if this is the last reference to the given iclog and
2873 * the WANT_SYNC bit is set.
2875 * When this function is entered, the iclog is not necessarily in the
2876 * WANT_SYNC state. It may be sitting around waiting to get filled.
2881 xlog_state_release_iclog(
2883 xlog_in_core_t *iclog)
2885 int sync = 0; /* do we sync? */
2887 if (iclog->ic_state & XLOG_STATE_IOERROR)
2888 return XFS_ERROR(EIO);
2890 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2891 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2894 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2895 spin_unlock(&log->l_icloglock);
2896 return XFS_ERROR(EIO);
2898 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2899 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2901 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2902 /* update tail before writing to iclog */
2903 xlog_assign_tail_lsn(log->l_mp);
2905 iclog->ic_state = XLOG_STATE_SYNCING;
2906 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2907 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2908 /* cycle incremented when incrementing curr_block */
2910 spin_unlock(&log->l_icloglock);
2913 * We let the log lock go, so it's possible that we hit a log I/O
2914 * error or some other SHUTDOWN condition that marks the iclog
2915 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2916 * this iclog has consistent data, so we ignore IOERROR
2917 * flags after this point.
2920 return xlog_sync(log, iclog);
2922 } /* xlog_state_release_iclog */
2926 * This routine will mark the current iclog in the ring as WANT_SYNC
2927 * and move the current iclog pointer to the next iclog in the ring.
2928 * When this routine is called from xlog_state_get_iclog_space(), the
2929 * exact size of the iclog has not yet been determined. All we know is
2930 * that every data block. We have run out of space in this log record.
2933 xlog_state_switch_iclogs(xlog_t *log,
2934 xlog_in_core_t *iclog,
2937 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2939 eventual_size = iclog->ic_offset;
2940 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2941 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2942 log->l_prev_block = log->l_curr_block;
2943 log->l_prev_cycle = log->l_curr_cycle;
2945 /* roll log?: ic_offset changed later */
2946 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2948 /* Round up to next log-sunit */
2949 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2950 log->l_mp->m_sb.sb_logsunit > 1) {
2951 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2952 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2955 if (log->l_curr_block >= log->l_logBBsize) {
2956 log->l_curr_cycle++;
2957 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2958 log->l_curr_cycle++;
2959 log->l_curr_block -= log->l_logBBsize;
2960 ASSERT(log->l_curr_block >= 0);
2962 ASSERT(iclog == log->l_iclog);
2963 log->l_iclog = iclog->ic_next;
2964 } /* xlog_state_switch_iclogs */
2967 * Write out all data in the in-core log as of this exact moment in time.
2969 * Data may be written to the in-core log during this call. However,
2970 * we don't guarantee this data will be written out. A change from past
2971 * implementation means this routine will *not* write out zero length LRs.
2973 * Basically, we try and perform an intelligent scan of the in-core logs.
2974 * If we determine there is no flushable data, we just return. There is no
2975 * flushable data if:
2977 * 1. the current iclog is active and has no data; the previous iclog
2978 * is in the active or dirty state.
2979 * 2. the current iclog is drity, and the previous iclog is in the
2980 * active or dirty state.
2984 * 1. the current iclog is not in the active nor dirty state.
2985 * 2. the current iclog dirty, and the previous iclog is not in the
2986 * active nor dirty state.
2987 * 3. the current iclog is active, and there is another thread writing
2988 * to this particular iclog.
2989 * 4. a) the current iclog is active and has no other writers
2990 * b) when we return from flushing out this iclog, it is still
2991 * not in the active nor dirty state.
2995 struct xfs_mount *mp,
2999 struct log *log = mp->m_log;
3000 struct xlog_in_core *iclog;
3003 XFS_STATS_INC(xs_log_force);
3006 xlog_cil_force(log);
3008 spin_lock(&log->l_icloglock);
3010 iclog = log->l_iclog;
3011 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3012 spin_unlock(&log->l_icloglock);
3013 return XFS_ERROR(EIO);
3016 /* If the head iclog is not active nor dirty, we just attach
3017 * ourselves to the head and go to sleep.
3019 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3020 iclog->ic_state == XLOG_STATE_DIRTY) {
3022 * If the head is dirty or (active and empty), then
3023 * we need to look at the previous iclog. If the previous
3024 * iclog is active or dirty we are done. There is nothing
3025 * to sync out. Otherwise, we attach ourselves to the
3026 * previous iclog and go to sleep.
