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// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_errortag.h"
14#include "xfs_error.h"
15#include "xfs_trans.h"
16#include "xfs_trans_priv.h"
17#include "xfs_log.h"
18#include "xfs_log_priv.h"
19#include "xfs_trace.h"
20#include "xfs_sysfs.h"
21#include "xfs_sb.h"
22#include "xfs_health.h"
23
24kmem_zone_t *xfs_log_ticket_zone;
25
26/* Local miscellaneous function prototypes */
27STATIC struct xlog *
28xlog_alloc_log(
29 struct xfs_mount *mp,
30 struct xfs_buftarg *log_target,
31 xfs_daddr_t blk_offset,
32 int num_bblks);
33STATIC int
34xlog_space_left(
35 struct xlog *log,
36 atomic64_t *head);
37STATIC void
38xlog_dealloc_log(
39 struct xlog *log);
40
41/* local state machine functions */
42STATIC void xlog_state_done_syncing(
43 struct xlog_in_core *iclog);
44STATIC int
45xlog_state_get_iclog_space(
46 struct xlog *log,
47 int len,
48 struct xlog_in_core **iclog,
49 struct xlog_ticket *ticket,
50 int *continued_write,
51 int *logoffsetp);
52STATIC void
53xlog_state_switch_iclogs(
54 struct xlog *log,
55 struct xlog_in_core *iclog,
56 int eventual_size);
57STATIC void
58xlog_grant_push_ail(
59 struct xlog *log,
60 int need_bytes);
61STATIC void
62xlog_sync(
63 struct xlog *log,
64 struct xlog_in_core *iclog);
65#if defined(DEBUG)
66STATIC void
67xlog_verify_dest_ptr(
68 struct xlog *log,
69 void *ptr);
70STATIC void
71xlog_verify_grant_tail(
72 struct xlog *log);
73STATIC void
74xlog_verify_iclog(
75 struct xlog *log,
76 struct xlog_in_core *iclog,
77 int count);
78STATIC void
79xlog_verify_tail_lsn(
80 struct xlog *log,
81 struct xlog_in_core *iclog,
82 xfs_lsn_t tail_lsn);
83#else
84#define xlog_verify_dest_ptr(a,b)
85#define xlog_verify_grant_tail(a)
86#define xlog_verify_iclog(a,b,c)
87#define xlog_verify_tail_lsn(a,b,c)
88#endif
89
90STATIC int
91xlog_iclogs_empty(
92 struct xlog *log);
93
94static void
95xlog_grant_sub_space(
96 struct xlog *log,
97 atomic64_t *head,
98 int bytes)
99{
100 int64_t head_val = atomic64_read(head);
101 int64_t new, old;
102
103 do {
104 int cycle, space;
105
106 xlog_crack_grant_head_val(head_val, &cycle, &space);
107
108 space -= bytes;
109 if (space < 0) {
110 space += log->l_logsize;
111 cycle--;
112 }
113
114 old = head_val;
115 new = xlog_assign_grant_head_val(cycle, space);
116 head_val = atomic64_cmpxchg(head, old, new);
117 } while (head_val != old);
118}
119
120static void
121xlog_grant_add_space(
122 struct xlog *log,
123 atomic64_t *head,
124 int bytes)
125{
126 int64_t head_val = atomic64_read(head);
127 int64_t new, old;
128
129 do {
130 int tmp;
131 int cycle, space;
132
133 xlog_crack_grant_head_val(head_val, &cycle, &space);
134
135 tmp = log->l_logsize - space;
136 if (tmp > bytes)
137 space += bytes;
138 else {
139 space = bytes - tmp;
140 cycle++;
141 }
142
143 old = head_val;
144 new = xlog_assign_grant_head_val(cycle, space);
145 head_val = atomic64_cmpxchg(head, old, new);
146 } while (head_val != old);
147}
148
149STATIC void
150xlog_grant_head_init(
151 struct xlog_grant_head *head)
152{
153 xlog_assign_grant_head(&head->grant, 1, 0);
154 INIT_LIST_HEAD(&head->waiters);
155 spin_lock_init(&head->lock);
156}
157
158STATIC void
159xlog_grant_head_wake_all(
160 struct xlog_grant_head *head)
161{
162 struct xlog_ticket *tic;
163
164 spin_lock(&head->lock);
165 list_for_each_entry(tic, &head->waiters, t_queue)
166 wake_up_process(tic->t_task);
167 spin_unlock(&head->lock);
168}
169
170static inline int
171xlog_ticket_reservation(
172 struct xlog *log,
173 struct xlog_grant_head *head,
174 struct xlog_ticket *tic)
175{
176 if (head == &log->l_write_head) {
177 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
178 return tic->t_unit_res;
179 } else {
180 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
181 return tic->t_unit_res * tic->t_cnt;
182 else
183 return tic->t_unit_res;
184 }
185}
186
187STATIC bool
188xlog_grant_head_wake(
189 struct xlog *log,
190 struct xlog_grant_head *head,
191 int *free_bytes)
192{
193 struct xlog_ticket *tic;
194 int need_bytes;
195 bool woken_task = false;
196
197 list_for_each_entry(tic, &head->waiters, t_queue) {
198
199 /*
200 * There is a chance that the size of the CIL checkpoints in
201 * progress at the last AIL push target calculation resulted in
202 * limiting the target to the log head (l_last_sync_lsn) at the
203 * time. This may not reflect where the log head is now as the
204 * CIL checkpoints may have completed.
205 *
206 * Hence when we are woken here, it may be that the head of the
207 * log that has moved rather than the tail. As the tail didn't
208 * move, there still won't be space available for the
209 * reservation we require. However, if the AIL has already
210 * pushed to the target defined by the old log head location, we
211 * will hang here waiting for something else to update the AIL
212 * push target.
213 *
214 * Therefore, if there isn't space to wake the first waiter on
215 * the grant head, we need to push the AIL again to ensure the
216 * target reflects both the current log tail and log head
217 * position before we wait for the tail to move again.
218 */
219
220 need_bytes = xlog_ticket_reservation(log, head, tic);
221 if (*free_bytes < need_bytes) {
222 if (!woken_task)
223 xlog_grant_push_ail(log, need_bytes);
224 return false;
225 }
226
227 *free_bytes -= need_bytes;
228 trace_xfs_log_grant_wake_up(log, tic);
229 wake_up_process(tic->t_task);
230 woken_task = true;
231 }
232
233 return true;
234}
235
236STATIC int
237xlog_grant_head_wait(
238 struct xlog *log,
239 struct xlog_grant_head *head,
240 struct xlog_ticket *tic,
241 int need_bytes) __releases(&head->lock)
242 __acquires(&head->lock)
243{
244 list_add_tail(&tic->t_queue, &head->waiters);
245
246 do {
247 if (XLOG_FORCED_SHUTDOWN(log))
248 goto shutdown;
249 xlog_grant_push_ail(log, need_bytes);
250
251 __set_current_state(TASK_UNINTERRUPTIBLE);
252 spin_unlock(&head->lock);
253
254 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
255
256 trace_xfs_log_grant_sleep(log, tic);
257 schedule();
258 trace_xfs_log_grant_wake(log, tic);
259
260 spin_lock(&head->lock);
261 if (XLOG_FORCED_SHUTDOWN(log))
262 goto shutdown;
263 } while (xlog_space_left(log, &head->grant) < need_bytes);
264
265 list_del_init(&tic->t_queue);
266 return 0;
267shutdown:
268 list_del_init(&tic->t_queue);
269 return -EIO;
270}
271
272/*
273 * Atomically get the log space required for a log ticket.
274 *
275 * Once a ticket gets put onto head->waiters, it will only return after the
276 * needed reservation is satisfied.
277 *
278 * This function is structured so that it has a lock free fast path. This is
279 * necessary because every new transaction reservation will come through this
280 * path. Hence any lock will be globally hot if we take it unconditionally on
281 * every pass.
282 *
283 * As tickets are only ever moved on and off head->waiters under head->lock, we
284 * only need to take that lock if we are going to add the ticket to the queue
285 * and sleep. We can avoid taking the lock if the ticket was never added to
286 * head->waiters because the t_queue list head will be empty and we hold the
287 * only reference to it so it can safely be checked unlocked.
288 */
289STATIC int
290xlog_grant_head_check(
291 struct xlog *log,
292 struct xlog_grant_head *head,
293 struct xlog_ticket *tic,
294 int *need_bytes)
295{
296 int free_bytes;
297 int error = 0;
298
299 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
300
301 /*
302 * If there are other waiters on the queue then give them a chance at
303 * logspace before us. Wake up the first waiters, if we do not wake
304 * up all the waiters then go to sleep waiting for more free space,
305 * otherwise try to get some space for this transaction.
306 */
307 *need_bytes = xlog_ticket_reservation(log, head, tic);
308 free_bytes = xlog_space_left(log, &head->grant);
309 if (!list_empty_careful(&head->waiters)) {
310 spin_lock(&head->lock);
311 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
312 free_bytes < *need_bytes) {
313 error = xlog_grant_head_wait(log, head, tic,
314 *need_bytes);
315 }
316 spin_unlock(&head->lock);
317 } else if (free_bytes < *need_bytes) {
318 spin_lock(&head->lock);
319 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
320 spin_unlock(&head->lock);
321 }
322
323 return error;
324}
325
326static void
327xlog_tic_reset_res(xlog_ticket_t *tic)
328{
329 tic->t_res_num = 0;
330 tic->t_res_arr_sum = 0;
331 tic->t_res_num_ophdrs = 0;
332}
333
334static void
335xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
336{
337 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
338 /* add to overflow and start again */
339 tic->t_res_o_flow += tic->t_res_arr_sum;
340 tic->t_res_num = 0;
341 tic->t_res_arr_sum = 0;
342 }
343
344 tic->t_res_arr[tic->t_res_num].r_len = len;
345 tic->t_res_arr[tic->t_res_num].r_type = type;
346 tic->t_res_arr_sum += len;
347 tic->t_res_num++;
348}
349
350/*
351 * Replenish the byte reservation required by moving the grant write head.
352 */
353int
354xfs_log_regrant(
355 struct xfs_mount *mp,
356 struct xlog_ticket *tic)
357{
358 struct xlog *log = mp->m_log;
359 int need_bytes;
360 int error = 0;
361
362 if (XLOG_FORCED_SHUTDOWN(log))
363 return -EIO;
364
365 XFS_STATS_INC(mp, xs_try_logspace);
366
367 /*
368 * This is a new transaction on the ticket, so we need to change the
369 * transaction ID so that the next transaction has a different TID in
370 * the log. Just add one to the existing tid so that we can see chains
371 * of rolling transactions in the log easily.
372 */
373 tic->t_tid++;
374
375 xlog_grant_push_ail(log, tic->t_unit_res);
376
377 tic->t_curr_res = tic->t_unit_res;
378 xlog_tic_reset_res(tic);
379
380 if (tic->t_cnt > 0)
381 return 0;
382
383 trace_xfs_log_regrant(log, tic);
384
385 error = xlog_grant_head_check(log, &log->l_write_head, tic,
386 &need_bytes);
387 if (error)
388 goto out_error;
389
390 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
391 trace_xfs_log_regrant_exit(log, tic);
392 xlog_verify_grant_tail(log);
393 return 0;
394
395out_error:
396 /*
397 * If we are failing, make sure the ticket doesn't have any current
398 * reservations. We don't want to add this back when the ticket/
399 * transaction gets cancelled.
400 */
401 tic->t_curr_res = 0;
402 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
403 return error;
404}
405
406/*
407 * Reserve log space and return a ticket corresponding to the reservation.
408 *
409 * Each reservation is going to reserve extra space for a log record header.
410 * When writes happen to the on-disk log, we don't subtract the length of the
411 * log record header from any reservation. By wasting space in each
412 * reservation, we prevent over allocation problems.
413 */
414int
415xfs_log_reserve(
416 struct xfs_mount *mp,
417 int unit_bytes,
418 int cnt,
419 struct xlog_ticket **ticp,
420 uint8_t client,
421 bool permanent)
422{
423 struct xlog *log = mp->m_log;
424 struct xlog_ticket *tic;
425 int need_bytes;
426 int error = 0;
427
428 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
429
430 if (XLOG_FORCED_SHUTDOWN(log))
431 return -EIO;
432
433 XFS_STATS_INC(mp, xs_try_logspace);
434
435 ASSERT(*ticp == NULL);
436 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
437 *ticp = tic;
438
439 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
440 : tic->t_unit_res);
441
442 trace_xfs_log_reserve(log, tic);
443
444 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
445 &need_bytes);
446 if (error)
447 goto out_error;
448
449 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
450 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
451 trace_xfs_log_reserve_exit(log, tic);
452 xlog_verify_grant_tail(log);
453 return 0;
454
455out_error:
456 /*
457 * If we are failing, make sure the ticket doesn't have any current
458 * reservations. We don't want to add this back when the ticket/
459 * transaction gets cancelled.
460 */
461 tic->t_curr_res = 0;
462 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
463 return error;
464}
465
466static bool
467__xlog_state_release_iclog(
468 struct xlog *log,
469 struct xlog_in_core *iclog)
470{
471 lockdep_assert_held(&log->l_icloglock);
472
473 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
474 /* update tail before writing to iclog */
475 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
476
477 iclog->ic_state = XLOG_STATE_SYNCING;
478 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
479 xlog_verify_tail_lsn(log, iclog, tail_lsn);
480 /* cycle incremented when incrementing curr_block */
481 return true;
482 }
483
484 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
485 return false;
486}
487
488/*
489 * Flush iclog to disk if this is the last reference to the given iclog and the
490 * it is in the WANT_SYNC state.
491 */
492static int
493xlog_state_release_iclog(
494 struct xlog *log,
495 struct xlog_in_core *iclog)
496{
497 lockdep_assert_held(&log->l_icloglock);
498
499 if (iclog->ic_state == XLOG_STATE_IOERROR)
500 return -EIO;
501
502 if (atomic_dec_and_test(&iclog->ic_refcnt) &&
503 __xlog_state_release_iclog(log, iclog)) {
504 spin_unlock(&log->l_icloglock);
505 xlog_sync(log, iclog);
506 spin_lock(&log->l_icloglock);
507 }
508
509 return 0;
510}
511
512void
513xfs_log_release_iclog(
514 struct xlog_in_core *iclog)
515{
516 struct xlog *log = iclog->ic_log;
517 bool sync = false;
518
519 if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
520 if (iclog->ic_state != XLOG_STATE_IOERROR)
521 sync = __xlog_state_release_iclog(log, iclog);
522 spin_unlock(&log->l_icloglock);
523 }
524
525 if (sync)
526 xlog_sync(log, iclog);
527}
528
529/*
530 * Mount a log filesystem
531 *
532 * mp - ubiquitous xfs mount point structure
533 * log_target - buftarg of on-disk log device
534 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
535 * num_bblocks - Number of BBSIZE blocks in on-disk log
536 *
537 * Return error or zero.
