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"
29 #include "xfs_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_error.h"
32 #include "xfs_log_priv.h"
33 #include "xfs_buf_item.h"
34 #include "xfs_bmap_btree.h"
35 #include "xfs_alloc_btree.h"
36 #include "xfs_ialloc_btree.h"
37 #include "xfs_log_recover.h"
38 #include "xfs_trans_priv.h"
39 #include "xfs_dir_sf.h"
40 #include "xfs_dir2_sf.h"
41 #include "xfs_attr_sf.h"
42 #include "xfs_dinode.h"
43 #include "xfs_inode.h"
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
52 /* Local miscellaneous function prototypes */
53 STATIC int xlog_bdstrat_cb(struct xfs_buf *);
54 STATIC int xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
55 xlog_in_core_t **, xfs_lsn_t *);
56 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
57 xfs_buftarg_t *log_target,
58 xfs_daddr_t blk_offset,
60 STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
61 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
62 STATIC void xlog_unalloc_log(xlog_t *log);
63 STATIC int xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
64 int nentries, xfs_log_ticket_t tic,
66 xlog_in_core_t **commit_iclog,
69 /* local state machine functions */
70 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
71 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
72 STATIC int xlog_state_get_iclog_space(xlog_t *log,
74 xlog_in_core_t **iclog,
75 xlog_ticket_t *ticket,
78 STATIC void xlog_state_put_ticket(xlog_t *log,
80 STATIC int xlog_state_release_iclog(xlog_t *log,
81 xlog_in_core_t *iclog);
82 STATIC void xlog_state_switch_iclogs(xlog_t *log,
83 xlog_in_core_t *iclog,
85 STATIC int xlog_state_sync(xlog_t *log,
89 STATIC int xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
90 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
92 /* local functions to manipulate grant head */
93 STATIC int xlog_grant_log_space(xlog_t *log,
95 STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
97 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
98 xlog_ticket_t *ticket);
99 STATIC int xlog_regrant_write_log_space(xlog_t *log,
100 xlog_ticket_t *ticket);
101 STATIC void xlog_ungrant_log_space(xlog_t *log,
102 xlog_ticket_t *ticket);
105 /* local ticket functions */
106 STATIC void xlog_state_ticket_alloc(xlog_t *log);
107 STATIC xlog_ticket_t *xlog_ticket_get(xlog_t *log,
112 STATIC void xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
115 STATIC void xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
116 STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
117 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
118 int count, boolean_t syncing);
119 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
122 #define xlog_verify_dest_ptr(a,b)
123 #define xlog_verify_grant_head(a,b)
124 #define xlog_verify_iclog(a,b,c,d)
125 #define xlog_verify_tail_lsn(a,b,c)
128 STATIC int xlog_iclogs_empty(xlog_t *log);
130 #if defined(XFS_LOG_TRACE)
132 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
136 if (!log->l_grant_trace) {
137 log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
138 if (!log->l_grant_trace)
141 /* ticket counts are 1 byte each */
142 cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
144 ktrace_enter(log->l_grant_trace,
146 (void *)log->l_reserve_headq,
147 (void *)log->l_write_headq,
148 (void *)((unsigned long)log->l_grant_reserve_cycle),
149 (void *)((unsigned long)log->l_grant_reserve_bytes),
150 (void *)((unsigned long)log->l_grant_write_cycle),
151 (void *)((unsigned long)log->l_grant_write_bytes),
152 (void *)((unsigned long)log->l_curr_cycle),
153 (void *)((unsigned long)log->l_curr_block),
154 (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
155 (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
157 (void *)((unsigned long)tic->t_trans_type),
159 (void *)((unsigned long)tic->t_curr_res),
160 (void *)((unsigned long)tic->t_unit_res));
164 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
166 if (!iclog->ic_trace)
167 iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
168 ktrace_enter(iclog->ic_trace,
169 (void *)((unsigned long)state),
170 (void *)((unsigned long)current_pid()),
171 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
173 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
174 (void *)NULL, (void *)NULL);
177 #define xlog_trace_loggrant(log,tic,string)
178 #define xlog_trace_iclog(iclog,state)
179 #endif /* XFS_LOG_TRACE */
183 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
187 tic->t_prev = (*qp)->t_prev;
188 (*qp)->t_prev->t_next = tic;
191 tic->t_prev = tic->t_next = tic;
195 tic->t_flags |= XLOG_TIC_IN_Q;
199 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
201 if (tic == tic->t_next) {
205 tic->t_next->t_prev = tic->t_prev;
206 tic->t_prev->t_next = tic->t_next;
209 tic->t_next = tic->t_prev = NULL;
210 tic->t_flags &= ~XLOG_TIC_IN_Q;
214 xlog_grant_sub_space(struct log *log, int bytes)
216 log->l_grant_write_bytes -= bytes;
217 if (log->l_grant_write_bytes < 0) {
218 log->l_grant_write_bytes += log->l_logsize;
219 log->l_grant_write_cycle--;
222 log->l_grant_reserve_bytes -= bytes;
223 if ((log)->l_grant_reserve_bytes < 0) {
224 log->l_grant_reserve_bytes += log->l_logsize;
225 log->l_grant_reserve_cycle--;
231 xlog_grant_add_space_write(struct log *log, int bytes)
233 log->l_grant_write_bytes += bytes;
234 if (log->l_grant_write_bytes > log->l_logsize) {
235 log->l_grant_write_bytes -= log->l_logsize;
236 log->l_grant_write_cycle++;
241 xlog_grant_add_space_reserve(struct log *log, int bytes)
243 log->l_grant_reserve_bytes += bytes;
244 if (log->l_grant_reserve_bytes > log->l_logsize) {
245 log->l_grant_reserve_bytes -= log->l_logsize;
246 log->l_grant_reserve_cycle++;
251 xlog_grant_add_space(struct log *log, int bytes)
253 xlog_grant_add_space_write(log, bytes);
254 xlog_grant_add_space_reserve(log, bytes);
261 * 1. currblock field gets updated at startup and after in-core logs
262 * marked as with WANT_SYNC.
266 * This routine is called when a user of a log manager ticket is done with
267 * the reservation. If the ticket was ever used, then a commit record for
268 * the associated transaction is written out as a log operation header with
269 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
270 * a given ticket. If the ticket was one with a permanent reservation, then
271 * a few operations are done differently. Permanent reservation tickets by
272 * default don't release the reservation. They just commit the current
273 * transaction with the belief that the reservation is still needed. A flag
274 * must be passed in before permanent reservations are actually released.
275 * When these type of tickets are not released, they need to be set into
276 * the inited state again. By doing this, a start record will be written
277 * out when the next write occurs.
280 xfs_log_done(xfs_mount_t *mp,
281 xfs_log_ticket_t xtic,
285 xlog_t *log = mp->m_log;
286 xlog_ticket_t *ticket = (xfs_log_ticket_t) xtic;
289 if (XLOG_FORCED_SHUTDOWN(log) ||
291 * If nothing was ever written, don't write out commit record.
292 * If we get an error, just continue and give back the log ticket.
294 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
295 (xlog_commit_record(mp, ticket,
296 (xlog_in_core_t **)iclog, &lsn)))) {
297 lsn = (xfs_lsn_t) -1;
298 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
299 flags |= XFS_LOG_REL_PERM_RESERV;
304 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
305 (flags & XFS_LOG_REL_PERM_RESERV)) {
307 * Release ticket if not permanent reservation or a specifc
308 * request has been made to release a permanent reservation.
310 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
311 xlog_ungrant_log_space(log, ticket);
312 xlog_state_put_ticket(log, ticket);
314 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
315 xlog_regrant_reserve_log_space(log, ticket);
318 /* If this ticket was a permanent reservation and we aren't
319 * trying to release it, reset the inited flags; so next time
320 * we write, a start record will be written out.
322 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
323 (flags & XFS_LOG_REL_PERM_RESERV) == 0)
324 ticket->t_flags |= XLOG_TIC_INITED;
331 * Force the in-core log to disk. If flags == XFS_LOG_SYNC,
332 * the force is done synchronously.
334 * Asynchronous forces are implemented by setting the WANT_SYNC
335 * bit in the appropriate in-core log and then returning.
337 * Synchronous forces are implemented with a semaphore. All callers
338 * to force a given lsn to disk will wait on a semaphore attached to the
339 * specific in-core log. When given in-core log finally completes its
340 * write to disk, that thread will wake up all threads waiting on the
350 xlog_t *log = mp->m_log;
354 log_flushed = &dummy;
356 ASSERT(flags & XFS_LOG_FORCE);
358 XFS_STATS_INC(xs_log_force);
360 if (log->l_flags & XLOG_IO_ERROR)
361 return XFS_ERROR(EIO);
363 return xlog_state_sync_all(log, flags, log_flushed);
365 return xlog_state_sync(log, lsn, flags, log_flushed);
366 } /* xfs_log_force */
369 * Attaches a new iclog I/O completion callback routine during
370 * transaction commit. If the log is in error state, a non-zero
371 * return code is handed back and the caller is responsible for
372 * executing the callback at an appropriate time.
375 xfs_log_notify(xfs_mount_t *mp, /* mount of partition */
376 void *iclog_hndl, /* iclog to hang callback off */
377 xfs_log_callback_t *cb)
379 xlog_t *log = mp->m_log;
380 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
385 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
387 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
388 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
390 *(iclog->ic_callback_tail) = cb;
391 iclog->ic_callback_tail = &(cb->cb_next);
393 LOG_UNLOCK(log, spl);
395 } /* xfs_log_notify */
398 xfs_log_release_iclog(xfs_mount_t *mp,
401 xlog_t *log = mp->m_log;
402 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
404 if (xlog_state_release_iclog(log, iclog)) {
405 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
413 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
414 * to the reservation.
415 * 2. Potentially, push buffers at tail of log to disk.
417 * Each reservation is going to reserve extra space for a log record header.
418 * When writes happen to the on-disk log, we don't subtract the length of the
419 * log record header from any reservation. By wasting space in each
420 * reservation, we prevent over allocation problems.
423 xfs_log_reserve(xfs_mount_t *mp,
426 xfs_log_ticket_t *ticket,
431 xlog_t *log = mp->m_log;
432 xlog_ticket_t *internal_ticket;
435 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
436 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
438 if (XLOG_FORCED_SHUTDOWN(log))
439 return XFS_ERROR(EIO);
441 XFS_STATS_INC(xs_try_logspace);
443 if (*ticket != NULL) {
444 ASSERT(flags & XFS_LOG_PERM_RESERV);
445 internal_ticket = (xlog_ticket_t *)*ticket;
446 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
447 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
448 retval = xlog_regrant_write_log_space(log, internal_ticket);
450 /* may sleep if need to allocate more tickets */
451 internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
453 internal_ticket->t_trans_type = t_type;
454 *ticket = internal_ticket;
455 xlog_trace_loggrant(log, internal_ticket,
456 (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
457 "xfs_log_reserve: create new ticket (permanent trans)" :
458 "xfs_log_reserve: create new ticket");
459 xlog_grant_push_ail(mp,
460 (internal_ticket->t_unit_res *
461 internal_ticket->t_cnt));
462 retval = xlog_grant_log_space(log, internal_ticket);
466 } /* xfs_log_reserve */
470 * Mount a log filesystem
472 * mp - ubiquitous xfs mount point structure
473 * log_target - buftarg of on-disk log device
474 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
475 * num_bblocks - Number of BBSIZE blocks in on-disk log
477 * Return error or zero.
