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"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
43 #include "xfs_trace.h"
45 kmem_zone_t *xfs_log_ticket_zone;
47 /* Local miscellaneous function prototypes */
48 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
49 xlog_in_core_t **, xfs_lsn_t *);
50 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
51 xfs_buftarg_t *log_target,
52 xfs_daddr_t blk_offset,
54 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
55 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
56 STATIC void xlog_dealloc_log(xlog_t *log);
57 STATIC int xlog_write(struct log *log, struct xfs_log_vec *log_vector,
58 struct xlog_ticket *tic, xfs_lsn_t *start_lsn,
59 xlog_in_core_t **commit_iclog, uint flags);
61 /* local state machine functions */
62 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
63 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
64 STATIC int xlog_state_get_iclog_space(xlog_t *log,
66 xlog_in_core_t **iclog,
67 xlog_ticket_t *ticket,
70 STATIC int xlog_state_release_iclog(xlog_t *log,
71 xlog_in_core_t *iclog);
72 STATIC void xlog_state_switch_iclogs(xlog_t *log,
73 xlog_in_core_t *iclog,
75 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
77 /* local functions to manipulate grant head */
78 STATIC int xlog_grant_log_space(xlog_t *log,
80 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
82 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
83 xlog_ticket_t *ticket);
84 STATIC int xlog_regrant_write_log_space(xlog_t *log,
85 xlog_ticket_t *ticket);
86 STATIC void xlog_ungrant_log_space(xlog_t *log,
87 xlog_ticket_t *ticket);
90 /* local ticket functions */
91 STATIC xlog_ticket_t *xlog_ticket_alloc(xlog_t *log, int unit_bytes, int count,
92 char clientid, uint flags,
96 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
97 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
98 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
99 int count, boolean_t syncing);
100 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
103 #define xlog_verify_dest_ptr(a,b)
104 #define xlog_verify_grant_head(a,b)
105 #define xlog_verify_iclog(a,b,c,d)
106 #define xlog_verify_tail_lsn(a,b,c)
109 STATIC int xlog_iclogs_empty(xlog_t *log);
113 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
117 tic->t_prev = (*qp)->t_prev;
118 (*qp)->t_prev->t_next = tic;
121 tic->t_prev = tic->t_next = tic;
125 tic->t_flags |= XLOG_TIC_IN_Q;
129 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
131 if (tic == tic->t_next) {
135 tic->t_next->t_prev = tic->t_prev;
136 tic->t_prev->t_next = tic->t_next;
139 tic->t_next = tic->t_prev = NULL;
140 tic->t_flags &= ~XLOG_TIC_IN_Q;
144 xlog_grant_sub_space(struct log *log, int bytes)
146 log->l_grant_write_bytes -= bytes;
147 if (log->l_grant_write_bytes < 0) {
148 log->l_grant_write_bytes += log->l_logsize;
149 log->l_grant_write_cycle--;
152 log->l_grant_reserve_bytes -= bytes;
153 if ((log)->l_grant_reserve_bytes < 0) {
154 log->l_grant_reserve_bytes += log->l_logsize;
155 log->l_grant_reserve_cycle--;
161 xlog_grant_add_space_write(struct log *log, int bytes)
163 int tmp = log->l_logsize - log->l_grant_write_bytes;
165 log->l_grant_write_bytes += bytes;
167 log->l_grant_write_cycle++;
168 log->l_grant_write_bytes = bytes - tmp;
173 xlog_grant_add_space_reserve(struct log *log, int bytes)
175 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
177 log->l_grant_reserve_bytes += bytes;
179 log->l_grant_reserve_cycle++;
180 log->l_grant_reserve_bytes = bytes - tmp;
185 xlog_grant_add_space(struct log *log, int bytes)
187 xlog_grant_add_space_write(log, bytes);
188 xlog_grant_add_space_reserve(log, bytes);
192 xlog_tic_reset_res(xlog_ticket_t *tic)
195 tic->t_res_arr_sum = 0;
196 tic->t_res_num_ophdrs = 0;
200 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
202 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
203 /* add to overflow and start again */
204 tic->t_res_o_flow += tic->t_res_arr_sum;
206 tic->t_res_arr_sum = 0;
209 tic->t_res_arr[tic->t_res_num].r_len = len;
210 tic->t_res_arr[tic->t_res_num].r_type = type;
211 tic->t_res_arr_sum += len;
218 * 1. currblock field gets updated at startup and after in-core logs
219 * marked as with WANT_SYNC.
223 * This routine is called when a user of a log manager ticket is done with
224 * the reservation. If the ticket was ever used, then a commit record for
225 * the associated transaction is written out as a log operation header with
226 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
227 * a given ticket. If the ticket was one with a permanent reservation, then
228 * a few operations are done differently. Permanent reservation tickets by
229 * default don't release the reservation. They just commit the current
230 * transaction with the belief that the reservation is still needed. A flag
231 * must be passed in before permanent reservations are actually released.
232 * When these type of tickets are not released, they need to be set into
233 * the inited state again. By doing this, a start record will be written
234 * out when the next write occurs.
238 struct xfs_mount *mp,
239 struct xlog_ticket *ticket,
240 struct xlog_in_core **iclog,
243 struct log *log = mp->m_log;
246 if (XLOG_FORCED_SHUTDOWN(log) ||
248 * If nothing was ever written, don't write out commit record.
249 * If we get an error, just continue and give back the log ticket.
251 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
252 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
253 lsn = (xfs_lsn_t) -1;
254 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
255 flags |= XFS_LOG_REL_PERM_RESERV;
260 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
261 (flags & XFS_LOG_REL_PERM_RESERV)) {
262 trace_xfs_log_done_nonperm(log, ticket);
265 * Release ticket if not permanent reservation or a specific
266 * request has been made to release a permanent reservation.
268 xlog_ungrant_log_space(log, ticket);
269 xfs_log_ticket_put(ticket);
271 trace_xfs_log_done_perm(log, ticket);
273 xlog_regrant_reserve_log_space(log, ticket);
274 /* If this ticket was a permanent reservation and we aren't
275 * trying to release it, reset the inited flags; so next time
276 * we write, a start record will be written out.
278 ticket->t_flags |= XLOG_TIC_INITED;
285 * Attaches a new iclog I/O completion callback routine during
286 * transaction commit. If the log is in error state, a non-zero
287 * return code is handed back and the caller is responsible for
288 * executing the callback at an appropriate time.
292 struct xfs_mount *mp,
293 struct xlog_in_core *iclog,
294 xfs_log_callback_t *cb)
298 spin_lock(&iclog->ic_callback_lock);
299 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
301 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
302 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
304 *(iclog->ic_callback_tail) = cb;
305 iclog->ic_callback_tail = &(cb->cb_next);
307 spin_unlock(&iclog->ic_callback_lock);
312 xfs_log_release_iclog(
313 struct xfs_mount *mp,
314 struct xlog_in_core *iclog)
316 if (xlog_state_release_iclog(mp->m_log, iclog)) {
317 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
325 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
326 * to the reservation.
327 * 2. Potentially, push buffers at tail of log to disk.
329 * Each reservation is going to reserve extra space for a log record header.
330 * When writes happen to the on-disk log, we don't subtract the length of the
331 * log record header from any reservation. By wasting space in each
332 * reservation, we prevent over allocation problems.
336 struct xfs_mount *mp,
339 struct xlog_ticket **ticket,
344 struct log *log = mp->m_log;
345 struct xlog_ticket *internal_ticket;
348 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
349 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
351 if (XLOG_FORCED_SHUTDOWN(log))
352 return XFS_ERROR(EIO);
354 XFS_STATS_INC(xs_try_logspace);
357 if (*ticket != NULL) {
358 ASSERT(flags & XFS_LOG_PERM_RESERV);
359 internal_ticket = *ticket;
362 * this is a new transaction on the ticket, so we need to
363 * change the transaction ID so that the next transaction has a
364 * different TID in the log. Just add one to the existing tid
365 * so that we can see chains of rolling transactions in the log
368 internal_ticket->t_tid++;
370 trace_xfs_log_reserve(log, internal_ticket);
372 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
373 retval = xlog_regrant_write_log_space(log, internal_ticket);
375 /* may sleep if need to allocate more tickets */
376 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
378 KM_SLEEP|KM_MAYFAIL);
379 if (!internal_ticket)
380 return XFS_ERROR(ENOMEM);
381 internal_ticket->t_trans_type = t_type;
382 *ticket = internal_ticket;
384 trace_xfs_log_reserve(log, internal_ticket);
386 xlog_grant_push_ail(mp,
387 (internal_ticket->t_unit_res *
388 internal_ticket->t_cnt));
389 retval = xlog_grant_log_space(log, internal_ticket);
393 } /* xfs_log_reserve */
397 * Mount a log filesystem
399 * mp - ubiquitous xfs mount point structure
400 * log_target - buftarg of on-disk log device
401 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
402 * num_bblocks - Number of BBSIZE blocks in on-disk log
404 * Return error or zero.
409 xfs_buftarg_t *log_target,
410 xfs_daddr_t blk_offset,
415 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
416 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
419 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
421 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
424 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
425 if (IS_ERR(mp->m_log)) {
426 error = -PTR_ERR(mp->m_log);
431 * Initialize the AIL now we have a log.
433 error = xfs_trans_ail_init(mp);
435 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
438 mp->m_log->l_ailp = mp->m_ail;
441 * skip log recovery on a norecovery mount. pretend it all
444 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
445 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
448 mp->m_flags &= ~XFS_MOUNT_RDONLY;
450 error = xlog_recover(mp->m_log);
453 mp->m_flags |= XFS_MOUNT_RDONLY;
455 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
456 goto out_destroy_ail;
460 /* Normal transactions can now occur */
461 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
466 xfs_trans_ail_destroy(mp);
468 xlog_dealloc_log(mp->m_log);
474 * Finish the recovery of the file system. This is separate from
475 * the xfs_log_mount() call, because it depends on the code in
476 * xfs_mountfs() to read in the root and real-time bitmap inodes
477 * between calling xfs_log_mount() and here.
479 * mp - ubiquitous xfs mount point structure
482 xfs_log_mount_finish(xfs_mount_t *mp)
486 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
487 error = xlog_recover_finish(mp->m_log);
490 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
497 * Final log writes as part of unmount.
499 * Mark the filesystem clean as unmount happens. Note that during relocation
500 * this routine needs to be executed as part of source-bag while the
501 * deallocation must not be done until source-end.
505 * Unmount record used to have a string "Unmount filesystem--" in the
506 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
507 * We just write the magic number now since that particular field isn't
508 * currently architecture converted and "nUmount" is a bit foo.
509 * As far as I know, there weren't any dependencies on the old behaviour.
513 xfs_log_unmount_write(xfs_mount_t *mp)
515 xlog_t *log = mp->m_log;
516 xlog_in_core_t *iclog;
518 xlog_in_core_t *first_iclog;
520 xlog_ticket_t *tic = NULL;
525 * Don't write out unmount record on read-only mounts.
526 * Or, if we are doing a forced umount (typically because of IO errors).