538 */
539int
540xfs_log_mount(
541 xfs_mount_t *mp,
542 xfs_buftarg_t *log_target,
543 xfs_daddr_t blk_offset,
544 int num_bblks)
545{
546 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
547 int error = 0;
548 int min_logfsbs;
549
550 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
551 xfs_notice(mp, "Mounting V%d Filesystem",
552 XFS_SB_VERSION_NUM(&mp->m_sb));
553 } else {
554 xfs_notice(mp,
555"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
556 XFS_SB_VERSION_NUM(&mp->m_sb));
557 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
558 }
559
560 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
561 if (IS_ERR(mp->m_log)) {
562 error = PTR_ERR(mp->m_log);
563 goto out;
564 }
565
566 /*
567 * Validate the given log space and drop a critical message via syslog
568 * if the log size is too small that would lead to some unexpected
569 * situations in transaction log space reservation stage.
570 *
571 * Note: we can't just reject the mount if the validation fails. This
572 * would mean that people would have to downgrade their kernel just to
573 * remedy the situation as there is no way to grow the log (short of
574 * black magic surgery with xfs_db).
575 *
576 * We can, however, reject mounts for CRC format filesystems, as the
577 * mkfs binary being used to make the filesystem should never create a
578 * filesystem with a log that is too small.
579 */
580 min_logfsbs = xfs_log_calc_minimum_size(mp);
581
582 if (mp->m_sb.sb_logblocks < min_logfsbs) {
583 xfs_warn(mp,
584 "Log size %d blocks too small, minimum size is %d blocks",
585 mp->m_sb.sb_logblocks, min_logfsbs);
586 error = -EINVAL;
587 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
588 xfs_warn(mp,
589 "Log size %d blocks too large, maximum size is %lld blocks",
590 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
591 error = -EINVAL;
592 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
593 xfs_warn(mp,
594 "log size %lld bytes too large, maximum size is %lld bytes",
595 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
596 XFS_MAX_LOG_BYTES);
597 error = -EINVAL;
598 } else if (mp->m_sb.sb_logsunit > 1 &&
599 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
600 xfs_warn(mp,
601 "log stripe unit %u bytes must be a multiple of block size",
602 mp->m_sb.sb_logsunit);
603 error = -EINVAL;
604 fatal = true;
605 }
606 if (error) {
607 /*
608 * Log check errors are always fatal on v5; or whenever bad
609 * metadata leads to a crash.
610 */
611 if (fatal) {
612 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
613 ASSERT(0);
614 goto out_free_log;
615 }
616 xfs_crit(mp, "Log size out of supported range.");
617 xfs_crit(mp,
618"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
619 }
620
621 /*
622 * Initialize the AIL now we have a log.
623 */
624 error = xfs_trans_ail_init(mp);
625 if (error) {
626 xfs_warn(mp, "AIL initialisation failed: error %d", error);
627 goto out_free_log;
628 }
629 mp->m_log->l_ailp = mp->m_ail;
630
631 /*
632 * skip log recovery on a norecovery mount. pretend it all
633 * just worked.
634 */
635 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
636 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
637
638 if (readonly)
639 mp->m_flags &= ~XFS_MOUNT_RDONLY;
640
641 error = xlog_recover(mp->m_log);
642
643 if (readonly)
644 mp->m_flags |= XFS_MOUNT_RDONLY;
645 if (error) {
646 xfs_warn(mp, "log mount/recovery failed: error %d",
647 error);
648 xlog_recover_cancel(mp->m_log);
649 goto out_destroy_ail;
650 }
651 }
652
653 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
654 "log");
655 if (error)
656 goto out_destroy_ail;
657
658 /* Normal transactions can now occur */
659 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
660
661 /*
662 * Now the log has been fully initialised and we know were our
663 * space grant counters are, we can initialise the permanent ticket
664 * needed for delayed logging to work.
665 */
666 xlog_cil_init_post_recovery(mp->m_log);
667
668 return 0;
669
670out_destroy_ail:
671 xfs_trans_ail_destroy(mp);
672out_free_log:
673 xlog_dealloc_log(mp->m_log);
674out:
675 return error;
676}
677
678/*
679 * Finish the recovery of the file system. This is separate from the
680 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
681 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
682 * here.
683 *
684 * If we finish recovery successfully, start the background log work. If we are
685 * not doing recovery, then we have a RO filesystem and we don't need to start
686 * it.
687 */
688int
689xfs_log_mount_finish(
690 struct xfs_mount *mp)
691{
692 int error = 0;
693 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
694 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
695
696 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
697 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
698 return 0;
699 } else if (readonly) {
700 /* Allow unlinked processing to proceed */
701 mp->m_flags &= ~XFS_MOUNT_RDONLY;
702 }
703
704 /*
705 * During the second phase of log recovery, we need iget and
706 * iput to behave like they do for an active filesystem.
707 * xfs_fs_drop_inode needs to be able to prevent the deletion
708 * of inodes before we're done replaying log items on those
709 * inodes. Turn it off immediately after recovery finishes
710 * so that we don't leak the quota inodes if subsequent mount
711 * activities fail.
712 *
713 * We let all inodes involved in redo item processing end up on
714 * the LRU instead of being evicted immediately so that if we do
715 * something to an unlinked inode, the irele won't cause
716 * premature truncation and freeing of the inode, which results
717 * in log recovery failure. We have to evict the unreferenced
718 * lru inodes after clearing SB_ACTIVE because we don't
719 * otherwise clean up the lru if there's a subsequent failure in
720 * xfs_mountfs, which leads to us leaking the inodes if nothing
721 * else (e.g. quotacheck) references the inodes before the
722 * mount failure occurs.
723 */
724 mp->m_super->s_flags |= SB_ACTIVE;
725 error = xlog_recover_finish(mp->m_log);
726 if (!error)
727 xfs_log_work_queue(mp);
728 mp->m_super->s_flags &= ~SB_ACTIVE;
729 evict_inodes(mp->m_super);
730
731 /*
732 * Drain the buffer LRU after log recovery. This is required for v4
733 * filesystems to avoid leaving around buffers with NULL verifier ops,
734 * but we do it unconditionally to make sure we're always in a clean
735 * cache state after mount.
736 *
737 * Don't push in the error case because the AIL may have pending intents
738 * that aren't removed until recovery is cancelled.
739 */
740 if (!error && recovered) {
741 xfs_log_force(mp, XFS_LOG_SYNC);
742 xfs_ail_push_all_sync(mp->m_ail);
743 }
744 xfs_wait_buftarg(mp->m_ddev_targp);
745
746 if (readonly)
747 mp->m_flags |= XFS_MOUNT_RDONLY;
748
749 return error;
750}
751
752/*
753 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
754 * the log.
755 */
756void
757xfs_log_mount_cancel(
758 struct xfs_mount *mp)
759{
760 xlog_recover_cancel(mp->m_log);
761 xfs_log_unmount(mp);
762}
763
764/*
765 * Wait for the iclog to be written disk, or return an error if the log has been
766 * shut down.
767 */
768static int
769xlog_wait_on_iclog(
770 struct xlog_in_core *iclog)
771 __releases(iclog->ic_log->l_icloglock)
772{
773 struct xlog *log = iclog->ic_log;
774
775 if (!XLOG_FORCED_SHUTDOWN(log) &&
776 iclog->ic_state != XLOG_STATE_ACTIVE &&
777 iclog->ic_state != XLOG_STATE_DIRTY) {
778 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
779 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
780 } else {
781 spin_unlock(&log->l_icloglock);
782 }
783
784 if (XLOG_FORCED_SHUTDOWN(log))
785 return -EIO;
786 return 0;
787}
788
789/*
790 * Write out an unmount record using the ticket provided. We have to account for
791 * the data space used in the unmount ticket as this write is not done from a
792 * transaction context that has already done the accounting for us.
793 */
794static int
795xlog_write_unmount_record(
796 struct xlog *log,
797 struct xlog_ticket *ticket,
798 xfs_lsn_t *lsn,
799 uint flags)
800{
801 struct xfs_unmount_log_format ulf = {
802 .magic = XLOG_UNMOUNT_TYPE,
803 };
804 struct xfs_log_iovec reg = {
805 .i_addr = &ulf,
806 .i_len = sizeof(ulf),
807 .i_type = XLOG_REG_TYPE_UNMOUNT,
808 };
809 struct xfs_log_vec vec = {
810 .lv_niovecs = 1,
811 .lv_iovecp = ®,
812 };
813
814 /* account for space used by record data */
815 ticket->t_curr_res -= sizeof(ulf);
816 return xlog_write(log, &vec, ticket, lsn, NULL, flags, false);
817}
818
819/*
820 * Mark the filesystem clean by writing an unmount record to the head of the
821 * log.
822 */
823static void
824xlog_unmount_write(
825 struct xlog *log)
826{
827 struct xfs_mount *mp = log->l_mp;
828 struct xlog_in_core *iclog;
829 struct xlog_ticket *tic = NULL;
830 xfs_lsn_t lsn;
831 uint flags = XLOG_UNMOUNT_TRANS;
832 int error;
833
834 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
835 if (error)
836 goto out_err;
837
838 error = xlog_write_unmount_record(log, tic, &lsn, flags);
839 /*
840 * At this point, we're umounting anyway, so there's no point in
841 * transitioning log state to IOERROR. Just continue...
842 */
843out_err:
844 if (error)
845 xfs_alert(mp, "%s: unmount record failed", __func__);
846
847 spin_lock(&log->l_icloglock);
848 iclog = log->l_iclog;
849 atomic_inc(&iclog->ic_refcnt);
850 if (iclog->ic_state == XLOG_STATE_ACTIVE)
851 xlog_state_switch_iclogs(log, iclog, 0);
852 else
853 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
854 iclog->ic_state == XLOG_STATE_IOERROR);
855 error = xlog_state_release_iclog(log, iclog);
856 xlog_wait_on_iclog(iclog);
857
858 if (tic) {
859 trace_xfs_log_umount_write(log, tic);
860 xfs_log_ticket_ungrant(log, tic);
861 }
862}
863
864static void
865xfs_log_unmount_verify_iclog(
866 struct xlog *log)
867{
868 struct xlog_in_core *iclog = log->l_iclog;
869
870 do {
871 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
872 ASSERT(iclog->ic_offset == 0);
873 } while ((iclog = iclog->ic_next) != log->l_iclog);
874}
875
876/*
877 * Unmount record used to have a string "Unmount filesystem--" in the
878 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
879 * We just write the magic number now since that particular field isn't
880 * currently architecture converted and "Unmount" is a bit foo.
881 * As far as I know, there weren't any dependencies on the old behaviour.
882 */
883static void
884xfs_log_unmount_write(
885 struct xfs_mount *mp)
886{
887 struct xlog *log = mp->m_log;
888
889 /*
890 * Don't write out unmount record on norecovery mounts or ro devices.
891 * Or, if we are doing a forced umount (typically because of IO errors).
892 */
893 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
894 xfs_readonly_buftarg(log->l_targ)) {
895 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
896 return;
897 }
898
899 xfs_log_force(mp, XFS_LOG_SYNC);
900
901 if (XLOG_FORCED_SHUTDOWN(log))
902 return;
903
904 /*
905 * If we think the summary counters are bad, avoid writing the unmount
906 * record to force log recovery at next mount, after which the summary
907 * counters will be recalculated. Refer to xlog_check_unmount_rec for
908 * more details.
909 */
910 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
911 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
912 xfs_alert(mp, "%s: will fix summary counters at next mount",
913 __func__);
914 return;
915 }
916
917 xfs_log_unmount_verify_iclog(log);
918 xlog_unmount_write(log);
919}
920
921/*
922 * Empty the log for unmount/freeze.
923 *
924 * To do this, we first need to shut down the background log work so it is not
925 * trying to cover the log as we clean up. We then need to unpin all objects in
926 * the log so we can then flush them out. Once they have completed their IO and
927 * run the callbacks removing themselves from the AIL, we can write the unmount
928 * record.
929 */
930void
931xfs_log_quiesce(
932 struct xfs_mount *mp)
933{
934 cancel_delayed_work_sync(&mp->m_log->l_work);
935 xfs_log_force(mp, XFS_LOG_SYNC);
936
937 /*
938 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
939 * will push it, xfs_wait_buftarg() will not wait for it. Further,
940 * xfs_buf_iowait() cannot be used because it was pushed with the
941 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
942 * the IO to complete.
943 */
944 xfs_ail_push_all_sync(mp->m_ail);
945 xfs_wait_buftarg(mp->m_ddev_targp);
946 xfs_buf_lock(mp->m_sb_bp);
947 xfs_buf_unlock(mp->m_sb_bp);
948
949 xfs_log_unmount_write(mp);
950}
951
952/*
953 * Shut down and release the AIL and Log.
954 *
955 * During unmount, we need to ensure we flush all the dirty metadata objects
956 * from the AIL so that the log is empty before we write the unmount record to
957 * the log. Once this is done, we can tear down the AIL and the log.
958 */
959void
960xfs_log_unmount(
961 struct xfs_mount *mp)
962{
963 xfs_log_quiesce(mp);
964
965 xfs_trans_ail_destroy(mp);
966
967 xfs_sysfs_del(&mp->m_log->l_kobj);
968
969 xlog_dealloc_log(mp->m_log);
970}
971
972void
973xfs_log_item_init(
974 struct xfs_mount *mp,
975 struct xfs_log_item *item,
976 int type,
977 const struct xfs_item_ops *ops)
978{
979 item->li_mountp = mp;
980 item->li_ailp = mp->m_ail;
981 item->li_type = type;
982 item->li_ops = ops;
983 item->li_lv = NULL;
984
985 INIT_LIST_HEAD(&item->li_ail);
986 INIT_LIST_HEAD(&item->li_cil);
987 INIT_LIST_HEAD(&item->li_bio_list);
988 INIT_LIST_HEAD(&item->li_trans);
989}
990
991/*
992 * Wake up processes waiting for log space after we have moved the log tail.
993 */
994void
995xfs_log_space_wake(
996 struct xfs_mount *mp)
997{
998 struct xlog *log = mp->m_log;
999 int free_bytes;
1000
1001 if (XLOG_FORCED_SHUTDOWN(log))
1002 return;
1003
1004 if (!list_empty_careful(&log->l_write_head.waiters)) {
1005 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1006
1007 spin_lock(&log->l_write_head.lock);
1008 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1009 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1010 spin_unlock(&log->l_write_head.lock);
1011 }
1012
1013 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1014 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1015
1016 spin_lock(&log->l_reserve_head.lock);
1017 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1018 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1019 spin_unlock(&log->l_reserve_head.lock);
1020 }
1021}
1022
1023/*
1024 * Determine if we have a transaction that has gone to disk that needs to be
1025 * covered. To begin the transition to the idle state firstly the log needs to
1026 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1027 * we start attempting to cover the log.
