Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / fs / xfs / xfs_log_cil.c
blob43aa42a3a5d319fffff98417bff98055083f1a75
1 /*
2 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
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.
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.
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
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_shared.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_error.h"
26 #include "xfs_alloc.h"
27 #include "xfs_extent_busy.h"
28 #include "xfs_discard.h"
29 #include "xfs_trans.h"
30 #include "xfs_trans_priv.h"
31 #include "xfs_log.h"
32 #include "xfs_log_priv.h"
33 #include "xfs_trace.h"
35 struct workqueue_struct *xfs_discard_wq;
38 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
39 * recover, so we don't allow failure here. Also, we allocate in a context that
40 * we don't want to be issuing transactions from, so we need to tell the
41 * allocation code this as well.
43 * We don't reserve any space for the ticket - we are going to steal whatever
44 * space we require from transactions as they commit. To ensure we reserve all
45 * the space required, we need to set the current reservation of the ticket to
46 * zero so that we know to steal the initial transaction overhead from the
47 * first transaction commit.
49 static struct xlog_ticket *
50 xlog_cil_ticket_alloc(
51 struct xlog *log)
53 struct xlog_ticket *tic;
55 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
56 KM_SLEEP|KM_NOFS);
59 * set the current reservation to zero so we know to steal the basic
60 * transaction overhead reservation from the first transaction commit.
62 tic->t_curr_res = 0;
63 return tic;
67 * After the first stage of log recovery is done, we know where the head and
68 * tail of the log are. We need this log initialisation done before we can
69 * initialise the first CIL checkpoint context.
71 * Here we allocate a log ticket to track space usage during a CIL push. This
72 * ticket is passed to xlog_write() directly so that we don't slowly leak log
73 * space by failing to account for space used by log headers and additional
74 * region headers for split regions.
76 void
77 xlog_cil_init_post_recovery(
78 struct xlog *log)
80 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
81 log->l_cilp->xc_ctx->sequence = 1;
84 static inline int
85 xlog_cil_iovec_space(
86 uint niovecs)
88 return round_up((sizeof(struct xfs_log_vec) +
89 niovecs * sizeof(struct xfs_log_iovec)),
90 sizeof(uint64_t));
94 * Allocate or pin log vector buffers for CIL insertion.
96 * The CIL currently uses disposable buffers for copying a snapshot of the
97 * modified items into the log during a push. The biggest problem with this is
98 * the requirement to allocate the disposable buffer during the commit if:
99 * a) does not exist; or
100 * b) it is too small
102 * If we do this allocation within xlog_cil_insert_format_items(), it is done
103 * under the xc_ctx_lock, which means that a CIL push cannot occur during
104 * the memory allocation. This means that we have a potential deadlock situation
105 * under low memory conditions when we have lots of dirty metadata pinned in
106 * the CIL and we need a CIL commit to occur to free memory.
108 * To avoid this, we need to move the memory allocation outside the
109 * xc_ctx_lock, but because the log vector buffers are disposable, that opens
110 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log
111 * vector buffers between the check and the formatting of the item into the
112 * log vector buffer within the xc_ctx_lock.
114 * Because the log vector buffer needs to be unchanged during the CIL push
115 * process, we cannot share the buffer between the transaction commit (which
116 * modifies the buffer) and the CIL push context that is writing the changes
117 * into the log. This means skipping preallocation of buffer space is
118 * unreliable, but we most definitely do not want to be allocating and freeing
119 * buffers unnecessarily during commits when overwrites can be done safely.
121 * The simplest solution to this problem is to allocate a shadow buffer when a
122 * log item is committed for the second time, and then to only use this buffer
123 * if necessary. The buffer can remain attached to the log item until such time
124 * it is needed, and this is the buffer that is reallocated to match the size of
125 * the incoming modification. Then during the formatting of the item we can swap
126 * the active buffer with the new one if we can't reuse the existing buffer. We
127 * don't free the old buffer as it may be reused on the next modification if
128 * it's size is right, otherwise we'll free and reallocate it at that point.
130 * This function builds a vector for the changes in each log item in the
131 * transaction. It then works out the length of the buffer needed for each log
132 * item, allocates them and attaches the vector to the log item in preparation
133 * for the formatting step which occurs under the xc_ctx_lock.
135 * While this means the memory footprint goes up, it avoids the repeated
136 * alloc/free pattern that repeated modifications of an item would otherwise
137 * cause, and hence minimises the CPU overhead of such behaviour.