3028 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3029 (atomic_read(&iclog->ic_refcnt) == 0
3030 && iclog->ic_offset == 0)) {
3031 iclog = iclog->ic_prev;
3032 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3033 iclog->ic_state == XLOG_STATE_DIRTY)
3038 if (atomic_read(&iclog->ic_refcnt) == 0) {
3039 /* We are the only one with access to this
3040 * iclog. Flush it out now. There should
3041 * be a roundoff of zero to show that someone
3042 * has already taken care of the roundoff from
3043 * the previous sync.
3045 atomic_inc(&iclog->ic_refcnt);
3046 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3047 xlog_state_switch_iclogs(log, iclog, 0);
3048 spin_unlock(&log->l_icloglock);
3050 if (xlog_state_release_iclog(log, iclog))
3051 return XFS_ERROR(EIO);
3055 spin_lock(&log->l_icloglock);
3056 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3057 iclog->ic_state != XLOG_STATE_DIRTY)
3062 /* Someone else is writing to this iclog.
3063 * Use its call to flush out the data. However,
3064 * the other thread may not force out this LR,
3065 * so we mark it WANT_SYNC.
3067 xlog_state_switch_iclogs(log, iclog, 0);
3073 /* By the time we come around again, the iclog could've been filled
3074 * which would give it another lsn. If we have a new lsn, just
3075 * return because the relevant data has been flushed.
3078 if (flags & XFS_LOG_SYNC) {
3080 * We must check if we're shutting down here, before
3081 * we wait, while we're holding the l_icloglock.
3082 * Then we check again after waking up, in case our
3083 * sleep was disturbed by a bad news.
3085 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3086 spin_unlock(&log->l_icloglock);
3087 return XFS_ERROR(EIO);
3089 XFS_STATS_INC(xs_log_force_sleep);
3090 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3092 * No need to grab the log lock here since we're
3093 * only deciding whether or not to return EIO
3094 * and the memory read should be atomic.
3096 if (iclog->ic_state & XLOG_STATE_IOERROR)
3097 return XFS_ERROR(EIO);
3103 spin_unlock(&log->l_icloglock);
3109 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3110 * about errors or whether the log was flushed or not. This is the normal
3111 * interface to use when trying to unpin items or move the log forward.
3120 error = _xfs_log_force(mp, flags, NULL);
3122 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3123 "error %d returned.", error);
3128 * Force the in-core log to disk for a specific LSN.
3130 * Find in-core log with lsn.
3131 * If it is in the DIRTY state, just return.
3132 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3133 * state and go to sleep or return.
3134 * If it is in any other state, go to sleep or return.
3136 * Synchronous forces are implemented with a signal variable. All callers
3137 * to force a given lsn to disk will wait on a the sv attached to the
3138 * specific in-core log. When given in-core log finally completes its
3139 * write to disk, that thread will wake up all threads waiting on the
3144 struct xfs_mount *mp,
3149 struct log *log = mp->m_log;
3150 struct xlog_in_core *iclog;
3151 int already_slept = 0;
3155 XFS_STATS_INC(xs_log_force);
3158 lsn = xlog_cil_force_lsn(log, lsn);
3159 if (lsn == NULLCOMMITLSN)
3164 spin_lock(&log->l_icloglock);
3165 iclog = log->l_iclog;
3166 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3167 spin_unlock(&log->l_icloglock);
3168 return XFS_ERROR(EIO);
3172 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3173 iclog = iclog->ic_next;
3177 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3178 spin_unlock(&log->l_icloglock);
3182 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3184 * We sleep here if we haven't already slept (e.g.
3185 * this is the first time we've looked at the correct
3186 * iclog buf) and the buffer before us is going to
3187 * be sync'ed. The reason for this is that if we
3188 * are doing sync transactions here, by waiting for
3189 * the previous I/O to complete, we can allow a few
3190 * more transactions into this iclog before we close
3193 * Otherwise, we mark the buffer WANT_SYNC, and bump
3194 * up the refcnt so we can release the log (which
3195 * drops the ref count). The state switch keeps new
3196 * transaction commits from using this buffer. When
3197 * the current commits finish writing into the buffer,
3198 * the refcount will drop to zero and the buffer will
3201 if (!already_slept &&
3202 (iclog->ic_prev->ic_state &
3203 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3204 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3206 XFS_STATS_INC(xs_log_force_sleep);
3208 sv_wait(&iclog->ic_prev->ic_write_wait,
3209 PSWP, &log->l_icloglock, s);
3215 atomic_inc(&iclog->ic_refcnt);
3216 xlog_state_switch_iclogs(log, iclog, 0);
3217 spin_unlock(&log->l_icloglock);
3218 if (xlog_state_release_iclog(log, iclog))
3219 return XFS_ERROR(EIO);
3222 spin_lock(&log->l_icloglock);
3225 if ((flags & XFS_LOG_SYNC) && /* sleep */
3227 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3229 * Don't wait on completion if we know that we've
3230 * gotten a log write error.
