]>
Commit | Line | Data |
---|---|---|
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 | ||
24 | kmem_zone_t *xfs_log_ticket_zone; | |
25 | ||
26 | /* Local miscellaneous function prototypes */ | |
27 | STATIC struct xlog * | |
28 | xlog_alloc_log( | |
29 | struct xfs_mount *mp, | |
30 | struct xfs_buftarg *log_target, | |
31 | xfs_daddr_t blk_offset, | |
32 | int num_bblks); | |
33 | STATIC int | |
34 | xlog_space_left( | |
35 | struct xlog *log, | |
36 | atomic64_t *head); | |
37 | STATIC void | |
38 | xlog_dealloc_log( | |
39 | struct xlog *log); | |
40 | ||
41 | /* local state machine functions */ | |
42 | STATIC void xlog_state_done_syncing( | |
43 | struct xlog_in_core *iclog); | |
44 | STATIC int | |
45 | xlog_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); | |
52 | STATIC void | |
53 | xlog_state_switch_iclogs( | |
54 | struct xlog *log, | |
55 | struct xlog_in_core *iclog, | |
56 | int eventual_size); | |
57 | STATIC void | |
58 | xlog_grant_push_ail( | |
59 | struct xlog *log, | |
60 | int need_bytes); | |
61 | STATIC void | |
62 | xlog_sync( | |
63 | struct xlog *log, | |
64 | struct xlog_in_core *iclog); | |
65 | #if defined(DEBUG) | |
66 | STATIC void | |
67 | xlog_verify_dest_ptr( | |
68 | struct xlog *log, | |
69 | void *ptr); | |
70 | STATIC void | |
71 | xlog_verify_grant_tail( | |
72 | struct xlog *log); | |
73 | STATIC void | |
74 | xlog_verify_iclog( | |
75 | struct xlog *log, | |
76 | struct xlog_in_core *iclog, | |
77 | int count); | |
78 | STATIC void | |
79 | xlog_verify_tail_lsn( | |
80 | struct xlog *log, | |
81 | struct xlog_in_core *iclog, | |
82 | xfs_lsn_t tail_lsn); | |
83 | #else | |
84 | #define xlog_verify_dest_ptr(a,b) | |
85 | #define xlog_verify_grant_tail(a) | |
86 | #define xlog_verify_iclog(a,b,c) | |
87 | #define xlog_verify_tail_lsn(a,b,c) | |
88 | #endif | |
89 | ||
90 | STATIC int | |
91 | xlog_iclogs_empty( | |
92 | struct xlog *log); | |
93 | ||
94 | static int | |
95 | xfs_log_cover(struct xfs_mount *); | |
96 | ||
97 | static void | |
98 | xlog_grant_sub_space( | |
99 | struct xlog *log, | |
100 | atomic64_t *head, | |
101 | int bytes) | |
102 | { | |
103 | int64_t head_val = atomic64_read(head); | |
104 | int64_t new, old; | |
105 | ||
106 | do { | |
107 | int cycle, space; | |
108 | ||
109 | xlog_crack_grant_head_val(head_val, &cycle, &space); | |
110 | ||
111 | space -= bytes; | |
112 | if (space < 0) { | |
113 | space += log->l_logsize; | |
114 | cycle--; | |
115 | } | |
116 | ||
117 | old = head_val; | |
118 | new = xlog_assign_grant_head_val(cycle, space); | |
119 | head_val = atomic64_cmpxchg(head, old, new); | |
120 | } while (head_val != old); | |
121 | } | |
122 | ||
123 | static void | |
124 | xlog_grant_add_space( | |
125 | struct xlog *log, | |
126 | atomic64_t *head, | |
127 | int bytes) | |
128 | { | |
129 | int64_t head_val = atomic64_read(head); | |
130 | int64_t new, old; | |
131 | ||
132 | do { | |
133 | int tmp; | |
134 | int cycle, space; | |
135 | ||
136 | xlog_crack_grant_head_val(head_val, &cycle, &space); | |
137 | ||
138 | tmp = log->l_logsize - space; | |
139 | if (tmp > bytes) | |
140 | space += bytes; | |
141 | else { | |
142 | space = bytes - tmp; | |
143 | cycle++; | |
144 | } | |
145 | ||
146 | old = head_val; | |
147 | new = xlog_assign_grant_head_val(cycle, space); | |
148 | head_val = atomic64_cmpxchg(head, old, new); | |
149 | } while (head_val != old); | |
150 | } | |
151 | ||
152 | STATIC void | |
153 | xlog_grant_head_init( | |
154 | struct xlog_grant_head *head) | |
155 | { | |
156 | xlog_assign_grant_head(&head->grant, 1, 0); | |
157 | INIT_LIST_HEAD(&head->waiters); | |
158 | spin_lock_init(&head->lock); | |
159 | } | |
160 | ||
161 | STATIC void | |
162 | xlog_grant_head_wake_all( | |
163 | struct xlog_grant_head *head) | |
164 | { | |
165 | struct xlog_ticket *tic; | |
166 | ||
167 | spin_lock(&head->lock); | |
168 | list_for_each_entry(tic, &head->waiters, t_queue) | |
169 | wake_up_process(tic->t_task); | |
170 | spin_unlock(&head->lock); | |
171 | } | |
172 | ||
173 | static inline int | |
174 | xlog_ticket_reservation( | |
175 | struct xlog *log, | |
176 | struct xlog_grant_head *head, | |
177 | struct xlog_ticket *tic) | |
178 | { | |
179 | if (head == &log->l_write_head) { | |
180 | ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV); | |
181 | return tic->t_unit_res; | |
182 | } else { | |
183 | if (tic->t_flags & XLOG_TIC_PERM_RESERV) | |
184 | return tic->t_unit_res * tic->t_cnt; | |
185 | else | |
186 | return tic->t_unit_res; | |
187 | } | |
188 | } | |
189 | ||
190 | STATIC bool | |
191 | xlog_grant_head_wake( | |
192 | struct xlog *log, | |
193 | struct xlog_grant_head *head, | |
194 | int *free_bytes) | |
195 | { | |
196 | struct xlog_ticket *tic; | |
197 | int need_bytes; | |
198 | bool woken_task = false; | |
199 | ||
200 | list_for_each_entry(tic, &head->waiters, t_queue) { | |
201 | ||
202 | /* | |
203 | * There is a chance that the size of the CIL checkpoints in | |
204 | * progress at the last AIL push target calculation resulted in | |
205 | * limiting the target to the log head (l_last_sync_lsn) at the | |
206 | * time. This may not reflect where the log head is now as the | |
207 | * CIL checkpoints may have completed. | |
208 | * | |
209 | * Hence when we are woken here, it may be that the head of the | |
210 | * log that has moved rather than the tail. As the tail didn't | |
211 | * move, there still won't be space available for the | |
212 | * reservation we require. However, if the AIL has already | |
213 | * pushed to the target defined by the old log head location, we | |
214 | * will hang here waiting for something else to update the AIL | |
215 | * push target. | |
216 | * | |
217 | * Therefore, if there isn't space to wake the first waiter on | |
218 | * the grant head, we need to push the AIL again to ensure the | |
219 | * target reflects both the current log tail and log head | |
220 | * position before we wait for the tail to move again. | |
221 | */ | |
222 | ||
223 | need_bytes = xlog_ticket_reservation(log, head, tic); | |
224 | if (*free_bytes < need_bytes) { | |
225 | if (!woken_task) | |
226 | xlog_grant_push_ail(log, need_bytes); | |
227 | return false; | |
228 | } | |
229 | ||
230 | *free_bytes -= need_bytes; | |
231 | trace_xfs_log_grant_wake_up(log, tic); | |
232 | wake_up_process(tic->t_task); | |
233 | woken_task = true; | |
234 | } | |
235 | ||
236 | return true; | |
237 | } | |
238 | ||
239 | STATIC int | |
240 | xlog_grant_head_wait( | |
241 | struct xlog *log, | |
242 | struct xlog_grant_head *head, | |
243 | struct xlog_ticket *tic, | |
244 | int need_bytes) __releases(&head->lock) | |
245 | __acquires(&head->lock) | |
246 | { | |
247 | list_add_tail(&tic->t_queue, &head->waiters); | |
248 | ||
249 | do { | |
250 | if (XLOG_FORCED_SHUTDOWN(log)) | |
251 | goto shutdown; | |
252 | xlog_grant_push_ail(log, need_bytes); | |
253 | ||
254 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
255 | spin_unlock(&head->lock); | |
256 | ||
257 | XFS_STATS_INC(log->l_mp, xs_sleep_logspace); | |
258 | ||
259 | trace_xfs_log_grant_sleep(log, tic); | |
260 | schedule(); | |
261 | trace_xfs_log_grant_wake(log, tic); | |
262 | ||
263 | spin_lock(&head->lock); | |
264 | if (XLOG_FORCED_SHUTDOWN(log)) | |
265 | goto shutdown; | |
266 | } while (xlog_space_left(log, &head->grant) < need_bytes); | |
267 | ||
268 | list_del_init(&tic->t_queue); | |
269 | return 0; | |
270 | shutdown: | |
271 | list_del_init(&tic->t_queue); | |
272 | return -EIO; | |
273 | } | |
274 | ||
275 | /* | |
276 | * Atomically get the log space required for a log ticket. | |
277 | * | |
278 | * Once a ticket gets put onto head->waiters, it will only return after the | |
279 | * needed reservation is satisfied. | |
280 | * | |
281 | * This function is structured so that it has a lock free fast path. This is | |
282 | * necessary because every new transaction reservation will come through this | |
283 | * path. Hence any lock will be globally hot if we take it unconditionally on | |
284 | * every pass. | |
285 | * | |
286 | * As tickets are only ever moved on and off head->waiters under head->lock, we | |
287 | * only need to take that lock if we are going to add the ticket to the queue | |
288 | * and sleep. We can avoid taking the lock if the ticket was never added to | |
289 | * head->waiters because the t_queue list head will be empty and we hold the | |
290 | * only reference to it so it can safely be checked unlocked. | |
291 | */ | |
292 | STATIC int | |
293 | xlog_grant_head_check( | |
294 | struct xlog *log, | |
295 | struct xlog_grant_head *head, | |
296 | struct xlog_ticket *tic, | |
297 | int *need_bytes) | |
298 | { | |
299 | int free_bytes; | |
300 | int error = 0; | |
301 | ||
302 | ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY)); | |
303 | ||
304 | /* | |
305 | * If there are other waiters on the queue then give them a chance at | |
306 | * logspace before us. Wake up the first waiters, if we do not wake | |
307 | * up all the waiters then go to sleep waiting for more free space, | |
308 | * otherwise try to get some space for this transaction. | |
309 | */ | |
310 | *need_bytes = xlog_ticket_reservation(log, head, tic); | |
311 | free_bytes = xlog_space_left(log, &head->grant); | |
312 | if (!list_empty_careful(&head->waiters)) { | |
313 | spin_lock(&head->lock); | |
314 | if (!xlog_grant_head_wake(log, head, &free_bytes) || | |
315 | free_bytes < *need_bytes) { | |
316 | error = xlog_grant_head_wait(log, head, tic, | |
317 | *need_bytes); | |
318 | } | |
319 | spin_unlock(&head->lock); | |
320 | } else if (free_bytes < *need_bytes) { | |
321 | spin_lock(&head->lock); | |
322 | error = xlog_grant_head_wait(log, head, tic, *need_bytes); | |
323 | spin_unlock(&head->lock); | |
324 | } | |
325 | ||
326 | return error; | |
327 | } | |
328 | ||
329 | static void | |
330 | xlog_tic_reset_res(xlog_ticket_t *tic) | |
331 | { | |
332 | tic->t_res_num = 0; | |
333 | tic->t_res_arr_sum = 0; | |
334 | tic->t_res_num_ophdrs = 0; | |
335 | } | |
336 | ||
337 | static void | |
338 | xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type) | |
339 | { | |
340 | if (tic->t_res_num == XLOG_TIC_LEN_MAX) { | |
341 | /* add to overflow and start again */ | |
342 | tic->t_res_o_flow += tic->t_res_arr_sum; | |
343 | tic->t_res_num = 0; | |
344 | tic->t_res_arr_sum = 0; | |
345 | } | |
346 | ||
347 | tic->t_res_arr[tic->t_res_num].r_len = len; | |
348 | tic->t_res_arr[tic->t_res_num].r_type = type; | |
349 | tic->t_res_arr_sum += len; | |
350 | tic->t_res_num++; | |
351 | } | |
352 | ||
353 | bool | |
354 | xfs_log_writable( | |
355 | struct xfs_mount *mp) | |
356 | { | |
357 | /* | |
358 | * Never write to the log on norecovery mounts, if the block device is | |
359 | * read-only, or if the filesystem is shutdown. Read-only mounts still | |
360 | * allow internal writes for log recovery and unmount purposes, so don't | |
361 | * restrict that case here. | |
362 | */ | |
363 | if (mp->m_flags & XFS_MOUNT_NORECOVERY) | |
364 | return false; | |
365 | if (xfs_readonly_buftarg(mp->m_log->l_targ)) | |
366 | return false; | |
367 | if (XFS_FORCED_SHUTDOWN(mp)) | |
368 | return false; | |
369 | return true; | |
370 | } | |
371 | ||
372 | /* | |
373 | * Replenish the byte reservation required by moving the grant write head. | |
374 | */ | |
375 | int | |
376 | xfs_log_regrant( | |
377 | struct xfs_mount *mp, | |
378 | struct xlog_ticket *tic) | |
379 | { | |
380 | struct xlog *log = mp->m_log; | |
381 | int need_bytes; | |
382 | int error = 0; | |
383 | ||
384 | if (XLOG_FORCED_SHUTDOWN(log)) | |
385 | return -EIO; | |
386 | ||
387 | XFS_STATS_INC(mp, xs_try_logspace); | |
388 | ||
389 | /* | |
390 | * This is a new transaction on the ticket, so we need to change the | |
391 | * transaction ID so that the next transaction has a different TID in | |
392 | * the log. Just add one to the existing tid so that we can see chains | |
393 | * of rolling transactions in the log easily. | |
394 | */ | |
395 | tic->t_tid++; | |
396 | ||
397 | xlog_grant_push_ail(log, tic->t_unit_res); | |
398 | ||
399 | tic->t_curr_res = tic->t_unit_res; | |
400 | xlog_tic_reset_res(tic); | |
401 | ||
402 | if (tic->t_cnt > 0) | |
403 | return 0; | |
404 | ||
405 | trace_xfs_log_regrant(log, tic); | |
406 | ||
407 | error = xlog_grant_head_check(log, &log->l_write_head, tic, | |
408 | &need_bytes); | |
409 | if (error) | |
410 | goto out_error; | |
411 | ||
412 | xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes); | |
413 | trace_xfs_log_regrant_exit(log, tic); | |
414 | xlog_verify_grant_tail(log); | |
415 | return 0; | |
416 | ||
417 | out_error: | |
418 | /* | |
419 | * If we are failing, make sure the ticket doesn't have any current | |
420 | * reservations. We don't want to add this back when the ticket/ | |
421 | * transaction gets cancelled. | |
422 | */ | |
423 | tic->t_curr_res = 0; | |
424 | tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ | |
425 | return error; | |
426 | } | |
427 | ||
428 | /* | |
429 | * Reserve log space and return a ticket corresponding to the reservation. | |
430 | * | |
431 | * Each reservation is going to reserve extra space for a log record header. | |
432 | * When writes happen to the on-disk log, we don't subtract the length of the | |
433 | * log record header from any reservation. By wasting space in each | |
434 | * reservation, we prevent over allocation problems. | |
435 | */ | |
436 | int | |
437 | xfs_log_reserve( | |
438 | struct xfs_mount *mp, | |
439 | int unit_bytes, | |
440 | int cnt, | |
441 | struct xlog_ticket **ticp, | |
442 | uint8_t client, | |
443 | bool permanent) | |
444 | { | |
445 | struct xlog *log = mp->m_log; | |
446 | struct xlog_ticket *tic; | |
447 | int need_bytes; | |
448 | int error = 0; | |
449 | ||
450 | ASSERT(client == XFS_TRANSACTION || client == XFS_LOG); | |
451 | ||
452 | if (XLOG_FORCED_SHUTDOWN(log)) | |
453 | return -EIO; | |
454 | ||
455 | XFS_STATS_INC(mp, xs_try_logspace); | |
456 | ||
457 | ASSERT(*ticp == NULL); | |
458 | tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent); | |
459 | *ticp = tic; | |
460 | ||
461 | xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt | |
462 | : tic->t_unit_res); | |
463 | ||
464 | trace_xfs_log_reserve(log, tic); | |
465 | ||
466 | error = xlog_grant_head_check(log, &log->l_reserve_head, tic, | |
467 | &need_bytes); | |
468 | if (error) | |
469 | goto out_error; | |
470 | ||
471 | xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes); | |
472 | xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes); | |
473 | trace_xfs_log_reserve_exit(log, tic); | |
474 | xlog_verify_grant_tail(log); | |
475 | return 0; | |
476 | ||
477 | out_error: | |
478 | /* | |
479 | * If we are failing, make sure the ticket doesn't have any current | |
480 | * reservations. We don't want to add this back when the ticket/ | |
481 | * transaction gets cancelled. | |
482 | */ | |
483 | tic->t_curr_res = 0; | |
484 | tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ | |
485 | return error; | |
486 | } | |
487 | ||
488 | static bool | |
489 | __xlog_state_release_iclog( | |
490 | struct xlog *log, | |
491 | struct xlog_in_core *iclog) | |
492 | { | |
493 | lockdep_assert_held(&log->l_icloglock); | |
494 | ||
495 | if (iclog->ic_state == XLOG_STATE_WANT_SYNC) { | |
496 | /* update tail before writing to iclog */ | |
497 | xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp); | |
498 | ||
499 | iclog->ic_state = XLOG_STATE_SYNCING; | |
500 | iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn); | |
501 | xlog_verify_tail_lsn(log, iclog, tail_lsn); | |
502 | /* cycle incremented when incrementing curr_block */ | |
503 | return true; | |
504 | } | |
505 | ||
506 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); | |
507 | return false; | |
508 | } | |
509 | ||
510 | /* | |
511 | * Flush iclog to disk if this is the last reference to the given iclog and the | |
512 | * it is in the WANT_SYNC state. | |
513 | */ | |
514 | static int | |
515 | xlog_state_release_iclog( | |
516 | struct xlog *log, | |
517 | struct xlog_in_core *iclog) | |
518 | { | |
519 | lockdep_assert_held(&log->l_icloglock); | |
520 | ||
521 | if (iclog->ic_state == XLOG_STATE_IOERROR) | |
522 | return -EIO; | |
523 | ||
524 | if (atomic_dec_and_test(&iclog->ic_refcnt) && | |
525 | __xlog_state_release_iclog(log, iclog)) { | |
526 | spin_unlock(&log->l_icloglock); | |
527 | xlog_sync(log, iclog); | |
528 | spin_lock(&log->l_icloglock); | |
529 | } | |
530 | ||
531 | return 0; | |
532 | } | |
533 | ||
534 | void | |
535 | xfs_log_release_iclog( | |
536 | struct xlog_in_core *iclog) | |
537 | { | |
538 | struct xlog *log = iclog->ic_log; | |
539 | bool sync = false; | |
540 | ||
541 | if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) { | |
542 | if (iclog->ic_state != XLOG_STATE_IOERROR) | |
543 | sync = __xlog_state_release_iclog(log, iclog); | |
544 | spin_unlock(&log->l_icloglock); | |
545 | } | |
546 | ||
547 | if (sync) | |
548 | xlog_sync(log, iclog); | |
549 | } | |
550 | ||
551 | /* | |
552 | * Mount a log filesystem | |
553 | * | |
554 | * mp - ubiquitous xfs mount point structure | |
555 | * log_target - buftarg of on-disk log device | |
556 | * blk_offset - Start block # where block size is 512 bytes (BBSIZE) | |
557 | * num_bblocks - Number of BBSIZE blocks in on-disk log | |
558 | * | |
559 | * Return error or zero. | |
560 | */ | |
561 | int | |
562 | xfs_log_mount( | |
563 | xfs_mount_t *mp, | |
564 | xfs_buftarg_t *log_target, | |
565 | xfs_daddr_t blk_offset, | |
566 | int num_bblks) | |
567 | { | |
568 | bool fatal = xfs_sb_version_hascrc(&mp->m_sb); | |
569 | int error = 0; | |
570 | int min_logfsbs; | |
571 | ||
572 | if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) { | |
573 | xfs_notice(mp, "Mounting V%d Filesystem", | |
574 | XFS_SB_VERSION_NUM(&mp->m_sb)); | |
575 | } else { | |
576 | xfs_notice(mp, | |
577 | "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.", | |
