Loading...
1/*
2 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it would be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write the Free Software Foundation,
15 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
16 */
17
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_bit.h"
22#include "xfs_log.h"
23#include "xfs_inum.h"
24#include "xfs_trans.h"
25#include "xfs_trans_priv.h"
26#include "xfs_log_priv.h"
27#include "xfs_sb.h"
28#include "xfs_ag.h"
29#include "xfs_mount.h"
30#include "xfs_error.h"
31#include "xfs_alloc.h"
32#include "xfs_discard.h"
33
34/*
35 * Perform initial CIL structure initialisation. If the CIL is not
36 * enabled in this filesystem, ensure the log->l_cilp is null so
37 * we can check this conditional to determine if we are doing delayed
38 * logging or not.
39 */
40int
41xlog_cil_init(
42 struct log *log)
43{
44 struct xfs_cil *cil;
45 struct xfs_cil_ctx *ctx;
46
47 log->l_cilp = NULL;
48 if (!(log->l_mp->m_flags & XFS_MOUNT_DELAYLOG))
49 return 0;
50
51 cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
52 if (!cil)
53 return ENOMEM;
54
55 ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
56 if (!ctx) {
57 kmem_free(cil);
58 return ENOMEM;
59 }
60
61 INIT_LIST_HEAD(&cil->xc_cil);
62 INIT_LIST_HEAD(&cil->xc_committing);
63 spin_lock_init(&cil->xc_cil_lock);
64 init_rwsem(&cil->xc_ctx_lock);
65 init_waitqueue_head(&cil->xc_commit_wait);
66
67 INIT_LIST_HEAD(&ctx->committing);
68 INIT_LIST_HEAD(&ctx->busy_extents);
69 ctx->sequence = 1;
70 ctx->cil = cil;
71 cil->xc_ctx = ctx;
72 cil->xc_current_sequence = ctx->sequence;
73
74 cil->xc_log = log;
75 log->l_cilp = cil;
76 return 0;
77}
78
79void
80xlog_cil_destroy(
81 struct log *log)
82{
83 if (!log->l_cilp)
84 return;
85
86 if (log->l_cilp->xc_ctx) {
87 if (log->l_cilp->xc_ctx->ticket)
88 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
89 kmem_free(log->l_cilp->xc_ctx);
90 }
91
92 ASSERT(list_empty(&log->l_cilp->xc_cil));
93 kmem_free(log->l_cilp);
94}
95
96/*
97 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
98 * recover, so we don't allow failure here. Also, we allocate in a context that
99 * we don't want to be issuing transactions from, so we need to tell the
100 * allocation code this as well.
101 *
102 * We don't reserve any space for the ticket - we are going to steal whatever
103 * space we require from transactions as they commit. To ensure we reserve all
104 * the space required, we need to set the current reservation of the ticket to
105 * zero so that we know to steal the initial transaction overhead from the
106 * first transaction commit.
107 */
108static struct xlog_ticket *
109xlog_cil_ticket_alloc(
110 struct log *log)
111{
112 struct xlog_ticket *tic;
113
114 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
115 KM_SLEEP|KM_NOFS);
116 tic->t_trans_type = XFS_TRANS_CHECKPOINT;
117
118 /*
119 * set the current reservation to zero so we know to steal the basic
120 * transaction overhead reservation from the first transaction commit.
121 */
122 tic->t_curr_res = 0;
123 return tic;
124}
125
126/*
127 * After the first stage of log recovery is done, we know where the head and
128 * tail of the log are. We need this log initialisation done before we can
129 * initialise the first CIL checkpoint context.
130 *
131 * Here we allocate a log ticket to track space usage during a CIL push. This
132 * ticket is passed to xlog_write() directly so that we don't slowly leak log
133 * space by failing to account for space used by log headers and additional
134 * region headers for split regions.
135 */
136void
137xlog_cil_init_post_recovery(
138 struct log *log)
139{
140 if (!log->l_cilp)
141 return;
142
143 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
144 log->l_cilp->xc_ctx->sequence = 1;
145 log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
146 log->l_curr_block);
147}
148
149/*
150 * Format log item into a flat buffers
151 *
152 * For delayed logging, we need to hold a formatted buffer containing all the
153 * changes on the log item. This enables us to relog the item in memory and
154 * write it out asynchronously without needing to relock the object that was
155 * modified at the time it gets written into the iclog.
156 *
157 * This function builds a vector for the changes in each log item in the
158 * transaction. It then works out the length of the buffer needed for each log
159 * item, allocates them and formats the vector for the item into the buffer.
160 * The buffer is then attached to the log item are then inserted into the
161 * Committed Item List for tracking until the next checkpoint is written out.
162 *
163 * We don't set up region headers during this process; we simply copy the
164 * regions into the flat buffer. We can do this because we still have to do a
165 * formatting step to write the regions into the iclog buffer. Writing the
166 * ophdrs during the iclog write means that we can support splitting large
167 * regions across iclog boundares without needing a change in the format of the
168 * item/region encapsulation.
169 *
170 * Hence what we need to do now is change the rewrite the vector array to point
171 * to the copied region inside the buffer we just allocated. This allows us to
172 * format the regions into the iclog as though they are being formatted
173 * directly out of the objects themselves.
174 */
175static void
176xlog_cil_format_items(
177 struct log *log,
178 struct xfs_log_vec *log_vector)
179{
180 struct xfs_log_vec *lv;
181
182 ASSERT(log_vector);
183 for (lv = log_vector; lv; lv = lv->lv_next) {
184 void *ptr;
185 int index;
186 int len = 0;
187
188 /* build the vector array and calculate it's length */
189 IOP_FORMAT(lv->lv_item, lv->lv_iovecp);
190 for (index = 0; index < lv->lv_niovecs; index++)
191 len += lv->lv_iovecp[index].i_len;
192
193 lv->lv_buf_len = len;
194 lv->lv_buf = kmem_alloc(lv->lv_buf_len, KM_SLEEP|KM_NOFS);
195 ptr = lv->lv_buf;
196
197 for (index = 0; index < lv->lv_niovecs; index++) {
198 struct xfs_log_iovec *vec = &lv->lv_iovecp[index];
199
200 memcpy(ptr, vec->i_addr, vec->i_len);
201 vec->i_addr = ptr;
202 ptr += vec->i_len;
203 }
204 ASSERT(ptr == lv->lv_buf + lv->lv_buf_len);
205 }
206}
207
208/*
209 * Prepare the log item for insertion into the CIL. Calculate the difference in
210 * log space and vectors it will consume, and if it is a new item pin it as
211 * well.
