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_dealloc_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 specific
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, SHUTDOWN_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)) {
501 bhv_vfs_t *vfsp = XFS_MTOVFS(mp);
502 int error, readonly = (vfsp->vfs_flag & VFS_RDONLY);
505 vfsp->vfs_flag &= ~VFS_RDONLY;
507 error = xlog_recover(mp->m_log);
510 vfsp->vfs_flag |= VFS_RDONLY;
512 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
513 xlog_dealloc_log(mp->m_log);
518 /* Normal transactions can now occur */
519 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
521 /* End mounting message in xfs_log_mount_finish */
523 } /* xfs_log_mount */
526 * Finish the recovery of the file system. This is separate from
527 * the xfs_log_mount() call, because it depends on the code in
528 * xfs_mountfs() to read in the root and real-time bitmap inodes
529 * between calling xfs_log_mount() and here.
531 * mp - ubiquitous xfs mount point structure
534 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
538 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
539 error = xlog_recover_finish(mp->m_log, mfsi_flags);
542 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
549 * Unmount processing for the log.
552 xfs_log_unmount(xfs_mount_t *mp)
556 error = xfs_log_unmount_write(mp);
557 xfs_log_unmount_dealloc(mp);
562 * Final log writes as part of unmount.
564 * Mark the filesystem clean as unmount happens. Note that during relocation
565 * this routine needs to be executed as part of source-bag while the
566 * deallocation must not be done until source-end.
570 * Unmount record used to have a string "Unmount filesystem--" in the
571 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
572 * We just write the magic number now since that particular field isn't
573 * currently architecture converted and "nUmount" is a bit foo.
574 * As far as I know, there weren't any dependencies on the old behaviour.
578 xfs_log_unmount_write(xfs_mount_t *mp)
580 xlog_t *log = mp->m_log;
581 xlog_in_core_t *iclog;
583 xlog_in_core_t *first_iclog;
585 xfs_log_iovec_t reg[1];
586 xfs_log_ticket_t tic = NULL;
591 /* the data section must be 32 bit size aligned */
595 __uint32_t pad2; /* may as well make it 64 bits */
596 } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
599 * Don't write out unmount record on read-only mounts.
600 * Or, if we are doing a forced umount (typically because of IO errors).
602 if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
605 xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
608 first_iclog = iclog = log->l_iclog;
610 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
611 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
612 ASSERT(iclog->ic_offset == 0);
614 iclog = iclog->ic_next;
615 } while (iclog != first_iclog);
617 if (! (XLOG_FORCED_SHUTDOWN(log))) {
618 reg[0].i_addr = (void*)&magic;
619 reg[0].i_len = sizeof(magic);
620 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_UNMOUNT);
622 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
624 /* remove inited flag */
625 ((xlog_ticket_t *)tic)->t_flags = 0;
626 error = xlog_write(mp, reg, 1, tic, &lsn,
627 NULL, XLOG_UNMOUNT_TRANS);
629 * At this point, we're umounting anyway,
630 * so there's no point in transitioning log state
631 * to IOERROR. Just continue...
636 xfs_fs_cmn_err(CE_ALERT, mp,
637 "xfs_log_unmount: unmount record failed");
642 iclog = log->l_iclog;
645 xlog_state_want_sync(log, iclog);
646 (void) xlog_state_release_iclog(log, iclog);
649 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
650 iclog->ic_state == XLOG_STATE_DIRTY)) {
651 if (!XLOG_FORCED_SHUTDOWN(log)) {
652 sv_wait(&iclog->ic_forcesema, PMEM,
653 &log->l_icloglock, s);
661 xlog_state_put_ticket(log, tic);
664 * We're already in forced_shutdown mode, couldn't
665 * even attempt to write out the unmount transaction.
667 * Go through the motions of sync'ing and releasing
668 * the iclog, even though no I/O will actually happen,
669 * we need to wait for other log I/Os that may already
670 * be in progress. Do this as a separate section of
671 * code so we'll know if we ever get stuck here that
672 * we're in this odd situation of trying to unmount
673 * a file system that went into forced_shutdown as
674 * the result of an unmount..
677 iclog = log->l_iclog;
681 xlog_state_want_sync(log, iclog);
682 (void) xlog_state_release_iclog(log, iclog);
686 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
687 || iclog->ic_state == XLOG_STATE_DIRTY
688 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
690 sv_wait(&iclog->ic_forcesema, PMEM,
691 &log->l_icloglock, s);
698 } /* xfs_log_unmount_write */
701 * Deallocate log structures for unmount/relocation.
704 xfs_log_unmount_dealloc(xfs_mount_t *mp)
706 xlog_dealloc_log(mp->m_log);
710 * Write region vectors to log. The write happens using the space reservation
711 * of the ticket (tic). It is not a requirement that all writes for a given
712 * transaction occur with one call to xfs_log_write().
715 xfs_log_write(xfs_mount_t * mp,
716 xfs_log_iovec_t reg[],
718 xfs_log_ticket_t tic,
719 xfs_lsn_t *start_lsn)
722 xlog_t *log = mp->m_log;
724 if (XLOG_FORCED_SHUTDOWN(log))
725 return XFS_ERROR(EIO);
727 if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
728 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
731 } /* xfs_log_write */
735 xfs_log_move_tail(xfs_mount_t *mp,
739 xlog_t *log = mp->m_log;
740 int need_bytes, free_bytes, cycle, bytes;
743 if (XLOG_FORCED_SHUTDOWN(log))
745 ASSERT(!XFS_FORCED_SHUTDOWN(mp));
748 /* needed since sync_lsn is 64 bits */
750 tail_lsn = log->l_last_sync_lsn;
756 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
760 log->l_tail_lsn = tail_lsn;
763 if ((tic = log->l_write_headq)) {
765 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
766 panic("Recovery problem");
768 cycle = log->l_grant_write_cycle;
769 bytes = log->l_grant_write_bytes;
770 free_bytes = xlog_space_left(log, cycle, bytes);
772 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
774 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
777 free_bytes -= tic->t_unit_res;
778 sv_signal(&tic->t_sema);
780 } while (tic != log->l_write_headq);
782 if ((tic = log->l_reserve_headq)) {
784 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
785 panic("Recovery problem");
787 cycle = log->l_grant_reserve_cycle;
788 bytes = log->l_grant_reserve_bytes;
789 free_bytes = xlog_space_left(log, cycle, bytes);
791 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
792 need_bytes = tic->t_unit_res*tic->t_cnt;
794 need_bytes = tic->t_unit_res;
795 if (free_bytes < need_bytes && tail_lsn != 1)
798 free_bytes -= need_bytes;
799 sv_signal(&tic->t_sema);
801 } while (tic != log->l_reserve_headq);
803 GRANT_UNLOCK(log, s);
804 } /* xfs_log_move_tail */
807 * Determine if we have a transaction that has gone to disk
808 * that needs to be covered. Log activity needs to be idle (no AIL and
809 * nothing in the iclogs). And, we need to be in the right state indicating
810 * something has gone out.
813 xfs_log_need_covered(xfs_mount_t *mp)
817 xlog_t *log = mp->m_log;
818 bhv_vfs_t *vfsp = XFS_MTOVFS(mp);
820 if (vfs_test_for_freeze(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
821 (vfsp->vfs_flag & VFS_RDONLY))
825 if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
826 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
827 && !xfs_trans_first_ail(mp, &gen)
828 && xlog_iclogs_empty(log)) {
829 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
830 log->l_covered_state = XLOG_STATE_COVER_DONE;
832 ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
833 log->l_covered_state = XLOG_STATE_COVER_DONE2;
841 /******************************************************************************
845 ******************************************************************************
848 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
849 * The log manager must keep track of the last LR which was committed
850 * to disk. The lsn of this LR will become the new tail_lsn whenever
851 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
852 * the situation where stuff could be written into the log but nothing
853 * was ever in the AIL when asked. Eventually, we panic since the
854 * tail hits the head.
856 * We may be holding the log iclog lock upon entering this routine.
859 xlog_assign_tail_lsn(xfs_mount_t *mp)
863 xlog_t *log = mp->m_log;
865 tail_lsn = xfs_trans_tail_ail(mp);
868 log->l_tail_lsn = tail_lsn;
870 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
872 GRANT_UNLOCK(log, s);
875 } /* xlog_assign_tail_lsn */
879 * Return the space in the log between the tail and the head. The head
880 * is passed in the cycle/bytes formal parms. In the special case where
881 * the reserve head has wrapped passed the tail, this calculation is no
882 * longer valid. In this case, just return 0 which means there is no space
883 * in the log. This works for all places where this function is called
884 * with the reserve head. Of course, if the write head were to ever
885 * wrap the tail, we should blow up. Rather than catch this case here,
886 * we depend on other ASSERTions in other parts of the code. XXXmiken
888 * This code also handles the case where the reservation head is behind
889 * the tail. The details of this case are described below, but the end
890 * result is that we return the size of the log as the amount of space left.
893 xlog_space_left(xlog_t *log, int cycle, int bytes)
899 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
900 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
901 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
902 free_bytes = log->l_logsize - (bytes - tail_bytes);
903 } else if ((tail_cycle + 1) < cycle) {
905 } else if (tail_cycle < cycle) {
906 ASSERT(tail_cycle == (cycle - 1));
907 free_bytes = tail_bytes - bytes;
910 * The reservation head is behind the tail.
911 * In this case we just want to return the size of the
912 * log as the amount of space left.
914 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
915 "xlog_space_left: head behind tail\n"
916 " tail_cycle = %d, tail_bytes = %d\n"
917 " GH cycle = %d, GH bytes = %d",
918 tail_cycle, tail_bytes, cycle, bytes);
920 free_bytes = log->l_logsize;
923 } /* xlog_space_left */
927 * Log function which is called when an io completes.
