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#else
83#define xlog_verify_dest_ptr(a,b)
84#define xlog_verify_grant_tail(a)
85#define xlog_verify_iclog(a,b,c)
86#define xlog_verify_tail_lsn(a,b)
87#endif
88
89STATIC int
90xlog_iclogs_empty(
91 struct xlog *log);
92
93static int
94xfs_log_cover(struct xfs_mount *);
95
96static void
97xlog_grant_sub_space(
98 struct xlog *log,
99 atomic64_t *head,
100 int bytes)
101{
102 int64_t head_val = atomic64_read(head);
103 int64_t new, old;
104
105 do {
106 int cycle, space;
107
108 xlog_crack_grant_head_val(head_val, &cycle, &space);
109
110 space -= bytes;
111 if (space < 0) {
112 space += log->l_logsize;
113 cycle--;
114 }
115
116 old = head_val;
117 new = xlog_assign_grant_head_val(cycle, space);
118 head_val = atomic64_cmpxchg(head, old, new);
119 } while (head_val != old);
120}
121
122static void
123xlog_grant_add_space(
124 struct xlog *log,
125 atomic64_t *head,
126 int bytes)
127{
128 int64_t head_val = atomic64_read(head);
129 int64_t new, old;
130
131 do {
132 int tmp;
133 int cycle, space;
134
135 xlog_crack_grant_head_val(head_val, &cycle, &space);
136
137 tmp = log->l_logsize - space;
138 if (tmp > bytes)
139 space += bytes;
140 else {
141 space = bytes - tmp;
142 cycle++;
143 }
144
145 old = head_val;
146 new = xlog_assign_grant_head_val(cycle, space);
147 head_val = atomic64_cmpxchg(head, old, new);
148 } while (head_val != old);
149}
150
151STATIC void
152xlog_grant_head_init(
153 struct xlog_grant_head *head)
154{
155 xlog_assign_grant_head(&head->grant, 1, 0);
156 INIT_LIST_HEAD(&head->waiters);
157 spin_lock_init(&head->lock);
158}
159
160STATIC void
161xlog_grant_head_wake_all(
162 struct xlog_grant_head *head)
163{
164 struct xlog_ticket *tic;
165
166 spin_lock(&head->lock);
167 list_for_each_entry(tic, &head->waiters, t_queue)
168 wake_up_process(tic->t_task);
169 spin_unlock(&head->lock);
170}
171
172static inline int
173xlog_ticket_reservation(
174 struct xlog *log,
175 struct xlog_grant_head *head,
176 struct xlog_ticket *tic)
177{
178 if (head == &log->l_write_head) {
179 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
180 return tic->t_unit_res;
181 } else {
182 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
183 return tic->t_unit_res * tic->t_cnt;
184 else
185 return tic->t_unit_res;
186 }
187}
188
189STATIC bool
190xlog_grant_head_wake(
191 struct xlog *log,
192 struct xlog_grant_head *head,
193 int *free_bytes)
194{
195 struct xlog_ticket *tic;
196 int need_bytes;
197 bool woken_task = false;
198
199 list_for_each_entry(tic, &head->waiters, t_queue) {
200
201 /*
202 * There is a chance that the size of the CIL checkpoints in
203 * progress at the last AIL push target calculation resulted in
204 * limiting the target to the log head (l_last_sync_lsn) at the
205 * time. This may not reflect where the log head is now as the
206 * CIL checkpoints may have completed.
207 *
208 * Hence when we are woken here, it may be that the head of the
209 * log that has moved rather than the tail. As the tail didn't
210 * move, there still won't be space available for the
211 * reservation we require. However, if the AIL has already
212 * pushed to the target defined by the old log head location, we
213 * will hang here waiting for something else to update the AIL
214 * push target.
215 *
216 * Therefore, if there isn't space to wake the first waiter on
217 * the grant head, we need to push the AIL again to ensure the
218 * target reflects both the current log tail and log head
219 * position before we wait for the tail to move again.
220 */
221
222 need_bytes = xlog_ticket_reservation(log, head, tic);
223 if (*free_bytes < need_bytes) {
224 if (!woken_task)
225 xlog_grant_push_ail(log, need_bytes);
226 return false;
227 }
228
229 *free_bytes -= need_bytes;
230 trace_xfs_log_grant_wake_up(log, tic);
231 wake_up_process(tic->t_task);
232 woken_task = true;
233 }
234
235 return true;
236}
237
238STATIC int
239xlog_grant_head_wait(
240 struct xlog *log,
241 struct xlog_grant_head *head,
242 struct xlog_ticket *tic,
243 int need_bytes) __releases(&head->lock)
244 __acquires(&head->lock)
245{
246 list_add_tail(&tic->t_queue, &head->waiters);
247
248 do {
249 if (XLOG_FORCED_SHUTDOWN(log))
250 goto shutdown;
251 xlog_grant_push_ail(log, need_bytes);
252
253 __set_current_state(TASK_UNINTERRUPTIBLE);
254 spin_unlock(&head->lock);
255
256 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
257
258 trace_xfs_log_grant_sleep(log, tic);
259 schedule();
260 trace_xfs_log_grant_wake(log, tic);
261
262 spin_lock(&head->lock);
263 if (XLOG_FORCED_SHUTDOWN(log))
264 goto shutdown;
265 } while (xlog_space_left(log, &head->grant) < need_bytes);
266
267 list_del_init(&tic->t_queue);
268 return 0;
269shutdown:
270 list_del_init(&tic->t_queue);
271 return -EIO;
272}
273
274/*
275 * Atomically get the log space required for a log ticket.
276 *
277 * Once a ticket gets put onto head->waiters, it will only return after the
278 * needed reservation is satisfied.
279 *
280 * This function is structured so that it has a lock free fast path. This is
281 * necessary because every new transaction reservation will come through this
282 * path. Hence any lock will be globally hot if we take it unconditionally on
283 * every pass.
284 *
285 * As tickets are only ever moved on and off head->waiters under head->lock, we
286 * only need to take that lock if we are going to add the ticket to the queue
287 * and sleep. We can avoid taking the lock if the ticket was never added to
288 * head->waiters because the t_queue list head will be empty and we hold the
289 * only reference to it so it can safely be checked unlocked.
290 */
291STATIC int
292xlog_grant_head_check(
293 struct xlog *log,
294 struct xlog_grant_head *head,
295 struct xlog_ticket *tic,
296 int *need_bytes)
297{
298 int free_bytes;
299 int error = 0;
300
301 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
302
303 /*
304 * If there are other waiters on the queue then give them a chance at
305 * logspace before us. Wake up the first waiters, if we do not wake
306 * up all the waiters then go to sleep waiting for more free space,
307 * otherwise try to get some space for this transaction.
308 */
309 *need_bytes = xlog_ticket_reservation(log, head, tic);
310 free_bytes = xlog_space_left(log, &head->grant);
311 if (!list_empty_careful(&head->waiters)) {
312 spin_lock(&head->lock);
313 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
314 free_bytes < *need_bytes) {
315 error = xlog_grant_head_wait(log, head, tic,
316 *need_bytes);
317 }
318 spin_unlock(&head->lock);
319 } else if (free_bytes < *need_bytes) {
320 spin_lock(&head->lock);
321 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
322 spin_unlock(&head->lock);
323 }
324
325 return error;
326}
327
328static void
329xlog_tic_reset_res(xlog_ticket_t *tic)
330{
331 tic->t_res_num = 0;
332 tic->t_res_arr_sum = 0;
333 tic->t_res_num_ophdrs = 0;
334}
335
336static void
337xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
338{
339 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
340 /* add to overflow and start again */
341 tic->t_res_o_flow += tic->t_res_arr_sum;
342 tic->t_res_num = 0;
343 tic->t_res_arr_sum = 0;
344 }
345
346 tic->t_res_arr[tic->t_res_num].r_len = len;
347 tic->t_res_arr[tic->t_res_num].r_type = type;
348 tic->t_res_arr_sum += len;
349 tic->t_res_num++;
350}
351
352bool
353xfs_log_writable(
354 struct xfs_mount *mp)
355{
356 /*
357 * Do not write to the log on norecovery mounts, if the data or log
358 * devices are read-only, or if the filesystem is shutdown. Read-only
359 * mounts allow internal writes for log recovery and unmount purposes,
360 * so don't restrict that case.
361 */
362 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
363 return false;
364 if (xfs_readonly_buftarg(mp->m_ddev_targp))
365 return false;
366 if (xfs_readonly_buftarg(mp->m_log->l_targ))
367 return false;
368 if (XFS_FORCED_SHUTDOWN(mp))
369 return false;
370 return true;
371}
372
373/*
374 * Replenish the byte reservation required by moving the grant write head.
375 */
376int
377xfs_log_regrant(
378 struct xfs_mount *mp,
379 struct xlog_ticket *tic)
380{
381 struct xlog *log = mp->m_log;
382 int need_bytes;
383 int error = 0;
384
385 if (XLOG_FORCED_SHUTDOWN(log))
386 return -EIO;
387
388 XFS_STATS_INC(mp, xs_try_logspace);
389
390 /*
391 * This is a new transaction on the ticket, so we need to change the
392 * transaction ID so that the next transaction has a different TID in
393 * the log. Just add one to the existing tid so that we can see chains
394 * of rolling transactions in the log easily.
395 */
396 tic->t_tid++;
397
398 xlog_grant_push_ail(log, tic->t_unit_res);
399
400 tic->t_curr_res = tic->t_unit_res;
401 xlog_tic_reset_res(tic);
402
403 if (tic->t_cnt > 0)
404 return 0;
405
406 trace_xfs_log_regrant(log, tic);
407
408 error = xlog_grant_head_check(log, &log->l_write_head, tic,
409 &need_bytes);
410 if (error)
411 goto out_error;
412
413 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
414 trace_xfs_log_regrant_exit(log, tic);
415 xlog_verify_grant_tail(log);
416 return 0;
417
418out_error:
419 /*
420 * If we are failing, make sure the ticket doesn't have any current
421 * reservations. We don't want to add this back when the ticket/
422 * transaction gets cancelled.
423 */
424 tic->t_curr_res = 0;
425 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
426 return error;
427}
428
429/*
430 * Reserve log space and return a ticket corresponding to the reservation.
431 *
432 * Each reservation is going to reserve extra space for a log record header.
433 * When writes happen to the on-disk log, we don't subtract the length of the
434 * log record header from any reservation. By wasting space in each
435 * reservation, we prevent over allocation problems.
436 */
437int
438xfs_log_reserve(
439 struct xfs_mount *mp,
440 int unit_bytes,
441 int cnt,
442 struct xlog_ticket **ticp,
443 uint8_t client,
444 bool permanent)
445{
446 struct xlog *log = mp->m_log;
447 struct xlog_ticket *tic;
448 int need_bytes;
449 int error = 0;
450
451 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
452
453 if (XLOG_FORCED_SHUTDOWN(log))
454 return -EIO;
455
456 XFS_STATS_INC(mp, xs_try_logspace);
457
458 ASSERT(*ticp == NULL);
459 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
460 *ticp = tic;
461
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
463 : tic->t_unit_res);
464
465 trace_xfs_log_reserve(log, tic);
466
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
468 &need_bytes);
469 if (error)
470 goto out_error;
471
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
476 return 0;
477
478out_error:
479 /*
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
483 */
484 tic->t_curr_res = 0;
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
486 return error;
487}
488
489/*
490 * Flush iclog to disk if this is the last reference to the given iclog and the
491 * it is in the WANT_SYNC state.
492 *
493 * If the caller passes in a non-zero @old_tail_lsn and the current log tail
494 * does not match, there may be metadata on disk that must be persisted before
495 * this iclog is written. To satisfy that requirement, set the
496 * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
497 * log tail value.
498 *
499 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
500 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
501 * written to stable storage, and implies that a commit record is contained
502 * within the iclog. We need to ensure that the log tail does not move beyond
503 * the tail that the first commit record in the iclog ordered against, otherwise
504 * correct recovery of that checkpoint becomes dependent on future operations
505 * performed on this iclog.
506 *
507 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
508 * current tail into iclog. Once the iclog tail is set, future operations must
509 * not modify it, otherwise they potentially violate ordering constraints for
510 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
511 * the iclog will get zeroed on activation of the iclog after sync, so we
512 * always capture the tail lsn on the iclog on the first NEED_FUA release
513 * regardless of the number of active reference counts on this iclog.
514 */
515
516int
517xlog_state_release_iclog(
518 struct xlog *log,
519 struct xlog_in_core *iclog,
520 xfs_lsn_t old_tail_lsn)
521{
522 xfs_lsn_t tail_lsn;
523 lockdep_assert_held(&log->l_icloglock);
524
525 trace_xlog_iclog_release(iclog, _RET_IP_);
526 if (iclog->ic_state == XLOG_STATE_IOERROR)
527 return -EIO;
528
529 /*
530 * Grabbing the current log tail needs to be atomic w.r.t. the writing
531 * of the tail LSN into the iclog so we guarantee that the log tail does
532 * not move between deciding if a cache flush is required and writing
533 * the LSN into the iclog below.
534 */
535 if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
536 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
537
538 if (old_tail_lsn && tail_lsn != old_tail_lsn)
539 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
540
541 if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
542 !iclog->ic_header.h_tail_lsn)
543 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
544 }
545
546 if (!atomic_dec_and_test(&iclog->ic_refcnt))
547 return 0;
548
549 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
550 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
551 return 0;
552 }
553
554 iclog->ic_state = XLOG_STATE_SYNCING;
555 if (!iclog->ic_header.h_tail_lsn)
556 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
557 xlog_verify_tail_lsn(log, iclog);
558 trace_xlog_iclog_syncing(iclog, _RET_IP_);
559
560 spin_unlock(&log->l_icloglock);
561 xlog_sync(log, iclog);
562 spin_lock(&log->l_icloglock);
563 return 0;
564}
565
566/*
567 * Mount a log filesystem
568 *
569 * mp - ubiquitous xfs mount point structure
570 * log_target - buftarg of on-disk log device
571 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
572 * num_bblocks - Number of BBSIZE blocks in on-disk log
573 *
574 * Return error or zero.
575 */
576int
577xfs_log_mount(
578 xfs_mount_t *mp,
579 xfs_buftarg_t *log_target,
580 xfs_daddr_t blk_offset,
581 int num_bblks)
582{
583 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
584 int error = 0;
585 int min_logfsbs;
586
587 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
588 xfs_notice(mp, "Mounting V%d Filesystem",
589 XFS_SB_VERSION_NUM(&mp->m_sb));
590 } else {
591 xfs_notice(mp,
592"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
593 XFS_SB_VERSION_NUM(&mp->m_sb));
594 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
595 }
596
597 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
598 if (IS_ERR(mp->m_log)) {
599 error = PTR_ERR(mp->m_log);
600 goto out;
601 }
602
603 /*
604 * Validate the given log space and drop a critical message via syslog
605 * if the log size is too small that would lead to some unexpected
606 * situations in transaction log space reservation stage.
