xfs_log_cil.c revision caa80090
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
4 */
5
6#include "xfs.h"
7#include "xfs_fs.h"
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"
16#include "xfs_log.h"
17#include "xfs_log_priv.h"
18#include "xfs_trace.h"
19
20struct workqueue_struct *xfs_discard_wq;
21
22/*
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.
27 *
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.
33 */
34static struct xlog_ticket *
35xlog_cil_ticket_alloc(
36	struct xlog	*log)
37{
38	struct xlog_ticket *tic;
39
40	tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0);
41
42	/*
43	 * set the current reservation to zero so we know to steal the basic
44	 * transaction overhead reservation from the first transaction commit.
45	 */
46	tic->t_curr_res = 0;
47	return tic;
48}
49
50/*
51 * After the first stage of log recovery is done, we know where the head and
52 * tail of the log are. We need this log initialisation done before we can
53 * initialise the first CIL checkpoint context.
54 *
55 * Here we allocate a log ticket to track space usage during a CIL push.  This
56 * ticket is passed to xlog_write() directly so that we don't slowly leak log
57 * space by failing to account for space used by log headers and additional
58 * region headers for split regions.
59 */
60void
61xlog_cil_init_post_recovery(
62	struct xlog	*log)
63{
64	log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
65	log->l_cilp->xc_ctx->sequence = 1;
66}
67
68static inline int
69xlog_cil_iovec_space(
70	uint	niovecs)
71{
72	return round_up((sizeof(struct xfs_log_vec) +
73					niovecs * sizeof(struct xfs_log_iovec)),
74			sizeof(uint64_t));
75}
76
77/*
78 * Allocate or pin log vector buffers for CIL insertion.
79 *
80 * The CIL currently uses disposable buffers for copying a snapshot of the
81 * modified items into the log during a push. The biggest problem with this is
82 * the requirement to allocate the disposable buffer during the commit if:
83 *	a) does not exist; or
84 *	b) it is too small
85 *
86 * If we do this allocation within xlog_cil_insert_format_items(), it is done
87 * under the xc_ctx_lock, which means that a CIL push cannot occur during
88 * the memory allocation. This means that we have a potential deadlock situation
89 * under low memory conditions when we have lots of dirty metadata pinned in
90 * the CIL and we need a CIL commit to occur to free memory.
91 *
92 * To avoid this, we need to move the memory allocation outside the
93 * xc_ctx_lock, but because the log vector buffers are disposable, that opens
94 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log
95 * vector buffers between the check and the formatting of the item into the
96 * log vector buffer within the xc_ctx_lock.
97 *
98 * Because the log vector buffer needs to be unchanged during the CIL push
99 * process, we cannot share the buffer between the transaction commit (which
100 * modifies the buffer) and the CIL push context that is writing the changes
101 * into the log. This means skipping preallocation of buffer space is
102 * unreliable, but we most definitely do not want to be allocating and freeing
103 * buffers unnecessarily during commits when overwrites can be done safely.
104 *
105 * The simplest solution to this problem is to allocate a shadow buffer when a
106 * log item is committed for the second time, and then to only use this buffer
107 * if necessary. The buffer can remain attached to the log item until such time
108 * it is needed, and this is the buffer that is reallocated to match the size of
109 * the incoming modification. Then during the formatting of the item we can swap
110 * the active buffer with the new one if we can't reuse the existing buffer. We
111 * don't free the old buffer as it may be reused on the next modification if
112 * it's size is right, otherwise we'll free and reallocate it at that point.
113 *
114 * This function builds a vector for the changes in each log item in the
115 * transaction. It then works out the length of the buffer needed for each log
116 * item, allocates them and attaches the vector to the log item in preparation
117 * for the formatting step which occurs under the xc_ctx_lock.
118 *
119 * While this means the memory footprint goes up, it avoids the repeated
120 * alloc/free pattern that repeated modifications of an item would otherwise
121 * cause, and hence minimises the CPU overhead of such behaviour.
122 */
123static void
124xlog_cil_alloc_shadow_bufs(
125	struct xlog		*log,
126	struct xfs_trans	*tp)
127{
128	struct xfs_log_item	*lip;
129
130	list_for_each_entry(lip, &tp->t_items, li_trans) {
131		struct xfs_log_vec *lv;
132		int	niovecs = 0;
133		int	nbytes = 0;
134		int	buf_size;
135		bool	ordered = false;
136
137		/* Skip items which aren't dirty in this transaction. */
138		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
139			continue;
140
141		/* get number of vecs and size of data to be stored */
142		lip->li_ops->iop_size(lip, &niovecs, &nbytes);
143
144		/*
145		 * Ordered items need to be tracked but we do not wish to write
146		 * them. We need a logvec to track the object, but we do not
147		 * need an iovec or buffer to be allocated for copying data.
