Loading...
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_log.h"
22#include "xfs_trans.h"
23#include "xfs_sb.h"
24#include "xfs_ag.h"
25#include "xfs_mount.h"
26#include "xfs_error.h"
27#include "xfs_log_priv.h"
28#include "xfs_buf_item.h"
29#include "xfs_bmap_btree.h"
30#include "xfs_alloc_btree.h"
31#include "xfs_ialloc_btree.h"
32#include "xfs_log_recover.h"
33#include "xfs_trans_priv.h"
34#include "xfs_dinode.h"
35#include "xfs_inode.h"
36#include "xfs_trace.h"
37
38kmem_zone_t *xfs_log_ticket_zone;
39
40/* Local miscellaneous function prototypes */
41STATIC int
42xlog_commit_record(
43 struct xlog *log,
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
47
48STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
49 xfs_buftarg_t *log_target,
50 xfs_daddr_t blk_offset,
51 int num_bblks);
52STATIC int
53xlog_space_left(
54 struct xlog *log,
55 atomic64_t *head);
56STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
57STATIC void xlog_dealloc_log(xlog_t *log);
58
59/* local state machine functions */
60STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
61STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
62STATIC int xlog_state_get_iclog_space(xlog_t *log,
63 int len,
64 xlog_in_core_t **iclog,
65 xlog_ticket_t *ticket,
66 int *continued_write,
67 int *logoffsetp);
68STATIC int xlog_state_release_iclog(xlog_t *log,
69 xlog_in_core_t *iclog);
70STATIC void xlog_state_switch_iclogs(xlog_t *log,
71 xlog_in_core_t *iclog,
72 int eventual_size);
73STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
74
75STATIC void
76xlog_grant_push_ail(
77 struct xlog *log,
78 int need_bytes);
79STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
81STATIC void xlog_ungrant_log_space(xlog_t *log,
82 xlog_ticket_t *ticket);
83
84#if defined(DEBUG)
85STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
86STATIC void
87xlog_verify_grant_tail(
88 struct xlog *log);
89STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
90 int count, boolean_t syncing);
91STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
92 xfs_lsn_t tail_lsn);
93#else
94#define xlog_verify_dest_ptr(a,b)
95#define xlog_verify_grant_tail(a)
96#define xlog_verify_iclog(a,b,c,d)
97#define xlog_verify_tail_lsn(a,b,c)
98#endif
99
100STATIC int xlog_iclogs_empty(xlog_t *log);
101
102static void
103xlog_grant_sub_space(
104 struct xlog *log,
105 atomic64_t *head,
106 int bytes)
107{
108 int64_t head_val = atomic64_read(head);
109 int64_t new, old;
110
111 do {
112 int cycle, space;
113
114 xlog_crack_grant_head_val(head_val, &cycle, &space);
115
116 space -= bytes;
117 if (space < 0) {
118 space += log->l_logsize;
119 cycle--;
120 }
121
122 old = head_val;
123 new = xlog_assign_grant_head_val(cycle, space);
124 head_val = atomic64_cmpxchg(head, old, new);
125 } while (head_val != old);
126}
127
128static void
129xlog_grant_add_space(
130 struct xlog *log,
131 atomic64_t *head,
132 int bytes)
133{
134 int64_t head_val = atomic64_read(head);
135 int64_t new, old;
136
137 do {
138 int tmp;
139 int cycle, space;
140
141 xlog_crack_grant_head_val(head_val, &cycle, &space);
142
143 tmp = log->l_logsize - space;
144 if (tmp > bytes)
145 space += bytes;
146 else {
147 space = bytes - tmp;
148 cycle++;
149 }
150
151 old = head_val;
152 new = xlog_assign_grant_head_val(cycle, space);
153 head_val = atomic64_cmpxchg(head, old, new);
154 } while (head_val != old);
155}
156
157STATIC void
158xlog_grant_head_init(
159 struct xlog_grant_head *head)
160{
161 xlog_assign_grant_head(&head->grant, 1, 0);
162 INIT_LIST_HEAD(&head->waiters);
163 spin_lock_init(&head->lock);
164}
165
166STATIC void
167xlog_grant_head_wake_all(
168 struct xlog_grant_head *head)
169{
170 struct xlog_ticket *tic;
171
172 spin_lock(&head->lock);
173 list_for_each_entry(tic, &head->waiters, t_queue)
174 wake_up_process(tic->t_task);
175 spin_unlock(&head->lock);
176}
177
178static inline int
179xlog_ticket_reservation(
180 struct xlog *log,
181 struct xlog_grant_head *head,
182 struct xlog_ticket *tic)
183{
184 if (head == &log->l_write_head) {
185 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
186 return tic->t_unit_res;
187 } else {
188 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
189 return tic->t_unit_res * tic->t_cnt;
190 else
191 return tic->t_unit_res;
192 }
193}
194
195STATIC bool
196xlog_grant_head_wake(
197 struct xlog *log,
198 struct xlog_grant_head *head,
199 int *free_bytes)
200{
201 struct xlog_ticket *tic;
202 int need_bytes;
203
204 list_for_each_entry(tic, &head->waiters, t_queue) {
205 need_bytes = xlog_ticket_reservation(log, head, tic);
206 if (*free_bytes < need_bytes)
207 return false;
208
209 *free_bytes -= need_bytes;
210 trace_xfs_log_grant_wake_up(log, tic);
211 wake_up_process(tic->t_task);
212 }
213
214 return true;
215}
216
217STATIC int
218xlog_grant_head_wait(
219 struct xlog *log,
220 struct xlog_grant_head *head,
221 struct xlog_ticket *tic,
222 int need_bytes)
223{
224 list_add_tail(&tic->t_queue, &head->waiters);
225
226 do {
227 if (XLOG_FORCED_SHUTDOWN(log))
228 goto shutdown;
229 xlog_grant_push_ail(log, need_bytes);
230
231 __set_current_state(TASK_UNINTERRUPTIBLE);
232 spin_unlock(&head->lock);
233
234 XFS_STATS_INC(xs_sleep_logspace);
235
236 trace_xfs_log_grant_sleep(log, tic);
237 schedule();
238 trace_xfs_log_grant_wake(log, tic);
239
240 spin_lock(&head->lock);
241 if (XLOG_FORCED_SHUTDOWN(log))
242 goto shutdown;
243 } while (xlog_space_left(log, &head->grant) < need_bytes);
244
245 list_del_init(&tic->t_queue);
246 return 0;
247shutdown:
248 list_del_init(&tic->t_queue);
249 return XFS_ERROR(EIO);
250}
251
252/*
253 * Atomically get the log space required for a log ticket.
254 *
255 * Once a ticket gets put onto head->waiters, it will only return after the
256 * needed reservation is satisfied.
257 *
258 * This function is structured so that it has a lock free fast path. This is
259 * necessary because every new transaction reservation will come through this
260 * path. Hence any lock will be globally hot if we take it unconditionally on
261 * every pass.
262 *
263 * As tickets are only ever moved on and off head->waiters under head->lock, we
264 * only need to take that lock if we are going to add the ticket to the queue
265 * and sleep. We can avoid taking the lock if the ticket was never added to
266 * head->waiters because the t_queue list head will be empty and we hold the
267 * only reference to it so it can safely be checked unlocked.
268 */
269STATIC int
270xlog_grant_head_check(
271 struct xlog *log,
272 struct xlog_grant_head *head,
273 struct xlog_ticket *tic,
274 int *need_bytes)
275{
276 int free_bytes;
277 int error = 0;
278
279 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
280
281 /*
282 * If there are other waiters on the queue then give them a chance at
283 * logspace before us. Wake up the first waiters, if we do not wake
284 * up all the waiters then go to sleep waiting for more free space,
285 * otherwise try to get some space for this transaction.
286 */
287 *need_bytes = xlog_ticket_reservation(log, head, tic);
288 free_bytes = xlog_space_left(log, &head->grant);
289 if (!list_empty_careful(&head->waiters)) {
290 spin_lock(&head->lock);
291 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
292 free_bytes < *need_bytes) {
293 error = xlog_grant_head_wait(log, head, tic,
294 *need_bytes);
295 }
296 spin_unlock(&head->lock);
297 } else if (free_bytes < *need_bytes) {
298 spin_lock(&head->lock);
299 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
300 spin_unlock(&head->lock);
301 }
302
303 return error;
304}
305
306static void
307xlog_tic_reset_res(xlog_ticket_t *tic)
308{
309 tic->t_res_num = 0;
310 tic->t_res_arr_sum = 0;
311 tic->t_res_num_ophdrs = 0;
312}
313
314static void
315xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
316{
317 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
318 /* add to overflow and start again */
319 tic->t_res_o_flow += tic->t_res_arr_sum;
320 tic->t_res_num = 0;
321 tic->t_res_arr_sum = 0;
322 }
323
324 tic->t_res_arr[tic->t_res_num].r_len = len;
325 tic->t_res_arr[tic->t_res_num].r_type = type;
326 tic->t_res_arr_sum += len;
327 tic->t_res_num++;
328}
329
330/*
331 * Replenish the byte reservation required by moving the grant write head.
332 */
333int
334xfs_log_regrant(
335 struct xfs_mount *mp,
336 struct xlog_ticket *tic)
337{
338 struct xlog *log = mp->m_log;
339 int need_bytes;
340 int error = 0;
341
342 if (XLOG_FORCED_SHUTDOWN(log))
343 return XFS_ERROR(EIO);
344
345 XFS_STATS_INC(xs_try_logspace);
346
347 /*
348 * This is a new transaction on the ticket, so we need to change the
349 * transaction ID so that the next transaction has a different TID in
350 * the log. Just add one to the existing tid so that we can see chains
351 * of rolling transactions in the log easily.
352 */
353 tic->t_tid++;
354
355 xlog_grant_push_ail(log, tic->t_unit_res);
356
357 tic->t_curr_res = tic->t_unit_res;
358 xlog_tic_reset_res(tic);
359
360 if (tic->t_cnt > 0)
361 return 0;
362
363 trace_xfs_log_regrant(log, tic);
364
365 error = xlog_grant_head_check(log, &log->l_write_head, tic,
366 &need_bytes);
367 if (error)
368 goto out_error;
369
370 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
371 trace_xfs_log_regrant_exit(log, tic);
372 xlog_verify_grant_tail(log);
373 return 0;
374
375out_error:
376 /*
377 * If we are failing, make sure the ticket doesn't have any current
378 * reservations. We don't want to add this back when the ticket/
379 * transaction gets cancelled.
380 */
381 tic->t_curr_res = 0;
382 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
383 return error;
384}
385
386/*
387 * Reserve log space and return a ticket corresponding the reservation.
388 *
389 * Each reservation is going to reserve extra space for a log record header.
390 * When writes happen to the on-disk log, we don't subtract the length of the
391 * log record header from any reservation. By wasting space in each
392 * reservation, we prevent over allocation problems.
393 */
394int
395xfs_log_reserve(
396 struct xfs_mount *mp,
397 int unit_bytes,
398 int cnt,
399 struct xlog_ticket **ticp,
400 __uint8_t client,
401 bool permanent,
402 uint t_type)
403{
404 struct xlog *log = mp->m_log;
405 struct xlog_ticket *tic;
406 int need_bytes;
407 int error = 0;
408
409 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
410
411 if (XLOG_FORCED_SHUTDOWN(log))
412 return XFS_ERROR(EIO);
413
414 XFS_STATS_INC(xs_try_logspace);
415
416 ASSERT(*ticp == NULL);
417 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
418 KM_SLEEP | KM_MAYFAIL);
419 if (!tic)
420 return XFS_ERROR(ENOMEM);
421
422 tic->t_trans_type = t_type;
423 *ticp = tic;
424
425 xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
426
427 trace_xfs_log_reserve(log, tic);
428
429 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
430 &need_bytes);
431 if (error)
432 goto out_error;
433
434 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
435 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 trace_xfs_log_reserve_exit(log, tic);
437 xlog_verify_grant_tail(log);
438 return 0;
439
440out_error:
441 /*
442 * If we are failing, make sure the ticket doesn't have any current
443 * reservations. We don't want to add this back when the ticket/
444 * transaction gets cancelled.
445 */
446 tic->t_curr_res = 0;
447 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
448 return error;
449}
450
451
452/*
453 * NOTES:
454 *
455 * 1. currblock field gets updated at startup and after in-core logs
456 * marked as with WANT_SYNC.
457 */
458
459/*
460 * This routine is called when a user of a log manager ticket is done with
461 * the reservation. If the ticket was ever used, then a commit record for
462 * the associated transaction is written out as a log operation header with
463 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
464 * a given ticket. If the ticket was one with a permanent reservation, then
465 * a few operations are done differently. Permanent reservation tickets by
466 * default don't release the reservation. They just commit the current
467 * transaction with the belief that the reservation is still needed. A flag
468 * must be passed in before permanent reservations are actually released.
469 * When these type of tickets are not released, they need to be set into
470 * the inited state again. By doing this, a start record will be written
471 * out when the next write occurs.
472 */
473xfs_lsn_t
474xfs_log_done(
475 struct xfs_mount *mp,
476 struct xlog_ticket *ticket,
477 struct xlog_in_core **iclog,
478 uint flags)
479{
480 struct xlog *log = mp->m_log;
481 xfs_lsn_t lsn = 0;
482
483 if (XLOG_FORCED_SHUTDOWN(log) ||
484 /*
485 * If nothing was ever written, don't write out commit record.
486 * If we get an error, just continue and give back the log ticket.
487 */
488 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
489 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
490 lsn = (xfs_lsn_t) -1;
491 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
492 flags |= XFS_LOG_REL_PERM_RESERV;
493 }
494 }
495
496
497 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
498 (flags & XFS_LOG_REL_PERM_RESERV)) {
499 trace_xfs_log_done_nonperm(log, ticket);
500
501 /*
502 * Release ticket if not permanent reservation or a specific
503 * request has been made to release a permanent reservation.
504 */
505 xlog_ungrant_log_space(log, ticket);
506 xfs_log_ticket_put(ticket);
507 } else {
508 trace_xfs_log_done_perm(log, ticket);
509
510 xlog_regrant_reserve_log_space(log, ticket);
511 /* If this ticket was a permanent reservation and we aren't
512 * trying to release it, reset the inited flags; so next time
513 * we write, a start record will be written out.
514 */
515 ticket->t_flags |= XLOG_TIC_INITED;
516 }
517
518 return lsn;
519}
520
521/*
522 * Attaches a new iclog I/O completion callback routine during
523 * transaction commit. If the log is in error state, a non-zero
524 * return code is handed back and the caller is responsible for
525 * executing the callback at an appropriate time.
526 */
527int
528xfs_log_notify(
529 struct xfs_mount *mp,
530 struct xlog_in_core *iclog,
531 xfs_log_callback_t *cb)
532{
533 int abortflg;
534
535 spin_lock(&iclog->ic_callback_lock);
536 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
537 if (!abortflg) {
538 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
539 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
540 cb->cb_next = NULL;
541 *(iclog->ic_callback_tail) = cb;
542 iclog->ic_callback_tail = &(cb->cb_next);
543 }
544 spin_unlock(&iclog->ic_callback_lock);
545 return abortflg;
546}
547
548int
549xfs_log_release_iclog(
550 struct xfs_mount *mp,
551 struct xlog_in_core *iclog)
552{
553 if (xlog_state_release_iclog(mp->m_log, iclog)) {
554 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
555 return EIO;
556 }
557
558 return 0;
559}
560
561/*
562 * Mount a log filesystem
563 *
564 * mp - ubiquitous xfs mount point structure
565 * log_target - buftarg of on-disk log device
566 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
567 * num_bblocks - Number of BBSIZE blocks in on-disk log
568 *
569 * Return error or zero.
570 */
571int
572xfs_log_mount(
573 xfs_mount_t *mp,
574 xfs_buftarg_t *log_target,
575 xfs_daddr_t blk_offset,
576 int num_bblks)
577{
578 int error;
579
580 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
581 xfs_notice(mp, "Mounting Filesystem");
582 else {
583 xfs_notice(mp,
584"Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
585 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
586 }
587
588 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
589 if (IS_ERR(mp->m_log)) {
590 error = -PTR_ERR(mp->m_log);
591 goto out;
592 }
593
594 /*
595 * Initialize the AIL now we have a log.
596 */
597 error = xfs_trans_ail_init(mp);
598 if (error) {
599 xfs_warn(mp, "AIL initialisation failed: error %d", error);
600 goto out_free_log;
601 }
602 mp->m_log->l_ailp = mp->m_ail;
603
604 /*
605 * skip log recovery on a norecovery mount. pretend it all
606 * just worked.
607 */
608 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
609 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
610
611 if (readonly)
612 mp->m_flags &= ~XFS_MOUNT_RDONLY;
613
614 error = xlog_recover(mp->m_log);
615
616 if (readonly)
617 mp->m_flags |= XFS_MOUNT_RDONLY;
618 if (error) {
619 xfs_warn(mp, "log mount/recovery failed: error %d",
620 error);
621 goto out_destroy_ail;
622 }
623 }
624
625 /* Normal transactions can now occur */
626 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
627
628 /*
629 * Now the log has been fully initialised and we know were our
630 * space grant counters are, we can initialise the permanent ticket
631 * needed for delayed logging to work.
632 */
633 xlog_cil_init_post_recovery(mp->m_log);
634
635 return 0;
636
637out_destroy_ail:
638 xfs_trans_ail_destroy(mp);
639out_free_log:
640 xlog_dealloc_log(mp->m_log);
641out:
642 return error;
643}
644
645/*
646 * Finish the recovery of the file system. This is separate from
647 * the xfs_log_mount() call, because it depends on the code in
648 * xfs_mountfs() to read in the root and real-time bitmap inodes
649 * between calling xfs_log_mount() and here.
650 *
651 * mp - ubiquitous xfs mount point structure
652 */
653int
654xfs_log_mount_finish(xfs_mount_t *mp)
655{
656 int error;
657
658 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
659 error = xlog_recover_finish(mp->m_log);
660 else {
661 error = 0;
662 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
663 }
664
665 return error;
666}
667
668/*
669 * Final log writes as part of unmount.
670 *
671 * Mark the filesystem clean as unmount happens. Note that during relocation
672 * this routine needs to be executed as part of source-bag while the
673 * deallocation must not be done until source-end.
674 */
675
676/*
677 * Unmount record used to have a string "Unmount filesystem--" in the
678 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
679 * We just write the magic number now since that particular field isn't
680 * currently architecture converted and "nUmount" is a bit foo.
681 * As far as I know, there weren't any dependencies on the old behaviour.
682 */
683
684int
685xfs_log_unmount_write(xfs_mount_t *mp)
686{
687 xlog_t *log = mp->m_log;
688 xlog_in_core_t *iclog;
689#ifdef DEBUG
690 xlog_in_core_t *first_iclog;
691#endif
692 xlog_ticket_t *tic = NULL;
693 xfs_lsn_t lsn;
694 int error;
695
696 /*
697 * Don't write out unmount record on read-only mounts.
698 * Or, if we are doing a forced umount (typically because of IO errors).
699 */
700 if (mp->m_flags & XFS_MOUNT_RDONLY)
701 return 0;
702
703 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
704 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
705
706#ifdef DEBUG
707 first_iclog = iclog = log->l_iclog;
708 do {
709 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
710 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
711 ASSERT(iclog->ic_offset == 0);
712 }
713 iclog = iclog->ic_next;
714 } while (iclog != first_iclog);
715#endif
716 if (! (XLOG_FORCED_SHUTDOWN(log))) {
717 error = xfs_log_reserve(mp, 600, 1, &tic,
718 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
719 if (!error) {
720 /* the data section must be 32 bit size aligned */
721 struct {
722 __uint16_t magic;
723 __uint16_t pad1;
724 __uint32_t pad2; /* may as well make it 64 bits */
725 } magic = {
726 .magic = XLOG_UNMOUNT_TYPE,
727 };
728 struct xfs_log_iovec reg = {
729 .i_addr = &magic,
730 .i_len = sizeof(magic),
731 .i_type = XLOG_REG_TYPE_UNMOUNT,
732 };
733 struct xfs_log_vec vec = {
734 .lv_niovecs = 1,
735 .lv_iovecp = ®,
736 };
737
738 /* remove inited flag, and account for space used */
739 tic->t_flags = 0;
740 tic->t_curr_res -= sizeof(magic);
741 error = xlog_write(log, &vec, tic, &lsn,
742 NULL, XLOG_UNMOUNT_TRANS);
743 /*
744 * At this point, we're umounting anyway,
745 * so there's no point in transitioning log state
746 * to IOERROR. Just continue...
747 */
748 }
749
750 if (error)
751 xfs_alert(mp, "%s: unmount record failed", __func__);
752
753
754 spin_lock(&log->l_icloglock);
755 iclog = log->l_iclog;
756 atomic_inc(&iclog->ic_refcnt);
757 xlog_state_want_sync(log, iclog);
758 spin_unlock(&log->l_icloglock);
759 error = xlog_state_release_iclog(log, iclog);
760
761 spin_lock(&log->l_icloglock);
762 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
763 iclog->ic_state == XLOG_STATE_DIRTY)) {
764 if (!XLOG_FORCED_SHUTDOWN(log)) {
765 xlog_wait(&iclog->ic_force_wait,
766 &log->l_icloglock);
767 } else {
768 spin_unlock(&log->l_icloglock);
769 }
770 } else {
771 spin_unlock(&log->l_icloglock);
772 }
773 if (tic) {
774 trace_xfs_log_umount_write(log, tic);
775 xlog_ungrant_log_space(log, tic);
776 xfs_log_ticket_put(tic);
777 }
778 } else {
779 /*
780 * We're already in forced_shutdown mode, couldn't
781 * even attempt to write out the unmount transaction.
782 *
783 * Go through the motions of sync'ing and releasing
784 * the iclog, even though no I/O will actually happen,
785 * we need to wait for other log I/Os that may already
786 * be in progress. Do this as a separate section of
787 * code so we'll know if we ever get stuck here that
788 * we're in this odd situation of trying to unmount
789 * a file system that went into forced_shutdown as
790 * the result of an unmount..
791 */
792 spin_lock(&log->l_icloglock);
793 iclog = log->l_iclog;
794 atomic_inc(&iclog->ic_refcnt);
795
796 xlog_state_want_sync(log, iclog);
797 spin_unlock(&log->l_icloglock);
798 error = xlog_state_release_iclog(log, iclog);
799
800 spin_lock(&log->l_icloglock);
801
802 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
803 || iclog->ic_state == XLOG_STATE_DIRTY
804 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
805
806 xlog_wait(&iclog->ic_force_wait,
807 &log->l_icloglock);
808 } else {
809 spin_unlock(&log->l_icloglock);
810 }
811 }
812
813 return error;
814} /* xfs_log_unmount_write */
815
816/*
817 * Deallocate log structures for unmount/relocation.
818 *
819 * We need to stop the aild from running before we destroy
820 * and deallocate the log as the aild references the log.
821 */
822void
823xfs_log_unmount(xfs_mount_t *mp)
824{
825 cancel_delayed_work_sync(&mp->m_sync_work);
826 xfs_trans_ail_destroy(mp);
827 xlog_dealloc_log(mp->m_log);
828}
829
830void
831xfs_log_item_init(
832 struct xfs_mount *mp,
833 struct xfs_log_item *item,
834 int type,
835 const struct xfs_item_ops *ops)
836{
837 item->li_mountp = mp;
838 item->li_ailp = mp->m_ail;
839 item->li_type = type;
840 item->li_ops = ops;
841 item->li_lv = NULL;
842
843 INIT_LIST_HEAD(&item->li_ail);
844 INIT_LIST_HEAD(&item->li_cil);
845}
846
847/*
848 * Wake up processes waiting for log space after we have moved the log tail.
849 */
850void
851xfs_log_space_wake(
852 struct xfs_mount *mp)
853{
854 struct xlog *log = mp->m_log;
855 int free_bytes;
856
857 if (XLOG_FORCED_SHUTDOWN(log))
858 return;
859
860 if (!list_empty_careful(&log->l_write_head.waiters)) {
861 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
862
863 spin_lock(&log->l_write_head.lock);
864 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
865 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
866 spin_unlock(&log->l_write_head.lock);
867 }
868
869 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
870 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
871
872 spin_lock(&log->l_reserve_head.lock);
873 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
874 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
875 spin_unlock(&log->l_reserve_head.lock);
876 }
877}
878
879/*
880 * Determine if we have a transaction that has gone to disk
881 * that needs to be covered. To begin the transition to the idle state
882 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
883 * If we are then in a state where covering is needed, the caller is informed
884 * that dummy transactions are required to move the log into the idle state.
885 *
886 * Because this is called as part of the sync process, we should also indicate
887 * that dummy transactions should be issued in anything but the covered or
888 * idle states. This ensures that the log tail is accurately reflected in
889 * the log at the end of the sync, hence if a crash occurrs avoids replay
890 * of transactions where the metadata is already on disk.
891 */
892int
893xfs_log_need_covered(xfs_mount_t *mp)
894{
895 int needed = 0;
896 xlog_t *log = mp->m_log;
897
898 if (!xfs_fs_writable(mp))
899 return 0;
900
901 spin_lock(&log->l_icloglock);
902 switch (log->l_covered_state) {
903 case XLOG_STATE_COVER_DONE:
904 case XLOG_STATE_COVER_DONE2:
905 case XLOG_STATE_COVER_IDLE:
906 break;
907 case XLOG_STATE_COVER_NEED:
908 case XLOG_STATE_COVER_NEED2:
909 if (!xfs_ail_min_lsn(log->l_ailp) &&
910 xlog_iclogs_empty(log)) {
911 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
912 log->l_covered_state = XLOG_STATE_COVER_DONE;
913 else
914 log->l_covered_state = XLOG_STATE_COVER_DONE2;
915 }
916 /* FALLTHRU */
917 default:
918 needed = 1;
919 break;
920 }
921 spin_unlock(&log->l_icloglock);
922 return needed;
923}
924
925/*
926 * We may be holding the log iclog lock upon entering this routine.
927 */
928xfs_lsn_t
929xlog_assign_tail_lsn_locked(
930 struct xfs_mount *mp)
931{
932 struct xlog *log = mp->m_log;
933 struct xfs_log_item *lip;
934 xfs_lsn_t tail_lsn;
935
936 assert_spin_locked(&mp->m_ail->xa_lock);
937
938 /*
939 * To make sure we always have a valid LSN for the log tail we keep
940 * track of the last LSN which was committed in log->l_last_sync_lsn,
941 * and use that when the AIL was empty.
942 */
943 lip = xfs_ail_min(mp->m_ail);
944 if (lip)
945 tail_lsn = lip->li_lsn;
946 else
947 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
948 atomic64_set(&log->l_tail_lsn, tail_lsn);
949 return tail_lsn;
950}
951
952xfs_lsn_t
953xlog_assign_tail_lsn(
954 struct xfs_mount *mp)
955{
956 xfs_lsn_t tail_lsn;
957
958 spin_lock(&mp->m_ail->xa_lock);
959 tail_lsn = xlog_assign_tail_lsn_locked(mp);
960 spin_unlock(&mp->m_ail->xa_lock);
961
962 return tail_lsn;
963}
964
965/*
966 * Return the space in the log between the tail and the head. The head
967 * is passed in the cycle/bytes formal parms. In the special case where
968 * the reserve head has wrapped passed the tail, this calculation is no
969 * longer valid. In this case, just return 0 which means there is no space
970 * in the log. This works for all places where this function is called
971 * with the reserve head. Of course, if the write head were to ever
972 * wrap the tail, we should blow up. Rather than catch this case here,
973 * we depend on other ASSERTions in other parts of the code. XXXmiken
974 *
975 * This code also handles the case where the reservation head is behind
976 * the tail. The details of this case are described below, but the end
977 * result is that we return the size of the log as the amount of space left.
978 */
979STATIC int
980xlog_space_left(
981 struct xlog *log,
982 atomic64_t *head)
983{
984 int free_bytes;
985 int tail_bytes;
986 int tail_cycle;
987 int head_cycle;
988 int head_bytes;
989
990 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
991 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
992 tail_bytes = BBTOB(tail_bytes);
993 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
994 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
995 else if (tail_cycle + 1 < head_cycle)
996 return 0;
997 else if (tail_cycle < head_cycle) {
998 ASSERT(tail_cycle == (head_cycle - 1));
999 free_bytes = tail_bytes - head_bytes;
1000 } else {
1001 /*
1002 * The reservation head is behind the tail.
1003 * In this case we just want to return the size of the
1004 * log as the amount of space left.
1005 */
1006 xfs_alert(log->l_mp,
1007 "xlog_space_left: head behind tail\n"
1008 " tail_cycle = %d, tail_bytes = %d\n"
1009 " GH cycle = %d, GH bytes = %d",
1010 tail_cycle, tail_bytes, head_cycle, head_bytes);
1011 ASSERT(0);
1012 free_bytes = log->l_logsize;
1013 }
1014 return free_bytes;
1015}
1016
1017
1018/*
1019 * Log function which is called when an io completes.
1020 *
1021 * The log manager needs its own routine, in order to control what
1022 * happens with the buffer after the write completes.