607 *
608 * Note: we can't just reject the mount if the validation fails. This
609 * would mean that people would have to downgrade their kernel just to
610 * remedy the situation as there is no way to grow the log (short of
611 * black magic surgery with xfs_db).
612 *
613 * We can, however, reject mounts for CRC format filesystems, as the
614 * mkfs binary being used to make the filesystem should never create a
615 * filesystem with a log that is too small.
616 */
617 min_logfsbs = xfs_log_calc_minimum_size(mp);
618
619 if (mp->m_sb.sb_logblocks < min_logfsbs) {
620 xfs_warn(mp,
621 "Log size %d blocks too small, minimum size is %d blocks",
622 mp->m_sb.sb_logblocks, min_logfsbs);
623 error = -EINVAL;
624 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
625 xfs_warn(mp,
626 "Log size %d blocks too large, maximum size is %lld blocks",
627 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
628 error = -EINVAL;
629 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
630 xfs_warn(mp,
631 "log size %lld bytes too large, maximum size is %lld bytes",
632 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
633 XFS_MAX_LOG_BYTES);
634 error = -EINVAL;
635 } else if (mp->m_sb.sb_logsunit > 1 &&
636 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
637 xfs_warn(mp,
638 "log stripe unit %u bytes must be a multiple of block size",
639 mp->m_sb.sb_logsunit);
640 error = -EINVAL;
641 fatal = true;
642 }
643 if (error) {
644 /*
645 * Log check errors are always fatal on v5; or whenever bad
646 * metadata leads to a crash.
647 */
648 if (fatal) {
649 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
650 ASSERT(0);
651 goto out_free_log;
652 }
653 xfs_crit(mp, "Log size out of supported range.");
654 xfs_crit(mp,
655"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
656 }
657
658 /*
659 * Initialize the AIL now we have a log.
660 */
661 error = xfs_trans_ail_init(mp);
662 if (error) {
663 xfs_warn(mp, "AIL initialisation failed: error %d", error);
664 goto out_free_log;
665 }
666 mp->m_log->l_ailp = mp->m_ail;
667
668 /*
669 * skip log recovery on a norecovery mount. pretend it all
670 * just worked.
671 */
672 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
673 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
674
675 if (readonly)
676 mp->m_flags &= ~XFS_MOUNT_RDONLY;
677
678 error = xlog_recover(mp->m_log);
679
680 if (readonly)
681 mp->m_flags |= XFS_MOUNT_RDONLY;
682 if (error) {
683 xfs_warn(mp, "log mount/recovery failed: error %d",
684 error);
685 xlog_recover_cancel(mp->m_log);
686 goto out_destroy_ail;
687 }
688 }
689
690 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
691 "log");
692 if (error)
693 goto out_destroy_ail;
694
695 /* Normal transactions can now occur */
696 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
697
698 /*
699 * Now the log has been fully initialised and we know were our
700 * space grant counters are, we can initialise the permanent ticket
701 * needed for delayed logging to work.
702 */
703 xlog_cil_init_post_recovery(mp->m_log);
704
705 return 0;
706
707out_destroy_ail:
708 xfs_trans_ail_destroy(mp);
709out_free_log:
710 xlog_dealloc_log(mp->m_log);
711out:
712 return error;
713}
714
715/*
716 * Finish the recovery of the file system. This is separate from the
717 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
718 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
719 * here.
720 *
721 * If we finish recovery successfully, start the background log work. If we are
722 * not doing recovery, then we have a RO filesystem and we don't need to start
723 * it.
724 */
725int
726xfs_log_mount_finish(
727 struct xfs_mount *mp)
728{
729 int error = 0;
730 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
731 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
732
733 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
734 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
735 return 0;
736 } else if (readonly) {
737 /* Allow unlinked processing to proceed */
738 mp->m_flags &= ~XFS_MOUNT_RDONLY;
739 }
740
741 /*
742 * During the second phase of log recovery, we need iget and
743 * iput to behave like they do for an active filesystem.
744 * xfs_fs_drop_inode needs to be able to prevent the deletion
745 * of inodes before we're done replaying log items on those
746 * inodes. Turn it off immediately after recovery finishes
747 * so that we don't leak the quota inodes if subsequent mount
748 * activities fail.
749 *
750 * We let all inodes involved in redo item processing end up on
751 * the LRU instead of being evicted immediately so that if we do
752 * something to an unlinked inode, the irele won't cause
753 * premature truncation and freeing of the inode, which results
754 * in log recovery failure. We have to evict the unreferenced
755 * lru inodes after clearing SB_ACTIVE because we don't
756 * otherwise clean up the lru if there's a subsequent failure in
757 * xfs_mountfs, which leads to us leaking the inodes if nothing
758 * else (e.g. quotacheck) references the inodes before the
759 * mount failure occurs.
760 */
761 mp->m_super->s_flags |= SB_ACTIVE;
762 error = xlog_recover_finish(mp->m_log);
763 if (!error)
764 xfs_log_work_queue(mp);
765 mp->m_super->s_flags &= ~SB_ACTIVE;
766 evict_inodes(mp->m_super);
767
768 /*
769 * Drain the buffer LRU after log recovery. This is required for v4
770 * filesystems to avoid leaving around buffers with NULL verifier ops,
771 * but we do it unconditionally to make sure we're always in a clean
772 * cache state after mount.
773 *
774 * Don't push in the error case because the AIL may have pending intents
775 * that aren't removed until recovery is cancelled.
776 */
777 if (!error && recovered) {
778 xfs_log_force(mp, XFS_LOG_SYNC);
779 xfs_ail_push_all_sync(mp->m_ail);
780 }
781 xfs_buftarg_drain(mp->m_ddev_targp);
782
783 if (readonly)
784 mp->m_flags |= XFS_MOUNT_RDONLY;
785
786 /* Make sure the log is dead if we're returning failure. */
787 ASSERT(!error || (mp->m_log->l_flags & XLOG_IO_ERROR));
788
789 return error;
790}
791
792/*
793 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
794 * the log.
795 */
796void
797xfs_log_mount_cancel(
798 struct xfs_mount *mp)
799{
800 xlog_recover_cancel(mp->m_log);
801 xfs_log_unmount(mp);
802}
803
804/*
805 * Flush out the iclog to disk ensuring that device caches are flushed and
806 * the iclog hits stable storage before any completion waiters are woken.
807 */
808static inline int
809xlog_force_iclog(
810 struct xlog_in_core *iclog)
811{
812 atomic_inc(&iclog->ic_refcnt);
813 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
814 if (iclog->ic_state == XLOG_STATE_ACTIVE)
815 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
816 return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
817}
818
819/*
820 * Wait for the iclog and all prior iclogs to be written disk as required by the
821 * log force state machine. Waiting on ic_force_wait ensures iclog completions
822 * have been ordered and callbacks run before we are woken here, hence
823 * guaranteeing that all the iclogs up to this one are on stable storage.
824 */
825int
826xlog_wait_on_iclog(
827 struct xlog_in_core *iclog)
828 __releases(iclog->ic_log->l_icloglock)
829{
830 struct xlog *log = iclog->ic_log;
831
832 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
833 if (!XLOG_FORCED_SHUTDOWN(log) &&
834 iclog->ic_state != XLOG_STATE_ACTIVE &&
835 iclog->ic_state != XLOG_STATE_DIRTY) {
836 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
837 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
838 } else {
839 spin_unlock(&log->l_icloglock);
840 }
841
842 if (XLOG_FORCED_SHUTDOWN(log))
843 return -EIO;
844 return 0;
845}
846
847/*
848 * Write out an unmount record using the ticket provided. We have to account for
849 * the data space used in the unmount ticket as this write is not done from a
850 * transaction context that has already done the accounting for us.
851 */
852static int
853xlog_write_unmount_record(
854 struct xlog *log,
855 struct xlog_ticket *ticket)
856{
857 struct xfs_unmount_log_format ulf = {
858 .magic = XLOG_UNMOUNT_TYPE,
859 };
860 struct xfs_log_iovec reg = {
861 .i_addr = &ulf,
862 .i_len = sizeof(ulf),
863 .i_type = XLOG_REG_TYPE_UNMOUNT,
864 };
865 struct xfs_log_vec vec = {
866 .lv_niovecs = 1,
867 .lv_iovecp = ®,
868 };
869
870 /* account for space used by record data */
871 ticket->t_curr_res -= sizeof(ulf);
872
873 return xlog_write(log, &vec, ticket, NULL, NULL, XLOG_UNMOUNT_TRANS);
874}
875
876/*
877 * Mark the filesystem clean by writing an unmount record to the head of the
878 * log.
879 */
880static void
881xlog_unmount_write(
882 struct xlog *log)
883{
884 struct xfs_mount *mp = log->l_mp;
885 struct xlog_in_core *iclog;
886 struct xlog_ticket *tic = NULL;
887 int error;
888
889 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
890 if (error)
891 goto out_err;
892
893 error = xlog_write_unmount_record(log, tic);
894 /*
895 * At this point, we're umounting anyway, so there's no point in
896 * transitioning log state to IOERROR. Just continue...
897 */
898out_err:
899 if (error)
900 xfs_alert(mp, "%s: unmount record failed", __func__);
901
902 spin_lock(&log->l_icloglock);
903 iclog = log->l_iclog;
904 error = xlog_force_iclog(iclog);
905 xlog_wait_on_iclog(iclog);
906
907 if (tic) {
908 trace_xfs_log_umount_write(log, tic);
909 xfs_log_ticket_ungrant(log, tic);
910 }
911}
912
913static void
914xfs_log_unmount_verify_iclog(
915 struct xlog *log)
916{
917 struct xlog_in_core *iclog = log->l_iclog;
918
919 do {
920 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
921 ASSERT(iclog->ic_offset == 0);
922 } while ((iclog = iclog->ic_next) != log->l_iclog);
923}
924
925/*
926 * Unmount record used to have a string "Unmount filesystem--" in the
927 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
928 * We just write the magic number now since that particular field isn't
929 * currently architecture converted and "Unmount" is a bit foo.
930 * As far as I know, there weren't any dependencies on the old behaviour.
931 */
932static void
933xfs_log_unmount_write(
934 struct xfs_mount *mp)
935{
936 struct xlog *log = mp->m_log;
937
938 if (!xfs_log_writable(mp))
939 return;
940
941 xfs_log_force(mp, XFS_LOG_SYNC);
942
943 if (XLOG_FORCED_SHUTDOWN(log))
944 return;
945
946 /*
947 * If we think the summary counters are bad, avoid writing the unmount
948 * record to force log recovery at next mount, after which the summary
949 * counters will be recalculated. Refer to xlog_check_unmount_rec for
950 * more details.
951 */
952 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
953 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
954 xfs_alert(mp, "%s: will fix summary counters at next mount",
955 __func__);
956 return;
957 }
958
959 xfs_log_unmount_verify_iclog(log);
960 xlog_unmount_write(log);
961}
962
963/*
964 * Empty the log for unmount/freeze.
965 *
966 * To do this, we first need to shut down the background log work so it is not
967 * trying to cover the log as we clean up. We then need to unpin all objects in
968 * the log so we can then flush them out. Once they have completed their IO and
969 * run the callbacks removing themselves from the AIL, we can cover the log.
970 */
971int
972xfs_log_quiesce(
973 struct xfs_mount *mp)
974{
975 cancel_delayed_work_sync(&mp->m_log->l_work);
976 xfs_log_force(mp, XFS_LOG_SYNC);
977
978 /*
979 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
980 * will push it, xfs_buftarg_wait() will not wait for it. Further,
981 * xfs_buf_iowait() cannot be used because it was pushed with the
982 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
983 * the IO to complete.
984 */
985 xfs_ail_push_all_sync(mp->m_ail);
986 xfs_buftarg_wait(mp->m_ddev_targp);
987 xfs_buf_lock(mp->m_sb_bp);
988 xfs_buf_unlock(mp->m_sb_bp);
989
990 return xfs_log_cover(mp);
991}
992
993void
994xfs_log_clean(
995 struct xfs_mount *mp)
996{
997 xfs_log_quiesce(mp);
998 xfs_log_unmount_write(mp);
999}
1000
1001/*
1002 * Shut down and release the AIL and Log.
1003 *
1004 * During unmount, we need to ensure we flush all the dirty metadata objects
1005 * from the AIL so that the log is empty before we write the unmount record to
1006 * the log. Once this is done, we can tear down the AIL and the log.
1007 */
1008void
1009xfs_log_unmount(
1010 struct xfs_mount *mp)
1011{
1012 xfs_log_clean(mp);
1013
1014 xfs_buftarg_drain(mp->m_ddev_targp);
1015
1016 xfs_trans_ail_destroy(mp);
1017
1018 xfs_sysfs_del(&mp->m_log->l_kobj);
1019
1020 xlog_dealloc_log(mp->m_log);
1021}
1022
1023void
1024xfs_log_item_init(
1025 struct xfs_mount *mp,
1026 struct xfs_log_item *item,
1027 int type,
1028 const struct xfs_item_ops *ops)
1029{
1030 item->li_mountp = mp;
1031 item->li_ailp = mp->m_ail;
1032 item->li_type = type;
1033 item->li_ops = ops;
1034 item->li_lv = NULL;
1035
1036 INIT_LIST_HEAD(&item->li_ail);
1037 INIT_LIST_HEAD(&item->li_cil);
1038 INIT_LIST_HEAD(&item->li_bio_list);
1039 INIT_LIST_HEAD(&item->li_trans);
1040}
1041
1042/*
1043 * Wake up processes waiting for log space after we have moved the log tail.
1044 */
1045void
1046xfs_log_space_wake(
1047 struct xfs_mount *mp)
1048{
1049 struct xlog *log = mp->m_log;
1050 int free_bytes;
1051
1052 if (XLOG_FORCED_SHUTDOWN(log))
1053 return;
1054
1055 if (!list_empty_careful(&log->l_write_head.waiters)) {
1056 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1057
1058 spin_lock(&log->l_write_head.lock);
1059 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1060 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1061 spin_unlock(&log->l_write_head.lock);
1062 }
1063
1064 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1065 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1066
1067 spin_lock(&log->l_reserve_head.lock);
1068 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1069 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1070 spin_unlock(&log->l_reserve_head.lock);
1071 }
1072}
1073
1074/*
1075 * Determine if we have a transaction that has gone to disk that needs to be
1076 * covered. To begin the transition to the idle state firstly the log needs to
1077 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1078 * we start attempting to cover the log.
1079 *
1080 * Only if we are then in a state where covering is needed, the caller is
1081 * informed that dummy transactions are required to move the log into the idle
1082 * state.
1083 *
1084 * If there are any items in the AIl or CIL, then we do not want to attempt to
1085 * cover the log as we may be in a situation where there isn't log space
1086 * available to run a dummy transaction and this can lead to deadlocks when the
1087 * tail of the log is pinned by an item that is modified in the CIL. Hence
1088 * there's no point in running a dummy transaction at this point because we
1089 * can't start trying to idle the log until both the CIL and AIL are empty.
