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 | struct kmem_cache *xfs_log_ticket_cache; |
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 void xlog_state_do_callback( |
45 | struct xlog *log); |
46 | STATIC int |
47 | xlog_state_get_iclog_space( |
48 | struct xlog *log, |
49 | int len, |
50 | struct xlog_in_core **iclog, |
51 | struct xlog_ticket *ticket, |
52 | int *logoffsetp); |
53 | STATIC void |
54 | xlog_grant_push_ail( |
55 | struct xlog *log, |
56 | int need_bytes); |
57 | STATIC void |
58 | xlog_sync( |
59 | struct xlog *log, |
60 | struct xlog_in_core *iclog, |
61 | struct xlog_ticket *ticket); |
62 | #if defined(DEBUG) |
63 | STATIC void |
64 | xlog_verify_grant_tail( |
65 | struct xlog *log); |
66 | STATIC void |
67 | xlog_verify_iclog( |
68 | struct xlog *log, |
69 | struct xlog_in_core *iclog, |
70 | int count); |
71 | STATIC void |
72 | xlog_verify_tail_lsn( |
73 | struct xlog *log, |
74 | struct xlog_in_core *iclog); |
75 | #else |
76 | #define xlog_verify_grant_tail(a) |
77 | #define xlog_verify_iclog(a,b,c) |
78 | #define xlog_verify_tail_lsn(a,b) |
79 | #endif |
80 | |
81 | STATIC int |
82 | xlog_iclogs_empty( |
83 | struct xlog *log); |
84 | |
85 | static int |
86 | xfs_log_cover(struct xfs_mount *); |
87 | |
88 | /* |
89 | * We need to make sure the buffer pointer returned is naturally aligned for the |
90 | * biggest basic data type we put into it. We have already accounted for this |
91 | * padding when sizing the buffer. |
92 | * |
93 | * However, this padding does not get written into the log, and hence we have to |
94 | * track the space used by the log vectors separately to prevent log space hangs |
95 | * due to inaccurate accounting (i.e. a leak) of the used log space through the |
96 | * CIL context ticket. |
97 | * |
98 | * We also add space for the xlog_op_header that describes this region in the |
99 | * log. This prepends the data region we return to the caller to copy their data |
100 | * into, so do all the static initialisation of the ophdr now. Because the ophdr |
101 | * is not 8 byte aligned, we have to be careful to ensure that we align the |
102 | * start of the buffer such that the region we return to the call is 8 byte |
103 | * aligned and packed against the tail of the ophdr. |
104 | */ |
105 | void * |
106 | xlog_prepare_iovec( |
107 | struct xfs_log_vec *lv, |
108 | struct xfs_log_iovec **vecp, |
109 | uint type) |
110 | { |
111 | struct xfs_log_iovec *vec = *vecp; |
112 | struct *oph; |
113 | uint32_t len; |
114 | void *buf; |
115 | |
116 | if (vec) { |
117 | ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs); |
118 | vec++; |
119 | } else { |
120 | vec = &lv->lv_iovecp[0]; |
121 | } |
122 | |
123 | len = lv->lv_buf_len + sizeof(struct xlog_op_header); |
124 | if (!IS_ALIGNED(len, sizeof(uint64_t))) { |
125 | lv->lv_buf_len = round_up(len, sizeof(uint64_t)) - |
126 | sizeof(struct xlog_op_header); |
127 | } |
128 | |
129 | vec->i_type = type; |
130 | vec->i_addr = lv->lv_buf + lv->lv_buf_len; |
131 | |
132 | oph = vec->i_addr; |
133 | oph->oh_clientid = XFS_TRANSACTION; |
134 | oph->oh_res2 = 0; |
135 | oph->oh_flags = 0; |
136 | |
137 | buf = vec->i_addr + sizeof(struct xlog_op_header); |
138 | ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t))); |
139 | |
140 | *vecp = vec; |
141 | return buf; |
142 | } |
143 | |
144 | static void |
145 | xlog_grant_sub_space( |
146 | struct xlog *log, |
147 | atomic64_t *head, |
148 | int bytes) |
149 | { |
150 | int64_t head_val = atomic64_read(v: head); |
151 | int64_t new, old; |
152 | |
153 | do { |
154 | int cycle, space; |
155 | |
156 | xlog_crack_grant_head_val(val: head_val, cycle: &cycle, space: &space); |
157 | |
158 | space -= bytes; |
159 | if (space < 0) { |
160 | space += log->l_logsize; |
161 | cycle--; |
162 | } |
163 | |
164 | old = head_val; |
165 | new = xlog_assign_grant_head_val(cycle, space); |
166 | head_val = atomic64_cmpxchg(v: head, old, new); |
167 | } while (head_val != old); |
168 | } |
169 | |
170 | static void |
171 | xlog_grant_add_space( |
172 | struct xlog *log, |
173 | atomic64_t *head, |
174 | int bytes) |
175 | { |
176 | int64_t head_val = atomic64_read(v: head); |
177 | int64_t new, old; |
178 | |
179 | do { |
180 | int tmp; |
181 | int cycle, space; |
182 | |
183 | xlog_crack_grant_head_val(val: head_val, cycle: &cycle, space: &space); |
184 | |
185 | tmp = log->l_logsize - space; |
186 | if (tmp > bytes) |
187 | space += bytes; |
188 | else { |
189 | space = bytes - tmp; |
190 | cycle++; |
191 | } |
192 | |
193 | old = head_val; |
194 | new = xlog_assign_grant_head_val(cycle, space); |
195 | head_val = atomic64_cmpxchg(v: head, old, new); |
196 | } while (head_val != old); |
197 | } |
198 | |
199 | STATIC void |
200 | xlog_grant_head_init( |
201 | struct xlog_grant_head *head) |
202 | { |
203 | xlog_assign_grant_head(head: &head->grant, cycle: 1, space: 0); |
204 | INIT_LIST_HEAD(list: &head->waiters); |
205 | spin_lock_init(&head->lock); |
206 | } |
207 | |
208 | STATIC void |
209 | xlog_grant_head_wake_all( |
210 | struct xlog_grant_head *head) |
211 | { |
212 | struct xlog_ticket *tic; |
213 | |
214 | spin_lock(lock: &head->lock); |
215 | list_for_each_entry(tic, &head->waiters, t_queue) |
216 | wake_up_process(tsk: tic->t_task); |
217 | spin_unlock(lock: &head->lock); |
218 | } |
219 | |
220 | static inline int |
221 | xlog_ticket_reservation( |
222 | struct xlog *log, |
223 | struct xlog_grant_head *head, |
224 | struct xlog_ticket *tic) |
225 | { |
226 | if (head == &log->l_write_head) { |
227 | ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV); |
228 | return tic->t_unit_res; |
229 | } |
230 | |
231 | if (tic->t_flags & XLOG_TIC_PERM_RESERV) |
232 | return tic->t_unit_res * tic->t_cnt; |
233 | |
234 | return tic->t_unit_res; |
235 | } |
236 | |
237 | STATIC bool |
238 | xlog_grant_head_wake( |
239 | struct xlog *log, |
240 | struct xlog_grant_head *head, |
241 | int *free_bytes) |
242 | { |
243 | struct xlog_ticket *tic; |
244 | int need_bytes; |
245 | bool woken_task = false; |
246 | |
247 | list_for_each_entry(tic, &head->waiters, t_queue) { |
248 | |
249 | /* |
250 | * There is a chance that the size of the CIL checkpoints in |
251 | * progress at the last AIL push target calculation resulted in |
252 | * limiting the target to the log head (l_last_sync_lsn) at the |
253 | * time. This may not reflect where the log head is now as the |
254 | * CIL checkpoints may have completed. |
255 | * |
256 | * Hence when we are woken here, it may be that the head of the |
257 | * log that has moved rather than the tail. As the tail didn't |
258 | * move, there still won't be space available for the |
259 | * reservation we require. However, if the AIL has already |
260 | * pushed to the target defined by the old log head location, we |
261 | * will hang here waiting for something else to update the AIL |
262 | * push target. |
263 | * |
264 | * Therefore, if there isn't space to wake the first waiter on |
265 | * the grant head, we need to push the AIL again to ensure the |
266 | * target reflects both the current log tail and log head |
267 | * position before we wait for the tail to move again. |
268 | */ |
269 | |
270 | need_bytes = xlog_ticket_reservation(log, head, tic); |
271 | if (*free_bytes < need_bytes) { |
272 | if (!woken_task) |
273 | xlog_grant_push_ail(log, need_bytes); |
274 | return false; |
275 | } |
276 | |
277 | *free_bytes -= need_bytes; |
278 | trace_xfs_log_grant_wake_up(log, tic); |
279 | wake_up_process(tsk: tic->t_task); |
280 | woken_task = true; |
281 | } |
282 | |
283 | return true; |
284 | } |
285 | |
286 | STATIC int |
287 | xlog_grant_head_wait( |
288 | struct xlog *log, |
289 | struct xlog_grant_head *head, |
290 | struct xlog_ticket *tic, |
291 | int need_bytes) __releases(&head->lock) |
292 | __acquires(&head->lock) |
293 | { |
294 | list_add_tail(new: &tic->t_queue, head: &head->waiters); |
295 | |
296 | do { |
297 | if (xlog_is_shutdown(log)) |
298 | goto shutdown; |
299 | xlog_grant_push_ail(log, need_bytes); |
300 | |
301 | __set_current_state(TASK_UNINTERRUPTIBLE); |
302 | spin_unlock(lock: &head->lock); |
303 | |
304 | XFS_STATS_INC(log->l_mp, xs_sleep_logspace); |
305 | |
306 | trace_xfs_log_grant_sleep(log, tic); |
307 | schedule(); |
308 | trace_xfs_log_grant_wake(log, tic); |
309 | |
310 | spin_lock(lock: &head->lock); |
311 | if (xlog_is_shutdown(log)) |
312 | goto shutdown; |
313 | } while (xlog_space_left(log, head: &head->grant) < need_bytes); |
314 | |
315 | list_del_init(entry: &tic->t_queue); |
316 | return 0; |
317 | shutdown: |
318 | list_del_init(entry: &tic->t_queue); |
319 | return -EIO; |
320 | } |
321 | |
322 | /* |
323 | * Atomically get the log space required for a log ticket. |
324 | * |
325 | * Once a ticket gets put onto head->waiters, it will only return after the |
326 | * needed reservation is satisfied. |
327 | * |
328 | * This function is structured so that it has a lock free fast path. This is |
329 | * necessary because every new transaction reservation will come through this |
330 | * path. Hence any lock will be globally hot if we take it unconditionally on |
331 | * every pass. |
332 | * |
333 | * As tickets are only ever moved on and off head->waiters under head->lock, we |
334 | * only need to take that lock if we are going to add the ticket to the queue |
335 | * and sleep. We can avoid taking the lock if the ticket was never added to |
336 | * head->waiters because the t_queue list head will be empty and we hold the |
337 | * only reference to it so it can safely be checked unlocked. |
338 | */ |
339 | STATIC int |
340 | xlog_grant_head_check( |
341 | struct xlog *log, |
342 | struct xlog_grant_head *head, |
343 | struct xlog_ticket *tic, |
344 | int *need_bytes) |
345 | { |
346 | int free_bytes; |
347 | int error = 0; |
348 | |
349 | ASSERT(!xlog_in_recovery(log)); |
350 | |
351 | /* |
352 | * If there are other waiters on the queue then give them a chance at |
353 | * logspace before us. Wake up the first waiters, if we do not wake |
354 | * up all the waiters then go to sleep waiting for more free space, |
355 | * otherwise try to get some space for this transaction. |
356 | */ |
357 | *need_bytes = xlog_ticket_reservation(log, head, tic); |
358 | free_bytes = xlog_space_left(log, head: &head->grant); |
359 | if (!list_empty_careful(head: &head->waiters)) { |
360 | spin_lock(lock: &head->lock); |
361 | if (!xlog_grant_head_wake(log, head, free_bytes: &free_bytes) || |
362 | free_bytes < *need_bytes) { |
363 | error = xlog_grant_head_wait(log, head, tic, |
364 | need_bytes: *need_bytes); |
365 | } |
366 | spin_unlock(lock: &head->lock); |
367 | } else if (free_bytes < *need_bytes) { |
368 | spin_lock(lock: &head->lock); |
369 | error = xlog_grant_head_wait(log, head, tic, need_bytes: *need_bytes); |
370 | spin_unlock(lock: &head->lock); |
371 | } |
372 | |
373 | return error; |
374 | } |
375 | |
376 | bool |
377 | xfs_log_writable( |
378 | struct xfs_mount *mp) |
379 | { |
380 | /* |
381 | * Do not write to the log on norecovery mounts, if the data or log |
382 | * devices are read-only, or if the filesystem is shutdown. Read-only |
383 | * mounts allow internal writes for log recovery and unmount purposes, |
384 | * so don't restrict that case. |
385 | */ |
386 | if (xfs_has_norecovery(mp)) |
387 | return false; |
388 | if (xfs_readonly_buftarg(mp->m_ddev_targp)) |
389 | return false; |
390 | if (xfs_readonly_buftarg(mp->m_log->l_targ)) |
391 | return false; |
392 | if (xlog_is_shutdown(log: mp->m_log)) |
393 | return false; |
394 | return true; |
395 | } |
396 | |
397 | /* |
398 | * Replenish the byte reservation required by moving the grant write head. |
399 | */ |
400 | int |
401 | xfs_log_regrant( |
402 | struct xfs_mount *mp, |
403 | struct xlog_ticket *tic) |
404 | { |
405 | struct xlog *log = mp->m_log; |
406 | int need_bytes; |
407 | int error = 0; |
408 | |
409 | if (xlog_is_shutdown(log)) |
410 | return -EIO; |
411 | |
412 | XFS_STATS_INC(mp, xs_try_logspace); |
413 | |
414 | /* |
415 | * This is a new transaction on the ticket, so we need to change the |
416 | * transaction ID so that the next transaction has a different TID in |
417 | * the log. Just add one to the existing tid so that we can see chains |
418 | * of rolling transactions in the log easily. |
419 | */ |
420 | tic->t_tid++; |
421 | |
422 | xlog_grant_push_ail(log, need_bytes: tic->t_unit_res); |
423 | |
424 | tic->t_curr_res = tic->t_unit_res; |
425 | if (tic->t_cnt > 0) |
426 | return 0; |
427 | |
428 | trace_xfs_log_regrant(log, tic); |
429 | |
430 | error = xlog_grant_head_check(log, head: &log->l_write_head, tic, |
431 | need_bytes: &need_bytes); |
432 | if (error) |
433 | goto out_error; |
434 | |
435 | xlog_grant_add_space(log, head: &log->l_write_head.grant, bytes: need_bytes); |
436 | trace_xfs_log_regrant_exit(log, tic); |
437 | xlog_verify_grant_tail(log); |
438 | return 0; |
439 | |
440 | out_error: |
441 | /* |
442 | * If we are failing, make sure the ticket doesn't have any current |
443 | * reservations. We don't want to add this back when the ticket/ |
444 | * transaction gets cancelled. |
445 | */ |
446 | tic->t_curr_res = 0; |
447 | tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ |
448 | return error; |
449 | } |
450 | |
451 | /* |
452 | * Reserve log space and return a ticket corresponding to the reservation. |
453 | * |
454 | * Each reservation is going to reserve extra space for a log record header. |
455 | * When writes happen to the on-disk log, we don't subtract the length of the |
456 | * log record header from any reservation. By wasting space in each |
457 | * reservation, we prevent over allocation problems. |
458 | */ |
459 | int |
460 | xfs_log_reserve( |
461 | struct xfs_mount *mp, |
462 | int unit_bytes, |
463 | int cnt, |
464 | struct xlog_ticket **ticp, |
465 | bool permanent) |
466 | { |
467 | struct xlog *log = mp->m_log; |
468 | struct xlog_ticket *tic; |
469 | int need_bytes; |
470 | int error = 0; |
471 | |
472 | if (xlog_is_shutdown(log)) |
473 | return -EIO; |
474 | |
475 | XFS_STATS_INC(mp, xs_try_logspace); |
476 | |
477 | ASSERT(*ticp == NULL); |
478 | tic = xlog_ticket_alloc(log, unit_bytes, count: cnt, permanent); |
479 | *ticp = tic; |
480 | |
481 | xlog_grant_push_ail(log, need_bytes: tic->t_cnt ? tic->t_unit_res * tic->t_cnt |
482 | : tic->t_unit_res); |
483 | |
484 | trace_xfs_log_reserve(log, tic); |
485 | |
486 | error = xlog_grant_head_check(log, head: &log->l_reserve_head, tic, |
487 | need_bytes: &need_bytes); |
488 | if (error) |
489 | goto out_error; |
490 | |
491 | xlog_grant_add_space(log, head: &log->l_reserve_head.grant, bytes: need_bytes); |
492 | xlog_grant_add_space(log, head: &log->l_write_head.grant, bytes: need_bytes); |
493 | trace_xfs_log_reserve_exit(log, tic); |
494 | xlog_verify_grant_tail(log); |
495 | return 0; |
496 | |
497 | out_error: |
498 | /* |
499 | * If we are failing, make sure the ticket doesn't have any current |
500 | * reservations. We don't want to add this back when the ticket/ |
501 | * transaction gets cancelled. |
502 | */ |
503 | tic->t_curr_res = 0; |
504 | tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ |
505 | return error; |
506 | } |
507 | |
508 | /* |
509 | * Run all the pending iclog callbacks and wake log force waiters and iclog |
510 | * space waiters so they can process the newly set shutdown state. We really |
511 | * don't care what order we process callbacks here because the log is shut down |
512 | * and so state cannot change on disk anymore. However, we cannot wake waiters |
513 | * until the callbacks have been processed because we may be in unmount and |
514 | * we must ensure that all AIL operations the callbacks perform have completed |
515 | * before we tear down the AIL. |
516 | * |
517 | * We avoid processing actively referenced iclogs so that we don't run callbacks |
518 | * while the iclog owner might still be preparing the iclog for IO submssion. |
519 | * These will be caught by xlog_state_iclog_release() and call this function |
520 | * again to process any callbacks that may have been added to that iclog. |
521 | */ |
522 | static void |
523 | xlog_state_shutdown_callbacks( |
524 | struct xlog *log) |
525 | { |
526 | struct xlog_in_core *iclog; |
527 | LIST_HEAD(cb_list); |
528 | |
529 | iclog = log->l_iclog; |
530 | do { |
531 | if (atomic_read(v: &iclog->ic_refcnt)) { |
532 | /* Reference holder will re-run iclog callbacks. */ |
533 | continue; |
534 | } |
535 | list_splice_init(list: &iclog->ic_callbacks, head: &cb_list); |
536 | spin_unlock(lock: &log->l_icloglock); |
537 | |
538 | xlog_cil_process_committed(list: &cb_list); |
539 | |
540 | spin_lock(lock: &log->l_icloglock); |
541 | wake_up_all(&iclog->ic_write_wait); |
542 | wake_up_all(&iclog->ic_force_wait); |
543 | } while ((iclog = iclog->ic_next) != log->l_iclog); |
544 | |
545 | wake_up_all(&log->l_flush_wait); |
546 | } |
547 | |
548 | /* |
549 | * Flush iclog to disk if this is the last reference to the given iclog and the |
550 | * it is in the WANT_SYNC state. |
551 | * |
552 | * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the |
553 | * log tail is updated correctly. NEED_FUA indicates that the iclog will be |
554 | * written to stable storage, and implies that a commit record is contained |
555 | * within the iclog. We need to ensure that the log tail does not move beyond |
556 | * the tail that the first commit record in the iclog ordered against, otherwise |
557 | * correct recovery of that checkpoint becomes dependent on future operations |
558 | * performed on this iclog. |
559 | * |
560 | * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the |