212 */
213STATIC void
214xfs_cil_prepare_item(
215 struct log *log,
216 struct xfs_log_vec *lv,
217 int *len,
218 int *diff_iovecs)
219{
220 struct xfs_log_vec *old = lv->lv_item->li_lv;
221
222 if (old) {
223 /* existing lv on log item, space used is a delta */
224 ASSERT(!list_empty(&lv->lv_item->li_cil));
225 ASSERT(old->lv_buf && old->lv_buf_len && old->lv_niovecs);
226
227 *len += lv->lv_buf_len - old->lv_buf_len;
228 *diff_iovecs += lv->lv_niovecs - old->lv_niovecs;
229 kmem_free(old->lv_buf);
230 kmem_free(old);
231 } else {
232 /* new lv, must pin the log item */
233 ASSERT(!lv->lv_item->li_lv);
234 ASSERT(list_empty(&lv->lv_item->li_cil));
235
236 *len += lv->lv_buf_len;
237 *diff_iovecs += lv->lv_niovecs;
238 IOP_PIN(lv->lv_item);
239
240 }
241
242 /* attach new log vector to log item */
243 lv->lv_item->li_lv = lv;
244
245 /*
246 * If this is the first time the item is being committed to the
247 * CIL, store the sequence number on the log item so we can
248 * tell in future commits whether this is the first checkpoint
249 * the item is being committed into.
250 */
251 if (!lv->lv_item->li_seq)
252 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
253}
254
255/*
256 * Insert the log items into the CIL and calculate the difference in space
257 * consumed by the item. Add the space to the checkpoint ticket and calculate
258 * if the change requires additional log metadata. If it does, take that space
259 * as well. Remove the amount of space we addded to the checkpoint ticket from
260 * the current transaction ticket so that the accounting works out correctly.
261 */
262static void
263xlog_cil_insert_items(
264 struct log *log,
265 struct xfs_log_vec *log_vector,
266 struct xlog_ticket *ticket)
267{
268 struct xfs_cil *cil = log->l_cilp;
269 struct xfs_cil_ctx *ctx = cil->xc_ctx;
270 struct xfs_log_vec *lv;
271 int len = 0;
272 int diff_iovecs = 0;
273 int iclog_space;
274
275 ASSERT(log_vector);
276
277 /*
278 * Do all the accounting aggregation and switching of log vectors
279 * around in a separate loop to the insertion of items into the CIL.
280 * Then we can do a separate loop to update the CIL within a single
281 * lock/unlock pair. This reduces the number of round trips on the CIL
282 * lock from O(nr_logvectors) to O(1) and greatly reduces the overall
283 * hold time for the transaction commit.
284 *
285 * If this is the first time the item is being placed into the CIL in
286 * this context, pin it so it can't be written to disk until the CIL is
287 * flushed to the iclog and the iclog written to disk.
288 *
289 * We can do this safely because the context can't checkpoint until we
290 * are done so it doesn't matter exactly how we update the CIL.
291 */
292 for (lv = log_vector; lv; lv = lv->lv_next)
293 xfs_cil_prepare_item(log, lv, &len, &diff_iovecs);
294
295 /* account for space used by new iovec headers */
296 len += diff_iovecs * sizeof(xlog_op_header_t);
297
298 spin_lock(&cil->xc_cil_lock);
299
300 /* move the items to the tail of the CIL */
301 for (lv = log_vector; lv; lv = lv->lv_next)
302 list_move_tail(&lv->lv_item->li_cil, &cil->xc_cil);
303
304 ctx->nvecs += diff_iovecs;
305
306 /*
307 * Now transfer enough transaction reservation to the context ticket
308 * for the checkpoint. The context ticket is special - the unit
309 * reservation has to grow as well as the current reservation as we
310 * steal from tickets so we can correctly determine the space used
311 * during the transaction commit.
312 */
313 if (ctx->ticket->t_curr_res == 0) {
314 /* first commit in checkpoint, steal the header reservation */
315 ASSERT(ticket->t_curr_res >= ctx->ticket->t_unit_res + len);
316 ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
317 ticket->t_curr_res -= ctx->ticket->t_unit_res;
318 }
319
320 /* do we need space for more log record headers? */
321 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
322 if (len > 0 && (ctx->space_used / iclog_space !=
323 (ctx->space_used + len) / iclog_space)) {
324 int hdrs;
325
326 hdrs = (len + iclog_space - 1) / iclog_space;
327 /* need to take into account split region headers, too */
328 hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
329 ctx->ticket->t_unit_res += hdrs;
330 ctx->ticket->t_curr_res += hdrs;
331 ticket->t_curr_res -= hdrs;
332 ASSERT(ticket->t_curr_res >= len);
333 }
334 ticket->t_curr_res -= len;
335 ctx->space_used += len;
336
337 spin_unlock(&cil->xc_cil_lock);
338}
339
340static void
341xlog_cil_free_logvec(
342 struct xfs_log_vec *log_vector)
343{
344 struct xfs_log_vec *lv;
345
346 for (lv = log_vector; lv; ) {
347 struct xfs_log_vec *next = lv->lv_next;
348 kmem_free(lv->lv_buf);
349 kmem_free(lv);
350 lv = next;
351 }
352}
353
354/*
355 * Mark all items committed and clear busy extents. We free the log vector
356 * chains in a separate pass so that we unpin the log items as quickly as
357 * possible.
358 */
359static void
360xlog_cil_committed(
361 void *args,
362 int abort)
363{
364 struct xfs_cil_ctx *ctx = args;
365 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
366
367 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
368 ctx->start_lsn, abort);
369
370 xfs_alloc_busy_sort(&ctx->busy_extents);
371 xfs_alloc_busy_clear(mp, &ctx->busy_extents,
372 (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
373
374 spin_lock(&ctx->cil->xc_cil_lock);
375 list_del(&ctx->committing);
376 spin_unlock(&ctx->cil->xc_cil_lock);
377
378 xlog_cil_free_logvec(ctx->lv_chain);
379
380 if (!list_empty(&ctx->busy_extents)) {
381 ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
382
383 xfs_discard_extents(mp, &ctx->busy_extents);
384 xfs_alloc_busy_clear(mp, &ctx->busy_extents, false);
385 }
386
387 kmem_free(ctx);
388}
389
390/*
391 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
392 * is a background flush and so we can chose to ignore it. Otherwise, if the
393 * current sequence is the same as @push_seq we need to do a flush. If
394 * @push_seq is less than the current sequence, then it has already been
395 * flushed and we don't need to do anything - the caller will wait for it to
396 * complete if necessary.
397 *
398 * @push_seq is a value rather than a flag because that allows us to do an
399 * unlocked check of the sequence number for a match. Hence we can allows log
400 * forces to run racily and not issue pushes for the same sequence twice. If we
401 * get a race between multiple pushes for the same sequence they will block on
402 * the first one and then abort, hence avoiding needless pushes.
403 */
404STATIC int
405xlog_cil_push(
406 struct log *log,
407 xfs_lsn_t push_seq)
408{
409 struct xfs_cil *cil = log->l_cilp;
410 struct xfs_log_vec *lv;
411 struct xfs_cil_ctx *ctx;
412 struct xfs_cil_ctx *new_ctx;
413 struct xlog_in_core *commit_iclog;
414 struct xlog_ticket *tic;
415 int num_lv;
416 int num_iovecs;
417 int len;
418 int error = 0;
419 struct xfs_trans_header thdr;
420 struct xfs_log_iovec lhdr;
421 struct xfs_log_vec lvhdr = { NULL };
422 xfs_lsn_t commit_lsn;
423
424 if (!cil)
425 return 0;
426
427 ASSERT(!push_seq || push_seq <= cil->xc_ctx->sequence);
428
429 new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
430 new_ctx->ticket = xlog_cil_ticket_alloc(log);
431
432 /*
433 * Lock out transaction commit, but don't block for background pushes
434 * unless we are well over the CIL space limit. See the definition of
435 * XLOG_CIL_HARD_SPACE_LIMIT() for the full explanation of the logic
436 * used here.