3232 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3233 spin_unlock(&log->l_icloglock);
3234 return XFS_ERROR(EIO);
3236 XFS_STATS_INC(xs_log_force_sleep);
3237 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3239 * No need to grab the log lock here since we're
3240 * only deciding whether or not to return EIO
3241 * and the memory read should be atomic.
3243 if (iclog->ic_state & XLOG_STATE_IOERROR)
3244 return XFS_ERROR(EIO);
3248 } else { /* just return */
3249 spin_unlock(&log->l_icloglock);
3253 } while (iclog != log->l_iclog);
3255 spin_unlock(&log->l_icloglock);
3260 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3261 * about errors or whether the log was flushed or not. This is the normal
3262 * interface to use when trying to unpin items or move the log forward.
3272 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3274 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3275 "error %d returned.", error);
3280 * Called when we want to mark the current iclog as being ready to sync to
3284 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3286 assert_spin_locked(&log->l_icloglock);
3288 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3289 xlog_state_switch_iclogs(log, iclog, 0);
3291 ASSERT(iclog->ic_state &
3292 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3297 /*****************************************************************************
3301 *****************************************************************************
3305 * Free a used ticket when its refcount falls to zero.
3309 xlog_ticket_t *ticket)
3311 ASSERT(atomic_read(&ticket->t_ref) > 0);
3312 if (atomic_dec_and_test(&ticket->t_ref)) {
3313 sv_destroy(&ticket->t_wait);
3314 kmem_zone_free(xfs_log_ticket_zone, ticket);
3320 xlog_ticket_t *ticket)
3322 ASSERT(atomic_read(&ticket->t_ref) > 0);
3323 atomic_inc(&ticket->t_ref);
3328 xfs_log_get_trans_ident(
3329 struct xfs_trans *tp)
3331 return tp->t_ticket->t_tid;
3335 * Allocate and initialise a new log ticket.
3346 struct xlog_ticket *tic;
3350 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3355 * Permanent reservations have up to 'cnt'-1 active log operations
3356 * in the log. A unit in this case is the amount of space for one
3357 * of these log operations. Normal reservations have a cnt of 1
3358 * and their unit amount is the total amount of space required.
3360 * The following lines of code account for non-transaction data
3361 * which occupy space in the on-disk log.
3363 * Normal form of a transaction is:
3364 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3365 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3367 * We need to account for all the leadup data and trailer data
3368 * around the transaction data.
3369 * And then we need to account for the worst case in terms of using
3371 * The worst case will happen if:
3372 * - the placement of the transaction happens to be such that the
3373 * roundoff is at its maximum
3374 * - the transaction data is synced before the commit record is synced
3375 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3376 * Therefore the commit record is in its own Log Record.
3377 * This can happen as the commit record is called with its
3378 * own region to xlog_write().
3379 * This then means that in the worst case, roundoff can happen for
3380 * the commit-rec as well.
3381 * The commit-rec is smaller than padding in this scenario and so it is
3382 * not added separately.
3385 /* for trans header */
3386 unit_bytes += sizeof(xlog_op_header_t);
3387 unit_bytes += sizeof(xfs_trans_header_t);
3390 unit_bytes += sizeof(xlog_op_header_t);
3393 * for LR headers - the space for data in an iclog is the size minus
3394 * the space used for the headers. If we use the iclog size, then we
3395 * undercalculate the number of headers required.
3397 * Furthermore - the addition of op headers for split-recs might
3398 * increase the space required enough to require more log and op
3399 * headers, so take that into account too.
3401 * IMPORTANT: This reservation makes the assumption that if this
3402 * transaction is the first in an iclog and hence has the LR headers
3403 * accounted to it, then the remaining space in the iclog is
3404 * exclusively for this transaction. i.e. if the transaction is larger
3405 * than the iclog, it will be the only thing in that iclog.
3406 * Fundamentally, this means we must pass the entire log vector to
3407 * xlog_write to guarantee this.
