2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_buf_item.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_trans_priv.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
39 #include "xfs_trace.h"
41 kmem_zone_t *xfs_log_ticket_zone;
43 /* Local miscellaneous function prototypes */
44 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
50 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
51 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52 STATIC void xlog_dealloc_log(xlog_t *log);
54 /* local state machine functions */
55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57 STATIC int xlog_state_get_iclog_space(xlog_t *log,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
63 STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65 STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
70 /* local functions to manipulate grant head */
71 STATIC int xlog_grant_log_space(xlog_t *log,
73 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
75 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
76 xlog_ticket_t *ticket);
77 STATIC int xlog_regrant_write_log_space(xlog_t *log,
78 xlog_ticket_t *ticket);
79 STATIC void xlog_ungrant_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
83 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
85 STATIC void xlog_verify_grant_tail(struct log *log);
86 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
87 int count, boolean_t syncing);
88 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
91 #define xlog_verify_dest_ptr(a,b)
92 #define xlog_verify_grant_head(a,b)
93 #define xlog_verify_grant_tail(a)
94 #define xlog_verify_iclog(a,b,c,d)
95 #define xlog_verify_tail_lsn(a,b,c)
98 STATIC int xlog_iclogs_empty(xlog_t *log);
101 xlog_grant_sub_space(
109 *space += log->l_logsize;
115 xlog_grant_add_space(
121 int tmp = log->l_logsize - *space;
125 *space = bytes - tmp;
130 xlog_tic_reset_res(xlog_ticket_t *tic)
133 tic->t_res_arr_sum = 0;
134 tic->t_res_num_ophdrs = 0;
138 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
140 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
141 /* add to overflow and start again */
142 tic->t_res_o_flow += tic->t_res_arr_sum;
144 tic->t_res_arr_sum = 0;
147 tic->t_res_arr[tic->t_res_num].r_len = len;
148 tic->t_res_arr[tic->t_res_num].r_type = type;
149 tic->t_res_arr_sum += len;
156 * 1. currblock field gets updated at startup and after in-core logs
157 * marked as with WANT_SYNC.
161 * This routine is called when a user of a log manager ticket is done with
162 * the reservation. If the ticket was ever used, then a commit record for
163 * the associated transaction is written out as a log operation header with
164 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
165 * a given ticket. If the ticket was one with a permanent reservation, then
166 * a few operations are done differently. Permanent reservation tickets by
167 * default don't release the reservation. They just commit the current
168 * transaction with the belief that the reservation is still needed. A flag
169 * must be passed in before permanent reservations are actually released.
170 * When these type of tickets are not released, they need to be set into
171 * the inited state again. By doing this, a start record will be written
172 * out when the next write occurs.
176 struct xfs_mount *mp,
177 struct xlog_ticket *ticket,
178 struct xlog_in_core **iclog,
181 struct log *log = mp->m_log;
184 if (XLOG_FORCED_SHUTDOWN(log) ||
186 * If nothing was ever written, don't write out commit record.
187 * If we get an error, just continue and give back the log ticket.
189 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
190 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
191 lsn = (xfs_lsn_t) -1;
192 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
193 flags |= XFS_LOG_REL_PERM_RESERV;
198 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
199 (flags & XFS_LOG_REL_PERM_RESERV)) {
200 trace_xfs_log_done_nonperm(log, ticket);
203 * Release ticket if not permanent reservation or a specific
204 * request has been made to release a permanent reservation.
206 xlog_ungrant_log_space(log, ticket);
207 xfs_log_ticket_put(ticket);
209 trace_xfs_log_done_perm(log, ticket);
211 xlog_regrant_reserve_log_space(log, ticket);
212 /* If this ticket was a permanent reservation and we aren't
213 * trying to release it, reset the inited flags; so next time
214 * we write, a start record will be written out.
216 ticket->t_flags |= XLOG_TIC_INITED;
223 * Attaches a new iclog I/O completion callback routine during
224 * transaction commit. If the log is in error state, a non-zero
225 * return code is handed back and the caller is responsible for
226 * executing the callback at an appropriate time.
230 struct xfs_mount *mp,
231 struct xlog_in_core *iclog,
232 xfs_log_callback_t *cb)
236 spin_lock(&iclog->ic_callback_lock);
237 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
239 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
240 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
242 *(iclog->ic_callback_tail) = cb;
243 iclog->ic_callback_tail = &(cb->cb_next);
245 spin_unlock(&iclog->ic_callback_lock);
250 xfs_log_release_iclog(
251 struct xfs_mount *mp,
252 struct xlog_in_core *iclog)
254 if (xlog_state_release_iclog(mp->m_log, iclog)) {
255 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
263 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
264 * to the reservation.
265 * 2. Potentially, push buffers at tail of log to disk.
267 * Each reservation is going to reserve extra space for a log record header.
268 * When writes happen to the on-disk log, we don't subtract the length of the
269 * log record header from any reservation. By wasting space in each
270 * reservation, we prevent over allocation problems.
274 struct xfs_mount *mp,
277 struct xlog_ticket **ticket,
282 struct log *log = mp->m_log;
283 struct xlog_ticket *internal_ticket;
286 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
288 if (XLOG_FORCED_SHUTDOWN(log))
289 return XFS_ERROR(EIO);
291 XFS_STATS_INC(xs_try_logspace);
294 if (*ticket != NULL) {
295 ASSERT(flags & XFS_LOG_PERM_RESERV);
296 internal_ticket = *ticket;
299 * this is a new transaction on the ticket, so we need to
300 * change the transaction ID so that the next transaction has a
301 * different TID in the log. Just add one to the existing tid
302 * so that we can see chains of rolling transactions in the log
305 internal_ticket->t_tid++;
307 trace_xfs_log_reserve(log, internal_ticket);
309 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
310 retval = xlog_regrant_write_log_space(log, internal_ticket);
312 /* may sleep if need to allocate more tickets */
313 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
315 KM_SLEEP|KM_MAYFAIL);
316 if (!internal_ticket)
317 return XFS_ERROR(ENOMEM);
318 internal_ticket->t_trans_type = t_type;
319 *ticket = internal_ticket;
321 trace_xfs_log_reserve(log, internal_ticket);
323 xlog_grant_push_ail(mp,
324 (internal_ticket->t_unit_res *
325 internal_ticket->t_cnt));
326 retval = xlog_grant_log_space(log, internal_ticket);
330 } /* xfs_log_reserve */
334 * Mount a log filesystem
336 * mp - ubiquitous xfs mount point structure
337 * log_target - buftarg of on-disk log device
338 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
339 * num_bblocks - Number of BBSIZE blocks in on-disk log
341 * Return error or zero.
346 xfs_buftarg_t *log_target,
347 xfs_daddr_t blk_offset,
352 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
353 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
356 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
358 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
361 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
362 if (IS_ERR(mp->m_log)) {
363 error = -PTR_ERR(mp->m_log);
368 * Initialize the AIL now we have a log.
370 error = xfs_trans_ail_init(mp);
372 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
375 mp->m_log->l_ailp = mp->m_ail;
378 * skip log recovery on a norecovery mount. pretend it all
381 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
382 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
385 mp->m_flags &= ~XFS_MOUNT_RDONLY;
387 error = xlog_recover(mp->m_log);
390 mp->m_flags |= XFS_MOUNT_RDONLY;
392 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
393 goto out_destroy_ail;
397 /* Normal transactions can now occur */
398 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
401 * Now the log has been fully initialised and we know were our
402 * space grant counters are, we can initialise the permanent ticket
403 * needed for delayed logging to work.
405 xlog_cil_init_post_recovery(mp->m_log);
410 xfs_trans_ail_destroy(mp);
412 xlog_dealloc_log(mp->m_log);
418 * Finish the recovery of the file system. This is separate from
419 * the xfs_log_mount() call, because it depends on the code in
420 * xfs_mountfs() to read in the root and real-time bitmap inodes
421 * between calling xfs_log_mount() and here.
423 * mp - ubiquitous xfs mount point structure
426 xfs_log_mount_finish(xfs_mount_t *mp)
430 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
431 error = xlog_recover_finish(mp->m_log);
434 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
441 * Final log writes as part of unmount.
443 * Mark the filesystem clean as unmount happens. Note that during relocation
444 * this routine needs to be executed as part of source-bag while the
445 * deallocation must not be done until source-end.
449 * Unmount record used to have a string "Unmount filesystem--" in the
450 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
451 * We just write the magic number now since that particular field isn't
452 * currently architecture converted and "nUmount" is a bit foo.
453 * As far as I know, there weren't any dependencies on the old behaviour.
457 xfs_log_unmount_write(xfs_mount_t *mp)
459 xlog_t *log = mp->m_log;
460 xlog_in_core_t *iclog;
462 xlog_in_core_t *first_iclog;
464 xlog_ticket_t *tic = NULL;
469 * Don't write out unmount record on read-only mounts.
470 * Or, if we are doing a forced umount (typically because of IO errors).
472 if (mp->m_flags & XFS_MOUNT_RDONLY)
475 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
476 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
479 first_iclog = iclog = log->l_iclog;
481 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
482 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
483 ASSERT(iclog->ic_offset == 0);
485 iclog = iclog->ic_next;
486 } while (iclog != first_iclog);
488 if (! (XLOG_FORCED_SHUTDOWN(log))) {
489 error = xfs_log_reserve(mp, 600, 1, &tic,
490 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
492 /* the data section must be 32 bit size aligned */
496 __uint32_t pad2; /* may as well make it 64 bits */
498 .magic = XLOG_UNMOUNT_TYPE,
500 struct xfs_log_iovec reg = {
502 .i_len = sizeof(magic),
503 .i_type = XLOG_REG_TYPE_UNMOUNT,
505 struct xfs_log_vec vec = {
510 /* remove inited flag */
512 error = xlog_write(log, &vec, tic, &lsn,
513 NULL, XLOG_UNMOUNT_TRANS);
515 * At this point, we're umounting anyway,
516 * so there's no point in transitioning log state
517 * to IOERROR. Just continue...
522 xfs_fs_cmn_err(CE_ALERT, mp,
523 "xfs_log_unmount: unmount record failed");
527 spin_lock(&log->l_icloglock);
528 iclog = log->l_iclog;
529 atomic_inc(&iclog->ic_refcnt);
530 xlog_state_want_sync(log, iclog);
531 spin_unlock(&log->l_icloglock);
532 error = xlog_state_release_iclog(log, iclog);
534 spin_lock(&log->l_icloglock);
535 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
536 iclog->ic_state == XLOG_STATE_DIRTY)) {
537 if (!XLOG_FORCED_SHUTDOWN(log)) {
538 sv_wait(&iclog->ic_force_wait, PMEM,
539 &log->l_icloglock, s);
541 spin_unlock(&log->l_icloglock);
544 spin_unlock(&log->l_icloglock);
547 trace_xfs_log_umount_write(log, tic);
548 xlog_ungrant_log_space(log, tic);
549 xfs_log_ticket_put(tic);
553 * We're already in forced_shutdown mode, couldn't
554 * even attempt to write out the unmount transaction.
556 * Go through the motions of sync'ing and releasing
557 * the iclog, even though no I/O will actually happen,
558 * we need to wait for other log I/Os that may already
559 * be in progress. Do this as a separate section of
560 * code so we'll know if we ever get stuck here that
561 * we're in this odd situation of trying to unmount
562 * a file system that went into forced_shutdown as
563 * the result of an unmount..
