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,
98 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
99 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
100 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
101 int count, boolean_t syncing);
102 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
105 #define xlog_verify_dest_ptr(a,b)
106 #define xlog_verify_grant_head(a,b)
107 #define xlog_verify_iclog(a,b,c,d)
108 #define xlog_verify_tail_lsn(a,b,c)
111 STATIC int xlog_iclogs_empty(xlog_t *log);
115 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
119 tic->t_prev = (*qp)->t_prev;
120 (*qp)->t_prev->t_next = tic;
123 tic->t_prev = tic->t_next = tic;
127 tic->t_flags |= XLOG_TIC_IN_Q;
131 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
133 if (tic == tic->t_next) {
137 tic->t_next->t_prev = tic->t_prev;
138 tic->t_prev->t_next = tic->t_next;
141 tic->t_next = tic->t_prev = NULL;
142 tic->t_flags &= ~XLOG_TIC_IN_Q;
146 xlog_grant_sub_space(struct log *log, int bytes)
148 log->l_grant_write_bytes -= bytes;
149 if (log->l_grant_write_bytes < 0) {
150 log->l_grant_write_bytes += log->l_logsize;
151 log->l_grant_write_cycle--;
154 log->l_grant_reserve_bytes -= bytes;
155 if ((log)->l_grant_reserve_bytes < 0) {
156 log->l_grant_reserve_bytes += log->l_logsize;
157 log->l_grant_reserve_cycle--;
163 xlog_grant_add_space_write(struct log *log, int bytes)
165 int tmp = log->l_logsize - log->l_grant_write_bytes;
167 log->l_grant_write_bytes += bytes;
169 log->l_grant_write_cycle++;
170 log->l_grant_write_bytes = bytes - tmp;
175 xlog_grant_add_space_reserve(struct log *log, int bytes)
177 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
179 log->l_grant_reserve_bytes += bytes;
181 log->l_grant_reserve_cycle++;
182 log->l_grant_reserve_bytes = bytes - tmp;
187 xlog_grant_add_space(struct log *log, int bytes)
189 xlog_grant_add_space_write(log, bytes);
190 xlog_grant_add_space_reserve(log, bytes);
194 xlog_tic_reset_res(xlog_ticket_t *tic)
197 tic->t_res_arr_sum = 0;
198 tic->t_res_num_ophdrs = 0;
202 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
204 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
205 /* add to overflow and start again */
206 tic->t_res_o_flow += tic->t_res_arr_sum;
208 tic->t_res_arr_sum = 0;
211 tic->t_res_arr[tic->t_res_num].r_len = len;
212 tic->t_res_arr[tic->t_res_num].r_type = type;
213 tic->t_res_arr_sum += len;
220 * 1. currblock field gets updated at startup and after in-core logs
221 * marked as with WANT_SYNC.
225 * This routine is called when a user of a log manager ticket is done with
226 * the reservation. If the ticket was ever used, then a commit record for
227 * the associated transaction is written out as a log operation header with
228 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
229 * a given ticket. If the ticket was one with a permanent reservation, then
230 * a few operations are done differently. Permanent reservation tickets by
231 * default don't release the reservation. They just commit the current
232 * transaction with the belief that the reservation is still needed. A flag
233 * must be passed in before permanent reservations are actually released.
234 * When these type of tickets are not released, they need to be set into
235 * the inited state again. By doing this, a start record will be written
236 * out when the next write occurs.
240 struct xfs_mount *mp,
241 struct xlog_ticket *ticket,
242 struct xlog_in_core **iclog,
245 struct log *log = mp->m_log;
248 if (XLOG_FORCED_SHUTDOWN(log) ||
250 * If nothing was ever written, don't write out commit record.
251 * If we get an error, just continue and give back the log ticket.
253 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
254 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
255 lsn = (xfs_lsn_t) -1;
256 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
257 flags |= XFS_LOG_REL_PERM_RESERV;
262 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
263 (flags & XFS_LOG_REL_PERM_RESERV)) {
264 trace_xfs_log_done_nonperm(log, ticket);
267 * Release ticket if not permanent reservation or a specific
268 * request has been made to release a permanent reservation.
270 xlog_ungrant_log_space(log, ticket);
271 xfs_log_ticket_put(ticket);
273 trace_xfs_log_done_perm(log, ticket);
275 xlog_regrant_reserve_log_space(log, ticket);
276 /* If this ticket was a permanent reservation and we aren't
277 * trying to release it, reset the inited flags; so next time
278 * we write, a start record will be written out.
280 ticket->t_flags |= XLOG_TIC_INITED;
287 * Attaches a new iclog I/O completion callback routine during
288 * transaction commit. If the log is in error state, a non-zero
289 * return code is handed back and the caller is responsible for
290 * executing the callback at an appropriate time.
294 struct xfs_mount *mp,
295 struct xlog_in_core *iclog,
296 xfs_log_callback_t *cb)
300 spin_lock(&iclog->ic_callback_lock);
301 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
303 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
304 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
306 *(iclog->ic_callback_tail) = cb;
307 iclog->ic_callback_tail = &(cb->cb_next);
309 spin_unlock(&iclog->ic_callback_lock);
314 xfs_log_release_iclog(
315 struct xfs_mount *mp,
316 struct xlog_in_core *iclog)
318 if (xlog_state_release_iclog(mp->m_log, iclog)) {
319 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
327 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
328 * to the reservation.
329 * 2. Potentially, push buffers at tail of log to disk.
331 * Each reservation is going to reserve extra space for a log record header.
332 * When writes happen to the on-disk log, we don't subtract the length of the
333 * log record header from any reservation. By wasting space in each
334 * reservation, we prevent over allocation problems.
338 struct xfs_mount *mp,
341 struct xlog_ticket **ticket,
346 struct log *log = mp->m_log;
347 struct xlog_ticket *internal_ticket;
350 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
351 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
353 if (XLOG_FORCED_SHUTDOWN(log))
354 return XFS_ERROR(EIO);
356 XFS_STATS_INC(xs_try_logspace);
359 if (*ticket != NULL) {
360 ASSERT(flags & XFS_LOG_PERM_RESERV);
361 internal_ticket = *ticket;
363 trace_xfs_log_reserve(log, internal_ticket);
365 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
366 retval = xlog_regrant_write_log_space(log, internal_ticket);
368 /* may sleep if need to allocate more tickets */
369 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
371 if (!internal_ticket)
372 return XFS_ERROR(ENOMEM);
373 internal_ticket->t_trans_type = t_type;
374 *ticket = internal_ticket;
376 trace_xfs_log_reserve(log, internal_ticket);
378 xlog_grant_push_ail(mp,
379 (internal_ticket->t_unit_res *
380 internal_ticket->t_cnt));
381 retval = xlog_grant_log_space(log, internal_ticket);
385 } /* xfs_log_reserve */
389 * Mount a log filesystem
391 * mp - ubiquitous xfs mount point structure
392 * log_target - buftarg of on-disk log device
393 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
394 * num_bblocks - Number of BBSIZE blocks in on-disk log
396 * Return error or zero.
401 xfs_buftarg_t *log_target,
402 xfs_daddr_t blk_offset,
407 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
408 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
411 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
413 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
416 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
417 if (IS_ERR(mp->m_log)) {
418 error = -PTR_ERR(mp->m_log);
423 * Initialize the AIL now we have a log.
425 error = xfs_trans_ail_init(mp);
427 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
430 mp->m_log->l_ailp = mp->m_ail;
433 * skip log recovery on a norecovery mount. pretend it all
436 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
437 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
440 mp->m_flags &= ~XFS_MOUNT_RDONLY;
442 error = xlog_recover(mp->m_log);
445 mp->m_flags |= XFS_MOUNT_RDONLY;
447 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
448 goto out_destroy_ail;
452 /* Normal transactions can now occur */
453 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
458 xfs_trans_ail_destroy(mp);
460 xlog_dealloc_log(mp->m_log);
466 * Finish the recovery of the file system. This is separate from
467 * the xfs_log_mount() call, because it depends on the code in
468 * xfs_mountfs() to read in the root and real-time bitmap inodes
469 * between calling xfs_log_mount() and here.
471 * mp - ubiquitous xfs mount point structure
474 xfs_log_mount_finish(xfs_mount_t *mp)
478 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
479 error = xlog_recover_finish(mp->m_log);
482 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
489 * Final log writes as part of unmount.
491 * Mark the filesystem clean as unmount happens. Note that during relocation
492 * this routine needs to be executed as part of source-bag while the
493 * deallocation must not be done until source-end.
497 * Unmount record used to have a string "Unmount filesystem--" in the
498 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
499 * We just write the magic number now since that particular field isn't
500 * currently architecture converted and "nUmount" is a bit foo.
501 * As far as I know, there weren't any dependencies on the old behaviour.
505 xfs_log_unmount_write(xfs_mount_t *mp)
507 xlog_t *log = mp->m_log;
508 xlog_in_core_t *iclog;
510 xlog_in_core_t *first_iclog;
512 xlog_ticket_t *tic = NULL;
517 * Don't write out unmount record on read-only mounts.
518 * Or, if we are doing a forced umount (typically because of IO errors).
520 if (mp->m_flags & XFS_MOUNT_RDONLY)
523 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
524 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
527 first_iclog = iclog = log->l_iclog;
529 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
530 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
531 ASSERT(iclog->ic_offset == 0);
533 iclog = iclog->ic_next;
534 } while (iclog != first_iclog);
536 if (! (XLOG_FORCED_SHUTDOWN(log))) {
537 error = xfs_log_reserve(mp, 600, 1, &tic,
538 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
540 /* the data section must be 32 bit size aligned */
544 __uint32_t pad2; /* may as well make it 64 bits */
546 .magic = XLOG_UNMOUNT_TYPE,
548 struct xfs_log_iovec reg = {
549 .i_addr = (void *)&magic,
550 .i_len = sizeof(magic),
551 .i_type = XLOG_REG_TYPE_UNMOUNT,
553 struct xfs_log_vec vec = {
558 /* remove inited flag */
560 error = xlog_write(log, &vec, tic, &lsn,
561 NULL, XLOG_UNMOUNT_TRANS);
563 * At this point, we're umounting anyway,
564 * so there's no point in transitioning log state
565 * to IOERROR. Just continue...
570 xfs_fs_cmn_err(CE_ALERT, mp,
571 "xfs_log_unmount: unmount record failed");
575 spin_lock(&log->l_icloglock);
576 iclog = log->l_iclog;
577 atomic_inc(&iclog->ic_refcnt);
578 xlog_state_want_sync(log, iclog);
579 spin_unlock(&log->l_icloglock);
580 error = xlog_state_release_iclog(log, iclog);
582 spin_lock(&log->l_icloglock);
583 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
584 iclog->ic_state == XLOG_STATE_DIRTY)) {
585 if (!XLOG_FORCED_SHUTDOWN(log)) {
586 sv_wait(&iclog->ic_force_wait, PMEM,
587 &log->l_icloglock, s);
589 spin_unlock(&log->l_icloglock);
592 spin_unlock(&log->l_icloglock);
595 trace_xfs_log_umount_write(log, tic);
596 xlog_ungrant_log_space(log, tic);
597 xfs_log_ticket_put(tic);
601 * We're already in forced_shutdown mode, couldn't
602 * even attempt to write out the unmount transaction.
