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 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
52 /* Local miscellaneous function prototypes */
53 STATIC int xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
54 xlog_in_core_t **, xfs_lsn_t *);
55 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
56 xfs_buftarg_t *log_target,
57 xfs_daddr_t blk_offset,
59 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
60 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
61 STATIC void xlog_dealloc_log(xlog_t *log);
62 STATIC int xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
63 int nentries, struct xlog_ticket *tic,
65 xlog_in_core_t **commit_iclog,
68 /* local state machine functions */
69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
70 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
71 STATIC int xlog_state_get_iclog_space(xlog_t *log,
73 xlog_in_core_t **iclog,
74 xlog_ticket_t *ticket,
77 STATIC int xlog_state_release_iclog(xlog_t *log,
78 xlog_in_core_t *iclog);
79 STATIC void xlog_state_switch_iclogs(xlog_t *log,
80 xlog_in_core_t *iclog,
82 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
84 /* local functions to manipulate grant head */
85 STATIC int xlog_grant_log_space(xlog_t *log,
87 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
89 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
90 xlog_ticket_t *ticket);
91 STATIC int xlog_regrant_write_log_space(xlog_t *log,
92 xlog_ticket_t *ticket);
93 STATIC void xlog_ungrant_log_space(xlog_t *log,
94 xlog_ticket_t *ticket);
97 /* local ticket functions */
98 STATIC xlog_ticket_t *xlog_ticket_alloc(xlog_t *log,
105 STATIC void xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
106 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
107 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
108 int count, boolean_t syncing);
109 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
112 #define xlog_verify_dest_ptr(a,b)
113 #define xlog_verify_grant_head(a,b)
114 #define xlog_verify_iclog(a,b,c,d)
115 #define xlog_verify_tail_lsn(a,b,c)
118 STATIC int xlog_iclogs_empty(xlog_t *log);
122 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
126 tic->t_prev = (*qp)->t_prev;
127 (*qp)->t_prev->t_next = tic;
130 tic->t_prev = tic->t_next = tic;
134 tic->t_flags |= XLOG_TIC_IN_Q;
138 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
140 if (tic == tic->t_next) {
144 tic->t_next->t_prev = tic->t_prev;
145 tic->t_prev->t_next = tic->t_next;
148 tic->t_next = tic->t_prev = NULL;
149 tic->t_flags &= ~XLOG_TIC_IN_Q;
153 xlog_grant_sub_space(struct log *log, int bytes)
155 log->l_grant_write_bytes -= bytes;
156 if (log->l_grant_write_bytes < 0) {
157 log->l_grant_write_bytes += log->l_logsize;
158 log->l_grant_write_cycle--;
161 log->l_grant_reserve_bytes -= bytes;
162 if ((log)->l_grant_reserve_bytes < 0) {
163 log->l_grant_reserve_bytes += log->l_logsize;
164 log->l_grant_reserve_cycle--;
170 xlog_grant_add_space_write(struct log *log, int bytes)
172 int tmp = log->l_logsize - log->l_grant_write_bytes;
174 log->l_grant_write_bytes += bytes;
176 log->l_grant_write_cycle++;
177 log->l_grant_write_bytes = bytes - tmp;
182 xlog_grant_add_space_reserve(struct log *log, int bytes)
184 int tmp = log->l_logsize - log->l_grant_reserve_bytes;
186 log->l_grant_reserve_bytes += bytes;
188 log->l_grant_reserve_cycle++;
189 log->l_grant_reserve_bytes = bytes - tmp;
194 xlog_grant_add_space(struct log *log, int bytes)
196 xlog_grant_add_space_write(log, bytes);
197 xlog_grant_add_space_reserve(log, bytes);
201 xlog_tic_reset_res(xlog_ticket_t *tic)
204 tic->t_res_arr_sum = 0;
205 tic->t_res_num_ophdrs = 0;
209 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
211 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
212 /* add to overflow and start again */
213 tic->t_res_o_flow += tic->t_res_arr_sum;
215 tic->t_res_arr_sum = 0;
218 tic->t_res_arr[tic->t_res_num].r_len = len;
219 tic->t_res_arr[tic->t_res_num].r_type = type;
220 tic->t_res_arr_sum += len;
227 * 1. currblock field gets updated at startup and after in-core logs
228 * marked as with WANT_SYNC.
232 * This routine is called when a user of a log manager ticket is done with
233 * the reservation. If the ticket was ever used, then a commit record for
234 * the associated transaction is written out as a log operation header with
235 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
236 * a given ticket. If the ticket was one with a permanent reservation, then
237 * a few operations are done differently. Permanent reservation tickets by
238 * default don't release the reservation. They just commit the current
239 * transaction with the belief that the reservation is still needed. A flag
240 * must be passed in before permanent reservations are actually released.
241 * When these type of tickets are not released, they need to be set into
242 * the inited state again. By doing this, a start record will be written
243 * out when the next write occurs.
247 struct xfs_mount *mp,
248 struct xlog_ticket *ticket,
249 struct xlog_in_core **iclog,
252 struct log *log = mp->m_log;
255 if (XLOG_FORCED_SHUTDOWN(log) ||
257 * If nothing was ever written, don't write out commit record.
258 * If we get an error, just continue and give back the log ticket.
260 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
261 (xlog_commit_record(mp, ticket, iclog, &lsn)))) {
262 lsn = (xfs_lsn_t) -1;
263 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
264 flags |= XFS_LOG_REL_PERM_RESERV;
269 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
270 (flags & XFS_LOG_REL_PERM_RESERV)) {
271 trace_xfs_log_done_nonperm(log, ticket);
274 * Release ticket if not permanent reservation or a specific
275 * request has been made to release a permanent reservation.
277 xlog_ungrant_log_space(log, ticket);
278 xfs_log_ticket_put(ticket);
280 trace_xfs_log_done_perm(log, ticket);
282 xlog_regrant_reserve_log_space(log, ticket);
283 /* If this ticket was a permanent reservation and we aren't
284 * trying to release it, reset the inited flags; so next time
285 * we write, a start record will be written out.
287 ticket->t_flags |= XLOG_TIC_INITED;
294 * Attaches a new iclog I/O completion callback routine during
295 * transaction commit. If the log is in error state, a non-zero
296 * return code is handed back and the caller is responsible for
297 * executing the callback at an appropriate time.
301 struct xfs_mount *mp,
302 struct xlog_in_core *iclog,
303 xfs_log_callback_t *cb)
307 spin_lock(&iclog->ic_callback_lock);
308 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
310 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
311 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
313 *(iclog->ic_callback_tail) = cb;
314 iclog->ic_callback_tail = &(cb->cb_next);
316 spin_unlock(&iclog->ic_callback_lock);
321 xfs_log_release_iclog(
322 struct xfs_mount *mp,
323 struct xlog_in_core *iclog)
325 if (xlog_state_release_iclog(mp->m_log, iclog)) {
326 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
334 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
335 * to the reservation.
336 * 2. Potentially, push buffers at tail of log to disk.
338 * Each reservation is going to reserve extra space for a log record header.
339 * When writes happen to the on-disk log, we don't subtract the length of the
340 * log record header from any reservation. By wasting space in each
341 * reservation, we prevent over allocation problems.
345 struct xfs_mount *mp,
348 struct xlog_ticket **ticket,
353 struct log *log = mp->m_log;
354 struct xlog_ticket *internal_ticket;
357 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
358 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
360 if (XLOG_FORCED_SHUTDOWN(log))
361 return XFS_ERROR(EIO);
363 XFS_STATS_INC(xs_try_logspace);
366 if (*ticket != NULL) {
367 ASSERT(flags & XFS_LOG_PERM_RESERV);
368 internal_ticket = *ticket;
370 trace_xfs_log_reserve(log, internal_ticket);
372 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
373 retval = xlog_regrant_write_log_space(log, internal_ticket);
375 /* may sleep if need to allocate more tickets */
376 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
378 if (!internal_ticket)
379 return XFS_ERROR(ENOMEM);
380 internal_ticket->t_trans_type = t_type;
381 *ticket = internal_ticket;
383 trace_xfs_log_reserve(log, internal_ticket);
385 xlog_grant_push_ail(mp,
386 (internal_ticket->t_unit_res *
387 internal_ticket->t_cnt));
388 retval = xlog_grant_log_space(log, internal_ticket);
392 } /* xfs_log_reserve */
396 * Mount a log filesystem
398 * mp - ubiquitous xfs mount point structure
399 * log_target - buftarg of on-disk log device
400 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
401 * num_bblocks - Number of BBSIZE blocks in on-disk log
403 * Return error or zero.
408 xfs_buftarg_t *log_target,
409 xfs_daddr_t blk_offset,
414 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
415 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
418 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
420 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
423 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
424 if (IS_ERR(mp->m_log)) {
425 error = -PTR_ERR(mp->m_log);
430 * Initialize the AIL now we have a log.
432 error = xfs_trans_ail_init(mp);
434 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
437 mp->m_log->l_ailp = mp->m_ail;
440 * skip log recovery on a norecovery mount. pretend it all
443 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
444 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
447 mp->m_flags &= ~XFS_MOUNT_RDONLY;
449 error = xlog_recover(mp->m_log);
452 mp->m_flags |= XFS_MOUNT_RDONLY;
454 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
455 goto out_destroy_ail;
459 /* Normal transactions can now occur */
460 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
465 xfs_trans_ail_destroy(mp);
467 xlog_dealloc_log(mp->m_log);
473 * Finish the recovery of the file system. This is separate from
474 * the xfs_log_mount() call, because it depends on the code in
475 * xfs_mountfs() to read in the root and real-time bitmap inodes
476 * between calling xfs_log_mount() and here.
478 * mp - ubiquitous xfs mount point structure
481 xfs_log_mount_finish(xfs_mount_t *mp)
485 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
486 error = xlog_recover_finish(mp->m_log);
489 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
496 * Final log writes as part of unmount.
498 * Mark the filesystem clean as unmount happens. Note that during relocation
499 * this routine needs to be executed as part of source-bag while the
500 * deallocation must not be done until source-end.
504 * Unmount record used to have a string "Unmount filesystem--" in the
505 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
506 * We just write the magic number now since that particular field isn't
507 * currently architecture converted and "nUmount" is a bit foo.
508 * As far as I know, there weren't any dependencies on the old behaviour.
512 xfs_log_unmount_write(xfs_mount_t *mp)
514 xlog_t *log = mp->m_log;
515 xlog_in_core_t *iclog;
517 xlog_in_core_t *first_iclog;
519 xfs_log_iovec_t reg[1];
520 xlog_ticket_t *tic = NULL;
524 /* the data section must be 32 bit size aligned */
528 __uint32_t pad2; /* may as well make it 64 bits */
529 } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
532 * Don't write out unmount record on read-only mounts.
533 * Or, if we are doing a forced umount (typically because of IO errors).