480 xfs_log_mount(xfs_mount_t *mp,
481 xfs_buftarg_t *log_target,
482 xfs_daddr_t blk_offset,
485 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
486 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
489 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
491 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
494 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
497 * skip log recovery on a norecovery mount. pretend it all
500 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
502 vfs_t *vfsp = XFS_MTOVFS(mp);
503 int readonly = (vfsp->vfs_flag & VFS_RDONLY);
506 vfsp->vfs_flag &= ~VFS_RDONLY;
508 error = xlog_recover(mp->m_log);
511 vfsp->vfs_flag |= VFS_RDONLY;
513 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
514 xlog_unalloc_log(mp->m_log);
519 /* Normal transactions can now occur */
520 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
522 /* End mounting message in xfs_log_mount_finish */
524 } /* xfs_log_mount */
527 * Finish the recovery of the file system. This is separate from
528 * the xfs_log_mount() call, because it depends on the code in
529 * xfs_mountfs() to read in the root and real-time bitmap inodes
530 * between calling xfs_log_mount() and here.
532 * mp - ubiquitous xfs mount point structure
535 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
539 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
540 error = xlog_recover_finish(mp->m_log, mfsi_flags);
543 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
550 * Unmount processing for the log.
553 xfs_log_unmount(xfs_mount_t *mp)
557 error = xfs_log_unmount_write(mp);
558 xfs_log_unmount_dealloc(mp);
563 * Final log writes as part of unmount.
565 * Mark the filesystem clean as unmount happens. Note that during relocation
566 * this routine needs to be executed as part of source-bag while the
567 * deallocation must not be done until source-end.
571 * Unmount record used to have a string "Unmount filesystem--" in the
572 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
573 * We just write the magic number now since that particular field isn't
574 * currently architecture converted and "nUmount" is a bit foo.
575 * As far as I know, there weren't any dependencies on the old behaviour.
579 xfs_log_unmount_write(xfs_mount_t *mp)
581 xlog_t *log = mp->m_log;
582 xlog_in_core_t *iclog;
584 xlog_in_core_t *first_iclog;
586 xfs_log_iovec_t reg[1];
587 xfs_log_ticket_t tic = NULL;
592 /* the data section must be 32 bit size aligned */
596 __uint32_t pad2; /* may as well make it 64 bits */
597 } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
600 * Don't write out unmount record on read-only mounts.
601 * Or, if we are doing a forced umount (typically because of IO errors).
603 if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
606 xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
609 first_iclog = iclog = log->l_iclog;
611 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
612 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
613 ASSERT(iclog->ic_offset == 0);
615 iclog = iclog->ic_next;
616 } while (iclog != first_iclog);
618 if (! (XLOG_FORCED_SHUTDOWN(log))) {
619 reg[0].i_addr = (void*)&magic;
620 reg[0].i_len = sizeof(magic);
621 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_UNMOUNT);
623 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
625 /* remove inited flag */
626 ((xlog_ticket_t *)tic)->t_flags = 0;
627 error = xlog_write(mp, reg, 1, tic, &lsn,
628 NULL, XLOG_UNMOUNT_TRANS);
630 * At this point, we're umounting anyway,
631 * so there's no point in transitioning log state
632 * to IOERROR. Just continue...
637 xfs_fs_cmn_err(CE_ALERT, mp,
638 "xfs_log_unmount: unmount record failed");
643 iclog = log->l_iclog;
646 xlog_state_want_sync(log, iclog);
647 (void) xlog_state_release_iclog(log, iclog);
650 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
651 iclog->ic_state == XLOG_STATE_DIRTY)) {
652 if (!XLOG_FORCED_SHUTDOWN(log)) {
653 sv_wait(&iclog->ic_forcesema, PMEM,
654 &log->l_icloglock, s);
662 xlog_state_put_ticket(log, tic);
665 * We're already in forced_shutdown mode, couldn't
666 * even attempt to write out the unmount transaction.
668 * Go through the motions of sync'ing and releasing
669 * the iclog, even though no I/O will actually happen,
670 * we need to wait for other log I/O's that may already
671 * be in progress. Do this as a separate section of
672 * code so we'll know if we ever get stuck here that
673 * we're in this odd situation of trying to unmount
674 * a file system that went into forced_shutdown as
675 * the result of an unmount..
678 iclog = log->l_iclog;
682 xlog_state_want_sync(log, iclog);
683 (void) xlog_state_release_iclog(log, iclog);
687 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
688 || iclog->ic_state == XLOG_STATE_DIRTY
689 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
691 sv_wait(&iclog->ic_forcesema, PMEM,
692 &log->l_icloglock, s);
699 } /* xfs_log_unmount_write */
702 * Deallocate log structures for unmount/relocation.
705 xfs_log_unmount_dealloc(xfs_mount_t *mp)
707 xlog_unalloc_log(mp->m_log);
711 * Write region vectors to log. The write happens using the space reservation
712 * of the ticket (tic). It is not a requirement that all writes for a given
713 * transaction occur with one call to xfs_log_write().
716 xfs_log_write(xfs_mount_t * mp,
717 xfs_log_iovec_t reg[],
719 xfs_log_ticket_t tic,
720 xfs_lsn_t *start_lsn)
723 xlog_t *log = mp->m_log;
725 if (XLOG_FORCED_SHUTDOWN(log))
726 return XFS_ERROR(EIO);
728 if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
729 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
732 } /* xfs_log_write */
736 xfs_log_move_tail(xfs_mount_t *mp,
740 xlog_t *log = mp->m_log;
741 int need_bytes, free_bytes, cycle, bytes;
744 if (XLOG_FORCED_SHUTDOWN(log))
746 ASSERT(!XFS_FORCED_SHUTDOWN(mp));
749 /* needed since sync_lsn is 64 bits */
751 tail_lsn = log->l_last_sync_lsn;
757 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
761 log->l_tail_lsn = tail_lsn;
764 if ((tic = log->l_write_headq)) {
766 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
767 panic("Recovery problem");
769 cycle = log->l_grant_write_cycle;
770 bytes = log->l_grant_write_bytes;
771 free_bytes = xlog_space_left(log, cycle, bytes);
773 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
775 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
778 free_bytes -= tic->t_unit_res;
779 sv_signal(&tic->t_sema);
781 } while (tic != log->l_write_headq);
783 if ((tic = log->l_reserve_headq)) {
785 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
786 panic("Recovery problem");
788 cycle = log->l_grant_reserve_cycle;
789 bytes = log->l_grant_reserve_bytes;
790 free_bytes = xlog_space_left(log, cycle, bytes);
792 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
793 need_bytes = tic->t_unit_res*tic->t_cnt;
795 need_bytes = tic->t_unit_res;
796 if (free_bytes < need_bytes && tail_lsn != 1)
799 free_bytes -= need_bytes;
800 sv_signal(&tic->t_sema);
802 } while (tic != log->l_reserve_headq);
804 GRANT_UNLOCK(log, s);
805 } /* xfs_log_move_tail */
808 * Determine if we have a transaction that has gone to disk
809 * that needs to be covered. Log activity needs to be idle (no AIL and
810 * nothing in the iclogs). And, we need to be in the right state indicating
811 * something has gone out.
814 xfs_log_need_covered(xfs_mount_t *mp)
818 xlog_t *log = mp->m_log;
819 vfs_t *vfsp = XFS_MTOVFS(mp);
821 if (fs_frozen(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
822 (vfsp->vfs_flag & VFS_RDONLY))
826 if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
827 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
828 && !xfs_trans_first_ail(mp, &gen)
829 && xlog_iclogs_empty(log)) {
830 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
831 log->l_covered_state = XLOG_STATE_COVER_DONE;
833 ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
834 log->l_covered_state = XLOG_STATE_COVER_DONE2;
842 /******************************************************************************
846 ******************************************************************************
849 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
850 * The log manager must keep track of the last LR which was committed
851 * to disk. The lsn of this LR will become the new tail_lsn whenever
852 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
853 * the situation where stuff could be written into the log but nothing
854 * was ever in the AIL when asked. Eventually, we panic since the
855 * tail hits the head.
857 * We may be holding the log iclog lock upon entering this routine.
860 xlog_assign_tail_lsn(xfs_mount_t *mp)
864 xlog_t *log = mp->m_log;
866 tail_lsn = xfs_trans_tail_ail(mp);
869 log->l_tail_lsn = tail_lsn;
871 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
873 GRANT_UNLOCK(log, s);
876 } /* xlog_assign_tail_lsn */
880 * Return the space in the log between the tail and the head. The head
881 * is passed in the cycle/bytes formal parms. In the special case where
882 * the reserve head has wrapped passed the tail, this calculation is no
883 * longer valid. In this case, just return 0 which means there is no space
884 * in the log. This works for all places where this function is called
885 * with the reserve head. Of course, if the write head were to ever
886 * wrap the tail, we should blow up. Rather than catch this case here,
887 * we depend on other ASSERTions in other parts of the code. XXXmiken
889 * This code also handles the case where the reservation head is behind
890 * the tail. The details of this case are described below, but the end
891 * result is that we return the size of the log as the amount of space left.
894 xlog_space_left(xlog_t *log, int cycle, int bytes)
900 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
901 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
902 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
903 free_bytes = log->l_logsize - (bytes - tail_bytes);
904 } else if ((tail_cycle + 1) < cycle) {
906 } else if (tail_cycle < cycle) {
907 ASSERT(tail_cycle == (cycle - 1));
908 free_bytes = tail_bytes - bytes;
911 * The reservation head is behind the tail.
912 * In this case we just want to return the size of the
913 * log as the amount of space left.
915 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
916 "xlog_space_left: head behind tail\n"
917 " tail_cycle = %d, tail_bytes = %d\n"
918 " GH cycle = %d, GH bytes = %d",
919 tail_cycle, tail_bytes, cycle, bytes);
921 free_bytes = log->l_logsize;
924 } /* xlog_space_left */
928 * Log function which is called when an io completes.
930 * The log manager needs its own routine, in order to control what
931 * happens with the buffer after the write completes.
934 xlog_iodone(xfs_buf_t *bp)
936 xlog_in_core_t *iclog;
940 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
941 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
942 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
946 * Some versions of cpp barf on the recursive definition of
947 * ic_log -> hic_fields.ic_log and expand ic_log twice when
948 * it is passed through two macros. Workaround broken cpp.
953 * Race to shutdown the filesystem if we see an error.
955 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
956 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
957 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
959 xfs_force_shutdown(l->l_mp, XFS_LOG_IO_ERROR);
961 * This flag will be propagated to the trans-committed
962 * callback routines to let them know that the log-commit
965 aborted = XFS_LI_ABORTED;
966 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
967 aborted = XFS_LI_ABORTED;
969 xlog_state_done_syncing(iclog, aborted);
970 if (!(XFS_BUF_ISASYNC(bp))) {
972 * Corresponding psema() will be done in bwrite(). If we don't
973 * vsema() here, panic.