565 spin_lock(&log->l_icloglock);
566 iclog = log->l_iclog;
567 atomic_inc(&iclog->ic_refcnt);
569 xlog_state_want_sync(log, iclog);
570 spin_unlock(&log->l_icloglock);
571 error = xlog_state_release_iclog(log, iclog);
573 spin_lock(&log->l_icloglock);
575 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
576 || iclog->ic_state == XLOG_STATE_DIRTY
577 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
579 sv_wait(&iclog->ic_force_wait, PMEM,
580 &log->l_icloglock, s);
582 spin_unlock(&log->l_icloglock);
587 } /* xfs_log_unmount_write */
590 * Deallocate log structures for unmount/relocation.
592 * We need to stop the aild from running before we destroy
593 * and deallocate the log as the aild references the log.
596 xfs_log_unmount(xfs_mount_t *mp)
598 xfs_trans_ail_destroy(mp);
599 xlog_dealloc_log(mp->m_log);
604 struct xfs_mount *mp,
605 struct xfs_log_item *item,
607 struct xfs_item_ops *ops)
609 item->li_mountp = mp;
610 item->li_ailp = mp->m_ail;
611 item->li_type = type;
615 INIT_LIST_HEAD(&item->li_ail);
616 INIT_LIST_HEAD(&item->li_cil);
620 * Write region vectors to log. The write happens using the space reservation
621 * of the ticket (tic). It is not a requirement that all writes for a given
622 * transaction occur with one call to xfs_log_write(). However, it is important
623 * to note that the transaction reservation code makes an assumption about the
624 * number of log headers a transaction requires that may be violated if you
625 * don't pass all the transaction vectors in one call....
629 struct xfs_mount *mp,
630 struct xfs_log_iovec reg[],
632 struct xlog_ticket *tic,
633 xfs_lsn_t *start_lsn)
635 struct log *log = mp->m_log;
637 struct xfs_log_vec vec = {
638 .lv_niovecs = nentries,
642 if (XLOG_FORCED_SHUTDOWN(log))
643 return XFS_ERROR(EIO);
645 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
647 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
652 xfs_log_move_tail(xfs_mount_t *mp,
656 xlog_t *log = mp->m_log;
657 int need_bytes, free_bytes, cycle, bytes;
659 if (XLOG_FORCED_SHUTDOWN(log))
663 /* needed since sync_lsn is 64 bits */
664 spin_lock(&log->l_icloglock);
665 tail_lsn = log->l_last_sync_lsn;
666 spin_unlock(&log->l_icloglock);
669 spin_lock(&log->l_grant_lock);
671 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
675 log->l_tail_lsn = tail_lsn;
678 if (!list_empty(&log->l_writeq)) {
680 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
681 panic("Recovery problem");
683 cycle = log->l_grant_write_cycle;
684 bytes = log->l_grant_write_bytes;
685 free_bytes = xlog_space_left(log, cycle, bytes);
686 list_for_each_entry(tic, &log->l_writeq, t_queue) {
687 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
689 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
692 free_bytes -= tic->t_unit_res;
693 sv_signal(&tic->t_wait);
697 if (!list_empty(&log->l_reserveq)) {
699 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
700 panic("Recovery problem");
702 cycle = log->l_grant_reserve_cycle;
703 bytes = log->l_grant_reserve_bytes;
704 free_bytes = xlog_space_left(log, cycle, bytes);
705 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
706 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
707 need_bytes = tic->t_unit_res*tic->t_cnt;
709 need_bytes = tic->t_unit_res;
710 if (free_bytes < need_bytes && tail_lsn != 1)
713 free_bytes -= need_bytes;
714 sv_signal(&tic->t_wait);
717 spin_unlock(&log->l_grant_lock);
718 } /* xfs_log_move_tail */
721 * Determine if we have a transaction that has gone to disk
722 * that needs to be covered. To begin the transition to the idle state
723 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
724 * If we are then in a state where covering is needed, the caller is informed
725 * that dummy transactions are required to move the log into the idle state.
727 * Because this is called as part of the sync process, we should also indicate
728 * that dummy transactions should be issued in anything but the covered or
729 * idle states. This ensures that the log tail is accurately reflected in
730 * the log at the end of the sync, hence if a crash occurrs avoids replay
731 * of transactions where the metadata is already on disk.
734 xfs_log_need_covered(xfs_mount_t *mp)
737 xlog_t *log = mp->m_log;
739 if (!xfs_fs_writable(mp))
742 spin_lock(&log->l_icloglock);
743 switch (log->l_covered_state) {
744 case XLOG_STATE_COVER_DONE:
745 case XLOG_STATE_COVER_DONE2:
746 case XLOG_STATE_COVER_IDLE:
748 case XLOG_STATE_COVER_NEED:
749 case XLOG_STATE_COVER_NEED2:
750 if (!xfs_trans_ail_tail(log->l_ailp) &&
751 xlog_iclogs_empty(log)) {
752 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
753 log->l_covered_state = XLOG_STATE_COVER_DONE;
755 log->l_covered_state = XLOG_STATE_COVER_DONE2;
762 spin_unlock(&log->l_icloglock);
766 /******************************************************************************
770 ******************************************************************************
773 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
774 * The log manager must keep track of the last LR which was committed
775 * to disk. The lsn of this LR will become the new tail_lsn whenever
776 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
777 * the situation where stuff could be written into the log but nothing
778 * was ever in the AIL when asked. Eventually, we panic since the
779 * tail hits the head.
781 * We may be holding the log iclog lock upon entering this routine.
784 xlog_assign_tail_lsn(xfs_mount_t *mp)
787 xlog_t *log = mp->m_log;
789 tail_lsn = xfs_trans_ail_tail(mp->m_ail);
790 spin_lock(&log->l_grant_lock);
792 log->l_tail_lsn = tail_lsn;
794 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
796 spin_unlock(&log->l_grant_lock);
799 } /* xlog_assign_tail_lsn */
803 * Return the space in the log between the tail and the head. The head
804 * is passed in the cycle/bytes formal parms. In the special case where
805 * the reserve head has wrapped passed the tail, this calculation is no
806 * longer valid. In this case, just return 0 which means there is no space
807 * in the log. This works for all places where this function is called
808 * with the reserve head. Of course, if the write head were to ever
809 * wrap the tail, we should blow up. Rather than catch this case here,
810 * we depend on other ASSERTions in other parts of the code. XXXmiken
812 * This code also handles the case where the reservation head is behind
813 * the tail. The details of this case are described below, but the end
814 * result is that we return the size of the log as the amount of space left.
817 xlog_space_left(xlog_t *log, int cycle, int bytes)
823 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
824 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
825 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
826 free_bytes = log->l_logsize - (bytes - tail_bytes);
827 } else if ((tail_cycle + 1) < cycle) {
829 } else if (tail_cycle < cycle) {
830 ASSERT(tail_cycle == (cycle - 1));
831 free_bytes = tail_bytes - bytes;
834 * The reservation head is behind the tail.
835 * In this case we just want to return the size of the
836 * log as the amount of space left.
838 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
839 "xlog_space_left: head behind tail\n"
840 " tail_cycle = %d, tail_bytes = %d\n"
841 " GH cycle = %d, GH bytes = %d",
842 tail_cycle, tail_bytes, cycle, bytes);
844 free_bytes = log->l_logsize;
847 } /* xlog_space_left */
851 * Log function which is called when an io completes.
853 * The log manager needs its own routine, in order to control what
854 * happens with the buffer after the write completes.
857 xlog_iodone(xfs_buf_t *bp)
859 xlog_in_core_t *iclog;
863 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
864 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
865 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
870 * Race to shutdown the filesystem if we see an error.
872 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
873 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
874 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
876 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
878 * This flag will be propagated to the trans-committed
879 * callback routines to let them know that the log-commit
882 aborted = XFS_LI_ABORTED;
883 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
884 aborted = XFS_LI_ABORTED;
887 /* log I/O is always issued ASYNC */
888 ASSERT(XFS_BUF_ISASYNC(bp));
889 xlog_state_done_syncing(iclog, aborted);
891 * do not reference the buffer (bp) here as we could race
892 * with it being freed after writing the unmount record to the
899 * Return size of each in-core log record buffer.
901 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
903 * If the filesystem blocksize is too large, we may need to choose a
904 * larger size since the directory code currently logs entire blocks.
908 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
914 if (mp->m_logbufs <= 0)
915 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
917 log->l_iclog_bufs = mp->m_logbufs;
920 * Buffer size passed in from mount system call.
922 if (mp->m_logbsize > 0) {
923 size = log->l_iclog_size = mp->m_logbsize;
924 log->l_iclog_size_log = 0;
926 log->l_iclog_size_log++;
930 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
931 /* # headers = size / 32k
932 * one header holds cycles from 32k of data
935 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
936 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
938 log->l_iclog_hsize = xhdrs << BBSHIFT;
939 log->l_iclog_heads = xhdrs;
941 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
942 log->l_iclog_hsize = BBSIZE;
943 log->l_iclog_heads = 1;
948 /* All machines use 32kB buffers by default. */
949 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
950 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
952 /* the default log size is 16k or 32k which is one header sector */
953 log->l_iclog_hsize = BBSIZE;
954 log->l_iclog_heads = 1;
957 /* are we being asked to make the sizes selected above visible? */
958 if (mp->m_logbufs == 0)
959 mp->m_logbufs = log->l_iclog_bufs;
960 if (mp->m_logbsize == 0)
961 mp->m_logbsize = log->l_iclog_size;
962 } /* xlog_get_iclog_buffer_size */
966 * This routine initializes some of the log structure for a given mount point.
967 * Its primary purpose is to fill in enough, so recovery can occur. However,
968 * some other stuff may be filled in too.
971 xlog_alloc_log(xfs_mount_t *mp,
972 xfs_buftarg_t *log_target,
973 xfs_daddr_t blk_offset,
977 xlog_rec_header_t *head;
978 xlog_in_core_t **iclogp;
979 xlog_in_core_t *iclog, *prev_iclog=NULL;
985 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
987 xlog_warn("XFS: Log allocation failed: No memory!");
992 log->l_targ = log_target;
993 log->l_logsize = BBTOB(num_bblks);
994 log->l_logBBstart = blk_offset;
995 log->l_logBBsize = num_bblks;
996 log->l_covered_state = XLOG_STATE_COVER_IDLE;
997 log->l_flags |= XLOG_ACTIVE_RECOVERY;
999 log->l_prev_block = -1;
1000 log->l_tail_lsn = xlog_assign_lsn(1, 0);
1001 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1002 log->l_last_sync_lsn = log->l_tail_lsn;
1003 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1004 log->l_grant_reserve_cycle = 1;
1005 log->l_grant_write_cycle = 1;
1006 INIT_LIST_HEAD(&log->l_reserveq);
1007 INIT_LIST_HEAD(&log->l_writeq);
1009 error = EFSCORRUPTED;
1010 if (xfs_sb_version_hassector(&mp->m_sb)) {
1011 log2_size = mp->m_sb.sb_logsectlog;
1012 if (log2_size < BBSHIFT) {
1013 xlog_warn("XFS: Log sector size too small "
1014 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1018 log2_size -= BBSHIFT;
1019 if (log2_size > mp->m_sectbb_log) {
1020 xlog_warn("XFS: Log sector size too large "
1021 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1025 /* for larger sector sizes, must have v2 or external log */
1026 if (log2_size && log->l_logBBstart > 0 &&
1027 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1029 xlog_warn("XFS: log sector size (0x%x) invalid "
1030 "for configuration.", log2_size);
1034 log->l_sectBBsize = 1 << log2_size;
1036 xlog_get_iclog_buffer_size(mp, log);
1039 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1042 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1043 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1044 ASSERT(XFS_BUF_ISBUSY(bp));
1045 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1048 spin_lock_init(&log->l_icloglock);
1049 spin_lock_init(&log->l_grant_lock);
1050 sv_init(&log->l_flush_wait, 0, "flush_wait");
1052 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1053 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1055 iclogp = &log->l_iclog;
1057 * The amount of memory to allocate for the iclog structure is
1058 * rather funky due to the way the structure is defined. It is
1059 * done this way so that we can use different sizes for machines
1060 * with different amounts of memory. See the definition of
1061 * xlog_in_core_t in xfs_log_priv.h for details.
