1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_shared.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_extent_busy.h"
14 #include "xfs_trans.h"
15 #include "xfs_trans_priv.h"
17 #include "xfs_log_priv.h"
18 #include "xfs_trace.h"
20 struct workqueue_struct *xfs_discard_wq;
23 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
24 * recover, so we don't allow failure here. Also, we allocate in a context that
25 * we don't want to be issuing transactions from, so we need to tell the
26 * allocation code this as well.
28 * We don't reserve any space for the ticket - we are going to steal whatever
29 * space we require from transactions as they commit. To ensure we reserve all
30 * the space required, we need to set the current reservation of the ticket to
31 * zero so that we know to steal the initial transaction overhead from the
32 * first transaction commit.
34 static struct xlog_ticket *
35 xlog_cil_ticket_alloc(
38 struct xlog_ticket *tic;
40 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0);
43 * set the current reservation to zero so we know to steal the basic
44 * transaction overhead reservation from the first transaction commit.
51 * Unavoidable forward declaration - xlog_cil_push_work() calls
52 * xlog_cil_ctx_alloc() itself.
54 static void xlog_cil_push_work(struct work_struct *work);
56 static struct xfs_cil_ctx *
57 xlog_cil_ctx_alloc(void)
59 struct xfs_cil_ctx *ctx;
61 ctx = kmem_zalloc(sizeof(*ctx), KM_NOFS);
62 INIT_LIST_HEAD(&ctx->committing);
63 INIT_LIST_HEAD(&ctx->busy_extents);
64 INIT_WORK(&ctx->push_work, xlog_cil_push_work);
71 struct xfs_cil_ctx *ctx)
73 ctx->sequence = ++cil->xc_current_sequence;
79 * After the first stage of log recovery is done, we know where the head and
80 * tail of the log are. We need this log initialisation done before we can
81 * initialise the first CIL checkpoint context.
83 * Here we allocate a log ticket to track space usage during a CIL push. This
84 * ticket is passed to xlog_write() directly so that we don't slowly leak log
85 * space by failing to account for space used by log headers and additional
86 * region headers for split regions.
89 xlog_cil_init_post_recovery(
92 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
93 log->l_cilp->xc_ctx->sequence = 1;
100 return round_up((sizeof(struct xfs_log_vec) +
101 niovecs * sizeof(struct xfs_log_iovec)),
106 * Allocate or pin log vector buffers for CIL insertion.
108 * The CIL currently uses disposable buffers for copying a snapshot of the
109 * modified items into the log during a push. The biggest problem with this is
110 * the requirement to allocate the disposable buffer during the commit if:
111 * a) does not exist; or
114 * If we do this allocation within xlog_cil_insert_format_items(), it is done
115 * under the xc_ctx_lock, which means that a CIL push cannot occur during
116 * the memory allocation. This means that we have a potential deadlock situation
117 * under low memory conditions when we have lots of dirty metadata pinned in
118 * the CIL and we need a CIL commit to occur to free memory.
120 * To avoid this, we need to move the memory allocation outside the
121 * xc_ctx_lock, but because the log vector buffers are disposable, that opens
122 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log
123 * vector buffers between the check and the formatting of the item into the
124 * log vector buffer within the xc_ctx_lock.
126 * Because the log vector buffer needs to be unchanged during the CIL push
127 * process, we cannot share the buffer between the transaction commit (which
128 * modifies the buffer) and the CIL push context that is writing the changes
129 * into the log. This means skipping preallocation of buffer space is
130 * unreliable, but we most definitely do not want to be allocating and freeing
131 * buffers unnecessarily during commits when overwrites can be done safely.
133 * The simplest solution to this problem is to allocate a shadow buffer when a
134 * log item is committed for the second time, and then to only use this buffer
135 * if necessary. The buffer can remain attached to the log item until such time
136 * it is needed, and this is the buffer that is reallocated to match the size of
137 * the incoming modification. Then during the formatting of the item we can swap
138 * the active buffer with the new one if we can't reuse the existing buffer. We
139 * don't free the old buffer as it may be reused on the next modification if
140 * it's size is right, otherwise we'll free and reallocate it at that point.
142 * This function builds a vector for the changes in each log item in the
143 * transaction. It then works out the length of the buffer needed for each log
144 * item, allocates them and attaches the vector to the log item in preparation
145 * for the formatting step which occurs under the xc_ctx_lock.
147 * While this means the memory footprint goes up, it avoids the repeated
148 * alloc/free pattern that repeated modifications of an item would otherwise
149 * cause, and hence minimises the CPU overhead of such behaviour.
152 xlog_cil_alloc_shadow_bufs(
154 struct xfs_trans *tp)
156 struct xfs_log_item *lip;
158 list_for_each_entry(lip, &tp->t_items, li_trans) {
159 struct xfs_log_vec *lv;
163 bool ordered = false;
165 /* Skip items which aren't dirty in this transaction. */
166 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
169 /* get number of vecs and size of data to be stored */
170 lip->li_ops->iop_size(lip, &niovecs, &nbytes);
173 * Ordered items need to be tracked but we do not wish to write
174 * them. We need a logvec to track the object, but we do not
175 * need an iovec or buffer to be allocated for copying data.
177 if (niovecs == XFS_LOG_VEC_ORDERED) {
184 * We 64-bit align the length of each iovec so that the start
185 * of the next one is naturally aligned. We'll need to
186 * account for that slack space here. Then round nbytes up
187 * to 64-bit alignment so that the initial buffer alignment is
188 * easy to calculate and verify.