535 if (mp->m_flags & XFS_MOUNT_RDONLY)
538 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
539 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
542 first_iclog = iclog = log->l_iclog;
544 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
545 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
546 ASSERT(iclog->ic_offset == 0);
548 iclog = iclog->ic_next;
549 } while (iclog != first_iclog);
551 if (! (XLOG_FORCED_SHUTDOWN(log))) {
552 reg[0].i_addr = (void*)&magic;
553 reg[0].i_len = sizeof(magic);
554 reg[0].i_type = XLOG_REG_TYPE_UNMOUNT;
556 error = xfs_log_reserve(mp, 600, 1, &tic,
557 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
559 /* remove inited flag */
560 ((xlog_ticket_t *)tic)->t_flags = 0;
561 error = xlog_write(mp, reg, 1, tic, &lsn,
562 NULL, XLOG_UNMOUNT_TRANS);
564 * At this point, we're umounting anyway,
565 * so there's no point in transitioning log state
566 * to IOERROR. Just continue...
571 xfs_fs_cmn_err(CE_ALERT, mp,
572 "xfs_log_unmount: unmount record failed");
576 spin_lock(&log->l_icloglock);
577 iclog = log->l_iclog;
578 atomic_inc(&iclog->ic_refcnt);
579 xlog_state_want_sync(log, iclog);
580 spin_unlock(&log->l_icloglock);
581 error = xlog_state_release_iclog(log, iclog);
583 spin_lock(&log->l_icloglock);
584 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
585 iclog->ic_state == XLOG_STATE_DIRTY)) {
586 if (!XLOG_FORCED_SHUTDOWN(log)) {
587 sv_wait(&iclog->ic_force_wait, PMEM,
588 &log->l_icloglock, s);
590 spin_unlock(&log->l_icloglock);
593 spin_unlock(&log->l_icloglock);
596 trace_xfs_log_umount_write(log, tic);
597 xlog_ungrant_log_space(log, tic);
598 xfs_log_ticket_put(tic);
602 * We're already in forced_shutdown mode, couldn't
603 * even attempt to write out the unmount transaction.
605 * Go through the motions of sync'ing and releasing
606 * the iclog, even though no I/O will actually happen,
607 * we need to wait for other log I/Os that may already
608 * be in progress. Do this as a separate section of
609 * code so we'll know if we ever get stuck here that
610 * we're in this odd situation of trying to unmount
611 * a file system that went into forced_shutdown as
612 * the result of an unmount..
614 spin_lock(&log->l_icloglock);
615 iclog = log->l_iclog;
616 atomic_inc(&iclog->ic_refcnt);
618 xlog_state_want_sync(log, iclog);
619 spin_unlock(&log->l_icloglock);
620 error = xlog_state_release_iclog(log, iclog);
622 spin_lock(&log->l_icloglock);
624 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
625 || iclog->ic_state == XLOG_STATE_DIRTY
626 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
628 sv_wait(&iclog->ic_force_wait, PMEM,
629 &log->l_icloglock, s);
631 spin_unlock(&log->l_icloglock);
636 } /* xfs_log_unmount_write */
639 * Deallocate log structures for unmount/relocation.
641 * We need to stop the aild from running before we destroy
642 * and deallocate the log as the aild references the log.
645 xfs_log_unmount(xfs_mount_t *mp)
647 xfs_trans_ail_destroy(mp);
648 xlog_dealloc_log(mp->m_log);
653 struct xfs_mount *mp,
654 struct xfs_log_item *item,
656 struct xfs_item_ops *ops)
658 item->li_mountp = mp;
659 item->li_ailp = mp->m_ail;
660 item->li_type = type;
665 * Write region vectors to log. The write happens using the space reservation
666 * of the ticket (tic). It is not a requirement that all writes for a given
667 * transaction occur with one call to xfs_log_write().
671 struct xfs_mount *mp,
672 struct xfs_log_iovec reg[],
674 struct xlog_ticket *tic,
675 xfs_lsn_t *start_lsn)
677 struct log *log = mp->m_log;
680 if (XLOG_FORCED_SHUTDOWN(log))
681 return XFS_ERROR(EIO);
683 error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0);
685 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
690 xfs_log_move_tail(xfs_mount_t *mp,
694 xlog_t *log = mp->m_log;
695 int need_bytes, free_bytes, cycle, bytes;
697 if (XLOG_FORCED_SHUTDOWN(log))
701 /* needed since sync_lsn is 64 bits */
702 spin_lock(&log->l_icloglock);
703 tail_lsn = log->l_last_sync_lsn;
704 spin_unlock(&log->l_icloglock);
707 spin_lock(&log->l_grant_lock);
709 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
713 log->l_tail_lsn = tail_lsn;
716 if ((tic = log->l_write_headq)) {
718 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
719 panic("Recovery problem");
721 cycle = log->l_grant_write_cycle;
722 bytes = log->l_grant_write_bytes;
723 free_bytes = xlog_space_left(log, cycle, bytes);
725 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
727 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
730 free_bytes -= tic->t_unit_res;
731 sv_signal(&tic->t_wait);
733 } while (tic != log->l_write_headq);
735 if ((tic = log->l_reserve_headq)) {
737 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
738 panic("Recovery problem");
740 cycle = log->l_grant_reserve_cycle;
741 bytes = log->l_grant_reserve_bytes;
742 free_bytes = xlog_space_left(log, cycle, bytes);
744 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
745 need_bytes = tic->t_unit_res*tic->t_cnt;
747 need_bytes = tic->t_unit_res;
748 if (free_bytes < need_bytes && tail_lsn != 1)
751 free_bytes -= need_bytes;
752 sv_signal(&tic->t_wait);
754 } while (tic != log->l_reserve_headq);
756 spin_unlock(&log->l_grant_lock);
757 } /* xfs_log_move_tail */
760 * Determine if we have a transaction that has gone to disk
761 * that needs to be covered. To begin the transition to the idle state
762 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
763 * If we are then in a state where covering is needed, the caller is informed
764 * that dummy transactions are required to move the log into the idle state.
766 * Because this is called as part of the sync process, we should also indicate
767 * that dummy transactions should be issued in anything but the covered or
768 * idle states. This ensures that the log tail is accurately reflected in
769 * the log at the end of the sync, hence if a crash occurrs avoids replay
770 * of transactions where the metadata is already on disk.
773 xfs_log_need_covered(xfs_mount_t *mp)
776 xlog_t *log = mp->m_log;
778 if (!xfs_fs_writable(mp))
781 spin_lock(&log->l_icloglock);
782 switch (log->l_covered_state) {
783 case XLOG_STATE_COVER_DONE:
784 case XLOG_STATE_COVER_DONE2:
785 case XLOG_STATE_COVER_IDLE:
787 case XLOG_STATE_COVER_NEED:
788 case XLOG_STATE_COVER_NEED2:
789 if (!xfs_trans_ail_tail(log->l_ailp) &&
790 xlog_iclogs_empty(log)) {
791 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
792 log->l_covered_state = XLOG_STATE_COVER_DONE;
794 log->l_covered_state = XLOG_STATE_COVER_DONE2;
801 spin_unlock(&log->l_icloglock);
805 /******************************************************************************
809 ******************************************************************************
812 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
813 * The log manager must keep track of the last LR which was committed
814 * to disk. The lsn of this LR will become the new tail_lsn whenever
815 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
816 * the situation where stuff could be written into the log but nothing
817 * was ever in the AIL when asked. Eventually, we panic since the
818 * tail hits the head.
820 * We may be holding the log iclog lock upon entering this routine.
823 xlog_assign_tail_lsn(xfs_mount_t *mp)
826 xlog_t *log = mp->m_log;
828 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
829 spin_lock(&log->l_grant_lock);
831 log->l_tail_lsn = tail_lsn;
833 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
835 spin_unlock(&log->l_grant_lock);
838 } /* xlog_assign_tail_lsn */
842 * Return the space in the log between the tail and the head. The head
843 * is passed in the cycle/bytes formal parms. In the special case where
844 * the reserve head has wrapped passed the tail, this calculation is no
845 * longer valid. In this case, just return 0 which means there is no space
846 * in the log. This works for all places where this function is called
847 * with the reserve head. Of course, if the write head were to ever
848 * wrap the tail, we should blow up. Rather than catch this case here,
849 * we depend on other ASSERTions in other parts of the code. XXXmiken
851 * This code also handles the case where the reservation head is behind
852 * the tail. The details of this case are described below, but the end
853 * result is that we return the size of the log as the amount of space left.
856 xlog_space_left(xlog_t *log, int cycle, int bytes)
862 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
863 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
864 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
865 free_bytes = log->l_logsize - (bytes - tail_bytes);
866 } else if ((tail_cycle + 1) < cycle) {
868 } else if (tail_cycle < cycle) {
869 ASSERT(tail_cycle == (cycle - 1));
870 free_bytes = tail_bytes - bytes;
873 * The reservation head is behind the tail.
874 * In this case we just want to return the size of the
875 * log as the amount of space left.
877 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
878 "xlog_space_left: head behind tail\n"
879 " tail_cycle = %d, tail_bytes = %d\n"
880 " GH cycle = %d, GH bytes = %d",
881 tail_cycle, tail_bytes, cycle, bytes);
883 free_bytes = log->l_logsize;
886 } /* xlog_space_left */
890 * Log function which is called when an io completes.
892 * The log manager needs its own routine, in order to control what
893 * happens with the buffer after the write completes.
896 xlog_iodone(xfs_buf_t *bp)
898 xlog_in_core_t *iclog;
902 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
903 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
904 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
909 * If the _XFS_BARRIER_FAILED flag was set by a lower
910 * layer, it means the underlying device no longer supports
911 * barrier I/O. Warn loudly and turn off barriers.
913 if (bp->b_flags & _XFS_BARRIER_FAILED) {
914 bp->b_flags &= ~_XFS_BARRIER_FAILED;
915 l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
916 xfs_fs_cmn_err(CE_WARN, l->l_mp,
917 "xlog_iodone: Barriers are no longer supported"
918 " by device. Disabling barriers\n");
922 * Race to shutdown the filesystem if we see an error.
924 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
925 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
926 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
928 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
930 * This flag will be propagated to the trans-committed
931 * callback routines to let them know that the log-commit
934 aborted = XFS_LI_ABORTED;
935 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
936 aborted = XFS_LI_ABORTED;
939 /* log I/O is always issued ASYNC */
940 ASSERT(XFS_BUF_ISASYNC(bp));
941 xlog_state_done_syncing(iclog, aborted);
943 * do not reference the buffer (bp) here as we could race
944 * with it being freed after writing the unmount record to the
951 * Return size of each in-core log record buffer.
953 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
955 * If the filesystem blocksize is too large, we may need to choose a
956 * larger size since the directory code currently logs entire blocks.
960 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
966 if (mp->m_logbufs <= 0)
967 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
969 log->l_iclog_bufs = mp->m_logbufs;
972 * Buffer size passed in from mount system call.