1028 *
1029 * Only if we are then in a state where covering is needed, the caller is
1030 * informed that dummy transactions are required to move the log into the idle
1031 * state.
1032 *
1033 * If there are any items in the AIl or CIL, then we do not want to attempt to
1034 * cover the log as we may be in a situation where there isn't log space
1035 * available to run a dummy transaction and this can lead to deadlocks when the
1036 * tail of the log is pinned by an item that is modified in the CIL. Hence
1037 * there's no point in running a dummy transaction at this point because we
1038 * can't start trying to idle the log until both the CIL and AIL are empty.
1039 */
1040static int
1041xfs_log_need_covered(xfs_mount_t *mp)
1042{
1043 struct xlog *log = mp->m_log;
1044 int needed = 0;
1045
1046 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1047 return 0;
1048
1049 if (!xlog_cil_empty(log))
1050 return 0;
1051
1052 spin_lock(&log->l_icloglock);
1053 switch (log->l_covered_state) {
1054 case XLOG_STATE_COVER_DONE:
1055 case XLOG_STATE_COVER_DONE2:
1056 case XLOG_STATE_COVER_IDLE:
1057 break;
1058 case XLOG_STATE_COVER_NEED:
1059 case XLOG_STATE_COVER_NEED2:
1060 if (xfs_ail_min_lsn(log->l_ailp))
1061 break;
1062 if (!xlog_iclogs_empty(log))
1063 break;
1064
1065 needed = 1;
1066 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1067 log->l_covered_state = XLOG_STATE_COVER_DONE;
1068 else
1069 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1070 break;
1071 default:
1072 needed = 1;
1073 break;
1074 }
1075 spin_unlock(&log->l_icloglock);
1076 return needed;
1077}
1078
1079/*
1080 * We may be holding the log iclog lock upon entering this routine.
1081 */
1082xfs_lsn_t
1083xlog_assign_tail_lsn_locked(
1084 struct xfs_mount *mp)
1085{
1086 struct xlog *log = mp->m_log;
1087 struct xfs_log_item *lip;
1088 xfs_lsn_t tail_lsn;
1089
1090 assert_spin_locked(&mp->m_ail->ail_lock);
1091
1092 /*
1093 * To make sure we always have a valid LSN for the log tail we keep
1094 * track of the last LSN which was committed in log->l_last_sync_lsn,
1095 * and use that when the AIL was empty.
1096 */
1097 lip = xfs_ail_min(mp->m_ail);
1098 if (lip)
1099 tail_lsn = lip->li_lsn;
1100 else
1101 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1102 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1103 atomic64_set(&log->l_tail_lsn, tail_lsn);
1104 return tail_lsn;
1105}
1106
1107xfs_lsn_t
1108xlog_assign_tail_lsn(
1109 struct xfs_mount *mp)
1110{
1111 xfs_lsn_t tail_lsn;
1112
1113 spin_lock(&mp->m_ail->ail_lock);
1114 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1115 spin_unlock(&mp->m_ail->ail_lock);
1116
1117 return tail_lsn;
1118}
1119
1120/*
1121 * Return the space in the log between the tail and the head. The head
1122 * is passed in the cycle/bytes formal parms. In the special case where
1123 * the reserve head has wrapped passed the tail, this calculation is no
1124 * longer valid. In this case, just return 0 which means there is no space
1125 * in the log. This works for all places where this function is called
1126 * with the reserve head. Of course, if the write head were to ever
1127 * wrap the tail, we should blow up. Rather than catch this case here,
1128 * we depend on other ASSERTions in other parts of the code. XXXmiken
1129 *
1130 * This code also handles the case where the reservation head is behind
1131 * the tail. The details of this case are described below, but the end
1132 * result is that we return the size of the log as the amount of space left.
1133 */
1134STATIC int
1135xlog_space_left(
1136 struct xlog *log,
1137 atomic64_t *head)
1138{
1139 int free_bytes;
1140 int tail_bytes;
1141 int tail_cycle;
1142 int head_cycle;
1143 int head_bytes;
1144
1145 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1146 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1147 tail_bytes = BBTOB(tail_bytes);
1148 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1149 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1150 else if (tail_cycle + 1 < head_cycle)
1151 return 0;
1152 else if (tail_cycle < head_cycle) {
1153 ASSERT(tail_cycle == (head_cycle - 1));
1154 free_bytes = tail_bytes - head_bytes;
1155 } else {
1156 /*
1157 * The reservation head is behind the tail.
1158 * In this case we just want to return the size of the
1159 * log as the amount of space left.
1160 */
1161 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1162 xfs_alert(log->l_mp,
1163 " tail_cycle = %d, tail_bytes = %d",
1164 tail_cycle, tail_bytes);
1165 xfs_alert(log->l_mp,
1166 " GH cycle = %d, GH bytes = %d",
1167 head_cycle, head_bytes);
1168 ASSERT(0);
1169 free_bytes = log->l_logsize;
1170 }
1171 return free_bytes;
1172}
1173
1174
1175static void
1176xlog_ioend_work(
1177 struct work_struct *work)
1178{
1179 struct xlog_in_core *iclog =
1180 container_of(work, struct xlog_in_core, ic_end_io_work);
1181 struct xlog *log = iclog->ic_log;
1182 int error;
1183
1184 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1185#ifdef DEBUG
1186 /* treat writes with injected CRC errors as failed */
1187 if (iclog->ic_fail_crc)
1188 error = -EIO;
1189#endif
1190
1191 /*
1192 * Race to shutdown the filesystem if we see an error.
1193 */
1194 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1195 xfs_alert(log->l_mp, "log I/O error %d", error);
1196 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1197 }
1198
1199 xlog_state_done_syncing(iclog);
1200 bio_uninit(&iclog->ic_bio);
1201
1202 /*
1203 * Drop the lock to signal that we are done. Nothing references the
1204 * iclog after this, so an unmount waiting on this lock can now tear it
1205 * down safely. As such, it is unsafe to reference the iclog after the
1206 * unlock as we could race with it being freed.
1207 */
1208 up(&iclog->ic_sema);
1209}
1210
1211/*
1212 * Return size of each in-core log record buffer.
1213 *
1214 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1215 *
1216 * If the filesystem blocksize is too large, we may need to choose a
1217 * larger size since the directory code currently logs entire blocks.
1218 */
1219STATIC void
1220xlog_get_iclog_buffer_size(
1221 struct xfs_mount *mp,
1222 struct xlog *log)
1223{
1224 if (mp->m_logbufs <= 0)
1225 mp->m_logbufs = XLOG_MAX_ICLOGS;
1226 if (mp->m_logbsize <= 0)
1227 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1228
1229 log->l_iclog_bufs = mp->m_logbufs;
1230 log->l_iclog_size = mp->m_logbsize;
1231
1232 /*
1233 * # headers = size / 32k - one header holds cycles from 32k of data.
1234 */
1235 log->l_iclog_heads =
1236 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1237 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1238}
1239
1240void
1241xfs_log_work_queue(
1242 struct xfs_mount *mp)
1243{
1244 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1245 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1246}
1247
1248/*
1249 * Every sync period we need to unpin all items in the AIL and push them to
1250 * disk. If there is nothing dirty, then we might need to cover the log to
1251 * indicate that the filesystem is idle.
1252 */
1253static void
1254xfs_log_worker(
1255 struct work_struct *work)
1256{
1257 struct xlog *log = container_of(to_delayed_work(work),
1258 struct xlog, l_work);
1259 struct xfs_mount *mp = log->l_mp;
1260
1261 /* dgc: errors ignored - not fatal and nowhere to report them */
1262 if (xfs_log_need_covered(mp)) {
1263 /*
1264 * Dump a transaction into the log that contains no real change.
1265 * This is needed to stamp the current tail LSN into the log
1266 * during the covering operation.
1267 *
1268 * We cannot use an inode here for this - that will push dirty
1269 * state back up into the VFS and then periodic inode flushing
1270 * will prevent log covering from making progress. Hence we
1271 * synchronously log the superblock instead to ensure the
1272 * superblock is immediately unpinned and can be written back.
1273 */
1274 xfs_sync_sb(mp, true);
1275 } else
1276 xfs_log_force(mp, 0);
1277
1278 /* start pushing all the metadata that is currently dirty */
1279 xfs_ail_push_all(mp->m_ail);
1280
1281 /* queue us up again */
1282 xfs_log_work_queue(mp);
1283}
1284
1285/*
1286 * This routine initializes some of the log structure for a given mount point.
1287 * Its primary purpose is to fill in enough, so recovery can occur. However,
1288 * some other stuff may be filled in too.
1289 */
1290STATIC struct xlog *
1291xlog_alloc_log(
1292 struct xfs_mount *mp,
1293 struct xfs_buftarg *log_target,
1294 xfs_daddr_t blk_offset,
1295 int num_bblks)
1296{
1297 struct xlog *log;
1298 xlog_rec_header_t *head;
1299 xlog_in_core_t **iclogp;
1300 xlog_in_core_t *iclog, *prev_iclog=NULL;
1301 int i;
1302 int error = -ENOMEM;
1303 uint log2_size = 0;
1304
1305 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1306 if (!log) {
1307 xfs_warn(mp, "Log allocation failed: No memory!");
1308 goto out;
1309 }
1310
1311 log->l_mp = mp;
1312 log->l_targ = log_target;
1313 log->l_logsize = BBTOB(num_bblks);
1314 log->l_logBBstart = blk_offset;
1315 log->l_logBBsize = num_bblks;
1316 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1317 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1318 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1319
1320 log->l_prev_block = -1;
1321 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1322 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1323 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1324 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1325
1326 xlog_grant_head_init(&log->l_reserve_head);
1327 xlog_grant_head_init(&log->l_write_head);
1328
1329 error = -EFSCORRUPTED;
1330 if (xfs_sb_version_hassector(&mp->m_sb)) {
1331 log2_size = mp->m_sb.sb_logsectlog;
1332 if (log2_size < BBSHIFT) {
1333 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1334 log2_size, BBSHIFT);
1335 goto out_free_log;
1336 }
1337
1338 log2_size -= BBSHIFT;
1339 if (log2_size > mp->m_sectbb_log) {
1340 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1341 log2_size, mp->m_sectbb_log);
1342 goto out_free_log;
1343 }
1344
1345 /* for larger sector sizes, must have v2 or external log */
1346 if (log2_size && log->l_logBBstart > 0 &&
1347 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1348 xfs_warn(mp,
1349 "log sector size (0x%x) invalid for configuration.",
1350 log2_size);
1351 goto out_free_log;
1352 }
1353 }
1354 log->l_sectBBsize = 1 << log2_size;
1355
1356 xlog_get_iclog_buffer_size(mp, log);
1357
1358 spin_lock_init(&log->l_icloglock);
1359 init_waitqueue_head(&log->l_flush_wait);
1360
1361 iclogp = &log->l_iclog;
1362 /*
1363 * The amount of memory to allocate for the iclog structure is
1364 * rather funky due to the way the structure is defined. It is
1365 * done this way so that we can use different sizes for machines
1366 * with different amounts of memory. See the definition of
1367 * xlog_in_core_t in xfs_log_priv.h for details.
1368 */
1369 ASSERT(log->l_iclog_size >= 4096);
1370 for (i = 0; i < log->l_iclog_bufs; i++) {
1371 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1372 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1373 sizeof(struct bio_vec);
1374
1375 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1376 if (!iclog)
1377 goto out_free_iclog;
1378
1379 *iclogp = iclog;
1380 iclog->ic_prev = prev_iclog;
1381 prev_iclog = iclog;
1382
1383 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1384 KM_MAYFAIL | KM_ZERO);
1385 if (!iclog->ic_data)
1386 goto out_free_iclog;
1387#ifdef DEBUG
1388 log->l_iclog_bak[i] = &iclog->ic_header;
1389#endif
1390 head = &iclog->ic_header;
1391 memset(head, 0, sizeof(xlog_rec_header_t));
1392 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1393 head->h_version = cpu_to_be32(
1394 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1395 head->h_size = cpu_to_be32(log->l_iclog_size);
1396 /* new fields */
1397 head->h_fmt = cpu_to_be32(XLOG_FMT);
1398 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1399
1400 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1401 iclog->ic_state = XLOG_STATE_ACTIVE;
1402 iclog->ic_log = log;
1403 atomic_set(&iclog->ic_refcnt, 0);
1404 spin_lock_init(&iclog->ic_callback_lock);
1405 INIT_LIST_HEAD(&iclog->ic_callbacks);
1406 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1407
1408 init_waitqueue_head(&iclog->ic_force_wait);
1409 init_waitqueue_head(&iclog->ic_write_wait);
1410 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1411 sema_init(&iclog->ic_sema, 1);
1412
1413 iclogp = &iclog->ic_next;
1414 }
1415 *iclogp = log->l_iclog; /* complete ring */
1416 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1417
1418 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1419 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1420 mp->m_super->s_id);
1421 if (!log->l_ioend_workqueue)
1422 goto out_free_iclog;
1423
1424 error = xlog_cil_init(log);
1425 if (error)
1426 goto out_destroy_workqueue;
1427 return log;
1428
1429out_destroy_workqueue:
1430 destroy_workqueue(log->l_ioend_workqueue);
1431out_free_iclog:
1432 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1433 prev_iclog = iclog->ic_next;
1434 kmem_free(iclog->ic_data);
1435 kmem_free(iclog);
1436 if (prev_iclog == log->l_iclog)
1437 break;
1438 }
1439out_free_log:
1440 kmem_free(log);
1441out:
1442 return ERR_PTR(error);
1443} /* xlog_alloc_log */
1444
1445/*
1446 * Write out the commit record of a transaction associated with the given
1447 * ticket to close off a running log write. Return the lsn of the commit record.
1448 */
1449int
1450xlog_commit_record(
1451 struct xlog *log,
1452 struct xlog_ticket *ticket,
1453 struct xlog_in_core **iclog,
1454 xfs_lsn_t *lsn)
1455{
1456 struct xfs_log_iovec reg = {
1457 .i_addr = NULL,
1458 .i_len = 0,
1459 .i_type = XLOG_REG_TYPE_COMMIT,
1460 };
1461 struct xfs_log_vec vec = {
1462 .lv_niovecs = 1,
1463 .lv_iovecp = ®,
1464 };
1465 int error;
1466
1467 if (XLOG_FORCED_SHUTDOWN(log))
1468 return -EIO;
1469
1470 error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS,
1471 false);
1472 if (error)
1473 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1474 return error;
1475}
1476
1477/*
1478 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1479 * log space. This code pushes on the lsn which would supposedly free up
1480 * the 25% which we want to leave free. We may need to adopt a policy which
1481 * pushes on an lsn which is further along in the log once we reach the high
1482 * water mark. In this manner, we would be creating a low water mark.