190 nbytes += niovecs * sizeof(uint64_t);
191 nbytes = round_up(nbytes, sizeof(uint64_t));
194 * The data buffer needs to start 64-bit aligned, so round up
195 * that space to ensure we can align it appropriately and not
196 * overrun the buffer.
198 buf_size = nbytes + xlog_cil_iovec_space(niovecs);
201 * if we have no shadow buffer, or it is too small, we need to
204 if (!lip->li_lv_shadow ||
205 buf_size > lip->li_lv_shadow->lv_size) {
208 * We free and allocate here as a realloc would copy
209 * unnecessary data. We don't use kmem_zalloc() for the
210 * same reason - we don't need to zero the data area in
211 * the buffer, only the log vector header and the iovec
214 kmem_free(lip->li_lv_shadow);
217 * We are in transaction context, which means this
218 * allocation will pick up GFP_NOFS from the
219 * memalloc_nofs_save/restore context the transaction
220 * holds. This means we can use GFP_KERNEL here so the
221 * generic kvmalloc() code will run vmalloc on
222 * contiguous page allocation failure as we require.
224 lv = kvmalloc(buf_size, GFP_KERNEL);
225 memset(lv, 0, xlog_cil_iovec_space(niovecs));
228 lv->lv_size = buf_size;
230 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
232 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
233 lip->li_lv_shadow = lv;
235 /* same or smaller, optimise common overwrite case */
236 lv = lip->li_lv_shadow;
238 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
245 /* Ensure the lv is set up according to ->iop_size */
246 lv->lv_niovecs = niovecs;
248 /* The allocated data region lies beyond the iovec region */
249 lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);
255 * Prepare the log item for insertion into the CIL. Calculate the difference in
256 * log space and vectors it will consume, and if it is a new item pin it as
260 xfs_cil_prepare_item(
262 struct xfs_log_vec *lv,
263 struct xfs_log_vec *old_lv,
267 /* Account for the new LV being passed in */
268 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
269 *diff_len += lv->lv_bytes;
270 *diff_iovecs += lv->lv_niovecs;
274 * If there is no old LV, this is the first time we've seen the item in
275 * this CIL context and so we need to pin it. If we are replacing the
276 * old_lv, then remove the space it accounts for and make it the shadow
277 * buffer for later freeing. In both cases we are now switching to the
278 * shadow buffer, so update the pointer to it appropriately.
281 if (lv->lv_item->li_ops->iop_pin)
282 lv->lv_item->li_ops->iop_pin(lv->lv_item);
283 lv->lv_item->li_lv_shadow = NULL;
284 } else if (old_lv != lv) {
285 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
287 *diff_len -= old_lv->lv_bytes;
288 *diff_iovecs -= old_lv->lv_niovecs;
289 lv->lv_item->li_lv_shadow = old_lv;
292 /* attach new log vector to log item */
293 lv->lv_item->li_lv = lv;
296 * If this is the first time the item is being committed to the
297 * CIL, store the sequence number on the log item so we can
298 * tell in future commits whether this is the first checkpoint
299 * the item is being committed into.
301 if (!lv->lv_item->li_seq)
302 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
306 * Format log item into a flat buffers
308 * For delayed logging, we need to hold a formatted buffer containing all the
309 * changes on the log item. This enables us to relog the item in memory and
310 * write it out asynchronously without needing to relock the object that was
311 * modified at the time it gets written into the iclog.
313 * This function takes the prepared log vectors attached to each log item, and
314 * formats the changes into the log vector buffer. The buffer it uses is
315 * dependent on the current state of the vector in the CIL - the shadow lv is
316 * guaranteed to be large enough for the current modification, but we will only
317 * use that if we can't reuse the existing lv. If we can't reuse the existing
318 * lv, then simple swap it out for the shadow lv. We don't free it - that is
319 * done lazily either by th enext modification or the freeing of the log item.
321 * We don't set up region headers during this process; we simply copy the
322 * regions into the flat buffer. We can do this because we still have to do a
323 * formatting step to write the regions into the iclog buffer. Writing the
324 * ophdrs during the iclog write means that we can support splitting large
325 * regions across iclog boundares without needing a change in the format of the
326 * item/region encapsulation.
328 * Hence what we need to do now is change the rewrite the vector array to point
329 * to the copied region inside the buffer we just allocated. This allows us to
330 * format the regions into the iclog as though they are being formatted
331 * directly out of the objects themselves.
334 xlog_cil_insert_format_items(
336 struct xfs_trans *tp,
340 struct xfs_log_item *lip;
343 /* Bail out if we didn't find a log item. */
344 if (list_empty(&tp->t_items)) {
349 list_for_each_entry(lip, &tp->t_items, li_trans) {
350 struct xfs_log_vec *lv;
351 struct xfs_log_vec *old_lv = NULL;
352 struct xfs_log_vec *shadow;
353 bool ordered = false;
355 /* Skip items which aren't dirty in this transaction. */
356 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
360 * The formatting size information is already attached to
361 * the shadow lv on the log item.
363 shadow = lip->li_lv_shadow;
364 if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)
367 /* Skip items that do not have any vectors for writing */
368 if (!shadow->lv_niovecs && !ordered)
371 /* compare to existing item size */
373 if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {
374 /* same or smaller, optimise common overwrite case */
382 * set the item up as though it is a new insertion so
383 * that the space reservation accounting is correct.
