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_log_format.h"
21#include "xfs_shared.h"
22#include "xfs_trans_resv.h"
23#include "xfs_sb.h"
24#include "xfs_ag.h"
25#include "xfs_mount.h"
26#include "xfs_error.h"
27#include "xfs_alloc.h"
28#include "xfs_extent_busy.h"
29#include "xfs_discard.h"
30#include "xfs_trans.h"
31#include "xfs_trans_priv.h"
32#include "xfs_log.h"
33#include "xfs_log_priv.h"
34
35/*
36 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
37 * recover, so we don't allow failure here. Also, we allocate in a context that
38 * we don't want to be issuing transactions from, so we need to tell the
39 * allocation code this as well.
40 *
41 * We don't reserve any space for the ticket - we are going to steal whatever
42 * space we require from transactions as they commit. To ensure we reserve all
43 * the space required, we need to set the current reservation of the ticket to
44 * zero so that we know to steal the initial transaction overhead from the
45 * first transaction commit.
46 */
47static struct xlog_ticket *
48xlog_cil_ticket_alloc(
49 struct xlog *log)
50{
51 struct xlog_ticket *tic;
52
53 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
54 KM_SLEEP|KM_NOFS);
55 tic->t_trans_type = XFS_TRANS_CHECKPOINT;
56
57 /*
58 * set the current reservation to zero so we know to steal the basic
59 * transaction overhead reservation from the first transaction commit.
60 */
61 tic->t_curr_res = 0;
62 return tic;
63}
64
65/*
66 * After the first stage of log recovery is done, we know where the head and
67 * tail of the log are. We need this log initialisation done before we can
68 * initialise the first CIL checkpoint context.
69 *
70 * Here we allocate a log ticket to track space usage during a CIL push. This
71 * ticket is passed to xlog_write() directly so that we don't slowly leak log
72 * space by failing to account for space used by log headers and additional
73 * region headers for split regions.
74 */
75void
76xlog_cil_init_post_recovery(
77 struct xlog *log)
78{
79 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
80 log->l_cilp->xc_ctx->sequence = 1;
81 log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
82 log->l_curr_block);
83}
84
85/*
86 * Prepare the log item for insertion into the CIL. Calculate the difference in
87 * log space and vectors it will consume, and if it is a new item pin it as
88 * well.
89 */
90STATIC void
91xfs_cil_prepare_item(
92 struct xlog *log,
93 struct xfs_log_vec *lv,
94 struct xfs_log_vec *old_lv,
95 int *diff_len,
96 int *diff_iovecs)
97{
98 /* Account for the new LV being passed in */
99 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
100 *diff_len += lv->lv_buf_len;
101 *diff_iovecs += lv->lv_niovecs;
102 }
103
104 /*
105 * If there is no old LV, this is the first time we've seen the item in
106 * this CIL context and so we need to pin it. If we are replacing the
107 * old_lv, then remove the space it accounts for and free it.
108 */
109 if (!old_lv)
110 lv->lv_item->li_ops->iop_pin(lv->lv_item);
111 else if (old_lv != lv) {
112 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
113
114 *diff_len -= old_lv->lv_buf_len;
115 *diff_iovecs -= old_lv->lv_niovecs;
116 kmem_free(old_lv);
117 }
118
119 /* attach new log vector to log item */
120 lv->lv_item->li_lv = lv;
121
122 /*
123 * If this is the first time the item is being committed to the
124 * CIL, store the sequence number on the log item so we can
125 * tell in future commits whether this is the first checkpoint
126 * the item is being committed into.
127 */
128 if (!lv->lv_item->li_seq)
129 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
130}
131
132/*
133 * Format log item into a flat buffers
134 *
135 * For delayed logging, we need to hold a formatted buffer containing all the
136 * changes on the log item. This enables us to relog the item in memory and
137 * write it out asynchronously without needing to relock the object that was
138 * modified at the time it gets written into the iclog.
139 *
140 * This function builds a vector for the changes in each log item in the
141 * transaction. It then works out the length of the buffer needed for each log
142 * item, allocates them and formats the vector for the item into the buffer.
143 * The buffer is then attached to the log item are then inserted into the
144 * Committed Item List for tracking until the next checkpoint is written out.