528 if (mp->m_flags & XFS_MOUNT_RDONLY)
531 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
532 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
535 first_iclog = iclog = log->l_iclog;
537 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
538 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
539 ASSERT(iclog->ic_offset == 0);
541 iclog = iclog->ic_next;
542 } while (iclog != first_iclog);
544 if (! (XLOG_FORCED_SHUTDOWN(log))) {
545 error = xfs_log_reserve(mp, 600, 1, &tic,
546 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
548 /* the data section must be 32 bit size aligned */
552 __uint32_t pad2; /* may as well make it 64 bits */
554 .magic = XLOG_UNMOUNT_TYPE,
556 struct xfs_log_iovec reg = {
557 .i_addr = (void *)&magic,
558 .i_len = sizeof(magic),
559 .i_type = XLOG_REG_TYPE_UNMOUNT,
561 struct xfs_log_vec vec = {
566 /* remove inited flag */
568 error = xlog_write(log, &vec, tic, &lsn,
569 NULL, XLOG_UNMOUNT_TRANS);
571 * At this point, we're umounting anyway,
572 * so there's no point in transitioning log state
573 * to IOERROR. Just continue...
578 xfs_fs_cmn_err(CE_ALERT, mp,
579 "xfs_log_unmount: unmount record failed");
583 spin_lock(&log->l_icloglock);
584 iclog = log->l_iclog;
585 atomic_inc(&iclog->ic_refcnt);
586 xlog_state_want_sync(log, iclog);
587 spin_unlock(&log->l_icloglock);
588 error = xlog_state_release_iclog(log, iclog);
590 spin_lock(&log->l_icloglock);
591 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
592 iclog->ic_state == XLOG_STATE_DIRTY)) {
593 if (!XLOG_FORCED_SHUTDOWN(log)) {
594 sv_wait(&iclog->ic_force_wait, PMEM,
595 &log->l_icloglock, s);
597 spin_unlock(&log->l_icloglock);
600 spin_unlock(&log->l_icloglock);
603 trace_xfs_log_umount_write(log, tic);
604 xlog_ungrant_log_space(log, tic);
605 xfs_log_ticket_put(tic);
609 * We're already in forced_shutdown mode, couldn't
610 * even attempt to write out the unmount transaction.
612 * Go through the motions of sync'ing and releasing
613 * the iclog, even though no I/O will actually happen,
614 * we need to wait for other log I/Os that may already
615 * be in progress. Do this as a separate section of
616 * code so we'll know if we ever get stuck here that
617 * we're in this odd situation of trying to unmount
618 * a file system that went into forced_shutdown as
619 * the result of an unmount..
621 spin_lock(&log->l_icloglock);
622 iclog = log->l_iclog;
623 atomic_inc(&iclog->ic_refcnt);
625 xlog_state_want_sync(log, iclog);
626 spin_unlock(&log->l_icloglock);
627 error = xlog_state_release_iclog(log, iclog);
629 spin_lock(&log->l_icloglock);
631 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
632 || iclog->ic_state == XLOG_STATE_DIRTY
633 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
635 sv_wait(&iclog->ic_force_wait, PMEM,
636 &log->l_icloglock, s);
638 spin_unlock(&log->l_icloglock);
643 } /* xfs_log_unmount_write */
646 * Deallocate log structures for unmount/relocation.
648 * We need to stop the aild from running before we destroy
649 * and deallocate the log as the aild references the log.
652 xfs_log_unmount(xfs_mount_t *mp)
654 xfs_trans_ail_destroy(mp);
655 xlog_dealloc_log(mp->m_log);
660 struct xfs_mount *mp,
661 struct xfs_log_item *item,
663 struct xfs_item_ops *ops)
665 item->li_mountp = mp;
666 item->li_ailp = mp->m_ail;
667 item->li_type = type;
672 * Write region vectors to log. The write happens using the space reservation
673 * of the ticket (tic). It is not a requirement that all writes for a given
674 * transaction occur with one call to xfs_log_write(). However, it is important
675 * to note that the transaction reservation code makes an assumption about the
676 * number of log headers a transaction requires that may be violated if you
677 * don't pass all the transaction vectors in one call....
681 struct xfs_mount *mp,
682 struct xfs_log_iovec reg[],
684 struct xlog_ticket *tic,
685 xfs_lsn_t *start_lsn)
687 struct log *log = mp->m_log;
689 struct xfs_log_vec vec = {
690 .lv_niovecs = nentries,
694 if (XLOG_FORCED_SHUTDOWN(log))
695 return XFS_ERROR(EIO);
697 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
699 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
704 xfs_log_move_tail(xfs_mount_t *mp,
708 xlog_t *log = mp->m_log;
709 int need_bytes, free_bytes, cycle, bytes;
711 if (XLOG_FORCED_SHUTDOWN(log))
715 /* needed since sync_lsn is 64 bits */
716 spin_lock(&log->l_icloglock);
717 tail_lsn = log->l_last_sync_lsn;
718 spin_unlock(&log->l_icloglock);
721 spin_lock(&log->l_grant_lock);
723 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
727 log->l_tail_lsn = tail_lsn;
730 if ((tic = log->l_write_headq)) {
732 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
733 panic("Recovery problem");
735 cycle = log->l_grant_write_cycle;
736 bytes = log->l_grant_write_bytes;
737 free_bytes = xlog_space_left(log, cycle, bytes);
739 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
741 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
744 free_bytes -= tic->t_unit_res;
745 sv_signal(&tic->t_wait);
747 } while (tic != log->l_write_headq);
749 if ((tic = log->l_reserve_headq)) {
751 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
752 panic("Recovery problem");
754 cycle = log->l_grant_reserve_cycle;
755 bytes = log->l_grant_reserve_bytes;
756 free_bytes = xlog_space_left(log, cycle, bytes);
758 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
759 need_bytes = tic->t_unit_res*tic->t_cnt;
761 need_bytes = tic->t_unit_res;
762 if (free_bytes < need_bytes && tail_lsn != 1)
765 free_bytes -= need_bytes;
766 sv_signal(&tic->t_wait);
768 } while (tic != log->l_reserve_headq);
770 spin_unlock(&log->l_grant_lock);
771 } /* xfs_log_move_tail */
774 * Determine if we have a transaction that has gone to disk
775 * that needs to be covered. To begin the transition to the idle state
776 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
777 * If we are then in a state where covering is needed, the caller is informed
778 * that dummy transactions are required to move the log into the idle state.
780 * Because this is called as part of the sync process, we should also indicate
781 * that dummy transactions should be issued in anything but the covered or
782 * idle states. This ensures that the log tail is accurately reflected in
783 * the log at the end of the sync, hence if a crash occurrs avoids replay
784 * of transactions where the metadata is already on disk.
787 xfs_log_need_covered(xfs_mount_t *mp)
790 xlog_t *log = mp->m_log;
792 if (!xfs_fs_writable(mp))
795 spin_lock(&log->l_icloglock);
796 switch (log->l_covered_state) {
797 case XLOG_STATE_COVER_DONE:
798 case XLOG_STATE_COVER_DONE2:
799 case XLOG_STATE_COVER_IDLE:
801 case XLOG_STATE_COVER_NEED:
802 case XLOG_STATE_COVER_NEED2:
803 if (!xfs_trans_ail_tail(log->l_ailp) &&
804 xlog_iclogs_empty(log)) {
805 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
806 log->l_covered_state = XLOG_STATE_COVER_DONE;
808 log->l_covered_state = XLOG_STATE_COVER_DONE2;
815 spin_unlock(&log->l_icloglock);
819 /******************************************************************************
823 ******************************************************************************
826 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
827 * The log manager must keep track of the last LR which was committed
828 * to disk. The lsn of this LR will become the new tail_lsn whenever
829 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
830 * the situation where stuff could be written into the log but nothing
831 * was ever in the AIL when asked. Eventually, we panic since the
832 * tail hits the head.
834 * We may be holding the log iclog lock upon entering this routine.
837 xlog_assign_tail_lsn(xfs_mount_t *mp)
840 xlog_t *log = mp->m_log;
842 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
843 spin_lock(&log->l_grant_lock);
845 log->l_tail_lsn = tail_lsn;
847 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
849 spin_unlock(&log->l_grant_lock);
852 } /* xlog_assign_tail_lsn */
856 * Return the space in the log between the tail and the head. The head
857 * is passed in the cycle/bytes formal parms. In the special case where
858 * the reserve head has wrapped passed the tail, this calculation is no
859 * longer valid. In this case, just return 0 which means there is no space
860 * in the log. This works for all places where this function is called
861 * with the reserve head. Of course, if the write head were to ever
862 * wrap the tail, we should blow up. Rather than catch this case here,
863 * we depend on other ASSERTions in other parts of the code. XXXmiken
865 * This code also handles the case where the reservation head is behind
866 * the tail. The details of this case are described below, but the end
867 * result is that we return the size of the log as the amount of space left.
870 xlog_space_left(xlog_t *log, int cycle, int bytes)
876 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
877 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
878 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
879 free_bytes = log->l_logsize - (bytes - tail_bytes);
880 } else if ((tail_cycle + 1) < cycle) {
882 } else if (tail_cycle < cycle) {
883 ASSERT(tail_cycle == (cycle - 1));
884 free_bytes = tail_bytes - bytes;
887 * The reservation head is behind the tail.
888 * In this case we just want to return the size of the
889 * log as the amount of space left.
891 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
892 "xlog_space_left: head behind tail\n"
893 " tail_cycle = %d, tail_bytes = %d\n"
894 " GH cycle = %d, GH bytes = %d",
895 tail_cycle, tail_bytes, cycle, bytes);
897 free_bytes = log->l_logsize;
900 } /* xlog_space_left */
904 * Log function which is called when an io completes.
906 * The log manager needs its own routine, in order to control what
907 * happens with the buffer after the write completes.
910 xlog_iodone(xfs_buf_t *bp)
912 xlog_in_core_t *iclog;
916 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
917 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
918 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
923 * If the _XFS_BARRIER_FAILED flag was set by a lower
924 * layer, it means the underlying device no longer supports
925 * barrier I/O. Warn loudly and turn off barriers.
927 if (bp->b_flags & _XFS_BARRIER_FAILED) {
928 bp->b_flags &= ~_XFS_BARRIER_FAILED;
929 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
930 xfs_fs_cmn_err(CE_WARN, l->l_mp,
931 "xlog_iodone: Barriers are no longer supported"
932 " by device. Disabling barriers\n");
936 * Race to shutdown the filesystem if we see an error.
938 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
939 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
940 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
942 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
944 * This flag will be propagated to the trans-committed
945 * callback routines to let them know that the log-commit
948 aborted = XFS_LI_ABORTED;
949 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
950 aborted = XFS_LI_ABORTED;
953 /* log I/O is always issued ASYNC */
954 ASSERT(XFS_BUF_ISASYNC(bp));
955 xlog_state_done_syncing(iclog, aborted);
957 * do not reference the buffer (bp) here as we could race
958 * with it being freed after writing the unmount record to the
965 * Return size of each in-core log record buffer.
967 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
969 * If the filesystem blocksize is too large, we may need to choose a
970 * larger size since the directory code currently logs entire blocks.
974 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
980 if (mp->m_logbufs <= 0)
981 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
983 log->l_iclog_bufs = mp->m_logbufs;
986 * Buffer size passed in from mount system call.