1090 */
1091static bool
1092xfs_log_need_covered(
1093 struct xfs_mount *mp)
1094{
1095 struct xlog *log = mp->m_log;
1096 bool needed = false;
1097
1098 if (!xlog_cil_empty(log))
1099 return false;
1100
1101 spin_lock(&log->l_icloglock);
1102 switch (log->l_covered_state) {
1103 case XLOG_STATE_COVER_DONE:
1104 case XLOG_STATE_COVER_DONE2:
1105 case XLOG_STATE_COVER_IDLE:
1106 break;
1107 case XLOG_STATE_COVER_NEED:
1108 case XLOG_STATE_COVER_NEED2:
1109 if (xfs_ail_min_lsn(log->l_ailp))
1110 break;
1111 if (!xlog_iclogs_empty(log))
1112 break;
1113
1114 needed = true;
1115 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1116 log->l_covered_state = XLOG_STATE_COVER_DONE;
1117 else
1118 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1119 break;
1120 default:
1121 needed = true;
1122 break;
1123 }
1124 spin_unlock(&log->l_icloglock);
1125 return needed;
1126}
1127
1128/*
1129 * Explicitly cover the log. This is similar to background log covering but
1130 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1131 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1132 * must all be empty.
1133 */
1134static int
1135xfs_log_cover(
1136 struct xfs_mount *mp)
1137{
1138 int error = 0;
1139 bool need_covered;
1140
1141 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1142 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1143 XFS_FORCED_SHUTDOWN(mp));
1144
1145 if (!xfs_log_writable(mp))
1146 return 0;
1147
1148 /*
1149 * xfs_log_need_covered() is not idempotent because it progresses the
1150 * state machine if the log requires covering. Therefore, we must call
1151 * this function once and use the result until we've issued an sb sync.
1152 * Do so first to make that abundantly clear.
1153 *
1154 * Fall into the covering sequence if the log needs covering or the
1155 * mount has lazy superblock accounting to sync to disk. The sb sync
1156 * used for covering accumulates the in-core counters, so covering
1157 * handles this for us.
1158 */
1159 need_covered = xfs_log_need_covered(mp);
1160 if (!need_covered && !xfs_sb_version_haslazysbcount(&mp->m_sb))
1161 return 0;
1162
1163 /*
1164 * To cover the log, commit the superblock twice (at most) in
1165 * independent checkpoints. The first serves as a reference for the
1166 * tail pointer. The sync transaction and AIL push empties the AIL and
1167 * updates the in-core tail to the LSN of the first checkpoint. The
1168 * second commit updates the on-disk tail with the in-core LSN,
1169 * covering the log. Push the AIL one more time to leave it empty, as
1170 * we found it.
1171 */
1172 do {
1173 error = xfs_sync_sb(mp, true);
1174 if (error)
1175 break;
1176 xfs_ail_push_all_sync(mp->m_ail);
1177 } while (xfs_log_need_covered(mp));
1178
1179 return error;
1180}
1181
1182/*
1183 * We may be holding the log iclog lock upon entering this routine.
1184 */
1185xfs_lsn_t
1186xlog_assign_tail_lsn_locked(
1187 struct xfs_mount *mp)
1188{
1189 struct xlog *log = mp->m_log;
1190 struct xfs_log_item *lip;
1191 xfs_lsn_t tail_lsn;
1192
1193 assert_spin_locked(&mp->m_ail->ail_lock);
1194
1195 /*
1196 * To make sure we always have a valid LSN for the log tail we keep
1197 * track of the last LSN which was committed in log->l_last_sync_lsn,
1198 * and use that when the AIL was empty.
1199 */
1200 lip = xfs_ail_min(mp->m_ail);
1201 if (lip)
1202 tail_lsn = lip->li_lsn;
1203 else
1204 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1205 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1206 atomic64_set(&log->l_tail_lsn, tail_lsn);
1207 return tail_lsn;
1208}
1209
1210xfs_lsn_t
1211xlog_assign_tail_lsn(
1212 struct xfs_mount *mp)
1213{
1214 xfs_lsn_t tail_lsn;
1215
1216 spin_lock(&mp->m_ail->ail_lock);
1217 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1218 spin_unlock(&mp->m_ail->ail_lock);
1219
1220 return tail_lsn;
1221}
1222
1223/*
1224 * Return the space in the log between the tail and the head. The head
1225 * is passed in the cycle/bytes formal parms. In the special case where
1226 * the reserve head has wrapped passed the tail, this calculation is no
1227 * longer valid. In this case, just return 0 which means there is no space
1228 * in the log. This works for all places where this function is called
1229 * with the reserve head. Of course, if the write head were to ever
1230 * wrap the tail, we should blow up. Rather than catch this case here,
1231 * we depend on other ASSERTions in other parts of the code. XXXmiken
1232 *
1233 * This code also handles the case where the reservation head is behind
1234 * the tail. The details of this case are described below, but the end
1235 * result is that we return the size of the log as the amount of space left.
1236 */
1237STATIC int
1238xlog_space_left(
1239 struct xlog *log,
1240 atomic64_t *head)
1241{
1242 int free_bytes;
1243 int tail_bytes;
1244 int tail_cycle;
1245 int head_cycle;
1246 int head_bytes;
1247
1248 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1249 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1250 tail_bytes = BBTOB(tail_bytes);
1251 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1252 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1253 else if (tail_cycle + 1 < head_cycle)
1254 return 0;
1255 else if (tail_cycle < head_cycle) {
1256 ASSERT(tail_cycle == (head_cycle - 1));
1257 free_bytes = tail_bytes - head_bytes;
1258 } else {
1259 /*
1260 * The reservation head is behind the tail.
1261 * In this case we just want to return the size of the
1262 * log as the amount of space left.
1263 */
1264 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1265 xfs_alert(log->l_mp,
1266 " tail_cycle = %d, tail_bytes = %d",
1267 tail_cycle, tail_bytes);
1268 xfs_alert(log->l_mp,
1269 " GH cycle = %d, GH bytes = %d",
1270 head_cycle, head_bytes);
1271 ASSERT(0);
1272 free_bytes = log->l_logsize;
1273 }
1274 return free_bytes;
1275}
1276
1277
1278static void
1279xlog_ioend_work(
1280 struct work_struct *work)
1281{
1282 struct xlog_in_core *iclog =
1283 container_of(work, struct xlog_in_core, ic_end_io_work);
1284 struct xlog *log = iclog->ic_log;
1285 int error;
1286
1287 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1288#ifdef DEBUG
1289 /* treat writes with injected CRC errors as failed */
1290 if (iclog->ic_fail_crc)
1291 error = -EIO;
1292#endif
1293
1294 /*
1295 * Race to shutdown the filesystem if we see an error.
1296 */
1297 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1298 xfs_alert(log->l_mp, "log I/O error %d", error);
1299 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1300 }
1301
1302 xlog_state_done_syncing(iclog);
1303 bio_uninit(&iclog->ic_bio);
1304
1305 /*
1306 * Drop the lock to signal that we are done. Nothing references the
1307 * iclog after this, so an unmount waiting on this lock can now tear it
1308 * down safely. As such, it is unsafe to reference the iclog after the
1309 * unlock as we could race with it being freed.
1310 */
1311 up(&iclog->ic_sema);
1312}
1313
1314/*
1315 * Return size of each in-core log record buffer.
1316 *
1317 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1318 *
1319 * If the filesystem blocksize is too large, we may need to choose a
1320 * larger size since the directory code currently logs entire blocks.
1321 */
1322STATIC void
1323xlog_get_iclog_buffer_size(
1324 struct xfs_mount *mp,
1325 struct xlog *log)
1326{
1327 if (mp->m_logbufs <= 0)
1328 mp->m_logbufs = XLOG_MAX_ICLOGS;
1329 if (mp->m_logbsize <= 0)
1330 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1331
1332 log->l_iclog_bufs = mp->m_logbufs;
1333 log->l_iclog_size = mp->m_logbsize;
1334
1335 /*
1336 * # headers = size / 32k - one header holds cycles from 32k of data.
1337 */
1338 log->l_iclog_heads =
1339 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1340 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1341}
1342
1343void
1344xfs_log_work_queue(
1345 struct xfs_mount *mp)
1346{
1347 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1348 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1349}
1350
1351/*
1352 * Every sync period we need to unpin all items in the AIL and push them to
1353 * disk. If there is nothing dirty, then we might need to cover the log to
1354 * indicate that the filesystem is idle.
1355 */
1356static void
1357xfs_log_worker(
1358 struct work_struct *work)
1359{
1360 struct xlog *log = container_of(to_delayed_work(work),
1361 struct xlog, l_work);
1362 struct xfs_mount *mp = log->l_mp;
1363
1364 /* dgc: errors ignored - not fatal and nowhere to report them */
1365 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1366 /*
1367 * Dump a transaction into the log that contains no real change.
1368 * This is needed to stamp the current tail LSN into the log
1369 * during the covering operation.
1370 *
1371 * We cannot use an inode here for this - that will push dirty
1372 * state back up into the VFS and then periodic inode flushing
1373 * will prevent log covering from making progress. Hence we
1374 * synchronously log the superblock instead to ensure the
1375 * superblock is immediately unpinned and can be written back.
1376 */
1377 xfs_sync_sb(mp, true);
1378 } else
1379 xfs_log_force(mp, 0);
1380
1381 /* start pushing all the metadata that is currently dirty */
1382 xfs_ail_push_all(mp->m_ail);
1383
1384 /* queue us up again */
1385 xfs_log_work_queue(mp);
1386}
1387
1388/*
1389 * This routine initializes some of the log structure for a given mount point.
1390 * Its primary purpose is to fill in enough, so recovery can occur. However,
1391 * some other stuff may be filled in too.
1392 */
1393STATIC struct xlog *
1394xlog_alloc_log(
1395 struct xfs_mount *mp,
1396 struct xfs_buftarg *log_target,
1397 xfs_daddr_t blk_offset,
1398 int num_bblks)
1399{
1400 struct xlog *log;
1401 xlog_rec_header_t *head;
1402 xlog_in_core_t **iclogp;
1403 xlog_in_core_t *iclog, *prev_iclog=NULL;
1404 int i;
1405 int error = -ENOMEM;
1406 uint log2_size = 0;
1407
1408 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1409 if (!log) {
1410 xfs_warn(mp, "Log allocation failed: No memory!");
1411 goto out;
1412 }
1413
1414 log->l_mp = mp;
1415 log->l_targ = log_target;
1416 log->l_logsize = BBTOB(num_bblks);
1417 log->l_logBBstart = blk_offset;
1418 log->l_logBBsize = num_bblks;
1419 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1420 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1421 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1422
1423 log->l_prev_block = -1;
1424 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1425 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1426 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1427 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1428
1429 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1)
1430 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1431 else
1432 log->l_iclog_roundoff = BBSIZE;
1433
1434 xlog_grant_head_init(&log->l_reserve_head);
1435 xlog_grant_head_init(&log->l_write_head);
1436
1437 error = -EFSCORRUPTED;
1438 if (xfs_sb_version_hassector(&mp->m_sb)) {
1439 log2_size = mp->m_sb.sb_logsectlog;
1440 if (log2_size < BBSHIFT) {
1441 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1442 log2_size, BBSHIFT);
1443 goto out_free_log;
1444 }
1445
1446 log2_size -= BBSHIFT;
1447 if (log2_size > mp->m_sectbb_log) {
1448 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1449 log2_size, mp->m_sectbb_log);
1450 goto out_free_log;
1451 }
1452
1453 /* for larger sector sizes, must have v2 or external log */
1454 if (log2_size && log->l_logBBstart > 0 &&
1455 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1456 xfs_warn(mp,
1457 "log sector size (0x%x) invalid for configuration.",
1458 log2_size);
1459 goto out_free_log;
1460 }
1461 }
1462 log->l_sectBBsize = 1 << log2_size;
1463
1464 xlog_get_iclog_buffer_size(mp, log);
1465
1466 spin_lock_init(&log->l_icloglock);
1467 init_waitqueue_head(&log->l_flush_wait);
1468
1469 iclogp = &log->l_iclog;
1470 /*
1471 * The amount of memory to allocate for the iclog structure is
1472 * rather funky due to the way the structure is defined. It is
1473 * done this way so that we can use different sizes for machines
1474 * with different amounts of memory. See the definition of
1475 * xlog_in_core_t in xfs_log_priv.h for details.
1476 */
1477 ASSERT(log->l_iclog_size >= 4096);
1478 for (i = 0; i < log->l_iclog_bufs; i++) {
1479 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1480 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1481 sizeof(struct bio_vec);
1482
1483 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1484 if (!iclog)
1485 goto out_free_iclog;
1486
1487 *iclogp = iclog;
1488 iclog->ic_prev = prev_iclog;
1489 prev_iclog = iclog;
1490
1491 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1492 KM_MAYFAIL | KM_ZERO);
1493 if (!iclog->ic_data)
1494 goto out_free_iclog;
1495#ifdef DEBUG
1496 log->l_iclog_bak[i] = &iclog->ic_header;
1497#endif
1498 head = &iclog->ic_header;
1499 memset(head, 0, sizeof(xlog_rec_header_t));
1500 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1501 head->h_version = cpu_to_be32(
1502 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1503 head->h_size = cpu_to_be32(log->l_iclog_size);
1504 /* new fields */
1505 head->h_fmt = cpu_to_be32(XLOG_FMT);
1506 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1507
1508 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1509 iclog->ic_state = XLOG_STATE_ACTIVE;
1510 iclog->ic_log = log;
1511 atomic_set(&iclog->ic_refcnt, 0);
1512 INIT_LIST_HEAD(&iclog->ic_callbacks);
1513 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1514
1515 init_waitqueue_head(&iclog->ic_force_wait);
1516 init_waitqueue_head(&iclog->ic_write_wait);
1517 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1518 sema_init(&iclog->ic_sema, 1);
1519
1520 iclogp = &iclog->ic_next;
1521 }
1522 *iclogp = log->l_iclog; /* complete ring */
1523 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1524
1525 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1526 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1527 WQ_HIGHPRI),
1528 0, mp->m_super->s_id);
1529 if (!log->l_ioend_workqueue)
1530 goto out_free_iclog;
1531
1532 error = xlog_cil_init(log);
1533 if (error)
1534 goto out_destroy_workqueue;
1535 return log;
1536
1537out_destroy_workqueue:
1538 destroy_workqueue(log->l_ioend_workqueue);
1539out_free_iclog:
1540 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1541 prev_iclog = iclog->ic_next;
1542 kmem_free(iclog->ic_data);
1543 kmem_free(iclog);
1544 if (prev_iclog == log->l_iclog)
1545 break;
1546 }
1547out_free_log:
1548 kmem_free(log);
1549out:
1550 return ERR_PTR(error);
1551} /* xlog_alloc_log */
1552
1553/*
1554 * Write out the commit record of a transaction associated with the given
1555 * ticket to close off a running log write. Return the lsn of the commit record.