145 *
146 * We don't set up region headers during this process; we simply copy the
147 * regions into the flat buffer. We can do this because we still have to do a
148 * formatting step to write the regions into the iclog buffer. Writing the
149 * ophdrs during the iclog write means that we can support splitting large
150 * regions across iclog boundares without needing a change in the format of the
151 * item/region encapsulation.
152 *
153 * Hence what we need to do now is change the rewrite the vector array to point
154 * to the copied region inside the buffer we just allocated. This allows us to
155 * format the regions into the iclog as though they are being formatted
156 * directly out of the objects themselves.
157 */
158static void
159xlog_cil_insert_format_items(
160 struct xlog *log,
161 struct xfs_trans *tp,
162 int *diff_len,
163 int *diff_iovecs)
164{
165 struct xfs_log_item_desc *lidp;
166
167
168 /* Bail out if we didn't find a log item. */
169 if (list_empty(&tp->t_items)) {
170 ASSERT(0);
171 return;
172 }
173
174 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
175 struct xfs_log_item *lip = lidp->lid_item;
176 struct xfs_log_vec *lv;
177 struct xfs_log_vec *old_lv;
178 int niovecs = 0;
179 int nbytes = 0;
180 int buf_size;
181 bool ordered = false;
182
183 /* Skip items which aren't dirty in this transaction. */
184 if (!(lidp->lid_flags & XFS_LID_DIRTY))
185 continue;
186
187 /* get number of vecs and size of data to be stored */
188 lip->li_ops->iop_size(lip, &niovecs, &nbytes);
189
190 /* Skip items that do not have any vectors for writing */
191 if (!niovecs)
192 continue;
193
194 /*
195 * Ordered items need to be tracked but we do not wish to write
196 * them. We need a logvec to track the object, but we do not
197 * need an iovec or buffer to be allocated for copying data.
198 */
199 if (niovecs == XFS_LOG_VEC_ORDERED) {
200 ordered = true;
201 niovecs = 0;
202 nbytes = 0;
203 }
204
205 /*
206 * We 64-bit align the length of each iovec so that the start
207 * of the next one is naturally aligned. We'll need to
208 * account for that slack space here. Then round nbytes up
209 * to 64-bit alignment so that the initial buffer alignment is
210 * easy to calculate and verify.
211 */
212 nbytes += niovecs * sizeof(uint64_t);
213 nbytes = round_up(nbytes, sizeof(uint64_t));
214
215 /* grab the old item if it exists for reservation accounting */
216 old_lv = lip->li_lv;
217
218 /*
219 * The data buffer needs to start 64-bit aligned, so round up
220 * that space to ensure we can align it appropriately and not
221 * overrun the buffer.
222 */
223 buf_size = nbytes +
224 round_up((sizeof(struct xfs_log_vec) +
225 niovecs * sizeof(struct xfs_log_iovec)),
226 sizeof(uint64_t));
227
228 /* compare to existing item size */
229 if (lip->li_lv && buf_size <= lip->li_lv->lv_size) {
230 /* same or smaller, optimise common overwrite case */
231 lv = lip->li_lv;
232 lv->lv_next = NULL;
233
234 if (ordered)
235 goto insert;
236
237 /*
238 * set the item up as though it is a new insertion so
239 * that the space reservation accounting is correct.
240 */
241 *diff_iovecs -= lv->lv_niovecs;
242 *diff_len -= lv->lv_buf_len;
243 } else {
244 /* allocate new data chunk */
245 lv = kmem_zalloc(buf_size, KM_SLEEP|KM_NOFS);
246 lv->lv_item = lip;
247 lv->lv_size = buf_size;
248 if (ordered) {
249 /* track as an ordered logvec */
250 ASSERT(lip->li_lv == NULL);
251 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
252 goto insert;
253 }
254 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
255 }
256
257 /* Ensure the lv is set up according to ->iop_size */
258 lv->lv_niovecs = niovecs;
259
260 /* The allocated data region lies beyond the iovec region */
261 lv->lv_buf_len = 0;
262 lv->lv_buf = (char *)lv + buf_size - nbytes;
263 ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
264
265 lip->li_ops->iop_format(lip, lv);
266insert:
267 ASSERT(lv->lv_buf_len <= nbytes);
268 xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
269 }
270}
271
272/*
273 * Insert the log items into the CIL and calculate the difference in space
274 * consumed by the item. Add the space to the checkpoint ticket and calculate
275 * if the change requires additional log metadata. If it does, take that space
276 * as well. Remove the amount of space we added to the checkpoint ticket from
277 * the current transaction ticket so that the accounting works out correctly.