139 static void
140 xlog_cil_alloc_shadow_bufs(
141 struct xlog *log,
142 struct xfs_trans *tp)
144 struct xfs_log_item_desc *lidp;
146 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
147 struct xfs_log_item *lip = lidp->lid_item;
148 struct xfs_log_vec *lv;
149 int niovecs = 0;
150 int nbytes = 0;
151 int buf_size;
152 bool ordered = false;
154 /* Skip items which aren't dirty in this transaction. */
155 if (!(lidp->lid_flags & XFS_LID_DIRTY))
156 continue;
158 /* get number of vecs and size of data to be stored */
159 lip->li_ops->iop_size(lip, &niovecs, &nbytes);
162 * Ordered items need to be tracked but we do not wish to write
163 * them. We need a logvec to track the object, but we do not
164 * need an iovec or buffer to be allocated for copying data.
166 if (niovecs == XFS_LOG_VEC_ORDERED) {
167 ordered = true;
168 niovecs = 0;
169 nbytes = 0;
173 * We 64-bit align the length of each iovec so that the start
174 * of the next one is naturally aligned. We'll need to
175 * account for that slack space here. Then round nbytes up
176 * to 64-bit alignment so that the initial buffer alignment is
177 * easy to calculate and verify.
179 nbytes += niovecs * sizeof(uint64_t);
180 nbytes = round_up(nbytes, sizeof(uint64_t));
183 * The data buffer needs to start 64-bit aligned, so round up
184 * that space to ensure we can align it appropriately and not
185 * overrun the buffer.
187 buf_size = nbytes + xlog_cil_iovec_space(niovecs);
190 * if we have no shadow buffer, or it is too small, we need to
191 * reallocate it.
193 if (!lip->li_lv_shadow ||
194 buf_size > lip->li_lv_shadow->lv_size) {
197 * We free and allocate here as a realloc would copy
198 * unecessary data. We don't use kmem_zalloc() for the
199 * same reason - we don't need to zero the data area in
200 * the buffer, only the log vector header and the iovec
201 * storage.
203 kmem_free(lip->li_lv_shadow);
205 lv = kmem_alloc(buf_size, KM_SLEEP|KM_NOFS);
206 memset(lv, 0, xlog_cil_iovec_space(niovecs));
208 lv->lv_item = lip;
209 lv->lv_size = buf_size;
210 if (ordered)
211 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
212 else
213 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
214 lip->li_lv_shadow = lv;
215 } else {
216 /* same or smaller, optimise common overwrite case */
217 lv = lip->li_lv_shadow;
218 if (ordered)
219 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
220 else
221 lv->lv_buf_len = 0;
222 lv->lv_bytes = 0;
223 lv->lv_next = NULL;
226 /* Ensure the lv is set up according to ->iop_size */
227 lv->lv_niovecs = niovecs;
229 /* The allocated data region lies beyond the iovec region */
230 lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);
236 * Prepare the log item for insertion into the CIL. Calculate the difference in
237 * log space and vectors it will consume, and if it is a new item pin it as
238 * well.
240 STATIC void
241 xfs_cil_prepare_item(
242 struct xlog *log,
243 struct xfs_log_vec *lv,
244 struct xfs_log_vec *old_lv,
245 int *diff_len,
246 int *diff_iovecs)
248 /* Account for the new LV being passed in */
249 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
250 *diff_len += lv->lv_bytes;
251 *diff_iovecs += lv->lv_niovecs;
255 * If there is no old LV, this is the first time we've seen the item in
256 * this CIL context and so we need to pin it. If we are replacing the
257 * old_lv, then remove the space it accounts for and make it the shadow
258 * buffer for later freeing. In both cases we are now switching to the
259 * shadow buffer, so update the the pointer to it appropriately.
261 if (!old_lv) {
262 lv->lv_item->li_ops->iop_pin(lv->lv_item);
263 lv->lv_item->li_lv_shadow = NULL;
264 } else if (old_lv != lv) {
265 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
267 *diff_len -= old_lv->lv_bytes;
268 *diff_iovecs -= old_lv->lv_niovecs;
269 lv->lv_item->li_lv_shadow = old_lv;
272 /* attach new log vector to log item */
273 lv->lv_item->li_lv = lv;
276 * If this is the first time the item is being committed to the
277 * CIL, store the sequence number on the log item so we can
278 * tell in future commits whether this is the first checkpoint
279 * the item is being committed into.
281 if (!lv->lv_item->li_seq)
282 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
286 * Format log item into a flat buffers
288 * For delayed logging, we need to hold a formatted buffer containing all the
289 * changes on the log item. This enables us to relog the item in memory and
290 * write it out asynchronously without needing to relock the object that was
291 * modified at the time it gets written into the iclog.
293 * This function takes the prepared log vectors attached to each log item, and
294 * formats the changes into the log vector buffer. The buffer it uses is
295 * dependent on the current state of the vector in the CIL - the shadow lv is
296 * guaranteed to be large enough for the current modification, but we will only
297 * use that if we can't reuse the existing lv. If we can't reuse the existing
298 * lv, then simple swap it out for the shadow lv. We don't free it - that is
299 * done lazily either by th enext modification or the freeing of the log item.