604 * Go through the motions of sync'ing and releasing
605 * the iclog, even though no I/O will actually happen,
606 * we need to wait for other log I/Os that may already
607 * be in progress. Do this as a separate section of
608 * code so we'll know if we ever get stuck here that
609 * we're in this odd situation of trying to unmount
610 * a file system that went into forced_shutdown as
611 * the result of an unmount..
613 spin_lock(&log->l_icloglock);
614 iclog = log->l_iclog;
615 atomic_inc(&iclog->ic_refcnt);
617 xlog_state_want_sync(log, iclog);
618 spin_unlock(&log->l_icloglock);
619 error = xlog_state_release_iclog(log, iclog);
621 spin_lock(&log->l_icloglock);
623 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
624 || iclog->ic_state == XLOG_STATE_DIRTY
625 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
627 sv_wait(&iclog->ic_force_wait, PMEM,
628 &log->l_icloglock, s);
630 spin_unlock(&log->l_icloglock);
635 } /* xfs_log_unmount_write */
638 * Deallocate log structures for unmount/relocation.
640 * We need to stop the aild from running before we destroy
641 * and deallocate the log as the aild references the log.
644 xfs_log_unmount(xfs_mount_t *mp)
646 xfs_trans_ail_destroy(mp);
647 xlog_dealloc_log(mp->m_log);
652 struct xfs_mount *mp,
653 struct xfs_log_item *item,
655 struct xfs_item_ops *ops)
657 item->li_mountp = mp;
658 item->li_ailp = mp->m_ail;
659 item->li_type = type;
664 * Write region vectors to log. The write happens using the space reservation
665 * of the ticket (tic). It is not a requirement that all writes for a given
666 * transaction occur with one call to xfs_log_write(). However, it is important
667 * to note that the transaction reservation code makes an assumption about the
668 * number of log headers a transaction requires that may be violated if you
669 * don't pass all the transaction vectors in one call....
673 struct xfs_mount *mp,
674 struct xfs_log_iovec reg[],
676 struct xlog_ticket *tic,
677 xfs_lsn_t *start_lsn)
679 struct log *log = mp->m_log;
681 struct xfs_log_vec vec = {
682 .lv_niovecs = nentries,
686 if (XLOG_FORCED_SHUTDOWN(log))
687 return XFS_ERROR(EIO);
689 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
691 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
696 xfs_log_move_tail(xfs_mount_t *mp,
700 xlog_t *log = mp->m_log;
701 int need_bytes, free_bytes, cycle, bytes;
703 if (XLOG_FORCED_SHUTDOWN(log))
707 /* needed since sync_lsn is 64 bits */
708 spin_lock(&log->l_icloglock);
709 tail_lsn = log->l_last_sync_lsn;
710 spin_unlock(&log->l_icloglock);
713 spin_lock(&log->l_grant_lock);
715 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
719 log->l_tail_lsn = tail_lsn;
722 if ((tic = log->l_write_headq)) {
724 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
725 panic("Recovery problem");
727 cycle = log->l_grant_write_cycle;
728 bytes = log->l_grant_write_bytes;
729 free_bytes = xlog_space_left(log, cycle, bytes);
731 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
733 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
736 free_bytes -= tic->t_unit_res;
737 sv_signal(&tic->t_wait);
739 } while (tic != log->l_write_headq);
741 if ((tic = log->l_reserve_headq)) {
743 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
744 panic("Recovery problem");
746 cycle = log->l_grant_reserve_cycle;
747 bytes = log->l_grant_reserve_bytes;
748 free_bytes = xlog_space_left(log, cycle, bytes);
750 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
751 need_bytes = tic->t_unit_res*tic->t_cnt;
753 need_bytes = tic->t_unit_res;
754 if (free_bytes < need_bytes && tail_lsn != 1)
757 free_bytes -= need_bytes;
758 sv_signal(&tic->t_wait);
760 } while (tic != log->l_reserve_headq);
762 spin_unlock(&log->l_grant_lock);
763 } /* xfs_log_move_tail */
766 * Determine if we have a transaction that has gone to disk
767 * that needs to be covered. To begin the transition to the idle state
768 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
769 * If we are then in a state where covering is needed, the caller is informed
770 * that dummy transactions are required to move the log into the idle state.
772 * Because this is called as part of the sync process, we should also indicate
773 * that dummy transactions should be issued in anything but the covered or
774 * idle states. This ensures that the log tail is accurately reflected in
775 * the log at the end of the sync, hence if a crash occurrs avoids replay
776 * of transactions where the metadata is already on disk.
779 xfs_log_need_covered(xfs_mount_t *mp)
782 xlog_t *log = mp->m_log;
784 if (!xfs_fs_writable(mp))
787 spin_lock(&log->l_icloglock);
788 switch (log->l_covered_state) {
789 case XLOG_STATE_COVER_DONE:
790 case XLOG_STATE_COVER_DONE2:
791 case XLOG_STATE_COVER_IDLE:
793 case XLOG_STATE_COVER_NEED:
794 case XLOG_STATE_COVER_NEED2:
795 if (!xfs_trans_ail_tail(log->l_ailp) &&
796 xlog_iclogs_empty(log)) {
797 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
798 log->l_covered_state = XLOG_STATE_COVER_DONE;
800 log->l_covered_state = XLOG_STATE_COVER_DONE2;
807 spin_unlock(&log->l_icloglock);
811 /******************************************************************************
815 ******************************************************************************
818 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
819 * The log manager must keep track of the last LR which was committed
820 * to disk. The lsn of this LR will become the new tail_lsn whenever
821 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
822 * the situation where stuff could be written into the log but nothing
823 * was ever in the AIL when asked. Eventually, we panic since the
824 * tail hits the head.
826 * We may be holding the log iclog lock upon entering this routine.
829 xlog_assign_tail_lsn(xfs_mount_t *mp)
832 xlog_t *log = mp->m_log;
834 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
835 spin_lock(&log->l_grant_lock);
837 log->l_tail_lsn = tail_lsn;
839 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
841 spin_unlock(&log->l_grant_lock);
844 } /* xlog_assign_tail_lsn */
848 * Return the space in the log between the tail and the head. The head
849 * is passed in the cycle/bytes formal parms. In the special case where
850 * the reserve head has wrapped passed the tail, this calculation is no
851 * longer valid. In this case, just return 0 which means there is no space
852 * in the log. This works for all places where this function is called
853 * with the reserve head. Of course, if the write head were to ever
854 * wrap the tail, we should blow up. Rather than catch this case here,
855 * we depend on other ASSERTions in other parts of the code. XXXmiken
857 * This code also handles the case where the reservation head is behind
858 * the tail. The details of this case are described below, but the end
859 * result is that we return the size of the log as the amount of space left.
862 xlog_space_left(xlog_t *log, int cycle, int bytes)
868 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
869 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
870 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
871 free_bytes = log->l_logsize - (bytes - tail_bytes);
872 } else if ((tail_cycle + 1) < cycle) {
874 } else if (tail_cycle < cycle) {
875 ASSERT(tail_cycle == (cycle - 1));
876 free_bytes = tail_bytes - bytes;
879 * The reservation head is behind the tail.
880 * In this case we just want to return the size of the
881 * log as the amount of space left.
883 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
884 "xlog_space_left: head behind tail\n"
885 " tail_cycle = %d, tail_bytes = %d\n"
886 " GH cycle = %d, GH bytes = %d",
887 tail_cycle, tail_bytes, cycle, bytes);
889 free_bytes = log->l_logsize;
892 } /* xlog_space_left */
896 * Log function which is called when an io completes.
898 * The log manager needs its own routine, in order to control what
899 * happens with the buffer after the write completes.
902 xlog_iodone(xfs_buf_t *bp)
904 xlog_in_core_t *iclog;
908 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
909 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
910 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
915 * If the _XFS_BARRIER_FAILED flag was set by a lower
916 * layer, it means the underlying device no longer supports
917 * barrier I/O. Warn loudly and turn off barriers.
919 if (bp->b_flags & _XFS_BARRIER_FAILED) {
920 bp->b_flags &= ~_XFS_BARRIER_FAILED;
921 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
922 xfs_fs_cmn_err(CE_WARN, l->l_mp,
923 "xlog_iodone: Barriers are no longer supported"
924 " by device. Disabling barriers\n");
928 * Race to shutdown the filesystem if we see an error.
930 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
931 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
932 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
934 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
936 * This flag will be propagated to the trans-committed
937 * callback routines to let them know that the log-commit
940 aborted = XFS_LI_ABORTED;
941 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
942 aborted = XFS_LI_ABORTED;
945 /* log I/O is always issued ASYNC */
946 ASSERT(XFS_BUF_ISASYNC(bp));
947 xlog_state_done_syncing(iclog, aborted);
949 * do not reference the buffer (bp) here as we could race
950 * with it being freed after writing the unmount record to the
957 * Return size of each in-core log record buffer.
959 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
961 * If the filesystem blocksize is too large, we may need to choose a
962 * larger size since the directory code currently logs entire blocks.
966 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
972 if (mp->m_logbufs <= 0)
973 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
975 log->l_iclog_bufs = mp->m_logbufs;
978 * Buffer size passed in from mount system call.
980 if (mp->m_logbsize > 0) {
981 size = log->l_iclog_size = mp->m_logbsize;
982 log->l_iclog_size_log = 0;
984 log->l_iclog_size_log++;
988 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
989 /* # headers = size / 32k
990 * one header holds cycles from 32k of data
993 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
994 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
996 log->l_iclog_hsize = xhdrs << BBSHIFT;
997 log->l_iclog_heads = xhdrs;
999 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1000 log->l_iclog_hsize = BBSIZE;
1001 log->l_iclog_heads = 1;
1006 /* All machines use 32kB buffers by default. */
1007 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1008 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1010 /* the default log size is 16k or 32k which is one header sector */
1011 log->l_iclog_hsize = BBSIZE;
1012 log->l_iclog_heads = 1;
1015 /* are we being asked to make the sizes selected above visible? */
1016 if (mp->m_logbufs == 0)
1017 mp->m_logbufs = log->l_iclog_bufs;
1018 if (mp->m_logbsize == 0)
1019 mp->m_logbsize = log->l_iclog_size;
1020 } /* xlog_get_iclog_buffer_size */
1024 * This routine initializes some of the log structure for a given mount point.
1025 * Its primary purpose is to fill in enough, so recovery can occur. However,
1026 * some other stuff may be filled in too.