1063 ASSERT(log->l_iclog_size >= 4096);
1064 for (i=0; i < log->l_iclog_bufs; i++) {
1065 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1067 goto out_free_iclog;
1070 iclog->ic_prev = prev_iclog;
1073 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1074 log->l_iclog_size, 0);
1076 goto out_free_iclog;
1077 if (!XFS_BUF_CPSEMA(bp))
1079 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1080 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1082 iclog->ic_data = bp->b_addr;
1084 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1086 head = &iclog->ic_header;
1087 memset(head, 0, sizeof(xlog_rec_header_t));
1088 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1089 head->h_version = cpu_to_be32(
1090 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1091 head->h_size = cpu_to_be32(log->l_iclog_size);
1093 head->h_fmt = cpu_to_be32(XLOG_FMT);
1094 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1096 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1097 iclog->ic_state = XLOG_STATE_ACTIVE;
1098 iclog->ic_log = log;
1099 atomic_set(&iclog->ic_refcnt, 0);
1100 spin_lock_init(&iclog->ic_callback_lock);
1101 iclog->ic_callback_tail = &(iclog->ic_callback);
1102 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1104 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1105 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1106 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1107 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1109 iclogp = &iclog->ic_next;
1111 *iclogp = log->l_iclog; /* complete ring */
1112 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1114 error = xlog_cil_init(log);
1116 goto out_free_iclog;
1120 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1121 prev_iclog = iclog->ic_next;
1123 sv_destroy(&iclog->ic_force_wait);
1124 sv_destroy(&iclog->ic_write_wait);
1125 xfs_buf_free(iclog->ic_bp);
1129 spinlock_destroy(&log->l_icloglock);
1130 spinlock_destroy(&log->l_grant_lock);
1131 xfs_buf_free(log->l_xbuf);
1135 return ERR_PTR(-error);
1136 } /* xlog_alloc_log */
1140 * Write out the commit record of a transaction associated with the given
1141 * ticket. Return the lsn of the commit record.
1146 struct xlog_ticket *ticket,
1147 struct xlog_in_core **iclog,
1148 xfs_lsn_t *commitlsnp)
1150 struct xfs_mount *mp = log->l_mp;
1152 struct xfs_log_iovec reg = {
1155 .i_type = XLOG_REG_TYPE_COMMIT,
1157 struct xfs_log_vec vec = {
1162 ASSERT_ALWAYS(iclog);
1163 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1166 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1171 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1172 * log space. This code pushes on the lsn which would supposedly free up
1173 * the 25% which we want to leave free. We may need to adopt a policy which
1174 * pushes on an lsn which is further along in the log once we reach the high
1175 * water mark. In this manner, we would be creating a low water mark.
1178 xlog_grant_push_ail(xfs_mount_t *mp,
1181 xlog_t *log = mp->m_log; /* pointer to the log */
1182 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1183 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1184 int free_blocks; /* free blocks left to write to */
1185 int free_bytes; /* free bytes left to write to */
1186 int threshold_block; /* block in lsn we'd like to be at */
1187 int threshold_cycle; /* lsn cycle we'd like to be at */
1190 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1192 spin_lock(&log->l_grant_lock);
1193 free_bytes = xlog_space_left(log,
1194 log->l_grant_reserve_cycle,
1195 log->l_grant_reserve_bytes);
1196 tail_lsn = log->l_tail_lsn;
1197 free_blocks = BTOBBT(free_bytes);
1200 * Set the threshold for the minimum number of free blocks in the
1201 * log to the maximum of what the caller needs, one quarter of the
1202 * log, and 256 blocks.
1204 free_threshold = BTOBB(need_bytes);
1205 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1206 free_threshold = MAX(free_threshold, 256);
1207 if (free_blocks < free_threshold) {
1208 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1209 threshold_cycle = CYCLE_LSN(tail_lsn);
1210 if (threshold_block >= log->l_logBBsize) {
1211 threshold_block -= log->l_logBBsize;
1212 threshold_cycle += 1;
1214 threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1216 /* Don't pass in an lsn greater than the lsn of the last
1217 * log record known to be on disk.
1219 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1220 threshold_lsn = log->l_last_sync_lsn;
1222 spin_unlock(&log->l_grant_lock);
1225 * Get the transaction layer to kick the dirty buffers out to
1226 * disk asynchronously. No point in trying to do this if
1227 * the filesystem is shutting down.
1229 if (threshold_lsn &&
1230 !XLOG_FORCED_SHUTDOWN(log))
1231 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1232 } /* xlog_grant_push_ail */
1235 * The bdstrat callback function for log bufs. This gives us a central
1236 * place to trap bufs in case we get hit by a log I/O error and need to
1237 * shutdown. Actually, in practice, even when we didn't get a log error,
1238 * we transition the iclogs to IOERROR state *after* flushing all existing
1239 * iclogs to disk. This is because we don't want anymore new transactions to be
1240 * started or completed afterwards.
1246 struct xlog_in_core *iclog;
1248 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1249 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1250 XFS_BUF_ERROR(bp, EIO);
1252 xfs_buf_ioend(bp, 0);
1254 * It would seem logical to return EIO here, but we rely on
1255 * the log state machine to propagate I/O errors instead of
1261 bp->b_flags |= _XBF_RUN_QUEUES;
1262 xfs_buf_iorequest(bp);
1267 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1268 * fashion. Previously, we should have moved the current iclog
1269 * ptr in the log to point to the next available iclog. This allows further
1270 * write to continue while this code syncs out an iclog ready to go.
1271 * Before an in-core log can be written out, the data section must be scanned
1272 * to save away the 1st word of each BBSIZE block into the header. We replace
1273 * it with the current cycle count. Each BBSIZE block is tagged with the
1274 * cycle count because there in an implicit assumption that drives will
1275 * guarantee that entire 512 byte blocks get written at once. In other words,
1276 * we can't have part of a 512 byte block written and part not written. By
1277 * tagging each block, we will know which blocks are valid when recovering
1278 * after an unclean shutdown.
1280 * This routine is single threaded on the iclog. No other thread can be in
1281 * this routine with the same iclog. Changing contents of iclog can there-
1282 * fore be done without grabbing the state machine lock. Updating the global
1283 * log will require grabbing the lock though.
1285 * The entire log manager uses a logical block numbering scheme. Only
1286 * log_sync (and then only bwrite()) know about the fact that the log may
1287 * not start with block zero on a given device. The log block start offset
1288 * is added immediately before calling bwrite().
1292 xlog_sync(xlog_t *log,
1293 xlog_in_core_t *iclog)
1295 xfs_caddr_t dptr; /* pointer to byte sized element */
1298 uint count; /* byte count of bwrite */
1299 uint count_init; /* initial count before roundup */
1300 int roundoff; /* roundoff to BB or stripe */
1301 int split = 0; /* split write into two regions */
1303 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1305 XFS_STATS_INC(xs_log_writes);
1306 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1308 /* Add for LR header */
1309 count_init = log->l_iclog_hsize + iclog->ic_offset;
1311 /* Round out the log write size */
1312 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1313 /* we have a v2 stripe unit to use */
1314 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1316 count = BBTOB(BTOBB(count_init));
1318 roundoff = count - count_init;
1319 ASSERT(roundoff >= 0);
1320 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1321 roundoff < log->l_mp->m_sb.sb_logsunit)
1323 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1324 roundoff < BBTOB(1)));
1326 /* move grant heads by roundoff in sync */
1327 spin_lock(&log->l_grant_lock);
1328 xlog_grant_add_space(log, &log->l_grant_reserve_cycle,
1329 &log->l_grant_reserve_bytes, roundoff);
1330 xlog_grant_add_space(log, &log->l_grant_write_cycle,
1331 &log->l_grant_write_bytes, roundoff);
1332 spin_unlock(&log->l_grant_lock);
1334 /* put cycle number in every block */
1335 xlog_pack_data(log, iclog, roundoff);
1337 /* real byte length */
1339 iclog->ic_header.h_len =
1340 cpu_to_be32(iclog->ic_offset + roundoff);
1342 iclog->ic_header.h_len =
1343 cpu_to_be32(iclog->ic_offset);
1347 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1348 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1349 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1351 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1353 /* Do we need to split this write into 2 parts? */
1354 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1355 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1356 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1357 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1359 iclog->ic_bwritecnt = 1;
1361 XFS_BUF_SET_COUNT(bp, count);
1362 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1363 XFS_BUF_ZEROFLAGS(bp);
1366 bp->b_flags |= XBF_LOG_BUFFER;
1368 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1369 XFS_BUF_ORDERED(bp);
1371 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1372 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1374 xlog_verify_iclog(log, iclog, count, B_TRUE);
1376 /* account for log which doesn't start at block #0 */
1377 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1379 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1384 if ((error = xlog_bdstrat(bp))) {
1385 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1390 bp = iclog->ic_log->l_xbuf;
1391 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1393 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1394 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1395 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1396 (__psint_t)count), split);
1397 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1398 XFS_BUF_ZEROFLAGS(bp);
1401 bp->b_flags |= XBF_LOG_BUFFER;
1402 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1403 XFS_BUF_ORDERED(bp);
1404 dptr = XFS_BUF_PTR(bp);
1406 * Bump the cycle numbers at the start of each block
1407 * since this part of the buffer is at the start of
1408 * a new cycle. Watch out for the header magic number
1411 for (i = 0; i < split; i += BBSIZE) {
1412 be32_add_cpu((__be32 *)dptr, 1);
1413 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1414 be32_add_cpu((__be32 *)dptr, 1);
1418 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1419 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1421 /* account for internal log which doesn't start at block #0 */
1422 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1424 if ((error = xlog_bdstrat(bp))) {
1425 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1426 bp, XFS_BUF_ADDR(bp));
1435 * Deallocate a log structure
1438 xlog_dealloc_log(xlog_t *log)
1440 xlog_in_core_t *iclog, *next_iclog;
1443 xlog_cil_destroy(log);
1445 iclog = log->l_iclog;
1446 for (i=0; i<log->l_iclog_bufs; i++) {
1447 sv_destroy(&iclog->ic_force_wait);
1448 sv_destroy(&iclog->ic_write_wait);
1449 xfs_buf_free(iclog->ic_bp);
1450 next_iclog = iclog->ic_next;
1454 spinlock_destroy(&log->l_icloglock);
1455 spinlock_destroy(&log->l_grant_lock);
1457 xfs_buf_free(log->l_xbuf);
1458 log->l_mp->m_log = NULL;
1460 } /* xlog_dealloc_log */
1463 * Update counters atomically now that memcpy is done.