988 if (mp->m_logbsize > 0) {
989 size = log->l_iclog_size = mp->m_logbsize;
990 log->l_iclog_size_log = 0;
992 log->l_iclog_size_log++;
996 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
997 /* # headers = size / 32k
998 * one header holds cycles from 32k of data
1001 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1002 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1004 log->l_iclog_hsize = xhdrs << BBSHIFT;
1005 log->l_iclog_heads = xhdrs;
1007 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1008 log->l_iclog_hsize = BBSIZE;
1009 log->l_iclog_heads = 1;
1014 /* All machines use 32kB buffers by default. */
1015 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1016 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1018 /* the default log size is 16k or 32k which is one header sector */
1019 log->l_iclog_hsize = BBSIZE;
1020 log->l_iclog_heads = 1;
1023 /* are we being asked to make the sizes selected above visible? */
1024 if (mp->m_logbufs == 0)
1025 mp->m_logbufs = log->l_iclog_bufs;
1026 if (mp->m_logbsize == 0)
1027 mp->m_logbsize = log->l_iclog_size;
1028 } /* xlog_get_iclog_buffer_size */
1032 * This routine initializes some of the log structure for a given mount point.
1033 * Its primary purpose is to fill in enough, so recovery can occur. However,
1034 * some other stuff may be filled in too.
1037 xlog_alloc_log(xfs_mount_t *mp,
1038 xfs_buftarg_t *log_target,
1039 xfs_daddr_t blk_offset,
1043 xlog_rec_header_t *head;
1044 xlog_in_core_t **iclogp;
1045 xlog_in_core_t *iclog, *prev_iclog=NULL;
1052 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1054 xlog_warn("XFS: Log allocation failed: No memory!");
1059 log->l_targ = log_target;
1060 log->l_logsize = BBTOB(num_bblks);
1061 log->l_logBBstart = blk_offset;
1062 log->l_logBBsize = num_bblks;
1063 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1064 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1066 log->l_prev_block = -1;
1067 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1068 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1069 log->l_last_sync_lsn = log->l_tail_lsn;
1070 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1071 log->l_grant_reserve_cycle = 1;
1072 log->l_grant_write_cycle = 1;
1074 error = EFSCORRUPTED;
1075 if (xfs_sb_version_hassector(&mp->m_sb)) {
1076 log2_size = mp->m_sb.sb_logsectlog;
1077 if (log2_size < BBSHIFT) {
1078 xlog_warn("XFS: Log sector size too small "
1079 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1083 log2_size -= BBSHIFT;
1084 if (log2_size > mp->m_sectbb_log) {
1085 xlog_warn("XFS: Log sector size too large "
1086 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1090 /* for larger sector sizes, must have v2 or external log */
1091 if (log2_size && log->l_logBBstart > 0 &&
1092 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1094 xlog_warn("XFS: log sector size (0x%x) invalid "
1095 "for configuration.", log2_size);
1099 log->l_sectBBsize = 1 << log2_size;
1101 xlog_get_iclog_buffer_size(mp, log);
1104 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1107 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1108 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1109 ASSERT(XFS_BUF_ISBUSY(bp));
1110 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1113 spin_lock_init(&log->l_icloglock);
1114 spin_lock_init(&log->l_grant_lock);
1115 sv_init(&log->l_flush_wait, 0, "flush_wait");
1117 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1118 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1120 iclogp = &log->l_iclog;
1122 * The amount of memory to allocate for the iclog structure is
1123 * rather funky due to the way the structure is defined. It is
1124 * done this way so that we can use different sizes for machines
1125 * with different amounts of memory. See the definition of
1126 * xlog_in_core_t in xfs_log_priv.h for details.
1128 iclogsize = log->l_iclog_size;
1129 ASSERT(log->l_iclog_size >= 4096);
1130 for (i=0; i < log->l_iclog_bufs; i++) {
1131 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1133 goto out_free_iclog;
1136 iclog->ic_prev = prev_iclog;
1139 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1141 goto out_free_iclog;
1142 if (!XFS_BUF_CPSEMA(bp))
1144 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1145 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1147 iclog->ic_data = bp->b_addr;
1149 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1151 head = &iclog->ic_header;
1152 memset(head, 0, sizeof(xlog_rec_header_t));
1153 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1154 head->h_version = cpu_to_be32(
1155 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1156 head->h_size = cpu_to_be32(log->l_iclog_size);
1158 head->h_fmt = cpu_to_be32(XLOG_FMT);
1159 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1161 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1162 iclog->ic_state = XLOG_STATE_ACTIVE;
1163 iclog->ic_log = log;
1164 atomic_set(&iclog->ic_refcnt, 0);
1165 spin_lock_init(&iclog->ic_callback_lock);
1166 iclog->ic_callback_tail = &(iclog->ic_callback);
1167 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1169 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1170 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1171 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1172 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1174 iclogp = &iclog->ic_next;
1176 *iclogp = log->l_iclog; /* complete ring */
1177 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1182 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1183 prev_iclog = iclog->ic_next;
1185 sv_destroy(&iclog->ic_force_wait);
1186 sv_destroy(&iclog->ic_write_wait);
1187 xfs_buf_free(iclog->ic_bp);
1191 spinlock_destroy(&log->l_icloglock);
1192 spinlock_destroy(&log->l_grant_lock);
1193 xfs_buf_free(log->l_xbuf);
1197 return ERR_PTR(-error);
1198 } /* xlog_alloc_log */
1202 * Write out the commit record of a transaction associated with the given
1203 * ticket. Return the lsn of the commit record.
1208 struct xlog_ticket *ticket,
1209 struct xlog_in_core **iclog,
1210 xfs_lsn_t *commitlsnp)
1212 struct xfs_mount *mp = log->l_mp;
1214 struct xfs_log_iovec reg = {
1217 .i_type = XLOG_REG_TYPE_COMMIT,
1219 struct xfs_log_vec vec = {
1224 ASSERT_ALWAYS(iclog);
1225 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1228 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1233 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1234 * log space. This code pushes on the lsn which would supposedly free up
1235 * the 25% which we want to leave free. We may need to adopt a policy which
1236 * pushes on an lsn which is further along in the log once we reach the high
1237 * water mark. In this manner, we would be creating a low water mark.
1240 xlog_grant_push_ail(xfs_mount_t *mp,
1243 xlog_t *log = mp->m_log; /* pointer to the log */
1244 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1245 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1246 int free_blocks; /* free blocks left to write to */
1247 int free_bytes; /* free bytes left to write to */
1248 int threshold_block; /* block in lsn we'd like to be at */
1249 int threshold_cycle; /* lsn cycle we'd like to be at */
1252 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1254 spin_lock(&log->l_grant_lock);
1255 free_bytes = xlog_space_left(log,
1256 log->l_grant_reserve_cycle,
1257 log->l_grant_reserve_bytes);
1258 tail_lsn = log->l_tail_lsn;
1259 free_blocks = BTOBBT(free_bytes);
1262 * Set the threshold for the minimum number of free blocks in the
1263 * log to the maximum of what the caller needs, one quarter of the
1264 * log, and 256 blocks.
1266 free_threshold = BTOBB(need_bytes);
1267 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1268 free_threshold = MAX(free_threshold, 256);
1269 if (free_blocks < free_threshold) {
1270 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1271 threshold_cycle = CYCLE_LSN(tail_lsn);
1272 if (threshold_block >= log->l_logBBsize) {
1273 threshold_block -= log->l_logBBsize;
1274 threshold_cycle += 1;
1276 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1278 /* Don't pass in an lsn greater than the lsn of the last
1279 * log record known to be on disk.
1281 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1282 threshold_lsn = log->l_last_sync_lsn;
1284 spin_unlock(&log->l_grant_lock);
1287 * Get the transaction layer to kick the dirty buffers out to
1288 * disk asynchronously. No point in trying to do this if
1289 * the filesystem is shutting down.
1291 if (threshold_lsn &&
1292 !XLOG_FORCED_SHUTDOWN(log))
1293 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1294 } /* xlog_grant_push_ail */
1297 * The bdstrat callback function for log bufs. This gives us a central
1298 * place to trap bufs in case we get hit by a log I/O error and need to
1299 * shutdown. Actually, in practice, even when we didn't get a log error,
1300 * we transition the iclogs to IOERROR state *after* flushing all existing
1301 * iclogs to disk. This is because we don't want anymore new transactions to be
1302 * started or completed afterwards.
1308 struct xlog_in_core *iclog;
1310 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1311 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1312 XFS_BUF_ERROR(bp, EIO);
1316 * It would seem logical to return EIO here, but we rely on
1317 * the log state machine to propagate I/O errors instead of
1323 bp->b_flags |= _XBF_RUN_QUEUES;
1324 xfs_buf_iorequest(bp);
1329 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1330 * fashion. Previously, we should have moved the current iclog
1331 * ptr in the log to point to the next available iclog. This allows further
1332 * write to continue while this code syncs out an iclog ready to go.
1333 * Before an in-core log can be written out, the data section must be scanned
1334 * to save away the 1st word of each BBSIZE block into the header. We replace
1335 * it with the current cycle count. Each BBSIZE block is tagged with the
1336 * cycle count because there in an implicit assumption that drives will
1337 * guarantee that entire 512 byte blocks get written at once. In other words,
1338 * we can't have part of a 512 byte block written and part not written. By
1339 * tagging each block, we will know which blocks are valid when recovering
1340 * after an unclean shutdown.
1342 * This routine is single threaded on the iclog. No other thread can be in
1343 * this routine with the same iclog. Changing contents of iclog can there-
1344 * fore be done without grabbing the state machine lock. Updating the global
1345 * log will require grabbing the lock though.
1347 * The entire log manager uses a logical block numbering scheme. Only
1348 * log_sync (and then only bwrite()) know about the fact that the log may
1349 * not start with block zero on a given device. The log block start offset
1350 * is added immediately before calling bwrite().
1354 xlog_sync(xlog_t *log,
1355 xlog_in_core_t *iclog)
1357 xfs_caddr_t dptr; /* pointer to byte sized element */
1360 uint count; /* byte count of bwrite */
1361 uint count_init; /* initial count before roundup */
1362 int roundoff; /* roundoff to BB or stripe */
1363 int split = 0; /* split write into two regions */
1365 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1367 XFS_STATS_INC(xs_log_writes);
1368 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1370 /* Add for LR header */
1371 count_init = log->l_iclog_hsize + iclog->ic_offset;
1373 /* Round out the log write size */
1374 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1375 /* we have a v2 stripe unit to use */
1376 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1378 count = BBTOB(BTOBB(count_init));
1380 roundoff = count - count_init;
1381 ASSERT(roundoff >= 0);
1382 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1383 roundoff < log->l_mp->m_sb.sb_logsunit)
1385 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1386 roundoff < BBTOB(1)));
1388 /* move grant heads by roundoff in sync */
1389 spin_lock(&log->l_grant_lock);
1390 xlog_grant_add_space(log, roundoff);
1391 spin_unlock(&log->l_grant_lock);
1393 /* put cycle number in every block */
1394 xlog_pack_data(log, iclog, roundoff);
1396 /* real byte length */
1398 iclog->ic_header.h_len =
1399 cpu_to_be32(iclog->ic_offset + roundoff);
1401 iclog->ic_header.h_len =
1402 cpu_to_be32(iclog->ic_offset);
1406 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1407 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1408 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1410 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1412 /* Do we need to split this write into 2 parts? */
1413 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1414 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1415 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1416 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1418 iclog->ic_bwritecnt = 1;
1420 XFS_BUF_SET_COUNT(bp, count);
1421 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1422 XFS_BUF_ZEROFLAGS(bp);
1425 bp->b_flags |= XBF_LOG_BUFFER;
1427 * Do an ordered write for the log block.