278 */
279static void
280xlog_cil_insert_items(
281 struct xlog *log,
282 struct xfs_trans *tp)
283{
284 struct xfs_cil *cil = log->l_cilp;
285 struct xfs_cil_ctx *ctx = cil->xc_ctx;
286 struct xfs_log_item_desc *lidp;
287 int len = 0;
288 int diff_iovecs = 0;
289 int iclog_space;
290
291 ASSERT(tp);
292
293 /*
294 * We can do this safely because the context can't checkpoint until we
295 * are done so it doesn't matter exactly how we update the CIL.
296 */
297 xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
298
299 /*
300 * Now (re-)position everything modified at the tail of the CIL.
301 * We do this here so we only need to take the CIL lock once during
302 * the transaction commit.
303 */
304 spin_lock(&cil->xc_cil_lock);
305 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
306 struct xfs_log_item *lip = lidp->lid_item;
307
308 /* Skip items which aren't dirty in this transaction. */
309 if (!(lidp->lid_flags & XFS_LID_DIRTY))
310 continue;
311
312 list_move_tail(&lip->li_cil, &cil->xc_cil);
313 }
314
315 /* account for space used by new iovec headers */
316 len += diff_iovecs * sizeof(xlog_op_header_t);
317 ctx->nvecs += diff_iovecs;
318
319 /* attach the transaction to the CIL if it has any busy extents */
320 if (!list_empty(&tp->t_busy))
321 list_splice_init(&tp->t_busy, &ctx->busy_extents);
322
323 /*
324 * Now transfer enough transaction reservation to the context ticket
325 * for the checkpoint. The context ticket is special - the unit
326 * reservation has to grow as well as the current reservation as we
327 * steal from tickets so we can correctly determine the space used
328 * during the transaction commit.
329 */
330 if (ctx->ticket->t_curr_res == 0) {
331 ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
332 tp->t_ticket->t_curr_res -= ctx->ticket->t_unit_res;
333 }
334
335 /* do we need space for more log record headers? */
336 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
337 if (len > 0 && (ctx->space_used / iclog_space !=
338 (ctx->space_used + len) / iclog_space)) {
339 int hdrs;
340
341 hdrs = (len + iclog_space - 1) / iclog_space;
342 /* need to take into account split region headers, too */
343 hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
344 ctx->ticket->t_unit_res += hdrs;
345 ctx->ticket->t_curr_res += hdrs;
346 tp->t_ticket->t_curr_res -= hdrs;
347 ASSERT(tp->t_ticket->t_curr_res >= len);
348 }
349 tp->t_ticket->t_curr_res -= len;
350 ctx->space_used += len;
351
352 spin_unlock(&cil->xc_cil_lock);
353}
354
355static void
356xlog_cil_free_logvec(
357 struct xfs_log_vec *log_vector)
358{
359 struct xfs_log_vec *lv;
360
361 for (lv = log_vector; lv; ) {
362 struct xfs_log_vec *next = lv->lv_next;
363 kmem_free(lv);
364 lv = next;
365 }
366}
367
368/*
369 * Mark all items committed and clear busy extents. We free the log vector
370 * chains in a separate pass so that we unpin the log items as quickly as
371 * possible.
372 */
373static void
374xlog_cil_committed(
375 void *args,
376 int abort)
377{
378 struct xfs_cil_ctx *ctx = args;
379 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
380
381 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
382 ctx->start_lsn, abort);
383
384 xfs_extent_busy_sort(&ctx->busy_extents);
385 xfs_extent_busy_clear(mp, &ctx->busy_extents,
386 (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
387
388 spin_lock(&ctx->cil->xc_push_lock);
389 list_del(&ctx->committing);
390 spin_unlock(&ctx->cil->xc_push_lock);
391
392 xlog_cil_free_logvec(ctx->lv_chain);
393
394 if (!list_empty(&ctx->busy_extents)) {
395 ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
396
397 xfs_discard_extents(mp, &ctx->busy_extents);
398 xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
399 }
400
401 kmem_free(ctx);
402}
403
404/*
405 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
406 * is a background flush and so we can chose to ignore it. Otherwise, if the
407 * current sequence is the same as @push_seq we need to do a flush. If
408 * @push_seq is less than the current sequence, then it has already been
409 * flushed and we don't need to do anything - the caller will wait for it to
410 * complete if necessary.