148		 */
149		if (niovecs == XFS_LOG_VEC_ORDERED) {
150			ordered = true;
151			niovecs = 0;
152			nbytes = 0;
153		}
154
155		/*
156		 * We 64-bit align the length of each iovec so that the start
157		 * of the next one is naturally aligned.  We'll need to
158		 * account for that slack space here. Then round nbytes up
159		 * to 64-bit alignment so that the initial buffer alignment is
160		 * easy to calculate and verify.
161		 */
162		nbytes += niovecs * sizeof(uint64_t);
163		nbytes = round_up(nbytes, sizeof(uint64_t));
164
165		/*
166		 * The data buffer needs to start 64-bit aligned, so round up
167		 * that space to ensure we can align it appropriately and not
168		 * overrun the buffer.
169		 */
170		buf_size = nbytes + xlog_cil_iovec_space(niovecs);
171
172		/*
173		 * if we have no shadow buffer, or it is too small, we need to
174		 * reallocate it.
175		 */
176		if (!lip->li_lv_shadow ||
177		    buf_size > lip->li_lv_shadow->lv_size) {
178
179			/*
180			 * We free and allocate here as a realloc would copy
181			 * unnecessary data. We don't use kmem_zalloc() for the
182			 * same reason - we don't need to zero the data area in
183			 * the buffer, only the log vector header and the iovec
184			 * storage.
185			 */
186			kmem_free(lip->li_lv_shadow);
187
188			/*
189			 * We are in transaction context, which means this
190			 * allocation will pick up GFP_NOFS from the
191			 * memalloc_nofs_save/restore context the transaction
192			 * holds. This means we can use GFP_KERNEL here so the
193			 * generic kvmalloc() code will run vmalloc on
194			 * contiguous page allocation failure as we require.
195			 */
196			lv = kvmalloc(buf_size, GFP_KERNEL);
197			memset(lv, 0, xlog_cil_iovec_space(niovecs));
198
199			lv->lv_item = lip;
200			lv->lv_size = buf_size;
201			if (ordered)
202				lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
203			else
204				lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
205			lip->li_lv_shadow = lv;
206		} else {
207			/* same or smaller, optimise common overwrite case */
208			lv = lip->li_lv_shadow;
209			if (ordered)
210				lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
211			else
212				lv->lv_buf_len = 0;
213			lv->lv_bytes = 0;
214			lv->lv_next = NULL;
215		}
216
217		/* Ensure the lv is set up according to ->iop_size */
218		lv->lv_niovecs = niovecs;
219
220		/* The allocated data region lies beyond the iovec region */
221		lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);
222	}
223
224}
225
226/*
227 * Prepare the log item for insertion into the CIL. Calculate the difference in
228 * log space and vectors it will consume, and if it is a new item pin it as
229 * well.
230 */
231STATIC void
232xfs_cil_prepare_item(
233	struct xlog		*log,
234	struct xfs_log_vec	*lv,
235	struct xfs_log_vec	*old_lv,
236	int			*diff_len,
237	int			*diff_iovecs)
238{
239	/* Account for the new LV being passed in */
240	if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
241		*diff_len += lv->lv_bytes;
242		*diff_iovecs += lv->lv_niovecs;
243	}
244
245	/*
246	 * If there is no old LV, this is the first time we've seen the item in
247	 * this CIL context and so we need to pin it. If we are replacing the
248	 * old_lv, then remove the space it accounts for and make it the shadow
249	 * buffer for later freeing. In both cases we are now switching to the
250	 * shadow buffer, so update the pointer to it appropriately.
251	 */
252	if (!old_lv) {
253		if (lv->lv_item->li_ops->iop_pin)
254			lv->lv_item->li_ops->iop_pin(lv->lv_item);
255		lv->lv_item->li_lv_shadow = NULL;
256	} else if (old_lv != lv) {
257		ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
258
259		*diff_len -= old_lv->lv_bytes;
260		*diff_iovecs -= old_lv->lv_niovecs;
261		lv->lv_item->li_lv_shadow = old_lv;
262	}
263
264	/* attach new log vector to log item */
265	lv->lv_item->li_lv = lv;
266
267	/*
268	 * If this is the first time the item is being committed to the
269	 * CIL, store the sequence number on the log item so we can
270	 * tell in future commits whether this is the first checkpoint
271	 * the item is being committed into.