974 if (mp->m_logbsize > 0) {
975 size = log->l_iclog_size = mp->m_logbsize;
976 log->l_iclog_size_log = 0;
978 log->l_iclog_size_log++;
982 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
983 /* # headers = size / 32k
984 * one header holds cycles from 32k of data
987 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
988 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
990 log->l_iclog_hsize = xhdrs << BBSHIFT;
991 log->l_iclog_heads = xhdrs;
993 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
994 log->l_iclog_hsize = BBSIZE;
995 log->l_iclog_heads = 1;
1000 /* All machines use 32kB buffers by default. */
1001 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1002 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1004 /* the default log size is 16k or 32k which is one header sector */
1005 log->l_iclog_hsize = BBSIZE;
1006 log->l_iclog_heads = 1;
1009 /* are we being asked to make the sizes selected above visible? */
1010 if (mp->m_logbufs == 0)
1011 mp->m_logbufs = log->l_iclog_bufs;
1012 if (mp->m_logbsize == 0)
1013 mp->m_logbsize = log->l_iclog_size;
1014 } /* xlog_get_iclog_buffer_size */
1018 * This routine initializes some of the log structure for a given mount point.
1019 * Its primary purpose is to fill in enough, so recovery can occur. However,
1020 * some other stuff may be filled in too.
1023 xlog_alloc_log(xfs_mount_t *mp,
1024 xfs_buftarg_t *log_target,
1025 xfs_daddr_t blk_offset,
1029 xlog_rec_header_t *head;
1030 xlog_in_core_t **iclogp;
1031 xlog_in_core_t *iclog, *prev_iclog=NULL;
1037 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1039 xlog_warn("XFS: Log allocation failed: No memory!");
1044 log->l_targ = log_target;
1045 log->l_logsize = BBTOB(num_bblks);
1046 log->l_logBBstart = blk_offset;
1047 log->l_logBBsize = num_bblks;
1048 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1049 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1051 log->l_prev_block = -1;
1052 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1053 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1054 log->l_last_sync_lsn = log->l_tail_lsn;
1055 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1056 log->l_grant_reserve_cycle = 1;
1057 log->l_grant_write_cycle = 1;
1059 error = EFSCORRUPTED;
1060 if (xfs_sb_version_hassector(&mp->m_sb)) {
1061 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1062 if (log->l_sectbb_log < 0 ||
1063 log->l_sectbb_log > mp->m_sectbb_log) {
1064 xlog_warn("XFS: Log sector size (0x%x) out of range.",
1069 /* for larger sector sizes, must have v2 or external log */
1070 if (log->l_sectbb_log != 0 &&
1071 (log->l_logBBstart != 0 &&
1072 !xfs_sb_version_haslogv2(&mp->m_sb))) {
1073 xlog_warn("XFS: log sector size (0x%x) invalid "
1074 "for configuration.", log->l_sectbb_log);
1077 if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1078 xlog_warn("XFS: Log sector log (0x%x) too small.",
1079 mp->m_sb.sb_logsectlog);
1083 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1085 xlog_get_iclog_buffer_size(mp, log);
1088 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1091 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1092 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1093 ASSERT(XFS_BUF_ISBUSY(bp));
1094 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1097 spin_lock_init(&log->l_icloglock);
1098 spin_lock_init(&log->l_grant_lock);
1099 sv_init(&log->l_flush_wait, 0, "flush_wait");
1101 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1102 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1104 iclogp = &log->l_iclog;
1106 * The amount of memory to allocate for the iclog structure is
1107 * rather funky due to the way the structure is defined. It is
1108 * done this way so that we can use different sizes for machines
1109 * with different amounts of memory. See the definition of
1110 * xlog_in_core_t in xfs_log_priv.h for details.
1112 iclogsize = log->l_iclog_size;
1113 ASSERT(log->l_iclog_size >= 4096);
1114 for (i=0; i < log->l_iclog_bufs; i++) {
1115 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1117 goto out_free_iclog;
1120 iclog->ic_prev = prev_iclog;
1123 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1125 goto out_free_iclog;
1126 if (!XFS_BUF_CPSEMA(bp))
1128 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1129 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1131 iclog->ic_data = bp->b_addr;
1133 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1135 head = &iclog->ic_header;
1136 memset(head, 0, sizeof(xlog_rec_header_t));
1137 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1138 head->h_version = cpu_to_be32(
1139 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1140 head->h_size = cpu_to_be32(log->l_iclog_size);
1142 head->h_fmt = cpu_to_be32(XLOG_FMT);
1143 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1145 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1146 iclog->ic_state = XLOG_STATE_ACTIVE;
1147 iclog->ic_log = log;
1148 atomic_set(&iclog->ic_refcnt, 0);
1149 spin_lock_init(&iclog->ic_callback_lock);
1150 iclog->ic_callback_tail = &(iclog->ic_callback);
1151 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1153 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1154 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1155 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1156 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1158 iclogp = &iclog->ic_next;
1160 *iclogp = log->l_iclog; /* complete ring */
1161 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1166 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1167 prev_iclog = iclog->ic_next;
1169 sv_destroy(&iclog->ic_force_wait);
1170 sv_destroy(&iclog->ic_write_wait);
1171 xfs_buf_free(iclog->ic_bp);
1175 spinlock_destroy(&log->l_icloglock);
1176 spinlock_destroy(&log->l_grant_lock);
1177 xfs_buf_free(log->l_xbuf);
1181 return ERR_PTR(-error);
1182 } /* xlog_alloc_log */
1186 * Write out the commit record of a transaction associated with the given
1187 * ticket. Return the lsn of the commit record.
1190 xlog_commit_record(xfs_mount_t *mp,
1191 xlog_ticket_t *ticket,
1192 xlog_in_core_t **iclog,
1193 xfs_lsn_t *commitlsnp)
1196 xfs_log_iovec_t reg[1];
1198 reg[0].i_addr = NULL;
1200 reg[0].i_type = XLOG_REG_TYPE_COMMIT;
1202 ASSERT_ALWAYS(iclog);
1203 if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1204 iclog, XLOG_COMMIT_TRANS))) {
1205 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1208 } /* xlog_commit_record */
1212 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1213 * log space. This code pushes on the lsn which would supposedly free up
1214 * the 25% which we want to leave free. We may need to adopt a policy which
1215 * pushes on an lsn which is further along in the log once we reach the high
1216 * water mark. In this manner, we would be creating a low water mark.
1219 xlog_grant_push_ail(xfs_mount_t *mp,
1222 xlog_t *log = mp->m_log; /* pointer to the log */
1223 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1224 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1225 int free_blocks; /* free blocks left to write to */
1226 int free_bytes; /* free bytes left to write to */
1227 int threshold_block; /* block in lsn we'd like to be at */
1228 int threshold_cycle; /* lsn cycle we'd like to be at */
1231 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1233 spin_lock(&log->l_grant_lock);
1234 free_bytes = xlog_space_left(log,
1235 log->l_grant_reserve_cycle,
1236 log->l_grant_reserve_bytes);
1237 tail_lsn = log->l_tail_lsn;
1238 free_blocks = BTOBBT(free_bytes);
1241 * Set the threshold for the minimum number of free blocks in the
1242 * log to the maximum of what the caller needs, one quarter of the
1243 * log, and 256 blocks.
1245 free_threshold = BTOBB(need_bytes);
1246 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1247 free_threshold = MAX(free_threshold, 256);
1248 if (free_blocks < free_threshold) {
1249 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1250 threshold_cycle = CYCLE_LSN(tail_lsn);
1251 if (threshold_block >= log->l_logBBsize) {
1252 threshold_block -= log->l_logBBsize;
1253 threshold_cycle += 1;
1255 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1257 /* Don't pass in an lsn greater than the lsn of the last
1258 * log record known to be on disk.
1260 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1261 threshold_lsn = log->l_last_sync_lsn;
1263 spin_unlock(&log->l_grant_lock);
1266 * Get the transaction layer to kick the dirty buffers out to
1267 * disk asynchronously. No point in trying to do this if
1268 * the filesystem is shutting down.
1270 if (threshold_lsn &&
1271 !XLOG_FORCED_SHUTDOWN(log))
1272 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1273 } /* xlog_grant_push_ail */
1276 * The bdstrat callback function for log bufs. This gives us a central
1277 * place to trap bufs in case we get hit by a log I/O error and need to
1278 * shutdown. Actually, in practice, even when we didn't get a log error,
1279 * we transition the iclogs to IOERROR state *after* flushing all existing
1280 * iclogs to disk. This is because we don't want anymore new transactions to be
1281 * started or completed afterwards.
1287 struct xlog_in_core *iclog;
1289 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1290 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1291 XFS_BUF_ERROR(bp, EIO);
1295 * It would seem logical to return EIO here, but we rely on
1296 * the log state machine to propagate I/O errors instead of
1302 bp->b_flags |= _XBF_RUN_QUEUES;
1303 xfs_buf_iorequest(bp);
1308 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1309 * fashion. Previously, we should have moved the current iclog
1310 * ptr in the log to point to the next available iclog. This allows further
1311 * write to continue while this code syncs out an iclog ready to go.
1312 * Before an in-core log can be written out, the data section must be scanned
1313 * to save away the 1st word of each BBSIZE block into the header. We replace
1314 * it with the current cycle count. Each BBSIZE block is tagged with the
1315 * cycle count because there in an implicit assumption that drives will
1316 * guarantee that entire 512 byte blocks get written at once. In other words,
1317 * we can't have part of a 512 byte block written and part not written. By
1318 * tagging each block, we will know which blocks are valid when recovering
1319 * after an unclean shutdown.
1321 * This routine is single threaded on the iclog. No other thread can be in
1322 * this routine with the same iclog. Changing contents of iclog can there-
1323 * fore be done without grabbing the state machine lock. Updating the global
1324 * log will require grabbing the lock though.
1326 * The entire log manager uses a logical block numbering scheme. Only
1327 * log_sync (and then only bwrite()) know about the fact that the log may
1328 * not start with block zero on a given device. The log block start offset
1329 * is added immediately before calling bwrite().
1333 xlog_sync(xlog_t *log,
1334 xlog_in_core_t *iclog)
1336 xfs_caddr_t dptr; /* pointer to byte sized element */
1339 uint count; /* byte count of bwrite */
1340 uint count_init; /* initial count before roundup */
1341 int roundoff; /* roundoff to BB or stripe */
1342 int split = 0; /* split write into two regions */
1344 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1346 XFS_STATS_INC(xs_log_writes);
1347 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1349 /* Add for LR header */
1350 count_init = log->l_iclog_hsize + iclog->ic_offset;
1352 /* Round out the log write size */
1353 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1354 /* we have a v2 stripe unit to use */
1355 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1357 count = BBTOB(BTOBB(count_init));
1359 roundoff = count - count_init;
1360 ASSERT(roundoff >= 0);
1361 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1362 roundoff < log->l_mp->m_sb.sb_logsunit)
1364 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1365 roundoff < BBTOB(1)));
1367 /* move grant heads by roundoff in sync */
1368 spin_lock(&log->l_grant_lock);
1369 xlog_grant_add_space(log, roundoff);
1370 spin_unlock(&log->l_grant_lock);
1372 /* put cycle number in every block */
1373 xlog_pack_data(log, iclog, roundoff);
1375 /* real byte length */
1377 iclog->ic_header.h_len =
1378 cpu_to_be32(iclog->ic_offset + roundoff);
1380 iclog->ic_header.h_len =
1381 cpu_to_be32(iclog->ic_offset);
1385 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1386 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1387 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1389 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1391 /* Do we need to split this write into 2 parts? */
1392 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1393 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1394 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1395 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1397 iclog->ic_bwritecnt = 1;
1399 XFS_BUF_SET_COUNT(bp, count);
1400 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1401 XFS_BUF_ZEROFLAGS(bp);
1404 bp->b_flags |= XBF_LOG_BUFFER;
1406 * Do an ordered write for the log block.