1467 xlog_state_finish_copy(xlog_t *log,
1468 xlog_in_core_t *iclog,
1472 spin_lock(&log->l_icloglock);
1474 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1475 iclog->ic_offset += copy_bytes;
1477 spin_unlock(&log->l_icloglock);
1478 } /* xlog_state_finish_copy */
1484 * print out info relating to regions written which consume
1489 struct xfs_mount *mp,
1490 struct xlog_ticket *ticket)
1493 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1495 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1496 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1517 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1560 xfs_fs_cmn_err(CE_WARN, mp,
1561 "xfs_log_write: reservation summary:\n"
1562 " trans type = %s (%u)\n"
1563 " unit res = %d bytes\n"
1564 " current res = %d bytes\n"
1565 " total reg = %u bytes (o/flow = %u bytes)\n"
1566 " ophdrs = %u (ophdr space = %u bytes)\n"
1567 " ophdr + reg = %u bytes\n"
1568 " num regions = %u\n",
1569 ((ticket->t_trans_type <= 0 ||
1570 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1571 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1572 ticket->t_trans_type,
1575 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1576 ticket->t_res_num_ophdrs, ophdr_spc,
1577 ticket->t_res_arr_sum +
1578 ticket->t_res_o_flow + ophdr_spc,
1581 for (i = 0; i < ticket->t_res_num; i++) {
1582 uint r_type = ticket->t_res_arr[i].r_type;
1584 "region[%u]: %s - %u bytes\n",
1586 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1587 "bad-rtype" : res_type_str[r_type-1]),
1588 ticket->t_res_arr[i].r_len);
1591 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1592 "xfs_log_write: reservation ran out. Need to up reservation");
1593 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1597 * Calculate the potential space needed by the log vector. Each region gets
1598 * its own xlog_op_header_t and may need to be double word aligned.
1601 xlog_write_calc_vec_length(
1602 struct xlog_ticket *ticket,
1603 struct xfs_log_vec *log_vector)
1605 struct xfs_log_vec *lv;
1610 /* acct for start rec of xact */
1611 if (ticket->t_flags & XLOG_TIC_INITED)
1614 for (lv = log_vector; lv; lv = lv->lv_next) {
1615 headers += lv->lv_niovecs;
1617 for (i = 0; i < lv->lv_niovecs; i++) {
1618 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1621 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1625 ticket->t_res_num_ophdrs += headers;
1626 len += headers * sizeof(struct xlog_op_header);
1632 * If first write for transaction, insert start record We can't be trying to
1633 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1636 xlog_write_start_rec(
1637 struct xlog_op_header *ophdr,
1638 struct xlog_ticket *ticket)
1640 if (!(ticket->t_flags & XLOG_TIC_INITED))
1643 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1644 ophdr->oh_clientid = ticket->t_clientid;
1646 ophdr->oh_flags = XLOG_START_TRANS;
1649 ticket->t_flags &= ~XLOG_TIC_INITED;
1651 return sizeof(struct xlog_op_header);
1654 static xlog_op_header_t *
1655 xlog_write_setup_ophdr(
1657 struct xlog_op_header *ophdr,
1658 struct xlog_ticket *ticket,
1661 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1662 ophdr->oh_clientid = ticket->t_clientid;
1665 /* are we copying a commit or unmount record? */
1666 ophdr->oh_flags = flags;
1669 * We've seen logs corrupted with bad transaction client ids. This
1670 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1671 * and shut down the filesystem.
1673 switch (ophdr->oh_clientid) {
1674 case XFS_TRANSACTION:
1679 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1680 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1681 ophdr->oh_clientid, ticket);
1689 * Set up the parameters of the region copy into the log. This has
1690 * to handle region write split across multiple log buffers - this
1691 * state is kept external to this function so that this code can
1692 * can be written in an obvious, self documenting manner.
1695 xlog_write_setup_copy(
1696 struct xlog_ticket *ticket,
1697 struct xlog_op_header *ophdr,
1698 int space_available,
1702 int *last_was_partial_copy,
1703 int *bytes_consumed)
1707 still_to_copy = space_required - *bytes_consumed;
1708 *copy_off = *bytes_consumed;
1710 if (still_to_copy <= space_available) {
1711 /* write of region completes here */
1712 *copy_len = still_to_copy;
1713 ophdr->oh_len = cpu_to_be32(*copy_len);
1714 if (*last_was_partial_copy)
1715 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1716 *last_was_partial_copy = 0;
1717 *bytes_consumed = 0;
1721 /* partial write of region, needs extra log op header reservation */
1722 *copy_len = space_available;
1723 ophdr->oh_len = cpu_to_be32(*copy_len);
1724 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1725 if (*last_was_partial_copy)
1726 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1727 *bytes_consumed += *copy_len;
1728 (*last_was_partial_copy)++;
1730 /* account for new log op header */
1731 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1732 ticket->t_res_num_ophdrs++;
1734 return sizeof(struct xlog_op_header);
1738 xlog_write_copy_finish(
1740 struct xlog_in_core *iclog,
1745 int *partial_copy_len,
1747 struct xlog_in_core **commit_iclog)
1749 if (*partial_copy) {
1751 * This iclog has already been marked WANT_SYNC by
1752 * xlog_state_get_iclog_space.
1754 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1757 return xlog_state_release_iclog(log, iclog);
1761 *partial_copy_len = 0;
1763 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1764 /* no more space in this iclog - push it. */
1765 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1769 spin_lock(&log->l_icloglock);
1770 xlog_state_want_sync(log, iclog);
1771 spin_unlock(&log->l_icloglock);
1774 return xlog_state_release_iclog(log, iclog);
1775 ASSERT(flags & XLOG_COMMIT_TRANS);
1776 *commit_iclog = iclog;
1783 * Write some region out to in-core log
1785 * This will be called when writing externally provided regions or when
1786 * writing out a commit record for a given transaction.
1788 * General algorithm:
1789 * 1. Find total length of this write. This may include adding to the
1790 * lengths passed in.
1791 * 2. Check whether we violate the tickets reservation.
1792 * 3. While writing to this iclog
1793 * A. Reserve as much space in this iclog as can get
1794 * B. If this is first write, save away start lsn
1795 * C. While writing this region:
1796 * 1. If first write of transaction, write start record
1797 * 2. Write log operation header (header per region)
1798 * 3. Find out if we can fit entire region into this iclog
1799 * 4. Potentially, verify destination memcpy ptr
1800 * 5. Memcpy (partial) region
1801 * 6. If partial copy, release iclog; otherwise, continue
1802 * copying more regions into current iclog
1803 * 4. Mark want sync bit (in simulation mode)
1804 * 5. Release iclog for potential flush to on-disk log.
1807 * 1. Panic if reservation is overrun. This should never happen since
1808 * reservation amounts are generated internal to the filesystem.
1810 * 1. Tickets are single threaded data structures.
1811 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1812 * syncing routine. When a single log_write region needs to span
1813 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1814 * on all log operation writes which don't contain the end of the
1815 * region. The XLOG_END_TRANS bit is used for the in-core log
1816 * operation which contains the end of the continued log_write region.
1817 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1818 * we don't really know exactly how much space will be used. As a result,
1819 * we don't update ic_offset until the end when we know exactly how many
1820 * bytes have been written out.
1825 struct xfs_log_vec *log_vector,
1826 struct xlog_ticket *ticket,
1827 xfs_lsn_t *start_lsn,
1828 struct xlog_in_core **commit_iclog,
1831 struct xlog_in_core *iclog = NULL;
1832 struct xfs_log_iovec *vecp;
1833 struct xfs_log_vec *lv;
1836 int partial_copy = 0;
1837 int partial_copy_len = 0;
1845 len = xlog_write_calc_vec_length(ticket, log_vector);
1848 * Region headers and bytes are already accounted for.
1849 * We only need to take into account start records and
1850 * split regions in this function.
1852 if (ticket->t_flags & XLOG_TIC_INITED)
1853 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1856 * Commit record headers need to be accounted for. These
1857 * come in as separate writes so are easy to detect.
1859 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1860 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1862 ticket->t_curr_res -= len;
1864 if (ticket->t_curr_res < 0)
1865 xlog_print_tic_res(log->l_mp, ticket);
1869 vecp = lv->lv_iovecp;
1870 while (lv && index < lv->lv_niovecs) {
1874 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1875 &contwr, &log_offset);
1879 ASSERT(log_offset <= iclog->ic_size - 1);
1880 ptr = iclog->ic_datap + log_offset;
1882 /* start_lsn is the first lsn written to. That's all we need. */
1884 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1887 * This loop writes out as many regions as can fit in the amount
1888 * of space which was allocated by xlog_state_get_iclog_space().
1890 while (lv && index < lv->lv_niovecs) {
1891 struct xfs_log_iovec *reg = &vecp[index];
1892 struct xlog_op_header *ophdr;
1897 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1898 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1900 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1901 if (start_rec_copy) {
1903 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1907 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1909 return XFS_ERROR(EIO);
1911 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1912 sizeof(struct xlog_op_header));
1914 len += xlog_write_setup_copy(ticket, ophdr,
1915 iclog->ic_size-log_offset,
1917 ©_off, ©_len,
1920 xlog_verify_dest_ptr(log, ptr);
1923 ASSERT(copy_len >= 0);
1924 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1925 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1927 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1929 data_cnt += contwr ? copy_len : 0;
1931 error = xlog_write_copy_finish(log, iclog, flags,
1932 &record_cnt, &data_cnt,
1941 * if we had a partial copy, we need to get more iclog
1942 * space but we don't want to increment the region
1943 * index because there is still more is this region to
1946 * If we completed writing this region, and we flushed
1947 * the iclog (indicated by resetting of the record
1948 * count), then we also need to get more log space. If
1949 * this was the last record, though, we are done and
1955 if (++index == lv->lv_niovecs) {
1959 vecp = lv->lv_iovecp;
1961 if (record_cnt == 0) {
1971 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1973 return xlog_state_release_iclog(log, iclog);
1975 ASSERT(flags & XLOG_COMMIT_TRANS);
1976 *commit_iclog = iclog;
1981 /*****************************************************************************
1983 * State Machine functions
1985 *****************************************************************************
1988 /* Clean iclogs starting from the head. This ordering must be
1989 * maintained, so an iclog doesn't become ACTIVE beyond one that
1990 * is SYNCING. This is also required to maintain the notion that we use
1991 * a ordered wait queue to hold off would be writers to the log when every
1992 * iclog is trying to sync to disk.
1994 * State Change: DIRTY -> ACTIVE
1997 xlog_state_clean_log(xlog_t *log)
1999 xlog_in_core_t *iclog;
2002 iclog = log->l_iclog;
2004 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2005 iclog->ic_state = XLOG_STATE_ACTIVE;
2006 iclog->ic_offset = 0;
2007 ASSERT(iclog->ic_callback == NULL);
2009 * If the number of ops in this iclog indicate it just
2010 * contains the dummy transaction, we can
2011 * change state into IDLE (the second time around).