1483 */
1484STATIC void
1485xlog_grant_push_ail(
1486 struct xlog *log,
1487 int need_bytes)
1488{
1489 xfs_lsn_t threshold_lsn = 0;
1490 xfs_lsn_t last_sync_lsn;
1491 int free_blocks;
1492 int free_bytes;
1493 int threshold_block;
1494 int threshold_cycle;
1495 int free_threshold;
1496
1497 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1498
1499 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1500 free_blocks = BTOBBT(free_bytes);
1501
1502 /*
1503 * Set the threshold for the minimum number of free blocks in the
1504 * log to the maximum of what the caller needs, one quarter of the
1505 * log, and 256 blocks.
1506 */
1507 free_threshold = BTOBB(need_bytes);
1508 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1509 free_threshold = max(free_threshold, 256);
1510 if (free_blocks >= free_threshold)
1511 return;
1512
1513 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1514 &threshold_block);
1515 threshold_block += free_threshold;
1516 if (threshold_block >= log->l_logBBsize) {
1517 threshold_block -= log->l_logBBsize;
1518 threshold_cycle += 1;
1519 }
1520 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1521 threshold_block);
1522 /*
1523 * Don't pass in an lsn greater than the lsn of the last
1524 * log record known to be on disk. Use a snapshot of the last sync lsn
1525 * so that it doesn't change between the compare and the set.
1526 */
1527 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1528 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1529 threshold_lsn = last_sync_lsn;
1530
1531 /*
1532 * Get the transaction layer to kick the dirty buffers out to
1533 * disk asynchronously. No point in trying to do this if
1534 * the filesystem is shutting down.
1535 */
1536 if (!XLOG_FORCED_SHUTDOWN(log))
1537 xfs_ail_push(log->l_ailp, threshold_lsn);
1538}
1539
1540/*
1541 * Stamp cycle number in every block
1542 */
1543STATIC void
1544xlog_pack_data(
1545 struct xlog *log,
1546 struct xlog_in_core *iclog,
1547 int roundoff)
1548{
1549 int i, j, k;
1550 int size = iclog->ic_offset + roundoff;
1551 __be32 cycle_lsn;
1552 char *dp;
1553
1554 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1555
1556 dp = iclog->ic_datap;
1557 for (i = 0; i < BTOBB(size); i++) {
1558 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1559 break;
1560 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1561 *(__be32 *)dp = cycle_lsn;
1562 dp += BBSIZE;
1563 }
1564
1565 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1566 xlog_in_core_2_t *xhdr = iclog->ic_data;
1567
1568 for ( ; i < BTOBB(size); i++) {
1569 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1570 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1571 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1572 *(__be32 *)dp = cycle_lsn;
1573 dp += BBSIZE;
1574 }
1575
1576 for (i = 1; i < log->l_iclog_heads; i++)
1577 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1578 }
1579}
1580
1581/*
1582 * Calculate the checksum for a log buffer.
1583 *
1584 * This is a little more complicated than it should be because the various
1585 * headers and the actual data are non-contiguous.
1586 */
1587__le32
1588xlog_cksum(
1589 struct xlog *log,
1590 struct xlog_rec_header *rhead,
1591 char *dp,
1592 int size)
1593{
1594 uint32_t crc;
1595
1596 /* first generate the crc for the record header ... */
1597 crc = xfs_start_cksum_update((char *)rhead,
1598 sizeof(struct xlog_rec_header),
1599 offsetof(struct xlog_rec_header, h_crc));
1600
1601 /* ... then for additional cycle data for v2 logs ... */
1602 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1603 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1604 int i;
1605 int xheads;
1606
1607 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1608 if (size % XLOG_HEADER_CYCLE_SIZE)
1609 xheads++;
1610
1611 for (i = 1; i < xheads; i++) {
1612 crc = crc32c(crc, &xhdr[i].hic_xheader,
1613 sizeof(struct xlog_rec_ext_header));
1614 }
1615 }
1616
1617 /* ... and finally for the payload */
1618 crc = crc32c(crc, dp, size);
1619
1620 return xfs_end_cksum(crc);
1621}
1622
1623static void
1624xlog_bio_end_io(
1625 struct bio *bio)
1626{
1627 struct xlog_in_core *iclog = bio->bi_private;
1628
1629 queue_work(iclog->ic_log->l_ioend_workqueue,
1630 &iclog->ic_end_io_work);
1631}
1632
1633static int
1634xlog_map_iclog_data(
1635 struct bio *bio,
1636 void *data,
1637 size_t count)
1638{
1639 do {
1640 struct page *page = kmem_to_page(data);
1641 unsigned int off = offset_in_page(data);
1642 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1643
1644 if (bio_add_page(bio, page, len, off) != len)
1645 return -EIO;
1646
1647 data += len;
1648 count -= len;
1649 } while (count);
1650
1651 return 0;
1652}
1653
1654STATIC void
1655xlog_write_iclog(
1656 struct xlog *log,
1657 struct xlog_in_core *iclog,
1658 uint64_t bno,
1659 unsigned int count,
1660 bool need_flush)
1661{
1662 ASSERT(bno < log->l_logBBsize);
1663
1664 /*
1665 * We lock the iclogbufs here so that we can serialise against I/O
1666 * completion during unmount. We might be processing a shutdown
1667 * triggered during unmount, and that can occur asynchronously to the
1668 * unmount thread, and hence we need to ensure that completes before
1669 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1670 * across the log IO to archieve that.
1671 */
1672 down(&iclog->ic_sema);
1673 if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1674 /*
1675 * It would seem logical to return EIO here, but we rely on
1676 * the log state machine to propagate I/O errors instead of
1677 * doing it here. We kick of the state machine and unlock
1678 * the buffer manually, the code needs to be kept in sync
1679 * with the I/O completion path.
1680 */
1681 xlog_state_done_syncing(iclog);
1682 up(&iclog->ic_sema);
1683 return;
1684 }
1685
1686 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1687 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1688 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1689 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1690 iclog->ic_bio.bi_private = iclog;
1691
1692 /*
1693 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1694 * IOs coming immediately after this one. This prevents the block layer
1695 * writeback throttle from throttling log writes behind background
1696 * metadata writeback and causing priority inversions.
1697 */
1698 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC |
1699 REQ_IDLE | REQ_FUA;
1700 if (need_flush)
1701 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1702
1703 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1704 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1705 return;
1706 }
1707 if (is_vmalloc_addr(iclog->ic_data))
1708 flush_kernel_vmap_range(iclog->ic_data, count);
1709
1710 /*
1711 * If this log buffer would straddle the end of the log we will have
1712 * to split it up into two bios, so that we can continue at the start.
1713 */
1714 if (bno + BTOBB(count) > log->l_logBBsize) {
1715 struct bio *split;
1716
1717 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1718 GFP_NOIO, &fs_bio_set);
1719 bio_chain(split, &iclog->ic_bio);
1720 submit_bio(split);
1721
1722 /* restart at logical offset zero for the remainder */
1723 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1724 }
1725
1726 submit_bio(&iclog->ic_bio);
1727}
1728
1729/*
1730 * We need to bump cycle number for the part of the iclog that is
1731 * written to the start of the log. Watch out for the header magic
1732 * number case, though.
1733 */
1734static void
1735xlog_split_iclog(
1736 struct xlog *log,
1737 void *data,
1738 uint64_t bno,
1739 unsigned int count)
1740{
1741 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1742 unsigned int i;
1743
1744 for (i = split_offset; i < count; i += BBSIZE) {
1745 uint32_t cycle = get_unaligned_be32(data + i);
1746
1747 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1748 cycle++;
1749 put_unaligned_be32(cycle, data + i);
1750 }
1751}
1752
1753static int
1754xlog_calc_iclog_size(
1755 struct xlog *log,
1756 struct xlog_in_core *iclog,
1757 uint32_t *roundoff)
1758{
1759 uint32_t count_init, count;
1760 bool use_lsunit;
1761
1762 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1763 log->l_mp->m_sb.sb_logsunit > 1;
1764
1765 /* Add for LR header */
1766 count_init = log->l_iclog_hsize + iclog->ic_offset;
1767
1768 /* Round out the log write size */
1769 if (use_lsunit) {
1770 /* we have a v2 stripe unit to use */
1771 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1772 } else {
1773 count = BBTOB(BTOBB(count_init));
1774 }
1775
1776 ASSERT(count >= count_init);
1777 *roundoff = count - count_init;
1778
1779 if (use_lsunit)
1780 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1781 else
1782 ASSERT(*roundoff < BBTOB(1));
1783 return count;
1784}
1785
1786/*
1787 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1788 * fashion. Previously, we should have moved the current iclog
1789 * ptr in the log to point to the next available iclog. This allows further
1790 * write to continue while this code syncs out an iclog ready to go.
1791 * Before an in-core log can be written out, the data section must be scanned
1792 * to save away the 1st word of each BBSIZE block into the header. We replace
1793 * it with the current cycle count. Each BBSIZE block is tagged with the
1794 * cycle count because there in an implicit assumption that drives will
1795 * guarantee that entire 512 byte blocks get written at once. In other words,
1796 * we can't have part of a 512 byte block written and part not written. By
1797 * tagging each block, we will know which blocks are valid when recovering
1798 * after an unclean shutdown.
1799 *
1800 * This routine is single threaded on the iclog. No other thread can be in
1801 * this routine with the same iclog. Changing contents of iclog can there-
1802 * fore be done without grabbing the state machine lock. Updating the global
1803 * log will require grabbing the lock though.
1804 *
1805 * The entire log manager uses a logical block numbering scheme. Only
1806 * xlog_write_iclog knows about the fact that the log may not start with
1807 * block zero on a given device.
1808 */
1809STATIC void
1810xlog_sync(
1811 struct xlog *log,
1812 struct xlog_in_core *iclog)
1813{
1814 unsigned int count; /* byte count of bwrite */
1815 unsigned int roundoff; /* roundoff to BB or stripe */
1816 uint64_t bno;
1817 unsigned int size;
1818 bool need_flush = true, split = false;
1819
1820 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1821
1822 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1823
1824 /* move grant heads by roundoff in sync */
1825 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1826 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1827
1828 /* put cycle number in every block */
1829 xlog_pack_data(log, iclog, roundoff);
1830
1831 /* real byte length */
1832 size = iclog->ic_offset;
1833 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1834 size += roundoff;
1835 iclog->ic_header.h_len = cpu_to_be32(size);
1836
1837 XFS_STATS_INC(log->l_mp, xs_log_writes);
1838 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1839
1840 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1841
1842 /* Do we need to split this write into 2 parts? */
1843 if (bno + BTOBB(count) > log->l_logBBsize) {
1844 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1845 split = true;
1846 }
1847
1848 /* calculcate the checksum */
1849 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1850 iclog->ic_datap, size);
1851 /*
1852 * Intentionally corrupt the log record CRC based on the error injection
1853 * frequency, if defined. This facilitates testing log recovery in the
1854 * event of torn writes. Hence, set the IOABORT state to abort the log
1855 * write on I/O completion and shutdown the fs. The subsequent mount
1856 * detects the bad CRC and attempts to recover.
1857 */
1858#ifdef DEBUG
1859 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1860 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1861 iclog->ic_fail_crc = true;
1862 xfs_warn(log->l_mp,
1863 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1864 be64_to_cpu(iclog->ic_header.h_lsn));
1865 }
1866#endif
1867
1868 /*
1869 * Flush the data device before flushing the log to make sure all meta
1870 * data written back from the AIL actually made it to disk before
1871 * stamping the new log tail LSN into the log buffer. For an external
1872 * log we need to issue the flush explicitly, and unfortunately
1873 * synchronously here; for an internal log we can simply use the block
1874 * layer state machine for preflushes.
1875 */
1876 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1877 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1878 need_flush = false;
1879 }
1880
1881 xlog_verify_iclog(log, iclog, count);
1882 xlog_write_iclog(log, iclog, bno, count, need_flush);
1883}
1884
1885/*
1886 * Deallocate a log structure
1887 */
1888STATIC void
1889xlog_dealloc_log(
1890 struct xlog *log)
1891{
1892 xlog_in_core_t *iclog, *next_iclog;
1893 int i;
1894
1895 xlog_cil_destroy(log);
1896
1897 /*
1898 * Cycle all the iclogbuf locks to make sure all log IO completion
1899 * is done before we tear down these buffers.
1900 */
1901 iclog = log->l_iclog;
1902 for (i = 0; i < log->l_iclog_bufs; i++) {
1903 down(&iclog->ic_sema);
1904 up(&iclog->ic_sema);
1905 iclog = iclog->ic_next;
1906 }
1907
1908 iclog = log->l_iclog;
1909 for (i = 0; i < log->l_iclog_bufs; i++) {
1910 next_iclog = iclog->ic_next;
1911 kmem_free(iclog->ic_data);
1912 kmem_free(iclog);
1913 iclog = next_iclog;
1914 }
1915
1916 log->l_mp->m_log = NULL;
1917 destroy_workqueue(log->l_ioend_workqueue);
1918 kmem_free(log);
1919}
1920
1921/*
1922 * Update counters atomically now that memcpy is done.