272	 */
273	if (!lv->lv_item->li_seq)
274		lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
275}
276
277/*
278 * Format log item into a flat buffers
279 *
280 * For delayed logging, we need to hold a formatted buffer containing all the
281 * changes on the log item. This enables us to relog the item in memory and
282 * write it out asynchronously without needing to relock the object that was
283 * modified at the time it gets written into the iclog.
284 *
285 * This function takes the prepared log vectors attached to each log item, and
286 * formats the changes into the log vector buffer. The buffer it uses is
287 * dependent on the current state of the vector in the CIL - the shadow lv is
288 * guaranteed to be large enough for the current modification, but we will only
289 * use that if we can't reuse the existing lv. If we can't reuse the existing
290 * lv, then simple swap it out for the shadow lv. We don't free it - that is
291 * done lazily either by th enext modification or the freeing of the log item.
292 *
293 * We don't set up region headers during this process; we simply copy the
294 * regions into the flat buffer. We can do this because we still have to do a
295 * formatting step to write the regions into the iclog buffer.  Writing the
296 * ophdrs during the iclog write means that we can support splitting large
297 * regions across iclog boundares without needing a change in the format of the
298 * item/region encapsulation.
299 *
300 * Hence what we need to do now is change the rewrite the vector array to point
301 * to the copied region inside the buffer we just allocated. This allows us to
302 * format the regions into the iclog as though they are being formatted
303 * directly out of the objects themselves.
304 */
305static void
306xlog_cil_insert_format_items(
307	struct xlog		*log,
308	struct xfs_trans	*tp,
309	int			*diff_len,
310	int			*diff_iovecs)
311{
312	struct xfs_log_item	*lip;
313
314
315	/* Bail out if we didn't find a log item.  */
316	if (list_empty(&tp->t_items)) {
317		ASSERT(0);
318		return;
319	}
320
321	list_for_each_entry(lip, &tp->t_items, li_trans) {
322		struct xfs_log_vec *lv;
323		struct xfs_log_vec *old_lv = NULL;
324		struct xfs_log_vec *shadow;
325		bool	ordered = false;
326
327		/* Skip items which aren't dirty in this transaction. */
328		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
329			continue;
330
331		/*
332		 * The formatting size information is already attached to
333		 * the shadow lv on the log item.
334		 */
335		shadow = lip->li_lv_shadow;
336		if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)
337			ordered = true;
338
339		/* Skip items that do not have any vectors for writing */
340		if (!shadow->lv_niovecs && !ordered)
341			continue;
342
343		/* compare to existing item size */
344		old_lv = lip->li_lv;
345		if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {
346			/* same or smaller, optimise common overwrite case */
347			lv = lip->li_lv;
348			lv->lv_next = NULL;
349
350			if (ordered)
351				goto insert;
352
353			/*
354			 * set the item up as though it is a new insertion so
355			 * that the space reservation accounting is correct.
356			 */
357			*diff_iovecs -= lv->lv_niovecs;
358			*diff_len -= lv->lv_bytes;
359
360			/* Ensure the lv is set up according to ->iop_size */
361			lv->lv_niovecs = shadow->lv_niovecs;
362
363			/* reset the lv buffer information for new formatting */
364			lv->lv_buf_len = 0;
365			lv->lv_bytes = 0;
366			lv->lv_buf = (char *)lv +
367					xlog_cil_iovec_space(lv->lv_niovecs);
368		} else {
369			/* switch to shadow buffer! */
370			lv = shadow;
371			lv->lv_item = lip;
372			if (ordered) {
373				/* track as an ordered logvec */
374				ASSERT(lip->li_lv == NULL);
375				goto insert;
376			}
377		}
378
379		ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
380		lip->li_ops->iop_format(lip, lv);
381insert:
382		xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
383	}
384}
385
386/*
387 * Insert the log items into the CIL and calculate the difference in space
388 * consumed by the item. Add the space to the checkpoint ticket and calculate
389 * if the change requires additional log metadata. If it does, take that space
390 * as well. Remove the amount of space we added to the checkpoint ticket from
391 * the current transaction ticket so that the accounting works out correctly.
392 */
393static void
394xlog_cil_insert_items(
395	struct xlog		*log,
396	struct xfs_trans	*tp)
397{
398	struct xfs_cil		*cil = log->l_cilp;
399	struct xfs_cil_ctx	*ctx = cil->xc_ctx;
400	struct xfs_log_item	*lip;
401	int			len = 0;
402	int			diff_iovecs = 0;
403	int			iclog_space;
404	int			iovhdr_res = 0, split_res = 0, ctx_res = 0;
405
406	ASSERT(tp);
407
408	/*
409	 * We can do this safely because the context can't checkpoint until we
410	 * are done so it doesn't matter exactly how we update the CIL.