1556 */
1557int
1558xlog_commit_record(
1559 struct xlog *log,
1560 struct xlog_ticket *ticket,
1561 struct xlog_in_core **iclog,
1562 xfs_lsn_t *lsn)
1563{
1564 struct xfs_log_iovec reg = {
1565 .i_addr = NULL,
1566 .i_len = 0,
1567 .i_type = XLOG_REG_TYPE_COMMIT,
1568 };
1569 struct xfs_log_vec vec = {
1570 .lv_niovecs = 1,
1571 .lv_iovecp = ®,
1572 };
1573 int error;
1574
1575 if (XLOG_FORCED_SHUTDOWN(log))
1576 return -EIO;
1577
1578 error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS);
1579 if (error)
1580 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1581 return error;
1582}
1583
1584/*
1585 * Compute the LSN that we'd need to push the log tail towards in order to have
1586 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1587 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1588 * log free space already meets all three thresholds, this function returns
1589 * NULLCOMMITLSN.
1590 */
1591xfs_lsn_t
1592xlog_grant_push_threshold(
1593 struct xlog *log,
1594 int need_bytes)
1595{
1596 xfs_lsn_t threshold_lsn = 0;
1597 xfs_lsn_t last_sync_lsn;
1598 int free_blocks;
1599 int free_bytes;
1600 int threshold_block;
1601 int threshold_cycle;
1602 int free_threshold;
1603
1604 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1605
1606 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1607 free_blocks = BTOBBT(free_bytes);
1608
1609 /*
1610 * Set the threshold for the minimum number of free blocks in the
1611 * log to the maximum of what the caller needs, one quarter of the
1612 * log, and 256 blocks.
1613 */
1614 free_threshold = BTOBB(need_bytes);
1615 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1616 free_threshold = max(free_threshold, 256);
1617 if (free_blocks >= free_threshold)
1618 return NULLCOMMITLSN;
1619
1620 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1621 &threshold_block);
1622 threshold_block += free_threshold;
1623 if (threshold_block >= log->l_logBBsize) {
1624 threshold_block -= log->l_logBBsize;
1625 threshold_cycle += 1;
1626 }
1627 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1628 threshold_block);
1629 /*
1630 * Don't pass in an lsn greater than the lsn of the last
1631 * log record known to be on disk. Use a snapshot of the last sync lsn
1632 * so that it doesn't change between the compare and the set.
1633 */
1634 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1635 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1636 threshold_lsn = last_sync_lsn;
1637
1638 return threshold_lsn;
1639}
1640
1641/*
1642 * Push the tail of the log if we need to do so to maintain the free log space
1643 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1644 * policy which pushes on an lsn which is further along in the log once we
1645 * reach the high water mark. In this manner, we would be creating a low water
1646 * mark.
1647 */
1648STATIC void
1649xlog_grant_push_ail(
1650 struct xlog *log,
1651 int need_bytes)
1652{
1653 xfs_lsn_t threshold_lsn;
1654
1655 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1656 if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1657 return;
1658
1659 /*
1660 * Get the transaction layer to kick the dirty buffers out to
1661 * disk asynchronously. No point in trying to do this if
1662 * the filesystem is shutting down.
1663 */
1664 xfs_ail_push(log->l_ailp, threshold_lsn);
1665}
1666
1667/*
1668 * Stamp cycle number in every block
1669 */
1670STATIC void
1671xlog_pack_data(
1672 struct xlog *log,
1673 struct xlog_in_core *iclog,
1674 int roundoff)
1675{
1676 int i, j, k;
1677 int size = iclog->ic_offset + roundoff;
1678 __be32 cycle_lsn;
1679 char *dp;
1680
1681 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1682
1683 dp = iclog->ic_datap;
1684 for (i = 0; i < BTOBB(size); i++) {
1685 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1686 break;
1687 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1688 *(__be32 *)dp = cycle_lsn;
1689 dp += BBSIZE;
1690 }
1691
1692 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1693 xlog_in_core_2_t *xhdr = iclog->ic_data;
1694
1695 for ( ; i < BTOBB(size); i++) {
1696 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1697 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1698 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1699 *(__be32 *)dp = cycle_lsn;
1700 dp += BBSIZE;
1701 }
1702
1703 for (i = 1; i < log->l_iclog_heads; i++)
1704 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1705 }
1706}
1707
1708/*
1709 * Calculate the checksum for a log buffer.
1710 *
1711 * This is a little more complicated than it should be because the various
1712 * headers and the actual data are non-contiguous.
1713 */
1714__le32
1715xlog_cksum(
1716 struct xlog *log,
1717 struct xlog_rec_header *rhead,
1718 char *dp,
1719 int size)
1720{
1721 uint32_t crc;
1722
1723 /* first generate the crc for the record header ... */
1724 crc = xfs_start_cksum_update((char *)rhead,
1725 sizeof(struct xlog_rec_header),
1726 offsetof(struct xlog_rec_header, h_crc));
1727
1728 /* ... then for additional cycle data for v2 logs ... */
1729 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1730 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1731 int i;
1732 int xheads;
1733
1734 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1735
1736 for (i = 1; i < xheads; i++) {
1737 crc = crc32c(crc, &xhdr[i].hic_xheader,
1738 sizeof(struct xlog_rec_ext_header));
1739 }
1740 }
1741
1742 /* ... and finally for the payload */
1743 crc = crc32c(crc, dp, size);
1744
1745 return xfs_end_cksum(crc);
1746}
1747
1748static void
1749xlog_bio_end_io(
1750 struct bio *bio)
1751{
1752 struct xlog_in_core *iclog = bio->bi_private;
1753
1754 queue_work(iclog->ic_log->l_ioend_workqueue,
1755 &iclog->ic_end_io_work);
1756}
1757
1758static int
1759xlog_map_iclog_data(
1760 struct bio *bio,
1761 void *data,
1762 size_t count)
1763{
1764 do {
1765 struct page *page = kmem_to_page(data);
1766 unsigned int off = offset_in_page(data);
1767 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1768
1769 if (bio_add_page(bio, page, len, off) != len)
1770 return -EIO;
1771
1772 data += len;
1773 count -= len;
1774 } while (count);
1775
1776 return 0;
1777}
1778
1779STATIC void
1780xlog_write_iclog(
1781 struct xlog *log,
1782 struct xlog_in_core *iclog,
1783 uint64_t bno,
1784 unsigned int count)
1785{
1786 ASSERT(bno < log->l_logBBsize);
1787 trace_xlog_iclog_write(iclog, _RET_IP_);
1788
1789 /*
1790 * We lock the iclogbufs here so that we can serialise against I/O
1791 * completion during unmount. We might be processing a shutdown
1792 * triggered during unmount, and that can occur asynchronously to the
1793 * unmount thread, and hence we need to ensure that completes before
1794 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1795 * across the log IO to archieve that.
1796 */
1797 down(&iclog->ic_sema);
1798 if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1799 /*
1800 * It would seem logical to return EIO here, but we rely on
1801 * the log state machine to propagate I/O errors instead of
1802 * doing it here. We kick of the state machine and unlock
1803 * the buffer manually, the code needs to be kept in sync
1804 * with the I/O completion path.
1805 */
1806 xlog_state_done_syncing(iclog);
1807 up(&iclog->ic_sema);
1808 return;
1809 }
1810
1811 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1812 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1813 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1814 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1815 iclog->ic_bio.bi_private = iclog;
1816
1817 /*
1818 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1819 * IOs coming immediately after this one. This prevents the block layer
1820 * writeback throttle from throttling log writes behind background
1821 * metadata writeback and causing priority inversions.
1822 */
1823 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
1824 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1825 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1826 /*
1827 * For external log devices, we also need to flush the data
1828 * device cache first to ensure all metadata writeback covered
1829 * by the LSN in this iclog is on stable storage. This is slow,
1830 * but it *must* complete before we issue the external log IO.
1831 */
1832 if (log->l_targ != log->l_mp->m_ddev_targp)
1833 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1834 }
1835 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1836 iclog->ic_bio.bi_opf |= REQ_FUA;
1837
1838 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1839
1840 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1841 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1842 return;
1843 }
1844 if (is_vmalloc_addr(iclog->ic_data))
1845 flush_kernel_vmap_range(iclog->ic_data, count);
1846
1847 /*
1848 * If this log buffer would straddle the end of the log we will have
1849 * to split it up into two bios, so that we can continue at the start.
1850 */
1851 if (bno + BTOBB(count) > log->l_logBBsize) {
1852 struct bio *split;
1853
1854 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1855 GFP_NOIO, &fs_bio_set);
1856 bio_chain(split, &iclog->ic_bio);
1857 submit_bio(split);
1858
1859 /* restart at logical offset zero for the remainder */
1860 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1861 }
1862
1863 submit_bio(&iclog->ic_bio);
1864}
1865
1866/*
1867 * We need to bump cycle number for the part of the iclog that is
1868 * written to the start of the log. Watch out for the header magic
1869 * number case, though.
1870 */
1871static void
1872xlog_split_iclog(
1873 struct xlog *log,
1874 void *data,
1875 uint64_t bno,
1876 unsigned int count)
1877{
1878 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1879 unsigned int i;
1880
1881 for (i = split_offset; i < count; i += BBSIZE) {
1882 uint32_t cycle = get_unaligned_be32(data + i);
1883
1884 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1885 cycle++;
1886 put_unaligned_be32(cycle, data + i);
1887 }
1888}
1889
1890static int
1891xlog_calc_iclog_size(
1892 struct xlog *log,
1893 struct xlog_in_core *iclog,
1894 uint32_t *roundoff)
1895{
1896 uint32_t count_init, count;
1897
1898 /* Add for LR header */
1899 count_init = log->l_iclog_hsize + iclog->ic_offset;
1900 count = roundup(count_init, log->l_iclog_roundoff);
1901
1902 *roundoff = count - count_init;
1903
1904 ASSERT(count >= count_init);
1905 ASSERT(*roundoff < log->l_iclog_roundoff);
1906 return count;
1907}
1908
1909/*
1910 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1911 * fashion. Previously, we should have moved the current iclog
1912 * ptr in the log to point to the next available iclog. This allows further
1913 * write to continue while this code syncs out an iclog ready to go.
1914 * Before an in-core log can be written out, the data section must be scanned
1915 * to save away the 1st word of each BBSIZE block into the header. We replace
1916 * it with the current cycle count. Each BBSIZE block is tagged with the
1917 * cycle count because there in an implicit assumption that drives will
1918 * guarantee that entire 512 byte blocks get written at once. In other words,
1919 * we can't have part of a 512 byte block written and part not written. By
1920 * tagging each block, we will know which blocks are valid when recovering
1921 * after an unclean shutdown.
1922 *
1923 * This routine is single threaded on the iclog. No other thread can be in
1924 * this routine with the same iclog. Changing contents of iclog can there-
1925 * fore be done without grabbing the state machine lock. Updating the global
1926 * log will require grabbing the lock though.
1927 *
1928 * The entire log manager uses a logical block numbering scheme. Only
1929 * xlog_write_iclog knows about the fact that the log may not start with
1930 * block zero on a given device.
1931 */
1932STATIC void
1933xlog_sync(
1934 struct xlog *log,
1935 struct xlog_in_core *iclog)
1936{
1937 unsigned int count; /* byte count of bwrite */
1938 unsigned int roundoff; /* roundoff to BB or stripe */
1939 uint64_t bno;
1940 unsigned int size;
1941
1942 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1943 trace_xlog_iclog_sync(iclog, _RET_IP_);
1944
1945 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1946
1947 /* move grant heads by roundoff in sync */
1948 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1949 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1950
1951 /* put cycle number in every block */
1952 xlog_pack_data(log, iclog, roundoff);
1953
1954 /* real byte length */
1955 size = iclog->ic_offset;
1956 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1957 size += roundoff;
1958 iclog->ic_header.h_len = cpu_to_be32(size);
1959
1960 XFS_STATS_INC(log->l_mp, xs_log_writes);
1961 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1962
1963 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1964
1965 /* Do we need to split this write into 2 parts? */
1966 if (bno + BTOBB(count) > log->l_logBBsize)
1967 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1968
1969 /* calculcate the checksum */
1970 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1971 iclog->ic_datap, size);
1972 /*
1973 * Intentionally corrupt the log record CRC based on the error injection
1974 * frequency, if defined. This facilitates testing log recovery in the
1975 * event of torn writes. Hence, set the IOABORT state to abort the log
1976 * write on I/O completion and shutdown the fs. The subsequent mount
1977 * detects the bad CRC and attempts to recover.
1978 */
1979#ifdef DEBUG
1980 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1981 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1982 iclog->ic_fail_crc = true;
1983 xfs_warn(log->l_mp,
1984 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1985 be64_to_cpu(iclog->ic_header.h_lsn));
1986 }
1987#endif
1988 xlog_verify_iclog(log, iclog, count);
1989 xlog_write_iclog(log, iclog, bno, count);
1990}
1991
1992/*
1993 * Deallocate a log structure
1994 */
1995STATIC void
1996xlog_dealloc_log(
1997 struct xlog *log)
1998{
1999 xlog_in_core_t *iclog, *next_iclog;
2000 int i;
2001
2002 xlog_cil_destroy(log);
2003
2004 /*
2005 * Cycle all the iclogbuf locks to make sure all log IO completion
2006 * is done before we tear down these buffers.
2007 */
2008 iclog = log->l_iclog;
2009 for (i = 0; i < log->l_iclog_bufs; i++) {
2010 down(&iclog->ic_sema);
2011 up(&iclog->ic_sema);
2012 iclog = iclog->ic_next;
2013 }
2014
2015 iclog = log->l_iclog;
2016 for (i = 0; i < log->l_iclog_bufs; i++) {
2017 next_iclog = iclog->ic_next;
2018 kmem_free(iclog->ic_data);
2019 kmem_free(iclog);
2020 iclog = next_iclog;
2021 }
2022
2023 log->l_mp->m_log = NULL;
2024 destroy_workqueue(log->l_ioend_workqueue);
2025 kmem_free(log);
2026}
2027
2028/*
2029 * Update counters atomically now that memcpy is done.