578 | XFS_SB_VERSION_NUM(&mp->m_sb)); | |
579 | ASSERT(mp->m_flags & XFS_MOUNT_RDONLY); | |
580 | } | |
581 | ||
582 | mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks); | |
583 | if (IS_ERR(mp->m_log)) { | |
584 | error = PTR_ERR(mp->m_log); | |
585 | goto out; | |
586 | } | |
587 | ||
588 | /* | |
589 | * Validate the given log space and drop a critical message via syslog | |
590 | * if the log size is too small that would lead to some unexpected | |
591 | * situations in transaction log space reservation stage. | |
592 | * | |
593 | * Note: we can't just reject the mount if the validation fails. This | |
594 | * would mean that people would have to downgrade their kernel just to | |
595 | * remedy the situation as there is no way to grow the log (short of | |
596 | * black magic surgery with xfs_db). | |
597 | * | |
598 | * We can, however, reject mounts for CRC format filesystems, as the | |
599 | * mkfs binary being used to make the filesystem should never create a | |
600 | * filesystem with a log that is too small. | |
601 | */ | |
602 | min_logfsbs = xfs_log_calc_minimum_size(mp); | |
603 | ||
604 | if (mp->m_sb.sb_logblocks < min_logfsbs) { | |
605 | xfs_warn(mp, | |
606 | "Log size %d blocks too small, minimum size is %d blocks", | |
607 | mp->m_sb.sb_logblocks, min_logfsbs); | |
608 | error = -EINVAL; | |
609 | } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) { | |
610 | xfs_warn(mp, | |
611 | "Log size %d blocks too large, maximum size is %lld blocks", | |
612 | mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS); | |
613 | error = -EINVAL; | |
614 | } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) { | |
615 | xfs_warn(mp, | |
616 | "log size %lld bytes too large, maximum size is %lld bytes", | |
617 | XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks), | |
618 | XFS_MAX_LOG_BYTES); | |
619 | error = -EINVAL; | |
620 | } else if (mp->m_sb.sb_logsunit > 1 && | |
621 | mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) { | |
622 | xfs_warn(mp, | |
623 | "log stripe unit %u bytes must be a multiple of block size", | |
624 | mp->m_sb.sb_logsunit); | |
625 | error = -EINVAL; | |
626 | fatal = true; | |
627 | } | |
628 | if (error) { | |
629 | /* | |
630 | * Log check errors are always fatal on v5; or whenever bad | |
631 | * metadata leads to a crash. | |
632 | */ | |
633 | if (fatal) { | |
634 | xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!"); | |
635 | ASSERT(0); | |
636 | goto out_free_log; | |
637 | } | |
638 | xfs_crit(mp, "Log size out of supported range."); | |
639 | xfs_crit(mp, | |
640 | "Continuing onwards, but if log hangs are experienced then please report this message in the bug report."); | |
641 | } | |
642 | ||
643 | /* | |
644 | * Initialize the AIL now we have a log. | |
645 | */ | |
646 | error = xfs_trans_ail_init(mp); | |
647 | if (error) { | |
648 | xfs_warn(mp, "AIL initialisation failed: error %d", error); | |
649 | goto out_free_log; | |
650 | } | |
651 | mp->m_log->l_ailp = mp->m_ail; | |
652 | ||
653 | /* | |
654 | * skip log recovery on a norecovery mount. pretend it all | |
655 | * just worked. | |
656 | */ | |
657 | if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) { | |
658 | int readonly = (mp->m_flags & XFS_MOUNT_RDONLY); | |
659 | ||
660 | if (readonly) | |
661 | mp->m_flags &= ~XFS_MOUNT_RDONLY; | |
662 | ||
663 | error = xlog_recover(mp->m_log); | |
664 | ||
665 | if (readonly) | |
666 | mp->m_flags |= XFS_MOUNT_RDONLY; | |
667 | if (error) { | |
668 | xfs_warn(mp, "log mount/recovery failed: error %d", | |
669 | error); | |
670 | xlog_recover_cancel(mp->m_log); | |
671 | goto out_destroy_ail; | |
672 | } | |
673 | } | |
674 | ||
675 | error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj, | |
676 | "log"); | |
677 | if (error) | |
678 | goto out_destroy_ail; | |
679 | ||
680 | /* Normal transactions can now occur */ | |
681 | mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY; | |
682 | ||
683 | /* | |
684 | * Now the log has been fully initialised and we know were our | |
685 | * space grant counters are, we can initialise the permanent ticket | |
686 | * needed for delayed logging to work. | |
687 | */ | |
688 | xlog_cil_init_post_recovery(mp->m_log); | |
689 | ||
690 | return 0; | |
691 | ||
692 | out_destroy_ail: | |
693 | xfs_trans_ail_destroy(mp); | |
694 | out_free_log: | |
695 | xlog_dealloc_log(mp->m_log); | |
696 | out: | |
697 | return error; | |
698 | } | |
699 | ||
700 | /* | |
701 | * Finish the recovery of the file system. This is separate from the | |
702 | * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read | |
703 | * in the root and real-time bitmap inodes between calling xfs_log_mount() and | |
704 | * here. | |
705 | * | |
706 | * If we finish recovery successfully, start the background log work. If we are | |
707 | * not doing recovery, then we have a RO filesystem and we don't need to start | |
708 | * it. | |
709 | */ | |
710 | int | |
711 | xfs_log_mount_finish( | |
712 | struct xfs_mount *mp) | |
713 | { | |
714 | int error = 0; | |
715 | bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY); | |
716 | bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED; | |
717 | ||
718 | if (mp->m_flags & XFS_MOUNT_NORECOVERY) { | |
719 | ASSERT(mp->m_flags & XFS_MOUNT_RDONLY); | |
720 | return 0; | |
721 | } else if (readonly) { | |
722 | /* Allow unlinked processing to proceed */ | |
723 | mp->m_flags &= ~XFS_MOUNT_RDONLY; | |
724 | } | |
725 | ||
726 | /* | |
727 | * During the second phase of log recovery, we need iget and | |
728 | * iput to behave like they do for an active filesystem. | |
729 | * xfs_fs_drop_inode needs to be able to prevent the deletion | |
730 | * of inodes before we're done replaying log items on those | |
731 | * inodes. Turn it off immediately after recovery finishes | |
732 | * so that we don't leak the quota inodes if subsequent mount | |
733 | * activities fail. | |
734 | * | |
735 | * We let all inodes involved in redo item processing end up on | |
736 | * the LRU instead of being evicted immediately so that if we do | |
737 | * something to an unlinked inode, the irele won't cause | |
738 | * premature truncation and freeing of the inode, which results | |
739 | * in log recovery failure. We have to evict the unreferenced | |
740 | * lru inodes after clearing SB_ACTIVE because we don't | |
741 | * otherwise clean up the lru if there's a subsequent failure in | |
742 | * xfs_mountfs, which leads to us leaking the inodes if nothing | |
743 | * else (e.g. quotacheck) references the inodes before the | |
744 | * mount failure occurs. | |
745 | */ | |
746 | mp->m_super->s_flags |= SB_ACTIVE; | |
747 | error = xlog_recover_finish(mp->m_log); | |
748 | if (!error) | |
749 | xfs_log_work_queue(mp); | |
750 | mp->m_super->s_flags &= ~SB_ACTIVE; | |
751 | evict_inodes(mp->m_super); | |
752 | ||
753 | /* | |
754 | * Drain the buffer LRU after log recovery. This is required for v4 | |
755 | * filesystems to avoid leaving around buffers with NULL verifier ops, | |
756 | * but we do it unconditionally to make sure we're always in a clean | |
757 | * cache state after mount. | |
758 | * | |
759 | * Don't push in the error case because the AIL may have pending intents | |
760 | * that aren't removed until recovery is cancelled. | |
761 | */ | |
762 | if (!error && recovered) { | |
763 | xfs_log_force(mp, XFS_LOG_SYNC); | |
764 | xfs_ail_push_all_sync(mp->m_ail); | |
765 | } | |
766 | xfs_buftarg_drain(mp->m_ddev_targp); | |
767 | ||
768 | if (readonly) | |
769 | mp->m_flags |= XFS_MOUNT_RDONLY; | |
770 | ||
771 | return error; | |
772 | } | |
773 | ||
774 | /* | |
775 | * The mount has failed. Cancel the recovery if it hasn't completed and destroy | |
776 | * the log. | |
777 | */ | |
778 | void | |
779 | xfs_log_mount_cancel( | |
780 | struct xfs_mount *mp) | |
781 | { | |
782 | xlog_recover_cancel(mp->m_log); | |
783 | xfs_log_unmount(mp); | |
784 | } | |
785 | ||
786 | /* | |
787 | * Wait for the iclog to be written disk, or return an error if the log has been | |
788 | * shut down. | |
789 | */ | |
790 | static int | |
791 | xlog_wait_on_iclog( | |
792 | struct xlog_in_core *iclog) | |
793 | __releases(iclog->ic_log->l_icloglock) | |
794 | { | |
795 | struct xlog *log = iclog->ic_log; | |
796 | ||
797 | if (!XLOG_FORCED_SHUTDOWN(log) && | |
798 | iclog->ic_state != XLOG_STATE_ACTIVE && | |
799 | iclog->ic_state != XLOG_STATE_DIRTY) { | |
800 | XFS_STATS_INC(log->l_mp, xs_log_force_sleep); | |
801 | xlog_wait(&iclog->ic_force_wait, &log->l_icloglock); | |
802 | } else { | |
803 | spin_unlock(&log->l_icloglock); | |
804 | } | |
805 | ||
806 | if (XLOG_FORCED_SHUTDOWN(log)) | |
807 | return -EIO; | |
808 | return 0; | |
809 | } | |
810 | ||
811 | /* | |
812 | * Write out an unmount record using the ticket provided. We have to account for | |
813 | * the data space used in the unmount ticket as this write is not done from a | |
814 | * transaction context that has already done the accounting for us. | |
815 | */ | |
816 | static int | |
817 | xlog_write_unmount_record( | |
818 | struct xlog *log, | |
819 | struct xlog_ticket *ticket, | |
820 | xfs_lsn_t *lsn, | |
821 | uint flags) | |
822 | { | |
823 | struct xfs_unmount_log_format ulf = { | |
824 | .magic = XLOG_UNMOUNT_TYPE, | |
825 | }; | |
826 | struct xfs_log_iovec reg = { | |
827 | .i_addr = &ulf, | |
828 | .i_len = sizeof(ulf), | |
829 | .i_type = XLOG_REG_TYPE_UNMOUNT, | |
830 | }; | |
831 | struct xfs_log_vec vec = { | |
832 | .lv_niovecs = 1, | |
833 | .lv_iovecp = ®, | |
834 | }; | |
835 | ||
836 | /* account for space used by record data */ | |
837 | ticket->t_curr_res -= sizeof(ulf); | |
838 | return xlog_write(log, &vec, ticket, lsn, NULL, flags, false); | |
839 | } | |
840 | ||
841 | /* | |
842 | * Mark the filesystem clean by writing an unmount record to the head of the | |
843 | * log. | |
844 | */ | |
845 | static void | |
846 | xlog_unmount_write( | |
847 | struct xlog *log) | |
848 | { | |
849 | struct xfs_mount *mp = log->l_mp; | |
850 | struct xlog_in_core *iclog; | |
851 | struct xlog_ticket *tic = NULL; | |
852 | xfs_lsn_t lsn; | |
853 | uint flags = XLOG_UNMOUNT_TRANS; | |
854 | int error; | |
855 | ||
856 | error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0); | |
857 | if (error) | |
858 | goto out_err; | |
859 | ||
860 | error = xlog_write_unmount_record(log, tic, &lsn, flags); | |
861 | /* | |
862 | * At this point, we're umounting anyway, so there's no point in | |
863 | * transitioning log state to IOERROR. Just continue... | |
864 | */ | |
865 | out_err: | |
866 | if (error) | |
867 | xfs_alert(mp, "%s: unmount record failed", __func__); | |
868 | ||
869 | spin_lock(&log->l_icloglock); | |
870 | iclog = log->l_iclog; | |
871 | atomic_inc(&iclog->ic_refcnt); | |
872 | if (iclog->ic_state == XLOG_STATE_ACTIVE) | |
873 | xlog_state_switch_iclogs(log, iclog, 0); | |
874 | else | |
875 | ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC || | |
876 | iclog->ic_state == XLOG_STATE_IOERROR); | |
877 | error = xlog_state_release_iclog(log, iclog); | |
878 | xlog_wait_on_iclog(iclog); | |
879 | ||
880 | if (tic) { | |
881 | trace_xfs_log_umount_write(log, tic); | |
882 | xfs_log_ticket_ungrant(log, tic); | |
883 | } | |
884 | } | |
885 | ||
886 | static void | |
887 | xfs_log_unmount_verify_iclog( | |
888 | struct xlog *log) | |
889 | { | |
890 | struct xlog_in_core *iclog = log->l_iclog; | |
891 | ||
892 | do { | |
893 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); | |
894 | ASSERT(iclog->ic_offset == 0); | |
895 | } while ((iclog = iclog->ic_next) != log->l_iclog); | |
896 | } | |
897 | ||
898 | /* | |
899 | * Unmount record used to have a string "Unmount filesystem--" in the | |
900 | * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE). | |
901 | * We just write the magic number now since that particular field isn't | |
902 | * currently architecture converted and "Unmount" is a bit foo. | |
903 | * As far as I know, there weren't any dependencies on the old behaviour. | |
904 | */ | |
905 | static void | |
906 | xfs_log_unmount_write( | |
907 | struct xfs_mount *mp) | |
908 | { | |
909 | struct xlog *log = mp->m_log; | |
910 | ||
911 | if (!xfs_log_writable(mp)) | |
912 | return; | |
913 | ||
914 | xfs_log_force(mp, XFS_LOG_SYNC); | |
915 | ||
916 | if (XLOG_FORCED_SHUTDOWN(log)) | |
917 | return; | |
918 | ||
919 | /* | |
920 | * If we think the summary counters are bad, avoid writing the unmount | |
921 | * record to force log recovery at next mount, after which the summary | |
922 | * counters will be recalculated. Refer to xlog_check_unmount_rec for | |
923 | * more details. | |
924 | */ | |
925 | if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp, | |
926 | XFS_ERRTAG_FORCE_SUMMARY_RECALC)) { | |
927 | xfs_alert(mp, "%s: will fix summary counters at next mount", | |
928 | __func__); | |
929 | return; | |
930 | } | |
931 | ||
932 | xfs_log_unmount_verify_iclog(log); | |
933 | xlog_unmount_write(log); | |
934 | } | |
935 | ||
936 | /* | |
937 | * Empty the log for unmount/freeze. | |
938 | * | |
939 | * To do this, we first need to shut down the background log work so it is not | |
940 | * trying to cover the log as we clean up. We then need to unpin all objects in | |
941 | * the log so we can then flush them out. Once they have completed their IO and | |
942 | * run the callbacks removing themselves from the AIL, we can cover the log. | |
943 | */ | |
944 | int | |
945 | xfs_log_quiesce( | |
946 | struct xfs_mount *mp) | |
947 | { | |
948 | cancel_delayed_work_sync(&mp->m_log->l_work); | |
949 | xfs_log_force(mp, XFS_LOG_SYNC); | |
950 | ||
951 | /* | |
952 | * The superblock buffer is uncached and while xfs_ail_push_all_sync() | |
953 | * will push it, xfs_buftarg_wait() will not wait for it. Further, | |
954 | * xfs_buf_iowait() cannot be used because it was pushed with the | |
955 | * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for | |
956 | * the IO to complete. | |
957 | */ | |
958 | xfs_ail_push_all_sync(mp->m_ail); | |
959 | xfs_buftarg_wait(mp->m_ddev_targp); | |
960 | xfs_buf_lock(mp->m_sb_bp); | |
961 | xfs_buf_unlock(mp->m_sb_bp); | |
962 | ||
963 | return xfs_log_cover(mp); | |
964 | } | |
965 | ||
966 | void | |
967 | xfs_log_clean( | |
968 | struct xfs_mount *mp) | |
969 | { | |
970 | xfs_log_quiesce(mp); | |
971 | xfs_log_unmount_write(mp); | |
972 | } | |
973 | ||
974 | /* | |
975 | * Shut down and release the AIL and Log. | |
976 | * | |
977 | * During unmount, we need to ensure we flush all the dirty metadata objects | |
978 | * from the AIL so that the log is empty before we write the unmount record to | |
979 | * the log. Once this is done, we can tear down the AIL and the log. | |
980 | */ | |
981 | void | |
982 | xfs_log_unmount( | |
983 | struct xfs_mount *mp) | |
984 | { | |
985 | xfs_log_clean(mp); | |
986 | ||
987 | xfs_buftarg_drain(mp->m_ddev_targp); | |
988 | ||
989 | xfs_trans_ail_destroy(mp); | |
990 | ||
991 | xfs_sysfs_del(&mp->m_log->l_kobj); | |
992 | ||
993 | xlog_dealloc_log(mp->m_log); | |
994 | } | |
995 | ||
996 | void | |
997 | xfs_log_item_init( | |
998 | struct xfs_mount *mp, | |
999 | struct xfs_log_item *item, | |
1000 | int type, | |
1001 | const struct xfs_item_ops *ops) | |
1002 | { | |
1003 | item->li_mountp = mp; | |
1004 | item->li_ailp = mp->m_ail; | |
1005 | item->li_type = type; | |
1006 | item->li_ops = ops; | |
1007 | item->li_lv = NULL; | |
1008 | ||
1009 | INIT_LIST_HEAD(&item->li_ail); | |
1010 | INIT_LIST_HEAD(&item->li_cil); | |
1011 | INIT_LIST_HEAD(&item->li_bio_list); | |
1012 | INIT_LIST_HEAD(&item->li_trans); | |
1013 | } | |
1014 | ||
1015 | /* | |
1016 | * Wake up processes waiting for log space after we have moved the log tail. | |
1017 | */ | |
1018 | void | |
1019 | xfs_log_space_wake( | |
1020 | struct xfs_mount *mp) | |
1021 | { | |
1022 | struct xlog *log = mp->m_log; | |
1023 | int free_bytes; | |
1024 | ||
1025 | if (XLOG_FORCED_SHUTDOWN(log)) | |
1026 | return; | |
1027 | ||
1028 | if (!list_empty_careful(&log->l_write_head.waiters)) { | |
1029 | ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY)); | |
1030 | ||
1031 | spin_lock(&log->l_write_head.lock); | |
1032 | free_bytes = xlog_space_left(log, &log->l_write_head.grant); | |
1033 | xlog_grant_head_wake(log, &log->l_write_head, &free_bytes); | |
1034 | spin_unlock(&log->l_write_head.lock); | |
1035 | } | |
1036 | ||
1037 | if (!list_empty_careful(&log->l_reserve_head.waiters)) { | |
1038 | ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY)); | |
1039 | ||
1040 | spin_lock(&log->l_reserve_head.lock); | |
1041 | free_bytes = xlog_space_left(log, &log->l_reserve_head.grant); | |
1042 | xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes); | |
1043 | spin_unlock(&log->l_reserve_head.lock); | |
1044 | } | |
1045 | } | |
1046 | ||
1047 | /* | |
1048 | * Determine if we have a transaction that has gone to disk that needs to be | |
1049 | * covered. To begin the transition to the idle state firstly the log needs to | |
1050 | * be idle. That means the CIL, the AIL and the iclogs needs to be empty before | |
1051 | * we start attempting to cover the log. | |
1052 | * | |
1053 | * Only if we are then in a state where covering is needed, the caller is | |
1054 | * informed that dummy transactions are required to move the log into the idle | |
1055 | * state. | |
1056 | * | |
1057 | * If there are any items in the AIl or CIL, then we do not want to attempt to | |
1058 | * cover the log as we may be in a situation where there isn't log space | |
1059 | * available to run a dummy transaction and this can lead to deadlocks when the | |
1060 | * tail of the log is pinned by an item that is modified in the CIL. Hence | |
1061 | * there's no point in running a dummy transaction at this point because we | |
1062 | * can't start trying to idle the log until both the CIL and AIL are empty. | |
1063 | */ | |
1064 | static bool | |
1065 | xfs_log_need_covered( | |