561 | * current tail into iclog. Once the iclog tail is set, future operations must |
562 | * not modify it, otherwise they potentially violate ordering constraints for |
563 | * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in |
564 | * the iclog will get zeroed on activation of the iclog after sync, so we |
565 | * always capture the tail lsn on the iclog on the first NEED_FUA release |
566 | * regardless of the number of active reference counts on this iclog. |
567 | */ |
568 | int |
569 | xlog_state_release_iclog( |
570 | struct xlog *log, |
571 | struct xlog_in_core *iclog, |
572 | struct xlog_ticket *ticket) |
573 | { |
574 | xfs_lsn_t tail_lsn; |
575 | bool last_ref; |
576 | |
577 | lockdep_assert_held(&log->l_icloglock); |
578 | |
579 | trace_xlog_iclog_release(iclog, _RET_IP_); |
580 | /* |
581 | * Grabbing the current log tail needs to be atomic w.r.t. the writing |
582 | * of the tail LSN into the iclog so we guarantee that the log tail does |
583 | * not move between the first time we know that the iclog needs to be |
584 | * made stable and when we eventually submit it. |
585 | */ |
586 | if ((iclog->ic_state == XLOG_STATE_WANT_SYNC || |
587 | (iclog->ic_flags & XLOG_ICL_NEED_FUA)) && |
588 | !iclog->ic_header.h_tail_lsn) { |
589 | tail_lsn = xlog_assign_tail_lsn(log->l_mp); |
590 | iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn); |
591 | } |
592 | |
593 | last_ref = atomic_dec_and_test(v: &iclog->ic_refcnt); |
594 | |
595 | if (xlog_is_shutdown(log)) { |
596 | /* |
597 | * If there are no more references to this iclog, process the |
598 | * pending iclog callbacks that were waiting on the release of |
599 | * this iclog. |
600 | */ |
601 | if (last_ref) |
602 | xlog_state_shutdown_callbacks(log); |
603 | return -EIO; |
604 | } |
605 | |
606 | if (!last_ref) |
607 | return 0; |
608 | |
609 | if (iclog->ic_state != XLOG_STATE_WANT_SYNC) { |
610 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); |
611 | return 0; |
612 | } |
613 | |
614 | iclog->ic_state = XLOG_STATE_SYNCING; |
615 | xlog_verify_tail_lsn(log, iclog); |
616 | trace_xlog_iclog_syncing(iclog, _RET_IP_); |
617 | |
618 | spin_unlock(lock: &log->l_icloglock); |
619 | xlog_sync(log, iclog, ticket); |
620 | spin_lock(lock: &log->l_icloglock); |
621 | return 0; |
622 | } |
623 | |
624 | /* |
625 | * Mount a log filesystem |
626 | * |
627 | * mp - ubiquitous xfs mount point structure |
628 | * log_target - buftarg of on-disk log device |
629 | * blk_offset - Start block # where block size is 512 bytes (BBSIZE) |
630 | * num_bblocks - Number of BBSIZE blocks in on-disk log |
631 | * |
632 | * Return error or zero. |
633 | */ |
634 | int |
635 | xfs_log_mount( |
636 | xfs_mount_t *mp, |
637 | struct xfs_buftarg *log_target, |
638 | xfs_daddr_t blk_offset, |
639 | int num_bblks) |
640 | { |
641 | struct xlog *log; |
642 | int error = 0; |
643 | int min_logfsbs; |
644 | |
645 | if (!xfs_has_norecovery(mp)) { |
646 | xfs_notice(mp, "Mounting V%d Filesystem %pU" , |
647 | XFS_SB_VERSION_NUM(&mp->m_sb), |
648 | &mp->m_sb.sb_uuid); |
649 | } else { |
650 | xfs_notice(mp, |
651 | "Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent." , |
652 | XFS_SB_VERSION_NUM(&mp->m_sb), |
653 | &mp->m_sb.sb_uuid); |
654 | ASSERT(xfs_is_readonly(mp)); |
655 | } |
656 | |
657 | log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks); |
658 | if (IS_ERR(ptr: log)) { |
659 | error = PTR_ERR(ptr: log); |
660 | goto out; |
661 | } |
662 | mp->m_log = log; |
663 | |
664 | /* |
665 | * Now that we have set up the log and it's internal geometry |
666 | * parameters, we can validate the given log space and drop a critical |
667 | * message via syslog if the log size is too small. A log that is too |
668 | * small can lead to unexpected situations in transaction log space |
669 | * reservation stage. The superblock verifier has already validated all |
670 | * the other log geometry constraints, so we don't have to check those |
671 | * here. |
672 | * |
673 | * Note: For v4 filesystems, we can't just reject the mount if the |
674 | * validation fails. This would mean that people would have to |
675 | * downgrade their kernel just to remedy the situation as there is no |
676 | * way to grow the log (short of black magic surgery with xfs_db). |
677 | * |
678 | * We can, however, reject mounts for V5 format filesystems, as the |
679 | * mkfs binary being used to make the filesystem should never create a |
680 | * filesystem with a log that is too small. |
681 | */ |
682 | min_logfsbs = xfs_log_calc_minimum_size(mp); |
683 | if (mp->m_sb.sb_logblocks < min_logfsbs) { |
684 | xfs_warn(mp, |
685 | "Log size %d blocks too small, minimum size is %d blocks" , |
686 | mp->m_sb.sb_logblocks, min_logfsbs); |
687 | |
688 | /* |
689 | * Log check errors are always fatal on v5; or whenever bad |
690 | * metadata leads to a crash. |
691 | */ |
692 | if (xfs_has_crc(mp)) { |
693 | xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!" ); |
694 | ASSERT(0); |
695 | error = -EINVAL; |
696 | goto out_free_log; |
697 | } |
698 | xfs_crit(mp, "Log size out of supported range." ); |
699 | xfs_crit(mp, |
700 | "Continuing onwards, but if log hangs are experienced then please report this message in the bug report." ); |
701 | } |
702 | |
703 | /* |
704 | * Initialize the AIL now we have a log. |
705 | */ |
706 | error = xfs_trans_ail_init(mp); |
707 | if (error) { |
708 | xfs_warn(mp, "AIL initialisation failed: error %d" , error); |
709 | goto out_free_log; |
710 | } |
711 | log->l_ailp = mp->m_ail; |
712 | |
713 | /* |
714 | * skip log recovery on a norecovery mount. pretend it all |
715 | * just worked. |
716 | */ |
717 | if (!xfs_has_norecovery(mp)) { |
718 | error = xlog_recover(log); |
719 | if (error) { |
720 | xfs_warn(mp, "log mount/recovery failed: error %d" , |
721 | error); |
722 | xlog_recover_cancel(log); |
723 | goto out_destroy_ail; |
724 | } |
725 | } |
726 | |
727 | error = xfs_sysfs_init(kobj: &log->l_kobj, ktype: &xfs_log_ktype, parent_kobj: &mp->m_kobj, |
728 | name: "log" ); |
729 | if (error) |
730 | goto out_destroy_ail; |
731 | |
732 | /* Normal transactions can now occur */ |
733 | clear_bit(XLOG_ACTIVE_RECOVERY, addr: &log->l_opstate); |
734 | |
735 | /* |
736 | * Now the log has been fully initialised and we know were our |
737 | * space grant counters are, we can initialise the permanent ticket |
738 | * needed for delayed logging to work. |
739 | */ |
740 | xlog_cil_init_post_recovery(log); |
741 | |
742 | return 0; |
743 | |
744 | out_destroy_ail: |
745 | xfs_trans_ail_destroy(mp); |
746 | out_free_log: |
747 | xlog_dealloc_log(log); |
748 | out: |
749 | return error; |
750 | } |
751 | |
752 | /* |
753 | * Finish the recovery of the file system. This is separate from the |
754 | * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read |
755 | * in the root and real-time bitmap inodes between calling xfs_log_mount() and |
756 | * here. |
757 | * |
758 | * If we finish recovery successfully, start the background log work. If we are |
759 | * not doing recovery, then we have a RO filesystem and we don't need to start |
760 | * it. |
761 | */ |
762 | int |
763 | xfs_log_mount_finish( |
764 | struct xfs_mount *mp) |
765 | { |
766 | struct xlog *log = mp->m_log; |
767 | int error = 0; |
768 | |
769 | if (xfs_has_norecovery(mp)) { |
770 | ASSERT(xfs_is_readonly(mp)); |
771 | return 0; |
772 | } |
773 | |
774 | /* |
775 | * During the second phase of log recovery, we need iget and |
776 | * iput to behave like they do for an active filesystem. |
777 | * xfs_fs_drop_inode needs to be able to prevent the deletion |
778 | * of inodes before we're done replaying log items on those |
779 | * inodes. Turn it off immediately after recovery finishes |
780 | * so that we don't leak the quota inodes if subsequent mount |
781 | * activities fail. |
782 | * |
783 | * We let all inodes involved in redo item processing end up on |
784 | * the LRU instead of being evicted immediately so that if we do |
785 | * something to an unlinked inode, the irele won't cause |
786 | * premature truncation and freeing of the inode, which results |
787 | * in log recovery failure. We have to evict the unreferenced |
788 | * lru inodes after clearing SB_ACTIVE because we don't |
789 | * otherwise clean up the lru if there's a subsequent failure in |
790 | * xfs_mountfs, which leads to us leaking the inodes if nothing |
791 | * else (e.g. quotacheck) references the inodes before the |
792 | * mount failure occurs. |
793 | */ |
794 | mp->m_super->s_flags |= SB_ACTIVE; |
795 | xfs_log_work_queue(mp); |
796 | if (xlog_recovery_needed(log)) |
797 | error = xlog_recover_finish(log); |
798 | mp->m_super->s_flags &= ~SB_ACTIVE; |
799 | evict_inodes(sb: mp->m_super); |
800 | |
801 | /* |
802 | * Drain the buffer LRU after log recovery. This is required for v4 |
803 | * filesystems to avoid leaving around buffers with NULL verifier ops, |
804 | * but we do it unconditionally to make sure we're always in a clean |
805 | * cache state after mount. |
806 | * |
807 | * Don't push in the error case because the AIL may have pending intents |
808 | * that aren't removed until recovery is cancelled. |
809 | */ |
810 | if (xlog_recovery_needed(log)) { |
811 | if (!error) { |
812 | xfs_log_force(mp, XFS_LOG_SYNC); |
813 | xfs_ail_push_all_sync(mp->m_ail); |
814 | } |
815 | xfs_notice(mp, "Ending recovery (logdev: %s)" , |
816 | mp->m_logname ? mp->m_logname : "internal" ); |
817 | } else { |
818 | xfs_info(mp, "Ending clean mount" ); |
819 | } |
820 | xfs_buftarg_drain(mp->m_ddev_targp); |
821 | |
822 | clear_bit(XLOG_RECOVERY_NEEDED, addr: &log->l_opstate); |
823 | |
824 | /* Make sure the log is dead if we're returning failure. */ |
825 | ASSERT(!error || xlog_is_shutdown(log)); |
826 | |
827 | return error; |
828 | } |
829 | |
830 | /* |
831 | * The mount has failed. Cancel the recovery if it hasn't completed and destroy |
832 | * the log. |
833 | */ |
834 | void |
835 | xfs_log_mount_cancel( |
836 | struct xfs_mount *mp) |
837 | { |
838 | xlog_recover_cancel(mp->m_log); |
839 | xfs_log_unmount(mp); |
840 | } |
841 | |
842 | /* |
843 | * Flush out the iclog to disk ensuring that device caches are flushed and |
844 | * the iclog hits stable storage before any completion waiters are woken. |
845 | */ |
846 | static inline int |
847 | xlog_force_iclog( |
848 | struct xlog_in_core *iclog) |
849 | { |
850 | atomic_inc(v: &iclog->ic_refcnt); |
851 | iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA; |
852 | if (iclog->ic_state == XLOG_STATE_ACTIVE) |
853 | xlog_state_switch_iclogs(log: iclog->ic_log, iclog, eventual_size: 0); |
854 | return xlog_state_release_iclog(log: iclog->ic_log, iclog, NULL); |
855 | } |
856 | |
857 | /* |
858 | * Cycle all the iclogbuf locks to make sure all log IO completion |
859 | * is done before we tear down these buffers. |
860 | */ |
861 | static void |
862 | xlog_wait_iclog_completion(struct xlog *log) |
863 | { |
864 | int i; |
865 | struct xlog_in_core *iclog = log->l_iclog; |
866 | |
867 | for (i = 0; i < log->l_iclog_bufs; i++) { |
868 | down(sem: &iclog->ic_sema); |
869 | up(sem: &iclog->ic_sema); |
870 | iclog = iclog->ic_next; |
871 | } |
872 | } |
873 | |
874 | /* |
875 | * Wait for the iclog and all prior iclogs to be written disk as required by the |
876 | * log force state machine. Waiting on ic_force_wait ensures iclog completions |
877 | * have been ordered and callbacks run before we are woken here, hence |
878 | * guaranteeing that all the iclogs up to this one are on stable storage. |
879 | */ |
880 | int |
881 | xlog_wait_on_iclog( |
882 | struct xlog_in_core *iclog) |
883 | __releases(iclog->ic_log->l_icloglock) |
884 | { |
885 | struct xlog *log = iclog->ic_log; |
886 | |
887 | trace_xlog_iclog_wait_on(iclog, _RET_IP_); |
888 | if (!xlog_is_shutdown(log) && |
889 | iclog->ic_state != XLOG_STATE_ACTIVE && |
890 | iclog->ic_state != XLOG_STATE_DIRTY) { |
891 | XFS_STATS_INC(log->l_mp, xs_log_force_sleep); |
892 | xlog_wait(wq: &iclog->ic_force_wait, lock: &log->l_icloglock); |
893 | } else { |
894 | spin_unlock(lock: &log->l_icloglock); |
895 | } |
896 | |
897 | if (xlog_is_shutdown(log)) |
898 | return -EIO; |
899 | return 0; |
900 | } |
901 | |
902 | /* |
903 | * Write out an unmount record using the ticket provided. We have to account for |
904 | * the data space used in the unmount ticket as this write is not done from a |
905 | * transaction context that has already done the accounting for us. |
906 | */ |
907 | static int |
908 | xlog_write_unmount_record( |
909 | struct xlog *log, |
910 | struct xlog_ticket *ticket) |
911 | { |
912 | struct { |
913 | struct ophdr; |
914 | struct xfs_unmount_log_format ulf; |
915 | } unmount_rec = { |
916 | .ophdr = { |
917 | .oh_clientid = XFS_LOG, |
918 | .oh_tid = cpu_to_be32(ticket->t_tid), |
919 | .oh_flags = XLOG_UNMOUNT_TRANS, |
920 | }, |
921 | .ulf = { |
922 | .magic = XLOG_UNMOUNT_TYPE, |
923 | }, |
924 | }; |
925 | struct xfs_log_iovec reg = { |
926 | .i_addr = &unmount_rec, |
927 | .i_len = sizeof(unmount_rec), |
928 | .i_type = XLOG_REG_TYPE_UNMOUNT, |
929 | }; |
930 | struct xfs_log_vec vec = { |
931 | .lv_niovecs = 1, |
932 | .lv_iovecp = ®, |
933 | }; |
934 | LIST_HEAD(lv_chain); |
935 | list_add(new: &vec.lv_list, head: &lv_chain); |
936 | |
937 | BUILD_BUG_ON((sizeof(struct xlog_op_header) + |
938 | sizeof(struct xfs_unmount_log_format)) != |
939 | sizeof(unmount_rec)); |
940 | |
941 | /* account for space used by record data */ |
942 | ticket->t_curr_res -= sizeof(unmount_rec); |
943 | |
944 | return xlog_write(log, NULL, lv_chain: &lv_chain, tic: ticket, len: reg.i_len); |
945 | } |
946 | |
947 | /* |
948 | * Mark the filesystem clean by writing an unmount record to the head of the |
949 | * log. |
950 | */ |
951 | static void |
952 | xlog_unmount_write( |
953 | struct xlog *log) |
954 | { |
955 | struct xfs_mount *mp = log->l_mp; |
956 | struct xlog_in_core *iclog; |
957 | struct xlog_ticket *tic = NULL; |
958 | int error; |
959 | |
960 | error = xfs_log_reserve(mp, unit_bytes: 600, cnt: 1, ticp: &tic, permanent: 0); |
961 | if (error) |
962 | goto out_err; |
963 | |
964 | error = xlog_write_unmount_record(log, ticket: tic); |
965 | /* |
966 | * At this point, we're umounting anyway, so there's no point in |
967 | * transitioning log state to shutdown. Just continue... |
968 | */ |
969 | out_err: |
970 | if (error) |
971 | xfs_alert(mp, "%s: unmount record failed" , __func__); |
972 | |
973 | spin_lock(lock: &log->l_icloglock); |
974 | iclog = log->l_iclog; |
975 | error = xlog_force_iclog(iclog); |
976 | xlog_wait_on_iclog(iclog); |
977 | |
978 | if (tic) { |
979 | trace_xfs_log_umount_write(log, tic); |
980 | xfs_log_ticket_ungrant(log, ticket: tic); |
981 | } |
982 | } |
983 | |
984 | static void |
985 | xfs_log_unmount_verify_iclog( |
986 | struct xlog *log) |
987 | { |
988 | struct xlog_in_core *iclog = log->l_iclog; |
989 | |
990 | do { |
991 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); |
992 | ASSERT(iclog->ic_offset == 0); |
993 | } while ((iclog = iclog->ic_next) != log->l_iclog); |
994 | } |
995 | |
996 | /* |
997 | * Unmount record used to have a string "Unmount filesystem--" in the |
998 | * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE). |
999 | * We just write the magic number now since that particular field isn't |
1000 | * currently architecture converted and "Unmount" is a bit foo. |
1001 | * As far as I know, there weren't any dependencies on the old behaviour. |
1002 | */ |
1003 | static void |
1004 | xfs_log_unmount_write( |
1005 | struct xfs_mount *mp) |
1006 | { |
1007 | struct xlog *log = mp->m_log; |
1008 | |
1009 | if (!xfs_log_writable(mp)) |
1010 | return; |
1011 | |
1012 | xfs_log_force(mp, XFS_LOG_SYNC); |
1013 | |
1014 | if (xlog_is_shutdown(log)) |
1015 | return; |
1016 | |
1017 | /* |
1018 | * If we think the summary counters are bad, avoid writing the unmount |
1019 | * record to force log recovery at next mount, after which the summary |
1020 | * counters will be recalculated. Refer to xlog_check_unmount_rec for |
1021 | * more details. |
1022 | */ |
1023 | if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp, |
1024 | XFS_ERRTAG_FORCE_SUMMARY_RECALC)) { |
1025 | xfs_alert(mp, "%s: will fix summary counters at next mount" , |
1026 | __func__); |
1027 | return; |
1028 | } |
1029 | |
1030 | xfs_log_unmount_verify_iclog(log); |
1031 | xlog_unmount_write(log); |
1032 | } |
1033 | |
1034 | /* |
1035 | * Empty the log for unmount/freeze. |
1036 | * |
1037 | * To do this, we first need to shut down the background log work so it is not |
1038 | * trying to cover the log as we clean up. We then need to unpin all objects in |
1039 | * the log so we can then flush them out. Once they have completed their IO and |
1040 | * run the callbacks removing themselves from the AIL, we can cover the log. |
1041 | */ |
1042 | int |
1043 | xfs_log_quiesce( |
1044 | struct xfs_mount *mp) |
1045 | { |
1046 | /* |
1047 | * Clear log incompat features since we're quiescing the log. Report |
1048 | * failures, though it's not fatal to have a higher log feature |
1049 | * protection level than the log contents actually require. |
1050 | */ |
1051 | if (xfs_clear_incompat_log_features(mp)) { |
1052 | int error; |
1053 | |
1054 | error = xfs_sync_sb(mp, false); |
1055 | if (error) |
1056 | xfs_warn(mp, |
1057 | "Failed to clear log incompat features on quiesce" ); |
1058 | } |
1059 | |
1060 | cancel_delayed_work_sync(dwork: &mp->m_log->l_work); |
1061 | xfs_log_force(mp, XFS_LOG_SYNC); |
1062 | |
1063 | /* |
1064 | * The superblock buffer is uncached and while xfs_ail_push_all_sync() |