385 *diff_iovecs -= lv->lv_niovecs;
386 *diff_len -= lv->lv_bytes;
388 /* Ensure the lv is set up according to ->iop_size */
389 lv->lv_niovecs = shadow->lv_niovecs;
391 /* reset the lv buffer information for new formatting */
394 lv->lv_buf = (char *)lv +
395 xlog_cil_iovec_space(lv->lv_niovecs);
397 /* switch to shadow buffer! */
401 /* track as an ordered logvec */
402 ASSERT(lip->li_lv == NULL);
407 ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
408 lip->li_ops->iop_format(lip, lv);
410 xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
415 * Insert the log items into the CIL and calculate the difference in space
416 * consumed by the item. Add the space to the checkpoint ticket and calculate
417 * if the change requires additional log metadata. If it does, take that space
418 * as well. Remove the amount of space we added to the checkpoint ticket from
419 * the current transaction ticket so that the accounting works out correctly.
422 xlog_cil_insert_items(
424 struct xfs_trans *tp)
426 struct xfs_cil *cil = log->l_cilp;
427 struct xfs_cil_ctx *ctx = cil->xc_ctx;
428 struct xfs_log_item *lip;
432 int iovhdr_res = 0, split_res = 0, ctx_res = 0;
437 * We can do this safely because the context can't checkpoint until we
438 * are done so it doesn't matter exactly how we update the CIL.
440 xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
442 spin_lock(&cil->xc_cil_lock);
444 /* account for space used by new iovec headers */
445 iovhdr_res = diff_iovecs * sizeof(xlog_op_header_t);
447 ctx->nvecs += diff_iovecs;
449 /* attach the transaction to the CIL if it has any busy extents */
450 if (!list_empty(&tp->t_busy))
451 list_splice_init(&tp->t_busy, &ctx->busy_extents);
454 * Now transfer enough transaction reservation to the context ticket
455 * for the checkpoint. The context ticket is special - the unit
456 * reservation has to grow as well as the current reservation as we
457 * steal from tickets so we can correctly determine the space used
458 * during the transaction commit.
460 if (ctx->ticket->t_curr_res == 0) {
461 ctx_res = ctx->ticket->t_unit_res;
462 ctx->ticket->t_curr_res = ctx_res;
463 tp->t_ticket->t_curr_res -= ctx_res;
466 /* do we need space for more log record headers? */
467 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
468 if (len > 0 && (ctx->space_used / iclog_space !=
469 (ctx->space_used + len) / iclog_space)) {
470 split_res = (len + iclog_space - 1) / iclog_space;
471 /* need to take into account split region headers, too */
472 split_res *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
473 ctx->ticket->t_unit_res += split_res;
474 ctx->ticket->t_curr_res += split_res;
475 tp->t_ticket->t_curr_res -= split_res;
476 ASSERT(tp->t_ticket->t_curr_res >= len);
478 tp->t_ticket->t_curr_res -= len;
479 ctx->space_used += len;
482 * If we've overrun the reservation, dump the tx details before we move
483 * the log items. Shutdown is imminent...
485 if (WARN_ON(tp->t_ticket->t_curr_res < 0)) {
486 xfs_warn(log->l_mp, "Transaction log reservation overrun:");
488 " log items: %d bytes (iov hdrs: %d bytes)",
490 xfs_warn(log->l_mp, " split region headers: %d bytes",
492 xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res);
493 xlog_print_trans(tp);
497 * Now (re-)position everything modified at the tail of the CIL.
498 * We do this here so we only need to take the CIL lock once during
499 * the transaction commit.
501 list_for_each_entry(lip, &tp->t_items, li_trans) {
503 /* Skip items which aren't dirty in this transaction. */
504 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
508 * Only move the item if it isn't already at the tail. This is
509 * to prevent a transient list_empty() state when reinserting
510 * an item that is already the only item in the CIL.
512 if (!list_is_last(&lip->li_cil, &cil->xc_cil))
513 list_move_tail(&lip->li_cil, &cil->xc_cil);
516 spin_unlock(&cil->xc_cil_lock);
518 if (tp->t_ticket->t_curr_res < 0)
519 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
523 xlog_cil_free_logvec(
524 struct xfs_log_vec *log_vector)
526 struct xfs_log_vec *lv;
528 for (lv = log_vector; lv; ) {
529 struct xfs_log_vec *next = lv->lv_next;
536 xlog_discard_endio_work(
537 struct work_struct *work)
539 struct xfs_cil_ctx *ctx =
540 container_of(work, struct xfs_cil_ctx, discard_endio_work);
541 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
543 xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
548 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
549 * pagb_lock. Note that we need a unbounded workqueue, otherwise we might
550 * get the execution delayed up to 30 seconds for weird reasons.
556 struct xfs_cil_ctx *ctx = bio->bi_private;
558 INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work);
559 queue_work(xfs_discard_wq, &ctx->discard_endio_work);
564 xlog_discard_busy_extents(
565 struct xfs_mount *mp,
566 struct xfs_cil_ctx *ctx)
568 struct list_head *list = &ctx->busy_extents;
569 struct xfs_extent_busy *busyp;
570 struct bio *bio = NULL;
571 struct blk_plug plug;
574 ASSERT(xfs_has_discard(mp));
576 blk_start_plug(&plug);
577 list_for_each_entry(busyp, list, list) {
578 trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
581 error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
582 XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
583 XFS_FSB_TO_BB(mp, busyp->length),
585 if (error && error != -EOPNOTSUPP) {
587 "discard failed for extent [0x%llx,%u], error %d",
588 (unsigned long long)busyp->bno,
596 bio->bi_private = ctx;
597 bio->bi_end_io = xlog_discard_endio;
600 xlog_discard_endio_work(&ctx->discard_endio_work);
602 blk_finish_plug(&plug);
606 * Mark all items committed and clear busy extents. We free the log vector
607 * chains in a separate pass so that we unpin the log items as quickly as
612 struct xfs_cil_ctx *ctx)
614 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
615 bool abort = xlog_is_shutdown(ctx->cil->xc_log);
618 * If the I/O failed, we're aborting the commit and already shutdown.