301 * We don't set up region headers during this process; we simply copy the
302 * regions into the flat buffer. We can do this because we still have to do a
303 * formatting step to write the regions into the iclog buffer. Writing the
304 * ophdrs during the iclog write means that we can support splitting large
305 * regions across iclog boundares without needing a change in the format of the
306 * item/region encapsulation.
308 * Hence what we need to do now is change the rewrite the vector array to point
309 * to the copied region inside the buffer we just allocated. This allows us to
310 * format the regions into the iclog as though they are being formatted
311 * directly out of the objects themselves.
313 static void
314 xlog_cil_insert_format_items(
315 struct xlog *log,
316 struct xfs_trans *tp,
317 int *diff_len,
318 int *diff_iovecs)
320 struct xfs_log_item_desc *lidp;
323 /* Bail out if we didn't find a log item. */
324 if (list_empty(&tp->t_items)) {
325 ASSERT(0);
326 return;
329 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
330 struct xfs_log_item *lip = lidp->lid_item;
331 struct xfs_log_vec *lv;
332 struct xfs_log_vec *old_lv = NULL;
333 struct xfs_log_vec *shadow;
334 bool ordered = false;
336 /* Skip items which aren't dirty in this transaction. */
337 if (!(lidp->lid_flags & XFS_LID_DIRTY))
338 continue;
341 * The formatting size information is already attached to
342 * the shadow lv on the log item.
344 shadow = lip->li_lv_shadow;
345 if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)
346 ordered = true;
348 /* Skip items that do not have any vectors for writing */
349 if (!shadow->lv_niovecs && !ordered)
350 continue;
352 /* compare to existing item size */
353 old_lv = lip->li_lv;
354 if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {
355 /* same or smaller, optimise common overwrite case */
356 lv = lip->li_lv;
357 lv->lv_next = NULL;
359 if (ordered)
360 goto insert;
363 * set the item up as though it is a new insertion so
364 * that the space reservation accounting is correct.
366 *diff_iovecs -= lv->lv_niovecs;
367 *diff_len -= lv->lv_bytes;
369 /* Ensure the lv is set up according to ->iop_size */
370 lv->lv_niovecs = shadow->lv_niovecs;
372 /* reset the lv buffer information for new formatting */
373 lv->lv_buf_len = 0;
374 lv->lv_bytes = 0;
375 lv->lv_buf = (char *)lv +
376 xlog_cil_iovec_space(lv->lv_niovecs);
377 } else {
378 /* switch to shadow buffer! */
379 lv = shadow;
380 lv->lv_item = lip;
381 if (ordered) {
382 /* track as an ordered logvec */
383 ASSERT(lip->li_lv == NULL);
384 goto insert;
388 ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
389 lip->li_ops->iop_format(lip, lv);
390 insert:
391 xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
396 * Insert the log items into the CIL and calculate the difference in space
397 * consumed by the item. Add the space to the checkpoint ticket and calculate
398 * if the change requires additional log metadata. If it does, take that space
399 * as well. Remove the amount of space we added to the checkpoint ticket from
400 * the current transaction ticket so that the accounting works out correctly.
402 static void
403 xlog_cil_insert_items(
404 struct xlog *log,
405 struct xfs_trans *tp)
407 struct xfs_cil *cil = log->l_cilp;
408 struct xfs_cil_ctx *ctx = cil->xc_ctx;
409 struct xfs_log_item_desc *lidp;
410 int len = 0;
411 int diff_iovecs = 0;
412 int iclog_space;
413 int iovhdr_res = 0, split_res = 0, ctx_res = 0;
415 ASSERT(tp);
418 * We can do this safely because the context can't checkpoint until we
419 * are done so it doesn't matter exactly how we update the CIL.
421 xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
423 spin_lock(&cil->xc_cil_lock);
425 /* account for space used by new iovec headers */
426 iovhdr_res = diff_iovecs * sizeof(xlog_op_header_t);
427 len += iovhdr_res;
428 ctx->nvecs += diff_iovecs;
430 /* attach the transaction to the CIL if it has any busy extents */
431 if (!list_empty(&tp->t_busy))
432 list_splice_init(&tp->t_busy, &ctx->busy_extents);
435 * Now transfer enough transaction reservation to the context ticket
436 * for the checkpoint. The context ticket is special - the unit
437 * reservation has to grow as well as the current reservation as we
438 * steal from tickets so we can correctly determine the space used
439 * during the transaction commit.