411 *
412 * @push_seq is a value rather than a flag because that allows us to do an
413 * unlocked check of the sequence number for a match. Hence we can allows log
414 * forces to run racily and not issue pushes for the same sequence twice. If we
415 * get a race between multiple pushes for the same sequence they will block on
416 * the first one and then abort, hence avoiding needless pushes.
417 */
418STATIC int
419xlog_cil_push(
420 struct xlog *log)
421{
422 struct xfs_cil *cil = log->l_cilp;
423 struct xfs_log_vec *lv;
424 struct xfs_cil_ctx *ctx;
425 struct xfs_cil_ctx *new_ctx;
426 struct xlog_in_core *commit_iclog;
427 struct xlog_ticket *tic;
428 int num_iovecs;
429 int error = 0;
430 struct xfs_trans_header thdr;
431 struct xfs_log_iovec lhdr;
432 struct xfs_log_vec lvhdr = { NULL };
433 xfs_lsn_t commit_lsn;
434 xfs_lsn_t push_seq;
435
436 if (!cil)
437 return 0;
438
439 new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
440 new_ctx->ticket = xlog_cil_ticket_alloc(log);
441
442 down_write(&cil->xc_ctx_lock);
443 ctx = cil->xc_ctx;
444
445 spin_lock(&cil->xc_push_lock);
446 push_seq = cil->xc_push_seq;
447 ASSERT(push_seq <= ctx->sequence);
448
449 /*
450 * Check if we've anything to push. If there is nothing, then we don't
451 * move on to a new sequence number and so we have to be able to push
452 * this sequence again later.
453 */
454 if (list_empty(&cil->xc_cil)) {
455 cil->xc_push_seq = 0;
456 spin_unlock(&cil->xc_push_lock);
457 goto out_skip;
458 }
459 spin_unlock(&cil->xc_push_lock);
460
461
462 /* check for a previously pushed seqeunce */
463 if (push_seq < cil->xc_ctx->sequence)
464 goto out_skip;
465
466 /*
467 * pull all the log vectors off the items in the CIL, and
468 * remove the items from the CIL. We don't need the CIL lock
469 * here because it's only needed on the transaction commit
470 * side which is currently locked out by the flush lock.
471 */
472 lv = NULL;
473 num_iovecs = 0;
474 while (!list_empty(&cil->xc_cil)) {
475 struct xfs_log_item *item;
476
477 item = list_first_entry(&cil->xc_cil,
478 struct xfs_log_item, li_cil);
479 list_del_init(&item->li_cil);
480 if (!ctx->lv_chain)
481 ctx->lv_chain = item->li_lv;
482 else
483 lv->lv_next = item->li_lv;
484 lv = item->li_lv;
485 item->li_lv = NULL;
486 num_iovecs += lv->lv_niovecs;
487 }
488
489 /*
490 * initialise the new context and attach it to the CIL. Then attach
491 * the current context to the CIL committing lsit so it can be found
492 * during log forces to extract the commit lsn of the sequence that
493 * needs to be forced.
494 */
495 INIT_LIST_HEAD(&new_ctx->committing);
496 INIT_LIST_HEAD(&new_ctx->busy_extents);
497 new_ctx->sequence = ctx->sequence + 1;
498 new_ctx->cil = cil;
499 cil->xc_ctx = new_ctx;
500
501 /*
502 * The switch is now done, so we can drop the context lock and move out
503 * of a shared context. We can't just go straight to the commit record,
504 * though - we need to synchronise with previous and future commits so
505 * that the commit records are correctly ordered in the log to ensure
506 * that we process items during log IO completion in the correct order.
507 *
508 * For example, if we get an EFI in one checkpoint and the EFD in the
509 * next (e.g. due to log forces), we do not want the checkpoint with
510 * the EFD to be committed before the checkpoint with the EFI. Hence
511 * we must strictly order the commit records of the checkpoints so
512 * that: a) the checkpoint callbacks are attached to the iclogs in the
513 * correct order; and b) the checkpoints are replayed in correct order
514 * in log recovery.
515 *
516 * Hence we need to add this context to the committing context list so
517 * that higher sequences will wait for us to write out a commit record
518 * before they do.