1023 */
1024void
1025xlog_iodone(xfs_buf_t *bp)
1026{
1027 xlog_in_core_t *iclog = bp->b_fspriv;
1028 xlog_t *l = iclog->ic_log;
1029 int aborted = 0;
1030
1031 /*
1032 * Race to shutdown the filesystem if we see an error.
1033 */
1034 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1035 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1036 xfs_buf_ioerror_alert(bp, __func__);
1037 xfs_buf_stale(bp);
1038 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1039 /*
1040 * This flag will be propagated to the trans-committed
1041 * callback routines to let them know that the log-commit
1042 * didn't succeed.
1043 */
1044 aborted = XFS_LI_ABORTED;
1045 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1046 aborted = XFS_LI_ABORTED;
1047 }
1048
1049 /* log I/O is always issued ASYNC */
1050 ASSERT(XFS_BUF_ISASYNC(bp));
1051 xlog_state_done_syncing(iclog, aborted);
1052 /*
1053 * do not reference the buffer (bp) here as we could race
1054 * with it being freed after writing the unmount record to the
1055 * log.
1056 */
1057
1058} /* xlog_iodone */
1059
1060/*
1061 * Return size of each in-core log record buffer.
1062 *
1063 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1064 *
1065 * If the filesystem blocksize is too large, we may need to choose a
1066 * larger size since the directory code currently logs entire blocks.
1067 */
1068
1069STATIC void
1070xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1071 xlog_t *log)
1072{
1073 int size;
1074 int xhdrs;
1075
1076 if (mp->m_logbufs <= 0)
1077 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1078 else
1079 log->l_iclog_bufs = mp->m_logbufs;
1080
1081 /*
1082 * Buffer size passed in from mount system call.
1083 */
1084 if (mp->m_logbsize > 0) {
1085 size = log->l_iclog_size = mp->m_logbsize;
1086 log->l_iclog_size_log = 0;
1087 while (size != 1) {
1088 log->l_iclog_size_log++;
1089 size >>= 1;
1090 }
1091
1092 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1093 /* # headers = size / 32k
1094 * one header holds cycles from 32k of data
1095 */
1096
1097 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1098 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1099 xhdrs++;
1100 log->l_iclog_hsize = xhdrs << BBSHIFT;
1101 log->l_iclog_heads = xhdrs;
1102 } else {
1103 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1104 log->l_iclog_hsize = BBSIZE;
1105 log->l_iclog_heads = 1;
1106 }
1107 goto done;
1108 }
1109
1110 /* All machines use 32kB buffers by default. */
1111 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1112 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1113
1114 /* the default log size is 16k or 32k which is one header sector */
1115 log->l_iclog_hsize = BBSIZE;
1116 log->l_iclog_heads = 1;
1117
1118done:
1119 /* are we being asked to make the sizes selected above visible? */
1120 if (mp->m_logbufs == 0)
1121 mp->m_logbufs = log->l_iclog_bufs;
1122 if (mp->m_logbsize == 0)
1123 mp->m_logbsize = log->l_iclog_size;
1124} /* xlog_get_iclog_buffer_size */
1125
1126
1127/*
1128 * This routine initializes some of the log structure for a given mount point.
1129 * Its primary purpose is to fill in enough, so recovery can occur. However,
1130 * some other stuff may be filled in too.
1131 */
1132STATIC xlog_t *
1133xlog_alloc_log(xfs_mount_t *mp,
1134 xfs_buftarg_t *log_target,
1135 xfs_daddr_t blk_offset,
1136 int num_bblks)
1137{
1138 xlog_t *log;
1139 xlog_rec_header_t *head;
1140 xlog_in_core_t **iclogp;
1141 xlog_in_core_t *iclog, *prev_iclog=NULL;
1142 xfs_buf_t *bp;
1143 int i;
1144 int error = ENOMEM;
1145 uint log2_size = 0;
1146
1147 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1148 if (!log) {
1149 xfs_warn(mp, "Log allocation failed: No memory!");
1150 goto out;
1151 }
1152
1153 log->l_mp = mp;
1154 log->l_targ = log_target;
1155 log->l_logsize = BBTOB(num_bblks);
1156 log->l_logBBstart = blk_offset;
1157 log->l_logBBsize = num_bblks;
1158 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1159 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1160
1161 log->l_prev_block = -1;
1162 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1163 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1164 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1165 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1166
1167 xlog_grant_head_init(&log->l_reserve_head);
1168 xlog_grant_head_init(&log->l_write_head);
1169
1170 error = EFSCORRUPTED;
1171 if (xfs_sb_version_hassector(&mp->m_sb)) {
1172 log2_size = mp->m_sb.sb_logsectlog;
1173 if (log2_size < BBSHIFT) {
1174 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1175 log2_size, BBSHIFT);
1176 goto out_free_log;
1177 }
1178
1179 log2_size -= BBSHIFT;
1180 if (log2_size > mp->m_sectbb_log) {
1181 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1182 log2_size, mp->m_sectbb_log);
1183 goto out_free_log;
1184 }
1185
1186 /* for larger sector sizes, must have v2 or external log */
1187 if (log2_size && log->l_logBBstart > 0 &&
1188 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1189 xfs_warn(mp,
1190 "log sector size (0x%x) invalid for configuration.",
1191 log2_size);
1192 goto out_free_log;
1193 }
1194 }
1195 log->l_sectBBsize = 1 << log2_size;
1196
1197 xlog_get_iclog_buffer_size(mp, log);
1198
1199 error = ENOMEM;
1200 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1201 if (!bp)
1202 goto out_free_log;
1203 bp->b_iodone = xlog_iodone;
1204 ASSERT(xfs_buf_islocked(bp));
1205 log->l_xbuf = bp;
1206
1207 spin_lock_init(&log->l_icloglock);
1208 init_waitqueue_head(&log->l_flush_wait);
1209
1210 iclogp = &log->l_iclog;
1211 /*
1212 * The amount of memory to allocate for the iclog structure is
1213 * rather funky due to the way the structure is defined. It is
1214 * done this way so that we can use different sizes for machines
1215 * with different amounts of memory. See the definition of
1216 * xlog_in_core_t in xfs_log_priv.h for details.
1217 */
1218 ASSERT(log->l_iclog_size >= 4096);
1219 for (i=0; i < log->l_iclog_bufs; i++) {
1220 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1221 if (!*iclogp)
1222 goto out_free_iclog;
1223
1224 iclog = *iclogp;
1225 iclog->ic_prev = prev_iclog;
1226 prev_iclog = iclog;
1227
1228 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1229 BTOBB(log->l_iclog_size), 0);
1230 if (!bp)
1231 goto out_free_iclog;
1232
1233 bp->b_iodone = xlog_iodone;
1234 iclog->ic_bp = bp;
1235 iclog->ic_data = bp->b_addr;
1236#ifdef DEBUG
1237 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1238#endif
1239 head = &iclog->ic_header;
1240 memset(head, 0, sizeof(xlog_rec_header_t));
1241 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1242 head->h_version = cpu_to_be32(
1243 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1244 head->h_size = cpu_to_be32(log->l_iclog_size);
1245 /* new fields */
1246 head->h_fmt = cpu_to_be32(XLOG_FMT);
1247 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1248
1249 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1250 iclog->ic_state = XLOG_STATE_ACTIVE;
1251 iclog->ic_log = log;
1252 atomic_set(&iclog->ic_refcnt, 0);
1253 spin_lock_init(&iclog->ic_callback_lock);
1254 iclog->ic_callback_tail = &(iclog->ic_callback);
1255 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1256
1257 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1258 init_waitqueue_head(&iclog->ic_force_wait);
1259 init_waitqueue_head(&iclog->ic_write_wait);
1260
1261 iclogp = &iclog->ic_next;
1262 }
1263 *iclogp = log->l_iclog; /* complete ring */
1264 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1265
1266 error = xlog_cil_init(log);
1267 if (error)
1268 goto out_free_iclog;
1269 return log;
1270
1271out_free_iclog:
1272 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1273 prev_iclog = iclog->ic_next;
1274 if (iclog->ic_bp)
1275 xfs_buf_free(iclog->ic_bp);
1276 kmem_free(iclog);
1277 }
1278 spinlock_destroy(&log->l_icloglock);
1279 xfs_buf_free(log->l_xbuf);
1280out_free_log:
1281 kmem_free(log);
1282out:
1283 return ERR_PTR(-error);
1284} /* xlog_alloc_log */
1285
1286
1287/*
1288 * Write out the commit record of a transaction associated with the given
1289 * ticket. Return the lsn of the commit record.
1290 */
1291STATIC int
1292xlog_commit_record(
1293 struct xlog *log,
1294 struct xlog_ticket *ticket,
1295 struct xlog_in_core **iclog,
1296 xfs_lsn_t *commitlsnp)
1297{
1298 struct xfs_mount *mp = log->l_mp;
1299 int error;
1300 struct xfs_log_iovec reg = {
1301 .i_addr = NULL,
1302 .i_len = 0,
1303 .i_type = XLOG_REG_TYPE_COMMIT,
1304 };
1305 struct xfs_log_vec vec = {
1306 .lv_niovecs = 1,
1307 .lv_iovecp = ®,
1308 };
1309
1310 ASSERT_ALWAYS(iclog);
1311 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1312 XLOG_COMMIT_TRANS);
1313 if (error)
1314 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1315 return error;
1316}
1317
1318/*
1319 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1320 * log space. This code pushes on the lsn which would supposedly free up
1321 * the 25% which we want to leave free. We may need to adopt a policy which
1322 * pushes on an lsn which is further along in the log once we reach the high
1323 * water mark. In this manner, we would be creating a low water mark.
1324 */
1325STATIC void
1326xlog_grant_push_ail(
1327 struct xlog *log,
1328 int need_bytes)
1329{
1330 xfs_lsn_t threshold_lsn = 0;
1331 xfs_lsn_t last_sync_lsn;
1332 int free_blocks;
1333 int free_bytes;
1334 int threshold_block;
1335 int threshold_cycle;
1336 int free_threshold;
1337
1338 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1339
1340 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1341 free_blocks = BTOBBT(free_bytes);
1342
1343 /*
1344 * Set the threshold for the minimum number of free blocks in the
1345 * log to the maximum of what the caller needs, one quarter of the
1346 * log, and 256 blocks.
1347 */
1348 free_threshold = BTOBB(need_bytes);
1349 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1350 free_threshold = MAX(free_threshold, 256);
1351 if (free_blocks >= free_threshold)
1352 return;
1353
1354 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1355 &threshold_block);
1356 threshold_block += free_threshold;
1357 if (threshold_block >= log->l_logBBsize) {
1358 threshold_block -= log->l_logBBsize;
1359 threshold_cycle += 1;
1360 }
1361 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1362 threshold_block);
1363 /*
1364 * Don't pass in an lsn greater than the lsn of the last
1365 * log record known to be on disk. Use a snapshot of the last sync lsn
1366 * so that it doesn't change between the compare and the set.
1367 */
1368 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1369 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1370 threshold_lsn = last_sync_lsn;
1371
1372 /*
1373 * Get the transaction layer to kick the dirty buffers out to
1374 * disk asynchronously. No point in trying to do this if
1375 * the filesystem is shutting down.
1376 */
1377 if (!XLOG_FORCED_SHUTDOWN(log))
1378 xfs_ail_push(log->l_ailp, threshold_lsn);
1379}
1380
1381/*
1382 * The bdstrat callback function for log bufs. This gives us a central
1383 * place to trap bufs in case we get hit by a log I/O error and need to
1384 * shutdown. Actually, in practice, even when we didn't get a log error,
1385 * we transition the iclogs to IOERROR state *after* flushing all existing
1386 * iclogs to disk. This is because we don't want anymore new transactions to be
1387 * started or completed afterwards.
1388 */
1389STATIC int
1390xlog_bdstrat(
1391 struct xfs_buf *bp)
1392{
1393 struct xlog_in_core *iclog = bp->b_fspriv;
1394
1395 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1396 xfs_buf_ioerror(bp, EIO);
1397 xfs_buf_stale(bp);
1398 xfs_buf_ioend(bp, 0);
1399 /*
1400 * It would seem logical to return EIO here, but we rely on
1401 * the log state machine to propagate I/O errors instead of
1402 * doing it here.
1403 */
1404 return 0;
1405 }
1406
1407 xfs_buf_iorequest(bp);
1408 return 0;
1409}
1410
1411/*
1412 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1413 * fashion. Previously, we should have moved the current iclog
1414 * ptr in the log to point to the next available iclog. This allows further
1415 * write to continue while this code syncs out an iclog ready to go.
1416 * Before an in-core log can be written out, the data section must be scanned
1417 * to save away the 1st word of each BBSIZE block into the header. We replace
1418 * it with the current cycle count. Each BBSIZE block is tagged with the
1419 * cycle count because there in an implicit assumption that drives will
1420 * guarantee that entire 512 byte blocks get written at once. In other words,
1421 * we can't have part of a 512 byte block written and part not written. By
1422 * tagging each block, we will know which blocks are valid when recovering
1423 * after an unclean shutdown.
1424 *
1425 * This routine is single threaded on the iclog. No other thread can be in
1426 * this routine with the same iclog. Changing contents of iclog can there-
1427 * fore be done without grabbing the state machine lock. Updating the global
1428 * log will require grabbing the lock though.
1429 *
1430 * The entire log manager uses a logical block numbering scheme. Only
1431 * log_sync (and then only bwrite()) know about the fact that the log may
1432 * not start with block zero on a given device. The log block start offset
1433 * is added immediately before calling bwrite().
1434 */
1435
1436STATIC int
1437xlog_sync(xlog_t *log,
1438 xlog_in_core_t *iclog)
1439{
1440 xfs_caddr_t dptr; /* pointer to byte sized element */
1441 xfs_buf_t *bp;
1442 int i;
1443 uint count; /* byte count of bwrite */
1444 uint count_init; /* initial count before roundup */
1445 int roundoff; /* roundoff to BB or stripe */
1446 int split = 0; /* split write into two regions */
1447 int error;
1448 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1449
1450 XFS_STATS_INC(xs_log_writes);
1451 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1452
1453 /* Add for LR header */
1454 count_init = log->l_iclog_hsize + iclog->ic_offset;
1455
1456 /* Round out the log write size */
1457 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1458 /* we have a v2 stripe unit to use */
1459 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1460 } else {
1461 count = BBTOB(BTOBB(count_init));
1462 }
1463 roundoff = count - count_init;
1464 ASSERT(roundoff >= 0);
1465 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1466 roundoff < log->l_mp->m_sb.sb_logsunit)
1467 ||
1468 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1469 roundoff < BBTOB(1)));
1470
1471 /* move grant heads by roundoff in sync */
1472 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1473 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1474
1475 /* put cycle number in every block */
1476 xlog_pack_data(log, iclog, roundoff);
1477
1478 /* real byte length */
1479 if (v2) {
1480 iclog->ic_header.h_len =
1481 cpu_to_be32(iclog->ic_offset + roundoff);
1482 } else {
1483 iclog->ic_header.h_len =
1484 cpu_to_be32(iclog->ic_offset);
1485 }
1486
1487 bp = iclog->ic_bp;
1488 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1489
1490 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1491
1492 /* Do we need to split this write into 2 parts? */
1493 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1494 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1495 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1496 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1497 } else {
1498 iclog->ic_bwritecnt = 1;
1499 }
1500 bp->b_io_length = BTOBB(count);
1501 bp->b_fspriv = iclog;
1502 XFS_BUF_ZEROFLAGS(bp);
1503 XFS_BUF_ASYNC(bp);
1504 bp->b_flags |= XBF_SYNCIO;
1505
1506 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1507 bp->b_flags |= XBF_FUA;
1508
1509 /*
1510 * Flush the data device before flushing the log to make
1511 * sure all meta data written back from the AIL actually made
1512 * it to disk before stamping the new log tail LSN into the
1513 * log buffer. For an external log we need to issue the
1514 * flush explicitly, and unfortunately synchronously here;
1515 * for an internal log we can simply use the block layer
1516 * state machine for preflushes.
1517 */
1518 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1519 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1520 else
1521 bp->b_flags |= XBF_FLUSH;
1522 }
1523
1524 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1525 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1526
1527 xlog_verify_iclog(log, iclog, count, B_TRUE);
1528
1529 /* account for log which doesn't start at block #0 */
1530 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1531 /*
1532 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1533 * is shutting down.
1534 */
1535 XFS_BUF_WRITE(bp);
1536
1537 error = xlog_bdstrat(bp);
1538 if (error) {
1539 xfs_buf_ioerror_alert(bp, "xlog_sync");
1540 return error;
1541 }
1542 if (split) {
1543 bp = iclog->ic_log->l_xbuf;
1544 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1545 xfs_buf_associate_memory(bp,
1546 (char *)&iclog->ic_header + count, split);
1547 bp->b_fspriv = iclog;
1548 XFS_BUF_ZEROFLAGS(bp);
1549 XFS_BUF_ASYNC(bp);
1550 bp->b_flags |= XBF_SYNCIO;
1551 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1552 bp->b_flags |= XBF_FUA;
1553 dptr = bp->b_addr;
1554 /*
1555 * Bump the cycle numbers at the start of each block
1556 * since this part of the buffer is at the start of
1557 * a new cycle. Watch out for the header magic number
1558 * case, though.
1559 */
1560 for (i = 0; i < split; i += BBSIZE) {
1561 be32_add_cpu((__be32 *)dptr, 1);
1562 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1563 be32_add_cpu((__be32 *)dptr, 1);
1564 dptr += BBSIZE;
1565 }
1566
1567 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1568 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1569
1570 /* account for internal log which doesn't start at block #0 */
1571 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1572 XFS_BUF_WRITE(bp);
1573 error = xlog_bdstrat(bp);
1574 if (error) {
1575 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1576 return error;
1577 }
1578 }
1579 return 0;
1580} /* xlog_sync */
1581
1582
1583/*
1584 * Deallocate a log structure
1585 */
1586STATIC void
1587xlog_dealloc_log(xlog_t *log)
1588{
1589 xlog_in_core_t *iclog, *next_iclog;
1590 int i;
1591
1592 xlog_cil_destroy(log);
1593
1594 /*
1595 * always need to ensure that the extra buffer does not point to memory
1596 * owned by another log buffer before we free it.
1597 */
1598 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1599 xfs_buf_free(log->l_xbuf);
1600
1601 iclog = log->l_iclog;
1602 for (i=0; i<log->l_iclog_bufs; i++) {
1603 xfs_buf_free(iclog->ic_bp);
1604 next_iclog = iclog->ic_next;
1605 kmem_free(iclog);
1606 iclog = next_iclog;
1607 }
1608 spinlock_destroy(&log->l_icloglock);
1609
1610 log->l_mp->m_log = NULL;
1611 kmem_free(log);
1612} /* xlog_dealloc_log */
1613
1614/*
1615 * Update counters atomically now that memcpy is done.
1616 */
1617/* ARGSUSED */
1618static inline void
1619xlog_state_finish_copy(xlog_t *log,
1620 xlog_in_core_t *iclog,
1621 int record_cnt,
1622 int copy_bytes)
1623{
1624 spin_lock(&log->l_icloglock);
1625
1626 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1627 iclog->ic_offset += copy_bytes;
1628
1629 spin_unlock(&log->l_icloglock);
1630} /* xlog_state_finish_copy */
1631
1632
1633
1634
1635/*
1636 * print out info relating to regions written which consume
1637 * the reservation
1638 */
1639void
1640xlog_print_tic_res(
1641 struct xfs_mount *mp,
1642 struct xlog_ticket *ticket)
1643{
1644 uint i;
1645 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1646
1647 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1648 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1649 "bformat",
1650 "bchunk",
1651 "efi_format",
1652 "efd_format",
1653 "iformat",
1654 "icore",
1655 "iext",
1656 "ibroot",
1657 "ilocal",
1658 "iattr_ext",
1659 "iattr_broot",
1660 "iattr_local",
1661 "qformat",
1662 "dquot",
1663 "quotaoff",
1664 "LR header",
1665 "unmount",
1666 "commit",
1667 "trans header"
1668 };
1669 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1670 "SETATTR_NOT_SIZE",
1671 "SETATTR_SIZE",
1672 "INACTIVE",
1673 "CREATE",
1674 "CREATE_TRUNC",
1675 "TRUNCATE_FILE",
1676 "REMOVE",
1677 "LINK",
1678 "RENAME",
1679 "MKDIR",
1680 "RMDIR",
1681 "SYMLINK",
1682 "SET_DMATTRS",
1683 "GROWFS",
1684 "STRAT_WRITE",
1685 "DIOSTRAT",
1686 "WRITE_SYNC",
1687 "WRITEID",
1688 "ADDAFORK",
1689 "ATTRINVAL",
1690 "ATRUNCATE",
1691 "ATTR_SET",
1692 "ATTR_RM",
1693 "ATTR_FLAG",
1694 "CLEAR_AGI_BUCKET",
1695 "QM_SBCHANGE",
1696 "DUMMY1",
1697 "DUMMY2",
1698 "QM_QUOTAOFF",
1699 "QM_DQALLOC",
1700 "QM_SETQLIM",
1701 "QM_DQCLUSTER",
1702 "QM_QINOCREATE",
1703 "QM_QUOTAOFF_END",
1704 "SB_UNIT",
1705 "FSYNC_TS",
1706 "GROWFSRT_ALLOC",
1707 "GROWFSRT_ZERO",
1708 "GROWFSRT_FREE",
1709 "SWAPEXT"
1710 };
1711
1712 xfs_warn(mp,
1713 "xlog_write: reservation summary:\n"
1714 " trans type = %s (%u)\n"
1715 " unit res = %d bytes\n"
1716 " current res = %d bytes\n"
1717 " total reg = %u bytes (o/flow = %u bytes)\n"
1718 " ophdrs = %u (ophdr space = %u bytes)\n"
1719 " ophdr + reg = %u bytes\n"
1720 " num regions = %u\n",
1721 ((ticket->t_trans_type <= 0 ||
1722 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1723 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1724 ticket->t_trans_type,
1725 ticket->t_unit_res,
1726 ticket->t_curr_res,
1727 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1728 ticket->t_res_num_ophdrs, ophdr_spc,
1729 ticket->t_res_arr_sum +
1730 ticket->t_res_o_flow + ophdr_spc,
1731 ticket->t_res_num);
1732
1733 for (i = 0; i < ticket->t_res_num; i++) {
1734 uint r_type = ticket->t_res_arr[i].r_type;
1735 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1736 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1737 "bad-rtype" : res_type_str[r_type-1]),
1738 ticket->t_res_arr[i].r_len);
1739 }
1740
1741 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1742 "xlog_write: reservation ran out. Need to up reservation");
1743 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1744}
1745
1746/*
1747 * Calculate the potential space needed by the log vector. Each region gets
1748 * its own xlog_op_header_t and may need to be double word aligned.
1749 */
1750static int
1751xlog_write_calc_vec_length(
1752 struct xlog_ticket *ticket,
1753 struct xfs_log_vec *log_vector)
1754{
1755 struct xfs_log_vec *lv;
1756 int headers = 0;
1757 int len = 0;
1758 int i;
1759
1760 /* acct for start rec of xact */
1761 if (ticket->t_flags & XLOG_TIC_INITED)
1762 headers++;
1763
1764 for (lv = log_vector; lv; lv = lv->lv_next) {
1765 headers += lv->lv_niovecs;
1766
1767 for (i = 0; i < lv->lv_niovecs; i++) {
1768 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1769
1770 len += vecp->i_len;
1771 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1772 }
1773 }
1774
1775 ticket->t_res_num_ophdrs += headers;
1776 len += headers * sizeof(struct xlog_op_header);
1777
1778 return len;
1779}
1780
1781/*
1782 * If first write for transaction, insert start record We can't be trying to
1783 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1784 */
1785static int
1786xlog_write_start_rec(
1787 struct xlog_op_header *ophdr,
1788 struct xlog_ticket *ticket)
1789{
1790 if (!(ticket->t_flags & XLOG_TIC_INITED))
1791 return 0;
1792
1793 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1794 ophdr->oh_clientid = ticket->t_clientid;
1795 ophdr->oh_len = 0;
1796 ophdr->oh_flags = XLOG_START_TRANS;
1797 ophdr->oh_res2 = 0;
1798
1799 ticket->t_flags &= ~XLOG_TIC_INITED;
1800
1801 return sizeof(struct xlog_op_header);
1802}
1803
1804static xlog_op_header_t *
1805xlog_write_setup_ophdr(
1806 struct xlog *log,
1807 struct xlog_op_header *ophdr,
1808 struct xlog_ticket *ticket,
1809 uint flags)
1810{
1811 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1812 ophdr->oh_clientid = ticket->t_clientid;
1813 ophdr->oh_res2 = 0;
1814
1815 /* are we copying a commit or unmount record? */
1816 ophdr->oh_flags = flags;
1817
1818 /*
1819 * We've seen logs corrupted with bad transaction client ids. This
1820 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1821 * and shut down the filesystem.
1822 */
1823 switch (ophdr->oh_clientid) {
1824 case XFS_TRANSACTION:
1825 case XFS_VOLUME:
1826 case XFS_LOG:
1827 break;
1828 default:
1829 xfs_warn(log->l_mp,
1830 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1831 ophdr->oh_clientid, ticket);
1832 return NULL;
1833 }
1834
1835 return ophdr;
1836}
1837
1838/*
1839 * Set up the parameters of the region copy into the log. This has
1840 * to handle region write split across multiple log buffers - this
1841 * state is kept external to this function so that this code can
1842 * can be written in an obvious, self documenting manner.
1843 */
1844static int
1845xlog_write_setup_copy(
1846 struct xlog_ticket *ticket,
1847 struct xlog_op_header *ophdr,
1848 int space_available,
1849 int space_required,
1850 int *copy_off,
1851 int *copy_len,
1852 int *last_was_partial_copy,
1853 int *bytes_consumed)
1854{
1855 int still_to_copy;
1856
1857 still_to_copy = space_required - *bytes_consumed;
1858 *copy_off = *bytes_consumed;
1859
1860 if (still_to_copy <= space_available) {
1861 /* write of region completes here */
1862 *copy_len = still_to_copy;
1863 ophdr->oh_len = cpu_to_be32(*copy_len);
1864 if (*last_was_partial_copy)
1865 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1866 *last_was_partial_copy = 0;
1867 *bytes_consumed = 0;
1868 return 0;
1869 }
1870
1871 /* partial write of region, needs extra log op header reservation */
1872 *copy_len = space_available;
1873 ophdr->oh_len = cpu_to_be32(*copy_len);
1874 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1875 if (*last_was_partial_copy)
1876 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1877 *bytes_consumed += *copy_len;
1878 (*last_was_partial_copy)++;
1879
1880 /* account for new log op header */
1881 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1882 ticket->t_res_num_ophdrs++;
1883
1884 return sizeof(struct xlog_op_header);
1885}
1886
1887static int
1888xlog_write_copy_finish(
1889 struct xlog *log,
1890 struct xlog_in_core *iclog,
1891 uint flags,
1892 int *record_cnt,
1893 int *data_cnt,
1894 int *partial_copy,
1895 int *partial_copy_len,
1896 int log_offset,
1897 struct xlog_in_core **commit_iclog)
1898{
1899 if (*partial_copy) {
1900 /*
1901 * This iclog has already been marked WANT_SYNC by
1902 * xlog_state_get_iclog_space.
1903 */
1904 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1905 *record_cnt = 0;
1906 *data_cnt = 0;
1907 return xlog_state_release_iclog(log, iclog);
1908 }
1909
1910 *partial_copy = 0;
1911 *partial_copy_len = 0;
1912
1913 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1914 /* no more space in this iclog - push it. */
1915 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1916 *record_cnt = 0;
1917 *data_cnt = 0;
1918
1919 spin_lock(&log->l_icloglock);
1920 xlog_state_want_sync(log, iclog);
1921 spin_unlock(&log->l_icloglock);
1922
1923 if (!commit_iclog)
1924 return xlog_state_release_iclog(log, iclog);
1925 ASSERT(flags & XLOG_COMMIT_TRANS);
1926 *commit_iclog = iclog;
1927 }
1928
1929 return 0;
1930}
1931
1932/*
1933 * Write some region out to in-core log
1934 *
1935 * This will be called when writing externally provided regions or when
1936 * writing out a commit record for a given transaction.
1937 *
1938 * General algorithm:
1939 * 1. Find total length of this write. This may include adding to the
1940 * lengths passed in.
1941 * 2. Check whether we violate the tickets reservation.
1942 * 3. While writing to this iclog
1943 * A. Reserve as much space in this iclog as can get
1944 * B. If this is first write, save away start lsn
1945 * C. While writing this region:
1946 * 1. If first write of transaction, write start record
1947 * 2. Write log operation header (header per region)
1948 * 3. Find out if we can fit entire region into this iclog
1949 * 4. Potentially, verify destination memcpy ptr
1950 * 5. Memcpy (partial) region
1951 * 6. If partial copy, release iclog; otherwise, continue
1952 * copying more regions into current iclog
1953 * 4. Mark want sync bit (in simulation mode)
1954 * 5. Release iclog for potential flush to on-disk log.
1955 *
1956 * ERRORS:
1957 * 1. Panic if reservation is overrun. This should never happen since
1958 * reservation amounts are generated internal to the filesystem.
1959 * NOTES:
1960 * 1. Tickets are single threaded data structures.