437 */
438 if (!down_write_trylock(&cil->xc_ctx_lock)) {
439 if (!push_seq &&
440 cil->xc_ctx->space_used < XLOG_CIL_HARD_SPACE_LIMIT(log))
441 goto out_free_ticket;
442 down_write(&cil->xc_ctx_lock);
443 }
444 ctx = cil->xc_ctx;
445
446 /* check if we've anything to push */
447 if (list_empty(&cil->xc_cil))
448 goto out_skip;
449
450 /* check for spurious background flush */
451 if (!push_seq && cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
452 goto out_skip;
453
454 /* check for a previously pushed seqeunce */
455 if (push_seq && push_seq < cil->xc_ctx->sequence)
456 goto out_skip;
457
458 /*
459 * pull all the log vectors off the items in the CIL, and
460 * remove the items from the CIL. We don't need the CIL lock
461 * here because it's only needed on the transaction commit
462 * side which is currently locked out by the flush lock.
463 */
464 lv = NULL;
465 num_lv = 0;
466 num_iovecs = 0;
467 len = 0;
468 while (!list_empty(&cil->xc_cil)) {
469 struct xfs_log_item *item;
470 int i;
471
472 item = list_first_entry(&cil->xc_cil,
473 struct xfs_log_item, li_cil);
474 list_del_init(&item->li_cil);
475 if (!ctx->lv_chain)
476 ctx->lv_chain = item->li_lv;
477 else
478 lv->lv_next = item->li_lv;
479 lv = item->li_lv;
480 item->li_lv = NULL;
481
482 num_lv++;
483 num_iovecs += lv->lv_niovecs;
484 for (i = 0; i < lv->lv_niovecs; i++)
485 len += lv->lv_iovecp[i].i_len;
486 }
487
488 /*
489 * initialise the new context and attach it to the CIL. Then attach
490 * the current context to the CIL committing lsit so it can be found
491 * during log forces to extract the commit lsn of the sequence that
492 * needs to be forced.
493 */
494 INIT_LIST_HEAD(&new_ctx->committing);
495 INIT_LIST_HEAD(&new_ctx->busy_extents);
496 new_ctx->sequence = ctx->sequence + 1;
497 new_ctx->cil = cil;
498 cil->xc_ctx = new_ctx;
499
500 /*
501 * mirror the new sequence into the cil structure so that we can do
502 * unlocked checks against the current sequence in log forces without
503 * risking deferencing a freed context pointer.
504 */
505 cil->xc_current_sequence = new_ctx->sequence;
506
507 /*
508 * The switch is now done, so we can drop the context lock and move out
509 * of a shared context. We can't just go straight to the commit record,
510 * though - we need to synchronise with previous and future commits so
511 * that the commit records are correctly ordered in the log to ensure
512 * that we process items during log IO completion in the correct order.
513 *
514 * For example, if we get an EFI in one checkpoint and the EFD in the
515 * next (e.g. due to log forces), we do not want the checkpoint with
516 * the EFD to be committed before the checkpoint with the EFI. Hence
517 * we must strictly order the commit records of the checkpoints so
518 * that: a) the checkpoint callbacks are attached to the iclogs in the
519 * correct order; and b) the checkpoints are replayed in correct order
520 * in log recovery.
521 *
522 * Hence we need to add this context to the committing context list so
523 * that higher sequences will wait for us to write out a commit record
524 * before they do.
525 */
526 spin_lock(&cil->xc_cil_lock);
527 list_add(&ctx->committing, &cil->xc_committing);
528 spin_unlock(&cil->xc_cil_lock);
529 up_write(&cil->xc_ctx_lock);
530
531 /*
532 * Build a checkpoint transaction header and write it to the log to
533 * begin the transaction. We need to account for the space used by the
534 * transaction header here as it is not accounted for in xlog_write().
535 *
536 * The LSN we need to pass to the log items on transaction commit is
537 * the LSN reported by the first log vector write. If we use the commit
538 * record lsn then we can move the tail beyond the grant write head.
539 */
540 tic = ctx->ticket;
541 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
542 thdr.th_type = XFS_TRANS_CHECKPOINT;
543 thdr.th_tid = tic->t_tid;
544 thdr.th_num_items = num_iovecs;
545 lhdr.i_addr = &thdr;
546 lhdr.i_len = sizeof(xfs_trans_header_t);
547 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
548 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
549
550 lvhdr.lv_niovecs = 1;
551 lvhdr.lv_iovecp = &lhdr;
552 lvhdr.lv_next = ctx->lv_chain;
553
554 error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
555 if (error)
556 goto out_abort_free_ticket;
557
558 /*
559 * now that we've written the checkpoint into the log, strictly
560 * order the commit records so replay will get them in the right order.
561 */
562restart:
563 spin_lock(&cil->xc_cil_lock);
564 list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
565 /*
566 * Higher sequences will wait for this one so skip them.
567 * Don't wait for own own sequence, either.
568 */
569 if (new_ctx->sequence >= ctx->sequence)
570 continue;
571 if (!new_ctx->commit_lsn) {
572 /*
573 * It is still being pushed! Wait for the push to
574 * complete, then start again from the beginning.
575 */
576 xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
577 goto restart;
578 }
579 }
580 spin_unlock(&cil->xc_cil_lock);
581
582 /* xfs_log_done always frees the ticket on error. */
583 commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
584 if (commit_lsn == -1)
585 goto out_abort;
586
587 /* attach all the transactions w/ busy extents to iclog */
588 ctx->log_cb.cb_func = xlog_cil_committed;
589 ctx->log_cb.cb_arg = ctx;
590 error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
591 if (error)
592 goto out_abort;
593
594 /*
595 * now the checkpoint commit is complete and we've attached the
596 * callbacks to the iclog we can assign the commit LSN to the context
597 * and wake up anyone who is waiting for the commit to complete.
598 */
599 spin_lock(&cil->xc_cil_lock);
600 ctx->commit_lsn = commit_lsn;
601 wake_up_all(&cil->xc_commit_wait);
602 spin_unlock(&cil->xc_cil_lock);
603
604 /* release the hounds! */
605 return xfs_log_release_iclog(log->l_mp, commit_iclog);
606
607out_skip:
608 up_write(&cil->xc_ctx_lock);
609out_free_ticket:
610 xfs_log_ticket_put(new_ctx->ticket);
611 kmem_free(new_ctx);
612 return 0;
613
614out_abort_free_ticket:
615 xfs_log_ticket_put(tic);
616out_abort:
617 xlog_cil_committed(ctx, XFS_LI_ABORTED);
618 return XFS_ERROR(EIO);
619}
620
621/*
622 * Commit a transaction with the given vector to the Committed Item List.
623 *
624 * To do this, we need to format the item, pin it in memory if required and
625 * account for the space used by the transaction. Once we have done that we
626 * need to release the unused reservation for the transaction, attach the
627 * transaction to the checkpoint context so we carry the busy extents through
628 * to checkpoint completion, and then unlock all the items in the transaction.