2012 * Otherwise we should change the state into
2014 * We don't need to cover the dummy.
2017 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2022 * We have two dirty iclogs so start over
2023 * This could also be num of ops indicates
2024 * this is not the dummy going out.
2028 iclog->ic_header.h_num_logops = 0;
2029 memset(iclog->ic_header.h_cycle_data, 0,
2030 sizeof(iclog->ic_header.h_cycle_data));
2031 iclog->ic_header.h_lsn = 0;
2032 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2035 break; /* stop cleaning */
2036 iclog = iclog->ic_next;
2037 } while (iclog != log->l_iclog);
2039 /* log is locked when we are called */
2041 * Change state for the dummy log recording.
2042 * We usually go to NEED. But we go to NEED2 if the changed indicates
2043 * we are done writing the dummy record.
2044 * If we are done with the second dummy recored (DONE2), then
2048 switch (log->l_covered_state) {
2049 case XLOG_STATE_COVER_IDLE:
2050 case XLOG_STATE_COVER_NEED:
2051 case XLOG_STATE_COVER_NEED2:
2052 log->l_covered_state = XLOG_STATE_COVER_NEED;
2055 case XLOG_STATE_COVER_DONE:
2057 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2059 log->l_covered_state = XLOG_STATE_COVER_NEED;
2062 case XLOG_STATE_COVER_DONE2:
2064 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2066 log->l_covered_state = XLOG_STATE_COVER_NEED;
2073 } /* xlog_state_clean_log */
2076 xlog_get_lowest_lsn(
2079 xlog_in_core_t *lsn_log;
2080 xfs_lsn_t lowest_lsn, lsn;
2082 lsn_log = log->l_iclog;
2085 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2086 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2087 if ((lsn && !lowest_lsn) ||
2088 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2092 lsn_log = lsn_log->ic_next;
2093 } while (lsn_log != log->l_iclog);
2099 xlog_state_do_callback(
2102 xlog_in_core_t *ciclog)
2104 xlog_in_core_t *iclog;
2105 xlog_in_core_t *first_iclog; /* used to know when we've
2106 * processed all iclogs once */
2107 xfs_log_callback_t *cb, *cb_next;
2109 xfs_lsn_t lowest_lsn;
2110 int ioerrors; /* counter: iclogs with errors */
2111 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2112 int funcdidcallbacks; /* flag: function did callbacks */
2113 int repeats; /* for issuing console warnings if
2114 * looping too many times */
2117 spin_lock(&log->l_icloglock);
2118 first_iclog = iclog = log->l_iclog;
2120 funcdidcallbacks = 0;
2125 * Scan all iclogs starting with the one pointed to by the
2126 * log. Reset this starting point each time the log is
2127 * unlocked (during callbacks).
2129 * Keep looping through iclogs until one full pass is made
2130 * without running any callbacks.
2132 first_iclog = log->l_iclog;
2133 iclog = log->l_iclog;
2134 loopdidcallbacks = 0;
2139 /* skip all iclogs in the ACTIVE & DIRTY states */
2140 if (iclog->ic_state &
2141 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2142 iclog = iclog->ic_next;
2147 * Between marking a filesystem SHUTDOWN and stopping
2148 * the log, we do flush all iclogs to disk (if there
2149 * wasn't a log I/O error). So, we do want things to
2150 * go smoothly in case of just a SHUTDOWN w/o a
2153 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2155 * Can only perform callbacks in order. Since
2156 * this iclog is not in the DONE_SYNC/
2157 * DO_CALLBACK state, we skip the rest and
2158 * just try to clean up. If we set our iclog
2159 * to DO_CALLBACK, we will not process it when
2160 * we retry since a previous iclog is in the
2161 * CALLBACK and the state cannot change since
2162 * we are holding the l_icloglock.
2164 if (!(iclog->ic_state &
2165 (XLOG_STATE_DONE_SYNC |
2166 XLOG_STATE_DO_CALLBACK))) {
2167 if (ciclog && (ciclog->ic_state ==
2168 XLOG_STATE_DONE_SYNC)) {
2169 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2174 * We now have an iclog that is in either the
2175 * DO_CALLBACK or DONE_SYNC states. The other
2176 * states (WANT_SYNC, SYNCING, or CALLBACK were
2177 * caught by the above if and are going to
2178 * clean (i.e. we aren't doing their callbacks)
2183 * We will do one more check here to see if we
2184 * have chased our tail around.
2187 lowest_lsn = xlog_get_lowest_lsn(log);
2189 XFS_LSN_CMP(lowest_lsn,
2190 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2191 iclog = iclog->ic_next;
2192 continue; /* Leave this iclog for
2196 iclog->ic_state = XLOG_STATE_CALLBACK;
2198 spin_unlock(&log->l_icloglock);
2200 /* l_last_sync_lsn field protected by
2201 * l_grant_lock. Don't worry about iclog's lsn.
2202 * No one else can be here except us.
2204 spin_lock(&log->l_grant_lock);
2205 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2206 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2207 log->l_last_sync_lsn =
2208 be64_to_cpu(iclog->ic_header.h_lsn);
2209 spin_unlock(&log->l_grant_lock);
2212 spin_unlock(&log->l_icloglock);
2217 * Keep processing entries in the callback list until
2218 * we come around and it is empty. We need to
2219 * atomically see that the list is empty and change the
2220 * state to DIRTY so that we don't miss any more
2221 * callbacks being added.
2223 spin_lock(&iclog->ic_callback_lock);
2224 cb = iclog->ic_callback;
2226 iclog->ic_callback_tail = &(iclog->ic_callback);
2227 iclog->ic_callback = NULL;
2228 spin_unlock(&iclog->ic_callback_lock);
2230 /* perform callbacks in the order given */
2231 for (; cb; cb = cb_next) {
2232 cb_next = cb->cb_next;
2233 cb->cb_func(cb->cb_arg, aborted);
2235 spin_lock(&iclog->ic_callback_lock);
2236 cb = iclog->ic_callback;
2242 spin_lock(&log->l_icloglock);
2243 ASSERT(iclog->ic_callback == NULL);
2244 spin_unlock(&iclog->ic_callback_lock);
2245 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2246 iclog->ic_state = XLOG_STATE_DIRTY;
2249 * Transition from DIRTY to ACTIVE if applicable.
2250 * NOP if STATE_IOERROR.
2252 xlog_state_clean_log(log);
2254 /* wake up threads waiting in xfs_log_force() */
2255 sv_broadcast(&iclog->ic_force_wait);
2257 iclog = iclog->ic_next;
2258 } while (first_iclog != iclog);
2260 if (repeats > 5000) {
2261 flushcnt += repeats;
2263 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2264 "%s: possible infinite loop (%d iterations)",
2265 __func__, flushcnt);
2267 } while (!ioerrors && loopdidcallbacks);
2270 * make one last gasp attempt to see if iclogs are being left in
2274 if (funcdidcallbacks) {
2275 first_iclog = iclog = log->l_iclog;
2277 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2279 * Terminate the loop if iclogs are found in states
2280 * which will cause other threads to clean up iclogs.
2282 * SYNCING - i/o completion will go through logs
2283 * DONE_SYNC - interrupt thread should be waiting for
2285 * IOERROR - give up hope all ye who enter here
2287 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2288 iclog->ic_state == XLOG_STATE_SYNCING ||
2289 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2290 iclog->ic_state == XLOG_STATE_IOERROR )
2292 iclog = iclog->ic_next;
2293 } while (first_iclog != iclog);
2297 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2299 spin_unlock(&log->l_icloglock);
2302 sv_broadcast(&log->l_flush_wait);
2307 * Finish transitioning this iclog to the dirty state.
2309 * Make sure that we completely execute this routine only when this is
2310 * the last call to the iclog. There is a good chance that iclog flushes,
2311 * when we reach the end of the physical log, get turned into 2 separate
2312 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2313 * routine. By using the reference count bwritecnt, we guarantee that only
2314 * the second completion goes through.
2316 * Callbacks could take time, so they are done outside the scope of the
2317 * global state machine log lock.
2320 xlog_state_done_syncing(
2321 xlog_in_core_t *iclog,
2324 xlog_t *log = iclog->ic_log;
2326 spin_lock(&log->l_icloglock);
2328 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2329 iclog->ic_state == XLOG_STATE_IOERROR);
2330 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2331 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2335 * If we got an error, either on the first buffer, or in the case of
2336 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2337 * and none should ever be attempted to be written to disk
2340 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2341 if (--iclog->ic_bwritecnt == 1) {
2342 spin_unlock(&log->l_icloglock);
2345 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2349 * Someone could be sleeping prior to writing out the next
2350 * iclog buffer, we wake them all, one will get to do the
2351 * I/O, the others get to wait for the result.
2353 sv_broadcast(&iclog->ic_write_wait);
2354 spin_unlock(&log->l_icloglock);
2355 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2356 } /* xlog_state_done_syncing */
2360 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2361 * sleep. We wait on the flush queue on the head iclog as that should be
2362 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2363 * we will wait here and all new writes will sleep until a sync completes.
2365 * The in-core logs are used in a circular fashion. They are not used
2366 * out-of-order even when an iclog past the head is free.
2369 * * log_offset where xlog_write() can start writing into the in-core
2371 * * in-core log pointer to which xlog_write() should write.
2372 * * boolean indicating this is a continued write to an in-core log.
2373 * If this is the last write, then the in-core log's offset field
2374 * needs to be incremented, depending on the amount of data which
2378 xlog_state_get_iclog_space(xlog_t *log,
2380 xlog_in_core_t **iclogp,
2381 xlog_ticket_t *ticket,
2382 int *continued_write,
2386 xlog_rec_header_t *head;
2387 xlog_in_core_t *iclog;
2391 spin_lock(&log->l_icloglock);
2392 if (XLOG_FORCED_SHUTDOWN(log)) {
2393 spin_unlock(&log->l_icloglock);
2394 return XFS_ERROR(EIO);
2397 iclog = log->l_iclog;
2398 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2399 XFS_STATS_INC(xs_log_noiclogs);
2401 /* Wait for log writes to have flushed */
2402 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2406 head = &iclog->ic_header;
2408 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2409 log_offset = iclog->ic_offset;
2411 /* On the 1st write to an iclog, figure out lsn. This works
2412 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2413 * committing to. If the offset is set, that's how many blocks
2416 if (log_offset == 0) {
2417 ticket->t_curr_res -= log->l_iclog_hsize;
2418 xlog_tic_add_region(ticket,
2420 XLOG_REG_TYPE_LRHEADER);
2421 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2422 head->h_lsn = cpu_to_be64(
2423 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2424 ASSERT(log->l_curr_block >= 0);
2427 /* If there is enough room to write everything, then do it. Otherwise,
2428 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2429 * bit is on, so this will get flushed out. Don't update ic_offset
2430 * until you know exactly how many bytes get copied. Therefore, wait
2431 * until later to update ic_offset.
2433 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2434 * can fit into remaining data section.
2436 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2437 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2440 * If I'm the only one writing to this iclog, sync it to disk.
2441 * We need to do an atomic compare and decrement here to avoid
2442 * racing with concurrent atomic_dec_and_lock() calls in
2443 * xlog_state_release_iclog() when there is more than one
2444 * reference to the iclog.