619 * Wake any commit waiters before aborting the log items so we don't
620 * block async log pushers on callbacks. Async log pushers explicitly do
621 * not wait on log force completion because they may be holding locks
622 * required to unpin items.
625 spin_lock(&ctx->cil->xc_push_lock);
626 wake_up_all(&ctx->cil->xc_start_wait);
627 wake_up_all(&ctx->cil->xc_commit_wait);
628 spin_unlock(&ctx->cil->xc_push_lock);
631 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
632 ctx->start_lsn, abort);
634 xfs_extent_busy_sort(&ctx->busy_extents);
635 xfs_extent_busy_clear(mp, &ctx->busy_extents,
636 xfs_has_discard(mp) && !abort);
638 spin_lock(&ctx->cil->xc_push_lock);
639 list_del(&ctx->committing);
640 spin_unlock(&ctx->cil->xc_push_lock);
642 xlog_cil_free_logvec(ctx->lv_chain);
644 if (!list_empty(&ctx->busy_extents))
645 xlog_discard_busy_extents(mp, ctx);
651 xlog_cil_process_committed(
652 struct list_head *list)
654 struct xfs_cil_ctx *ctx;
656 while ((ctx = list_first_entry_or_null(list,
657 struct xfs_cil_ctx, iclog_entry))) {
658 list_del(&ctx->iclog_entry);
659 xlog_cil_committed(ctx);
664 * Record the LSN of the iclog we were just granted space to start writing into.
665 * If the context doesn't have a start_lsn recorded, then this iclog will
666 * contain the start record for the checkpoint. Otherwise this write contains
667 * the commit record for the checkpoint.
670 xlog_cil_set_ctx_write_state(
671 struct xfs_cil_ctx *ctx,
672 struct xlog_in_core *iclog)
674 struct xfs_cil *cil = ctx->cil;
675 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
677 ASSERT(!ctx->commit_lsn);
678 if (!ctx->start_lsn) {
679 spin_lock(&cil->xc_push_lock);
681 * The LSN we need to pass to the log items on transaction
682 * commit is the LSN reported by the first log vector write, not
683 * the commit lsn. If we use the commit record lsn then we can
684 * move the grant write head beyond the tail LSN and overwrite
687 ctx->start_lsn = lsn;
688 wake_up_all(&cil->xc_start_wait);
689 spin_unlock(&cil->xc_push_lock);
692 * Make sure the metadata we are about to overwrite in the log
693 * has been flushed to stable storage before this iclog is
696 spin_lock(&cil->xc_log->l_icloglock);
697 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
698 spin_unlock(&cil->xc_log->l_icloglock);
703 * Take a reference to the iclog for the context so that we still hold
704 * it when xlog_write is done and has released it. This means the
705 * context controls when the iclog is released for IO.
707 atomic_inc(&iclog->ic_refcnt);
710 * xlog_state_get_iclog_space() guarantees there is enough space in the
711 * iclog for an entire commit record, so we can attach the context
712 * callbacks now. This needs to be done before we make the commit_lsn
713 * visible to waiters so that checkpoints with commit records in the
714 * same iclog order their IO completion callbacks in the same order that
715 * the commit records appear in the iclog.
717 spin_lock(&cil->xc_log->l_icloglock);
718 list_add_tail(&ctx->iclog_entry, &iclog->ic_callbacks);
719 spin_unlock(&cil->xc_log->l_icloglock);
722 * Now we can record the commit LSN and wake anyone waiting for this
723 * sequence to have the ordered commit record assigned to a physical
724 * location in the log.
726 spin_lock(&cil->xc_push_lock);
727 ctx->commit_iclog = iclog;
728 ctx->commit_lsn = lsn;
729 wake_up_all(&cil->xc_commit_wait);
730 spin_unlock(&cil->xc_push_lock);
735 * Ensure that the order of log writes follows checkpoint sequence order. This
736 * relies on the context LSN being zero until the log write has guaranteed the
737 * LSN that the log write will start at via xlog_state_get_iclog_space().
745 xlog_cil_order_write(
748 enum _record_type record)
750 struct xfs_cil_ctx *ctx;
753 spin_lock(&cil->xc_push_lock);
754 list_for_each_entry(ctx, &cil->xc_committing, committing) {
756 * Avoid getting stuck in this loop because we were woken by the
757 * shutdown, but then went back to sleep once already in the
760 if (xlog_is_shutdown(cil->xc_log)) {
761 spin_unlock(&cil->xc_push_lock);
766 * Higher sequences will wait for this one so skip them.
767 * Don't wait for our own sequence, either.
769 if (ctx->sequence >= sequence)
772 /* Wait until the LSN for the record has been recorded. */
775 if (!ctx->start_lsn) {
776 xlog_wait(&cil->xc_start_wait, &cil->xc_push_lock);
781 if (!ctx->commit_lsn) {
782 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
788 spin_unlock(&cil->xc_push_lock);
793 * Write out the log vector change now attached to the CIL context. This will
794 * write a start record that needs to be strictly ordered in ascending CIL
795 * sequence order so that log recovery will always use in-order start LSNs when
796 * replaying checkpoints.