629 *
630 * For more specific information about the order of operations in
631 * xfs_log_commit_cil() please refer to the comments in
632 * xfs_trans_commit_iclog().
633 *
634 * Called with the context lock already held in read mode to lock out
635 * background commit, returns without it held once background commits are
636 * allowed again.
637 */
638void
639xfs_log_commit_cil(
640 struct xfs_mount *mp,
641 struct xfs_trans *tp,
642 struct xfs_log_vec *log_vector,
643 xfs_lsn_t *commit_lsn,
644 int flags)
645{
646 struct log *log = mp->m_log;
647 int log_flags = 0;
648 int push = 0;
649
650 if (flags & XFS_TRANS_RELEASE_LOG_RES)
651 log_flags = XFS_LOG_REL_PERM_RESERV;
652
653 /*
654 * do all the hard work of formatting items (including memory
655 * allocation) outside the CIL context lock. This prevents stalling CIL
656 * pushes when we are low on memory and a transaction commit spends a
657 * lot of time in memory reclaim.
658 */
659 xlog_cil_format_items(log, log_vector);
660
661 /* lock out background commit */
662 down_read(&log->l_cilp->xc_ctx_lock);
663 if (commit_lsn)
664 *commit_lsn = log->l_cilp->xc_ctx->sequence;
665
666 xlog_cil_insert_items(log, log_vector, tp->t_ticket);
667
668 /* check we didn't blow the reservation */
669 if (tp->t_ticket->t_curr_res < 0)
670 xlog_print_tic_res(log->l_mp, tp->t_ticket);
671
672 /* attach the transaction to the CIL if it has any busy extents */
673 if (!list_empty(&tp->t_busy)) {
674 spin_lock(&log->l_cilp->xc_cil_lock);
675 list_splice_init(&tp->t_busy,
676 &log->l_cilp->xc_ctx->busy_extents);
677 spin_unlock(&log->l_cilp->xc_cil_lock);
678 }
679
680 tp->t_commit_lsn = *commit_lsn;
681 xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
682 xfs_trans_unreserve_and_mod_sb(tp);
683
684 /*
685 * Once all the items of the transaction have been copied to the CIL,
686 * the items can be unlocked and freed.
687 *
688 * This needs to be done before we drop the CIL context lock because we
689 * have to update state in the log items and unlock them before they go
690 * to disk. If we don't, then the CIL checkpoint can race with us and
691 * we can run checkpoint completion before we've updated and unlocked
692 * the log items. This affects (at least) processing of stale buffers,
693 * inodes and EFIs.
694 */
695 xfs_trans_free_items(tp, *commit_lsn, 0);
696
697 /* check for background commit before unlock */
698 if (log->l_cilp->xc_ctx->space_used > XLOG_CIL_SPACE_LIMIT(log))
699 push = 1;
700
701 up_read(&log->l_cilp->xc_ctx_lock);
702
703 /*
704 * We need to push CIL every so often so we don't cache more than we
705 * can fit in the log. The limit really is that a checkpoint can't be
706 * more than half the log (the current checkpoint is not allowed to
707 * overwrite the previous checkpoint), but commit latency and memory
708 * usage limit this to a smaller size in most cases.
709 */
710 if (push)
711 xlog_cil_push(log, 0);
712}
713
714/*
715 * Conditionally push the CIL based on the sequence passed in.
716 *
717 * We only need to push if we haven't already pushed the sequence
718 * number given. Hence the only time we will trigger a push here is
719 * if the push sequence is the same as the current context.
720 *
721 * We return the current commit lsn to allow the callers to determine if a
722 * iclog flush is necessary following this call.
723 *
724 * XXX: Initially, just push the CIL unconditionally and return whatever
725 * commit lsn is there. It'll be empty, so this is broken for now.
726 */
727xfs_lsn_t
728xlog_cil_force_lsn(
729 struct log *log,
730 xfs_lsn_t sequence)
731{
732 struct xfs_cil *cil = log->l_cilp;
733 struct xfs_cil_ctx *ctx;
734 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
735
736 ASSERT(sequence <= cil->xc_current_sequence);
737
738 /*
739 * check to see if we need to force out the current context.
740 * xlog_cil_push() handles racing pushes for the same sequence,
741 * so no need to deal with it here.
742 */
743 if (sequence == cil->xc_current_sequence)
744 xlog_cil_push(log, sequence);
745
746 /*
747 * See if we can find a previous sequence still committing.
748 * We need to wait for all previous sequence commits to complete
749 * before allowing the force of push_seq to go ahead. Hence block
750 * on commits for those as well.
751 */
752restart:
753 spin_lock(&cil->xc_cil_lock);
754 list_for_each_entry(ctx, &cil->xc_committing, committing) {
755 if (ctx->sequence > sequence)
756 continue;
757 if (!ctx->commit_lsn) {
758 /*
759 * It is still being pushed! Wait for the push to
760 * complete, then start again from the beginning.
761 */
762 xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
763 goto restart;
764 }
765 if (ctx->sequence != sequence)
766 continue;
767 /* found it! */
768 commit_lsn = ctx->commit_lsn;
769 }
770 spin_unlock(&cil->xc_cil_lock);
771 return commit_lsn;
772}
773
774/*
775 * Check if the current log item was first committed in this sequence.
776 * We can't rely on just the log item being in the CIL, we have to check
777 * the recorded commit sequence number.
778 *
779 * Note: for this to be used in a non-racy manner, it has to be called with
780 * CIL flushing locked out. As a result, it should only be used during the
781 * transaction commit process when deciding what to format into the item.
782 */
783bool
784xfs_log_item_in_current_chkpt(
785 struct xfs_log_item *lip)
786{
787 struct xfs_cil_ctx *ctx;
788
789 if (!(lip->li_mountp->m_flags & XFS_MOUNT_DELAYLOG))
790 return false;
791 if (list_empty(&lip->li_cil))
792 return false;
793
794 ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
795
796 /*
797 * li_seq is written on the first commit of a log item to record the
798 * first checkpoint it is written to. Hence if it is different to the
799 * current sequence, we're in a new checkpoint.
800 */
801 if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
802 return false;
803 return true;
804}
1/*
2 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it would be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write the Free Software Foundation,
15 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
16 */
17
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_log.h"
22#include "xfs_trans.h"
23#include "xfs_trans_priv.h"
24#include "xfs_log_priv.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
27#include "xfs_mount.h"
28#include "xfs_error.h"
29#include "xfs_alloc.h"
30#include "xfs_extent_busy.h"
31#include "xfs_discard.h"
32
33/*
34 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
35 * recover, so we don't allow failure here. Also, we allocate in a context that
36 * we don't want to be issuing transactions from, so we need to tell the
37 * allocation code this as well.
38 *
39 * We don't reserve any space for the ticket - we are going to steal whatever
40 * space we require from transactions as they commit. To ensure we reserve all
41 * the space required, we need to set the current reservation of the ticket to
42 * zero so that we know to steal the initial transaction overhead from the
43 * first transaction commit.