1961 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1962 * syncing routine. When a single log_write region needs to span
1963 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1964 * on all log operation writes which don't contain the end of the
1965 * region. The XLOG_END_TRANS bit is used for the in-core log
1966 * operation which contains the end of the continued log_write region.
1967 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1968 * we don't really know exactly how much space will be used. As a result,
1969 * we don't update ic_offset until the end when we know exactly how many
1970 * bytes have been written out.
1971 */
1972int
1973xlog_write(
1974 struct xlog *log,
1975 struct xfs_log_vec *log_vector,
1976 struct xlog_ticket *ticket,
1977 xfs_lsn_t *start_lsn,
1978 struct xlog_in_core **commit_iclog,
1979 uint flags)
1980{
1981 struct xlog_in_core *iclog = NULL;
1982 struct xfs_log_iovec *vecp;
1983 struct xfs_log_vec *lv;
1984 int len;
1985 int index;
1986 int partial_copy = 0;
1987 int partial_copy_len = 0;
1988 int contwr = 0;
1989 int record_cnt = 0;
1990 int data_cnt = 0;
1991 int error;
1992
1993 *start_lsn = 0;
1994
1995 len = xlog_write_calc_vec_length(ticket, log_vector);
1996
1997 /*
1998 * Region headers and bytes are already accounted for.
1999 * We only need to take into account start records and
2000 * split regions in this function.
2001 */
2002 if (ticket->t_flags & XLOG_TIC_INITED)
2003 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2004
2005 /*
2006 * Commit record headers need to be accounted for. These
2007 * come in as separate writes so are easy to detect.
2008 */
2009 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2010 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2011
2012 if (ticket->t_curr_res < 0)
2013 xlog_print_tic_res(log->l_mp, ticket);
2014
2015 index = 0;
2016 lv = log_vector;
2017 vecp = lv->lv_iovecp;
2018 while (lv && index < lv->lv_niovecs) {
2019 void *ptr;
2020 int log_offset;
2021
2022 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2023 &contwr, &log_offset);
2024 if (error)
2025 return error;
2026
2027 ASSERT(log_offset <= iclog->ic_size - 1);
2028 ptr = iclog->ic_datap + log_offset;
2029
2030 /* start_lsn is the first lsn written to. That's all we need. */
2031 if (!*start_lsn)
2032 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2033
2034 /*
2035 * This loop writes out as many regions as can fit in the amount
2036 * of space which was allocated by xlog_state_get_iclog_space().
2037 */
2038 while (lv && index < lv->lv_niovecs) {
2039 struct xfs_log_iovec *reg = &vecp[index];
2040 struct xlog_op_header *ophdr;
2041 int start_rec_copy;
2042 int copy_len;
2043 int copy_off;
2044
2045 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2046 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2047
2048 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2049 if (start_rec_copy) {
2050 record_cnt++;
2051 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2052 start_rec_copy);
2053 }
2054
2055 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2056 if (!ophdr)
2057 return XFS_ERROR(EIO);
2058
2059 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2060 sizeof(struct xlog_op_header));
2061
2062 len += xlog_write_setup_copy(ticket, ophdr,
2063 iclog->ic_size-log_offset,
2064 reg->i_len,
2065 ©_off, ©_len,
2066 &partial_copy,
2067 &partial_copy_len);
2068 xlog_verify_dest_ptr(log, ptr);
2069
2070 /* copy region */
2071 ASSERT(copy_len >= 0);
2072 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2073 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2074
2075 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2076 record_cnt++;
2077 data_cnt += contwr ? copy_len : 0;
2078
2079 error = xlog_write_copy_finish(log, iclog, flags,
2080 &record_cnt, &data_cnt,
2081 &partial_copy,
2082 &partial_copy_len,
2083 log_offset,
2084 commit_iclog);
2085 if (error)
2086 return error;
2087
2088 /*
2089 * if we had a partial copy, we need to get more iclog
2090 * space but we don't want to increment the region
2091 * index because there is still more is this region to
2092 * write.
2093 *
2094 * If we completed writing this region, and we flushed
2095 * the iclog (indicated by resetting of the record
2096 * count), then we also need to get more log space. If
2097 * this was the last record, though, we are done and
2098 * can just return.
2099 */
2100 if (partial_copy)
2101 break;
2102
2103 if (++index == lv->lv_niovecs) {
2104 lv = lv->lv_next;
2105 index = 0;
2106 if (lv)
2107 vecp = lv->lv_iovecp;
2108 }
2109 if (record_cnt == 0) {
2110 if (!lv)
2111 return 0;
2112 break;
2113 }
2114 }
2115 }
2116
2117 ASSERT(len == 0);
2118
2119 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2120 if (!commit_iclog)
2121 return xlog_state_release_iclog(log, iclog);
2122
2123 ASSERT(flags & XLOG_COMMIT_TRANS);
2124 *commit_iclog = iclog;
2125 return 0;
2126}
2127
2128
2129/*****************************************************************************
2130 *
2131 * State Machine functions
2132 *
2133 *****************************************************************************
2134 */
2135
2136/* Clean iclogs starting from the head. This ordering must be
2137 * maintained, so an iclog doesn't become ACTIVE beyond one that
2138 * is SYNCING. This is also required to maintain the notion that we use
2139 * a ordered wait queue to hold off would be writers to the log when every
2140 * iclog is trying to sync to disk.
2141 *
2142 * State Change: DIRTY -> ACTIVE
2143 */
2144STATIC void
2145xlog_state_clean_log(xlog_t *log)
2146{
2147 xlog_in_core_t *iclog;
2148 int changed = 0;
2149
2150 iclog = log->l_iclog;
2151 do {
2152 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2153 iclog->ic_state = XLOG_STATE_ACTIVE;
2154 iclog->ic_offset = 0;
2155 ASSERT(iclog->ic_callback == NULL);
2156 /*
2157 * If the number of ops in this iclog indicate it just
2158 * contains the dummy transaction, we can
2159 * change state into IDLE (the second time around).
2160 * Otherwise we should change the state into
2161 * NEED a dummy.
2162 * We don't need to cover the dummy.
2163 */
2164 if (!changed &&
2165 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2166 XLOG_COVER_OPS)) {
2167 changed = 1;
2168 } else {
2169 /*
2170 * We have two dirty iclogs so start over
2171 * This could also be num of ops indicates
2172 * this is not the dummy going out.
2173 */
2174 changed = 2;
2175 }
2176 iclog->ic_header.h_num_logops = 0;
2177 memset(iclog->ic_header.h_cycle_data, 0,
2178 sizeof(iclog->ic_header.h_cycle_data));
2179 iclog->ic_header.h_lsn = 0;
2180 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2181 /* do nothing */;
2182 else
2183 break; /* stop cleaning */
2184 iclog = iclog->ic_next;
2185 } while (iclog != log->l_iclog);
2186
2187 /* log is locked when we are called */
2188 /*
2189 * Change state for the dummy log recording.
2190 * We usually go to NEED. But we go to NEED2 if the changed indicates
2191 * we are done writing the dummy record.
2192 * If we are done with the second dummy recored (DONE2), then
2193 * we go to IDLE.
2194 */
2195 if (changed) {
2196 switch (log->l_covered_state) {
2197 case XLOG_STATE_COVER_IDLE:
2198 case XLOG_STATE_COVER_NEED:
2199 case XLOG_STATE_COVER_NEED2:
2200 log->l_covered_state = XLOG_STATE_COVER_NEED;
2201 break;
2202
2203 case XLOG_STATE_COVER_DONE:
2204 if (changed == 1)
2205 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2206 else
2207 log->l_covered_state = XLOG_STATE_COVER_NEED;
2208 break;
2209
2210 case XLOG_STATE_COVER_DONE2:
2211 if (changed == 1)
2212 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2213 else
2214 log->l_covered_state = XLOG_STATE_COVER_NEED;
2215 break;
2216
2217 default:
2218 ASSERT(0);
2219 }
2220 }
2221} /* xlog_state_clean_log */
2222
2223STATIC xfs_lsn_t
2224xlog_get_lowest_lsn(
2225 xlog_t *log)
2226{
2227 xlog_in_core_t *lsn_log;
2228 xfs_lsn_t lowest_lsn, lsn;
2229
2230 lsn_log = log->l_iclog;
2231 lowest_lsn = 0;
2232 do {
2233 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2234 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2235 if ((lsn && !lowest_lsn) ||
2236 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2237 lowest_lsn = lsn;
2238 }
2239 }
2240 lsn_log = lsn_log->ic_next;
2241 } while (lsn_log != log->l_iclog);
2242 return lowest_lsn;
2243}
2244
2245
2246STATIC void
2247xlog_state_do_callback(
2248 xlog_t *log,
2249 int aborted,
2250 xlog_in_core_t *ciclog)
2251{
2252 xlog_in_core_t *iclog;
2253 xlog_in_core_t *first_iclog; /* used to know when we've
2254 * processed all iclogs once */
2255 xfs_log_callback_t *cb, *cb_next;
2256 int flushcnt = 0;
2257 xfs_lsn_t lowest_lsn;
2258 int ioerrors; /* counter: iclogs with errors */
2259 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2260 int funcdidcallbacks; /* flag: function did callbacks */
2261 int repeats; /* for issuing console warnings if
2262 * looping too many times */
2263 int wake = 0;
2264
2265 spin_lock(&log->l_icloglock);
2266 first_iclog = iclog = log->l_iclog;
2267 ioerrors = 0;
2268 funcdidcallbacks = 0;
2269 repeats = 0;
2270
2271 do {
2272 /*
2273 * Scan all iclogs starting with the one pointed to by the
2274 * log. Reset this starting point each time the log is
2275 * unlocked (during callbacks).
2276 *
2277 * Keep looping through iclogs until one full pass is made
2278 * without running any callbacks.
2279 */
2280 first_iclog = log->l_iclog;
2281 iclog = log->l_iclog;
2282 loopdidcallbacks = 0;
2283 repeats++;
2284
2285 do {
2286
2287 /* skip all iclogs in the ACTIVE & DIRTY states */
2288 if (iclog->ic_state &
2289 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2290 iclog = iclog->ic_next;
2291 continue;
2292 }
2293
2294 /*
2295 * Between marking a filesystem SHUTDOWN and stopping
2296 * the log, we do flush all iclogs to disk (if there
2297 * wasn't a log I/O error). So, we do want things to
2298 * go smoothly in case of just a SHUTDOWN w/o a
2299 * LOG_IO_ERROR.
2300 */
2301 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2302 /*
2303 * Can only perform callbacks in order. Since
2304 * this iclog is not in the DONE_SYNC/
2305 * DO_CALLBACK state, we skip the rest and
2306 * just try to clean up. If we set our iclog
2307 * to DO_CALLBACK, we will not process it when
2308 * we retry since a previous iclog is in the
2309 * CALLBACK and the state cannot change since
2310 * we are holding the l_icloglock.
2311 */
2312 if (!(iclog->ic_state &
2313 (XLOG_STATE_DONE_SYNC |
2314 XLOG_STATE_DO_CALLBACK))) {
2315 if (ciclog && (ciclog->ic_state ==
2316 XLOG_STATE_DONE_SYNC)) {
2317 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2318 }
2319 break;
2320 }
2321 /*
2322 * We now have an iclog that is in either the
2323 * DO_CALLBACK or DONE_SYNC states. The other
2324 * states (WANT_SYNC, SYNCING, or CALLBACK were
2325 * caught by the above if and are going to
2326 * clean (i.e. we aren't doing their callbacks)
2327 * see the above if.
2328 */
2329
2330 /*
2331 * We will do one more check here to see if we
2332 * have chased our tail around.
2333 */
2334
2335 lowest_lsn = xlog_get_lowest_lsn(log);
2336 if (lowest_lsn &&
2337 XFS_LSN_CMP(lowest_lsn,
2338 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2339 iclog = iclog->ic_next;
2340 continue; /* Leave this iclog for
2341 * another thread */
2342 }
2343
2344 iclog->ic_state = XLOG_STATE_CALLBACK;
2345
2346
2347 /*
2348 * update the last_sync_lsn before we drop the
2349 * icloglock to ensure we are the only one that
2350 * can update it.
2351 */
2352 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2353 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2354 atomic64_set(&log->l_last_sync_lsn,
2355 be64_to_cpu(iclog->ic_header.h_lsn));
2356
2357 } else
2358 ioerrors++;
2359
2360 spin_unlock(&log->l_icloglock);
2361
2362 /*
2363 * Keep processing entries in the callback list until
2364 * we come around and it is empty. We need to
2365 * atomically see that the list is empty and change the
2366 * state to DIRTY so that we don't miss any more
2367 * callbacks being added.
2368 */
2369 spin_lock(&iclog->ic_callback_lock);
2370 cb = iclog->ic_callback;
2371 while (cb) {
2372 iclog->ic_callback_tail = &(iclog->ic_callback);
2373 iclog->ic_callback = NULL;
2374 spin_unlock(&iclog->ic_callback_lock);
2375
2376 /* perform callbacks in the order given */
2377 for (; cb; cb = cb_next) {
2378 cb_next = cb->cb_next;
2379 cb->cb_func(cb->cb_arg, aborted);
2380 }
2381 spin_lock(&iclog->ic_callback_lock);
2382 cb = iclog->ic_callback;
2383 }
2384
2385 loopdidcallbacks++;
2386 funcdidcallbacks++;
2387
2388 spin_lock(&log->l_icloglock);
2389 ASSERT(iclog->ic_callback == NULL);
2390 spin_unlock(&iclog->ic_callback_lock);
2391 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2392 iclog->ic_state = XLOG_STATE_DIRTY;
2393
2394 /*
2395 * Transition from DIRTY to ACTIVE if applicable.
2396 * NOP if STATE_IOERROR.
2397 */
2398 xlog_state_clean_log(log);
2399
2400 /* wake up threads waiting in xfs_log_force() */
2401 wake_up_all(&iclog->ic_force_wait);
2402
2403 iclog = iclog->ic_next;
2404 } while (first_iclog != iclog);
2405
2406 if (repeats > 5000) {
2407 flushcnt += repeats;
2408 repeats = 0;
2409 xfs_warn(log->l_mp,
2410 "%s: possible infinite loop (%d iterations)",
2411 __func__, flushcnt);
2412 }
2413 } while (!ioerrors && loopdidcallbacks);
2414
2415 /*
2416 * make one last gasp attempt to see if iclogs are being left in
2417 * limbo..
2418 */
2419#ifdef DEBUG
2420 if (funcdidcallbacks) {
2421 first_iclog = iclog = log->l_iclog;
2422 do {
2423 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2424 /*
2425 * Terminate the loop if iclogs are found in states
2426 * which will cause other threads to clean up iclogs.
2427 *
2428 * SYNCING - i/o completion will go through logs
2429 * DONE_SYNC - interrupt thread should be waiting for
2430 * l_icloglock
2431 * IOERROR - give up hope all ye who enter here
2432 */
2433 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2434 iclog->ic_state == XLOG_STATE_SYNCING ||
2435 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2436 iclog->ic_state == XLOG_STATE_IOERROR )
2437 break;
2438 iclog = iclog->ic_next;
2439 } while (first_iclog != iclog);
2440 }
2441#endif
2442
2443 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2444 wake = 1;
2445 spin_unlock(&log->l_icloglock);
2446
2447 if (wake)
2448 wake_up_all(&log->l_flush_wait);
2449}
2450
2451
2452/*
2453 * Finish transitioning this iclog to the dirty state.
2454 *
2455 * Make sure that we completely execute this routine only when this is
2456 * the last call to the iclog. There is a good chance that iclog flushes,
2457 * when we reach the end of the physical log, get turned into 2 separate
2458 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2459 * routine. By using the reference count bwritecnt, we guarantee that only
2460 * the second completion goes through.
2461 *
2462 * Callbacks could take time, so they are done outside the scope of the
2463 * global state machine log lock.
2464 */
2465STATIC void
2466xlog_state_done_syncing(
2467 xlog_in_core_t *iclog,
2468 int aborted)
2469{
2470 xlog_t *log = iclog->ic_log;
2471
2472 spin_lock(&log->l_icloglock);
2473
2474 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2475 iclog->ic_state == XLOG_STATE_IOERROR);
2476 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2477 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2478
2479
2480 /*
2481 * If we got an error, either on the first buffer, or in the case of
2482 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2483 * and none should ever be attempted to be written to disk
2484 * again.
2485 */
2486 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2487 if (--iclog->ic_bwritecnt == 1) {
2488 spin_unlock(&log->l_icloglock);
2489 return;
2490 }
2491 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2492 }
2493
2494 /*
2495 * Someone could be sleeping prior to writing out the next
2496 * iclog buffer, we wake them all, one will get to do the
2497 * I/O, the others get to wait for the result.
2498 */
2499 wake_up_all(&iclog->ic_write_wait);
2500 spin_unlock(&log->l_icloglock);
2501 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2502} /* xlog_state_done_syncing */
2503
2504
2505/*
2506 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2507 * sleep. We wait on the flush queue on the head iclog as that should be
2508 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2509 * we will wait here and all new writes will sleep until a sync completes.
2510 *
2511 * The in-core logs are used in a circular fashion. They are not used
2512 * out-of-order even when an iclog past the head is free.
2513 *
2514 * return:
2515 * * log_offset where xlog_write() can start writing into the in-core
2516 * log's data space.
2517 * * in-core log pointer to which xlog_write() should write.
2518 * * boolean indicating this is a continued write to an in-core log.
2519 * If this is the last write, then the in-core log's offset field
2520 * needs to be incremented, depending on the amount of data which
2521 * is copied.
2522 */
2523STATIC int
2524xlog_state_get_iclog_space(xlog_t *log,
2525 int len,
2526 xlog_in_core_t **iclogp,
2527 xlog_ticket_t *ticket,
2528 int *continued_write,
2529 int *logoffsetp)
2530{
2531 int log_offset;
2532 xlog_rec_header_t *head;
2533 xlog_in_core_t *iclog;
2534 int error;
2535
2536restart:
2537 spin_lock(&log->l_icloglock);
2538 if (XLOG_FORCED_SHUTDOWN(log)) {
2539 spin_unlock(&log->l_icloglock);
2540 return XFS_ERROR(EIO);
2541 }
2542
2543 iclog = log->l_iclog;
2544 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2545 XFS_STATS_INC(xs_log_noiclogs);
2546
2547 /* Wait for log writes to have flushed */
2548 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2549 goto restart;
2550 }
2551
2552 head = &iclog->ic_header;
2553
2554 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2555 log_offset = iclog->ic_offset;
2556
2557 /* On the 1st write to an iclog, figure out lsn. This works
2558 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2559 * committing to. If the offset is set, that's how many blocks
2560 * must be written.
2561 */
2562 if (log_offset == 0) {
2563 ticket->t_curr_res -= log->l_iclog_hsize;
2564 xlog_tic_add_region(ticket,
2565 log->l_iclog_hsize,
2566 XLOG_REG_TYPE_LRHEADER);
2567 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2568 head->h_lsn = cpu_to_be64(
2569 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2570 ASSERT(log->l_curr_block >= 0);
2571 }
2572
2573 /* If there is enough room to write everything, then do it. Otherwise,
2574 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2575 * bit is on, so this will get flushed out. Don't update ic_offset
2576 * until you know exactly how many bytes get copied. Therefore, wait
2577 * until later to update ic_offset.
2578 *
2579 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2580 * can fit into remaining data section.
2581 */
2582 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2583 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2584
2585 /*
2586 * If I'm the only one writing to this iclog, sync it to disk.
2587 * We need to do an atomic compare and decrement here to avoid
2588 * racing with concurrent atomic_dec_and_lock() calls in
2589 * xlog_state_release_iclog() when there is more than one
2590 * reference to the iclog.
2591 */
2592 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2593 /* we are the only one */
2594 spin_unlock(&log->l_icloglock);
2595 error = xlog_state_release_iclog(log, iclog);
2596 if (error)
2597 return error;
2598 } else {
2599 spin_unlock(&log->l_icloglock);
2600 }
2601 goto restart;
2602 }
2603
2604 /* Do we have enough room to write the full amount in the remainder
2605 * of this iclog? Or must we continue a write on the next iclog and
2606 * mark this iclog as completely taken? In the case where we switch
2607 * iclogs (to mark it taken), this particular iclog will release/sync
2608 * to disk in xlog_write().
2609 */
2610 if (len <= iclog->ic_size - iclog->ic_offset) {
2611 *continued_write = 0;
2612 iclog->ic_offset += len;
2613 } else {
2614 *continued_write = 1;
2615 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2616 }
2617 *iclogp = iclog;
2618
2619 ASSERT(iclog->ic_offset <= iclog->ic_size);
2620 spin_unlock(&log->l_icloglock);
2621
2622 *logoffsetp = log_offset;
2623 return 0;
2624} /* xlog_state_get_iclog_space */
2625
2626/* The first cnt-1 times through here we don't need to
2627 * move the grant write head because the permanent
2628 * reservation has reserved cnt times the unit amount.
2629 * Release part of current permanent unit reservation and
2630 * reset current reservation to be one units worth. Also
2631 * move grant reservation head forward.
2632 */
2633STATIC void
2634xlog_regrant_reserve_log_space(xlog_t *log,
2635 xlog_ticket_t *ticket)
2636{
2637 trace_xfs_log_regrant_reserve_enter(log, ticket);
2638
2639 if (ticket->t_cnt > 0)
2640 ticket->t_cnt--;
2641
2642 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2643 ticket->t_curr_res);
2644 xlog_grant_sub_space(log, &log->l_write_head.grant,
2645 ticket->t_curr_res);
2646 ticket->t_curr_res = ticket->t_unit_res;
2647 xlog_tic_reset_res(ticket);
2648
2649 trace_xfs_log_regrant_reserve_sub(log, ticket);
2650
2651 /* just return if we still have some of the pre-reserved space */
2652 if (ticket->t_cnt > 0)
2653 return;
2654
2655 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2656 ticket->t_unit_res);
2657
2658 trace_xfs_log_regrant_reserve_exit(log, ticket);
2659
2660 ticket->t_curr_res = ticket->t_unit_res;
2661 xlog_tic_reset_res(ticket);
2662} /* xlog_regrant_reserve_log_space */
2663
2664
2665/*
2666 * Give back the space left from a reservation.
2667 *
2668 * All the information we need to make a correct determination of space left
2669 * is present. For non-permanent reservations, things are quite easy. The
2670 * count should have been decremented to zero. We only need to deal with the
2671 * space remaining in the current reservation part of the ticket. If the
2672 * ticket contains a permanent reservation, there may be left over space which
2673 * needs to be released. A count of N means that N-1 refills of the current
2674 * reservation can be done before we need to ask for more space. The first
2675 * one goes to fill up the first current reservation. Once we run out of
2676 * space, the count will stay at zero and the only space remaining will be
2677 * in the current reservation field.
2678 */
2679STATIC void
2680xlog_ungrant_log_space(xlog_t *log,
2681 xlog_ticket_t *ticket)
2682{
2683 int bytes;
2684
2685 if (ticket->t_cnt > 0)
2686 ticket->t_cnt--;
2687
2688 trace_xfs_log_ungrant_enter(log, ticket);
2689 trace_xfs_log_ungrant_sub(log, ticket);
2690
2691 /*
2692 * If this is a permanent reservation ticket, we may be able to free
2693 * up more space based on the remaining count.
2694 */
2695 bytes = ticket->t_curr_res;
2696 if (ticket->t_cnt > 0) {
2697 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2698 bytes += ticket->t_unit_res*ticket->t_cnt;
2699 }
2700
2701 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2702 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2703
2704 trace_xfs_log_ungrant_exit(log, ticket);
2705
2706 xfs_log_space_wake(log->l_mp);
2707}
2708
2709/*
2710 * Flush iclog to disk if this is the last reference to the given iclog and
2711 * the WANT_SYNC bit is set.
2712 *
2713 * When this function is entered, the iclog is not necessarily in the
2714 * WANT_SYNC state. It may be sitting around waiting to get filled.
2715 *
2716 *
2717 */
2718STATIC int
2719xlog_state_release_iclog(
2720 xlog_t *log,
2721 xlog_in_core_t *iclog)
2722{
2723 int sync = 0; /* do we sync? */
2724
2725 if (iclog->ic_state & XLOG_STATE_IOERROR)
2726 return XFS_ERROR(EIO);
2727
2728 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2729 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2730 return 0;
2731
2732 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2733 spin_unlock(&log->l_icloglock);
2734 return XFS_ERROR(EIO);
2735 }
2736 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2737 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2738
2739 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2740 /* update tail before writing to iclog */
2741 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2742 sync++;
2743 iclog->ic_state = XLOG_STATE_SYNCING;
2744 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2745 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2746 /* cycle incremented when incrementing curr_block */
2747 }
2748 spin_unlock(&log->l_icloglock);
2749
2750 /*
2751 * We let the log lock go, so it's possible that we hit a log I/O
2752 * error or some other SHUTDOWN condition that marks the iclog
2753 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2754 * this iclog has consistent data, so we ignore IOERROR
2755 * flags after this point.
2756 */
2757 if (sync)
2758 return xlog_sync(log, iclog);
2759 return 0;
2760} /* xlog_state_release_iclog */
2761
2762
2763/*
2764 * This routine will mark the current iclog in the ring as WANT_SYNC
2765 * and move the current iclog pointer to the next iclog in the ring.
2766 * When this routine is called from xlog_state_get_iclog_space(), the
2767 * exact size of the iclog has not yet been determined. All we know is
2768 * that every data block. We have run out of space in this log record.
2769 */
2770STATIC void
2771xlog_state_switch_iclogs(xlog_t *log,
2772 xlog_in_core_t *iclog,
2773 int eventual_size)
2774{
2775 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2776 if (!eventual_size)
2777 eventual_size = iclog->ic_offset;
2778 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2779 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2780 log->l_prev_block = log->l_curr_block;
2781 log->l_prev_cycle = log->l_curr_cycle;
2782
2783 /* roll log?: ic_offset changed later */
2784 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2785
2786 /* Round up to next log-sunit */
2787 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2788 log->l_mp->m_sb.sb_logsunit > 1) {
2789 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2790 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2791 }
2792
2793 if (log->l_curr_block >= log->l_logBBsize) {
2794 log->l_curr_cycle++;
2795 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2796 log->l_curr_cycle++;
2797 log->l_curr_block -= log->l_logBBsize;
2798 ASSERT(log->l_curr_block >= 0);
2799 }
2800 ASSERT(iclog == log->l_iclog);
2801 log->l_iclog = iclog->ic_next;
2802} /* xlog_state_switch_iclogs */
2803
2804/*
2805 * Write out all data in the in-core log as of this exact moment in time.
2806 *
2807 * Data may be written to the in-core log during this call. However,
2808 * we don't guarantee this data will be written out. A change from past
2809 * implementation means this routine will *not* write out zero length LRs.
2810 *
2811 * Basically, we try and perform an intelligent scan of the in-core logs.
2812 * If we determine there is no flushable data, we just return. There is no
2813 * flushable data if:
2814 *
2815 * 1. the current iclog is active and has no data; the previous iclog
2816 * is in the active or dirty state.
2817 * 2. the current iclog is drity, and the previous iclog is in the
2818 * active or dirty state.
2819 *
2820 * We may sleep if:
2821 *
2822 * 1. the current iclog is not in the active nor dirty state.
2823 * 2. the current iclog dirty, and the previous iclog is not in the
2824 * active nor dirty state.
2825 * 3. the current iclog is active, and there is another thread writing
2826 * to this particular iclog.
2827 * 4. a) the current iclog is active and has no other writers
2828 * b) when we return from flushing out this iclog, it is still
2829 * not in the active nor dirty state.
2830 */
2831int
2832_xfs_log_force(
2833 struct xfs_mount *mp,
2834 uint flags,
2835 int *log_flushed)
2836{
2837 struct xlog *log = mp->m_log;
2838 struct xlog_in_core *iclog;
2839 xfs_lsn_t lsn;
2840
2841 XFS_STATS_INC(xs_log_force);
2842
2843 xlog_cil_force(log);
2844
2845 spin_lock(&log->l_icloglock);
2846
2847 iclog = log->l_iclog;
2848 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2849 spin_unlock(&log->l_icloglock);
2850 return XFS_ERROR(EIO);
2851 }
2852
2853 /* If the head iclog is not active nor dirty, we just attach
2854 * ourselves to the head and go to sleep.
2855 */
2856 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2857 iclog->ic_state == XLOG_STATE_DIRTY) {
2858 /*
2859 * If the head is dirty or (active and empty), then
2860 * we need to look at the previous iclog. If the previous
2861 * iclog is active or dirty we are done. There is nothing
2862 * to sync out. Otherwise, we attach ourselves to the
2863 * previous iclog and go to sleep.
2864 */
2865 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2866 (atomic_read(&iclog->ic_refcnt) == 0
2867 && iclog->ic_offset == 0)) {
2868 iclog = iclog->ic_prev;
2869 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2870 iclog->ic_state == XLOG_STATE_DIRTY)
2871 goto no_sleep;
2872 else
2873 goto maybe_sleep;
2874 } else {
2875 if (atomic_read(&iclog->ic_refcnt) == 0) {
2876 /* We are the only one with access to this
2877 * iclog. Flush it out now. There should
2878 * be a roundoff of zero to show that someone
2879 * has already taken care of the roundoff from
2880 * the previous sync.