799 xlog_cil_write_chain(
800 struct xfs_cil_ctx *ctx,
801 struct xfs_log_vec *chain)
803 struct xlog *log = ctx->cil->xc_log;
806 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _START_RECORD);
809 return xlog_write(log, ctx, chain, ctx->ticket, XLOG_START_TRANS);
813 * Write out the commit record of a checkpoint transaction to close off a
814 * running log write. These commit records are strictly ordered in ascending CIL
815 * sequence order so that log recovery will always replay the checkpoints in the
819 xlog_cil_write_commit_record(
820 struct xfs_cil_ctx *ctx)
822 struct xlog *log = ctx->cil->xc_log;
823 struct xfs_log_iovec reg = {
826 .i_type = XLOG_REG_TYPE_COMMIT,
828 struct xfs_log_vec vec = {
834 if (xlog_is_shutdown(log))
837 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _COMMIT_RECORD);
841 error = xlog_write(log, ctx, &vec, ctx->ticket, XLOG_COMMIT_TRANS);
843 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
848 * Push the Committed Item List to the log.
850 * If the current sequence is the same as xc_push_seq we need to do a flush. If
851 * xc_push_seq is less than the current sequence, then it has already been
852 * flushed and we don't need to do anything - the caller will wait for it to
853 * complete if necessary.
855 * xc_push_seq is checked unlocked against the sequence number for a match.
856 * Hence we can allow log forces to run racily and not issue pushes for the
857 * same sequence twice. If we get a race between multiple pushes for the same
858 * sequence they will block on the first one and then abort, hence avoiding
863 struct work_struct *work)
865 struct xfs_cil_ctx *ctx =
866 container_of(work, struct xfs_cil_ctx, push_work);
867 struct xfs_cil *cil = ctx->cil;
868 struct xlog *log = cil->xc_log;
869 struct xfs_log_vec *lv;
870 struct xfs_cil_ctx *new_ctx;
871 struct xlog_ticket *tic;
874 struct xfs_trans_header thdr;
875 struct xfs_log_iovec lhdr;
876 struct xfs_log_vec lvhdr = { NULL };
878 bool push_commit_stable;
880 new_ctx = xlog_cil_ctx_alloc();
881 new_ctx->ticket = xlog_cil_ticket_alloc(log);
883 down_write(&cil->xc_ctx_lock);
885 spin_lock(&cil->xc_push_lock);
886 push_seq = cil->xc_push_seq;
887 ASSERT(push_seq <= ctx->sequence);
888 push_commit_stable = cil->xc_push_commit_stable;
889 cil->xc_push_commit_stable = false;
892 * As we are about to switch to a new, empty CIL context, we no longer
893 * need to throttle tasks on CIL space overruns. Wake any waiters that
894 * the hard push throttle may have caught so they can start committing
895 * to the new context. The ctx->xc_push_lock provides the serialisation
896 * necessary for safely using the lockless waitqueue_active() check in
899 if (waitqueue_active(&cil->xc_push_wait))
900 wake_up_all(&cil->xc_push_wait);
903 * Check if we've anything to push. If there is nothing, then we don't
904 * move on to a new sequence number and so we have to be able to push
905 * this sequence again later.
907 if (list_empty(&cil->xc_cil)) {
908 cil->xc_push_seq = 0;
909 spin_unlock(&cil->xc_push_lock);
914 /* check for a previously pushed sequence */
915 if (push_seq < ctx->sequence) {
916 spin_unlock(&cil->xc_push_lock);
921 * We are now going to push this context, so add it to the committing
922 * list before we do anything else. This ensures that anyone waiting on
923 * this push can easily detect the difference between a "push in
924 * progress" and "CIL is empty, nothing to do".
926 * IOWs, a wait loop can now check for:
927 * the current sequence not being found on the committing list;
929 * an unchanged sequence number
930 * to detect a push that had nothing to do and therefore does not need
931 * waiting on. If the CIL is not empty, we get put on the committing
932 * list before emptying the CIL and bumping the sequence number. Hence
933 * an empty CIL and an unchanged sequence number means we jumped out
934 * above after doing nothing.
936 * Hence the waiter will either find the commit sequence on the
937 * committing list or the sequence number will be unchanged and the CIL
938 * still dirty. In that latter case, the push has not yet started, and
939 * so the waiter will have to continue trying to check the CIL
940 * committing list until it is found. In extreme cases of delay, the
941 * sequence may fully commit between the attempts the wait makes to wait
942 * on the commit sequence.
944 list_add(&ctx->committing, &cil->xc_committing);
945 spin_unlock(&cil->xc_push_lock);
948 * Pull all the log vectors off the items in the CIL, and remove the
949 * items from the CIL. We don't need the CIL lock here because it's only
950 * needed on the transaction commit side which is currently locked out
955 while (!list_empty(&cil->xc_cil)) {
956 struct xfs_log_item *item;
958 item = list_first_entry(&cil->xc_cil,
959 struct xfs_log_item, li_cil);
960 list_del_init(&item->li_cil);
962 ctx->lv_chain = item->li_lv;
964 lv->lv_next = item->li_lv;
967 num_iovecs += lv->lv_niovecs;
971 * Switch the contexts so we can drop the context lock and move out
972 * of a shared context. We can't just go straight to the commit record,
973 * though - we need to synchronise with previous and future commits so
974 * that the commit records are correctly ordered in the log to ensure
975 * that we process items during log IO completion in the correct order.