441 if (ctx->ticket->t_curr_res == 0) {
442 ctx_res = ctx->ticket->t_unit_res;
443 ctx->ticket->t_curr_res = ctx_res;
444 tp->t_ticket->t_curr_res -= ctx_res;
447 /* do we need space for more log record headers? */
448 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
449 if (len > 0 && (ctx->space_used / iclog_space !=
450 (ctx->space_used + len) / iclog_space)) {
451 split_res = (len + iclog_space - 1) / iclog_space;
452 /* need to take into account split region headers, too */
453 split_res *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
454 ctx->ticket->t_unit_res += split_res;
455 ctx->ticket->t_curr_res += split_res;
456 tp->t_ticket->t_curr_res -= split_res;
457 ASSERT(tp->t_ticket->t_curr_res >= len);
459 tp->t_ticket->t_curr_res -= len;
460 ctx->space_used += len;
463 * If we've overrun the reservation, dump the tx details before we move
464 * the log items. Shutdown is imminent...
466 if (WARN_ON(tp->t_ticket->t_curr_res < 0)) {
467 xfs_warn(log->l_mp, "Transaction log reservation overrun:");
468 xfs_warn(log->l_mp,
469 " log items: %d bytes (iov hdrs: %d bytes)",
470 len, iovhdr_res);
471 xfs_warn(log->l_mp, " split region headers: %d bytes",
472 split_res);
473 xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res);
474 xlog_print_trans(tp);
478 * Now (re-)position everything modified at the tail of the CIL.
479 * We do this here so we only need to take the CIL lock once during
480 * the transaction commit.
482 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
483 struct xfs_log_item *lip = lidp->lid_item;
485 /* Skip items which aren't dirty in this transaction. */
486 if (!(lidp->lid_flags & XFS_LID_DIRTY))
487 continue;
490 * Only move the item if it isn't already at the tail. This is
491 * to prevent a transient list_empty() state when reinserting
492 * an item that is already the only item in the CIL.
494 if (!list_is_last(&lip->li_cil, &cil->xc_cil))
495 list_move_tail(&lip->li_cil, &cil->xc_cil);
498 spin_unlock(&cil->xc_cil_lock);
500 if (tp->t_ticket->t_curr_res < 0)
501 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
504 static void
505 xlog_cil_free_logvec(
506 struct xfs_log_vec *log_vector)
508 struct xfs_log_vec *lv;
510 for (lv = log_vector; lv; ) {
511 struct xfs_log_vec *next = lv->lv_next;
512 kmem_free(lv);
513 lv = next;
517 static void
518 xlog_discard_endio_work(
519 struct work_struct *work)
521 struct xfs_cil_ctx *ctx =
522 container_of(work, struct xfs_cil_ctx, discard_endio_work);
523 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
525 xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
526 kmem_free(ctx);
530 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
531 * pagb_lock. Note that we need a unbounded workqueue, otherwise we might
532 * get the execution delayed up to 30 seconds for weird reasons.
534 static void
535 xlog_discard_endio(
536 struct bio *bio)
538 struct xfs_cil_ctx *ctx = bio->bi_private;
540 INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work);
541 queue_work(xfs_discard_wq, &ctx->discard_endio_work);
542 bio_put(bio);
545 static void
546 xlog_discard_busy_extents(
547 struct xfs_mount *mp,
548 struct xfs_cil_ctx *ctx)
550 struct list_head *list = &ctx->busy_extents;
551 struct xfs_extent_busy *busyp;
552 struct bio *bio = NULL;
553 struct blk_plug plug;
554 int error = 0;
556 ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
558 blk_start_plug(&plug);
559 list_for_each_entry(busyp, list, list) {
560 trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
561 busyp->length);
563 error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
564 XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
565 XFS_FSB_TO_BB(mp, busyp->length),
566 GFP_NOFS, 0, &bio);
567 if (error && error != -EOPNOTSUPP) {
568 xfs_info(mp,
569 "discard failed for extent [0x%llx,%u], error %d",
570 (unsigned long long)busyp->bno,
571 busyp->length,
572 error);
573 break;
577 if (bio) {
578 bio->bi_private = ctx;
579 bio->bi_end_io = xlog_discard_endio;
580 submit_bio(bio);
581 } else {
582 xlog_discard_endio_work(&ctx->discard_endio_work);
584 blk_finish_plug(&plug);
588 * Mark all items committed and clear busy extents. We free the log vector
589 * chains in a separate pass so that we unpin the log items as quickly as
590 * possible.
592 static void
593 xlog_cil_committed(
594 void *args,
595 int abort)
597 struct xfs_cil_ctx *ctx = args;
598 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
600 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
601 ctx->start_lsn, abort);
603 xfs_extent_busy_sort(&ctx->busy_extents);
604 xfs_extent_busy_clear(mp, &ctx->busy_extents,
605 (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
608 * If we are aborting the commit, wake up anyone waiting on the
609 * committing list. If we don't, then a shutdown we can leave processes
610 * waiting in xlog_cil_force_lsn() waiting on a sequence commit that
611 * will never happen because we aborted it.