975 XFS_BUF_V_IODONESEMA(bp);
980 * The bdstrat callback function for log bufs. This gives us a central
981 * place to trap bufs in case we get hit by a log I/O error and need to
982 * shutdown. Actually, in practice, even when we didn't get a log error,
983 * we transition the iclogs to IOERROR state *after* flushing all existing
984 * iclogs to disk. This is because we don't want anymore new transactions to be
985 * started or completed afterwards.
988 xlog_bdstrat_cb(struct xfs_buf *bp)
990 xlog_in_core_t *iclog;
992 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
994 if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
995 /* note for irix bstrat will need struct bdevsw passed
996 * Fix the following macro if the code ever is merged
1002 xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1003 XFS_BUF_ERROR(bp, EIO);
1006 return (XFS_ERROR(EIO));
1012 * Return size of each in-core log record buffer.
1014 * Low memory machines only get 2 16KB buffers. We don't want to waste
1015 * memory here. However, all other machines get at least 2 32KB buffers.
1016 * The number is hard coded because we don't care about the minimum
1017 * memory size, just 32MB systems.
1019 * If the filesystem blocksize is too large, we may need to choose a
1020 * larger size since the directory code currently logs entire blocks.
1024 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1030 if (mp->m_logbufs <= 0) {
1031 if (xfs_physmem <= btoc(128*1024*1024)) {
1032 log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1033 } else if (xfs_physmem <= btoc(400*1024*1024)) {
1034 log->l_iclog_bufs = XLOG_MED_ICLOGS;
1035 } else { /* 256K with 32K bufs */
1036 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1039 log->l_iclog_bufs = mp->m_logbufs;
1043 * Buffer size passed in from mount system call.
1045 if (mp->m_logbsize > 0) {
1046 size = log->l_iclog_size = mp->m_logbsize;
1047 log->l_iclog_size_log = 0;
1049 log->l_iclog_size_log++;
1053 if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1054 /* # headers = size / 32K
1055 * one header holds cycles from 32K of data
1058 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1059 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1061 log->l_iclog_hsize = xhdrs << BBSHIFT;
1062 log->l_iclog_heads = xhdrs;
1064 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1065 log->l_iclog_hsize = BBSIZE;
1066 log->l_iclog_heads = 1;
1072 * Special case machines that have less than 32MB of memory.
1073 * All machines with more memory use 32KB buffers.
1075 if (xfs_physmem <= btoc(32*1024*1024)) {
1076 /* Don't change; min configuration */
1077 log->l_iclog_size = XLOG_RECORD_BSIZE; /* 16k */
1078 log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1080 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE; /* 32k */
1081 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1084 /* the default log size is 16k or 32k which is one header sector */
1085 log->l_iclog_hsize = BBSIZE;
1086 log->l_iclog_heads = 1;
1089 * For 16KB, we use 3 32KB buffers. For 32KB block sizes, we use
1090 * 4 32KB buffers. For 64KB block sizes, we use 8 32KB buffers.
1092 if (mp->m_sb.sb_blocksize >= 16*1024) {
1093 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1094 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1095 if (mp->m_logbufs <= 0) {
1096 switch (mp->m_sb.sb_blocksize) {
1097 case 16*1024: /* 16 KB */
1098 log->l_iclog_bufs = 3;
1100 case 32*1024: /* 32 KB */
1101 log->l_iclog_bufs = 4;
1103 case 64*1024: /* 64 KB */
1104 log->l_iclog_bufs = 8;
1107 xlog_panic("XFS: Invalid blocksize");
1113 done: /* are we being asked to make the sizes selected above visible? */
1114 if (mp->m_logbufs == 0)
1115 mp->m_logbufs = log->l_iclog_bufs;
1116 if (mp->m_logbsize == 0)
1117 mp->m_logbsize = log->l_iclog_size;
1118 } /* xlog_get_iclog_buffer_size */
1122 * This routine initializes some of the log structure for a given mount point.
1123 * Its primary purpose is to fill in enough, so recovery can occur. However,
1124 * some other stuff may be filled in too.
1127 xlog_alloc_log(xfs_mount_t *mp,
1128 xfs_buftarg_t *log_target,
1129 xfs_daddr_t blk_offset,
1133 xlog_rec_header_t *head;
1134 xlog_in_core_t **iclogp;
1135 xlog_in_core_t *iclog, *prev_iclog=NULL;
1140 log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1143 log->l_targ = log_target;
1144 log->l_logsize = BBTOB(num_bblks);
1145 log->l_logBBstart = blk_offset;
1146 log->l_logBBsize = num_bblks;
1147 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1148 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1150 log->l_prev_block = -1;
1151 ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1152 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1153 log->l_last_sync_lsn = log->l_tail_lsn;
1154 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1155 log->l_grant_reserve_cycle = 1;
1156 log->l_grant_write_cycle = 1;
1158 if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1159 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1160 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1161 /* for larger sector sizes, must have v2 or external log */
1162 ASSERT(log->l_sectbb_log == 0 ||
1163 log->l_logBBstart == 0 ||
1164 XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1165 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1167 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1169 xlog_get_iclog_buffer_size(mp, log);
1171 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1172 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1173 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1174 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1175 ASSERT(XFS_BUF_ISBUSY(bp));
1176 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1179 spinlock_init(&log->l_icloglock, "iclog");
1180 spinlock_init(&log->l_grant_lock, "grhead_iclog");
1181 initnsema(&log->l_flushsema, 0, "ic-flush");
1182 xlog_state_ticket_alloc(log); /* wait until after icloglock inited */
1184 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1185 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1187 iclogp = &log->l_iclog;
1189 * The amount of memory to allocate for the iclog structure is
1190 * rather funky due to the way the structure is defined. It is
1191 * done this way so that we can use different sizes for machines
1192 * with different amounts of memory. See the definition of
1193 * xlog_in_core_t in xfs_log_priv.h for details.
1195 iclogsize = log->l_iclog_size;
1196 ASSERT(log->l_iclog_size >= 4096);
1197 for (i=0; i < log->l_iclog_bufs; i++) {
1198 *iclogp = (xlog_in_core_t *)
1199 kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1201 iclog->hic_data = (xlog_in_core_2_t *)
1202 kmem_zalloc(iclogsize, KM_SLEEP);
1204 iclog->ic_prev = prev_iclog;
1206 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1208 head = &iclog->ic_header;
1209 memset(head, 0, sizeof(xlog_rec_header_t));
1210 INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1211 INT_SET(head->h_version, ARCH_CONVERT,
1212 XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1213 INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1215 INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1216 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1218 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1219 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1220 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1221 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1224 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1225 iclog->ic_state = XLOG_STATE_ACTIVE;
1226 iclog->ic_log = log;
1227 iclog->ic_callback_tail = &(iclog->ic_callback);
1228 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1230 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1231 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1232 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1233 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1235 iclogp = &iclog->ic_next;
1237 *iclogp = log->l_iclog; /* complete ring */
1238 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1241 } /* xlog_alloc_log */
1245 * Write out the commit record of a transaction associated with the given
1246 * ticket. Return the lsn of the commit record.
1249 xlog_commit_record(xfs_mount_t *mp,
1250 xlog_ticket_t *ticket,
1251 xlog_in_core_t **iclog,
1252 xfs_lsn_t *commitlsnp)
1255 xfs_log_iovec_t reg[1];
1257 reg[0].i_addr = NULL;
1259 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_COMMIT);
1261 ASSERT_ALWAYS(iclog);
1262 if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1263 iclog, XLOG_COMMIT_TRANS))) {
1264 xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
1267 } /* xlog_commit_record */
1271 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1272 * log space. This code pushes on the lsn which would supposedly free up
1273 * the 25% which we want to leave free. We may need to adopt a policy which
1274 * pushes on an lsn which is further along in the log once we reach the high
1275 * water mark. In this manner, we would be creating a low water mark.
1278 xlog_grant_push_ail(xfs_mount_t *mp,
1281 xlog_t *log = mp->m_log; /* pointer to the log */
1282 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1283 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1284 int free_blocks; /* free blocks left to write to */
1285 int free_bytes; /* free bytes left to write to */
1286 int threshold_block; /* block in lsn we'd like to be at */
1287 int threshold_cycle; /* lsn cycle we'd like to be at */
1291 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1293 s = GRANT_LOCK(log);
1294 free_bytes = xlog_space_left(log,
1295 log->l_grant_reserve_cycle,
1296 log->l_grant_reserve_bytes);
1297 tail_lsn = log->l_tail_lsn;
1298 free_blocks = BTOBBT(free_bytes);
1301 * Set the threshold for the minimum number of free blocks in the
1302 * log to the maximum of what the caller needs, one quarter of the
1303 * log, and 256 blocks.
1305 free_threshold = BTOBB(need_bytes);
1306 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1307 free_threshold = MAX(free_threshold, 256);
1308 if (free_blocks < free_threshold) {
1309 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1310 threshold_cycle = CYCLE_LSN(tail_lsn);
1311 if (threshold_block >= log->l_logBBsize) {
1312 threshold_block -= log->l_logBBsize;
1313 threshold_cycle += 1;
1315 ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1318 /* Don't pass in an lsn greater than the lsn of the last
1319 * log record known to be on disk.
1321 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1322 threshold_lsn = log->l_last_sync_lsn;
1324 GRANT_UNLOCK(log, s);
1327 * Get the transaction layer to kick the dirty buffers out to
1328 * disk asynchronously. No point in trying to do this if
1329 * the filesystem is shutting down.
1331 if (threshold_lsn &&
1332 !XLOG_FORCED_SHUTDOWN(log))
1333 xfs_trans_push_ail(mp, threshold_lsn);
1334 } /* xlog_grant_push_ail */
1338 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1339 * fashion. Previously, we should have moved the current iclog
1340 * ptr in the log to point to the next available iclog. This allows further
1341 * write to continue while this code syncs out an iclog ready to go.
1342 * Before an in-core log can be written out, the data section must be scanned
1343 * to save away the 1st word of each BBSIZE block into the header. We replace
1344 * it with the current cycle count. Each BBSIZE block is tagged with the
1345 * cycle count because there in an implicit assumption that drives will
1346 * guarantee that entire 512 byte blocks get written at once. In other words,
1347 * we can't have part of a 512 byte block written and part not written. By
1348 * tagging each block, we will know which blocks are valid when recovering
1349 * after an unclean shutdown.
1351 * This routine is single threaded on the iclog. No other thread can be in
1352 * this routine with the same iclog. Changing contents of iclog can there-
1353 * fore be done without grabbing the state machine lock. Updating the global
1354 * log will require grabbing the lock though.
1356 * The entire log manager uses a logical block numbering scheme. Only
1357 * log_sync (and then only bwrite()) know about the fact that the log may
1358 * not start with block zero on a given device. The log block start offset
1359 * is added immediately before calling bwrite().