1923 */
1924static inline void
1925xlog_state_finish_copy(
1926 struct xlog *log,
1927 struct xlog_in_core *iclog,
1928 int record_cnt,
1929 int copy_bytes)
1930{
1931 lockdep_assert_held(&log->l_icloglock);
1932
1933 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1934 iclog->ic_offset += copy_bytes;
1935}
1936
1937/*
1938 * print out info relating to regions written which consume
1939 * the reservation
1940 */
1941void
1942xlog_print_tic_res(
1943 struct xfs_mount *mp,
1944 struct xlog_ticket *ticket)
1945{
1946 uint i;
1947 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1948
1949 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1950#define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
1951 static char *res_type_str[] = {
1952 REG_TYPE_STR(BFORMAT, "bformat"),
1953 REG_TYPE_STR(BCHUNK, "bchunk"),
1954 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
1955 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
1956 REG_TYPE_STR(IFORMAT, "iformat"),
1957 REG_TYPE_STR(ICORE, "icore"),
1958 REG_TYPE_STR(IEXT, "iext"),
1959 REG_TYPE_STR(IBROOT, "ibroot"),
1960 REG_TYPE_STR(ILOCAL, "ilocal"),
1961 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
1962 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
1963 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
1964 REG_TYPE_STR(QFORMAT, "qformat"),
1965 REG_TYPE_STR(DQUOT, "dquot"),
1966 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
1967 REG_TYPE_STR(LRHEADER, "LR header"),
1968 REG_TYPE_STR(UNMOUNT, "unmount"),
1969 REG_TYPE_STR(COMMIT, "commit"),
1970 REG_TYPE_STR(TRANSHDR, "trans header"),
1971 REG_TYPE_STR(ICREATE, "inode create"),
1972 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
1973 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
1974 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
1975 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
1976 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
1977 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
1978 };
1979 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
1980#undef REG_TYPE_STR
1981
1982 xfs_warn(mp, "ticket reservation summary:");
1983 xfs_warn(mp, " unit res = %d bytes",
1984 ticket->t_unit_res);
1985 xfs_warn(mp, " current res = %d bytes",
1986 ticket->t_curr_res);
1987 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
1988 ticket->t_res_arr_sum, ticket->t_res_o_flow);
1989 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
1990 ticket->t_res_num_ophdrs, ophdr_spc);
1991 xfs_warn(mp, " ophdr + reg = %u bytes",
1992 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
1993 xfs_warn(mp, " num regions = %u",
1994 ticket->t_res_num);
1995
1996 for (i = 0; i < ticket->t_res_num; i++) {
1997 uint r_type = ticket->t_res_arr[i].r_type;
1998 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
1999 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2000 "bad-rtype" : res_type_str[r_type]),
2001 ticket->t_res_arr[i].r_len);
2002 }
2003}
2004
2005/*
2006 * Print a summary of the transaction.
2007 */
2008void
2009xlog_print_trans(
2010 struct xfs_trans *tp)
2011{
2012 struct xfs_mount *mp = tp->t_mountp;
2013 struct xfs_log_item *lip;
2014
2015 /* dump core transaction and ticket info */
2016 xfs_warn(mp, "transaction summary:");
2017 xfs_warn(mp, " log res = %d", tp->t_log_res);
2018 xfs_warn(mp, " log count = %d", tp->t_log_count);
2019 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2020
2021 xlog_print_tic_res(mp, tp->t_ticket);
2022
2023 /* dump each log item */
2024 list_for_each_entry(lip, &tp->t_items, li_trans) {
2025 struct xfs_log_vec *lv = lip->li_lv;
2026 struct xfs_log_iovec *vec;
2027 int i;
2028
2029 xfs_warn(mp, "log item: ");
2030 xfs_warn(mp, " type = 0x%x", lip->li_type);
2031 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2032 if (!lv)
2033 continue;
2034 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2035 xfs_warn(mp, " size = %d", lv->lv_size);
2036 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2037 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2038
2039 /* dump each iovec for the log item */
2040 vec = lv->lv_iovecp;
2041 for (i = 0; i < lv->lv_niovecs; i++) {
2042 int dumplen = min(vec->i_len, 32);
2043
2044 xfs_warn(mp, " iovec[%d]", i);
2045 xfs_warn(mp, " type = 0x%x", vec->i_type);
2046 xfs_warn(mp, " len = %d", vec->i_len);
2047 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2048 xfs_hex_dump(vec->i_addr, dumplen);
2049
2050 vec++;
2051 }
2052 }
2053}
2054
2055/*
2056 * Calculate the potential space needed by the log vector. We may need a start
2057 * record, and each region gets its own struct xlog_op_header and may need to be
2058 * double word aligned.
2059 */
2060static int
2061xlog_write_calc_vec_length(
2062 struct xlog_ticket *ticket,
2063 struct xfs_log_vec *log_vector,
2064 bool need_start_rec)
2065{
2066 struct xfs_log_vec *lv;
2067 int headers = need_start_rec ? 1 : 0;
2068 int len = 0;
2069 int i;
2070
2071 for (lv = log_vector; lv; lv = lv->lv_next) {
2072 /* we don't write ordered log vectors */
2073 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2074 continue;
2075
2076 headers += lv->lv_niovecs;
2077
2078 for (i = 0; i < lv->lv_niovecs; i++) {
2079 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2080
2081 len += vecp->i_len;
2082 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2083 }
2084 }
2085
2086 ticket->t_res_num_ophdrs += headers;
2087 len += headers * sizeof(struct xlog_op_header);
2088
2089 return len;
2090}
2091
2092static void
2093xlog_write_start_rec(
2094 struct xlog_op_header *ophdr,
2095 struct xlog_ticket *ticket)
2096{
2097 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2098 ophdr->oh_clientid = ticket->t_clientid;
2099 ophdr->oh_len = 0;
2100 ophdr->oh_flags = XLOG_START_TRANS;
2101 ophdr->oh_res2 = 0;
2102}
2103
2104static xlog_op_header_t *
2105xlog_write_setup_ophdr(
2106 struct xlog *log,
2107 struct xlog_op_header *ophdr,
2108 struct xlog_ticket *ticket,
2109 uint flags)
2110{
2111 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2112 ophdr->oh_clientid = ticket->t_clientid;
2113 ophdr->oh_res2 = 0;
2114
2115 /* are we copying a commit or unmount record? */
2116 ophdr->oh_flags = flags;
2117
2118 /*
2119 * We've seen logs corrupted with bad transaction client ids. This
2120 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2121 * and shut down the filesystem.
2122 */
2123 switch (ophdr->oh_clientid) {
2124 case XFS_TRANSACTION:
2125 case XFS_VOLUME:
2126 case XFS_LOG:
2127 break;
2128 default:
2129 xfs_warn(log->l_mp,
2130 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2131 ophdr->oh_clientid, ticket);
2132 return NULL;
2133 }
2134
2135 return ophdr;
2136}
2137
2138/*
2139 * Set up the parameters of the region copy into the log. This has
2140 * to handle region write split across multiple log buffers - this
2141 * state is kept external to this function so that this code can
2142 * be written in an obvious, self documenting manner.
2143 */
2144static int
2145xlog_write_setup_copy(
2146 struct xlog_ticket *ticket,
2147 struct xlog_op_header *ophdr,
2148 int space_available,
2149 int space_required,
2150 int *copy_off,
2151 int *copy_len,
2152 int *last_was_partial_copy,
2153 int *bytes_consumed)
2154{
2155 int still_to_copy;
2156
2157 still_to_copy = space_required - *bytes_consumed;
2158 *copy_off = *bytes_consumed;
2159
2160 if (still_to_copy <= space_available) {
2161 /* write of region completes here */
2162 *copy_len = still_to_copy;
2163 ophdr->oh_len = cpu_to_be32(*copy_len);
2164 if (*last_was_partial_copy)
2165 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2166 *last_was_partial_copy = 0;
2167 *bytes_consumed = 0;
2168 return 0;
2169 }
2170
2171 /* partial write of region, needs extra log op header reservation */
2172 *copy_len = space_available;
2173 ophdr->oh_len = cpu_to_be32(*copy_len);
2174 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2175 if (*last_was_partial_copy)
2176 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2177 *bytes_consumed += *copy_len;
2178 (*last_was_partial_copy)++;
2179
2180 /* account for new log op header */
2181 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2182 ticket->t_res_num_ophdrs++;
2183
2184 return sizeof(struct xlog_op_header);
2185}
2186
2187static int
2188xlog_write_copy_finish(
2189 struct xlog *log,
2190 struct xlog_in_core *iclog,
2191 uint flags,
2192 int *record_cnt,
2193 int *data_cnt,
2194 int *partial_copy,
2195 int *partial_copy_len,
2196 int log_offset,
2197 struct xlog_in_core **commit_iclog)
2198{
2199 int error;
2200
2201 if (*partial_copy) {
2202 /*
2203 * This iclog has already been marked WANT_SYNC by
2204 * xlog_state_get_iclog_space.
2205 */
2206 spin_lock(&log->l_icloglock);
2207 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2208 *record_cnt = 0;
2209 *data_cnt = 0;
2210 goto release_iclog;
2211 }
2212
2213 *partial_copy = 0;
2214 *partial_copy_len = 0;
2215
2216 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2217 /* no more space in this iclog - push it. */
2218 spin_lock(&log->l_icloglock);
2219 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2220 *record_cnt = 0;
2221 *data_cnt = 0;
2222
2223 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2224 xlog_state_switch_iclogs(log, iclog, 0);
2225 else
2226 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2227 iclog->ic_state == XLOG_STATE_IOERROR);
2228 if (!commit_iclog)
2229 goto release_iclog;
2230 spin_unlock(&log->l_icloglock);
2231 ASSERT(flags & XLOG_COMMIT_TRANS);
2232 *commit_iclog = iclog;
2233 }
2234
2235 return 0;
2236
2237release_iclog:
2238 error = xlog_state_release_iclog(log, iclog);
2239 spin_unlock(&log->l_icloglock);
2240 return error;
2241}
2242
2243/*
2244 * Write some region out to in-core log
2245 *
2246 * This will be called when writing externally provided regions or when
2247 * writing out a commit record for a given transaction.
2248 *
2249 * General algorithm:
2250 * 1. Find total length of this write. This may include adding to the
2251 * lengths passed in.
2252 * 2. Check whether we violate the tickets reservation.
2253 * 3. While writing to this iclog
2254 * A. Reserve as much space in this iclog as can get
2255 * B. If this is first write, save away start lsn
2256 * C. While writing this region:
2257 * 1. If first write of transaction, write start record
2258 * 2. Write log operation header (header per region)
2259 * 3. Find out if we can fit entire region into this iclog
2260 * 4. Potentially, verify destination memcpy ptr
2261 * 5. Memcpy (partial) region
2262 * 6. If partial copy, release iclog; otherwise, continue
2263 * copying more regions into current iclog
2264 * 4. Mark want sync bit (in simulation mode)
2265 * 5. Release iclog for potential flush to on-disk log.
2266 *
2267 * ERRORS:
2268 * 1. Panic if reservation is overrun. This should never happen since
2269 * reservation amounts are generated internal to the filesystem.
2270 * NOTES:
2271 * 1. Tickets are single threaded data structures.
2272 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2273 * syncing routine. When a single log_write region needs to span
2274 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2275 * on all log operation writes which don't contain the end of the
2276 * region. The XLOG_END_TRANS bit is used for the in-core log
2277 * operation which contains the end of the continued log_write region.
2278 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2279 * we don't really know exactly how much space will be used. As a result,
2280 * we don't update ic_offset until the end when we know exactly how many
2281 * bytes have been written out.
2282 */
2283int
2284xlog_write(
2285 struct xlog *log,
2286 struct xfs_log_vec *log_vector,
2287 struct xlog_ticket *ticket,
2288 xfs_lsn_t *start_lsn,
2289 struct xlog_in_core **commit_iclog,
2290 uint flags,
2291 bool need_start_rec)
2292{
2293 struct xlog_in_core *iclog = NULL;
2294 struct xfs_log_vec *lv = log_vector;
2295 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2296 int index = 0;
2297 int len;
2298 int partial_copy = 0;
2299 int partial_copy_len = 0;
2300 int contwr = 0;
2301 int record_cnt = 0;
2302 int data_cnt = 0;
2303 int error = 0;
2304
2305 /*
2306 * If this is a commit or unmount transaction, we don't need a start
2307 * record to be written. We do, however, have to account for the
2308 * commit or unmount header that gets written. Hence we always have
2309 * to account for an extra xlog_op_header here.
2310 */
2311 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2312 if (ticket->t_curr_res < 0) {
2313 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2314 "ctx ticket reservation ran out. Need to up reservation");
2315 xlog_print_tic_res(log->l_mp, ticket);
2316 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2317 }
2318
2319 len = xlog_write_calc_vec_length(ticket, log_vector, need_start_rec);
2320 *start_lsn = 0;
2321 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2322 void *ptr;
2323 int log_offset;
2324
2325 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2326 &contwr, &log_offset);
2327 if (error)
2328 return error;
2329
2330 ASSERT(log_offset <= iclog->ic_size - 1);
2331 ptr = iclog->ic_datap + log_offset;
2332
2333 /* start_lsn is the first lsn written to. That's all we need. */
2334 if (!*start_lsn)
2335 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2336
2337 /*
2338 * This loop writes out as many regions as can fit in the amount
2339 * of space which was allocated by xlog_state_get_iclog_space().
2340 */
2341 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2342 struct xfs_log_iovec *reg;
2343 struct xlog_op_header *ophdr;
2344 int copy_len;
2345 int copy_off;
2346 bool ordered = false;
2347
2348 /* ordered log vectors have no regions to write */
2349 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2350 ASSERT(lv->lv_niovecs == 0);
2351 ordered = true;
2352 goto next_lv;
2353 }
2354
2355 reg = &vecp[index];
2356 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2357 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2358
2359 /*
2360 * Before we start formatting log vectors, we need to
2361 * write a start record. Only do this for the first
2362 * iclog we write to.
2363 */
2364 if (need_start_rec) {
2365 xlog_write_start_rec(ptr, ticket);
2366 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2367 sizeof(struct xlog_op_header));
2368 }
2369
2370 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2371 if (!ophdr)
2372 return -EIO;
2373
2374 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2375 sizeof(struct xlog_op_header));
2376
2377 len += xlog_write_setup_copy(ticket, ophdr,
2378 iclog->ic_size-log_offset,
2379 reg->i_len,
2380 ©_off, ©_len,
2381 &partial_copy,
2382 &partial_copy_len);
2383 xlog_verify_dest_ptr(log, ptr);
2384
2385 /*
2386 * Copy region.
2387 *
2388 * Unmount records just log an opheader, so can have
2389 * empty payloads with no data region to copy. Hence we
2390 * only copy the payload if the vector says it has data
2391 * to copy.
2392 */
2393 ASSERT(copy_len >= 0);
2394 if (copy_len > 0) {
2395 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2396 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2397 copy_len);
2398 }
2399 copy_len += sizeof(struct xlog_op_header);
2400 record_cnt++;
2401 if (need_start_rec) {
2402 copy_len += sizeof(struct xlog_op_header);
2403 record_cnt++;
2404 need_start_rec = false;
2405 }
2406 data_cnt += contwr ? copy_len : 0;
2407
2408 error = xlog_write_copy_finish(log, iclog, flags,
2409 &record_cnt, &data_cnt,
2410 &partial_copy,
2411 &partial_copy_len,
2412 log_offset,
2413 commit_iclog);
2414 if (error)
2415 return error;
2416
2417 /*
2418 * if we had a partial copy, we need to get more iclog
2419 * space but we don't want to increment the region
2420 * index because there is still more is this region to
2421 * write.
2422 *
2423 * If we completed writing this region, and we flushed
2424 * the iclog (indicated by resetting of the record
2425 * count), then we also need to get more log space. If
2426 * this was the last record, though, we are done and
2427 * can just return.