3409 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3410 num_headers = howmany(unit_bytes, iclog_space);
3412 /* for split-recs - ophdrs added when data split over LRs */
3413 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3415 /* add extra header reservations if we overrun */
3416 while (!num_headers ||
3417 howmany(unit_bytes, iclog_space) > num_headers) {
3418 unit_bytes += sizeof(xlog_op_header_t);
3421 unit_bytes += log->l_iclog_hsize * num_headers;
3423 /* for commit-rec LR header - note: padding will subsume the ophdr */
3424 unit_bytes += log->l_iclog_hsize;
3426 /* for roundoff padding for transaction data and one for commit record */
3427 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3428 log->l_mp->m_sb.sb_logsunit > 1) {
3429 /* log su roundoff */
3430 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3433 unit_bytes += 2*BBSIZE;
3436 atomic_set(&tic->t_ref, 1);
3437 tic->t_unit_res = unit_bytes;
3438 tic->t_curr_res = unit_bytes;
3441 tic->t_tid = random32();
3442 tic->t_clientid = client;
3443 tic->t_flags = XLOG_TIC_INITED;
3444 tic->t_trans_type = 0;
3445 if (xflags & XFS_LOG_PERM_RESERV)
3446 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3447 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3449 xlog_tic_reset_res(tic);
3455 /******************************************************************************
3457 * Log debug routines
3459 ******************************************************************************
3463 * Make sure that the destination ptr is within the valid data region of
3464 * one of the iclogs. This uses backup pointers stored in a different
3465 * part of the log in case we trash the log structure.
3468 xlog_verify_dest_ptr(
3475 for (i = 0; i < log->l_iclog_bufs; i++) {
3476 if (ptr >= log->l_iclog_bak[i] &&
3477 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3482 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3486 xlog_verify_grant_head(xlog_t *log, int equals)
3488 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3490 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3492 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3494 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3495 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3497 } /* xlog_verify_grant_head */
3499 /* check if it will fit */
3501 xlog_verify_tail_lsn(xlog_t *log,
3502 xlog_in_core_t *iclog,
3507 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3509 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3510 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3511 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3513 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3515 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3516 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3518 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3519 if (blocks < BTOBB(iclog->ic_offset) + 1)
3520 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3522 } /* xlog_verify_tail_lsn */
3525 * Perform a number of checks on the iclog before writing to disk.
3527 * 1. Make sure the iclogs are still circular
3528 * 2. Make sure we have a good magic number
3529 * 3. Make sure we don't have magic numbers in the data
3530 * 4. Check fields of each log operation header for:
3531 * A. Valid client identifier
3532 * B. tid ptr value falls in valid ptr space (user space code)
3533 * C. Length in log record header is correct according to the
3534 * individual operation headers within record.
3535 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3536 * log, check the preceding blocks of the physical log to make sure all
3537 * the cycle numbers agree with the current cycle number.
3540 xlog_verify_iclog(xlog_t *log,
3541 xlog_in_core_t *iclog,
3545 xlog_op_header_t *ophead;
3546 xlog_in_core_t *icptr;
3547 xlog_in_core_2_t *xhdr;
3549 xfs_caddr_t base_ptr;
3550 __psint_t field_offset;
3552 int len, i, j, k, op_len;
3555 /* check validity of iclog pointers */
3556 spin_lock(&log->l_icloglock);
3557 icptr = log->l_iclog;
3558 for (i=0; i < log->l_iclog_bufs; i++) {
3560 xlog_panic("xlog_verify_iclog: invalid ptr");
3561 icptr = icptr->ic_next;
3563 if (icptr != log->l_iclog)
3564 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3565 spin_unlock(&log->l_icloglock);
3567 /* check log magic numbers */
3568 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3569 xlog_panic("xlog_verify_iclog: invalid magic num");
3571 ptr = (xfs_caddr_t) &iclog->ic_header;
3572 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3574 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3575 xlog_panic("xlog_verify_iclog: unexpected magic num");
3579 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3580 ptr = iclog->ic_datap;
3582 ophead = (xlog_op_header_t *)ptr;
3583 xhdr = iclog->ic_data;
3584 for (i = 0; i < len; i++) {
3585 ophead = (xlog_op_header_t *)ptr;
3587 /* clientid is only 1 byte */
3588 field_offset = (__psint_t)
3589 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3590 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3591 clientid = ophead->oh_clientid;
3593 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3594 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3595 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3596 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3597 clientid = xlog_get_client_id(
3598 xhdr[j].hic_xheader.xh_cycle_data[k]);
3600 clientid = xlog_get_client_id(
3601 iclog->ic_header.h_cycle_data[idx]);
3604 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3605 cmn_err(CE_WARN, "xlog_verify_iclog: "
3606 "invalid clientid %d op 0x%p offset 0x%lx",
3607 clientid, ophead, (unsigned long)field_offset);
3610 field_offset = (__psint_t)
3611 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3612 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3613 op_len = be32_to_cpu(ophead->oh_len);
3615 idx = BTOBBT((__psint_t)&ophead->oh_len -
3616 (__psint_t)iclog->ic_datap);
3617 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3618 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3619 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3620 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3622 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3625 ptr += sizeof(xlog_op_header_t) + op_len;
3627 } /* xlog_verify_iclog */
3631 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3637 xlog_in_core_t *iclog, *ic;
3639 iclog = log->l_iclog;
3640 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3642 * Mark all the incore logs IOERROR.