613 spin_lock(&ctx->cil->xc_push_lock);
614 if (abort)
615 wake_up_all(&ctx->cil->xc_commit_wait);
616 list_del(&ctx->committing);
617 spin_unlock(&ctx->cil->xc_push_lock);
619 xlog_cil_free_logvec(ctx->lv_chain);
621 if (!list_empty(&ctx->busy_extents))
622 xlog_discard_busy_extents(mp, ctx);
623 else
624 kmem_free(ctx);
628 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
629 * is a background flush and so we can chose to ignore it. Otherwise, if the
630 * current sequence is the same as @push_seq we need to do a flush. If
631 * @push_seq is less than the current sequence, then it has already been
632 * flushed and we don't need to do anything - the caller will wait for it to
633 * complete if necessary.
635 * @push_seq is a value rather than a flag because that allows us to do an
636 * unlocked check of the sequence number for a match. Hence we can allows log
637 * forces to run racily and not issue pushes for the same sequence twice. If we
638 * get a race between multiple pushes for the same sequence they will block on
639 * the first one and then abort, hence avoiding needless pushes.
641 STATIC int
642 xlog_cil_push(
643 struct xlog *log)
645 struct xfs_cil *cil = log->l_cilp;
646 struct xfs_log_vec *lv;
647 struct xfs_cil_ctx *ctx;
648 struct xfs_cil_ctx *new_ctx;
649 struct xlog_in_core *commit_iclog;
650 struct xlog_ticket *tic;
651 int num_iovecs;
652 int error = 0;
653 struct xfs_trans_header thdr;
654 struct xfs_log_iovec lhdr;
655 struct xfs_log_vec lvhdr = { NULL };
656 xfs_lsn_t commit_lsn;
657 xfs_lsn_t push_seq;
659 if (!cil)
660 return 0;
662 new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
663 new_ctx->ticket = xlog_cil_ticket_alloc(log);
665 down_write(&cil->xc_ctx_lock);
666 ctx = cil->xc_ctx;
668 spin_lock(&cil->xc_push_lock);
669 push_seq = cil->xc_push_seq;
670 ASSERT(push_seq <= ctx->sequence);
673 * Check if we've anything to push. If there is nothing, then we don't
674 * move on to a new sequence number and so we have to be able to push
675 * this sequence again later.
677 if (list_empty(&cil->xc_cil)) {
678 cil->xc_push_seq = 0;
679 spin_unlock(&cil->xc_push_lock);
680 goto out_skip;
684 /* check for a previously pushed seqeunce */
685 if (push_seq < cil->xc_ctx->sequence) {
686 spin_unlock(&cil->xc_push_lock);
687 goto out_skip;
691 * We are now going to push this context, so add it to the committing
692 * list before we do anything else. This ensures that anyone waiting on
693 * this push can easily detect the difference between a "push in
694 * progress" and "CIL is empty, nothing to do".
696 * IOWs, a wait loop can now check for:
697 * the current sequence not being found on the committing list;
698 * an empty CIL; and
699 * an unchanged sequence number
700 * to detect a push that had nothing to do and therefore does not need
701 * waiting on. If the CIL is not empty, we get put on the committing
702 * list before emptying the CIL and bumping the sequence number. Hence
703 * an empty CIL and an unchanged sequence number means we jumped out
704 * above after doing nothing.
706 * Hence the waiter will either find the commit sequence on the
707 * committing list or the sequence number will be unchanged and the CIL
708 * still dirty. In that latter case, the push has not yet started, and
709 * so the waiter will have to continue trying to check the CIL
710 * committing list until it is found. In extreme cases of delay, the
711 * sequence may fully commit between the attempts the wait makes to wait
712 * on the commit sequence.
714 list_add(&ctx->committing, &cil->xc_committing);
715 spin_unlock(&cil->xc_push_lock);
718 * pull all the log vectors off the items in the CIL, and
719 * remove the items from the CIL. We don't need the CIL lock
720 * here because it's only needed on the transaction commit
721 * side which is currently locked out by the flush lock.
723 lv = NULL;
724 num_iovecs = 0;
725 while (!list_empty(&cil->xc_cil)) {
726 struct xfs_log_item *item;
728 item = list_first_entry(&cil->xc_cil,
729 struct xfs_log_item, li_cil);
730 list_del_init(&item->li_cil);
731 if (!ctx->lv_chain)
732 ctx->lv_chain = item->li_lv;
733 else
734 lv->lv_next = item->li_lv;
735 lv = item->li_lv;
736 item->li_lv = NULL;
737 num_iovecs += lv->lv_niovecs;
741 * initialise the new context and attach it to the CIL. Then attach
742 * the current context to the CIL committing lsit so it can be found
743 * during log forces to extract the commit lsn of the sequence that
744 * needs to be forced.
746 INIT_LIST_HEAD(&new_ctx->committing);
747 INIT_LIST_HEAD(&new_ctx->busy_extents);
748 new_ctx->sequence = ctx->sequence + 1;
749 new_ctx->cil = cil;
750 cil->xc_ctx = new_ctx;
753 * The switch is now done, so we can drop the context lock and move out
754 * of a shared context. We can't just go straight to the commit record,
755 * though - we need to synchronise with previous and future commits so
756 * that the commit records are correctly ordered in the log to ensure
757 * that we process items during log IO completion in the correct order.