2881 */
2882 atomic_inc(&iclog->ic_refcnt);
2883 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2884 xlog_state_switch_iclogs(log, iclog, 0);
2885 spin_unlock(&log->l_icloglock);
2886
2887 if (xlog_state_release_iclog(log, iclog))
2888 return XFS_ERROR(EIO);
2889
2890 if (log_flushed)
2891 *log_flushed = 1;
2892 spin_lock(&log->l_icloglock);
2893 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2894 iclog->ic_state != XLOG_STATE_DIRTY)
2895 goto maybe_sleep;
2896 else
2897 goto no_sleep;
2898 } else {
2899 /* Someone else is writing to this iclog.
2900 * Use its call to flush out the data. However,
2901 * the other thread may not force out this LR,
2902 * so we mark it WANT_SYNC.
2903 */
2904 xlog_state_switch_iclogs(log, iclog, 0);
2905 goto maybe_sleep;
2906 }
2907 }
2908 }
2909
2910 /* By the time we come around again, the iclog could've been filled
2911 * which would give it another lsn. If we have a new lsn, just
2912 * return because the relevant data has been flushed.
2913 */
2914maybe_sleep:
2915 if (flags & XFS_LOG_SYNC) {
2916 /*
2917 * We must check if we're shutting down here, before
2918 * we wait, while we're holding the l_icloglock.
2919 * Then we check again after waking up, in case our
2920 * sleep was disturbed by a bad news.
2921 */
2922 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2923 spin_unlock(&log->l_icloglock);
2924 return XFS_ERROR(EIO);
2925 }
2926 XFS_STATS_INC(xs_log_force_sleep);
2927 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2928 /*
2929 * No need to grab the log lock here since we're
2930 * only deciding whether or not to return EIO
2931 * and the memory read should be atomic.
2932 */
2933 if (iclog->ic_state & XLOG_STATE_IOERROR)
2934 return XFS_ERROR(EIO);
2935 if (log_flushed)
2936 *log_flushed = 1;
2937 } else {
2938
2939no_sleep:
2940 spin_unlock(&log->l_icloglock);
2941 }
2942 return 0;
2943}
2944
2945/*
2946 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2947 * about errors or whether the log was flushed or not. This is the normal
2948 * interface to use when trying to unpin items or move the log forward.
2949 */
2950void
2951xfs_log_force(
2952 xfs_mount_t *mp,
2953 uint flags)
2954{
2955 int error;
2956
2957 trace_xfs_log_force(mp, 0);
2958 error = _xfs_log_force(mp, flags, NULL);
2959 if (error)
2960 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2961}
2962
2963/*
2964 * Force the in-core log to disk for a specific LSN.
2965 *
2966 * Find in-core log with lsn.
2967 * If it is in the DIRTY state, just return.
2968 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2969 * state and go to sleep or return.
2970 * If it is in any other state, go to sleep or return.
2971 *
2972 * Synchronous forces are implemented with a signal variable. All callers
2973 * to force a given lsn to disk will wait on a the sv attached to the
2974 * specific in-core log. When given in-core log finally completes its
2975 * write to disk, that thread will wake up all threads waiting on the
2976 * sv.
2977 */
2978int
2979_xfs_log_force_lsn(
2980 struct xfs_mount *mp,
2981 xfs_lsn_t lsn,
2982 uint flags,
2983 int *log_flushed)
2984{
2985 struct xlog *log = mp->m_log;
2986 struct xlog_in_core *iclog;
2987 int already_slept = 0;
2988
2989 ASSERT(lsn != 0);
2990
2991 XFS_STATS_INC(xs_log_force);
2992
2993 lsn = xlog_cil_force_lsn(log, lsn);
2994 if (lsn == NULLCOMMITLSN)
2995 return 0;
2996
2997try_again:
2998 spin_lock(&log->l_icloglock);
2999 iclog = log->l_iclog;
3000 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3001 spin_unlock(&log->l_icloglock);
3002 return XFS_ERROR(EIO);
3003 }
3004
3005 do {
3006 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3007 iclog = iclog->ic_next;
3008 continue;
3009 }
3010
3011 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3012 spin_unlock(&log->l_icloglock);
3013 return 0;
3014 }
3015
3016 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3017 /*
3018 * We sleep here if we haven't already slept (e.g.
3019 * this is the first time we've looked at the correct
3020 * iclog buf) and the buffer before us is going to
3021 * be sync'ed. The reason for this is that if we
3022 * are doing sync transactions here, by waiting for
3023 * the previous I/O to complete, we can allow a few
3024 * more transactions into this iclog before we close
3025 * it down.
3026 *
3027 * Otherwise, we mark the buffer WANT_SYNC, and bump
3028 * up the refcnt so we can release the log (which
3029 * drops the ref count). The state switch keeps new
3030 * transaction commits from using this buffer. When
3031 * the current commits finish writing into the buffer,
3032 * the refcount will drop to zero and the buffer will
3033 * go out then.
3034 */
3035 if (!already_slept &&
3036 (iclog->ic_prev->ic_state &
3037 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3038 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3039
3040 XFS_STATS_INC(xs_log_force_sleep);
3041
3042 xlog_wait(&iclog->ic_prev->ic_write_wait,
3043 &log->l_icloglock);
3044 if (log_flushed)
3045 *log_flushed = 1;
3046 already_slept = 1;
3047 goto try_again;
3048 }
3049 atomic_inc(&iclog->ic_refcnt);
3050 xlog_state_switch_iclogs(log, iclog, 0);
3051 spin_unlock(&log->l_icloglock);
3052 if (xlog_state_release_iclog(log, iclog))
3053 return XFS_ERROR(EIO);
3054 if (log_flushed)
3055 *log_flushed = 1;
3056 spin_lock(&log->l_icloglock);
3057 }
3058
3059 if ((flags & XFS_LOG_SYNC) && /* sleep */
3060 !(iclog->ic_state &
3061 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3062 /*
3063 * Don't wait on completion if we know that we've
3064 * gotten a log write error.
3065 */
3066 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3067 spin_unlock(&log->l_icloglock);
3068 return XFS_ERROR(EIO);
3069 }
3070 XFS_STATS_INC(xs_log_force_sleep);
3071 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3072 /*
3073 * No need to grab the log lock here since we're
3074 * only deciding whether or not to return EIO
3075 * and the memory read should be atomic.
3076 */
3077 if (iclog->ic_state & XLOG_STATE_IOERROR)
3078 return XFS_ERROR(EIO);
3079
3080 if (log_flushed)
3081 *log_flushed = 1;
3082 } else { /* just return */
3083 spin_unlock(&log->l_icloglock);
3084 }
3085
3086 return 0;
3087 } while (iclog != log->l_iclog);
3088
3089 spin_unlock(&log->l_icloglock);
3090 return 0;
3091}
3092
3093/*
3094 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3095 * about errors or whether the log was flushed or not. This is the normal
3096 * interface to use when trying to unpin items or move the log forward.
3097 */
3098void
3099xfs_log_force_lsn(
3100 xfs_mount_t *mp,
3101 xfs_lsn_t lsn,
3102 uint flags)
3103{
3104 int error;
3105
3106 trace_xfs_log_force(mp, lsn);
3107 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3108 if (error)
3109 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3110}
3111
3112/*
3113 * Called when we want to mark the current iclog as being ready to sync to
3114 * disk.
3115 */
3116STATIC void
3117xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3118{
3119 assert_spin_locked(&log->l_icloglock);
3120
3121 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3122 xlog_state_switch_iclogs(log, iclog, 0);
3123 } else {
3124 ASSERT(iclog->ic_state &
3125 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3126 }
3127}
3128
3129
3130/*****************************************************************************
3131 *
3132 * TICKET functions
3133 *
3134 *****************************************************************************
3135 */
3136
3137/*
3138 * Free a used ticket when its refcount falls to zero.
3139 */
3140void
3141xfs_log_ticket_put(
3142 xlog_ticket_t *ticket)
3143{
3144 ASSERT(atomic_read(&ticket->t_ref) > 0);
3145 if (atomic_dec_and_test(&ticket->t_ref))
3146 kmem_zone_free(xfs_log_ticket_zone, ticket);
3147}
3148
3149xlog_ticket_t *
3150xfs_log_ticket_get(
3151 xlog_ticket_t *ticket)
3152{
3153 ASSERT(atomic_read(&ticket->t_ref) > 0);
3154 atomic_inc(&ticket->t_ref);
3155 return ticket;
3156}
3157
3158/*
3159 * Allocate and initialise a new log ticket.
3160 */
3161xlog_ticket_t *
3162xlog_ticket_alloc(
3163 struct xlog *log,
3164 int unit_bytes,
3165 int cnt,
3166 char client,
3167 bool permanent,
3168 xfs_km_flags_t alloc_flags)
3169{
3170 struct xlog_ticket *tic;
3171 uint num_headers;
3172 int iclog_space;
3173
3174 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3175 if (!tic)
3176 return NULL;
3177
3178 /*
3179 * Permanent reservations have up to 'cnt'-1 active log operations
3180 * in the log. A unit in this case is the amount of space for one
3181 * of these log operations. Normal reservations have a cnt of 1
3182 * and their unit amount is the total amount of space required.
3183 *
3184 * The following lines of code account for non-transaction data
3185 * which occupy space in the on-disk log.
3186 *
3187 * Normal form of a transaction is:
3188 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3189 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3190 *
3191 * We need to account for all the leadup data and trailer data
3192 * around the transaction data.
3193 * And then we need to account for the worst case in terms of using
3194 * more space.
3195 * The worst case will happen if:
3196 * - the placement of the transaction happens to be such that the
3197 * roundoff is at its maximum
3198 * - the transaction data is synced before the commit record is synced
3199 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3200 * Therefore the commit record is in its own Log Record.
3201 * This can happen as the commit record is called with its
3202 * own region to xlog_write().
3203 * This then means that in the worst case, roundoff can happen for
3204 * the commit-rec as well.
3205 * The commit-rec is smaller than padding in this scenario and so it is
3206 * not added separately.
3207 */
3208
3209 /* for trans header */
3210 unit_bytes += sizeof(xlog_op_header_t);
3211 unit_bytes += sizeof(xfs_trans_header_t);
3212
3213 /* for start-rec */
3214 unit_bytes += sizeof(xlog_op_header_t);
3215
3216 /*
3217 * for LR headers - the space for data in an iclog is the size minus
3218 * the space used for the headers. If we use the iclog size, then we
3219 * undercalculate the number of headers required.
3220 *
3221 * Furthermore - the addition of op headers for split-recs might
3222 * increase the space required enough to require more log and op
3223 * headers, so take that into account too.
3224 *
3225 * IMPORTANT: This reservation makes the assumption that if this
3226 * transaction is the first in an iclog and hence has the LR headers
3227 * accounted to it, then the remaining space in the iclog is
3228 * exclusively for this transaction. i.e. if the transaction is larger
3229 * than the iclog, it will be the only thing in that iclog.
3230 * Fundamentally, this means we must pass the entire log vector to
3231 * xlog_write to guarantee this.
3232 */
3233 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3234 num_headers = howmany(unit_bytes, iclog_space);
3235
3236 /* for split-recs - ophdrs added when data split over LRs */
3237 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3238
3239 /* add extra header reservations if we overrun */
3240 while (!num_headers ||
3241 howmany(unit_bytes, iclog_space) > num_headers) {
3242 unit_bytes += sizeof(xlog_op_header_t);
3243 num_headers++;
3244 }
3245 unit_bytes += log->l_iclog_hsize * num_headers;
3246
3247 /* for commit-rec LR header - note: padding will subsume the ophdr */
3248 unit_bytes += log->l_iclog_hsize;
3249
3250 /* for roundoff padding for transaction data and one for commit record */
3251 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3252 log->l_mp->m_sb.sb_logsunit > 1) {
3253 /* log su roundoff */
3254 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3255 } else {
3256 /* BB roundoff */
3257 unit_bytes += 2*BBSIZE;
3258 }
3259
3260 atomic_set(&tic->t_ref, 1);
3261 tic->t_task = current;
3262 INIT_LIST_HEAD(&tic->t_queue);
3263 tic->t_unit_res = unit_bytes;
3264 tic->t_curr_res = unit_bytes;
3265 tic->t_cnt = cnt;
3266 tic->t_ocnt = cnt;
3267 tic->t_tid = random32();
3268 tic->t_clientid = client;
3269 tic->t_flags = XLOG_TIC_INITED;
3270 tic->t_trans_type = 0;
3271 if (permanent)
3272 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3273
3274 xlog_tic_reset_res(tic);
3275
3276 return tic;
3277}
3278
3279
3280/******************************************************************************
3281 *
3282 * Log debug routines
3283 *
3284 ******************************************************************************
3285 */
3286#if defined(DEBUG)
3287/*
3288 * Make sure that the destination ptr is within the valid data region of
3289 * one of the iclogs. This uses backup pointers stored in a different
3290 * part of the log in case we trash the log structure.
3291 */
3292void
3293xlog_verify_dest_ptr(
3294 struct xlog *log,
3295 char *ptr)
3296{
3297 int i;
3298 int good_ptr = 0;
3299
3300 for (i = 0; i < log->l_iclog_bufs; i++) {
3301 if (ptr >= log->l_iclog_bak[i] &&
3302 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3303 good_ptr++;
3304 }
3305
3306 if (!good_ptr)
3307 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3308}
3309
3310/*
3311 * Check to make sure the grant write head didn't just over lap the tail. If
3312 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3313 * the cycles differ by exactly one and check the byte count.
3314 *
3315 * This check is run unlocked, so can give false positives. Rather than assert
3316 * on failures, use a warn-once flag and a panic tag to allow the admin to
3317 * determine if they want to panic the machine when such an error occurs. For
3318 * debug kernels this will have the same effect as using an assert but, unlinke
3319 * an assert, it can be turned off at runtime.
3320 */
3321STATIC void
3322xlog_verify_grant_tail(
3323 struct xlog *log)
3324{
3325 int tail_cycle, tail_blocks;
3326 int cycle, space;
3327
3328 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3329 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3330 if (tail_cycle != cycle) {
3331 if (cycle - 1 != tail_cycle &&
3332 !(log->l_flags & XLOG_TAIL_WARN)) {
3333 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3334 "%s: cycle - 1 != tail_cycle", __func__);
3335 log->l_flags |= XLOG_TAIL_WARN;
3336 }
3337
3338 if (space > BBTOB(tail_blocks) &&
3339 !(log->l_flags & XLOG_TAIL_WARN)) {
3340 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3341 "%s: space > BBTOB(tail_blocks)", __func__);
3342 log->l_flags |= XLOG_TAIL_WARN;
3343 }
3344 }
3345}
3346
3347/* check if it will fit */
3348STATIC void
3349xlog_verify_tail_lsn(xlog_t *log,
3350 xlog_in_core_t *iclog,
3351 xfs_lsn_t tail_lsn)
3352{
3353 int blocks;
3354
3355 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3356 blocks =
3357 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3358 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3359 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3360 } else {
3361 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3362
3363 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3364 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3365
3366 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3367 if (blocks < BTOBB(iclog->ic_offset) + 1)
3368 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3369 }
3370} /* xlog_verify_tail_lsn */
3371
3372/*
3373 * Perform a number of checks on the iclog before writing to disk.
3374 *
3375 * 1. Make sure the iclogs are still circular
3376 * 2. Make sure we have a good magic number
3377 * 3. Make sure we don't have magic numbers in the data
3378 * 4. Check fields of each log operation header for:
3379 * A. Valid client identifier
3380 * B. tid ptr value falls in valid ptr space (user space code)
3381 * C. Length in log record header is correct according to the
3382 * individual operation headers within record.
3383 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3384 * log, check the preceding blocks of the physical log to make sure all
3385 * the cycle numbers agree with the current cycle number.
3386 */
3387STATIC void
3388xlog_verify_iclog(xlog_t *log,
3389 xlog_in_core_t *iclog,
3390 int count,
3391 boolean_t syncing)
3392{
3393 xlog_op_header_t *ophead;
3394 xlog_in_core_t *icptr;
3395 xlog_in_core_2_t *xhdr;
3396 xfs_caddr_t ptr;
3397 xfs_caddr_t base_ptr;
3398 __psint_t field_offset;
3399 __uint8_t clientid;
3400 int len, i, j, k, op_len;
3401 int idx;
3402
3403 /* check validity of iclog pointers */
3404 spin_lock(&log->l_icloglock);
3405 icptr = log->l_iclog;
3406 for (i=0; i < log->l_iclog_bufs; i++) {
3407 if (icptr == NULL)
3408 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3409 icptr = icptr->ic_next;
3410 }
3411 if (icptr != log->l_iclog)
3412 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3413 spin_unlock(&log->l_icloglock);
3414
3415 /* check log magic numbers */
3416 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3417 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3418
3419 ptr = (xfs_caddr_t) &iclog->ic_header;
3420 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3421 ptr += BBSIZE) {
3422 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3423 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3424 __func__);
3425 }
3426
3427 /* check fields */
3428 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3429 ptr = iclog->ic_datap;
3430 base_ptr = ptr;
3431 ophead = (xlog_op_header_t *)ptr;
3432 xhdr = iclog->ic_data;
3433 for (i = 0; i < len; i++) {
3434 ophead = (xlog_op_header_t *)ptr;
3435
3436 /* clientid is only 1 byte */
3437 field_offset = (__psint_t)
3438 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3439 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3440 clientid = ophead->oh_clientid;
3441 } else {
3442 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3443 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3444 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3445 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3446 clientid = xlog_get_client_id(
3447 xhdr[j].hic_xheader.xh_cycle_data[k]);
3448 } else {
3449 clientid = xlog_get_client_id(
3450 iclog->ic_header.h_cycle_data[idx]);
3451 }
3452 }
3453 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3454 xfs_warn(log->l_mp,
3455 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3456 __func__, clientid, ophead,
3457 (unsigned long)field_offset);
3458
3459 /* check length */
3460 field_offset = (__psint_t)
3461 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3462 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3463 op_len = be32_to_cpu(ophead->oh_len);
3464 } else {
3465 idx = BTOBBT((__psint_t)&ophead->oh_len -
3466 (__psint_t)iclog->ic_datap);
3467 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3468 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3469 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3470 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3471 } else {
3472 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3473 }
3474 }
3475 ptr += sizeof(xlog_op_header_t) + op_len;
3476 }
3477} /* xlog_verify_iclog */
3478#endif
3479
3480/*
3481 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3482 */
3483STATIC int
3484xlog_state_ioerror(
3485 xlog_t *log)
3486{
3487 xlog_in_core_t *iclog, *ic;
3488
3489 iclog = log->l_iclog;
3490 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3491 /*
3492 * Mark all the incore logs IOERROR.
3493 * From now on, no log flushes will result.
3494 */
3495 ic = iclog;
3496 do {
3497 ic->ic_state = XLOG_STATE_IOERROR;
3498 ic = ic->ic_next;
3499 } while (ic != iclog);
3500 return 0;
3501 }
3502 /*
3503 * Return non-zero, if state transition has already happened.
3504 */
3505 return 1;
3506}
3507
3508/*
3509 * This is called from xfs_force_shutdown, when we're forcibly
3510 * shutting down the filesystem, typically because of an IO error.
3511 * Our main objectives here are to make sure that:
3512 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3513 * parties to find out, 'atomically'.
3514 * b. those who're sleeping on log reservations, pinned objects and
3515 * other resources get woken up, and be told the bad news.
3516 * c. nothing new gets queued up after (a) and (b) are done.
3517 * d. if !logerror, flush the iclogs to disk, then seal them off
3518 * for business.
3519 *
3520 * Note: for delayed logging the !logerror case needs to flush the regions
3521 * held in memory out to the iclogs before flushing them to disk. This needs
3522 * to be done before the log is marked as shutdown, otherwise the flush to the
3523 * iclogs will fail.
3524 */
3525int
3526xfs_log_force_umount(
3527 struct xfs_mount *mp,
3528 int logerror)
3529{
3530 xlog_t *log;
3531 int retval;
3532
3533 log = mp->m_log;
3534
3535 /*
3536 * If this happens during log recovery, don't worry about
3537 * locking; the log isn't open for business yet.
3538 */
3539 if (!log ||
3540 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3541 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3542 if (mp->m_sb_bp)
3543 XFS_BUF_DONE(mp->m_sb_bp);
3544 return 0;
3545 }
3546
3547 /*
3548 * Somebody could've already done the hard work for us.
3549 * No need to get locks for this.
3550 */
3551 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3552 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3553 return 1;
3554 }
3555 retval = 0;
3556
3557 /*
3558 * Flush the in memory commit item list before marking the log as
3559 * being shut down. We need to do it in this order to ensure all the
3560 * completed transactions are flushed to disk with the xfs_log_force()
3561 * call below.
3562 */
3563 if (!logerror)
3564 xlog_cil_force(log);
3565
3566 /*
3567 * mark the filesystem and the as in a shutdown state and wake
3568 * everybody up to tell them the bad news.
3569 */
3570 spin_lock(&log->l_icloglock);
3571 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3572 if (mp->m_sb_bp)
3573 XFS_BUF_DONE(mp->m_sb_bp);
3574
3575 /*
3576 * This flag is sort of redundant because of the mount flag, but
3577 * it's good to maintain the separation between the log and the rest
3578 * of XFS.
3579 */
3580 log->l_flags |= XLOG_IO_ERROR;
3581
3582 /*
3583 * If we hit a log error, we want to mark all the iclogs IOERROR
3584 * while we're still holding the loglock.
3585 */
3586 if (logerror)
3587 retval = xlog_state_ioerror(log);
3588 spin_unlock(&log->l_icloglock);
3589
3590 /*
3591 * We don't want anybody waiting for log reservations after this. That
3592 * means we have to wake up everybody queued up on reserveq as well as
3593 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3594 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3595 * action is protected by the grant locks.
3596 */
3597 xlog_grant_head_wake_all(&log->l_reserve_head);
3598 xlog_grant_head_wake_all(&log->l_write_head);
3599
3600 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3601 ASSERT(!logerror);
3602 /*
3603 * Force the incore logs to disk before shutting the
3604 * log down completely.
3605 */
3606 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3607
3608 spin_lock(&log->l_icloglock);
3609 retval = xlog_state_ioerror(log);
3610 spin_unlock(&log->l_icloglock);
3611 }
3612 /*
3613 * Wake up everybody waiting on xfs_log_force.
3614 * Callback all log item committed functions as if the
3615 * log writes were completed.
3616 */
3617 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3618
3619#ifdef XFSERRORDEBUG
3620 {
3621 xlog_in_core_t *iclog;
3622
3623 spin_lock(&log->l_icloglock);
3624 iclog = log->l_iclog;
3625 do {
3626 ASSERT(iclog->ic_callback == 0);
3627 iclog = iclog->ic_next;
3628 } while (iclog != log->l_iclog);
3629 spin_unlock(&log->l_icloglock);
3630 }
3631#endif
3632 /* return non-zero if log IOERROR transition had already happened */
3633 return retval;
3634}
3635
3636STATIC int
3637xlog_iclogs_empty(xlog_t *log)
3638{
3639 xlog_in_core_t *iclog;
3640
3641 iclog = log->l_iclog;
3642 do {
3643 /* endianness does not matter here, zero is zero in
3644 * any language.
3645 */
3646 if (iclog->ic_header.h_num_logops)
3647 return 0;
3648 iclog = iclog->ic_next;
3649 } while (iclog != log->l_iclog);
3650 return 1;
3651}
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_shared.h"
21#include "xfs_format.h"
22#include "xfs_log_format.h"
23#include "xfs_trans_resv.h"
24#include "xfs_mount.h"
25#include "xfs_error.h"
26#include "xfs_trans.h"
27#include "xfs_trans_priv.h"
28#include "xfs_log.h"
29#include "xfs_log_priv.h"
30#include "xfs_log_recover.h"
31#include "xfs_inode.h"
32#include "xfs_trace.h"
33#include "xfs_fsops.h"
34#include "xfs_cksum.h"
35#include "xfs_sysfs.h"
36#include "xfs_sb.h"
37
38kmem_zone_t *xfs_log_ticket_zone;
39
40/* Local miscellaneous function prototypes */
41STATIC int
42xlog_commit_record(
43 struct xlog *log,
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
47
48STATIC struct xlog *
49xlog_alloc_log(
50 struct xfs_mount *mp,
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
53 int num_bblks);
54STATIC int
55xlog_space_left(
56 struct xlog *log,
57 atomic64_t *head);
58STATIC int
59xlog_sync(
60 struct xlog *log,
61 struct xlog_in_core *iclog);
62STATIC void
63xlog_dealloc_log(
64 struct xlog *log);
65
66/* local state machine functions */
67STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
68STATIC void
69xlog_state_do_callback(
70 struct xlog *log,
71 int aborted,
72 struct xlog_in_core *iclog);
73STATIC int
74xlog_state_get_iclog_space(
75 struct xlog *log,
76 int len,
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
79 int *continued_write,
80 int *logoffsetp);
81STATIC int
82xlog_state_release_iclog(
83 struct xlog *log,
84 struct xlog_in_core *iclog);
85STATIC void
86xlog_state_switch_iclogs(
87 struct xlog *log,
88 struct xlog_in_core *iclog,
89 int eventual_size);
90STATIC void
91xlog_state_want_sync(
92 struct xlog *log,
93 struct xlog_in_core *iclog);
94
95STATIC void
96xlog_grant_push_ail(
97 struct xlog *log,
98 int need_bytes);
99STATIC void
100xlog_regrant_reserve_log_space(
101 struct xlog *log,
102 struct xlog_ticket *ticket);
103STATIC void
104xlog_ungrant_log_space(
105 struct xlog *log,
106 struct xlog_ticket *ticket);
107
108#if defined(DEBUG)
109STATIC void
110xlog_verify_dest_ptr(
111 struct xlog *log,
112 void *ptr);
113STATIC void
114xlog_verify_grant_tail(
115 struct xlog *log);
116STATIC void
117xlog_verify_iclog(
118 struct xlog *log,
119 struct xlog_in_core *iclog,
120 int count,
121 bool syncing);
122STATIC void
123xlog_verify_tail_lsn(
124 struct xlog *log,
125 struct xlog_in_core *iclog,
126 xfs_lsn_t tail_lsn);
127#else
128#define xlog_verify_dest_ptr(a,b)
129#define xlog_verify_grant_tail(a)
130#define xlog_verify_iclog(a,b,c,d)
131#define xlog_verify_tail_lsn(a,b,c)
132#endif
133
134STATIC int
135xlog_iclogs_empty(
136 struct xlog *log);
137
138static void
139xlog_grant_sub_space(
140 struct xlog *log,
141 atomic64_t *head,
142 int bytes)
143{
144 int64_t head_val = atomic64_read(head);
145 int64_t new, old;
146
147 do {
148 int cycle, space;
149
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
151
152 space -= bytes;
153 if (space < 0) {
154 space += log->l_logsize;
155 cycle--;
156 }
157
158 old = head_val;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
162}
163
164static void
165xlog_grant_add_space(
166 struct xlog *log,
167 atomic64_t *head,
168 int bytes)
169{
170 int64_t head_val = atomic64_read(head);
171 int64_t new, old;
172
173 do {
174 int tmp;
175 int cycle, space;
176
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
178
179 tmp = log->l_logsize - space;
180 if (tmp > bytes)
181 space += bytes;
182 else {
183 space = bytes - tmp;
184 cycle++;
185 }
186
187 old = head_val;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
191}
192
193STATIC void
194xlog_grant_head_init(
195 struct xlog_grant_head *head)
196{
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
200}
201
202STATIC void
203xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
205{
206 struct xlog_ticket *tic;
207
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
212}
213
214static inline int
215xlog_ticket_reservation(
216 struct xlog *log,
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
219{
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
223 } else {
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
226 else
227 return tic->t_unit_res;
228 }
229}
230
231STATIC bool
232xlog_grant_head_wake(
233 struct xlog *log,
234 struct xlog_grant_head *head,
235 int *free_bytes)
236{
237 struct xlog_ticket *tic;
238 int need_bytes;
239
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
243 return false;
244
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
248 }
249
250 return true;
251}
252
253STATIC int
254xlog_grant_head_wait(
255 struct xlog *log,
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
260{
261 list_add_tail(&tic->t_queue, &head->waiters);
262
263 do {
264 if (XLOG_FORCED_SHUTDOWN(log))
265 goto shutdown;
266 xlog_grant_push_ail(log, need_bytes);
267
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
270
271 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
272
273 trace_xfs_log_grant_sleep(log, tic);
274 schedule();
275 trace_xfs_log_grant_wake(log, tic);
276
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
279 goto shutdown;
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
281
282 list_del_init(&tic->t_queue);
283 return 0;
284shutdown:
285 list_del_init(&tic->t_queue);
286 return -EIO;
287}
288
289/*
290 * Atomically get the log space required for a log ticket.
291 *
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
294 *
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
298 * every pass.