2428 */
2429 if (partial_copy)
2430 break;
2431
2432 if (++index == lv->lv_niovecs) {
2433next_lv:
2434 lv = lv->lv_next;
2435 index = 0;
2436 if (lv)
2437 vecp = lv->lv_iovecp;
2438 }
2439 if (record_cnt == 0 && !ordered) {
2440 if (!lv)
2441 return 0;
2442 break;
2443 }
2444 }
2445 }
2446
2447 ASSERT(len == 0);
2448
2449 spin_lock(&log->l_icloglock);
2450 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2451 if (commit_iclog) {
2452 ASSERT(flags & XLOG_COMMIT_TRANS);
2453 *commit_iclog = iclog;
2454 } else {
2455 error = xlog_state_release_iclog(log, iclog);
2456 }
2457 spin_unlock(&log->l_icloglock);
2458
2459 return error;
2460}
2461
2462static void
2463xlog_state_activate_iclog(
2464 struct xlog_in_core *iclog,
2465 int *iclogs_changed)
2466{
2467 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2468
2469 /*
2470 * If the number of ops in this iclog indicate it just contains the
2471 * dummy transaction, we can change state into IDLE (the second time
2472 * around). Otherwise we should change the state into NEED a dummy.
2473 * We don't need to cover the dummy.
2474 */
2475 if (*iclogs_changed == 0 &&
2476 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2477 *iclogs_changed = 1;
2478 } else {
2479 /*
2480 * We have two dirty iclogs so start over. This could also be
2481 * num of ops indicating this is not the dummy going out.
2482 */
2483 *iclogs_changed = 2;
2484 }
2485
2486 iclog->ic_state = XLOG_STATE_ACTIVE;
2487 iclog->ic_offset = 0;
2488 iclog->ic_header.h_num_logops = 0;
2489 memset(iclog->ic_header.h_cycle_data, 0,
2490 sizeof(iclog->ic_header.h_cycle_data));
2491 iclog->ic_header.h_lsn = 0;
2492}
2493
2494/*
2495 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2496 * ACTIVE after iclog I/O has completed.
2497 */
2498static void
2499xlog_state_activate_iclogs(
2500 struct xlog *log,
2501 int *iclogs_changed)
2502{
2503 struct xlog_in_core *iclog = log->l_iclog;
2504
2505 do {
2506 if (iclog->ic_state == XLOG_STATE_DIRTY)
2507 xlog_state_activate_iclog(iclog, iclogs_changed);
2508 /*
2509 * The ordering of marking iclogs ACTIVE must be maintained, so
2510 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2511 */
2512 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2513 break;
2514 } while ((iclog = iclog->ic_next) != log->l_iclog);
2515}
2516
2517static int
2518xlog_covered_state(
2519 int prev_state,
2520 int iclogs_changed)
2521{
2522 /*
2523 * We usually go to NEED. But we go to NEED2 if the changed indicates we
2524 * are done writing the dummy record. If we are done with the second
2525 * dummy recored (DONE2), then we go to IDLE.
2526 */
2527 switch (prev_state) {
2528 case XLOG_STATE_COVER_IDLE:
2529 case XLOG_STATE_COVER_NEED:
2530 case XLOG_STATE_COVER_NEED2:
2531 break;
2532 case XLOG_STATE_COVER_DONE:
2533 if (iclogs_changed == 1)
2534 return XLOG_STATE_COVER_NEED2;
2535 break;
2536 case XLOG_STATE_COVER_DONE2:
2537 if (iclogs_changed == 1)
2538 return XLOG_STATE_COVER_IDLE;
2539 break;
2540 default:
2541 ASSERT(0);
2542 }
2543
2544 return XLOG_STATE_COVER_NEED;
2545}
2546
2547STATIC void
2548xlog_state_clean_iclog(
2549 struct xlog *log,
2550 struct xlog_in_core *dirty_iclog)
2551{
2552 int iclogs_changed = 0;
2553
2554 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2555
2556 xlog_state_activate_iclogs(log, &iclogs_changed);
2557 wake_up_all(&dirty_iclog->ic_force_wait);
2558
2559 if (iclogs_changed) {
2560 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2561 iclogs_changed);
2562 }
2563}
2564
2565STATIC xfs_lsn_t
2566xlog_get_lowest_lsn(
2567 struct xlog *log)
2568{
2569 struct xlog_in_core *iclog = log->l_iclog;
2570 xfs_lsn_t lowest_lsn = 0, lsn;
2571
2572 do {
2573 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2574 iclog->ic_state == XLOG_STATE_DIRTY)
2575 continue;
2576
2577 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2578 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2579 lowest_lsn = lsn;
2580 } while ((iclog = iclog->ic_next) != log->l_iclog);
2581
2582 return lowest_lsn;
2583}
2584
2585/*
2586 * Completion of a iclog IO does not imply that a transaction has completed, as
2587 * transactions can be large enough to span many iclogs. We cannot change the
2588 * tail of the log half way through a transaction as this may be the only
2589 * transaction in the log and moving the tail to point to the middle of it
2590 * will prevent recovery from finding the start of the transaction. Hence we
2591 * should only update the last_sync_lsn if this iclog contains transaction
2592 * completion callbacks on it.
2593 *
2594 * We have to do this before we drop the icloglock to ensure we are the only one
2595 * that can update it.
2596 *
2597 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2598 * the reservation grant head pushing. This is due to the fact that the push
2599 * target is bound by the current last_sync_lsn value. Hence if we have a large
2600 * amount of log space bound up in this committing transaction then the
2601 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2602 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2603 * should push the AIL to ensure the push target (and hence the grant head) is
2604 * no longer bound by the old log head location and can move forwards and make
2605 * progress again.
2606 */
2607static void
2608xlog_state_set_callback(
2609 struct xlog *log,
2610 struct xlog_in_core *iclog,
2611 xfs_lsn_t header_lsn)
2612{
2613 iclog->ic_state = XLOG_STATE_CALLBACK;
2614
2615 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2616 header_lsn) <= 0);
2617
2618 if (list_empty_careful(&iclog->ic_callbacks))
2619 return;
2620
2621 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2622 xlog_grant_push_ail(log, 0);
2623}
2624
2625/*
2626 * Return true if we need to stop processing, false to continue to the next
2627 * iclog. The caller will need to run callbacks if the iclog is returned in the
2628 * XLOG_STATE_CALLBACK state.
2629 */
2630static bool
2631xlog_state_iodone_process_iclog(
2632 struct xlog *log,
2633 struct xlog_in_core *iclog,
2634 bool *ioerror)
2635{
2636 xfs_lsn_t lowest_lsn;
2637 xfs_lsn_t header_lsn;
2638
2639 switch (iclog->ic_state) {
2640 case XLOG_STATE_ACTIVE:
2641 case XLOG_STATE_DIRTY:
2642 /*
2643 * Skip all iclogs in the ACTIVE & DIRTY states:
2644 */
2645 return false;
2646 case XLOG_STATE_IOERROR:
2647 /*
2648 * Between marking a filesystem SHUTDOWN and stopping the log,
2649 * we do flush all iclogs to disk (if there wasn't a log I/O
2650 * error). So, we do want things to go smoothly in case of just
2651 * a SHUTDOWN w/o a LOG_IO_ERROR.
2652 */
2653 *ioerror = true;
2654 return false;
2655 case XLOG_STATE_DONE_SYNC:
2656 /*
2657 * Now that we have an iclog that is in the DONE_SYNC state, do
2658 * one more check here to see if we have chased our tail around.
2659 * If this is not the lowest lsn iclog, then we will leave it
2660 * for another completion to process.
2661 */
2662 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2663 lowest_lsn = xlog_get_lowest_lsn(log);
2664 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2665 return false;
2666 xlog_state_set_callback(log, iclog, header_lsn);
2667 return false;
2668 default:
2669 /*
2670 * Can only perform callbacks in order. Since this iclog is not
2671 * in the DONE_SYNC state, we skip the rest and just try to
2672 * clean up.
2673 */
2674 return true;
2675 }
2676}
2677
2678/*
2679 * Keep processing entries in the iclog callback list until we come around and
2680 * it is empty. We need to atomically see that the list is empty and change the
2681 * state to DIRTY so that we don't miss any more callbacks being added.
2682 *
2683 * This function is called with the icloglock held and returns with it held. We
2684 * drop it while running callbacks, however, as holding it over thousands of
2685 * callbacks is unnecessary and causes excessive contention if we do.
2686 */
2687static void
2688xlog_state_do_iclog_callbacks(
2689 struct xlog *log,
2690 struct xlog_in_core *iclog)
2691 __releases(&log->l_icloglock)
2692 __acquires(&log->l_icloglock)
2693{
2694 spin_unlock(&log->l_icloglock);
2695 spin_lock(&iclog->ic_callback_lock);
2696 while (!list_empty(&iclog->ic_callbacks)) {
2697 LIST_HEAD(tmp);
2698
2699 list_splice_init(&iclog->ic_callbacks, &tmp);
2700
2701 spin_unlock(&iclog->ic_callback_lock);
2702 xlog_cil_process_committed(&tmp);
2703 spin_lock(&iclog->ic_callback_lock);
2704 }
2705
2706 /*
2707 * Pick up the icloglock while still holding the callback lock so we
2708 * serialise against anyone trying to add more callbacks to this iclog
2709 * now we've finished processing.
2710 */
2711 spin_lock(&log->l_icloglock);
2712 spin_unlock(&iclog->ic_callback_lock);
2713}
2714
2715STATIC void
2716xlog_state_do_callback(
2717 struct xlog *log)
2718{
2719 struct xlog_in_core *iclog;
2720 struct xlog_in_core *first_iclog;
2721 bool cycled_icloglock;
2722 bool ioerror;
2723 int flushcnt = 0;
2724 int repeats = 0;
2725
2726 spin_lock(&log->l_icloglock);
2727 do {
2728 /*
2729 * Scan all iclogs starting with the one pointed to by the
2730 * log. Reset this starting point each time the log is
2731 * unlocked (during callbacks).
2732 *
2733 * Keep looping through iclogs until one full pass is made
2734 * without running any callbacks.
2735 */
2736 first_iclog = log->l_iclog;
2737 iclog = log->l_iclog;
2738 cycled_icloglock = false;
2739 ioerror = false;
2740 repeats++;
2741
2742 do {
2743 if (xlog_state_iodone_process_iclog(log, iclog,
2744 &ioerror))
2745 break;
2746
2747 if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2748 iclog->ic_state != XLOG_STATE_IOERROR) {
2749 iclog = iclog->ic_next;
2750 continue;
2751 }
2752
2753 /*
2754 * Running callbacks will drop the icloglock which means
2755 * we'll have to run at least one more complete loop.
2756 */
2757 cycled_icloglock = true;
2758 xlog_state_do_iclog_callbacks(log, iclog);
2759 if (XLOG_FORCED_SHUTDOWN(log))
2760 wake_up_all(&iclog->ic_force_wait);
2761 else
2762 xlog_state_clean_iclog(log, iclog);
2763 iclog = iclog->ic_next;
2764 } while (first_iclog != iclog);
2765
2766 if (repeats > 5000) {
2767 flushcnt += repeats;
2768 repeats = 0;
2769 xfs_warn(log->l_mp,
2770 "%s: possible infinite loop (%d iterations)",
2771 __func__, flushcnt);
2772 }
2773 } while (!ioerror && cycled_icloglock);
2774
2775 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2776 log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2777 wake_up_all(&log->l_flush_wait);
2778
2779 spin_unlock(&log->l_icloglock);
2780}
2781
2782
2783/*
2784 * Finish transitioning this iclog to the dirty state.
2785 *
2786 * Make sure that we completely execute this routine only when this is
2787 * the last call to the iclog. There is a good chance that iclog flushes,
2788 * when we reach the end of the physical log, get turned into 2 separate
2789 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2790 * routine. By using the reference count bwritecnt, we guarantee that only
2791 * the second completion goes through.
2792 *
2793 * Callbacks could take time, so they are done outside the scope of the
2794 * global state machine log lock.
2795 */
2796STATIC void
2797xlog_state_done_syncing(
2798 struct xlog_in_core *iclog)
2799{
2800 struct xlog *log = iclog->ic_log;
2801
2802 spin_lock(&log->l_icloglock);
2803 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2804
2805 /*
2806 * If we got an error, either on the first buffer, or in the case of
2807 * split log writes, on the second, we shut down the file system and
2808 * no iclogs should ever be attempted to be written to disk again.
2809 */
2810 if (!XLOG_FORCED_SHUTDOWN(log)) {
2811 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2812 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2813 }
2814
2815 /*
2816 * Someone could be sleeping prior to writing out the next
2817 * iclog buffer, we wake them all, one will get to do the
2818 * I/O, the others get to wait for the result.
2819 */
2820 wake_up_all(&iclog->ic_write_wait);
2821 spin_unlock(&log->l_icloglock);
2822 xlog_state_do_callback(log);
2823}
2824
2825/*
2826 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2827 * sleep. We wait on the flush queue on the head iclog as that should be
2828 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2829 * we will wait here and all new writes will sleep until a sync completes.
2830 *
2831 * The in-core logs are used in a circular fashion. They are not used
2832 * out-of-order even when an iclog past the head is free.
2833 *
2834 * return:
2835 * * log_offset where xlog_write() can start writing into the in-core
2836 * log's data space.
2837 * * in-core log pointer to which xlog_write() should write.
2838 * * boolean indicating this is a continued write to an in-core log.
2839 * If this is the last write, then the in-core log's offset field
2840 * needs to be incremented, depending on the amount of data which
2841 * is copied.
2842 */
2843STATIC int
2844xlog_state_get_iclog_space(
2845 struct xlog *log,
2846 int len,
2847 struct xlog_in_core **iclogp,
2848 struct xlog_ticket *ticket,
2849 int *continued_write,
2850 int *logoffsetp)
2851{
2852 int log_offset;
2853 xlog_rec_header_t *head;
2854 xlog_in_core_t *iclog;
2855
2856restart:
2857 spin_lock(&log->l_icloglock);
2858 if (XLOG_FORCED_SHUTDOWN(log)) {
2859 spin_unlock(&log->l_icloglock);
2860 return -EIO;
2861 }
2862
2863 iclog = log->l_iclog;
2864 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2865 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2866
2867 /* Wait for log writes to have flushed */
2868 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2869 goto restart;
2870 }
2871
2872 head = &iclog->ic_header;
2873
2874 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2875 log_offset = iclog->ic_offset;
2876
2877 /* On the 1st write to an iclog, figure out lsn. This works
2878 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2879 * committing to. If the offset is set, that's how many blocks
2880 * must be written.