2446 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2447 /* we are the only one */
2448 spin_unlock(&log->l_icloglock);
2449 error = xlog_state_release_iclog(log, iclog);
2453 spin_unlock(&log->l_icloglock);
2458 /* Do we have enough room to write the full amount in the remainder
2459 * of this iclog? Or must we continue a write on the next iclog and
2460 * mark this iclog as completely taken? In the case where we switch
2461 * iclogs (to mark it taken), this particular iclog will release/sync
2462 * to disk in xlog_write().
2464 if (len <= iclog->ic_size - iclog->ic_offset) {
2465 *continued_write = 0;
2466 iclog->ic_offset += len;
2468 *continued_write = 1;
2469 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2473 ASSERT(iclog->ic_offset <= iclog->ic_size);
2474 spin_unlock(&log->l_icloglock);
2476 *logoffsetp = log_offset;
2478 } /* xlog_state_get_iclog_space */
2481 * Atomically get the log space required for a log ticket.
2483 * Once a ticket gets put onto the reserveq, it will only return after
2484 * the needed reservation is satisfied.
2487 xlog_grant_log_space(xlog_t *log,
2494 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2495 panic("grant Recovery problem");
2498 /* Is there space or do we need to sleep? */
2499 spin_lock(&log->l_grant_lock);
2501 trace_xfs_log_grant_enter(log, tic);
2503 /* something is already sleeping; insert new transaction at end */
2504 if (!list_empty(&log->l_reserveq)) {
2505 list_add_tail(&tic->t_queue, &log->l_reserveq);
2507 trace_xfs_log_grant_sleep1(log, tic);
2510 * Gotta check this before going to sleep, while we're
2511 * holding the grant lock.
2513 if (XLOG_FORCED_SHUTDOWN(log))
2516 XFS_STATS_INC(xs_sleep_logspace);
2517 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2519 * If we got an error, and the filesystem is shutting down,
2520 * we'll catch it down below. So just continue...
2522 trace_xfs_log_grant_wake1(log, tic);
2523 spin_lock(&log->l_grant_lock);
2525 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2526 need_bytes = tic->t_unit_res*tic->t_ocnt;
2528 need_bytes = tic->t_unit_res;
2531 if (XLOG_FORCED_SHUTDOWN(log))
2534 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2535 log->l_grant_reserve_bytes);
2536 if (free_bytes < need_bytes) {
2537 if (list_empty(&tic->t_queue))
2538 list_add_tail(&tic->t_queue, &log->l_reserveq);
2540 trace_xfs_log_grant_sleep2(log, tic);
2542 spin_unlock(&log->l_grant_lock);
2543 xlog_grant_push_ail(log->l_mp, need_bytes);
2544 spin_lock(&log->l_grant_lock);
2546 XFS_STATS_INC(xs_sleep_logspace);
2547 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2549 spin_lock(&log->l_grant_lock);
2550 if (XLOG_FORCED_SHUTDOWN(log))
2553 trace_xfs_log_grant_wake2(log, tic);
2558 list_del_init(&tic->t_queue);
2560 /* we've got enough space */
2561 xlog_grant_add_space(log, &log->l_grant_reserve_cycle,
2562 &log->l_grant_reserve_bytes, need_bytes);
2563 xlog_grant_add_space(log, &log->l_grant_write_cycle,
2564 &log->l_grant_write_bytes, need_bytes);
2565 trace_xfs_log_grant_exit(log, tic);
2566 xlog_verify_grant_head(log, 1);
2567 xlog_verify_grant_tail(log);
2568 spin_unlock(&log->l_grant_lock);
2572 list_del_init(&tic->t_queue);
2573 trace_xfs_log_grant_error(log, tic);
2576 * If we are failing, make sure the ticket doesn't have any
2577 * current reservations. We don't want to add this back when
2578 * the ticket/transaction gets cancelled.
2580 tic->t_curr_res = 0;
2581 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2582 spin_unlock(&log->l_grant_lock);
2583 return XFS_ERROR(EIO);
2584 } /* xlog_grant_log_space */
2588 * Replenish the byte reservation required by moving the grant write head.
2593 xlog_regrant_write_log_space(xlog_t *log,
2596 int free_bytes, need_bytes;
2598 tic->t_curr_res = tic->t_unit_res;
2599 xlog_tic_reset_res(tic);
2605 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2606 panic("regrant Recovery problem");
2609 spin_lock(&log->l_grant_lock);
2611 trace_xfs_log_regrant_write_enter(log, tic);
2613 if (XLOG_FORCED_SHUTDOWN(log))
2616 /* If there are other waiters on the queue then give them a
2617 * chance at logspace before us. Wake up the first waiters,
2618 * if we do not wake up all the waiters then go to sleep waiting
2619 * for more free space, otherwise try to get some space for
2622 need_bytes = tic->t_unit_res;
2623 if (!list_empty(&log->l_writeq)) {
2624 struct xlog_ticket *ntic;
2625 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2626 log->l_grant_write_bytes);
2627 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2628 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2630 if (free_bytes < ntic->t_unit_res)
2632 free_bytes -= ntic->t_unit_res;
2633 sv_signal(&ntic->t_wait);
2636 if (ntic != list_first_entry(&log->l_writeq,
2637 struct xlog_ticket, t_queue)) {
2638 if (list_empty(&tic->t_queue))
2639 list_add_tail(&tic->t_queue, &log->l_writeq);
2641 trace_xfs_log_regrant_write_sleep1(log, tic);
2643 spin_unlock(&log->l_grant_lock);
2644 xlog_grant_push_ail(log->l_mp, need_bytes);
2645 spin_lock(&log->l_grant_lock);
2647 XFS_STATS_INC(xs_sleep_logspace);
2648 sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2649 &log->l_grant_lock, s);
2651 /* If we're shutting down, this tic is already
2653 spin_lock(&log->l_grant_lock);
2654 if (XLOG_FORCED_SHUTDOWN(log))
2657 trace_xfs_log_regrant_write_wake1(log, tic);
2662 if (XLOG_FORCED_SHUTDOWN(log))
2665 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2666 log->l_grant_write_bytes);
2667 if (free_bytes < need_bytes) {
2668 if (list_empty(&tic->t_queue))
2669 list_add_tail(&tic->t_queue, &log->l_writeq);
2670 spin_unlock(&log->l_grant_lock);
2671 xlog_grant_push_ail(log->l_mp, need_bytes);
2672 spin_lock(&log->l_grant_lock);
2674 XFS_STATS_INC(xs_sleep_logspace);
2675 trace_xfs_log_regrant_write_sleep2(log, tic);
2677 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2679 /* If we're shutting down, this tic is already off the queue */
2680 spin_lock(&log->l_grant_lock);
2681 if (XLOG_FORCED_SHUTDOWN(log))
2684 trace_xfs_log_regrant_write_wake2(log, tic);
2688 list_del_init(&tic->t_queue);
2690 /* we've got enough space */
2691 xlog_grant_add_space(log, &log->l_grant_write_cycle,
2692 &log->l_grant_write_bytes, need_bytes);
2693 trace_xfs_log_regrant_write_exit(log, tic);
2694 xlog_verify_grant_head(log, 1);
2695 xlog_verify_grant_tail(log);
2696 spin_unlock(&log->l_grant_lock);
2701 list_del_init(&tic->t_queue);
2702 trace_xfs_log_regrant_write_error(log, tic);
2705 * If we are failing, make sure the ticket doesn't have any
2706 * current reservations. We don't want to add this back when
2707 * the ticket/transaction gets cancelled.
2709 tic->t_curr_res = 0;
2710 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2711 spin_unlock(&log->l_grant_lock);
2712 return XFS_ERROR(EIO);
2713 } /* xlog_regrant_write_log_space */
2716 /* The first cnt-1 times through here we don't need to
2717 * move the grant write head because the permanent
2718 * reservation has reserved cnt times the unit amount.
2719 * Release part of current permanent unit reservation and
2720 * reset current reservation to be one units worth. Also
2721 * move grant reservation head forward.
2724 xlog_regrant_reserve_log_space(xlog_t *log,
2725 xlog_ticket_t *ticket)
2727 trace_xfs_log_regrant_reserve_enter(log, ticket);
2729 if (ticket->t_cnt > 0)
2732 spin_lock(&log->l_grant_lock);
2733 xlog_grant_sub_space(log, &log->l_grant_reserve_cycle,
2734 &log->l_grant_reserve_bytes,
2735 ticket->t_curr_res);
2736 xlog_grant_sub_space(log, &log->l_grant_write_cycle,
2737 &log->l_grant_write_bytes,
2738 ticket->t_curr_res);
2739 ticket->t_curr_res = ticket->t_unit_res;
2740 xlog_tic_reset_res(ticket);
2742 trace_xfs_log_regrant_reserve_sub(log, ticket);
2744 xlog_verify_grant_head(log, 1);
2746 /* just return if we still have some of the pre-reserved space */
2747 if (ticket->t_cnt > 0) {
2748 spin_unlock(&log->l_grant_lock);
2752 xlog_grant_add_space(log, &log->l_grant_reserve_cycle,
2753 &log->l_grant_reserve_bytes,
2754 ticket->t_unit_res);
2756 trace_xfs_log_regrant_reserve_exit(log, ticket);
2758 xlog_verify_grant_head(log, 0);
2759 spin_unlock(&log->l_grant_lock);
2760 ticket->t_curr_res = ticket->t_unit_res;
2761 xlog_tic_reset_res(ticket);
2762 } /* xlog_regrant_reserve_log_space */
2766 * Give back the space left from a reservation.
2768 * All the information we need to make a correct determination of space left
2769 * is present. For non-permanent reservations, things are quite easy. The
2770 * count should have been decremented to zero. We only need to deal with the
2771 * space remaining in the current reservation part of the ticket. If the
2772 * ticket contains a permanent reservation, there may be left over space which
2773 * needs to be released. A count of N means that N-1 refills of the current
2774 * reservation can be done before we need to ask for more space. The first
2775 * one goes to fill up the first current reservation. Once we run out of
2776 * space, the count will stay at zero and the only space remaining will be
2777 * in the current reservation field.
2780 xlog_ungrant_log_space(xlog_t *log,
2781 xlog_ticket_t *ticket)
2785 if (ticket->t_cnt > 0)
2788 spin_lock(&log->l_grant_lock);
2789 trace_xfs_log_ungrant_enter(log, ticket);
2790 trace_xfs_log_ungrant_sub(log, ticket);
2793 * If this is a permanent reservation ticket, we may be able to free
2794 * up more space based on the remaining count.
2796 bytes = ticket->t_curr_res;
2797 if (ticket->t_cnt > 0) {
2798 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2799 bytes += ticket->t_unit_res*ticket->t_cnt;
2802 xlog_grant_sub_space(log, &log->l_grant_reserve_cycle,
2803 &log->l_grant_reserve_bytes, bytes);
2804 xlog_grant_sub_space(log, &log->l_grant_write_cycle,
2805 &log->l_grant_write_bytes, bytes);
2807 trace_xfs_log_ungrant_exit(log, ticket);
2809 xlog_verify_grant_head(log, 1);
2810 spin_unlock(&log->l_grant_lock);
2811 xfs_log_move_tail(log->l_mp, 1);
2812 } /* xlog_ungrant_log_space */
2816 * Flush iclog to disk if this is the last reference to the given iclog and
2817 * the WANT_SYNC bit is set.