299 *
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
305 */
306STATIC int
307xlog_grant_head_check(
308 struct xlog *log,
309 struct xlog_grant_head *head,
310 struct xlog_ticket *tic,
311 int *need_bytes)
312{
313 int free_bytes;
314 int error = 0;
315
316 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
317
318 /*
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
323 */
324 *need_bytes = xlog_ticket_reservation(log, head, tic);
325 free_bytes = xlog_space_left(log, &head->grant);
326 if (!list_empty_careful(&head->waiters)) {
327 spin_lock(&head->lock);
328 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
329 free_bytes < *need_bytes) {
330 error = xlog_grant_head_wait(log, head, tic,
331 *need_bytes);
332 }
333 spin_unlock(&head->lock);
334 } else if (free_bytes < *need_bytes) {
335 spin_lock(&head->lock);
336 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
337 spin_unlock(&head->lock);
338 }
339
340 return error;
341}
342
343static void
344xlog_tic_reset_res(xlog_ticket_t *tic)
345{
346 tic->t_res_num = 0;
347 tic->t_res_arr_sum = 0;
348 tic->t_res_num_ophdrs = 0;
349}
350
351static void
352xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
353{
354 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
355 /* add to overflow and start again */
356 tic->t_res_o_flow += tic->t_res_arr_sum;
357 tic->t_res_num = 0;
358 tic->t_res_arr_sum = 0;
359 }
360
361 tic->t_res_arr[tic->t_res_num].r_len = len;
362 tic->t_res_arr[tic->t_res_num].r_type = type;
363 tic->t_res_arr_sum += len;
364 tic->t_res_num++;
365}
366
367/*
368 * Replenish the byte reservation required by moving the grant write head.
369 */
370int
371xfs_log_regrant(
372 struct xfs_mount *mp,
373 struct xlog_ticket *tic)
374{
375 struct xlog *log = mp->m_log;
376 int need_bytes;
377 int error = 0;
378
379 if (XLOG_FORCED_SHUTDOWN(log))
380 return -EIO;
381
382 XFS_STATS_INC(mp, xs_try_logspace);
383
384 /*
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
389 */
390 tic->t_tid++;
391
392 xlog_grant_push_ail(log, tic->t_unit_res);
393
394 tic->t_curr_res = tic->t_unit_res;
395 xlog_tic_reset_res(tic);
396
397 if (tic->t_cnt > 0)
398 return 0;
399
400 trace_xfs_log_regrant(log, tic);
401
402 error = xlog_grant_head_check(log, &log->l_write_head, tic,
403 &need_bytes);
404 if (error)
405 goto out_error;
406
407 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
408 trace_xfs_log_regrant_exit(log, tic);
409 xlog_verify_grant_tail(log);
410 return 0;
411
412out_error:
413 /*
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
417 */
418 tic->t_curr_res = 0;
419 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
420 return error;
421}
422
423/*
424 * Reserve log space and return a ticket corresponding the reservation.
425 *
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
430 */
431int
432xfs_log_reserve(
433 struct xfs_mount *mp,
434 int unit_bytes,
435 int cnt,
436 struct xlog_ticket **ticp,
437 __uint8_t client,
438 bool permanent,
439 uint t_type)
440{
441 struct xlog *log = mp->m_log;
442 struct xlog_ticket *tic;
443 int need_bytes;
444 int error = 0;
445
446 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
447
448 if (XLOG_FORCED_SHUTDOWN(log))
449 return -EIO;
450
451 XFS_STATS_INC(mp, xs_try_logspace);
452
453 ASSERT(*ticp == NULL);
454 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
455 KM_SLEEP | KM_MAYFAIL);
456 if (!tic)
457 return -ENOMEM;
458
459 tic->t_trans_type = t_type;
460 *ticp = tic;
461
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
463 : tic->t_unit_res);
464
465 trace_xfs_log_reserve(log, tic);
466
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
468 &need_bytes);
469 if (error)
470 goto out_error;
471
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
476 return 0;
477
478out_error:
479 /*
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
483 */
484 tic->t_curr_res = 0;
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
486 return error;
487}
488
489
490/*
491 * NOTES:
492 *
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
495 */
496
497/*
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
510 */
511xfs_lsn_t
512xfs_log_done(
513 struct xfs_mount *mp,
514 struct xlog_ticket *ticket,
515 struct xlog_in_core **iclog,
516 bool regrant)
517{
518 struct xlog *log = mp->m_log;
519 xfs_lsn_t lsn = 0;
520
521 if (XLOG_FORCED_SHUTDOWN(log) ||
522 /*
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
525 */
526 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
527 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
528 lsn = (xfs_lsn_t) -1;
529 regrant = false;
530 }
531
532
533 if (!regrant) {
534 trace_xfs_log_done_nonperm(log, ticket);
535
536 /*
537 * Release ticket if not permanent reservation or a specific
538 * request has been made to release a permanent reservation.
539 */
540 xlog_ungrant_log_space(log, ticket);
541 } else {
542 trace_xfs_log_done_perm(log, ticket);
543
544 xlog_regrant_reserve_log_space(log, ticket);
545 /* If this ticket was a permanent reservation and we aren't
546 * trying to release it, reset the inited flags; so next time
547 * we write, a start record will be written out.
548 */
549 ticket->t_flags |= XLOG_TIC_INITED;
550 }
551
552 xfs_log_ticket_put(ticket);
553 return lsn;
554}
555
556/*
557 * Attaches a new iclog I/O completion callback routine during
558 * transaction commit. If the log is in error state, a non-zero
559 * return code is handed back and the caller is responsible for
560 * executing the callback at an appropriate time.
561 */
562int
563xfs_log_notify(
564 struct xfs_mount *mp,
565 struct xlog_in_core *iclog,
566 xfs_log_callback_t *cb)
567{
568 int abortflg;
569
570 spin_lock(&iclog->ic_callback_lock);
571 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
572 if (!abortflg) {
573 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
574 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
575 cb->cb_next = NULL;
576 *(iclog->ic_callback_tail) = cb;
577 iclog->ic_callback_tail = &(cb->cb_next);
578 }
579 spin_unlock(&iclog->ic_callback_lock);
580 return abortflg;
581}
582
583int
584xfs_log_release_iclog(
585 struct xfs_mount *mp,
586 struct xlog_in_core *iclog)
587{
588 if (xlog_state_release_iclog(mp->m_log, iclog)) {
589 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
590 return -EIO;
591 }
592
593 return 0;
594}
595
596/*
597 * Mount a log filesystem
598 *
599 * mp - ubiquitous xfs mount point structure
600 * log_target - buftarg of on-disk log device
601 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
602 * num_bblocks - Number of BBSIZE blocks in on-disk log
603 *
604 * Return error or zero.
605 */
606int
607xfs_log_mount(
608 xfs_mount_t *mp,
609 xfs_buftarg_t *log_target,
610 xfs_daddr_t blk_offset,
611 int num_bblks)
612{
613 int error = 0;
614 int min_logfsbs;
615
616 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
617 xfs_notice(mp, "Mounting V%d Filesystem",
618 XFS_SB_VERSION_NUM(&mp->m_sb));
619 } else {
620 xfs_notice(mp,
621"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
622 XFS_SB_VERSION_NUM(&mp->m_sb));
623 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
624 }
625
626 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
627 if (IS_ERR(mp->m_log)) {
628 error = PTR_ERR(mp->m_log);
629 goto out;
630 }
631
632 /*
633 * Validate the given log space and drop a critical message via syslog
634 * if the log size is too small that would lead to some unexpected
635 * situations in transaction log space reservation stage.
636 *
637 * Note: we can't just reject the mount if the validation fails. This
638 * would mean that people would have to downgrade their kernel just to
639 * remedy the situation as there is no way to grow the log (short of
640 * black magic surgery with xfs_db).
641 *
642 * We can, however, reject mounts for CRC format filesystems, as the
643 * mkfs binary being used to make the filesystem should never create a
644 * filesystem with a log that is too small.
645 */
646 min_logfsbs = xfs_log_calc_minimum_size(mp);
647
648 if (mp->m_sb.sb_logblocks < min_logfsbs) {
649 xfs_warn(mp,
650 "Log size %d blocks too small, minimum size is %d blocks",
651 mp->m_sb.sb_logblocks, min_logfsbs);
652 error = -EINVAL;
653 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
654 xfs_warn(mp,
655 "Log size %d blocks too large, maximum size is %lld blocks",
656 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
657 error = -EINVAL;
658 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
659 xfs_warn(mp,
660 "log size %lld bytes too large, maximum size is %lld bytes",
661 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
662 XFS_MAX_LOG_BYTES);
663 error = -EINVAL;
664 }
665 if (error) {
666 if (xfs_sb_version_hascrc(&mp->m_sb)) {
667 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
668 ASSERT(0);
669 goto out_free_log;
670 }
671 xfs_crit(mp, "Log size out of supported range.");
672 xfs_crit(mp,
673"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
674 }
675
676 /*
677 * Initialize the AIL now we have a log.
678 */
679 error = xfs_trans_ail_init(mp);
680 if (error) {
681 xfs_warn(mp, "AIL initialisation failed: error %d", error);
682 goto out_free_log;
683 }
684 mp->m_log->l_ailp = mp->m_ail;
685
686 /*
687 * skip log recovery on a norecovery mount. pretend it all
688 * just worked.
689 */
690 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
691 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
692
693 if (readonly)
694 mp->m_flags &= ~XFS_MOUNT_RDONLY;
695
696 error = xlog_recover(mp->m_log);
697
698 if (readonly)
699 mp->m_flags |= XFS_MOUNT_RDONLY;
700 if (error) {
701 xfs_warn(mp, "log mount/recovery failed: error %d",
702 error);
703 xlog_recover_cancel(mp->m_log);
704 goto out_destroy_ail;
705 }
706 }
707
708 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
709 "log");
710 if (error)
711 goto out_destroy_ail;
712
713 /* Normal transactions can now occur */
714 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
715
716 /*
717 * Now the log has been fully initialised and we know were our
718 * space grant counters are, we can initialise the permanent ticket
719 * needed for delayed logging to work.
720 */
721 xlog_cil_init_post_recovery(mp->m_log);
722
723 return 0;
724
725out_destroy_ail:
726 xfs_trans_ail_destroy(mp);
727out_free_log:
728 xlog_dealloc_log(mp->m_log);
729out:
730 return error;
731}
732
733/*
734 * Finish the recovery of the file system. This is separate from the
735 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
736 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
737 * here.
738 *
739 * If we finish recovery successfully, start the background log work. If we are
740 * not doing recovery, then we have a RO filesystem and we don't need to start
741 * it.
742 */
743int
744xfs_log_mount_finish(
745 struct xfs_mount *mp)
746{
747 int error = 0;
748
749 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
750 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
751 return 0;
752 }
753
754 error = xlog_recover_finish(mp->m_log);
755 if (!error)
756 xfs_log_work_queue(mp);
757
758 return error;
759}
760
761/*
762 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
763 * the log.
764 */
765int
766xfs_log_mount_cancel(
767 struct xfs_mount *mp)
768{
769 int error;
770
771 error = xlog_recover_cancel(mp->m_log);
772 xfs_log_unmount(mp);
773
774 return error;
775}
776
777/*
778 * Final log writes as part of unmount.
779 *
780 * Mark the filesystem clean as unmount happens. Note that during relocation
781 * this routine needs to be executed as part of source-bag while the
782 * deallocation must not be done until source-end.
783 */
784
785/*
786 * Unmount record used to have a string "Unmount filesystem--" in the
787 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
788 * We just write the magic number now since that particular field isn't
789 * currently architecture converted and "Unmount" is a bit foo.
790 * As far as I know, there weren't any dependencies on the old behaviour.
791 */
792
793int
794xfs_log_unmount_write(xfs_mount_t *mp)
795{
796 struct xlog *log = mp->m_log;
797 xlog_in_core_t *iclog;
798#ifdef DEBUG
799 xlog_in_core_t *first_iclog;
800#endif
801 xlog_ticket_t *tic = NULL;
802 xfs_lsn_t lsn;
803 int error;
804
805 /*
806 * Don't write out unmount record on read-only mounts.
807 * Or, if we are doing a forced umount (typically because of IO errors).
808 */
809 if (mp->m_flags & XFS_MOUNT_RDONLY)
810 return 0;
811
812 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
813 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
814
815#ifdef DEBUG
816 first_iclog = iclog = log->l_iclog;
817 do {
818 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
819 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
820 ASSERT(iclog->ic_offset == 0);
821 }
822 iclog = iclog->ic_next;
823 } while (iclog != first_iclog);
824#endif
825 if (! (XLOG_FORCED_SHUTDOWN(log))) {
826 error = xfs_log_reserve(mp, 600, 1, &tic,
827 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
828 if (!error) {
829 /* the data section must be 32 bit size aligned */
830 struct {
831 __uint16_t magic;
832 __uint16_t pad1;
833 __uint32_t pad2; /* may as well make it 64 bits */
834 } magic = {
835 .magic = XLOG_UNMOUNT_TYPE,
836 };
837 struct xfs_log_iovec reg = {
838 .i_addr = &magic,
839 .i_len = sizeof(magic),
840 .i_type = XLOG_REG_TYPE_UNMOUNT,
841 };
842 struct xfs_log_vec vec = {
843 .lv_niovecs = 1,
844 .lv_iovecp = ®,
845 };
846
847 /* remove inited flag, and account for space used */
848 tic->t_flags = 0;
849 tic->t_curr_res -= sizeof(magic);
850 error = xlog_write(log, &vec, tic, &lsn,
851 NULL, XLOG_UNMOUNT_TRANS);
852 /*
853 * At this point, we're umounting anyway,
854 * so there's no point in transitioning log state
855 * to IOERROR. Just continue...
856 */
857 }
858
859 if (error)
860 xfs_alert(mp, "%s: unmount record failed", __func__);
861
862
863 spin_lock(&log->l_icloglock);
864 iclog = log->l_iclog;
865 atomic_inc(&iclog->ic_refcnt);
866 xlog_state_want_sync(log, iclog);
867 spin_unlock(&log->l_icloglock);
868 error = xlog_state_release_iclog(log, iclog);
869
870 spin_lock(&log->l_icloglock);
871 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
872 iclog->ic_state == XLOG_STATE_DIRTY)) {
873 if (!XLOG_FORCED_SHUTDOWN(log)) {
874 xlog_wait(&iclog->ic_force_wait,
875 &log->l_icloglock);
876 } else {
877 spin_unlock(&log->l_icloglock);
878 }
879 } else {
880 spin_unlock(&log->l_icloglock);
881 }
882 if (tic) {
883 trace_xfs_log_umount_write(log, tic);
884 xlog_ungrant_log_space(log, tic);
885 xfs_log_ticket_put(tic);
886 }
887 } else {
888 /*
889 * We're already in forced_shutdown mode, couldn't
890 * even attempt to write out the unmount transaction.
891 *
892 * Go through the motions of sync'ing and releasing
893 * the iclog, even though no I/O will actually happen,
894 * we need to wait for other log I/Os that may already
895 * be in progress. Do this as a separate section of
896 * code so we'll know if we ever get stuck here that
897 * we're in this odd situation of trying to unmount
898 * a file system that went into forced_shutdown as
899 * the result of an unmount..
900 */
901 spin_lock(&log->l_icloglock);
902 iclog = log->l_iclog;
903 atomic_inc(&iclog->ic_refcnt);
904
905 xlog_state_want_sync(log, iclog);
906 spin_unlock(&log->l_icloglock);
907 error = xlog_state_release_iclog(log, iclog);
908
909 spin_lock(&log->l_icloglock);
910
911 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
912 || iclog->ic_state == XLOG_STATE_DIRTY
913 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
914
915 xlog_wait(&iclog->ic_force_wait,
916 &log->l_icloglock);
917 } else {
918 spin_unlock(&log->l_icloglock);
919 }
920 }
921
922 return error;
923} /* xfs_log_unmount_write */
924
925/*
926 * Empty the log for unmount/freeze.
927 *
928 * To do this, we first need to shut down the background log work so it is not
929 * trying to cover the log as we clean up. We then need to unpin all objects in
930 * the log so we can then flush them out. Once they have completed their IO and
931 * run the callbacks removing themselves from the AIL, we can write the unmount
932 * record.
933 */
934void
935xfs_log_quiesce(
936 struct xfs_mount *mp)
937{
938 cancel_delayed_work_sync(&mp->m_log->l_work);
939 xfs_log_force(mp, XFS_LOG_SYNC);
940
941 /*
942 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
943 * will push it, xfs_wait_buftarg() will not wait for it. Further,
944 * xfs_buf_iowait() cannot be used because it was pushed with the
945 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
946 * the IO to complete.
947 */
948 xfs_ail_push_all_sync(mp->m_ail);
949 xfs_wait_buftarg(mp->m_ddev_targp);
950 xfs_buf_lock(mp->m_sb_bp);
951 xfs_buf_unlock(mp->m_sb_bp);
952
953 xfs_log_unmount_write(mp);
954}
955
956/*
957 * Shut down and release the AIL and Log.
958 *
959 * During unmount, we need to ensure we flush all the dirty metadata objects
960 * from the AIL so that the log is empty before we write the unmount record to
961 * the log. Once this is done, we can tear down the AIL and the log.
962 */
963void
964xfs_log_unmount(
965 struct xfs_mount *mp)
966{
967 xfs_log_quiesce(mp);
968
969 xfs_trans_ail_destroy(mp);
970
971 xfs_sysfs_del(&mp->m_log->l_kobj);
972
973 xlog_dealloc_log(mp->m_log);
974}
975
976void
977xfs_log_item_init(
978 struct xfs_mount *mp,
979 struct xfs_log_item *item,
980 int type,
981 const struct xfs_item_ops *ops)
982{
983 item->li_mountp = mp;
984 item->li_ailp = mp->m_ail;
985 item->li_type = type;
986 item->li_ops = ops;
987 item->li_lv = NULL;
988
989 INIT_LIST_HEAD(&item->li_ail);
990 INIT_LIST_HEAD(&item->li_cil);
991}
992
993/*
994 * Wake up processes waiting for log space after we have moved the log tail.
995 */
996void
997xfs_log_space_wake(
998 struct xfs_mount *mp)
999{
1000 struct xlog *log = mp->m_log;
1001 int free_bytes;
1002
1003 if (XLOG_FORCED_SHUTDOWN(log))
1004 return;
1005
1006 if (!list_empty_careful(&log->l_write_head.waiters)) {
1007 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1008
1009 spin_lock(&log->l_write_head.lock);
1010 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1011 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1012 spin_unlock(&log->l_write_head.lock);
1013 }
1014
1015 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1016 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1017
1018 spin_lock(&log->l_reserve_head.lock);
1019 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1020 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1021 spin_unlock(&log->l_reserve_head.lock);
1022 }
1023}
1024
1025/*
1026 * Determine if we have a transaction that has gone to disk that needs to be
1027 * covered. To begin the transition to the idle state firstly the log needs to
1028 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1029 * we start attempting to cover the log.
1030 *
1031 * Only if we are then in a state where covering is needed, the caller is
1032 * informed that dummy transactions are required to move the log into the idle
1033 * state.
1034 *
1035 * If there are any items in the AIl or CIL, then we do not want to attempt to
1036 * cover the log as we may be in a situation where there isn't log space
1037 * available to run a dummy transaction and this can lead to deadlocks when the
1038 * tail of the log is pinned by an item that is modified in the CIL. Hence
1039 * there's no point in running a dummy transaction at this point because we
1040 * can't start trying to idle the log until both the CIL and AIL are empty.
1041 */
1042int
1043xfs_log_need_covered(xfs_mount_t *mp)
1044{
1045 struct xlog *log = mp->m_log;
1046 int needed = 0;
1047
1048 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1049 return 0;
1050
1051 if (!xlog_cil_empty(log))
1052 return 0;
1053
1054 spin_lock(&log->l_icloglock);
1055 switch (log->l_covered_state) {
1056 case XLOG_STATE_COVER_DONE:
1057 case XLOG_STATE_COVER_DONE2:
1058 case XLOG_STATE_COVER_IDLE:
1059 break;
1060 case XLOG_STATE_COVER_NEED:
1061 case XLOG_STATE_COVER_NEED2:
1062 if (xfs_ail_min_lsn(log->l_ailp))
1063 break;
1064 if (!xlog_iclogs_empty(log))
1065 break;
1066
1067 needed = 1;
1068 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1069 log->l_covered_state = XLOG_STATE_COVER_DONE;
1070 else
1071 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1072 break;
1073 default:
1074 needed = 1;
1075 break;
1076 }
1077 spin_unlock(&log->l_icloglock);
1078 return needed;
1079}
1080
1081/*
1082 * We may be holding the log iclog lock upon entering this routine.
1083 */
1084xfs_lsn_t
1085xlog_assign_tail_lsn_locked(
1086 struct xfs_mount *mp)
1087{
1088 struct xlog *log = mp->m_log;
1089 struct xfs_log_item *lip;
1090 xfs_lsn_t tail_lsn;
1091
1092 assert_spin_locked(&mp->m_ail->xa_lock);
1093
1094 /*
1095 * To make sure we always have a valid LSN for the log tail we keep
1096 * track of the last LSN which was committed in log->l_last_sync_lsn,
1097 * and use that when the AIL was empty.
1098 */
1099 lip = xfs_ail_min(mp->m_ail);
1100 if (lip)
1101 tail_lsn = lip->li_lsn;
1102 else
1103 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1104 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1105 atomic64_set(&log->l_tail_lsn, tail_lsn);
1106 return tail_lsn;
1107}
1108
1109xfs_lsn_t
1110xlog_assign_tail_lsn(
1111 struct xfs_mount *mp)
1112{
1113 xfs_lsn_t tail_lsn;
1114
1115 spin_lock(&mp->m_ail->xa_lock);
1116 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1117 spin_unlock(&mp->m_ail->xa_lock);
1118
1119 return tail_lsn;
1120}
1121
1122/*
1123 * Return the space in the log between the tail and the head. The head
1124 * is passed in the cycle/bytes formal parms. In the special case where
1125 * the reserve head has wrapped passed the tail, this calculation is no
1126 * longer valid. In this case, just return 0 which means there is no space
1127 * in the log. This works for all places where this function is called
1128 * with the reserve head. Of course, if the write head were to ever
1129 * wrap the tail, we should blow up. Rather than catch this case here,
1130 * we depend on other ASSERTions in other parts of the code. XXXmiken
1131 *
1132 * This code also handles the case where the reservation head is behind
1133 * the tail. The details of this case are described below, but the end
1134 * result is that we return the size of the log as the amount of space left.
1135 */
1136STATIC int
1137xlog_space_left(
1138 struct xlog *log,
1139 atomic64_t *head)
1140{
1141 int free_bytes;
1142 int tail_bytes;
1143 int tail_cycle;
1144 int head_cycle;
1145 int head_bytes;
1146
1147 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1148 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1149 tail_bytes = BBTOB(tail_bytes);
1150 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1151 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1152 else if (tail_cycle + 1 < head_cycle)
1153 return 0;
1154 else if (tail_cycle < head_cycle) {
1155 ASSERT(tail_cycle == (head_cycle - 1));
1156 free_bytes = tail_bytes - head_bytes;
1157 } else {
1158 /*
1159 * The reservation head is behind the tail.
1160 * In this case we just want to return the size of the
1161 * log as the amount of space left.
1162 */
1163 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1164 xfs_alert(log->l_mp,
1165 " tail_cycle = %d, tail_bytes = %d",
1166 tail_cycle, tail_bytes);
1167 xfs_alert(log->l_mp,
1168 " GH cycle = %d, GH bytes = %d",
1169 head_cycle, head_bytes);
1170 ASSERT(0);
1171 free_bytes = log->l_logsize;
1172 }
1173 return free_bytes;
1174}
1175
1176
1177/*
1178 * Log function which is called when an io completes.
1179 *
1180 * The log manager needs its own routine, in order to control what
1181 * happens with the buffer after the write completes.
1182 */
1183void
1184xlog_iodone(xfs_buf_t *bp)
1185{
1186 struct xlog_in_core *iclog = bp->b_fspriv;
1187 struct xlog *l = iclog->ic_log;
1188 int aborted = 0;
1189
1190 /*
1191 * Race to shutdown the filesystem if we see an error or the iclog is in
1192 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1193 * CRC errors into log recovery.
1194 */
1195 if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
1196 XFS_RANDOM_IODONE_IOERR) ||
1197 iclog->ic_state & XLOG_STATE_IOABORT) {
1198 if (iclog->ic_state & XLOG_STATE_IOABORT)
1199 iclog->ic_state &= ~XLOG_STATE_IOABORT;
1200
1201 xfs_buf_ioerror_alert(bp, __func__);
1202 xfs_buf_stale(bp);
1203 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1204 /*
1205 * This flag will be propagated to the trans-committed
1206 * callback routines to let them know that the log-commit
1207 * didn't succeed.
1208 */
1209 aborted = XFS_LI_ABORTED;
1210 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1211 aborted = XFS_LI_ABORTED;
1212 }
1213
1214 /* log I/O is always issued ASYNC */
1215 ASSERT(bp->b_flags & XBF_ASYNC);
1216 xlog_state_done_syncing(iclog, aborted);
1217
1218 /*
1219 * drop the buffer lock now that we are done. Nothing references
1220 * the buffer after this, so an unmount waiting on this lock can now
1221 * tear it down safely. As such, it is unsafe to reference the buffer
1222 * (bp) after the unlock as we could race with it being freed.
1223 */
1224 xfs_buf_unlock(bp);
1225}
1226
1227/*
1228 * Return size of each in-core log record buffer.
1229 *
1230 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1231 *
1232 * If the filesystem blocksize is too large, we may need to choose a
1233 * larger size since the directory code currently logs entire blocks.
1234 */
1235
1236STATIC void
1237xlog_get_iclog_buffer_size(
1238 struct xfs_mount *mp,
1239 struct xlog *log)
1240{
1241 int size;
1242 int xhdrs;
1243
1244 if (mp->m_logbufs <= 0)
1245 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1246 else
1247 log->l_iclog_bufs = mp->m_logbufs;
1248
1249 /*
1250 * Buffer size passed in from mount system call.
1251 */
1252 if (mp->m_logbsize > 0) {
1253 size = log->l_iclog_size = mp->m_logbsize;
1254 log->l_iclog_size_log = 0;
1255 while (size != 1) {
1256 log->l_iclog_size_log++;
1257 size >>= 1;
1258 }
1259
1260 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1261 /* # headers = size / 32k
1262 * one header holds cycles from 32k of data
1263 */
1264
1265 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1266 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1267 xhdrs++;
1268 log->l_iclog_hsize = xhdrs << BBSHIFT;
1269 log->l_iclog_heads = xhdrs;
1270 } else {
1271 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1272 log->l_iclog_hsize = BBSIZE;
1273 log->l_iclog_heads = 1;
1274 }
1275 goto done;
1276 }
1277
1278 /* All machines use 32kB buffers by default. */
1279 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1280 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1281
1282 /* the default log size is 16k or 32k which is one header sector */
1283 log->l_iclog_hsize = BBSIZE;
1284 log->l_iclog_heads = 1;
1285
1286done:
1287 /* are we being asked to make the sizes selected above visible? */
1288 if (mp->m_logbufs == 0)
1289 mp->m_logbufs = log->l_iclog_bufs;
1290 if (mp->m_logbsize == 0)
1291 mp->m_logbsize = log->l_iclog_size;
1292} /* xlog_get_iclog_buffer_size */
1293
1294
1295void
1296xfs_log_work_queue(
1297 struct xfs_mount *mp)
1298{
1299 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1300 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1301}
1302
1303/*
1304 * Every sync period we need to unpin all items in the AIL and push them to
1305 * disk. If there is nothing dirty, then we might need to cover the log to
1306 * indicate that the filesystem is idle.
1307 */
1308void
1309xfs_log_worker(
1310 struct work_struct *work)
1311{
1312 struct xlog *log = container_of(to_delayed_work(work),
1313 struct xlog, l_work);
1314 struct xfs_mount *mp = log->l_mp;
1315
1316 /* dgc: errors ignored - not fatal and nowhere to report them */
1317 if (xfs_log_need_covered(mp)) {
1318 /*
1319 * Dump a transaction into the log that contains no real change.
1320 * This is needed to stamp the current tail LSN into the log
1321 * during the covering operation.
1322 *
1323 * We cannot use an inode here for this - that will push dirty
1324 * state back up into the VFS and then periodic inode flushing
1325 * will prevent log covering from making progress. Hence we
1326 * synchronously log the superblock instead to ensure the
1327 * superblock is immediately unpinned and can be written back.
1328 */
1329 xfs_sync_sb(mp, true);
1330 } else
1331 xfs_log_force(mp, 0);
1332
1333 /* start pushing all the metadata that is currently dirty */
1334 xfs_ail_push_all(mp->m_ail);
1335
1336 /* queue us up again */
1337 xfs_log_work_queue(mp);
1338}
1339
1340/*
1341 * This routine initializes some of the log structure for a given mount point.
1342 * Its primary purpose is to fill in enough, so recovery can occur. However,
1343 * some other stuff may be filled in too.