411	 */
412	xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
413
414	spin_lock(&cil->xc_cil_lock);
415
416	/* account for space used by new iovec headers  */
417	iovhdr_res = diff_iovecs * sizeof(xlog_op_header_t);
418	len += iovhdr_res;
419	ctx->nvecs += diff_iovecs;
420
421	/* attach the transaction to the CIL if it has any busy extents */
422	if (!list_empty(&tp->t_busy))
423		list_splice_init(&tp->t_busy, &ctx->busy_extents);
424
425	/*
426	 * Now transfer enough transaction reservation to the context ticket
427	 * for the checkpoint. The context ticket is special - the unit
428	 * reservation has to grow as well as the current reservation as we
429	 * steal from tickets so we can correctly determine the space used
430	 * during the transaction commit.
431	 */
432	if (ctx->ticket->t_curr_res == 0) {
433		ctx_res = ctx->ticket->t_unit_res;
434		ctx->ticket->t_curr_res = ctx_res;
435		tp->t_ticket->t_curr_res -= ctx_res;
436	}
437
438	/* do we need space for more log record headers? */
439	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
440	if (len > 0 && (ctx->space_used / iclog_space !=
441				(ctx->space_used + len) / iclog_space)) {
442		split_res = (len + iclog_space - 1) / iclog_space;
443		/* need to take into account split region headers, too */
444		split_res *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
445		ctx->ticket->t_unit_res += split_res;
446		ctx->ticket->t_curr_res += split_res;
447		tp->t_ticket->t_curr_res -= split_res;
448		ASSERT(tp->t_ticket->t_curr_res >= len);
449	}
450	tp->t_ticket->t_curr_res -= len;
451	ctx->space_used += len;
452
453	/*
454	 * If we've overrun the reservation, dump the tx details before we move
455	 * the log items. Shutdown is imminent...
456	 */
457	if (WARN_ON(tp->t_ticket->t_curr_res < 0)) {
458		xfs_warn(log->l_mp, "Transaction log reservation overrun:");
459		xfs_warn(log->l_mp,
460			 "  log items: %d bytes (iov hdrs: %d bytes)",
461			 len, iovhdr_res);
462		xfs_warn(log->l_mp, "  split region headers: %d bytes",
463			 split_res);
464		xfs_warn(log->l_mp, "  ctx ticket: %d bytes", ctx_res);
465		xlog_print_trans(tp);
466	}
467
468	/*
469	 * Now (re-)position everything modified at the tail of the CIL.
470	 * We do this here so we only need to take the CIL lock once during
471	 * the transaction commit.
472	 */
473	list_for_each_entry(lip, &tp->t_items, li_trans) {
474
475		/* Skip items which aren't dirty in this transaction. */
476		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
477			continue;
478
479		/*
480		 * Only move the item if it isn't already at the tail. This is
481		 * to prevent a transient list_empty() state when reinserting
482		 * an item that is already the only item in the CIL.
483		 */
484		if (!list_is_last(&lip->li_cil, &cil->xc_cil))
485			list_move_tail(&lip->li_cil, &cil->xc_cil);
486	}
487
488	spin_unlock(&cil->xc_cil_lock);
489
490	if (tp->t_ticket->t_curr_res < 0)
491		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
492}
493
494static void
495xlog_cil_free_logvec(
496	struct xfs_log_vec	*log_vector)
497{
498	struct xfs_log_vec	*lv;
499
500	for (lv = log_vector; lv; ) {
501		struct xfs_log_vec *next = lv->lv_next;
502		kmem_free(lv);
503		lv = next;
504	}
505}
506
507static void
508xlog_discard_endio_work(
509	struct work_struct	*work)
510{
511	struct xfs_cil_ctx	*ctx =
512		container_of(work, struct xfs_cil_ctx, discard_endio_work);
513	struct xfs_mount	*mp = ctx->cil->xc_log->l_mp;
514
515	xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
516	kmem_free(ctx);
517}
518
519/*
520 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
521 * pagb_lock.  Note that we need a unbounded workqueue, otherwise we might
522 * get the execution delayed up to 30 seconds for weird reasons.