2819 * When this function is entered, the iclog is not necessarily in the
2820 * WANT_SYNC state. It may be sitting around waiting to get filled.
2825 xlog_state_release_iclog(
2827 xlog_in_core_t *iclog)
2829 int sync = 0; /* do we sync? */
2831 if (iclog->ic_state & XLOG_STATE_IOERROR)
2832 return XFS_ERROR(EIO);
2834 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2835 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2838 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2839 spin_unlock(&log->l_icloglock);
2840 return XFS_ERROR(EIO);
2842 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2843 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2845 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2846 /* update tail before writing to iclog */
2847 xlog_assign_tail_lsn(log->l_mp);
2849 iclog->ic_state = XLOG_STATE_SYNCING;
2850 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2851 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2852 /* cycle incremented when incrementing curr_block */
2854 spin_unlock(&log->l_icloglock);
2857 * We let the log lock go, so it's possible that we hit a log I/O
2858 * error or some other SHUTDOWN condition that marks the iclog
2859 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2860 * this iclog has consistent data, so we ignore IOERROR
2861 * flags after this point.
2864 return xlog_sync(log, iclog);
2866 } /* xlog_state_release_iclog */
2870 * This routine will mark the current iclog in the ring as WANT_SYNC
2871 * and move the current iclog pointer to the next iclog in the ring.
2872 * When this routine is called from xlog_state_get_iclog_space(), the
2873 * exact size of the iclog has not yet been determined. All we know is
2874 * that every data block. We have run out of space in this log record.
2877 xlog_state_switch_iclogs(xlog_t *log,
2878 xlog_in_core_t *iclog,
2881 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2883 eventual_size = iclog->ic_offset;
2884 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2885 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2886 log->l_prev_block = log->l_curr_block;
2887 log->l_prev_cycle = log->l_curr_cycle;
2889 /* roll log?: ic_offset changed later */
2890 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2892 /* Round up to next log-sunit */
2893 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2894 log->l_mp->m_sb.sb_logsunit > 1) {
2895 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2896 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2899 if (log->l_curr_block >= log->l_logBBsize) {
2900 log->l_curr_cycle++;
2901 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2902 log->l_curr_cycle++;
2903 log->l_curr_block -= log->l_logBBsize;
2904 ASSERT(log->l_curr_block >= 0);
2906 ASSERT(iclog == log->l_iclog);
2907 log->l_iclog = iclog->ic_next;
2908 } /* xlog_state_switch_iclogs */
2911 * Write out all data in the in-core log as of this exact moment in time.
2913 * Data may be written to the in-core log during this call. However,
2914 * we don't guarantee this data will be written out. A change from past
2915 * implementation means this routine will *not* write out zero length LRs.
2917 * Basically, we try and perform an intelligent scan of the in-core logs.
2918 * If we determine there is no flushable data, we just return. There is no
2919 * flushable data if:
2921 * 1. the current iclog is active and has no data; the previous iclog
2922 * is in the active or dirty state.
2923 * 2. the current iclog is drity, and the previous iclog is in the
2924 * active or dirty state.
2928 * 1. the current iclog is not in the active nor dirty state.
2929 * 2. the current iclog dirty, and the previous iclog is not in the
2930 * active nor dirty state.
2931 * 3. the current iclog is active, and there is another thread writing
2932 * to this particular iclog.
2933 * 4. a) the current iclog is active and has no other writers
2934 * b) when we return from flushing out this iclog, it is still
2935 * not in the active nor dirty state.
2939 struct xfs_mount *mp,
2943 struct log *log = mp->m_log;
2944 struct xlog_in_core *iclog;
2947 XFS_STATS_INC(xs_log_force);
2950 xlog_cil_force(log);
2952 spin_lock(&log->l_icloglock);
2954 iclog = log->l_iclog;
2955 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2956 spin_unlock(&log->l_icloglock);
2957 return XFS_ERROR(EIO);
2960 /* If the head iclog is not active nor dirty, we just attach
2961 * ourselves to the head and go to sleep.
2963 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2964 iclog->ic_state == XLOG_STATE_DIRTY) {
2966 * If the head is dirty or (active and empty), then
2967 * we need to look at the previous iclog. If the previous
2968 * iclog is active or dirty we are done. There is nothing
2969 * to sync out. Otherwise, we attach ourselves to the
2970 * previous iclog and go to sleep.
2972 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2973 (atomic_read(&iclog->ic_refcnt) == 0
2974 && iclog->ic_offset == 0)) {
2975 iclog = iclog->ic_prev;
2976 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2977 iclog->ic_state == XLOG_STATE_DIRTY)
2982 if (atomic_read(&iclog->ic_refcnt) == 0) {
2983 /* We are the only one with access to this
2984 * iclog. Flush it out now. There should
2985 * be a roundoff of zero to show that someone
2986 * has already taken care of the roundoff from
2987 * the previous sync.
2989 atomic_inc(&iclog->ic_refcnt);
2990 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2991 xlog_state_switch_iclogs(log, iclog, 0);
2992 spin_unlock(&log->l_icloglock);
2994 if (xlog_state_release_iclog(log, iclog))
2995 return XFS_ERROR(EIO);
2999 spin_lock(&log->l_icloglock);
3000 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3001 iclog->ic_state != XLOG_STATE_DIRTY)
3006 /* Someone else is writing to this iclog.
3007 * Use its call to flush out the data. However,
3008 * the other thread may not force out this LR,
3009 * so we mark it WANT_SYNC.
3011 xlog_state_switch_iclogs(log, iclog, 0);
3017 /* By the time we come around again, the iclog could've been filled
3018 * which would give it another lsn. If we have a new lsn, just
3019 * return because the relevant data has been flushed.
3022 if (flags & XFS_LOG_SYNC) {
3024 * We must check if we're shutting down here, before
3025 * we wait, while we're holding the l_icloglock.
3026 * Then we check again after waking up, in case our
3027 * sleep was disturbed by a bad news.
3029 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3030 spin_unlock(&log->l_icloglock);
3031 return XFS_ERROR(EIO);
3033 XFS_STATS_INC(xs_log_force_sleep);
3034 sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3036 * No need to grab the log lock here since we're
3037 * only deciding whether or not to return EIO
3038 * and the memory read should be atomic.
3040 if (iclog->ic_state & XLOG_STATE_IOERROR)
3041 return XFS_ERROR(EIO);
3047 spin_unlock(&log->l_icloglock);
3053 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3054 * about errors or whether the log was flushed or not. This is the normal
3055 * interface to use when trying to unpin items or move the log forward.
3064 error = _xfs_log_force(mp, flags, NULL);
3066 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3067 "error %d returned.", error);
3072 * Force the in-core log to disk for a specific LSN.
3074 * Find in-core log with lsn.
3075 * If it is in the DIRTY state, just return.
3076 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3077 * state and go to sleep or return.
3078 * If it is in any other state, go to sleep or return.
3080 * Synchronous forces are implemented with a signal variable. All callers
3081 * to force a given lsn to disk will wait on a the sv attached to the
3082 * specific in-core log. When given in-core log finally completes its
3083 * write to disk, that thread will wake up all threads waiting on the
3088 struct xfs_mount *mp,
3093 struct log *log = mp->m_log;
3094 struct xlog_in_core *iclog;
3095 int already_slept = 0;
3099 XFS_STATS_INC(xs_log_force);
3102 lsn = xlog_cil_force_lsn(log, lsn);
3103 if (lsn == NULLCOMMITLSN)
3108 spin_lock(&log->l_icloglock);
3109 iclog = log->l_iclog;
3110 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3111 spin_unlock(&log->l_icloglock);
3112 return XFS_ERROR(EIO);
3116 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3117 iclog = iclog->ic_next;
3121 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3122 spin_unlock(&log->l_icloglock);
3126 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3128 * We sleep here if we haven't already slept (e.g.
3129 * this is the first time we've looked at the correct
3130 * iclog buf) and the buffer before us is going to
3131 * be sync'ed. The reason for this is that if we
3132 * are doing sync transactions here, by waiting for
3133 * the previous I/O to complete, we can allow a few
3134 * more transactions into this iclog before we close
3137 * Otherwise, we mark the buffer WANT_SYNC, and bump
3138 * up the refcnt so we can release the log (which
3139 * drops the ref count). The state switch keeps new
3140 * transaction commits from using this buffer. When
3141 * the current commits finish writing into the buffer,
3142 * the refcount will drop to zero and the buffer will
3145 if (!already_slept &&
3146 (iclog->ic_prev->ic_state &
3147 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3148 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3150 XFS_STATS_INC(xs_log_force_sleep);
3152 sv_wait(&iclog->ic_prev->ic_write_wait,
3153 PSWP, &log->l_icloglock, s);
3159 atomic_inc(&iclog->ic_refcnt);
3160 xlog_state_switch_iclogs(log, iclog, 0);
3161 spin_unlock(&log->l_icloglock);
3162 if (xlog_state_release_iclog(log, iclog))
3163 return XFS_ERROR(EIO);
3166 spin_lock(&log->l_icloglock);
3169 if ((flags & XFS_LOG_SYNC) && /* sleep */
3171 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3173 * Don't wait on completion if we know that we've
3174 * gotten a log write error.
3176 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3177 spin_unlock(&log->l_icloglock);
3178 return XFS_ERROR(EIO);
3180 XFS_STATS_INC(xs_log_force_sleep);
3181 sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3183 * No need to grab the log lock here since we're
3184 * only deciding whether or not to return EIO
3185 * and the memory read should be atomic.
3187 if (iclog->ic_state & XLOG_STATE_IOERROR)
3188 return XFS_ERROR(EIO);
3192 } else { /* just return */
3193 spin_unlock(&log->l_icloglock);
3197 } while (iclog != log->l_iclog);
3199 spin_unlock(&log->l_icloglock);
3204 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3205 * about errors or whether the log was flushed or not. This is the normal
3206 * interface to use when trying to unpin items or move the log forward.
3216 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3218 xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3219 "error %d returned.", error);
3224 * Called when we want to mark the current iclog as being ready to sync to
3228 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3230 assert_spin_locked(&log->l_icloglock);
3232 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3233 xlog_state_switch_iclogs(log, iclog, 0);
3235 ASSERT(iclog->ic_state &
3236 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3241 /*****************************************************************************
3245 *****************************************************************************
3249 * Free a used ticket when its refcount falls to zero.
3253 xlog_ticket_t *ticket)
3255 ASSERT(atomic_read(&ticket->t_ref) > 0);
3256 if (atomic_dec_and_test(&ticket->t_ref)) {
3257 sv_destroy(&ticket->t_wait);
3258 kmem_zone_free(xfs_log_ticket_zone, ticket);
3264 xlog_ticket_t *ticket)
3266 ASSERT(atomic_read(&ticket->t_ref) > 0);
3267 atomic_inc(&ticket->t_ref);
3272 xfs_log_get_trans_ident(
3273 struct xfs_trans *tp)
3275 return tp->t_ticket->t_tid;
3279 * Allocate and initialise a new log ticket.
3290 struct xlog_ticket *tic;
3294 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3299 * Permanent reservations have up to 'cnt'-1 active log operations
3300 * in the log. A unit in this case is the amount of space for one
3301 * of these log operations. Normal reservations have a cnt of 1
3302 * and their unit amount is the total amount of space required.
3304 * The following lines of code account for non-transaction data
3305 * which occupy space in the on-disk log.