1065 | * will push it, xfs_buftarg_wait() will not wait for it. Further, |
1066 | * xfs_buf_iowait() cannot be used because it was pushed with the |
1067 | * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for |
1068 | * the IO to complete. |
1069 | */ |
1070 | xfs_ail_push_all_sync(mp->m_ail); |
1071 | xfs_buftarg_wait(mp->m_ddev_targp); |
1072 | xfs_buf_lock(mp->m_sb_bp); |
1073 | xfs_buf_unlock(mp->m_sb_bp); |
1074 | |
1075 | return xfs_log_cover(mp); |
1076 | } |
1077 | |
1078 | void |
1079 | xfs_log_clean( |
1080 | struct xfs_mount *mp) |
1081 | { |
1082 | xfs_log_quiesce(mp); |
1083 | xfs_log_unmount_write(mp); |
1084 | } |
1085 | |
1086 | /* |
1087 | * Shut down and release the AIL and Log. |
1088 | * |
1089 | * During unmount, we need to ensure we flush all the dirty metadata objects |
1090 | * from the AIL so that the log is empty before we write the unmount record to |
1091 | * the log. Once this is done, we can tear down the AIL and the log. |
1092 | */ |
1093 | void |
1094 | xfs_log_unmount( |
1095 | struct xfs_mount *mp) |
1096 | { |
1097 | xfs_log_clean(mp); |
1098 | |
1099 | /* |
1100 | * If shutdown has come from iclog IO context, the log |
1101 | * cleaning will have been skipped and so we need to wait |
1102 | * for the iclog to complete shutdown processing before we |
1103 | * tear anything down. |
1104 | */ |
1105 | xlog_wait_iclog_completion(log: mp->m_log); |
1106 | |
1107 | xfs_buftarg_drain(mp->m_ddev_targp); |
1108 | |
1109 | xfs_trans_ail_destroy(mp); |
1110 | |
1111 | xfs_sysfs_del(kobj: &mp->m_log->l_kobj); |
1112 | |
1113 | xlog_dealloc_log(log: mp->m_log); |
1114 | } |
1115 | |
1116 | void |
1117 | xfs_log_item_init( |
1118 | struct xfs_mount *mp, |
1119 | struct xfs_log_item *item, |
1120 | int type, |
1121 | const struct xfs_item_ops *ops) |
1122 | { |
1123 | item->li_log = mp->m_log; |
1124 | item->li_ailp = mp->m_ail; |
1125 | item->li_type = type; |
1126 | item->li_ops = ops; |
1127 | item->li_lv = NULL; |
1128 | |
1129 | INIT_LIST_HEAD(list: &item->li_ail); |
1130 | INIT_LIST_HEAD(list: &item->li_cil); |
1131 | INIT_LIST_HEAD(list: &item->li_bio_list); |
1132 | INIT_LIST_HEAD(list: &item->li_trans); |
1133 | } |
1134 | |
1135 | /* |
1136 | * Wake up processes waiting for log space after we have moved the log tail. |
1137 | */ |
1138 | void |
1139 | xfs_log_space_wake( |
1140 | struct xfs_mount *mp) |
1141 | { |
1142 | struct xlog *log = mp->m_log; |
1143 | int free_bytes; |
1144 | |
1145 | if (xlog_is_shutdown(log)) |
1146 | return; |
1147 | |
1148 | if (!list_empty_careful(head: &log->l_write_head.waiters)) { |
1149 | ASSERT(!xlog_in_recovery(log)); |
1150 | |
1151 | spin_lock(lock: &log->l_write_head.lock); |
1152 | free_bytes = xlog_space_left(log, head: &log->l_write_head.grant); |
1153 | xlog_grant_head_wake(log, head: &log->l_write_head, free_bytes: &free_bytes); |
1154 | spin_unlock(lock: &log->l_write_head.lock); |
1155 | } |
1156 | |
1157 | if (!list_empty_careful(head: &log->l_reserve_head.waiters)) { |
1158 | ASSERT(!xlog_in_recovery(log)); |
1159 | |
1160 | spin_lock(lock: &log->l_reserve_head.lock); |
1161 | free_bytes = xlog_space_left(log, head: &log->l_reserve_head.grant); |
1162 | xlog_grant_head_wake(log, head: &log->l_reserve_head, free_bytes: &free_bytes); |
1163 | spin_unlock(lock: &log->l_reserve_head.lock); |
1164 | } |
1165 | } |
1166 | |
1167 | /* |
1168 | * Determine if we have a transaction that has gone to disk that needs to be |
1169 | * covered. To begin the transition to the idle state firstly the log needs to |
1170 | * be idle. That means the CIL, the AIL and the iclogs needs to be empty before |
1171 | * we start attempting to cover the log. |
1172 | * |
1173 | * Only if we are then in a state where covering is needed, the caller is |
1174 | * informed that dummy transactions are required to move the log into the idle |
1175 | * state. |
1176 | * |
1177 | * If there are any items in the AIl or CIL, then we do not want to attempt to |
1178 | * cover the log as we may be in a situation where there isn't log space |
1179 | * available to run a dummy transaction and this can lead to deadlocks when the |
1180 | * tail of the log is pinned by an item that is modified in the CIL. Hence |
1181 | * there's no point in running a dummy transaction at this point because we |
1182 | * can't start trying to idle the log until both the CIL and AIL are empty. |
1183 | */ |
1184 | static bool |
1185 | xfs_log_need_covered( |
1186 | struct xfs_mount *mp) |
1187 | { |
1188 | struct xlog *log = mp->m_log; |
1189 | bool needed = false; |
1190 | |
1191 | if (!xlog_cil_empty(log)) |
1192 | return false; |
1193 | |
1194 | spin_lock(lock: &log->l_icloglock); |
1195 | switch (log->l_covered_state) { |
1196 | case XLOG_STATE_COVER_DONE: |
1197 | case XLOG_STATE_COVER_DONE2: |
1198 | case XLOG_STATE_COVER_IDLE: |
1199 | break; |
1200 | case XLOG_STATE_COVER_NEED: |
1201 | case XLOG_STATE_COVER_NEED2: |
1202 | if (xfs_ail_min_lsn(log->l_ailp)) |
1203 | break; |
1204 | if (!xlog_iclogs_empty(log)) |
1205 | break; |
1206 | |
1207 | needed = true; |
1208 | if (log->l_covered_state == XLOG_STATE_COVER_NEED) |
1209 | log->l_covered_state = XLOG_STATE_COVER_DONE; |
1210 | else |
1211 | log->l_covered_state = XLOG_STATE_COVER_DONE2; |
1212 | break; |
1213 | default: |
1214 | needed = true; |
1215 | break; |
1216 | } |
1217 | spin_unlock(lock: &log->l_icloglock); |
1218 | return needed; |
1219 | } |
1220 | |
1221 | /* |
1222 | * Explicitly cover the log. This is similar to background log covering but |
1223 | * intended for usage in quiesce codepaths. The caller is responsible to ensure |
1224 | * the log is idle and suitable for covering. The CIL, iclog buffers and AIL |
1225 | * must all be empty. |
1226 | */ |
1227 | static int |
1228 | xfs_log_cover( |
1229 | struct xfs_mount *mp) |
1230 | { |
1231 | int error = 0; |
1232 | bool need_covered; |
1233 | |
1234 | ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) && |
1235 | !xfs_ail_min_lsn(mp->m_log->l_ailp)) || |
1236 | xlog_is_shutdown(mp->m_log)); |
1237 | |
1238 | if (!xfs_log_writable(mp)) |
1239 | return 0; |
1240 | |
1241 | /* |
1242 | * xfs_log_need_covered() is not idempotent because it progresses the |
1243 | * state machine if the log requires covering. Therefore, we must call |
1244 | * this function once and use the result until we've issued an sb sync. |
1245 | * Do so first to make that abundantly clear. |
1246 | * |
1247 | * Fall into the covering sequence if the log needs covering or the |
1248 | * mount has lazy superblock accounting to sync to disk. The sb sync |
1249 | * used for covering accumulates the in-core counters, so covering |
1250 | * handles this for us. |
1251 | */ |
1252 | need_covered = xfs_log_need_covered(mp); |
1253 | if (!need_covered && !xfs_has_lazysbcount(mp)) |
1254 | return 0; |
1255 | |
1256 | /* |
1257 | * To cover the log, commit the superblock twice (at most) in |
1258 | * independent checkpoints. The first serves as a reference for the |
1259 | * tail pointer. The sync transaction and AIL push empties the AIL and |
1260 | * updates the in-core tail to the LSN of the first checkpoint. The |
1261 | * second commit updates the on-disk tail with the in-core LSN, |
1262 | * covering the log. Push the AIL one more time to leave it empty, as |
1263 | * we found it. |
1264 | */ |
1265 | do { |
1266 | error = xfs_sync_sb(mp, true); |
1267 | if (error) |
1268 | break; |
1269 | xfs_ail_push_all_sync(mp->m_ail); |
1270 | } while (xfs_log_need_covered(mp)); |
1271 | |
1272 | return error; |
1273 | } |
1274 | |
1275 | /* |
1276 | * We may be holding the log iclog lock upon entering this routine. |
1277 | */ |
1278 | xfs_lsn_t |
1279 | xlog_assign_tail_lsn_locked( |
1280 | struct xfs_mount *mp) |
1281 | { |
1282 | struct xlog *log = mp->m_log; |
1283 | struct xfs_log_item *lip; |
1284 | xfs_lsn_t tail_lsn; |
1285 | |
1286 | assert_spin_locked(&mp->m_ail->ail_lock); |
1287 | |
1288 | /* |
1289 | * To make sure we always have a valid LSN for the log tail we keep |
1290 | * track of the last LSN which was committed in log->l_last_sync_lsn, |
1291 | * and use that when the AIL was empty. |
1292 | */ |
1293 | lip = xfs_ail_min(ailp: mp->m_ail); |
1294 | if (lip) |
1295 | tail_lsn = lip->li_lsn; |
1296 | else |
1297 | tail_lsn = atomic64_read(&log->l_last_sync_lsn); |
1298 | trace_xfs_log_assign_tail_lsn(log, tail_lsn); |
1299 | atomic64_set(&log->l_tail_lsn, tail_lsn); |
1300 | return tail_lsn; |
1301 | } |
1302 | |
1303 | xfs_lsn_t |
1304 | xlog_assign_tail_lsn( |
1305 | struct xfs_mount *mp) |
1306 | { |
1307 | xfs_lsn_t tail_lsn; |
1308 | |
1309 | spin_lock(lock: &mp->m_ail->ail_lock); |
1310 | tail_lsn = xlog_assign_tail_lsn_locked(mp); |
1311 | spin_unlock(lock: &mp->m_ail->ail_lock); |
1312 | |
1313 | return tail_lsn; |
1314 | } |
1315 | |
1316 | /* |
1317 | * Return the space in the log between the tail and the head. The head |
1318 | * is passed in the cycle/bytes formal parms. In the special case where |
1319 | * the reserve head has wrapped passed the tail, this calculation is no |
1320 | * longer valid. In this case, just return 0 which means there is no space |
1321 | * in the log. This works for all places where this function is called |
1322 | * with the reserve head. Of course, if the write head were to ever |
1323 | * wrap the tail, we should blow up. Rather than catch this case here, |
1324 | * we depend on other ASSERTions in other parts of the code. XXXmiken |
1325 | * |
1326 | * If reservation head is behind the tail, we have a problem. Warn about it, |
1327 | * but then treat it as if the log is empty. |
1328 | * |
1329 | * If the log is shut down, the head and tail may be invalid or out of whack, so |
1330 | * shortcut invalidity asserts in this case so that we don't trigger them |
1331 | * falsely. |
1332 | */ |
1333 | STATIC int |
1334 | xlog_space_left( |
1335 | struct xlog *log, |
1336 | atomic64_t *head) |
1337 | { |
1338 | int tail_bytes; |
1339 | int tail_cycle; |
1340 | int head_cycle; |
1341 | int head_bytes; |
1342 | |
1343 | xlog_crack_grant_head(head, cycle: &head_cycle, space: &head_bytes); |
1344 | xlog_crack_atomic_lsn(lsn: &log->l_tail_lsn, cycle: &tail_cycle, block: &tail_bytes); |
1345 | tail_bytes = BBTOB(tail_bytes); |
1346 | if (tail_cycle == head_cycle && head_bytes >= tail_bytes) |
1347 | return log->l_logsize - (head_bytes - tail_bytes); |
1348 | if (tail_cycle + 1 < head_cycle) |
1349 | return 0; |
1350 | |
1351 | /* Ignore potential inconsistency when shutdown. */ |
1352 | if (xlog_is_shutdown(log)) |
1353 | return log->l_logsize; |
1354 | |
1355 | if (tail_cycle < head_cycle) { |
1356 | ASSERT(tail_cycle == (head_cycle - 1)); |
1357 | return tail_bytes - head_bytes; |
1358 | } |
1359 | |
1360 | /* |
1361 | * The reservation head is behind the tail. In this case we just want to |
1362 | * return the size of the log as the amount of space left. |
1363 | */ |
1364 | xfs_alert(log->l_mp, "xlog_space_left: head behind tail" ); |
1365 | xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d" , |
1366 | tail_cycle, tail_bytes); |
1367 | xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d" , |
1368 | head_cycle, head_bytes); |
1369 | ASSERT(0); |
1370 | return log->l_logsize; |
1371 | } |
1372 | |
1373 | |
1374 | static void |
1375 | xlog_ioend_work( |
1376 | struct work_struct *work) |
1377 | { |
1378 | struct xlog_in_core *iclog = |
1379 | container_of(work, struct xlog_in_core, ic_end_io_work); |
1380 | struct xlog *log = iclog->ic_log; |
1381 | int error; |
1382 | |
1383 | error = blk_status_to_errno(status: iclog->ic_bio.bi_status); |
1384 | #ifdef DEBUG |
1385 | /* treat writes with injected CRC errors as failed */ |
1386 | if (iclog->ic_fail_crc) |
1387 | error = -EIO; |
1388 | #endif |
1389 | |
1390 | /* |
1391 | * Race to shutdown the filesystem if we see an error. |
1392 | */ |
1393 | if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) { |
1394 | xfs_alert(log->l_mp, "log I/O error %d" , error); |
1395 | xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); |
1396 | } |
1397 | |
1398 | xlog_state_done_syncing(iclog); |
1399 | bio_uninit(&iclog->ic_bio); |
1400 | |
1401 | /* |
1402 | * Drop the lock to signal that we are done. Nothing references the |
1403 | * iclog after this, so an unmount waiting on this lock can now tear it |
1404 | * down safely. As such, it is unsafe to reference the iclog after the |
1405 | * unlock as we could race with it being freed. |
1406 | */ |
1407 | up(sem: &iclog->ic_sema); |
1408 | } |
1409 | |
1410 | /* |
1411 | * Return size of each in-core log record buffer. |
1412 | * |
1413 | * All machines get 8 x 32kB buffers by default, unless tuned otherwise. |
1414 | * |
1415 | * If the filesystem blocksize is too large, we may need to choose a |
1416 | * larger size since the directory code currently logs entire blocks. |
1417 | */ |
1418 | STATIC void |
1419 | xlog_get_iclog_buffer_size( |
1420 | struct xfs_mount *mp, |
1421 | struct xlog *log) |
1422 | { |
1423 | if (mp->m_logbufs <= 0) |
1424 | mp->m_logbufs = XLOG_MAX_ICLOGS; |
1425 | if (mp->m_logbsize <= 0) |
1426 | mp->m_logbsize = XLOG_BIG_RECORD_BSIZE; |
1427 | |
1428 | log->l_iclog_bufs = mp->m_logbufs; |
1429 | log->l_iclog_size = mp->m_logbsize; |
1430 | |
1431 | /* |
1432 | * # headers = size / 32k - one header holds cycles from 32k of data. |
1433 | */ |
1434 | log->l_iclog_heads = |
1435 | DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE); |
1436 | log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT; |
1437 | } |
1438 | |
1439 | void |
1440 | xfs_log_work_queue( |
1441 | struct xfs_mount *mp) |
1442 | { |
1443 | queue_delayed_work(wq: mp->m_sync_workqueue, dwork: &mp->m_log->l_work, |
1444 | delay: msecs_to_jiffies(xfs_syncd_centisecs * 10)); |
1445 | } |
1446 | |
1447 | /* |
1448 | * Clear the log incompat flags if we have the opportunity. |
1449 | * |
1450 | * This only happens if we're about to log the second dummy transaction as part |
1451 | * of covering the log and we can get the log incompat feature usage lock. |
1452 | */ |
1453 | static inline void |
1454 | xlog_clear_incompat( |
1455 | struct xlog *log) |
1456 | { |
1457 | struct xfs_mount *mp = log->l_mp; |
1458 | |
1459 | if (!xfs_sb_has_incompat_log_feature(&mp->m_sb, |
1460 | XFS_SB_FEAT_INCOMPAT_LOG_ALL)) |
1461 | return; |
1462 | |
1463 | if (log->l_covered_state != XLOG_STATE_COVER_DONE2) |
1464 | return; |
1465 | |
1466 | if (!down_write_trylock(sem: &log->l_incompat_users)) |
1467 | return; |
1468 | |
1469 | xfs_clear_incompat_log_features(mp); |
1470 | up_write(sem: &log->l_incompat_users); |
1471 | } |
1472 | |
1473 | /* |
1474 | * Every sync period we need to unpin all items in the AIL and push them to |
1475 | * disk. If there is nothing dirty, then we might need to cover the log to |
1476 | * indicate that the filesystem is idle. |
1477 | */ |
1478 | static void |
1479 | xfs_log_worker( |
1480 | struct work_struct *work) |
1481 | { |
1482 | struct xlog *log = container_of(to_delayed_work(work), |
1483 | struct xlog, l_work); |
1484 | struct xfs_mount *mp = log->l_mp; |
1485 | |
1486 | /* dgc: errors ignored - not fatal and nowhere to report them */ |
1487 | if (xfs_fs_writable(mp, level: SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) { |
1488 | /* |
1489 | * Dump a transaction into the log that contains no real change. |
1490 | * This is needed to stamp the current tail LSN into the log |
1491 | * during the covering operation. |
1492 | * |
1493 | * We cannot use an inode here for this - that will push dirty |
1494 | * state back up into the VFS and then periodic inode flushing |
1495 | * will prevent log covering from making progress. Hence we |
1496 | * synchronously log the superblock instead to ensure the |
1497 | * superblock is immediately unpinned and can be written back. |
1498 | */ |
1499 | xlog_clear_incompat(log); |
1500 | xfs_sync_sb(mp, true); |
1501 | } else |
1502 | xfs_log_force(mp, flags: 0); |
1503 | |
1504 | /* start pushing all the metadata that is currently dirty */ |
1505 | xfs_ail_push_all(mp->m_ail); |
1506 | |
1507 | /* queue us up again */ |
1508 | xfs_log_work_queue(mp); |
1509 | } |
1510 | |
1511 | /* |
1512 | * This routine initializes some of the log structure for a given mount point. |
1513 | * Its primary purpose is to fill in enough, so recovery can occur. However, |
1514 | * some other stuff may be filled in too. |
1515 | */ |
1516 | STATIC struct xlog * |
1517 | xlog_alloc_log( |
1518 | struct xfs_mount *mp, |
1519 | struct xfs_buftarg *log_target, |
1520 | xfs_daddr_t blk_offset, |
1521 | int num_bblks) |
1522 | { |
1523 | struct xlog *log; |
1524 | xlog_rec_header_t *head; |
1525 | xlog_in_core_t **iclogp; |
1526 | xlog_in_core_t *iclog, *prev_iclog=NULL; |
1527 | int i; |
1528 | int error = -ENOMEM; |
1529 | uint log2_size = 0; |
1530 | |
1531 | log = kzalloc(size: sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL); |
1532 | if (!log) { |
1533 | xfs_warn(mp, "Log allocation failed: No memory!" ); |
1534 | goto out; |
1535 | } |
1536 | |
1537 | log->l_mp = mp; |
1538 | log->l_targ = log_target; |
1539 | log->l_logsize = BBTOB(num_bblks); |
1540 | log->l_logBBstart = blk_offset; |
1541 | log->l_logBBsize = num_bblks; |
1542 | log->l_covered_state = XLOG_STATE_COVER_IDLE; |
1543 | set_bit(XLOG_ACTIVE_RECOVERY, addr: &log->l_opstate); |
1544 | INIT_DELAYED_WORK(&log->l_work, xfs_log_worker); |
1545 | INIT_LIST_HEAD(list: &log->r_dfops); |
1546 | |
1547 | log->l_prev_block = -1; |
1548 | /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */ |
1549 | xlog_assign_atomic_lsn(lsn: &log->l_tail_lsn, cycle: 1, block: 0); |