523 */
524static void
525xlog_discard_endio(
526	struct bio		*bio)
527{
528	struct xfs_cil_ctx	*ctx = bio->bi_private;
529
530	INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work);
531	queue_work(xfs_discard_wq, &ctx->discard_endio_work);
532	bio_put(bio);
533}
534
535static void
536xlog_discard_busy_extents(
537	struct xfs_mount	*mp,
538	struct xfs_cil_ctx	*ctx)
539{
540	struct list_head	*list = &ctx->busy_extents;
541	struct xfs_extent_busy	*busyp;
542	struct bio		*bio = NULL;
543	struct blk_plug		plug;
544	int			error = 0;
545
546	ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
547
548	blk_start_plug(&plug);
549	list_for_each_entry(busyp, list, list) {
550		trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
551					 busyp->length);
552
553		error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
554				XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
555				XFS_FSB_TO_BB(mp, busyp->length),
556				GFP_NOFS, 0, &bio);
557		if (error && error != -EOPNOTSUPP) {
558			xfs_info(mp,
559	 "discard failed for extent [0x%llx,%u], error %d",
560				 (unsigned long long)busyp->bno,
561				 busyp->length,
562				 error);
563			break;
564		}
565	}
566
567	if (bio) {
568		bio->bi_private = ctx;
569		bio->bi_end_io = xlog_discard_endio;
570		submit_bio(bio);
571	} else {
572		xlog_discard_endio_work(&ctx->discard_endio_work);
573	}
574	blk_finish_plug(&plug);
575}
576
577/*
578 * Mark all items committed and clear busy extents. We free the log vector
579 * chains in a separate pass so that we unpin the log items as quickly as
580 * possible.
581 */
582static void
583xlog_cil_committed(
584	struct xfs_cil_ctx	*ctx)
585{
586	struct xfs_mount	*mp = ctx->cil->xc_log->l_mp;
587	bool			abort = xlog_is_shutdown(ctx->cil->xc_log);
588
589	/*
590	 * If the I/O failed, we're aborting the commit and already shutdown.
591	 * Wake any commit waiters before aborting the log items so we don't
592	 * block async log pushers on callbacks. Async log pushers explicitly do
593	 * not wait on log force completion because they may be holding locks
594	 * required to unpin items.
595	 */
596	if (abort) {
597		spin_lock(&ctx->cil->xc_push_lock);
598		wake_up_all(&ctx->cil->xc_commit_wait);
599		spin_unlock(&ctx->cil->xc_push_lock);
600	}
601
602	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
603					ctx->start_lsn, abort);
604
605	xfs_extent_busy_sort(&ctx->busy_extents);
606	xfs_extent_busy_clear(mp, &ctx->busy_extents,
607			     (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
608
609	spin_lock(&ctx->cil->xc_push_lock);
610	list_del(&ctx->committing);
611	spin_unlock(&ctx->cil->xc_push_lock);
612
613	xlog_cil_free_logvec(ctx->lv_chain);
614
615	if (!list_empty(&ctx->busy_extents))
616		xlog_discard_busy_extents(mp, ctx);
617	else
618		kmem_free(ctx);
619}
620
621void
622xlog_cil_process_committed(
623	struct list_head	*list)
624{
625	struct xfs_cil_ctx	*ctx;
626
627	while ((ctx = list_first_entry_or_null(list,
628			struct xfs_cil_ctx, iclog_entry))) {
629		list_del(&ctx->iclog_entry);
630		xlog_cil_committed(ctx);
631	}
632}
633
634/*
635* Record the LSN of the iclog we were just granted space to start writing into.
636* If the context doesn't have a start_lsn recorded, then this iclog will
637* contain the start record for the checkpoint. Otherwise this write contains
638* the commit record for the checkpoint.
639*/
640void
641xlog_cil_set_ctx_write_state(
642	struct xfs_cil_ctx	*ctx,
643	struct xlog_in_core	*iclog)
644{
645	struct xfs_cil		*cil = ctx->cil;
646	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
647
648	ASSERT(!ctx->commit_lsn);
649	if (!ctx->start_lsn) {
650		spin_lock(&cil->xc_push_lock);
651		ctx->start_lsn = lsn;
652		spin_unlock(&cil->xc_push_lock);
653		return;
654	}
655
656	/*
657	 * Take a reference to the iclog for the context so that we still hold
658	 * it when xlog_write is done and has released it. This means the
659	 * context controls when the iclog is released for IO.
660	 */
661	atomic_inc(&iclog->ic_refcnt);
662
663	/*
664	 * xlog_state_get_iclog_space() guarantees there is enough space in the
665	 * iclog for an entire commit record, so we can attach the context
666	 * callbacks now.  This needs to be done before we make the commit_lsn
667	 * visible to waiters so that checkpoints with commit records in the
668	 * same iclog order their IO completion callbacks in the same order that
669	 * the commit records appear in the iclog.
670	 */
671	spin_lock(&cil->xc_log->l_icloglock);
672	list_add_tail(&ctx->iclog_entry, &iclog->ic_callbacks);
673	spin_unlock(&cil->xc_log->l_icloglock);
674
675	/*
676	 * Now we can record the commit LSN and wake anyone waiting for this
677	 * sequence to have the ordered commit record assigned to a physical
678	 * location in the log.