1029 xlog_alloc_log(xfs_mount_t *mp,
1030 xfs_buftarg_t *log_target,
1031 xfs_daddr_t blk_offset,
1035 xlog_rec_header_t *head;
1036 xlog_in_core_t **iclogp;
1037 xlog_in_core_t *iclog, *prev_iclog=NULL;
1043 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1045 xlog_warn("XFS: Log allocation failed: No memory!");
1050 log->l_targ = log_target;
1051 log->l_logsize = BBTOB(num_bblks);
1052 log->l_logBBstart = blk_offset;
1053 log->l_logBBsize = num_bblks;
1054 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1055 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1057 log->l_prev_block = -1;
1058 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1059 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1060 log->l_last_sync_lsn = log->l_tail_lsn;
1061 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1062 log->l_grant_reserve_cycle = 1;
1063 log->l_grant_write_cycle = 1;
1065 error = EFSCORRUPTED;
1066 if (xfs_sb_version_hassector(&mp->m_sb)) {
1067 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1068 if (log->l_sectbb_log < 0 ||
1069 log->l_sectbb_log > mp->m_sectbb_log) {
1070 xlog_warn("XFS: Log sector size (0x%x) out of range.",
1075 /* for larger sector sizes, must have v2 or external log */
1076 if (log->l_sectbb_log != 0 &&
1077 (log->l_logBBstart != 0 &&
1078 !xfs_sb_version_haslogv2(&mp->m_sb))) {
1079 xlog_warn("XFS: log sector size (0x%x) invalid "
1080 "for configuration.", log->l_sectbb_log);
1083 if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1084 xlog_warn("XFS: Log sector log (0x%x) too small.",
1085 mp->m_sb.sb_logsectlog);
1089 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1091 xlog_get_iclog_buffer_size(mp, log);
1094 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1097 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1098 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1099 ASSERT(XFS_BUF_ISBUSY(bp));
1100 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1103 spin_lock_init(&log->l_icloglock);
1104 spin_lock_init(&log->l_grant_lock);
1105 sv_init(&log->l_flush_wait, 0, "flush_wait");
1107 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1108 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1110 iclogp = &log->l_iclog;
1112 * The amount of memory to allocate for the iclog structure is
1113 * rather funky due to the way the structure is defined. It is
1114 * done this way so that we can use different sizes for machines
1115 * with different amounts of memory. See the definition of
1116 * xlog_in_core_t in xfs_log_priv.h for details.
1118 iclogsize = log->l_iclog_size;
1119 ASSERT(log->l_iclog_size >= 4096);
1120 for (i=0; i < log->l_iclog_bufs; i++) {
1121 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1123 goto out_free_iclog;
1126 iclog->ic_prev = prev_iclog;
1129 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1131 goto out_free_iclog;
1132 if (!XFS_BUF_CPSEMA(bp))
1134 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1135 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1137 iclog->ic_data = bp->b_addr;
1139 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1141 head = &iclog->ic_header;
1142 memset(head, 0, sizeof(xlog_rec_header_t));
1143 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1144 head->h_version = cpu_to_be32(
1145 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1146 head->h_size = cpu_to_be32(log->l_iclog_size);
1148 head->h_fmt = cpu_to_be32(XLOG_FMT);
1149 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1151 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1152 iclog->ic_state = XLOG_STATE_ACTIVE;
1153 iclog->ic_log = log;
1154 atomic_set(&iclog->ic_refcnt, 0);
1155 spin_lock_init(&iclog->ic_callback_lock);
1156 iclog->ic_callback_tail = &(iclog->ic_callback);
1157 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1159 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1160 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1161 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1162 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1164 iclogp = &iclog->ic_next;
1166 *iclogp = log->l_iclog; /* complete ring */
1167 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1172 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1173 prev_iclog = iclog->ic_next;
1175 sv_destroy(&iclog->ic_force_wait);
1176 sv_destroy(&iclog->ic_write_wait);
1177 xfs_buf_free(iclog->ic_bp);
1181 spinlock_destroy(&log->l_icloglock);
1182 spinlock_destroy(&log->l_grant_lock);
1183 xfs_buf_free(log->l_xbuf);
1187 return ERR_PTR(-error);
1188 } /* xlog_alloc_log */
1192 * Write out the commit record of a transaction associated with the given
1193 * ticket. Return the lsn of the commit record.
1198 struct xlog_ticket *ticket,
1199 struct xlog_in_core **iclog,
1200 xfs_lsn_t *commitlsnp)
1202 struct xfs_mount *mp = log->l_mp;
1204 struct xfs_log_iovec reg = {
1207 .i_type = XLOG_REG_TYPE_COMMIT,
1209 struct xfs_log_vec vec = {
1214 ASSERT_ALWAYS(iclog);
1215 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1218 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1223 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1224 * log space. This code pushes on the lsn which would supposedly free up
1225 * the 25% which we want to leave free. We may need to adopt a policy which
1226 * pushes on an lsn which is further along in the log once we reach the high
1227 * water mark. In this manner, we would be creating a low water mark.
1230 xlog_grant_push_ail(xfs_mount_t *mp,
1233 xlog_t *log = mp->m_log; /* pointer to the log */
1234 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1235 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1236 int free_blocks; /* free blocks left to write to */
1237 int free_bytes; /* free bytes left to write to */
1238 int threshold_block; /* block in lsn we'd like to be at */
1239 int threshold_cycle; /* lsn cycle we'd like to be at */
1242 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1244 spin_lock(&log->l_grant_lock);
1245 free_bytes = xlog_space_left(log,
1246 log->l_grant_reserve_cycle,
1247 log->l_grant_reserve_bytes);
1248 tail_lsn = log->l_tail_lsn;
1249 free_blocks = BTOBBT(free_bytes);
1252 * Set the threshold for the minimum number of free blocks in the
1253 * log to the maximum of what the caller needs, one quarter of the
1254 * log, and 256 blocks.
1256 free_threshold = BTOBB(need_bytes);
1257 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1258 free_threshold = MAX(free_threshold, 256);
1259 if (free_blocks < free_threshold) {
1260 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1261 threshold_cycle = CYCLE_LSN(tail_lsn);
1262 if (threshold_block >= log->l_logBBsize) {
1263 threshold_block -= log->l_logBBsize;
1264 threshold_cycle += 1;
1266 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1268 /* Don't pass in an lsn greater than the lsn of the last
1269 * log record known to be on disk.
1271 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1272 threshold_lsn = log->l_last_sync_lsn;
1274 spin_unlock(&log->l_grant_lock);
1277 * Get the transaction layer to kick the dirty buffers out to
1278 * disk asynchronously. No point in trying to do this if
1279 * the filesystem is shutting down.
1281 if (threshold_lsn &&
1282 !XLOG_FORCED_SHUTDOWN(log))
1283 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1284 } /* xlog_grant_push_ail */
1287 * The bdstrat callback function for log bufs. This gives us a central
1288 * place to trap bufs in case we get hit by a log I/O error and need to
1289 * shutdown. Actually, in practice, even when we didn't get a log error,
1290 * we transition the iclogs to IOERROR state *after* flushing all existing
1291 * iclogs to disk. This is because we don't want anymore new transactions to be
1292 * started or completed afterwards.
1298 struct xlog_in_core *iclog;
1300 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1301 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1302 XFS_BUF_ERROR(bp, EIO);
1306 * It would seem logical to return EIO here, but we rely on
1307 * the log state machine to propagate I/O errors instead of
1313 bp->b_flags |= _XBF_RUN_QUEUES;
1314 xfs_buf_iorequest(bp);
1319 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1320 * fashion. Previously, we should have moved the current iclog
1321 * ptr in the log to point to the next available iclog. This allows further
1322 * write to continue while this code syncs out an iclog ready to go.
1323 * Before an in-core log can be written out, the data section must be scanned
1324 * to save away the 1st word of each BBSIZE block into the header. We replace
1325 * it with the current cycle count. Each BBSIZE block is tagged with the
1326 * cycle count because there in an implicit assumption that drives will
1327 * guarantee that entire 512 byte blocks get written at once. In other words,
1328 * we can't have part of a 512 byte block written and part not written. By
1329 * tagging each block, we will know which blocks are valid when recovering
1330 * after an unclean shutdown.
1332 * This routine is single threaded on the iclog. No other thread can be in
1333 * this routine with the same iclog. Changing contents of iclog can there-
1334 * fore be done without grabbing the state machine lock. Updating the global
1335 * log will require grabbing the lock though.
1337 * The entire log manager uses a logical block numbering scheme. Only
1338 * log_sync (and then only bwrite()) know about the fact that the log may
1339 * not start with block zero on a given device. The log block start offset
1340 * is added immediately before calling bwrite().
1344 xlog_sync(xlog_t *log,
1345 xlog_in_core_t *iclog)
1347 xfs_caddr_t dptr; /* pointer to byte sized element */
1350 uint count; /* byte count of bwrite */
1351 uint count_init; /* initial count before roundup */
1352 int roundoff; /* roundoff to BB or stripe */
1353 int split = 0; /* split write into two regions */
1355 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1357 XFS_STATS_INC(xs_log_writes);
1358 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1360 /* Add for LR header */
1361 count_init = log->l_iclog_hsize + iclog->ic_offset;
1363 /* Round out the log write size */
1364 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1365 /* we have a v2 stripe unit to use */
1366 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1368 count = BBTOB(BTOBB(count_init));
1370 roundoff = count - count_init;
1371 ASSERT(roundoff >= 0);
1372 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1373 roundoff < log->l_mp->m_sb.sb_logsunit)
1375 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1376 roundoff < BBTOB(1)));
1378 /* move grant heads by roundoff in sync */
1379 spin_lock(&log->l_grant_lock);
1380 xlog_grant_add_space(log, roundoff);
1381 spin_unlock(&log->l_grant_lock);
1383 /* put cycle number in every block */
1384 xlog_pack_data(log, iclog, roundoff);
1386 /* real byte length */
1388 iclog->ic_header.h_len =
1389 cpu_to_be32(iclog->ic_offset + roundoff);
1391 iclog->ic_header.h_len =
1392 cpu_to_be32(iclog->ic_offset);
1396 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1397 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1398 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1400 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1402 /* Do we need to split this write into 2 parts? */
1403 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1404 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1405 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1406 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1408 iclog->ic_bwritecnt = 1;
1410 XFS_BUF_SET_COUNT(bp, count);
1411 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1412 XFS_BUF_ZEROFLAGS(bp);
1415 bp->b_flags |= XBF_LOG_BUFFER;
1417 * Do an ordered write for the log block.
1418 * Its unnecessary to flush the first split block in the log wrap case.