44 */
45static struct xlog_ticket *
46xlog_cil_ticket_alloc(
47 struct xlog *log)
48{
49 struct xlog_ticket *tic;
50
51 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
52 KM_SLEEP|KM_NOFS);
53 tic->t_trans_type = XFS_TRANS_CHECKPOINT;
54
55 /*
56 * set the current reservation to zero so we know to steal the basic
57 * transaction overhead reservation from the first transaction commit.
58 */
59 tic->t_curr_res = 0;
60 return tic;
61}
62
63/*
64 * After the first stage of log recovery is done, we know where the head and
65 * tail of the log are. We need this log initialisation done before we can
66 * initialise the first CIL checkpoint context.
67 *
68 * Here we allocate a log ticket to track space usage during a CIL push. This
69 * ticket is passed to xlog_write() directly so that we don't slowly leak log
70 * space by failing to account for space used by log headers and additional
71 * region headers for split regions.
72 */
73void
74xlog_cil_init_post_recovery(
75 struct xlog *log)
76{
77 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
78 log->l_cilp->xc_ctx->sequence = 1;
79 log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
80 log->l_curr_block);
81}
82
83/*
84 * Format log item into a flat buffers
85 *
86 * For delayed logging, we need to hold a formatted buffer containing all the
87 * changes on the log item. This enables us to relog the item in memory and
88 * write it out asynchronously without needing to relock the object that was
89 * modified at the time it gets written into the iclog.
90 *
91 * This function builds a vector for the changes in each log item in the
92 * transaction. It then works out the length of the buffer needed for each log
93 * item, allocates them and formats the vector for the item into the buffer.
94 * The buffer is then attached to the log item are then inserted into the
95 * Committed Item List for tracking until the next checkpoint is written out.
96 *
97 * We don't set up region headers during this process; we simply copy the
98 * regions into the flat buffer. We can do this because we still have to do a
99 * formatting step to write the regions into the iclog buffer. Writing the
100 * ophdrs during the iclog write means that we can support splitting large
101 * regions across iclog boundares without needing a change in the format of the
102 * item/region encapsulation.
103 *
104 * Hence what we need to do now is change the rewrite the vector array to point
105 * to the copied region inside the buffer we just allocated. This allows us to
106 * format the regions into the iclog as though they are being formatted
107 * directly out of the objects themselves.
108 */
109static struct xfs_log_vec *
110xlog_cil_prepare_log_vecs(
111 struct xfs_trans *tp)
112{
113 struct xfs_log_item_desc *lidp;
114 struct xfs_log_vec *lv = NULL;
115 struct xfs_log_vec *ret_lv = NULL;
116
117
118 /* Bail out if we didn't find a log item. */
119 if (list_empty(&tp->t_items)) {
120 ASSERT(0);
121 return NULL;
122 }
123
124 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
125 struct xfs_log_vec *new_lv;
126 void *ptr;
127 int index;
128 int len = 0;
129 uint niovecs;
130
131 /* Skip items which aren't dirty in this transaction. */
132 if (!(lidp->lid_flags & XFS_LID_DIRTY))
133 continue;
134
135 /* Skip items that do not have any vectors for writing */
136 niovecs = IOP_SIZE(lidp->lid_item);
137 if (!niovecs)
138 continue;
139
140 new_lv = kmem_zalloc(sizeof(*new_lv) +
141 niovecs * sizeof(struct xfs_log_iovec),
142 KM_SLEEP);
143
144 /* The allocated iovec region lies beyond the log vector. */
145 new_lv->lv_iovecp = (struct xfs_log_iovec *)&new_lv[1];
146 new_lv->lv_niovecs = niovecs;
147 new_lv->lv_item = lidp->lid_item;
148
149 /* build the vector array and calculate it's length */
150 IOP_FORMAT(new_lv->lv_item, new_lv->lv_iovecp);
151 for (index = 0; index < new_lv->lv_niovecs; index++)
152 len += new_lv->lv_iovecp[index].i_len;
153
154 new_lv->lv_buf_len = len;
155 new_lv->lv_buf = kmem_alloc(new_lv->lv_buf_len,
156 KM_SLEEP|KM_NOFS);
157 ptr = new_lv->lv_buf;
158
159 for (index = 0; index < new_lv->lv_niovecs; index++) {
160 struct xfs_log_iovec *vec = &new_lv->lv_iovecp[index];
161
162 memcpy(ptr, vec->i_addr, vec->i_len);
163 vec->i_addr = ptr;
164 ptr += vec->i_len;
165 }
166 ASSERT(ptr == new_lv->lv_buf + new_lv->lv_buf_len);
167
168 if (!ret_lv)
169 ret_lv = new_lv;
170 else
171 lv->lv_next = new_lv;
172 lv = new_lv;
173 }
174
175 return ret_lv;
176}
177
178/*
179 * Prepare the log item for insertion into the CIL. Calculate the difference in
180 * log space and vectors it will consume, and if it is a new item pin it as
181 * well.
182 */
183STATIC void
184xfs_cil_prepare_item(
185 struct xlog *log,
186 struct xfs_log_vec *lv,
187 int *len,
188 int *diff_iovecs)
189{
190 struct xfs_log_vec *old = lv->lv_item->li_lv;
191
192 if (old) {
193 /* existing lv on log item, space used is a delta */
194 ASSERT(!list_empty(&lv->lv_item->li_cil));
195 ASSERT(old->lv_buf && old->lv_buf_len && old->lv_niovecs);
196
197 *len += lv->lv_buf_len - old->lv_buf_len;
198 *diff_iovecs += lv->lv_niovecs - old->lv_niovecs;
199 kmem_free(old->lv_buf);
200 kmem_free(old);
201 } else {
202 /* new lv, must pin the log item */
203 ASSERT(!lv->lv_item->li_lv);
204 ASSERT(list_empty(&lv->lv_item->li_cil));
205
206 *len += lv->lv_buf_len;
207 *diff_iovecs += lv->lv_niovecs;
208 IOP_PIN(lv->lv_item);
209
210 }
211
212 /* attach new log vector to log item */
213 lv->lv_item->li_lv = lv;
214
215 /*
216 * If this is the first time the item is being committed to the
217 * CIL, store the sequence number on the log item so we can
218 * tell in future commits whether this is the first checkpoint
219 * the item is being committed into.
220 */
221 if (!lv->lv_item->li_seq)
222 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
223}
224
225/*
226 * Insert the log items into the CIL and calculate the difference in space
227 * consumed by the item. Add the space to the checkpoint ticket and calculate
228 * if the change requires additional log metadata. If it does, take that space
229 * as well. Remove the amount of space we added to the checkpoint ticket from
230 * the current transaction ticket so that the accounting works out correctly.
231 */
232static void
233xlog_cil_insert_items(
234 struct xlog *log,
235 struct xfs_log_vec *log_vector,
236 struct xlog_ticket *ticket)
237{
238 struct xfs_cil *cil = log->l_cilp;
239 struct xfs_cil_ctx *ctx = cil->xc_ctx;
240 struct xfs_log_vec *lv;
241 int len = 0;
242 int diff_iovecs = 0;
243 int iclog_space;
244
245 ASSERT(log_vector);
246
247 /*
248 * Do all the accounting aggregation and switching of log vectors
249 * around in a separate loop to the insertion of items into the CIL.