1550 | xlog_assign_atomic_lsn(lsn: &log->l_last_sync_lsn, cycle: 1, block: 0); |
1551 | log->l_curr_cycle = 1; /* 0 is bad since this is initial value */ |
1552 | |
1553 | if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1) |
1554 | log->l_iclog_roundoff = mp->m_sb.sb_logsunit; |
1555 | else |
1556 | log->l_iclog_roundoff = BBSIZE; |
1557 | |
1558 | xlog_grant_head_init(head: &log->l_reserve_head); |
1559 | xlog_grant_head_init(head: &log->l_write_head); |
1560 | |
1561 | error = -EFSCORRUPTED; |
1562 | if (xfs_has_sector(mp)) { |
1563 | log2_size = mp->m_sb.sb_logsectlog; |
1564 | if (log2_size < BBSHIFT) { |
1565 | xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)" , |
1566 | log2_size, BBSHIFT); |
1567 | goto out_free_log; |
1568 | } |
1569 | |
1570 | log2_size -= BBSHIFT; |
1571 | if (log2_size > mp->m_sectbb_log) { |
1572 | xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)" , |
1573 | log2_size, mp->m_sectbb_log); |
1574 | goto out_free_log; |
1575 | } |
1576 | |
1577 | /* for larger sector sizes, must have v2 or external log */ |
1578 | if (log2_size && log->l_logBBstart > 0 && |
1579 | !xfs_has_logv2(mp)) { |
1580 | xfs_warn(mp, |
1581 | "log sector size (0x%x) invalid for configuration." , |
1582 | log2_size); |
1583 | goto out_free_log; |
1584 | } |
1585 | } |
1586 | log->l_sectBBsize = 1 << log2_size; |
1587 | |
1588 | init_rwsem(&log->l_incompat_users); |
1589 | |
1590 | xlog_get_iclog_buffer_size(mp, log); |
1591 | |
1592 | spin_lock_init(&log->l_icloglock); |
1593 | init_waitqueue_head(&log->l_flush_wait); |
1594 | |
1595 | iclogp = &log->l_iclog; |
1596 | /* |
1597 | * The amount of memory to allocate for the iclog structure is |
1598 | * rather funky due to the way the structure is defined. It is |
1599 | * done this way so that we can use different sizes for machines |
1600 | * with different amounts of memory. See the definition of |
1601 | * xlog_in_core_t in xfs_log_priv.h for details. |
1602 | */ |
1603 | ASSERT(log->l_iclog_size >= 4096); |
1604 | for (i = 0; i < log->l_iclog_bufs; i++) { |
1605 | size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) * |
1606 | sizeof(struct bio_vec); |
1607 | |
1608 | iclog = kzalloc(size: sizeof(*iclog) + bvec_size, |
1609 | GFP_KERNEL | __GFP_RETRY_MAYFAIL); |
1610 | if (!iclog) |
1611 | goto out_free_iclog; |
1612 | |
1613 | *iclogp = iclog; |
1614 | iclog->ic_prev = prev_iclog; |
1615 | prev_iclog = iclog; |
1616 | |
1617 | iclog->ic_data = kvzalloc(size: log->l_iclog_size, |
1618 | GFP_KERNEL | __GFP_RETRY_MAYFAIL); |
1619 | if (!iclog->ic_data) |
1620 | goto out_free_iclog; |
1621 | head = &iclog->ic_header; |
1622 | memset(head, 0, sizeof(xlog_rec_header_t)); |
1623 | head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM); |
1624 | head->h_version = cpu_to_be32( |
1625 | xfs_has_logv2(log->l_mp) ? 2 : 1); |
1626 | head->h_size = cpu_to_be32(log->l_iclog_size); |
1627 | /* new fields */ |
1628 | head->h_fmt = cpu_to_be32(XLOG_FMT); |
1629 | memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t)); |
1630 | |
1631 | iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize; |
1632 | iclog->ic_state = XLOG_STATE_ACTIVE; |
1633 | iclog->ic_log = log; |
1634 | atomic_set(v: &iclog->ic_refcnt, i: 0); |
1635 | INIT_LIST_HEAD(list: &iclog->ic_callbacks); |
1636 | iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize; |
1637 | |
1638 | init_waitqueue_head(&iclog->ic_force_wait); |
1639 | init_waitqueue_head(&iclog->ic_write_wait); |
1640 | INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work); |
1641 | sema_init(sem: &iclog->ic_sema, val: 1); |
1642 | |
1643 | iclogp = &iclog->ic_next; |
1644 | } |
1645 | *iclogp = log->l_iclog; /* complete ring */ |
1646 | log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */ |
1647 | |
1648 | log->l_ioend_workqueue = alloc_workqueue(fmt: "xfs-log/%s" , |
1649 | XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | |
1650 | WQ_HIGHPRI), |
1651 | max_active: 0, mp->m_super->s_id); |
1652 | if (!log->l_ioend_workqueue) |
1653 | goto out_free_iclog; |
1654 | |
1655 | error = xlog_cil_init(log); |
1656 | if (error) |
1657 | goto out_destroy_workqueue; |
1658 | return log; |
1659 | |
1660 | out_destroy_workqueue: |
1661 | destroy_workqueue(wq: log->l_ioend_workqueue); |
1662 | out_free_iclog: |
1663 | for (iclog = log->l_iclog; iclog; iclog = prev_iclog) { |
1664 | prev_iclog = iclog->ic_next; |
1665 | kvfree(addr: iclog->ic_data); |
1666 | kfree(objp: iclog); |
1667 | if (prev_iclog == log->l_iclog) |
1668 | break; |
1669 | } |
1670 | out_free_log: |
1671 | kfree(objp: log); |
1672 | out: |
1673 | return ERR_PTR(error); |
1674 | } /* xlog_alloc_log */ |
1675 | |
1676 | /* |
1677 | * Compute the LSN that we'd need to push the log tail towards in order to have |
1678 | * (a) enough on-disk log space to log the number of bytes specified, (b) at |
1679 | * least 25% of the log space free, and (c) at least 256 blocks free. If the |
1680 | * log free space already meets all three thresholds, this function returns |
1681 | * NULLCOMMITLSN. |
1682 | */ |
1683 | xfs_lsn_t |
1684 | xlog_grant_push_threshold( |
1685 | struct xlog *log, |
1686 | int need_bytes) |
1687 | { |
1688 | xfs_lsn_t threshold_lsn = 0; |
1689 | xfs_lsn_t last_sync_lsn; |
1690 | int free_blocks; |
1691 | int free_bytes; |
1692 | int threshold_block; |
1693 | int threshold_cycle; |
1694 | int free_threshold; |
1695 | |
1696 | ASSERT(BTOBB(need_bytes) < log->l_logBBsize); |
1697 | |
1698 | free_bytes = xlog_space_left(log, head: &log->l_reserve_head.grant); |
1699 | free_blocks = BTOBBT(free_bytes); |
1700 | |
1701 | /* |
1702 | * Set the threshold for the minimum number of free blocks in the |
1703 | * log to the maximum of what the caller needs, one quarter of the |
1704 | * log, and 256 blocks. |
1705 | */ |
1706 | free_threshold = BTOBB(need_bytes); |
1707 | free_threshold = max(free_threshold, (log->l_logBBsize >> 2)); |
1708 | free_threshold = max(free_threshold, 256); |
1709 | if (free_blocks >= free_threshold) |
1710 | return NULLCOMMITLSN; |
1711 | |
1712 | xlog_crack_atomic_lsn(lsn: &log->l_tail_lsn, cycle: &threshold_cycle, |
1713 | block: &threshold_block); |
1714 | threshold_block += free_threshold; |
1715 | if (threshold_block >= log->l_logBBsize) { |
1716 | threshold_block -= log->l_logBBsize; |
1717 | threshold_cycle += 1; |
1718 | } |
1719 | threshold_lsn = xlog_assign_lsn(threshold_cycle, |
1720 | threshold_block); |
1721 | /* |
1722 | * Don't pass in an lsn greater than the lsn of the last |
1723 | * log record known to be on disk. Use a snapshot of the last sync lsn |
1724 | * so that it doesn't change between the compare and the set. |
1725 | */ |
1726 | last_sync_lsn = atomic64_read(&log->l_last_sync_lsn); |
1727 | if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0) |
1728 | threshold_lsn = last_sync_lsn; |
1729 | |
1730 | return threshold_lsn; |
1731 | } |
1732 | |
1733 | /* |
1734 | * Push the tail of the log if we need to do so to maintain the free log space |
1735 | * thresholds set out by xlog_grant_push_threshold. We may need to adopt a |
1736 | * policy which pushes on an lsn which is further along in the log once we |
1737 | * reach the high water mark. In this manner, we would be creating a low water |
1738 | * mark. |
1739 | */ |
1740 | STATIC void |
1741 | xlog_grant_push_ail( |
1742 | struct xlog *log, |
1743 | int need_bytes) |
1744 | { |
1745 | xfs_lsn_t threshold_lsn; |
1746 | |
1747 | threshold_lsn = xlog_grant_push_threshold(log, need_bytes); |
1748 | if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log)) |
1749 | return; |
1750 | |
1751 | /* |
1752 | * Get the transaction layer to kick the dirty buffers out to |
1753 | * disk asynchronously. No point in trying to do this if |
1754 | * the filesystem is shutting down. |
1755 | */ |
1756 | xfs_ail_push(log->l_ailp, threshold_lsn); |
1757 | } |
1758 | |
1759 | /* |
1760 | * Stamp cycle number in every block |
1761 | */ |
1762 | STATIC void |
1763 | xlog_pack_data( |
1764 | struct xlog *log, |
1765 | struct xlog_in_core *iclog, |
1766 | int roundoff) |
1767 | { |
1768 | int i, j, k; |
1769 | int size = iclog->ic_offset + roundoff; |
1770 | __be32 cycle_lsn; |
1771 | char *dp; |
1772 | |
1773 | cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn); |
1774 | |
1775 | dp = iclog->ic_datap; |
1776 | for (i = 0; i < BTOBB(size); i++) { |
1777 | if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) |
1778 | break; |
1779 | iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp; |
1780 | *(__be32 *)dp = cycle_lsn; |
1781 | dp += BBSIZE; |
1782 | } |
1783 | |
1784 | if (xfs_has_logv2(mp: log->l_mp)) { |
1785 | xlog_in_core_2_t *xhdr = iclog->ic_data; |
1786 | |
1787 | for ( ; i < BTOBB(size); i++) { |
1788 | j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
1789 | k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
1790 | xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp; |
1791 | *(__be32 *)dp = cycle_lsn; |
1792 | dp += BBSIZE; |
1793 | } |
1794 | |
1795 | for (i = 1; i < log->l_iclog_heads; i++) |
1796 | xhdr[i].hic_xheader.xh_cycle = cycle_lsn; |
1797 | } |
1798 | } |
1799 | |
1800 | /* |
1801 | * Calculate the checksum for a log buffer. |
1802 | * |
1803 | * This is a little more complicated than it should be because the various |
1804 | * headers and the actual data are non-contiguous. |
1805 | */ |
1806 | __le32 |
1807 | xlog_cksum( |
1808 | struct xlog *log, |
1809 | struct xlog_rec_header *rhead, |
1810 | char *dp, |
1811 | int size) |
1812 | { |
1813 | uint32_t crc; |
1814 | |
1815 | /* first generate the crc for the record header ... */ |
1816 | crc = xfs_start_cksum_update((char *)rhead, |
1817 | sizeof(struct xlog_rec_header), |
1818 | offsetof(struct xlog_rec_header, h_crc)); |
1819 | |
1820 | /* ... then for additional cycle data for v2 logs ... */ |
1821 | if (xfs_has_logv2(mp: log->l_mp)) { |
1822 | union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead; |
1823 | int i; |
1824 | int xheads; |
1825 | |
1826 | xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE); |
1827 | |
1828 | for (i = 1; i < xheads; i++) { |
1829 | crc = crc32c(crc, &xhdr[i].hic_xheader, |
1830 | sizeof(struct xlog_rec_ext_header)); |
1831 | } |
1832 | } |
1833 | |
1834 | /* ... and finally for the payload */ |
1835 | crc = crc32c(crc, address: dp, length: size); |
1836 | |
1837 | return xfs_end_cksum(crc); |
1838 | } |
1839 | |
1840 | static void |
1841 | xlog_bio_end_io( |
1842 | struct bio *bio) |
1843 | { |
1844 | struct xlog_in_core *iclog = bio->bi_private; |
1845 | |
1846 | queue_work(wq: iclog->ic_log->l_ioend_workqueue, |
1847 | work: &iclog->ic_end_io_work); |
1848 | } |
1849 | |
1850 | static int |
1851 | xlog_map_iclog_data( |
1852 | struct bio *bio, |
1853 | void *data, |
1854 | size_t count) |
1855 | { |
1856 | do { |
1857 | struct page *page = kmem_to_page(addr: data); |
1858 | unsigned int off = offset_in_page(data); |
1859 | size_t len = min_t(size_t, count, PAGE_SIZE - off); |
1860 | |
1861 | if (bio_add_page(bio, page, len, off) != len) |
1862 | return -EIO; |
1863 | |
1864 | data += len; |
1865 | count -= len; |
1866 | } while (count); |
1867 | |
1868 | return 0; |
1869 | } |
1870 | |
1871 | STATIC void |
1872 | xlog_write_iclog( |
1873 | struct xlog *log, |
1874 | struct xlog_in_core *iclog, |
1875 | uint64_t bno, |
1876 | unsigned int count) |
1877 | { |
1878 | ASSERT(bno < log->l_logBBsize); |
1879 | trace_xlog_iclog_write(iclog, _RET_IP_); |
1880 | |
1881 | /* |
1882 | * We lock the iclogbufs here so that we can serialise against I/O |
1883 | * completion during unmount. We might be processing a shutdown |
1884 | * triggered during unmount, and that can occur asynchronously to the |
1885 | * unmount thread, and hence we need to ensure that completes before |
1886 | * tearing down the iclogbufs. Hence we need to hold the buffer lock |
1887 | * across the log IO to archieve that. |
1888 | */ |
1889 | down(sem: &iclog->ic_sema); |
1890 | if (xlog_is_shutdown(log)) { |
1891 | /* |
1892 | * It would seem logical to return EIO here, but we rely on |
1893 | * the log state machine to propagate I/O errors instead of |
1894 | * doing it here. We kick of the state machine and unlock |
1895 | * the buffer manually, the code needs to be kept in sync |
1896 | * with the I/O completion path. |
1897 | */ |
1898 | goto sync; |
1899 | } |
1900 | |
1901 | /* |
1902 | * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more |
1903 | * IOs coming immediately after this one. This prevents the block layer |
1904 | * writeback throttle from throttling log writes behind background |
1905 | * metadata writeback and causing priority inversions. |
1906 | */ |
1907 | bio_init(bio: &iclog->ic_bio, bdev: log->l_targ->bt_bdev, table: iclog->ic_bvec, |
1908 | howmany(count, PAGE_SIZE), |
1909 | opf: REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE); |
1910 | iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno; |
1911 | iclog->ic_bio.bi_end_io = xlog_bio_end_io; |
1912 | iclog->ic_bio.bi_private = iclog; |
1913 | |
1914 | if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) { |
1915 | iclog->ic_bio.bi_opf |= REQ_PREFLUSH; |
1916 | /* |
1917 | * For external log devices, we also need to flush the data |
1918 | * device cache first to ensure all metadata writeback covered |
1919 | * by the LSN in this iclog is on stable storage. This is slow, |
1920 | * but it *must* complete before we issue the external log IO. |
1921 | * |
1922 | * If the flush fails, we cannot conclude that past metadata |
1923 | * writeback from the log succeeded. Repeating the flush is |
1924 | * not possible, hence we must shut down with log IO error to |
1925 | * avoid shutdown re-entering this path and erroring out again. |
1926 | */ |
1927 | if (log->l_targ != log->l_mp->m_ddev_targp && |
1928 | blkdev_issue_flush(bdev: log->l_mp->m_ddev_targp->bt_bdev)) |
1929 | goto shutdown; |
1930 | } |
1931 | if (iclog->ic_flags & XLOG_ICL_NEED_FUA) |
1932 | iclog->ic_bio.bi_opf |= REQ_FUA; |
1933 | |
1934 | iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA); |
1935 | |
1936 | if (xlog_map_iclog_data(bio: &iclog->ic_bio, data: iclog->ic_data, count)) |
1937 | goto shutdown; |
1938 | |
1939 | if (is_vmalloc_addr(x: iclog->ic_data)) |
1940 | flush_kernel_vmap_range(vaddr: iclog->ic_data, size: count); |
1941 | |
1942 | /* |
1943 | * If this log buffer would straddle the end of the log we will have |
1944 | * to split it up into two bios, so that we can continue at the start. |
1945 | */ |
1946 | if (bno + BTOBB(count) > log->l_logBBsize) { |
1947 | struct bio *split; |
1948 | |
1949 | split = bio_split(bio: &iclog->ic_bio, sectors: log->l_logBBsize - bno, |
1950 | GFP_NOIO, bs: &fs_bio_set); |
1951 | bio_chain(split, &iclog->ic_bio); |
1952 | submit_bio(bio: split); |
1953 | |
1954 | /* restart at logical offset zero for the remainder */ |
1955 | iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart; |
1956 | } |
1957 | |
1958 | submit_bio(bio: &iclog->ic_bio); |
1959 | return; |
1960 | shutdown: |
1961 | xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); |
1962 | sync: |
1963 | xlog_state_done_syncing(iclog); |
1964 | up(sem: &iclog->ic_sema); |
1965 | } |
1966 | |
1967 | /* |
1968 | * We need to bump cycle number for the part of the iclog that is |
1969 | * written to the start of the log. Watch out for the header magic |
1970 | * number case, though. |
1971 | */ |
1972 | static void |
1973 | xlog_split_iclog( |
1974 | struct xlog *log, |
1975 | void *data, |
1976 | uint64_t bno, |
1977 | unsigned int count) |
1978 | { |
1979 | unsigned int split_offset = BBTOB(log->l_logBBsize - bno); |
1980 | unsigned int i; |
1981 | |
1982 | for (i = split_offset; i < count; i += BBSIZE) { |
1983 | uint32_t cycle = get_unaligned_be32(p: data + i); |
1984 | |
1985 | if (++cycle == XLOG_HEADER_MAGIC_NUM) |
1986 | cycle++; |
1987 | put_unaligned_be32(val: cycle, p: data + i); |
1988 | } |
1989 | } |
1990 | |
1991 | static int |
1992 | xlog_calc_iclog_size( |
1993 | struct xlog *log, |
1994 | struct xlog_in_core *iclog, |
1995 | uint32_t *roundoff) |
1996 | { |
1997 | uint32_t count_init, count; |
1998 | |
1999 | /* Add for LR header */ |
2000 | count_init = log->l_iclog_hsize + iclog->ic_offset; |
2001 | count = roundup(count_init, log->l_iclog_roundoff); |
2002 | |
2003 | *roundoff = count - count_init; |
2004 | |
2005 | ASSERT(count >= count_init); |
2006 | ASSERT(*roundoff < log->l_iclog_roundoff); |
2007 | return count; |
2008 | } |
2009 | |
2010 | /* |
2011 | * Flush out the in-core log (iclog) to the on-disk log in an asynchronous |
2012 | * fashion. Previously, we should have moved the current iclog |
2013 | * ptr in the log to point to the next available iclog. This allows further |
2014 | * write to continue while this code syncs out an iclog ready to go. |
2015 | * Before an in-core log can be written out, the data section must be scanned |
2016 | * to save away the 1st word of each BBSIZE block into the header. We replace |
2017 | * it with the current cycle count. Each BBSIZE block is tagged with the |
2018 | * cycle count because there in an implicit assumption that drives will |
2019 | * guarantee that entire 512 byte blocks get written at once. In other words, |
2020 | * we can't have part of a 512 byte block written and part not written. By |
2021 | * tagging each block, we will know which blocks are valid when recovering |
2022 | * after an unclean shutdown. |
2023 | * |
2024 | * This routine is single threaded on the iclog. No other thread can be in |
2025 | * this routine with the same iclog. Changing contents of iclog can there- |