759 * For example, if we get an EFI in one checkpoint and the EFD in the
760 * next (e.g. due to log forces), we do not want the checkpoint with
761 * the EFD to be committed before the checkpoint with the EFI. Hence
762 * we must strictly order the commit records of the checkpoints so
763 * that: a) the checkpoint callbacks are attached to the iclogs in the
764 * correct order; and b) the checkpoints are replayed in correct order
765 * in log recovery.
767 * Hence we need to add this context to the committing context list so
768 * that higher sequences will wait for us to write out a commit record
769 * before they do.
771 * xfs_log_force_lsn requires us to mirror the new sequence into the cil
772 * structure atomically with the addition of this sequence to the
773 * committing list. This also ensures that we can do unlocked checks
774 * against the current sequence in log forces without risking
775 * deferencing a freed context pointer.
777 spin_lock(&cil->xc_push_lock);
778 cil->xc_current_sequence = new_ctx->sequence;
779 spin_unlock(&cil->xc_push_lock);
780 up_write(&cil->xc_ctx_lock);
783 * Build a checkpoint transaction header and write it to the log to
784 * begin the transaction. We need to account for the space used by the
785 * transaction header here as it is not accounted for in xlog_write().
787 * The LSN we need to pass to the log items on transaction commit is
788 * the LSN reported by the first log vector write. If we use the commit
789 * record lsn then we can move the tail beyond the grant write head.
791 tic = ctx->ticket;
792 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
793 thdr.th_type = XFS_TRANS_CHECKPOINT;
794 thdr.th_tid = tic->t_tid;
795 thdr.th_num_items = num_iovecs;
796 lhdr.i_addr = &thdr;
797 lhdr.i_len = sizeof(xfs_trans_header_t);
798 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
799 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
801 lvhdr.lv_niovecs = 1;
802 lvhdr.lv_iovecp = &lhdr;
803 lvhdr.lv_next = ctx->lv_chain;
805 error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
806 if (error)
807 goto out_abort_free_ticket;
810 * now that we've written the checkpoint into the log, strictly
811 * order the commit records so replay will get them in the right order.
813 restart:
814 spin_lock(&cil->xc_push_lock);
815 list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
817 * Avoid getting stuck in this loop because we were woken by the
818 * shutdown, but then went back to sleep once already in the
819 * shutdown state.
821 if (XLOG_FORCED_SHUTDOWN(log)) {
822 spin_unlock(&cil->xc_push_lock);
823 goto out_abort_free_ticket;
827 * Higher sequences will wait for this one so skip them.
828 * Don't wait for our own sequence, either.
830 if (new_ctx->sequence >= ctx->sequence)
831 continue;
832 if (!new_ctx->commit_lsn) {
834 * It is still being pushed! Wait for the push to
835 * complete, then start again from the beginning.
837 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
838 goto restart;
841 spin_unlock(&cil->xc_push_lock);
843 /* xfs_log_done always frees the ticket on error. */
844 commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, false);
845 if (commit_lsn == -1)
846 goto out_abort;
848 /* attach all the transactions w/ busy extents to iclog */
849 ctx->log_cb.cb_func = xlog_cil_committed;
850 ctx->log_cb.cb_arg = ctx;
851 error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
852 if (error)
853 goto out_abort;
856 * now the checkpoint commit is complete and we've attached the
857 * callbacks to the iclog we can assign the commit LSN to the context
858 * and wake up anyone who is waiting for the commit to complete.
860 spin_lock(&cil->xc_push_lock);
861 ctx->commit_lsn = commit_lsn;
862 wake_up_all(&cil->xc_commit_wait);
863 spin_unlock(&cil->xc_push_lock);
865 /* release the hounds! */
866 return xfs_log_release_iclog(log->l_mp, commit_iclog);
868 out_skip:
869 up_write(&cil->xc_ctx_lock);
870 xfs_log_ticket_put(new_ctx->ticket);
871 kmem_free(new_ctx);
872 return 0;
874 out_abort_free_ticket:
875 xfs_log_ticket_put(tic);
876 out_abort:
877 xlog_cil_committed(ctx, XFS_LI_ABORTED);
878 return -EIO;
881 static void
882 xlog_cil_push_work(
883 struct work_struct *work)
885 struct xfs_cil *cil = container_of(work, struct xfs_cil,
886 xc_push_work);
887 xlog_cil_push(cil->xc_log);
891 * We need to push CIL every so often so we don't cache more than we can fit in
892 * the log. The limit really is that a checkpoint can't be more than half the
893 * log (the current checkpoint is not allowed to overwrite the previous
894 * checkpoint), but commit latency and memory usage limit this to a smaller
895 * size.