1344 */
1345STATIC struct xlog *
1346xlog_alloc_log(
1347 struct xfs_mount *mp,
1348 struct xfs_buftarg *log_target,
1349 xfs_daddr_t blk_offset,
1350 int num_bblks)
1351{
1352 struct xlog *log;
1353 xlog_rec_header_t *head;
1354 xlog_in_core_t **iclogp;
1355 xlog_in_core_t *iclog, *prev_iclog=NULL;
1356 xfs_buf_t *bp;
1357 int i;
1358 int error = -ENOMEM;
1359 uint log2_size = 0;
1360
1361 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1362 if (!log) {
1363 xfs_warn(mp, "Log allocation failed: No memory!");
1364 goto out;
1365 }
1366
1367 log->l_mp = mp;
1368 log->l_targ = log_target;
1369 log->l_logsize = BBTOB(num_bblks);
1370 log->l_logBBstart = blk_offset;
1371 log->l_logBBsize = num_bblks;
1372 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1373 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1374 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1375
1376 log->l_prev_block = -1;
1377 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1378 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1379 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1380 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1381
1382 xlog_grant_head_init(&log->l_reserve_head);
1383 xlog_grant_head_init(&log->l_write_head);
1384
1385 error = -EFSCORRUPTED;
1386 if (xfs_sb_version_hassector(&mp->m_sb)) {
1387 log2_size = mp->m_sb.sb_logsectlog;
1388 if (log2_size < BBSHIFT) {
1389 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1390 log2_size, BBSHIFT);
1391 goto out_free_log;
1392 }
1393
1394 log2_size -= BBSHIFT;
1395 if (log2_size > mp->m_sectbb_log) {
1396 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1397 log2_size, mp->m_sectbb_log);
1398 goto out_free_log;
1399 }
1400
1401 /* for larger sector sizes, must have v2 or external log */
1402 if (log2_size && log->l_logBBstart > 0 &&
1403 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1404 xfs_warn(mp,
1405 "log sector size (0x%x) invalid for configuration.",
1406 log2_size);
1407 goto out_free_log;
1408 }
1409 }
1410 log->l_sectBBsize = 1 << log2_size;
1411
1412 xlog_get_iclog_buffer_size(mp, log);
1413
1414 /*
1415 * Use a NULL block for the extra log buffer used during splits so that
1416 * it will trigger errors if we ever try to do IO on it without first
1417 * having set it up properly.
1418 */
1419 error = -ENOMEM;
1420 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1421 BTOBB(log->l_iclog_size), 0);
1422 if (!bp)
1423 goto out_free_log;
1424
1425 /*
1426 * The iclogbuf buffer locks are held over IO but we are not going to do
1427 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1428 * when appropriately.
1429 */
1430 ASSERT(xfs_buf_islocked(bp));
1431 xfs_buf_unlock(bp);
1432
1433 /* use high priority wq for log I/O completion */
1434 bp->b_ioend_wq = mp->m_log_workqueue;
1435 bp->b_iodone = xlog_iodone;
1436 log->l_xbuf = bp;
1437
1438 spin_lock_init(&log->l_icloglock);
1439 init_waitqueue_head(&log->l_flush_wait);
1440
1441 iclogp = &log->l_iclog;
1442 /*
1443 * The amount of memory to allocate for the iclog structure is
1444 * rather funky due to the way the structure is defined. It is
1445 * done this way so that we can use different sizes for machines
1446 * with different amounts of memory. See the definition of
1447 * xlog_in_core_t in xfs_log_priv.h for details.
1448 */
1449 ASSERT(log->l_iclog_size >= 4096);
1450 for (i=0; i < log->l_iclog_bufs; i++) {
1451 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1452 if (!*iclogp)
1453 goto out_free_iclog;
1454
1455 iclog = *iclogp;
1456 iclog->ic_prev = prev_iclog;
1457 prev_iclog = iclog;
1458
1459 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1460 BTOBB(log->l_iclog_size), 0);
1461 if (!bp)
1462 goto out_free_iclog;
1463
1464 ASSERT(xfs_buf_islocked(bp));
1465 xfs_buf_unlock(bp);
1466
1467 /* use high priority wq for log I/O completion */
1468 bp->b_ioend_wq = mp->m_log_workqueue;
1469 bp->b_iodone = xlog_iodone;
1470 iclog->ic_bp = bp;
1471 iclog->ic_data = bp->b_addr;
1472#ifdef DEBUG
1473 log->l_iclog_bak[i] = &iclog->ic_header;
1474#endif
1475 head = &iclog->ic_header;
1476 memset(head, 0, sizeof(xlog_rec_header_t));
1477 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1478 head->h_version = cpu_to_be32(
1479 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1480 head->h_size = cpu_to_be32(log->l_iclog_size);
1481 /* new fields */
1482 head->h_fmt = cpu_to_be32(XLOG_FMT);
1483 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1484
1485 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1486 iclog->ic_state = XLOG_STATE_ACTIVE;
1487 iclog->ic_log = log;
1488 atomic_set(&iclog->ic_refcnt, 0);
1489 spin_lock_init(&iclog->ic_callback_lock);
1490 iclog->ic_callback_tail = &(iclog->ic_callback);
1491 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1492
1493 init_waitqueue_head(&iclog->ic_force_wait);
1494 init_waitqueue_head(&iclog->ic_write_wait);
1495
1496 iclogp = &iclog->ic_next;
1497 }
1498 *iclogp = log->l_iclog; /* complete ring */
1499 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1500
1501 error = xlog_cil_init(log);
1502 if (error)
1503 goto out_free_iclog;
1504 return log;
1505
1506out_free_iclog:
1507 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1508 prev_iclog = iclog->ic_next;
1509 if (iclog->ic_bp)
1510 xfs_buf_free(iclog->ic_bp);
1511 kmem_free(iclog);
1512 }
1513 spinlock_destroy(&log->l_icloglock);
1514 xfs_buf_free(log->l_xbuf);
1515out_free_log:
1516 kmem_free(log);
1517out:
1518 return ERR_PTR(error);
1519} /* xlog_alloc_log */
1520
1521
1522/*
1523 * Write out the commit record of a transaction associated with the given
1524 * ticket. Return the lsn of the commit record.
1525 */
1526STATIC int
1527xlog_commit_record(
1528 struct xlog *log,
1529 struct xlog_ticket *ticket,
1530 struct xlog_in_core **iclog,
1531 xfs_lsn_t *commitlsnp)
1532{
1533 struct xfs_mount *mp = log->l_mp;
1534 int error;
1535 struct xfs_log_iovec reg = {
1536 .i_addr = NULL,
1537 .i_len = 0,
1538 .i_type = XLOG_REG_TYPE_COMMIT,
1539 };
1540 struct xfs_log_vec vec = {
1541 .lv_niovecs = 1,
1542 .lv_iovecp = ®,
1543 };
1544
1545 ASSERT_ALWAYS(iclog);
1546 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1547 XLOG_COMMIT_TRANS);
1548 if (error)
1549 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1550 return error;
1551}
1552
1553/*
1554 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1555 * log space. This code pushes on the lsn which would supposedly free up
1556 * the 25% which we want to leave free. We may need to adopt a policy which
1557 * pushes on an lsn which is further along in the log once we reach the high
1558 * water mark. In this manner, we would be creating a low water mark.
1559 */
1560STATIC void
1561xlog_grant_push_ail(
1562 struct xlog *log,
1563 int need_bytes)
1564{
1565 xfs_lsn_t threshold_lsn = 0;
1566 xfs_lsn_t last_sync_lsn;
1567 int free_blocks;
1568 int free_bytes;
1569 int threshold_block;
1570 int threshold_cycle;
1571 int free_threshold;
1572
1573 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1574
1575 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1576 free_blocks = BTOBBT(free_bytes);
1577
1578 /*
1579 * Set the threshold for the minimum number of free blocks in the
1580 * log to the maximum of what the caller needs, one quarter of the
1581 * log, and 256 blocks.
1582 */
1583 free_threshold = BTOBB(need_bytes);
1584 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1585 free_threshold = MAX(free_threshold, 256);
1586 if (free_blocks >= free_threshold)
1587 return;
1588
1589 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1590 &threshold_block);
1591 threshold_block += free_threshold;
1592 if (threshold_block >= log->l_logBBsize) {
1593 threshold_block -= log->l_logBBsize;
1594 threshold_cycle += 1;
1595 }
1596 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1597 threshold_block);
1598 /*
1599 * Don't pass in an lsn greater than the lsn of the last
1600 * log record known to be on disk. Use a snapshot of the last sync lsn
1601 * so that it doesn't change between the compare and the set.
1602 */
1603 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1604 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1605 threshold_lsn = last_sync_lsn;
1606
1607 /*
1608 * Get the transaction layer to kick the dirty buffers out to
1609 * disk asynchronously. No point in trying to do this if
1610 * the filesystem is shutting down.
1611 */
1612 if (!XLOG_FORCED_SHUTDOWN(log))
1613 xfs_ail_push(log->l_ailp, threshold_lsn);
1614}
1615
1616/*
1617 * Stamp cycle number in every block
1618 */
1619STATIC void
1620xlog_pack_data(
1621 struct xlog *log,
1622 struct xlog_in_core *iclog,
1623 int roundoff)
1624{
1625 int i, j, k;
1626 int size = iclog->ic_offset + roundoff;
1627 __be32 cycle_lsn;
1628 char *dp;
1629
1630 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1631
1632 dp = iclog->ic_datap;
1633 for (i = 0; i < BTOBB(size); i++) {
1634 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1635 break;
1636 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1637 *(__be32 *)dp = cycle_lsn;
1638 dp += BBSIZE;
1639 }
1640
1641 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1642 xlog_in_core_2_t *xhdr = iclog->ic_data;
1643
1644 for ( ; i < BTOBB(size); i++) {
1645 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1646 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1647 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1648 *(__be32 *)dp = cycle_lsn;
1649 dp += BBSIZE;
1650 }
1651
1652 for (i = 1; i < log->l_iclog_heads; i++)
1653 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1654 }
1655}
1656
1657/*
1658 * Calculate the checksum for a log buffer.
1659 *
1660 * This is a little more complicated than it should be because the various
1661 * headers and the actual data are non-contiguous.
1662 */
1663__le32
1664xlog_cksum(
1665 struct xlog *log,
1666 struct xlog_rec_header *rhead,
1667 char *dp,
1668 int size)
1669{
1670 __uint32_t crc;
1671
1672 /* first generate the crc for the record header ... */
1673 crc = xfs_start_cksum((char *)rhead,
1674 sizeof(struct xlog_rec_header),
1675 offsetof(struct xlog_rec_header, h_crc));
1676
1677 /* ... then for additional cycle data for v2 logs ... */
1678 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1679 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1680 int i;
1681 int xheads;
1682
1683 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1684 if (size % XLOG_HEADER_CYCLE_SIZE)
1685 xheads++;
1686
1687 for (i = 1; i < xheads; i++) {
1688 crc = crc32c(crc, &xhdr[i].hic_xheader,
1689 sizeof(struct xlog_rec_ext_header));
1690 }
1691 }
1692
1693 /* ... and finally for the payload */
1694 crc = crc32c(crc, dp, size);
1695
1696 return xfs_end_cksum(crc);
1697}
1698
1699/*
1700 * The bdstrat callback function for log bufs. This gives us a central
1701 * place to trap bufs in case we get hit by a log I/O error and need to
1702 * shutdown. Actually, in practice, even when we didn't get a log error,
1703 * we transition the iclogs to IOERROR state *after* flushing all existing
1704 * iclogs to disk. This is because we don't want anymore new transactions to be
1705 * started or completed afterwards.
1706 *
1707 * We lock the iclogbufs here so that we can serialise against IO completion
1708 * during unmount. We might be processing a shutdown triggered during unmount,
1709 * and that can occur asynchronously to the unmount thread, and hence we need to
1710 * ensure that completes before tearing down the iclogbufs. Hence we need to
1711 * hold the buffer lock across the log IO to acheive that.
1712 */
1713STATIC int
1714xlog_bdstrat(
1715 struct xfs_buf *bp)
1716{
1717 struct xlog_in_core *iclog = bp->b_fspriv;
1718
1719 xfs_buf_lock(bp);
1720 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1721 xfs_buf_ioerror(bp, -EIO);
1722 xfs_buf_stale(bp);
1723 xfs_buf_ioend(bp);
1724 /*
1725 * It would seem logical to return EIO here, but we rely on
1726 * the log state machine to propagate I/O errors instead of
1727 * doing it here. Similarly, IO completion will unlock the
1728 * buffer, so we don't do it here.
1729 */
1730 return 0;
1731 }
1732
1733 xfs_buf_submit(bp);
1734 return 0;
1735}
1736
1737/*
1738 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1739 * fashion. Previously, we should have moved the current iclog
1740 * ptr in the log to point to the next available iclog. This allows further
1741 * write to continue while this code syncs out an iclog ready to go.
1742 * Before an in-core log can be written out, the data section must be scanned
1743 * to save away the 1st word of each BBSIZE block into the header. We replace
1744 * it with the current cycle count. Each BBSIZE block is tagged with the
1745 * cycle count because there in an implicit assumption that drives will
1746 * guarantee that entire 512 byte blocks get written at once. In other words,
1747 * we can't have part of a 512 byte block written and part not written. By
1748 * tagging each block, we will know which blocks are valid when recovering
1749 * after an unclean shutdown.
1750 *
1751 * This routine is single threaded on the iclog. No other thread can be in
1752 * this routine with the same iclog. Changing contents of iclog can there-
1753 * fore be done without grabbing the state machine lock. Updating the global
1754 * log will require grabbing the lock though.
1755 *
1756 * The entire log manager uses a logical block numbering scheme. Only
1757 * log_sync (and then only bwrite()) know about the fact that the log may
1758 * not start with block zero on a given device. The log block start offset
1759 * is added immediately before calling bwrite().
1760 */
1761
1762STATIC int
1763xlog_sync(
1764 struct xlog *log,
1765 struct xlog_in_core *iclog)
1766{
1767 xfs_buf_t *bp;
1768 int i;
1769 uint count; /* byte count of bwrite */
1770 uint count_init; /* initial count before roundup */
1771 int roundoff; /* roundoff to BB or stripe */
1772 int split = 0; /* split write into two regions */
1773 int error;
1774 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1775 int size;
1776
1777 XFS_STATS_INC(log->l_mp, xs_log_writes);
1778 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1779
1780 /* Add for LR header */
1781 count_init = log->l_iclog_hsize + iclog->ic_offset;
1782
1783 /* Round out the log write size */
1784 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1785 /* we have a v2 stripe unit to use */
1786 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1787 } else {
1788 count = BBTOB(BTOBB(count_init));
1789 }
1790 roundoff = count - count_init;
1791 ASSERT(roundoff >= 0);
1792 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1793 roundoff < log->l_mp->m_sb.sb_logsunit)
1794 ||
1795 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1796 roundoff < BBTOB(1)));
1797
1798 /* move grant heads by roundoff in sync */
1799 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1800 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1801
1802 /* put cycle number in every block */
1803 xlog_pack_data(log, iclog, roundoff);
1804
1805 /* real byte length */
1806 size = iclog->ic_offset;
1807 if (v2)
1808 size += roundoff;
1809 iclog->ic_header.h_len = cpu_to_be32(size);
1810
1811 bp = iclog->ic_bp;
1812 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1813
1814 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1815
1816 /* Do we need to split this write into 2 parts? */
1817 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1818 char *dptr;
1819
1820 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1821 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1822 iclog->ic_bwritecnt = 2;
1823
1824 /*
1825 * Bump the cycle numbers at the start of each block in the
1826 * part of the iclog that ends up in the buffer that gets
1827 * written to the start of the log.
1828 *
1829 * Watch out for the header magic number case, though.
1830 */
1831 dptr = (char *)&iclog->ic_header + count;
1832 for (i = 0; i < split; i += BBSIZE) {
1833 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1834 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1835 cycle++;
1836 *(__be32 *)dptr = cpu_to_be32(cycle);
1837
1838 dptr += BBSIZE;
1839 }
1840 } else {
1841 iclog->ic_bwritecnt = 1;
1842 }
1843
1844 /* calculcate the checksum */
1845 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1846 iclog->ic_datap, size);
1847#ifdef DEBUG
1848 /*
1849 * Intentionally corrupt the log record CRC based on the error injection
1850 * frequency, if defined. This facilitates testing log recovery in the
1851 * event of torn writes. Hence, set the IOABORT state to abort the log
1852 * write on I/O completion and shutdown the fs. The subsequent mount
1853 * detects the bad CRC and attempts to recover.
1854 */
1855 if (log->l_badcrc_factor &&
1856 (prandom_u32() % log->l_badcrc_factor == 0)) {
1857 iclog->ic_header.h_crc &= 0xAAAAAAAA;
1858 iclog->ic_state |= XLOG_STATE_IOABORT;
1859 xfs_warn(log->l_mp,
1860 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1861 be64_to_cpu(iclog->ic_header.h_lsn));
1862 }
1863#endif
1864
1865 bp->b_io_length = BTOBB(count);
1866 bp->b_fspriv = iclog;
1867 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1868 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1869
1870 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1871 bp->b_flags |= XBF_FUA;
1872
1873 /*
1874 * Flush the data device before flushing the log to make
1875 * sure all meta data written back from the AIL actually made
1876 * it to disk before stamping the new log tail LSN into the
1877 * log buffer. For an external log we need to issue the
1878 * flush explicitly, and unfortunately synchronously here;
1879 * for an internal log we can simply use the block layer
1880 * state machine for preflushes.
1881 */
1882 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1883 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1884 else
1885 bp->b_flags |= XBF_FLUSH;
1886 }
1887
1888 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1889 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1890
1891 xlog_verify_iclog(log, iclog, count, true);
1892
1893 /* account for log which doesn't start at block #0 */
1894 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1895
1896 /*
1897 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1898 * is shutting down.
1899 */
1900 error = xlog_bdstrat(bp);
1901 if (error) {
1902 xfs_buf_ioerror_alert(bp, "xlog_sync");
1903 return error;
1904 }
1905 if (split) {
1906 bp = iclog->ic_log->l_xbuf;
1907 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1908 xfs_buf_associate_memory(bp,
1909 (char *)&iclog->ic_header + count, split);
1910 bp->b_fspriv = iclog;
1911 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1912 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1913 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1914 bp->b_flags |= XBF_FUA;
1915
1916 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1917 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1918
1919 /* account for internal log which doesn't start at block #0 */
1920 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1921 error = xlog_bdstrat(bp);
1922 if (error) {
1923 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1924 return error;
1925 }
1926 }
1927 return 0;
1928} /* xlog_sync */
1929
1930/*
1931 * Deallocate a log structure
1932 */
1933STATIC void
1934xlog_dealloc_log(
1935 struct xlog *log)
1936{
1937 xlog_in_core_t *iclog, *next_iclog;
1938 int i;
1939
1940 xlog_cil_destroy(log);
1941
1942 /*
1943 * Cycle all the iclogbuf locks to make sure all log IO completion
1944 * is done before we tear down these buffers.
1945 */
1946 iclog = log->l_iclog;
1947 for (i = 0; i < log->l_iclog_bufs; i++) {
1948 xfs_buf_lock(iclog->ic_bp);
1949 xfs_buf_unlock(iclog->ic_bp);
1950 iclog = iclog->ic_next;
1951 }
1952
1953 /*
1954 * Always need to ensure that the extra buffer does not point to memory
1955 * owned by another log buffer before we free it. Also, cycle the lock
1956 * first to ensure we've completed IO on it.
1957 */
1958 xfs_buf_lock(log->l_xbuf);
1959 xfs_buf_unlock(log->l_xbuf);
1960 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1961 xfs_buf_free(log->l_xbuf);
1962
1963 iclog = log->l_iclog;
1964 for (i = 0; i < log->l_iclog_bufs; i++) {
1965 xfs_buf_free(iclog->ic_bp);
1966 next_iclog = iclog->ic_next;
1967 kmem_free(iclog);
1968 iclog = next_iclog;
1969 }
1970 spinlock_destroy(&log->l_icloglock);
1971
1972 log->l_mp->m_log = NULL;
1973 kmem_free(log);
1974} /* xlog_dealloc_log */
1975
1976/*
1977 * Update counters atomically now that memcpy is done.
1978 */
1979/* ARGSUSED */
1980static inline void
1981xlog_state_finish_copy(
1982 struct xlog *log,
1983 struct xlog_in_core *iclog,
1984 int record_cnt,
1985 int copy_bytes)
1986{
1987 spin_lock(&log->l_icloglock);
1988
1989 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1990 iclog->ic_offset += copy_bytes;
1991
1992 spin_unlock(&log->l_icloglock);
1993} /* xlog_state_finish_copy */
1994
1995
1996
1997
1998/*
1999 * print out info relating to regions written which consume
2000 * the reservation
2001 */
2002void
2003xlog_print_tic_res(
2004 struct xfs_mount *mp,
2005 struct xlog_ticket *ticket)
2006{
2007 uint i;
2008 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2009
2010 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2011#define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2012 static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2013 REG_TYPE_STR(BFORMAT, "bformat"),
2014 REG_TYPE_STR(BCHUNK, "bchunk"),
2015 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2016 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2017 REG_TYPE_STR(IFORMAT, "iformat"),
2018 REG_TYPE_STR(ICORE, "icore"),
2019 REG_TYPE_STR(IEXT, "iext"),
2020 REG_TYPE_STR(IBROOT, "ibroot"),
2021 REG_TYPE_STR(ILOCAL, "ilocal"),
2022 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2023 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2024 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2025 REG_TYPE_STR(QFORMAT, "qformat"),
2026 REG_TYPE_STR(DQUOT, "dquot"),
2027 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2028 REG_TYPE_STR(LRHEADER, "LR header"),
2029 REG_TYPE_STR(UNMOUNT, "unmount"),
2030 REG_TYPE_STR(COMMIT, "commit"),
2031 REG_TYPE_STR(TRANSHDR, "trans header"),
2032 REG_TYPE_STR(ICREATE, "inode create")
2033 };
2034#undef REG_TYPE_STR
2035#define TRANS_TYPE_STR(type) [XFS_TRANS_##type] = #type
2036 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2037 TRANS_TYPE_STR(SETATTR_NOT_SIZE),
2038 TRANS_TYPE_STR(SETATTR_SIZE),
2039 TRANS_TYPE_STR(INACTIVE),
2040 TRANS_TYPE_STR(CREATE),
2041 TRANS_TYPE_STR(CREATE_TRUNC),
2042 TRANS_TYPE_STR(TRUNCATE_FILE),
2043 TRANS_TYPE_STR(REMOVE),
2044 TRANS_TYPE_STR(LINK),
2045 TRANS_TYPE_STR(RENAME),
2046 TRANS_TYPE_STR(MKDIR),
2047 TRANS_TYPE_STR(RMDIR),
2048 TRANS_TYPE_STR(SYMLINK),
2049 TRANS_TYPE_STR(SET_DMATTRS),
2050 TRANS_TYPE_STR(GROWFS),
2051 TRANS_TYPE_STR(STRAT_WRITE),
2052 TRANS_TYPE_STR(DIOSTRAT),
2053 TRANS_TYPE_STR(WRITEID),
2054 TRANS_TYPE_STR(ADDAFORK),
2055 TRANS_TYPE_STR(ATTRINVAL),
2056 TRANS_TYPE_STR(ATRUNCATE),
2057 TRANS_TYPE_STR(ATTR_SET),
2058 TRANS_TYPE_STR(ATTR_RM),
2059 TRANS_TYPE_STR(ATTR_FLAG),
2060 TRANS_TYPE_STR(CLEAR_AGI_BUCKET),
2061 TRANS_TYPE_STR(SB_CHANGE),
2062 TRANS_TYPE_STR(DUMMY1),
2063 TRANS_TYPE_STR(DUMMY2),
2064 TRANS_TYPE_STR(QM_QUOTAOFF),
2065 TRANS_TYPE_STR(QM_DQALLOC),
2066 TRANS_TYPE_STR(QM_SETQLIM),
2067 TRANS_TYPE_STR(QM_DQCLUSTER),
2068 TRANS_TYPE_STR(QM_QINOCREATE),
2069 TRANS_TYPE_STR(QM_QUOTAOFF_END),
2070 TRANS_TYPE_STR(FSYNC_TS),
2071 TRANS_TYPE_STR(GROWFSRT_ALLOC),
2072 TRANS_TYPE_STR(GROWFSRT_ZERO),
2073 TRANS_TYPE_STR(GROWFSRT_FREE),
2074 TRANS_TYPE_STR(SWAPEXT),
2075 TRANS_TYPE_STR(CHECKPOINT),
2076 TRANS_TYPE_STR(ICREATE),
2077 TRANS_TYPE_STR(CREATE_TMPFILE)
2078 };
2079#undef TRANS_TYPE_STR
2080
2081 xfs_warn(mp, "xlog_write: reservation summary:");
2082 xfs_warn(mp, " trans type = %s (%u)",
2083 ((ticket->t_trans_type <= 0 ||
2084 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2085 "bad-trans-type" : trans_type_str[ticket->t_trans_type]),
2086 ticket->t_trans_type);
2087 xfs_warn(mp, " unit res = %d bytes",
2088 ticket->t_unit_res);
2089 xfs_warn(mp, " current res = %d bytes",
2090 ticket->t_curr_res);
2091 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2092 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2093 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2094 ticket->t_res_num_ophdrs, ophdr_spc);
2095 xfs_warn(mp, " ophdr + reg = %u bytes",
2096 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2097 xfs_warn(mp, " num regions = %u",
2098 ticket->t_res_num);
2099
2100 for (i = 0; i < ticket->t_res_num; i++) {
2101 uint r_type = ticket->t_res_arr[i].r_type;
2102 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2103 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2104 "bad-rtype" : res_type_str[r_type]),
2105 ticket->t_res_arr[i].r_len);
2106 }
2107
2108 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2109 "xlog_write: reservation ran out. Need to up reservation");
2110 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2111}
2112
2113/*
2114 * Calculate the potential space needed by the log vector. Each region gets
2115 * its own xlog_op_header_t and may need to be double word aligned.
2116 */
2117static int
2118xlog_write_calc_vec_length(
2119 struct xlog_ticket *ticket,
2120 struct xfs_log_vec *log_vector)
2121{
2122 struct xfs_log_vec *lv;
2123 int headers = 0;
2124 int len = 0;
2125 int i;
2126
2127 /* acct for start rec of xact */
2128 if (ticket->t_flags & XLOG_TIC_INITED)
2129 headers++;
2130
2131 for (lv = log_vector; lv; lv = lv->lv_next) {
2132 /* we don't write ordered log vectors */
2133 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2134 continue;
2135
2136 headers += lv->lv_niovecs;
2137
2138 for (i = 0; i < lv->lv_niovecs; i++) {
2139 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2140
2141 len += vecp->i_len;
2142 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2143 }
2144 }
2145
2146 ticket->t_res_num_ophdrs += headers;
2147 len += headers * sizeof(struct xlog_op_header);
2148
2149 return len;
2150}
2151
2152/*
2153 * If first write for transaction, insert start record We can't be trying to
2154 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2155 */
2156static int
2157xlog_write_start_rec(
2158 struct xlog_op_header *ophdr,
2159 struct xlog_ticket *ticket)
2160{
2161 if (!(ticket->t_flags & XLOG_TIC_INITED))
2162 return 0;
2163
2164 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2165 ophdr->oh_clientid = ticket->t_clientid;
2166 ophdr->oh_len = 0;
2167 ophdr->oh_flags = XLOG_START_TRANS;
2168 ophdr->oh_res2 = 0;
2169
2170 ticket->t_flags &= ~XLOG_TIC_INITED;
2171
2172 return sizeof(struct xlog_op_header);
2173}
2174
2175static xlog_op_header_t *
2176xlog_write_setup_ophdr(
2177 struct xlog *log,
2178 struct xlog_op_header *ophdr,
2179 struct xlog_ticket *ticket,
2180 uint flags)
2181{
2182 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2183 ophdr->oh_clientid = ticket->t_clientid;
2184 ophdr->oh_res2 = 0;
2185
2186 /* are we copying a commit or unmount record? */
2187 ophdr->oh_flags = flags;
2188
2189 /*
2190 * We've seen logs corrupted with bad transaction client ids. This
2191 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2192 * and shut down the filesystem.
2193 */
2194 switch (ophdr->oh_clientid) {
2195 case XFS_TRANSACTION:
2196 case XFS_VOLUME:
2197 case XFS_LOG:
2198 break;
2199 default:
2200 xfs_warn(log->l_mp,
2201 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2202 ophdr->oh_clientid, ticket);
2203 return NULL;
2204 }
2205
2206 return ophdr;
2207}
2208
2209/*
2210 * Set up the parameters of the region copy into the log. This has
2211 * to handle region write split across multiple log buffers - this
2212 * state is kept external to this function so that this code can
2213 * be written in an obvious, self documenting manner.