1407 * Its unnecessary to flush the first split block in the log wrap case.
1409 if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1410 XFS_BUF_ORDERED(bp);
1412 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1413 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1415 xlog_verify_iclog(log, iclog, count, B_TRUE);
1417 /* account for log which doesn't start at block #0 */
1418 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1420 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1425 if ((error = xlog_bdstrat(bp))) {
1426 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1431 bp = iclog->ic_log->l_xbuf;
1432 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1434 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1435 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1436 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1437 (__psint_t)count), split);
1438 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1439 XFS_BUF_ZEROFLAGS(bp);
1442 bp->b_flags |= XBF_LOG_BUFFER;
1443 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1444 XFS_BUF_ORDERED(bp);
1445 dptr = XFS_BUF_PTR(bp);
1447 * Bump the cycle numbers at the start of each block
1448 * since this part of the buffer is at the start of
1449 * a new cycle. Watch out for the header magic number
1452 for (i = 0; i < split; i += BBSIZE) {
1453 be32_add_cpu((__be32 *)dptr, 1);
1454 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1455 be32_add_cpu((__be32 *)dptr, 1);
1459 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1460 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1462 /* account for internal log which doesn't start at block #0 */
1463 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1465 if ((error = xlog_bdstrat(bp))) {
1466 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1467 bp, XFS_BUF_ADDR(bp));
1476 * Deallocate a log structure
1479 xlog_dealloc_log(xlog_t *log)
1481 xlog_in_core_t *iclog, *next_iclog;
1484 iclog = log->l_iclog;
1485 for (i=0; i<log->l_iclog_bufs; i++) {
1486 sv_destroy(&iclog->ic_force_wait);
1487 sv_destroy(&iclog->ic_write_wait);
1488 xfs_buf_free(iclog->ic_bp);
1489 next_iclog = iclog->ic_next;
1493 spinlock_destroy(&log->l_icloglock);
1494 spinlock_destroy(&log->l_grant_lock);
1496 xfs_buf_free(log->l_xbuf);
1497 log->l_mp->m_log = NULL;
1499 } /* xlog_dealloc_log */
1502 * Update counters atomically now that memcpy is done.
1506 xlog_state_finish_copy(xlog_t *log,
1507 xlog_in_core_t *iclog,
1511 spin_lock(&log->l_icloglock);
1513 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1514 iclog->ic_offset += copy_bytes;
1516 spin_unlock(&log->l_icloglock);
1517 } /* xlog_state_finish_copy */
1523 * print out info relating to regions written which consume
1527 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1530 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1532 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1533 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1554 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1597 xfs_fs_cmn_err(CE_WARN, mp,
1598 "xfs_log_write: reservation summary:\n"
1599 " trans type = %s (%u)\n"
1600 " unit res = %d bytes\n"
1601 " current res = %d bytes\n"
1602 " total reg = %u bytes (o/flow = %u bytes)\n"
1603 " ophdrs = %u (ophdr space = %u bytes)\n"
1604 " ophdr + reg = %u bytes\n"
1605 " num regions = %u\n",
1606 ((ticket->t_trans_type <= 0 ||
1607 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1608 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1609 ticket->t_trans_type,
1612 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1613 ticket->t_res_num_ophdrs, ophdr_spc,
1614 ticket->t_res_arr_sum +
1615 ticket->t_res_o_flow + ophdr_spc,
1618 for (i = 0; i < ticket->t_res_num; i++) {
1619 uint r_type = ticket->t_res_arr[i].r_type;
1621 "region[%u]: %s - %u bytes\n",
1623 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1624 "bad-rtype" : res_type_str[r_type-1]),
1625 ticket->t_res_arr[i].r_len);
1630 * Write some region out to in-core log
1632 * This will be called when writing externally provided regions or when
1633 * writing out a commit record for a given transaction.
1635 * General algorithm:
1636 * 1. Find total length of this write. This may include adding to the
1637 * lengths passed in.
1638 * 2. Check whether we violate the tickets reservation.
1639 * 3. While writing to this iclog
1640 * A. Reserve as much space in this iclog as can get
1641 * B. If this is first write, save away start lsn
1642 * C. While writing this region:
1643 * 1. If first write of transaction, write start record
1644 * 2. Write log operation header (header per region)
1645 * 3. Find out if we can fit entire region into this iclog
1646 * 4. Potentially, verify destination memcpy ptr
1647 * 5. Memcpy (partial) region
1648 * 6. If partial copy, release iclog; otherwise, continue
1649 * copying more regions into current iclog
1650 * 4. Mark want sync bit (in simulation mode)
1651 * 5. Release iclog for potential flush to on-disk log.
1654 * 1. Panic if reservation is overrun. This should never happen since
1655 * reservation amounts are generated internal to the filesystem.
1657 * 1. Tickets are single threaded data structures.
1658 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1659 * syncing routine. When a single log_write region needs to span
1660 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1661 * on all log operation writes which don't contain the end of the
1662 * region. The XLOG_END_TRANS bit is used for the in-core log
1663 * operation which contains the end of the continued log_write region.
1664 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1665 * we don't really know exactly how much space will be used. As a result,
1666 * we don't update ic_offset until the end when we know exactly how many
1667 * bytes have been written out.
1671 struct xfs_mount *mp,
1672 struct xfs_log_iovec reg[],
1674 struct xlog_ticket *ticket,
1675 xfs_lsn_t *start_lsn,
1676 struct xlog_in_core **commit_iclog,
1679 xlog_t *log = mp->m_log;
1680 xlog_in_core_t *iclog = NULL; /* ptr to current in-core log */
1681 xlog_op_header_t *logop_head; /* ptr to log operation header */
1682 __psint_t ptr; /* copy address into data region */
1683 int len; /* # xlog_write() bytes 2 still copy */
1684 int index; /* region index currently copying */
1685 int log_offset; /* offset (from 0) into data region */
1686 int start_rec_copy; /* # bytes to copy for start record */
1687 int partial_copy; /* did we split a region? */
1688 int partial_copy_len;/* # bytes copied if split region */
1689 int need_copy; /* # bytes need to memcpy this region */
1690 int copy_len; /* # bytes actually memcpy'ing */
1691 int copy_off; /* # bytes from entry start */
1692 int contwr; /* continued write of in-core log? */
1694 int record_cnt = 0, data_cnt = 0;
1696 partial_copy_len = partial_copy = 0;
1698 /* Calculate potential maximum space. Each region gets its own
1699 * xlog_op_header_t and may need to be double word aligned.
1702 if (ticket->t_flags & XLOG_TIC_INITED) { /* acct for start rec of xact */
1703 len += sizeof(xlog_op_header_t);
1704 ticket->t_res_num_ophdrs++;
1707 for (index = 0; index < nentries; index++) {
1708 len += sizeof(xlog_op_header_t); /* each region gets >= 1 */
1709 ticket->t_res_num_ophdrs++;
1710 len += reg[index].i_len;
1711 xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1713 contwr = *start_lsn = 0;
1715 if (ticket->t_curr_res < len) {
1716 xlog_print_tic_res(mp, ticket);
1719 "xfs_log_write: reservation ran out. Need to up reservation");
1721 /* Customer configurable panic */
1722 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1723 "xfs_log_write: reservation ran out. Need to up reservation");
1724 /* If we did not panic, shutdown the filesystem */
1725 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1728 ticket->t_curr_res -= len;
1730 for (index = 0; index < nentries; ) {
1731 if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1732 &contwr, &log_offset)))
1735 ASSERT(log_offset <= iclog->ic_size - 1);
1736 ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1738 /* start_lsn is the first lsn written to. That's all we need. */
1740 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1742 /* This loop writes out as many regions as can fit in the amount
1743 * of space which was allocated by xlog_state_get_iclog_space().
1745 while (index < nentries) {
1746 ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1747 ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1750 /* If first write for transaction, insert start record.
1751 * We can't be trying to commit if we are inited. We can't
1752 * have any "partial_copy" if we are inited.
1754 if (ticket->t_flags & XLOG_TIC_INITED) {
1755 logop_head = (xlog_op_header_t *)ptr;
1756 logop_head->oh_tid = cpu_to_be32(ticket->t_tid);
1757 logop_head->oh_clientid = ticket->t_clientid;
1758 logop_head->oh_len = 0;
1759 logop_head->oh_flags = XLOG_START_TRANS;
1760 logop_head->oh_res2 = 0;
1761 ticket->t_flags &= ~XLOG_TIC_INITED; /* clear bit */
1764 start_rec_copy = sizeof(xlog_op_header_t);
1765 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1768 /* Copy log operation header directly into data section */
1769 logop_head = (xlog_op_header_t *)ptr;
1770 logop_head->oh_tid = cpu_to_be32(ticket->t_tid);
1771 logop_head->oh_clientid = ticket->t_clientid;
1772 logop_head->oh_res2 = 0;
1774 /* header copied directly */
1775 xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1777 /* are we copying a commit or unmount record? */
1778 logop_head->oh_flags = flags;
1781 * We've seen logs corrupted with bad transaction client
1782 * ids. This makes sure that XFS doesn't generate them on.
1783 * Turn this into an EIO and shut down the filesystem.
1785 switch (logop_head->oh_clientid) {
1786 case XFS_TRANSACTION:
1791 xfs_fs_cmn_err(CE_WARN, mp,
1792 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1793 logop_head->oh_clientid, ticket);
1794 return XFS_ERROR(EIO);
1797 /* Partial write last time? => (partial_copy != 0)
1798 * need_copy is the amount we'd like to copy if everything could
1799 * fit in the current memcpy.
1801 need_copy = reg[index].i_len - partial_copy_len;
1803 copy_off = partial_copy_len;
1804 if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1805 copy_len = need_copy;
1806 logop_head->oh_len = cpu_to_be32(copy_len);
1808 logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1809 partial_copy_len = partial_copy = 0;
1810 } else { /* partial write */
1811 copy_len = iclog->ic_size - log_offset;
1812 logop_head->oh_len = cpu_to_be32(copy_len);
1813 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1815 logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1816 partial_copy_len += copy_len;
1818 len += sizeof(xlog_op_header_t); /* from splitting of region */
1819 /* account for new log op header */
1820 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1821 ticket->t_res_num_ophdrs++;
1823 xlog_verify_dest_ptr(log, ptr);
1826 ASSERT(copy_len >= 0);
1827 memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1828 xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1830 /* make copy_len total bytes copied, including headers */
1831 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1833 data_cnt += contwr ? copy_len : 0;
1834 if (partial_copy) { /* copied partial region */
1835 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1836 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1837 record_cnt = data_cnt = 0;
1838 if ((error = xlog_state_release_iclog(log, iclog)))
1840 break; /* don't increment index */
1841 } else { /* copied entire region */
1843 partial_copy_len = partial_copy = 0;
1845 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1846 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1847 record_cnt = data_cnt = 0;
1848 spin_lock(&log->l_icloglock);
1849 xlog_state_want_sync(log, iclog);
1850 spin_unlock(&log->l_icloglock);
1852 ASSERT(flags & XLOG_COMMIT_TRANS);
1853 *commit_iclog = iclog;
1854 } else if ((error = xlog_state_release_iclog(log, iclog)))
1856 if (index == nentries)
1857 return 0; /* we are done */
1861 } /* if (partial_copy) */
1862 } /* while (index < nentries) */
1863 } /* for (index = 0; index < nentries; ) */
1866 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1868 ASSERT(flags & XLOG_COMMIT_TRANS);
1869 *commit_iclog = iclog;
1872 return xlog_state_release_iclog(log, iclog);
1876 /*****************************************************************************
1878 * State Machine functions
1880 *****************************************************************************
1883 /* Clean iclogs starting from the head. This ordering must be
1884 * maintained, so an iclog doesn't become ACTIVE beyond one that
1885 * is SYNCING. This is also required to maintain the notion that we use
1886 * a ordered wait queue to hold off would be writers to the log when every
1887 * iclog is trying to sync to disk.