2881 */
2882 if (log_offset == 0) {
2883 ticket->t_curr_res -= log->l_iclog_hsize;
2884 xlog_tic_add_region(ticket,
2885 log->l_iclog_hsize,
2886 XLOG_REG_TYPE_LRHEADER);
2887 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2888 head->h_lsn = cpu_to_be64(
2889 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2890 ASSERT(log->l_curr_block >= 0);
2891 }
2892
2893 /* If there is enough room to write everything, then do it. Otherwise,
2894 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2895 * bit is on, so this will get flushed out. Don't update ic_offset
2896 * until you know exactly how many bytes get copied. Therefore, wait
2897 * until later to update ic_offset.
2898 *
2899 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2900 * can fit into remaining data section.
2901 */
2902 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2903 int error = 0;
2904
2905 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2906
2907 /*
2908 * If we are the only one writing to this iclog, sync it to
2909 * disk. We need to do an atomic compare and decrement here to
2910 * avoid racing with concurrent atomic_dec_and_lock() calls in
2911 * xlog_state_release_iclog() when there is more than one
2912 * reference to the iclog.
2913 */
2914 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2915 error = xlog_state_release_iclog(log, iclog);
2916 spin_unlock(&log->l_icloglock);
2917 if (error)
2918 return error;
2919 goto restart;
2920 }
2921
2922 /* Do we have enough room to write the full amount in the remainder
2923 * of this iclog? Or must we continue a write on the next iclog and
2924 * mark this iclog as completely taken? In the case where we switch
2925 * iclogs (to mark it taken), this particular iclog will release/sync
2926 * to disk in xlog_write().
2927 */
2928 if (len <= iclog->ic_size - iclog->ic_offset) {
2929 *continued_write = 0;
2930 iclog->ic_offset += len;
2931 } else {
2932 *continued_write = 1;
2933 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2934 }
2935 *iclogp = iclog;
2936
2937 ASSERT(iclog->ic_offset <= iclog->ic_size);
2938 spin_unlock(&log->l_icloglock);
2939
2940 *logoffsetp = log_offset;
2941 return 0;
2942}
2943
2944/*
2945 * The first cnt-1 times a ticket goes through here we don't need to move the
2946 * grant write head because the permanent reservation has reserved cnt times the
2947 * unit amount. Release part of current permanent unit reservation and reset
2948 * current reservation to be one units worth. Also move grant reservation head
2949 * forward.
2950 */
2951void
2952xfs_log_ticket_regrant(
2953 struct xlog *log,
2954 struct xlog_ticket *ticket)
2955{
2956 trace_xfs_log_ticket_regrant(log, ticket);
2957
2958 if (ticket->t_cnt > 0)
2959 ticket->t_cnt--;
2960
2961 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2962 ticket->t_curr_res);
2963 xlog_grant_sub_space(log, &log->l_write_head.grant,
2964 ticket->t_curr_res);
2965 ticket->t_curr_res = ticket->t_unit_res;
2966 xlog_tic_reset_res(ticket);
2967
2968 trace_xfs_log_ticket_regrant_sub(log, ticket);
2969
2970 /* just return if we still have some of the pre-reserved space */
2971 if (!ticket->t_cnt) {
2972 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2973 ticket->t_unit_res);
2974 trace_xfs_log_ticket_regrant_exit(log, ticket);
2975
2976 ticket->t_curr_res = ticket->t_unit_res;
2977 xlog_tic_reset_res(ticket);
2978 }
2979
2980 xfs_log_ticket_put(ticket);
2981}
2982
2983/*
2984 * Give back the space left from a reservation.
2985 *
2986 * All the information we need to make a correct determination of space left
2987 * is present. For non-permanent reservations, things are quite easy. The
2988 * count should have been decremented to zero. We only need to deal with the
2989 * space remaining in the current reservation part of the ticket. If the
2990 * ticket contains a permanent reservation, there may be left over space which
2991 * needs to be released. A count of N means that N-1 refills of the current
2992 * reservation can be done before we need to ask for more space. The first
2993 * one goes to fill up the first current reservation. Once we run out of
2994 * space, the count will stay at zero and the only space remaining will be
2995 * in the current reservation field.
2996 */
2997void
2998xfs_log_ticket_ungrant(
2999 struct xlog *log,
3000 struct xlog_ticket *ticket)
3001{
3002 int bytes;
3003
3004 trace_xfs_log_ticket_ungrant(log, ticket);
3005
3006 if (ticket->t_cnt > 0)
3007 ticket->t_cnt--;
3008
3009 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3010
3011 /*
3012 * If this is a permanent reservation ticket, we may be able to free
3013 * up more space based on the remaining count.
3014 */
3015 bytes = ticket->t_curr_res;
3016 if (ticket->t_cnt > 0) {
3017 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3018 bytes += ticket->t_unit_res*ticket->t_cnt;
3019 }
3020
3021 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3022 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3023
3024 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3025
3026 xfs_log_space_wake(log->l_mp);
3027 xfs_log_ticket_put(ticket);
3028}
3029
3030/*
3031 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3032 * the current iclog pointer to the next iclog in the ring.
3033 */
3034STATIC void
3035xlog_state_switch_iclogs(
3036 struct xlog *log,
3037 struct xlog_in_core *iclog,
3038 int eventual_size)
3039{
3040 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3041 assert_spin_locked(&log->l_icloglock);
3042
3043 if (!eventual_size)
3044 eventual_size = iclog->ic_offset;
3045 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3046 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3047 log->l_prev_block = log->l_curr_block;
3048 log->l_prev_cycle = log->l_curr_cycle;
3049
3050 /* roll log?: ic_offset changed later */
3051 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3052
3053 /* Round up to next log-sunit */
3054 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3055 log->l_mp->m_sb.sb_logsunit > 1) {
3056 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3057 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3058 }
3059
3060 if (log->l_curr_block >= log->l_logBBsize) {
3061 /*
3062 * Rewind the current block before the cycle is bumped to make
3063 * sure that the combined LSN never transiently moves forward
3064 * when the log wraps to the next cycle. This is to support the
3065 * unlocked sample of these fields from xlog_valid_lsn(). Most
3066 * other cases should acquire l_icloglock.
3067 */
3068 log->l_curr_block -= log->l_logBBsize;
3069 ASSERT(log->l_curr_block >= 0);
3070 smp_wmb();
3071 log->l_curr_cycle++;
3072 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3073 log->l_curr_cycle++;
3074 }
3075 ASSERT(iclog == log->l_iclog);
3076 log->l_iclog = iclog->ic_next;
3077}
3078
3079/*
3080 * Write out all data in the in-core log as of this exact moment in time.
3081 *
3082 * Data may be written to the in-core log during this call. However,
3083 * we don't guarantee this data will be written out. A change from past
3084 * implementation means this routine will *not* write out zero length LRs.
3085 *
3086 * Basically, we try and perform an intelligent scan of the in-core logs.
3087 * If we determine there is no flushable data, we just return. There is no
3088 * flushable data if:
3089 *
3090 * 1. the current iclog is active and has no data; the previous iclog
3091 * is in the active or dirty state.
3092 * 2. the current iclog is drity, and the previous iclog is in the
3093 * active or dirty state.
3094 *
3095 * We may sleep if:
3096 *
3097 * 1. the current iclog is not in the active nor dirty state.
3098 * 2. the current iclog dirty, and the previous iclog is not in the
3099 * active nor dirty state.
3100 * 3. the current iclog is active, and there is another thread writing
3101 * to this particular iclog.
3102 * 4. a) the current iclog is active and has no other writers
3103 * b) when we return from flushing out this iclog, it is still
3104 * not in the active nor dirty state.
3105 */
3106int
3107xfs_log_force(
3108 struct xfs_mount *mp,
3109 uint flags)
3110{
3111 struct xlog *log = mp->m_log;
3112 struct xlog_in_core *iclog;
3113 xfs_lsn_t lsn;
3114
3115 XFS_STATS_INC(mp, xs_log_force);
3116 trace_xfs_log_force(mp, 0, _RET_IP_);
3117
3118 xlog_cil_force(log);
3119
3120 spin_lock(&log->l_icloglock);
3121 iclog = log->l_iclog;
3122 if (iclog->ic_state == XLOG_STATE_IOERROR)
3123 goto out_error;
3124
3125 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3126 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3127 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3128 /*
3129 * If the head is dirty or (active and empty), then we need to
3130 * look at the previous iclog.
3131 *
3132 * If the previous iclog is active or dirty we are done. There
3133 * is nothing to sync out. Otherwise, we attach ourselves to the
3134 * previous iclog and go to sleep.
3135 */
3136 iclog = iclog->ic_prev;
3137 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3138 if (atomic_read(&iclog->ic_refcnt) == 0) {
3139 /*
3140 * We are the only one with access to this iclog.
3141 *
3142 * Flush it out now. There should be a roundoff of zero
3143 * to show that someone has already taken care of the
3144 * roundoff from the previous sync.
3145 */
3146 atomic_inc(&iclog->ic_refcnt);
3147 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3148 xlog_state_switch_iclogs(log, iclog, 0);
3149 if (xlog_state_release_iclog(log, iclog))
3150 goto out_error;
3151
3152 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3153 goto out_unlock;
3154 } else {
3155 /*
3156 * Someone else is writing to this iclog.
3157 *
3158 * Use its call to flush out the data. However, the
3159 * other thread may not force out this LR, so we mark
3160 * it WANT_SYNC.
3161 */
3162 xlog_state_switch_iclogs(log, iclog, 0);
3163 }
3164 } else {
3165 /*
3166 * If the head iclog is not active nor dirty, we just attach
3167 * ourselves to the head and go to sleep if necessary.
3168 */
3169 ;
3170 }
3171
3172 if (flags & XFS_LOG_SYNC)
3173 return xlog_wait_on_iclog(iclog);
3174out_unlock:
3175 spin_unlock(&log->l_icloglock);
3176 return 0;
3177out_error:
3178 spin_unlock(&log->l_icloglock);
3179 return -EIO;
3180}
3181
3182static int
3183__xfs_log_force_lsn(
3184 struct xfs_mount *mp,
3185 xfs_lsn_t lsn,
3186 uint flags,
3187 int *log_flushed,
3188 bool already_slept)
3189{
3190 struct xlog *log = mp->m_log;
3191 struct xlog_in_core *iclog;
3192
3193 spin_lock(&log->l_icloglock);
3194 iclog = log->l_iclog;
3195 if (iclog->ic_state == XLOG_STATE_IOERROR)
3196 goto out_error;
3197
3198 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3199 iclog = iclog->ic_next;
3200 if (iclog == log->l_iclog)
3201 goto out_unlock;
3202 }
3203
3204 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3205 /*
3206 * We sleep here if we haven't already slept (e.g. this is the
3207 * first time we've looked at the correct iclog buf) and the
3208 * buffer before us is going to be sync'ed. The reason for this
3209 * is that if we are doing sync transactions here, by waiting
3210 * for the previous I/O to complete, we can allow a few more
3211 * transactions into this iclog before we close it down.
3212 *
3213 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3214 * refcnt so we can release the log (which drops the ref count).
3215 * The state switch keeps new transaction commits from using
3216 * this buffer. When the current commits finish writing into
3217 * the buffer, the refcount will drop to zero and the buffer
3218 * will go out then.
3219 */
3220 if (!already_slept &&
3221 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3222 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3223 XFS_STATS_INC(mp, xs_log_force_sleep);
3224
3225 xlog_wait(&iclog->ic_prev->ic_write_wait,
3226 &log->l_icloglock);
3227 return -EAGAIN;
3228 }
3229 atomic_inc(&iclog->ic_refcnt);
3230 xlog_state_switch_iclogs(log, iclog, 0);
3231 if (xlog_state_release_iclog(log, iclog))
3232 goto out_error;
3233 if (log_flushed)
3234 *log_flushed = 1;
3235 }
3236
3237 if (flags & XFS_LOG_SYNC)
3238 return xlog_wait_on_iclog(iclog);
3239out_unlock:
3240 spin_unlock(&log->l_icloglock);
3241 return 0;
3242out_error:
3243 spin_unlock(&log->l_icloglock);
3244 return -EIO;
3245}
3246
3247/*
3248 * Force the in-core log to disk for a specific LSN.
3249 *
3250 * Find in-core log with lsn.
3251 * If it is in the DIRTY state, just return.
3252 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3253 * state and go to sleep or return.
3254 * If it is in any other state, go to sleep or return.
3255 *
3256 * Synchronous forces are implemented with a wait queue. All callers trying
3257 * to force a given lsn to disk must wait on the queue attached to the
3258 * specific in-core log. When given in-core log finally completes its write
3259 * to disk, that thread will wake up all threads waiting on the queue.
3260 */
3261int
3262xfs_log_force_lsn(
3263 struct xfs_mount *mp,
3264 xfs_lsn_t lsn,
3265 uint flags,
3266 int *log_flushed)
3267{
3268 int ret;
3269 ASSERT(lsn != 0);
3270
3271 XFS_STATS_INC(mp, xs_log_force);
3272 trace_xfs_log_force(mp, lsn, _RET_IP_);
3273
3274 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3275 if (lsn == NULLCOMMITLSN)
3276 return 0;
3277
3278 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3279 if (ret == -EAGAIN)
3280 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3281 return ret;
3282}
3283
3284/*
3285 * Free a used ticket when its refcount falls to zero.
3286 */
3287void
3288xfs_log_ticket_put(
3289 xlog_ticket_t *ticket)
3290{
3291 ASSERT(atomic_read(&ticket->t_ref) > 0);
3292 if (atomic_dec_and_test(&ticket->t_ref))
3293 kmem_cache_free(xfs_log_ticket_zone, ticket);
3294}
3295
3296xlog_ticket_t *
3297xfs_log_ticket_get(
3298 xlog_ticket_t *ticket)
3299{
3300 ASSERT(atomic_read(&ticket->t_ref) > 0);
3301 atomic_inc(&ticket->t_ref);
3302 return ticket;
3303}
3304
3305/*
3306 * Figure out the total log space unit (in bytes) that would be
3307 * required for a log ticket.
3308 */
3309int
3310xfs_log_calc_unit_res(
3311 struct xfs_mount *mp,
3312 int unit_bytes)
3313{
3314 struct xlog *log = mp->m_log;
3315 int iclog_space;
3316 uint num_headers;
3317
3318 /*
3319 * Permanent reservations have up to 'cnt'-1 active log operations
3320 * in the log. A unit in this case is the amount of space for one
3321 * of these log operations. Normal reservations have a cnt of 1
3322 * and their unit amount is the total amount of space required.
3323 *
3324 * The following lines of code account for non-transaction data
3325 * which occupy space in the on-disk log.
3326 *
3327 * Normal form of a transaction is:
3328 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3329 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3330 *
3331 * We need to account for all the leadup data and trailer data
3332 * around the transaction data.
3333 * And then we need to account for the worst case in terms of using
3334 * more space.
3335 * The worst case will happen if:
3336 * - the placement of the transaction happens to be such that the
3337 * roundoff is at its maximum
3338 * - the transaction data is synced before the commit record is synced
3339 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3340 * Therefore the commit record is in its own Log Record.