250 * Then we can do a separate loop to update the CIL within a single
251 * lock/unlock pair. This reduces the number of round trips on the CIL
252 * lock from O(nr_logvectors) to O(1) and greatly reduces the overall
253 * hold time for the transaction commit.
254 *
255 * If this is the first time the item is being placed into the CIL in
256 * this context, pin it so it can't be written to disk until the CIL is
257 * flushed to the iclog and the iclog written to disk.
258 *
259 * We can do this safely because the context can't checkpoint until we
260 * are done so it doesn't matter exactly how we update the CIL.
261 */
262 for (lv = log_vector; lv; lv = lv->lv_next)
263 xfs_cil_prepare_item(log, lv, &len, &diff_iovecs);
264
265 /* account for space used by new iovec headers */
266 len += diff_iovecs * sizeof(xlog_op_header_t);
267
268 spin_lock(&cil->xc_cil_lock);
269
270 /* move the items to the tail of the CIL */
271 for (lv = log_vector; lv; lv = lv->lv_next)
272 list_move_tail(&lv->lv_item->li_cil, &cil->xc_cil);
273
274 ctx->nvecs += diff_iovecs;
275
276 /*
277 * Now transfer enough transaction reservation to the context ticket
278 * for the checkpoint. The context ticket is special - the unit
279 * reservation has to grow as well as the current reservation as we
280 * steal from tickets so we can correctly determine the space used
281 * during the transaction commit.
282 */
283 if (ctx->ticket->t_curr_res == 0) {
284 /* first commit in checkpoint, steal the header reservation */
285 ASSERT(ticket->t_curr_res >= ctx->ticket->t_unit_res + len);
286 ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
287 ticket->t_curr_res -= ctx->ticket->t_unit_res;
288 }
289
290 /* do we need space for more log record headers? */
291 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
292 if (len > 0 && (ctx->space_used / iclog_space !=
293 (ctx->space_used + len) / iclog_space)) {
294 int hdrs;
295
296 hdrs = (len + iclog_space - 1) / iclog_space;
297 /* need to take into account split region headers, too */
298 hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
299 ctx->ticket->t_unit_res += hdrs;
300 ctx->ticket->t_curr_res += hdrs;
301 ticket->t_curr_res -= hdrs;
302 ASSERT(ticket->t_curr_res >= len);
303 }
304 ticket->t_curr_res -= len;
305 ctx->space_used += len;
306
307 spin_unlock(&cil->xc_cil_lock);
308}
309
310static void
311xlog_cil_free_logvec(
312 struct xfs_log_vec *log_vector)
313{
314 struct xfs_log_vec *lv;
315
316 for (lv = log_vector; lv; ) {
317 struct xfs_log_vec *next = lv->lv_next;
318 kmem_free(lv->lv_buf);
319 kmem_free(lv);
320 lv = next;
321 }
322}
323
324/*
325 * Mark all items committed and clear busy extents. We free the log vector
326 * chains in a separate pass so that we unpin the log items as quickly as
327 * possible.
328 */
329static void
330xlog_cil_committed(
331 void *args,
332 int abort)
333{
334 struct xfs_cil_ctx *ctx = args;
335 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
336
337 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
338 ctx->start_lsn, abort);
339
340 xfs_extent_busy_sort(&ctx->busy_extents);
341 xfs_extent_busy_clear(mp, &ctx->busy_extents,
342 (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
343
344 spin_lock(&ctx->cil->xc_cil_lock);
345 list_del(&ctx->committing);
346 spin_unlock(&ctx->cil->xc_cil_lock);
347
348 xlog_cil_free_logvec(ctx->lv_chain);
349
350 if (!list_empty(&ctx->busy_extents)) {
351 ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
352
353 xfs_discard_extents(mp, &ctx->busy_extents);
354 xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
355 }
356
357 kmem_free(ctx);
358}
359
360/*
361 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
362 * is a background flush and so we can chose to ignore it. Otherwise, if the
363 * current sequence is the same as @push_seq we need to do a flush. If
364 * @push_seq is less than the current sequence, then it has already been
365 * flushed and we don't need to do anything - the caller will wait for it to
366 * complete if necessary.
367 *
368 * @push_seq is a value rather than a flag because that allows us to do an
369 * unlocked check of the sequence number for a match. Hence we can allows log
370 * forces to run racily and not issue pushes for the same sequence twice. If we
371 * get a race between multiple pushes for the same sequence they will block on
372 * the first one and then abort, hence avoiding needless pushes.
373 */
374STATIC int
375xlog_cil_push(
376 struct xlog *log)
377{
378 struct xfs_cil *cil = log->l_cilp;
379 struct xfs_log_vec *lv;
380 struct xfs_cil_ctx *ctx;
381 struct xfs_cil_ctx *new_ctx;
382 struct xlog_in_core *commit_iclog;
383 struct xlog_ticket *tic;
384 int num_lv;
385 int num_iovecs;
386 int len;
387 int error = 0;
388 struct xfs_trans_header thdr;
389 struct xfs_log_iovec lhdr;
390 struct xfs_log_vec lvhdr = { NULL };
391 xfs_lsn_t commit_lsn;
392 xfs_lsn_t push_seq;
393
394 if (!cil)
395 return 0;
396
397 new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
398 new_ctx->ticket = xlog_cil_ticket_alloc(log);
399
400 down_write(&cil->xc_ctx_lock);
401 ctx = cil->xc_ctx;
402
403 spin_lock(&cil->xc_cil_lock);
404 push_seq = cil->xc_push_seq;
405 ASSERT(push_seq <= ctx->sequence);
406
407 /*
408 * Check if we've anything to push. If there is nothing, then we don't
409 * move on to a new sequence number and so we have to be able to push
410 * this sequence again later.
411 */
412 if (list_empty(&cil->xc_cil)) {
413 cil->xc_push_seq = 0;
414 spin_unlock(&cil->xc_cil_lock);
415 goto out_skip;
416 }
417 spin_unlock(&cil->xc_cil_lock);
418
419
420 /* check for a previously pushed seqeunce */
421 if (push_seq < cil->xc_ctx->sequence)
422 goto out_skip;
423
424 /*
425 * pull all the log vectors off the items in the CIL, and
426 * remove the items from the CIL. We don't need the CIL lock
427 * here because it's only needed on the transaction commit
428 * side which is currently locked out by the flush lock.
429 */
430 lv = NULL;
431 num_lv = 0;
432 num_iovecs = 0;
433 len = 0;
434 while (!list_empty(&cil->xc_cil)) {
435 struct xfs_log_item *item;
436 int i;
437
438 item = list_first_entry(&cil->xc_cil,
439 struct xfs_log_item, li_cil);
440 list_del_init(&item->li_cil);
441 if (!ctx->lv_chain)
442 ctx->lv_chain = item->li_lv;
443 else
444 lv->lv_next = item->li_lv;
445 lv = item->li_lv;
446 item->li_lv = NULL;
447
448 num_lv++;
449 num_iovecs += lv->lv_niovecs;
450 for (i = 0; i < lv->lv_niovecs; i++)
451 len += lv->lv_iovecp[i].i_len;
452 }
453
454 /*
455 * initialise the new context and attach it to the CIL. Then attach
456 * the current context to the CIL committing lsit so it can be found
457 * during log forces to extract the commit lsn of the sequence that
458 * needs to be forced.