1420 if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1421 XFS_BUF_ORDERED(bp);
1423 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1424 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1426 xlog_verify_iclog(log, iclog, count, B_TRUE);
1428 /* account for log which doesn't start at block #0 */
1429 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1431 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1436 if ((error = xlog_bdstrat(bp))) {
1437 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1442 bp = iclog->ic_log->l_xbuf;
1443 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1445 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1446 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1447 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1448 (__psint_t)count), split);
1449 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1450 XFS_BUF_ZEROFLAGS(bp);
1453 bp->b_flags |= XBF_LOG_BUFFER;
1454 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1455 XFS_BUF_ORDERED(bp);
1456 dptr = XFS_BUF_PTR(bp);
1458 * Bump the cycle numbers at the start of each block
1459 * since this part of the buffer is at the start of
1460 * a new cycle. Watch out for the header magic number
1463 for (i = 0; i < split; i += BBSIZE) {
1464 be32_add_cpu((__be32 *)dptr, 1);
1465 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1466 be32_add_cpu((__be32 *)dptr, 1);
1470 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1471 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1473 /* account for internal log which doesn't start at block #0 */
1474 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1476 if ((error = xlog_bdstrat(bp))) {
1477 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1478 bp, XFS_BUF_ADDR(bp));
1487 * Deallocate a log structure
1490 xlog_dealloc_log(xlog_t *log)
1492 xlog_in_core_t *iclog, *next_iclog;
1495 iclog = log->l_iclog;
1496 for (i=0; i<log->l_iclog_bufs; i++) {
1497 sv_destroy(&iclog->ic_force_wait);
1498 sv_destroy(&iclog->ic_write_wait);
1499 xfs_buf_free(iclog->ic_bp);
1500 next_iclog = iclog->ic_next;
1504 spinlock_destroy(&log->l_icloglock);
1505 spinlock_destroy(&log->l_grant_lock);
1507 xfs_buf_free(log->l_xbuf);
1508 log->l_mp->m_log = NULL;
1510 } /* xlog_dealloc_log */
1513 * Update counters atomically now that memcpy is done.
1517 xlog_state_finish_copy(xlog_t *log,
1518 xlog_in_core_t *iclog,
1522 spin_lock(&log->l_icloglock);
1524 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1525 iclog->ic_offset += copy_bytes;
1527 spin_unlock(&log->l_icloglock);
1528 } /* xlog_state_finish_copy */
1534 * print out info relating to regions written which consume
1538 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1541 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1543 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1544 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1565 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1608 xfs_fs_cmn_err(CE_WARN, mp,
1609 "xfs_log_write: reservation summary:\n"
1610 " trans type = %s (%u)\n"
1611 " unit res = %d bytes\n"
1612 " current res = %d bytes\n"
1613 " total reg = %u bytes (o/flow = %u bytes)\n"
1614 " ophdrs = %u (ophdr space = %u bytes)\n"
1615 " ophdr + reg = %u bytes\n"
1616 " num regions = %u\n",
1617 ((ticket->t_trans_type <= 0 ||
1618 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1619 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1620 ticket->t_trans_type,
1623 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1624 ticket->t_res_num_ophdrs, ophdr_spc,
1625 ticket->t_res_arr_sum +
1626 ticket->t_res_o_flow + ophdr_spc,
1629 for (i = 0; i < ticket->t_res_num; i++) {
1630 uint r_type = ticket->t_res_arr[i].r_type;
1632 "region[%u]: %s - %u bytes\n",
1634 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1635 "bad-rtype" : res_type_str[r_type-1]),
1636 ticket->t_res_arr[i].r_len);
1641 * Calculate the potential space needed by the log vector. Each region gets
1642 * its own xlog_op_header_t and may need to be double word aligned.
1645 xlog_write_calc_vec_length(
1646 struct xlog_ticket *ticket,
1647 struct xfs_log_vec *log_vector)
1649 struct xfs_log_vec *lv;
1654 /* acct for start rec of xact */
1655 if (ticket->t_flags & XLOG_TIC_INITED)
1658 for (lv = log_vector; lv; lv = lv->lv_next) {
1659 headers += lv->lv_niovecs;
1661 for (i = 0; i < lv->lv_niovecs; i++) {
1662 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1665 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1669 ticket->t_res_num_ophdrs += headers;
1670 len += headers * sizeof(struct xlog_op_header);
1676 * If first write for transaction, insert start record We can't be trying to
1677 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1680 xlog_write_start_rec(
1681 struct xlog_op_header *ophdr,
1682 struct xlog_ticket *ticket)
1684 if (!(ticket->t_flags & XLOG_TIC_INITED))
1687 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1688 ophdr->oh_clientid = ticket->t_clientid;
1690 ophdr->oh_flags = XLOG_START_TRANS;
1693 ticket->t_flags &= ~XLOG_TIC_INITED;
1695 return sizeof(struct xlog_op_header);
1698 static xlog_op_header_t *
1699 xlog_write_setup_ophdr(
1701 struct xlog_op_header *ophdr,
1702 struct xlog_ticket *ticket,
1705 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1706 ophdr->oh_clientid = ticket->t_clientid;
1709 /* are we copying a commit or unmount record? */
1710 ophdr->oh_flags = flags;
1713 * We've seen logs corrupted with bad transaction client ids. This
1714 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1715 * and shut down the filesystem.
1717 switch (ophdr->oh_clientid) {
1718 case XFS_TRANSACTION:
1723 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1724 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1725 ophdr->oh_clientid, ticket);
1733 * Set up the parameters of the region copy into the log. This has
1734 * to handle region write split across multiple log buffers - this
1735 * state is kept external to this function so that this code can
1736 * can be written in an obvious, self documenting manner.
1739 xlog_write_setup_copy(
1740 struct xlog_ticket *ticket,
1741 struct xlog_op_header *ophdr,
1742 int space_available,
1746 int *last_was_partial_copy,
1747 int *bytes_consumed)
1751 still_to_copy = space_required - *bytes_consumed;
1752 *copy_off = *bytes_consumed;
1754 if (still_to_copy <= space_available) {
1755 /* write of region completes here */
1756 *copy_len = still_to_copy;
1757 ophdr->oh_len = cpu_to_be32(*copy_len);
1758 if (*last_was_partial_copy)
1759 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1760 *last_was_partial_copy = 0;
1761 *bytes_consumed = 0;
1765 /* partial write of region, needs extra log op header reservation */
1766 *copy_len = space_available;
1767 ophdr->oh_len = cpu_to_be32(*copy_len);
1768 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1769 if (*last_was_partial_copy)
1770 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1771 *bytes_consumed += *copy_len;
1772 (*last_was_partial_copy)++;
1774 /* account for new log op header */
1775 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1776 ticket->t_res_num_ophdrs++;
1778 return sizeof(struct xlog_op_header);
1782 xlog_write_copy_finish(
1784 struct xlog_in_core *iclog,
1789 int *partial_copy_len,
1791 struct xlog_in_core **commit_iclog)
1793 if (*partial_copy) {
1795 * This iclog has already been marked WANT_SYNC by
1796 * xlog_state_get_iclog_space.
1798 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1801 return xlog_state_release_iclog(log, iclog);
1805 *partial_copy_len = 0;
1807 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1808 /* no more space in this iclog - push it. */
1809 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1813 spin_lock(&log->l_icloglock);
1814 xlog_state_want_sync(log, iclog);
1815 spin_unlock(&log->l_icloglock);
1818 return xlog_state_release_iclog(log, iclog);
1819 ASSERT(flags & XLOG_COMMIT_TRANS);
1820 *commit_iclog = iclog;
1827 * Write some region out to in-core log
1829 * This will be called when writing externally provided regions or when
1830 * writing out a commit record for a given transaction.
1832 * General algorithm:
1833 * 1. Find total length of this write. This may include adding to the
1834 * lengths passed in.
1835 * 2. Check whether we violate the tickets reservation.
1836 * 3. While writing to this iclog
1837 * A. Reserve as much space in this iclog as can get
1838 * B. If this is first write, save away start lsn
1839 * C. While writing this region:
1840 * 1. If first write of transaction, write start record
1841 * 2. Write log operation header (header per region)
1842 * 3. Find out if we can fit entire region into this iclog
1843 * 4. Potentially, verify destination memcpy ptr
1844 * 5. Memcpy (partial) region
1845 * 6. If partial copy, release iclog; otherwise, continue
1846 * copying more regions into current iclog
1847 * 4. Mark want sync bit (in simulation mode)
1848 * 5. Release iclog for potential flush to on-disk log.
1851 * 1. Panic if reservation is overrun. This should never happen since
1852 * reservation amounts are generated internal to the filesystem.
1854 * 1. Tickets are single threaded data structures.
1855 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1856 * syncing routine. When a single log_write region needs to span
1857 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1858 * on all log operation writes which don't contain the end of the
1859 * region. The XLOG_END_TRANS bit is used for the in-core log
1860 * operation which contains the end of the continued log_write region.
1861 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1862 * we don't really know exactly how much space will be used. As a result,
1863 * we don't update ic_offset until the end when we know exactly how many
1864 * bytes have been written out.
1869 struct xfs_log_vec *log_vector,
1870 struct xlog_ticket *ticket,
1871 xfs_lsn_t *start_lsn,
1872 struct xlog_in_core **commit_iclog,
1875 struct xlog_in_core *iclog = NULL;
1876 struct xfs_log_iovec *vecp;
1877 struct xfs_log_vec *lv;
1880 int partial_copy = 0;
1881 int partial_copy_len = 0;
1889 len = xlog_write_calc_vec_length(ticket, log_vector);
1890 if (ticket->t_curr_res < len) {
1891 xlog_print_tic_res(log->l_mp, ticket);
1894 "xfs_log_write: reservation ran out. Need to up reservation");
1896 /* Customer configurable panic */
1897 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, log->l_mp,
1898 "xfs_log_write: reservation ran out. Need to up reservation");
1900 /* If we did not panic, shutdown the filesystem */
1901 xfs_force_shutdown(log->l_mp, SHUTDOWN_CORRUPT_INCORE);
1905 ticket->t_curr_res -= len;
1909 vecp = lv->lv_iovecp;
1910 while (lv && index < lv->lv_niovecs) {
1914 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1915 &contwr, &log_offset);
1919 ASSERT(log_offset <= iclog->ic_size - 1);
1920 ptr = iclog->ic_datap + log_offset;
1922 /* start_lsn is the first lsn written to. That's all we need. */
1924 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1927 * This loop writes out as many regions as can fit in the amount
1928 * of space which was allocated by xlog_state_get_iclog_space().
1930 while (lv && index < lv->lv_niovecs) {
1931 struct xfs_log_iovec *reg = &vecp[index];
1932 struct xlog_op_header *ophdr;
1937 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1938 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1940 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1941 if (start_rec_copy) {
1943 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1947 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1949 return XFS_ERROR(EIO);
1951 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1952 sizeof(struct xlog_op_header));
1954 len += xlog_write_setup_copy(ticket, ophdr,
1955 iclog->ic_size-log_offset,
1957 ©_off, ©_len,
1960 xlog_verify_dest_ptr(log, ptr);
1963 ASSERT(copy_len >= 0);
1964 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1965 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1967 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1969 data_cnt += contwr ? copy_len : 0;
1971 error = xlog_write_copy_finish(log, iclog, flags,
1972 &record_cnt, &data_cnt,
1981 * if we had a partial copy, we need to get more iclog
1982 * space but we don't want to increment the region
1983 * index because there is still more is this region to
1986 * If we completed writing this region, and we flushed
1987 * the iclog (indicated by resetting of the record
1988 * count), then we also need to get more log space. If
1989 * this was the last record, though, we are done and
1995 if (++index == lv->lv_niovecs) {
1999 vecp = lv->lv_iovecp;
2001 if (record_cnt == 0) {
2011 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2013 return xlog_state_release_iclog(log, iclog);
2015 ASSERT(flags & XLOG_COMMIT_TRANS);
2016 *commit_iclog = iclog;
2021 /*****************************************************************************
2023 * State Machine functions
2025 *****************************************************************************
2028 /* Clean iclogs starting from the head. This ordering must be
2029 * maintained, so an iclog doesn't become ACTIVE beyond one that
2030 * is SYNCING. This is also required to maintain the notion that we use
2031 * a ordered wait queue to hold off would be writers to the log when every
2032 * iclog is trying to sync to disk.