929 * The log manager needs its own routine, in order to control what
930 * happens with the buffer after the write completes.
933 xlog_iodone(xfs_buf_t *bp)
935 xlog_in_core_t *iclog;
939 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
940 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
941 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
945 * Some versions of cpp barf on the recursive definition of
946 * ic_log -> hic_fields.ic_log and expand ic_log twice when
947 * it is passed through two macros. Workaround broken cpp.
952 * Race to shutdown the filesystem if we see an error.
954 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
955 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
956 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
958 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
960 * This flag will be propagated to the trans-committed
961 * callback routines to let them know that the log-commit
964 aborted = XFS_LI_ABORTED;
965 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
966 aborted = XFS_LI_ABORTED;
968 xlog_state_done_syncing(iclog, aborted);
969 if (!(XFS_BUF_ISASYNC(bp))) {
971 * Corresponding psema() will be done in bwrite(). If we don't
972 * vsema() here, panic.
974 XFS_BUF_V_IODONESEMA(bp);
979 * The bdstrat callback function for log bufs. This gives us a central
980 * place to trap bufs in case we get hit by a log I/O error and need to
981 * shutdown. Actually, in practice, even when we didn't get a log error,
982 * we transition the iclogs to IOERROR state *after* flushing all existing
983 * iclogs to disk. This is because we don't want anymore new transactions to be
984 * started or completed afterwards.
987 xlog_bdstrat_cb(struct xfs_buf *bp)
989 xlog_in_core_t *iclog;
991 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
993 if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
994 /* note for irix bstrat will need struct bdevsw passed
995 * Fix the following macro if the code ever is merged
1001 xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1002 XFS_BUF_ERROR(bp, EIO);
1005 return XFS_ERROR(EIO);
1011 * Return size of each in-core log record buffer.
1013 * Low memory machines only get 2 16KB buffers. We don't want to waste
1014 * memory here. However, all other machines get at least 2 32KB buffers.
1015 * The number is hard coded because we don't care about the minimum
1016 * memory size, just 32MB systems.
1018 * If the filesystem blocksize is too large, we may need to choose a
1019 * larger size since the directory code currently logs entire blocks.
1023 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1029 if (mp->m_logbufs <= 0) {
1030 if (xfs_physmem <= btoc(128*1024*1024)) {
1031 log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1032 } else if (xfs_physmem <= btoc(400*1024*1024)) {
1033 log->l_iclog_bufs = XLOG_MED_ICLOGS;
1034 } else { /* 256K with 32K bufs */
1035 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1038 log->l_iclog_bufs = mp->m_logbufs;
1042 * Buffer size passed in from mount system call.
1044 if (mp->m_logbsize > 0) {
1045 size = log->l_iclog_size = mp->m_logbsize;
1046 log->l_iclog_size_log = 0;
1048 log->l_iclog_size_log++;
1052 if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1053 /* # headers = size / 32K
1054 * one header holds cycles from 32K of data
1057 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1058 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1060 log->l_iclog_hsize = xhdrs << BBSHIFT;
1061 log->l_iclog_heads = xhdrs;
1063 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1064 log->l_iclog_hsize = BBSIZE;
1065 log->l_iclog_heads = 1;
1071 * Special case machines that have less than 32MB of memory.
1072 * All machines with more memory use 32KB buffers.
1074 if (xfs_physmem <= btoc(32*1024*1024)) {
1075 /* Don't change; min configuration */
1076 log->l_iclog_size = XLOG_RECORD_BSIZE; /* 16k */
1077 log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1079 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE; /* 32k */
1080 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1083 /* the default log size is 16k or 32k which is one header sector */
1084 log->l_iclog_hsize = BBSIZE;
1085 log->l_iclog_heads = 1;
1088 * For 16KB, we use 3 32KB buffers. For 32KB block sizes, we use
1089 * 4 32KB buffers. For 64KB block sizes, we use 8 32KB buffers.
1091 if (mp->m_sb.sb_blocksize >= 16*1024) {
1092 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1093 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1094 if (mp->m_logbufs <= 0) {
1095 switch (mp->m_sb.sb_blocksize) {
1096 case 16*1024: /* 16 KB */
1097 log->l_iclog_bufs = 3;
1099 case 32*1024: /* 32 KB */
1100 log->l_iclog_bufs = 4;
1102 case 64*1024: /* 64 KB */
1103 log->l_iclog_bufs = 8;
1106 xlog_panic("XFS: Invalid blocksize");
1112 done: /* are we being asked to make the sizes selected above visible? */
1113 if (mp->m_logbufs == 0)
1114 mp->m_logbufs = log->l_iclog_bufs;
1115 if (mp->m_logbsize == 0)
1116 mp->m_logbsize = log->l_iclog_size;
1117 } /* xlog_get_iclog_buffer_size */
1121 * This routine initializes some of the log structure for a given mount point.
1122 * Its primary purpose is to fill in enough, so recovery can occur. However,
1123 * some other stuff may be filled in too.
1126 xlog_alloc_log(xfs_mount_t *mp,
1127 xfs_buftarg_t *log_target,
1128 xfs_daddr_t blk_offset,
1132 xlog_rec_header_t *head;
1133 xlog_in_core_t **iclogp;
1134 xlog_in_core_t *iclog, *prev_iclog=NULL;
1139 log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1142 log->l_targ = log_target;
1143 log->l_logsize = BBTOB(num_bblks);
1144 log->l_logBBstart = blk_offset;
1145 log->l_logBBsize = num_bblks;
1146 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1147 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1149 log->l_prev_block = -1;
1150 ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1151 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1152 log->l_last_sync_lsn = log->l_tail_lsn;
1153 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1154 log->l_grant_reserve_cycle = 1;
1155 log->l_grant_write_cycle = 1;
1157 if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1158 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1159 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1160 /* for larger sector sizes, must have v2 or external log */
1161 ASSERT(log->l_sectbb_log == 0 ||
1162 log->l_logBBstart == 0 ||
1163 XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1164 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1166 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1168 xlog_get_iclog_buffer_size(mp, log);
1170 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1171 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1172 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1173 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1174 ASSERT(XFS_BUF_ISBUSY(bp));
1175 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1178 spinlock_init(&log->l_icloglock, "iclog");
1179 spinlock_init(&log->l_grant_lock, "grhead_iclog");
1180 initnsema(&log->l_flushsema, 0, "ic-flush");
1181 xlog_state_ticket_alloc(log); /* wait until after icloglock inited */
1183 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1184 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1186 iclogp = &log->l_iclog;
1188 * The amount of memory to allocate for the iclog structure is
1189 * rather funky due to the way the structure is defined. It is
1190 * done this way so that we can use different sizes for machines
1191 * with different amounts of memory. See the definition of
1192 * xlog_in_core_t in xfs_log_priv.h for details.
1194 iclogsize = log->l_iclog_size;
1195 ASSERT(log->l_iclog_size >= 4096);
1196 for (i=0; i < log->l_iclog_bufs; i++) {
1197 *iclogp = (xlog_in_core_t *)
1198 kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1200 iclog->hic_data = (xlog_in_core_2_t *)
1201 kmem_zalloc(iclogsize, KM_SLEEP);
1203 iclog->ic_prev = prev_iclog;
1205 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1207 head = &iclog->ic_header;
1208 memset(head, 0, sizeof(xlog_rec_header_t));
1209 INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1210 INT_SET(head->h_version, ARCH_CONVERT,
1211 XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1212 INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1214 INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1215 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1217 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1218 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1219 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1220 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1223 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1224 iclog->ic_state = XLOG_STATE_ACTIVE;
1225 iclog->ic_log = log;
1226 iclog->ic_callback_tail = &(iclog->ic_callback);
1227 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1229 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1230 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1231 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1232 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1234 iclogp = &iclog->ic_next;
1236 *iclogp = log->l_iclog; /* complete ring */
1237 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1240 } /* xlog_alloc_log */
1244 * Write out the commit record of a transaction associated with the given
1245 * ticket. Return the lsn of the commit record.
1248 xlog_commit_record(xfs_mount_t *mp,
1249 xlog_ticket_t *ticket,
1250 xlog_in_core_t **iclog,
1251 xfs_lsn_t *commitlsnp)
1254 xfs_log_iovec_t reg[1];
1256 reg[0].i_addr = NULL;
1258 XLOG_VEC_SET_TYPE(®[0], XLOG_REG_TYPE_COMMIT);
1260 ASSERT_ALWAYS(iclog);
1261 if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1262 iclog, XLOG_COMMIT_TRANS))) {
1263 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1266 } /* xlog_commit_record */
1270 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1271 * log space. This code pushes on the lsn which would supposedly free up
1272 * the 25% which we want to leave free. We may need to adopt a policy which
1273 * pushes on an lsn which is further along in the log once we reach the high
1274 * water mark. In this manner, we would be creating a low water mark.
1277 xlog_grant_push_ail(xfs_mount_t *mp,
1280 xlog_t *log = mp->m_log; /* pointer to the log */
1281 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1282 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1283 int free_blocks; /* free blocks left to write to */
1284 int free_bytes; /* free bytes left to write to */
1285 int threshold_block; /* block in lsn we'd like to be at */
1286 int threshold_cycle; /* lsn cycle we'd like to be at */
1290 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1292 s = GRANT_LOCK(log);
1293 free_bytes = xlog_space_left(log,
1294 log->l_grant_reserve_cycle,
1295 log->l_grant_reserve_bytes);
1296 tail_lsn = log->l_tail_lsn;
1297 free_blocks = BTOBBT(free_bytes);
1300 * Set the threshold for the minimum number of free blocks in the
1301 * log to the maximum of what the caller needs, one quarter of the
1302 * log, and 256 blocks.