1066 | struct xfs_mount *mp) | |
1067 | { | |
1068 | struct xlog *log = mp->m_log; | |
1069 | bool needed = false; | |
1070 | ||
1071 | if (!xlog_cil_empty(log)) | |
1072 | return 0; | |
1073 | ||
1074 | spin_lock(&log->l_icloglock); | |
1075 | switch (log->l_covered_state) { | |
1076 | case XLOG_STATE_COVER_DONE: | |
1077 | case XLOG_STATE_COVER_DONE2: | |
1078 | case XLOG_STATE_COVER_IDLE: | |
1079 | break; | |
1080 | case XLOG_STATE_COVER_NEED: | |
1081 | case XLOG_STATE_COVER_NEED2: | |
1082 | if (xfs_ail_min_lsn(log->l_ailp)) | |
1083 | break; | |
1084 | if (!xlog_iclogs_empty(log)) | |
1085 | break; | |
1086 | ||
1087 | needed = true; | |
1088 | if (log->l_covered_state == XLOG_STATE_COVER_NEED) | |
1089 | log->l_covered_state = XLOG_STATE_COVER_DONE; | |
1090 | else | |
1091 | log->l_covered_state = XLOG_STATE_COVER_DONE2; | |
1092 | break; | |
1093 | default: | |
1094 | needed = true; | |
1095 | break; | |
1096 | } | |
1097 | spin_unlock(&log->l_icloglock); | |
1098 | return needed; | |
1099 | } | |
1100 | ||
1101 | /* | |
1102 | * Explicitly cover the log. This is similar to background log covering but | |
1103 | * intended for usage in quiesce codepaths. The caller is responsible to ensure | |
1104 | * the log is idle and suitable for covering. The CIL, iclog buffers and AIL | |
1105 | * must all be empty. | |
1106 | */ | |
1107 | static int | |
1108 | xfs_log_cover( | |
1109 | struct xfs_mount *mp) | |
1110 | { | |
1111 | struct xlog *log = mp->m_log; | |
1112 | int error = 0; | |
1113 | ||
1114 | ASSERT((xlog_cil_empty(log) && xlog_iclogs_empty(log) && | |
1115 | !xfs_ail_min_lsn(log->l_ailp)) || | |
1116 | XFS_FORCED_SHUTDOWN(mp)); | |
1117 | ||
1118 | if (!xfs_log_writable(mp)) | |
1119 | return 0; | |
1120 | ||
1121 | /* | |
1122 | * To cover the log, commit the superblock twice (at most) in | |
1123 | * independent checkpoints. The first serves as a reference for the | |
1124 | * tail pointer. The sync transaction and AIL push empties the AIL and | |
1125 | * updates the in-core tail to the LSN of the first checkpoint. The | |
1126 | * second commit updates the on-disk tail with the in-core LSN, | |
1127 | * covering the log. Push the AIL one more time to leave it empty, as | |
1128 | * we found it. | |
1129 | */ | |
1130 | while (xfs_log_need_covered(mp)) { | |
1131 | error = xfs_sync_sb(mp, true); | |
1132 | if (error) | |
1133 | break; | |
1134 | xfs_ail_push_all_sync(mp->m_ail); | |
1135 | } | |
1136 | ||
1137 | return error; | |
1138 | } | |
1139 | ||
1140 | /* | |
1141 | * We may be holding the log iclog lock upon entering this routine. | |
1142 | */ | |
1143 | xfs_lsn_t | |
1144 | xlog_assign_tail_lsn_locked( | |
1145 | struct xfs_mount *mp) | |
1146 | { | |
1147 | struct xlog *log = mp->m_log; | |
1148 | struct xfs_log_item *lip; | |
1149 | xfs_lsn_t tail_lsn; | |
1150 | ||
1151 | assert_spin_locked(&mp->m_ail->ail_lock); | |
1152 | ||
1153 | /* | |
1154 | * To make sure we always have a valid LSN for the log tail we keep | |
1155 | * track of the last LSN which was committed in log->l_last_sync_lsn, | |
1156 | * and use that when the AIL was empty. | |
1157 | */ | |
1158 | lip = xfs_ail_min(mp->m_ail); | |
1159 | if (lip) | |
1160 | tail_lsn = lip->li_lsn; | |
1161 | else | |
1162 | tail_lsn = atomic64_read(&log->l_last_sync_lsn); | |
1163 | trace_xfs_log_assign_tail_lsn(log, tail_lsn); | |
1164 | atomic64_set(&log->l_tail_lsn, tail_lsn); | |
1165 | return tail_lsn; | |
1166 | } | |
1167 | ||
1168 | xfs_lsn_t | |
1169 | xlog_assign_tail_lsn( | |
1170 | struct xfs_mount *mp) | |
1171 | { | |
1172 | xfs_lsn_t tail_lsn; | |
1173 | ||
1174 | spin_lock(&mp->m_ail->ail_lock); | |
1175 | tail_lsn = xlog_assign_tail_lsn_locked(mp); | |
1176 | spin_unlock(&mp->m_ail->ail_lock); | |
1177 | ||
1178 | return tail_lsn; | |
1179 | } | |
1180 | ||
1181 | /* | |
1182 | * Return the space in the log between the tail and the head. The head | |
1183 | * is passed in the cycle/bytes formal parms. In the special case where | |
1184 | * the reserve head has wrapped passed the tail, this calculation is no | |
1185 | * longer valid. In this case, just return 0 which means there is no space | |
1186 | * in the log. This works for all places where this function is called | |
1187 | * with the reserve head. Of course, if the write head were to ever | |
1188 | * wrap the tail, we should blow up. Rather than catch this case here, | |
1189 | * we depend on other ASSERTions in other parts of the code. XXXmiken | |
1190 | * | |
1191 | * This code also handles the case where the reservation head is behind | |
1192 | * the tail. The details of this case are described below, but the end | |
1193 | * result is that we return the size of the log as the amount of space left. | |
1194 | */ | |
1195 | STATIC int | |
1196 | xlog_space_left( | |
1197 | struct xlog *log, | |
1198 | atomic64_t *head) | |
1199 | { | |
1200 | int free_bytes; | |
1201 | int tail_bytes; | |
1202 | int tail_cycle; | |
1203 | int head_cycle; | |
1204 | int head_bytes; | |
1205 | ||
1206 | xlog_crack_grant_head(head, &head_cycle, &head_bytes); | |
1207 | xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes); | |
1208 | tail_bytes = BBTOB(tail_bytes); | |
1209 | if (tail_cycle == head_cycle && head_bytes >= tail_bytes) | |
1210 | free_bytes = log->l_logsize - (head_bytes - tail_bytes); | |
1211 | else if (tail_cycle + 1 < head_cycle) | |
1212 | return 0; | |
1213 | else if (tail_cycle < head_cycle) { | |
1214 | ASSERT(tail_cycle == (head_cycle - 1)); | |
1215 | free_bytes = tail_bytes - head_bytes; | |
1216 | } else { | |
1217 | /* | |
1218 | * The reservation head is behind the tail. | |
1219 | * In this case we just want to return the size of the | |
1220 | * log as the amount of space left. | |
1221 | */ | |
1222 | xfs_alert(log->l_mp, "xlog_space_left: head behind tail"); | |
1223 | xfs_alert(log->l_mp, | |
1224 | " tail_cycle = %d, tail_bytes = %d", | |
1225 | tail_cycle, tail_bytes); | |
1226 | xfs_alert(log->l_mp, | |
1227 | " GH cycle = %d, GH bytes = %d", | |
1228 | head_cycle, head_bytes); | |
1229 | ASSERT(0); | |
1230 | free_bytes = log->l_logsize; | |
1231 | } | |
1232 | return free_bytes; | |
1233 | } | |
1234 | ||
1235 | ||
1236 | static void | |
1237 | xlog_ioend_work( | |
1238 | struct work_struct *work) | |
1239 | { | |
1240 | struct xlog_in_core *iclog = | |
1241 | container_of(work, struct xlog_in_core, ic_end_io_work); | |
1242 | struct xlog *log = iclog->ic_log; | |
1243 | int error; | |
1244 | ||
1245 | error = blk_status_to_errno(iclog->ic_bio.bi_status); | |
1246 | #ifdef DEBUG | |
1247 | /* treat writes with injected CRC errors as failed */ | |
1248 | if (iclog->ic_fail_crc) | |
1249 | error = -EIO; | |
1250 | #endif | |
1251 | ||
1252 | /* | |
1253 | * Race to shutdown the filesystem if we see an error. | |
1254 | */ | |
1255 | if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) { | |
1256 | xfs_alert(log->l_mp, "log I/O error %d", error); | |
1257 | xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR); | |
1258 | } | |
1259 | ||
1260 | xlog_state_done_syncing(iclog); | |
1261 | bio_uninit(&iclog->ic_bio); | |
1262 | ||
1263 | /* | |
1264 | * Drop the lock to signal that we are done. Nothing references the | |
1265 | * iclog after this, so an unmount waiting on this lock can now tear it | |
1266 | * down safely. As such, it is unsafe to reference the iclog after the | |
1267 | * unlock as we could race with it being freed. | |
1268 | */ | |
1269 | up(&iclog->ic_sema); | |
1270 | } | |
1271 | ||
1272 | /* | |
1273 | * Return size of each in-core log record buffer. | |
1274 | * | |
1275 | * All machines get 8 x 32kB buffers by default, unless tuned otherwise. | |
1276 | * | |
1277 | * If the filesystem blocksize is too large, we may need to choose a | |
1278 | * larger size since the directory code currently logs entire blocks. | |
1279 | */ | |
1280 | STATIC void | |
1281 | xlog_get_iclog_buffer_size( | |
1282 | struct xfs_mount *mp, | |
1283 | struct xlog *log) | |
1284 | { | |
1285 | if (mp->m_logbufs <= 0) | |
1286 | mp->m_logbufs = XLOG_MAX_ICLOGS; | |
1287 | if (mp->m_logbsize <= 0) | |
1288 | mp->m_logbsize = XLOG_BIG_RECORD_BSIZE; | |
1289 | ||
1290 | log->l_iclog_bufs = mp->m_logbufs; | |
1291 | log->l_iclog_size = mp->m_logbsize; | |
1292 | ||
1293 | /* | |
1294 | * # headers = size / 32k - one header holds cycles from 32k of data. | |
1295 | */ | |
1296 | log->l_iclog_heads = | |
1297 | DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE); | |
1298 | log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT; | |
1299 | } | |
1300 | ||
1301 | void | |
1302 | xfs_log_work_queue( | |
1303 | struct xfs_mount *mp) | |
1304 | { | |
1305 | queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work, | |
1306 | msecs_to_jiffies(xfs_syncd_centisecs * 10)); | |
1307 | } | |
1308 | ||
1309 | /* | |
1310 | * Every sync period we need to unpin all items in the AIL and push them to | |
1311 | * disk. If there is nothing dirty, then we might need to cover the log to | |
1312 | * indicate that the filesystem is idle. | |
1313 | */ | |
1314 | static void | |
1315 | xfs_log_worker( | |
1316 | struct work_struct *work) | |
1317 | { | |
1318 | struct xlog *log = container_of(to_delayed_work(work), | |
1319 | struct xlog, l_work); | |
1320 | struct xfs_mount *mp = log->l_mp; | |
1321 | ||
1322 | /* dgc: errors ignored - not fatal and nowhere to report them */ | |
1323 | if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) { | |
1324 | /* | |
1325 | * Dump a transaction into the log that contains no real change. | |
1326 | * This is needed to stamp the current tail LSN into the log | |
1327 | * during the covering operation. | |
1328 | * | |
1329 | * We cannot use an inode here for this - that will push dirty | |
1330 | * state back up into the VFS and then periodic inode flushing | |
1331 | * will prevent log covering from making progress. Hence we | |
1332 | * synchronously log the superblock instead to ensure the | |
1333 | * superblock is immediately unpinned and can be written back. | |
1334 | */ | |
1335 | xfs_sync_sb(mp, true); | |
1336 | } else | |
1337 | xfs_log_force(mp, 0); | |
1338 | ||
1339 | /* start pushing all the metadata that is currently dirty */ | |
1340 | xfs_ail_push_all(mp->m_ail); | |
1341 | ||
1342 | /* queue us up again */ | |
1343 | xfs_log_work_queue(mp); | |
1344 | } | |
1345 | ||
1346 | /* | |
1347 | * This routine initializes some of the log structure for a given mount point. | |
1348 | * Its primary purpose is to fill in enough, so recovery can occur. However, | |
1349 | * some other stuff may be filled in too. | |
1350 | */ | |
1351 | STATIC struct xlog * | |
1352 | xlog_alloc_log( | |
1353 | struct xfs_mount *mp, | |
1354 | struct xfs_buftarg *log_target, | |
1355 | xfs_daddr_t blk_offset, | |
1356 | int num_bblks) | |
1357 | { | |
1358 | struct xlog *log; | |
1359 | xlog_rec_header_t *head; | |
1360 | xlog_in_core_t **iclogp; | |
1361 | xlog_in_core_t *iclog, *prev_iclog=NULL; | |
1362 | int i; | |
1363 | int error = -ENOMEM; | |
1364 | uint log2_size = 0; | |
1365 | ||
1366 | log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL); | |
1367 | if (!log) { | |
1368 | xfs_warn(mp, "Log allocation failed: No memory!"); | |
1369 | goto out; | |
1370 | } | |
1371 | ||
1372 | log->l_mp = mp; | |
1373 | log->l_targ = log_target; | |
1374 | log->l_logsize = BBTOB(num_bblks); | |
1375 | log->l_logBBstart = blk_offset; | |
1376 | log->l_logBBsize = num_bblks; | |
1377 | log->l_covered_state = XLOG_STATE_COVER_IDLE; | |
1378 | log->l_flags |= XLOG_ACTIVE_RECOVERY; | |
1379 | INIT_DELAYED_WORK(&log->l_work, xfs_log_worker); | |
1380 | ||
1381 | log->l_prev_block = -1; | |
1382 | /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */ | |
1383 | xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0); | |
1384 | xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0); | |
1385 | log->l_curr_cycle = 1; /* 0 is bad since this is initial value */ | |
1386 | ||
1387 | xlog_grant_head_init(&log->l_reserve_head); | |
1388 | xlog_grant_head_init(&log->l_write_head); | |
1389 | ||
1390 | error = -EFSCORRUPTED; | |
1391 | if (xfs_sb_version_hassector(&mp->m_sb)) { | |
1392 | log2_size = mp->m_sb.sb_logsectlog; | |
1393 | if (log2_size < BBSHIFT) { | |
1394 | xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)", | |
1395 | log2_size, BBSHIFT); | |
1396 | goto out_free_log; | |
1397 | } | |
1398 | ||
1399 | log2_size -= BBSHIFT; | |
1400 | if (log2_size > mp->m_sectbb_log) { | |
1401 | xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)", | |
1402 | log2_size, mp->m_sectbb_log); | |
1403 | goto out_free_log; | |
1404 | } | |
1405 | ||
1406 | /* for larger sector sizes, must have v2 or external log */ | |
1407 | if (log2_size && log->l_logBBstart > 0 && | |
1408 | !xfs_sb_version_haslogv2(&mp->m_sb)) { | |
1409 | xfs_warn(mp, | |
1410 | "log sector size (0x%x) invalid for configuration.", | |
1411 | log2_size); | |
1412 | goto out_free_log; | |
1413 | } | |
1414 | } | |
1415 | log->l_sectBBsize = 1 << log2_size; | |
1416 | ||
1417 | xlog_get_iclog_buffer_size(mp, log); | |
1418 | ||
1419 | spin_lock_init(&log->l_icloglock); | |
1420 | init_waitqueue_head(&log->l_flush_wait); | |
1421 | ||
1422 | iclogp = &log->l_iclog; | |
1423 | /* | |
1424 | * The amount of memory to allocate for the iclog structure is | |
1425 | * rather funky due to the way the structure is defined. It is | |
1426 | * done this way so that we can use different sizes for machines | |
1427 | * with different amounts of memory. See the definition of | |
1428 | * xlog_in_core_t in xfs_log_priv.h for details. | |
1429 | */ | |
1430 | ASSERT(log->l_iclog_size >= 4096); | |
1431 | for (i = 0; i < log->l_iclog_bufs; i++) { | |
1432 | int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp); | |
1433 | size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) * | |
1434 | sizeof(struct bio_vec); | |
1435 | ||
1436 | iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL); | |
1437 | if (!iclog) | |
1438 | goto out_free_iclog; | |
1439 | ||
1440 | *iclogp = iclog; | |
1441 | iclog->ic_prev = prev_iclog; | |
1442 | prev_iclog = iclog; | |
1443 | ||
1444 | iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask, | |
1445 | KM_MAYFAIL | KM_ZERO); | |
1446 | if (!iclog->ic_data) | |
1447 | goto out_free_iclog; | |
1448 | #ifdef DEBUG | |
1449 | log->l_iclog_bak[i] = &iclog->ic_header; | |
1450 | #endif | |
1451 | head = &iclog->ic_header; | |
1452 | memset(head, 0, sizeof(xlog_rec_header_t)); | |
1453 | head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM); | |
1454 | head->h_version = cpu_to_be32( | |
1455 | xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1); | |
1456 | head->h_size = cpu_to_be32(log->l_iclog_size); | |
1457 | /* new fields */ | |
1458 | head->h_fmt = cpu_to_be32(XLOG_FMT); | |
1459 | memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t)); | |
1460 | ||
1461 | iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize; | |
1462 | iclog->ic_state = XLOG_STATE_ACTIVE; | |
1463 | iclog->ic_log = log; | |
1464 | atomic_set(&iclog->ic_refcnt, 0); | |
1465 | spin_lock_init(&iclog->ic_callback_lock); | |
1466 | INIT_LIST_HEAD(&iclog->ic_callbacks); | |
1467 | iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize; | |
1468 | ||
1469 | init_waitqueue_head(&iclog->ic_force_wait); | |
1470 | init_waitqueue_head(&iclog->ic_write_wait); | |
1471 | INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work); | |
1472 | sema_init(&iclog->ic_sema, 1); | |
1473 | ||
1474 | iclogp = &iclog->ic_next; | |
1475 | } | |
1476 | *iclogp = log->l_iclog; /* complete ring */ | |
1477 | log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */ | |
1478 | ||
1479 | log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s", | |
1480 | WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0, | |
1481 | mp->m_super->s_id); | |
1482 | if (!log->l_ioend_workqueue) | |
1483 | goto out_free_iclog; | |
1484 | ||
1485 | error = xlog_cil_init(log); | |
1486 | if (error) | |
1487 | goto out_destroy_workqueue; | |
1488 | return log; | |
1489 | ||
1490 | out_destroy_workqueue: | |
1491 | destroy_workqueue(log->l_ioend_workqueue); | |
1492 | out_free_iclog: | |
1493 | for (iclog = log->l_iclog; iclog; iclog = prev_iclog) { | |
1494 | prev_iclog = iclog->ic_next; | |
1495 | kmem_free(iclog->ic_data); | |
1496 | kmem_free(iclog); | |
1497 | if (prev_iclog == log->l_iclog) | |
1498 | break; | |
1499 | } | |
1500 | out_free_log: | |
1501 | kmem_free(log); | |
1502 | out: | |
1503 | return ERR_PTR(error); | |
1504 | } /* xlog_alloc_log */ | |
1505 | ||
1506 | /* | |
1507 | * Write out the commit record of a transaction associated with the given | |
1508 | * ticket to close off a running log write. Return the lsn of the commit record. | |
1509 | */ | |
1510 | int | |
1511 | xlog_commit_record( | |
1512 | struct xlog *log, | |
1513 | struct xlog_ticket *ticket, | |
1514 | struct xlog_in_core **iclog, | |
1515 | xfs_lsn_t *lsn) | |
1516 | { | |
1517 | struct xfs_log_iovec reg = { | |
1518 | .i_addr = NULL, | |
1519 | .i_len = 0, | |
1520 | .i_type = XLOG_REG_TYPE_COMMIT, | |
1521 | }; | |
1522 | struct xfs_log_vec vec = { | |
1523 | .lv_niovecs = 1, | |
1524 | .lv_iovecp = ®, | |
1525 | }; | |
1526 | int error; | |
1527 | ||
1528 | if (XLOG_FORCED_SHUTDOWN(log)) | |
1529 | return -EIO; | |
1530 | ||
1531 | error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS, | |
1532 | false); | |
1533 | if (error) | |
1534 | xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR); | |
1535 | return error; | |
1536 | } | |
1537 | ||
1538 | /* | |
1539 | * Compute the LSN that we'd need to push the log tail towards in order to have | |
1540 | * (a) enough on-disk log space to log the number of bytes specified, (b) at | |
1541 | * least 25% of the log space free, and (c) at least 256 blocks free. If the | |
1542 | * log free space already meets all three thresholds, this function returns | |
1543 | * NULLCOMMITLSN. | |