1428 * Its unnecessary to flush the first split block in the log wrap case.
1430 if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1431 XFS_BUF_ORDERED(bp);
1433 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1434 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1436 xlog_verify_iclog(log, iclog, count, B_TRUE);
1438 /* account for log which doesn't start at block #0 */
1439 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1441 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1446 if ((error = xlog_bdstrat(bp))) {
1447 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1452 bp = iclog->ic_log->l_xbuf;
1453 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1455 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1456 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1457 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1458 (__psint_t)count), split);
1459 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1460 XFS_BUF_ZEROFLAGS(bp);
1463 bp->b_flags |= XBF_LOG_BUFFER;
1464 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1465 XFS_BUF_ORDERED(bp);
1466 dptr = XFS_BUF_PTR(bp);
1468 * Bump the cycle numbers at the start of each block
1469 * since this part of the buffer is at the start of
1470 * a new cycle. Watch out for the header magic number
1473 for (i = 0; i < split; i += BBSIZE) {
1474 be32_add_cpu((__be32 *)dptr, 1);
1475 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1476 be32_add_cpu((__be32 *)dptr, 1);
1480 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1481 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1483 /* account for internal log which doesn't start at block #0 */
1484 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1486 if ((error = xlog_bdstrat(bp))) {
1487 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1488 bp, XFS_BUF_ADDR(bp));
1497 * Deallocate a log structure
1500 xlog_dealloc_log(xlog_t *log)
1502 xlog_in_core_t *iclog, *next_iclog;
1505 iclog = log->l_iclog;
1506 for (i=0; i<log->l_iclog_bufs; i++) {
1507 sv_destroy(&iclog->ic_force_wait);
1508 sv_destroy(&iclog->ic_write_wait);
1509 xfs_buf_free(iclog->ic_bp);
1510 next_iclog = iclog->ic_next;
1514 spinlock_destroy(&log->l_icloglock);
1515 spinlock_destroy(&log->l_grant_lock);
1517 xfs_buf_free(log->l_xbuf);
1518 log->l_mp->m_log = NULL;
1520 } /* xlog_dealloc_log */
1523 * Update counters atomically now that memcpy is done.
1527 xlog_state_finish_copy(xlog_t *log,
1528 xlog_in_core_t *iclog,
1532 spin_lock(&log->l_icloglock);
1534 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1535 iclog->ic_offset += copy_bytes;
1537 spin_unlock(&log->l_icloglock);
1538 } /* xlog_state_finish_copy */
1544 * print out info relating to regions written which consume
1548 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1551 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1553 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1554 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1575 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1618 xfs_fs_cmn_err(CE_WARN, mp,
1619 "xfs_log_write: reservation summary:\n"
1620 " trans type = %s (%u)\n"
1621 " unit res = %d bytes\n"
1622 " current res = %d bytes\n"
1623 " total reg = %u bytes (o/flow = %u bytes)\n"
1624 " ophdrs = %u (ophdr space = %u bytes)\n"
1625 " ophdr + reg = %u bytes\n"
1626 " num regions = %u\n",
1627 ((ticket->t_trans_type <= 0 ||
1628 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1629 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1630 ticket->t_trans_type,
1633 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1634 ticket->t_res_num_ophdrs, ophdr_spc,
1635 ticket->t_res_arr_sum +
1636 ticket->t_res_o_flow + ophdr_spc,
1639 for (i = 0; i < ticket->t_res_num; i++) {
1640 uint r_type = ticket->t_res_arr[i].r_type;
1642 "region[%u]: %s - %u bytes\n",
1644 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1645 "bad-rtype" : res_type_str[r_type-1]),
1646 ticket->t_res_arr[i].r_len);
1651 * Calculate the potential space needed by the log vector. Each region gets
1652 * its own xlog_op_header_t and may need to be double word aligned.
1655 xlog_write_calc_vec_length(
1656 struct xlog_ticket *ticket,
1657 struct xfs_log_vec *log_vector)
1659 struct xfs_log_vec *lv;
1664 /* acct for start rec of xact */
1665 if (ticket->t_flags & XLOG_TIC_INITED)
1668 for (lv = log_vector; lv; lv = lv->lv_next) {
1669 headers += lv->lv_niovecs;
1671 for (i = 0; i < lv->lv_niovecs; i++) {
1672 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1675 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1679 ticket->t_res_num_ophdrs += headers;
1680 len += headers * sizeof(struct xlog_op_header);
1686 * If first write for transaction, insert start record We can't be trying to
1687 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1690 xlog_write_start_rec(
1691 struct xlog_op_header *ophdr,
1692 struct xlog_ticket *ticket)
1694 if (!(ticket->t_flags & XLOG_TIC_INITED))
1697 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1698 ophdr->oh_clientid = ticket->t_clientid;
1700 ophdr->oh_flags = XLOG_START_TRANS;
1703 ticket->t_flags &= ~XLOG_TIC_INITED;
1705 return sizeof(struct xlog_op_header);
1708 static xlog_op_header_t *
1709 xlog_write_setup_ophdr(
1711 struct xlog_op_header *ophdr,
1712 struct xlog_ticket *ticket,
1715 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1716 ophdr->oh_clientid = ticket->t_clientid;
1719 /* are we copying a commit or unmount record? */
1720 ophdr->oh_flags = flags;
1723 * We've seen logs corrupted with bad transaction client ids. This
1724 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1725 * and shut down the filesystem.
1727 switch (ophdr->oh_clientid) {
1728 case XFS_TRANSACTION:
1733 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1734 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1735 ophdr->oh_clientid, ticket);
1743 * Set up the parameters of the region copy into the log. This has
1744 * to handle region write split across multiple log buffers - this
1745 * state is kept external to this function so that this code can
1746 * can be written in an obvious, self documenting manner.
1749 xlog_write_setup_copy(
1750 struct xlog_ticket *ticket,
1751 struct xlog_op_header *ophdr,
1752 int space_available,
1756 int *last_was_partial_copy,
1757 int *bytes_consumed)
1761 still_to_copy = space_required - *bytes_consumed;
1762 *copy_off = *bytes_consumed;
1764 if (still_to_copy <= space_available) {
1765 /* write of region completes here */
1766 *copy_len = still_to_copy;
1767 ophdr->oh_len = cpu_to_be32(*copy_len);
1768 if (*last_was_partial_copy)
1769 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1770 *last_was_partial_copy = 0;
1771 *bytes_consumed = 0;
1775 /* partial write of region, needs extra log op header reservation */
1776 *copy_len = space_available;
1777 ophdr->oh_len = cpu_to_be32(*copy_len);
1778 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1779 if (*last_was_partial_copy)
1780 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1781 *bytes_consumed += *copy_len;
1782 (*last_was_partial_copy)++;
1784 /* account for new log op header */
1785 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1786 ticket->t_res_num_ophdrs++;
1788 return sizeof(struct xlog_op_header);
1792 xlog_write_copy_finish(
1794 struct xlog_in_core *iclog,
1799 int *partial_copy_len,
1801 struct xlog_in_core **commit_iclog)
1803 if (*partial_copy) {
1805 * This iclog has already been marked WANT_SYNC by
1806 * xlog_state_get_iclog_space.
1808 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1811 return xlog_state_release_iclog(log, iclog);
1815 *partial_copy_len = 0;
1817 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1818 /* no more space in this iclog - push it. */
1819 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1823 spin_lock(&log->l_icloglock);
1824 xlog_state_want_sync(log, iclog);
1825 spin_unlock(&log->l_icloglock);
1828 return xlog_state_release_iclog(log, iclog);
1829 ASSERT(flags & XLOG_COMMIT_TRANS);
1830 *commit_iclog = iclog;
1837 * Write some region out to in-core log
1839 * This will be called when writing externally provided regions or when
1840 * writing out a commit record for a given transaction.
1842 * General algorithm:
1843 * 1. Find total length of this write. This may include adding to the
1844 * lengths passed in.
1845 * 2. Check whether we violate the tickets reservation.
1846 * 3. While writing to this iclog
1847 * A. Reserve as much space in this iclog as can get
1848 * B. If this is first write, save away start lsn
1849 * C. While writing this region:
1850 * 1. If first write of transaction, write start record
1851 * 2. Write log operation header (header per region)
1852 * 3. Find out if we can fit entire region into this iclog
1853 * 4. Potentially, verify destination memcpy ptr
1854 * 5. Memcpy (partial) region
1855 * 6. If partial copy, release iclog; otherwise, continue
1856 * copying more regions into current iclog
1857 * 4. Mark want sync bit (in simulation mode)
1858 * 5. Release iclog for potential flush to on-disk log.
1861 * 1. Panic if reservation is overrun. This should never happen since
1862 * reservation amounts are generated internal to the filesystem.
1864 * 1. Tickets are single threaded data structures.
1865 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1866 * syncing routine. When a single log_write region needs to span
1867 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1868 * on all log operation writes which don't contain the end of the
1869 * region. The XLOG_END_TRANS bit is used for the in-core log
1870 * operation which contains the end of the continued log_write region.
1871 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1872 * we don't really know exactly how much space will be used. As a result,
1873 * we don't update ic_offset until the end when we know exactly how many
1874 * bytes have been written out.
1879 struct xfs_log_vec *log_vector,
1880 struct xlog_ticket *ticket,
1881 xfs_lsn_t *start_lsn,
1882 struct xlog_in_core **commit_iclog,
1885 struct xlog_in_core *iclog = NULL;
1886 struct xfs_log_iovec *vecp;
1887 struct xfs_log_vec *lv;
1890 int partial_copy = 0;
1891 int partial_copy_len = 0;
1899 len = xlog_write_calc_vec_length(ticket, log_vector);
1900 if (ticket->t_curr_res < len) {
1901 xlog_print_tic_res(log->l_mp, ticket);
1904 "xfs_log_write: reservation ran out. Need to up reservation");
1906 /* Customer configurable panic */
1907 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, log->l_mp,
1908 "xfs_log_write: reservation ran out. Need to up reservation");
1910 /* If we did not panic, shutdown the filesystem */
1911 xfs_force_shutdown(log->l_mp, SHUTDOWN_CORRUPT_INCORE);
1915 ticket->t_curr_res -= len;
1919 vecp = lv->lv_iovecp;
1920 while (lv && index < lv->lv_niovecs) {
1924 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1925 &contwr, &log_offset);
1929 ASSERT(log_offset <= iclog->ic_size - 1);
1930 ptr = iclog->ic_datap + log_offset;
1932 /* start_lsn is the first lsn written to. That's all we need. */
1934 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1937 * This loop writes out as many regions as can fit in the amount
1938 * of space which was allocated by xlog_state_get_iclog_space().