1304 free_threshold = BTOBB(need_bytes);
1305 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1306 free_threshold = MAX(free_threshold, 256);
1307 if (free_blocks < free_threshold) {
1308 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1309 threshold_cycle = CYCLE_LSN(tail_lsn);
1310 if (threshold_block >= log->l_logBBsize) {
1311 threshold_block -= log->l_logBBsize;
1312 threshold_cycle += 1;
1314 ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1317 /* Don't pass in an lsn greater than the lsn of the last
1318 * log record known to be on disk.
1320 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1321 threshold_lsn = log->l_last_sync_lsn;
1323 GRANT_UNLOCK(log, s);
1326 * Get the transaction layer to kick the dirty buffers out to
1327 * disk asynchronously. No point in trying to do this if
1328 * the filesystem is shutting down.
1330 if (threshold_lsn &&
1331 !XLOG_FORCED_SHUTDOWN(log))
1332 xfs_trans_push_ail(mp, threshold_lsn);
1333 } /* xlog_grant_push_ail */
1337 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1338 * fashion. Previously, we should have moved the current iclog
1339 * ptr in the log to point to the next available iclog. This allows further
1340 * write to continue while this code syncs out an iclog ready to go.
1341 * Before an in-core log can be written out, the data section must be scanned
1342 * to save away the 1st word of each BBSIZE block into the header. We replace
1343 * it with the current cycle count. Each BBSIZE block is tagged with the
1344 * cycle count because there in an implicit assumption that drives will
1345 * guarantee that entire 512 byte blocks get written at once. In other words,
1346 * we can't have part of a 512 byte block written and part not written. By
1347 * tagging each block, we will know which blocks are valid when recovering
1348 * after an unclean shutdown.
1350 * This routine is single threaded on the iclog. No other thread can be in
1351 * this routine with the same iclog. Changing contents of iclog can there-
1352 * fore be done without grabbing the state machine lock. Updating the global
1353 * log will require grabbing the lock though.
1355 * The entire log manager uses a logical block numbering scheme. Only
1356 * log_sync (and then only bwrite()) know about the fact that the log may
1357 * not start with block zero on a given device. The log block start offset
1358 * is added immediately before calling bwrite().
1362 xlog_sync(xlog_t *log,
1363 xlog_in_core_t *iclog)
1365 xfs_caddr_t dptr; /* pointer to byte sized element */
1368 uint count; /* byte count of bwrite */
1369 uint count_init; /* initial count before roundup */
1370 int roundoff; /* roundoff to BB or stripe */
1371 int split = 0; /* split write into two regions */
1374 int v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1376 XFS_STATS_INC(xs_log_writes);
1377 ASSERT(iclog->ic_refcnt == 0);
1379 /* Add for LR header */
1380 count_init = log->l_iclog_hsize + iclog->ic_offset;
1382 /* Round out the log write size */
1383 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1384 /* we have a v2 stripe unit to use */
1385 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1387 count = BBTOB(BTOBB(count_init));
1389 roundoff = count - count_init;
1390 ASSERT(roundoff >= 0);
1391 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1392 roundoff < log->l_mp->m_sb.sb_logsunit)
1394 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1395 roundoff < BBTOB(1)));
1397 /* move grant heads by roundoff in sync */
1398 s = GRANT_LOCK(log);
1399 xlog_grant_add_space(log, roundoff);
1400 GRANT_UNLOCK(log, s);
1402 /* put cycle number in every block */
1403 xlog_pack_data(log, iclog, roundoff);
1405 /* real byte length */
1407 INT_SET(iclog->ic_header.h_len,
1409 iclog->ic_offset + roundoff);
1411 INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1414 /* put ops count in correct order */
1415 ops = iclog->ic_header.h_num_logops;
1416 INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1419 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1420 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1421 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1423 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1425 /* Do we need to split this write into 2 parts? */
1426 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1427 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1428 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1429 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1431 iclog->ic_bwritecnt = 1;
1433 XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1434 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
1438 * Do an ordered write for the log block.
1440 * It may not be needed to flush the first split block in the log wrap
1441 * case, but do it anyways to be safe -AK
1443 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1444 XFS_BUF_ORDERED(bp);
1446 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1447 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1449 xlog_verify_iclog(log, iclog, count, B_TRUE);
1451 /* account for log which doesn't start at block #0 */
1452 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1454 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1459 if ((error = XFS_bwrite(bp))) {
1460 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1465 bp = iclog->ic_log->l_xbuf;
1466 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1468 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1469 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1470 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1471 (__psint_t)count), split);
1472 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1475 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1476 XFS_BUF_ORDERED(bp);
1477 dptr = XFS_BUF_PTR(bp);
1479 * Bump the cycle numbers at the start of each block
1480 * since this part of the buffer is at the start of
1481 * a new cycle. Watch out for the header magic number
1484 for (i=0; i<split; i += BBSIZE) {
1485 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1486 if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1487 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1491 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1492 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1494 /* account for internal log which doesn't start at block #0 */
1495 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1497 if ((error = XFS_bwrite(bp))) {
1498 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1499 bp, XFS_BUF_ADDR(bp));
1508 * Deallocate a log structure
1511 xlog_dealloc_log(xlog_t *log)
1513 xlog_in_core_t *iclog, *next_iclog;
1514 xlog_ticket_t *tic, *next_tic;
1518 iclog = log->l_iclog;
1519 for (i=0; i<log->l_iclog_bufs; i++) {
1520 sv_destroy(&iclog->ic_forcesema);
1521 sv_destroy(&iclog->ic_writesema);
1522 xfs_buf_free(iclog->ic_bp);
1523 #ifdef XFS_LOG_TRACE
1524 if (iclog->ic_trace != NULL) {
1525 ktrace_free(iclog->ic_trace);
1528 next_iclog = iclog->ic_next;
1529 kmem_free(iclog->hic_data, log->l_iclog_size);
1530 kmem_free(iclog, sizeof(xlog_in_core_t));
1533 freesema(&log->l_flushsema);
1534 spinlock_destroy(&log->l_icloglock);
1535 spinlock_destroy(&log->l_grant_lock);
1537 /* XXXsup take a look at this again. */
1538 if ((log->l_ticket_cnt != log->l_ticket_tcnt) &&
1539 !XLOG_FORCED_SHUTDOWN(log)) {
1540 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1541 "xlog_dealloc_log: (cnt: %d, total: %d)",
1542 log->l_ticket_cnt, log->l_ticket_tcnt);
1543 /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1546 tic = log->l_unmount_free;
1548 next_tic = tic->t_next;
1549 kmem_free(tic, NBPP);
1553 xfs_buf_free(log->l_xbuf);
1554 #ifdef XFS_LOG_TRACE
1555 if (log->l_trace != NULL) {
1556 ktrace_free(log->l_trace);
1558 if (log->l_grant_trace != NULL) {
1559 ktrace_free(log->l_grant_trace);
1562 log->l_mp->m_log = NULL;
1563 kmem_free(log, sizeof(xlog_t));
1564 } /* xlog_dealloc_log */
1567 * Update counters atomically now that memcpy is done.
1571 xlog_state_finish_copy(xlog_t *log,
1572 xlog_in_core_t *iclog,
1580 iclog->ic_header.h_num_logops += record_cnt;
1581 iclog->ic_offset += copy_bytes;
1584 } /* xlog_state_finish_copy */
1590 * print out info relating to regions written which consume
1594 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1597 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1599 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1600 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1621 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1664 xfs_fs_cmn_err(CE_WARN, mp,
1665 "xfs_log_write: reservation summary:\n"
1666 " trans type = %s (%u)\n"
1667 " unit res = %d bytes\n"
1668 " current res = %d bytes\n"
1669 " total reg = %u bytes (o/flow = %u bytes)\n"
1670 " ophdrs = %u (ophdr space = %u bytes)\n"
1671 " ophdr + reg = %u bytes\n"
1672 " num regions = %u\n",
1673 ((ticket->t_trans_type <= 0 ||
1674 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1675 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1676 ticket->t_trans_type,
1679 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1680 ticket->t_res_num_ophdrs, ophdr_spc,
1681 ticket->t_res_arr_sum +
1682 ticket->t_res_o_flow + ophdr_spc,
1685 for (i = 0; i < ticket->t_res_num; i++) {
1686 uint r_type = ticket->t_res_arr[i].r_type;
1688 "region[%u]: %s - %u bytes\n",
1690 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1691 "bad-rtype" : res_type_str[r_type-1]),
1692 ticket->t_res_arr[i].r_len);
1697 * Write some region out to in-core log
1699 * This will be called when writing externally provided regions or when
1700 * writing out a commit record for a given transaction.
1702 * General algorithm:
1703 * 1. Find total length of this write. This may include adding to the
1704 * lengths passed in.
1705 * 2. Check whether we violate the tickets reservation.
1706 * 3. While writing to this iclog
1707 * A. Reserve as much space in this iclog as can get
1708 * B. If this is first write, save away start lsn
1709 * C. While writing this region:
1710 * 1. If first write of transaction, write start record
1711 * 2. Write log operation header (header per region)
1712 * 3. Find out if we can fit entire region into this iclog
1713 * 4. Potentially, verify destination memcpy ptr
1714 * 5. Memcpy (partial) region
1715 * 6. If partial copy, release iclog; otherwise, continue
1716 * copying more regions into current iclog
1717 * 4. Mark want sync bit (in simulation mode)
1718 * 5. Release iclog for potential flush to on-disk log.
1721 * 1. Panic if reservation is overrun. This should never happen since
1722 * reservation amounts are generated internal to the filesystem.
1724 * 1. Tickets are single threaded data structures.
1725 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1726 * syncing routine. When a single log_write region needs to span
1727 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1728 * on all log operation writes which don't contain the end of the
1729 * region. The XLOG_END_TRANS bit is used for the in-core log
1730 * operation which contains the end of the continued log_write region.