1544 | */ | |
1545 | xfs_lsn_t | |
1546 | xlog_grant_push_threshold( | |
1547 | struct xlog *log, | |
1548 | int need_bytes) | |
1549 | { | |
1550 | xfs_lsn_t threshold_lsn = 0; | |
1551 | xfs_lsn_t last_sync_lsn; | |
1552 | int free_blocks; | |
1553 | int free_bytes; | |
1554 | int threshold_block; | |
1555 | int threshold_cycle; | |
1556 | int free_threshold; | |
1557 | ||
1558 | ASSERT(BTOBB(need_bytes) < log->l_logBBsize); | |
1559 | ||
1560 | free_bytes = xlog_space_left(log, &log->l_reserve_head.grant); | |
1561 | free_blocks = BTOBBT(free_bytes); | |
1562 | ||
1563 | /* | |
1564 | * Set the threshold for the minimum number of free blocks in the | |
1565 | * log to the maximum of what the caller needs, one quarter of the | |
1566 | * log, and 256 blocks. | |
1567 | */ | |
1568 | free_threshold = BTOBB(need_bytes); | |
1569 | free_threshold = max(free_threshold, (log->l_logBBsize >> 2)); | |
1570 | free_threshold = max(free_threshold, 256); | |
1571 | if (free_blocks >= free_threshold) | |
1572 | return NULLCOMMITLSN; | |
1573 | ||
1574 | xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle, | |
1575 | &threshold_block); | |
1576 | threshold_block += free_threshold; | |
1577 | if (threshold_block >= log->l_logBBsize) { | |
1578 | threshold_block -= log->l_logBBsize; | |
1579 | threshold_cycle += 1; | |
1580 | } | |
1581 | threshold_lsn = xlog_assign_lsn(threshold_cycle, | |
1582 | threshold_block); | |
1583 | /* | |
1584 | * Don't pass in an lsn greater than the lsn of the last | |
1585 | * log record known to be on disk. Use a snapshot of the last sync lsn | |
1586 | * so that it doesn't change between the compare and the set. | |
1587 | */ | |
1588 | last_sync_lsn = atomic64_read(&log->l_last_sync_lsn); | |
1589 | if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0) | |
1590 | threshold_lsn = last_sync_lsn; | |
1591 | ||
1592 | return threshold_lsn; | |
1593 | } | |
1594 | ||
1595 | /* | |
1596 | * Push the tail of the log if we need to do so to maintain the free log space | |
1597 | * thresholds set out by xlog_grant_push_threshold. We may need to adopt a | |
1598 | * policy which pushes on an lsn which is further along in the log once we | |
1599 | * reach the high water mark. In this manner, we would be creating a low water | |
1600 | * mark. | |
1601 | */ | |
1602 | STATIC void | |
1603 | xlog_grant_push_ail( | |
1604 | struct xlog *log, | |
1605 | int need_bytes) | |
1606 | { | |
1607 | xfs_lsn_t threshold_lsn; | |
1608 | ||
1609 | threshold_lsn = xlog_grant_push_threshold(log, need_bytes); | |
1610 | if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log)) | |
1611 | return; | |
1612 | ||
1613 | /* | |
1614 | * Get the transaction layer to kick the dirty buffers out to | |
1615 | * disk asynchronously. No point in trying to do this if | |
1616 | * the filesystem is shutting down. | |
1617 | */ | |
1618 | xfs_ail_push(log->l_ailp, threshold_lsn); | |
1619 | } | |
1620 | ||
1621 | /* | |
1622 | * Stamp cycle number in every block | |
1623 | */ | |
1624 | STATIC void | |
1625 | xlog_pack_data( | |
1626 | struct xlog *log, | |
1627 | struct xlog_in_core *iclog, | |
1628 | int roundoff) | |
1629 | { | |
1630 | int i, j, k; | |
1631 | int size = iclog->ic_offset + roundoff; | |
1632 | __be32 cycle_lsn; | |
1633 | char *dp; | |
1634 | ||
1635 | cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn); | |
1636 | ||
1637 | dp = iclog->ic_datap; | |
1638 | for (i = 0; i < BTOBB(size); i++) { | |
1639 | if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) | |
1640 | break; | |
1641 | iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp; | |
1642 | *(__be32 *)dp = cycle_lsn; | |
1643 | dp += BBSIZE; | |
1644 | } | |
1645 | ||
1646 | if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) { | |
1647 | xlog_in_core_2_t *xhdr = iclog->ic_data; | |
1648 | ||
1649 | for ( ; i < BTOBB(size); i++) { | |
1650 | j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
1651 | k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
1652 | xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp; | |
1653 | *(__be32 *)dp = cycle_lsn; | |
1654 | dp += BBSIZE; | |
1655 | } | |
1656 | ||
1657 | for (i = 1; i < log->l_iclog_heads; i++) | |
1658 | xhdr[i].hic_xheader.xh_cycle = cycle_lsn; | |
1659 | } | |
1660 | } | |
1661 | ||
1662 | /* | |
1663 | * Calculate the checksum for a log buffer. | |
1664 | * | |
1665 | * This is a little more complicated than it should be because the various | |
1666 | * headers and the actual data are non-contiguous. | |
1667 | */ | |
1668 | __le32 | |
1669 | xlog_cksum( | |
1670 | struct xlog *log, | |
1671 | struct xlog_rec_header *rhead, | |
1672 | char *dp, | |
1673 | int size) | |
1674 | { | |
1675 | uint32_t crc; | |
1676 | ||
1677 | /* first generate the crc for the record header ... */ | |
1678 | crc = xfs_start_cksum_update((char *)rhead, | |
1679 | sizeof(struct xlog_rec_header), | |
1680 | offsetof(struct xlog_rec_header, h_crc)); | |
1681 | ||
1682 | /* ... then for additional cycle data for v2 logs ... */ | |
1683 | if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) { | |
1684 | union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead; | |
1685 | int i; | |
1686 | int xheads; | |
1687 | ||
1688 | xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE); | |
1689 | ||
1690 | for (i = 1; i < xheads; i++) { | |
1691 | crc = crc32c(crc, &xhdr[i].hic_xheader, | |
1692 | sizeof(struct xlog_rec_ext_header)); | |
1693 | } | |
1694 | } | |
1695 | ||
1696 | /* ... and finally for the payload */ | |
1697 | crc = crc32c(crc, dp, size); | |
1698 | ||
1699 | return xfs_end_cksum(crc); | |
1700 | } | |
1701 | ||
1702 | static void | |
1703 | xlog_bio_end_io( | |
1704 | struct bio *bio) | |
1705 | { | |
1706 | struct xlog_in_core *iclog = bio->bi_private; | |
1707 | ||
1708 | queue_work(iclog->ic_log->l_ioend_workqueue, | |
1709 | &iclog->ic_end_io_work); | |
1710 | } | |
1711 | ||
1712 | static int | |
1713 | xlog_map_iclog_data( | |
1714 | struct bio *bio, | |
1715 | void *data, | |
1716 | size_t count) | |
1717 | { | |
1718 | do { | |
1719 | struct page *page = kmem_to_page(data); | |
1720 | unsigned int off = offset_in_page(data); | |
1721 | size_t len = min_t(size_t, count, PAGE_SIZE - off); | |
1722 | ||
1723 | if (bio_add_page(bio, page, len, off) != len) | |
1724 | return -EIO; | |
1725 | ||
1726 | data += len; | |
1727 | count -= len; | |
1728 | } while (count); | |
1729 | ||
1730 | return 0; | |
1731 | } | |
1732 | ||
1733 | STATIC void | |
1734 | xlog_write_iclog( | |
1735 | struct xlog *log, | |
1736 | struct xlog_in_core *iclog, | |
1737 | uint64_t bno, | |
1738 | unsigned int count, | |
1739 | bool need_flush) | |
1740 | { | |
1741 | ASSERT(bno < log->l_logBBsize); | |
1742 | ||
1743 | /* | |
1744 | * We lock the iclogbufs here so that we can serialise against I/O | |
1745 | * completion during unmount. We might be processing a shutdown | |
1746 | * triggered during unmount, and that can occur asynchronously to the | |
1747 | * unmount thread, and hence we need to ensure that completes before | |
1748 | * tearing down the iclogbufs. Hence we need to hold the buffer lock | |
1749 | * across the log IO to archieve that. | |
1750 | */ | |
1751 | down(&iclog->ic_sema); | |
1752 | if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) { | |
1753 | /* | |
1754 | * It would seem logical to return EIO here, but we rely on | |
1755 | * the log state machine to propagate I/O errors instead of | |
1756 | * doing it here. We kick of the state machine and unlock | |
1757 | * the buffer manually, the code needs to be kept in sync | |
1758 | * with the I/O completion path. | |
1759 | */ | |
1760 | xlog_state_done_syncing(iclog); | |
1761 | up(&iclog->ic_sema); | |
1762 | return; | |
1763 | } | |
1764 | ||
1765 | bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE)); | |
1766 | bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev); | |
1767 | iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno; | |
1768 | iclog->ic_bio.bi_end_io = xlog_bio_end_io; | |
1769 | iclog->ic_bio.bi_private = iclog; | |
1770 | ||
1771 | /* | |
1772 | * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more | |
1773 | * IOs coming immediately after this one. This prevents the block layer | |
1774 | * writeback throttle from throttling log writes behind background | |
1775 | * metadata writeback and causing priority inversions. | |
1776 | */ | |
1777 | iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | | |
1778 | REQ_IDLE | REQ_FUA; | |
1779 | if (need_flush) | |
1780 | iclog->ic_bio.bi_opf |= REQ_PREFLUSH; | |
1781 | ||
1782 | if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) { | |
1783 | xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR); | |
1784 | return; | |
1785 | } | |
1786 | if (is_vmalloc_addr(iclog->ic_data)) | |
1787 | flush_kernel_vmap_range(iclog->ic_data, count); | |
1788 | ||
1789 | /* | |
1790 | * If this log buffer would straddle the end of the log we will have | |
1791 | * to split it up into two bios, so that we can continue at the start. | |
1792 | */ | |
1793 | if (bno + BTOBB(count) > log->l_logBBsize) { | |
1794 | struct bio *split; | |
1795 | ||
1796 | split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno, | |
1797 | GFP_NOIO, &fs_bio_set); | |
1798 | bio_chain(split, &iclog->ic_bio); | |
1799 | submit_bio(split); | |
1800 | ||
1801 | /* restart at logical offset zero for the remainder */ | |
1802 | iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart; | |
1803 | } | |
1804 | ||
1805 | submit_bio(&iclog->ic_bio); | |
1806 | } | |
1807 | ||
1808 | /* | |
1809 | * We need to bump cycle number for the part of the iclog that is | |
1810 | * written to the start of the log. Watch out for the header magic | |
1811 | * number case, though. | |
1812 | */ | |
1813 | static void | |
1814 | xlog_split_iclog( | |
1815 | struct xlog *log, | |
1816 | void *data, | |
1817 | uint64_t bno, | |
1818 | unsigned int count) | |
1819 | { | |
1820 | unsigned int split_offset = BBTOB(log->l_logBBsize - bno); | |
1821 | unsigned int i; | |
1822 | ||
1823 | for (i = split_offset; i < count; i += BBSIZE) { | |
1824 | uint32_t cycle = get_unaligned_be32(data + i); | |
1825 | ||
1826 | if (++cycle == XLOG_HEADER_MAGIC_NUM) | |
1827 | cycle++; | |
1828 | put_unaligned_be32(cycle, data + i); | |
1829 | } | |
1830 | } | |
1831 | ||
1832 | static int | |
1833 | xlog_calc_iclog_size( | |
1834 | struct xlog *log, | |
1835 | struct xlog_in_core *iclog, | |
1836 | uint32_t *roundoff) | |
1837 | { | |
1838 | uint32_t count_init, count; | |
1839 | bool use_lsunit; | |
1840 | ||
1841 | use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) && | |
1842 | log->l_mp->m_sb.sb_logsunit > 1; | |
1843 | ||
1844 | /* Add for LR header */ | |
1845 | count_init = log->l_iclog_hsize + iclog->ic_offset; | |
1846 | ||
1847 | /* Round out the log write size */ | |
1848 | if (use_lsunit) { | |
1849 | /* we have a v2 stripe unit to use */ | |
1850 | count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init)); | |
1851 | } else { | |
1852 | count = BBTOB(BTOBB(count_init)); | |
1853 | } | |
1854 | ||
1855 | ASSERT(count >= count_init); | |
1856 | *roundoff = count - count_init; | |
1857 | ||
1858 | if (use_lsunit) | |
1859 | ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit); | |
1860 | else | |
1861 | ASSERT(*roundoff < BBTOB(1)); | |
1862 | return count; | |
1863 | } | |
1864 | ||
1865 | /* | |
1866 | * Flush out the in-core log (iclog) to the on-disk log in an asynchronous | |
1867 | * fashion. Previously, we should have moved the current iclog | |
1868 | * ptr in the log to point to the next available iclog. This allows further | |
1869 | * write to continue while this code syncs out an iclog ready to go. | |
1870 | * Before an in-core log can be written out, the data section must be scanned | |
1871 | * to save away the 1st word of each BBSIZE block into the header. We replace | |
1872 | * it with the current cycle count. Each BBSIZE block is tagged with the | |
1873 | * cycle count because there in an implicit assumption that drives will | |
1874 | * guarantee that entire 512 byte blocks get written at once. In other words, | |
1875 | * we can't have part of a 512 byte block written and part not written. By | |
1876 | * tagging each block, we will know which blocks are valid when recovering | |
1877 | * after an unclean shutdown. | |
1878 | * | |
1879 | * This routine is single threaded on the iclog. No other thread can be in | |
1880 | * this routine with the same iclog. Changing contents of iclog can there- | |
1881 | * fore be done without grabbing the state machine lock. Updating the global | |
1882 | * log will require grabbing the lock though. | |
1883 | * | |
1884 | * The entire log manager uses a logical block numbering scheme. Only | |
1885 | * xlog_write_iclog knows about the fact that the log may not start with | |
1886 | * block zero on a given device. | |
1887 | */ | |
1888 | STATIC void | |
1889 | xlog_sync( | |
1890 | struct xlog *log, | |
1891 | struct xlog_in_core *iclog) | |
1892 | { | |
1893 | unsigned int count; /* byte count of bwrite */ | |
1894 | unsigned int roundoff; /* roundoff to BB or stripe */ | |
1895 | uint64_t bno; | |
1896 | unsigned int size; | |
1897 | bool need_flush = true, split = false; | |
1898 | ||
1899 | ASSERT(atomic_read(&iclog->ic_refcnt) == 0); | |
1900 | ||
1901 | count = xlog_calc_iclog_size(log, iclog, &roundoff); | |
1902 | ||
1903 | /* move grant heads by roundoff in sync */ | |
1904 | xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff); | |
1905 | xlog_grant_add_space(log, &log->l_write_head.grant, roundoff); | |
1906 | ||
1907 | /* put cycle number in every block */ | |
1908 | xlog_pack_data(log, iclog, roundoff); | |
1909 | ||
1910 | /* real byte length */ | |
1911 | size = iclog->ic_offset; | |
1912 | if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) | |
1913 | size += roundoff; | |
1914 | iclog->ic_header.h_len = cpu_to_be32(size); | |
1915 | ||
1916 | XFS_STATS_INC(log->l_mp, xs_log_writes); | |
1917 | XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count)); | |
1918 | ||
1919 | bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)); | |
1920 | ||
1921 | /* Do we need to split this write into 2 parts? */ | |
1922 | if (bno + BTOBB(count) > log->l_logBBsize) { | |
1923 | xlog_split_iclog(log, &iclog->ic_header, bno, count); | |
1924 | split = true; | |
1925 | } | |
1926 | ||
1927 | /* calculcate the checksum */ | |
1928 | iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header, | |
1929 | iclog->ic_datap, size); | |
1930 | /* | |
1931 | * Intentionally corrupt the log record CRC based on the error injection | |
1932 | * frequency, if defined. This facilitates testing log recovery in the | |
1933 | * event of torn writes. Hence, set the IOABORT state to abort the log | |
1934 | * write on I/O completion and shutdown the fs. The subsequent mount | |
1935 | * detects the bad CRC and attempts to recover. | |
1936 | */ | |
1937 | #ifdef DEBUG | |
1938 | if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) { | |
1939 | iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA); | |
1940 | iclog->ic_fail_crc = true; | |
1941 | xfs_warn(log->l_mp, | |
1942 | "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.", | |
1943 | be64_to_cpu(iclog->ic_header.h_lsn)); | |
1944 | } | |
1945 | #endif | |
1946 | ||
1947 | /* | |
1948 | * Flush the data device before flushing the log to make sure all meta | |
1949 | * data written back from the AIL actually made it to disk before | |
1950 | * stamping the new log tail LSN into the log buffer. For an external | |
1951 | * log we need to issue the flush explicitly, and unfortunately | |
1952 | * synchronously here; for an internal log we can simply use the block | |
1953 | * layer state machine for preflushes. | |
1954 | */ | |
1955 | if (log->l_targ != log->l_mp->m_ddev_targp || split) { | |
1956 | xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp); | |
1957 | need_flush = false; | |
1958 | } | |
1959 | ||
1960 | xlog_verify_iclog(log, iclog, count); | |
1961 | xlog_write_iclog(log, iclog, bno, count, need_flush); | |
1962 | } | |
1963 | ||
1964 | /* | |
1965 | * Deallocate a log structure | |
1966 | */ | |
1967 | STATIC void | |
1968 | xlog_dealloc_log( | |
1969 | struct xlog *log) | |
1970 | { | |
1971 | xlog_in_core_t *iclog, *next_iclog; | |
1972 | int i; | |
1973 | ||
1974 | xlog_cil_destroy(log); | |
1975 | ||
1976 | /* | |
1977 | * Cycle all the iclogbuf locks to make sure all log IO completion | |
1978 | * is done before we tear down these buffers. | |
1979 | */ | |
1980 | iclog = log->l_iclog; | |
1981 | for (i = 0; i < log->l_iclog_bufs; i++) { | |
1982 | down(&iclog->ic_sema); | |
1983 | up(&iclog->ic_sema); | |
1984 | iclog = iclog->ic_next; | |
1985 | } | |
1986 | ||
1987 | iclog = log->l_iclog; | |
1988 | for (i = 0; i < log->l_iclog_bufs; i++) { | |
1989 | next_iclog = iclog->ic_next; | |
1990 | kmem_free(iclog->ic_data); | |
1991 | kmem_free(iclog); | |
1992 | iclog = next_iclog; | |
1993 | } | |
1994 | ||
1995 | log->l_mp->m_log = NULL; | |
1996 | destroy_workqueue(log->l_ioend_workqueue); | |
1997 | kmem_free(log); | |
1998 | } | |
1999 | ||
2000 | /* | |
2001 | * Update counters atomically now that memcpy is done. | |
2002 | */ | |
2003 | static inline void | |
2004 | xlog_state_finish_copy( | |
2005 | struct xlog *log, | |
2006 | struct xlog_in_core *iclog, | |
2007 | int record_cnt, | |
2008 | int copy_bytes) | |
2009 | { | |
2010 | lockdep_assert_held(&log->l_icloglock); | |
2011 | ||
2012 | be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt); | |
2013 | iclog->ic_offset += copy_bytes; | |
2014 | } | |
2015 | ||
2016 | /* | |
2017 | * print out info relating to regions written which consume | |
2018 | * the reservation | |
2019 | */ | |
2020 | void | |
2021 | xlog_print_tic_res( | |
2022 | struct xfs_mount *mp, | |
2023 | struct xlog_ticket *ticket) | |
2024 | { | |