1363 xlog_sync(xlog_t *log,
1364 xlog_in_core_t *iclog)
1366 xfs_caddr_t dptr; /* pointer to byte sized element */
1369 uint count; /* byte count of bwrite */
1370 uint count_init; /* initial count before roundup */
1371 int roundoff; /* roundoff to BB or stripe */
1372 int split = 0; /* split write into two regions */
1375 int v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1377 XFS_STATS_INC(xs_log_writes);
1378 ASSERT(iclog->ic_refcnt == 0);
1380 /* Add for LR header */
1381 count_init = log->l_iclog_hsize + iclog->ic_offset;
1383 /* Round out the log write size */
1384 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1385 /* we have a v2 stripe unit to use */
1386 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1388 count = BBTOB(BTOBB(count_init));
1390 roundoff = count - count_init;
1391 ASSERT(roundoff >= 0);
1392 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1393 roundoff < log->l_mp->m_sb.sb_logsunit)
1395 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1396 roundoff < BBTOB(1)));
1398 /* move grant heads by roundoff in sync */
1399 s = GRANT_LOCK(log);
1400 xlog_grant_add_space(log, roundoff);
1401 GRANT_UNLOCK(log, s);
1403 /* put cycle number in every block */
1404 xlog_pack_data(log, iclog, roundoff);
1406 /* real byte length */
1408 INT_SET(iclog->ic_header.h_len,
1410 iclog->ic_offset + roundoff);
1412 INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1415 /* put ops count in correct order */
1416 ops = iclog->ic_header.h_num_logops;
1417 INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1420 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1421 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1422 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1424 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1426 /* Do we need to split this write into 2 parts? */
1427 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1428 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1429 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1430 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1432 iclog->ic_bwritecnt = 1;
1434 XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1435 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1439 * Do an ordered write for the log block.
1441 * It may not be needed to flush the first split block in the log wrap
1442 * case, but do it anyways to be safe -AK
1444 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1445 XFS_BUF_ORDERED(bp);
1447 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1448 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1450 xlog_verify_iclog(log, iclog, count, B_TRUE);
1452 /* account for log which doesn't start at block #0 */
1453 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1455 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1460 if ((error = XFS_bwrite(bp))) {
1461 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1466 bp = iclog->ic_log->l_xbuf;
1467 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1469 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1470 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1471 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1472 (__psint_t)count), split);
1473 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1476 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1477 XFS_BUF_ORDERED(bp);
1478 dptr = XFS_BUF_PTR(bp);
1480 * Bump the cycle numbers at the start of each block
1481 * since this part of the buffer is at the start of
1482 * a new cycle. Watch out for the header magic number
1485 for (i=0; i<split; i += BBSIZE) {
1486 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1487 if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1488 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1492 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1493 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1495 /* account for internal log which does't start at block #0 */
1496 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1498 if ((error = XFS_bwrite(bp))) {
1499 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1500 bp, XFS_BUF_ADDR(bp));
1509 * Unallocate a log structure
1512 xlog_unalloc_log(xlog_t *log)
1514 xlog_in_core_t *iclog, *next_iclog;
1515 xlog_ticket_t *tic, *next_tic;
1519 iclog = log->l_iclog;
1520 for (i=0; i<log->l_iclog_bufs; i++) {
1521 sv_destroy(&iclog->ic_forcesema);
1522 sv_destroy(&iclog->ic_writesema);
1523 xfs_buf_free(iclog->ic_bp);
1524 #ifdef XFS_LOG_TRACE
1525 if (iclog->ic_trace != NULL) {
1526 ktrace_free(iclog->ic_trace);
1529 next_iclog = iclog->ic_next;
1530 kmem_free(iclog->hic_data, log->l_iclog_size);
1531 kmem_free(iclog, sizeof(xlog_in_core_t));
1534 freesema(&log->l_flushsema);
1535 spinlock_destroy(&log->l_icloglock);
1536 spinlock_destroy(&log->l_grant_lock);
1538 /* XXXsup take a look at this again. */
1539 if ((log->l_ticket_cnt != log->l_ticket_tcnt) &&
1540 !XLOG_FORCED_SHUTDOWN(log)) {
1541 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1542 "xlog_unalloc_log: (cnt: %d, total: %d)",
1543 log->l_ticket_cnt, log->l_ticket_tcnt);
1544 /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1547 tic = log->l_unmount_free;
1549 next_tic = tic->t_next;
1550 kmem_free(tic, NBPP);
1554 xfs_buf_free(log->l_xbuf);
1555 #ifdef XFS_LOG_TRACE
1556 if (log->l_trace != NULL) {
1557 ktrace_free(log->l_trace);
1559 if (log->l_grant_trace != NULL) {
1560 ktrace_free(log->l_grant_trace);
1563 log->l_mp->m_log = NULL;
1564 kmem_free(log, sizeof(xlog_t));
1565 } /* xlog_unalloc_log */
1568 * Update counters atomically now that memcpy is done.
1572 xlog_state_finish_copy(xlog_t *log,
1573 xlog_in_core_t *iclog,
1581 iclog->ic_header.h_num_logops += record_cnt;
1582 iclog->ic_offset += copy_bytes;
1585 } /* xlog_state_finish_copy */
1591 * print out info relating to regions written which consume
1594 #if defined(XFS_LOG_RES_DEBUG)
1596 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1599 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1601 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1602 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1623 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1666 xfs_fs_cmn_err(CE_WARN, mp,
1667 "xfs_log_write: reservation summary:\n"
1668 " trans type = %s (%u)\n"
1669 " unit res = %d bytes\n"
1670 " current res = %d bytes\n"
1671 " total reg = %u bytes (o/flow = %u bytes)\n"
1672 " ophdrs = %u (ophdr space = %u bytes)\n"
1673 " ophdr + reg = %u bytes\n"
1674 " num regions = %u\n",
1675 ((ticket->t_trans_type <= 0 ||
1676 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1677 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1678 ticket->t_trans_type,
1681 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1682 ticket->t_res_num_ophdrs, ophdr_spc,
1683 ticket->t_res_arr_sum +
1684 ticket->t_res_o_flow + ophdr_spc,
1687 for (i = 0; i < ticket->t_res_num; i++) {
1688 uint r_type = ticket->t_res_arr[i].r_type;
1690 "region[%u]: %s - %u bytes\n",
1692 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1693 "bad-rtype" : res_type_str[r_type-1]),
1694 ticket->t_res_arr[i].r_len);
1698 #define xlog_print_tic_res(mp, ticket)
1702 * Write some region out to in-core log
1704 * This will be called when writing externally provided regions or when
1705 * writing out a commit record for a given transaction.
1707 * General algorithm:
1708 * 1. Find total length of this write. This may include adding to the
1709 * lengths passed in.
1710 * 2. Check whether we violate the tickets reservation.
1711 * 3. While writing to this iclog
1712 * A. Reserve as much space in this iclog as can get
1713 * B. If this is first write, save away start lsn
1714 * C. While writing this region:
1715 * 1. If first write of transaction, write start record
1716 * 2. Write log operation header (header per region)
1717 * 3. Find out if we can fit entire region into this iclog
1718 * 4. Potentially, verify destination memcpy ptr
1719 * 5. Memcpy (partial) region
1720 * 6. If partial copy, release iclog; otherwise, continue
1721 * copying more regions into current iclog
1722 * 4. Mark want sync bit (in simulation mode)
1723 * 5. Release iclog for potential flush to on-disk log.
1726 * 1. Panic if reservation is overrun. This should never happen since
1727 * reservation amounts are generated internal to the filesystem.
1729 * 1. Tickets are single threaded data structures.
1730 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1731 * syncing routine. When a single log_write region needs to span
1732 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1733 * on all log operation writes which don't contain the end of the
1734 * region. The XLOG_END_TRANS bit is used for the in-core log
1735 * operation which contains the end of the continued log_write region.
1736 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1737 * we don't really know exactly how much space will be used. As a result,
1738 * we don't update ic_offset until the end when we know exactly how many
1739 * bytes have been written out.
1742 xlog_write(xfs_mount_t * mp,
1743 xfs_log_iovec_t reg[],
1745 xfs_log_ticket_t tic,
1746 xfs_lsn_t *start_lsn,
1747 xlog_in_core_t **commit_iclog,
1750 xlog_t *log = mp->m_log;
1751 xlog_ticket_t *ticket = (xlog_ticket_t *)tic;
1752 xlog_op_header_t *logop_head; /* ptr to log operation header */
1753 xlog_in_core_t *iclog; /* ptr to current in-core log */
1754 __psint_t ptr; /* copy address into data region */
1755 int len; /* # xlog_write() bytes 2 still copy */
1756 int index; /* region index currently copying */
1757 int log_offset; /* offset (from 0) into data region */
1758 int start_rec_copy; /* # bytes to copy for start record */
1759 int partial_copy; /* did we split a region? */
1760 int partial_copy_len;/* # bytes copied if split region */
1761 int need_copy; /* # bytes need to memcpy this region */
1762 int copy_len; /* # bytes actually memcpy'ing */
1763 int copy_off; /* # bytes from entry start */
1764 int contwr; /* continued write of in-core log? */
1766 int record_cnt = 0, data_cnt = 0;
1768 partial_copy_len = partial_copy = 0;
1770 /* Calculate potential maximum space. Each region gets its own
1771 * xlog_op_header_t and may need to be double word aligned.
1774 if (ticket->t_flags & XLOG_TIC_INITED) { /* acct for start rec of xact */
1775 len += sizeof(xlog_op_header_t);
1776 XLOG_TIC_ADD_OPHDR(ticket);
1779 for (index = 0; index < nentries; index++) {
1780 len += sizeof(xlog_op_header_t); /* each region gets >= 1 */
1781 XLOG_TIC_ADD_OPHDR(ticket);
1782 len += reg[index].i_len;
1783 XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1785 contwr = *start_lsn = 0;
1787 if (ticket->t_curr_res < len) {
1788 xlog_print_tic_res(mp, ticket);
1791 "xfs_log_write: reservation ran out. Need to up reservation");
1793 /* Customer configurable panic */
1794 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1795 "xfs_log_write: reservation ran out. Need to up reservation");
1796 /* If we did not panic, shutdown the filesystem */
1797 xfs_force_shutdown(mp, XFS_CORRUPT_INCORE);
1800 ticket->t_curr_res -= len;
1802 for (index = 0; index < nentries; ) {
1803 if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1804 &contwr, &log_offset)))
1807 ASSERT(log_offset <= iclog->ic_size - 1);
1808 ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1810 /* start_lsn is the first lsn written to. That's all we need. */
1812 *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1814 /* This loop writes out as many regions as can fit in the amount
1815 * of space which was allocated by xlog_state_get_iclog_space().
1817 while (index < nentries) {
1818 ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1819 ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1822 /* If first write for transaction, insert start record.
1823 * We can't be trying to commit if we are inited. We can't
1824 * have any "partial_copy" if we are inited.
1826 if (ticket->t_flags & XLOG_TIC_INITED) {
1827 logop_head = (xlog_op_header_t *)ptr;
1828 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1829 logop_head->oh_clientid = ticket->t_clientid;
1830 logop_head->oh_len = 0;
1831 logop_head->oh_flags = XLOG_START_TRANS;
1832 logop_head->oh_res2 = 0;
1833 ticket->t_flags &= ~XLOG_TIC_INITED; /* clear bit */
1836 start_rec_copy = sizeof(xlog_op_header_t);
1837 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1840 /* Copy log operation header directly into data section */
1841 logop_head = (xlog_op_header_t *)ptr;
1842 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1843 logop_head->oh_clientid = ticket->t_clientid;
1844 logop_head->oh_res2 = 0;
1846 /* header copied directly */
1847 xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1849 /* are we copying a commit or unmount record? */
1850 logop_head->oh_flags = flags;
1853 * We've seen logs corrupted with bad transaction client
1854 * ids. This makes sure that XFS doesn't generate them on.