1889 * State Change: DIRTY -> ACTIVE
1892 xlog_state_clean_log(xlog_t *log)
1894 xlog_in_core_t *iclog;
1897 iclog = log->l_iclog;
1899 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1900 iclog->ic_state = XLOG_STATE_ACTIVE;
1901 iclog->ic_offset = 0;
1902 ASSERT(iclog->ic_callback == NULL);
1904 * If the number of ops in this iclog indicate it just
1905 * contains the dummy transaction, we can
1906 * change state into IDLE (the second time around).
1907 * Otherwise we should change the state into
1909 * We don't need to cover the dummy.
1912 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
1917 * We have two dirty iclogs so start over
1918 * This could also be num of ops indicates
1919 * this is not the dummy going out.
1923 iclog->ic_header.h_num_logops = 0;
1924 memset(iclog->ic_header.h_cycle_data, 0,
1925 sizeof(iclog->ic_header.h_cycle_data));
1926 iclog->ic_header.h_lsn = 0;
1927 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1930 break; /* stop cleaning */
1931 iclog = iclog->ic_next;
1932 } while (iclog != log->l_iclog);
1934 /* log is locked when we are called */
1936 * Change state for the dummy log recording.
1937 * We usually go to NEED. But we go to NEED2 if the changed indicates
1938 * we are done writing the dummy record.
1939 * If we are done with the second dummy recored (DONE2), then
1943 switch (log->l_covered_state) {
1944 case XLOG_STATE_COVER_IDLE:
1945 case XLOG_STATE_COVER_NEED:
1946 case XLOG_STATE_COVER_NEED2:
1947 log->l_covered_state = XLOG_STATE_COVER_NEED;
1950 case XLOG_STATE_COVER_DONE:
1952 log->l_covered_state = XLOG_STATE_COVER_NEED2;
1954 log->l_covered_state = XLOG_STATE_COVER_NEED;
1957 case XLOG_STATE_COVER_DONE2:
1959 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1961 log->l_covered_state = XLOG_STATE_COVER_NEED;
1968 } /* xlog_state_clean_log */
1971 xlog_get_lowest_lsn(
1974 xlog_in_core_t *lsn_log;
1975 xfs_lsn_t lowest_lsn, lsn;
1977 lsn_log = log->l_iclog;
1980 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
1981 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
1982 if ((lsn && !lowest_lsn) ||
1983 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
1987 lsn_log = lsn_log->ic_next;
1988 } while (lsn_log != log->l_iclog);
1994 xlog_state_do_callback(
1997 xlog_in_core_t *ciclog)
1999 xlog_in_core_t *iclog;
2000 xlog_in_core_t *first_iclog; /* used to know when we've
2001 * processed all iclogs once */
2002 xfs_log_callback_t *cb, *cb_next;
2004 xfs_lsn_t lowest_lsn;
2005 int ioerrors; /* counter: iclogs with errors */
2006 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2007 int funcdidcallbacks; /* flag: function did callbacks */
2008 int repeats; /* for issuing console warnings if
2009 * looping too many times */
2012 spin_lock(&log->l_icloglock);
2013 first_iclog = iclog = log->l_iclog;
2015 funcdidcallbacks = 0;
2020 * Scan all iclogs starting with the one pointed to by the
2021 * log. Reset this starting point each time the log is
2022 * unlocked (during callbacks).
2024 * Keep looping through iclogs until one full pass is made
2025 * without running any callbacks.
2027 first_iclog = log->l_iclog;
2028 iclog = log->l_iclog;
2029 loopdidcallbacks = 0;
2034 /* skip all iclogs in the ACTIVE & DIRTY states */
2035 if (iclog->ic_state &
2036 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2037 iclog = iclog->ic_next;
2042 * Between marking a filesystem SHUTDOWN and stopping
2043 * the log, we do flush all iclogs to disk (if there
2044 * wasn't a log I/O error). So, we do want things to
2045 * go smoothly in case of just a SHUTDOWN w/o a
2048 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2050 * Can only perform callbacks in order. Since
2051 * this iclog is not in the DONE_SYNC/
2052 * DO_CALLBACK state, we skip the rest and
2053 * just try to clean up. If we set our iclog
2054 * to DO_CALLBACK, we will not process it when
2055 * we retry since a previous iclog is in the
2056 * CALLBACK and the state cannot change since
2057 * we are holding the l_icloglock.
2059 if (!(iclog->ic_state &
2060 (XLOG_STATE_DONE_SYNC |
2061 XLOG_STATE_DO_CALLBACK))) {
2062 if (ciclog && (ciclog->ic_state ==
2063 XLOG_STATE_DONE_SYNC)) {
2064 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2069 * We now have an iclog that is in either the
2070 * DO_CALLBACK or DONE_SYNC states. The other
2071 * states (WANT_SYNC, SYNCING, or CALLBACK were
2072 * caught by the above if and are going to
2073 * clean (i.e. we aren't doing their callbacks)
2078 * We will do one more check here to see if we
2079 * have chased our tail around.
2082 lowest_lsn = xlog_get_lowest_lsn(log);
2084 XFS_LSN_CMP(lowest_lsn,
2085 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2086 iclog = iclog->ic_next;
2087 continue; /* Leave this iclog for
2091 iclog->ic_state = XLOG_STATE_CALLBACK;
2093 spin_unlock(&log->l_icloglock);
2095 /* l_last_sync_lsn field protected by
2096 * l_grant_lock. Don't worry about iclog's lsn.
2097 * No one else can be here except us.
2099 spin_lock(&log->l_grant_lock);
2100 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2101 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2102 log->l_last_sync_lsn =
2103 be64_to_cpu(iclog->ic_header.h_lsn);
2104 spin_unlock(&log->l_grant_lock);
2107 spin_unlock(&log->l_icloglock);
2112 * Keep processing entries in the callback list until
2113 * we come around and it is empty. We need to
2114 * atomically see that the list is empty and change the
2115 * state to DIRTY so that we don't miss any more
2116 * callbacks being added.
2118 spin_lock(&iclog->ic_callback_lock);
2119 cb = iclog->ic_callback;
2121 iclog->ic_callback_tail = &(iclog->ic_callback);
2122 iclog->ic_callback = NULL;
2123 spin_unlock(&iclog->ic_callback_lock);
2125 /* perform callbacks in the order given */
2126 for (; cb; cb = cb_next) {
2127 cb_next = cb->cb_next;
2128 cb->cb_func(cb->cb_arg, aborted);
2130 spin_lock(&iclog->ic_callback_lock);
2131 cb = iclog->ic_callback;
2137 spin_lock(&log->l_icloglock);
2138 ASSERT(iclog->ic_callback == NULL);
2139 spin_unlock(&iclog->ic_callback_lock);
2140 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2141 iclog->ic_state = XLOG_STATE_DIRTY;
2144 * Transition from DIRTY to ACTIVE if applicable.
2145 * NOP if STATE_IOERROR.
2147 xlog_state_clean_log(log);
2149 /* wake up threads waiting in xfs_log_force() */
2150 sv_broadcast(&iclog->ic_force_wait);
2152 iclog = iclog->ic_next;
2153 } while (first_iclog != iclog);
2155 if (repeats > 5000) {
2156 flushcnt += repeats;
2158 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2159 "%s: possible infinite loop (%d iterations)",
2160 __func__, flushcnt);
2162 } while (!ioerrors && loopdidcallbacks);
2165 * make one last gasp attempt to see if iclogs are being left in
2169 if (funcdidcallbacks) {
2170 first_iclog = iclog = log->l_iclog;
2172 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2174 * Terminate the loop if iclogs are found in states
2175 * which will cause other threads to clean up iclogs.
2177 * SYNCING - i/o completion will go through logs
2178 * DONE_SYNC - interrupt thread should be waiting for
2180 * IOERROR - give up hope all ye who enter here
2182 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2183 iclog->ic_state == XLOG_STATE_SYNCING ||
2184 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2185 iclog->ic_state == XLOG_STATE_IOERROR )
2187 iclog = iclog->ic_next;
2188 } while (first_iclog != iclog);
2192 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2194 spin_unlock(&log->l_icloglock);
2197 sv_broadcast(&log->l_flush_wait);
2202 * Finish transitioning this iclog to the dirty state.
2204 * Make sure that we completely execute this routine only when this is
2205 * the last call to the iclog. There is a good chance that iclog flushes,
2206 * when we reach the end of the physical log, get turned into 2 separate
2207 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2208 * routine. By using the reference count bwritecnt, we guarantee that only
2209 * the second completion goes through.
2211 * Callbacks could take time, so they are done outside the scope of the
2212 * global state machine log lock.
2215 xlog_state_done_syncing(
2216 xlog_in_core_t *iclog,
2219 xlog_t *log = iclog->ic_log;
2221 spin_lock(&log->l_icloglock);
2223 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2224 iclog->ic_state == XLOG_STATE_IOERROR);
2225 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2226 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2230 * If we got an error, either on the first buffer, or in the case of
2231 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2232 * and none should ever be attempted to be written to disk
2235 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2236 if (--iclog->ic_bwritecnt == 1) {
2237 spin_unlock(&log->l_icloglock);
2240 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2244 * Someone could be sleeping prior to writing out the next
2245 * iclog buffer, we wake them all, one will get to do the
2246 * I/O, the others get to wait for the result.
2248 sv_broadcast(&iclog->ic_write_wait);
2249 spin_unlock(&log->l_icloglock);
2250 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2251 } /* xlog_state_done_syncing */
2255 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2256 * sleep. We wait on the flush queue on the head iclog as that should be
2257 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2258 * we will wait here and all new writes will sleep until a sync completes.
2260 * The in-core logs are used in a circular fashion. They are not used
2261 * out-of-order even when an iclog past the head is free.
2264 * * log_offset where xlog_write() can start writing into the in-core
2266 * * in-core log pointer to which xlog_write() should write.
2267 * * boolean indicating this is a continued write to an in-core log.