3643 * From now on, no log flushes will result.
3647 ic->ic_state = XLOG_STATE_IOERROR;
3649 } while (ic != iclog);
3653 * Return non-zero, if state transition has already happened.
3659 * This is called from xfs_force_shutdown, when we're forcibly
3660 * shutting down the filesystem, typically because of an IO error.
3661 * Our main objectives here are to make sure that:
3662 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3663 * parties to find out, 'atomically'.
3664 * b. those who're sleeping on log reservations, pinned objects and
3665 * other resources get woken up, and be told the bad news.
3666 * c. nothing new gets queued up after (a) and (b) are done.
3667 * d. if !logerror, flush the iclogs to disk, then seal them off
3670 * Note: for delayed logging the !logerror case needs to flush the regions
3671 * held in memory out to the iclogs before flushing them to disk. This needs
3672 * to be done before the log is marked as shutdown, otherwise the flush to the
3676 xfs_log_force_umount(
3677 struct xfs_mount *mp,
3687 * If this happens during log recovery, don't worry about
3688 * locking; the log isn't open for business yet.
3691 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3692 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3694 XFS_BUF_DONE(mp->m_sb_bp);
3699 * Somebody could've already done the hard work for us.
3700 * No need to get locks for this.
3702 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3703 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3709 * Flush the in memory commit item list before marking the log as
3710 * being shut down. We need to do it in this order to ensure all the
3711 * completed transactions are flushed to disk with the xfs_log_force()
3714 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3715 xlog_cil_force(log);
3718 * We must hold both the GRANT lock and the LOG lock,
3719 * before we mark the filesystem SHUTDOWN and wake
3720 * everybody up to tell the bad news.
3722 spin_lock(&log->l_icloglock);
3723 spin_lock(&log->l_grant_lock);
3724 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3726 XFS_BUF_DONE(mp->m_sb_bp);
3729 * This flag is sort of redundant because of the mount flag, but
3730 * it's good to maintain the separation between the log and the rest
3733 log->l_flags |= XLOG_IO_ERROR;
3736 * If we hit a log error, we want to mark all the iclogs IOERROR
3737 * while we're still holding the loglock.
3740 retval = xlog_state_ioerror(log);
3741 spin_unlock(&log->l_icloglock);
3744 * We don't want anybody waiting for log reservations
3745 * after this. That means we have to wake up everybody
3746 * queued up on reserve_headq as well as write_headq.
3747 * In addition, we make sure in xlog_{re}grant_log_space
3748 * that we don't enqueue anything once the SHUTDOWN flag
3749 * is set, and this action is protected by the GRANTLOCK.
3751 if ((tic = log->l_reserve_headq)) {
3753 sv_signal(&tic->t_wait);
3755 } while (tic != log->l_reserve_headq);
3758 if ((tic = log->l_write_headq)) {
3760 sv_signal(&tic->t_wait);
3762 } while (tic != log->l_write_headq);
3764 spin_unlock(&log->l_grant_lock);
3766 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3769 * Force the incore logs to disk before shutting the
3770 * log down completely.
3772 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3774 spin_lock(&log->l_icloglock);
3775 retval = xlog_state_ioerror(log);
3776 spin_unlock(&log->l_icloglock);
3779 * Wake up everybody waiting on xfs_log_force.
3780 * Callback all log item committed functions as if the
3781 * log writes were completed.
3783 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3785 #ifdef XFSERRORDEBUG
3787 xlog_in_core_t *iclog;
3789 spin_lock(&log->l_icloglock);
3790 iclog = log->l_iclog;
3792 ASSERT(iclog->ic_callback == 0);
3793 iclog = iclog->ic_next;
3794 } while (iclog != log->l_iclog);
3795 spin_unlock(&log->l_icloglock);
3798 /* return non-zero if log IOERROR transition had already happened */
3803 xlog_iclogs_empty(xlog_t *log)
3805 xlog_in_core_t *iclog;
3807 iclog = log->l_iclog;
3809 /* endianness does not matter here, zero is zero in
3812 if (iclog->ic_header.h_num_logops)
3814 iclog = iclog->ic_next;
3815 } while (iclog != log->l_iclog);