2025 | uint i; | |
2026 | uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t); | |
2027 | ||
2028 | /* match with XLOG_REG_TYPE_* in xfs_log.h */ | |
2029 | #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str | |
2030 | static char *res_type_str[] = { | |
2031 | REG_TYPE_STR(BFORMAT, "bformat"), | |
2032 | REG_TYPE_STR(BCHUNK, "bchunk"), | |
2033 | REG_TYPE_STR(EFI_FORMAT, "efi_format"), | |
2034 | REG_TYPE_STR(EFD_FORMAT, "efd_format"), | |
2035 | REG_TYPE_STR(IFORMAT, "iformat"), | |
2036 | REG_TYPE_STR(ICORE, "icore"), | |
2037 | REG_TYPE_STR(IEXT, "iext"), | |
2038 | REG_TYPE_STR(IBROOT, "ibroot"), | |
2039 | REG_TYPE_STR(ILOCAL, "ilocal"), | |
2040 | REG_TYPE_STR(IATTR_EXT, "iattr_ext"), | |
2041 | REG_TYPE_STR(IATTR_BROOT, "iattr_broot"), | |
2042 | REG_TYPE_STR(IATTR_LOCAL, "iattr_local"), | |
2043 | REG_TYPE_STR(QFORMAT, "qformat"), | |
2044 | REG_TYPE_STR(DQUOT, "dquot"), | |
2045 | REG_TYPE_STR(QUOTAOFF, "quotaoff"), | |
2046 | REG_TYPE_STR(LRHEADER, "LR header"), | |
2047 | REG_TYPE_STR(UNMOUNT, "unmount"), | |
2048 | REG_TYPE_STR(COMMIT, "commit"), | |
2049 | REG_TYPE_STR(TRANSHDR, "trans header"), | |
2050 | REG_TYPE_STR(ICREATE, "inode create"), | |
2051 | REG_TYPE_STR(RUI_FORMAT, "rui_format"), | |
2052 | REG_TYPE_STR(RUD_FORMAT, "rud_format"), | |
2053 | REG_TYPE_STR(CUI_FORMAT, "cui_format"), | |
2054 | REG_TYPE_STR(CUD_FORMAT, "cud_format"), | |
2055 | REG_TYPE_STR(BUI_FORMAT, "bui_format"), | |
2056 | REG_TYPE_STR(BUD_FORMAT, "bud_format"), | |
2057 | }; | |
2058 | BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1); | |
2059 | #undef REG_TYPE_STR | |
2060 | ||
2061 | xfs_warn(mp, "ticket reservation summary:"); | |
2062 | xfs_warn(mp, " unit res = %d bytes", | |
2063 | ticket->t_unit_res); | |
2064 | xfs_warn(mp, " current res = %d bytes", | |
2065 | ticket->t_curr_res); | |
2066 | xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)", | |
2067 | ticket->t_res_arr_sum, ticket->t_res_o_flow); | |
2068 | xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)", | |
2069 | ticket->t_res_num_ophdrs, ophdr_spc); | |
2070 | xfs_warn(mp, " ophdr + reg = %u bytes", | |
2071 | ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc); | |
2072 | xfs_warn(mp, " num regions = %u", | |
2073 | ticket->t_res_num); | |
2074 | ||
2075 | for (i = 0; i < ticket->t_res_num; i++) { | |
2076 | uint r_type = ticket->t_res_arr[i].r_type; | |
2077 | xfs_warn(mp, "region[%u]: %s - %u bytes", i, | |
2078 | ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ? | |
2079 | "bad-rtype" : res_type_str[r_type]), | |
2080 | ticket->t_res_arr[i].r_len); | |
2081 | } | |
2082 | } | |
2083 | ||
2084 | /* | |
2085 | * Print a summary of the transaction. | |
2086 | */ | |
2087 | void | |
2088 | xlog_print_trans( | |
2089 | struct xfs_trans *tp) | |
2090 | { | |
2091 | struct xfs_mount *mp = tp->t_mountp; | |
2092 | struct xfs_log_item *lip; | |
2093 | ||
2094 | /* dump core transaction and ticket info */ | |
2095 | xfs_warn(mp, "transaction summary:"); | |
2096 | xfs_warn(mp, " log res = %d", tp->t_log_res); | |
2097 | xfs_warn(mp, " log count = %d", tp->t_log_count); | |
2098 | xfs_warn(mp, " flags = 0x%x", tp->t_flags); | |
2099 | ||
2100 | xlog_print_tic_res(mp, tp->t_ticket); | |
2101 | ||
2102 | /* dump each log item */ | |
2103 | list_for_each_entry(lip, &tp->t_items, li_trans) { | |
2104 | struct xfs_log_vec *lv = lip->li_lv; | |
2105 | struct xfs_log_iovec *vec; | |
2106 | int i; | |
2107 | ||
2108 | xfs_warn(mp, "log item: "); | |
2109 | xfs_warn(mp, " type = 0x%x", lip->li_type); | |
2110 | xfs_warn(mp, " flags = 0x%lx", lip->li_flags); | |
2111 | if (!lv) | |
2112 | continue; | |
2113 | xfs_warn(mp, " niovecs = %d", lv->lv_niovecs); | |
2114 | xfs_warn(mp, " size = %d", lv->lv_size); | |
2115 | xfs_warn(mp, " bytes = %d", lv->lv_bytes); | |
2116 | xfs_warn(mp, " buf len = %d", lv->lv_buf_len); | |
2117 | ||
2118 | /* dump each iovec for the log item */ | |
2119 | vec = lv->lv_iovecp; | |
2120 | for (i = 0; i < lv->lv_niovecs; i++) { | |
2121 | int dumplen = min(vec->i_len, 32); | |
2122 | ||
2123 | xfs_warn(mp, " iovec[%d]", i); | |
2124 | xfs_warn(mp, " type = 0x%x", vec->i_type); | |
2125 | xfs_warn(mp, " len = %d", vec->i_len); | |
2126 | xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i); | |
2127 | xfs_hex_dump(vec->i_addr, dumplen); | |
2128 | ||
2129 | vec++; | |
2130 | } | |
2131 | } | |
2132 | } | |
2133 | ||
2134 | /* | |
2135 | * Calculate the potential space needed by the log vector. We may need a start | |
2136 | * record, and each region gets its own struct xlog_op_header and may need to be | |
2137 | * double word aligned. | |
2138 | */ | |
2139 | static int | |
2140 | xlog_write_calc_vec_length( | |
2141 | struct xlog_ticket *ticket, | |
2142 | struct xfs_log_vec *log_vector, | |
2143 | bool need_start_rec) | |
2144 | { | |
2145 | struct xfs_log_vec *lv; | |
2146 | int headers = need_start_rec ? 1 : 0; | |
2147 | int len = 0; | |
2148 | int i; | |
2149 | ||
2150 | for (lv = log_vector; lv; lv = lv->lv_next) { | |
2151 | /* we don't write ordered log vectors */ | |
2152 | if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) | |
2153 | continue; | |
2154 | ||
2155 | headers += lv->lv_niovecs; | |
2156 | ||
2157 | for (i = 0; i < lv->lv_niovecs; i++) { | |
2158 | struct xfs_log_iovec *vecp = &lv->lv_iovecp[i]; | |
2159 | ||
2160 | len += vecp->i_len; | |
2161 | xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type); | |
2162 | } | |
2163 | } | |
2164 | ||
2165 | ticket->t_res_num_ophdrs += headers; | |
2166 | len += headers * sizeof(struct xlog_op_header); | |
2167 | ||
2168 | return len; | |
2169 | } | |
2170 | ||
2171 | static void | |
2172 | xlog_write_start_rec( | |
2173 | struct xlog_op_header *ophdr, | |
2174 | struct xlog_ticket *ticket) | |
2175 | { | |
2176 | ophdr->oh_tid = cpu_to_be32(ticket->t_tid); | |
2177 | ophdr->oh_clientid = ticket->t_clientid; | |
2178 | ophdr->oh_len = 0; | |
2179 | ophdr->oh_flags = XLOG_START_TRANS; | |
2180 | ophdr->oh_res2 = 0; | |
2181 | } | |
2182 | ||
2183 | static xlog_op_header_t * | |
2184 | xlog_write_setup_ophdr( | |
2185 | struct xlog *log, | |
2186 | struct xlog_op_header *ophdr, | |
2187 | struct xlog_ticket *ticket, | |
2188 | uint flags) | |
2189 | { | |
2190 | ophdr->oh_tid = cpu_to_be32(ticket->t_tid); | |
2191 | ophdr->oh_clientid = ticket->t_clientid; | |
2192 | ophdr->oh_res2 = 0; | |
2193 | ||
2194 | /* are we copying a commit or unmount record? */ | |
2195 | ophdr->oh_flags = flags; | |
2196 | ||
2197 | /* | |
2198 | * We've seen logs corrupted with bad transaction client ids. This | |
2199 | * makes sure that XFS doesn't generate them on. Turn this into an EIO | |
2200 | * and shut down the filesystem. | |
2201 | */ | |
2202 | switch (ophdr->oh_clientid) { | |
2203 | case XFS_TRANSACTION: | |
2204 | case XFS_VOLUME: | |
2205 | case XFS_LOG: | |
2206 | break; | |
2207 | default: | |
2208 | xfs_warn(log->l_mp, | |
2209 | "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT, | |
2210 | ophdr->oh_clientid, ticket); | |
2211 | return NULL; | |
2212 | } | |
2213 | ||
2214 | return ophdr; | |
2215 | } | |
2216 | ||
2217 | /* | |
2218 | * Set up the parameters of the region copy into the log. This has | |
2219 | * to handle region write split across multiple log buffers - this | |
2220 | * state is kept external to this function so that this code can | |
2221 | * be written in an obvious, self documenting manner. | |
2222 | */ | |
2223 | static int | |
2224 | xlog_write_setup_copy( | |
2225 | struct xlog_ticket *ticket, | |
2226 | struct xlog_op_header *ophdr, | |
2227 | int space_available, | |
2228 | int space_required, | |
2229 | int *copy_off, | |
2230 | int *copy_len, | |
2231 | int *last_was_partial_copy, | |
2232 | int *bytes_consumed) | |
2233 | { | |
2234 | int still_to_copy; | |
2235 | ||
2236 | still_to_copy = space_required - *bytes_consumed; | |
2237 | *copy_off = *bytes_consumed; | |
2238 | ||
2239 | if (still_to_copy <= space_available) { | |
2240 | /* write of region completes here */ | |
2241 | *copy_len = still_to_copy; | |
2242 | ophdr->oh_len = cpu_to_be32(*copy_len); | |
2243 | if (*last_was_partial_copy) | |
2244 | ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS); | |
2245 | *last_was_partial_copy = 0; | |
2246 | *bytes_consumed = 0; | |
2247 | return 0; | |
2248 | } | |
2249 | ||
2250 | /* partial write of region, needs extra log op header reservation */ | |
2251 | *copy_len = space_available; | |
2252 | ophdr->oh_len = cpu_to_be32(*copy_len); | |
2253 | ophdr->oh_flags |= XLOG_CONTINUE_TRANS; | |
2254 | if (*last_was_partial_copy) | |
2255 | ophdr->oh_flags |= XLOG_WAS_CONT_TRANS; | |
2256 | *bytes_consumed += *copy_len; | |
2257 | (*last_was_partial_copy)++; | |
2258 | ||
2259 | /* account for new log op header */ | |
2260 | ticket->t_curr_res -= sizeof(struct xlog_op_header); | |
2261 | ticket->t_res_num_ophdrs++; | |
2262 | ||
2263 | return sizeof(struct xlog_op_header); | |
2264 | } | |
2265 | ||
2266 | static int | |
2267 | xlog_write_copy_finish( | |
2268 | struct xlog *log, | |
2269 | struct xlog_in_core *iclog, | |
2270 | uint flags, | |
2271 | int *record_cnt, | |
2272 | int *data_cnt, | |
2273 | int *partial_copy, | |
2274 | int *partial_copy_len, | |
2275 | int log_offset, | |
2276 | struct xlog_in_core **commit_iclog) | |
2277 | { | |
2278 | int error; | |
2279 | ||
2280 | if (*partial_copy) { | |
2281 | /* | |
2282 | * This iclog has already been marked WANT_SYNC by | |
2283 | * xlog_state_get_iclog_space. | |
2284 | */ | |
2285 | spin_lock(&log->l_icloglock); | |
2286 | xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt); | |
2287 | *record_cnt = 0; | |
2288 | *data_cnt = 0; | |
2289 | goto release_iclog; | |
2290 | } | |
2291 | ||
2292 | *partial_copy = 0; | |
2293 | *partial_copy_len = 0; | |
2294 | ||
2295 | if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) { | |
2296 | /* no more space in this iclog - push it. */ | |
2297 | spin_lock(&log->l_icloglock); | |
2298 | xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt); | |
2299 | *record_cnt = 0; | |
2300 | *data_cnt = 0; | |
2301 | ||
2302 | if (iclog->ic_state == XLOG_STATE_ACTIVE) | |
2303 | xlog_state_switch_iclogs(log, iclog, 0); | |
2304 | else | |
2305 | ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC || | |
2306 | iclog->ic_state == XLOG_STATE_IOERROR); | |
2307 | if (!commit_iclog) | |
2308 | goto release_iclog; | |
2309 | spin_unlock(&log->l_icloglock); | |
2310 | ASSERT(flags & XLOG_COMMIT_TRANS); | |
2311 | *commit_iclog = iclog; | |
2312 | } | |
2313 | ||
2314 | return 0; | |
2315 | ||
2316 | release_iclog: | |
2317 | error = xlog_state_release_iclog(log, iclog); | |
2318 | spin_unlock(&log->l_icloglock); | |
2319 | return error; | |
2320 | } | |
2321 | ||
2322 | /* | |
2323 | * Write some region out to in-core log | |
2324 | * | |
2325 | * This will be called when writing externally provided regions or when | |
2326 | * writing out a commit record for a given transaction. | |
2327 | * | |
2328 | * General algorithm: | |
2329 | * 1. Find total length of this write. This may include adding to the | |
2330 | * lengths passed in. | |
2331 | * 2. Check whether we violate the tickets reservation. | |
2332 | * 3. While writing to this iclog | |
2333 | * A. Reserve as much space in this iclog as can get | |
2334 | * B. If this is first write, save away start lsn | |
2335 | * C. While writing this region: | |
2336 | * 1. If first write of transaction, write start record | |
2337 | * 2. Write log operation header (header per region) | |
2338 | * 3. Find out if we can fit entire region into this iclog | |
2339 | * 4. Potentially, verify destination memcpy ptr | |
2340 | * 5. Memcpy (partial) region | |
2341 | * 6. If partial copy, release iclog; otherwise, continue | |
2342 | * copying more regions into current iclog | |
2343 | * 4. Mark want sync bit (in simulation mode) | |
2344 | * 5. Release iclog for potential flush to on-disk log. | |
2345 | * | |
2346 | * ERRORS: | |
2347 | * 1. Panic if reservation is overrun. This should never happen since | |
2348 | * reservation amounts are generated internal to the filesystem. | |
2349 | * NOTES: | |
2350 | * 1. Tickets are single threaded data structures. | |
2351 | * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the | |
2352 | * syncing routine. When a single log_write region needs to span | |
2353 | * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set | |
2354 | * on all log operation writes which don't contain the end of the | |
2355 | * region. The XLOG_END_TRANS bit is used for the in-core log | |
2356 | * operation which contains the end of the continued log_write region. | |
2357 | * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog, | |
2358 | * we don't really know exactly how much space will be used. As a result, | |
2359 | * we don't update ic_offset until the end when we know exactly how many | |
2360 | * bytes have been written out. | |
2361 | */ | |
2362 | int | |
2363 | xlog_write( | |
2364 | struct xlog *log, | |
2365 | struct xfs_log_vec *log_vector, | |
2366 | struct xlog_ticket *ticket, | |
2367 | xfs_lsn_t *start_lsn, | |
2368 | struct xlog_in_core **commit_iclog, | |
2369 | uint flags, | |
2370 | bool need_start_rec) | |
2371 | { | |
2372 | struct xlog_in_core *iclog = NULL; | |
2373 | struct xfs_log_vec *lv = log_vector; | |
2374 | struct xfs_log_iovec *vecp = lv->lv_iovecp; | |
2375 | int index = 0; | |
2376 | int len; | |
2377 | int partial_copy = 0; | |
2378 | int partial_copy_len = 0; | |
2379 | int contwr = 0; | |
2380 | int record_cnt = 0; | |
2381 | int data_cnt = 0; | |
2382 | int error = 0; | |
2383 | ||
2384 | /* | |
2385 | * If this is a commit or unmount transaction, we don't need a start | |
2386 | * record to be written. We do, however, have to account for the | |
2387 | * commit or unmount header that gets written. Hence we always have | |
2388 | * to account for an extra xlog_op_header here. | |
2389 | */ | |
2390 | ticket->t_curr_res -= sizeof(struct xlog_op_header); | |
2391 | if (ticket->t_curr_res < 0) { | |
2392 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, | |
2393 | "ctx ticket reservation ran out. Need to up reservation"); | |
2394 | xlog_print_tic_res(log->l_mp, ticket); | |
2395 | xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR); | |
2396 | } | |
2397 | ||
2398 | len = xlog_write_calc_vec_length(ticket, log_vector, need_start_rec); | |
2399 | *start_lsn = 0; | |
2400 | while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) { | |
2401 | void *ptr; | |
2402 | int log_offset; | |
2403 | ||
2404 | error = xlog_state_get_iclog_space(log, len, &iclog, ticket, | |
2405 | &contwr, &log_offset); | |
2406 | if (error) | |
2407 | return error; | |
2408 | ||
2409 | ASSERT(log_offset <= iclog->ic_size - 1); | |
2410 | ptr = iclog->ic_datap + log_offset; | |
2411 | ||
2412 | /* start_lsn is the first lsn written to. That's all we need. */ | |
2413 | if (!*start_lsn) | |
2414 | *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn); | |
2415 | ||
2416 | /* | |
2417 | * This loop writes out as many regions as can fit in the amount | |
2418 | * of space which was allocated by xlog_state_get_iclog_space(). | |
2419 | */ | |
2420 | while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) { | |
2421 | struct xfs_log_iovec *reg; | |
2422 | struct xlog_op_header *ophdr; | |
2423 | int copy_len; | |
2424 | int copy_off; | |
2425 | bool ordered = false; | |
2426 | ||
2427 | /* ordered log vectors have no regions to write */ | |
2428 | if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) { | |
2429 | ASSERT(lv->lv_niovecs == 0); | |
2430 | ordered = true; | |
2431 | goto next_lv; | |
2432 | } | |
2433 | ||
2434 | reg = &vecp[index]; | |
2435 | ASSERT(reg->i_len % sizeof(int32_t) == 0); | |
2436 | ASSERT((unsigned long)ptr % sizeof(int32_t) == 0); | |
2437 | ||
2438 | /* | |
2439 | * Before we start formatting log vectors, we need to | |
2440 | * write a start record. Only do this for the first | |
2441 | * iclog we write to. | |
2442 | */ | |
2443 | if (need_start_rec) { | |
2444 | xlog_write_start_rec(ptr, ticket); | |
2445 | xlog_write_adv_cnt(&ptr, &len, &log_offset, | |
2446 | sizeof(struct xlog_op_header)); | |
2447 | } | |
2448 | ||
2449 | ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags); | |
2450 | if (!ophdr) | |
2451 | return -EIO; | |
2452 | ||
2453 | xlog_write_adv_cnt(&ptr, &len, &log_offset, | |
2454 | sizeof(struct xlog_op_header)); | |
2455 | ||
2456 | len += xlog_write_setup_copy(ticket, ophdr, | |
2457 | iclog->ic_size-log_offset, | |
2458 | reg->i_len, | |
2459 | ©_off, ©_len, | |
2460 | &partial_copy, | |
2461 | &partial_copy_len); | |
2462 | xlog_verify_dest_ptr(log, ptr); | |
2463 | ||
2464 | /* | |
2465 | * Copy region. | |
2466 | * | |
2467 | * Unmount records just log an opheader, so can have | |
2468 | * empty payloads with no data region to copy. Hence we | |
2469 | * only copy the payload if the vector says it has data | |
2470 | * to copy. | |
2471 | */ | |
2472 | ASSERT(copy_len >= 0); | |
2473 | if (copy_len > 0) { | |
2474 | memcpy(ptr, reg->i_addr + copy_off, copy_len); | |
2475 | xlog_write_adv_cnt(&ptr, &len, &log_offset, | |
2476 | copy_len); | |
2477 | } | |
2478 | copy_len += sizeof(struct xlog_op_header); | |
2479 | record_cnt++; | |
2480 | if (need_start_rec) { | |
2481 | copy_len += sizeof(struct xlog_op_header); | |
2482 | record_cnt++; | |
2483 | need_start_rec = false; | |
2484 | } | |
2485 | data_cnt += contwr ? copy_len : 0; | |
2486 | ||
2487 | error = xlog_write_copy_finish(log, iclog, flags, | |
2488 | &record_cnt, &data_cnt, | |
2489 | &partial_copy, | |
2490 | &partial_copy_len, | |
2491 | log_offset, | |
2492 | commit_iclog); | |
2493 | if (error) | |
2494 | return error; | |
2495 | ||
2496 | /* | |
2497 | * if we had a partial copy, we need to get more iclog | |
2498 | * space but we don't want to increment the region | |
2499 | * index because there is still more is this region to | |