2026 | * fore be done without grabbing the state machine lock. Updating the global |
2027 | * log will require grabbing the lock though. |
2028 | * |
2029 | * The entire log manager uses a logical block numbering scheme. Only |
2030 | * xlog_write_iclog knows about the fact that the log may not start with |
2031 | * block zero on a given device. |
2032 | */ |
2033 | STATIC void |
2034 | xlog_sync( |
2035 | struct xlog *log, |
2036 | struct xlog_in_core *iclog, |
2037 | struct xlog_ticket *ticket) |
2038 | { |
2039 | unsigned int count; /* byte count of bwrite */ |
2040 | unsigned int roundoff; /* roundoff to BB or stripe */ |
2041 | uint64_t bno; |
2042 | unsigned int size; |
2043 | |
2044 | ASSERT(atomic_read(&iclog->ic_refcnt) == 0); |
2045 | trace_xlog_iclog_sync(iclog, _RET_IP_); |
2046 | |
2047 | count = xlog_calc_iclog_size(log, iclog, roundoff: &roundoff); |
2048 | |
2049 | /* |
2050 | * If we have a ticket, account for the roundoff via the ticket |
2051 | * reservation to avoid touching the hot grant heads needlessly. |
2052 | * Otherwise, we have to move grant heads directly. |
2053 | */ |
2054 | if (ticket) { |
2055 | ticket->t_curr_res -= roundoff; |
2056 | } else { |
2057 | xlog_grant_add_space(log, head: &log->l_reserve_head.grant, bytes: roundoff); |
2058 | xlog_grant_add_space(log, head: &log->l_write_head.grant, bytes: roundoff); |
2059 | } |
2060 | |
2061 | /* put cycle number in every block */ |
2062 | xlog_pack_data(log, iclog, roundoff); |
2063 | |
2064 | /* real byte length */ |
2065 | size = iclog->ic_offset; |
2066 | if (xfs_has_logv2(mp: log->l_mp)) |
2067 | size += roundoff; |
2068 | iclog->ic_header.h_len = cpu_to_be32(size); |
2069 | |
2070 | XFS_STATS_INC(log->l_mp, xs_log_writes); |
2071 | XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count)); |
2072 | |
2073 | bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)); |
2074 | |
2075 | /* Do we need to split this write into 2 parts? */ |
2076 | if (bno + BTOBB(count) > log->l_logBBsize) |
2077 | xlog_split_iclog(log, data: &iclog->ic_header, bno, count); |
2078 | |
2079 | /* calculcate the checksum */ |
2080 | iclog->ic_header.h_crc = xlog_cksum(log, rhead: &iclog->ic_header, |
2081 | dp: iclog->ic_datap, size); |
2082 | /* |
2083 | * Intentionally corrupt the log record CRC based on the error injection |
2084 | * frequency, if defined. This facilitates testing log recovery in the |
2085 | * event of torn writes. Hence, set the IOABORT state to abort the log |
2086 | * write on I/O completion and shutdown the fs. The subsequent mount |
2087 | * detects the bad CRC and attempts to recover. |
2088 | */ |
2089 | #ifdef DEBUG |
2090 | if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) { |
2091 | iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA); |
2092 | iclog->ic_fail_crc = true; |
2093 | xfs_warn(log->l_mp, |
2094 | "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent." , |
2095 | be64_to_cpu(iclog->ic_header.h_lsn)); |
2096 | } |
2097 | #endif |
2098 | xlog_verify_iclog(log, iclog, count); |
2099 | xlog_write_iclog(log, iclog, bno, count); |
2100 | } |
2101 | |
2102 | /* |
2103 | * Deallocate a log structure |
2104 | */ |
2105 | STATIC void |
2106 | xlog_dealloc_log( |
2107 | struct xlog *log) |
2108 | { |
2109 | xlog_in_core_t *iclog, *next_iclog; |
2110 | int i; |
2111 | |
2112 | /* |
2113 | * Destroy the CIL after waiting for iclog IO completion because an |
2114 | * iclog EIO error will try to shut down the log, which accesses the |
2115 | * CIL to wake up the waiters. |
2116 | */ |
2117 | xlog_cil_destroy(log); |
2118 | |
2119 | iclog = log->l_iclog; |
2120 | for (i = 0; i < log->l_iclog_bufs; i++) { |
2121 | next_iclog = iclog->ic_next; |
2122 | kvfree(addr: iclog->ic_data); |
2123 | kfree(objp: iclog); |
2124 | iclog = next_iclog; |
2125 | } |
2126 | |
2127 | log->l_mp->m_log = NULL; |
2128 | destroy_workqueue(wq: log->l_ioend_workqueue); |
2129 | kfree(objp: log); |
2130 | } |
2131 | |
2132 | /* |
2133 | * Update counters atomically now that memcpy is done. |
2134 | */ |
2135 | static inline void |
2136 | xlog_state_finish_copy( |
2137 | struct xlog *log, |
2138 | struct xlog_in_core *iclog, |
2139 | int record_cnt, |
2140 | int copy_bytes) |
2141 | { |
2142 | lockdep_assert_held(&log->l_icloglock); |
2143 | |
2144 | be32_add_cpu(var: &iclog->ic_header.h_num_logops, val: record_cnt); |
2145 | iclog->ic_offset += copy_bytes; |
2146 | } |
2147 | |
2148 | /* |
2149 | * print out info relating to regions written which consume |
2150 | * the reservation |
2151 | */ |
2152 | void |
2153 | xlog_print_tic_res( |
2154 | struct xfs_mount *mp, |
2155 | struct xlog_ticket *ticket) |
2156 | { |
2157 | xfs_warn(mp, "ticket reservation summary:" ); |
2158 | xfs_warn(mp, " unit res = %d bytes" , ticket->t_unit_res); |
2159 | xfs_warn(mp, " current res = %d bytes" , ticket->t_curr_res); |
2160 | xfs_warn(mp, " original count = %d" , ticket->t_ocnt); |
2161 | xfs_warn(mp, " remaining count = %d" , ticket->t_cnt); |
2162 | } |
2163 | |
2164 | /* |
2165 | * Print a summary of the transaction. |
2166 | */ |
2167 | void |
2168 | xlog_print_trans( |
2169 | struct xfs_trans *tp) |
2170 | { |
2171 | struct xfs_mount *mp = tp->t_mountp; |
2172 | struct xfs_log_item *lip; |
2173 | |
2174 | /* dump core transaction and ticket info */ |
2175 | xfs_warn(mp, "transaction summary:" ); |
2176 | xfs_warn(mp, " log res = %d" , tp->t_log_res); |
2177 | xfs_warn(mp, " log count = %d" , tp->t_log_count); |
2178 | xfs_warn(mp, " flags = 0x%x" , tp->t_flags); |
2179 | |
2180 | xlog_print_tic_res(mp, ticket: tp->t_ticket); |
2181 | |
2182 | /* dump each log item */ |
2183 | list_for_each_entry(lip, &tp->t_items, li_trans) { |
2184 | struct xfs_log_vec *lv = lip->li_lv; |
2185 | struct xfs_log_iovec *vec; |
2186 | int i; |
2187 | |
2188 | xfs_warn(mp, "log item: " ); |
2189 | xfs_warn(mp, " type = 0x%x" , lip->li_type); |
2190 | xfs_warn(mp, " flags = 0x%lx" , lip->li_flags); |
2191 | if (!lv) |
2192 | continue; |
2193 | xfs_warn(mp, " niovecs = %d" , lv->lv_niovecs); |
2194 | xfs_warn(mp, " size = %d" , lv->lv_size); |
2195 | xfs_warn(mp, " bytes = %d" , lv->lv_bytes); |
2196 | xfs_warn(mp, " buf len = %d" , lv->lv_buf_len); |
2197 | |
2198 | /* dump each iovec for the log item */ |
2199 | vec = lv->lv_iovecp; |
2200 | for (i = 0; i < lv->lv_niovecs; i++) { |
2201 | int dumplen = min(vec->i_len, 32); |
2202 | |
2203 | xfs_warn(mp, " iovec[%d]" , i); |
2204 | xfs_warn(mp, " type = 0x%x" , vec->i_type); |
2205 | xfs_warn(mp, " len = %d" , vec->i_len); |
2206 | xfs_warn(mp, " first %d bytes of iovec[%d]:" , dumplen, i); |
2207 | xfs_hex_dump(p: vec->i_addr, length: dumplen); |
2208 | |
2209 | vec++; |
2210 | } |
2211 | } |
2212 | } |
2213 | |
2214 | static inline void |
2215 | xlog_write_iovec( |
2216 | struct xlog_in_core *iclog, |
2217 | uint32_t *log_offset, |
2218 | void *data, |
2219 | uint32_t write_len, |
2220 | int *bytes_left, |
2221 | uint32_t *record_cnt, |
2222 | uint32_t *data_cnt) |
2223 | { |
2224 | ASSERT(*log_offset < iclog->ic_log->l_iclog_size); |
2225 | ASSERT(*log_offset % sizeof(int32_t) == 0); |
2226 | ASSERT(write_len % sizeof(int32_t) == 0); |
2227 | |
2228 | memcpy(iclog->ic_datap + *log_offset, data, write_len); |
2229 | *log_offset += write_len; |
2230 | *bytes_left -= write_len; |
2231 | (*record_cnt)++; |
2232 | *data_cnt += write_len; |
2233 | } |
2234 | |
2235 | /* |
2236 | * Write log vectors into a single iclog which is guaranteed by the caller |
2237 | * to have enough space to write the entire log vector into. |
2238 | */ |
2239 | static void |
2240 | xlog_write_full( |
2241 | struct xfs_log_vec *lv, |
2242 | struct xlog_ticket *ticket, |
2243 | struct xlog_in_core *iclog, |
2244 | uint32_t *log_offset, |
2245 | uint32_t *len, |
2246 | uint32_t *record_cnt, |
2247 | uint32_t *data_cnt) |
2248 | { |
2249 | int index; |
2250 | |
2251 | ASSERT(*log_offset + *len <= iclog->ic_size || |
2252 | iclog->ic_state == XLOG_STATE_WANT_SYNC); |
2253 | |
2254 | /* |
2255 | * Ordered log vectors have no regions to write so this |
2256 | * loop will naturally skip them. |
2257 | */ |
2258 | for (index = 0; index < lv->lv_niovecs; index++) { |
2259 | struct xfs_log_iovec *reg = &lv->lv_iovecp[index]; |
2260 | struct *ophdr = reg->i_addr; |
2261 | |
2262 | ophdr->oh_tid = cpu_to_be32(ticket->t_tid); |
2263 | xlog_write_iovec(iclog, log_offset, data: reg->i_addr, |
2264 | write_len: reg->i_len, bytes_left: len, record_cnt, data_cnt); |
2265 | } |
2266 | } |
2267 | |
2268 | static int |
2269 | xlog_write_get_more_iclog_space( |
2270 | struct xlog_ticket *ticket, |
2271 | struct xlog_in_core **iclogp, |
2272 | uint32_t *log_offset, |
2273 | uint32_t len, |
2274 | uint32_t *record_cnt, |
2275 | uint32_t *data_cnt) |
2276 | { |
2277 | struct xlog_in_core *iclog = *iclogp; |
2278 | struct xlog *log = iclog->ic_log; |
2279 | int error; |
2280 | |
2281 | spin_lock(lock: &log->l_icloglock); |
2282 | ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC); |
2283 | xlog_state_finish_copy(log, iclog, record_cnt: *record_cnt, copy_bytes: *data_cnt); |
2284 | error = xlog_state_release_iclog(log, iclog, ticket); |
2285 | spin_unlock(lock: &log->l_icloglock); |
2286 | if (error) |
2287 | return error; |
2288 | |
2289 | error = xlog_state_get_iclog_space(log, len, iclog: &iclog, ticket, |
2290 | logoffsetp: log_offset); |
2291 | if (error) |
2292 | return error; |
2293 | *record_cnt = 0; |
2294 | *data_cnt = 0; |
2295 | *iclogp = iclog; |
2296 | return 0; |
2297 | } |
2298 | |
2299 | /* |
2300 | * Write log vectors into a single iclog which is smaller than the current chain |
2301 | * length. We write until we cannot fit a full record into the remaining space |
2302 | * and then stop. We return the log vector that is to be written that cannot |
2303 | * wholly fit in the iclog. |
2304 | */ |
2305 | static int |
2306 | xlog_write_partial( |
2307 | struct xfs_log_vec *lv, |
2308 | struct xlog_ticket *ticket, |
2309 | struct xlog_in_core **iclogp, |
2310 | uint32_t *log_offset, |
2311 | uint32_t *len, |
2312 | uint32_t *record_cnt, |
2313 | uint32_t *data_cnt) |
2314 | { |
2315 | struct xlog_in_core *iclog = *iclogp; |
2316 | struct *ophdr; |
2317 | int index = 0; |
2318 | uint32_t rlen; |
2319 | int error; |
2320 | |
2321 | /* walk the logvec, copying until we run out of space in the iclog */ |
2322 | for (index = 0; index < lv->lv_niovecs; index++) { |
2323 | struct xfs_log_iovec *reg = &lv->lv_iovecp[index]; |
2324 | uint32_t reg_offset = 0; |
2325 | |
2326 | /* |
2327 | * The first region of a continuation must have a non-zero |
2328 | * length otherwise log recovery will just skip over it and |
2329 | * start recovering from the next opheader it finds. Because we |
2330 | * mark the next opheader as a continuation, recovery will then |
2331 | * incorrectly add the continuation to the previous region and |
2332 | * that breaks stuff. |
2333 | * |
2334 | * Hence if there isn't space for region data after the |
2335 | * opheader, then we need to start afresh with a new iclog. |
2336 | */ |
2337 | if (iclog->ic_size - *log_offset <= |
2338 | sizeof(struct xlog_op_header)) { |
2339 | error = xlog_write_get_more_iclog_space(ticket, |
2340 | iclogp: &iclog, log_offset, len: *len, record_cnt, |
2341 | data_cnt); |
2342 | if (error) |
2343 | return error; |
2344 | } |
2345 | |
2346 | ophdr = reg->i_addr; |
2347 | rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset); |
2348 | |
2349 | ophdr->oh_tid = cpu_to_be32(ticket->t_tid); |
2350 | ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header)); |
2351 | if (rlen != reg->i_len) |
2352 | ophdr->oh_flags |= XLOG_CONTINUE_TRANS; |
2353 | |
2354 | xlog_write_iovec(iclog, log_offset, data: reg->i_addr, |
2355 | write_len: rlen, bytes_left: len, record_cnt, data_cnt); |
2356 | |
2357 | /* If we wrote the whole region, move to the next. */ |
2358 | if (rlen == reg->i_len) |
2359 | continue; |
2360 | |
2361 | /* |
2362 | * We now have a partially written iovec, but it can span |
2363 | * multiple iclogs so we loop here. First we release the iclog |
2364 | * we currently have, then we get a new iclog and add a new |
2365 | * opheader. Then we continue copying from where we were until |
2366 | * we either complete the iovec or fill the iclog. If we |
2367 | * complete the iovec, then we increment the index and go right |
2368 | * back to the top of the outer loop. if we fill the iclog, we |
2369 | * run the inner loop again. |
2370 | * |
2371 | * This is complicated by the tail of a region using all the |
2372 | * space in an iclog and hence requiring us to release the iclog |
2373 | * and get a new one before returning to the outer loop. We must |
2374 | * always guarantee that we exit this inner loop with at least |
2375 | * space for log transaction opheaders left in the current |
2376 | * iclog, hence we cannot just terminate the loop at the end |
2377 | * of the of the continuation. So we loop while there is no |
2378 | * space left in the current iclog, and check for the end of the |
2379 | * continuation after getting a new iclog. |
2380 | */ |
2381 | do { |
2382 | /* |
2383 | * Ensure we include the continuation opheader in the |
2384 | * space we need in the new iclog by adding that size |
2385 | * to the length we require. This continuation opheader |
2386 | * needs to be accounted to the ticket as the space it |
2387 | * consumes hasn't been accounted to the lv we are |
2388 | * writing. |
2389 | */ |
2390 | error = xlog_write_get_more_iclog_space(ticket, |
2391 | &iclog, log_offset, |
2392 | *len + sizeof(struct xlog_op_header), |
2393 | record_cnt, data_cnt); |
2394 | if (error) |
2395 | return error; |
2396 | |
2397 | ophdr = iclog->ic_datap + *log_offset; |
2398 | ophdr->oh_tid = cpu_to_be32(ticket->t_tid); |
2399 | ophdr->oh_clientid = XFS_TRANSACTION; |
2400 | ophdr->oh_res2 = 0; |
2401 | ophdr->oh_flags = XLOG_WAS_CONT_TRANS; |
2402 | |
2403 | ticket->t_curr_res -= sizeof(struct xlog_op_header); |
2404 | *log_offset += sizeof(struct xlog_op_header); |
2405 | *data_cnt += sizeof(struct xlog_op_header); |
2406 | |
2407 | /* |
2408 | * If rlen fits in the iclog, then end the region |
2409 | * continuation. Otherwise we're going around again. |
2410 | */ |
2411 | reg_offset += rlen; |
2412 | rlen = reg->i_len - reg_offset; |
2413 | if (rlen <= iclog->ic_size - *log_offset) |
2414 | ophdr->oh_flags |= XLOG_END_TRANS; |
2415 | else |
2416 | ophdr->oh_flags |= XLOG_CONTINUE_TRANS; |
2417 | |
2418 | rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset); |
2419 | ophdr->oh_len = cpu_to_be32(rlen); |
2420 | |
2421 | xlog_write_iovec(iclog, log_offset, |
2422 | data: reg->i_addr + reg_offset, |
2423 | write_len: rlen, bytes_left: len, record_cnt, data_cnt); |
2424 | |
2425 | } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS); |
2426 | } |
2427 | |
2428 | /* |
2429 | * No more iovecs remain in this logvec so return the next log vec to |
2430 | * the caller so it can go back to fast path copying. |
2431 | */ |
2432 | *iclogp = iclog; |
2433 | return 0; |
2434 | } |
2435 | |
2436 | /* |
2437 | * Write some region out to in-core log |
2438 | * |
2439 | * This will be called when writing externally provided regions or when |
2440 | * writing out a commit record for a given transaction. |
2441 | * |
2442 | * General algorithm: |
2443 | * 1. Find total length of this write. This may include adding to the |
2444 | * lengths passed in. |
2445 | * 2. Check whether we violate the tickets reservation. |
2446 | * 3. While writing to this iclog |
2447 | * A. Reserve as much space in this iclog as can get |
2448 | * B. If this is first write, save away start lsn |
2449 | * C. While writing this region: |
2450 | * 1. If first write of transaction, write start record |
2451 | * 2. Write log operation header (header per region) |
2452 | * 3. Find out if we can fit entire region into this iclog |
2453 | * 4. Potentially, verify destination memcpy ptr |
2454 | * 5. Memcpy (partial) region |
2455 | * 6. If partial copy, release iclog; otherwise, continue |
2456 | * copying more regions into current iclog |
2457 | * 4. Mark want sync bit (in simulation mode) |
2458 | * 5. Release iclog for potential flush to on-disk log. |
2459 | * |
2460 | * ERRORS: |
2461 | * 1. Panic if reservation is overrun. This should never happen since |
2462 | * reservation amounts are generated internal to the filesystem. |
2463 | * NOTES: |
2464 | * 1. Tickets are single threaded data structures. |
2465 | * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the |
2466 | * syncing routine. When a single log_write region needs to span |
2467 | * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set |
2468 | * on all log operation writes which don't contain the end of the |
2469 | * region. The XLOG_END_TRANS bit is used for the in-core log |
2470 | * operation which contains the end of the continued log_write region. |
2471 | * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog, |
2472 | * we don't really know exactly how much space will be used. As a result, |
2473 | * we don't update ic_offset until the end when we know exactly how many |
2474 | * bytes have been written out. |
2475 | */ |
2476 | int |
2477 | xlog_write( |
2478 | struct xlog *log, |
2479 | struct xfs_cil_ctx *ctx, |
2480 | struct list_head *lv_chain, |
2481 | struct xlog_ticket *ticket, |
2482 | uint32_t len) |
2483 | |
2484 | { |
2485 | struct xlog_in_core *iclog = NULL; |
2486 | struct xfs_log_vec *lv; |
2487 | uint32_t record_cnt = 0; |
2488 | uint32_t data_cnt = 0; |
2489 | int error = 0; |
2490 | int log_offset; |
2491 | |
2492 | if (ticket->t_curr_res < 0) { |
2493 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, |