2030 */
2031static inline void
2032xlog_state_finish_copy(
2033 struct xlog *log,
2034 struct xlog_in_core *iclog,
2035 int record_cnt,
2036 int copy_bytes)
2037{
2038 lockdep_assert_held(&log->l_icloglock);
2039
2040 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2041 iclog->ic_offset += copy_bytes;
2042}
2043
2044/*
2045 * print out info relating to regions written which consume
2046 * the reservation
2047 */
2048void
2049xlog_print_tic_res(
2050 struct xfs_mount *mp,
2051 struct xlog_ticket *ticket)
2052{
2053 uint i;
2054 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2055
2056 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2057#define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2058 static char *res_type_str[] = {
2059 REG_TYPE_STR(BFORMAT, "bformat"),
2060 REG_TYPE_STR(BCHUNK, "bchunk"),
2061 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2062 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2063 REG_TYPE_STR(IFORMAT, "iformat"),
2064 REG_TYPE_STR(ICORE, "icore"),
2065 REG_TYPE_STR(IEXT, "iext"),
2066 REG_TYPE_STR(IBROOT, "ibroot"),
2067 REG_TYPE_STR(ILOCAL, "ilocal"),
2068 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2069 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2070 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2071 REG_TYPE_STR(QFORMAT, "qformat"),
2072 REG_TYPE_STR(DQUOT, "dquot"),
2073 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2074 REG_TYPE_STR(LRHEADER, "LR header"),
2075 REG_TYPE_STR(UNMOUNT, "unmount"),
2076 REG_TYPE_STR(COMMIT, "commit"),
2077 REG_TYPE_STR(TRANSHDR, "trans header"),
2078 REG_TYPE_STR(ICREATE, "inode create"),
2079 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2080 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2081 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2082 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2083 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2084 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2085 };
2086 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2087#undef REG_TYPE_STR
2088
2089 xfs_warn(mp, "ticket reservation summary:");
2090 xfs_warn(mp, " unit res = %d bytes",
2091 ticket->t_unit_res);
2092 xfs_warn(mp, " current res = %d bytes",
2093 ticket->t_curr_res);
2094 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2095 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2096 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2097 ticket->t_res_num_ophdrs, ophdr_spc);
2098 xfs_warn(mp, " ophdr + reg = %u bytes",
2099 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2100 xfs_warn(mp, " num regions = %u",
2101 ticket->t_res_num);
2102
2103 for (i = 0; i < ticket->t_res_num; i++) {
2104 uint r_type = ticket->t_res_arr[i].r_type;
2105 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2106 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2107 "bad-rtype" : res_type_str[r_type]),
2108 ticket->t_res_arr[i].r_len);
2109 }
2110}
2111
2112/*
2113 * Print a summary of the transaction.
2114 */
2115void
2116xlog_print_trans(
2117 struct xfs_trans *tp)
2118{
2119 struct xfs_mount *mp = tp->t_mountp;
2120 struct xfs_log_item *lip;
2121
2122 /* dump core transaction and ticket info */
2123 xfs_warn(mp, "transaction summary:");
2124 xfs_warn(mp, " log res = %d", tp->t_log_res);
2125 xfs_warn(mp, " log count = %d", tp->t_log_count);
2126 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2127
2128 xlog_print_tic_res(mp, tp->t_ticket);
2129
2130 /* dump each log item */
2131 list_for_each_entry(lip, &tp->t_items, li_trans) {
2132 struct xfs_log_vec *lv = lip->li_lv;
2133 struct xfs_log_iovec *vec;
2134 int i;
2135
2136 xfs_warn(mp, "log item: ");
2137 xfs_warn(mp, " type = 0x%x", lip->li_type);
2138 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2139 if (!lv)
2140 continue;
2141 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2142 xfs_warn(mp, " size = %d", lv->lv_size);
2143 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2144 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2145
2146 /* dump each iovec for the log item */
2147 vec = lv->lv_iovecp;
2148 for (i = 0; i < lv->lv_niovecs; i++) {
2149 int dumplen = min(vec->i_len, 32);
2150
2151 xfs_warn(mp, " iovec[%d]", i);
2152 xfs_warn(mp, " type = 0x%x", vec->i_type);
2153 xfs_warn(mp, " len = %d", vec->i_len);
2154 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2155 xfs_hex_dump(vec->i_addr, dumplen);
2156
2157 vec++;
2158 }
2159 }
2160}
2161
2162/*
2163 * Calculate the potential space needed by the log vector. We may need a start
2164 * record, and each region gets its own struct xlog_op_header and may need to be
2165 * double word aligned.
2166 */
2167static int
2168xlog_write_calc_vec_length(
2169 struct xlog_ticket *ticket,
2170 struct xfs_log_vec *log_vector,
2171 uint optype)
2172{
2173 struct xfs_log_vec *lv;
2174 int headers = 0;
2175 int len = 0;
2176 int i;
2177
2178 if (optype & XLOG_START_TRANS)
2179 headers++;
2180
2181 for (lv = log_vector; lv; lv = lv->lv_next) {
2182 /* we don't write ordered log vectors */
2183 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2184 continue;
2185
2186 headers += lv->lv_niovecs;
2187
2188 for (i = 0; i < lv->lv_niovecs; i++) {
2189 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2190
2191 len += vecp->i_len;
2192 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2193 }
2194 }
2195
2196 ticket->t_res_num_ophdrs += headers;
2197 len += headers * sizeof(struct xlog_op_header);
2198
2199 return len;
2200}
2201
2202static void
2203xlog_write_start_rec(
2204 struct xlog_op_header *ophdr,
2205 struct xlog_ticket *ticket)
2206{
2207 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2208 ophdr->oh_clientid = ticket->t_clientid;
2209 ophdr->oh_len = 0;
2210 ophdr->oh_flags = XLOG_START_TRANS;
2211 ophdr->oh_res2 = 0;
2212}
2213
2214static xlog_op_header_t *
2215xlog_write_setup_ophdr(
2216 struct xlog *log,
2217 struct xlog_op_header *ophdr,
2218 struct xlog_ticket *ticket,
2219 uint flags)
2220{
2221 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2222 ophdr->oh_clientid = ticket->t_clientid;
2223 ophdr->oh_res2 = 0;
2224
2225 /* are we copying a commit or unmount record? */
2226 ophdr->oh_flags = flags;
2227
2228 /*
2229 * We've seen logs corrupted with bad transaction client ids. This
2230 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2231 * and shut down the filesystem.
2232 */
2233 switch (ophdr->oh_clientid) {
2234 case XFS_TRANSACTION:
2235 case XFS_VOLUME:
2236 case XFS_LOG:
2237 break;
2238 default:
2239 xfs_warn(log->l_mp,
2240 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2241 ophdr->oh_clientid, ticket);
2242 return NULL;
2243 }
2244
2245 return ophdr;
2246}
2247
2248/*
2249 * Set up the parameters of the region copy into the log. This has
2250 * to handle region write split across multiple log buffers - this
2251 * state is kept external to this function so that this code can
2252 * be written in an obvious, self documenting manner.
2253 */
2254static int
2255xlog_write_setup_copy(
2256 struct xlog_ticket *ticket,
2257 struct xlog_op_header *ophdr,
2258 int space_available,
2259 int space_required,
2260 int *copy_off,
2261 int *copy_len,
2262 int *last_was_partial_copy,
2263 int *bytes_consumed)
2264{
2265 int still_to_copy;
2266
2267 still_to_copy = space_required - *bytes_consumed;
2268 *copy_off = *bytes_consumed;
2269
2270 if (still_to_copy <= space_available) {
2271 /* write of region completes here */
2272 *copy_len = still_to_copy;
2273 ophdr->oh_len = cpu_to_be32(*copy_len);
2274 if (*last_was_partial_copy)
2275 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2276 *last_was_partial_copy = 0;
2277 *bytes_consumed = 0;
2278 return 0;
2279 }
2280
2281 /* partial write of region, needs extra log op header reservation */
2282 *copy_len = space_available;
2283 ophdr->oh_len = cpu_to_be32(*copy_len);
2284 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2285 if (*last_was_partial_copy)
2286 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2287 *bytes_consumed += *copy_len;
2288 (*last_was_partial_copy)++;
2289
2290 /* account for new log op header */
2291 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2292 ticket->t_res_num_ophdrs++;
2293
2294 return sizeof(struct xlog_op_header);
2295}
2296
2297static int
2298xlog_write_copy_finish(
2299 struct xlog *log,
2300 struct xlog_in_core *iclog,
2301 uint flags,
2302 int *record_cnt,
2303 int *data_cnt,
2304 int *partial_copy,
2305 int *partial_copy_len,
2306 int log_offset,
2307 struct xlog_in_core **commit_iclog)
2308{
2309 int error;
2310
2311 if (*partial_copy) {
2312 /*
2313 * This iclog has already been marked WANT_SYNC by
2314 * xlog_state_get_iclog_space.
2315 */
2316 spin_lock(&log->l_icloglock);
2317 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2318 *record_cnt = 0;
2319 *data_cnt = 0;
2320 goto release_iclog;
2321 }
2322
2323 *partial_copy = 0;
2324 *partial_copy_len = 0;
2325
2326 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2327 /* no more space in this iclog - push it. */
2328 spin_lock(&log->l_icloglock);
2329 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2330 *record_cnt = 0;
2331 *data_cnt = 0;
2332
2333 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2334 xlog_state_switch_iclogs(log, iclog, 0);
2335 else
2336 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2337 iclog->ic_state == XLOG_STATE_IOERROR);
2338 if (!commit_iclog)
2339 goto release_iclog;
2340 spin_unlock(&log->l_icloglock);
2341 ASSERT(flags & XLOG_COMMIT_TRANS);
2342 *commit_iclog = iclog;
2343 }
2344
2345 return 0;
2346
2347release_iclog:
2348 error = xlog_state_release_iclog(log, iclog, 0);
2349 spin_unlock(&log->l_icloglock);
2350 return error;
2351}
2352
2353/*
2354 * Write some region out to in-core log
2355 *
2356 * This will be called when writing externally provided regions or when
2357 * writing out a commit record for a given transaction.
2358 *
2359 * General algorithm:
2360 * 1. Find total length of this write. This may include adding to the
2361 * lengths passed in.
2362 * 2. Check whether we violate the tickets reservation.
2363 * 3. While writing to this iclog
2364 * A. Reserve as much space in this iclog as can get
2365 * B. If this is first write, save away start lsn
2366 * C. While writing this region:
2367 * 1. If first write of transaction, write start record
2368 * 2. Write log operation header (header per region)
2369 * 3. Find out if we can fit entire region into this iclog
2370 * 4. Potentially, verify destination memcpy ptr
2371 * 5. Memcpy (partial) region
2372 * 6. If partial copy, release iclog; otherwise, continue
2373 * copying more regions into current iclog
2374 * 4. Mark want sync bit (in simulation mode)
2375 * 5. Release iclog for potential flush to on-disk log.
2376 *
2377 * ERRORS:
2378 * 1. Panic if reservation is overrun. This should never happen since
2379 * reservation amounts are generated internal to the filesystem.
2380 * NOTES:
2381 * 1. Tickets are single threaded data structures.
2382 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2383 * syncing routine. When a single log_write region needs to span
2384 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2385 * on all log operation writes which don't contain the end of the
2386 * region. The XLOG_END_TRANS bit is used for the in-core log
2387 * operation which contains the end of the continued log_write region.
2388 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2389 * we don't really know exactly how much space will be used. As a result,
2390 * we don't update ic_offset until the end when we know exactly how many
2391 * bytes have been written out.
2392 */
2393int
2394xlog_write(
2395 struct xlog *log,
2396 struct xfs_log_vec *log_vector,
2397 struct xlog_ticket *ticket,
2398 xfs_lsn_t *start_lsn,
2399 struct xlog_in_core **commit_iclog,
2400 uint optype)
2401{
2402 struct xlog_in_core *iclog = NULL;
2403 struct xfs_log_vec *lv = log_vector;
2404 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2405 int index = 0;
2406 int len;
2407 int partial_copy = 0;
2408 int partial_copy_len = 0;
2409 int contwr = 0;
2410 int record_cnt = 0;
2411 int data_cnt = 0;
2412 int error = 0;
2413
2414 /*
2415 * If this is a commit or unmount transaction, we don't need a start
2416 * record to be written. We do, however, have to account for the
2417 * commit or unmount header that gets written. Hence we always have
2418 * to account for an extra xlog_op_header here.
2419 */
2420 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2421 if (ticket->t_curr_res < 0) {
2422 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2423 "ctx ticket reservation ran out. Need to up reservation");
2424 xlog_print_tic_res(log->l_mp, ticket);
2425 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2426 }
2427
2428 len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2429 if (start_lsn)
2430 *start_lsn = 0;
2431 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2432 void *ptr;
2433 int log_offset;
2434
2435 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2436 &contwr, &log_offset);
2437 if (error)
2438 return error;
2439
2440 ASSERT(log_offset <= iclog->ic_size - 1);
2441 ptr = iclog->ic_datap + log_offset;
2442
2443 /* Start_lsn is the first lsn written to. */
2444 if (start_lsn && !*start_lsn)
2445 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2446
2447 /*
2448 * This loop writes out as many regions as can fit in the amount
2449 * of space which was allocated by xlog_state_get_iclog_space().
2450 */
2451 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2452 struct xfs_log_iovec *reg;
2453 struct xlog_op_header *ophdr;
2454 int copy_len;
2455 int copy_off;
2456 bool ordered = false;
2457 bool wrote_start_rec = false;
2458
2459 /* ordered log vectors have no regions to write */
2460 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2461 ASSERT(lv->lv_niovecs == 0);
2462 ordered = true;
2463 goto next_lv;
2464 }
2465
2466 reg = &vecp[index];
2467 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2468 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2469
2470 /*
2471 * Before we start formatting log vectors, we need to
2472 * write a start record. Only do this for the first
2473 * iclog we write to.
2474 */
2475 if (optype & XLOG_START_TRANS) {
2476 xlog_write_start_rec(ptr, ticket);
2477 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2478 sizeof(struct xlog_op_header));
2479 optype &= ~XLOG_START_TRANS;
2480 wrote_start_rec = true;
2481 }
2482
2483 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2484 if (!ophdr)
2485 return -EIO;
2486
2487 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2488 sizeof(struct xlog_op_header));
2489
2490 len += xlog_write_setup_copy(ticket, ophdr,
2491 iclog->ic_size-log_offset,
2492 reg->i_len,
2493 ©_off, ©_len,
2494 &partial_copy,
2495 &partial_copy_len);
2496 xlog_verify_dest_ptr(log, ptr);
2497
2498 /*
2499 * Copy region.
2500 *
2501 * Unmount records just log an opheader, so can have
2502 * empty payloads with no data region to copy. Hence we
2503 * only copy the payload if the vector says it has data
2504 * to copy.
2505 */
2506 ASSERT(copy_len >= 0);
2507 if (copy_len > 0) {
2508 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2509 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2510 copy_len);
2511 }
2512 copy_len += sizeof(struct xlog_op_header);
2513 record_cnt++;
2514 if (wrote_start_rec) {
2515 copy_len += sizeof(struct xlog_op_header);
2516 record_cnt++;
2517 }
2518 data_cnt += contwr ? copy_len : 0;
2519
2520 error = xlog_write_copy_finish(log, iclog, optype,
2521 &record_cnt, &data_cnt,
2522 &partial_copy,
2523 &partial_copy_len,
2524 log_offset,
2525 commit_iclog);
2526 if (error)
2527 return error;
2528
2529 /*
2530 * if we had a partial copy, we need to get more iclog
2531 * space but we don't want to increment the region
2532 * index because there is still more is this region to
2533 * write.
2534 *
2535 * If we completed writing this region, and we flushed
2536 * the iclog (indicated by resetting of the record
2537 * count), then we also need to get more log space. If
2538 * this was the last record, though, we are done and
2539 * can just return.