679	 */
680	spin_lock(&cil->xc_push_lock);
681	ctx->commit_iclog = iclog;
682	ctx->commit_lsn = lsn;
683	wake_up_all(&cil->xc_commit_wait);
684	spin_unlock(&cil->xc_push_lock);
685}
686
687
688/*
689 * Ensure that the order of log writes follows checkpoint sequence order. This
690 * relies on the context LSN being zero until the log write has guaranteed the
691 * LSN that the log write will start at via xlog_state_get_iclog_space().
692 */
693static int
694xlog_cil_order_write(
695	struct xfs_cil		*cil,
696	xfs_csn_t		sequence)
697{
698	struct xfs_cil_ctx	*ctx;
699
700restart:
701	spin_lock(&cil->xc_push_lock);
702	list_for_each_entry(ctx, &cil->xc_committing, committing) {
703		/*
704		 * Avoid getting stuck in this loop because we were woken by the
705		 * shutdown, but then went back to sleep once already in the
706		 * shutdown state.
707		 */
708		if (xlog_is_shutdown(cil->xc_log)) {
709			spin_unlock(&cil->xc_push_lock);
710			return -EIO;
711		}
712
713		/*
714		 * Higher sequences will wait for this one so skip them.
715		 * Don't wait for our own sequence, either.
716		 */
717		if (ctx->sequence >= sequence)
718			continue;
719		if (!ctx->commit_lsn) {
720			/*
721			 * It is still being pushed! Wait for the push to
722			 * complete, then start again from the beginning.
723			 */
724			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
725			goto restart;
726		}
727	}
728	spin_unlock(&cil->xc_push_lock);
729	return 0;
730}
731
732/*
733 * Write out the commit record of a checkpoint transaction to close off a
734 * running log write. These commit records are strictly ordered in ascending CIL
735 * sequence order so that log recovery will always replay the checkpoints in the
736 * correct order.
737 */
738static int
739xlog_cil_write_commit_record(
740	struct xfs_cil_ctx	*ctx)
741{
742	struct xlog		*log = ctx->cil->xc_log;
743	struct xfs_log_iovec	reg = {
744		.i_addr = NULL,
745		.i_len = 0,
746		.i_type = XLOG_REG_TYPE_COMMIT,
747	};
748	struct xfs_log_vec	vec = {
749		.lv_niovecs = 1,
750		.lv_iovecp = &reg,
751	};
752	int			error;
753
754	if (xlog_is_shutdown(log))
755		return -EIO;
756
757	error = xlog_write(log, ctx, &vec, ctx->ticket, XLOG_COMMIT_TRANS);
758	if (error)
759		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
760	return error;
761}
762
763/*
764 * Push the Committed Item List to the log.
765 *
766 * If the current sequence is the same as xc_push_seq we need to do a flush. If
767 * xc_push_seq is less than the current sequence, then it has already been
768 * flushed and we don't need to do anything - the caller will wait for it to
769 * complete if necessary.
770 *
771 * xc_push_seq is checked unlocked against the sequence number for a match.
772 * Hence we can allow log forces to run racily and not issue pushes for the
773 * same sequence twice.  If we get a race between multiple pushes for the same
774 * sequence they will block on the first one and then abort, hence avoiding
775 * needless pushes.
776 */
777static void
778xlog_cil_push_work(
779	struct work_struct	*work)
780{
781	struct xfs_cil		*cil =
782		container_of(work, struct xfs_cil, xc_push_work);
783	struct xlog		*log = cil->xc_log;
784	struct xfs_log_vec	*lv;
785	struct xfs_cil_ctx	*ctx;
786	struct xfs_cil_ctx	*new_ctx;
787	struct xlog_ticket	*tic;
788	int			num_iovecs;
789	int			error = 0;
790	struct xfs_trans_header thdr;
791	struct xfs_log_iovec	lhdr;
792	struct xfs_log_vec	lvhdr = { NULL };
793	xfs_lsn_t		preflush_tail_lsn;
794	xfs_csn_t		push_seq;
795	struct bio		bio;
796	DECLARE_COMPLETION_ONSTACK(bdev_flush);
797
798	new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_NOFS);
799	new_ctx->ticket = xlog_cil_ticket_alloc(log);
800
801	down_write(&cil->xc_ctx_lock);
802	ctx = cil->xc_ctx;
803
804	spin_lock(&cil->xc_push_lock);
805	push_seq = cil->xc_push_seq;
806	ASSERT(push_seq <= ctx->sequence);
807
808	/*
809	 * As we are about to switch to a new, empty CIL context, we no longer
810	 * need to throttle tasks on CIL space overruns. Wake any waiters that
811	 * the hard push throttle may have caught so they can start committing
812	 * to the new context. The ctx->xc_push_lock provides the serialisation
813	 * necessary for safely using the lockless waitqueue_active() check in
814	 * this context.