2494 | "ctx ticket reservation ran out. Need to up reservation" ); |
2495 | xlog_print_tic_res(mp: log->l_mp, ticket); |
2496 | xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); |
2497 | } |
2498 | |
2499 | error = xlog_state_get_iclog_space(log, len, iclog: &iclog, ticket, |
2500 | logoffsetp: &log_offset); |
2501 | if (error) |
2502 | return error; |
2503 | |
2504 | ASSERT(log_offset <= iclog->ic_size - 1); |
2505 | |
2506 | /* |
2507 | * If we have a context pointer, pass it the first iclog we are |
2508 | * writing to so it can record state needed for iclog write |
2509 | * ordering. |
2510 | */ |
2511 | if (ctx) |
2512 | xlog_cil_set_ctx_write_state(ctx, iclog); |
2513 | |
2514 | list_for_each_entry(lv, lv_chain, lv_list) { |
2515 | /* |
2516 | * If the entire log vec does not fit in the iclog, punt it to |
2517 | * the partial copy loop which can handle this case. |
2518 | */ |
2519 | if (lv->lv_niovecs && |
2520 | lv->lv_bytes > iclog->ic_size - log_offset) { |
2521 | error = xlog_write_partial(lv, ticket, iclogp: &iclog, |
2522 | log_offset: &log_offset, len: &len, record_cnt: &record_cnt, |
2523 | data_cnt: &data_cnt); |
2524 | if (error) { |
2525 | /* |
2526 | * We have no iclog to release, so just return |
2527 | * the error immediately. |
2528 | */ |
2529 | return error; |
2530 | } |
2531 | } else { |
2532 | xlog_write_full(lv, ticket, iclog, log_offset: &log_offset, |
2533 | len: &len, record_cnt: &record_cnt, data_cnt: &data_cnt); |
2534 | } |
2535 | } |
2536 | ASSERT(len == 0); |
2537 | |
2538 | /* |
2539 | * We've already been guaranteed that the last writes will fit inside |
2540 | * the current iclog, and hence it will already have the space used by |
2541 | * those writes accounted to it. Hence we do not need to update the |
2542 | * iclog with the number of bytes written here. |
2543 | */ |
2544 | spin_lock(lock: &log->l_icloglock); |
2545 | xlog_state_finish_copy(log, iclog, record_cnt, copy_bytes: 0); |
2546 | error = xlog_state_release_iclog(log, iclog, ticket); |
2547 | spin_unlock(lock: &log->l_icloglock); |
2548 | |
2549 | return error; |
2550 | } |
2551 | |
2552 | static void |
2553 | xlog_state_activate_iclog( |
2554 | struct xlog_in_core *iclog, |
2555 | int *iclogs_changed) |
2556 | { |
2557 | ASSERT(list_empty_careful(&iclog->ic_callbacks)); |
2558 | trace_xlog_iclog_activate(iclog, _RET_IP_); |
2559 | |
2560 | /* |
2561 | * If the number of ops in this iclog indicate it just contains the |
2562 | * dummy transaction, we can change state into IDLE (the second time |
2563 | * around). Otherwise we should change the state into NEED a dummy. |
2564 | * We don't need to cover the dummy. |
2565 | */ |
2566 | if (*iclogs_changed == 0 && |
2567 | iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) { |
2568 | *iclogs_changed = 1; |
2569 | } else { |
2570 | /* |
2571 | * We have two dirty iclogs so start over. This could also be |
2572 | * num of ops indicating this is not the dummy going out. |
2573 | */ |
2574 | *iclogs_changed = 2; |
2575 | } |
2576 | |
2577 | iclog->ic_state = XLOG_STATE_ACTIVE; |
2578 | iclog->ic_offset = 0; |
2579 | iclog->ic_header.h_num_logops = 0; |
2580 | memset(iclog->ic_header.h_cycle_data, 0, |
2581 | sizeof(iclog->ic_header.h_cycle_data)); |
2582 | iclog->ic_header.h_lsn = 0; |
2583 | iclog->ic_header.h_tail_lsn = 0; |
2584 | } |
2585 | |
2586 | /* |
2587 | * Loop through all iclogs and mark all iclogs currently marked DIRTY as |
2588 | * ACTIVE after iclog I/O has completed. |
2589 | */ |
2590 | static void |
2591 | xlog_state_activate_iclogs( |
2592 | struct xlog *log, |
2593 | int *iclogs_changed) |
2594 | { |
2595 | struct xlog_in_core *iclog = log->l_iclog; |
2596 | |
2597 | do { |
2598 | if (iclog->ic_state == XLOG_STATE_DIRTY) |
2599 | xlog_state_activate_iclog(iclog, iclogs_changed); |
2600 | /* |
2601 | * The ordering of marking iclogs ACTIVE must be maintained, so |
2602 | * an iclog doesn't become ACTIVE beyond one that is SYNCING. |
2603 | */ |
2604 | else if (iclog->ic_state != XLOG_STATE_ACTIVE) |
2605 | break; |
2606 | } while ((iclog = iclog->ic_next) != log->l_iclog); |
2607 | } |
2608 | |
2609 | static int |
2610 | xlog_covered_state( |
2611 | int prev_state, |
2612 | int iclogs_changed) |
2613 | { |
2614 | /* |
2615 | * We go to NEED for any non-covering writes. We go to NEED2 if we just |
2616 | * wrote the first covering record (DONE). We go to IDLE if we just |
2617 | * wrote the second covering record (DONE2) and remain in IDLE until a |
2618 | * non-covering write occurs. |
2619 | */ |
2620 | switch (prev_state) { |
2621 | case XLOG_STATE_COVER_IDLE: |
2622 | if (iclogs_changed == 1) |
2623 | return XLOG_STATE_COVER_IDLE; |
2624 | fallthrough; |
2625 | case XLOG_STATE_COVER_NEED: |
2626 | case XLOG_STATE_COVER_NEED2: |
2627 | break; |
2628 | case XLOG_STATE_COVER_DONE: |
2629 | if (iclogs_changed == 1) |
2630 | return XLOG_STATE_COVER_NEED2; |
2631 | break; |
2632 | case XLOG_STATE_COVER_DONE2: |
2633 | if (iclogs_changed == 1) |
2634 | return XLOG_STATE_COVER_IDLE; |
2635 | break; |
2636 | default: |
2637 | ASSERT(0); |
2638 | } |
2639 | |
2640 | return XLOG_STATE_COVER_NEED; |
2641 | } |
2642 | |
2643 | STATIC void |
2644 | xlog_state_clean_iclog( |
2645 | struct xlog *log, |
2646 | struct xlog_in_core *dirty_iclog) |
2647 | { |
2648 | int iclogs_changed = 0; |
2649 | |
2650 | trace_xlog_iclog_clean(iclog: dirty_iclog, _RET_IP_); |
2651 | |
2652 | dirty_iclog->ic_state = XLOG_STATE_DIRTY; |
2653 | |
2654 | xlog_state_activate_iclogs(log, iclogs_changed: &iclogs_changed); |
2655 | wake_up_all(&dirty_iclog->ic_force_wait); |
2656 | |
2657 | if (iclogs_changed) { |
2658 | log->l_covered_state = xlog_covered_state(prev_state: log->l_covered_state, |
2659 | iclogs_changed); |
2660 | } |
2661 | } |
2662 | |
2663 | STATIC xfs_lsn_t |
2664 | xlog_get_lowest_lsn( |
2665 | struct xlog *log) |
2666 | { |
2667 | struct xlog_in_core *iclog = log->l_iclog; |
2668 | xfs_lsn_t lowest_lsn = 0, lsn; |
2669 | |
2670 | do { |
2671 | if (iclog->ic_state == XLOG_STATE_ACTIVE || |
2672 | iclog->ic_state == XLOG_STATE_DIRTY) |
2673 | continue; |
2674 | |
2675 | lsn = be64_to_cpu(iclog->ic_header.h_lsn); |
2676 | if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0) |
2677 | lowest_lsn = lsn; |
2678 | } while ((iclog = iclog->ic_next) != log->l_iclog); |
2679 | |
2680 | return lowest_lsn; |
2681 | } |
2682 | |
2683 | /* |
2684 | * Completion of a iclog IO does not imply that a transaction has completed, as |
2685 | * transactions can be large enough to span many iclogs. We cannot change the |
2686 | * tail of the log half way through a transaction as this may be the only |
2687 | * transaction in the log and moving the tail to point to the middle of it |
2688 | * will prevent recovery from finding the start of the transaction. Hence we |
2689 | * should only update the last_sync_lsn if this iclog contains transaction |
2690 | * completion callbacks on it. |
2691 | * |
2692 | * We have to do this before we drop the icloglock to ensure we are the only one |
2693 | * that can update it. |
2694 | * |
2695 | * If we are moving the last_sync_lsn forwards, we also need to ensure we kick |
2696 | * the reservation grant head pushing. This is due to the fact that the push |
2697 | * target is bound by the current last_sync_lsn value. Hence if we have a large |
2698 | * amount of log space bound up in this committing transaction then the |
2699 | * last_sync_lsn value may be the limiting factor preventing tail pushing from |
2700 | * freeing space in the log. Hence once we've updated the last_sync_lsn we |
2701 | * should push the AIL to ensure the push target (and hence the grant head) is |
2702 | * no longer bound by the old log head location and can move forwards and make |
2703 | * progress again. |
2704 | */ |
2705 | static void |
2706 | xlog_state_set_callback( |
2707 | struct xlog *log, |
2708 | struct xlog_in_core *iclog, |
2709 | xfs_lsn_t ) |
2710 | { |
2711 | trace_xlog_iclog_callback(iclog, _RET_IP_); |
2712 | iclog->ic_state = XLOG_STATE_CALLBACK; |
2713 | |
2714 | ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn), |
2715 | header_lsn) <= 0); |
2716 | |
2717 | if (list_empty_careful(head: &iclog->ic_callbacks)) |
2718 | return; |
2719 | |
2720 | atomic64_set(v: &log->l_last_sync_lsn, i: header_lsn); |
2721 | xlog_grant_push_ail(log, need_bytes: 0); |
2722 | } |
2723 | |
2724 | /* |
2725 | * Return true if we need to stop processing, false to continue to the next |
2726 | * iclog. The caller will need to run callbacks if the iclog is returned in the |
2727 | * XLOG_STATE_CALLBACK state. |
2728 | */ |
2729 | static bool |
2730 | xlog_state_iodone_process_iclog( |
2731 | struct xlog *log, |
2732 | struct xlog_in_core *iclog) |
2733 | { |
2734 | xfs_lsn_t lowest_lsn; |
2735 | xfs_lsn_t header_lsn; |
2736 | |
2737 | switch (iclog->ic_state) { |
2738 | case XLOG_STATE_ACTIVE: |
2739 | case XLOG_STATE_DIRTY: |
2740 | /* |
2741 | * Skip all iclogs in the ACTIVE & DIRTY states: |
2742 | */ |
2743 | return false; |
2744 | case XLOG_STATE_DONE_SYNC: |
2745 | /* |
2746 | * Now that we have an iclog that is in the DONE_SYNC state, do |
2747 | * one more check here to see if we have chased our tail around. |
2748 | * If this is not the lowest lsn iclog, then we will leave it |
2749 | * for another completion to process. |
2750 | */ |
2751 | header_lsn = be64_to_cpu(iclog->ic_header.h_lsn); |
2752 | lowest_lsn = xlog_get_lowest_lsn(log); |
2753 | if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0) |
2754 | return false; |
2755 | xlog_state_set_callback(log, iclog, header_lsn); |
2756 | return false; |
2757 | default: |
2758 | /* |
2759 | * Can only perform callbacks in order. Since this iclog is not |
2760 | * in the DONE_SYNC state, we skip the rest and just try to |
2761 | * clean up. |
2762 | */ |
2763 | return true; |
2764 | } |
2765 | } |
2766 | |
2767 | /* |
2768 | * Loop over all the iclogs, running attached callbacks on them. Return true if |
2769 | * we ran any callbacks, indicating that we dropped the icloglock. We don't need |
2770 | * to handle transient shutdown state here at all because |
2771 | * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown |
2772 | * cleanup of the callbacks. |
2773 | */ |
2774 | static bool |
2775 | xlog_state_do_iclog_callbacks( |
2776 | struct xlog *log) |
2777 | __releases(&log->l_icloglock) |
2778 | __acquires(&log->l_icloglock) |
2779 | { |
2780 | struct xlog_in_core *first_iclog = log->l_iclog; |
2781 | struct xlog_in_core *iclog = first_iclog; |
2782 | bool ran_callback = false; |
2783 | |
2784 | do { |
2785 | LIST_HEAD(cb_list); |
2786 | |
2787 | if (xlog_state_iodone_process_iclog(log, iclog)) |
2788 | break; |
2789 | if (iclog->ic_state != XLOG_STATE_CALLBACK) { |
2790 | iclog = iclog->ic_next; |
2791 | continue; |
2792 | } |
2793 | list_splice_init(list: &iclog->ic_callbacks, head: &cb_list); |
2794 | spin_unlock(lock: &log->l_icloglock); |
2795 | |
2796 | trace_xlog_iclog_callbacks_start(iclog, _RET_IP_); |
2797 | xlog_cil_process_committed(list: &cb_list); |
2798 | trace_xlog_iclog_callbacks_done(iclog, _RET_IP_); |
2799 | ran_callback = true; |
2800 | |
2801 | spin_lock(lock: &log->l_icloglock); |
2802 | xlog_state_clean_iclog(log, dirty_iclog: iclog); |
2803 | iclog = iclog->ic_next; |
2804 | } while (iclog != first_iclog); |
2805 | |
2806 | return ran_callback; |
2807 | } |
2808 | |
2809 | |
2810 | /* |
2811 | * Loop running iclog completion callbacks until there are no more iclogs in a |
2812 | * state that can run callbacks. |
2813 | */ |
2814 | STATIC void |
2815 | xlog_state_do_callback( |
2816 | struct xlog *log) |
2817 | { |
2818 | int flushcnt = 0; |
2819 | int repeats = 0; |
2820 | |
2821 | spin_lock(lock: &log->l_icloglock); |
2822 | while (xlog_state_do_iclog_callbacks(log)) { |
2823 | if (xlog_is_shutdown(log)) |
2824 | break; |
2825 | |
2826 | if (++repeats > 5000) { |
2827 | flushcnt += repeats; |
2828 | repeats = 0; |
2829 | xfs_warn(log->l_mp, |
2830 | "%s: possible infinite loop (%d iterations)" , |
2831 | __func__, flushcnt); |
2832 | } |
2833 | } |
2834 | |
2835 | if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE) |
2836 | wake_up_all(&log->l_flush_wait); |
2837 | |
2838 | spin_unlock(lock: &log->l_icloglock); |
2839 | } |
2840 | |
2841 | |
2842 | /* |
2843 | * Finish transitioning this iclog to the dirty state. |
2844 | * |
2845 | * Callbacks could take time, so they are done outside the scope of the |
2846 | * global state machine log lock. |
2847 | */ |
2848 | STATIC void |
2849 | xlog_state_done_syncing( |
2850 | struct xlog_in_core *iclog) |
2851 | { |
2852 | struct xlog *log = iclog->ic_log; |
2853 | |
2854 | spin_lock(lock: &log->l_icloglock); |
2855 | ASSERT(atomic_read(&iclog->ic_refcnt) == 0); |
2856 | trace_xlog_iclog_sync_done(iclog, _RET_IP_); |
2857 | |
2858 | /* |
2859 | * If we got an error, either on the first buffer, or in the case of |
2860 | * split log writes, on the second, we shut down the file system and |
2861 | * no iclogs should ever be attempted to be written to disk again. |
2862 | */ |
2863 | if (!xlog_is_shutdown(log)) { |
2864 | ASSERT(iclog->ic_state == XLOG_STATE_SYNCING); |
2865 | iclog->ic_state = XLOG_STATE_DONE_SYNC; |
2866 | } |
2867 | |
2868 | /* |
2869 | * Someone could be sleeping prior to writing out the next |
2870 | * iclog buffer, we wake them all, one will get to do the |
2871 | * I/O, the others get to wait for the result. |
2872 | */ |
2873 | wake_up_all(&iclog->ic_write_wait); |
2874 | spin_unlock(lock: &log->l_icloglock); |
2875 | xlog_state_do_callback(log); |
2876 | } |
2877 | |
2878 | /* |
2879 | * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must |
2880 | * sleep. We wait on the flush queue on the head iclog as that should be |
2881 | * the first iclog to complete flushing. Hence if all iclogs are syncing, |
2882 | * we will wait here and all new writes will sleep until a sync completes. |
2883 | * |
2884 | * The in-core logs are used in a circular fashion. They are not used |
2885 | * out-of-order even when an iclog past the head is free. |
2886 | * |
2887 | * return: |
2888 | * * log_offset where xlog_write() can start writing into the in-core |
2889 | * log's data space. |
2890 | * * in-core log pointer to which xlog_write() should write. |
2891 | * * boolean indicating this is a continued write to an in-core log. |
2892 | * If this is the last write, then the in-core log's offset field |
2893 | * needs to be incremented, depending on the amount of data which |
2894 | * is copied. |
2895 | */ |
2896 | STATIC int |
2897 | xlog_state_get_iclog_space( |
2898 | struct xlog *log, |
2899 | int len, |
2900 | struct xlog_in_core **iclogp, |
2901 | struct xlog_ticket *ticket, |
2902 | int *logoffsetp) |
2903 | { |
2904 | int log_offset; |
2905 | xlog_rec_header_t *head; |
2906 | xlog_in_core_t *iclog; |
2907 | |
2908 | restart: |
2909 | spin_lock(lock: &log->l_icloglock); |
2910 | if (xlog_is_shutdown(log)) { |
2911 | spin_unlock(lock: &log->l_icloglock); |
2912 | return -EIO; |
2913 | } |
2914 | |
2915 | iclog = log->l_iclog; |
2916 | if (iclog->ic_state != XLOG_STATE_ACTIVE) { |
2917 | XFS_STATS_INC(log->l_mp, xs_log_noiclogs); |
2918 | |
2919 | /* Wait for log writes to have flushed */ |
2920 | xlog_wait(wq: &log->l_flush_wait, lock: &log->l_icloglock); |
2921 | goto restart; |
2922 | } |
2923 | |
2924 | head = &iclog->ic_header; |
2925 | |
2926 | atomic_inc(v: &iclog->ic_refcnt); /* prevents sync */ |
2927 | log_offset = iclog->ic_offset; |
2928 | |
2929 | trace_xlog_iclog_get_space(iclog, _RET_IP_); |
2930 | |
2931 | /* On the 1st write to an iclog, figure out lsn. This works |
2932 | * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are |
2933 | * committing to. If the offset is set, that's how many blocks |
2934 | * must be written. |
2935 | */ |
2936 | if (log_offset == 0) { |
2937 | ticket->t_curr_res -= log->l_iclog_hsize; |
2938 | head->h_cycle = cpu_to_be32(log->l_curr_cycle); |
2939 | head->h_lsn = cpu_to_be64( |
2940 | xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block)); |
2941 | ASSERT(log->l_curr_block >= 0); |
2942 | } |
2943 | |
2944 | /* If there is enough room to write everything, then do it. Otherwise, |
2945 | * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC |
2946 | * bit is on, so this will get flushed out. Don't update ic_offset |
2947 | * until you know exactly how many bytes get copied. Therefore, wait |
2948 | * until later to update ic_offset. |
2949 | * |
2950 | * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's |
2951 | * can fit into remaining data section. |
2952 | */ |
2953 | if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) { |
2954 | int error = 0; |
2955 | |
2956 | xlog_state_switch_iclogs(log, iclog, eventual_size: iclog->ic_size); |
2957 | |
2958 | /* |
2959 | * If we are the only one writing to this iclog, sync it to |
2960 | * disk. We need to do an atomic compare and decrement here to |
2961 | * avoid racing with concurrent atomic_dec_and_lock() calls in |
2962 | * xlog_state_release_iclog() when there is more than one |
2963 | * reference to the iclog. |
2964 | */ |
2965 | if (!atomic_add_unless(v: &iclog->ic_refcnt, a: -1, u: 1)) |
2966 | error = xlog_state_release_iclog(log, iclog, ticket); |
2967 | spin_unlock(lock: &log->l_icloglock); |
2968 | if (error) |
2969 | return error; |
2970 | goto restart; |
2971 | } |
2972 | |
2973 | /* Do we have enough room to write the full amount in the remainder |
2974 | * of this iclog? Or must we continue a write on the next iclog and |
2975 | * mark this iclog as completely taken? In the case where we switch |