897 static void
898 xlog_cil_push_background(
899 struct xlog *log)
901 struct xfs_cil *cil = log->l_cilp;
904 * The cil won't be empty because we are called while holding the
905 * context lock so whatever we added to the CIL will still be there
907 ASSERT(!list_empty(&cil->xc_cil));
910 * don't do a background push if we haven't used up all the
911 * space available yet.
913 if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
914 return;
916 spin_lock(&cil->xc_push_lock);
917 if (cil->xc_push_seq < cil->xc_current_sequence) {
918 cil->xc_push_seq = cil->xc_current_sequence;
919 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
921 spin_unlock(&cil->xc_push_lock);
926 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
927 * number that is passed. When it returns, the work will be queued for
928 * @push_seq, but it won't be completed. The caller is expected to do any
929 * waiting for push_seq to complete if it is required.
931 static void
932 xlog_cil_push_now(
933 struct xlog *log,
934 xfs_lsn_t push_seq)
936 struct xfs_cil *cil = log->l_cilp;
938 if (!cil)
939 return;
941 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
943 /* start on any pending background push to minimise wait time on it */
944 flush_work(&cil->xc_push_work);
947 * If the CIL is empty or we've already pushed the sequence then
948 * there's no work we need to do.
950 spin_lock(&cil->xc_push_lock);
951 if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
952 spin_unlock(&cil->xc_push_lock);
953 return;
956 cil->xc_push_seq = push_seq;
957 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
958 spin_unlock(&cil->xc_push_lock);
961 bool
962 xlog_cil_empty(
963 struct xlog *log)
965 struct xfs_cil *cil = log->l_cilp;
966 bool empty = false;
968 spin_lock(&cil->xc_push_lock);
969 if (list_empty(&cil->xc_cil))
970 empty = true;
971 spin_unlock(&cil->xc_push_lock);
972 return empty;
976 * Commit a transaction with the given vector to the Committed Item List.
978 * To do this, we need to format the item, pin it in memory if required and
979 * account for the space used by the transaction. Once we have done that we
980 * need to release the unused reservation for the transaction, attach the
981 * transaction to the checkpoint context so we carry the busy extents through
982 * to checkpoint completion, and then unlock all the items in the transaction.
984 * Called with the context lock already held in read mode to lock out
985 * background commit, returns without it held once background commits are
986 * allowed again.
988 void
989 xfs_log_commit_cil(
990 struct xfs_mount *mp,
991 struct xfs_trans *tp,
992 xfs_lsn_t *commit_lsn,
993 bool regrant)
995 struct xlog *log = mp->m_log;
996 struct xfs_cil *cil = log->l_cilp;
997 xfs_lsn_t xc_commit_lsn;
1000 * Do all necessary memory allocation before we lock the CIL.
1001 * This ensures the allocation does not deadlock with a CIL
1002 * push in memory reclaim (e.g. from kswapd).
1004 xlog_cil_alloc_shadow_bufs(log, tp);
1006 /* lock out background commit */
1007 down_read(&cil->xc_ctx_lock);
1009 xlog_cil_insert_items(log, tp);
1011 xc_commit_lsn = cil->xc_ctx->sequence;
1012 if (commit_lsn)
1013 *commit_lsn = xc_commit_lsn;
1015 xfs_log_done(mp, tp->t_ticket, NULL, regrant);
1016 xfs_trans_unreserve_and_mod_sb(tp);
1019 * Once all the items of the transaction have been copied to the CIL,
1020 * the items can be unlocked and freed.
1022 * This needs to be done before we drop the CIL context lock because we
1023 * have to update state in the log items and unlock them before they go
1024 * to disk. If we don't, then the CIL checkpoint can race with us and
1025 * we can run checkpoint completion before we've updated and unlocked
1026 * the log items. This affects (at least) processing of stale buffers,
1027 * inodes and EFIs.
1029 xfs_trans_free_items(tp, xc_commit_lsn, false);
1031 xlog_cil_push_background(log);
1033 up_read(&cil->xc_ctx_lock);
1037 * Conditionally push the CIL based on the sequence passed in.
1039 * We only need to push if we haven't already pushed the sequence
1040 * number given. Hence the only time we will trigger a push here is
1041 * if the push sequence is the same as the current context.
1043 * We return the current commit lsn to allow the callers to determine if a
1044 * iclog flush is necessary following this call.
1046 xfs_lsn_t
1047 xlog_cil_force_lsn(
1048 struct xlog *log,
1049 xfs_lsn_t sequence)
1051 struct xfs_cil *cil = log->l_cilp;
1052 struct xfs_cil_ctx *ctx;
1053 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
1055 ASSERT(sequence <= cil->xc_current_sequence);
1058 * check to see if we need to force out the current context.