459 */
460 INIT_LIST_HEAD(&new_ctx->committing);
461 INIT_LIST_HEAD(&new_ctx->busy_extents);
462 new_ctx->sequence = ctx->sequence + 1;
463 new_ctx->cil = cil;
464 cil->xc_ctx = new_ctx;
465
466 /*
467 * mirror the new sequence into the cil structure so that we can do
468 * unlocked checks against the current sequence in log forces without
469 * risking deferencing a freed context pointer.
470 */
471 cil->xc_current_sequence = new_ctx->sequence;
472
473 /*
474 * The switch is now done, so we can drop the context lock and move out
475 * of a shared context. We can't just go straight to the commit record,
476 * though - we need to synchronise with previous and future commits so
477 * that the commit records are correctly ordered in the log to ensure
478 * that we process items during log IO completion in the correct order.
479 *
480 * For example, if we get an EFI in one checkpoint and the EFD in the
481 * next (e.g. due to log forces), we do not want the checkpoint with
482 * the EFD to be committed before the checkpoint with the EFI. Hence
483 * we must strictly order the commit records of the checkpoints so
484 * that: a) the checkpoint callbacks are attached to the iclogs in the
485 * correct order; and b) the checkpoints are replayed in correct order
486 * in log recovery.
487 *
488 * Hence we need to add this context to the committing context list so
489 * that higher sequences will wait for us to write out a commit record
490 * before they do.
491 */
492 spin_lock(&cil->xc_cil_lock);
493 list_add(&ctx->committing, &cil->xc_committing);
494 spin_unlock(&cil->xc_cil_lock);
495 up_write(&cil->xc_ctx_lock);
496
497 /*
498 * Build a checkpoint transaction header and write it to the log to
499 * begin the transaction. We need to account for the space used by the
500 * transaction header here as it is not accounted for in xlog_write().
501 *
502 * The LSN we need to pass to the log items on transaction commit is
503 * the LSN reported by the first log vector write. If we use the commit
504 * record lsn then we can move the tail beyond the grant write head.
505 */
506 tic = ctx->ticket;
507 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
508 thdr.th_type = XFS_TRANS_CHECKPOINT;
509 thdr.th_tid = tic->t_tid;
510 thdr.th_num_items = num_iovecs;
511 lhdr.i_addr = &thdr;
512 lhdr.i_len = sizeof(xfs_trans_header_t);
513 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
514 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
515
516 lvhdr.lv_niovecs = 1;
517 lvhdr.lv_iovecp = &lhdr;
518 lvhdr.lv_next = ctx->lv_chain;
519
520 error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
521 if (error)
522 goto out_abort_free_ticket;
523
524 /*
525 * now that we've written the checkpoint into the log, strictly
526 * order the commit records so replay will get them in the right order.
527 */
528restart:
529 spin_lock(&cil->xc_cil_lock);
530 list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
531 /*
532 * Higher sequences will wait for this one so skip them.
533 * Don't wait for own own sequence, either.
534 */
535 if (new_ctx->sequence >= ctx->sequence)
536 continue;
537 if (!new_ctx->commit_lsn) {
538 /*
539 * It is still being pushed! Wait for the push to
540 * complete, then start again from the beginning.
541 */
542 xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
543 goto restart;
544 }
545 }
546 spin_unlock(&cil->xc_cil_lock);
547
548 /* xfs_log_done always frees the ticket on error. */
549 commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
550 if (commit_lsn == -1)
551 goto out_abort;
552
553 /* attach all the transactions w/ busy extents to iclog */
554 ctx->log_cb.cb_func = xlog_cil_committed;
555 ctx->log_cb.cb_arg = ctx;
556 error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
557 if (error)
558 goto out_abort;
559
560 /*
561 * now the checkpoint commit is complete and we've attached the
562 * callbacks to the iclog we can assign the commit LSN to the context
563 * and wake up anyone who is waiting for the commit to complete.
564 */
565 spin_lock(&cil->xc_cil_lock);
566 ctx->commit_lsn = commit_lsn;
567 wake_up_all(&cil->xc_commit_wait);
568 spin_unlock(&cil->xc_cil_lock);
569
570 /* release the hounds! */
571 return xfs_log_release_iclog(log->l_mp, commit_iclog);
572
573out_skip:
574 up_write(&cil->xc_ctx_lock);
575 xfs_log_ticket_put(new_ctx->ticket);
576 kmem_free(new_ctx);
577 return 0;
578
579out_abort_free_ticket:
580 xfs_log_ticket_put(tic);
581out_abort:
582 xlog_cil_committed(ctx, XFS_LI_ABORTED);
583 return XFS_ERROR(EIO);
584}
585
586static void
587xlog_cil_push_work(
588 struct work_struct *work)
589{
590 struct xfs_cil *cil = container_of(work, struct xfs_cil,
591 xc_push_work);
592 xlog_cil_push(cil->xc_log);
593}
594
595/*
596 * We need to push CIL every so often so we don't cache more than we can fit in
597 * the log. The limit really is that a checkpoint can't be more than half the
598 * log (the current checkpoint is not allowed to overwrite the previous
599 * checkpoint), but commit latency and memory usage limit this to a smaller
600 * size.
601 */
602static void
603xlog_cil_push_background(
604 struct xlog *log)
605{
606 struct xfs_cil *cil = log->l_cilp;
607
608 /*
609 * The cil won't be empty because we are called while holding the
610 * context lock so whatever we added to the CIL will still be there
611 */
612 ASSERT(!list_empty(&cil->xc_cil));
613
614 /*
615 * don't do a background push if we haven't used up all the
616 * space available yet.
617 */
618 if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
619 return;
620
621 spin_lock(&cil->xc_cil_lock);
622 if (cil->xc_push_seq < cil->xc_current_sequence) {
623 cil->xc_push_seq = cil->xc_current_sequence;
624 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
625 }
626 spin_unlock(&cil->xc_cil_lock);
627
628}
629
630static void
631xlog_cil_push_foreground(
632 struct xlog *log,
633 xfs_lsn_t push_seq)
634{
635 struct xfs_cil *cil = log->l_cilp;
636
637 if (!cil)
638 return;
639
640 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
641
642 /* start on any pending background push to minimise wait time on it */
643 flush_work(&cil->xc_push_work);
644
645 /*
646 * If the CIL is empty or we've already pushed the sequence then
647 * there's no work we need to do.
648 */
649 spin_lock(&cil->xc_cil_lock);
650 if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
651 spin_unlock(&cil->xc_cil_lock);
652 return;
653 }
654
655 cil->xc_push_seq = push_seq;
656 spin_unlock(&cil->xc_cil_lock);
657
658 /* do the push now */
659 xlog_cil_push(log);
660}
661
662/*
663 * Commit a transaction with the given vector to the Committed Item List.