1731 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1732 * we don't really know exactly how much space will be used. As a result,
1733 * we don't update ic_offset until the end when we know exactly how many
1734 * bytes have been written out.
1737 xlog_write(xfs_mount_t * mp,
1738 xfs_log_iovec_t reg[],
1740 xfs_log_ticket_t tic,
1741 xfs_lsn_t *start_lsn,
1742 xlog_in_core_t **commit_iclog,
1745 xlog_t *log = mp->m_log;
1746 xlog_ticket_t *ticket = (xlog_ticket_t *)tic;
1747 xlog_op_header_t *logop_head; /* ptr to log operation header */
1748 xlog_in_core_t *iclog; /* ptr to current in-core log */
1749 __psint_t ptr; /* copy address into data region */
1750 int len; /* # xlog_write() bytes 2 still copy */
1751 int index; /* region index currently copying */
1752 int log_offset; /* offset (from 0) into data region */
1753 int start_rec_copy; /* # bytes to copy for start record */
1754 int partial_copy; /* did we split a region? */
1755 int partial_copy_len;/* # bytes copied if split region */
1756 int need_copy; /* # bytes need to memcpy this region */
1757 int copy_len; /* # bytes actually memcpy'ing */
1758 int copy_off; /* # bytes from entry start */
1759 int contwr; /* continued write of in-core log? */
1761 int record_cnt = 0, data_cnt = 0;
1763 partial_copy_len = partial_copy = 0;
1765 /* Calculate potential maximum space. Each region gets its own
1766 * xlog_op_header_t and may need to be double word aligned.
1769 if (ticket->t_flags & XLOG_TIC_INITED) { /* acct for start rec of xact */
1770 len += sizeof(xlog_op_header_t);
1771 XLOG_TIC_ADD_OPHDR(ticket);
1774 for (index = 0; index < nentries; index++) {
1775 len += sizeof(xlog_op_header_t); /* each region gets >= 1 */
1776 XLOG_TIC_ADD_OPHDR(ticket);
1777 len += reg[index].i_len;
1778 XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1780 contwr = *start_lsn = 0;
1782 if (ticket->t_curr_res < len) {
1783 xlog_print_tic_res(mp, ticket);
1786 "xfs_log_write: reservation ran out. Need to up reservation");
1788 /* Customer configurable panic */
1789 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1790 "xfs_log_write: reservation ran out. Need to up reservation");
1791 /* If we did not panic, shutdown the filesystem */
1792 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1795 ticket->t_curr_res -= len;
1797 for (index = 0; index < nentries; ) {
1798 if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1799 &contwr, &log_offset)))
1802 ASSERT(log_offset <= iclog->ic_size - 1);
1803 ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1805 /* start_lsn is the first lsn written to. That's all we need. */
1807 *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1809 /* This loop writes out as many regions as can fit in the amount
1810 * of space which was allocated by xlog_state_get_iclog_space().
1812 while (index < nentries) {
1813 ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1814 ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1817 /* If first write for transaction, insert start record.
1818 * We can't be trying to commit if we are inited. We can't
1819 * have any "partial_copy" if we are inited.
1821 if (ticket->t_flags & XLOG_TIC_INITED) {
1822 logop_head = (xlog_op_header_t *)ptr;
1823 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1824 logop_head->oh_clientid = ticket->t_clientid;
1825 logop_head->oh_len = 0;
1826 logop_head->oh_flags = XLOG_START_TRANS;
1827 logop_head->oh_res2 = 0;
1828 ticket->t_flags &= ~XLOG_TIC_INITED; /* clear bit */
1831 start_rec_copy = sizeof(xlog_op_header_t);
1832 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1835 /* Copy log operation header directly into data section */
1836 logop_head = (xlog_op_header_t *)ptr;
1837 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1838 logop_head->oh_clientid = ticket->t_clientid;
1839 logop_head->oh_res2 = 0;
1841 /* header copied directly */
1842 xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1844 /* are we copying a commit or unmount record? */
1845 logop_head->oh_flags = flags;
1848 * We've seen logs corrupted with bad transaction client
1849 * ids. This makes sure that XFS doesn't generate them on.
1850 * Turn this into an EIO and shut down the filesystem.
1852 switch (logop_head->oh_clientid) {
1853 case XFS_TRANSACTION:
1858 xfs_fs_cmn_err(CE_WARN, mp,
1859 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1860 logop_head->oh_clientid, tic);
1861 return XFS_ERROR(EIO);
1864 /* Partial write last time? => (partial_copy != 0)
1865 * need_copy is the amount we'd like to copy if everything could
1866 * fit in the current memcpy.
1868 need_copy = reg[index].i_len - partial_copy_len;
1870 copy_off = partial_copy_len;
1871 if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1872 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1874 logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1875 partial_copy_len = partial_copy = 0;
1876 } else { /* partial write */
1877 copy_len = iclog->ic_size - log_offset;
1878 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1879 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1881 logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1882 partial_copy_len += copy_len;
1884 len += sizeof(xlog_op_header_t); /* from splitting of region */
1885 /* account for new log op header */
1886 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1887 XLOG_TIC_ADD_OPHDR(ticket);
1889 xlog_verify_dest_ptr(log, ptr);
1892 ASSERT(copy_len >= 0);
1893 memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1894 xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1896 /* make copy_len total bytes copied, including headers */
1897 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1899 data_cnt += contwr ? copy_len : 0;
1900 if (partial_copy) { /* copied partial region */
1901 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1902 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1903 record_cnt = data_cnt = 0;
1904 if ((error = xlog_state_release_iclog(log, iclog)))
1906 break; /* don't increment index */
1907 } else { /* copied entire region */
1909 partial_copy_len = partial_copy = 0;
1911 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1912 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1913 record_cnt = data_cnt = 0;
1914 xlog_state_want_sync(log, iclog);
1916 ASSERT(flags & XLOG_COMMIT_TRANS);
1917 *commit_iclog = iclog;
1918 } else if ((error = xlog_state_release_iclog(log, iclog)))
1920 if (index == nentries)
1921 return 0; /* we are done */
1925 } /* if (partial_copy) */
1926 } /* while (index < nentries) */
1927 } /* for (index = 0; index < nentries; ) */
1930 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1932 ASSERT(flags & XLOG_COMMIT_TRANS);
1933 *commit_iclog = iclog;
1936 return xlog_state_release_iclog(log, iclog);
1940 /*****************************************************************************
1942 * State Machine functions
1944 *****************************************************************************
1947 /* Clean iclogs starting from the head. This ordering must be
1948 * maintained, so an iclog doesn't become ACTIVE beyond one that
1949 * is SYNCING. This is also required to maintain the notion that we use
1950 * a counting semaphore to hold off would be writers to the log when every
1951 * iclog is trying to sync to disk.
1953 * State Change: DIRTY -> ACTIVE
1956 xlog_state_clean_log(xlog_t *log)
1958 xlog_in_core_t *iclog;
1961 iclog = log->l_iclog;
1963 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1964 iclog->ic_state = XLOG_STATE_ACTIVE;
1965 iclog->ic_offset = 0;
1966 iclog->ic_callback = NULL; /* don't need to free */
1968 * If the number of ops in this iclog indicate it just
1969 * contains the dummy transaction, we can
1970 * change state into IDLE (the second time around).
1971 * Otherwise we should change the state into
1973 * We don't need to cover the dummy.
1976 (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1980 * We have two dirty iclogs so start over
1981 * This could also be num of ops indicates
1982 * this is not the dummy going out.
1986 iclog->ic_header.h_num_logops = 0;
1987 memset(iclog->ic_header.h_cycle_data, 0,
1988 sizeof(iclog->ic_header.h_cycle_data));
1989 iclog->ic_header.h_lsn = 0;
1990 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1993 break; /* stop cleaning */
1994 iclog = iclog->ic_next;
1995 } while (iclog != log->l_iclog);
1997 /* log is locked when we are called */
1999 * Change state for the dummy log recording.
2000 * We usually go to NEED. But we go to NEED2 if the changed indicates
2001 * we are done writing the dummy record.
2002 * If we are done with the second dummy recored (DONE2), then
2006 switch (log->l_covered_state) {
2007 case XLOG_STATE_COVER_IDLE:
2008 case XLOG_STATE_COVER_NEED:
2009 case XLOG_STATE_COVER_NEED2:
2010 log->l_covered_state = XLOG_STATE_COVER_NEED;
2013 case XLOG_STATE_COVER_DONE:
2015 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2017 log->l_covered_state = XLOG_STATE_COVER_NEED;
2020 case XLOG_STATE_COVER_DONE2:
2022 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2024 log->l_covered_state = XLOG_STATE_COVER_NEED;
2031 } /* xlog_state_clean_log */
2034 xlog_get_lowest_lsn(
2037 xlog_in_core_t *lsn_log;
2038 xfs_lsn_t lowest_lsn, lsn;
2040 lsn_log = log->l_iclog;
2043 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2044 lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2045 if ((lsn && !lowest_lsn) ||
2046 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2050 lsn_log = lsn_log->ic_next;
2051 } while (lsn_log != log->l_iclog);
2057 xlog_state_do_callback(
2060 xlog_in_core_t *ciclog)
2062 xlog_in_core_t *iclog;
2063 xlog_in_core_t *first_iclog; /* used to know when we've
2064 * processed all iclogs once */
2065 xfs_log_callback_t *cb, *cb_next;
2067 xfs_lsn_t lowest_lsn;
2068 int ioerrors; /* counter: iclogs with errors */
2069 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2070 int funcdidcallbacks; /* flag: function did callbacks */
2071 int repeats; /* for issuing console warnings if
2072 * looping too many times */
2076 first_iclog = iclog = log->l_iclog;
2078 funcdidcallbacks = 0;
2083 * Scan all iclogs starting with the one pointed to by the
2084 * log. Reset this starting point each time the log is
2085 * unlocked (during callbacks).