2500 | * write. | |
2501 | * | |
2502 | * If we completed writing this region, and we flushed | |
2503 | * the iclog (indicated by resetting of the record | |
2504 | * count), then we also need to get more log space. If | |
2505 | * this was the last record, though, we are done and | |
2506 | * can just return. | |
2507 | */ | |
2508 | if (partial_copy) | |
2509 | break; | |
2510 | ||
2511 | if (++index == lv->lv_niovecs) { | |
2512 | next_lv: | |
2513 | lv = lv->lv_next; | |
2514 | index = 0; | |
2515 | if (lv) | |
2516 | vecp = lv->lv_iovecp; | |
2517 | } | |
2518 | if (record_cnt == 0 && !ordered) { | |
2519 | if (!lv) | |
2520 | return 0; | |
2521 | break; | |
2522 | } | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | ASSERT(len == 0); | |
2527 | ||
2528 | spin_lock(&log->l_icloglock); | |
2529 | xlog_state_finish_copy(log, iclog, record_cnt, data_cnt); | |
2530 | if (commit_iclog) { | |
2531 | ASSERT(flags & XLOG_COMMIT_TRANS); | |
2532 | *commit_iclog = iclog; | |
2533 | } else { | |
2534 | error = xlog_state_release_iclog(log, iclog); | |
2535 | } | |
2536 | spin_unlock(&log->l_icloglock); | |
2537 | ||
2538 | return error; | |
2539 | } | |
2540 | ||
2541 | static void | |
2542 | xlog_state_activate_iclog( | |
2543 | struct xlog_in_core *iclog, | |
2544 | int *iclogs_changed) | |
2545 | { | |
2546 | ASSERT(list_empty_careful(&iclog->ic_callbacks)); | |
2547 | ||
2548 | /* | |
2549 | * If the number of ops in this iclog indicate it just contains the | |
2550 | * dummy transaction, we can change state into IDLE (the second time | |
2551 | * around). Otherwise we should change the state into NEED a dummy. | |
2552 | * We don't need to cover the dummy. | |
2553 | */ | |
2554 | if (*iclogs_changed == 0 && | |
2555 | iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) { | |
2556 | *iclogs_changed = 1; | |
2557 | } else { | |
2558 | /* | |
2559 | * We have two dirty iclogs so start over. This could also be | |
2560 | * num of ops indicating this is not the dummy going out. | |
2561 | */ | |
2562 | *iclogs_changed = 2; | |
2563 | } | |
2564 | ||
2565 | iclog->ic_state = XLOG_STATE_ACTIVE; | |
2566 | iclog->ic_offset = 0; | |
2567 | iclog->ic_header.h_num_logops = 0; | |
2568 | memset(iclog->ic_header.h_cycle_data, 0, | |
2569 | sizeof(iclog->ic_header.h_cycle_data)); | |
2570 | iclog->ic_header.h_lsn = 0; | |
2571 | } | |
2572 | ||
2573 | /* | |
2574 | * Loop through all iclogs and mark all iclogs currently marked DIRTY as | |
2575 | * ACTIVE after iclog I/O has completed. | |
2576 | */ | |
2577 | static void | |
2578 | xlog_state_activate_iclogs( | |
2579 | struct xlog *log, | |
2580 | int *iclogs_changed) | |
2581 | { | |
2582 | struct xlog_in_core *iclog = log->l_iclog; | |
2583 | ||
2584 | do { | |
2585 | if (iclog->ic_state == XLOG_STATE_DIRTY) | |
2586 | xlog_state_activate_iclog(iclog, iclogs_changed); | |
2587 | /* | |
2588 | * The ordering of marking iclogs ACTIVE must be maintained, so | |
2589 | * an iclog doesn't become ACTIVE beyond one that is SYNCING. | |
2590 | */ | |
2591 | else if (iclog->ic_state != XLOG_STATE_ACTIVE) | |
2592 | break; | |
2593 | } while ((iclog = iclog->ic_next) != log->l_iclog); | |
2594 | } | |
2595 | ||
2596 | static int | |
2597 | xlog_covered_state( | |
2598 | int prev_state, | |
2599 | int iclogs_changed) | |
2600 | { | |
2601 | /* | |
2602 | * We go to NEED for any non-covering writes. We go to NEED2 if we just | |
2603 | * wrote the first covering record (DONE). We go to IDLE if we just | |
2604 | * wrote the second covering record (DONE2) and remain in IDLE until a | |
2605 | * non-covering write occurs. | |
2606 | */ | |
2607 | switch (prev_state) { | |
2608 | case XLOG_STATE_COVER_IDLE: | |
2609 | if (iclogs_changed == 1) | |
2610 | return XLOG_STATE_COVER_IDLE; | |
2611 | case XLOG_STATE_COVER_NEED: | |
2612 | case XLOG_STATE_COVER_NEED2: | |
2613 | break; | |
2614 | case XLOG_STATE_COVER_DONE: | |
2615 | if (iclogs_changed == 1) | |
2616 | return XLOG_STATE_COVER_NEED2; | |
2617 | break; | |
2618 | case XLOG_STATE_COVER_DONE2: | |
2619 | if (iclogs_changed == 1) | |
2620 | return XLOG_STATE_COVER_IDLE; | |
2621 | break; | |
2622 | default: | |
2623 | ASSERT(0); | |
2624 | } | |
2625 | ||
2626 | return XLOG_STATE_COVER_NEED; | |
2627 | } | |
2628 | ||
2629 | STATIC void | |
2630 | xlog_state_clean_iclog( | |
2631 | struct xlog *log, | |
2632 | struct xlog_in_core *dirty_iclog) | |
2633 | { | |
2634 | int iclogs_changed = 0; | |
2635 | ||
2636 | dirty_iclog->ic_state = XLOG_STATE_DIRTY; | |
2637 | ||
2638 | xlog_state_activate_iclogs(log, &iclogs_changed); | |
2639 | wake_up_all(&dirty_iclog->ic_force_wait); | |
2640 | ||
2641 | if (iclogs_changed) { | |
2642 | log->l_covered_state = xlog_covered_state(log->l_covered_state, | |
2643 | iclogs_changed); | |
2644 | } | |
2645 | } | |
2646 | ||
2647 | STATIC xfs_lsn_t | |
2648 | xlog_get_lowest_lsn( | |
2649 | struct xlog *log) | |
2650 | { | |
2651 | struct xlog_in_core *iclog = log->l_iclog; | |
2652 | xfs_lsn_t lowest_lsn = 0, lsn; | |
2653 | ||
2654 | do { | |
2655 | if (iclog->ic_state == XLOG_STATE_ACTIVE || | |
2656 | iclog->ic_state == XLOG_STATE_DIRTY) | |
2657 | continue; | |
2658 | ||
2659 | lsn = be64_to_cpu(iclog->ic_header.h_lsn); | |
2660 | if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0) | |
2661 | lowest_lsn = lsn; | |
2662 | } while ((iclog = iclog->ic_next) != log->l_iclog); | |
2663 | ||
2664 | return lowest_lsn; | |
2665 | } | |
2666 | ||
2667 | /* | |
2668 | * Completion of a iclog IO does not imply that a transaction has completed, as | |
2669 | * transactions can be large enough to span many iclogs. We cannot change the | |
2670 | * tail of the log half way through a transaction as this may be the only | |
2671 | * transaction in the log and moving the tail to point to the middle of it | |
2672 | * will prevent recovery from finding the start of the transaction. Hence we | |
2673 | * should only update the last_sync_lsn if this iclog contains transaction | |
2674 | * completion callbacks on it. | |
2675 | * | |
2676 | * We have to do this before we drop the icloglock to ensure we are the only one | |
2677 | * that can update it. | |
2678 | * | |
2679 | * If we are moving the last_sync_lsn forwards, we also need to ensure we kick | |
2680 | * the reservation grant head pushing. This is due to the fact that the push | |
2681 | * target is bound by the current last_sync_lsn value. Hence if we have a large | |
2682 | * amount of log space bound up in this committing transaction then the | |
2683 | * last_sync_lsn value may be the limiting factor preventing tail pushing from | |
2684 | * freeing space in the log. Hence once we've updated the last_sync_lsn we | |
2685 | * should push the AIL to ensure the push target (and hence the grant head) is | |
2686 | * no longer bound by the old log head location and can move forwards and make | |
2687 | * progress again. | |
2688 | */ | |
2689 | static void | |
2690 | xlog_state_set_callback( | |
2691 | struct xlog *log, | |
2692 | struct xlog_in_core *iclog, | |
2693 | xfs_lsn_t header_lsn) | |
2694 | { | |
2695 | iclog->ic_state = XLOG_STATE_CALLBACK; | |
2696 | ||
2697 | ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn), | |
2698 | header_lsn) <= 0); | |
2699 | ||
2700 | if (list_empty_careful(&iclog->ic_callbacks)) | |
2701 | return; | |
2702 | ||
2703 | atomic64_set(&log->l_last_sync_lsn, header_lsn); | |
2704 | xlog_grant_push_ail(log, 0); | |
2705 | } | |
2706 | ||
2707 | /* | |
2708 | * Return true if we need to stop processing, false to continue to the next | |
2709 | * iclog. The caller will need to run callbacks if the iclog is returned in the | |
2710 | * XLOG_STATE_CALLBACK state. | |
2711 | */ | |
2712 | static bool | |
2713 | xlog_state_iodone_process_iclog( | |
2714 | struct xlog *log, | |
2715 | struct xlog_in_core *iclog, | |
2716 | bool *ioerror) | |
2717 | { | |
2718 | xfs_lsn_t lowest_lsn; | |
2719 | xfs_lsn_t header_lsn; | |
2720 | ||
2721 | switch (iclog->ic_state) { | |
2722 | case XLOG_STATE_ACTIVE: | |
2723 | case XLOG_STATE_DIRTY: | |
2724 | /* | |
2725 | * Skip all iclogs in the ACTIVE & DIRTY states: | |
2726 | */ | |
2727 | return false; | |
2728 | case XLOG_STATE_IOERROR: | |
2729 | /* | |
2730 | * Between marking a filesystem SHUTDOWN and stopping the log, | |
2731 | * we do flush all iclogs to disk (if there wasn't a log I/O | |
2732 | * error). So, we do want things to go smoothly in case of just | |
2733 | * a SHUTDOWN w/o a LOG_IO_ERROR. | |
2734 | */ | |
2735 | *ioerror = true; | |
2736 | return false; | |
2737 | case XLOG_STATE_DONE_SYNC: | |
2738 | /* | |
2739 | * Now that we have an iclog that is in the DONE_SYNC state, do | |
2740 | * one more check here to see if we have chased our tail around. | |
2741 | * If this is not the lowest lsn iclog, then we will leave it | |
2742 | * for another completion to process. | |
2743 | */ | |
2744 | header_lsn = be64_to_cpu(iclog->ic_header.h_lsn); | |
2745 | lowest_lsn = xlog_get_lowest_lsn(log); | |
2746 | if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0) | |
2747 | return false; | |
2748 | xlog_state_set_callback(log, iclog, header_lsn); | |
2749 | return false; | |
2750 | default: | |
2751 | /* | |
2752 | * Can only perform callbacks in order. Since this iclog is not | |
2753 | * in the DONE_SYNC state, we skip the rest and just try to | |
2754 | * clean up. | |
2755 | */ | |
2756 | return true; | |
2757 | } | |
2758 | } | |
2759 | ||
2760 | /* | |
2761 | * Keep processing entries in the iclog callback list until we come around and | |
2762 | * it is empty. We need to atomically see that the list is empty and change the | |
2763 | * state to DIRTY so that we don't miss any more callbacks being added. | |
2764 | * | |
2765 | * This function is called with the icloglock held and returns with it held. We | |
2766 | * drop it while running callbacks, however, as holding it over thousands of | |
2767 | * callbacks is unnecessary and causes excessive contention if we do. | |
2768 | */ | |
2769 | static void | |
2770 | xlog_state_do_iclog_callbacks( | |
2771 | struct xlog *log, | |
2772 | struct xlog_in_core *iclog) | |
2773 | __releases(&log->l_icloglock) | |
2774 | __acquires(&log->l_icloglock) | |
2775 | { | |
2776 | spin_unlock(&log->l_icloglock); | |
2777 | spin_lock(&iclog->ic_callback_lock); | |
2778 | while (!list_empty(&iclog->ic_callbacks)) { | |
2779 | LIST_HEAD(tmp); | |
2780 | ||
2781 | list_splice_init(&iclog->ic_callbacks, &tmp); | |
2782 | ||
2783 | spin_unlock(&iclog->ic_callback_lock); | |
2784 | xlog_cil_process_committed(&tmp); | |
2785 | spin_lock(&iclog->ic_callback_lock); | |
2786 | } | |
2787 | ||
2788 | /* | |
2789 | * Pick up the icloglock while still holding the callback lock so we | |
2790 | * serialise against anyone trying to add more callbacks to this iclog | |
2791 | * now we've finished processing. | |
2792 | */ | |
2793 | spin_lock(&log->l_icloglock); | |
2794 | spin_unlock(&iclog->ic_callback_lock); | |
2795 | } | |
2796 | ||
2797 | STATIC void | |
2798 | xlog_state_do_callback( | |
2799 | struct xlog *log) | |
2800 | { | |
2801 | struct xlog_in_core *iclog; | |
2802 | struct xlog_in_core *first_iclog; | |
2803 | bool cycled_icloglock; | |
2804 | bool ioerror; | |
2805 | int flushcnt = 0; | |
2806 | int repeats = 0; | |
2807 | ||
2808 | spin_lock(&log->l_icloglock); | |
2809 | do { | |
2810 | /* | |
2811 | * Scan all iclogs starting with the one pointed to by the | |
2812 | * log. Reset this starting point each time the log is | |
2813 | * unlocked (during callbacks). | |
2814 | * | |
2815 | * Keep looping through iclogs until one full pass is made | |
2816 | * without running any callbacks. | |
2817 | */ | |
2818 | first_iclog = log->l_iclog; | |
2819 | iclog = log->l_iclog; | |
2820 | cycled_icloglock = false; | |
2821 | ioerror = false; | |
2822 | repeats++; | |
2823 | ||
2824 | do { | |
2825 | if (xlog_state_iodone_process_iclog(log, iclog, | |
2826 | &ioerror)) | |
2827 | break; | |
2828 | ||
2829 | if (iclog->ic_state != XLOG_STATE_CALLBACK && | |
2830 | iclog->ic_state != XLOG_STATE_IOERROR) { | |
2831 | iclog = iclog->ic_next; | |
2832 | continue; | |
2833 | } | |
2834 | ||
2835 | /* | |
2836 | * Running callbacks will drop the icloglock which means | |
2837 | * we'll have to run at least one more complete loop. | |
2838 | */ | |
2839 | cycled_icloglock = true; | |
2840 | xlog_state_do_iclog_callbacks(log, iclog); | |
2841 | if (XLOG_FORCED_SHUTDOWN(log)) | |
2842 | wake_up_all(&iclog->ic_force_wait); | |
2843 | else | |
2844 | xlog_state_clean_iclog(log, iclog); | |
2845 | iclog = iclog->ic_next; | |
2846 | } while (first_iclog != iclog); | |
2847 | ||
2848 | if (repeats > 5000) { | |
2849 | flushcnt += repeats; | |
2850 | repeats = 0; | |
2851 | xfs_warn(log->l_mp, | |
2852 | "%s: possible infinite loop (%d iterations)", | |
2853 | __func__, flushcnt); | |
2854 | } | |
2855 | } while (!ioerror && cycled_icloglock); | |
2856 | ||
2857 | if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE || | |
2858 | log->l_iclog->ic_state == XLOG_STATE_IOERROR) | |
2859 | wake_up_all(&log->l_flush_wait); | |
2860 | ||
2861 | spin_unlock(&log->l_icloglock); | |
2862 | } | |
2863 | ||
2864 | ||
2865 | /* | |
2866 | * Finish transitioning this iclog to the dirty state. | |
2867 | * | |
2868 | * Make sure that we completely execute this routine only when this is | |
2869 | * the last call to the iclog. There is a good chance that iclog flushes, | |
2870 | * when we reach the end of the physical log, get turned into 2 separate | |
2871 | * calls to bwrite. Hence, one iclog flush could generate two calls to this | |
2872 | * routine. By using the reference count bwritecnt, we guarantee that only | |
2873 | * the second completion goes through. | |
2874 | * | |
2875 | * Callbacks could take time, so they are done outside the scope of the | |
2876 | * global state machine log lock. | |
2877 | */ | |
2878 | STATIC void | |
2879 | xlog_state_done_syncing( | |
2880 | struct xlog_in_core *iclog) | |
2881 | { | |
2882 | struct xlog *log = iclog->ic_log; | |
2883 | ||
2884 | spin_lock(&log->l_icloglock); | |
2885 | ASSERT(atomic_read(&iclog->ic_refcnt) == 0); | |
2886 | ||
2887 | /* | |
2888 | * If we got an error, either on the first buffer, or in the case of | |
2889 | * split log writes, on the second, we shut down the file system and | |
2890 | * no iclogs should ever be attempted to be written to disk again. | |
2891 | */ | |
2892 | if (!XLOG_FORCED_SHUTDOWN(log)) { | |
2893 | ASSERT(iclog->ic_state == XLOG_STATE_SYNCING); | |
2894 | iclog->ic_state = XLOG_STATE_DONE_SYNC; | |
2895 | } | |
2896 | ||
2897 | /* | |
2898 | * Someone could be sleeping prior to writing out the next | |
2899 | * iclog buffer, we wake them all, one will get to do the | |
2900 | * I/O, the others get to wait for the result. | |
2901 | */ | |
2902 | wake_up_all(&iclog->ic_write_wait); | |
2903 | spin_unlock(&log->l_icloglock); | |
2904 | xlog_state_do_callback(log); | |
2905 | } | |
2906 | ||
2907 | /* | |
2908 | * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must | |
2909 | * sleep. We wait on the flush queue on the head iclog as that should be | |
2910 | * the first iclog to complete flushing. Hence if all iclogs are syncing, | |
2911 | * we will wait here and all new writes will sleep until a sync completes. | |
2912 | * | |
2913 | * The in-core logs are used in a circular fashion. They are not used | |
2914 | * out-of-order even when an iclog past the head is free. | |
2915 | * | |
2916 | * return: | |
2917 | * * log_offset where xlog_write() can start writing into the in-core | |
2918 | * log's data space. | |
2919 | * * in-core log pointer to which xlog_write() should write. | |
2920 | * * boolean indicating this is a continued write to an in-core log. | |
2921 | * If this is the last write, then the in-core log's offset field | |
2922 | * needs to be incremented, depending on the amount of data which | |
2923 | * is copied. | |
2924 | */ | |
2925 | STATIC int | |
2926 | xlog_state_get_iclog_space( | |
2927 | struct xlog *log, | |
2928 | int len, | |
2929 | struct xlog_in_core **iclogp, | |
2930 | struct xlog_ticket *ticket, | |
2931 | int *continued_write, | |
2932 | int *logoffsetp) | |
2933 | { | |
2934 | int log_offset; | |
2935 | xlog_rec_header_t *head; | |
2936 | xlog_in_core_t *iclog; | |
2937 | ||
2938 | restart: | |
2939 | spin_lock(&log->l_icloglock); | |
2940 | if (XLOG_FORCED_SHUTDOWN(log)) { | |
2941 | spin_unlock(&log->l_icloglock); | |
2942 | return -EIO; | |
2943 | } | |
2944 | ||
2945 | iclog = log->l_iclog; | |
2946 | if (iclog->ic_state != XLOG_STATE_ACTIVE) { | |
2947 | XFS_STATS_INC(log->l_mp, xs_log_noiclogs); | |
2948 | ||
2949 | /* Wait for log writes to have flushed */ | |
2950 | xlog_wait(&log->l_flush_wait, &log->l_icloglock); | |
2951 | goto restart; | |
2952 | } | |
2953 | ||
2954 | head = &iclog->ic_header; | |
2955 | ||
2956 | atomic_inc(&iclog->ic_refcnt); /* prevents sync */ | |
2957 | log_offset = iclog->ic_offset; | |
2958 | ||
2959 | /* On the 1st write to an iclog, figure out lsn. This works | |
2960 | * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are | |
2961 | * committing to. If the offset is set, that's how many blocks | |
2962 | * must be written. | |
2963 | */ | |
2964 | if (log_offset == 0) { | |
2965 | ticket->t_curr_res -= log->l_iclog_hsize; | |
2966 | xlog_tic_add_region(ticket, | |
2967 | log->l_iclog_hsize, | |
2968 | XLOG_REG_TYPE_LRHEADER); | |
2969 | head->h_cycle = cpu_to_be32(log->l_curr_cycle); | |
2970 | head->h_lsn = cpu_to_be64( | |
2971 | xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block)); | |
2972 | ASSERT(log->l_curr_block >= 0); | |
2973 | } | |
2974 | ||
2975 | /* If there is enough room to write everything, then do it. Otherwise, | |
2976 | * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC | |
2977 | * bit is on, so this will get flushed out. Don't update ic_offset | |
2978 | * until you know exactly how many bytes get copied. Therefore, wait | |