815	 */
816	if (waitqueue_active(&cil->xc_push_wait))
817		wake_up_all(&cil->xc_push_wait);
818
819	/*
820	 * Check if we've anything to push. If there is nothing, then we don't
821	 * move on to a new sequence number and so we have to be able to push
822	 * this sequence again later.
823	 */
824	if (list_empty(&cil->xc_cil)) {
825		cil->xc_push_seq = 0;
826		spin_unlock(&cil->xc_push_lock);
827		goto out_skip;
828	}
829
830
831	/* check for a previously pushed sequence */
832	if (push_seq < cil->xc_ctx->sequence) {
833		spin_unlock(&cil->xc_push_lock);
834		goto out_skip;
835	}
836
837	/*
838	 * We are now going to push this context, so add it to the committing
839	 * list before we do anything else. This ensures that anyone waiting on
840	 * this push can easily detect the difference between a "push in
841	 * progress" and "CIL is empty, nothing to do".
842	 *
843	 * IOWs, a wait loop can now check for:
844	 *	the current sequence not being found on the committing list;
845	 *	an empty CIL; and
846	 *	an unchanged sequence number
847	 * to detect a push that had nothing to do and therefore does not need
848	 * waiting on. If the CIL is not empty, we get put on the committing
849	 * list before emptying the CIL and bumping the sequence number. Hence
850	 * an empty CIL and an unchanged sequence number means we jumped out
851	 * above after doing nothing.
852	 *
853	 * Hence the waiter will either find the commit sequence on the
854	 * committing list or the sequence number will be unchanged and the CIL
855	 * still dirty. In that latter case, the push has not yet started, and
856	 * so the waiter will have to continue trying to check the CIL
857	 * committing list until it is found. In extreme cases of delay, the
858	 * sequence may fully commit between the attempts the wait makes to wait
859	 * on the commit sequence.
860	 */
861	list_add(&ctx->committing, &cil->xc_committing);
862	spin_unlock(&cil->xc_push_lock);
863
864	/*
865	 * The CIL is stable at this point - nothing new will be added to it
866	 * because we hold the flush lock exclusively. Hence we can now issue
867	 * a cache flush to ensure all the completed metadata in the journal we
868	 * are about to overwrite is on stable storage.
869	 *
870	 * Because we are issuing this cache flush before we've written the
871	 * tail lsn to the iclog, we can have metadata IO completions move the
872	 * tail forwards between the completion of this flush and the iclog
873	 * being written. In this case, we need to re-issue the cache flush
874	 * before the iclog write. To detect whether the log tail moves, sample
875	 * the tail LSN *before* we issue the flush.
876	 */
877	preflush_tail_lsn = atomic64_read(&log->l_tail_lsn);
878	xfs_flush_bdev_async(&bio, log->l_mp->m_ddev_targp->bt_bdev,
879				&bdev_flush);
880
881	/*
882	 * Pull all the log vectors off the items in the CIL, and remove the
883	 * items from the CIL. We don't need the CIL lock here because it's only
884	 * needed on the transaction commit side which is currently locked out
885	 * by the flush lock.
886	 */
887	lv = NULL;
888	num_iovecs = 0;
889	while (!list_empty(&cil->xc_cil)) {
890		struct xfs_log_item	*item;
891
892		item = list_first_entry(&cil->xc_cil,
893					struct xfs_log_item, li_cil);
894		list_del_init(&item->li_cil);
895		if (!ctx->lv_chain)
896			ctx->lv_chain = item->li_lv;
897		else
898			lv->lv_next = item->li_lv;
899		lv = item->li_lv;
900		item->li_lv = NULL;
901		num_iovecs += lv->lv_niovecs;
902	}
903
904	/*
905	 * initialise the new context and attach it to the CIL. Then attach
906	 * the current context to the CIL committing list so it can be found
907	 * during log forces to extract the commit lsn of the sequence that
908	 * needs to be forced.
909	 */
910	INIT_LIST_HEAD(&new_ctx->committing);
911	INIT_LIST_HEAD(&new_ctx->busy_extents);
912	new_ctx->sequence = ctx->sequence + 1;
913	new_ctx->cil = cil;
914	cil->xc_ctx = new_ctx;
915
916	/*
917	 * The switch is now done, so we can drop the context lock and move out
918	 * of a shared context. We can't just go straight to the commit record,
919	 * though - we need to synchronise with previous and future commits so
920	 * that the commit records are correctly ordered in the log to ensure
921	 * that we process items during log IO completion in the correct order.