2268 * If this is the last write, then the in-core log's offset field
2269 * needs to be incremented, depending on the amount of data which
2273 xlog_state_get_iclog_space(xlog_t *log,
2275 xlog_in_core_t **iclogp,
2276 xlog_ticket_t *ticket,
2277 int *continued_write,
2281 xlog_rec_header_t *head;
2282 xlog_in_core_t *iclog;
2286 spin_lock(&log->l_icloglock);
2287 if (XLOG_FORCED_SHUTDOWN(log)) {
2288 spin_unlock(&log->l_icloglock);
2289 return XFS_ERROR(EIO);
2292 iclog = log->l_iclog;
2293 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2294 XFS_STATS_INC(xs_log_noiclogs);
2296 /* Wait for log writes to have flushed */
2297 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2301 head = &iclog->ic_header;
2303 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2304 log_offset = iclog->ic_offset;
2306 /* On the 1st write to an iclog, figure out lsn. This works
2307 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2308 * committing to. If the offset is set, that's how many blocks
2311 if (log_offset == 0) {
2312 ticket->t_curr_res -= log->l_iclog_hsize;
2313 xlog_tic_add_region(ticket,
2315 XLOG_REG_TYPE_LRHEADER);
2316 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2317 head->h_lsn = cpu_to_be64(
2318 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2319 ASSERT(log->l_curr_block >= 0);
2322 /* If there is enough room to write everything, then do it. Otherwise,
2323 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2324 * bit is on, so this will get flushed out. Don't update ic_offset
2325 * until you know exactly how many bytes get copied. Therefore, wait
2326 * until later to update ic_offset.
2328 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2329 * can fit into remaining data section.
2331 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2332 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2335 * If I'm the only one writing to this iclog, sync it to disk.
2336 * We need to do an atomic compare and decrement here to avoid
2337 * racing with concurrent atomic_dec_and_lock() calls in
2338 * xlog_state_release_iclog() when there is more than one
2339 * reference to the iclog.
2341 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2342 /* we are the only one */
2343 spin_unlock(&log->l_icloglock);
2344 error = xlog_state_release_iclog(log, iclog);
2348 spin_unlock(&log->l_icloglock);
2353 /* Do we have enough room to write the full amount in the remainder
2354 * of this iclog? Or must we continue a write on the next iclog and
2355 * mark this iclog as completely taken? In the case where we switch
2356 * iclogs (to mark it taken), this particular iclog will release/sync
2357 * to disk in xlog_write().
2359 if (len <= iclog->ic_size - iclog->ic_offset) {
2360 *continued_write = 0;
2361 iclog->ic_offset += len;
2363 *continued_write = 1;
2364 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2368 ASSERT(iclog->ic_offset <= iclog->ic_size);
2369 spin_unlock(&log->l_icloglock);
2371 *logoffsetp = log_offset;
2373 } /* xlog_state_get_iclog_space */
2376 * Atomically get the log space required for a log ticket.
2378 * Once a ticket gets put onto the reserveq, it will only return after
2379 * the needed reservation is satisfied.
2382 xlog_grant_log_space(xlog_t *log,
2393 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2394 panic("grant Recovery problem");
2397 /* Is there space or do we need to sleep? */
2398 spin_lock(&log->l_grant_lock);
2400 trace_xfs_log_grant_enter(log, tic);
2402 /* something is already sleeping; insert new transaction at end */
2403 if (log->l_reserve_headq) {
2404 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2406 trace_xfs_log_grant_sleep1(log, tic);
2409 * Gotta check this before going to sleep, while we're
2410 * holding the grant lock.
2412 if (XLOG_FORCED_SHUTDOWN(log))
2415 XFS_STATS_INC(xs_sleep_logspace);
2416 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2418 * If we got an error, and the filesystem is shutting down,
2419 * we'll catch it down below. So just continue...
2421 trace_xfs_log_grant_wake1(log, tic);
2422 spin_lock(&log->l_grant_lock);
2424 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2425 need_bytes = tic->t_unit_res*tic->t_ocnt;
2427 need_bytes = tic->t_unit_res;
2430 if (XLOG_FORCED_SHUTDOWN(log))
2433 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2434 log->l_grant_reserve_bytes);
2435 if (free_bytes < need_bytes) {
2436 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2437 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2439 trace_xfs_log_grant_sleep2(log, tic);
2441 spin_unlock(&log->l_grant_lock);
2442 xlog_grant_push_ail(log->l_mp, need_bytes);
2443 spin_lock(&log->l_grant_lock);
2445 XFS_STATS_INC(xs_sleep_logspace);
2446 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2448 spin_lock(&log->l_grant_lock);
2449 if (XLOG_FORCED_SHUTDOWN(log))
2452 trace_xfs_log_grant_wake2(log, tic);
2455 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2456 xlog_del_ticketq(&log->l_reserve_headq, tic);
2458 /* we've got enough space */
2459 xlog_grant_add_space(log, need_bytes);
2461 tail_lsn = log->l_tail_lsn;
2463 * Check to make sure the grant write head didn't just over lap the
2464 * tail. If the cycles are the same, we can't be overlapping.
2465 * Otherwise, make sure that the cycles differ by exactly one and
2466 * check the byte count.
2468 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2469 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2470 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2473 trace_xfs_log_grant_exit(log, tic);
2474 xlog_verify_grant_head(log, 1);
2475 spin_unlock(&log->l_grant_lock);
2479 if (tic->t_flags & XLOG_TIC_IN_Q)
2480 xlog_del_ticketq(&log->l_reserve_headq, tic);
2482 trace_xfs_log_grant_error(log, tic);
2485 * If we are failing, make sure the ticket doesn't have any
2486 * current reservations. We don't want to add this back when
2487 * the ticket/transaction gets cancelled.
2489 tic->t_curr_res = 0;
2490 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2491 spin_unlock(&log->l_grant_lock);
2492 return XFS_ERROR(EIO);
2493 } /* xlog_grant_log_space */
2497 * Replenish the byte reservation required by moving the grant write head.
2502 xlog_regrant_write_log_space(xlog_t *log,
2505 int free_bytes, need_bytes;
2506 xlog_ticket_t *ntic;
2511 tic->t_curr_res = tic->t_unit_res;
2512 xlog_tic_reset_res(tic);
2518 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2519 panic("regrant Recovery problem");
2522 spin_lock(&log->l_grant_lock);
2524 trace_xfs_log_regrant_write_enter(log, tic);
2526 if (XLOG_FORCED_SHUTDOWN(log))
2529 /* If there are other waiters on the queue then give them a
2530 * chance at logspace before us. Wake up the first waiters,
2531 * if we do not wake up all the waiters then go to sleep waiting
2532 * for more free space, otherwise try to get some space for
2535 need_bytes = tic->t_unit_res;
2536 if ((ntic = log->l_write_headq)) {
2537 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2538 log->l_grant_write_bytes);
2540 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2542 if (free_bytes < ntic->t_unit_res)
2544 free_bytes -= ntic->t_unit_res;
2545 sv_signal(&ntic->t_wait);
2546 ntic = ntic->t_next;
2547 } while (ntic != log->l_write_headq);
2549 if (ntic != log->l_write_headq) {
2550 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2551 xlog_ins_ticketq(&log->l_write_headq, tic);
2553 trace_xfs_log_regrant_write_sleep1(log, tic);
2555 spin_unlock(&log->l_grant_lock);
2556 xlog_grant_push_ail(log->l_mp, need_bytes);
2557 spin_lock(&log->l_grant_lock);
2559 XFS_STATS_INC(xs_sleep_logspace);
2560 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2561 &log->l_grant_lock, s);
2563 /* If we're shutting down, this tic is already
2565 spin_lock(&log->l_grant_lock);
2566 if (XLOG_FORCED_SHUTDOWN(log))
2569 trace_xfs_log_regrant_write_wake1(log, tic);
2574 if (XLOG_FORCED_SHUTDOWN(log))
2577 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2578 log->l_grant_write_bytes);
2579 if (free_bytes < need_bytes) {
2580 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2581 xlog_ins_ticketq(&log->l_write_headq, tic);
2582 spin_unlock(&log->l_grant_lock);
2583 xlog_grant_push_ail(log->l_mp, need_bytes);
2584 spin_lock(&log->l_grant_lock);
2586 XFS_STATS_INC(xs_sleep_logspace);
2587 trace_xfs_log_regrant_write_sleep2(log, tic);
2589 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2591 /* If we're shutting down, this tic is already off the queue */
2592 spin_lock(&log->l_grant_lock);
2593 if (XLOG_FORCED_SHUTDOWN(log))
2596 trace_xfs_log_regrant_write_wake2(log, tic);
2598 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2599 xlog_del_ticketq(&log->l_write_headq, tic);
2601 /* we've got enough space */
2602 xlog_grant_add_space_write(log, need_bytes);
2604 tail_lsn = log->l_tail_lsn;
2605 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2606 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2607 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2611 trace_xfs_log_regrant_write_exit(log, tic);
2613 xlog_verify_grant_head(log, 1);
2614 spin_unlock(&log->l_grant_lock);
2619 if (tic->t_flags & XLOG_TIC_IN_Q)
2620 xlog_del_ticketq(&log->l_reserve_headq, tic);
2622 trace_xfs_log_regrant_write_error(log, tic);
2625 * If we are failing, make sure the ticket doesn't have any
2626 * current reservations. We don't want to add this back when
2627 * the ticket/transaction gets cancelled.
2629 tic->t_curr_res = 0;
2630 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2631 spin_unlock(&log->l_grant_lock);
2632 return XFS_ERROR(EIO);
2633 } /* xlog_regrant_write_log_space */
2636 /* The first cnt-1 times through here we don't need to
2637 * move the grant write head because the permanent
2638 * reservation has reserved cnt times the unit amount.
2639 * Release part of current permanent unit reservation and
2640 * reset current reservation to be one units worth. Also
2641 * move grant reservation head forward.
2644 xlog_regrant_reserve_log_space(xlog_t *log,
2645 xlog_ticket_t *ticket)
2647 trace_xfs_log_regrant_reserve_enter(log, ticket);
2649 if (ticket->t_cnt > 0)
2652 spin_lock(&log->l_grant_lock);
2653 xlog_grant_sub_space(log, ticket->t_curr_res);
2654 ticket->t_curr_res = ticket->t_unit_res;
2655 xlog_tic_reset_res(ticket);
2657 trace_xfs_log_regrant_reserve_sub(log, ticket);
2659 xlog_verify_grant_head(log, 1);
2661 /* just return if we still have some of the pre-reserved space */
2662 if (ticket->t_cnt > 0) {
2663 spin_unlock(&log->l_grant_lock);
2667 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2669 trace_xfs_log_regrant_reserve_exit(log, ticket);
2671 xlog_verify_grant_head(log, 0);
2672 spin_unlock(&log->l_grant_lock);
2673 ticket->t_curr_res = ticket->t_unit_res;
2674 xlog_tic_reset_res(ticket);
2675 } /* xlog_regrant_reserve_log_space */
2679 * Give back the space left from a reservation.