2214 */
2215static int
2216xlog_write_setup_copy(
2217 struct xlog_ticket *ticket,
2218 struct xlog_op_header *ophdr,
2219 int space_available,
2220 int space_required,
2221 int *copy_off,
2222 int *copy_len,
2223 int *last_was_partial_copy,
2224 int *bytes_consumed)
2225{
2226 int still_to_copy;
2227
2228 still_to_copy = space_required - *bytes_consumed;
2229 *copy_off = *bytes_consumed;
2230
2231 if (still_to_copy <= space_available) {
2232 /* write of region completes here */
2233 *copy_len = still_to_copy;
2234 ophdr->oh_len = cpu_to_be32(*copy_len);
2235 if (*last_was_partial_copy)
2236 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2237 *last_was_partial_copy = 0;
2238 *bytes_consumed = 0;
2239 return 0;
2240 }
2241
2242 /* partial write of region, needs extra log op header reservation */
2243 *copy_len = space_available;
2244 ophdr->oh_len = cpu_to_be32(*copy_len);
2245 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2246 if (*last_was_partial_copy)
2247 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2248 *bytes_consumed += *copy_len;
2249 (*last_was_partial_copy)++;
2250
2251 /* account for new log op header */
2252 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2253 ticket->t_res_num_ophdrs++;
2254
2255 return sizeof(struct xlog_op_header);
2256}
2257
2258static int
2259xlog_write_copy_finish(
2260 struct xlog *log,
2261 struct xlog_in_core *iclog,
2262 uint flags,
2263 int *record_cnt,
2264 int *data_cnt,
2265 int *partial_copy,
2266 int *partial_copy_len,
2267 int log_offset,
2268 struct xlog_in_core **commit_iclog)
2269{
2270 if (*partial_copy) {
2271 /*
2272 * This iclog has already been marked WANT_SYNC by
2273 * xlog_state_get_iclog_space.
2274 */
2275 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2276 *record_cnt = 0;
2277 *data_cnt = 0;
2278 return xlog_state_release_iclog(log, iclog);
2279 }
2280
2281 *partial_copy = 0;
2282 *partial_copy_len = 0;
2283
2284 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2285 /* no more space in this iclog - push it. */
2286 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2287 *record_cnt = 0;
2288 *data_cnt = 0;
2289
2290 spin_lock(&log->l_icloglock);
2291 xlog_state_want_sync(log, iclog);
2292 spin_unlock(&log->l_icloglock);
2293
2294 if (!commit_iclog)
2295 return xlog_state_release_iclog(log, iclog);
2296 ASSERT(flags & XLOG_COMMIT_TRANS);
2297 *commit_iclog = iclog;
2298 }
2299
2300 return 0;
2301}
2302
2303/*
2304 * Write some region out to in-core log
2305 *
2306 * This will be called when writing externally provided regions or when
2307 * writing out a commit record for a given transaction.
2308 *
2309 * General algorithm:
2310 * 1. Find total length of this write. This may include adding to the
2311 * lengths passed in.
2312 * 2. Check whether we violate the tickets reservation.
2313 * 3. While writing to this iclog
2314 * A. Reserve as much space in this iclog as can get
2315 * B. If this is first write, save away start lsn
2316 * C. While writing this region:
2317 * 1. If first write of transaction, write start record
2318 * 2. Write log operation header (header per region)
2319 * 3. Find out if we can fit entire region into this iclog
2320 * 4. Potentially, verify destination memcpy ptr
2321 * 5. Memcpy (partial) region
2322 * 6. If partial copy, release iclog; otherwise, continue
2323 * copying more regions into current iclog
2324 * 4. Mark want sync bit (in simulation mode)
2325 * 5. Release iclog for potential flush to on-disk log.
2326 *
2327 * ERRORS:
2328 * 1. Panic if reservation is overrun. This should never happen since
2329 * reservation amounts are generated internal to the filesystem.
2330 * NOTES:
2331 * 1. Tickets are single threaded data structures.
2332 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2333 * syncing routine. When a single log_write region needs to span
2334 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2335 * on all log operation writes which don't contain the end of the
2336 * region. The XLOG_END_TRANS bit is used for the in-core log
2337 * operation which contains the end of the continued log_write region.
2338 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2339 * we don't really know exactly how much space will be used. As a result,
2340 * we don't update ic_offset until the end when we know exactly how many
2341 * bytes have been written out.
2342 */
2343int
2344xlog_write(
2345 struct xlog *log,
2346 struct xfs_log_vec *log_vector,
2347 struct xlog_ticket *ticket,
2348 xfs_lsn_t *start_lsn,
2349 struct xlog_in_core **commit_iclog,
2350 uint flags)
2351{
2352 struct xlog_in_core *iclog = NULL;
2353 struct xfs_log_iovec *vecp;
2354 struct xfs_log_vec *lv;
2355 int len;
2356 int index;
2357 int partial_copy = 0;
2358 int partial_copy_len = 0;
2359 int contwr = 0;
2360 int record_cnt = 0;
2361 int data_cnt = 0;
2362 int error;
2363
2364 *start_lsn = 0;
2365
2366 len = xlog_write_calc_vec_length(ticket, log_vector);
2367
2368 /*
2369 * Region headers and bytes are already accounted for.
2370 * We only need to take into account start records and
2371 * split regions in this function.
2372 */
2373 if (ticket->t_flags & XLOG_TIC_INITED)
2374 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2375
2376 /*
2377 * Commit record headers need to be accounted for. These
2378 * come in as separate writes so are easy to detect.
2379 */
2380 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2381 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2382
2383 if (ticket->t_curr_res < 0)
2384 xlog_print_tic_res(log->l_mp, ticket);
2385
2386 index = 0;
2387 lv = log_vector;
2388 vecp = lv->lv_iovecp;
2389 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2390 void *ptr;
2391 int log_offset;
2392
2393 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2394 &contwr, &log_offset);
2395 if (error)
2396 return error;
2397
2398 ASSERT(log_offset <= iclog->ic_size - 1);
2399 ptr = iclog->ic_datap + log_offset;
2400
2401 /* start_lsn is the first lsn written to. That's all we need. */
2402 if (!*start_lsn)
2403 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2404
2405 /*
2406 * This loop writes out as many regions as can fit in the amount
2407 * of space which was allocated by xlog_state_get_iclog_space().
2408 */
2409 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2410 struct xfs_log_iovec *reg;
2411 struct xlog_op_header *ophdr;
2412 int start_rec_copy;
2413 int copy_len;
2414 int copy_off;
2415 bool ordered = false;
2416
2417 /* ordered log vectors have no regions to write */
2418 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2419 ASSERT(lv->lv_niovecs == 0);
2420 ordered = true;
2421 goto next_lv;
2422 }
2423
2424 reg = &vecp[index];
2425 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2426 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2427
2428 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2429 if (start_rec_copy) {
2430 record_cnt++;
2431 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2432 start_rec_copy);
2433 }
2434
2435 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2436 if (!ophdr)
2437 return -EIO;
2438
2439 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2440 sizeof(struct xlog_op_header));
2441
2442 len += xlog_write_setup_copy(ticket, ophdr,
2443 iclog->ic_size-log_offset,
2444 reg->i_len,
2445 ©_off, ©_len,
2446 &partial_copy,
2447 &partial_copy_len);
2448 xlog_verify_dest_ptr(log, ptr);
2449
2450 /*
2451 * Copy region.
2452 *
2453 * Unmount records just log an opheader, so can have
2454 * empty payloads with no data region to copy. Hence we
2455 * only copy the payload if the vector says it has data
2456 * to copy.
2457 */
2458 ASSERT(copy_len >= 0);
2459 if (copy_len > 0) {
2460 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2461 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2462 copy_len);
2463 }
2464 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2465 record_cnt++;
2466 data_cnt += contwr ? copy_len : 0;
2467
2468 error = xlog_write_copy_finish(log, iclog, flags,
2469 &record_cnt, &data_cnt,
2470 &partial_copy,
2471 &partial_copy_len,
2472 log_offset,
2473 commit_iclog);
2474 if (error)
2475 return error;
2476
2477 /*
2478 * if we had a partial copy, we need to get more iclog
2479 * space but we don't want to increment the region
2480 * index because there is still more is this region to
2481 * write.
2482 *
2483 * If we completed writing this region, and we flushed
2484 * the iclog (indicated by resetting of the record
2485 * count), then we also need to get more log space. If
2486 * this was the last record, though, we are done and
2487 * can just return.
2488 */
2489 if (partial_copy)
2490 break;
2491
2492 if (++index == lv->lv_niovecs) {
2493next_lv:
2494 lv = lv->lv_next;
2495 index = 0;
2496 if (lv)
2497 vecp = lv->lv_iovecp;
2498 }
2499 if (record_cnt == 0 && ordered == false) {
2500 if (!lv)
2501 return 0;
2502 break;
2503 }
2504 }
2505 }
2506
2507 ASSERT(len == 0);
2508
2509 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2510 if (!commit_iclog)
2511 return xlog_state_release_iclog(log, iclog);
2512
2513 ASSERT(flags & XLOG_COMMIT_TRANS);
2514 *commit_iclog = iclog;
2515 return 0;
2516}
2517
2518
2519/*****************************************************************************
2520 *
2521 * State Machine functions
2522 *
2523 *****************************************************************************
2524 */
2525
2526/* Clean iclogs starting from the head. This ordering must be
2527 * maintained, so an iclog doesn't become ACTIVE beyond one that
2528 * is SYNCING. This is also required to maintain the notion that we use
2529 * a ordered wait queue to hold off would be writers to the log when every
2530 * iclog is trying to sync to disk.
2531 *
2532 * State Change: DIRTY -> ACTIVE
2533 */
2534STATIC void
2535xlog_state_clean_log(
2536 struct xlog *log)
2537{
2538 xlog_in_core_t *iclog;
2539 int changed = 0;
2540
2541 iclog = log->l_iclog;
2542 do {
2543 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2544 iclog->ic_state = XLOG_STATE_ACTIVE;
2545 iclog->ic_offset = 0;
2546 ASSERT(iclog->ic_callback == NULL);
2547 /*
2548 * If the number of ops in this iclog indicate it just
2549 * contains the dummy transaction, we can
2550 * change state into IDLE (the second time around).
2551 * Otherwise we should change the state into
2552 * NEED a dummy.
2553 * We don't need to cover the dummy.
2554 */
2555 if (!changed &&
2556 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2557 XLOG_COVER_OPS)) {
2558 changed = 1;
2559 } else {
2560 /*
2561 * We have two dirty iclogs so start over
2562 * This could also be num of ops indicates
2563 * this is not the dummy going out.
2564 */
2565 changed = 2;
2566 }
2567 iclog->ic_header.h_num_logops = 0;
2568 memset(iclog->ic_header.h_cycle_data, 0,
2569 sizeof(iclog->ic_header.h_cycle_data));
2570 iclog->ic_header.h_lsn = 0;
2571 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2572 /* do nothing */;
2573 else
2574 break; /* stop cleaning */
2575 iclog = iclog->ic_next;
2576 } while (iclog != log->l_iclog);
2577
2578 /* log is locked when we are called */
2579 /*
2580 * Change state for the dummy log recording.
2581 * We usually go to NEED. But we go to NEED2 if the changed indicates
2582 * we are done writing the dummy record.
2583 * If we are done with the second dummy recored (DONE2), then
2584 * we go to IDLE.
2585 */
2586 if (changed) {
2587 switch (log->l_covered_state) {
2588 case XLOG_STATE_COVER_IDLE:
2589 case XLOG_STATE_COVER_NEED:
2590 case XLOG_STATE_COVER_NEED2:
2591 log->l_covered_state = XLOG_STATE_COVER_NEED;
2592 break;
2593
2594 case XLOG_STATE_COVER_DONE:
2595 if (changed == 1)
2596 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2597 else
2598 log->l_covered_state = XLOG_STATE_COVER_NEED;
2599 break;
2600
2601 case XLOG_STATE_COVER_DONE2:
2602 if (changed == 1)
2603 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2604 else
2605 log->l_covered_state = XLOG_STATE_COVER_NEED;
2606 break;
2607
2608 default:
2609 ASSERT(0);
2610 }
2611 }
2612} /* xlog_state_clean_log */
2613
2614STATIC xfs_lsn_t
2615xlog_get_lowest_lsn(
2616 struct xlog *log)
2617{
2618 xlog_in_core_t *lsn_log;
2619 xfs_lsn_t lowest_lsn, lsn;
2620
2621 lsn_log = log->l_iclog;
2622 lowest_lsn = 0;
2623 do {
2624 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2625 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2626 if ((lsn && !lowest_lsn) ||
2627 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2628 lowest_lsn = lsn;
2629 }
2630 }
2631 lsn_log = lsn_log->ic_next;
2632 } while (lsn_log != log->l_iclog);
2633 return lowest_lsn;
2634}
2635
2636
2637STATIC void
2638xlog_state_do_callback(
2639 struct xlog *log,
2640 int aborted,
2641 struct xlog_in_core *ciclog)
2642{
2643 xlog_in_core_t *iclog;
2644 xlog_in_core_t *first_iclog; /* used to know when we've
2645 * processed all iclogs once */
2646 xfs_log_callback_t *cb, *cb_next;
2647 int flushcnt = 0;
2648 xfs_lsn_t lowest_lsn;
2649 int ioerrors; /* counter: iclogs with errors */
2650 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2651 int funcdidcallbacks; /* flag: function did callbacks */
2652 int repeats; /* for issuing console warnings if
2653 * looping too many times */
2654 int wake = 0;
2655
2656 spin_lock(&log->l_icloglock);
2657 first_iclog = iclog = log->l_iclog;
2658 ioerrors = 0;
2659 funcdidcallbacks = 0;
2660 repeats = 0;
2661
2662 do {
2663 /*
2664 * Scan all iclogs starting with the one pointed to by the
2665 * log. Reset this starting point each time the log is
2666 * unlocked (during callbacks).
2667 *
2668 * Keep looping through iclogs until one full pass is made
2669 * without running any callbacks.
2670 */
2671 first_iclog = log->l_iclog;
2672 iclog = log->l_iclog;
2673 loopdidcallbacks = 0;
2674 repeats++;
2675
2676 do {
2677
2678 /* skip all iclogs in the ACTIVE & DIRTY states */
2679 if (iclog->ic_state &
2680 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2681 iclog = iclog->ic_next;
2682 continue;
2683 }
2684
2685 /*
2686 * Between marking a filesystem SHUTDOWN and stopping
2687 * the log, we do flush all iclogs to disk (if there
2688 * wasn't a log I/O error). So, we do want things to
2689 * go smoothly in case of just a SHUTDOWN w/o a
2690 * LOG_IO_ERROR.
2691 */
2692 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2693 /*
2694 * Can only perform callbacks in order. Since
2695 * this iclog is not in the DONE_SYNC/
2696 * DO_CALLBACK state, we skip the rest and
2697 * just try to clean up. If we set our iclog
2698 * to DO_CALLBACK, we will not process it when
2699 * we retry since a previous iclog is in the
2700 * CALLBACK and the state cannot change since
2701 * we are holding the l_icloglock.
2702 */
2703 if (!(iclog->ic_state &
2704 (XLOG_STATE_DONE_SYNC |
2705 XLOG_STATE_DO_CALLBACK))) {
2706 if (ciclog && (ciclog->ic_state ==
2707 XLOG_STATE_DONE_SYNC)) {
2708 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2709 }
2710 break;
2711 }
2712 /*
2713 * We now have an iclog that is in either the
2714 * DO_CALLBACK or DONE_SYNC states. The other
2715 * states (WANT_SYNC, SYNCING, or CALLBACK were
2716 * caught by the above if and are going to
2717 * clean (i.e. we aren't doing their callbacks)
2718 * see the above if.
2719 */
2720
2721 /*
2722 * We will do one more check here to see if we
2723 * have chased our tail around.
2724 */
2725
2726 lowest_lsn = xlog_get_lowest_lsn(log);
2727 if (lowest_lsn &&
2728 XFS_LSN_CMP(lowest_lsn,
2729 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2730 iclog = iclog->ic_next;
2731 continue; /* Leave this iclog for
2732 * another thread */
2733 }
2734
2735 iclog->ic_state = XLOG_STATE_CALLBACK;
2736
2737
2738 /*
2739 * Completion of a iclog IO does not imply that
2740 * a transaction has completed, as transactions
2741 * can be large enough to span many iclogs. We
2742 * cannot change the tail of the log half way
2743 * through a transaction as this may be the only
2744 * transaction in the log and moving th etail to
2745 * point to the middle of it will prevent
2746 * recovery from finding the start of the
2747 * transaction. Hence we should only update the
2748 * last_sync_lsn if this iclog contains
2749 * transaction completion callbacks on it.
2750 *
2751 * We have to do this before we drop the
2752 * icloglock to ensure we are the only one that
2753 * can update it.
2754 */
2755 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2756 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2757 if (iclog->ic_callback)
2758 atomic64_set(&log->l_last_sync_lsn,
2759 be64_to_cpu(iclog->ic_header.h_lsn));
2760
2761 } else
2762 ioerrors++;
2763
2764 spin_unlock(&log->l_icloglock);
2765
2766 /*
2767 * Keep processing entries in the callback list until
2768 * we come around and it is empty. We need to
2769 * atomically see that the list is empty and change the
2770 * state to DIRTY so that we don't miss any more
2771 * callbacks being added.
2772 */
2773 spin_lock(&iclog->ic_callback_lock);
2774 cb = iclog->ic_callback;
2775 while (cb) {
2776 iclog->ic_callback_tail = &(iclog->ic_callback);
2777 iclog->ic_callback = NULL;
2778 spin_unlock(&iclog->ic_callback_lock);
2779
2780 /* perform callbacks in the order given */
2781 for (; cb; cb = cb_next) {
2782 cb_next = cb->cb_next;
2783 cb->cb_func(cb->cb_arg, aborted);
2784 }
2785 spin_lock(&iclog->ic_callback_lock);
2786 cb = iclog->ic_callback;
2787 }
2788
2789 loopdidcallbacks++;
2790 funcdidcallbacks++;
2791
2792 spin_lock(&log->l_icloglock);
2793 ASSERT(iclog->ic_callback == NULL);
2794 spin_unlock(&iclog->ic_callback_lock);
2795 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2796 iclog->ic_state = XLOG_STATE_DIRTY;
2797
2798 /*
2799 * Transition from DIRTY to ACTIVE if applicable.
2800 * NOP if STATE_IOERROR.
2801 */
2802 xlog_state_clean_log(log);
2803
2804 /* wake up threads waiting in xfs_log_force() */
2805 wake_up_all(&iclog->ic_force_wait);
2806
2807 iclog = iclog->ic_next;
2808 } while (first_iclog != iclog);
2809
2810 if (repeats > 5000) {
2811 flushcnt += repeats;
2812 repeats = 0;
2813 xfs_warn(log->l_mp,
2814 "%s: possible infinite loop (%d iterations)",
2815 __func__, flushcnt);
2816 }
2817 } while (!ioerrors && loopdidcallbacks);
2818
2819#ifdef DEBUG
2820 /*
2821 * Make one last gasp attempt to see if iclogs are being left in limbo.
2822 * If the above loop finds an iclog earlier than the current iclog and
2823 * in one of the syncing states, the current iclog is put into
2824 * DO_CALLBACK and the callbacks are deferred to the completion of the
2825 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2826 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2827 * states.
2828 *
2829 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2830 * for ic_state == SYNCING.
2831 */
2832 if (funcdidcallbacks) {
2833 first_iclog = iclog = log->l_iclog;
2834 do {
2835 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2836 /*
2837 * Terminate the loop if iclogs are found in states
2838 * which will cause other threads to clean up iclogs.
2839 *
2840 * SYNCING - i/o completion will go through logs
2841 * DONE_SYNC - interrupt thread should be waiting for
2842 * l_icloglock
2843 * IOERROR - give up hope all ye who enter here
2844 */
2845 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2846 iclog->ic_state & XLOG_STATE_SYNCING ||
2847 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2848 iclog->ic_state == XLOG_STATE_IOERROR )
2849 break;
2850 iclog = iclog->ic_next;
2851 } while (first_iclog != iclog);
2852 }
2853#endif
2854
2855 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2856 wake = 1;
2857 spin_unlock(&log->l_icloglock);
2858
2859 if (wake)
2860 wake_up_all(&log->l_flush_wait);
2861}
2862
2863
2864/*
2865 * Finish transitioning this iclog to the dirty state.
2866 *
2867 * Make sure that we completely execute this routine only when this is
2868 * the last call to the iclog. There is a good chance that iclog flushes,
2869 * when we reach the end of the physical log, get turned into 2 separate
2870 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2871 * routine. By using the reference count bwritecnt, we guarantee that only
2872 * the second completion goes through.
2873 *
2874 * Callbacks could take time, so they are done outside the scope of the
2875 * global state machine log lock.
2876 */
2877STATIC void
2878xlog_state_done_syncing(
2879 xlog_in_core_t *iclog,
2880 int aborted)
2881{
2882 struct xlog *log = iclog->ic_log;
2883
2884 spin_lock(&log->l_icloglock);
2885
2886 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2887 iclog->ic_state == XLOG_STATE_IOERROR);
2888 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2889 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2890
2891
2892 /*
2893 * If we got an error, either on the first buffer, or in the case of
2894 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2895 * and none should ever be attempted to be written to disk
2896 * again.
2897 */
2898 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2899 if (--iclog->ic_bwritecnt == 1) {
2900 spin_unlock(&log->l_icloglock);
2901 return;
2902 }
2903 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2904 }
2905
2906 /*
2907 * Someone could be sleeping prior to writing out the next
2908 * iclog buffer, we wake them all, one will get to do the
2909 * I/O, the others get to wait for the result.
2910 */
2911 wake_up_all(&iclog->ic_write_wait);
2912 spin_unlock(&log->l_icloglock);
2913 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2914} /* xlog_state_done_syncing */
2915
2916
2917/*
2918 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2919 * sleep. We wait on the flush queue on the head iclog as that should be
2920 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2921 * we will wait here and all new writes will sleep until a sync completes.
2922 *
2923 * The in-core logs are used in a circular fashion. They are not used
2924 * out-of-order even when an iclog past the head is free.
2925 *
2926 * return:
2927 * * log_offset where xlog_write() can start writing into the in-core
2928 * log's data space.
2929 * * in-core log pointer to which xlog_write() should write.
2930 * * boolean indicating this is a continued write to an in-core log.
2931 * If this is the last write, then the in-core log's offset field
2932 * needs to be incremented, depending on the amount of data which
2933 * is copied.
2934 */
2935STATIC int
2936xlog_state_get_iclog_space(
2937 struct xlog *log,
2938 int len,
2939 struct xlog_in_core **iclogp,
2940 struct xlog_ticket *ticket,
2941 int *continued_write,
2942 int *logoffsetp)
2943{
2944 int log_offset;
2945 xlog_rec_header_t *head;
2946 xlog_in_core_t *iclog;
2947 int error;
2948
2949restart:
2950 spin_lock(&log->l_icloglock);
2951 if (XLOG_FORCED_SHUTDOWN(log)) {
2952 spin_unlock(&log->l_icloglock);
2953 return -EIO;
2954 }
2955
2956 iclog = log->l_iclog;
2957 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2958 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2959
2960 /* Wait for log writes to have flushed */
2961 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2962 goto restart;
2963 }
2964
2965 head = &iclog->ic_header;
2966
2967 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2968 log_offset = iclog->ic_offset;
2969
2970 /* On the 1st write to an iclog, figure out lsn. This works
2971 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2972 * committing to. If the offset is set, that's how many blocks
2973 * must be written.
2974 */
2975 if (log_offset == 0) {
2976 ticket->t_curr_res -= log->l_iclog_hsize;
2977 xlog_tic_add_region(ticket,
2978 log->l_iclog_hsize,
2979 XLOG_REG_TYPE_LRHEADER);
2980 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2981 head->h_lsn = cpu_to_be64(
2982 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2983 ASSERT(log->l_curr_block >= 0);
2984 }
2985
2986 /* If there is enough room to write everything, then do it. Otherwise,
2987 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2988 * bit is on, so this will get flushed out. Don't update ic_offset
2989 * until you know exactly how many bytes get copied. Therefore, wait
2990 * until later to update ic_offset.
2991 *
2992 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2993 * can fit into remaining data section.
2994 */
2995 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2996 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2997
2998 /*
2999 * If I'm the only one writing to this iclog, sync it to disk.
3000 * We need to do an atomic compare and decrement here to avoid
3001 * racing with concurrent atomic_dec_and_lock() calls in
3002 * xlog_state_release_iclog() when there is more than one
3003 * reference to the iclog.
3004 */
3005 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3006 /* we are the only one */
3007 spin_unlock(&log->l_icloglock);
3008 error = xlog_state_release_iclog(log, iclog);
3009 if (error)
3010 return error;
3011 } else {
3012 spin_unlock(&log->l_icloglock);
3013 }
3014 goto restart;
3015 }
3016
3017 /* Do we have enough room to write the full amount in the remainder
3018 * of this iclog? Or must we continue a write on the next iclog and
3019 * mark this iclog as completely taken? In the case where we switch
3020 * iclogs (to mark it taken), this particular iclog will release/sync
3021 * to disk in xlog_write().
3022 */
3023 if (len <= iclog->ic_size - iclog->ic_offset) {
3024 *continued_write = 0;
3025 iclog->ic_offset += len;
3026 } else {
3027 *continued_write = 1;
3028 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3029 }
3030 *iclogp = iclog;
3031
3032 ASSERT(iclog->ic_offset <= iclog->ic_size);
3033 spin_unlock(&log->l_icloglock);
3034
3035 *logoffsetp = log_offset;
3036 return 0;
3037} /* xlog_state_get_iclog_space */
3038
3039/* The first cnt-1 times through here we don't need to
3040 * move the grant write head because the permanent
3041 * reservation has reserved cnt times the unit amount.
3042 * Release part of current permanent unit reservation and
3043 * reset current reservation to be one units worth. Also
3044 * move grant reservation head forward.
3045 */
3046STATIC void
3047xlog_regrant_reserve_log_space(
3048 struct xlog *log,
3049 struct xlog_ticket *ticket)
3050{
3051 trace_xfs_log_regrant_reserve_enter(log, ticket);
3052
3053 if (ticket->t_cnt > 0)
3054 ticket->t_cnt--;
3055
3056 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3057 ticket->t_curr_res);
3058 xlog_grant_sub_space(log, &log->l_write_head.grant,
3059 ticket->t_curr_res);
3060 ticket->t_curr_res = ticket->t_unit_res;
3061 xlog_tic_reset_res(ticket);
3062
3063 trace_xfs_log_regrant_reserve_sub(log, ticket);
3064
3065 /* just return if we still have some of the pre-reserved space */
3066 if (ticket->t_cnt > 0)
3067 return;
3068
3069 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3070 ticket->t_unit_res);
3071
3072 trace_xfs_log_regrant_reserve_exit(log, ticket);
3073
3074 ticket->t_curr_res = ticket->t_unit_res;
3075 xlog_tic_reset_res(ticket);
3076} /* xlog_regrant_reserve_log_space */
3077
3078
3079/*
3080 * Give back the space left from a reservation.
3081 *
3082 * All the information we need to make a correct determination of space left
3083 * is present. For non-permanent reservations, things are quite easy. The
3084 * count should have been decremented to zero. We only need to deal with the
3085 * space remaining in the current reservation part of the ticket. If the
3086 * ticket contains a permanent reservation, there may be left over space which
3087 * needs to be released. A count of N means that N-1 refills of the current
3088 * reservation can be done before we need to ask for more space. The first
3089 * one goes to fill up the first current reservation. Once we run out of
3090 * space, the count will stay at zero and the only space remaining will be
3091 * in the current reservation field.
3092 */
3093STATIC void
3094xlog_ungrant_log_space(
3095 struct xlog *log,
3096 struct xlog_ticket *ticket)
3097{
3098 int bytes;
3099
3100 if (ticket->t_cnt > 0)
3101 ticket->t_cnt--;
3102
3103 trace_xfs_log_ungrant_enter(log, ticket);
3104 trace_xfs_log_ungrant_sub(log, ticket);
3105
3106 /*
3107 * If this is a permanent reservation ticket, we may be able to free
3108 * up more space based on the remaining count.
3109 */
3110 bytes = ticket->t_curr_res;
3111 if (ticket->t_cnt > 0) {
3112 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3113 bytes += ticket->t_unit_res*ticket->t_cnt;
3114 }
3115
3116 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3117 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3118
3119 trace_xfs_log_ungrant_exit(log, ticket);
3120
3121 xfs_log_space_wake(log->l_mp);
3122}
3123
3124/*
3125 * Flush iclog to disk if this is the last reference to the given iclog and
3126 * the WANT_SYNC bit is set.
3127 *
3128 * When this function is entered, the iclog is not necessarily in the
3129 * WANT_SYNC state. It may be sitting around waiting to get filled.
3130 *
3131 *
3132 */
3133STATIC int
3134xlog_state_release_iclog(
3135 struct xlog *log,
3136 struct xlog_in_core *iclog)
3137{
3138 int sync = 0; /* do we sync? */
3139
3140 if (iclog->ic_state & XLOG_STATE_IOERROR)
3141 return -EIO;
3142
3143 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3144 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3145 return 0;
3146
3147 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3148 spin_unlock(&log->l_icloglock);
3149 return -EIO;
3150 }
3151 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3152 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3153
3154 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3155 /* update tail before writing to iclog */
3156 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3157 sync++;
3158 iclog->ic_state = XLOG_STATE_SYNCING;
3159 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3160 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3161 /* cycle incremented when incrementing curr_block */
3162 }
3163 spin_unlock(&log->l_icloglock);
3164
3165 /*
3166 * We let the log lock go, so it's possible that we hit a log I/O
3167 * error or some other SHUTDOWN condition that marks the iclog
3168 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3169 * this iclog has consistent data, so we ignore IOERROR
3170 * flags after this point.