2540 */
2541 if (partial_copy)
2542 break;
2543
2544 if (++index == lv->lv_niovecs) {
2545next_lv:
2546 lv = lv->lv_next;
2547 index = 0;
2548 if (lv)
2549 vecp = lv->lv_iovecp;
2550 }
2551 if (record_cnt == 0 && !ordered) {
2552 if (!lv)
2553 return 0;
2554 break;
2555 }
2556 }
2557 }
2558
2559 ASSERT(len == 0);
2560
2561 spin_lock(&log->l_icloglock);
2562 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2563 if (commit_iclog) {
2564 ASSERT(optype & XLOG_COMMIT_TRANS);
2565 *commit_iclog = iclog;
2566 } else {
2567 error = xlog_state_release_iclog(log, iclog, 0);
2568 }
2569 spin_unlock(&log->l_icloglock);
2570
2571 return error;
2572}
2573
2574static void
2575xlog_state_activate_iclog(
2576 struct xlog_in_core *iclog,
2577 int *iclogs_changed)
2578{
2579 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2580 trace_xlog_iclog_activate(iclog, _RET_IP_);
2581
2582 /*
2583 * If the number of ops in this iclog indicate it just contains the
2584 * dummy transaction, we can change state into IDLE (the second time
2585 * around). Otherwise we should change the state into NEED a dummy.
2586 * We don't need to cover the dummy.
2587 */
2588 if (*iclogs_changed == 0 &&
2589 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2590 *iclogs_changed = 1;
2591 } else {
2592 /*
2593 * We have two dirty iclogs so start over. This could also be
2594 * num of ops indicating this is not the dummy going out.
2595 */
2596 *iclogs_changed = 2;
2597 }
2598
2599 iclog->ic_state = XLOG_STATE_ACTIVE;
2600 iclog->ic_offset = 0;
2601 iclog->ic_header.h_num_logops = 0;
2602 memset(iclog->ic_header.h_cycle_data, 0,
2603 sizeof(iclog->ic_header.h_cycle_data));
2604 iclog->ic_header.h_lsn = 0;
2605 iclog->ic_header.h_tail_lsn = 0;
2606}
2607
2608/*
2609 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2610 * ACTIVE after iclog I/O has completed.
2611 */
2612static void
2613xlog_state_activate_iclogs(
2614 struct xlog *log,
2615 int *iclogs_changed)
2616{
2617 struct xlog_in_core *iclog = log->l_iclog;
2618
2619 do {
2620 if (iclog->ic_state == XLOG_STATE_DIRTY)
2621 xlog_state_activate_iclog(iclog, iclogs_changed);
2622 /*
2623 * The ordering of marking iclogs ACTIVE must be maintained, so
2624 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2625 */
2626 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2627 break;
2628 } while ((iclog = iclog->ic_next) != log->l_iclog);
2629}
2630
2631static int
2632xlog_covered_state(
2633 int prev_state,
2634 int iclogs_changed)
2635{
2636 /*
2637 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2638 * wrote the first covering record (DONE). We go to IDLE if we just
2639 * wrote the second covering record (DONE2) and remain in IDLE until a
2640 * non-covering write occurs.
2641 */
2642 switch (prev_state) {
2643 case XLOG_STATE_COVER_IDLE:
2644 if (iclogs_changed == 1)
2645 return XLOG_STATE_COVER_IDLE;
2646 fallthrough;
2647 case XLOG_STATE_COVER_NEED:
2648 case XLOG_STATE_COVER_NEED2:
2649 break;
2650 case XLOG_STATE_COVER_DONE:
2651 if (iclogs_changed == 1)
2652 return XLOG_STATE_COVER_NEED2;
2653 break;
2654 case XLOG_STATE_COVER_DONE2:
2655 if (iclogs_changed == 1)
2656 return XLOG_STATE_COVER_IDLE;
2657 break;
2658 default:
2659 ASSERT(0);
2660 }
2661
2662 return XLOG_STATE_COVER_NEED;
2663}
2664
2665STATIC void
2666xlog_state_clean_iclog(
2667 struct xlog *log,
2668 struct xlog_in_core *dirty_iclog)
2669{
2670 int iclogs_changed = 0;
2671
2672 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2673
2674 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2675
2676 xlog_state_activate_iclogs(log, &iclogs_changed);
2677 wake_up_all(&dirty_iclog->ic_force_wait);
2678
2679 if (iclogs_changed) {
2680 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2681 iclogs_changed);
2682 }
2683}
2684
2685STATIC xfs_lsn_t
2686xlog_get_lowest_lsn(
2687 struct xlog *log)
2688{
2689 struct xlog_in_core *iclog = log->l_iclog;
2690 xfs_lsn_t lowest_lsn = 0, lsn;
2691
2692 do {
2693 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2694 iclog->ic_state == XLOG_STATE_DIRTY)
2695 continue;
2696
2697 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2698 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2699 lowest_lsn = lsn;
2700 } while ((iclog = iclog->ic_next) != log->l_iclog);
2701
2702 return lowest_lsn;
2703}
2704
2705/*
2706 * Completion of a iclog IO does not imply that a transaction has completed, as
2707 * transactions can be large enough to span many iclogs. We cannot change the
2708 * tail of the log half way through a transaction as this may be the only
2709 * transaction in the log and moving the tail to point to the middle of it
2710 * will prevent recovery from finding the start of the transaction. Hence we
2711 * should only update the last_sync_lsn if this iclog contains transaction
2712 * completion callbacks on it.
2713 *
2714 * We have to do this before we drop the icloglock to ensure we are the only one
2715 * that can update it.
2716 *
2717 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2718 * the reservation grant head pushing. This is due to the fact that the push
2719 * target is bound by the current last_sync_lsn value. Hence if we have a large
2720 * amount of log space bound up in this committing transaction then the
2721 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2722 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2723 * should push the AIL to ensure the push target (and hence the grant head) is
2724 * no longer bound by the old log head location and can move forwards and make
2725 * progress again.
2726 */
2727static void
2728xlog_state_set_callback(
2729 struct xlog *log,
2730 struct xlog_in_core *iclog,
2731 xfs_lsn_t header_lsn)
2732{
2733 trace_xlog_iclog_callback(iclog, _RET_IP_);
2734 iclog->ic_state = XLOG_STATE_CALLBACK;
2735
2736 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2737 header_lsn) <= 0);
2738
2739 if (list_empty_careful(&iclog->ic_callbacks))
2740 return;
2741
2742 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2743 xlog_grant_push_ail(log, 0);
2744}
2745
2746/*
2747 * Return true if we need to stop processing, false to continue to the next
2748 * iclog. The caller will need to run callbacks if the iclog is returned in the
2749 * XLOG_STATE_CALLBACK state.
2750 */
2751static bool
2752xlog_state_iodone_process_iclog(
2753 struct xlog *log,
2754 struct xlog_in_core *iclog,
2755 bool *ioerror)
2756{
2757 xfs_lsn_t lowest_lsn;
2758 xfs_lsn_t header_lsn;
2759
2760 switch (iclog->ic_state) {
2761 case XLOG_STATE_ACTIVE:
2762 case XLOG_STATE_DIRTY:
2763 /*
2764 * Skip all iclogs in the ACTIVE & DIRTY states:
2765 */
2766 return false;
2767 case XLOG_STATE_IOERROR:
2768 /*
2769 * Between marking a filesystem SHUTDOWN and stopping the log,
2770 * we do flush all iclogs to disk (if there wasn't a log I/O
2771 * error). So, we do want things to go smoothly in case of just
2772 * a SHUTDOWN w/o a LOG_IO_ERROR.
2773 */
2774 *ioerror = true;
2775 return false;
2776 case XLOG_STATE_DONE_SYNC:
2777 /*
2778 * Now that we have an iclog that is in the DONE_SYNC state, do
2779 * one more check here to see if we have chased our tail around.
2780 * If this is not the lowest lsn iclog, then we will leave it
2781 * for another completion to process.
2782 */
2783 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2784 lowest_lsn = xlog_get_lowest_lsn(log);
2785 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2786 return false;
2787 xlog_state_set_callback(log, iclog, header_lsn);
2788 return false;
2789 default:
2790 /*
2791 * Can only perform callbacks in order. Since this iclog is not
2792 * in the DONE_SYNC state, we skip the rest and just try to
2793 * clean up.
2794 */
2795 return true;
2796 }
2797}
2798
2799STATIC void
2800xlog_state_do_callback(
2801 struct xlog *log)
2802{
2803 struct xlog_in_core *iclog;
2804 struct xlog_in_core *first_iclog;
2805 bool cycled_icloglock;
2806 bool ioerror;
2807 int flushcnt = 0;
2808 int repeats = 0;
2809
2810 spin_lock(&log->l_icloglock);
2811 do {
2812 /*
2813 * Scan all iclogs starting with the one pointed to by the
2814 * log. Reset this starting point each time the log is
2815 * unlocked (during callbacks).
2816 *
2817 * Keep looping through iclogs until one full pass is made
2818 * without running any callbacks.
2819 */
2820 first_iclog = log->l_iclog;
2821 iclog = log->l_iclog;
2822 cycled_icloglock = false;
2823 ioerror = false;
2824 repeats++;
2825
2826 do {
2827 LIST_HEAD(cb_list);
2828
2829 if (xlog_state_iodone_process_iclog(log, iclog,
2830 &ioerror))
2831 break;
2832
2833 if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2834 iclog->ic_state != XLOG_STATE_IOERROR) {
2835 iclog = iclog->ic_next;
2836 continue;
2837 }
2838 list_splice_init(&iclog->ic_callbacks, &cb_list);
2839 spin_unlock(&log->l_icloglock);
2840
2841 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2842 xlog_cil_process_committed(&cb_list);
2843 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2844 cycled_icloglock = true;
2845
2846 spin_lock(&log->l_icloglock);
2847 if (XLOG_FORCED_SHUTDOWN(log))
2848 wake_up_all(&iclog->ic_force_wait);
2849 else
2850 xlog_state_clean_iclog(log, iclog);
2851 iclog = iclog->ic_next;
2852 } while (first_iclog != iclog);
2853
2854 if (repeats > 5000) {
2855 flushcnt += repeats;
2856 repeats = 0;
2857 xfs_warn(log->l_mp,
2858 "%s: possible infinite loop (%d iterations)",
2859 __func__, flushcnt);
2860 }
2861 } while (!ioerror && cycled_icloglock);
2862
2863 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2864 log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2865 wake_up_all(&log->l_flush_wait);
2866
2867 spin_unlock(&log->l_icloglock);
2868}
2869
2870
2871/*
2872 * Finish transitioning this iclog to the dirty state.
2873 *
2874 * Make sure that we completely execute this routine only when this is
2875 * the last call to the iclog. There is a good chance that iclog flushes,
2876 * when we reach the end of the physical log, get turned into 2 separate
2877 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2878 * routine. By using the reference count bwritecnt, we guarantee that only
2879 * the second completion goes through.
2880 *
2881 * Callbacks could take time, so they are done outside the scope of the
2882 * global state machine log lock.
2883 */
2884STATIC void
2885xlog_state_done_syncing(
2886 struct xlog_in_core *iclog)
2887{
2888 struct xlog *log = iclog->ic_log;
2889
2890 spin_lock(&log->l_icloglock);
2891 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2892 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2893
2894 /*
2895 * If we got an error, either on the first buffer, or in the case of
2896 * split log writes, on the second, we shut down the file system and
2897 * no iclogs should ever be attempted to be written to disk again.
2898 */
2899 if (!XLOG_FORCED_SHUTDOWN(log)) {
2900 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2901 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2902 }
2903
2904 /*
2905 * Someone could be sleeping prior to writing out the next
2906 * iclog buffer, we wake them all, one will get to do the
2907 * I/O, the others get to wait for the result.
2908 */
2909 wake_up_all(&iclog->ic_write_wait);
2910 spin_unlock(&log->l_icloglock);
2911 xlog_state_do_callback(log);
2912}
2913
2914/*
2915 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2916 * sleep. We wait on the flush queue on the head iclog as that should be
2917 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2918 * we will wait here and all new writes will sleep until a sync completes.
2919 *
2920 * The in-core logs are used in a circular fashion. They are not used
2921 * out-of-order even when an iclog past the head is free.
2922 *
2923 * return:
2924 * * log_offset where xlog_write() can start writing into the in-core
2925 * log's data space.
2926 * * in-core log pointer to which xlog_write() should write.
2927 * * boolean indicating this is a continued write to an in-core log.
2928 * If this is the last write, then the in-core log's offset field
2929 * needs to be incremented, depending on the amount of data which
2930 * is copied.
2931 */
2932STATIC int
2933xlog_state_get_iclog_space(
2934 struct xlog *log,
2935 int len,
2936 struct xlog_in_core **iclogp,
2937 struct xlog_ticket *ticket,
2938 int *continued_write,
2939 int *logoffsetp)
2940{
2941 int log_offset;
2942 xlog_rec_header_t *head;
2943 xlog_in_core_t *iclog;
2944
2945restart:
2946 spin_lock(&log->l_icloglock);
2947 if (XLOG_FORCED_SHUTDOWN(log)) {
2948 spin_unlock(&log->l_icloglock);
2949 return -EIO;
2950 }
2951
2952 iclog = log->l_iclog;
2953 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2954 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2955
2956 /* Wait for log writes to have flushed */
2957 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2958 goto restart;
2959 }
2960
2961 head = &iclog->ic_header;
2962
2963 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2964 log_offset = iclog->ic_offset;
2965
2966 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2967
2968 /* On the 1st write to an iclog, figure out lsn. This works
2969 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2970 * committing to. If the offset is set, that's how many blocks
2971 * must be written.
2972 */
2973 if (log_offset == 0) {
2974 ticket->t_curr_res -= log->l_iclog_hsize;
2975 xlog_tic_add_region(ticket,
2976 log->l_iclog_hsize,
2977 XLOG_REG_TYPE_LRHEADER);
2978 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2979 head->h_lsn = cpu_to_be64(
2980 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2981 ASSERT(log->l_curr_block >= 0);
2982 }
2983
2984 /* If there is enough room to write everything, then do it. Otherwise,
2985 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2986 * bit is on, so this will get flushed out. Don't update ic_offset
2987 * until you know exactly how many bytes get copied. Therefore, wait
2988 * until later to update ic_offset.
2989 *
2990 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2991 * can fit into remaining data section.
2992 */
2993 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2994 int error = 0;
2995
2996 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2997
2998 /*
2999 * If we are the only one writing to this iclog, sync it to
3000 * disk. We need to do an atomic compare and decrement here to
3001 * avoid racing with concurrent atomic_dec_and_lock() calls in
3002 * xlog_state_release_iclog() when there is more than one
3003 * reference to the iclog.