2087 * Keep looping through iclogs until one full pass is made
2088 * without running any callbacks.
2090 first_iclog = log->l_iclog;
2091 iclog = log->l_iclog;
2092 loopdidcallbacks = 0;
2097 /* skip all iclogs in the ACTIVE & DIRTY states */
2098 if (iclog->ic_state &
2099 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2100 iclog = iclog->ic_next;
2105 * Between marking a filesystem SHUTDOWN and stopping
2106 * the log, we do flush all iclogs to disk (if there
2107 * wasn't a log I/O error). So, we do want things to
2108 * go smoothly in case of just a SHUTDOWN w/o a
2111 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2113 * Can only perform callbacks in order. Since
2114 * this iclog is not in the DONE_SYNC/
2115 * DO_CALLBACK state, we skip the rest and
2116 * just try to clean up. If we set our iclog
2117 * to DO_CALLBACK, we will not process it when
2118 * we retry since a previous iclog is in the
2119 * CALLBACK and the state cannot change since
2120 * we are holding the LOG_LOCK.
2122 if (!(iclog->ic_state &
2123 (XLOG_STATE_DONE_SYNC |
2124 XLOG_STATE_DO_CALLBACK))) {
2125 if (ciclog && (ciclog->ic_state ==
2126 XLOG_STATE_DONE_SYNC)) {
2127 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2132 * We now have an iclog that is in either the
2133 * DO_CALLBACK or DONE_SYNC states. The other
2134 * states (WANT_SYNC, SYNCING, or CALLBACK were
2135 * caught by the above if and are going to
2136 * clean (i.e. we aren't doing their callbacks)
2141 * We will do one more check here to see if we
2142 * have chased our tail around.
2145 lowest_lsn = xlog_get_lowest_lsn(log);
2149 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2151 iclog = iclog->ic_next;
2152 continue; /* Leave this iclog for
2156 iclog->ic_state = XLOG_STATE_CALLBACK;
2160 /* l_last_sync_lsn field protected by
2161 * GRANT_LOCK. Don't worry about iclog's lsn.
2162 * No one else can be here except us.
2164 s = GRANT_LOCK(log);
2166 log->l_last_sync_lsn,
2167 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2169 log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2170 GRANT_UNLOCK(log, s);
2173 * Keep processing entries in the callback list
2174 * until we come around and it is empty. We
2175 * need to atomically see that the list is
2176 * empty and change the state to DIRTY so that
2177 * we don't miss any more callbacks being added.
2183 cb = iclog->ic_callback;
2186 iclog->ic_callback_tail = &(iclog->ic_callback);
2187 iclog->ic_callback = NULL;
2190 /* perform callbacks in the order given */
2191 for (; cb != 0; cb = cb_next) {
2192 cb_next = cb->cb_next;
2193 cb->cb_func(cb->cb_arg, aborted);
2196 cb = iclog->ic_callback;
2202 ASSERT(iclog->ic_callback == 0);
2203 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2204 iclog->ic_state = XLOG_STATE_DIRTY;
2207 * Transition from DIRTY to ACTIVE if applicable.
2208 * NOP if STATE_IOERROR.
2210 xlog_state_clean_log(log);
2212 /* wake up threads waiting in xfs_log_force() */
2213 sv_broadcast(&iclog->ic_forcesema);
2215 iclog = iclog->ic_next;
2216 } while (first_iclog != iclog);
2217 if (repeats && (repeats % 10) == 0) {
2218 xfs_fs_cmn_err(CE_WARN, log->l_mp,
2219 "xlog_state_do_callback: looping %d", repeats);
2221 } while (!ioerrors && loopdidcallbacks);
2224 * make one last gasp attempt to see if iclogs are being left in
2228 if (funcdidcallbacks) {
2229 first_iclog = iclog = log->l_iclog;
2231 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2233 * Terminate the loop if iclogs are found in states
2234 * which will cause other threads to clean up iclogs.
2236 * SYNCING - i/o completion will go through logs
2237 * DONE_SYNC - interrupt thread should be waiting for
2239 * IOERROR - give up hope all ye who enter here
2241 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2242 iclog->ic_state == XLOG_STATE_SYNCING ||
2243 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2244 iclog->ic_state == XLOG_STATE_IOERROR )
2246 iclog = iclog->ic_next;
2247 } while (first_iclog != iclog);
2251 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2252 flushcnt = log->l_flushcnt;
2253 log->l_flushcnt = 0;
2257 vsema(&log->l_flushsema);
2258 } /* xlog_state_do_callback */
2262 * Finish transitioning this iclog to the dirty state.
2264 * Make sure that we completely execute this routine only when this is
2265 * the last call to the iclog. There is a good chance that iclog flushes,
2266 * when we reach the end of the physical log, get turned into 2 separate
2267 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2268 * routine. By using the reference count bwritecnt, we guarantee that only
2269 * the second completion goes through.
2271 * Callbacks could take time, so they are done outside the scope of the
2272 * global state machine log lock. Assume that the calls to cvsema won't
2273 * take a long time. At least we know it won't sleep.
2276 xlog_state_done_syncing(
2277 xlog_in_core_t *iclog,
2280 xlog_t *log = iclog->ic_log;
2285 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2286 iclog->ic_state == XLOG_STATE_IOERROR);
2287 ASSERT(iclog->ic_refcnt == 0);
2288 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2292 * If we got an error, either on the first buffer, or in the case of
2293 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2294 * and none should ever be attempted to be written to disk
2297 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2298 if (--iclog->ic_bwritecnt == 1) {
2302 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2306 * Someone could be sleeping prior to writing out the next
2307 * iclog buffer, we wake them all, one will get to do the
2308 * I/O, the others get to wait for the result.
2310 sv_broadcast(&iclog->ic_writesema);
2312 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2313 } /* xlog_state_done_syncing */
2317 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2318 * sleep. The flush semaphore is set to the number of in-core buffers and
2319 * decremented around disk syncing. Therefore, if all buffers are syncing,
2320 * this semaphore will cause new writes to sleep until a sync completes.
2321 * Otherwise, this code just does p() followed by v(). This approximates
2322 * a sleep/wakeup except we can't race.
2324 * The in-core logs are used in a circular fashion. They are not used
2325 * out-of-order even when an iclog past the head is free.
2328 * * log_offset where xlog_write() can start writing into the in-core
2330 * * in-core log pointer to which xlog_write() should write.
2331 * * boolean indicating this is a continued write to an in-core log.
2332 * If this is the last write, then the in-core log's offset field
2333 * needs to be incremented, depending on the amount of data which
2337 xlog_state_get_iclog_space(xlog_t *log,
2339 xlog_in_core_t **iclogp,
2340 xlog_ticket_t *ticket,
2341 int *continued_write,
2346 xlog_rec_header_t *head;
2347 xlog_in_core_t *iclog;
2352 if (XLOG_FORCED_SHUTDOWN(log)) {
2354 return XFS_ERROR(EIO);
2357 iclog = log->l_iclog;
2358 if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2361 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2362 XFS_STATS_INC(xs_log_noiclogs);
2363 /* Ensure that log writes happen */
2364 psema(&log->l_flushsema, PINOD);
2367 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2368 head = &iclog->ic_header;
2370 iclog->ic_refcnt++; /* prevents sync */
2371 log_offset = iclog->ic_offset;
2373 /* On the 1st write to an iclog, figure out lsn. This works
2374 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2375 * committing to. If the offset is set, that's how many blocks
2378 if (log_offset == 0) {
2379 ticket->t_curr_res -= log->l_iclog_hsize;
2380 XLOG_TIC_ADD_REGION(ticket,
2382 XLOG_REG_TYPE_LRHEADER);
2383 INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2384 ASSIGN_LSN(head->h_lsn, log);
2385 ASSERT(log->l_curr_block >= 0);
2388 /* If there is enough room to write everything, then do it. Otherwise,
2389 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2390 * bit is on, so this will get flushed out. Don't update ic_offset
2391 * until you know exactly how many bytes get copied. Therefore, wait
2392 * until later to update ic_offset.
2394 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2395 * can fit into remaining data section.
2397 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2398 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2400 /* If I'm the only one writing to this iclog, sync it to disk */
2401 if (iclog->ic_refcnt == 1) {
2403 if ((error = xlog_state_release_iclog(log, iclog)))
2412 /* Do we have enough room to write the full amount in the remainder
2413 * of this iclog? Or must we continue a write on the next iclog and
2414 * mark this iclog as completely taken? In the case where we switch
2415 * iclogs (to mark it taken), this particular iclog will release/sync
2416 * to disk in xlog_write().
2418 if (len <= iclog->ic_size - iclog->ic_offset) {
2419 *continued_write = 0;
2420 iclog->ic_offset += len;
2422 *continued_write = 1;
2423 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2427 ASSERT(iclog->ic_offset <= iclog->ic_size);
2430 *logoffsetp = log_offset;
2432 } /* xlog_state_get_iclog_space */
2435 * Atomically get the log space required for a log ticket.
2437 * Once a ticket gets put onto the reserveq, it will only return after
2438 * the needed reservation is satisfied.
2441 xlog_grant_log_space(xlog_t *log,
2453 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2454 panic("grant Recovery problem");
2457 /* Is there space or do we need to sleep? */
2458 s = GRANT_LOCK(log);
2459 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2461 /* something is already sleeping; insert new transaction at end */
2462 if (log->l_reserve_headq) {
2463 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2464 xlog_trace_loggrant(log, tic,
2465 "xlog_grant_log_space: sleep 1");
2467 * Gotta check this before going to sleep, while we're
2468 * holding the grant lock.