1855 * Turn this into an EIO and shut down the filesystem.
1857 switch (logop_head->oh_clientid) {
1858 case XFS_TRANSACTION:
1863 xfs_fs_cmn_err(CE_WARN, mp,
1864 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1865 logop_head->oh_clientid, tic);
1866 return XFS_ERROR(EIO);
1869 /* Partial write last time? => (partial_copy != 0)
1870 * need_copy is the amount we'd like to copy if everything could
1871 * fit in the current memcpy.
1873 need_copy = reg[index].i_len - partial_copy_len;
1875 copy_off = partial_copy_len;
1876 if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1877 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1879 logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1880 partial_copy_len = partial_copy = 0;
1881 } else { /* partial write */
1882 copy_len = iclog->ic_size - log_offset;
1883 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1884 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1886 logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1887 partial_copy_len += copy_len;
1889 len += sizeof(xlog_op_header_t); /* from splitting of region */
1890 /* account for new log op header */
1891 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1892 XLOG_TIC_ADD_OPHDR(ticket);
1894 xlog_verify_dest_ptr(log, ptr);
1897 ASSERT(copy_len >= 0);
1898 memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1899 xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1901 /* make copy_len total bytes copied, including headers */
1902 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1904 data_cnt += contwr ? copy_len : 0;
1905 if (partial_copy) { /* copied partial region */
1906 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1907 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1908 record_cnt = data_cnt = 0;
1909 if ((error = xlog_state_release_iclog(log, iclog)))
1911 break; /* don't increment index */
1912 } else { /* copied entire region */
1914 partial_copy_len = partial_copy = 0;
1916 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1917 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1918 record_cnt = data_cnt = 0;
1919 xlog_state_want_sync(log, iclog);
1921 ASSERT(flags & XLOG_COMMIT_TRANS);
1922 *commit_iclog = iclog;
1923 } else if ((error = xlog_state_release_iclog(log, iclog)))
1925 if (index == nentries)
1926 return 0; /* we are done */
1930 } /* if (partial_copy) */
1931 } /* while (index < nentries) */
1932 } /* for (index = 0; index < nentries; ) */
1935 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1937 ASSERT(flags & XLOG_COMMIT_TRANS);
1938 *commit_iclog = iclog;
1941 return (xlog_state_release_iclog(log, iclog));
1945 /*****************************************************************************
1947 * State Machine functions
1949 *****************************************************************************
1952 /* Clean iclogs starting from the head. This ordering must be
1953 * maintained, so an iclog doesn't become ACTIVE beyond one that
1954 * is SYNCING. This is also required to maintain the notion that we use
1955 * a counting semaphore to hold off would be writers to the log when every
1956 * iclog is trying to sync to disk.
1958 * State Change: DIRTY -> ACTIVE
1961 xlog_state_clean_log(xlog_t *log)
1963 xlog_in_core_t *iclog;
1966 iclog = log->l_iclog;
1968 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1969 iclog->ic_state = XLOG_STATE_ACTIVE;
1970 iclog->ic_offset = 0;
1971 iclog->ic_callback = NULL; /* don't need to free */
1973 * If the number of ops in this iclog indicate it just
1974 * contains the dummy transaction, we can
1975 * change state into IDLE (the second time around).
1976 * Otherwise we should change the state into
1978 * We don't need to cover the dummy.
1981 (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1985 * We have two dirty iclogs so start over
1986 * This could also be num of ops indicates
1987 * this is not the dummy going out.
1991 iclog->ic_header.h_num_logops = 0;
1992 memset(iclog->ic_header.h_cycle_data, 0,
1993 sizeof(iclog->ic_header.h_cycle_data));
1994 iclog->ic_header.h_lsn = 0;
1995 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1998 break; /* stop cleaning */
1999 iclog = iclog->ic_next;
2000 } while (iclog != log->l_iclog);
2002 /* log is locked when we are called */
2004 * Change state for the dummy log recording.
2005 * We usually go to NEED. But we go to NEED2 if the changed indicates
2006 * we are done writing the dummy record.
2007 * If we are done with the second dummy recored (DONE2), then
2011 switch (log->l_covered_state) {
2012 case XLOG_STATE_COVER_IDLE:
2013 case XLOG_STATE_COVER_NEED:
2014 case XLOG_STATE_COVER_NEED2:
2015 log->l_covered_state = XLOG_STATE_COVER_NEED;
2018 case XLOG_STATE_COVER_DONE:
2020 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2022 log->l_covered_state = XLOG_STATE_COVER_NEED;
2025 case XLOG_STATE_COVER_DONE2:
2027 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2029 log->l_covered_state = XLOG_STATE_COVER_NEED;
2036 } /* xlog_state_clean_log */
2039 xlog_get_lowest_lsn(
2042 xlog_in_core_t *lsn_log;
2043 xfs_lsn_t lowest_lsn, lsn;
2045 lsn_log = log->l_iclog;
2048 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2049 lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2050 if ((lsn && !lowest_lsn) ||
2051 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2055 lsn_log = lsn_log->ic_next;
2056 } while (lsn_log != log->l_iclog);
2062 xlog_state_do_callback(
2065 xlog_in_core_t *ciclog)
2067 xlog_in_core_t *iclog;
2068 xlog_in_core_t *first_iclog; /* used to know when we've
2069 * processed all iclogs once */
2070 xfs_log_callback_t *cb, *cb_next;
2072 xfs_lsn_t lowest_lsn;
2073 int ioerrors; /* counter: iclogs with errors */
2074 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2075 int funcdidcallbacks; /* flag: function did callbacks */
2076 int repeats; /* for issuing console warnings if
2077 * looping too many times */
2081 first_iclog = iclog = log->l_iclog;
2083 funcdidcallbacks = 0;
2088 * Scan all iclogs starting with the one pointed to by the
2089 * log. Reset this starting point each time the log is
2090 * unlocked (during callbacks).
2092 * Keep looping through iclogs until one full pass is made
2093 * without running any callbacks.
2095 first_iclog = log->l_iclog;
2096 iclog = log->l_iclog;
2097 loopdidcallbacks = 0;
2102 /* skip all iclogs in the ACTIVE & DIRTY states */
2103 if (iclog->ic_state &
2104 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2105 iclog = iclog->ic_next;
2110 * Between marking a filesystem SHUTDOWN and stopping
2111 * the log, we do flush all iclogs to disk (if there
2112 * wasn't a log I/O error). So, we do want things to
2113 * go smoothly in case of just a SHUTDOWN w/o a
2116 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2118 * Can only perform callbacks in order. Since
2119 * this iclog is not in the DONE_SYNC/
2120 * DO_CALLBACK state, we skip the rest and
2121 * just try to clean up. If we set our iclog
2122 * to DO_CALLBACK, we will not process it when
2123 * we retry since a previous iclog is in the
2124 * CALLBACK and the state cannot change since
2125 * we are holding the LOG_LOCK.
2127 if (!(iclog->ic_state &
2128 (XLOG_STATE_DONE_SYNC |
2129 XLOG_STATE_DO_CALLBACK))) {
2130 if (ciclog && (ciclog->ic_state ==
2131 XLOG_STATE_DONE_SYNC)) {
2132 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2137 * We now have an iclog that is in either the
2138 * DO_CALLBACK or DONE_SYNC states. The other
2139 * states (WANT_SYNC, SYNCING, or CALLBACK were
2140 * caught by the above if and are going to
2141 * clean (i.e. we aren't doing their callbacks)
2146 * We will do one more check here to see if we
2147 * have chased our tail around.
2150 lowest_lsn = xlog_get_lowest_lsn(log);
2154 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2156 iclog = iclog->ic_next;
2157 continue; /* Leave this iclog for
2161 iclog->ic_state = XLOG_STATE_CALLBACK;
2165 /* l_last_sync_lsn field protected by
2166 * GRANT_LOCK. Don't worry about iclog's lsn.
2167 * No one else can be here except us.
2169 s = GRANT_LOCK(log);
2171 log->l_last_sync_lsn,
2172 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2174 log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2175 GRANT_UNLOCK(log, s);
2178 * Keep processing entries in the callback list
2179 * until we come around and it is empty. We
2180 * need to atomically see that the list is
2181 * empty and change the state to DIRTY so that
2182 * we don't miss any more callbacks being added.
2188 cb = iclog->ic_callback;
2191 iclog->ic_callback_tail = &(iclog->ic_callback);
2192 iclog->ic_callback = NULL;
2195 /* perform callbacks in the order given */
2196 for (; cb != 0; cb = cb_next) {
2197 cb_next = cb->cb_next;
2198 cb->cb_func(cb->cb_arg, aborted);
2201 cb = iclog->ic_callback;
2207 ASSERT(iclog->ic_callback == 0);
2208 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2209 iclog->ic_state = XLOG_STATE_DIRTY;
2212 * Transition from DIRTY to ACTIVE if applicable.
2213 * NOP if STATE_IOERROR.
2215 xlog_state_clean_log(log);
2217 /* wake up threads waiting in xfs_log_force() */
2218 sv_broadcast(&iclog->ic_forcesema);
2220 iclog = iclog->ic_next;
2221 } while (first_iclog != iclog);
2222 if (repeats && (repeats % 10) == 0) {
2223 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2224 "xlog_state_do_callback: looping %d", repeats);
2226 } while (!ioerrors && loopdidcallbacks);
2229 * make one last gasp attempt to see if iclogs are being left in
2233 if (funcdidcallbacks) {
2234 first_iclog = iclog = log->l_iclog;
2236 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2238 * Terminate the loop if iclogs are found in states
2239 * which will cause other threads to clean up iclogs.
2241 * SYNCING - i/o completion will go through logs
2242 * DONE_SYNC - interrupt thread should be waiting for
2244 * IOERROR - give up hope all ye who enter here
2246 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2247 iclog->ic_state == XLOG_STATE_SYNCING ||
2248 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2249 iclog->ic_state == XLOG_STATE_IOERROR )
2251 iclog = iclog->ic_next;
2252 } while (first_iclog != iclog);
2256 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2257 flushcnt = log->l_flushcnt;
2258 log->l_flushcnt = 0;
2262 vsema(&log->l_flushsema);
2263 } /* xlog_state_do_callback */
2267 * Finish transitioning this iclog to the dirty state.
2269 * Make sure that we completely execute this routine only when this is
2270 * the last call to the iclog. There is a good chance that iclog flushes,
2271 * when we reach the end of the physical log, get turned into 2 separate
2272 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2273 * routine. By using the reference count bwritecnt, we guarantee that only
2274 * the second completion goes through.
2276 * Callbacks could take time, so they are done outside the scope of the
2277 * global state machine log lock. Assume that the calls to cvsema won't
2278 * take a long time. At least we know it won't sleep.