2034 * State Change: DIRTY -> ACTIVE
2037 xlog_state_clean_log(xlog_t *log)
2039 xlog_in_core_t *iclog;
2042 iclog = log->l_iclog;
2044 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2045 iclog->ic_state = XLOG_STATE_ACTIVE;
2046 iclog->ic_offset = 0;
2047 ASSERT(iclog->ic_callback == NULL);
2049 * If the number of ops in this iclog indicate it just
2050 * contains the dummy transaction, we can
2051 * change state into IDLE (the second time around).
2052 * Otherwise we should change the state into
2054 * We don't need to cover the dummy.
2057 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2062 * We have two dirty iclogs so start over
2063 * This could also be num of ops indicates
2064 * this is not the dummy going out.
2068 iclog->ic_header.h_num_logops = 0;
2069 memset(iclog->ic_header.h_cycle_data, 0,
2070 sizeof(iclog->ic_header.h_cycle_data));
2071 iclog->ic_header.h_lsn = 0;
2072 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2075 break; /* stop cleaning */
2076 iclog = iclog->ic_next;
2077 } while (iclog != log->l_iclog);
2079 /* log is locked when we are called */
2081 * Change state for the dummy log recording.
2082 * We usually go to NEED. But we go to NEED2 if the changed indicates
2083 * we are done writing the dummy record.
2084 * If we are done with the second dummy recored (DONE2), then
2088 switch (log->l_covered_state) {
2089 case XLOG_STATE_COVER_IDLE:
2090 case XLOG_STATE_COVER_NEED:
2091 case XLOG_STATE_COVER_NEED2:
2092 log->l_covered_state = XLOG_STATE_COVER_NEED;
2095 case XLOG_STATE_COVER_DONE:
2097 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2099 log->l_covered_state = XLOG_STATE_COVER_NEED;
2102 case XLOG_STATE_COVER_DONE2:
2104 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2106 log->l_covered_state = XLOG_STATE_COVER_NEED;
2113 } /* xlog_state_clean_log */
2116 xlog_get_lowest_lsn(
2119 xlog_in_core_t *lsn_log;
2120 xfs_lsn_t lowest_lsn, lsn;
2122 lsn_log = log->l_iclog;
2125 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2126 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2127 if ((lsn && !lowest_lsn) ||
2128 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2132 lsn_log = lsn_log->ic_next;
2133 } while (lsn_log != log->l_iclog);
2139 xlog_state_do_callback(
2142 xlog_in_core_t *ciclog)
2144 xlog_in_core_t *iclog;
2145 xlog_in_core_t *first_iclog; /* used to know when we've
2146 * processed all iclogs once */
2147 xfs_log_callback_t *cb, *cb_next;
2149 xfs_lsn_t lowest_lsn;
2150 int ioerrors; /* counter: iclogs with errors */
2151 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2152 int funcdidcallbacks; /* flag: function did callbacks */
2153 int repeats; /* for issuing console warnings if
2154 * looping too many times */
2157 spin_lock(&log->l_icloglock);
2158 first_iclog = iclog = log->l_iclog;
2160 funcdidcallbacks = 0;
2165 * Scan all iclogs starting with the one pointed to by the
2166 * log. Reset this starting point each time the log is
2167 * unlocked (during callbacks).
2169 * Keep looping through iclogs until one full pass is made
2170 * without running any callbacks.
2172 first_iclog = log->l_iclog;
2173 iclog = log->l_iclog;
2174 loopdidcallbacks = 0;
2179 /* skip all iclogs in the ACTIVE & DIRTY states */
2180 if (iclog->ic_state &
2181 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2182 iclog = iclog->ic_next;
2187 * Between marking a filesystem SHUTDOWN and stopping
2188 * the log, we do flush all iclogs to disk (if there
2189 * wasn't a log I/O error). So, we do want things to
2190 * go smoothly in case of just a SHUTDOWN w/o a
2193 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2195 * Can only perform callbacks in order. Since
2196 * this iclog is not in the DONE_SYNC/
2197 * DO_CALLBACK state, we skip the rest and
2198 * just try to clean up. If we set our iclog
2199 * to DO_CALLBACK, we will not process it when
2200 * we retry since a previous iclog is in the
2201 * CALLBACK and the state cannot change since
2202 * we are holding the l_icloglock.
2204 if (!(iclog->ic_state &
2205 (XLOG_STATE_DONE_SYNC |
2206 XLOG_STATE_DO_CALLBACK))) {
2207 if (ciclog && (ciclog->ic_state ==
2208 XLOG_STATE_DONE_SYNC)) {
2209 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2214 * We now have an iclog that is in either the
2215 * DO_CALLBACK or DONE_SYNC states. The other
2216 * states (WANT_SYNC, SYNCING, or CALLBACK were
2217 * caught by the above if and are going to
2218 * clean (i.e. we aren't doing their callbacks)
2223 * We will do one more check here to see if we
2224 * have chased our tail around.
2227 lowest_lsn = xlog_get_lowest_lsn(log);
2229 XFS_LSN_CMP(lowest_lsn,
2230 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2231 iclog = iclog->ic_next;
2232 continue; /* Leave this iclog for
2236 iclog->ic_state = XLOG_STATE_CALLBACK;
2238 spin_unlock(&log->l_icloglock);
2240 /* l_last_sync_lsn field protected by
2241 * l_grant_lock. Don't worry about iclog's lsn.
2242 * No one else can be here except us.
2244 spin_lock(&log->l_grant_lock);
2245 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2246 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2247 log->l_last_sync_lsn =
2248 be64_to_cpu(iclog->ic_header.h_lsn);
2249 spin_unlock(&log->l_grant_lock);
2252 spin_unlock(&log->l_icloglock);
2257 * Keep processing entries in the callback list until
2258 * we come around and it is empty. We need to
2259 * atomically see that the list is empty and change the
2260 * state to DIRTY so that we don't miss any more
2261 * callbacks being added.
2263 spin_lock(&iclog->ic_callback_lock);
2264 cb = iclog->ic_callback;
2266 iclog->ic_callback_tail = &(iclog->ic_callback);
2267 iclog->ic_callback = NULL;
2268 spin_unlock(&iclog->ic_callback_lock);
2270 /* perform callbacks in the order given */
2271 for (; cb; cb = cb_next) {
2272 cb_next = cb->cb_next;
2273 cb->cb_func(cb->cb_arg, aborted);
2275 spin_lock(&iclog->ic_callback_lock);
2276 cb = iclog->ic_callback;
2282 spin_lock(&log->l_icloglock);
2283 ASSERT(iclog->ic_callback == NULL);
2284 spin_unlock(&iclog->ic_callback_lock);
2285 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2286 iclog->ic_state = XLOG_STATE_DIRTY;
2289 * Transition from DIRTY to ACTIVE if applicable.
2290 * NOP if STATE_IOERROR.
2292 xlog_state_clean_log(log);
2294 /* wake up threads waiting in xfs_log_force() */
2295 sv_broadcast(&iclog->ic_force_wait);
2297 iclog = iclog->ic_next;
2298 } while (first_iclog != iclog);
2300 if (repeats > 5000) {
2301 flushcnt += repeats;
2303 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2304 "%s: possible infinite loop (%d iterations)",
2305 __func__, flushcnt);
2307 } while (!ioerrors && loopdidcallbacks);
2310 * make one last gasp attempt to see if iclogs are being left in
2314 if (funcdidcallbacks) {
2315 first_iclog = iclog = log->l_iclog;
2317 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2319 * Terminate the loop if iclogs are found in states
2320 * which will cause other threads to clean up iclogs.
2322 * SYNCING - i/o completion will go through logs
2323 * DONE_SYNC - interrupt thread should be waiting for
2325 * IOERROR - give up hope all ye who enter here
2327 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2328 iclog->ic_state == XLOG_STATE_SYNCING ||
2329 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2330 iclog->ic_state == XLOG_STATE_IOERROR )
2332 iclog = iclog->ic_next;
2333 } while (first_iclog != iclog);
2337 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2339 spin_unlock(&log->l_icloglock);
2342 sv_broadcast(&log->l_flush_wait);
2347 * Finish transitioning this iclog to the dirty state.
2349 * Make sure that we completely execute this routine only when this is
2350 * the last call to the iclog. There is a good chance that iclog flushes,
2351 * when we reach the end of the physical log, get turned into 2 separate
2352 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2353 * routine. By using the reference count bwritecnt, we guarantee that only
2354 * the second completion goes through.
2356 * Callbacks could take time, so they are done outside the scope of the
2357 * global state machine log lock.
2360 xlog_state_done_syncing(
2361 xlog_in_core_t *iclog,
2364 xlog_t *log = iclog->ic_log;
2366 spin_lock(&log->l_icloglock);
2368 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2369 iclog->ic_state == XLOG_STATE_IOERROR);
2370 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2371 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2375 * If we got an error, either on the first buffer, or in the case of
2376 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2377 * and none should ever be attempted to be written to disk
2380 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2381 if (--iclog->ic_bwritecnt == 1) {
2382 spin_unlock(&log->l_icloglock);
2385 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2389 * Someone could be sleeping prior to writing out the next
2390 * iclog buffer, we wake them all, one will get to do the
2391 * I/O, the others get to wait for the result.
2393 sv_broadcast(&iclog->ic_write_wait);
2394 spin_unlock(&log->l_icloglock);
2395 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2396 } /* xlog_state_done_syncing */
2400 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2401 * sleep. We wait on the flush queue on the head iclog as that should be
2402 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2403 * we will wait here and all new writes will sleep until a sync completes.
2405 * The in-core logs are used in a circular fashion. They are not used
2406 * out-of-order even when an iclog past the head is free.
2409 * * log_offset where xlog_write() can start writing into the in-core
2411 * * in-core log pointer to which xlog_write() should write.
2412 * * boolean indicating this is a continued write to an in-core log.
2413 * If this is the last write, then the in-core log's offset field
2414 * needs to be incremented, depending on the amount of data which
2418 xlog_state_get_iclog_space(xlog_t *log,
2420 xlog_in_core_t **iclogp,
2421 xlog_ticket_t *ticket,
2422 int *continued_write,
2426 xlog_rec_header_t *head;
2427 xlog_in_core_t *iclog;
2431 spin_lock(&log->l_icloglock);
2432 if (XLOG_FORCED_SHUTDOWN(log)) {
2433 spin_unlock(&log->l_icloglock);
2434 return XFS_ERROR(EIO);
2437 iclog = log->l_iclog;
2438 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2439 XFS_STATS_INC(xs_log_noiclogs);
2441 /* Wait for log writes to have flushed */
2442 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2446 head = &iclog->ic_header;
2448 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2449 log_offset = iclog->ic_offset;
2451 /* On the 1st write to an iclog, figure out lsn. This works
2452 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2453 * committing to. If the offset is set, that's how many blocks
2456 if (log_offset == 0) {
2457 ticket->t_curr_res -= log->l_iclog_hsize;
2458 xlog_tic_add_region(ticket,
2460 XLOG_REG_TYPE_LRHEADER);
2461 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2462 head->h_lsn = cpu_to_be64(
2463 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2464 ASSERT(log->l_curr_block >= 0);
2467 /* If there is enough room to write everything, then do it. Otherwise,
2468 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2469 * bit is on, so this will get flushed out. Don't update ic_offset
2470 * until you know exactly how many bytes get copied. Therefore, wait
2471 * until later to update ic_offset.