3004 */
3005 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3006 error = xlog_state_release_iclog(log, iclog, 0);
3007 spin_unlock(&log->l_icloglock);
3008 if (error)
3009 return error;
3010 goto restart;
3011 }
3012
3013 /* Do we have enough room to write the full amount in the remainder
3014 * of this iclog? Or must we continue a write on the next iclog and
3015 * mark this iclog as completely taken? In the case where we switch
3016 * iclogs (to mark it taken), this particular iclog will release/sync
3017 * to disk in xlog_write().
3018 */
3019 if (len <= iclog->ic_size - iclog->ic_offset) {
3020 *continued_write = 0;
3021 iclog->ic_offset += len;
3022 } else {
3023 *continued_write = 1;
3024 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3025 }
3026 *iclogp = iclog;
3027
3028 ASSERT(iclog->ic_offset <= iclog->ic_size);
3029 spin_unlock(&log->l_icloglock);
3030
3031 *logoffsetp = log_offset;
3032 return 0;
3033}
3034
3035/*
3036 * The first cnt-1 times a ticket goes through here we don't need to move the
3037 * grant write head because the permanent reservation has reserved cnt times the
3038 * unit amount. Release part of current permanent unit reservation and reset
3039 * current reservation to be one units worth. Also move grant reservation head
3040 * forward.
3041 */
3042void
3043xfs_log_ticket_regrant(
3044 struct xlog *log,
3045 struct xlog_ticket *ticket)
3046{
3047 trace_xfs_log_ticket_regrant(log, ticket);
3048
3049 if (ticket->t_cnt > 0)
3050 ticket->t_cnt--;
3051
3052 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3053 ticket->t_curr_res);
3054 xlog_grant_sub_space(log, &log->l_write_head.grant,
3055 ticket->t_curr_res);
3056 ticket->t_curr_res = ticket->t_unit_res;
3057 xlog_tic_reset_res(ticket);
3058
3059 trace_xfs_log_ticket_regrant_sub(log, ticket);
3060
3061 /* just return if we still have some of the pre-reserved space */
3062 if (!ticket->t_cnt) {
3063 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3064 ticket->t_unit_res);
3065 trace_xfs_log_ticket_regrant_exit(log, ticket);
3066
3067 ticket->t_curr_res = ticket->t_unit_res;
3068 xlog_tic_reset_res(ticket);
3069 }
3070
3071 xfs_log_ticket_put(ticket);
3072}
3073
3074/*
3075 * Give back the space left from a reservation.
3076 *
3077 * All the information we need to make a correct determination of space left
3078 * is present. For non-permanent reservations, things are quite easy. The
3079 * count should have been decremented to zero. We only need to deal with the
3080 * space remaining in the current reservation part of the ticket. If the
3081 * ticket contains a permanent reservation, there may be left over space which
3082 * needs to be released. A count of N means that N-1 refills of the current
3083 * reservation can be done before we need to ask for more space. The first
3084 * one goes to fill up the first current reservation. Once we run out of
3085 * space, the count will stay at zero and the only space remaining will be
3086 * in the current reservation field.
3087 */
3088void
3089xfs_log_ticket_ungrant(
3090 struct xlog *log,
3091 struct xlog_ticket *ticket)
3092{
3093 int bytes;
3094
3095 trace_xfs_log_ticket_ungrant(log, ticket);
3096
3097 if (ticket->t_cnt > 0)
3098 ticket->t_cnt--;
3099
3100 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3101
3102 /*
3103 * If this is a permanent reservation ticket, we may be able to free
3104 * up more space based on the remaining count.
3105 */
3106 bytes = ticket->t_curr_res;
3107 if (ticket->t_cnt > 0) {
3108 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3109 bytes += ticket->t_unit_res*ticket->t_cnt;
3110 }
3111
3112 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3113 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3114
3115 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3116
3117 xfs_log_space_wake(log->l_mp);
3118 xfs_log_ticket_put(ticket);
3119}
3120
3121/*
3122 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3123 * the current iclog pointer to the next iclog in the ring.
3124 */
3125STATIC void
3126xlog_state_switch_iclogs(
3127 struct xlog *log,
3128 struct xlog_in_core *iclog,
3129 int eventual_size)
3130{
3131 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3132 assert_spin_locked(&log->l_icloglock);
3133 trace_xlog_iclog_switch(iclog, _RET_IP_);
3134
3135 if (!eventual_size)
3136 eventual_size = iclog->ic_offset;
3137 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3138 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3139 log->l_prev_block = log->l_curr_block;
3140 log->l_prev_cycle = log->l_curr_cycle;
3141
3142 /* roll log?: ic_offset changed later */
3143 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3144
3145 /* Round up to next log-sunit */
3146 if (log->l_iclog_roundoff > BBSIZE) {
3147 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3148 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3149 }
3150
3151 if (log->l_curr_block >= log->l_logBBsize) {
3152 /*
3153 * Rewind the current block before the cycle is bumped to make
3154 * sure that the combined LSN never transiently moves forward
3155 * when the log wraps to the next cycle. This is to support the
3156 * unlocked sample of these fields from xlog_valid_lsn(). Most
3157 * other cases should acquire l_icloglock.
3158 */
3159 log->l_curr_block -= log->l_logBBsize;
3160 ASSERT(log->l_curr_block >= 0);
3161 smp_wmb();
3162 log->l_curr_cycle++;
3163 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3164 log->l_curr_cycle++;
3165 }
3166 ASSERT(iclog == log->l_iclog);
3167 log->l_iclog = iclog->ic_next;
3168}
3169
3170/*
3171 * Force the iclog to disk and check if the iclog has been completed before
3172 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3173 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3174 * If completion has already occurred, tell the caller so that it can avoid an
3175 * unnecessary wait on the iclog.
3176 */
3177static int
3178xlog_force_and_check_iclog(
3179 struct xlog_in_core *iclog,
3180 bool *completed)
3181{
3182 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3183 int error;
3184
3185 *completed = false;
3186 error = xlog_force_iclog(iclog);
3187 if (error)
3188 return error;
3189
3190 /*
3191 * If the iclog has already been completed and reused the header LSN
3192 * will have been rewritten by completion
3193 */
3194 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3195 *completed = true;
3196 return 0;
3197}
3198
3199/*
3200 * Write out all data in the in-core log as of this exact moment in time.
3201 *
3202 * Data may be written to the in-core log during this call. However,
3203 * we don't guarantee this data will be written out. A change from past
3204 * implementation means this routine will *not* write out zero length LRs.
3205 *
3206 * Basically, we try and perform an intelligent scan of the in-core logs.
3207 * If we determine there is no flushable data, we just return. There is no
3208 * flushable data if:
3209 *
3210 * 1. the current iclog is active and has no data; the previous iclog
3211 * is in the active or dirty state.
3212 * 2. the current iclog is drity, and the previous iclog is in the
3213 * active or dirty state.
3214 *
3215 * We may sleep if:
3216 *
3217 * 1. the current iclog is not in the active nor dirty state.
3218 * 2. the current iclog dirty, and the previous iclog is not in the
3219 * active nor dirty state.
3220 * 3. the current iclog is active, and there is another thread writing
3221 * to this particular iclog.
3222 * 4. a) the current iclog is active and has no other writers
3223 * b) when we return from flushing out this iclog, it is still
3224 * not in the active nor dirty state.
3225 */
3226int
3227xfs_log_force(
3228 struct xfs_mount *mp,
3229 uint flags)
3230{
3231 struct xlog *log = mp->m_log;
3232 struct xlog_in_core *iclog;
3233
3234 XFS_STATS_INC(mp, xs_log_force);
3235 trace_xfs_log_force(mp, 0, _RET_IP_);
3236
3237 xlog_cil_force(log);
3238
3239 spin_lock(&log->l_icloglock);
3240 iclog = log->l_iclog;
3241 if (iclog->ic_state == XLOG_STATE_IOERROR)
3242 goto out_error;
3243
3244 trace_xlog_iclog_force(iclog, _RET_IP_);
3245
3246 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3247 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3248 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3249 /*
3250 * If the head is dirty or (active and empty), then we need to
3251 * look at the previous iclog.
3252 *
3253 * If the previous iclog is active or dirty we are done. There
3254 * is nothing to sync out. Otherwise, we attach ourselves to the
3255 * previous iclog and go to sleep.
3256 */
3257 iclog = iclog->ic_prev;
3258 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3259 if (atomic_read(&iclog->ic_refcnt) == 0) {
3260 /* We have exclusive access to this iclog. */
3261 bool completed;
3262
3263 if (xlog_force_and_check_iclog(iclog, &completed))
3264 goto out_error;
3265
3266 if (completed)
3267 goto out_unlock;
3268 } else {
3269 /*
3270 * Someone else is still writing to this iclog, so we
3271 * need to ensure that when they release the iclog it
3272 * gets synced immediately as we may be waiting on it.
3273 */
3274 xlog_state_switch_iclogs(log, iclog, 0);
3275 }
3276 }
3277
3278 /*
3279 * The iclog we are about to wait on may contain the checkpoint pushed
3280 * by the above xlog_cil_force() call, but it may not have been pushed
3281 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3282 * are flushed when this iclog is written.
3283 */
3284 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3285 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3286
3287 if (flags & XFS_LOG_SYNC)
3288 return xlog_wait_on_iclog(iclog);
3289out_unlock:
3290 spin_unlock(&log->l_icloglock);
3291 return 0;
3292out_error:
3293 spin_unlock(&log->l_icloglock);
3294 return -EIO;
3295}
3296
3297static int
3298xlog_force_lsn(
3299 struct xlog *log,
3300 xfs_lsn_t lsn,
3301 uint flags,
3302 int *log_flushed,
3303 bool already_slept)
3304{
3305 struct xlog_in_core *iclog;
3306 bool completed;
3307
3308 spin_lock(&log->l_icloglock);
3309 iclog = log->l_iclog;
3310 if (iclog->ic_state == XLOG_STATE_IOERROR)
3311 goto out_error;
3312
3313 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3314 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3315 iclog = iclog->ic_next;
3316 if (iclog == log->l_iclog)
3317 goto out_unlock;
3318 }
3319
3320 switch (iclog->ic_state) {
3321 case XLOG_STATE_ACTIVE:
3322 /*
3323 * We sleep here if we haven't already slept (e.g. this is the
3324 * first time we've looked at the correct iclog buf) and the
3325 * buffer before us is going to be sync'ed. The reason for this
3326 * is that if we are doing sync transactions here, by waiting
3327 * for the previous I/O to complete, we can allow a few more
3328 * transactions into this iclog before we close it down.
3329 *
3330 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3331 * refcnt so we can release the log (which drops the ref count).
3332 * The state switch keeps new transaction commits from using
3333 * this buffer. When the current commits finish writing into
3334 * the buffer, the refcount will drop to zero and the buffer
3335 * will go out then.
3336 */
3337 if (!already_slept &&
3338 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3339 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3340 xlog_wait(&iclog->ic_prev->ic_write_wait,
3341 &log->l_icloglock);
3342 return -EAGAIN;
3343 }
3344 if (xlog_force_and_check_iclog(iclog, &completed))
3345 goto out_error;
3346 if (log_flushed)
3347 *log_flushed = 1;
3348 if (completed)
3349 goto out_unlock;
3350 break;
3351 case XLOG_STATE_WANT_SYNC:
3352 /*
3353 * This iclog may contain the checkpoint pushed by the
3354 * xlog_cil_force_seq() call, but there are other writers still
3355 * accessing it so it hasn't been pushed to disk yet. Like the
3356 * ACTIVE case above, we need to make sure caches are flushed
3357 * when this iclog is written.
3358 */
3359 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3360 break;
3361 default:
3362 /*
3363 * The entire checkpoint was written by the CIL force and is on
3364 * its way to disk already. It will be stable when it
3365 * completes, so we don't need to manipulate caches here at all.
3366 * We just need to wait for completion if necessary.
3367 */
3368 break;
3369 }
3370
3371 if (flags & XFS_LOG_SYNC)
3372 return xlog_wait_on_iclog(iclog);
3373out_unlock:
3374 spin_unlock(&log->l_icloglock);
3375 return 0;
3376out_error:
3377 spin_unlock(&log->l_icloglock);
3378 return -EIO;
3379}
3380
3381/*
3382 * Force the in-core log to disk for a specific LSN.
3383 *
3384 * Find in-core log with lsn.
3385 * If it is in the DIRTY state, just return.
3386 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3387 * state and go to sleep or return.
3388 * If it is in any other state, go to sleep or return.
3389 *
3390 * Synchronous forces are implemented with a wait queue. All callers trying
3391 * to force a given lsn to disk must wait on the queue attached to the
3392 * specific in-core log. When given in-core log finally completes its write
3393 * to disk, that thread will wake up all threads waiting on the queue.
3394 */
3395int
3396xfs_log_force_seq(
3397 struct xfs_mount *mp,
3398 xfs_csn_t seq,
3399 uint flags,
3400 int *log_flushed)
3401{
3402 struct xlog *log = mp->m_log;
3403 xfs_lsn_t lsn;
3404 int ret;
3405 ASSERT(seq != 0);
3406
3407 XFS_STATS_INC(mp, xs_log_force);
3408 trace_xfs_log_force(mp, seq, _RET_IP_);
3409
3410 lsn = xlog_cil_force_seq(log, seq);
3411 if (lsn == NULLCOMMITLSN)
3412 return 0;
3413
3414 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3415 if (ret == -EAGAIN) {
3416 XFS_STATS_INC(mp, xs_log_force_sleep);
3417 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3418 }
3419 return ret;
3420}
3421
3422/*
3423 * Free a used ticket when its refcount falls to zero.
3424 */
3425void
3426xfs_log_ticket_put(
3427 xlog_ticket_t *ticket)
3428{
3429 ASSERT(atomic_read(&ticket->t_ref) > 0);
3430 if (atomic_dec_and_test(&ticket->t_ref))
3431 kmem_cache_free(xfs_log_ticket_zone, ticket);
3432}
3433
3434xlog_ticket_t *
3435xfs_log_ticket_get(
3436 xlog_ticket_t *ticket)
3437{
3438 ASSERT(atomic_read(&ticket->t_ref) > 0);
3439 atomic_inc(&ticket->t_ref);
3440 return ticket;
3441}
3442
3443/*
3444 * Figure out the total log space unit (in bytes) that would be
3445 * required for a log ticket.
3446 */
3447static int
3448xlog_calc_unit_res(
3449 struct xlog *log,
3450 int unit_bytes)
3451{
3452 int iclog_space;
3453 uint num_headers;
3454
3455 /*
3456 * Permanent reservations have up to 'cnt'-1 active log operations
3457 * in the log. A unit in this case is the amount of space for one
3458 * of these log operations. Normal reservations have a cnt of 1
3459 * and their unit amount is the total amount of space required.
3460 *
3461 * The following lines of code account for non-transaction data
3462 * which occupy space in the on-disk log.
3463 *
3464 * Normal form of a transaction is:
3465 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3466 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3467 *
3468 * We need to account for all the leadup data and trailer data
3469 * around the transaction data.
3470 * And then we need to account for the worst case in terms of using
3471 * more space.