2470 if (XLOG_FORCED_SHUTDOWN(log))
2473 XFS_STATS_INC(xs_sleep_logspace);
2474 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2476 * If we got an error, and the filesystem is shutting down,
2477 * we'll catch it down below. So just continue...
2479 xlog_trace_loggrant(log, tic,
2480 "xlog_grant_log_space: wake 1");
2481 s = GRANT_LOCK(log);
2483 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2484 need_bytes = tic->t_unit_res*tic->t_ocnt;
2486 need_bytes = tic->t_unit_res;
2489 if (XLOG_FORCED_SHUTDOWN(log))
2492 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2493 log->l_grant_reserve_bytes);
2494 if (free_bytes < need_bytes) {
2495 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2496 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2497 xlog_trace_loggrant(log, tic,
2498 "xlog_grant_log_space: sleep 2");
2499 XFS_STATS_INC(xs_sleep_logspace);
2500 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2502 if (XLOG_FORCED_SHUTDOWN(log)) {
2503 s = GRANT_LOCK(log);
2507 xlog_trace_loggrant(log, tic,
2508 "xlog_grant_log_space: wake 2");
2509 xlog_grant_push_ail(log->l_mp, need_bytes);
2510 s = GRANT_LOCK(log);
2512 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2513 xlog_del_ticketq(&log->l_reserve_headq, tic);
2515 /* we've got enough space */
2516 xlog_grant_add_space(log, need_bytes);
2518 tail_lsn = log->l_tail_lsn;
2520 * Check to make sure the grant write head didn't just over lap the
2521 * tail. If the cycles are the same, we can't be overlapping.
2522 * Otherwise, make sure that the cycles differ by exactly one and
2523 * check the byte count.
2525 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2526 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2527 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2530 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2531 xlog_verify_grant_head(log, 1);
2532 GRANT_UNLOCK(log, s);
2536 if (tic->t_flags & XLOG_TIC_IN_Q)
2537 xlog_del_ticketq(&log->l_reserve_headq, tic);
2538 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2540 * If we are failing, make sure the ticket doesn't have any
2541 * current reservations. We don't want to add this back when
2542 * the ticket/transaction gets cancelled.
2544 tic->t_curr_res = 0;
2545 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2546 GRANT_UNLOCK(log, s);
2547 return XFS_ERROR(EIO);
2548 } /* xlog_grant_log_space */
2552 * Replenish the byte reservation required by moving the grant write head.
2557 xlog_regrant_write_log_space(xlog_t *log,
2561 int free_bytes, need_bytes;
2562 xlog_ticket_t *ntic;
2567 tic->t_curr_res = tic->t_unit_res;
2568 XLOG_TIC_RESET_RES(tic);
2574 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2575 panic("regrant Recovery problem");
2578 s = GRANT_LOCK(log);
2579 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2581 if (XLOG_FORCED_SHUTDOWN(log))
2584 /* If there are other waiters on the queue then give them a
2585 * chance at logspace before us. Wake up the first waiters,
2586 * if we do not wake up all the waiters then go to sleep waiting
2587 * for more free space, otherwise try to get some space for
2591 if ((ntic = log->l_write_headq)) {
2592 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2593 log->l_grant_write_bytes);
2595 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2597 if (free_bytes < ntic->t_unit_res)
2599 free_bytes -= ntic->t_unit_res;
2600 sv_signal(&ntic->t_sema);
2601 ntic = ntic->t_next;
2602 } while (ntic != log->l_write_headq);
2604 if (ntic != log->l_write_headq) {
2605 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2606 xlog_ins_ticketq(&log->l_write_headq, tic);
2608 xlog_trace_loggrant(log, tic,
2609 "xlog_regrant_write_log_space: sleep 1");
2610 XFS_STATS_INC(xs_sleep_logspace);
2611 sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2612 &log->l_grant_lock, s);
2614 /* If we're shutting down, this tic is already
2616 if (XLOG_FORCED_SHUTDOWN(log)) {
2617 s = GRANT_LOCK(log);
2621 xlog_trace_loggrant(log, tic,
2622 "xlog_regrant_write_log_space: wake 1");
2623 xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2624 s = GRANT_LOCK(log);
2628 need_bytes = tic->t_unit_res;
2631 if (XLOG_FORCED_SHUTDOWN(log))
2634 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2635 log->l_grant_write_bytes);
2636 if (free_bytes < need_bytes) {
2637 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2638 xlog_ins_ticketq(&log->l_write_headq, tic);
2639 XFS_STATS_INC(xs_sleep_logspace);
2640 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2642 /* If we're shutting down, this tic is already off the queue */
2643 if (XLOG_FORCED_SHUTDOWN(log)) {
2644 s = GRANT_LOCK(log);
2648 xlog_trace_loggrant(log, tic,
2649 "xlog_regrant_write_log_space: wake 2");
2650 xlog_grant_push_ail(log->l_mp, need_bytes);
2651 s = GRANT_LOCK(log);
2653 } else if (tic->t_flags & XLOG_TIC_IN_Q)
2654 xlog_del_ticketq(&log->l_write_headq, tic);
2656 /* we've got enough space */
2657 xlog_grant_add_space_write(log, need_bytes);
2659 tail_lsn = log->l_tail_lsn;
2660 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2661 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2662 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2666 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2667 xlog_verify_grant_head(log, 1);
2668 GRANT_UNLOCK(log, s);
2673 if (tic->t_flags & XLOG_TIC_IN_Q)
2674 xlog_del_ticketq(&log->l_reserve_headq, tic);
2675 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2677 * If we are failing, make sure the ticket doesn't have any
2678 * current reservations. We don't want to add this back when
2679 * the ticket/transaction gets cancelled.
2681 tic->t_curr_res = 0;
2682 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2683 GRANT_UNLOCK(log, s);
2684 return XFS_ERROR(EIO);
2685 } /* xlog_regrant_write_log_space */
2688 /* The first cnt-1 times through here we don't need to
2689 * move the grant write head because the permanent
2690 * reservation has reserved cnt times the unit amount.
2691 * Release part of current permanent unit reservation and
2692 * reset current reservation to be one units worth. Also
2693 * move grant reservation head forward.
2696 xlog_regrant_reserve_log_space(xlog_t *log,
2697 xlog_ticket_t *ticket)
2701 xlog_trace_loggrant(log, ticket,
2702 "xlog_regrant_reserve_log_space: enter");
2703 if (ticket->t_cnt > 0)
2706 s = GRANT_LOCK(log);
2707 xlog_grant_sub_space(log, ticket->t_curr_res);
2708 ticket->t_curr_res = ticket->t_unit_res;
2709 XLOG_TIC_RESET_RES(ticket);
2710 xlog_trace_loggrant(log, ticket,
2711 "xlog_regrant_reserve_log_space: sub current res");
2712 xlog_verify_grant_head(log, 1);
2714 /* just return if we still have some of the pre-reserved space */
2715 if (ticket->t_cnt > 0) {
2716 GRANT_UNLOCK(log, s);
2720 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2721 xlog_trace_loggrant(log, ticket,
2722 "xlog_regrant_reserve_log_space: exit");
2723 xlog_verify_grant_head(log, 0);
2724 GRANT_UNLOCK(log, s);
2725 ticket->t_curr_res = ticket->t_unit_res;
2726 XLOG_TIC_RESET_RES(ticket);
2727 } /* xlog_regrant_reserve_log_space */
2731 * Give back the space left from a reservation.
2733 * All the information we need to make a correct determination of space left
2734 * is present. For non-permanent reservations, things are quite easy. The
2735 * count should have been decremented to zero. We only need to deal with the
2736 * space remaining in the current reservation part of the ticket. If the
2737 * ticket contains a permanent reservation, there may be left over space which
2738 * needs to be released. A count of N means that N-1 refills of the current
2739 * reservation can be done before we need to ask for more space. The first
2740 * one goes to fill up the first current reservation. Once we run out of
2741 * space, the count will stay at zero and the only space remaining will be
2742 * in the current reservation field.
2745 xlog_ungrant_log_space(xlog_t *log,
2746 xlog_ticket_t *ticket)
2750 if (ticket->t_cnt > 0)
2753 s = GRANT_LOCK(log);
2754 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2756 xlog_grant_sub_space(log, ticket->t_curr_res);
2758 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2760 /* If this is a permanent reservation ticket, we may be able to free
2761 * up more space based on the remaining count.
2763 if (ticket->t_cnt > 0) {
2764 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2765 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2768 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2769 xlog_verify_grant_head(log, 1);
2770 GRANT_UNLOCK(log, s);
2771 xfs_log_move_tail(log->l_mp, 1);
2772 } /* xlog_ungrant_log_space */
2776 * Atomically put back used ticket.
2779 xlog_state_put_ticket(xlog_t *log,
2785 xlog_ticket_put(log, tic);
2787 } /* xlog_state_put_ticket */
2790 * Flush iclog to disk if this is the last reference to the given iclog and
2791 * the WANT_SYNC bit is set.
2793 * When this function is entered, the iclog is not necessarily in the
2794 * WANT_SYNC state. It may be sitting around waiting to get filled.
2799 xlog_state_release_iclog(xlog_t *log,
2800 xlog_in_core_t *iclog)
2803 int sync = 0; /* do we sync? */
2805 xlog_assign_tail_lsn(log->l_mp);
2809 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2811 return XFS_ERROR(EIO);
2814 ASSERT(iclog->ic_refcnt > 0);
2815 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2816 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2818 if (--iclog->ic_refcnt == 0 &&
2819 iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2821 iclog->ic_state = XLOG_STATE_SYNCING;
2822 INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2823 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2824 /* cycle incremented when incrementing curr_block */
2830 * We let the log lock go, so it's possible that we hit a log I/O
2831 * error or some other SHUTDOWN condition that marks the iclog
2832 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2833 * this iclog has consistent data, so we ignore IOERROR
2834 * flags after this point.