2473 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2474 * can fit into remaining data section.
2476 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2477 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2480 * If I'm the only one writing to this iclog, sync it to disk.
2481 * We need to do an atomic compare and decrement here to avoid
2482 * racing with concurrent atomic_dec_and_lock() calls in
2483 * xlog_state_release_iclog() when there is more than one
2484 * reference to the iclog.
2486 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2487 /* we are the only one */
2488 spin_unlock(&log->l_icloglock);
2489 error = xlog_state_release_iclog(log, iclog);
2493 spin_unlock(&log->l_icloglock);
2498 /* Do we have enough room to write the full amount in the remainder
2499 * of this iclog? Or must we continue a write on the next iclog and
2500 * mark this iclog as completely taken? In the case where we switch
2501 * iclogs (to mark it taken), this particular iclog will release/sync
2502 * to disk in xlog_write().
2504 if (len <= iclog->ic_size - iclog->ic_offset) {
2505 *continued_write = 0;
2506 iclog->ic_offset += len;
2508 *continued_write = 1;
2509 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2513 ASSERT(iclog->ic_offset <= iclog->ic_size);
2514 spin_unlock(&log->l_icloglock);
2516 *logoffsetp = log_offset;
2518 } /* xlog_state_get_iclog_space */
2521 * Atomically get the log space required for a log ticket.
2523 * Once a ticket gets put onto the reserveq, it will only return after
2524 * the needed reservation is satisfied.
2527 xlog_grant_log_space(xlog_t *log,
2538 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2539 panic("grant Recovery problem");
2542 /* Is there space or do we need to sleep? */
2543 spin_lock(&log->l_grant_lock);
2545 trace_xfs_log_grant_enter(log, tic);
2547 /* something is already sleeping; insert new transaction at end */
2548 if (log->l_reserve_headq) {
2549 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2551 trace_xfs_log_grant_sleep1(log, tic);
2554 * Gotta check this before going to sleep, while we're
2555 * holding the grant lock.
2557 if (XLOG_FORCED_SHUTDOWN(log))
2560 XFS_STATS_INC(xs_sleep_logspace);
2561 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2563 * If we got an error, and the filesystem is shutting down,
2564 * we'll catch it down below. So just continue...
2566 trace_xfs_log_grant_wake1(log, tic);
2567 spin_lock(&log->l_grant_lock);
2569 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2570 need_bytes = tic->t_unit_res*tic->t_ocnt;
2572 need_bytes = tic->t_unit_res;
2575 if (XLOG_FORCED_SHUTDOWN(log))
2578 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2579 log->l_grant_reserve_bytes);
2580 if (free_bytes < need_bytes) {
2581 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2582 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2584 trace_xfs_log_grant_sleep2(log, tic);
2586 spin_unlock(&log->l_grant_lock);
2587 xlog_grant_push_ail(log->l_mp, need_bytes);
2588 spin_lock(&log->l_grant_lock);
2590 XFS_STATS_INC(xs_sleep_logspace);
2591 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2593 spin_lock(&log->l_grant_lock);
2594 if (XLOG_FORCED_SHUTDOWN(log))
2597 trace_xfs_log_grant_wake2(log, tic);
2600 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2601 xlog_del_ticketq(&log->l_reserve_headq, tic);
2603 /* we've got enough space */
2604 xlog_grant_add_space(log, need_bytes);
2606 tail_lsn = log->l_tail_lsn;
2608 * Check to make sure the grant write head didn't just over lap the
2609 * tail. If the cycles are the same, we can't be overlapping.
2610 * Otherwise, make sure that the cycles differ by exactly one and
2611 * check the byte count.
2613 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2614 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2615 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2618 trace_xfs_log_grant_exit(log, tic);
2619 xlog_verify_grant_head(log, 1);
2620 spin_unlock(&log->l_grant_lock);
2624 if (tic->t_flags & XLOG_TIC_IN_Q)
2625 xlog_del_ticketq(&log->l_reserve_headq, tic);
2627 trace_xfs_log_grant_error(log, tic);
2630 * If we are failing, make sure the ticket doesn't have any
2631 * current reservations. We don't want to add this back when
2632 * the ticket/transaction gets cancelled.
2634 tic->t_curr_res = 0;
2635 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2636 spin_unlock(&log->l_grant_lock);
2637 return XFS_ERROR(EIO);
2638 } /* xlog_grant_log_space */
2642 * Replenish the byte reservation required by moving the grant write head.
2647 xlog_regrant_write_log_space(xlog_t *log,
2650 int free_bytes, need_bytes;
2651 xlog_ticket_t *ntic;
2656 tic->t_curr_res = tic->t_unit_res;
2657 xlog_tic_reset_res(tic);
2663 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2664 panic("regrant Recovery problem");
2667 spin_lock(&log->l_grant_lock);
2669 trace_xfs_log_regrant_write_enter(log, tic);
2671 if (XLOG_FORCED_SHUTDOWN(log))
2674 /* If there are other waiters on the queue then give them a
2675 * chance at logspace before us. Wake up the first waiters,
2676 * if we do not wake up all the waiters then go to sleep waiting
2677 * for more free space, otherwise try to get some space for
2680 need_bytes = tic->t_unit_res;
2681 if ((ntic = log->l_write_headq)) {
2682 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2683 log->l_grant_write_bytes);
2685 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2687 if (free_bytes < ntic->t_unit_res)
2689 free_bytes -= ntic->t_unit_res;
2690 sv_signal(&ntic->t_wait);
2691 ntic = ntic->t_next;
2692 } while (ntic != log->l_write_headq);
2694 if (ntic != log->l_write_headq) {
2695 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2696 xlog_ins_ticketq(&log->l_write_headq, tic);
2698 trace_xfs_log_regrant_write_sleep1(log, tic);
2700 spin_unlock(&log->l_grant_lock);
2701 xlog_grant_push_ail(log->l_mp, need_bytes);
2702 spin_lock(&log->l_grant_lock);
2704 XFS_STATS_INC(xs_sleep_logspace);
2705 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2706 &log->l_grant_lock, s);
2708 /* If we're shutting down, this tic is already
2710 spin_lock(&log->l_grant_lock);
2711 if (XLOG_FORCED_SHUTDOWN(log))
2714 trace_xfs_log_regrant_write_wake1(log, tic);
2719 if (XLOG_FORCED_SHUTDOWN(log))
2722 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2723 log->l_grant_write_bytes);
2724 if (free_bytes < need_bytes) {
2725 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2726 xlog_ins_ticketq(&log->l_write_headq, tic);
2727 spin_unlock(&log->l_grant_lock);
2728 xlog_grant_push_ail(log->l_mp, need_bytes);
2729 spin_lock(&log->l_grant_lock);
2731 XFS_STATS_INC(xs_sleep_logspace);
2732 trace_xfs_log_regrant_write_sleep2(log, tic);
2734 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2736 /* If we're shutting down, this tic is already off the queue */
2737 spin_lock(&log->l_grant_lock);
2738 if (XLOG_FORCED_SHUTDOWN(log))
2741 trace_xfs_log_regrant_write_wake2(log, tic);
2743 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2744 xlog_del_ticketq(&log->l_write_headq, tic);
2746 /* we've got enough space */
2747 xlog_grant_add_space_write(log, need_bytes);
2749 tail_lsn = log->l_tail_lsn;
2750 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2751 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2752 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2756 trace_xfs_log_regrant_write_exit(log, tic);
2758 xlog_verify_grant_head(log, 1);
2759 spin_unlock(&log->l_grant_lock);
2764 if (tic->t_flags & XLOG_TIC_IN_Q)
2765 xlog_del_ticketq(&log->l_reserve_headq, tic);
2767 trace_xfs_log_regrant_write_error(log, tic);
2770 * If we are failing, make sure the ticket doesn't have any
2771 * current reservations. We don't want to add this back when
2772 * the ticket/transaction gets cancelled.
2774 tic->t_curr_res = 0;
2775 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2776 spin_unlock(&log->l_grant_lock);
2777 return XFS_ERROR(EIO);
2778 } /* xlog_regrant_write_log_space */
2781 /* The first cnt-1 times through here we don't need to
2782 * move the grant write head because the permanent
2783 * reservation has reserved cnt times the unit amount.
2784 * Release part of current permanent unit reservation and
2785 * reset current reservation to be one units worth. Also
2786 * move grant reservation head forward.
2789 xlog_regrant_reserve_log_space(xlog_t *log,
2790 xlog_ticket_t *ticket)
2792 trace_xfs_log_regrant_reserve_enter(log, ticket);
2794 if (ticket->t_cnt > 0)
2797 spin_lock(&log->l_grant_lock);
2798 xlog_grant_sub_space(log, ticket->t_curr_res);
2799 ticket->t_curr_res = ticket->t_unit_res;
2800 xlog_tic_reset_res(ticket);
2802 trace_xfs_log_regrant_reserve_sub(log, ticket);
2804 xlog_verify_grant_head(log, 1);
2806 /* just return if we still have some of the pre-reserved space */
2807 if (ticket->t_cnt > 0) {
2808 spin_unlock(&log->l_grant_lock);
2812 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2814 trace_xfs_log_regrant_reserve_exit(log, ticket);
2816 xlog_verify_grant_head(log, 0);
2817 spin_unlock(&log->l_grant_lock);
2818 ticket->t_curr_res = ticket->t_unit_res;
2819 xlog_tic_reset_res(ticket);
2820 } /* xlog_regrant_reserve_log_space */
2824 * Give back the space left from a reservation.
2826 * All the information we need to make a correct determination of space left
2827 * is present. For non-permanent reservations, things are quite easy. The
2828 * count should have been decremented to zero. We only need to deal with the
2829 * space remaining in the current reservation part of the ticket. If the
2830 * ticket contains a permanent reservation, there may be left over space which
2831 * needs to be released. A count of N means that N-1 refills of the current
2832 * reservation can be done before we need to ask for more space. The first
2833 * one goes to fill up the first current reservation. Once we run out of
2834 * space, the count will stay at zero and the only space remaining will be
2835 * in the current reservation field.
2838 xlog_ungrant_log_space(xlog_t *log,
2839 xlog_ticket_t *ticket)
2841 if (ticket->t_cnt > 0)
2844 spin_lock(&log->l_grant_lock);
2845 trace_xfs_log_ungrant_enter(log, ticket);
2847 xlog_grant_sub_space(log, ticket->t_curr_res);
2849 trace_xfs_log_ungrant_sub(log, ticket);
2851 /* If this is a permanent reservation ticket, we may be able to free
2852 * up more space based on the remaining count.