3472 * The worst case will happen if:
3473 * - the placement of the transaction happens to be such that the
3474 * roundoff is at its maximum
3475 * - the transaction data is synced before the commit record is synced
3476 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3477 * Therefore the commit record is in its own Log Record.
3478 * This can happen as the commit record is called with its
3479 * own region to xlog_write().
3480 * This then means that in the worst case, roundoff can happen for
3481 * the commit-rec as well.
3482 * The commit-rec is smaller than padding in this scenario and so it is
3483 * not added separately.
3484 */
3485
3486 /* for trans header */
3487 unit_bytes += sizeof(xlog_op_header_t);
3488 unit_bytes += sizeof(xfs_trans_header_t);
3489
3490 /* for start-rec */
3491 unit_bytes += sizeof(xlog_op_header_t);
3492
3493 /*
3494 * for LR headers - the space for data in an iclog is the size minus
3495 * the space used for the headers. If we use the iclog size, then we
3496 * undercalculate the number of headers required.
3497 *
3498 * Furthermore - the addition of op headers for split-recs might
3499 * increase the space required enough to require more log and op
3500 * headers, so take that into account too.
3501 *
3502 * IMPORTANT: This reservation makes the assumption that if this
3503 * transaction is the first in an iclog and hence has the LR headers
3504 * accounted to it, then the remaining space in the iclog is
3505 * exclusively for this transaction. i.e. if the transaction is larger
3506 * than the iclog, it will be the only thing in that iclog.
3507 * Fundamentally, this means we must pass the entire log vector to
3508 * xlog_write to guarantee this.
3509 */
3510 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3511 num_headers = howmany(unit_bytes, iclog_space);
3512
3513 /* for split-recs - ophdrs added when data split over LRs */
3514 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3515
3516 /* add extra header reservations if we overrun */
3517 while (!num_headers ||
3518 howmany(unit_bytes, iclog_space) > num_headers) {
3519 unit_bytes += sizeof(xlog_op_header_t);
3520 num_headers++;
3521 }
3522 unit_bytes += log->l_iclog_hsize * num_headers;
3523
3524 /* for commit-rec LR header - note: padding will subsume the ophdr */
3525 unit_bytes += log->l_iclog_hsize;
3526
3527 /* roundoff padding for transaction data and one for commit record */
3528 unit_bytes += 2 * log->l_iclog_roundoff;
3529
3530 return unit_bytes;
3531}
3532
3533int
3534xfs_log_calc_unit_res(
3535 struct xfs_mount *mp,
3536 int unit_bytes)
3537{
3538 return xlog_calc_unit_res(mp->m_log, unit_bytes);
3539}
3540
3541/*
3542 * Allocate and initialise a new log ticket.
3543 */
3544struct xlog_ticket *
3545xlog_ticket_alloc(
3546 struct xlog *log,
3547 int unit_bytes,
3548 int cnt,
3549 char client,
3550 bool permanent)
3551{
3552 struct xlog_ticket *tic;
3553 int unit_res;
3554
3555 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3556
3557 unit_res = xlog_calc_unit_res(log, unit_bytes);
3558
3559 atomic_set(&tic->t_ref, 1);
3560 tic->t_task = current;
3561 INIT_LIST_HEAD(&tic->t_queue);
3562 tic->t_unit_res = unit_res;
3563 tic->t_curr_res = unit_res;
3564 tic->t_cnt = cnt;
3565 tic->t_ocnt = cnt;
3566 tic->t_tid = prandom_u32();
3567 tic->t_clientid = client;
3568 if (permanent)
3569 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3570
3571 xlog_tic_reset_res(tic);
3572
3573 return tic;
3574}
3575
3576#if defined(DEBUG)
3577/*
3578 * Make sure that the destination ptr is within the valid data region of
3579 * one of the iclogs. This uses backup pointers stored in a different
3580 * part of the log in case we trash the log structure.
3581 */
3582STATIC void
3583xlog_verify_dest_ptr(
3584 struct xlog *log,
3585 void *ptr)
3586{
3587 int i;
3588 int good_ptr = 0;
3589
3590 for (i = 0; i < log->l_iclog_bufs; i++) {
3591 if (ptr >= log->l_iclog_bak[i] &&
3592 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3593 good_ptr++;
3594 }
3595
3596 if (!good_ptr)
3597 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3598}
3599
3600/*
3601 * Check to make sure the grant write head didn't just over lap the tail. If
3602 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3603 * the cycles differ by exactly one and check the byte count.
3604 *
3605 * This check is run unlocked, so can give false positives. Rather than assert
3606 * on failures, use a warn-once flag and a panic tag to allow the admin to
3607 * determine if they want to panic the machine when such an error occurs. For
3608 * debug kernels this will have the same effect as using an assert but, unlinke
3609 * an assert, it can be turned off at runtime.
3610 */
3611STATIC void
3612xlog_verify_grant_tail(
3613 struct xlog *log)
3614{
3615 int tail_cycle, tail_blocks;
3616 int cycle, space;
3617
3618 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3619 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3620 if (tail_cycle != cycle) {
3621 if (cycle - 1 != tail_cycle &&
3622 !(log->l_flags & XLOG_TAIL_WARN)) {
3623 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3624 "%s: cycle - 1 != tail_cycle", __func__);
3625 log->l_flags |= XLOG_TAIL_WARN;
3626 }
3627
3628 if (space > BBTOB(tail_blocks) &&
3629 !(log->l_flags & XLOG_TAIL_WARN)) {
3630 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3631 "%s: space > BBTOB(tail_blocks)", __func__);
3632 log->l_flags |= XLOG_TAIL_WARN;
3633 }
3634 }
3635}
3636
3637/* check if it will fit */
3638STATIC void
3639xlog_verify_tail_lsn(
3640 struct xlog *log,
3641 struct xlog_in_core *iclog)
3642{
3643 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3644 int blocks;
3645
3646 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3647 blocks =
3648 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3649 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3650 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3651 } else {
3652 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3653
3654 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3655 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3656
3657 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3658 if (blocks < BTOBB(iclog->ic_offset) + 1)
3659 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3660 }
3661}
3662
3663/*
3664 * Perform a number of checks on the iclog before writing to disk.
3665 *
3666 * 1. Make sure the iclogs are still circular
3667 * 2. Make sure we have a good magic number
3668 * 3. Make sure we don't have magic numbers in the data
3669 * 4. Check fields of each log operation header for:
3670 * A. Valid client identifier
3671 * B. tid ptr value falls in valid ptr space (user space code)
3672 * C. Length in log record header is correct according to the
3673 * individual operation headers within record.
3674 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3675 * log, check the preceding blocks of the physical log to make sure all
3676 * the cycle numbers agree with the current cycle number.
3677 */
3678STATIC void
3679xlog_verify_iclog(
3680 struct xlog *log,
3681 struct xlog_in_core *iclog,
3682 int count)
3683{
3684 xlog_op_header_t *ophead;
3685 xlog_in_core_t *icptr;
3686 xlog_in_core_2_t *xhdr;
3687 void *base_ptr, *ptr, *p;
3688 ptrdiff_t field_offset;
3689 uint8_t clientid;
3690 int len, i, j, k, op_len;
3691 int idx;
3692
3693 /* check validity of iclog pointers */
3694 spin_lock(&log->l_icloglock);
3695 icptr = log->l_iclog;
3696 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3697 ASSERT(icptr);
3698
3699 if (icptr != log->l_iclog)
3700 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3701 spin_unlock(&log->l_icloglock);
3702
3703 /* check log magic numbers */
3704 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3705 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3706
3707 base_ptr = ptr = &iclog->ic_header;
3708 p = &iclog->ic_header;
3709 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3710 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3711 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3712 __func__);
3713 }
3714
3715 /* check fields */
3716 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3717 base_ptr = ptr = iclog->ic_datap;
3718 ophead = ptr;
3719 xhdr = iclog->ic_data;
3720 for (i = 0; i < len; i++) {
3721 ophead = ptr;
3722
3723 /* clientid is only 1 byte */
3724 p = &ophead->oh_clientid;
3725 field_offset = p - base_ptr;
3726 if (field_offset & 0x1ff) {
3727 clientid = ophead->oh_clientid;
3728 } else {
3729 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3730 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3731 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3732 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3733 clientid = xlog_get_client_id(
3734 xhdr[j].hic_xheader.xh_cycle_data[k]);
3735 } else {
3736 clientid = xlog_get_client_id(
3737 iclog->ic_header.h_cycle_data[idx]);
3738 }
3739 }
3740 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3741 xfs_warn(log->l_mp,
3742 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3743 __func__, clientid, ophead,
3744 (unsigned long)field_offset);
3745
3746 /* check length */
3747 p = &ophead->oh_len;
3748 field_offset = p - base_ptr;
3749 if (field_offset & 0x1ff) {
3750 op_len = be32_to_cpu(ophead->oh_len);
3751 } else {
3752 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3753 (uintptr_t)iclog->ic_datap);
3754 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3755 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3756 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3757 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3758 } else {
3759 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3760 }
3761 }
3762 ptr += sizeof(xlog_op_header_t) + op_len;
3763 }
3764}
3765#endif
3766
3767/*
3768 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3769 */
3770STATIC int
3771xlog_state_ioerror(
3772 struct xlog *log)
3773{
3774 xlog_in_core_t *iclog, *ic;
3775
3776 iclog = log->l_iclog;
3777 if (iclog->ic_state != XLOG_STATE_IOERROR) {
3778 /*
3779 * Mark all the incore logs IOERROR.
3780 * From now on, no log flushes will result.
3781 */
3782 ic = iclog;
3783 do {
3784 ic->ic_state = XLOG_STATE_IOERROR;
3785 ic = ic->ic_next;
3786 } while (ic != iclog);
3787 return 0;
3788 }
3789 /*
3790 * Return non-zero, if state transition has already happened.
3791 */
3792 return 1;
3793}
3794
3795/*
3796 * This is called from xfs_force_shutdown, when we're forcibly
3797 * shutting down the filesystem, typically because of an IO error.
3798 * Our main objectives here are to make sure that:
3799 * a. if !logerror, flush the logs to disk. Anything modified
3800 * after this is ignored.
3801 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3802 * parties to find out, 'atomically'.
3803 * c. those who're sleeping on log reservations, pinned objects and
3804 * other resources get woken up, and be told the bad news.
3805 * d. nothing new gets queued up after (b) and (c) are done.
3806 *
3807 * Note: for the !logerror case we need to flush the regions held in memory out
3808 * to disk first. This needs to be done before the log is marked as shutdown,
3809 * otherwise the iclog writes will fail.
3810 */
3811int
3812xfs_log_force_umount(
3813 struct xfs_mount *mp,
3814 int logerror)
3815{
3816 struct xlog *log;
3817 int retval;
3818
3819 log = mp->m_log;
3820
3821 /*
3822 * If this happens during log recovery, don't worry about
3823 * locking; the log isn't open for business yet.
3824 */
3825 if (!log ||
3826 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3827 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3828 if (mp->m_sb_bp)
3829 mp->m_sb_bp->b_flags |= XBF_DONE;
3830 return 0;
3831 }
3832
3833 /*
3834 * Somebody could've already done the hard work for us.
3835 * No need to get locks for this.
3836 */
3837 if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3838 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3839 return 1;
3840 }
3841
3842 /*
3843 * Flush all the completed transactions to disk before marking the log
3844 * being shut down. We need to do it in this order to ensure that
3845 * completed operations are safely on disk before we shut down, and that
3846 * we don't have to issue any buffer IO after the shutdown flags are set
3847 * to guarantee this.
3848 */
3849 if (!logerror)
3850 xfs_log_force(mp, XFS_LOG_SYNC);
3851
3852 /*
3853 * mark the filesystem and the as in a shutdown state and wake
3854 * everybody up to tell them the bad news.
3855 */
3856 spin_lock(&log->l_icloglock);
3857 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3858 if (mp->m_sb_bp)
3859 mp->m_sb_bp->b_flags |= XBF_DONE;
3860
3861 /*
3862 * Mark the log and the iclogs with IO error flags to prevent any
3863 * further log IO from being issued or completed.
3864 */
3865 log->l_flags |= XLOG_IO_ERROR;
3866 retval = xlog_state_ioerror(log);
3867 spin_unlock(&log->l_icloglock);
3868
3869 /*
3870 * We don't want anybody waiting for log reservations after this. That
3871 * means we have to wake up everybody queued up on reserveq as well as
3872 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3873 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3874 * action is protected by the grant locks.
3875 */
3876 xlog_grant_head_wake_all(&log->l_reserve_head);
3877 xlog_grant_head_wake_all(&log->l_write_head);
3878
3879 /*
3880 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3881 * as if the log writes were completed. The abort handling in the log
3882 * item committed callback functions will do this again under lock to
3883 * avoid races.
3884 */
3885 spin_lock(&log->l_cilp->xc_push_lock);
3886 wake_up_all(&log->l_cilp->xc_commit_wait);
3887 spin_unlock(&log->l_cilp->xc_push_lock);
3888 xlog_state_do_callback(log);
3889
3890 /* return non-zero if log IOERROR transition had already happened */
3891 return retval;
3892}
3893
3894STATIC int
3895xlog_iclogs_empty(
3896 struct xlog *log)
3897{
3898 xlog_in_core_t *iclog;
3899
3900 iclog = log->l_iclog;
3901 do {
3902 /* endianness does not matter here, zero is zero in
3903 * any language.
3904 */
3905 if (iclog->ic_header.h_num_logops)
3906 return 0;
3907 iclog = iclog->ic_next;
3908 } while (iclog != log->l_iclog);
3909 return 1;
3910}
3911
3912/*
3913 * Verify that an LSN stamped into a piece of metadata is valid. This is
3914 * intended for use in read verifiers on v5 superblocks.
3915 */
3916bool
3917xfs_log_check_lsn(
3918 struct xfs_mount *mp,
3919 xfs_lsn_t lsn)
3920{
3921 struct xlog *log = mp->m_log;
3922 bool valid;
3923
3924 /*
3925 * norecovery mode skips mount-time log processing and unconditionally
3926 * resets the in-core LSN. We can't validate in this mode, but
3927 * modifications are not allowed anyways so just return true.
3928 */
3929 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3930 return true;
3931
3932 /*
3933 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3934 * handled by recovery and thus safe to ignore here.
3935 */
3936 if (lsn == NULLCOMMITLSN)
3937 return true;
3938
3939 valid = xlog_valid_lsn(mp->m_log, lsn);
3940
3941 /* warn the user about what's gone wrong before verifier failure */
3942 if (!valid) {
3943 spin_lock(&log->l_icloglock);
3944 xfs_warn(mp,
3945"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3946"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3947 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3948 log->l_curr_cycle, log->l_curr_block);
3949 spin_unlock(&log->l_icloglock);
3950 }
3951
3952 return valid;
3953}
3954
3955bool
3956xfs_log_in_recovery(
3957 struct xfs_mount *mp)
3958{
3959 struct xlog *log = mp->m_log;
3960
3961 return log->l_flags & XLOG_ACTIVE_RECOVERY;
3962}