519 *
520 * xfs_log_force_lsn requires us to mirror the new sequence into the cil
521 * structure atomically with the addition of this sequence to the
522 * committing list. This also ensures that we can do unlocked checks
523 * against the current sequence in log forces without risking
524 * deferencing a freed context pointer.
525 */
526 spin_lock(&cil->xc_push_lock);
527 cil->xc_current_sequence = new_ctx->sequence;
528 list_add(&ctx->committing, &cil->xc_committing);
529 spin_unlock(&cil->xc_push_lock);
530 up_write(&cil->xc_ctx_lock);
531
532 /*
533 * Build a checkpoint transaction header and write it to the log to
534 * begin the transaction. We need to account for the space used by the
535 * transaction header here as it is not accounted for in xlog_write().
536 *
537 * The LSN we need to pass to the log items on transaction commit is
538 * the LSN reported by the first log vector write. If we use the commit
539 * record lsn then we can move the tail beyond the grant write head.
540 */
541 tic = ctx->ticket;
542 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
543 thdr.th_type = XFS_TRANS_CHECKPOINT;
544 thdr.th_tid = tic->t_tid;
545 thdr.th_num_items = num_iovecs;
546 lhdr.i_addr = &thdr;
547 lhdr.i_len = sizeof(xfs_trans_header_t);
548 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
549 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
550
551 lvhdr.lv_niovecs = 1;
552 lvhdr.lv_iovecp = &lhdr;
553 lvhdr.lv_next = ctx->lv_chain;
554
555 error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
556 if (error)
557 goto out_abort_free_ticket;
558
559 /*
560 * now that we've written the checkpoint into the log, strictly
561 * order the commit records so replay will get them in the right order.
562 */
563restart:
564 spin_lock(&cil->xc_push_lock);
565 list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
566 /*
567 * Higher sequences will wait for this one so skip them.
568 * Don't wait for own own sequence, either.
569 */
570 if (new_ctx->sequence >= ctx->sequence)
571 continue;
572 if (!new_ctx->commit_lsn) {
573 /*
574 * It is still being pushed! Wait for the push to
575 * complete, then start again from the beginning.
576 */
577 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
578 goto restart;
579 }
580 }
581 spin_unlock(&cil->xc_push_lock);
582
583 /* xfs_log_done always frees the ticket on error. */
584 commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
585 if (commit_lsn == -1)
586 goto out_abort;
587
588 /* attach all the transactions w/ busy extents to iclog */
589 ctx->log_cb.cb_func = xlog_cil_committed;
590 ctx->log_cb.cb_arg = ctx;
591 error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
592 if (error)
593 goto out_abort;
594
595 /*
596 * now the checkpoint commit is complete and we've attached the
597 * callbacks to the iclog we can assign the commit LSN to the context
598 * and wake up anyone who is waiting for the commit to complete.
599 */
600 spin_lock(&cil->xc_push_lock);
601 ctx->commit_lsn = commit_lsn;
602 wake_up_all(&cil->xc_commit_wait);
603 spin_unlock(&cil->xc_push_lock);
604
605 /* release the hounds! */
606 return xfs_log_release_iclog(log->l_mp, commit_iclog);
607
608out_skip:
609 up_write(&cil->xc_ctx_lock);
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
621static void
622xlog_cil_push_work(
623 struct work_struct *work)
624{
625 struct xfs_cil *cil = container_of(work, struct xfs_cil,
626 xc_push_work);
627 xlog_cil_push(cil->xc_log);
628}
629
630/*
631 * We need to push CIL every so often so we don't cache more than we can fit in
632 * the log. The limit really is that a checkpoint can't be more than half the
633 * log (the current checkpoint is not allowed to overwrite the previous
634 * checkpoint), but commit latency and memory usage limit this to a smaller
635 * size.
636 */
637static void
638xlog_cil_push_background(
639 struct xlog *log)
640{
641 struct xfs_cil *cil = log->l_cilp;
642
643 /*
644 * The cil won't be empty because we are called while holding the
645 * context lock so whatever we added to the CIL will still be there
646 */
647 ASSERT(!list_empty(&cil->xc_cil));
648
649 /*
650 * don't do a background push if we haven't used up all the
651 * space available yet.