2854 if (ticket->t_cnt > 0) {
2855 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2856 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2859 trace_xfs_log_ungrant_exit(log, ticket);
2861 xlog_verify_grant_head(log, 1);
2862 spin_unlock(&log->l_grant_lock);
2863 xfs_log_move_tail(log->l_mp, 1);
2864 } /* xlog_ungrant_log_space */
2868 * Flush iclog to disk if this is the last reference to the given iclog and
2869 * the WANT_SYNC bit is set.
2871 * When this function is entered, the iclog is not necessarily in the
2872 * WANT_SYNC state. It may be sitting around waiting to get filled.
2877 xlog_state_release_iclog(
2879 xlog_in_core_t *iclog)
2881 int sync = 0; /* do we sync? */
2883 if (iclog->ic_state & XLOG_STATE_IOERROR)
2884 return XFS_ERROR(EIO);
2886 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2887 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2890 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2891 spin_unlock(&log->l_icloglock);
2892 return XFS_ERROR(EIO);
2894 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2895 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2897 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2898 /* update tail before writing to iclog */
2899 xlog_assign_tail_lsn(log->l_mp);
2901 iclog->ic_state = XLOG_STATE_SYNCING;
2902 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2903 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2904 /* cycle incremented when incrementing curr_block */
2906 spin_unlock(&log->l_icloglock);
2909 * We let the log lock go, so it's possible that we hit a log I/O
2910 * error or some other SHUTDOWN condition that marks the iclog
2911 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2912 * this iclog has consistent data, so we ignore IOERROR
2913 * flags after this point.
2916 return xlog_sync(log, iclog);
2918 } /* xlog_state_release_iclog */
2922 * This routine will mark the current iclog in the ring as WANT_SYNC
2923 * and move the current iclog pointer to the next iclog in the ring.
2924 * When this routine is called from xlog_state_get_iclog_space(), the
2925 * exact size of the iclog has not yet been determined. All we know is
2926 * that every data block. We have run out of space in this log record.
2929 xlog_state_switch_iclogs(xlog_t *log,
2930 xlog_in_core_t *iclog,
2933 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2935 eventual_size = iclog->ic_offset;
2936 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2937 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2938 log->l_prev_block = log->l_curr_block;
2939 log->l_prev_cycle = log->l_curr_cycle;
2941 /* roll log?: ic_offset changed later */
2942 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2944 /* Round up to next log-sunit */
2945 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2946 log->l_mp->m_sb.sb_logsunit > 1) {
2947 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2948 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2951 if (log->l_curr_block >= log->l_logBBsize) {
2952 log->l_curr_cycle++;
2953 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2954 log->l_curr_cycle++;
2955 log->l_curr_block -= log->l_logBBsize;
2956 ASSERT(log->l_curr_block >= 0);
2958 ASSERT(iclog == log->l_iclog);
2959 log->l_iclog = iclog->ic_next;
2960 } /* xlog_state_switch_iclogs */
2963 * Write out all data in the in-core log as of this exact moment in time.
2965 * Data may be written to the in-core log during this call. However,
2966 * we don't guarantee this data will be written out. A change from past
2967 * implementation means this routine will *not* write out zero length LRs.
2969 * Basically, we try and perform an intelligent scan of the in-core logs.
2970 * If we determine there is no flushable data, we just return. There is no
2971 * flushable data if:
2973 * 1. the current iclog is active and has no data; the previous iclog
2974 * is in the active or dirty state.
2975 * 2. the current iclog is drity, and the previous iclog is in the
2976 * active or dirty state.
2980 * 1. the current iclog is not in the active nor dirty state.
2981 * 2. the current iclog dirty, and the previous iclog is not in the
2982 * active nor dirty state.
2983 * 3. the current iclog is active, and there is another thread writing
2984 * to this particular iclog.
2985 * 4. a) the current iclog is active and has no other writers
2986 * b) when we return from flushing out this iclog, it is still
2987 * not in the active nor dirty state.
2991 struct xfs_mount *mp,
2995 struct log *log = mp->m_log;
2996 struct xlog_in_core *iclog;
2999 XFS_STATS_INC(xs_log_force);
3001 spin_lock(&log->l_icloglock);
3003 iclog = log->l_iclog;
3004 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3005 spin_unlock(&log->l_icloglock);
3006 return XFS_ERROR(EIO);
3009 /* If the head iclog is not active nor dirty, we just attach
3010 * ourselves to the head and go to sleep.
3012 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3013 iclog->ic_state == XLOG_STATE_DIRTY) {
3015 * If the head is dirty or (active and empty), then
3016 * we need to look at the previous iclog. If the previous
3017 * iclog is active or dirty we are done. There is nothing
3018 * to sync out. Otherwise, we attach ourselves to the
3019 * previous iclog and go to sleep.
3021 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3022 (atomic_read(&iclog->ic_refcnt) == 0
3023 && iclog->ic_offset == 0)) {
3024 iclog = iclog->ic_prev;
3025 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3026 iclog->ic_state == XLOG_STATE_DIRTY)
3031 if (atomic_read(&iclog->ic_refcnt) == 0) {
3032 /* We are the only one with access to this
3033 * iclog. Flush it out now. There should
3034 * be a roundoff of zero to show that someone
3035 * has already taken care of the roundoff from
3036 * the previous sync.
3038 atomic_inc(&iclog->ic_refcnt);
3039 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3040 xlog_state_switch_iclogs(log, iclog, 0);
3041 spin_unlock(&log->l_icloglock);
3043 if (xlog_state_release_iclog(log, iclog))
3044 return XFS_ERROR(EIO);
3048 spin_lock(&log->l_icloglock);
3049 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3050 iclog->ic_state != XLOG_STATE_DIRTY)
3055 /* Someone else is writing to this iclog.
3056 * Use its call to flush out the data. However,
3057 * the other thread may not force out this LR,
3058 * so we mark it WANT_SYNC.
3060 xlog_state_switch_iclogs(log, iclog, 0);
3066 /* By the time we come around again, the iclog could've been filled
3067 * which would give it another lsn. If we have a new lsn, just
3068 * return because the relevant data has been flushed.
3071 if (flags & XFS_LOG_SYNC) {
3073 * We must check if we're shutting down here, before
3074 * we wait, while we're holding the l_icloglock.
3075 * Then we check again after waking up, in case our
3076 * sleep was disturbed by a bad news.
3078 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3079 spin_unlock(&log->l_icloglock);
3080 return XFS_ERROR(EIO);
3082 XFS_STATS_INC(xs_log_force_sleep);
3083 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3085 * No need to grab the log lock here since we're
3086 * only deciding whether or not to return EIO
3087 * and the memory read should be atomic.
3089 if (iclog->ic_state & XLOG_STATE_IOERROR)
3090 return XFS_ERROR(EIO);
3096 spin_unlock(&log->l_icloglock);
3102 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3103 * about errors or whether the log was flushed or not. This is the normal
3104 * interface to use when trying to unpin items or move the log forward.
3113 error = _xfs_log_force(mp, flags, NULL);
3115 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3116 "error %d returned.", error);
3121 * Force the in-core log to disk for a specific LSN.
3123 * Find in-core log with lsn.
3124 * If it is in the DIRTY state, just return.
3125 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3126 * state and go to sleep or return.
3127 * If it is in any other state, go to sleep or return.
3129 * Synchronous forces are implemented with a signal variable. All callers
3130 * to force a given lsn to disk will wait on a the sv attached to the
3131 * specific in-core log. When given in-core log finally completes its
3132 * write to disk, that thread will wake up all threads waiting on the
3137 struct xfs_mount *mp,
3142 struct log *log = mp->m_log;
3143 struct xlog_in_core *iclog;
3144 int already_slept = 0;
3148 XFS_STATS_INC(xs_log_force);
3151 spin_lock(&log->l_icloglock);
3152 iclog = log->l_iclog;
3153 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3154 spin_unlock(&log->l_icloglock);
3155 return XFS_ERROR(EIO);
3159 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3160 iclog = iclog->ic_next;
3164 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3165 spin_unlock(&log->l_icloglock);
3169 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3171 * We sleep here if we haven't already slept (e.g.
3172 * this is the first time we've looked at the correct
3173 * iclog buf) and the buffer before us is going to
3174 * be sync'ed. The reason for this is that if we
3175 * are doing sync transactions here, by waiting for
3176 * the previous I/O to complete, we can allow a few
3177 * more transactions into this iclog before we close
3180 * Otherwise, we mark the buffer WANT_SYNC, and bump
3181 * up the refcnt so we can release the log (which
3182 * drops the ref count). The state switch keeps new
3183 * transaction commits from using this buffer. When
3184 * the current commits finish writing into the buffer,
3185 * the refcount will drop to zero and the buffer will
3188 if (!already_slept &&
3189 (iclog->ic_prev->ic_state &
3190 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3191 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3193 XFS_STATS_INC(xs_log_force_sleep);
3195 sv_wait(&iclog->ic_prev->ic_write_wait,
3196 PSWP, &log->l_icloglock, s);
3202 atomic_inc(&iclog->ic_refcnt);
3203 xlog_state_switch_iclogs(log, iclog, 0);
3204 spin_unlock(&log->l_icloglock);
3205 if (xlog_state_release_iclog(log, iclog))
3206 return XFS_ERROR(EIO);
3209 spin_lock(&log->l_icloglock);
3212 if ((flags & XFS_LOG_SYNC) && /* sleep */
3214 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3216 * Don't wait on completion if we know that we've
3217 * gotten a log write error.
3219 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3220 spin_unlock(&log->l_icloglock);
3221 return XFS_ERROR(EIO);
3223 XFS_STATS_INC(xs_log_force_sleep);
3224 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3226 * No need to grab the log lock here since we're
3227 * only deciding whether or not to return EIO
3228 * and the memory read should be atomic.
3230 if (iclog->ic_state & XLOG_STATE_IOERROR)
3231 return XFS_ERROR(EIO);
3235 } else { /* just return */
3236 spin_unlock(&log->l_icloglock);
3240 } while (iclog != log->l_iclog);
3242 spin_unlock(&log->l_icloglock);
3247 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3248 * about errors or whether the log was flushed or not. This is the normal
3249 * interface to use when trying to unpin items or move the log forward.
3259 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3261 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3262 "error %d returned.", error);
3267 * Called when we want to mark the current iclog as being ready to sync to
3271 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3273 assert_spin_locked(&log->l_icloglock);
3275 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3276 xlog_state_switch_iclogs(log, iclog, 0);
3278 ASSERT(iclog->ic_state &
3279 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3284 /*****************************************************************************
3288 *****************************************************************************
3292 * Free a used ticket when its refcount falls to zero.
3296 xlog_ticket_t *ticket)
3298 ASSERT(atomic_read(&ticket->t_ref) > 0);
3299 if (atomic_dec_and_test(&ticket->t_ref)) {
3300 sv_destroy(&ticket->t_wait);
3301 kmem_zone_free(xfs_log_ticket_zone, ticket);
3307 xlog_ticket_t *ticket)
3309 ASSERT(atomic_read(&ticket->t_ref) > 0);
3310 atomic_inc(&ticket->t_ref);
3315 * Allocate and initialise a new log ticket.