2837 return xlog_sync(log, iclog);
2841 } /* xlog_state_release_iclog */
2845 * This routine will mark the current iclog in the ring as WANT_SYNC
2846 * and move the current iclog pointer to the next iclog in the ring.
2847 * When this routine is called from xlog_state_get_iclog_space(), the
2848 * exact size of the iclog has not yet been determined. All we know is
2849 * that every data block. We have run out of space in this log record.
2852 xlog_state_switch_iclogs(xlog_t *log,
2853 xlog_in_core_t *iclog,
2856 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2858 eventual_size = iclog->ic_offset;
2859 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2860 INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2861 log->l_prev_block = log->l_curr_block;
2862 log->l_prev_cycle = log->l_curr_cycle;
2864 /* roll log?: ic_offset changed later */
2865 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2867 /* Round up to next log-sunit */
2868 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2869 log->l_mp->m_sb.sb_logsunit > 1) {
2870 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2871 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2874 if (log->l_curr_block >= log->l_logBBsize) {
2875 log->l_curr_cycle++;
2876 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2877 log->l_curr_cycle++;
2878 log->l_curr_block -= log->l_logBBsize;
2879 ASSERT(log->l_curr_block >= 0);
2881 ASSERT(iclog == log->l_iclog);
2882 log->l_iclog = iclog->ic_next;
2883 } /* xlog_state_switch_iclogs */
2887 * Write out all data in the in-core log as of this exact moment in time.
2889 * Data may be written to the in-core log during this call. However,
2890 * we don't guarantee this data will be written out. A change from past
2891 * implementation means this routine will *not* write out zero length LRs.
2893 * Basically, we try and perform an intelligent scan of the in-core logs.
2894 * If we determine there is no flushable data, we just return. There is no
2895 * flushable data if:
2897 * 1. the current iclog is active and has no data; the previous iclog
2898 * is in the active or dirty state.
2899 * 2. the current iclog is drity, and the previous iclog is in the
2900 * active or dirty state.
2902 * We may sleep (call psema) if:
2904 * 1. the current iclog is not in the active nor dirty state.
2905 * 2. the current iclog dirty, and the previous iclog is not in the
2906 * active nor dirty state.
2907 * 3. the current iclog is active, and there is another thread writing
2908 * to this particular iclog.
2909 * 4. a) the current iclog is active and has no other writers
2910 * b) when we return from flushing out this iclog, it is still
2911 * not in the active nor dirty state.
2914 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2916 xlog_in_core_t *iclog;
2922 iclog = log->l_iclog;
2923 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2925 return XFS_ERROR(EIO);
2928 /* If the head iclog is not active nor dirty, we just attach
2929 * ourselves to the head and go to sleep.
2931 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2932 iclog->ic_state == XLOG_STATE_DIRTY) {
2934 * If the head is dirty or (active and empty), then
2935 * we need to look at the previous iclog. If the previous
2936 * iclog is active or dirty we are done. There is nothing
2937 * to sync out. Otherwise, we attach ourselves to the
2938 * previous iclog and go to sleep.
2940 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2941 (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2942 iclog = iclog->ic_prev;
2943 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2944 iclog->ic_state == XLOG_STATE_DIRTY)
2949 if (iclog->ic_refcnt == 0) {
2950 /* We are the only one with access to this
2951 * iclog. Flush it out now. There should
2952 * be a roundoff of zero to show that someone
2953 * has already taken care of the roundoff from
2954 * the previous sync.
2957 lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2958 xlog_state_switch_iclogs(log, iclog, 0);
2961 if (xlog_state_release_iclog(log, iclog))
2962 return XFS_ERROR(EIO);
2965 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2966 iclog->ic_state != XLOG_STATE_DIRTY)
2971 /* Someone else is writing to this iclog.
2972 * Use its call to flush out the data. However,
2973 * the other thread may not force out this LR,
2974 * so we mark it WANT_SYNC.
2976 xlog_state_switch_iclogs(log, iclog, 0);
2982 /* By the time we come around again, the iclog could've been filled
2983 * which would give it another lsn. If we have a new lsn, just
2984 * return because the relevant data has been flushed.
2987 if (flags & XFS_LOG_SYNC) {
2989 * We must check if we're shutting down here, before
2990 * we wait, while we're holding the LOG_LOCK.
2991 * Then we check again after waking up, in case our
2992 * sleep was disturbed by a bad news.
2994 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2996 return XFS_ERROR(EIO);
2998 XFS_STATS_INC(xs_log_force_sleep);
2999 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3001 * No need to grab the log lock here since we're
3002 * only deciding whether or not to return EIO
3003 * and the memory read should be atomic.
3005 if (iclog->ic_state & XLOG_STATE_IOERROR)
3006 return XFS_ERROR(EIO);
3015 } /* xlog_state_sync_all */
3019 * Used by code which implements synchronous log forces.
3021 * Find in-core log with lsn.
3022 * If it is in the DIRTY state, just return.
3023 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3024 * state and go to sleep or return.
3025 * If it is in any other state, go to sleep or return.
3027 * If filesystem activity goes to zero, the iclog will get flushed only by
3031 xlog_state_sync(xlog_t *log,
3036 xlog_in_core_t *iclog;
3037 int already_slept = 0;
3043 iclog = log->l_iclog;
3045 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3047 return XFS_ERROR(EIO);
3051 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3052 iclog = iclog->ic_next;
3056 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3061 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3063 * We sleep here if we haven't already slept (e.g.
3064 * this is the first time we've looked at the correct
3065 * iclog buf) and the buffer before us is going to
3066 * be sync'ed. The reason for this is that if we
3067 * are doing sync transactions here, by waiting for
3068 * the previous I/O to complete, we can allow a few
3069 * more transactions into this iclog before we close
3072 * Otherwise, we mark the buffer WANT_SYNC, and bump
3073 * up the refcnt so we can release the log (which drops
3074 * the ref count). The state switch keeps new transaction
3075 * commits from using this buffer. When the current commits
3076 * finish writing into the buffer, the refcount will drop to
3077 * zero and the buffer will go out then.
3079 if (!already_slept &&
3080 (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3081 XLOG_STATE_SYNCING))) {
3082 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3083 XFS_STATS_INC(xs_log_force_sleep);
3084 sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3085 &log->l_icloglock, s);
3091 xlog_state_switch_iclogs(log, iclog, 0);
3093 if (xlog_state_release_iclog(log, iclog))
3094 return XFS_ERROR(EIO);
3100 if ((flags & XFS_LOG_SYNC) && /* sleep */
3101 !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3104 * Don't wait on the forcesema if we know that we've
3105 * gotten a log write error.
3107 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3109 return XFS_ERROR(EIO);
3111 XFS_STATS_INC(xs_log_force_sleep);
3112 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3114 * No need to grab the log lock here since we're
3115 * only deciding whether or not to return EIO
3116 * and the memory read should be atomic.
3118 if (iclog->ic_state & XLOG_STATE_IOERROR)
3119 return XFS_ERROR(EIO);
3121 } else { /* just return */
3126 } while (iclog != log->l_iclog);
3130 } /* xlog_state_sync */
3134 * Called when we want to mark the current iclog as being ready to sync to
3138 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3144 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3145 xlog_state_switch_iclogs(log, iclog, 0);
3147 ASSERT(iclog->ic_state &
3148 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3152 } /* xlog_state_want_sync */
3156 /*****************************************************************************
3160 *****************************************************************************
3164 * Algorithm doesn't take into account page size. ;-(
3167 xlog_state_ticket_alloc(xlog_t *log)
3169 xlog_ticket_t *t_list;
3170 xlog_ticket_t *next;
3172 uint i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3176 * The kmem_zalloc may sleep, so we shouldn't be holding the
3177 * global lock. XXXmiken: may want to use zone allocator.
3179 buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3183 /* Attach 1st ticket to Q, so we can keep track of allocated memory */
3184 t_list = (xlog_ticket_t *)buf;
3185 t_list->t_next = log->l_unmount_free;
3186 log->l_unmount_free = t_list++;
3187 log->l_ticket_cnt++;
3188 log->l_ticket_tcnt++;
3190 /* Next ticket becomes first ticket attached to ticket free list */
3191 if (log->l_freelist != NULL) {
3192 ASSERT(log->l_tail != NULL);
3193 log->l_tail->t_next = t_list;
3195 log->l_freelist = t_list;
3197 log->l_ticket_cnt++;
3198 log->l_ticket_tcnt++;
3200 /* Cycle through rest of alloc'ed memory, building up free Q */
3201 for ( ; i > 0; i--) {
3203 t_list->t_next = next;
3205 log->l_ticket_cnt++;
3206 log->l_ticket_tcnt++;
3208 t_list->t_next = NULL;
3209 log->l_tail = t_list;
3211 } /* xlog_state_ticket_alloc */
3215 * Put ticket into free list
3217 * Assumption: log lock is held around this call.
3220 xlog_ticket_put(xlog_t *log,
3221 xlog_ticket_t *ticket)
3223 sv_destroy(&ticket->t_sema);
3226 * Don't think caching will make that much difference. It's
3227 * more important to make debug easier.
3230 /* real code will want to use LIFO for caching */
3231 ticket->t_next = log->l_freelist;
3232 log->l_freelist = ticket;
3233 /* no need to clear fields */
3235 /* When we debug, it is easier if tickets are cycled */
3236 ticket->t_next = NULL;
3237 if (log->l_tail != 0) {
3238 log->l_tail->t_next = ticket;
3240 ASSERT(log->l_freelist == 0);
3241 log->l_freelist = ticket;
3243 log->l_tail = ticket;
3245 log->l_ticket_cnt++;
3246 } /* xlog_ticket_put */
3250 * Grab ticket off freelist or allocation some more
3253 xlog_ticket_get(xlog_t *log,
3264 if (log->l_freelist == NULL)
3265 xlog_state_ticket_alloc(log); /* potentially sleep */
3268 if (log->l_freelist == NULL) {
3272 tic = log->l_freelist;
3273 log->l_freelist = tic->t_next;
3274 if (log->l_freelist == NULL)
3276 log->l_ticket_cnt--;
3280 * Permanent reservations have up to 'cnt'-1 active log operations
3281 * in the log. A unit in this case is the amount of space for one
3282 * of these log operations. Normal reservations have a cnt of 1
3283 * and their unit amount is the total amount of space required.