1059 * xlog_cil_push() handles racing pushes for the same sequence,
1060 * so no need to deal with it here.
1062 restart:
1063 xlog_cil_push_now(log, sequence);
1066 * See if we can find a previous sequence still committing.
1067 * We need to wait for all previous sequence commits to complete
1068 * before allowing the force of push_seq to go ahead. Hence block
1069 * on commits for those as well.
1071 spin_lock(&cil->xc_push_lock);
1072 list_for_each_entry(ctx, &cil->xc_committing, committing) {
1074 * Avoid getting stuck in this loop because we were woken by the
1075 * shutdown, but then went back to sleep once already in the
1076 * shutdown state.
1078 if (XLOG_FORCED_SHUTDOWN(log))
1079 goto out_shutdown;
1080 if (ctx->sequence > sequence)
1081 continue;
1082 if (!ctx->commit_lsn) {
1084 * It is still being pushed! Wait for the push to
1085 * complete, then start again from the beginning.
1087 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1088 goto restart;
1090 if (ctx->sequence != sequence)
1091 continue;
1092 /* found it! */
1093 commit_lsn = ctx->commit_lsn;
1097 * The call to xlog_cil_push_now() executes the push in the background.
1098 * Hence by the time we have got here it our sequence may not have been
1099 * pushed yet. This is true if the current sequence still matches the
1100 * push sequence after the above wait loop and the CIL still contains
1101 * dirty objects. This is guaranteed by the push code first adding the
1102 * context to the committing list before emptying the CIL.
1104 * Hence if we don't find the context in the committing list and the
1105 * current sequence number is unchanged then the CIL contents are
1106 * significant. If the CIL is empty, if means there was nothing to push
1107 * and that means there is nothing to wait for. If the CIL is not empty,
1108 * it means we haven't yet started the push, because if it had started
1109 * we would have found the context on the committing list.
1111 if (sequence == cil->xc_current_sequence &&
1112 !list_empty(&cil->xc_cil)) {
1113 spin_unlock(&cil->xc_push_lock);
1114 goto restart;
1117 spin_unlock(&cil->xc_push_lock);
1118 return commit_lsn;
1121 * We detected a shutdown in progress. We need to trigger the log force
1122 * to pass through it's iclog state machine error handling, even though
1123 * we are already in a shutdown state. Hence we can't return
1124 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
1125 * LSN is already stable), so we return a zero LSN instead.
1127 out_shutdown:
1128 spin_unlock(&cil->xc_push_lock);
1129 return 0;
1133 * Check if the current log item was first committed in this sequence.
1134 * We can't rely on just the log item being in the CIL, we have to check
1135 * the recorded commit sequence number.
1137 * Note: for this to be used in a non-racy manner, it has to be called with
1138 * CIL flushing locked out. As a result, it should only be used during the
1139 * transaction commit process when deciding what to format into the item.
1141 bool
1142 xfs_log_item_in_current_chkpt(
1143 struct xfs_log_item *lip)
1145 struct xfs_cil_ctx *ctx;
1147 if (list_empty(&lip->li_cil))
1148 return false;
1150 ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
1153 * li_seq is written on the first commit of a log item to record the
1154 * first checkpoint it is written to. Hence if it is different to the
1155 * current sequence, we're in a new checkpoint.
1157 if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
1158 return false;
1159 return true;
1163 * Perform initial CIL structure initialisation.
1166 xlog_cil_init(
1167 struct xlog *log)
1169 struct xfs_cil *cil;
1170 struct xfs_cil_ctx *ctx;
1172 cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
1173 if (!cil)
1174 return -ENOMEM;
1176 ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
1177 if (!ctx) {
1178 kmem_free(cil);
1179 return -ENOMEM;
1182 INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
1183 INIT_LIST_HEAD(&cil->xc_cil);
1184 INIT_LIST_HEAD(&cil->xc_committing);
1185 spin_lock_init(&cil->xc_cil_lock);
1186 spin_lock_init(&cil->xc_push_lock);
1187 init_rwsem(&cil->xc_ctx_lock);
1188 init_waitqueue_head(&cil->xc_commit_wait);
1190 INIT_LIST_HEAD(&ctx->committing);
1191 INIT_LIST_HEAD(&ctx->busy_extents);
1192 ctx->sequence = 1;
1193 ctx->cil = cil;
1194 cil->xc_ctx = ctx;
1195 cil->xc_current_sequence = ctx->sequence;
1197 cil->xc_log = log;
1198 log->l_cilp = cil;
1199 return 0;
1202 void
1203 xlog_cil_destroy(
1204 struct xlog *log)
1206 if (log->l_cilp->xc_ctx) {
1207 if (log->l_cilp->xc_ctx->ticket)
1208 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
1209 kmem_free(log->l_cilp->xc_ctx);
1212 ASSERT(list_empty(&log->l_cilp->xc_cil));
1213 kmem_free(log->l_cilp);