977 * For example, if we get an EFI in one checkpoint and the EFD in the
978 * next (e.g. due to log forces), we do not want the checkpoint with
979 * the EFD to be committed before the checkpoint with the EFI. Hence
980 * we must strictly order the commit records of the checkpoints so
981 * that: a) the checkpoint callbacks are attached to the iclogs in the
982 * correct order; and b) the checkpoints are replayed in correct order
985 * Hence we need to add this context to the committing context list so
986 * that higher sequences will wait for us to write out a commit record
989 * xfs_log_force_seq requires us to mirror the new sequence into the cil
990 * structure atomically with the addition of this sequence to the
991 * committing list. This also ensures that we can do unlocked checks
992 * against the current sequence in log forces without risking
993 * deferencing a freed context pointer.
995 spin_lock(&cil->xc_push_lock);
996 xlog_cil_ctx_switch(cil, new_ctx);
997 spin_unlock(&cil->xc_push_lock);
998 up_write(&cil->xc_ctx_lock);
1001 * Build a checkpoint transaction header and write it to the log to
1002 * begin the transaction. We need to account for the space used by the
1003 * transaction header here as it is not accounted for in xlog_write().
1005 * The LSN we need to pass to the log items on transaction commit is
1006 * the LSN reported by the first log vector write. If we use the commit
1007 * record lsn then we can move the tail beyond the grant write head.
1010 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
1011 thdr.th_type = XFS_TRANS_CHECKPOINT;
1012 thdr.th_tid = tic->t_tid;
1013 thdr.th_num_items = num_iovecs;
1014 lhdr.i_addr = &thdr;
1015 lhdr.i_len = sizeof(xfs_trans_header_t);
1016 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
1017 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
1019 lvhdr.lv_niovecs = 1;
1020 lvhdr.lv_iovecp = &lhdr;
1021 lvhdr.lv_next = ctx->lv_chain;
1023 error = xlog_cil_write_chain(ctx, &lvhdr);
1025 goto out_abort_free_ticket;
1027 error = xlog_cil_write_commit_record(ctx);
1029 goto out_abort_free_ticket;
1031 xfs_log_ticket_ungrant(log, tic);
1034 * If the checkpoint spans multiple iclogs, wait for all previous iclogs
1035 * to complete before we submit the commit_iclog. We can't use state
1036 * checks for this - ACTIVE can be either a past completed iclog or a
1037 * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a
1038 * past or future iclog awaiting IO or ordered IO completion to be run.
1039 * In the latter case, if it's a future iclog and we wait on it, the we
1040 * will hang because it won't get processed through to ic_force_wait
1041 * wakeup until this commit_iclog is written to disk. Hence we use the
1042 * iclog header lsn and compare it to the commit lsn to determine if we
1043 * need to wait on iclogs or not.
1045 spin_lock(&log->l_icloglock);
1046 if (ctx->start_lsn != ctx->commit_lsn) {
1049 plsn = be64_to_cpu(ctx->commit_iclog->ic_prev->ic_header.h_lsn);
1050 if (plsn && XFS_LSN_CMP(plsn, ctx->commit_lsn) < 0) {
1052 * Waiting on ic_force_wait orders the completion of
1053 * iclogs older than ic_prev. Hence we only need to wait
1054 * on the most recent older iclog here.
1056 xlog_wait_on_iclog(ctx->commit_iclog->ic_prev);
1057 spin_lock(&log->l_icloglock);
1061 * We need to issue a pre-flush so that the ordering for this
1062 * checkpoint is correctly preserved down to stable storage.
1064 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
1068 * The commit iclog must be written to stable storage to guarantee
1069 * journal IO vs metadata writeback IO is correctly ordered on stable
1072 * If the push caller needs the commit to be immediately stable and the
1073 * commit_iclog is not yet marked as XLOG_STATE_WANT_SYNC to indicate it
1074 * will be written when released, switch it's state to WANT_SYNC right
1077 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA;
1078 if (push_commit_stable &&
1079 ctx->commit_iclog->ic_state == XLOG_STATE_ACTIVE)
1080 xlog_state_switch_iclogs(log, ctx->commit_iclog, 0);
1081 xlog_state_release_iclog(log, ctx->commit_iclog);
1083 /* Not safe to reference ctx now! */
1085 spin_unlock(&log->l_icloglock);
1089 up_write(&cil->xc_ctx_lock);
1090 xfs_log_ticket_put(new_ctx->ticket);
1094 out_abort_free_ticket:
1095 xfs_log_ticket_ungrant(log, tic);
1096 ASSERT(xlog_is_shutdown(log));
1097 if (!ctx->commit_iclog) {
1098 xlog_cil_committed(ctx);
1101 spin_lock(&log->l_icloglock);
1102 xlog_state_release_iclog(log, ctx->commit_iclog);
1103 /* Not safe to reference ctx now! */
1104 spin_unlock(&log->l_icloglock);
1108 * We need to push CIL every so often so we don't cache more than we can fit in
1109 * the log. The limit really is that a checkpoint can't be more than half the
1110 * log (the current checkpoint is not allowed to overwrite the previous
1111 * checkpoint), but commit latency and memory usage limit this to a smaller
1115 xlog_cil_push_background(
1116 struct xlog *log) __releases(cil->xc_ctx_lock)
1118 struct xfs_cil *cil = log->l_cilp;
1121 * The cil won't be empty because we are called while holding the
1122 * context lock so whatever we added to the CIL will still be there
1124 ASSERT(!list_empty(&cil->xc_cil));
1127 * Don't do a background push if we haven't used up all the
1128 * space available yet.