2681 * All the information we need to make a correct determination of space left
2682 * is present. For non-permanent reservations, things are quite easy. The
2683 * count should have been decremented to zero. We only need to deal with the
2684 * space remaining in the current reservation part of the ticket. If the
2685 * ticket contains a permanent reservation, there may be left over space which
2686 * needs to be released. A count of N means that N-1 refills of the current
2687 * reservation can be done before we need to ask for more space. The first
2688 * one goes to fill up the first current reservation. Once we run out of
2689 * space, the count will stay at zero and the only space remaining will be
2690 * in the current reservation field.
2693 xlog_ungrant_log_space(xlog_t *log,
2694 xlog_ticket_t *ticket)
2696 if (ticket->t_cnt > 0)
2699 spin_lock(&log->l_grant_lock);
2700 trace_xfs_log_ungrant_enter(log, ticket);
2702 xlog_grant_sub_space(log, ticket->t_curr_res);
2704 trace_xfs_log_ungrant_sub(log, ticket);
2706 /* If this is a permanent reservation ticket, we may be able to free
2707 * up more space based on the remaining count.
2709 if (ticket->t_cnt > 0) {
2710 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2711 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2714 trace_xfs_log_ungrant_exit(log, ticket);
2716 xlog_verify_grant_head(log, 1);
2717 spin_unlock(&log->l_grant_lock);
2718 xfs_log_move_tail(log->l_mp, 1);
2719 } /* xlog_ungrant_log_space */
2723 * Flush iclog to disk if this is the last reference to the given iclog and
2724 * the WANT_SYNC bit is set.
2726 * When this function is entered, the iclog is not necessarily in the
2727 * WANT_SYNC state. It may be sitting around waiting to get filled.
2732 xlog_state_release_iclog(
2734 xlog_in_core_t *iclog)
2736 int sync = 0; /* do we sync? */
2738 if (iclog->ic_state & XLOG_STATE_IOERROR)
2739 return XFS_ERROR(EIO);
2741 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2742 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2745 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2746 spin_unlock(&log->l_icloglock);
2747 return XFS_ERROR(EIO);
2749 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2750 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2752 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2753 /* update tail before writing to iclog */
2754 xlog_assign_tail_lsn(log->l_mp);
2756 iclog->ic_state = XLOG_STATE_SYNCING;
2757 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2758 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2759 /* cycle incremented when incrementing curr_block */
2761 spin_unlock(&log->l_icloglock);
2764 * We let the log lock go, so it's possible that we hit a log I/O
2765 * error or some other SHUTDOWN condition that marks the iclog
2766 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2767 * this iclog has consistent data, so we ignore IOERROR
2768 * flags after this point.
2771 return xlog_sync(log, iclog);
2773 } /* xlog_state_release_iclog */
2777 * This routine will mark the current iclog in the ring as WANT_SYNC
2778 * and move the current iclog pointer to the next iclog in the ring.
2779 * When this routine is called from xlog_state_get_iclog_space(), the
2780 * exact size of the iclog has not yet been determined. All we know is
2781 * that every data block. We have run out of space in this log record.
2784 xlog_state_switch_iclogs(xlog_t *log,
2785 xlog_in_core_t *iclog,
2788 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2790 eventual_size = iclog->ic_offset;
2791 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2792 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2793 log->l_prev_block = log->l_curr_block;
2794 log->l_prev_cycle = log->l_curr_cycle;
2796 /* roll log?: ic_offset changed later */
2797 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2799 /* Round up to next log-sunit */
2800 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2801 log->l_mp->m_sb.sb_logsunit > 1) {
2802 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2803 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2806 if (log->l_curr_block >= log->l_logBBsize) {
2807 log->l_curr_cycle++;
2808 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2809 log->l_curr_cycle++;
2810 log->l_curr_block -= log->l_logBBsize;
2811 ASSERT(log->l_curr_block >= 0);
2813 ASSERT(iclog == log->l_iclog);
2814 log->l_iclog = iclog->ic_next;
2815 } /* xlog_state_switch_iclogs */
2818 * Write out all data in the in-core log as of this exact moment in time.
2820 * Data may be written to the in-core log during this call. However,
2821 * we don't guarantee this data will be written out. A change from past
2822 * implementation means this routine will *not* write out zero length LRs.
2824 * Basically, we try and perform an intelligent scan of the in-core logs.
2825 * If we determine there is no flushable data, we just return. There is no
2826 * flushable data if:
2828 * 1. the current iclog is active and has no data; the previous iclog
2829 * is in the active or dirty state.
2830 * 2. the current iclog is drity, and the previous iclog is in the
2831 * active or dirty state.
2835 * 1. the current iclog is not in the active nor dirty state.
2836 * 2. the current iclog dirty, and the previous iclog is not in the
2837 * active nor dirty state.
2838 * 3. the current iclog is active, and there is another thread writing
2839 * to this particular iclog.
2840 * 4. a) the current iclog is active and has no other writers
2841 * b) when we return from flushing out this iclog, it is still
2842 * not in the active nor dirty state.
2846 struct xfs_mount *mp,
2850 struct log *log = mp->m_log;
2851 struct xlog_in_core *iclog;
2854 XFS_STATS_INC(xs_log_force);
2856 spin_lock(&log->l_icloglock);
2858 iclog = log->l_iclog;
2859 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2860 spin_unlock(&log->l_icloglock);
2861 return XFS_ERROR(EIO);
2864 /* If the head iclog is not active nor dirty, we just attach
2865 * ourselves to the head and go to sleep.
2867 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2868 iclog->ic_state == XLOG_STATE_DIRTY) {
2870 * If the head is dirty or (active and empty), then
2871 * we need to look at the previous iclog. If the previous
2872 * iclog is active or dirty we are done. There is nothing
2873 * to sync out. Otherwise, we attach ourselves to the
2874 * previous iclog and go to sleep.
2876 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2877 (atomic_read(&iclog->ic_refcnt) == 0
2878 && iclog->ic_offset == 0)) {
2879 iclog = iclog->ic_prev;
2880 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2881 iclog->ic_state == XLOG_STATE_DIRTY)
2886 if (atomic_read(&iclog->ic_refcnt) == 0) {
2887 /* We are the only one with access to this
2888 * iclog. Flush it out now. There should
2889 * be a roundoff of zero to show that someone
2890 * has already taken care of the roundoff from
2891 * the previous sync.
2893 atomic_inc(&iclog->ic_refcnt);
2894 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2895 xlog_state_switch_iclogs(log, iclog, 0);
2896 spin_unlock(&log->l_icloglock);
2898 if (xlog_state_release_iclog(log, iclog))
2899 return XFS_ERROR(EIO);
2903 spin_lock(&log->l_icloglock);
2904 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2905 iclog->ic_state != XLOG_STATE_DIRTY)
2910 /* Someone else is writing to this iclog.
2911 * Use its call to flush out the data. However,
2912 * the other thread may not force out this LR,
2913 * so we mark it WANT_SYNC.
2915 xlog_state_switch_iclogs(log, iclog, 0);
2921 /* By the time we come around again, the iclog could've been filled
2922 * which would give it another lsn. If we have a new lsn, just
2923 * return because the relevant data has been flushed.
2926 if (flags & XFS_LOG_SYNC) {
2928 * We must check if we're shutting down here, before
2929 * we wait, while we're holding the l_icloglock.
2930 * Then we check again after waking up, in case our
2931 * sleep was disturbed by a bad news.
2933 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2934 spin_unlock(&log->l_icloglock);
2935 return XFS_ERROR(EIO);
2937 XFS_STATS_INC(xs_log_force_sleep);
2938 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
2940 * No need to grab the log lock here since we're
2941 * only deciding whether or not to return EIO
2942 * and the memory read should be atomic.
2944 if (iclog->ic_state & XLOG_STATE_IOERROR)
2945 return XFS_ERROR(EIO);
2951 spin_unlock(&log->l_icloglock);
2957 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2958 * about errors or whether the log was flushed or not. This is the normal
2959 * interface to use when trying to unpin items or move the log forward.
2968 error = _xfs_log_force(mp, flags, NULL);
2970 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
2971 "error %d returned.", error);
2976 * Force the in-core log to disk for a specific LSN.
2978 * Find in-core log with lsn.
2979 * If it is in the DIRTY state, just return.
2980 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2981 * state and go to sleep or return.
2982 * If it is in any other state, go to sleep or return.
2984 * Synchronous forces are implemented with a signal variable. All callers
2985 * to force a given lsn to disk will wait on a the sv attached to the
2986 * specific in-core log. When given in-core log finally completes its
2987 * write to disk, that thread will wake up all threads waiting on the
2992 struct xfs_mount *mp,
2997 struct log *log = mp->m_log;
2998 struct xlog_in_core *iclog;
2999 int already_slept = 0;
3003 XFS_STATS_INC(xs_log_force);
3006 spin_lock(&log->l_icloglock);
3007 iclog = log->l_iclog;
3008 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3009 spin_unlock(&log->l_icloglock);
3010 return XFS_ERROR(EIO);
3014 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3015 iclog = iclog->ic_next;
3019 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3020 spin_unlock(&log->l_icloglock);
3024 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3026 * We sleep here if we haven't already slept (e.g.
3027 * this is the first time we've looked at the correct
3028 * iclog buf) and the buffer before us is going to
3029 * be sync'ed. The reason for this is that if we
3030 * are doing sync transactions here, by waiting for
3031 * the previous I/O to complete, we can allow a few
3032 * more transactions into this iclog before we close
3035 * Otherwise, we mark the buffer WANT_SYNC, and bump
3036 * up the refcnt so we can release the log (which
3037 * drops the ref count). The state switch keeps new
3038 * transaction commits from using this buffer. When
3039 * the current commits finish writing into the buffer,
3040 * the refcount will drop to zero and the buffer will
3043 if (!already_slept &&
3044 (iclog->ic_prev->ic_state &
3045 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3046 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3048 XFS_STATS_INC(xs_log_force_sleep);
3050 sv_wait(&iclog->ic_prev->ic_write_wait,
3051 PSWP, &log->l_icloglock, s);
3057 atomic_inc(&iclog->ic_refcnt);
3058 xlog_state_switch_iclogs(log, iclog, 0);
3059 spin_unlock(&log->l_icloglock);
3060 if (xlog_state_release_iclog(log, iclog))
3061 return XFS_ERROR(EIO);
3064 spin_lock(&log->l_icloglock);
3067 if ((flags & XFS_LOG_SYNC) && /* sleep */
3069 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3071 * Don't wait on completion if we know that we've
3072 * gotten a log write error.
3074 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3075 spin_unlock(&log->l_icloglock);
3076 return XFS_ERROR(EIO);
3078 XFS_STATS_INC(xs_log_force_sleep);
3079 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3081 * No need to grab the log lock here since we're
3082 * only deciding whether or not to return EIO
3083 * and the memory read should be atomic.
3085 if (iclog->ic_state & XLOG_STATE_IOERROR)
3086 return XFS_ERROR(EIO);
3090 } else { /* just return */
3091 spin_unlock(&log->l_icloglock);
3095 } while (iclog != log->l_iclog);
3097 spin_unlock(&log->l_icloglock);
3102 * Wrapper for _xfs_log_force_lsn(), 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.