1940 while (lv && index < lv->lv_niovecs) {
1941 struct xfs_log_iovec *reg = &vecp[index];
1942 struct xlog_op_header *ophdr;
1947 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1948 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1950 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1951 if (start_rec_copy) {
1953 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1957 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1959 return XFS_ERROR(EIO);
1961 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1962 sizeof(struct xlog_op_header));
1964 len += xlog_write_setup_copy(ticket, ophdr,
1965 iclog->ic_size-log_offset,
1967 ©_off, ©_len,
1970 xlog_verify_dest_ptr(log, ptr);
1973 ASSERT(copy_len >= 0);
1974 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1975 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1977 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1979 data_cnt += contwr ? copy_len : 0;
1981 error = xlog_write_copy_finish(log, iclog, flags,
1982 &record_cnt, &data_cnt,
1991 * if we had a partial copy, we need to get more iclog
1992 * space but we don't want to increment the region
1993 * index because there is still more is this region to
1996 * If we completed writing this region, and we flushed
1997 * the iclog (indicated by resetting of the record
1998 * count), then we also need to get more log space. If
1999 * this was the last record, though, we are done and
2005 if (++index == lv->lv_niovecs) {
2009 vecp = lv->lv_iovecp;
2011 if (record_cnt == 0) {
2021 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2023 return xlog_state_release_iclog(log, iclog);
2025 ASSERT(flags & XLOG_COMMIT_TRANS);
2026 *commit_iclog = iclog;
2031 /*****************************************************************************
2033 * State Machine functions
2035 *****************************************************************************
2038 /* Clean iclogs starting from the head. This ordering must be
2039 * maintained, so an iclog doesn't become ACTIVE beyond one that
2040 * is SYNCING. This is also required to maintain the notion that we use
2041 * a ordered wait queue to hold off would be writers to the log when every
2042 * iclog is trying to sync to disk.
2044 * State Change: DIRTY -> ACTIVE
2047 xlog_state_clean_log(xlog_t *log)
2049 xlog_in_core_t *iclog;
2052 iclog = log->l_iclog;
2054 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2055 iclog->ic_state = XLOG_STATE_ACTIVE;
2056 iclog->ic_offset = 0;
2057 ASSERT(iclog->ic_callback == NULL);
2059 * If the number of ops in this iclog indicate it just
2060 * contains the dummy transaction, we can
2061 * change state into IDLE (the second time around).
2062 * Otherwise we should change the state into
2064 * We don't need to cover the dummy.
2067 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2072 * We have two dirty iclogs so start over
2073 * This could also be num of ops indicates
2074 * this is not the dummy going out.
2078 iclog->ic_header.h_num_logops = 0;
2079 memset(iclog->ic_header.h_cycle_data, 0,
2080 sizeof(iclog->ic_header.h_cycle_data));
2081 iclog->ic_header.h_lsn = 0;
2082 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2085 break; /* stop cleaning */
2086 iclog = iclog->ic_next;
2087 } while (iclog != log->l_iclog);
2089 /* log is locked when we are called */
2091 * Change state for the dummy log recording.
2092 * We usually go to NEED. But we go to NEED2 if the changed indicates
2093 * we are done writing the dummy record.
2094 * If we are done with the second dummy recored (DONE2), then
2098 switch (log->l_covered_state) {
2099 case XLOG_STATE_COVER_IDLE:
2100 case XLOG_STATE_COVER_NEED:
2101 case XLOG_STATE_COVER_NEED2:
2102 log->l_covered_state = XLOG_STATE_COVER_NEED;
2105 case XLOG_STATE_COVER_DONE:
2107 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2109 log->l_covered_state = XLOG_STATE_COVER_NEED;
2112 case XLOG_STATE_COVER_DONE2:
2114 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2116 log->l_covered_state = XLOG_STATE_COVER_NEED;
2123 } /* xlog_state_clean_log */
2126 xlog_get_lowest_lsn(
2129 xlog_in_core_t *lsn_log;
2130 xfs_lsn_t lowest_lsn, lsn;
2132 lsn_log = log->l_iclog;
2135 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2136 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2137 if ((lsn && !lowest_lsn) ||
2138 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2142 lsn_log = lsn_log->ic_next;
2143 } while (lsn_log != log->l_iclog);
2149 xlog_state_do_callback(
2152 xlog_in_core_t *ciclog)
2154 xlog_in_core_t *iclog;
2155 xlog_in_core_t *first_iclog; /* used to know when we've
2156 * processed all iclogs once */
2157 xfs_log_callback_t *cb, *cb_next;
2159 xfs_lsn_t lowest_lsn;
2160 int ioerrors; /* counter: iclogs with errors */
2161 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2162 int funcdidcallbacks; /* flag: function did callbacks */
2163 int repeats; /* for issuing console warnings if
2164 * looping too many times */
2167 spin_lock(&log->l_icloglock);
2168 first_iclog = iclog = log->l_iclog;
2170 funcdidcallbacks = 0;
2175 * Scan all iclogs starting with the one pointed to by the
2176 * log. Reset this starting point each time the log is
2177 * unlocked (during callbacks).
2179 * Keep looping through iclogs until one full pass is made
2180 * without running any callbacks.
2182 first_iclog = log->l_iclog;
2183 iclog = log->l_iclog;
2184 loopdidcallbacks = 0;
2189 /* skip all iclogs in the ACTIVE & DIRTY states */
2190 if (iclog->ic_state &
2191 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2192 iclog = iclog->ic_next;
2197 * Between marking a filesystem SHUTDOWN and stopping
2198 * the log, we do flush all iclogs to disk (if there
2199 * wasn't a log I/O error). So, we do want things to
2200 * go smoothly in case of just a SHUTDOWN w/o a
2203 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2205 * Can only perform callbacks in order. Since
2206 * this iclog is not in the DONE_SYNC/
2207 * DO_CALLBACK state, we skip the rest and
2208 * just try to clean up. If we set our iclog
2209 * to DO_CALLBACK, we will not process it when
2210 * we retry since a previous iclog is in the
2211 * CALLBACK and the state cannot change since
2212 * we are holding the l_icloglock.
2214 if (!(iclog->ic_state &
2215 (XLOG_STATE_DONE_SYNC |
2216 XLOG_STATE_DO_CALLBACK))) {
2217 if (ciclog && (ciclog->ic_state ==
2218 XLOG_STATE_DONE_SYNC)) {
2219 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2224 * We now have an iclog that is in either the
2225 * DO_CALLBACK or DONE_SYNC states. The other
2226 * states (WANT_SYNC, SYNCING, or CALLBACK were
2227 * caught by the above if and are going to
2228 * clean (i.e. we aren't doing their callbacks)
2233 * We will do one more check here to see if we
2234 * have chased our tail around.
2237 lowest_lsn = xlog_get_lowest_lsn(log);
2239 XFS_LSN_CMP(lowest_lsn,
2240 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2241 iclog = iclog->ic_next;
2242 continue; /* Leave this iclog for
2246 iclog->ic_state = XLOG_STATE_CALLBACK;
2248 spin_unlock(&log->l_icloglock);
2250 /* l_last_sync_lsn field protected by
2251 * l_grant_lock. Don't worry about iclog's lsn.
2252 * No one else can be here except us.
2254 spin_lock(&log->l_grant_lock);
2255 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2256 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2257 log->l_last_sync_lsn =
2258 be64_to_cpu(iclog->ic_header.h_lsn);
2259 spin_unlock(&log->l_grant_lock);
2262 spin_unlock(&log->l_icloglock);
2267 * Keep processing entries in the callback list until
2268 * we come around and it is empty. We need to
2269 * atomically see that the list is empty and change the
2270 * state to DIRTY so that we don't miss any more
2271 * callbacks being added.
2273 spin_lock(&iclog->ic_callback_lock);
2274 cb = iclog->ic_callback;
2276 iclog->ic_callback_tail = &(iclog->ic_callback);
2277 iclog->ic_callback = NULL;
2278 spin_unlock(&iclog->ic_callback_lock);
2280 /* perform callbacks in the order given */
2281 for (; cb; cb = cb_next) {
2282 cb_next = cb->cb_next;
2283 cb->cb_func(cb->cb_arg, aborted);
2285 spin_lock(&iclog->ic_callback_lock);
2286 cb = iclog->ic_callback;
2292 spin_lock(&log->l_icloglock);
2293 ASSERT(iclog->ic_callback == NULL);
2294 spin_unlock(&iclog->ic_callback_lock);
2295 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2296 iclog->ic_state = XLOG_STATE_DIRTY;
2299 * Transition from DIRTY to ACTIVE if applicable.
2300 * NOP if STATE_IOERROR.
2302 xlog_state_clean_log(log);
2304 /* wake up threads waiting in xfs_log_force() */
2305 sv_broadcast(&iclog->ic_force_wait);
2307 iclog = iclog->ic_next;
2308 } while (first_iclog != iclog);
2310 if (repeats > 5000) {
2311 flushcnt += repeats;
2313 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2314 "%s: possible infinite loop (%d iterations)",
2315 __func__, flushcnt);
2317 } while (!ioerrors && loopdidcallbacks);
2320 * make one last gasp attempt to see if iclogs are being left in
2324 if (funcdidcallbacks) {
2325 first_iclog = iclog = log->l_iclog;
2327 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2329 * Terminate the loop if iclogs are found in states
2330 * which will cause other threads to clean up iclogs.
2332 * SYNCING - i/o completion will go through logs
2333 * DONE_SYNC - interrupt thread should be waiting for
2335 * IOERROR - give up hope all ye who enter here
2337 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2338 iclog->ic_state == XLOG_STATE_SYNCING ||
2339 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2340 iclog->ic_state == XLOG_STATE_IOERROR )
2342 iclog = iclog->ic_next;
2343 } while (first_iclog != iclog);
2347 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2349 spin_unlock(&log->l_icloglock);
2352 sv_broadcast(&log->l_flush_wait);
2357 * Finish transitioning this iclog to the dirty state.
2359 * Make sure that we completely execute this routine only when this is
2360 * the last call to the iclog. There is a good chance that iclog flushes,
2361 * when we reach the end of the physical log, get turned into 2 separate
2362 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2363 * routine. By using the reference count bwritecnt, we guarantee that only
2364 * the second completion goes through.
2366 * Callbacks could take time, so they are done outside the scope of the
2367 * global state machine log lock.
2370 xlog_state_done_syncing(
2371 xlog_in_core_t *iclog,
2374 xlog_t *log = iclog->ic_log;
2376 spin_lock(&log->l_icloglock);
2378 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2379 iclog->ic_state == XLOG_STATE_IOERROR);
2380 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2381 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2385 * If we got an error, either on the first buffer, or in the case of
2386 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2387 * and none should ever be attempted to be written to disk
2390 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2391 if (--iclog->ic_bwritecnt == 1) {
2392 spin_unlock(&log->l_icloglock);
2395 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2399 * Someone could be sleeping prior to writing out the next
2400 * iclog buffer, we wake them all, one will get to do the
2401 * I/O, the others get to wait for the result.
2403 sv_broadcast(&iclog->ic_write_wait);
2404 spin_unlock(&log->l_icloglock);
2405 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2406 } /* xlog_state_done_syncing */
2410 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2411 * sleep. We wait on the flush queue on the head iclog as that should be
2412 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2413 * we will wait here and all new writes will sleep until a sync completes.
2415 * The in-core logs are used in a circular fashion. They are not used
2416 * out-of-order even when an iclog past the head is free.
2419 * * log_offset where xlog_write() can start writing into the in-core
2421 * * in-core log pointer to which xlog_write() should write.
2422 * * boolean indicating this is a continued write to an in-core log.