3307 * Normal form of a transaction is:
3308 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3309 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3311 * We need to account for all the leadup data and trailer data
3312 * around the transaction data.
3313 * And then we need to account for the worst case in terms of using
3315 * The worst case will happen if:
3316 * - the placement of the transaction happens to be such that the
3317 * roundoff is at its maximum
3318 * - the transaction data is synced before the commit record is synced
3319 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3320 * Therefore the commit record is in its own Log Record.
3321 * This can happen as the commit record is called with its
3322 * own region to xlog_write().
3323 * This then means that in the worst case, roundoff can happen for
3324 * the commit-rec as well.
3325 * The commit-rec is smaller than padding in this scenario and so it is
3326 * not added separately.
3329 /* for trans header */
3330 unit_bytes += sizeof(xlog_op_header_t);
3331 unit_bytes += sizeof(xfs_trans_header_t);
3334 unit_bytes += sizeof(xlog_op_header_t);
3337 * for LR headers - the space for data in an iclog is the size minus
3338 * the space used for the headers. If we use the iclog size, then we
3339 * undercalculate the number of headers required.
3341 * Furthermore - the addition of op headers for split-recs might
3342 * increase the space required enough to require more log and op
3343 * headers, so take that into account too.
3345 * IMPORTANT: This reservation makes the assumption that if this
3346 * transaction is the first in an iclog and hence has the LR headers
3347 * accounted to it, then the remaining space in the iclog is
3348 * exclusively for this transaction. i.e. if the transaction is larger
3349 * than the iclog, it will be the only thing in that iclog.
3350 * Fundamentally, this means we must pass the entire log vector to
3351 * xlog_write to guarantee this.
3353 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3354 num_headers = howmany(unit_bytes, iclog_space);
3356 /* for split-recs - ophdrs added when data split over LRs */
3357 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3359 /* add extra header reservations if we overrun */
3360 while (!num_headers ||
3361 howmany(unit_bytes, iclog_space) > num_headers) {
3362 unit_bytes += sizeof(xlog_op_header_t);
3365 unit_bytes += log->l_iclog_hsize * num_headers;
3367 /* for commit-rec LR header - note: padding will subsume the ophdr */
3368 unit_bytes += log->l_iclog_hsize;
3370 /* for roundoff padding for transaction data and one for commit record */
3371 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3372 log->l_mp->m_sb.sb_logsunit > 1) {
3373 /* log su roundoff */
3374 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3377 unit_bytes += 2*BBSIZE;
3380 atomic_set(&tic->t_ref, 1);
3381 INIT_LIST_HEAD(&tic->t_queue);
3382 tic->t_unit_res = unit_bytes;
3383 tic->t_curr_res = unit_bytes;
3386 tic->t_tid = random32();
3387 tic->t_clientid = client;
3388 tic->t_flags = XLOG_TIC_INITED;
3389 tic->t_trans_type = 0;
3390 if (xflags & XFS_LOG_PERM_RESERV)
3391 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3392 sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3394 xlog_tic_reset_res(tic);
3400 /******************************************************************************
3402 * Log debug routines
3404 ******************************************************************************
3408 * Make sure that the destination ptr is within the valid data region of
3409 * one of the iclogs. This uses backup pointers stored in a different
3410 * part of the log in case we trash the log structure.
3413 xlog_verify_dest_ptr(
3420 for (i = 0; i < log->l_iclog_bufs; i++) {
3421 if (ptr >= log->l_iclog_bak[i] &&
3422 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3427 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3431 xlog_verify_grant_head(xlog_t *log, int equals)
3433 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3435 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3437 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3439 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3440 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3442 } /* xlog_verify_grant_head */
3445 xlog_verify_grant_tail(
3448 xfs_lsn_t tail_lsn = log->l_tail_lsn;
3451 * Check to make sure the grant write head didn't just over lap the
3452 * tail. If the cycles are the same, we can't be overlapping.
3453 * Otherwise, make sure that the cycles differ by exactly one and
3454 * check the byte count.
3456 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
3457 ASSERT(log->l_grant_write_cycle - 1 == CYCLE_LSN(tail_lsn));
3458 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
3462 /* check if it will fit */
3464 xlog_verify_tail_lsn(xlog_t *log,
3465 xlog_in_core_t *iclog,
3470 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3472 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3473 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3474 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3476 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3478 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3479 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3481 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3482 if (blocks < BTOBB(iclog->ic_offset) + 1)
3483 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3485 } /* xlog_verify_tail_lsn */
3488 * Perform a number of checks on the iclog before writing to disk.
3490 * 1. Make sure the iclogs are still circular
3491 * 2. Make sure we have a good magic number
3492 * 3. Make sure we don't have magic numbers in the data
3493 * 4. Check fields of each log operation header for:
3494 * A. Valid client identifier
3495 * B. tid ptr value falls in valid ptr space (user space code)
3496 * C. Length in log record header is correct according to the
3497 * individual operation headers within record.
3498 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3499 * log, check the preceding blocks of the physical log to make sure all
3500 * the cycle numbers agree with the current cycle number.
3503 xlog_verify_iclog(xlog_t *log,
3504 xlog_in_core_t *iclog,
3508 xlog_op_header_t *ophead;
3509 xlog_in_core_t *icptr;
3510 xlog_in_core_2_t *xhdr;
3512 xfs_caddr_t base_ptr;
3513 __psint_t field_offset;
3515 int len, i, j, k, op_len;
3518 /* check validity of iclog pointers */
3519 spin_lock(&log->l_icloglock);
3520 icptr = log->l_iclog;
3521 for (i=0; i < log->l_iclog_bufs; i++) {
3523 xlog_panic("xlog_verify_iclog: invalid ptr");
3524 icptr = icptr->ic_next;
3526 if (icptr != log->l_iclog)
3527 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3528 spin_unlock(&log->l_icloglock);
3530 /* check log magic numbers */
3531 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3532 xlog_panic("xlog_verify_iclog: invalid magic num");
3534 ptr = (xfs_caddr_t) &iclog->ic_header;
3535 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3537 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3538 xlog_panic("xlog_verify_iclog: unexpected magic num");
3542 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3543 ptr = iclog->ic_datap;
3545 ophead = (xlog_op_header_t *)ptr;
3546 xhdr = iclog->ic_data;
3547 for (i = 0; i < len; i++) {
3548 ophead = (xlog_op_header_t *)ptr;
3550 /* clientid is only 1 byte */
3551 field_offset = (__psint_t)
3552 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3553 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3554 clientid = ophead->oh_clientid;
3556 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3557 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3558 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3559 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3560 clientid = xlog_get_client_id(
3561 xhdr[j].hic_xheader.xh_cycle_data[k]);
3563 clientid = xlog_get_client_id(
3564 iclog->ic_header.h_cycle_data[idx]);
3567 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3568 cmn_err(CE_WARN, "xlog_verify_iclog: "
3569 "invalid clientid %d op 0x%p offset 0x%lx",
3570 clientid, ophead, (unsigned long)field_offset);
3573 field_offset = (__psint_t)
3574 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3575 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3576 op_len = be32_to_cpu(ophead->oh_len);
3578 idx = BTOBBT((__psint_t)&ophead->oh_len -
3579 (__psint_t)iclog->ic_datap);
3580 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3581 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3582 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3583 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3585 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3588 ptr += sizeof(xlog_op_header_t) + op_len;
3590 } /* xlog_verify_iclog */
3594 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3600 xlog_in_core_t *iclog, *ic;
3602 iclog = log->l_iclog;
3603 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3605 * Mark all the incore logs IOERROR.
3606 * From now on, no log flushes will result.
3610 ic->ic_state = XLOG_STATE_IOERROR;
3612 } while (ic != iclog);
3616 * Return non-zero, if state transition has already happened.
3622 * This is called from xfs_force_shutdown, when we're forcibly
3623 * shutting down the filesystem, typically because of an IO error.
3624 * Our main objectives here are to make sure that:
3625 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3626 * parties to find out, 'atomically'.
3627 * b. those who're sleeping on log reservations, pinned objects and
3628 * other resources get woken up, and be told the bad news.
3629 * c. nothing new gets queued up after (a) and (b) are done.
3630 * d. if !logerror, flush the iclogs to disk, then seal them off
3633 * Note: for delayed logging the !logerror case needs to flush the regions
3634 * held in memory out to the iclogs before flushing them to disk. This needs
3635 * to be done before the log is marked as shutdown, otherwise the flush to the
3639 xfs_log_force_umount(
3640 struct xfs_mount *mp,
3650 * If this happens during log recovery, don't worry about
3651 * locking; the log isn't open for business yet.
3654 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3655 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3657 XFS_BUF_DONE(mp->m_sb_bp);
3662 * Somebody could've already done the hard work for us.
3663 * No need to get locks for this.
3665 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3666 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3672 * Flush the in memory commit item list before marking the log as
3673 * being shut down. We need to do it in this order to ensure all the
3674 * completed transactions are flushed to disk with the xfs_log_force()
3677 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3678 xlog_cil_force(log);
3681 * We must hold both the GRANT lock and the LOG lock,
3682 * before we mark the filesystem SHUTDOWN and wake
3683 * everybody up to tell the bad news.
3685 spin_lock(&log->l_icloglock);
3686 spin_lock(&log->l_grant_lock);
3687 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3689 XFS_BUF_DONE(mp->m_sb_bp);
3692 * This flag is sort of redundant because of the mount flag, but
3693 * it's good to maintain the separation between the log and the rest
3696 log->l_flags |= XLOG_IO_ERROR;
3699 * If we hit a log error, we want to mark all the iclogs IOERROR
3700 * while we're still holding the loglock.
3703 retval = xlog_state_ioerror(log);
3704 spin_unlock(&log->l_icloglock);
3707 * We don't want anybody waiting for log reservations after this. That
3708 * means we have to wake up everybody queued up on reserveq as well as
3709 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3710 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3711 * action is protected by the GRANTLOCK.
3713 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3714 sv_signal(&tic->t_wait);
3716 list_for_each_entry(tic, &log->l_writeq, t_queue)
3717 sv_signal(&tic->t_wait);
3718 spin_unlock(&log->l_grant_lock);
3720 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3723 * Force the incore logs to disk before shutting the
3724 * log down completely.
3726 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3728 spin_lock(&log->l_icloglock);
3729 retval = xlog_state_ioerror(log);
3730 spin_unlock(&log->l_icloglock);
3733 * Wake up everybody waiting on xfs_log_force.
3734 * Callback all log item committed functions as if the
3735 * log writes were completed.
3737 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3739 #ifdef XFSERRORDEBUG
3741 xlog_in_core_t *iclog;
3743 spin_lock(&log->l_icloglock);
3744 iclog = log->l_iclog;
3746 ASSERT(iclog->ic_callback == 0);
3747 iclog = iclog->ic_next;
3748 } while (iclog != log->l_iclog);
3749 spin_unlock(&log->l_icloglock);
3752 /* return non-zero if log IOERROR transition had already happened */
3757 xlog_iclogs_empty(xlog_t *log)
3759 xlog_in_core_t *iclog;
3761 iclog = log->l_iclog;
3763 /* endianness does not matter here, zero is zero in
3766 if (iclog->ic_header.h_num_logops)
3768 iclog = iclog->ic_next;
3769 } while (iclog != log->l_iclog);