2976 | * iclogs (to mark it taken), this particular iclog will release/sync |
2977 | * to disk in xlog_write(). |
2978 | */ |
2979 | if (len <= iclog->ic_size - iclog->ic_offset) |
2980 | iclog->ic_offset += len; |
2981 | else |
2982 | xlog_state_switch_iclogs(log, iclog, eventual_size: iclog->ic_size); |
2983 | *iclogp = iclog; |
2984 | |
2985 | ASSERT(iclog->ic_offset <= iclog->ic_size); |
2986 | spin_unlock(lock: &log->l_icloglock); |
2987 | |
2988 | *logoffsetp = log_offset; |
2989 | return 0; |
2990 | } |
2991 | |
2992 | /* |
2993 | * The first cnt-1 times a ticket goes through here we don't need to move the |
2994 | * grant write head because the permanent reservation has reserved cnt times the |
2995 | * unit amount. Release part of current permanent unit reservation and reset |
2996 | * current reservation to be one units worth. Also move grant reservation head |
2997 | * forward. |
2998 | */ |
2999 | void |
3000 | xfs_log_ticket_regrant( |
3001 | struct xlog *log, |
3002 | struct xlog_ticket *ticket) |
3003 | { |
3004 | trace_xfs_log_ticket_regrant(log, tic: ticket); |
3005 | |
3006 | if (ticket->t_cnt > 0) |
3007 | ticket->t_cnt--; |
3008 | |
3009 | xlog_grant_sub_space(log, head: &log->l_reserve_head.grant, |
3010 | bytes: ticket->t_curr_res); |
3011 | xlog_grant_sub_space(log, head: &log->l_write_head.grant, |
3012 | bytes: ticket->t_curr_res); |
3013 | ticket->t_curr_res = ticket->t_unit_res; |
3014 | |
3015 | trace_xfs_log_ticket_regrant_sub(log, tic: ticket); |
3016 | |
3017 | /* just return if we still have some of the pre-reserved space */ |
3018 | if (!ticket->t_cnt) { |
3019 | xlog_grant_add_space(log, head: &log->l_reserve_head.grant, |
3020 | bytes: ticket->t_unit_res); |
3021 | trace_xfs_log_ticket_regrant_exit(log, tic: ticket); |
3022 | |
3023 | ticket->t_curr_res = ticket->t_unit_res; |
3024 | } |
3025 | |
3026 | xfs_log_ticket_put(ticket); |
3027 | } |
3028 | |
3029 | /* |
3030 | * Give back the space left from a reservation. |
3031 | * |
3032 | * All the information we need to make a correct determination of space left |
3033 | * is present. For non-permanent reservations, things are quite easy. The |
3034 | * count should have been decremented to zero. We only need to deal with the |
3035 | * space remaining in the current reservation part of the ticket. If the |
3036 | * ticket contains a permanent reservation, there may be left over space which |
3037 | * needs to be released. A count of N means that N-1 refills of the current |
3038 | * reservation can be done before we need to ask for more space. The first |
3039 | * one goes to fill up the first current reservation. Once we run out of |
3040 | * space, the count will stay at zero and the only space remaining will be |
3041 | * in the current reservation field. |
3042 | */ |
3043 | void |
3044 | xfs_log_ticket_ungrant( |
3045 | struct xlog *log, |
3046 | struct xlog_ticket *ticket) |
3047 | { |
3048 | int bytes; |
3049 | |
3050 | trace_xfs_log_ticket_ungrant(log, tic: ticket); |
3051 | |
3052 | if (ticket->t_cnt > 0) |
3053 | ticket->t_cnt--; |
3054 | |
3055 | trace_xfs_log_ticket_ungrant_sub(log, tic: ticket); |
3056 | |
3057 | /* |
3058 | * If this is a permanent reservation ticket, we may be able to free |
3059 | * up more space based on the remaining count. |
3060 | */ |
3061 | bytes = ticket->t_curr_res; |
3062 | if (ticket->t_cnt > 0) { |
3063 | ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV); |
3064 | bytes += ticket->t_unit_res*ticket->t_cnt; |
3065 | } |
3066 | |
3067 | xlog_grant_sub_space(log, head: &log->l_reserve_head.grant, bytes); |
3068 | xlog_grant_sub_space(log, head: &log->l_write_head.grant, bytes); |
3069 | |
3070 | trace_xfs_log_ticket_ungrant_exit(log, tic: ticket); |
3071 | |
3072 | xfs_log_space_wake(mp: log->l_mp); |
3073 | xfs_log_ticket_put(ticket); |
3074 | } |
3075 | |
3076 | /* |
3077 | * This routine will mark the current iclog in the ring as WANT_SYNC and move |
3078 | * the current iclog pointer to the next iclog in the ring. |
3079 | */ |
3080 | void |
3081 | xlog_state_switch_iclogs( |
3082 | struct xlog *log, |
3083 | struct xlog_in_core *iclog, |
3084 | int eventual_size) |
3085 | { |
3086 | ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); |
3087 | assert_spin_locked(&log->l_icloglock); |
3088 | trace_xlog_iclog_switch(iclog, _RET_IP_); |
3089 | |
3090 | if (!eventual_size) |
3091 | eventual_size = iclog->ic_offset; |
3092 | iclog->ic_state = XLOG_STATE_WANT_SYNC; |
3093 | iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block); |
3094 | log->l_prev_block = log->l_curr_block; |
3095 | log->l_prev_cycle = log->l_curr_cycle; |
3096 | |
3097 | /* roll log?: ic_offset changed later */ |
3098 | log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize); |
3099 | |
3100 | /* Round up to next log-sunit */ |
3101 | if (log->l_iclog_roundoff > BBSIZE) { |
3102 | uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff); |
3103 | log->l_curr_block = roundup(log->l_curr_block, sunit_bb); |
3104 | } |
3105 | |
3106 | if (log->l_curr_block >= log->l_logBBsize) { |
3107 | /* |
3108 | * Rewind the current block before the cycle is bumped to make |
3109 | * sure that the combined LSN never transiently moves forward |
3110 | * when the log wraps to the next cycle. This is to support the |
3111 | * unlocked sample of these fields from xlog_valid_lsn(). Most |
3112 | * other cases should acquire l_icloglock. |
3113 | */ |
3114 | log->l_curr_block -= log->l_logBBsize; |
3115 | ASSERT(log->l_curr_block >= 0); |
3116 | smp_wmb(); |
3117 | log->l_curr_cycle++; |
3118 | if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM) |
3119 | log->l_curr_cycle++; |
3120 | } |
3121 | ASSERT(iclog == log->l_iclog); |
3122 | log->l_iclog = iclog->ic_next; |
3123 | } |
3124 | |
3125 | /* |
3126 | * Force the iclog to disk and check if the iclog has been completed before |
3127 | * xlog_force_iclog() returns. This can happen on synchronous (e.g. |
3128 | * pmem) or fast async storage because we drop the icloglock to issue the IO. |
3129 | * If completion has already occurred, tell the caller so that it can avoid an |
3130 | * unnecessary wait on the iclog. |
3131 | */ |
3132 | static int |
3133 | xlog_force_and_check_iclog( |
3134 | struct xlog_in_core *iclog, |
3135 | bool *completed) |
3136 | { |
3137 | xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn); |
3138 | int error; |
3139 | |
3140 | *completed = false; |
3141 | error = xlog_force_iclog(iclog); |
3142 | if (error) |
3143 | return error; |
3144 | |
3145 | /* |
3146 | * If the iclog has already been completed and reused the header LSN |
3147 | * will have been rewritten by completion |
3148 | */ |
3149 | if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) |
3150 | *completed = true; |
3151 | return 0; |
3152 | } |
3153 | |
3154 | /* |
3155 | * Write out all data in the in-core log as of this exact moment in time. |
3156 | * |
3157 | * Data may be written to the in-core log during this call. However, |
3158 | * we don't guarantee this data will be written out. A change from past |
3159 | * implementation means this routine will *not* write out zero length LRs. |
3160 | * |
3161 | * Basically, we try and perform an intelligent scan of the in-core logs. |
3162 | * If we determine there is no flushable data, we just return. There is no |
3163 | * flushable data if: |
3164 | * |
3165 | * 1. the current iclog is active and has no data; the previous iclog |
3166 | * is in the active or dirty state. |
3167 | * 2. the current iclog is drity, and the previous iclog is in the |
3168 | * active or dirty state. |
3169 | * |
3170 | * We may sleep if: |
3171 | * |
3172 | * 1. the current iclog is not in the active nor dirty state. |
3173 | * 2. the current iclog dirty, and the previous iclog is not in the |
3174 | * active nor dirty state. |
3175 | * 3. the current iclog is active, and there is another thread writing |
3176 | * to this particular iclog. |
3177 | * 4. a) the current iclog is active and has no other writers |
3178 | * b) when we return from flushing out this iclog, it is still |
3179 | * not in the active nor dirty state. |
3180 | */ |
3181 | int |
3182 | xfs_log_force( |
3183 | struct xfs_mount *mp, |
3184 | uint flags) |
3185 | { |
3186 | struct xlog *log = mp->m_log; |
3187 | struct xlog_in_core *iclog; |
3188 | |
3189 | XFS_STATS_INC(mp, xs_log_force); |
3190 | trace_xfs_log_force(mp, 0, _RET_IP_); |
3191 | |
3192 | xlog_cil_force(log); |
3193 | |
3194 | spin_lock(lock: &log->l_icloglock); |
3195 | if (xlog_is_shutdown(log)) |
3196 | goto out_error; |
3197 | |
3198 | iclog = log->l_iclog; |
3199 | trace_xlog_iclog_force(iclog, _RET_IP_); |
3200 | |
3201 | if (iclog->ic_state == XLOG_STATE_DIRTY || |
3202 | (iclog->ic_state == XLOG_STATE_ACTIVE && |
3203 | atomic_read(v: &iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) { |
3204 | /* |
3205 | * If the head is dirty or (active and empty), then we need to |
3206 | * look at the previous iclog. |
3207 | * |
3208 | * If the previous iclog is active or dirty we are done. There |
3209 | * is nothing to sync out. Otherwise, we attach ourselves to the |
3210 | * previous iclog and go to sleep. |
3211 | */ |
3212 | iclog = iclog->ic_prev; |
3213 | } else if (iclog->ic_state == XLOG_STATE_ACTIVE) { |
3214 | if (atomic_read(v: &iclog->ic_refcnt) == 0) { |
3215 | /* We have exclusive access to this iclog. */ |
3216 | bool completed; |
3217 | |
3218 | if (xlog_force_and_check_iclog(iclog, completed: &completed)) |
3219 | goto out_error; |
3220 | |
3221 | if (completed) |
3222 | goto out_unlock; |
3223 | } else { |
3224 | /* |
3225 | * Someone else is still writing to this iclog, so we |
3226 | * need to ensure that when they release the iclog it |
3227 | * gets synced immediately as we may be waiting on it. |
3228 | */ |
3229 | xlog_state_switch_iclogs(log, iclog, eventual_size: 0); |
3230 | } |
3231 | } |
3232 | |
3233 | /* |
3234 | * The iclog we are about to wait on may contain the checkpoint pushed |
3235 | * by the above xlog_cil_force() call, but it may not have been pushed |
3236 | * to disk yet. Like the ACTIVE case above, we need to make sure caches |
3237 | * are flushed when this iclog is written. |
3238 | */ |
3239 | if (iclog->ic_state == XLOG_STATE_WANT_SYNC) |
3240 | iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA; |
3241 | |
3242 | if (flags & XFS_LOG_SYNC) |
3243 | return xlog_wait_on_iclog(iclog); |
3244 | out_unlock: |
3245 | spin_unlock(lock: &log->l_icloglock); |
3246 | return 0; |
3247 | out_error: |
3248 | spin_unlock(lock: &log->l_icloglock); |
3249 | return -EIO; |
3250 | } |
3251 | |
3252 | /* |
3253 | * Force the log to a specific LSN. |
3254 | * |
3255 | * If an iclog with that lsn can be found: |
3256 | * If it is in the DIRTY state, just return. |
3257 | * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC |
3258 | * state and go to sleep or return. |
3259 | * If it is in any other state, go to sleep or return. |
3260 | * |
3261 | * Synchronous forces are implemented with a wait queue. All callers trying |
3262 | * to force a given lsn to disk must wait on the queue attached to the |
3263 | * specific in-core log. When given in-core log finally completes its write |
3264 | * to disk, that thread will wake up all threads waiting on the queue. |
3265 | */ |
3266 | static int |
3267 | xlog_force_lsn( |
3268 | struct xlog *log, |
3269 | xfs_lsn_t lsn, |
3270 | uint flags, |
3271 | int *log_flushed, |
3272 | bool already_slept) |
3273 | { |
3274 | struct xlog_in_core *iclog; |
3275 | bool completed; |
3276 | |
3277 | spin_lock(lock: &log->l_icloglock); |
3278 | if (xlog_is_shutdown(log)) |
3279 | goto out_error; |
3280 | |
3281 | iclog = log->l_iclog; |
3282 | while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) { |
3283 | trace_xlog_iclog_force_lsn(iclog, _RET_IP_); |
3284 | iclog = iclog->ic_next; |
3285 | if (iclog == log->l_iclog) |
3286 | goto out_unlock; |
3287 | } |
3288 | |
3289 | switch (iclog->ic_state) { |
3290 | case XLOG_STATE_ACTIVE: |
3291 | /* |
3292 | * We sleep here if we haven't already slept (e.g. this is the |
3293 | * first time we've looked at the correct iclog buf) and the |
3294 | * buffer before us is going to be sync'ed. The reason for this |
3295 | * is that if we are doing sync transactions here, by waiting |
3296 | * for the previous I/O to complete, we can allow a few more |
3297 | * transactions into this iclog before we close it down. |
3298 | * |
3299 | * Otherwise, we mark the buffer WANT_SYNC, and bump up the |
3300 | * refcnt so we can release the log (which drops the ref count). |
3301 | * The state switch keeps new transaction commits from using |
3302 | * this buffer. When the current commits finish writing into |
3303 | * the buffer, the refcount will drop to zero and the buffer |
3304 | * will go out then. |
3305 | */ |
3306 | if (!already_slept && |
3307 | (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC || |
3308 | iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) { |
3309 | xlog_wait(wq: &iclog->ic_prev->ic_write_wait, |
3310 | lock: &log->l_icloglock); |
3311 | return -EAGAIN; |
3312 | } |
3313 | if (xlog_force_and_check_iclog(iclog, completed: &completed)) |
3314 | goto out_error; |
3315 | if (log_flushed) |
3316 | *log_flushed = 1; |
3317 | if (completed) |
3318 | goto out_unlock; |
3319 | break; |
3320 | case XLOG_STATE_WANT_SYNC: |
3321 | /* |
3322 | * This iclog may contain the checkpoint pushed by the |
3323 | * xlog_cil_force_seq() call, but there are other writers still |
3324 | * accessing it so it hasn't been pushed to disk yet. Like the |
3325 | * ACTIVE case above, we need to make sure caches are flushed |
3326 | * when this iclog is written. |
3327 | */ |
3328 | iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA; |
3329 | break; |
3330 | default: |
3331 | /* |
3332 | * The entire checkpoint was written by the CIL force and is on |
3333 | * its way to disk already. It will be stable when it |
3334 | * completes, so we don't need to manipulate caches here at all. |
3335 | * We just need to wait for completion if necessary. |
3336 | */ |
3337 | break; |
3338 | } |
3339 | |
3340 | if (flags & XFS_LOG_SYNC) |
3341 | return xlog_wait_on_iclog(iclog); |
3342 | out_unlock: |
3343 | spin_unlock(lock: &log->l_icloglock); |
3344 | return 0; |
3345 | out_error: |
3346 | spin_unlock(lock: &log->l_icloglock); |
3347 | return -EIO; |
3348 | } |
3349 | |
3350 | /* |
3351 | * Force the log to a specific checkpoint sequence. |
3352 | * |
3353 | * First force the CIL so that all the required changes have been flushed to the |
3354 | * iclogs. If the CIL force completed it will return a commit LSN that indicates |
3355 | * the iclog that needs to be flushed to stable storage. If the caller needs |
3356 | * a synchronous log force, we will wait on the iclog with the LSN returned by |
3357 | * xlog_cil_force_seq() to be completed. |
3358 | */ |
3359 | int |
3360 | xfs_log_force_seq( |
3361 | struct xfs_mount *mp, |
3362 | xfs_csn_t seq, |
3363 | uint flags, |
3364 | int *log_flushed) |
3365 | { |
3366 | struct xlog *log = mp->m_log; |
3367 | xfs_lsn_t lsn; |
3368 | int ret; |
3369 | ASSERT(seq != 0); |
3370 | |
3371 | XFS_STATS_INC(mp, xs_log_force); |
3372 | trace_xfs_log_force(mp, seq, _RET_IP_); |
3373 | |
3374 | lsn = xlog_cil_force_seq(log, seq); |
3375 | if (lsn == NULLCOMMITLSN) |
3376 | return 0; |
3377 | |
3378 | ret = xlog_force_lsn(log, lsn, flags, log_flushed, false); |
3379 | if (ret == -EAGAIN) { |
3380 | XFS_STATS_INC(mp, xs_log_force_sleep); |
3381 | ret = xlog_force_lsn(log, lsn, flags, log_flushed, true); |
3382 | } |
3383 | return ret; |
3384 | } |
3385 | |
3386 | /* |
3387 | * Free a used ticket when its refcount falls to zero. |
3388 | */ |
3389 | void |
3390 | xfs_log_ticket_put( |
3391 | xlog_ticket_t *ticket) |
3392 | { |
3393 | ASSERT(atomic_read(&ticket->t_ref) > 0); |
3394 | if (atomic_dec_and_test(v: &ticket->t_ref)) |
3395 | kmem_cache_free(s: xfs_log_ticket_cache, objp: ticket); |
3396 | } |
3397 | |
3398 | xlog_ticket_t * |
3399 | xfs_log_ticket_get( |
3400 | xlog_ticket_t *ticket) |
3401 | { |
3402 | ASSERT(atomic_read(&ticket->t_ref) > 0); |
3403 | atomic_inc(v: &ticket->t_ref); |
3404 | return ticket; |
3405 | } |
3406 | |
3407 | /* |
3408 | * Figure out the total log space unit (in bytes) that would be |
3409 | * required for a log ticket. |
3410 | */ |
3411 | static int |
3412 | xlog_calc_unit_res( |
3413 | struct xlog *log, |
3414 | int unit_bytes, |
3415 | int *niclogs) |
3416 | { |
3417 | int iclog_space; |
3418 | uint ; |
3419 | |
3420 | /* |
3421 | * Permanent reservations have up to 'cnt'-1 active log operations |
3422 | * in the log. A unit in this case is the amount of space for one |
3423 | * of these log operations. Normal reservations have a cnt of 1 |
3424 | * and their unit amount is the total amount of space required. |
3425 | * |
3426 | * The following lines of code account for non-transaction data |
3427 | * which occupy space in the on-disk log. |
3428 | * |
3429 | * Normal form of a transaction is: |
3430 | * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph> |
3431 | * and then there are LR hdrs, split-recs and roundoff at end of syncs. |
3432 | * |
3433 | * We need to account for all the leadup data and trailer data |
3434 | * around the transaction data. |
3435 | * And then we need to account for the worst case in terms of using |
3436 | * more space. |
3437 | * The worst case will happen if: |
3438 | * - the placement of the transaction happens to be such that the |
3439 | * roundoff is at its maximum |
3440 | * - the transaction data is synced before the commit record is synced |