652 */
653 if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
654 return;
655
656 spin_lock(&cil->xc_push_lock);
657 if (cil->xc_push_seq < cil->xc_current_sequence) {
658 cil->xc_push_seq = cil->xc_current_sequence;
659 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
660 }
661 spin_unlock(&cil->xc_push_lock);
662
663}
664
665/*
666 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
667 * number that is passed. When it returns, the work will be queued for
668 * @push_seq, but it won't be completed. The caller is expected to do any
669 * waiting for push_seq to complete if it is required.
670 */
671static void
672xlog_cil_push_now(
673 struct xlog *log,
674 xfs_lsn_t push_seq)
675{
676 struct xfs_cil *cil = log->l_cilp;
677
678 if (!cil)
679 return;
680
681 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
682
683 /* start on any pending background push to minimise wait time on it */
684 flush_work(&cil->xc_push_work);
685
686 /*
687 * If the CIL is empty or we've already pushed the sequence then
688 * there's no work we need to do.
689 */
690 spin_lock(&cil->xc_push_lock);
691 if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
692 spin_unlock(&cil->xc_push_lock);
693 return;
694 }
695
696 cil->xc_push_seq = push_seq;
697 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
698 spin_unlock(&cil->xc_push_lock);
699}
700
701bool
702xlog_cil_empty(
703 struct xlog *log)
704{
705 struct xfs_cil *cil = log->l_cilp;
706 bool empty = false;
707
708 spin_lock(&cil->xc_push_lock);
709 if (list_empty(&cil->xc_cil))
710 empty = true;
711 spin_unlock(&cil->xc_push_lock);
712 return empty;
713}
714
715/*
716 * Commit a transaction with the given vector to the Committed Item List.
717 *
718 * To do this, we need to format the item, pin it in memory if required and
719 * account for the space used by the transaction. Once we have done that we
720 * need to release the unused reservation for the transaction, attach the
721 * transaction to the checkpoint context so we carry the busy extents through
722 * to checkpoint completion, and then unlock all the items in the transaction.
723 *
724 * Called with the context lock already held in read mode to lock out
725 * background commit, returns without it held once background commits are
726 * allowed again.
727 */
728void
729xfs_log_commit_cil(
730 struct xfs_mount *mp,
731 struct xfs_trans *tp,
732 xfs_lsn_t *commit_lsn,
733 int flags)
734{
735 struct xlog *log = mp->m_log;
736 struct xfs_cil *cil = log->l_cilp;
737 int log_flags = 0;
738
739 if (flags & XFS_TRANS_RELEASE_LOG_RES)
740 log_flags = XFS_LOG_REL_PERM_RESERV;
741
742 /* lock out background commit */
743 down_read(&cil->xc_ctx_lock);
744
745 xlog_cil_insert_items(log, tp);
746
747 /* check we didn't blow the reservation */
748 if (tp->t_ticket->t_curr_res < 0)
749 xlog_print_tic_res(mp, tp->t_ticket);
750
751 tp->t_commit_lsn = cil->xc_ctx->sequence;
752 if (commit_lsn)
753 *commit_lsn = tp->t_commit_lsn;
754
755 xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
756 xfs_trans_unreserve_and_mod_sb(tp);
757
758 /*
759 * Once all the items of the transaction have been copied to the CIL,
760 * the items can be unlocked and freed.
761 *
762 * This needs to be done before we drop the CIL context lock because we
763 * have to update state in the log items and unlock them before they go
764 * to disk. If we don't, then the CIL checkpoint can race with us and
765 * we can run checkpoint completion before we've updated and unlocked
766 * the log items. This affects (at least) processing of stale buffers,
767 * inodes and EFIs.
768 */
769 xfs_trans_free_items(tp, tp->t_commit_lsn, 0);
770
771 xlog_cil_push_background(log);
772
773 up_read(&cil->xc_ctx_lock);
774}
775
776/*
777 * Conditionally push the CIL based on the sequence passed in.
778 *
779 * We only need to push if we haven't already pushed the sequence
780 * number given. Hence the only time we will trigger a push here is
781 * if the push sequence is the same as the current context.
782 *
783 * We return the current commit lsn to allow the callers to determine if a
784 * iclog flush is necessary following this call.
785 */
786xfs_lsn_t
787xlog_cil_force_lsn(
788 struct xlog *log,
789 xfs_lsn_t sequence)
790{
791 struct xfs_cil *cil = log->l_cilp;
792 struct xfs_cil_ctx *ctx;
793 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
794
795 ASSERT(sequence <= cil->xc_current_sequence);
796
797 /*
798 * check to see if we need to force out the current context.