2281 xlog_state_done_syncing(
2282 xlog_in_core_t *iclog,
2285 xlog_t *log = iclog->ic_log;
2290 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2291 iclog->ic_state == XLOG_STATE_IOERROR);
2292 ASSERT(iclog->ic_refcnt == 0);
2293 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2297 * If we got an error, either on the first buffer, or in the case of
2298 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2299 * and none should ever be attempted to be written to disk
2302 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2303 if (--iclog->ic_bwritecnt == 1) {
2307 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2311 * Someone could be sleeping prior to writing out the next
2312 * iclog buffer, we wake them all, one will get to do the
2313 * I/O, the others get to wait for the result.
2315 sv_broadcast(&iclog->ic_writesema);
2317 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2318 } /* xlog_state_done_syncing */
2322 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2323 * sleep. The flush semaphore is set to the number of in-core buffers and
2324 * decremented around disk syncing. Therefore, if all buffers are syncing,
2325 * this semaphore will cause new writes to sleep until a sync completes.
2326 * Otherwise, this code just does p() followed by v(). This approximates
2327 * a sleep/wakeup except we can't race.
2329 * The in-core logs are used in a circular fashion. They are not used
2330 * out-of-order even when an iclog past the head is free.
2333 * * log_offset where xlog_write() can start writing into the in-core
2335 * * in-core log pointer to which xlog_write() should write.
2336 * * boolean indicating this is a continued write to an in-core log.
2337 * If this is the last write, then the in-core log's offset field
2338 * needs to be incremented, depending on the amount of data which
2342 xlog_state_get_iclog_space(xlog_t *log,
2344 xlog_in_core_t **iclogp,
2345 xlog_ticket_t *ticket,
2346 int *continued_write,
2351 xlog_rec_header_t *head;
2352 xlog_in_core_t *iclog;
2357 if (XLOG_FORCED_SHUTDOWN(log)) {
2359 return XFS_ERROR(EIO);
2362 iclog = log->l_iclog;
2363 if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2366 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2367 XFS_STATS_INC(xs_log_noiclogs);
2368 /* Ensure that log writes happen */
2369 psema(&log->l_flushsema, PINOD);
2372 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2373 head = &iclog->ic_header;
2375 iclog->ic_refcnt++; /* prevents sync */
2376 log_offset = iclog->ic_offset;
2378 /* On the 1st write to an iclog, figure out lsn. This works
2379 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2380 * committing to. If the offset is set, that's how many blocks
2383 if (log_offset == 0) {
2384 ticket->t_curr_res -= log->l_iclog_hsize;
2385 XLOG_TIC_ADD_REGION(ticket,
2387 XLOG_REG_TYPE_LRHEADER);
2388 INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2389 ASSIGN_LSN(head->h_lsn, log);
2390 ASSERT(log->l_curr_block >= 0);
2393 /* If there is enough room to write everything, then do it. Otherwise,
2394 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2395 * bit is on, so this will get flushed out. Don't update ic_offset
2396 * until you know exactly how many bytes get copied. Therefore, wait
2397 * until later to update ic_offset.
2399 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2400 * can fit into remaining data section.
2402 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2403 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2405 /* If I'm the only one writing to this iclog, sync it to disk */
2406 if (iclog->ic_refcnt == 1) {
2408 if ((error = xlog_state_release_iclog(log, iclog)))
2417 /* Do we have enough room to write the full amount in the remainder
2418 * of this iclog? Or must we continue a write on the next iclog and
2419 * mark this iclog as completely taken? In the case where we switch
2420 * iclogs (to mark it taken), this particular iclog will release/sync
2421 * to disk in xlog_write().
2423 if (len <= iclog->ic_size - iclog->ic_offset) {
2424 *continued_write = 0;
2425 iclog->ic_offset += len;
2427 *continued_write = 1;
2428 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2432 ASSERT(iclog->ic_offset <= iclog->ic_size);
2435 *logoffsetp = log_offset;
2437 } /* xlog_state_get_iclog_space */
2440 * Atomically get the log space required for a log ticket.
2442 * Once a ticket gets put onto the reserveq, it will only return after
2443 * the needed reservation is satisfied.
2446 xlog_grant_log_space(xlog_t *log,
2458 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2459 panic("grant Recovery problem");
2462 /* Is there space or do we need to sleep? */
2463 s = GRANT_LOCK(log);
2464 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2466 /* something is already sleeping; insert new transaction at end */
2467 if (log->l_reserve_headq) {
2468 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2469 xlog_trace_loggrant(log, tic,
2470 "xlog_grant_log_space: sleep 1");
2472 * Gotta check this before going to sleep, while we're
2473 * holding the grant lock.
2475 if (XLOG_FORCED_SHUTDOWN(log))
2478 XFS_STATS_INC(xs_sleep_logspace);
2479 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2481 * If we got an error, and the filesystem is shutting down,
2482 * we'll catch it down below. So just continue...
2484 xlog_trace_loggrant(log, tic,
2485 "xlog_grant_log_space: wake 1");
2486 s = GRANT_LOCK(log);
2488 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2489 need_bytes = tic->t_unit_res*tic->t_ocnt;
2491 need_bytes = tic->t_unit_res;
2494 if (XLOG_FORCED_SHUTDOWN(log))
2497 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2498 log->l_grant_reserve_bytes);
2499 if (free_bytes < need_bytes) {
2500 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2501 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2502 xlog_trace_loggrant(log, tic,
2503 "xlog_grant_log_space: sleep 2");
2504 XFS_STATS_INC(xs_sleep_logspace);
2505 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2507 if (XLOG_FORCED_SHUTDOWN(log)) {
2508 s = GRANT_LOCK(log);
2512 xlog_trace_loggrant(log, tic,
2513 "xlog_grant_log_space: wake 2");
2514 xlog_grant_push_ail(log->l_mp, need_bytes);
2515 s = GRANT_LOCK(log);
2517 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2518 xlog_del_ticketq(&log->l_reserve_headq, tic);
2520 /* we've got enough space */
2521 xlog_grant_add_space(log, need_bytes);
2523 tail_lsn = log->l_tail_lsn;
2525 * Check to make sure the grant write head didn't just over lap the
2526 * tail. If the cycles are the same, we can't be overlapping.
2527 * Otherwise, make sure that the cycles differ by exactly one and
2528 * check the byte count.
2530 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2531 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2532 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2535 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2536 xlog_verify_grant_head(log, 1);
2537 GRANT_UNLOCK(log, s);
2541 if (tic->t_flags & XLOG_TIC_IN_Q)
2542 xlog_del_ticketq(&log->l_reserve_headq, tic);
2543 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2545 * If we are failing, make sure the ticket doesn't have any
2546 * current reservations. We don't want to add this back when
2547 * the ticket/transaction gets cancelled.
2549 tic->t_curr_res = 0;
2550 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2551 GRANT_UNLOCK(log, s);
2552 return XFS_ERROR(EIO);
2553 } /* xlog_grant_log_space */
2557 * Replenish the byte reservation required by moving the grant write head.
2562 xlog_regrant_write_log_space(xlog_t *log,
2566 int free_bytes, need_bytes;
2567 xlog_ticket_t *ntic;
2572 tic->t_curr_res = tic->t_unit_res;
2573 XLOG_TIC_RESET_RES(tic);
2579 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2580 panic("regrant Recovery problem");
2583 s = GRANT_LOCK(log);
2584 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2586 if (XLOG_FORCED_SHUTDOWN(log))
2589 /* If there are other waiters on the queue then give them a
2590 * chance at logspace before us. Wake up the first waiters,
2591 * if we do not wake up all the waiters then go to sleep waiting
2592 * for more free space, otherwise try to get some space for
2596 if ((ntic = log->l_write_headq)) {
2597 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2598 log->l_grant_write_bytes);
2600 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2602 if (free_bytes < ntic->t_unit_res)
2604 free_bytes -= ntic->t_unit_res;
2605 sv_signal(&ntic->t_sema);
2606 ntic = ntic->t_next;
2607 } while (ntic != log->l_write_headq);
2609 if (ntic != log->l_write_headq) {
2610 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2611 xlog_ins_ticketq(&log->l_write_headq, tic);
2613 xlog_trace_loggrant(log, tic,
2614 "xlog_regrant_write_log_space: sleep 1");
2615 XFS_STATS_INC(xs_sleep_logspace);
2616 sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2617 &log->l_grant_lock, s);
2619 /* If we're shutting down, this tic is already
2621 if (XLOG_FORCED_SHUTDOWN(log)) {
2622 s = GRANT_LOCK(log);
2626 xlog_trace_loggrant(log, tic,
2627 "xlog_regrant_write_log_space: wake 1");
2628 xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2629 s = GRANT_LOCK(log);
2633 need_bytes = tic->t_unit_res;
2636 if (XLOG_FORCED_SHUTDOWN(log))
2639 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2640 log->l_grant_write_bytes);
2641 if (free_bytes < need_bytes) {
2642 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2643 xlog_ins_ticketq(&log->l_write_headq, tic);
2644 XFS_STATS_INC(xs_sleep_logspace);
2645 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2647 /* If we're shutting down, this tic is already off the queue */
2648 if (XLOG_FORCED_SHUTDOWN(log)) {
2649 s = GRANT_LOCK(log);
2653 xlog_trace_loggrant(log, tic,
2654 "xlog_regrant_write_log_space: wake 2");
2655 xlog_grant_push_ail(log->l_mp, need_bytes);
2656 s = GRANT_LOCK(log);
2658 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2659 xlog_del_ticketq(&log->l_write_headq, tic);
2661 /* we've got enough space */
2662 xlog_grant_add_space_write(log, need_bytes);
2664 tail_lsn = log->l_tail_lsn;
2665 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2666 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2667 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2671 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2672 xlog_verify_grant_head(log, 1);
2673 GRANT_UNLOCK(log, s);
2678 if (tic->t_flags & XLOG_TIC_IN_Q)
2679 xlog_del_ticketq(&log->l_reserve_headq, tic);
2680 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2682 * If we are failing, make sure the ticket doesn't have any
2683 * current reservations. We don't want to add this back when
2684 * the ticket/transaction gets cancelled.
2686 tic->t_curr_res = 0;
2687 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2688 GRANT_UNLOCK(log, s);
2689 return XFS_ERROR(EIO);
2690 } /* xlog_regrant_write_log_space */
2693 /* The first cnt-1 times through here we don't need to
2694 * move the grant write head because the permanent
2695 * reservation has reserved cnt times the unit amount.
2696 * Release part of current permanent unit reservation and
2697 * reset current reservation to be one units worth. Also
2698 * move grant reservation head forward.
2701 xlog_regrant_reserve_log_space(xlog_t *log,
2702 xlog_ticket_t *ticket)
2706 xlog_trace_loggrant(log, ticket,
2707 "xlog_regrant_reserve_log_space: enter");
2708 if (ticket->t_cnt > 0)
2711 s = GRANT_LOCK(log);
2712 xlog_grant_sub_space(log, ticket->t_curr_res);
2713 ticket->t_curr_res = ticket->t_unit_res;
2714 XLOG_TIC_RESET_RES(ticket);
2715 xlog_trace_loggrant(log, ticket,
2716 "xlog_regrant_reserve_log_space: sub current res");
2717 xlog_verify_grant_head(log, 1);
2719 /* just return if we still have some of the pre-reserved space */
2720 if (ticket->t_cnt > 0) {
2721 GRANT_UNLOCK(log, s);
2725 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2726 xlog_trace_loggrant(log, ticket,
2727 "xlog_regrant_reserve_log_space: exit");
2728 xlog_verify_grant_head(log, 0);
2729 GRANT_UNLOCK(log, s);
2730 ticket->t_curr_res = ticket->t_unit_res;
2731 XLOG_TIC_RESET_RES(ticket);
2732 } /* xlog_regrant_reserve_log_space */
2736 * Give back the space left from a reservation.