3285 * The following lines of code account for non-transaction data
3286 * which occupy space in the on-disk log.
3288 * Normal form of a transaction is:
3289 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3290 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3292 * We need to account for all the leadup data and trailer data
3293 * around the transaction data.
3294 * And then we need to account for the worst case in terms of using
3296 * The worst case will happen if:
3297 * - the placement of the transaction happens to be such that the
3298 * roundoff is at its maximum
3299 * - the transaction data is synced before the commit record is synced
3300 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3301 * Therefore the commit record is in its own Log Record.
3302 * This can happen as the commit record is called with its
3303 * own region to xlog_write().
3304 * This then means that in the worst case, roundoff can happen for
3305 * the commit-rec as well.
3306 * The commit-rec is smaller than padding in this scenario and so it is
3307 * not added separately.
3310 /* for trans header */
3311 unit_bytes += sizeof(xlog_op_header_t);
3312 unit_bytes += sizeof(xfs_trans_header_t);
3315 unit_bytes += sizeof(xlog_op_header_t);
3317 /* for LR headers */
3318 num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3319 unit_bytes += log->l_iclog_hsize * num_headers;
3321 /* for commit-rec LR header - note: padding will subsume the ophdr */
3322 unit_bytes += log->l_iclog_hsize;
3324 /* for split-recs - ophdrs added when data split over LRs */
3325 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3327 /* for roundoff padding for transaction data and one for commit record */
3328 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3329 log->l_mp->m_sb.sb_logsunit > 1) {
3330 /* log su roundoff */
3331 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3334 unit_bytes += 2*BBSIZE;
3337 tic->t_unit_res = unit_bytes;
3338 tic->t_curr_res = unit_bytes;
3341 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3342 tic->t_clientid = client;
3343 tic->t_flags = XLOG_TIC_INITED;
3344 tic->t_trans_type = 0;
3345 if (xflags & XFS_LOG_PERM_RESERV)
3346 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3347 sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3349 XLOG_TIC_RESET_RES(tic);
3352 } /* xlog_ticket_get */
3355 /******************************************************************************
3357 * Log debug routines
3359 ******************************************************************************
3363 * Make sure that the destination ptr is within the valid data region of
3364 * one of the iclogs. This uses backup pointers stored in a different
3365 * part of the log in case we trash the log structure.
3368 xlog_verify_dest_ptr(xlog_t *log,
3374 for (i=0; i < log->l_iclog_bufs; i++) {
3375 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3376 ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3380 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3381 } /* xlog_verify_dest_ptr */
3384 xlog_verify_grant_head(xlog_t *log, int equals)
3386 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3388 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3390 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3392 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3393 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3395 } /* xlog_verify_grant_head */
3397 /* check if it will fit */
3399 xlog_verify_tail_lsn(xlog_t *log,
3400 xlog_in_core_t *iclog,
3405 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3407 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3408 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3409 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3411 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3413 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3414 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3416 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3417 if (blocks < BTOBB(iclog->ic_offset) + 1)
3418 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3420 } /* xlog_verify_tail_lsn */
3423 * Perform a number of checks on the iclog before writing to disk.
3425 * 1. Make sure the iclogs are still circular
3426 * 2. Make sure we have a good magic number
3427 * 3. Make sure we don't have magic numbers in the data
3428 * 4. Check fields of each log operation header for:
3429 * A. Valid client identifier
3430 * B. tid ptr value falls in valid ptr space (user space code)
3431 * C. Length in log record header is correct according to the
3432 * individual operation headers within record.
3433 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3434 * log, check the preceding blocks of the physical log to make sure all
3435 * the cycle numbers agree with the current cycle number.
3438 xlog_verify_iclog(xlog_t *log,
3439 xlog_in_core_t *iclog,
3443 xlog_op_header_t *ophead;
3444 xlog_in_core_t *icptr;
3445 xlog_in_core_2_t *xhdr;
3447 xfs_caddr_t base_ptr;
3448 __psint_t field_offset;
3450 int len, i, j, k, op_len;
3454 /* check validity of iclog pointers */
3456 icptr = log->l_iclog;
3457 for (i=0; i < log->l_iclog_bufs; i++) {
3459 xlog_panic("xlog_verify_iclog: invalid ptr");
3460 icptr = icptr->ic_next;
3462 if (icptr != log->l_iclog)
3463 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3466 /* check log magic numbers */
3467 ptr = (xfs_caddr_t) &(iclog->ic_header);
3468 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3469 xlog_panic("xlog_verify_iclog: invalid magic num");
3471 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3473 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3474 xlog_panic("xlog_verify_iclog: unexpected magic num");
3478 len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3479 ptr = iclog->ic_datap;
3481 ophead = (xlog_op_header_t *)ptr;
3482 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3483 for (i = 0; i < len; i++) {
3484 ophead = (xlog_op_header_t *)ptr;
3486 /* clientid is only 1 byte */
3487 field_offset = (__psint_t)
3488 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3489 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3490 clientid = ophead->oh_clientid;
3492 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3493 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3494 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3495 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3496 clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3498 clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3501 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3502 cmn_err(CE_WARN, "xlog_verify_iclog: "
3503 "invalid clientid %d op 0x%p offset 0x%lx",
3504 clientid, ophead, (unsigned long)field_offset);
3507 field_offset = (__psint_t)
3508 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3509 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3510 op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3512 idx = BTOBBT((__psint_t)&ophead->oh_len -
3513 (__psint_t)iclog->ic_datap);
3514 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3515 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3516 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3517 op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3519 op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3522 ptr += sizeof(xlog_op_header_t) + op_len;
3524 } /* xlog_verify_iclog */
3528 * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3534 xlog_in_core_t *iclog, *ic;
3536 iclog = log->l_iclog;
3537 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3539 * Mark all the incore logs IOERROR.
3540 * From now on, no log flushes will result.
3544 ic->ic_state = XLOG_STATE_IOERROR;
3546 } while (ic != iclog);
3550 * Return non-zero, if state transition has already happened.
3556 * This is called from xfs_force_shutdown, when we're forcibly
3557 * shutting down the filesystem, typically because of an IO error.
3558 * Our main objectives here are to make sure that:
3559 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3560 * parties to find out, 'atomically'.
3561 * b. those who're sleeping on log reservations, pinned objects and
3562 * other resources get woken up, and be told the bad news.
3563 * c. nothing new gets queued up after (a) and (b) are done.
3564 * d. if !logerror, flush the iclogs to disk, then seal them off
3568 xfs_log_force_umount(
3569 struct xfs_mount *mp,
3582 * If this happens during log recovery, don't worry about
3583 * locking; the log isn't open for business yet.
3586 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3587 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3588 XFS_BUF_DONE(mp->m_sb_bp);
3593 * Somebody could've already done the hard work for us.
3594 * No need to get locks for this.
3596 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3597 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3602 * We must hold both the GRANT lock and the LOG lock,
3603 * before we mark the filesystem SHUTDOWN and wake
3604 * everybody up to tell the bad news.
3606 s = GRANT_LOCK(log);
3608 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3609 XFS_BUF_DONE(mp->m_sb_bp);
3611 * This flag is sort of redundant because of the mount flag, but
3612 * it's good to maintain the separation between the log and the rest
3615 log->l_flags |= XLOG_IO_ERROR;
3618 * If we hit a log error, we want to mark all the iclogs IOERROR
3619 * while we're still holding the loglock.
3622 retval = xlog_state_ioerror(log);
3623 LOG_UNLOCK(log, s2);
3626 * We don't want anybody waiting for log reservations
3627 * after this. That means we have to wake up everybody
3628 * queued up on reserve_headq as well as write_headq.
3629 * In addition, we make sure in xlog_{re}grant_log_space
3630 * that we don't enqueue anything once the SHUTDOWN flag
3631 * is set, and this action is protected by the GRANTLOCK.
3633 if ((tic = log->l_reserve_headq)) {
3635 sv_signal(&tic->t_sema);
3637 } while (tic != log->l_reserve_headq);
3640 if ((tic = log->l_write_headq)) {
3642 sv_signal(&tic->t_sema);
3644 } while (tic != log->l_write_headq);
3646 GRANT_UNLOCK(log, s);
3648 if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3651 * Force the incore logs to disk before shutting the
3652 * log down completely.
3654 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3656 retval = xlog_state_ioerror(log);
3657 LOG_UNLOCK(log, s2);
3660 * Wake up everybody waiting on xfs_log_force.
3661 * Callback all log item committed functions as if the
3662 * log writes were completed.
3664 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3666 #ifdef XFSERRORDEBUG
3668 xlog_in_core_t *iclog;
3671 iclog = log->l_iclog;
3673 ASSERT(iclog->ic_callback == 0);
3674 iclog = iclog->ic_next;
3675 } while (iclog != log->l_iclog);
3679 /* return non-zero if log IOERROR transition had already happened */
3684 xlog_iclogs_empty(xlog_t *log)
3686 xlog_in_core_t *iclog;
3688 iclog = log->l_iclog;
3690 /* endianness does not matter here, zero is zero in
3693 if (iclog->ic_header.h_num_logops)
3695 iclog = iclog->ic_next;
3696 } while (iclog != log->l_iclog);