1130 if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log)) {
1131 up_read(&cil->xc_ctx_lock);
1135 spin_lock(&cil->xc_push_lock);
1136 if (cil->xc_push_seq < cil->xc_current_sequence) {
1137 cil->xc_push_seq = cil->xc_current_sequence;
1138 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);
1142 * Drop the context lock now, we can't hold that if we need to sleep
1143 * because we are over the blocking threshold. The push_lock is still
1144 * held, so blocking threshold sleep/wakeup is still correctly
1147 up_read(&cil->xc_ctx_lock);
1150 * If we are well over the space limit, throttle the work that is being
1151 * done until the push work on this context has begun. Enforce the hard
1152 * throttle on all transaction commits once it has been activated, even
1153 * if the committing transactions have resulted in the space usage
1154 * dipping back down under the hard limit.
1156 * The ctx->xc_push_lock provides the serialisation necessary for safely
1157 * using the lockless waitqueue_active() check in this context.
1159 if (cil->xc_ctx->space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log) ||
1160 waitqueue_active(&cil->xc_push_wait)) {
1161 trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket);
1162 ASSERT(cil->xc_ctx->space_used < log->l_logsize);
1163 xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock);
1167 spin_unlock(&cil->xc_push_lock);
1172 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
1173 * number that is passed. When it returns, the work will be queued for
1174 * @push_seq, but it won't be completed.
1176 * If the caller is performing a synchronous force, we will flush the workqueue
1177 * to get previously queued work moving to minimise the wait time they will
1178 * undergo waiting for all outstanding pushes to complete. The caller is
1179 * expected to do the required waiting for push_seq to complete.
1181 * If the caller is performing an async push, we need to ensure that the
1182 * checkpoint is fully flushed out of the iclogs when we finish the push. If we
1183 * don't do this, then the commit record may remain sitting in memory in an
1184 * ACTIVE iclog. This then requires another full log force to push to disk,
1185 * which defeats the purpose of having an async, non-blocking CIL force
1186 * mechanism. Hence in this case we need to pass a flag to the push work to
1187 * indicate it needs to flush the commit record itself.
1195 struct xfs_cil *cil = log->l_cilp;
1200 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
1202 /* start on any pending background push to minimise wait time on it */
1204 flush_workqueue(cil->xc_push_wq);
1207 * If the CIL is empty or we've already pushed the sequence then
1208 * there's no work we need to do.
1210 spin_lock(&cil->xc_push_lock);
1211 if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
1212 spin_unlock(&cil->xc_push_lock);
1216 cil->xc_push_seq = push_seq;
1217 cil->xc_push_commit_stable = async;
1218 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);
1219 spin_unlock(&cil->xc_push_lock);
1226 struct xfs_cil *cil = log->l_cilp;
1229 spin_lock(&cil->xc_push_lock);
1230 if (list_empty(&cil->xc_cil))
1232 spin_unlock(&cil->xc_push_lock);
1237 * Commit a transaction with the given vector to the Committed Item List.
1239 * To do this, we need to format the item, pin it in memory if required and
1240 * account for the space used by the transaction. Once we have done that we
1241 * need to release the unused reservation for the transaction, attach the
1242 * transaction to the checkpoint context so we carry the busy extents through
1243 * to checkpoint completion, and then unlock all the items in the transaction.
1245 * Called with the context lock already held in read mode to lock out
1246 * background commit, returns without it held once background commits are
1252 struct xfs_trans *tp,
1253 xfs_csn_t *commit_seq,
1256 struct xfs_cil *cil = log->l_cilp;
1257 struct xfs_log_item *lip, *next;
1260 * Do all necessary memory allocation before we lock the CIL.
1261 * This ensures the allocation does not deadlock with a CIL
1262 * push in memory reclaim (e.g. from kswapd).
1264 xlog_cil_alloc_shadow_bufs(log, tp);
1266 /* lock out background commit */
1267 down_read(&cil->xc_ctx_lock);
1269 xlog_cil_insert_items(log, tp);
1271 if (regrant && !xlog_is_shutdown(log))
1272 xfs_log_ticket_regrant(log, tp->t_ticket);
1274 xfs_log_ticket_ungrant(log, tp->t_ticket);
1275 tp->t_ticket = NULL;
1276 xfs_trans_unreserve_and_mod_sb(tp);
1279 * Once all the items of the transaction have been copied to the CIL,
1280 * the items can be unlocked and possibly freed.
1282 * This needs to be done before we drop the CIL context lock because we
1283 * have to update state in the log items and unlock them before they go
1284 * to disk. If we don't, then the CIL checkpoint can race with us and
1285 * we can run checkpoint completion before we've updated and unlocked
1286 * the log items. This affects (at least) processing of stale buffers,
1289 trace_xfs_trans_commit_items(tp, _RET_IP_);
1290 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
1291 xfs_trans_del_item(lip);
1292 if (lip->li_ops->iop_committing)
1293 lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence);
1296 *commit_seq = cil->xc_ctx->sequence;
1298 /* xlog_cil_push_background() releases cil->xc_ctx_lock */
1299 xlog_cil_push_background(log);
1303 * Flush the CIL to stable storage but don't wait for it to complete. This
1304 * requires the CIL push to ensure the commit record for the push hits the disk,
1305 * but otherwise is no different to a push done from a log force.