2738 * All the information we need to make a correct determination of space left
2739 * is present. For non-permanent reservations, things are quite easy. The
2740 * count should have been decremented to zero. We only need to deal with the
2741 * space remaining in the current reservation part of the ticket. If the
2742 * ticket contains a permanent reservation, there may be left over space which
2743 * needs to be released. A count of N means that N-1 refills of the current
2744 * reservation can be done before we need to ask for more space. The first
2745 * one goes to fill up the first current reservation. Once we run out of
2746 * space, the count will stay at zero and the only space remaining will be
2747 * in the current reservation field.
2750 xlog_ungrant_log_space(xlog_t *log,
2751 xlog_ticket_t *ticket)
2755 if (ticket->t_cnt > 0)
2758 s = GRANT_LOCK(log);
2759 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2761 xlog_grant_sub_space(log, ticket->t_curr_res);
2763 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2765 /* If this is a permanent reservation ticket, we may be able to free
2766 * up more space based on the remaining count.
2768 if (ticket->t_cnt > 0) {
2769 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2770 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2773 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2774 xlog_verify_grant_head(log, 1);
2775 GRANT_UNLOCK(log, s);
2776 xfs_log_move_tail(log->l_mp, 1);
2777 } /* xlog_ungrant_log_space */
2781 * Atomically put back used ticket.
2784 xlog_state_put_ticket(xlog_t *log,
2790 xlog_ticket_put(log, tic);
2792 } /* xlog_state_put_ticket */
2795 * Flush iclog to disk if this is the last reference to the given iclog and
2796 * the WANT_SYNC bit is set.
2798 * When this function is entered, the iclog is not necessarily in the
2799 * WANT_SYNC state. It may be sitting around waiting to get filled.
2804 xlog_state_release_iclog(xlog_t *log,
2805 xlog_in_core_t *iclog)
2808 int sync = 0; /* do we sync? */
2810 xlog_assign_tail_lsn(log->l_mp);
2814 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2816 return XFS_ERROR(EIO);
2819 ASSERT(iclog->ic_refcnt > 0);
2820 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2821 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2823 if (--iclog->ic_refcnt == 0 &&
2824 iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2826 iclog->ic_state = XLOG_STATE_SYNCING;
2827 INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2828 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2829 /* cycle incremented when incrementing curr_block */
2835 * We let the log lock go, so it's possible that we hit a log I/O
2836 * error or someother SHUTDOWN condition that marks the iclog
2837 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2838 * this iclog has consistent data, so we ignore IOERROR
2839 * flags after this point.
2842 return xlog_sync(log, iclog);
2846 } /* xlog_state_release_iclog */
2850 * This routine will mark the current iclog in the ring as WANT_SYNC
2851 * and move the current iclog pointer to the next iclog in the ring.
2852 * When this routine is called from xlog_state_get_iclog_space(), the
2853 * exact size of the iclog has not yet been determined. All we know is
2854 * that every data block. We have run out of space in this log record.
2857 xlog_state_switch_iclogs(xlog_t *log,
2858 xlog_in_core_t *iclog,
2861 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2863 eventual_size = iclog->ic_offset;
2864 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2865 INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2866 log->l_prev_block = log->l_curr_block;
2867 log->l_prev_cycle = log->l_curr_cycle;
2869 /* roll log?: ic_offset changed later */
2870 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2872 /* Round up to next log-sunit */
2873 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2874 log->l_mp->m_sb.sb_logsunit > 1) {
2875 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2876 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2879 if (log->l_curr_block >= log->l_logBBsize) {
2880 log->l_curr_cycle++;
2881 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2882 log->l_curr_cycle++;
2883 log->l_curr_block -= log->l_logBBsize;
2884 ASSERT(log->l_curr_block >= 0);
2886 ASSERT(iclog == log->l_iclog);
2887 log->l_iclog = iclog->ic_next;
2888 } /* xlog_state_switch_iclogs */
2892 * Write out all data in the in-core log as of this exact moment in time.
2894 * Data may be written to the in-core log during this call. However,
2895 * we don't guarantee this data will be written out. A change from past
2896 * implementation means this routine will *not* write out zero length LRs.
2898 * Basically, we try and perform an intelligent scan of the in-core logs.
2899 * If we determine there is no flushable data, we just return. There is no
2900 * flushable data if:
2902 * 1. the current iclog is active and has no data; the previous iclog
2903 * is in the active or dirty state.
2904 * 2. the current iclog is drity, and the previous iclog is in the
2905 * active or dirty state.
2907 * We may sleep (call psema) if:
2909 * 1. the current iclog is not in the active nor dirty state.
2910 * 2. the current iclog dirty, and the previous iclog is not in the
2911 * active nor dirty state.
2912 * 3. the current iclog is active, and there is another thread writing
2913 * to this particular iclog.
2914 * 4. a) the current iclog is active and has no other writers
2915 * b) when we return from flushing out this iclog, it is still
2916 * not in the active nor dirty state.
2919 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2921 xlog_in_core_t *iclog;
2927 iclog = log->l_iclog;
2928 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2930 return XFS_ERROR(EIO);
2933 /* If the head iclog is not active nor dirty, we just attach
2934 * ourselves to the head and go to sleep.
2936 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2937 iclog->ic_state == XLOG_STATE_DIRTY) {
2939 * If the head is dirty or (active and empty), then
2940 * we need to look at the previous iclog. If the previous
2941 * iclog is active or dirty we are done. There is nothing
2942 * to sync out. Otherwise, we attach ourselves to the
2943 * previous iclog and go to sleep.
2945 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2946 (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2947 iclog = iclog->ic_prev;
2948 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2949 iclog->ic_state == XLOG_STATE_DIRTY)
2954 if (iclog->ic_refcnt == 0) {
2955 /* We are the only one with access to this
2956 * iclog. Flush it out now. There should
2957 * be a roundoff of zero to show that someone
2958 * has already taken care of the roundoff from
2959 * the previous sync.
2962 lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2963 xlog_state_switch_iclogs(log, iclog, 0);
2966 if (xlog_state_release_iclog(log, iclog))
2967 return XFS_ERROR(EIO);
2970 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2971 iclog->ic_state != XLOG_STATE_DIRTY)
2976 /* Someone else is writing to this iclog.
2977 * Use its call to flush out the data. However,
2978 * the other thread may not force out this LR,
2979 * so we mark it WANT_SYNC.
2981 xlog_state_switch_iclogs(log, iclog, 0);
2987 /* By the time we come around again, the iclog could've been filled
2988 * which would give it another lsn. If we have a new lsn, just
2989 * return because the relevant data has been flushed.
2992 if (flags & XFS_LOG_SYNC) {
2994 * We must check if we're shutting down here, before
2995 * we wait, while we're holding the LOG_LOCK.
2996 * Then we check again after waking up, in case our
2997 * sleep was disturbed by a bad news.
2999 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3001 return XFS_ERROR(EIO);
3003 XFS_STATS_INC(xs_log_force_sleep);
3004 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3006 * No need to grab the log lock here since we're
3007 * only deciding whether or not to return EIO
3008 * and the memory read should be atomic.
3010 if (iclog->ic_state & XLOG_STATE_IOERROR)
3011 return XFS_ERROR(EIO);
3020 } /* xlog_state_sync_all */
3024 * Used by code which implements synchronous log forces.
3026 * Find in-core log with lsn.
3027 * If it is in the DIRTY state, just return.
3028 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3029 * state and go to sleep or return.
3030 * If it is in any other state, go to sleep or return.
3032 * If filesystem activity goes to zero, the iclog will get flushed only by
3036 xlog_state_sync(xlog_t *log,
3041 xlog_in_core_t *iclog;
3042 int already_slept = 0;
3048 iclog = log->l_iclog;
3050 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3052 return XFS_ERROR(EIO);
3056 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3057 iclog = iclog->ic_next;
3061 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3066 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3068 * We sleep here if we haven't already slept (e.g.
3069 * this is the first time we've looked at the correct
3070 * iclog buf) and the buffer before us is going to
3071 * be sync'ed. The reason for this is that if we
3072 * are doing sync transactions here, by waiting for
3073 * the previous I/O to complete, we can allow a few
3074 * more transactions into this iclog before we close
3077 * Otherwise, we mark the buffer WANT_SYNC, and bump
3078 * up the refcnt so we can release the log (which drops
3079 * the ref count). The state switch keeps new transaction
3080 * commits from using this buffer. When the current commits
3081 * finish writing into the buffer, the refcount will drop to
3082 * zero and the buffer will go out then.
3084 if (!already_slept &&
3085 (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3086 XLOG_STATE_SYNCING))) {
3087 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3088 XFS_STATS_INC(xs_log_force_sleep);
3089 sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3090 &log->l_icloglock, s);
3096 xlog_state_switch_iclogs(log, iclog, 0);
3098 if (xlog_state_release_iclog(log, iclog))
3099 return XFS_ERROR(EIO);
3105 if ((flags & XFS_LOG_SYNC) && /* sleep */
3106 !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3109 * Don't wait on the forcesema if we know that we've
3110 * gotten a log write error.
3112 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3114 return XFS_ERROR(EIO);
3116 XFS_STATS_INC(xs_log_force_sleep);
3117 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3119 * No need to grab the log lock here since we're
3120 * only deciding whether or not to return EIO
3121 * and the memory read should be atomic.
3123 if (iclog->ic_state & XLOG_STATE_IOERROR)
3124 return XFS_ERROR(EIO);
3126 } else { /* just return */
3131 } while (iclog != log->l_iclog);
3135 } /* xlog_state_sync */
3139 * Called when we want to mark the current iclog as being ready to sync to
3143 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3149 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3150 xlog_state_switch_iclogs(log, iclog, 0);
3152 ASSERT(iclog->ic_state &
3153 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3157 } /* xlog_state_want_sync */
3161 /*****************************************************************************
3165 *****************************************************************************
3169 * Algorithm doesn't take into account page size. ;-(
3172 xlog_state_ticket_alloc(xlog_t *log)
3174 xlog_ticket_t *t_list;
3175 xlog_ticket_t *next;
3177 uint i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3181 * The kmem_zalloc may sleep, so we shouldn't be holding the
3182 * global lock. XXXmiken: may want to use zone allocator.