2423 * If this is the last write, then the in-core log's offset field
2424 * needs to be incremented, depending on the amount of data which
2428 xlog_state_get_iclog_space(xlog_t *log,
2430 xlog_in_core_t **iclogp,
2431 xlog_ticket_t *ticket,
2432 int *continued_write,
2436 xlog_rec_header_t *head;
2437 xlog_in_core_t *iclog;
2441 spin_lock(&log->l_icloglock);
2442 if (XLOG_FORCED_SHUTDOWN(log)) {
2443 spin_unlock(&log->l_icloglock);
2444 return XFS_ERROR(EIO);
2447 iclog = log->l_iclog;
2448 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2449 XFS_STATS_INC(xs_log_noiclogs);
2451 /* Wait for log writes to have flushed */
2452 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2456 head = &iclog->ic_header;
2458 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2459 log_offset = iclog->ic_offset;
2461 /* On the 1st write to an iclog, figure out lsn. This works
2462 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2463 * committing to. If the offset is set, that's how many blocks
2466 if (log_offset == 0) {
2467 ticket->t_curr_res -= log->l_iclog_hsize;
2468 xlog_tic_add_region(ticket,
2470 XLOG_REG_TYPE_LRHEADER);
2471 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2472 head->h_lsn = cpu_to_be64(
2473 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2474 ASSERT(log->l_curr_block >= 0);
2477 /* If there is enough room to write everything, then do it. Otherwise,
2478 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2479 * bit is on, so this will get flushed out. Don't update ic_offset
2480 * until you know exactly how many bytes get copied. Therefore, wait
2481 * until later to update ic_offset.
2483 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2484 * can fit into remaining data section.
2486 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2487 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2490 * If I'm the only one writing to this iclog, sync it to disk.
2491 * We need to do an atomic compare and decrement here to avoid
2492 * racing with concurrent atomic_dec_and_lock() calls in
2493 * xlog_state_release_iclog() when there is more than one
2494 * reference to the iclog.
2496 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2497 /* we are the only one */
2498 spin_unlock(&log->l_icloglock);
2499 error = xlog_state_release_iclog(log, iclog);
2503 spin_unlock(&log->l_icloglock);
2508 /* Do we have enough room to write the full amount in the remainder
2509 * of this iclog? Or must we continue a write on the next iclog and
2510 * mark this iclog as completely taken? In the case where we switch
2511 * iclogs (to mark it taken), this particular iclog will release/sync
2512 * to disk in xlog_write().
2514 if (len <= iclog->ic_size - iclog->ic_offset) {
2515 *continued_write = 0;
2516 iclog->ic_offset += len;
2518 *continued_write = 1;
2519 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2523 ASSERT(iclog->ic_offset <= iclog->ic_size);
2524 spin_unlock(&log->l_icloglock);
2526 *logoffsetp = log_offset;
2528 } /* xlog_state_get_iclog_space */
2531 * Atomically get the log space required for a log ticket.
2533 * Once a ticket gets put onto the reserveq, it will only return after
2534 * the needed reservation is satisfied.
2537 xlog_grant_log_space(xlog_t *log,
2548 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2549 panic("grant Recovery problem");
2552 /* Is there space or do we need to sleep? */
2553 spin_lock(&log->l_grant_lock);
2555 trace_xfs_log_grant_enter(log, tic);
2557 /* something is already sleeping; insert new transaction at end */
2558 if (log->l_reserve_headq) {
2559 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2561 trace_xfs_log_grant_sleep1(log, tic);
2564 * Gotta check this before going to sleep, while we're
2565 * holding the grant lock.
2567 if (XLOG_FORCED_SHUTDOWN(log))
2570 XFS_STATS_INC(xs_sleep_logspace);
2571 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2573 * If we got an error, and the filesystem is shutting down,
2574 * we'll catch it down below. So just continue...
2576 trace_xfs_log_grant_wake1(log, tic);
2577 spin_lock(&log->l_grant_lock);
2579 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2580 need_bytes = tic->t_unit_res*tic->t_ocnt;
2582 need_bytes = tic->t_unit_res;
2585 if (XLOG_FORCED_SHUTDOWN(log))
2588 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2589 log->l_grant_reserve_bytes);
2590 if (free_bytes < need_bytes) {
2591 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2592 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2594 trace_xfs_log_grant_sleep2(log, tic);
2596 spin_unlock(&log->l_grant_lock);
2597 xlog_grant_push_ail(log->l_mp, need_bytes);
2598 spin_lock(&log->l_grant_lock);
2600 XFS_STATS_INC(xs_sleep_logspace);
2601 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2603 spin_lock(&log->l_grant_lock);
2604 if (XLOG_FORCED_SHUTDOWN(log))
2607 trace_xfs_log_grant_wake2(log, tic);
2610 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2611 xlog_del_ticketq(&log->l_reserve_headq, tic);
2613 /* we've got enough space */
2614 xlog_grant_add_space(log, need_bytes);
2616 tail_lsn = log->l_tail_lsn;
2618 * Check to make sure the grant write head didn't just over lap the
2619 * tail. If the cycles are the same, we can't be overlapping.
2620 * Otherwise, make sure that the cycles differ by exactly one and
2621 * check the byte count.
2623 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2624 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2625 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2628 trace_xfs_log_grant_exit(log, tic);
2629 xlog_verify_grant_head(log, 1);
2630 spin_unlock(&log->l_grant_lock);
2634 if (tic->t_flags & XLOG_TIC_IN_Q)
2635 xlog_del_ticketq(&log->l_reserve_headq, tic);
2637 trace_xfs_log_grant_error(log, tic);
2640 * If we are failing, make sure the ticket doesn't have any
2641 * current reservations. We don't want to add this back when
2642 * the ticket/transaction gets cancelled.
2644 tic->t_curr_res = 0;
2645 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2646 spin_unlock(&log->l_grant_lock);
2647 return XFS_ERROR(EIO);
2648 } /* xlog_grant_log_space */
2652 * Replenish the byte reservation required by moving the grant write head.
2657 xlog_regrant_write_log_space(xlog_t *log,
2660 int free_bytes, need_bytes;
2661 xlog_ticket_t *ntic;
2666 tic->t_curr_res = tic->t_unit_res;
2667 xlog_tic_reset_res(tic);
2673 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2674 panic("regrant Recovery problem");
2677 spin_lock(&log->l_grant_lock);
2679 trace_xfs_log_regrant_write_enter(log, tic);
2681 if (XLOG_FORCED_SHUTDOWN(log))
2684 /* If there are other waiters on the queue then give them a
2685 * chance at logspace before us. Wake up the first waiters,
2686 * if we do not wake up all the waiters then go to sleep waiting
2687 * for more free space, otherwise try to get some space for
2690 need_bytes = tic->t_unit_res;
2691 if ((ntic = log->l_write_headq)) {
2692 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2693 log->l_grant_write_bytes);
2695 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2697 if (free_bytes < ntic->t_unit_res)
2699 free_bytes -= ntic->t_unit_res;
2700 sv_signal(&ntic->t_wait);
2701 ntic = ntic->t_next;
2702 } while (ntic != log->l_write_headq);
2704 if (ntic != log->l_write_headq) {
2705 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2706 xlog_ins_ticketq(&log->l_write_headq, tic);
2708 trace_xfs_log_regrant_write_sleep1(log, tic);
2710 spin_unlock(&log->l_grant_lock);
2711 xlog_grant_push_ail(log->l_mp, need_bytes);
2712 spin_lock(&log->l_grant_lock);
2714 XFS_STATS_INC(xs_sleep_logspace);
2715 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2716 &log->l_grant_lock, s);
2718 /* If we're shutting down, this tic is already
2720 spin_lock(&log->l_grant_lock);
2721 if (XLOG_FORCED_SHUTDOWN(log))
2724 trace_xfs_log_regrant_write_wake1(log, tic);
2729 if (XLOG_FORCED_SHUTDOWN(log))
2732 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2733 log->l_grant_write_bytes);
2734 if (free_bytes < need_bytes) {
2735 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2736 xlog_ins_ticketq(&log->l_write_headq, tic);
2737 spin_unlock(&log->l_grant_lock);
2738 xlog_grant_push_ail(log->l_mp, need_bytes);
2739 spin_lock(&log->l_grant_lock);
2741 XFS_STATS_INC(xs_sleep_logspace);
2742 trace_xfs_log_regrant_write_sleep2(log, tic);
2744 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2746 /* If we're shutting down, this tic is already off the queue */
2747 spin_lock(&log->l_grant_lock);
2748 if (XLOG_FORCED_SHUTDOWN(log))
2751 trace_xfs_log_regrant_write_wake2(log, tic);
2753 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2754 xlog_del_ticketq(&log->l_write_headq, tic);
2756 /* we've got enough space */
2757 xlog_grant_add_space_write(log, need_bytes);
2759 tail_lsn = log->l_tail_lsn;
2760 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2761 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2762 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2766 trace_xfs_log_regrant_write_exit(log, tic);
2768 xlog_verify_grant_head(log, 1);
2769 spin_unlock(&log->l_grant_lock);
2774 if (tic->t_flags & XLOG_TIC_IN_Q)
2775 xlog_del_ticketq(&log->l_reserve_headq, tic);
2777 trace_xfs_log_regrant_write_error(log, tic);
2780 * If we are failing, make sure the ticket doesn't have any
2781 * current reservations. We don't want to add this back when
2782 * the ticket/transaction gets cancelled.
2784 tic->t_curr_res = 0;
2785 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2786 spin_unlock(&log->l_grant_lock);
2787 return XFS_ERROR(EIO);
2788 } /* xlog_regrant_write_log_space */
2791 /* The first cnt-1 times through here we don't need to
2792 * move the grant write head because the permanent
2793 * reservation has reserved cnt times the unit amount.
2794 * Release part of current permanent unit reservation and
2795 * reset current reservation to be one units worth. Also
2796 * move grant reservation head forward.
2799 xlog_regrant_reserve_log_space(xlog_t *log,
2800 xlog_ticket_t *ticket)
2802 trace_xfs_log_regrant_reserve_enter(log, ticket);
2804 if (ticket->t_cnt > 0)
2807 spin_lock(&log->l_grant_lock);
2808 xlog_grant_sub_space(log, ticket->t_curr_res);
2809 ticket->t_curr_res = ticket->t_unit_res;
2810 xlog_tic_reset_res(ticket);
2812 trace_xfs_log_regrant_reserve_sub(log, ticket);
2814 xlog_verify_grant_head(log, 1);
2816 /* just return if we still have some of the pre-reserved space */
2817 if (ticket->t_cnt > 0) {
2818 spin_unlock(&log->l_grant_lock);
2822 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2824 trace_xfs_log_regrant_reserve_exit(log, ticket);
2826 xlog_verify_grant_head(log, 0);
2827 spin_unlock(&log->l_grant_lock);
2828 ticket->t_curr_res = ticket->t_unit_res;
2829 xlog_tic_reset_res(ticket);
2830 } /* xlog_regrant_reserve_log_space */
2834 * Give back the space left from a reservation.
2836 * All the information we need to make a correct determination of space left
2837 * is present. For non-permanent reservations, things are quite easy. The
2838 * count should have been decremented to zero. We only need to deal with the
2839 * space remaining in the current reservation part of the ticket. If the
2840 * ticket contains a permanent reservation, there may be left over space which
2841 * needs to be released. A count of N means that N-1 refills of the current
2842 * reservation can be done before we need to ask for more space. The first
2843 * one goes to fill up the first current reservation. Once we run out of
2844 * space, the count will stay at zero and the only space remaining will be
2845 * in the current reservation field.