3441 | * i.e. <transaction-data><roundoff> | <commit-rec><roundoff> |
3442 | * Therefore the commit record is in its own Log Record. |
3443 | * This can happen as the commit record is called with its |
3444 | * own region to xlog_write(). |
3445 | * This then means that in the worst case, roundoff can happen for |
3446 | * the commit-rec as well. |
3447 | * The commit-rec is smaller than padding in this scenario and so it is |
3448 | * not added separately. |
3449 | */ |
3450 | |
3451 | /* for trans header */ |
3452 | unit_bytes += sizeof(xlog_op_header_t); |
3453 | unit_bytes += sizeof(xfs_trans_header_t); |
3454 | |
3455 | /* for start-rec */ |
3456 | unit_bytes += sizeof(xlog_op_header_t); |
3457 | |
3458 | /* |
3459 | * for LR headers - the space for data in an iclog is the size minus |
3460 | * the space used for the headers. If we use the iclog size, then we |
3461 | * undercalculate the number of headers required. |
3462 | * |
3463 | * Furthermore - the addition of op headers for split-recs might |
3464 | * increase the space required enough to require more log and op |
3465 | * headers, so take that into account too. |
3466 | * |
3467 | * IMPORTANT: This reservation makes the assumption that if this |
3468 | * transaction is the first in an iclog and hence has the LR headers |
3469 | * accounted to it, then the remaining space in the iclog is |
3470 | * exclusively for this transaction. i.e. if the transaction is larger |
3471 | * than the iclog, it will be the only thing in that iclog. |
3472 | * Fundamentally, this means we must pass the entire log vector to |
3473 | * xlog_write to guarantee this. |
3474 | */ |
3475 | iclog_space = log->l_iclog_size - log->l_iclog_hsize; |
3476 | num_headers = howmany(unit_bytes, iclog_space); |
3477 | |
3478 | /* for split-recs - ophdrs added when data split over LRs */ |
3479 | unit_bytes += sizeof(xlog_op_header_t) * num_headers; |
3480 | |
3481 | /* add extra header reservations if we overrun */ |
3482 | while (!num_headers || |
3483 | howmany(unit_bytes, iclog_space) > num_headers) { |
3484 | unit_bytes += sizeof(xlog_op_header_t); |
3485 | num_headers++; |
3486 | } |
3487 | unit_bytes += log->l_iclog_hsize * num_headers; |
3488 | |
3489 | /* for commit-rec LR header - note: padding will subsume the ophdr */ |
3490 | unit_bytes += log->l_iclog_hsize; |
3491 | |
3492 | /* roundoff padding for transaction data and one for commit record */ |
3493 | unit_bytes += 2 * log->l_iclog_roundoff; |
3494 | |
3495 | if (niclogs) |
3496 | *niclogs = num_headers; |
3497 | return unit_bytes; |
3498 | } |
3499 | |
3500 | int |
3501 | xfs_log_calc_unit_res( |
3502 | struct xfs_mount *mp, |
3503 | int unit_bytes) |
3504 | { |
3505 | return xlog_calc_unit_res(log: mp->m_log, unit_bytes, NULL); |
3506 | } |
3507 | |
3508 | /* |
3509 | * Allocate and initialise a new log ticket. |
3510 | */ |
3511 | struct xlog_ticket * |
3512 | xlog_ticket_alloc( |
3513 | struct xlog *log, |
3514 | int unit_bytes, |
3515 | int cnt, |
3516 | bool permanent) |
3517 | { |
3518 | struct xlog_ticket *tic; |
3519 | int unit_res; |
3520 | |
3521 | tic = kmem_cache_zalloc(k: xfs_log_ticket_cache, |
3522 | GFP_KERNEL | __GFP_NOFAIL); |
3523 | |
3524 | unit_res = xlog_calc_unit_res(log, unit_bytes, niclogs: &tic->t_iclog_hdrs); |
3525 | |
3526 | atomic_set(v: &tic->t_ref, i: 1); |
3527 | tic->t_task = current; |
3528 | INIT_LIST_HEAD(list: &tic->t_queue); |
3529 | tic->t_unit_res = unit_res; |
3530 | tic->t_curr_res = unit_res; |
3531 | tic->t_cnt = cnt; |
3532 | tic->t_ocnt = cnt; |
3533 | tic->t_tid = get_random_u32(); |
3534 | if (permanent) |
3535 | tic->t_flags |= XLOG_TIC_PERM_RESERV; |
3536 | |
3537 | return tic; |
3538 | } |
3539 | |
3540 | #if defined(DEBUG) |
3541 | /* |
3542 | * Check to make sure the grant write head didn't just over lap the tail. If |
3543 | * the cycles are the same, we can't be overlapping. Otherwise, make sure that |
3544 | * the cycles differ by exactly one and check the byte count. |
3545 | * |
3546 | * This check is run unlocked, so can give false positives. Rather than assert |
3547 | * on failures, use a warn-once flag and a panic tag to allow the admin to |
3548 | * determine if they want to panic the machine when such an error occurs. For |
3549 | * debug kernels this will have the same effect as using an assert but, unlinke |
3550 | * an assert, it can be turned off at runtime. |
3551 | */ |
3552 | STATIC void |
3553 | xlog_verify_grant_tail( |
3554 | struct xlog *log) |
3555 | { |
3556 | int tail_cycle, tail_blocks; |
3557 | int cycle, space; |
3558 | |
3559 | xlog_crack_grant_head(head: &log->l_write_head.grant, cycle: &cycle, space: &space); |
3560 | xlog_crack_atomic_lsn(lsn: &log->l_tail_lsn, cycle: &tail_cycle, block: &tail_blocks); |
3561 | if (tail_cycle != cycle) { |
3562 | if (cycle - 1 != tail_cycle && |
3563 | !test_and_set_bit(XLOG_TAIL_WARN, addr: &log->l_opstate)) { |
3564 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, |
3565 | "%s: cycle - 1 != tail_cycle" , __func__); |
3566 | } |
3567 | |
3568 | if (space > BBTOB(tail_blocks) && |
3569 | !test_and_set_bit(XLOG_TAIL_WARN, addr: &log->l_opstate)) { |
3570 | xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, |
3571 | "%s: space > BBTOB(tail_blocks)" , __func__); |
3572 | } |
3573 | } |
3574 | } |
3575 | |
3576 | /* check if it will fit */ |
3577 | STATIC void |
3578 | xlog_verify_tail_lsn( |
3579 | struct xlog *log, |
3580 | struct xlog_in_core *iclog) |
3581 | { |
3582 | xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn); |
3583 | int blocks; |
3584 | |
3585 | if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) { |
3586 | blocks = |
3587 | log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn)); |
3588 | if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize)) |
3589 | xfs_emerg(log->l_mp, "%s: ran out of log space" , __func__); |
3590 | } else { |
3591 | ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle); |
3592 | |
3593 | if (BLOCK_LSN(tail_lsn) == log->l_prev_block) |
3594 | xfs_emerg(log->l_mp, "%s: tail wrapped" , __func__); |
3595 | |
3596 | blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block; |
3597 | if (blocks < BTOBB(iclog->ic_offset) + 1) |
3598 | xfs_emerg(log->l_mp, "%s: ran out of log space" , __func__); |
3599 | } |
3600 | } |
3601 | |
3602 | /* |
3603 | * Perform a number of checks on the iclog before writing to disk. |
3604 | * |
3605 | * 1. Make sure the iclogs are still circular |
3606 | * 2. Make sure we have a good magic number |
3607 | * 3. Make sure we don't have magic numbers in the data |
3608 | * 4. Check fields of each log operation header for: |
3609 | * A. Valid client identifier |
3610 | * B. tid ptr value falls in valid ptr space (user space code) |
3611 | * C. Length in log record header is correct according to the |
3612 | * individual operation headers within record. |
3613 | * 5. When a bwrite will occur within 5 blocks of the front of the physical |
3614 | * log, check the preceding blocks of the physical log to make sure all |
3615 | * the cycle numbers agree with the current cycle number. |
3616 | */ |
3617 | STATIC void |
3618 | xlog_verify_iclog( |
3619 | struct xlog *log, |
3620 | struct xlog_in_core *iclog, |
3621 | int count) |
3622 | { |
3623 | xlog_op_header_t *ophead; |
3624 | xlog_in_core_t *icptr; |
3625 | xlog_in_core_2_t *xhdr; |
3626 | void *base_ptr, *ptr, *p; |
3627 | ptrdiff_t field_offset; |
3628 | uint8_t clientid; |
3629 | int len, i, j, k, op_len; |
3630 | int idx; |
3631 | |
3632 | /* check validity of iclog pointers */ |
3633 | spin_lock(lock: &log->l_icloglock); |
3634 | icptr = log->l_iclog; |
3635 | for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next) |
3636 | ASSERT(icptr); |
3637 | |
3638 | if (icptr != log->l_iclog) |
3639 | xfs_emerg(log->l_mp, "%s: corrupt iclog ring" , __func__); |
3640 | spin_unlock(lock: &log->l_icloglock); |
3641 | |
3642 | /* check log magic numbers */ |
3643 | if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) |
3644 | xfs_emerg(log->l_mp, "%s: invalid magic num" , __func__); |
3645 | |
3646 | base_ptr = ptr = &iclog->ic_header; |
3647 | p = &iclog->ic_header; |
3648 | for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) { |
3649 | if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) |
3650 | xfs_emerg(log->l_mp, "%s: unexpected magic num" , |
3651 | __func__); |
3652 | } |
3653 | |
3654 | /* check fields */ |
3655 | len = be32_to_cpu(iclog->ic_header.h_num_logops); |
3656 | base_ptr = ptr = iclog->ic_datap; |
3657 | ophead = ptr; |
3658 | xhdr = iclog->ic_data; |
3659 | for (i = 0; i < len; i++) { |
3660 | ophead = ptr; |
3661 | |
3662 | /* clientid is only 1 byte */ |
3663 | p = &ophead->oh_clientid; |
3664 | field_offset = p - base_ptr; |
3665 | if (field_offset & 0x1ff) { |
3666 | clientid = ophead->oh_clientid; |
3667 | } else { |
3668 | idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap); |
3669 | if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { |
3670 | j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
3671 | k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
3672 | clientid = xlog_get_client_id( |
3673 | xhdr[j].hic_xheader.xh_cycle_data[k]); |
3674 | } else { |
3675 | clientid = xlog_get_client_id( |
3676 | i: iclog->ic_header.h_cycle_data[idx]); |
3677 | } |
3678 | } |
3679 | if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) { |
3680 | xfs_warn(log->l_mp, |
3681 | "%s: op %d invalid clientid %d op " PTR_FMT" offset 0x%lx" , |
3682 | __func__, i, clientid, ophead, |
3683 | (unsigned long)field_offset); |
3684 | } |
3685 | |
3686 | /* check length */ |
3687 | p = &ophead->oh_len; |
3688 | field_offset = p - base_ptr; |
3689 | if (field_offset & 0x1ff) { |
3690 | op_len = be32_to_cpu(ophead->oh_len); |
3691 | } else { |
3692 | idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap); |
3693 | if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { |
3694 | j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
3695 | k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); |
3696 | op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]); |
3697 | } else { |
3698 | op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]); |
3699 | } |
3700 | } |
3701 | ptr += sizeof(xlog_op_header_t) + op_len; |
3702 | } |
3703 | } |
3704 | #endif |
3705 | |
3706 | /* |
3707 | * Perform a forced shutdown on the log. |
3708 | * |
3709 | * This can be called from low level log code to trigger a shutdown, or from the |
3710 | * high level mount shutdown code when the mount shuts down. |
3711 | * |
3712 | * Our main objectives here are to make sure that: |
3713 | * a. if the shutdown was not due to a log IO error, flush the logs to |
3714 | * disk. Anything modified after this is ignored. |
3715 | * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested |
3716 | * parties to find out. Nothing new gets queued after this is done. |
3717 | * c. Tasks sleeping on log reservations, pinned objects and |
3718 | * other resources get woken up. |
3719 | * d. The mount is also marked as shut down so that log triggered shutdowns |
3720 | * still behave the same as if they called xfs_forced_shutdown(). |
3721 | * |
3722 | * Return true if the shutdown cause was a log IO error and we actually shut the |
3723 | * log down. |
3724 | */ |
3725 | bool |
3726 | xlog_force_shutdown( |
3727 | struct xlog *log, |
3728 | uint32_t shutdown_flags) |
3729 | { |
3730 | bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR); |
3731 | |
3732 | if (!log) |
3733 | return false; |
3734 | |
3735 | /* |
3736 | * Flush all the completed transactions to disk before marking the log |
3737 | * being shut down. We need to do this first as shutting down the log |
3738 | * before the force will prevent the log force from flushing the iclogs |
3739 | * to disk. |
3740 | * |
3741 | * When we are in recovery, there are no transactions to flush, and |
3742 | * we don't want to touch the log because we don't want to perturb the |
3743 | * current head/tail for future recovery attempts. Hence we need to |
3744 | * avoid a log force in this case. |
3745 | * |
3746 | * If we are shutting down due to a log IO error, then we must avoid |
3747 | * trying to write the log as that may just result in more IO errors and |
3748 | * an endless shutdown/force loop. |
3749 | */ |
3750 | if (!log_error && !xlog_in_recovery(log)) |
3751 | xfs_log_force(mp: log->l_mp, XFS_LOG_SYNC); |
3752 | |
3753 | /* |
3754 | * Atomically set the shutdown state. If the shutdown state is already |
3755 | * set, there someone else is performing the shutdown and so we are done |
3756 | * here. This should never happen because we should only ever get called |
3757 | * once by the first shutdown caller. |
3758 | * |
3759 | * Much of the log state machine transitions assume that shutdown state |
3760 | * cannot change once they hold the log->l_icloglock. Hence we need to |
3761 | * hold that lock here, even though we use the atomic test_and_set_bit() |
3762 | * operation to set the shutdown state. |
3763 | */ |
3764 | spin_lock(lock: &log->l_icloglock); |
3765 | if (test_and_set_bit(XLOG_IO_ERROR, addr: &log->l_opstate)) { |
3766 | spin_unlock(lock: &log->l_icloglock); |
3767 | return false; |
3768 | } |
3769 | spin_unlock(lock: &log->l_icloglock); |
3770 | |
3771 | /* |
3772 | * If this log shutdown also sets the mount shutdown state, issue a |
3773 | * shutdown warning message. |
3774 | */ |
3775 | if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, addr: &log->l_mp->m_opstate)) { |
3776 | xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR, |
3777 | "Filesystem has been shut down due to log error (0x%x)." , |
3778 | shutdown_flags); |
3779 | xfs_alert(log->l_mp, |
3780 | "Please unmount the filesystem and rectify the problem(s)." ); |
3781 | if (xfs_error_level >= XFS_ERRLEVEL_HIGH) |
3782 | xfs_stack_trace(); |
3783 | } |
3784 | |
3785 | /* |
3786 | * We don't want anybody waiting for log reservations after this. That |
3787 | * means we have to wake up everybody queued up on reserveq as well as |
3788 | * writeq. In addition, we make sure in xlog_{re}grant_log_space that |
3789 | * we don't enqueue anything once the SHUTDOWN flag is set, and this |
3790 | * action is protected by the grant locks. |
3791 | */ |
3792 | xlog_grant_head_wake_all(head: &log->l_reserve_head); |
3793 | xlog_grant_head_wake_all(head: &log->l_write_head); |
3794 | |
3795 | /* |
3796 | * Wake up everybody waiting on xfs_log_force. Wake the CIL push first |
3797 | * as if the log writes were completed. The abort handling in the log |
3798 | * item committed callback functions will do this again under lock to |
3799 | * avoid races. |
3800 | */ |
3801 | spin_lock(lock: &log->l_cilp->xc_push_lock); |
3802 | wake_up_all(&log->l_cilp->xc_start_wait); |
3803 | wake_up_all(&log->l_cilp->xc_commit_wait); |
3804 | spin_unlock(lock: &log->l_cilp->xc_push_lock); |
3805 | |
3806 | spin_lock(lock: &log->l_icloglock); |
3807 | xlog_state_shutdown_callbacks(log); |
3808 | spin_unlock(lock: &log->l_icloglock); |
3809 | |
3810 | wake_up_var(var: &log->l_opstate); |
3811 | return log_error; |
3812 | } |
3813 | |
3814 | STATIC int |
3815 | xlog_iclogs_empty( |
3816 | struct xlog *log) |
3817 | { |
3818 | xlog_in_core_t *iclog; |
3819 | |
3820 | iclog = log->l_iclog; |
3821 | do { |
3822 | /* endianness does not matter here, zero is zero in |
3823 | * any language. |
3824 | */ |
3825 | if (iclog->ic_header.h_num_logops) |
3826 | return 0; |
3827 | iclog = iclog->ic_next; |
3828 | } while (iclog != log->l_iclog); |
3829 | return 1; |
3830 | } |
3831 | |
3832 | /* |
3833 | * Verify that an LSN stamped into a piece of metadata is valid. This is |
3834 | * intended for use in read verifiers on v5 superblocks. |
3835 | */ |
3836 | bool |
3837 | xfs_log_check_lsn( |
3838 | struct xfs_mount *mp, |
3839 | xfs_lsn_t lsn) |
3840 | { |
3841 | struct xlog *log = mp->m_log; |
3842 | bool valid; |
3843 | |
3844 | /* |
3845 | * norecovery mode skips mount-time log processing and unconditionally |
3846 | * resets the in-core LSN. We can't validate in this mode, but |
3847 | * modifications are not allowed anyways so just return true. |
3848 | */ |
3849 | if (xfs_has_norecovery(mp)) |
3850 | return true; |
3851 | |
3852 | /* |
3853 | * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is |
3854 | * handled by recovery and thus safe to ignore here. |
3855 | */ |
3856 | if (lsn == NULLCOMMITLSN) |
3857 | return true; |
3858 | |
3859 | valid = xlog_valid_lsn(mp->m_log, lsn); |
3860 | |
3861 | /* warn the user about what's gone wrong before verifier failure */ |
3862 | if (!valid) { |
3863 | spin_lock(lock: &log->l_icloglock); |
3864 | xfs_warn(mp, |
3865 | "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). " |
3866 | "Please unmount and run xfs_repair (>= v4.3) to resolve." , |
3867 | CYCLE_LSN(lsn), BLOCK_LSN(lsn), |
3868 | log->l_curr_cycle, log->l_curr_block); |
3869 | spin_unlock(lock: &log->l_icloglock); |
3870 | } |
3871 | |
3872 | return valid; |
3873 | } |
3874 | |
3875 | /* |
3876 | * Notify the log that we're about to start using a feature that is protected |
3877 | * by a log incompat feature flag. This will prevent log covering from |
3878 | * clearing those flags. |
3879 | */ |
3880 | void |
3881 | xlog_use_incompat_feat( |
3882 | struct xlog *log) |
3883 | { |
3884 | down_read(sem: &log->l_incompat_users); |
3885 | } |
3886 | |
3887 | /* Notify the log that we've finished using log incompat features. */ |
3888 | void |
3889 | xlog_drop_incompat_feat( |
3890 | struct xlog *log) |
3891 | { |
3892 | up_read(sem: &log->l_incompat_users); |
3893 | } |
3894 | |