3184 buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3188 /* Attach 1st ticket to Q, so we can keep track of allocated memory */
3189 t_list = (xlog_ticket_t *)buf;
3190 t_list->t_next = log->l_unmount_free;
3191 log->l_unmount_free = t_list++;
3192 log->l_ticket_cnt++;
3193 log->l_ticket_tcnt++;
3195 /* Next ticket becomes first ticket attached to ticket free list */
3196 if (log->l_freelist != NULL) {
3197 ASSERT(log->l_tail != NULL);
3198 log->l_tail->t_next = t_list;
3200 log->l_freelist = t_list;
3202 log->l_ticket_cnt++;
3203 log->l_ticket_tcnt++;
3205 /* Cycle through rest of alloc'ed memory, building up free Q */
3206 for ( ; i > 0; i--) {
3208 t_list->t_next = next;
3210 log->l_ticket_cnt++;
3211 log->l_ticket_tcnt++;
3213 t_list->t_next = NULL;
3214 log->l_tail = t_list;
3216 } /* xlog_state_ticket_alloc */
3220 * Put ticket into free list
3222 * Assumption: log lock is held around this call.
3225 xlog_ticket_put(xlog_t *log,
3226 xlog_ticket_t *ticket)
3228 sv_destroy(&ticket->t_sema);
3231 * Don't think caching will make that much difference. It's
3232 * more important to make debug easier.
3235 /* real code will want to use LIFO for caching */
3236 ticket->t_next = log->l_freelist;
3237 log->l_freelist = ticket;
3238 /* no need to clear fields */
3240 /* When we debug, it is easier if tickets are cycled */
3241 ticket->t_next = NULL;
3242 if (log->l_tail != 0) {
3243 log->l_tail->t_next = ticket;
3245 ASSERT(log->l_freelist == 0);
3246 log->l_freelist = ticket;
3248 log->l_tail = ticket;
3250 log->l_ticket_cnt++;
3251 } /* xlog_ticket_put */
3255 * Grab ticket off freelist or allocation some more
3258 xlog_ticket_get(xlog_t *log,
3269 if (log->l_freelist == NULL)
3270 xlog_state_ticket_alloc(log); /* potentially sleep */
3273 if (log->l_freelist == NULL) {
3277 tic = log->l_freelist;
3278 log->l_freelist = tic->t_next;
3279 if (log->l_freelist == NULL)
3281 log->l_ticket_cnt--;
3285 * Permanent reservations have up to 'cnt'-1 active log operations
3286 * in the log. A unit in this case is the amount of space for one
3287 * of these log operations. Normal reservations have a cnt of 1
3288 * and their unit amount is the total amount of space required.
3290 * The following lines of code account for non-transaction data
3291 * which occupy space in the on-disk log.
3293 * Normal form of a transaction is:
3294 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3295 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3297 * We need to account for all the leadup data and trailer data
3298 * around the transaction data.
3299 * And then we need to account for the worst case in terms of using
3301 * The worst case will happen if:
3302 * - the placement of the transaction happens to be such that the
3303 * roundoff is at its maximum
3304 * - the transaction data is synced before the commit record is synced
3305 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3306 * Therefore the commit record is in its own Log Record.
3307 * This can happen as the commit record is called with its
3308 * own region to xlog_write().
3309 * This then means that in the worst case, roundoff can happen for
3310 * the commit-rec as well.
3311 * The commit-rec is smaller than padding in this scenario and so it is
3312 * not added separately.
3315 /* for trans header */
3316 unit_bytes += sizeof(xlog_op_header_t);
3317 unit_bytes += sizeof(xfs_trans_header_t);
3320 unit_bytes += sizeof(xlog_op_header_t);
3322 /* for LR headers */
3323 num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3324 unit_bytes += log->l_iclog_hsize * num_headers;
3326 /* for commit-rec LR header - note: padding will subsume the ophdr */
3327 unit_bytes += log->l_iclog_hsize;
3329 /* for split-recs - ophdrs added when data split over LRs */
3330 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3332 /* for roundoff padding for transaction data and one for commit record */
3333 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3334 log->l_mp->m_sb.sb_logsunit > 1) {
3335 /* log su roundoff */
3336 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3339 unit_bytes += 2*BBSIZE;
3342 tic->t_unit_res = unit_bytes;
3343 tic->t_curr_res = unit_bytes;
3346 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3347 tic->t_clientid = client;
3348 tic->t_flags = XLOG_TIC_INITED;
3349 tic->t_trans_type = 0;
3350 if (xflags & XFS_LOG_PERM_RESERV)
3351 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3352 sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3354 XLOG_TIC_RESET_RES(tic);
3357 } /* xlog_ticket_get */
3360 /******************************************************************************
3362 * Log debug routines
3364 ******************************************************************************
3368 * Make sure that the destination ptr is within the valid data region of
3369 * one of the iclogs. This uses backup pointers stored in a different
3370 * part of the log in case we trash the log structure.
3373 xlog_verify_dest_ptr(xlog_t *log,
3379 for (i=0; i < log->l_iclog_bufs; i++) {
3380 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3381 ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3385 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3386 } /* xlog_verify_dest_ptr */
3389 xlog_verify_grant_head(xlog_t *log, int equals)
3391 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3393 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3395 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3397 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3398 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3400 } /* xlog_verify_grant_head */
3402 /* check if it will fit */
3404 xlog_verify_tail_lsn(xlog_t *log,
3405 xlog_in_core_t *iclog,
3410 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3412 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3413 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3414 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3416 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3418 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3419 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3421 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3422 if (blocks < BTOBB(iclog->ic_offset) + 1)
3423 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3425 } /* xlog_verify_tail_lsn */
3428 * Perform a number of checks on the iclog before writing to disk.
3430 * 1. Make sure the iclogs are still circular
3431 * 2. Make sure we have a good magic number
3432 * 3. Make sure we don't have magic numbers in the data
3433 * 4. Check fields of each log operation header for:
3434 * A. Valid client identifier
3435 * B. tid ptr value falls in valid ptr space (user space code)
3436 * C. Length in log record header is correct according to the
3437 * individual operation headers within record.
3438 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3439 * log, check the preceding blocks of the physical log to make sure all
3440 * the cycle numbers agree with the current cycle number.
3443 xlog_verify_iclog(xlog_t *log,
3444 xlog_in_core_t *iclog,
3448 xlog_op_header_t *ophead;
3449 xlog_in_core_t *icptr;
3450 xlog_in_core_2_t *xhdr;
3452 xfs_caddr_t base_ptr;
3453 __psint_t field_offset;
3455 int len, i, j, k, op_len;
3459 /* check validity of iclog pointers */
3461 icptr = log->l_iclog;
3462 for (i=0; i < log->l_iclog_bufs; i++) {
3464 xlog_panic("xlog_verify_iclog: invalid ptr");
3465 icptr = icptr->ic_next;
3467 if (icptr != log->l_iclog)
3468 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3471 /* check log magic numbers */
3472 ptr = (xfs_caddr_t) &(iclog->ic_header);
3473 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3474 xlog_panic("xlog_verify_iclog: invalid magic num");
3476 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3478 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3479 xlog_panic("xlog_verify_iclog: unexpected magic num");
3483 len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3484 ptr = iclog->ic_datap;
3486 ophead = (xlog_op_header_t *)ptr;
3487 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3488 for (i = 0; i < len; i++) {
3489 ophead = (xlog_op_header_t *)ptr;
3491 /* clientid is only 1 byte */
3492 field_offset = (__psint_t)
3493 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3494 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3495 clientid = ophead->oh_clientid;
3497 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3498 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3499 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3500 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3501 clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3503 clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3506 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3507 cmn_err(CE_WARN, "xlog_verify_iclog: "
3508 "invalid clientid %d op 0x%p offset 0x%lx",
3509 clientid, ophead, (unsigned long)field_offset);
3512 field_offset = (__psint_t)
3513 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3514 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3515 op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3517 idx = BTOBBT((__psint_t)&ophead->oh_len -
3518 (__psint_t)iclog->ic_datap);
3519 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3520 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3521 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3522 op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3524 op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3527 ptr += sizeof(xlog_op_header_t) + op_len;
3529 } /* xlog_verify_iclog */
3533 * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3539 xlog_in_core_t *iclog, *ic;
3541 iclog = log->l_iclog;
3542 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3544 * Mark all the incore logs IOERROR.
3545 * From now on, no log flushes will result.
3549 ic->ic_state = XLOG_STATE_IOERROR;
3551 } while (ic != iclog);
3555 * Return non-zero, if state transition has already happened.
3561 * This is called from xfs_force_shutdown, when we're forcibly
3562 * shutting down the filesystem, typically because of an IO error.
3563 * Our main objectives here are to make sure that:
3564 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3565 * parties to find out, 'atomically'.
3566 * b. those who're sleeping on log reservations, pinned objects and
3567 * other resources get woken up, and be told the bad news.
3568 * c. nothing new gets queued up after (a) and (b) are done.
3569 * d. if !logerror, flush the iclogs to disk, then seal them off
3573 xfs_log_force_umount(
3574 struct xfs_mount *mp,
3587 * If this happens during log recovery, don't worry about
3588 * locking; the log isn't open for business yet.
3591 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3592 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3593 XFS_BUF_DONE(mp->m_sb_bp);
3598 * Somebody could've already done the hard work for us.
3599 * No need to get locks for this.
3601 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3602 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3607 * We must hold both the GRANT lock and the LOG lock,
3608 * before we mark the filesystem SHUTDOWN and wake
3609 * everybody up to tell the bad news.
3611 s = GRANT_LOCK(log);
3613 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3614 XFS_BUF_DONE(mp->m_sb_bp);
3616 * This flag is sort of redundant because of the mount flag, but
3617 * it's good to maintain the separation between the log and the rest
3620 log->l_flags |= XLOG_IO_ERROR;
3623 * If we hit a log error, we want to mark all the iclogs IOERROR
3624 * while we're still holding the loglock.
3627 retval = xlog_state_ioerror(log);
3628 LOG_UNLOCK(log, s2);
3631 * We don't want anybody waiting for log reservations
3632 * after this. That means we have to wake up everybody
3633 * queued up on reserve_headq as well as write_headq.
3634 * In addition, we make sure in xlog_{re}grant_log_space
3635 * that we don't enqueue anything once the SHUTDOWN flag
3636 * is set, and this action is protected by the GRANTLOCK.
3638 if ((tic = log->l_reserve_headq)) {
3640 sv_signal(&tic->t_sema);
3642 } while (tic != log->l_reserve_headq);
3645 if ((tic = log->l_write_headq)) {
3647 sv_signal(&tic->t_sema);
3649 } while (tic != log->l_write_headq);
3651 GRANT_UNLOCK(log, s);
3653 if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3656 * Force the incore logs to disk before shutting the
3657 * log down completely.
3659 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3661 retval = xlog_state_ioerror(log);
3662 LOG_UNLOCK(log, s2);
3665 * Wake up everybody waiting on xfs_log_force.
3666 * Callback all log item committed functions as if the
3667 * log writes were completed.
3669 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3671 #ifdef XFSERRORDEBUG
3673 xlog_in_core_t *iclog;
3676 iclog = log->l_iclog;
3678 ASSERT(iclog->ic_callback == 0);
3679 iclog = iclog->ic_next;
3680 } while (iclog != log->l_iclog);
3684 /* return non-zero if log IOERROR transition had already happened */
3689 xlog_iclogs_empty(xlog_t *log)
3691 xlog_in_core_t *iclog;
3693 iclog = log->l_iclog;
3695 /* endianness does not matter here, zero is zero in
3698 if (iclog->ic_header.h_num_logops)
3700 iclog = iclog->ic_next;
3701 } while (iclog != log->l_iclog);