3341 * This can happen as the commit record is called with its
3342 * own region to xlog_write().
3343 * This then means that in the worst case, roundoff can happen for
3344 * the commit-rec as well.
3345 * The commit-rec is smaller than padding in this scenario and so it is
3346 * not added separately.
3347 */
3348
3349 /* for trans header */
3350 unit_bytes += sizeof(xlog_op_header_t);
3351 unit_bytes += sizeof(xfs_trans_header_t);
3352
3353 /* for start-rec */
3354 unit_bytes += sizeof(xlog_op_header_t);
3355
3356 /*
3357 * for LR headers - the space for data in an iclog is the size minus
3358 * the space used for the headers. If we use the iclog size, then we
3359 * undercalculate the number of headers required.
3360 *
3361 * Furthermore - the addition of op headers for split-recs might
3362 * increase the space required enough to require more log and op
3363 * headers, so take that into account too.
3364 *
3365 * IMPORTANT: This reservation makes the assumption that if this
3366 * transaction is the first in an iclog and hence has the LR headers
3367 * accounted to it, then the remaining space in the iclog is
3368 * exclusively for this transaction. i.e. if the transaction is larger
3369 * than the iclog, it will be the only thing in that iclog.
3370 * Fundamentally, this means we must pass the entire log vector to
3371 * xlog_write to guarantee this.
3372 */
3373 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3374 num_headers = howmany(unit_bytes, iclog_space);
3375
3376 /* for split-recs - ophdrs added when data split over LRs */
3377 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3378
3379 /* add extra header reservations if we overrun */
3380 while (!num_headers ||
3381 howmany(unit_bytes, iclog_space) > num_headers) {
3382 unit_bytes += sizeof(xlog_op_header_t);
3383 num_headers++;
3384 }
3385 unit_bytes += log->l_iclog_hsize * num_headers;
3386
3387 /* for commit-rec LR header - note: padding will subsume the ophdr */
3388 unit_bytes += log->l_iclog_hsize;
3389
3390 /* for roundoff padding for transaction data and one for commit record */
3391 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3392 /* log su roundoff */
3393 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3394 } else {
3395 /* BB roundoff */
3396 unit_bytes += 2 * BBSIZE;
3397 }
3398
3399 return unit_bytes;
3400}
3401
3402/*
3403 * Allocate and initialise a new log ticket.
3404 */
3405struct xlog_ticket *
3406xlog_ticket_alloc(
3407 struct xlog *log,
3408 int unit_bytes,
3409 int cnt,
3410 char client,
3411 bool permanent)
3412{
3413 struct xlog_ticket *tic;
3414 int unit_res;
3415
3416 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3417
3418 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3419
3420 atomic_set(&tic->t_ref, 1);
3421 tic->t_task = current;
3422 INIT_LIST_HEAD(&tic->t_queue);
3423 tic->t_unit_res = unit_res;
3424 tic->t_curr_res = unit_res;
3425 tic->t_cnt = cnt;
3426 tic->t_ocnt = cnt;
3427 tic->t_tid = prandom_u32();
3428 tic->t_clientid = client;
3429 if (permanent)
3430 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3431
3432 xlog_tic_reset_res(tic);
3433
3434 return tic;
3435}
3436
3437#if defined(DEBUG)
3438/*
3439 * Make sure that the destination ptr is within the valid data region of
3440 * one of the iclogs. This uses backup pointers stored in a different
3441 * part of the log in case we trash the log structure.
3442 */
3443STATIC void
3444xlog_verify_dest_ptr(
3445 struct xlog *log,
3446 void *ptr)
3447{
3448 int i;
3449 int good_ptr = 0;
3450
3451 for (i = 0; i < log->l_iclog_bufs; i++) {
3452 if (ptr >= log->l_iclog_bak[i] &&
3453 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3454 good_ptr++;
3455 }
3456
3457 if (!good_ptr)
3458 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3459}
3460
3461/*
3462 * Check to make sure the grant write head didn't just over lap the tail. If
3463 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3464 * the cycles differ by exactly one and check the byte count.
3465 *
3466 * This check is run unlocked, so can give false positives. Rather than assert
3467 * on failures, use a warn-once flag and a panic tag to allow the admin to
3468 * determine if they want to panic the machine when such an error occurs. For
3469 * debug kernels this will have the same effect as using an assert but, unlinke
3470 * an assert, it can be turned off at runtime.
3471 */
3472STATIC void
3473xlog_verify_grant_tail(
3474 struct xlog *log)
3475{
3476 int tail_cycle, tail_blocks;
3477 int cycle, space;
3478
3479 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3480 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3481 if (tail_cycle != cycle) {
3482 if (cycle - 1 != tail_cycle &&
3483 !(log->l_flags & XLOG_TAIL_WARN)) {
3484 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3485 "%s: cycle - 1 != tail_cycle", __func__);
3486 log->l_flags |= XLOG_TAIL_WARN;
3487 }
3488
3489 if (space > BBTOB(tail_blocks) &&
3490 !(log->l_flags & XLOG_TAIL_WARN)) {
3491 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3492 "%s: space > BBTOB(tail_blocks)", __func__);
3493 log->l_flags |= XLOG_TAIL_WARN;
3494 }
3495 }
3496}
3497
3498/* check if it will fit */
3499STATIC void
3500xlog_verify_tail_lsn(
3501 struct xlog *log,
3502 struct xlog_in_core *iclog,
3503 xfs_lsn_t tail_lsn)
3504{
3505 int blocks;
3506
3507 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3508 blocks =
3509 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3510 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3511 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3512 } else {
3513 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3514
3515 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3516 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3517
3518 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3519 if (blocks < BTOBB(iclog->ic_offset) + 1)
3520 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3521 }
3522}
3523
3524/*
3525 * Perform a number of checks on the iclog before writing to disk.
3526 *
3527 * 1. Make sure the iclogs are still circular
3528 * 2. Make sure we have a good magic number
3529 * 3. Make sure we don't have magic numbers in the data
3530 * 4. Check fields of each log operation header for:
3531 * A. Valid client identifier
3532 * B. tid ptr value falls in valid ptr space (user space code)
3533 * C. Length in log record header is correct according to the
3534 * individual operation headers within record.
3535 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3536 * log, check the preceding blocks of the physical log to make sure all
3537 * the cycle numbers agree with the current cycle number.
3538 */
3539STATIC void
3540xlog_verify_iclog(
3541 struct xlog *log,
3542 struct xlog_in_core *iclog,
3543 int count)
3544{
3545 xlog_op_header_t *ophead;
3546 xlog_in_core_t *icptr;
3547 xlog_in_core_2_t *xhdr;
3548 void *base_ptr, *ptr, *p;
3549 ptrdiff_t field_offset;
3550 uint8_t clientid;
3551 int len, i, j, k, op_len;
3552 int idx;
3553
3554 /* check validity of iclog pointers */
3555 spin_lock(&log->l_icloglock);
3556 icptr = log->l_iclog;
3557 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3558 ASSERT(icptr);
3559
3560 if (icptr != log->l_iclog)
3561 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3562 spin_unlock(&log->l_icloglock);
3563
3564 /* check log magic numbers */
3565 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3566 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3567
3568 base_ptr = ptr = &iclog->ic_header;
3569 p = &iclog->ic_header;
3570 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3571 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3572 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3573 __func__);
3574 }
3575
3576 /* check fields */
3577 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3578 base_ptr = ptr = iclog->ic_datap;
3579 ophead = ptr;
3580 xhdr = iclog->ic_data;
3581 for (i = 0; i < len; i++) {
3582 ophead = ptr;
3583
3584 /* clientid is only 1 byte */
3585 p = &ophead->oh_clientid;
3586 field_offset = p - base_ptr;
3587 if (field_offset & 0x1ff) {
3588 clientid = ophead->oh_clientid;
3589 } else {
3590 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3591 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3592 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3593 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3594 clientid = xlog_get_client_id(
3595 xhdr[j].hic_xheader.xh_cycle_data[k]);
3596 } else {
3597 clientid = xlog_get_client_id(
3598 iclog->ic_header.h_cycle_data[idx]);
3599 }
3600 }
3601 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3602 xfs_warn(log->l_mp,
3603 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3604 __func__, clientid, ophead,
3605 (unsigned long)field_offset);
3606
3607 /* check length */
3608 p = &ophead->oh_len;
3609 field_offset = p - base_ptr;
3610 if (field_offset & 0x1ff) {
3611 op_len = be32_to_cpu(ophead->oh_len);
3612 } else {
3613 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3614 (uintptr_t)iclog->ic_datap);
3615 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3616 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3617 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3618 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3619 } else {
3620 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3621 }
3622 }
3623 ptr += sizeof(xlog_op_header_t) + op_len;
3624 }
3625}
3626#endif
3627
3628/*
3629 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3630 */
3631STATIC int
3632xlog_state_ioerror(
3633 struct xlog *log)
3634{
3635 xlog_in_core_t *iclog, *ic;
3636
3637 iclog = log->l_iclog;
3638 if (iclog->ic_state != XLOG_STATE_IOERROR) {
3639 /*
3640 * Mark all the incore logs IOERROR.
3641 * From now on, no log flushes will result.
3642 */
3643 ic = iclog;
3644 do {
3645 ic->ic_state = XLOG_STATE_IOERROR;
3646 ic = ic->ic_next;
3647 } while (ic != iclog);
3648 return 0;
3649 }
3650 /*
3651 * Return non-zero, if state transition has already happened.
3652 */
3653 return 1;
3654}
3655
3656/*
3657 * This is called from xfs_force_shutdown, when we're forcibly
3658 * shutting down the filesystem, typically because of an IO error.
3659 * Our main objectives here are to make sure that:
3660 * a. if !logerror, flush the logs to disk. Anything modified
3661 * after this is ignored.
3662 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3663 * parties to find out, 'atomically'.
3664 * c. those who're sleeping on log reservations, pinned objects and
3665 * other resources get woken up, and be told the bad news.
3666 * d. nothing new gets queued up after (b) and (c) are done.
3667 *
3668 * Note: for the !logerror case we need to flush the regions held in memory out
3669 * to disk first. This needs to be done before the log is marked as shutdown,
3670 * otherwise the iclog writes will fail.
3671 */
3672int
3673xfs_log_force_umount(
3674 struct xfs_mount *mp,
3675 int logerror)
3676{
3677 struct xlog *log;
3678 int retval;
3679
3680 log = mp->m_log;
3681
3682 /*
3683 * If this happens during log recovery, don't worry about
3684 * locking; the log isn't open for business yet.
3685 */
3686 if (!log ||
3687 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3688 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3689 if (mp->m_sb_bp)
3690 mp->m_sb_bp->b_flags |= XBF_DONE;
3691 return 0;
3692 }
3693
3694 /*
3695 * Somebody could've already done the hard work for us.
3696 * No need to get locks for this.
3697 */
3698 if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3699 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3700 return 1;
3701 }
3702
3703 /*
3704 * Flush all the completed transactions to disk before marking the log
3705 * being shut down. We need to do it in this order to ensure that
3706 * completed operations are safely on disk before we shut down, and that
3707 * we don't have to issue any buffer IO after the shutdown flags are set
3708 * to guarantee this.
3709 */
3710 if (!logerror)
3711 xfs_log_force(mp, XFS_LOG_SYNC);
3712
3713 /*
3714 * mark the filesystem and the as in a shutdown state and wake
3715 * everybody up to tell them the bad news.
3716 */
3717 spin_lock(&log->l_icloglock);
3718 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3719 if (mp->m_sb_bp)
3720 mp->m_sb_bp->b_flags |= XBF_DONE;
3721
3722 /*
3723 * Mark the log and the iclogs with IO error flags to prevent any
3724 * further log IO from being issued or completed.
3725 */
3726 log->l_flags |= XLOG_IO_ERROR;
3727 retval = xlog_state_ioerror(log);
3728 spin_unlock(&log->l_icloglock);
3729
3730 /*
3731 * We don't want anybody waiting for log reservations after this. That
3732 * means we have to wake up everybody queued up on reserveq as well as
3733 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3734 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3735 * action is protected by the grant locks.
3736 */
3737 xlog_grant_head_wake_all(&log->l_reserve_head);
3738 xlog_grant_head_wake_all(&log->l_write_head);
3739
3740 /*
3741 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3742 * as if the log writes were completed. The abort handling in the log
3743 * item committed callback functions will do this again under lock to
3744 * avoid races.
3745 */
3746 spin_lock(&log->l_cilp->xc_push_lock);
3747 wake_up_all(&log->l_cilp->xc_commit_wait);
3748 spin_unlock(&log->l_cilp->xc_push_lock);
3749 xlog_state_do_callback(log);
3750
3751 /* return non-zero if log IOERROR transition had already happened */
3752 return retval;
3753}
3754
3755STATIC int
3756xlog_iclogs_empty(
3757 struct xlog *log)
3758{
3759 xlog_in_core_t *iclog;
3760
3761 iclog = log->l_iclog;
3762 do {
3763 /* endianness does not matter here, zero is zero in
3764 * any language.
3765 */
3766 if (iclog->ic_header.h_num_logops)
3767 return 0;
3768 iclog = iclog->ic_next;
3769 } while (iclog != log->l_iclog);
3770 return 1;
3771}
3772
3773/*
3774 * Verify that an LSN stamped into a piece of metadata is valid. This is
3775 * intended for use in read verifiers on v5 superblocks.
3776 */
3777bool
3778xfs_log_check_lsn(
3779 struct xfs_mount *mp,
3780 xfs_lsn_t lsn)
3781{
3782 struct xlog *log = mp->m_log;
3783 bool valid;
3784
3785 /*
3786 * norecovery mode skips mount-time log processing and unconditionally
3787 * resets the in-core LSN. We can't validate in this mode, but
3788 * modifications are not allowed anyways so just return true.
3789 */
3790 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3791 return true;
3792
3793 /*
3794 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3795 * handled by recovery and thus safe to ignore here.
3796 */
3797 if (lsn == NULLCOMMITLSN)
3798 return true;
3799
3800 valid = xlog_valid_lsn(mp->m_log, lsn);
3801
3802 /* warn the user about what's gone wrong before verifier failure */
3803 if (!valid) {
3804 spin_lock(&log->l_icloglock);
3805 xfs_warn(mp,
3806"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3807"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3808 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3809 log->l_curr_cycle, log->l_curr_block);
3810 spin_unlock(&log->l_icloglock);
3811 }
3812
3813 return valid;
3814}
3815
3816bool
3817xfs_log_in_recovery(
3818 struct xfs_mount *mp)
3819{
3820 struct xlog *log = mp->m_log;
3821
3822 return log->l_flags & XLOG_ACTIVE_RECOVERY;
3823}