2979 | * until later to update ic_offset. | |
2980 | * | |
2981 | * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's | |
2982 | * can fit into remaining data section. | |
2983 | */ | |
2984 | if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) { | |
2985 | int error = 0; | |
2986 | ||
2987 | xlog_state_switch_iclogs(log, iclog, iclog->ic_size); | |
2988 | ||
2989 | /* | |
2990 | * If we are the only one writing to this iclog, sync it to | |
2991 | * disk. We need to do an atomic compare and decrement here to | |
2992 | * avoid racing with concurrent atomic_dec_and_lock() calls in | |
2993 | * xlog_state_release_iclog() when there is more than one | |
2994 | * reference to the iclog. | |
2995 | */ | |
2996 | if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) | |
2997 | error = xlog_state_release_iclog(log, iclog); | |
2998 | spin_unlock(&log->l_icloglock); | |
2999 | if (error) | |
3000 | return error; | |
3001 | goto restart; | |
3002 | } | |
3003 | ||
3004 | /* Do we have enough room to write the full amount in the remainder | |
3005 | * of this iclog? Or must we continue a write on the next iclog and | |
3006 | * mark this iclog as completely taken? In the case where we switch | |
3007 | * iclogs (to mark it taken), this particular iclog will release/sync | |
3008 | * to disk in xlog_write(). | |
3009 | */ | |
3010 | if (len <= iclog->ic_size - iclog->ic_offset) { | |
3011 | *continued_write = 0; | |
3012 | iclog->ic_offset += len; | |
3013 | } else { | |
3014 | *continued_write = 1; | |
3015 | xlog_state_switch_iclogs(log, iclog, iclog->ic_size); | |
3016 | } | |
3017 | *iclogp = iclog; | |
3018 | ||
3019 | ASSERT(iclog->ic_offset <= iclog->ic_size); | |
3020 | spin_unlock(&log->l_icloglock); | |
3021 | ||
3022 | *logoffsetp = log_offset; | |
3023 | return 0; | |
3024 | } | |
3025 | ||
3026 | /* | |
3027 | * The first cnt-1 times a ticket goes through here we don't need to move the | |
3028 | * grant write head because the permanent reservation has reserved cnt times the | |
3029 | * unit amount. Release part of current permanent unit reservation and reset | |
3030 | * current reservation to be one units worth. Also move grant reservation head | |
3031 | * forward. | |
3032 | */ | |
3033 | void | |
3034 | xfs_log_ticket_regrant( | |
3035 | struct xlog *log, | |
3036 | struct xlog_ticket *ticket) | |
3037 | { | |
3038 | trace_xfs_log_ticket_regrant(log, ticket); | |
3039 | ||
3040 | if (ticket->t_cnt > 0) | |
3041 | ticket->t_cnt--; | |
3042 | ||
3043 | xlog_grant_sub_space(log, &log->l_reserve_head.grant, | |
3044 | ticket->t_curr_res); | |
3045 | xlog_grant_sub_space(log, &log->l_write_head.grant, | |
3046 | ticket->t_curr_res); | |
3047 | ticket->t_curr_res = ticket->t_unit_res; | |
3048 | xlog_tic_reset_res(ticket); | |
3049 | ||
3050 | trace_xfs_log_ticket_regrant_sub(log, ticket); | |
3051 | ||
3052 | /* just return if we still have some of the pre-reserved space */ | |
3053 | if (!ticket->t_cnt) { | |
3054 | xlog_grant_add_space(log, &log->l_reserve_head.grant, | |
3055 | ticket->t_unit_res); | |
3056 | trace_xfs_log_ticket_regrant_exit(log, ticket); | |
3057 | ||
3058 | ticket->t_curr_res = ticket->t_unit_res; | |
3059 | xlog_tic_reset_res(ticket); | |
3060 | } | |
3061 | ||
3062 | xfs_log_ticket_put(ticket); | |
3063 | } | |
3064 | ||
3065 | /* | |
3066 | * Give back the space left from a reservation. | |
3067 | * | |
3068 | * All the information we need to make a correct determination of space left | |
3069 | * is present. For non-permanent reservations, things are quite easy. The | |
3070 | * count should have been decremented to zero. We only need to deal with the | |
3071 | * space remaining in the current reservation part of the ticket. If the | |
3072 | * ticket contains a permanent reservation, there may be left over space which | |
3073 | * needs to be released. A count of N means that N-1 refills of the current | |
3074 | * reservation can be done before we need to ask for more space. The first | |
3075 | * one goes to fill up the first current reservation. Once we run out of | |
3076 | * space, the count will stay at zero and the only space remaining will be | |
3077 | * in the current reservation field. | |
3078 | */ | |
3079 | void | |
3080 | xfs_log_ticket_ungrant( | |
3081 | struct xlog *log, | |
3082 | struct xlog_ticket *ticket) | |
3083 | { | |
3084 | int bytes; | |
3085 | ||
3086 | trace_xfs_log_ticket_ungrant(log, ticket); | |
3087 | ||
3088 | if (ticket->t_cnt > 0) | |
3089 | ticket->t_cnt--; | |
3090 | ||
3091 | trace_xfs_log_ticket_ungrant_sub(log, ticket); | |
3092 | ||
3093 | /* | |
3094 | * If this is a permanent reservation ticket, we may be able to free | |
3095 | * up more space based on the remaining count. | |
3096 | */ | |
3097 | bytes = ticket->t_curr_res; | |
3098 | if (ticket->t_cnt > 0) { | |
3099 | ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV); | |
3100 | bytes += ticket->t_unit_res*ticket->t_cnt; | |
3101 | } | |
3102 | ||
3103 | xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes); | |
3104 | xlog_grant_sub_space(log, &log->l_write_head.grant, bytes); | |
3105 | ||
3106 | trace_xfs_log_ticket_ungrant_exit(log, ticket); | |
3107 | ||
3108 | xfs_log_space_wake(log->l_mp); | |
3109 | xfs_log_ticket_put(ticket); | |
3110 | } | |
3111 | ||
3112 | /* | |
3113 | * This routine will mark the current iclog in the ring as WANT_SYNC and move | |
3114 | * the current iclog pointer to the next iclog in the ring. | |
3115 | */ | |
3116 | STATIC void | |
3117 | xlog_state_switch_iclogs( | |
3118 | struct xlog *log, | |
3119 | struct xlog_in_core *iclog, | |
3120 | int eventual_size) | |
3121 | { | |
3122 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); | |
3123 | assert_spin_locked(&log->l_icloglock); | |
3124 | ||
3125 | if (!eventual_size) | |
3126 | eventual_size = iclog->ic_offset; | |
3127 | iclog->ic_state = XLOG_STATE_WANT_SYNC; | |
3128 | iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block); | |
3129 | log->l_prev_block = log->l_curr_block; | |
3130 | log->l_prev_cycle = log->l_curr_cycle; | |
3131 | ||
3132 | /* roll log?: ic_offset changed later */ | |
3133 | log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize); | |
3134 | ||
3135 | /* Round up to next log-sunit */ | |
3136 | if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) && | |
3137 | log->l_mp->m_sb.sb_logsunit > 1) { | |
3138 | uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit); | |
3139 | log->l_curr_block = roundup(log->l_curr_block, sunit_bb); | |
3140 | } | |
3141 | ||
3142 | if (log->l_curr_block >= log->l_logBBsize) { | |
3143 | /* | |
3144 | * Rewind the current block before the cycle is bumped to make | |
3145 | * sure that the combined LSN never transiently moves forward | |
3146 | * when the log wraps to the next cycle. This is to support the | |
3147 | * unlocked sample of these fields from xlog_valid_lsn(). Most | |
3148 | * other cases should acquire l_icloglock. | |
3149 | */ | |
3150 | log->l_curr_block -= log->l_logBBsize; | |
3151 | ASSERT(log->l_curr_block >= 0); | |
3152 | smp_wmb(); | |
3153 | log->l_curr_cycle++; | |
3154 | if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM) | |
3155 | log->l_curr_cycle++; | |
3156 | } | |
3157 | ASSERT(iclog == log->l_iclog); | |
3158 | log->l_iclog = iclog->ic_next; | |
3159 | } | |
3160 | ||
3161 | /* | |
3162 | * Write out all data in the in-core log as of this exact moment in time. | |
3163 | * | |
3164 | * Data may be written to the in-core log during this call. However, | |
3165 | * we don't guarantee this data will be written out. A change from past | |
3166 | * implementation means this routine will *not* write out zero length LRs. | |
3167 | * | |
3168 | * Basically, we try and perform an intelligent scan of the in-core logs. | |
3169 | * If we determine there is no flushable data, we just return. There is no | |
3170 | * flushable data if: | |
3171 | * | |
3172 | * 1. the current iclog is active and has no data; the previous iclog | |
3173 | * is in the active or dirty state. | |
3174 | * 2. the current iclog is drity, and the previous iclog is in the | |
3175 | * active or dirty state. | |
3176 | * | |
3177 | * We may sleep if: | |
3178 | * | |
3179 | * 1. the current iclog is not in the active nor dirty state. | |
3180 | * 2. the current iclog dirty, and the previous iclog is not in the | |
3181 | * active nor dirty state. | |
3182 | * 3. the current iclog is active, and there is another thread writing | |
3183 | * to this particular iclog. | |
3184 | * 4. a) the current iclog is active and has no other writers | |
3185 | * b) when we return from flushing out this iclog, it is still | |
3186 | * not in the active nor dirty state. | |
3187 | */ | |
3188 | int | |
3189 | xfs_log_force( | |
3190 | struct xfs_mount *mp, | |
3191 | uint flags) | |
3192 | { | |
3193 | struct xlog *log = mp->m_log; | |
3194 | struct xlog_in_core *iclog; | |
3195 | xfs_lsn_t lsn; | |
3196 | ||
3197 | XFS_STATS_INC(mp, xs_log_force); | |
3198 | trace_xfs_log_force(mp, 0, _RET_IP_); | |
3199 | ||
3200 | xlog_cil_force(log); | |
3201 | ||
3202 | spin_lock(&log->l_icloglock); | |
3203 | iclog = log->l_iclog; | |
3204 | if (iclog->ic_state == XLOG_STATE_IOERROR) | |
3205 | goto out_error; | |
3206 | ||
3207 | if (iclog->ic_state == XLOG_STATE_DIRTY || | |
3208 | (iclog->ic_state == XLOG_STATE_ACTIVE && | |
3209 | atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) { | |
3210 | /* | |
3211 | * If the head is dirty or (active and empty), then we need to | |
3212 | * look at the previous iclog. | |
3213 | * | |
3214 | * If the previous iclog is active or dirty we are done. There | |
3215 | * is nothing to sync out. Otherwise, we attach ourselves to the | |
3216 | * previous iclog and go to sleep. | |
3217 | */ | |
3218 | iclog = iclog->ic_prev; | |
3219 | } else if (iclog->ic_state == XLOG_STATE_ACTIVE) { | |
3220 | if (atomic_read(&iclog->ic_refcnt) == 0) { | |
3221 | /* | |
3222 | * We are the only one with access to this iclog. | |
3223 | * | |
3224 | * Flush it out now. There should be a roundoff of zero | |
3225 | * to show that someone has already taken care of the | |
3226 | * roundoff from the previous sync. | |
3227 | */ | |
3228 | atomic_inc(&iclog->ic_refcnt); | |
3229 | lsn = be64_to_cpu(iclog->ic_header.h_lsn); | |
3230 | xlog_state_switch_iclogs(log, iclog, 0); | |
3231 | if (xlog_state_release_iclog(log, iclog)) | |
3232 | goto out_error; | |
3233 | ||
3234 | if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) | |
3235 | goto out_unlock; | |
3236 | } else { | |
3237 | /* | |
3238 | * Someone else is writing to this iclog. | |
3239 | * | |
3240 | * Use its call to flush out the data. However, the | |
3241 | * other thread may not force out this LR, so we mark | |
3242 | * it WANT_SYNC. | |
3243 | */ | |
3244 | xlog_state_switch_iclogs(log, iclog, 0); | |
3245 | } | |
3246 | } else { | |
3247 | /* | |
3248 | * If the head iclog is not active nor dirty, we just attach | |
3249 | * ourselves to the head and go to sleep if necessary. | |
3250 | */ | |
3251 | ; | |
3252 | } | |
3253 | ||
3254 | if (flags & XFS_LOG_SYNC) | |
3255 | return xlog_wait_on_iclog(iclog); | |
3256 | out_unlock: | |
3257 | spin_unlock(&log->l_icloglock); | |
3258 | return 0; | |
3259 | out_error: | |
3260 | spin_unlock(&log->l_icloglock); | |
3261 | return -EIO; | |
3262 | } | |
3263 | ||
3264 | static int | |
3265 | __xfs_log_force_lsn( | |
3266 | struct xfs_mount *mp, | |
3267 | xfs_lsn_t lsn, | |
3268 | uint flags, | |
3269 | int *log_flushed, | |
3270 | bool already_slept) | |
3271 | { | |
3272 | struct xlog *log = mp->m_log; | |
3273 | struct xlog_in_core *iclog; | |
3274 | ||
3275 | spin_lock(&log->l_icloglock); | |
3276 | iclog = log->l_iclog; | |
3277 | if (iclog->ic_state == XLOG_STATE_IOERROR) | |
3278 | goto out_error; | |
3279 | ||
3280 | while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) { | |
3281 | iclog = iclog->ic_next; | |
3282 | if (iclog == log->l_iclog) | |
3283 | goto out_unlock; | |
3284 | } | |
3285 | ||
3286 | if (iclog->ic_state == XLOG_STATE_ACTIVE) { | |
3287 | /* | |
3288 | * We sleep here if we haven't already slept (e.g. this is the | |
3289 | * first time we've looked at the correct iclog buf) and the | |
3290 | * buffer before us is going to be sync'ed. The reason for this | |
3291 | * is that if we are doing sync transactions here, by waiting | |
3292 | * for the previous I/O to complete, we can allow a few more | |
3293 | * transactions into this iclog before we close it down. | |
3294 | * | |
3295 | * Otherwise, we mark the buffer WANT_SYNC, and bump up the | |
3296 | * refcnt so we can release the log (which drops the ref count). | |
3297 | * The state switch keeps new transaction commits from using | |
3298 | * this buffer. When the current commits finish writing into | |
3299 | * the buffer, the refcount will drop to zero and the buffer | |
3300 | * will go out then. | |
3301 | */ | |
3302 | if (!already_slept && | |
3303 | (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC || | |
3304 | iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) { | |
3305 | XFS_STATS_INC(mp, xs_log_force_sleep); | |
3306 | ||
3307 | xlog_wait(&iclog->ic_prev->ic_write_wait, | |
3308 | &log->l_icloglock); | |
3309 | return -EAGAIN; | |
3310 | } | |
3311 | atomic_inc(&iclog->ic_refcnt); | |
3312 | xlog_state_switch_iclogs(log, iclog, 0); | |
3313 | if (xlog_state_release_iclog(log, iclog)) | |
3314 | goto out_error; | |
3315 | if (log_flushed) | |
3316 | *log_flushed = 1; | |
3317 | } | |
3318 | ||
3319 | if (flags & XFS_LOG_SYNC) | |
3320 | return xlog_wait_on_iclog(iclog); | |
3321 | out_unlock: | |
3322 | spin_unlock(&log->l_icloglock); | |
3323 | return 0; | |
3324 | out_error: | |
3325 | spin_unlock(&log->l_icloglock); | |
3326 | return -EIO; | |
3327 | } | |
3328 | ||
3329 | /* | |
3330 | * Force the in-core log to disk for a specific LSN. | |
3331 | * | |
3332 | * Find in-core log with lsn. | |
3333 | * If it is in the DIRTY state, just return. | |
3334 | * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC | |
3335 | * state and go to sleep or return. | |
3336 | * If it is in any other state, go to sleep or return. | |
3337 | * | |
3338 | * Synchronous forces are implemented with a wait queue. All callers trying | |
3339 | * to force a given lsn to disk must wait on the queue attached to the | |
3340 | * specific in-core log. When given in-core log finally completes its write | |
3341 | * to disk, that thread will wake up all threads waiting on the queue. | |
3342 | */ | |
3343 | int | |
3344 | xfs_log_force_lsn( | |
3345 | struct xfs_mount *mp, | |
3346 | xfs_lsn_t lsn, | |
3347 | uint flags, | |
3348 | int *log_flushed) | |
3349 | { | |
3350 | int ret; | |
3351 | ASSERT(lsn != 0); | |
3352 | ||
3353 | XFS_STATS_INC(mp, xs_log_force); | |
3354 | trace_xfs_log_force(mp, lsn, _RET_IP_); | |
3355 | ||
3356 | lsn = xlog_cil_force_lsn(mp->m_log, lsn); | |
3357 | if (lsn == NULLCOMMITLSN) | |
3358 | return 0; | |
3359 | ||
3360 | ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false); | |
3361 | if (ret == -EAGAIN) | |
3362 | ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true); | |
3363 | return ret; | |
3364 | } | |
3365 | ||
3366 | /* | |
3367 | * Free a used ticket when its refcount falls to zero. | |
3368 | */ | |
3369 | void | |
3370 | xfs_log_ticket_put( | |
3371 | xlog_ticket_t *ticket) | |
3372 | { | |
3373 | ASSERT(atomic_read(&ticket->t_ref) > 0); | |
3374 | if (atomic_dec_and_test(&ticket->t_ref)) | |
3375 | kmem_cache_free(xfs_log_ticket_zone, ticket); | |
3376 | } | |
3377 | ||
3378 | xlog_ticket_t * | |
3379 | xfs_log_ticket_get( | |
3380 | xlog_ticket_t *ticket) | |
3381 | { | |
3382 | ASSERT(atomic_read(&ticket->t_ref) > 0); | |
3383 | atomic_inc(&ticket->t_ref); | |
3384 | return ticket; | |
3385 | } | |
3386 | ||
3387 | /* | |
3388 | * Figure out the total log space unit (in bytes) that would be | |
3389 | * required for a log ticket. | |
3390 | */ | |
3391 | int | |
3392 | xfs_log_calc_unit_res( | |
3393 | struct xfs_mount *mp, | |
3394 | int unit_bytes) | |
3395 | { | |
3396 | struct xlog *log = mp->m_log; | |
3397 | int iclog_space; | |
3398 | uint num_headers; | |
3399 | ||
3400 | /* | |
3401 | * Permanent reservations have up to 'cnt'-1 active log operations | |
3402 | * in the log. A unit in this case is the amount of space for one | |
3403 | * of these log operations. Normal reservations have a cnt of 1 | |
3404 | * and their unit amount is the total amount of space required. | |
3405 | * | |
3406 | * The following lines of code account for non-transaction data | |
3407 | * which occupy space in the on-disk log. | |
3408 | * | |
3409 | * Normal form of a transaction is: | |
3410 | * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph> | |
3411 | * and then there are LR hdrs, split-recs and roundoff at end of syncs. | |
3412 | * | |
3413 | * We need to account for all the leadup data and trailer data | |
3414 | * around the transaction data. | |
3415 | * And then we need to account for the worst case in terms of using | |
3416 | * more space. | |
3417 | * The worst case will happen if: | |
3418 | * - the placement of the transaction happens to be such that the | |
3419 | * roundoff is at its maximum | |
3420 | * - the transaction data is synced before the commit record is synced | |
3421 | * i.e. <transaction-data><roundoff> | <commit-rec><roundoff> | |
3422 | * Therefore the commit record is in its own Log Record. | |
3423 | * This can happen as the commit record is called with its | |
3424 | * own region to xlog_write(). | |
3425 | * This then means that in the worst case, roundoff can happen for | |
3426 | * the commit-rec as well. | |
3427 | * The commit-rec is smaller than padding in this scenario and so it is | |
3428 | * not added separately. | |
3429 | */ | |
3430 | ||
3431 | /* for trans header */ | |
3432 | unit_bytes += sizeof(xlog_op_header_t); | |