1311 xfs_csn_t seq = log->l_cilp->xc_current_sequence;
1313 trace_xfs_log_force(log->l_mp, seq, _RET_IP_);
1314 xlog_cil_push_now(log, seq, true);
1318 * Conditionally push the CIL based on the sequence passed in.
1320 * We only need to push if we haven't already pushed the sequence number given.
1321 * Hence the only time we will trigger a push here is if the push sequence is
1322 * the same as the current context.
1324 * We return the current commit lsn to allow the callers to determine if a
1325 * iclog flush is necessary following this call.
1332 struct xfs_cil *cil = log->l_cilp;
1333 struct xfs_cil_ctx *ctx;
1334 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
1336 ASSERT(sequence <= cil->xc_current_sequence);
1339 sequence = cil->xc_current_sequence;
1340 trace_xfs_log_force(log->l_mp, sequence, _RET_IP_);
1343 * check to see if we need to force out the current context.
1344 * xlog_cil_push() handles racing pushes for the same sequence,
1345 * so no need to deal with it here.
1348 xlog_cil_push_now(log, sequence, false);
1351 * See if we can find a previous sequence still committing.
1352 * We need to wait for all previous sequence commits to complete
1353 * before allowing the force of push_seq to go ahead. Hence block
1354 * on commits for those as well.
1356 spin_lock(&cil->xc_push_lock);
1357 list_for_each_entry(ctx, &cil->xc_committing, committing) {
1359 * Avoid getting stuck in this loop because we were woken by the
1360 * shutdown, but then went back to sleep once already in the
1363 if (xlog_is_shutdown(log))
1365 if (ctx->sequence > sequence)
1367 if (!ctx->commit_lsn) {
1369 * It is still being pushed! Wait for the push to
1370 * complete, then start again from the beginning.
1372 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
1373 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1376 if (ctx->sequence != sequence)
1379 commit_lsn = ctx->commit_lsn;
1383 * The call to xlog_cil_push_now() executes the push in the background.
1384 * Hence by the time we have got here it our sequence may not have been
1385 * pushed yet. This is true if the current sequence still matches the
1386 * push sequence after the above wait loop and the CIL still contains
1387 * dirty objects. This is guaranteed by the push code first adding the
1388 * context to the committing list before emptying the CIL.
1390 * Hence if we don't find the context in the committing list and the
1391 * current sequence number is unchanged then the CIL contents are
1392 * significant. If the CIL is empty, if means there was nothing to push
1393 * and that means there is nothing to wait for. If the CIL is not empty,
1394 * it means we haven't yet started the push, because if it had started
1395 * we would have found the context on the committing list.
1397 if (sequence == cil->xc_current_sequence &&
1398 !list_empty(&cil->xc_cil)) {
1399 spin_unlock(&cil->xc_push_lock);
1403 spin_unlock(&cil->xc_push_lock);
1407 * We detected a shutdown in progress. We need to trigger the log force
1408 * to pass through it's iclog state machine error handling, even though
1409 * we are already in a shutdown state. Hence we can't return
1410 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
1411 * LSN is already stable), so we return a zero LSN instead.
1414 spin_unlock(&cil->xc_push_lock);
1419 * Check if the current log item was first committed in this sequence.
1420 * We can't rely on just the log item being in the CIL, we have to check
1421 * the recorded commit sequence number.
1423 * Note: for this to be used in a non-racy manner, it has to be called with
1424 * CIL flushing locked out. As a result, it should only be used during the
1425 * transaction commit process when deciding what to format into the item.
1428 xfs_log_item_in_current_chkpt(
1429 struct xfs_log_item *lip)
1431 struct xfs_cil *cil = lip->li_mountp->m_log->l_cilp;
1433 if (list_empty(&lip->li_cil))
1437 * li_seq is written on the first commit of a log item to record the
1438 * first checkpoint it is written to. Hence if it is different to the
1439 * current sequence, we're in a new checkpoint.
1441 return lip->li_seq == READ_ONCE(cil->xc_current_sequence);
1445 * Perform initial CIL structure initialisation.
1451 struct xfs_cil *cil;
1452 struct xfs_cil_ctx *ctx;
1454 cil = kmem_zalloc(sizeof(*cil), KM_MAYFAIL);
1458 * Limit the CIL pipeline depth to 4 concurrent works to bound the
1459 * concurrency the log spinlocks will be exposed to.
1461 cil->xc_push_wq = alloc_workqueue("xfs-cil/%s",
1462 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_UNBOUND),
1463 4, log->l_mp->m_super->s_id);
1464 if (!cil->xc_push_wq)
1465 goto out_destroy_cil;
1467 INIT_LIST_HEAD(&cil->xc_cil);
1468 INIT_LIST_HEAD(&cil->xc_committing);
1469 spin_lock_init(&cil->xc_cil_lock);
1470 spin_lock_init(&cil->xc_push_lock);
1471 init_waitqueue_head(&cil->xc_push_wait);
1472 init_rwsem(&cil->xc_ctx_lock);
1473 init_waitqueue_head(&cil->xc_start_wait);
1474 init_waitqueue_head(&cil->xc_commit_wait);
1478 ctx = xlog_cil_ctx_alloc();
1479 xlog_cil_ctx_switch(cil, ctx);
1492 if (log->l_cilp->xc_ctx) {
1493 if (log->l_cilp->xc_ctx->ticket)
1494 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
1495 kmem_free(log->l_cilp->xc_ctx);
1498 ASSERT(list_empty(&log->l_cilp->xc_cil));
1499 destroy_workqueue(log->l_cilp->xc_push_wq);
1500 kmem_free(log->l_cilp);