799 * xlog_cil_push() handles racing pushes for the same sequence,
800 * so no need to deal with it here.
801 */
802restart:
803 xlog_cil_push_now(log, sequence);
804
805 /*
806 * See if we can find a previous sequence still committing.
807 * We need to wait for all previous sequence commits to complete
808 * before allowing the force of push_seq to go ahead. Hence block
809 * on commits for those as well.
810 */
811 spin_lock(&cil->xc_push_lock);
812 list_for_each_entry(ctx, &cil->xc_committing, committing) {
813 if (ctx->sequence > sequence)
814 continue;
815 if (!ctx->commit_lsn) {
816 /*
817 * It is still being pushed! Wait for the push to
818 * complete, then start again from the beginning.
819 */
820 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
821 goto restart;
822 }
823 if (ctx->sequence != sequence)
824 continue;
825 /* found it! */
826 commit_lsn = ctx->commit_lsn;
827 }
828
829 /*
830 * The call to xlog_cil_push_now() executes the push in the background.
831 * Hence by the time we have got here it our sequence may not have been
832 * pushed yet. This is true if the current sequence still matches the
833 * push sequence after the above wait loop and the CIL still contains
834 * dirty objects.
835 *
836 * When the push occurs, it will empty the CIL and
837 * atomically increment the currect sequence past the push sequence and
838 * move it into the committing list. Of course, if the CIL is clean at
839 * the time of the push, it won't have pushed the CIL at all, so in that
840 * case we should try the push for this sequence again from the start
841 * just in case.
842 */
843
844 if (sequence == cil->xc_current_sequence &&
845 !list_empty(&cil->xc_cil)) {
846 spin_unlock(&cil->xc_push_lock);
847 goto restart;
848 }
849
850 spin_unlock(&cil->xc_push_lock);
851 return commit_lsn;
852}
853
854/*
855 * Check if the current log item was first committed in this sequence.
856 * We can't rely on just the log item being in the CIL, we have to check
857 * the recorded commit sequence number.
858 *
859 * Note: for this to be used in a non-racy manner, it has to be called with
860 * CIL flushing locked out. As a result, it should only be used during the
861 * transaction commit process when deciding what to format into the item.
862 */
863bool
864xfs_log_item_in_current_chkpt(
865 struct xfs_log_item *lip)
866{
867 struct xfs_cil_ctx *ctx;
868
869 if (list_empty(&lip->li_cil))
870 return false;
871
872 ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
873
874 /*
875 * li_seq is written on the first commit of a log item to record the
876 * first checkpoint it is written to. Hence if it is different to the
877 * current sequence, we're in a new checkpoint.
878 */
879 if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
880 return false;
881 return true;
882}
883
884/*
885 * Perform initial CIL structure initialisation.
886 */
887int
888xlog_cil_init(
889 struct xlog *log)
890{
891 struct xfs_cil *cil;
892 struct xfs_cil_ctx *ctx;
893
894 cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
895 if (!cil)
896 return ENOMEM;
897
898 ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
899 if (!ctx) {
900 kmem_free(cil);
901 return ENOMEM;
902 }
903
904 INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
905 INIT_LIST_HEAD(&cil->xc_cil);
906 INIT_LIST_HEAD(&cil->xc_committing);
907 spin_lock_init(&cil->xc_cil_lock);
908 spin_lock_init(&cil->xc_push_lock);
909 init_rwsem(&cil->xc_ctx_lock);
910 init_waitqueue_head(&cil->xc_commit_wait);
911
912 INIT_LIST_HEAD(&ctx->committing);
913 INIT_LIST_HEAD(&ctx->busy_extents);
914 ctx->sequence = 1;
915 ctx->cil = cil;
916 cil->xc_ctx = ctx;
917 cil->xc_current_sequence = ctx->sequence;
918
919 cil->xc_log = log;
920 log->l_cilp = cil;
921 return 0;
922}
923
924void
925xlog_cil_destroy(
926 struct xlog *log)
927{
928 if (log->l_cilp->xc_ctx) {
929 if (log->l_cilp->xc_ctx->ticket)
930 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
931 kmem_free(log->l_cilp->xc_ctx);
932 }
933
934 ASSERT(list_empty(&log->l_cilp->xc_cil));
935 kmem_free(log->l_cilp);
936}
937