664 *
665 * To do this, we need to format the item, pin it in memory if required and
666 * account for the space used by the transaction. Once we have done that we
667 * need to release the unused reservation for the transaction, attach the
668 * transaction to the checkpoint context so we carry the busy extents through
669 * to checkpoint completion, and then unlock all the items in the transaction.
670 *
671 * For more specific information about the order of operations in
672 * xfs_log_commit_cil() please refer to the comments in
673 * xfs_trans_commit_iclog().
674 *
675 * Called with the context lock already held in read mode to lock out
676 * background commit, returns without it held once background commits are
677 * allowed again.
678 */
679int
680xfs_log_commit_cil(
681 struct xfs_mount *mp,
682 struct xfs_trans *tp,
683 xfs_lsn_t *commit_lsn,
684 int flags)
685{
686 struct xlog *log = mp->m_log;
687 int log_flags = 0;
688 struct xfs_log_vec *log_vector;
689
690 if (flags & XFS_TRANS_RELEASE_LOG_RES)
691 log_flags = XFS_LOG_REL_PERM_RESERV;
692
693 /*
694 * Do all the hard work of formatting items (including memory
695 * allocation) outside the CIL context lock. This prevents stalling CIL
696 * pushes when we are low on memory and a transaction commit spends a
697 * lot of time in memory reclaim.
698 */
699 log_vector = xlog_cil_prepare_log_vecs(tp);
700 if (!log_vector)
701 return ENOMEM;
702
703 /* lock out background commit */
704 down_read(&log->l_cilp->xc_ctx_lock);
705 if (commit_lsn)
706 *commit_lsn = log->l_cilp->xc_ctx->sequence;
707
708 xlog_cil_insert_items(log, log_vector, tp->t_ticket);
709
710 /* check we didn't blow the reservation */
711 if (tp->t_ticket->t_curr_res < 0)
712 xlog_print_tic_res(log->l_mp, tp->t_ticket);
713
714 /* attach the transaction to the CIL if it has any busy extents */
715 if (!list_empty(&tp->t_busy)) {
716 spin_lock(&log->l_cilp->xc_cil_lock);
717 list_splice_init(&tp->t_busy,
718 &log->l_cilp->xc_ctx->busy_extents);
719 spin_unlock(&log->l_cilp->xc_cil_lock);
720 }
721
722 tp->t_commit_lsn = *commit_lsn;
723 xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
724 xfs_trans_unreserve_and_mod_sb(tp);
725
726 /*
727 * Once all the items of the transaction have been copied to the CIL,
728 * the items can be unlocked and freed.
729 *
730 * This needs to be done before we drop the CIL context lock because we
731 * have to update state in the log items and unlock them before they go
732 * to disk. If we don't, then the CIL checkpoint can race with us and
733 * we can run checkpoint completion before we've updated and unlocked
734 * the log items. This affects (at least) processing of stale buffers,
735 * inodes and EFIs.
736 */
737 xfs_trans_free_items(tp, *commit_lsn, 0);
738
739 xlog_cil_push_background(log);
740
741 up_read(&log->l_cilp->xc_ctx_lock);
742 return 0;
743}
744
745/*
746 * Conditionally push the CIL based on the sequence passed in.
747 *
748 * We only need to push if we haven't already pushed the sequence
749 * number given. Hence the only time we will trigger a push here is
750 * if the push sequence is the same as the current context.
751 *
752 * We return the current commit lsn to allow the callers to determine if a
753 * iclog flush is necessary following this call.
754 */
755xfs_lsn_t
756xlog_cil_force_lsn(
757 struct xlog *log,
758 xfs_lsn_t sequence)
759{
760 struct xfs_cil *cil = log->l_cilp;
761 struct xfs_cil_ctx *ctx;
762 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
763
764 ASSERT(sequence <= cil->xc_current_sequence);
765
766 /*
767 * check to see if we need to force out the current context.
768 * xlog_cil_push() handles racing pushes for the same sequence,
769 * so no need to deal with it here.
770 */
771 xlog_cil_push_foreground(log, sequence);
772
773 /*
774 * See if we can find a previous sequence still committing.
775 * We need to wait for all previous sequence commits to complete
776 * before allowing the force of push_seq to go ahead. Hence block
777 * on commits for those as well.
778 */
779restart:
780 spin_lock(&cil->xc_cil_lock);
781 list_for_each_entry(ctx, &cil->xc_committing, committing) {
782 if (ctx->sequence > sequence)
783 continue;
784 if (!ctx->commit_lsn) {
785 /*
786 * It is still being pushed! Wait for the push to
787 * complete, then start again from the beginning.
788 */
789 xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
790 goto restart;
791 }
792 if (ctx->sequence != sequence)
793 continue;
794 /* found it! */
795 commit_lsn = ctx->commit_lsn;
796 }
797 spin_unlock(&cil->xc_cil_lock);
798 return commit_lsn;
799}
800
801/*
802 * Check if the current log item was first committed in this sequence.
803 * We can't rely on just the log item being in the CIL, we have to check
804 * the recorded commit sequence number.
805 *
806 * Note: for this to be used in a non-racy manner, it has to be called with
807 * CIL flushing locked out. As a result, it should only be used during the
808 * transaction commit process when deciding what to format into the item.
809 */
810bool
811xfs_log_item_in_current_chkpt(
812 struct xfs_log_item *lip)
813{
814 struct xfs_cil_ctx *ctx;
815
816 if (list_empty(&lip->li_cil))
817 return false;
818
819 ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
820
821 /*
822 * li_seq is written on the first commit of a log item to record the
823 * first checkpoint it is written to. Hence if it is different to the
824 * current sequence, we're in a new checkpoint.
825 */
826 if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
827 return false;
828 return true;
829}
830
831/*
832 * Perform initial CIL structure initialisation.
833 */
834int
835xlog_cil_init(
836 struct xlog *log)
837{
838 struct xfs_cil *cil;
839 struct xfs_cil_ctx *ctx;
840
841 cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
842 if (!cil)
843 return ENOMEM;
844
845 ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
846 if (!ctx) {
847 kmem_free(cil);
848 return ENOMEM;
849 }
850
851 INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
852 INIT_LIST_HEAD(&cil->xc_cil);
853 INIT_LIST_HEAD(&cil->xc_committing);
854 spin_lock_init(&cil->xc_cil_lock);
855 init_rwsem(&cil->xc_ctx_lock);
856 init_waitqueue_head(&cil->xc_commit_wait);
857
858 INIT_LIST_HEAD(&ctx->committing);
859 INIT_LIST_HEAD(&ctx->busy_extents);
860 ctx->sequence = 1;
861 ctx->cil = cil;
862 cil->xc_ctx = ctx;
863 cil->xc_current_sequence = ctx->sequence;
864
865 cil->xc_log = log;
866 log->l_cilp = cil;
867 return 0;
868}
869
870void
871xlog_cil_destroy(
872 struct xlog *log)
873{
874 if (log->l_cilp->xc_ctx) {
875 if (log->l_cilp->xc_ctx->ticket)
876 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
877 kmem_free(log->l_cilp->xc_ctx);
878 }
879
880 ASSERT(list_empty(&log->l_cilp->xc_cil));
881 kmem_free(log->l_cilp);
882}
883