3433 | unit_bytes += sizeof(xfs_trans_header_t); | |
3434 | ||
3435 | /* for start-rec */ | |
3436 | unit_bytes += sizeof(xlog_op_header_t); | |
3437 | ||
3438 | /* | |
3439 | * for LR headers - the space for data in an iclog is the size minus | |
3440 | * the space used for the headers. If we use the iclog size, then we | |
3441 | * undercalculate the number of headers required. | |
3442 | * | |
3443 | * Furthermore - the addition of op headers for split-recs might | |
3444 | * increase the space required enough to require more log and op | |
3445 | * headers, so take that into account too. | |
3446 | * | |
3447 | * IMPORTANT: This reservation makes the assumption that if this | |
3448 | * transaction is the first in an iclog and hence has the LR headers | |
3449 | * accounted to it, then the remaining space in the iclog is | |
3450 | * exclusively for this transaction. i.e. if the transaction is larger | |
3451 | * than the iclog, it will be the only thing in that iclog. | |
3452 | * Fundamentally, this means we must pass the entire log vector to | |
3453 | * xlog_write to guarantee this. | |
3454 | */ | |
3455 | iclog_space = log->l_iclog_size - log->l_iclog_hsize; | |
3456 | num_headers = howmany(unit_bytes, iclog_space); | |
3457 | ||
3458 | /* for split-recs - ophdrs added when data split over LRs */ | |
3459 | unit_bytes += sizeof(xlog_op_header_t) * num_headers; | |
3460 | ||
3461 | /* add extra header reservations if we overrun */ | |
3462 | while (!num_headers || | |
3463 | howmany(unit_bytes, iclog_space) > num_headers) { | |
3464 | unit_bytes += sizeof(xlog_op_header_t); | |
3465 | num_headers++; | |
3466 | } | |
3467 | unit_bytes += log->l_iclog_hsize * num_headers; | |
3468 | ||
3469 | /* for commit-rec LR header - note: padding will subsume the ophdr */ | |
3470 | unit_bytes += log->l_iclog_hsize; | |
3471 | ||
3472 | /* for roundoff padding for transaction data and one for commit record */ | |
3473 | if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) { | |
3474 | /* log su roundoff */ | |
3475 | unit_bytes += 2 * mp->m_sb.sb_logsunit; | |
3476 | } else { | |
3477 | /* BB roundoff */ | |
3478 | unit_bytes += 2 * BBSIZE; | |
3479 | } | |
3480 | ||
3481 | return unit_bytes; | |
3482 | } | |
3483 | ||
3484 | /* | |
3485 | * Allocate and initialise a new log ticket. | |
3486 | */ | |
3487 | struct xlog_ticket * | |
3488 | xlog_ticket_alloc( | |
3489 | struct xlog *log, | |
3490 | int unit_bytes, | |
3491 | int cnt, | |
3492 | char client, | |
3493 | bool permanent) | |
3494 | { | |
3495 | struct xlog_ticket *tic; | |
3496 | int unit_res; | |
3497 | ||
3498 | tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL); | |
3499 | ||
3500 | unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes); | |
3501 | ||
3502 | atomic_set(&tic->t_ref, 1); | |
3503 | tic->t_task = current; | |
3504 | INIT_LIST_HEAD(&tic->t_queue); | |
3505 | tic->t_unit_res = unit_res; | |
3506 | tic->t_curr_res = unit_res; | |
3507 | tic->t_cnt = cnt; | |
3508 | tic->t_ocnt = cnt; | |
3509 | tic->t_tid = prandom_u32(); | |
3510 | tic->t_clientid = client; | |
3511 | if (permanent) | |
3512 | tic->t_flags |= XLOG_TIC_PERM_RESERV; | |
3513 | ||
3514 | xlog_tic_reset_res(tic); | |
3515 | ||
3516 | return tic; | |
3517 | } | |
3518 | ||
3519 | #if defined(DEBUG) | |
3520 | /* | |
3521 | * Make sure that the destination ptr is within the valid data region of | |
3522 | * one of the iclogs. This uses backup pointers stored in a different | |
3523 | * part of the log in case we trash the log structure. | |
3524 | */ | |
3525 | STATIC void | |
3526 | xlog_verify_dest_ptr( | |
3527 | struct xlog *log, | |
3528 | void *ptr) | |
3529 | { | |
3530 | int i; | |
3531 | int good_ptr = 0; | |
3532 | ||
3533 | for (i = 0; i < log->l_iclog_bufs; i++) { | |
3534 | if (ptr >= log->l_iclog_bak[i] && | |
3535 | ptr <= log->l_iclog_bak[i] + log->l_iclog_size) | |
3536 | good_ptr++; | |
3537 | } | |
3538 | ||
3539 | if (!good_ptr) | |
3540 | xfs_emerg(log->l_mp, "%s: invalid ptr", __func__); | |
3541 | } | |
3542 | ||
3543 | /* | |
3544 | * Check to make sure the grant write head didn't just over lap the tail. If | |
3545 | * the cycles are the same, we can't be overlapping. Otherwise, make sure that | |
3546 | * the cycles differ by exactly one and check the byte count. | |
3547 | * | |
3548 | * This check is run unlocked, so can give false positives. Rather than assert | |
3549 | * on failures, use a warn-once flag and a panic tag to allow the admin to | |
3550 | * determine if they want to panic the machine when such an error occurs. For | |
3551 | * debug kernels this will have the same effect as using an assert but, unlinke | |
3552 | * an assert, it can be turned off at runtime. | |
3553 | */ | |
3554 | STATIC void | |
3555 | xlog_verify_grant_tail( | |
3556 | struct xlog *log) | |
3557 | { | |
3558 | int tail_cycle, tail_blocks; | |
3559 | int cycle, space; | |
3560 | ||
3561 | xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space); | |
3562 | xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks); | |
3563 | if (tail_cycle != cycle) { | |
3564 | if (cycle - 1 != tail_cycle && | |
3565 | !(log->l_flags & XLOG_TAIL_WARN)) { | |
3566 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, | |
3567 | "%s: cycle - 1 != tail_cycle", __func__); | |
3568 | log->l_flags |= XLOG_TAIL_WARN; | |
3569 | } | |
3570 | ||
3571 | if (space > BBTOB(tail_blocks) && | |
3572 | !(log->l_flags & XLOG_TAIL_WARN)) { | |
3573 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, | |
3574 | "%s: space > BBTOB(tail_blocks)", __func__); | |
3575 | log->l_flags |= XLOG_TAIL_WARN; | |
3576 | } | |
3577 | } | |
3578 | } | |
3579 | ||
3580 | /* check if it will fit */ | |
3581 | STATIC void | |
3582 | xlog_verify_tail_lsn( | |
3583 | struct xlog *log, | |
3584 | struct xlog_in_core *iclog, | |
3585 | xfs_lsn_t tail_lsn) | |
3586 | { | |
3587 | int blocks; | |
3588 | ||
3589 | if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) { | |
3590 | blocks = | |
3591 | log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn)); | |
3592 | if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize)) | |
3593 | xfs_emerg(log->l_mp, "%s: ran out of log space", __func__); | |
3594 | } else { | |
3595 | ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle); | |
3596 | ||
3597 | if (BLOCK_LSN(tail_lsn) == log->l_prev_block) | |
3598 | xfs_emerg(log->l_mp, "%s: tail wrapped", __func__); | |
3599 | ||
3600 | blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block; | |
3601 | if (blocks < BTOBB(iclog->ic_offset) + 1) | |
3602 | xfs_emerg(log->l_mp, "%s: ran out of log space", __func__); | |
3603 | } | |
3604 | } | |
3605 | ||
3606 | /* | |
3607 | * Perform a number of checks on the iclog before writing to disk. | |
3608 | * | |
3609 | * 1. Make sure the iclogs are still circular | |
3610 | * 2. Make sure we have a good magic number | |
3611 | * 3. Make sure we don't have magic numbers in the data | |
3612 | * 4. Check fields of each log operation header for: | |
3613 | * A. Valid client identifier | |
3614 | * B. tid ptr value falls in valid ptr space (user space code) | |
3615 | * C. Length in log record header is correct according to the | |
3616 | * individual operation headers within record. | |
3617 | * 5. When a bwrite will occur within 5 blocks of the front of the physical | |
3618 | * log, check the preceding blocks of the physical log to make sure all | |
3619 | * the cycle numbers agree with the current cycle number. | |
3620 | */ | |
3621 | STATIC void | |
3622 | xlog_verify_iclog( | |
3623 | struct xlog *log, | |
3624 | struct xlog_in_core *iclog, | |
3625 | int count) | |
3626 | { | |
3627 | xlog_op_header_t *ophead; | |
3628 | xlog_in_core_t *icptr; | |
3629 | xlog_in_core_2_t *xhdr; | |
3630 | void *base_ptr, *ptr, *p; | |
3631 | ptrdiff_t field_offset; | |
3632 | uint8_t clientid; | |
3633 | int len, i, j, k, op_len; | |
3634 | int idx; | |
3635 | ||
3636 | /* check validity of iclog pointers */ | |
3637 | spin_lock(&log->l_icloglock); | |
3638 | icptr = log->l_iclog; | |
3639 | for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next) | |
3640 | ASSERT(icptr); | |
3641 | ||
3642 | if (icptr != log->l_iclog) | |
3643 | xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__); | |
3644 | spin_unlock(&log->l_icloglock); | |
3645 | ||
3646 | /* check log magic numbers */ | |
3647 | if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) | |
3648 | xfs_emerg(log->l_mp, "%s: invalid magic num", __func__); | |
3649 | ||
3650 | base_ptr = ptr = &iclog->ic_header; | |
3651 | p = &iclog->ic_header; | |
3652 | for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) { | |
3653 | if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) | |
3654 | xfs_emerg(log->l_mp, "%s: unexpected magic num", | |
3655 | __func__); | |
3656 | } | |
3657 | ||
3658 | /* check fields */ | |
3659 | len = be32_to_cpu(iclog->ic_header.h_num_logops); | |
3660 | base_ptr = ptr = iclog->ic_datap; | |
3661 | ophead = ptr; | |
3662 | xhdr = iclog->ic_data; | |
3663 | for (i = 0; i < len; i++) { | |
3664 | ophead = ptr; | |
3665 | ||
3666 | /* clientid is only 1 byte */ | |
3667 | p = &ophead->oh_clientid; | |
3668 | field_offset = p - base_ptr; | |
3669 | if (field_offset & 0x1ff) { | |
3670 | clientid = ophead->oh_clientid; | |
3671 | } else { | |
3672 | idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap); | |
3673 | if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { | |
3674 | j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
3675 | k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
3676 | clientid = xlog_get_client_id( | |
3677 | xhdr[j].hic_xheader.xh_cycle_data[k]); | |
3678 | } else { | |
3679 | clientid = xlog_get_client_id( | |
3680 | iclog->ic_header.h_cycle_data[idx]); | |
3681 | } | |
3682 | } | |
3683 | if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) | |
3684 | xfs_warn(log->l_mp, | |
3685 | "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx", | |
3686 | __func__, clientid, ophead, | |
3687 | (unsigned long)field_offset); | |
3688 | ||
3689 | /* check length */ | |
3690 | p = &ophead->oh_len; | |
3691 | field_offset = p - base_ptr; | |
3692 | if (field_offset & 0x1ff) { | |
3693 | op_len = be32_to_cpu(ophead->oh_len); | |
3694 | } else { | |
3695 | idx = BTOBBT((uintptr_t)&ophead->oh_len - | |
3696 | (uintptr_t)iclog->ic_datap); | |
3697 | if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { | |
3698 | j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
3699 | k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); | |
3700 | op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]); | |
3701 | } else { | |
3702 | op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]); | |
3703 | } | |
3704 | } | |
3705 | ptr += sizeof(xlog_op_header_t) + op_len; | |
3706 | } | |
3707 | } | |
3708 | #endif | |
3709 | ||
3710 | /* | |
3711 | * Mark all iclogs IOERROR. l_icloglock is held by the caller. | |
3712 | */ | |
3713 | STATIC int | |
3714 | xlog_state_ioerror( | |
3715 | struct xlog *log) | |
3716 | { | |
3717 | xlog_in_core_t *iclog, *ic; | |
3718 | ||
3719 | iclog = log->l_iclog; | |
3720 | if (iclog->ic_state != XLOG_STATE_IOERROR) { | |
3721 | /* | |
3722 | * Mark all the incore logs IOERROR. | |
3723 | * From now on, no log flushes will result. | |
3724 | */ | |
3725 | ic = iclog; | |
3726 | do { | |
3727 | ic->ic_state = XLOG_STATE_IOERROR; | |
3728 | ic = ic->ic_next; | |
3729 | } while (ic != iclog); | |
3730 | return 0; | |
3731 | } | |
3732 | /* | |
3733 | * Return non-zero, if state transition has already happened. | |
3734 | */ | |
3735 | return 1; | |
3736 | } | |
3737 | ||
3738 | /* | |
3739 | * This is called from xfs_force_shutdown, when we're forcibly | |
3740 | * shutting down the filesystem, typically because of an IO error. | |
3741 | * Our main objectives here are to make sure that: | |
3742 | * a. if !logerror, flush the logs to disk. Anything modified | |
3743 | * after this is ignored. | |
3744 | * b. the filesystem gets marked 'SHUTDOWN' for all interested | |
3745 | * parties to find out, 'atomically'. | |
3746 | * c. those who're sleeping on log reservations, pinned objects and | |
3747 | * other resources get woken up, and be told the bad news. | |
3748 | * d. nothing new gets queued up after (b) and (c) are done. | |
3749 | * | |
3750 | * Note: for the !logerror case we need to flush the regions held in memory out | |
3751 | * to disk first. This needs to be done before the log is marked as shutdown, | |
3752 | * otherwise the iclog writes will fail. | |
3753 | */ | |
3754 | int | |
3755 | xfs_log_force_umount( | |
3756 | struct xfs_mount *mp, | |
3757 | int logerror) | |
3758 | { | |
3759 | struct xlog *log; | |
3760 | int retval; | |
3761 | ||
3762 | log = mp->m_log; | |
3763 | ||
3764 | /* | |
3765 | * If this happens during log recovery, don't worry about | |
3766 | * locking; the log isn't open for business yet. | |
3767 | */ | |
3768 | if (!log || | |
3769 | log->l_flags & XLOG_ACTIVE_RECOVERY) { | |
3770 | mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN; | |
3771 | if (mp->m_sb_bp) | |
3772 | mp->m_sb_bp->b_flags |= XBF_DONE; | |
3773 | return 0; | |
3774 | } | |
3775 | ||
3776 | /* | |
3777 | * Somebody could've already done the hard work for us. | |
3778 | * No need to get locks for this. | |
3779 | */ | |
3780 | if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) { | |
3781 | ASSERT(XLOG_FORCED_SHUTDOWN(log)); | |
3782 | return 1; | |
3783 | } | |
3784 | ||
3785 | /* | |
3786 | * Flush all the completed transactions to disk before marking the log | |
3787 | * being shut down. We need to do it in this order to ensure that | |
3788 | * completed operations are safely on disk before we shut down, and that | |
3789 | * we don't have to issue any buffer IO after the shutdown flags are set | |
3790 | * to guarantee this. | |
3791 | */ | |
3792 | if (!logerror) | |
3793 | xfs_log_force(mp, XFS_LOG_SYNC); | |
3794 | ||
3795 | /* | |
3796 | * mark the filesystem and the as in a shutdown state and wake | |
3797 | * everybody up to tell them the bad news. | |
3798 | */ | |
3799 | spin_lock(&log->l_icloglock); | |
3800 | mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN; | |
3801 | if (mp->m_sb_bp) | |
3802 | mp->m_sb_bp->b_flags |= XBF_DONE; | |
3803 | ||
3804 | /* | |
3805 | * Mark the log and the iclogs with IO error flags to prevent any | |
3806 | * further log IO from being issued or completed. | |
3807 | */ | |
3808 | log->l_flags |= XLOG_IO_ERROR; | |
3809 | retval = xlog_state_ioerror(log); | |
3810 | spin_unlock(&log->l_icloglock); | |
3811 | ||
3812 | /* | |
3813 | * We don't want anybody waiting for log reservations after this. That | |
3814 | * means we have to wake up everybody queued up on reserveq as well as | |
3815 | * writeq. In addition, we make sure in xlog_{re}grant_log_space that | |
3816 | * we don't enqueue anything once the SHUTDOWN flag is set, and this | |
3817 | * action is protected by the grant locks. | |
3818 | */ | |
3819 | xlog_grant_head_wake_all(&log->l_reserve_head); | |
3820 | xlog_grant_head_wake_all(&log->l_write_head); | |
3821 | ||
3822 | /* | |
3823 | * Wake up everybody waiting on xfs_log_force. Wake the CIL push first | |
3824 | * as if the log writes were completed. The abort handling in the log | |
3825 | * item committed callback functions will do this again under lock to | |
3826 | * avoid races. | |
3827 | */ | |
3828 | spin_lock(&log->l_cilp->xc_push_lock); | |
3829 | wake_up_all(&log->l_cilp->xc_commit_wait); | |
3830 | spin_unlock(&log->l_cilp->xc_push_lock); | |
3831 | xlog_state_do_callback(log); | |
3832 | ||
3833 | /* return non-zero if log IOERROR transition had already happened */ | |
3834 | return retval; | |
3835 | } | |
3836 | ||
3837 | STATIC int | |
3838 | xlog_iclogs_empty( | |
3839 | struct xlog *log) | |
3840 | { | |
3841 | xlog_in_core_t *iclog; | |
3842 | ||
3843 | iclog = log->l_iclog; | |
3844 | do { | |
3845 | /* endianness does not matter here, zero is zero in | |
3846 | * any language. | |
3847 | */ | |
3848 | if (iclog->ic_header.h_num_logops) | |
3849 | return 0; | |
3850 | iclog = iclog->ic_next; | |
3851 | } while (iclog != log->l_iclog); | |
3852 | return 1; | |
3853 | } | |
3854 | ||
3855 | /* | |
3856 | * Verify that an LSN stamped into a piece of metadata is valid. This is | |
3857 | * intended for use in read verifiers on v5 superblocks. | |
3858 | */ | |
3859 | bool | |
3860 | xfs_log_check_lsn( | |
3861 | struct xfs_mount *mp, | |
3862 | xfs_lsn_t lsn) | |
3863 | { | |
3864 | struct xlog *log = mp->m_log; | |
3865 | bool valid; | |
3866 | ||
3867 | /* | |
3868 | * norecovery mode skips mount-time log processing and unconditionally | |
3869 | * resets the in-core LSN. We can't validate in this mode, but | |
3870 | * modifications are not allowed anyways so just return true. | |
3871 | */ | |
3872 | if (mp->m_flags & XFS_MOUNT_NORECOVERY) | |
3873 | return true; | |
3874 | ||
3875 | /* | |
3876 | * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is | |
3877 | * handled by recovery and thus safe to ignore here. | |
3878 | */ | |
3879 | if (lsn == NULLCOMMITLSN) | |
3880 | return true; | |
3881 | ||
3882 | valid = xlog_valid_lsn(mp->m_log, lsn); | |
3883 | ||
3884 | /* warn the user about what's gone wrong before verifier failure */ | |
3885 | if (!valid) { | |
3886 | spin_lock(&log->l_icloglock); | |
3887 | xfs_warn(mp, | |
3888 | "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). " | |
3889 | "Please unmount and run xfs_repair (>= v4.3) to resolve.", | |
3890 | CYCLE_LSN(lsn), BLOCK_LSN(lsn), | |
3891 | log->l_curr_cycle, log->l_curr_block); | |
3892 | spin_unlock(&log->l_icloglock); | |
3893 | } | |
3894 | ||
3895 | return valid; | |
3896 | } | |
3897 | ||
3898 | bool | |
3899 | xfs_log_in_recovery( | |
3900 | struct xfs_mount *mp) | |
3901 | { | |
3902 | struct xlog *log = mp->m_log; | |
3903 | ||
3904 | return log->l_flags & XLOG_ACTIVE_RECOVERY; | |
3905 | } |