3114 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3116 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3117 "error %d returned.", error);
3122 * Called when we want to mark the current iclog as being ready to sync to
3126 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3128 assert_spin_locked(&log->l_icloglock);
3130 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3131 xlog_state_switch_iclogs(log, iclog, 0);
3133 ASSERT(iclog->ic_state &
3134 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3139 /*****************************************************************************
3143 *****************************************************************************
3147 * Free a used ticket when its refcount falls to zero.
3151 xlog_ticket_t *ticket)
3153 ASSERT(atomic_read(&ticket->t_ref) > 0);
3154 if (atomic_dec_and_test(&ticket->t_ref)) {
3155 sv_destroy(&ticket->t_wait);
3156 kmem_zone_free(xfs_log_ticket_zone, ticket);
3162 xlog_ticket_t *ticket)
3164 ASSERT(atomic_read(&ticket->t_ref) > 0);
3165 atomic_inc(&ticket->t_ref);
3170 * Allocate and initialise a new log ticket.
3172 STATIC xlog_ticket_t *
3173 xlog_ticket_alloc(xlog_t *log,
3182 tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3187 * Permanent reservations have up to 'cnt'-1 active log operations
3188 * in the log. A unit in this case is the amount of space for one
3189 * of these log operations. Normal reservations have a cnt of 1
3190 * and their unit amount is the total amount of space required.
3192 * The following lines of code account for non-transaction data
3193 * which occupy space in the on-disk log.
3195 * Normal form of a transaction is:
3196 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3197 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3199 * We need to account for all the leadup data and trailer data
3200 * around the transaction data.
3201 * And then we need to account for the worst case in terms of using
3203 * The worst case will happen if:
3204 * - the placement of the transaction happens to be such that the
3205 * roundoff is at its maximum
3206 * - the transaction data is synced before the commit record is synced
3207 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3208 * Therefore the commit record is in its own Log Record.
3209 * This can happen as the commit record is called with its
3210 * own region to xlog_write().
3211 * This then means that in the worst case, roundoff can happen for
3212 * the commit-rec as well.
3213 * The commit-rec is smaller than padding in this scenario and so it is
3214 * not added separately.
3217 /* for trans header */
3218 unit_bytes += sizeof(xlog_op_header_t);
3219 unit_bytes += sizeof(xfs_trans_header_t);
3222 unit_bytes += sizeof(xlog_op_header_t);
3224 /* for LR headers */
3225 num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3226 unit_bytes += log->l_iclog_hsize * num_headers;
3228 /* for commit-rec LR header - note: padding will subsume the ophdr */
3229 unit_bytes += log->l_iclog_hsize;
3231 /* for split-recs - ophdrs added when data split over LRs */
3232 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3234 /* for roundoff padding for transaction data and one for commit record */
3235 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3236 log->l_mp->m_sb.sb_logsunit > 1) {
3237 /* log su roundoff */
3238 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3241 unit_bytes += 2*BBSIZE;
3244 atomic_set(&tic->t_ref, 1);
3245 tic->t_unit_res = unit_bytes;
3246 tic->t_curr_res = unit_bytes;
3249 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3250 tic->t_clientid = client;
3251 tic->t_flags = XLOG_TIC_INITED;
3252 tic->t_trans_type = 0;
3253 if (xflags & XFS_LOG_PERM_RESERV)
3254 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3255 sv_init(&(tic->t_wait), SV_DEFAULT, "logtick");
3257 xlog_tic_reset_res(tic);
3263 /******************************************************************************
3265 * Log debug routines
3267 ******************************************************************************
3271 * Make sure that the destination ptr is within the valid data region of
3272 * one of the iclogs. This uses backup pointers stored in a different
3273 * part of the log in case we trash the log structure.
3276 xlog_verify_dest_ptr(xlog_t *log,
3282 for (i=0; i < log->l_iclog_bufs; i++) {
3283 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3284 ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3288 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3289 } /* xlog_verify_dest_ptr */
3292 xlog_verify_grant_head(xlog_t *log, int equals)
3294 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3296 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3298 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3300 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3301 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3303 } /* xlog_verify_grant_head */
3305 /* check if it will fit */
3307 xlog_verify_tail_lsn(xlog_t *log,
3308 xlog_in_core_t *iclog,
3313 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3315 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3316 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3317 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3319 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3321 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3322 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3324 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3325 if (blocks < BTOBB(iclog->ic_offset) + 1)
3326 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3328 } /* xlog_verify_tail_lsn */
3331 * Perform a number of checks on the iclog before writing to disk.
3333 * 1. Make sure the iclogs are still circular
3334 * 2. Make sure we have a good magic number
3335 * 3. Make sure we don't have magic numbers in the data
3336 * 4. Check fields of each log operation header for:
3337 * A. Valid client identifier
3338 * B. tid ptr value falls in valid ptr space (user space code)
3339 * C. Length in log record header is correct according to the
3340 * individual operation headers within record.
3341 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3342 * log, check the preceding blocks of the physical log to make sure all
3343 * the cycle numbers agree with the current cycle number.
3346 xlog_verify_iclog(xlog_t *log,
3347 xlog_in_core_t *iclog,
3351 xlog_op_header_t *ophead;
3352 xlog_in_core_t *icptr;
3353 xlog_in_core_2_t *xhdr;
3355 xfs_caddr_t base_ptr;
3356 __psint_t field_offset;
3358 int len, i, j, k, op_len;
3361 /* check validity of iclog pointers */
3362 spin_lock(&log->l_icloglock);
3363 icptr = log->l_iclog;
3364 for (i=0; i < log->l_iclog_bufs; i++) {
3366 xlog_panic("xlog_verify_iclog: invalid ptr");
3367 icptr = icptr->ic_next;
3369 if (icptr != log->l_iclog)
3370 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3371 spin_unlock(&log->l_icloglock);
3373 /* check log magic numbers */
3374 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3375 xlog_panic("xlog_verify_iclog: invalid magic num");
3377 ptr = (xfs_caddr_t) &iclog->ic_header;
3378 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3380 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3381 xlog_panic("xlog_verify_iclog: unexpected magic num");
3385 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3386 ptr = iclog->ic_datap;
3388 ophead = (xlog_op_header_t *)ptr;
3389 xhdr = iclog->ic_data;
3390 for (i = 0; i < len; i++) {
3391 ophead = (xlog_op_header_t *)ptr;
3393 /* clientid is only 1 byte */
3394 field_offset = (__psint_t)
3395 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3396 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3397 clientid = ophead->oh_clientid;
3399 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3400 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3401 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3402 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3403 clientid = xlog_get_client_id(
3404 xhdr[j].hic_xheader.xh_cycle_data[k]);
3406 clientid = xlog_get_client_id(
3407 iclog->ic_header.h_cycle_data[idx]);
3410 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3411 cmn_err(CE_WARN, "xlog_verify_iclog: "
3412 "invalid clientid %d op 0x%p offset 0x%lx",
3413 clientid, ophead, (unsigned long)field_offset);
3416 field_offset = (__psint_t)
3417 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3418 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3419 op_len = be32_to_cpu(ophead->oh_len);
3421 idx = BTOBBT((__psint_t)&ophead->oh_len -
3422 (__psint_t)iclog->ic_datap);
3423 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3424 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3425 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3426 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3428 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3431 ptr += sizeof(xlog_op_header_t) + op_len;
3433 } /* xlog_verify_iclog */
3437 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3443 xlog_in_core_t *iclog, *ic;
3445 iclog = log->l_iclog;
3446 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3448 * Mark all the incore logs IOERROR.
3449 * From now on, no log flushes will result.
3453 ic->ic_state = XLOG_STATE_IOERROR;
3455 } while (ic != iclog);
3459 * Return non-zero, if state transition has already happened.
3465 * This is called from xfs_force_shutdown, when we're forcibly
3466 * shutting down the filesystem, typically because of an IO error.
3467 * Our main objectives here are to make sure that:
3468 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3469 * parties to find out, 'atomically'.
3470 * b. those who're sleeping on log reservations, pinned objects and
3471 * other resources get woken up, and be told the bad news.
3472 * c. nothing new gets queued up after (a) and (b) are done.
3473 * d. if !logerror, flush the iclogs to disk, then seal them off
3477 xfs_log_force_umount(
3478 struct xfs_mount *mp,
3488 * If this happens during log recovery, don't worry about
3489 * locking; the log isn't open for business yet.
3492 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3493 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3495 XFS_BUF_DONE(mp->m_sb_bp);
3500 * Somebody could've already done the hard work for us.
3501 * No need to get locks for this.
3503 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3504 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3509 * We must hold both the GRANT lock and the LOG lock,
3510 * before we mark the filesystem SHUTDOWN and wake
3511 * everybody up to tell the bad news.
3513 spin_lock(&log->l_icloglock);
3514 spin_lock(&log->l_grant_lock);
3515 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3517 XFS_BUF_DONE(mp->m_sb_bp);
3520 * This flag is sort of redundant because of the mount flag, but
3521 * it's good to maintain the separation between the log and the rest
3524 log->l_flags |= XLOG_IO_ERROR;
3527 * If we hit a log error, we want to mark all the iclogs IOERROR
3528 * while we're still holding the loglock.
3531 retval = xlog_state_ioerror(log);
3532 spin_unlock(&log->l_icloglock);
3535 * We don't want anybody waiting for log reservations
3536 * after this. That means we have to wake up everybody
3537 * queued up on reserve_headq as well as write_headq.
3538 * In addition, we make sure in xlog_{re}grant_log_space
3539 * that we don't enqueue anything once the SHUTDOWN flag
3540 * is set, and this action is protected by the GRANTLOCK.
3542 if ((tic = log->l_reserve_headq)) {
3544 sv_signal(&tic->t_wait);
3546 } while (tic != log->l_reserve_headq);
3549 if ((tic = log->l_write_headq)) {
3551 sv_signal(&tic->t_wait);
3553 } while (tic != log->l_write_headq);
3555 spin_unlock(&log->l_grant_lock);
3557 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3560 * Force the incore logs to disk before shutting the
3561 * log down completely.
3563 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3565 spin_lock(&log->l_icloglock);
3566 retval = xlog_state_ioerror(log);
3567 spin_unlock(&log->l_icloglock);
3570 * Wake up everybody waiting on xfs_log_force.
3571 * Callback all log item committed functions as if the
3572 * log writes were completed.
3574 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3576 #ifdef XFSERRORDEBUG
3578 xlog_in_core_t *iclog;
3580 spin_lock(&log->l_icloglock);
3581 iclog = log->l_iclog;
3583 ASSERT(iclog->ic_callback == 0);
3584 iclog = iclog->ic_next;
3585 } while (iclog != log->l_iclog);
3586 spin_unlock(&log->l_icloglock);
3589 /* return non-zero if log IOERROR transition had already happened */
3594 xlog_iclogs_empty(xlog_t *log)
3596 xlog_in_core_t *iclog;
3598 iclog = log->l_iclog;
3600 /* endianness does not matter here, zero is zero in
3603 if (iclog->ic_header.h_num_logops)
3605 iclog = iclog->ic_next;
3606 } while (iclog != log->l_iclog);