3317 STATIC xlog_ticket_t *
3325 struct xlog_ticket *tic;
3329 tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3334 * Permanent reservations have up to 'cnt'-1 active log operations
3335 * in the log. A unit in this case is the amount of space for one
3336 * of these log operations. Normal reservations have a cnt of 1
3337 * and their unit amount is the total amount of space required.
3339 * The following lines of code account for non-transaction data
3340 * which occupy space in the on-disk log.
3342 * Normal form of a transaction is:
3343 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3344 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3346 * We need to account for all the leadup data and trailer data
3347 * around the transaction data.
3348 * And then we need to account for the worst case in terms of using
3350 * The worst case will happen if:
3351 * - the placement of the transaction happens to be such that the
3352 * roundoff is at its maximum
3353 * - the transaction data is synced before the commit record is synced
3354 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3355 * Therefore the commit record is in its own Log Record.
3356 * This can happen as the commit record is called with its
3357 * own region to xlog_write().
3358 * This then means that in the worst case, roundoff can happen for
3359 * the commit-rec as well.
3360 * The commit-rec is smaller than padding in this scenario and so it is
3361 * not added separately.
3364 /* for trans header */
3365 unit_bytes += sizeof(xlog_op_header_t);
3366 unit_bytes += sizeof(xfs_trans_header_t);
3369 unit_bytes += sizeof(xlog_op_header_t);
3372 * for LR headers - the space for data in an iclog is the size minus
3373 * the space used for the headers. If we use the iclog size, then we
3374 * undercalculate the number of headers required.
3376 * Furthermore - the addition of op headers for split-recs might
3377 * increase the space required enough to require more log and op
3378 * headers, so take that into account too.
3380 * IMPORTANT: This reservation makes the assumption that if this
3381 * transaction is the first in an iclog and hence has the LR headers
3382 * accounted to it, then the remaining space in the iclog is
3383 * exclusively for this transaction. i.e. if the transaction is larger
3384 * than the iclog, it will be the only thing in that iclog.
3385 * Fundamentally, this means we must pass the entire log vector to
3386 * xlog_write to guarantee this.
3388 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3389 num_headers = howmany(unit_bytes, iclog_space);
3391 /* for split-recs - ophdrs added when data split over LRs */
3392 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3394 /* add extra header reservations if we overrun */
3395 while (!num_headers ||
3396 howmany(unit_bytes, iclog_space) > num_headers) {
3397 unit_bytes += sizeof(xlog_op_header_t);
3400 unit_bytes += log->l_iclog_hsize * num_headers;
3402 /* for commit-rec LR header - note: padding will subsume the ophdr */
3403 unit_bytes += log->l_iclog_hsize;
3405 /* for roundoff padding for transaction data and one for commit record */
3406 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3407 log->l_mp->m_sb.sb_logsunit > 1) {
3408 /* log su roundoff */
3409 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3412 unit_bytes += 2*BBSIZE;
3415 atomic_set(&tic->t_ref, 1);
3416 tic->t_unit_res = unit_bytes;
3417 tic->t_curr_res = unit_bytes;
3420 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3421 tic->t_clientid = client;
3422 tic->t_flags = XLOG_TIC_INITED;
3423 tic->t_trans_type = 0;
3424 if (xflags & XFS_LOG_PERM_RESERV)
3425 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3426 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3428 xlog_tic_reset_res(tic);
3434 /******************************************************************************
3436 * Log debug routines
3438 ******************************************************************************
3442 * Make sure that the destination ptr is within the valid data region of
3443 * one of the iclogs. This uses backup pointers stored in a different
3444 * part of the log in case we trash the log structure.
3447 xlog_verify_dest_ptr(
3454 for (i = 0; i < log->l_iclog_bufs; i++) {
3455 if (ptr >= log->l_iclog_bak[i] &&
3456 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3461 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3465 xlog_verify_grant_head(xlog_t *log, int equals)
3467 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3469 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3471 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3473 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3474 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3476 } /* xlog_verify_grant_head */
3478 /* check if it will fit */
3480 xlog_verify_tail_lsn(xlog_t *log,
3481 xlog_in_core_t *iclog,
3486 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3488 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3489 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3490 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3492 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3494 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3495 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3497 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3498 if (blocks < BTOBB(iclog->ic_offset) + 1)
3499 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3501 } /* xlog_verify_tail_lsn */
3504 * Perform a number of checks on the iclog before writing to disk.
3506 * 1. Make sure the iclogs are still circular
3507 * 2. Make sure we have a good magic number
3508 * 3. Make sure we don't have magic numbers in the data
3509 * 4. Check fields of each log operation header for:
3510 * A. Valid client identifier
3511 * B. tid ptr value falls in valid ptr space (user space code)
3512 * C. Length in log record header is correct according to the
3513 * individual operation headers within record.
3514 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3515 * log, check the preceding blocks of the physical log to make sure all
3516 * the cycle numbers agree with the current cycle number.
3519 xlog_verify_iclog(xlog_t *log,
3520 xlog_in_core_t *iclog,
3524 xlog_op_header_t *ophead;
3525 xlog_in_core_t *icptr;
3526 xlog_in_core_2_t *xhdr;
3528 xfs_caddr_t base_ptr;
3529 __psint_t field_offset;
3531 int len, i, j, k, op_len;
3534 /* check validity of iclog pointers */
3535 spin_lock(&log->l_icloglock);
3536 icptr = log->l_iclog;
3537 for (i=0; i < log->l_iclog_bufs; i++) {
3539 xlog_panic("xlog_verify_iclog: invalid ptr");
3540 icptr = icptr->ic_next;
3542 if (icptr != log->l_iclog)
3543 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3544 spin_unlock(&log->l_icloglock);
3546 /* check log magic numbers */
3547 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3548 xlog_panic("xlog_verify_iclog: invalid magic num");
3550 ptr = (xfs_caddr_t) &iclog->ic_header;
3551 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3553 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3554 xlog_panic("xlog_verify_iclog: unexpected magic num");
3558 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3559 ptr = iclog->ic_datap;
3561 ophead = (xlog_op_header_t *)ptr;
3562 xhdr = iclog->ic_data;
3563 for (i = 0; i < len; i++) {
3564 ophead = (xlog_op_header_t *)ptr;
3566 /* clientid is only 1 byte */
3567 field_offset = (__psint_t)
3568 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3569 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3570 clientid = ophead->oh_clientid;
3572 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3573 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3574 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3575 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3576 clientid = xlog_get_client_id(
3577 xhdr[j].hic_xheader.xh_cycle_data[k]);
3579 clientid = xlog_get_client_id(
3580 iclog->ic_header.h_cycle_data[idx]);
3583 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3584 cmn_err(CE_WARN, "xlog_verify_iclog: "
3585 "invalid clientid %d op 0x%p offset 0x%lx",
3586 clientid, ophead, (unsigned long)field_offset);
3589 field_offset = (__psint_t)
3590 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3591 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3592 op_len = be32_to_cpu(ophead->oh_len);
3594 idx = BTOBBT((__psint_t)&ophead->oh_len -
3595 (__psint_t)iclog->ic_datap);
3596 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3597 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3598 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3599 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3601 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3604 ptr += sizeof(xlog_op_header_t) + op_len;
3606 } /* xlog_verify_iclog */
3610 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3616 xlog_in_core_t *iclog, *ic;
3618 iclog = log->l_iclog;
3619 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3621 * Mark all the incore logs IOERROR.
3622 * From now on, no log flushes will result.
3626 ic->ic_state = XLOG_STATE_IOERROR;
3628 } while (ic != iclog);
3632 * Return non-zero, if state transition has already happened.
3638 * This is called from xfs_force_shutdown, when we're forcibly
3639 * shutting down the filesystem, typically because of an IO error.
3640 * Our main objectives here are to make sure that:
3641 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3642 * parties to find out, 'atomically'.
3643 * b. those who're sleeping on log reservations, pinned objects and
3644 * other resources get woken up, and be told the bad news.
3645 * c. nothing new gets queued up after (a) and (b) are done.
3646 * d. if !logerror, flush the iclogs to disk, then seal them off
3650 xfs_log_force_umount(
3651 struct xfs_mount *mp,
3661 * If this happens during log recovery, don't worry about
3662 * locking; the log isn't open for business yet.
3665 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3666 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3668 XFS_BUF_DONE(mp->m_sb_bp);
3673 * Somebody could've already done the hard work for us.
3674 * No need to get locks for this.
3676 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3677 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3682 * We must hold both the GRANT lock and the LOG lock,
3683 * before we mark the filesystem SHUTDOWN and wake
3684 * everybody up to tell the bad news.
3686 spin_lock(&log->l_icloglock);
3687 spin_lock(&log->l_grant_lock);
3688 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3690 XFS_BUF_DONE(mp->m_sb_bp);
3693 * This flag is sort of redundant because of the mount flag, but
3694 * it's good to maintain the separation between the log and the rest
3697 log->l_flags |= XLOG_IO_ERROR;
3700 * If we hit a log error, we want to mark all the iclogs IOERROR
3701 * while we're still holding the loglock.
3704 retval = xlog_state_ioerror(log);
3705 spin_unlock(&log->l_icloglock);
3708 * We don't want anybody waiting for log reservations
3709 * after this. That means we have to wake up everybody
3710 * queued up on reserve_headq as well as write_headq.
3711 * In addition, we make sure in xlog_{re}grant_log_space
3712 * that we don't enqueue anything once the SHUTDOWN flag
3713 * is set, and this action is protected by the GRANTLOCK.
3715 if ((tic = log->l_reserve_headq)) {
3717 sv_signal(&tic->t_wait);
3719 } while (tic != log->l_reserve_headq);
3722 if ((tic = log->l_write_headq)) {
3724 sv_signal(&tic->t_wait);
3726 } while (tic != log->l_write_headq);
3728 spin_unlock(&log->l_grant_lock);
3730 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3733 * Force the incore logs to disk before shutting the
3734 * log down completely.
3736 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3738 spin_lock(&log->l_icloglock);
3739 retval = xlog_state_ioerror(log);
3740 spin_unlock(&log->l_icloglock);
3743 * Wake up everybody waiting on xfs_log_force.
3744 * Callback all log item committed functions as if the
3745 * log writes were completed.
3747 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3749 #ifdef XFSERRORDEBUG
3751 xlog_in_core_t *iclog;
3753 spin_lock(&log->l_icloglock);
3754 iclog = log->l_iclog;
3756 ASSERT(iclog->ic_callback == 0);
3757 iclog = iclog->ic_next;
3758 } while (iclog != log->l_iclog);
3759 spin_unlock(&log->l_icloglock);
3762 /* return non-zero if log IOERROR transition had already happened */
3767 xlog_iclogs_empty(xlog_t *log)
3769 xlog_in_core_t *iclog;
3771 iclog = log->l_iclog;
3773 /* endianness does not matter here, zero is zero in
3776 if (iclog->ic_header.h_num_logops)
3778 iclog = iclog->ic_next;
3779 } while (iclog != log->l_iclog);