3171 */
3172 if (sync)
3173 return xlog_sync(log, iclog);
3174 return 0;
3175} /* xlog_state_release_iclog */
3176
3177
3178/*
3179 * This routine will mark the current iclog in the ring as WANT_SYNC
3180 * and move the current iclog pointer to the next iclog in the ring.
3181 * When this routine is called from xlog_state_get_iclog_space(), the
3182 * exact size of the iclog has not yet been determined. All we know is
3183 * that every data block. We have run out of space in this log record.
3184 */
3185STATIC void
3186xlog_state_switch_iclogs(
3187 struct xlog *log,
3188 struct xlog_in_core *iclog,
3189 int eventual_size)
3190{
3191 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3192 if (!eventual_size)
3193 eventual_size = iclog->ic_offset;
3194 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3195 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3196 log->l_prev_block = log->l_curr_block;
3197 log->l_prev_cycle = log->l_curr_cycle;
3198
3199 /* roll log?: ic_offset changed later */
3200 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3201
3202 /* Round up to next log-sunit */
3203 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3204 log->l_mp->m_sb.sb_logsunit > 1) {
3205 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3206 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3207 }
3208
3209 if (log->l_curr_block >= log->l_logBBsize) {
3210 /*
3211 * Rewind the current block before the cycle is bumped to make
3212 * sure that the combined LSN never transiently moves forward
3213 * when the log wraps to the next cycle. This is to support the
3214 * unlocked sample of these fields from xlog_valid_lsn(). Most
3215 * other cases should acquire l_icloglock.
3216 */
3217 log->l_curr_block -= log->l_logBBsize;
3218 ASSERT(log->l_curr_block >= 0);
3219 smp_wmb();
3220 log->l_curr_cycle++;
3221 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3222 log->l_curr_cycle++;
3223 }
3224 ASSERT(iclog == log->l_iclog);
3225 log->l_iclog = iclog->ic_next;
3226} /* xlog_state_switch_iclogs */
3227
3228/*
3229 * Write out all data in the in-core log as of this exact moment in time.
3230 *
3231 * Data may be written to the in-core log during this call. However,
3232 * we don't guarantee this data will be written out. A change from past
3233 * implementation means this routine will *not* write out zero length LRs.
3234 *
3235 * Basically, we try and perform an intelligent scan of the in-core logs.
3236 * If we determine there is no flushable data, we just return. There is no
3237 * flushable data if:
3238 *
3239 * 1. the current iclog is active and has no data; the previous iclog
3240 * is in the active or dirty state.
3241 * 2. the current iclog is drity, and the previous iclog is in the
3242 * active or dirty state.
3243 *
3244 * We may sleep if:
3245 *
3246 * 1. the current iclog is not in the active nor dirty state.
3247 * 2. the current iclog dirty, and the previous iclog is not in the
3248 * active nor dirty state.
3249 * 3. the current iclog is active, and there is another thread writing
3250 * to this particular iclog.
3251 * 4. a) the current iclog is active and has no other writers
3252 * b) when we return from flushing out this iclog, it is still
3253 * not in the active nor dirty state.
3254 */
3255int
3256_xfs_log_force(
3257 struct xfs_mount *mp,
3258 uint flags,
3259 int *log_flushed)
3260{
3261 struct xlog *log = mp->m_log;
3262 struct xlog_in_core *iclog;
3263 xfs_lsn_t lsn;
3264
3265 XFS_STATS_INC(mp, xs_log_force);
3266
3267 xlog_cil_force(log);
3268
3269 spin_lock(&log->l_icloglock);
3270
3271 iclog = log->l_iclog;
3272 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3273 spin_unlock(&log->l_icloglock);
3274 return -EIO;
3275 }
3276
3277 /* If the head iclog is not active nor dirty, we just attach
3278 * ourselves to the head and go to sleep.
3279 */
3280 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3281 iclog->ic_state == XLOG_STATE_DIRTY) {
3282 /*
3283 * If the head is dirty or (active and empty), then
3284 * we need to look at the previous iclog. If the previous
3285 * iclog is active or dirty we are done. There is nothing
3286 * to sync out. Otherwise, we attach ourselves to the
3287 * previous iclog and go to sleep.
3288 */
3289 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3290 (atomic_read(&iclog->ic_refcnt) == 0
3291 && iclog->ic_offset == 0)) {
3292 iclog = iclog->ic_prev;
3293 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3294 iclog->ic_state == XLOG_STATE_DIRTY)
3295 goto no_sleep;
3296 else
3297 goto maybe_sleep;
3298 } else {
3299 if (atomic_read(&iclog->ic_refcnt) == 0) {
3300 /* We are the only one with access to this
3301 * iclog. Flush it out now. There should
3302 * be a roundoff of zero to show that someone
3303 * has already taken care of the roundoff from
3304 * the previous sync.
3305 */
3306 atomic_inc(&iclog->ic_refcnt);
3307 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3308 xlog_state_switch_iclogs(log, iclog, 0);
3309 spin_unlock(&log->l_icloglock);
3310
3311 if (xlog_state_release_iclog(log, iclog))
3312 return -EIO;
3313
3314 if (log_flushed)
3315 *log_flushed = 1;
3316 spin_lock(&log->l_icloglock);
3317 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3318 iclog->ic_state != XLOG_STATE_DIRTY)
3319 goto maybe_sleep;
3320 else
3321 goto no_sleep;
3322 } else {
3323 /* Someone else is writing to this iclog.
3324 * Use its call to flush out the data. However,
3325 * the other thread may not force out this LR,
3326 * so we mark it WANT_SYNC.
3327 */
3328 xlog_state_switch_iclogs(log, iclog, 0);
3329 goto maybe_sleep;
3330 }
3331 }
3332 }
3333
3334 /* By the time we come around again, the iclog could've been filled
3335 * which would give it another lsn. If we have a new lsn, just
3336 * return because the relevant data has been flushed.
3337 */
3338maybe_sleep:
3339 if (flags & XFS_LOG_SYNC) {
3340 /*
3341 * We must check if we're shutting down here, before
3342 * we wait, while we're holding the l_icloglock.
3343 * Then we check again after waking up, in case our
3344 * sleep was disturbed by a bad news.
3345 */
3346 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3347 spin_unlock(&log->l_icloglock);
3348 return -EIO;
3349 }
3350 XFS_STATS_INC(mp, xs_log_force_sleep);
3351 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3352 /*
3353 * No need to grab the log lock here since we're
3354 * only deciding whether or not to return EIO
3355 * and the memory read should be atomic.
3356 */
3357 if (iclog->ic_state & XLOG_STATE_IOERROR)
3358 return -EIO;
3359 if (log_flushed)
3360 *log_flushed = 1;
3361 } else {
3362
3363no_sleep:
3364 spin_unlock(&log->l_icloglock);
3365 }
3366 return 0;
3367}
3368
3369/*
3370 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3371 * about errors or whether the log was flushed or not. This is the normal
3372 * interface to use when trying to unpin items or move the log forward.
3373 */
3374void
3375xfs_log_force(
3376 xfs_mount_t *mp,
3377 uint flags)
3378{
3379 int error;
3380
3381 trace_xfs_log_force(mp, 0);
3382 error = _xfs_log_force(mp, flags, NULL);
3383 if (error)
3384 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3385}
3386
3387/*
3388 * Force the in-core log to disk for a specific LSN.
3389 *
3390 * Find in-core log with lsn.
3391 * If it is in the DIRTY state, just return.
3392 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3393 * state and go to sleep or return.
3394 * If it is in any other state, go to sleep or return.
3395 *
3396 * Synchronous forces are implemented with a signal variable. All callers
3397 * to force a given lsn to disk will wait on a the sv attached to the
3398 * specific in-core log. When given in-core log finally completes its
3399 * write to disk, that thread will wake up all threads waiting on the
3400 * sv.
3401 */
3402int
3403_xfs_log_force_lsn(
3404 struct xfs_mount *mp,
3405 xfs_lsn_t lsn,
3406 uint flags,
3407 int *log_flushed)
3408{
3409 struct xlog *log = mp->m_log;
3410 struct xlog_in_core *iclog;
3411 int already_slept = 0;
3412
3413 ASSERT(lsn != 0);
3414
3415 XFS_STATS_INC(mp, xs_log_force);
3416
3417 lsn = xlog_cil_force_lsn(log, lsn);
3418 if (lsn == NULLCOMMITLSN)
3419 return 0;
3420
3421try_again:
3422 spin_lock(&log->l_icloglock);
3423 iclog = log->l_iclog;
3424 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3425 spin_unlock(&log->l_icloglock);
3426 return -EIO;
3427 }
3428
3429 do {
3430 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3431 iclog = iclog->ic_next;
3432 continue;
3433 }
3434
3435 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3436 spin_unlock(&log->l_icloglock);
3437 return 0;
3438 }
3439
3440 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3441 /*
3442 * We sleep here if we haven't already slept (e.g.
3443 * this is the first time we've looked at the correct
3444 * iclog buf) and the buffer before us is going to
3445 * be sync'ed. The reason for this is that if we
3446 * are doing sync transactions here, by waiting for
3447 * the previous I/O to complete, we can allow a few
3448 * more transactions into this iclog before we close
3449 * it down.
3450 *
3451 * Otherwise, we mark the buffer WANT_SYNC, and bump
3452 * up the refcnt so we can release the log (which
3453 * drops the ref count). The state switch keeps new
3454 * transaction commits from using this buffer. When
3455 * the current commits finish writing into the buffer,
3456 * the refcount will drop to zero and the buffer will
3457 * go out then.
3458 */
3459 if (!already_slept &&
3460 (iclog->ic_prev->ic_state &
3461 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3462 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3463
3464 XFS_STATS_INC(mp, xs_log_force_sleep);
3465
3466 xlog_wait(&iclog->ic_prev->ic_write_wait,
3467 &log->l_icloglock);
3468 if (log_flushed)
3469 *log_flushed = 1;
3470 already_slept = 1;
3471 goto try_again;
3472 }
3473 atomic_inc(&iclog->ic_refcnt);
3474 xlog_state_switch_iclogs(log, iclog, 0);
3475 spin_unlock(&log->l_icloglock);
3476 if (xlog_state_release_iclog(log, iclog))
3477 return -EIO;
3478 if (log_flushed)
3479 *log_flushed = 1;
3480 spin_lock(&log->l_icloglock);
3481 }
3482
3483 if ((flags & XFS_LOG_SYNC) && /* sleep */
3484 !(iclog->ic_state &
3485 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3486 /*
3487 * Don't wait on completion if we know that we've
3488 * gotten a log write error.
3489 */
3490 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3491 spin_unlock(&log->l_icloglock);
3492 return -EIO;
3493 }
3494 XFS_STATS_INC(mp, xs_log_force_sleep);
3495 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3496 /*
3497 * No need to grab the log lock here since we're
3498 * only deciding whether or not to return EIO
3499 * and the memory read should be atomic.
3500 */
3501 if (iclog->ic_state & XLOG_STATE_IOERROR)
3502 return -EIO;
3503
3504 if (log_flushed)
3505 *log_flushed = 1;
3506 } else { /* just return */
3507 spin_unlock(&log->l_icloglock);
3508 }
3509
3510 return 0;
3511 } while (iclog != log->l_iclog);
3512
3513 spin_unlock(&log->l_icloglock);
3514 return 0;
3515}
3516
3517/*
3518 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3519 * about errors or whether the log was flushed or not. This is the normal
3520 * interface to use when trying to unpin items or move the log forward.
3521 */
3522void
3523xfs_log_force_lsn(
3524 xfs_mount_t *mp,
3525 xfs_lsn_t lsn,
3526 uint flags)
3527{
3528 int error;
3529
3530 trace_xfs_log_force(mp, lsn);
3531 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3532 if (error)
3533 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3534}
3535
3536/*
3537 * Called when we want to mark the current iclog as being ready to sync to
3538 * disk.
3539 */
3540STATIC void
3541xlog_state_want_sync(
3542 struct xlog *log,
3543 struct xlog_in_core *iclog)
3544{
3545 assert_spin_locked(&log->l_icloglock);
3546
3547 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3548 xlog_state_switch_iclogs(log, iclog, 0);
3549 } else {
3550 ASSERT(iclog->ic_state &
3551 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3552 }
3553}
3554
3555
3556/*****************************************************************************
3557 *
3558 * TICKET functions
3559 *
3560 *****************************************************************************
3561 */
3562
3563/*
3564 * Free a used ticket when its refcount falls to zero.
3565 */
3566void
3567xfs_log_ticket_put(
3568 xlog_ticket_t *ticket)
3569{
3570 ASSERT(atomic_read(&ticket->t_ref) > 0);
3571 if (atomic_dec_and_test(&ticket->t_ref))
3572 kmem_zone_free(xfs_log_ticket_zone, ticket);
3573}
3574
3575xlog_ticket_t *
3576xfs_log_ticket_get(
3577 xlog_ticket_t *ticket)
3578{
3579 ASSERT(atomic_read(&ticket->t_ref) > 0);
3580 atomic_inc(&ticket->t_ref);
3581 return ticket;
3582}
3583
3584/*
3585 * Figure out the total log space unit (in bytes) that would be
3586 * required for a log ticket.
3587 */
3588int
3589xfs_log_calc_unit_res(
3590 struct xfs_mount *mp,
3591 int unit_bytes)
3592{
3593 struct xlog *log = mp->m_log;
3594 int iclog_space;
3595 uint num_headers;
3596
3597 /*
3598 * Permanent reservations have up to 'cnt'-1 active log operations
3599 * in the log. A unit in this case is the amount of space for one
3600 * of these log operations. Normal reservations have a cnt of 1
3601 * and their unit amount is the total amount of space required.
3602 *
3603 * The following lines of code account for non-transaction data
3604 * which occupy space in the on-disk log.
3605 *
3606 * Normal form of a transaction is:
3607 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3608 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3609 *
3610 * We need to account for all the leadup data and trailer data
3611 * around the transaction data.
3612 * And then we need to account for the worst case in terms of using
3613 * more space.
3614 * The worst case will happen if:
3615 * - the placement of the transaction happens to be such that the
3616 * roundoff is at its maximum
3617 * - the transaction data is synced before the commit record is synced
3618 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3619 * Therefore the commit record is in its own Log Record.
3620 * This can happen as the commit record is called with its
3621 * own region to xlog_write().
3622 * This then means that in the worst case, roundoff can happen for
3623 * the commit-rec as well.
3624 * The commit-rec is smaller than padding in this scenario and so it is
3625 * not added separately.
3626 */
3627
3628 /* for trans header */
3629 unit_bytes += sizeof(xlog_op_header_t);
3630 unit_bytes += sizeof(xfs_trans_header_t);
3631
3632 /* for start-rec */
3633 unit_bytes += sizeof(xlog_op_header_t);
3634
3635 /*
3636 * for LR headers - the space for data in an iclog is the size minus
3637 * the space used for the headers. If we use the iclog size, then we
3638 * undercalculate the number of headers required.
3639 *
3640 * Furthermore - the addition of op headers for split-recs might
3641 * increase the space required enough to require more log and op
3642 * headers, so take that into account too.
3643 *
3644 * IMPORTANT: This reservation makes the assumption that if this
3645 * transaction is the first in an iclog and hence has the LR headers
3646 * accounted to it, then the remaining space in the iclog is
3647 * exclusively for this transaction. i.e. if the transaction is larger
3648 * than the iclog, it will be the only thing in that iclog.
3649 * Fundamentally, this means we must pass the entire log vector to
3650 * xlog_write to guarantee this.
3651 */
3652 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3653 num_headers = howmany(unit_bytes, iclog_space);
3654
3655 /* for split-recs - ophdrs added when data split over LRs */
3656 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3657
3658 /* add extra header reservations if we overrun */
3659 while (!num_headers ||
3660 howmany(unit_bytes, iclog_space) > num_headers) {
3661 unit_bytes += sizeof(xlog_op_header_t);
3662 num_headers++;
3663 }
3664 unit_bytes += log->l_iclog_hsize * num_headers;
3665
3666 /* for commit-rec LR header - note: padding will subsume the ophdr */
3667 unit_bytes += log->l_iclog_hsize;
3668
3669 /* for roundoff padding for transaction data and one for commit record */
3670 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3671 /* log su roundoff */
3672 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3673 } else {
3674 /* BB roundoff */
3675 unit_bytes += 2 * BBSIZE;
3676 }
3677
3678 return unit_bytes;
3679}
3680
3681/*
3682 * Allocate and initialise a new log ticket.
3683 */
3684struct xlog_ticket *
3685xlog_ticket_alloc(
3686 struct xlog *log,
3687 int unit_bytes,
3688 int cnt,
3689 char client,
3690 bool permanent,
3691 xfs_km_flags_t alloc_flags)
3692{
3693 struct xlog_ticket *tic;
3694 int unit_res;
3695
3696 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3697 if (!tic)
3698 return NULL;
3699
3700 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3701
3702 atomic_set(&tic->t_ref, 1);
3703 tic->t_task = current;
3704 INIT_LIST_HEAD(&tic->t_queue);
3705 tic->t_unit_res = unit_res;
3706 tic->t_curr_res = unit_res;
3707 tic->t_cnt = cnt;
3708 tic->t_ocnt = cnt;
3709 tic->t_tid = prandom_u32();
3710 tic->t_clientid = client;
3711 tic->t_flags = XLOG_TIC_INITED;
3712 tic->t_trans_type = 0;
3713 if (permanent)
3714 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3715
3716 xlog_tic_reset_res(tic);
3717
3718 return tic;
3719}
3720
3721
3722/******************************************************************************
3723 *
3724 * Log debug routines
3725 *
3726 ******************************************************************************
3727 */
3728#if defined(DEBUG)
3729/*
3730 * Make sure that the destination ptr is within the valid data region of
3731 * one of the iclogs. This uses backup pointers stored in a different
3732 * part of the log in case we trash the log structure.
3733 */
3734void
3735xlog_verify_dest_ptr(
3736 struct xlog *log,
3737 void *ptr)
3738{
3739 int i;
3740 int good_ptr = 0;
3741
3742 for (i = 0; i < log->l_iclog_bufs; i++) {
3743 if (ptr >= log->l_iclog_bak[i] &&
3744 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3745 good_ptr++;
3746 }
3747
3748 if (!good_ptr)
3749 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3750}
3751
3752/*
3753 * Check to make sure the grant write head didn't just over lap the tail. If
3754 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3755 * the cycles differ by exactly one and check the byte count.
3756 *
3757 * This check is run unlocked, so can give false positives. Rather than assert
3758 * on failures, use a warn-once flag and a panic tag to allow the admin to
3759 * determine if they want to panic the machine when such an error occurs. For
3760 * debug kernels this will have the same effect as using an assert but, unlinke
3761 * an assert, it can be turned off at runtime.
3762 */
3763STATIC void
3764xlog_verify_grant_tail(
3765 struct xlog *log)
3766{
3767 int tail_cycle, tail_blocks;
3768 int cycle, space;
3769
3770 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3771 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3772 if (tail_cycle != cycle) {
3773 if (cycle - 1 != tail_cycle &&
3774 !(log->l_flags & XLOG_TAIL_WARN)) {
3775 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3776 "%s: cycle - 1 != tail_cycle", __func__);
3777 log->l_flags |= XLOG_TAIL_WARN;
3778 }
3779
3780 if (space > BBTOB(tail_blocks) &&
3781 !(log->l_flags & XLOG_TAIL_WARN)) {
3782 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3783 "%s: space > BBTOB(tail_blocks)", __func__);
3784 log->l_flags |= XLOG_TAIL_WARN;
3785 }
3786 }
3787}
3788
3789/* check if it will fit */
3790STATIC void
3791xlog_verify_tail_lsn(
3792 struct xlog *log,
3793 struct xlog_in_core *iclog,
3794 xfs_lsn_t tail_lsn)
3795{
3796 int blocks;
3797
3798 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3799 blocks =
3800 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3801 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3802 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3803 } else {
3804 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3805
3806 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3807 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3808
3809 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3810 if (blocks < BTOBB(iclog->ic_offset) + 1)
3811 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3812 }
3813} /* xlog_verify_tail_lsn */
3814
3815/*
3816 * Perform a number of checks on the iclog before writing to disk.
3817 *
3818 * 1. Make sure the iclogs are still circular
3819 * 2. Make sure we have a good magic number
3820 * 3. Make sure we don't have magic numbers in the data
3821 * 4. Check fields of each log operation header for:
3822 * A. Valid client identifier
3823 * B. tid ptr value falls in valid ptr space (user space code)
3824 * C. Length in log record header is correct according to the
3825 * individual operation headers within record.
3826 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3827 * log, check the preceding blocks of the physical log to make sure all
3828 * the cycle numbers agree with the current cycle number.
3829 */
3830STATIC void
3831xlog_verify_iclog(
3832 struct xlog *log,
3833 struct xlog_in_core *iclog,
3834 int count,
3835 bool syncing)
3836{
3837 xlog_op_header_t *ophead;
3838 xlog_in_core_t *icptr;
3839 xlog_in_core_2_t *xhdr;
3840 void *base_ptr, *ptr, *p;
3841 ptrdiff_t field_offset;
3842 __uint8_t clientid;
3843 int len, i, j, k, op_len;
3844 int idx;
3845
3846 /* check validity of iclog pointers */
3847 spin_lock(&log->l_icloglock);
3848 icptr = log->l_iclog;
3849 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3850 ASSERT(icptr);
3851
3852 if (icptr != log->l_iclog)
3853 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3854 spin_unlock(&log->l_icloglock);
3855
3856 /* check log magic numbers */
3857 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3858 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3859
3860 base_ptr = ptr = &iclog->ic_header;
3861 p = &iclog->ic_header;
3862 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3863 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3864 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3865 __func__);
3866 }
3867
3868 /* check fields */
3869 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3870 base_ptr = ptr = iclog->ic_datap;
3871 ophead = ptr;
3872 xhdr = iclog->ic_data;
3873 for (i = 0; i < len; i++) {
3874 ophead = ptr;
3875
3876 /* clientid is only 1 byte */
3877 p = &ophead->oh_clientid;
3878 field_offset = p - base_ptr;
3879 if (!syncing || (field_offset & 0x1ff)) {
3880 clientid = ophead->oh_clientid;
3881 } else {
3882 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3883 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3884 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3885 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3886 clientid = xlog_get_client_id(
3887 xhdr[j].hic_xheader.xh_cycle_data[k]);
3888 } else {
3889 clientid = xlog_get_client_id(
3890 iclog->ic_header.h_cycle_data[idx]);
3891 }
3892 }
3893 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3894 xfs_warn(log->l_mp,
3895 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3896 __func__, clientid, ophead,
3897 (unsigned long)field_offset);
3898
3899 /* check length */
3900 p = &ophead->oh_len;
3901 field_offset = p - base_ptr;
3902 if (!syncing || (field_offset & 0x1ff)) {
3903 op_len = be32_to_cpu(ophead->oh_len);
3904 } else {
3905 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3906 (uintptr_t)iclog->ic_datap);
3907 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3908 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3909 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3910 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3911 } else {
3912 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3913 }
3914 }
3915 ptr += sizeof(xlog_op_header_t) + op_len;
3916 }
3917} /* xlog_verify_iclog */
3918#endif
3919
3920/*
3921 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3922 */
3923STATIC int
3924xlog_state_ioerror(
3925 struct xlog *log)
3926{
3927 xlog_in_core_t *iclog, *ic;
3928
3929 iclog = log->l_iclog;
3930 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3931 /*
3932 * Mark all the incore logs IOERROR.
3933 * From now on, no log flushes will result.
3934 */
3935 ic = iclog;
3936 do {
3937 ic->ic_state = XLOG_STATE_IOERROR;
3938 ic = ic->ic_next;
3939 } while (ic != iclog);
3940 return 0;
3941 }
3942 /*
3943 * Return non-zero, if state transition has already happened.
3944 */
3945 return 1;
3946}
3947
3948/*
3949 * This is called from xfs_force_shutdown, when we're forcibly
3950 * shutting down the filesystem, typically because of an IO error.
3951 * Our main objectives here are to make sure that:
3952 * a. if !logerror, flush the logs to disk. Anything modified
3953 * after this is ignored.
3954 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3955 * parties to find out, 'atomically'.
3956 * c. those who're sleeping on log reservations, pinned objects and
3957 * other resources get woken up, and be told the bad news.
3958 * d. nothing new gets queued up after (b) and (c) are done.
3959 *
3960 * Note: for the !logerror case we need to flush the regions held in memory out
3961 * to disk first. This needs to be done before the log is marked as shutdown,
3962 * otherwise the iclog writes will fail.
3963 */
3964int
3965xfs_log_force_umount(
3966 struct xfs_mount *mp,
3967 int logerror)
3968{
3969 struct xlog *log;
3970 int retval;
3971
3972 log = mp->m_log;
3973
3974 /*
3975 * If this happens during log recovery, don't worry about
3976 * locking; the log isn't open for business yet.
3977 */
3978 if (!log ||
3979 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3980 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3981 if (mp->m_sb_bp)
3982 mp->m_sb_bp->b_flags |= XBF_DONE;
3983 return 0;
3984 }
3985
3986 /*
3987 * Somebody could've already done the hard work for us.
3988 * No need to get locks for this.
3989 */
3990 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3991 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3992 return 1;
3993 }
3994
3995 /*
3996 * Flush all the completed transactions to disk before marking the log
3997 * being shut down. We need to do it in this order to ensure that
3998 * completed operations are safely on disk before we shut down, and that
3999 * we don't have to issue any buffer IO after the shutdown flags are set
4000 * to guarantee this.
4001 */
4002 if (!logerror)
4003 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
4004
4005 /*
4006 * mark the filesystem and the as in a shutdown state and wake
4007 * everybody up to tell them the bad news.
4008 */
4009 spin_lock(&log->l_icloglock);
4010 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4011 if (mp->m_sb_bp)
4012 mp->m_sb_bp->b_flags |= XBF_DONE;
4013
4014 /*
4015 * Mark the log and the iclogs with IO error flags to prevent any
4016 * further log IO from being issued or completed.
4017 */
4018 log->l_flags |= XLOG_IO_ERROR;
4019 retval = xlog_state_ioerror(log);
4020 spin_unlock(&log->l_icloglock);
4021
4022 /*
4023 * We don't want anybody waiting for log reservations after this. That
4024 * means we have to wake up everybody queued up on reserveq as well as
4025 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
4026 * we don't enqueue anything once the SHUTDOWN flag is set, and this
4027 * action is protected by the grant locks.
4028 */
4029 xlog_grant_head_wake_all(&log->l_reserve_head);
4030 xlog_grant_head_wake_all(&log->l_write_head);
4031
4032 /*
4033 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4034 * as if the log writes were completed. The abort handling in the log
4035 * item committed callback functions will do this again under lock to
4036 * avoid races.
4037 */
4038 wake_up_all(&log->l_cilp->xc_commit_wait);
4039 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4040
4041#ifdef XFSERRORDEBUG
4042 {
4043 xlog_in_core_t *iclog;
4044
4045 spin_lock(&log->l_icloglock);
4046 iclog = log->l_iclog;
4047 do {
4048 ASSERT(iclog->ic_callback == 0);
4049 iclog = iclog->ic_next;
4050 } while (iclog != log->l_iclog);
4051 spin_unlock(&log->l_icloglock);
4052 }
4053#endif
4054 /* return non-zero if log IOERROR transition had already happened */
4055 return retval;
4056}
4057
4058STATIC int
4059xlog_iclogs_empty(
4060 struct xlog *log)
4061{
4062 xlog_in_core_t *iclog;
4063
4064 iclog = log->l_iclog;
4065 do {
4066 /* endianness does not matter here, zero is zero in
4067 * any language.
4068 */
4069 if (iclog->ic_header.h_num_logops)
4070 return 0;
4071 iclog = iclog->ic_next;
4072 } while (iclog != log->l_iclog);
4073 return 1;
4074}
4075
4076/*
4077 * Verify that an LSN stamped into a piece of metadata is valid. This is
4078 * intended for use in read verifiers on v5 superblocks.
4079 */
4080bool
4081xfs_log_check_lsn(
4082 struct xfs_mount *mp,
4083 xfs_lsn_t lsn)
4084{
4085 struct xlog *log = mp->m_log;
4086 bool valid;
4087
4088 /*
4089 * norecovery mode skips mount-time log processing and unconditionally
4090 * resets the in-core LSN. We can't validate in this mode, but
4091 * modifications are not allowed anyways so just return true.
4092 */
4093 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4094 return true;
4095
4096 /*
4097 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4098 * handled by recovery and thus safe to ignore here.
4099 */
4100 if (lsn == NULLCOMMITLSN)
4101 return true;
4102
4103 valid = xlog_valid_lsn(mp->m_log, lsn);
4104
4105 /* warn the user about what's gone wrong before verifier failure */
4106 if (!valid) {
4107 spin_lock(&log->l_icloglock);
4108 xfs_warn(mp,
4109"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4110"Please unmount and run xfs_repair (>= v4.3) to resolve.",
4111 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4112 log->l_curr_cycle, log->l_curr_block);
4113 spin_unlock(&log->l_icloglock);
4114 }
4115
4116 return valid;
4117}