2848 xlog_ungrant_log_space(xlog_t *log,
2849 xlog_ticket_t *ticket)
2851 if (ticket->t_cnt > 0)
2854 spin_lock(&log->l_grant_lock);
2855 trace_xfs_log_ungrant_enter(log, ticket);
2857 xlog_grant_sub_space(log, ticket->t_curr_res);
2859 trace_xfs_log_ungrant_sub(log, ticket);
2861 /* If this is a permanent reservation ticket, we may be able to free
2862 * up more space based on the remaining count.
2864 if (ticket->t_cnt > 0) {
2865 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2866 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2869 trace_xfs_log_ungrant_exit(log, ticket);
2871 xlog_verify_grant_head(log, 1);
2872 spin_unlock(&log->l_grant_lock);
2873 xfs_log_move_tail(log->l_mp, 1);
2874 } /* xlog_ungrant_log_space */
2878 * Flush iclog to disk if this is the last reference to the given iclog and
2879 * the WANT_SYNC bit is set.
2881 * When this function is entered, the iclog is not necessarily in the
2882 * WANT_SYNC state. It may be sitting around waiting to get filled.
2887 xlog_state_release_iclog(
2889 xlog_in_core_t *iclog)
2891 int sync = 0; /* do we sync? */
2893 if (iclog->ic_state & XLOG_STATE_IOERROR)
2894 return XFS_ERROR(EIO);
2896 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2897 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2900 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2901 spin_unlock(&log->l_icloglock);
2902 return XFS_ERROR(EIO);
2904 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2905 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2907 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2908 /* update tail before writing to iclog */
2909 xlog_assign_tail_lsn(log->l_mp);
2911 iclog->ic_state = XLOG_STATE_SYNCING;
2912 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2913 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2914 /* cycle incremented when incrementing curr_block */
2916 spin_unlock(&log->l_icloglock);
2919 * We let the log lock go, so it's possible that we hit a log I/O
2920 * error or some other SHUTDOWN condition that marks the iclog
2921 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2922 * this iclog has consistent data, so we ignore IOERROR
2923 * flags after this point.
2926 return xlog_sync(log, iclog);
2928 } /* xlog_state_release_iclog */
2932 * This routine will mark the current iclog in the ring as WANT_SYNC
2933 * and move the current iclog pointer to the next iclog in the ring.
2934 * When this routine is called from xlog_state_get_iclog_space(), the
2935 * exact size of the iclog has not yet been determined. All we know is
2936 * that every data block. We have run out of space in this log record.
2939 xlog_state_switch_iclogs(xlog_t *log,
2940 xlog_in_core_t *iclog,
2943 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2945 eventual_size = iclog->ic_offset;
2946 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2947 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2948 log->l_prev_block = log->l_curr_block;
2949 log->l_prev_cycle = log->l_curr_cycle;
2951 /* roll log?: ic_offset changed later */
2952 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2954 /* Round up to next log-sunit */
2955 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2956 log->l_mp->m_sb.sb_logsunit > 1) {
2957 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2958 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2961 if (log->l_curr_block >= log->l_logBBsize) {
2962 log->l_curr_cycle++;
2963 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2964 log->l_curr_cycle++;
2965 log->l_curr_block -= log->l_logBBsize;
2966 ASSERT(log->l_curr_block >= 0);
2968 ASSERT(iclog == log->l_iclog);
2969 log->l_iclog = iclog->ic_next;
2970 } /* xlog_state_switch_iclogs */
2973 * Write out all data in the in-core log as of this exact moment in time.
2975 * Data may be written to the in-core log during this call. However,
2976 * we don't guarantee this data will be written out. A change from past
2977 * implementation means this routine will *not* write out zero length LRs.
2979 * Basically, we try and perform an intelligent scan of the in-core logs.
2980 * If we determine there is no flushable data, we just return. There is no
2981 * flushable data if:
2983 * 1. the current iclog is active and has no data; the previous iclog
2984 * is in the active or dirty state.
2985 * 2. the current iclog is drity, and the previous iclog is in the
2986 * active or dirty state.
2990 * 1. the current iclog is not in the active nor dirty state.
2991 * 2. the current iclog dirty, and the previous iclog is not in the
2992 * active nor dirty state.
2993 * 3. the current iclog is active, and there is another thread writing
2994 * to this particular iclog.
2995 * 4. a) the current iclog is active and has no other writers
2996 * b) when we return from flushing out this iclog, it is still
2997 * not in the active nor dirty state.
3001 struct xfs_mount *mp,
3005 struct log *log = mp->m_log;
3006 struct xlog_in_core *iclog;
3009 XFS_STATS_INC(xs_log_force);
3011 spin_lock(&log->l_icloglock);
3013 iclog = log->l_iclog;
3014 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3015 spin_unlock(&log->l_icloglock);
3016 return XFS_ERROR(EIO);
3019 /* If the head iclog is not active nor dirty, we just attach
3020 * ourselves to the head and go to sleep.
3022 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3023 iclog->ic_state == XLOG_STATE_DIRTY) {
3025 * If the head is dirty or (active and empty), then
3026 * we need to look at the previous iclog. If the previous
3027 * iclog is active or dirty we are done. There is nothing
3028 * to sync out. Otherwise, we attach ourselves to the
3029 * previous iclog and go to sleep.
3031 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3032 (atomic_read(&iclog->ic_refcnt) == 0
3033 && iclog->ic_offset == 0)) {
3034 iclog = iclog->ic_prev;
3035 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3036 iclog->ic_state == XLOG_STATE_DIRTY)
3041 if (atomic_read(&iclog->ic_refcnt) == 0) {
3042 /* We are the only one with access to this
3043 * iclog. Flush it out now. There should
3044 * be a roundoff of zero to show that someone
3045 * has already taken care of the roundoff from
3046 * the previous sync.
3048 atomic_inc(&iclog->ic_refcnt);
3049 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3050 xlog_state_switch_iclogs(log, iclog, 0);
3051 spin_unlock(&log->l_icloglock);
3053 if (xlog_state_release_iclog(log, iclog))
3054 return XFS_ERROR(EIO);
3058 spin_lock(&log->l_icloglock);
3059 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3060 iclog->ic_state != XLOG_STATE_DIRTY)
3065 /* Someone else is writing to this iclog.
3066 * Use its call to flush out the data. However,
3067 * the other thread may not force out this LR,
3068 * so we mark it WANT_SYNC.
3070 xlog_state_switch_iclogs(log, iclog, 0);
3076 /* By the time we come around again, the iclog could've been filled
3077 * which would give it another lsn. If we have a new lsn, just
3078 * return because the relevant data has been flushed.
3081 if (flags & XFS_LOG_SYNC) {
3083 * We must check if we're shutting down here, before
3084 * we wait, while we're holding the l_icloglock.
3085 * Then we check again after waking up, in case our
3086 * sleep was disturbed by a bad news.
3088 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3089 spin_unlock(&log->l_icloglock);
3090 return XFS_ERROR(EIO);
3092 XFS_STATS_INC(xs_log_force_sleep);
3093 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3095 * No need to grab the log lock here since we're
3096 * only deciding whether or not to return EIO
3097 * and the memory read should be atomic.
3099 if (iclog->ic_state & XLOG_STATE_IOERROR)
3100 return XFS_ERROR(EIO);
3106 spin_unlock(&log->l_icloglock);
3112 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3113 * about errors or whether the log was flushed or not. This is the normal
3114 * interface to use when trying to unpin items or move the log forward.
3123 error = _xfs_log_force(mp, flags, NULL);
3125 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3126 "error %d returned.", error);
3131 * Force the in-core log to disk for a specific LSN.
3133 * Find in-core log with lsn.
3134 * If it is in the DIRTY state, just return.
3135 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3136 * state and go to sleep or return.
3137 * If it is in any other state, go to sleep or return.
3139 * Synchronous forces are implemented with a signal variable. All callers
3140 * to force a given lsn to disk will wait on a the sv attached to the
3141 * specific in-core log. When given in-core log finally completes its
3142 * write to disk, that thread will wake up all threads waiting on the
3147 struct xfs_mount *mp,
3152 struct log *log = mp->m_log;
3153 struct xlog_in_core *iclog;
3154 int already_slept = 0;
3158 XFS_STATS_INC(xs_log_force);
3161 spin_lock(&log->l_icloglock);
3162 iclog = log->l_iclog;
3163 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3164 spin_unlock(&log->l_icloglock);
3165 return XFS_ERROR(EIO);
3169 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3170 iclog = iclog->ic_next;
3174 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3175 spin_unlock(&log->l_icloglock);
3179 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3181 * We sleep here if we haven't already slept (e.g.
3182 * this is the first time we've looked at the correct
3183 * iclog buf) and the buffer before us is going to
3184 * be sync'ed. The reason for this is that if we
3185 * are doing sync transactions here, by waiting for
3186 * the previous I/O to complete, we can allow a few
3187 * more transactions into this iclog before we close
3190 * Otherwise, we mark the buffer WANT_SYNC, and bump
3191 * up the refcnt so we can release the log (which
3192 * drops the ref count). The state switch keeps new
3193 * transaction commits from using this buffer. When
3194 * the current commits finish writing into the buffer,
3195 * the refcount will drop to zero and the buffer will
3198 if (!already_slept &&
3199 (iclog->ic_prev->ic_state &
3200 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3201 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3203 XFS_STATS_INC(xs_log_force_sleep);
3205 sv_wait(&iclog->ic_prev->ic_write_wait,
3206 PSWP, &log->l_icloglock, s);
3212 atomic_inc(&iclog->ic_refcnt);
3213 xlog_state_switch_iclogs(log, iclog, 0);
3214 spin_unlock(&log->l_icloglock);
3215 if (xlog_state_release_iclog(log, iclog))
3216 return XFS_ERROR(EIO);
3219 spin_lock(&log->l_icloglock);
3222 if ((flags & XFS_LOG_SYNC) && /* sleep */
3224 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3226 * Don't wait on completion if we know that we've
3227 * gotten a log write error.
3229 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3230 spin_unlock(&log->l_icloglock);
3231 return XFS_ERROR(EIO);
3233 XFS_STATS_INC(xs_log_force_sleep);
3234 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3236 * No need to grab the log lock here since we're
3237 * only deciding whether or not to return EIO
3238 * and the memory read should be atomic.
3240 if (iclog->ic_state & XLOG_STATE_IOERROR)
3241 return XFS_ERROR(EIO);
3245 } else { /* just return */
3246 spin_unlock(&log->l_icloglock);
3250 } while (iclog != log->l_iclog);
3252 spin_unlock(&log->l_icloglock);
3257 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3258 * about errors or whether the log was flushed or not. This is the normal
3259 * interface to use when trying to unpin items or move the log forward.
3269 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3271 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3272 "error %d returned.", error);
3277 * Called when we want to mark the current iclog as being ready to sync to
3281 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3283 assert_spin_locked(&log->l_icloglock);
3285 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3286 xlog_state_switch_iclogs(log, iclog, 0);
3288 ASSERT(iclog->ic_state &
3289 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3294 /*****************************************************************************
3298 *****************************************************************************
3302 * Free a used ticket when its refcount falls to zero.
3306 xlog_ticket_t *ticket)
3308 ASSERT(atomic_read(&ticket->t_ref) > 0);
3309 if (atomic_dec_and_test(&ticket->t_ref)) {
3310 sv_destroy(&ticket->t_wait);
3311 kmem_zone_free(xfs_log_ticket_zone, ticket);
3317 xlog_ticket_t *ticket)
3319 ASSERT(atomic_read(&ticket->t_ref) > 0);
3320 atomic_inc(&ticket->t_ref);
3325 * Allocate and initialise a new log ticket.
3327 STATIC xlog_ticket_t *
3336 struct xlog_ticket *tic;
3340 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3345 * Permanent reservations have up to 'cnt'-1 active log operations
3346 * in the log. A unit in this case is the amount of space for one
3347 * of these log operations. Normal reservations have a cnt of 1
3348 * and their unit amount is the total amount of space required.
3350 * The following lines of code account for non-transaction data
3351 * which occupy space in the on-disk log.
3353 * Normal form of a transaction is:
3354 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3355 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3357 * We need to account for all the leadup data and trailer data
3358 * around the transaction data.
3359 * And then we need to account for the worst case in terms of using
3361 * The worst case will happen if:
3362 * - the placement of the transaction happens to be such that the
3363 * roundoff is at its maximum
3364 * - the transaction data is synced before the commit record is synced
3365 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3366 * Therefore the commit record is in its own Log Record.
3367 * This can happen as the commit record is called with its
3368 * own region to xlog_write().
3369 * This then means that in the worst case, roundoff can happen for
3370 * the commit-rec as well.
3371 * The commit-rec is smaller than padding in this scenario and so it is
3372 * not added separately.
3375 /* for trans header */
3376 unit_bytes += sizeof(xlog_op_header_t);
3377 unit_bytes += sizeof(xfs_trans_header_t);
3380 unit_bytes += sizeof(xlog_op_header_t);
3383 * for LR headers - the space for data in an iclog is the size minus
3384 * the space used for the headers. If we use the iclog size, then we
3385 * undercalculate the number of headers required.
3387 * Furthermore - the addition of op headers for split-recs might
3388 * increase the space required enough to require more log and op
3389 * headers, so take that into account too.
3391 * IMPORTANT: This reservation makes the assumption that if this
3392 * transaction is the first in an iclog and hence has the LR headers
3393 * accounted to it, then the remaining space in the iclog is
3394 * exclusively for this transaction. i.e. if the transaction is larger
3395 * than the iclog, it will be the only thing in that iclog.
3396 * Fundamentally, this means we must pass the entire log vector to
3397 * xlog_write to guarantee this.
3399 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3400 num_headers = howmany(unit_bytes, iclog_space);
3402 /* for split-recs - ophdrs added when data split over LRs */
3403 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3405 /* add extra header reservations if we overrun */
3406 while (!num_headers ||
3407 howmany(unit_bytes, iclog_space) > num_headers) {
3408 unit_bytes += sizeof(xlog_op_header_t);
3411 unit_bytes += log->l_iclog_hsize * num_headers;
3413 /* for commit-rec LR header - note: padding will subsume the ophdr */
3414 unit_bytes += log->l_iclog_hsize;
3416 /* for roundoff padding for transaction data and one for commit record */
3417 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3418 log->l_mp->m_sb.sb_logsunit > 1) {
3419 /* log su roundoff */
3420 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3423 unit_bytes += 2*BBSIZE;
3426 atomic_set(&tic->t_ref, 1);
3427 tic->t_unit_res = unit_bytes;
3428 tic->t_curr_res = unit_bytes;
3431 tic->t_tid = random32();
3432 tic->t_clientid = client;
3433 tic->t_flags = XLOG_TIC_INITED;
3434 tic->t_trans_type = 0;
3435 if (xflags & XFS_LOG_PERM_RESERV)
3436 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3437 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3439 xlog_tic_reset_res(tic);
3445 /******************************************************************************
3447 * Log debug routines
3449 ******************************************************************************
3453 * Make sure that the destination ptr is within the valid data region of
3454 * one of the iclogs. This uses backup pointers stored in a different
3455 * part of the log in case we trash the log structure.
3458 xlog_verify_dest_ptr(
3465 for (i = 0; i < log->l_iclog_bufs; i++) {
3466 if (ptr >= log->l_iclog_bak[i] &&
3467 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3472 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3476 xlog_verify_grant_head(xlog_t *log, int equals)
3478 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3480 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3482 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3484 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3485 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3487 } /* xlog_verify_grant_head */
3489 /* check if it will fit */
3491 xlog_verify_tail_lsn(xlog_t *log,
3492 xlog_in_core_t *iclog,
3497 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3499 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3500 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3501 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3503 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3505 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3506 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3508 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3509 if (blocks < BTOBB(iclog->ic_offset) + 1)
3510 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3512 } /* xlog_verify_tail_lsn */
3515 * Perform a number of checks on the iclog before writing to disk.
3517 * 1. Make sure the iclogs are still circular
3518 * 2. Make sure we have a good magic number
3519 * 3. Make sure we don't have magic numbers in the data
3520 * 4. Check fields of each log operation header for:
3521 * A. Valid client identifier
3522 * B. tid ptr value falls in valid ptr space (user space code)
3523 * C. Length in log record header is correct according to the
3524 * individual operation headers within record.
3525 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3526 * log, check the preceding blocks of the physical log to make sure all
3527 * the cycle numbers agree with the current cycle number.
3530 xlog_verify_iclog(xlog_t *log,
3531 xlog_in_core_t *iclog,
3535 xlog_op_header_t *ophead;
3536 xlog_in_core_t *icptr;
3537 xlog_in_core_2_t *xhdr;
3539 xfs_caddr_t base_ptr;
3540 __psint_t field_offset;
3542 int len, i, j, k, op_len;
3545 /* check validity of iclog pointers */
3546 spin_lock(&log->l_icloglock);
3547 icptr = log->l_iclog;
3548 for (i=0; i < log->l_iclog_bufs; i++) {
3550 xlog_panic("xlog_verify_iclog: invalid ptr");
3551 icptr = icptr->ic_next;
3553 if (icptr != log->l_iclog)
3554 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3555 spin_unlock(&log->l_icloglock);
3557 /* check log magic numbers */
3558 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3559 xlog_panic("xlog_verify_iclog: invalid magic num");
3561 ptr = (xfs_caddr_t) &iclog->ic_header;
3562 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3564 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3565 xlog_panic("xlog_verify_iclog: unexpected magic num");
3569 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3570 ptr = iclog->ic_datap;
3572 ophead = (xlog_op_header_t *)ptr;
3573 xhdr = iclog->ic_data;
3574 for (i = 0; i < len; i++) {
3575 ophead = (xlog_op_header_t *)ptr;
3577 /* clientid is only 1 byte */
3578 field_offset = (__psint_t)
3579 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3580 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3581 clientid = ophead->oh_clientid;
3583 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3584 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3585 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3586 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3587 clientid = xlog_get_client_id(
3588 xhdr[j].hic_xheader.xh_cycle_data[k]);
3590 clientid = xlog_get_client_id(
3591 iclog->ic_header.h_cycle_data[idx]);
3594 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3595 cmn_err(CE_WARN, "xlog_verify_iclog: "
3596 "invalid clientid %d op 0x%p offset 0x%lx",
3597 clientid, ophead, (unsigned long)field_offset);
3600 field_offset = (__psint_t)
3601 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3602 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3603 op_len = be32_to_cpu(ophead->oh_len);
3605 idx = BTOBBT((__psint_t)&ophead->oh_len -
3606 (__psint_t)iclog->ic_datap);
3607 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3608 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3609 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3610 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3612 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3615 ptr += sizeof(xlog_op_header_t) + op_len;
3617 } /* xlog_verify_iclog */
3621 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3627 xlog_in_core_t *iclog, *ic;
3629 iclog = log->l_iclog;
3630 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3632 * Mark all the incore logs IOERROR.
3633 * From now on, no log flushes will result.
3637 ic->ic_state = XLOG_STATE_IOERROR;
3639 } while (ic != iclog);
3643 * Return non-zero, if state transition has already happened.
3649 * This is called from xfs_force_shutdown, when we're forcibly
3650 * shutting down the filesystem, typically because of an IO error.
3651 * Our main objectives here are to make sure that:
3652 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3653 * parties to find out, 'atomically'.
3654 * b. those who're sleeping on log reservations, pinned objects and
3655 * other resources get woken up, and be told the bad news.
3656 * c. nothing new gets queued up after (a) and (b) are done.
3657 * d. if !logerror, flush the iclogs to disk, then seal them off
3661 xfs_log_force_umount(
3662 struct xfs_mount *mp,
3672 * If this happens during log recovery, don't worry about
3673 * locking; the log isn't open for business yet.
3676 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3677 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3679 XFS_BUF_DONE(mp->m_sb_bp);
3684 * Somebody could've already done the hard work for us.
3685 * No need to get locks for this.
3687 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3688 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3693 * We must hold both the GRANT lock and the LOG lock,
3694 * before we mark the filesystem SHUTDOWN and wake
3695 * everybody up to tell the bad news.
3697 spin_lock(&log->l_icloglock);
3698 spin_lock(&log->l_grant_lock);
3699 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3701 XFS_BUF_DONE(mp->m_sb_bp);
3704 * This flag is sort of redundant because of the mount flag, but
3705 * it's good to maintain the separation between the log and the rest
3708 log->l_flags |= XLOG_IO_ERROR;
3711 * If we hit a log error, we want to mark all the iclogs IOERROR
3712 * while we're still holding the loglock.
3715 retval = xlog_state_ioerror(log);
3716 spin_unlock(&log->l_icloglock);
3719 * We don't want anybody waiting for log reservations
3720 * after this. That means we have to wake up everybody
3721 * queued up on reserve_headq as well as write_headq.
3722 * In addition, we make sure in xlog_{re}grant_log_space
3723 * that we don't enqueue anything once the SHUTDOWN flag
3724 * is set, and this action is protected by the GRANTLOCK.
3726 if ((tic = log->l_reserve_headq)) {
3728 sv_signal(&tic->t_wait);
3730 } while (tic != log->l_reserve_headq);
3733 if ((tic = log->l_write_headq)) {
3735 sv_signal(&tic->t_wait);
3737 } while (tic != log->l_write_headq);
3739 spin_unlock(&log->l_grant_lock);
3741 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3744 * Force the incore logs to disk before shutting the
3745 * log down completely.
3747 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3749 spin_lock(&log->l_icloglock);
3750 retval = xlog_state_ioerror(log);
3751 spin_unlock(&log->l_icloglock);
3754 * Wake up everybody waiting on xfs_log_force.
3755 * Callback all log item committed functions as if the
3756 * log writes were completed.
3758 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3760 #ifdef XFSERRORDEBUG
3762 xlog_in_core_t *iclog;
3764 spin_lock(&log->l_icloglock);
3765 iclog = log->l_iclog;
3767 ASSERT(iclog->ic_callback == 0);
3768 iclog = iclog->ic_next;
3769 } while (iclog != log->l_iclog);
3770 spin_unlock(&log->l_icloglock);
3773 /* return non-zero if log IOERROR transition had already happened */
3778 xlog_iclogs_empty(xlog_t *log)
3780 xlog_in_core_t *iclog;
3782 iclog = log->l_iclog;
3784 /* endianness does not matter here, zero is zero in
3787 if (iclog->ic_header.h_num_logops)
3789 iclog = iclog->ic_next;
3790 } while (iclog != log->l_iclog);