922	 *
923	 * For example, if we get an EFI in one checkpoint and the EFD in the
924	 * next (e.g. due to log forces), we do not want the checkpoint with
925	 * the EFD to be committed before the checkpoint with the EFI.  Hence
926	 * we must strictly order the commit records of the checkpoints so
927	 * that: a) the checkpoint callbacks are attached to the iclogs in the
928	 * correct order; and b) the checkpoints are replayed in correct order
929	 * in log recovery.
930	 *
931	 * Hence we need to add this context to the committing context list so
932	 * that higher sequences will wait for us to write out a commit record
933	 * before they do.
934	 *
935	 * xfs_log_force_seq requires us to mirror the new sequence into the cil
936	 * structure atomically with the addition of this sequence to the
937	 * committing list. This also ensures that we can do unlocked checks
938	 * against the current sequence in log forces without risking
939	 * deferencing a freed context pointer.
940	 */
941	spin_lock(&cil->xc_push_lock);
942	cil->xc_current_sequence = new_ctx->sequence;
943	spin_unlock(&cil->xc_push_lock);
944	up_write(&cil->xc_ctx_lock);
945
946	/*
947	 * Build a checkpoint transaction header and write it to the log to
948	 * begin the transaction. We need to account for the space used by the
949	 * transaction header here as it is not accounted for in xlog_write().
950	 *
951	 * The LSN we need to pass to the log items on transaction commit is
952	 * the LSN reported by the first log vector write. If we use the commit
953	 * record lsn then we can move the tail beyond the grant write head.
954	 */
955	tic = ctx->ticket;
956	thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
957	thdr.th_type = XFS_TRANS_CHECKPOINT;
958	thdr.th_tid = tic->t_tid;
959	thdr.th_num_items = num_iovecs;
960	lhdr.i_addr = &thdr;
961	lhdr.i_len = sizeof(xfs_trans_header_t);
962	lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
963	tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
964
965	lvhdr.lv_niovecs = 1;
966	lvhdr.lv_iovecp = &lhdr;
967	lvhdr.lv_next = ctx->lv_chain;
968
969	/*
970	 * Before we format and submit the first iclog, we have to ensure that
971	 * the metadata writeback ordering cache flush is complete.
972	 */
973	wait_for_completion(&bdev_flush);
974
975	error = xlog_write(log, ctx, &lvhdr, tic, XLOG_START_TRANS);
976	if (error)
977		goto out_abort_free_ticket;
978
979	error = xlog_cil_order_write(ctx->cil, ctx->sequence);
980	if (error)
981		goto out_abort_free_ticket;
982
983	error = xlog_cil_write_commit_record(ctx);
984	if (error)
985		goto out_abort_free_ticket;
986
987	xfs_log_ticket_ungrant(log, tic);
988
989	/*
990	 * If the checkpoint spans multiple iclogs, wait for all previous iclogs
991	 * to complete before we submit the commit_iclog. We can't use state
992	 * checks for this - ACTIVE can be either a past completed iclog or a
993	 * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a
994	 * past or future iclog awaiting IO or ordered IO completion to be run.
995	 * In the latter case, if it's a future iclog and we wait on it, the we
996	 * will hang because it won't get processed through to ic_force_wait
997	 * wakeup until this commit_iclog is written to disk.  Hence we use the
998	 * iclog header lsn and compare it to the commit lsn to determine if we
999	 * need to wait on iclogs or not.
1000	 */
1001	spin_lock(&log->l_icloglock);
1002	if (ctx->start_lsn != ctx->commit_lsn) {
1003		xfs_lsn_t	plsn;
1004
1005		plsn = be64_to_cpu(ctx->commit_iclog->ic_prev->ic_header.h_lsn);
1006		if (plsn && XFS_LSN_CMP(plsn, ctx->commit_lsn) < 0) {
1007			/*
1008			 * Waiting on ic_force_wait orders the completion of
1009			 * iclogs older than ic_prev. Hence we only need to wait
1010			 * on the most recent older iclog here.
1011			 */
1012			xlog_wait_on_iclog(ctx->commit_iclog->ic_prev);
1013			spin_lock(&log->l_icloglock);
1014		}
1015
1016		/*
1017		 * We need to issue a pre-flush so that the ordering for this
1018		 * checkpoint is correctly preserved down to stable storage.
1019		 */
1020		ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
1021	}
1022
1023	/*
1024	 * The commit iclog must be written to stable storage to guarantee
1025	 * journal IO vs metadata writeback IO is correctly ordered on stable
1026	 * storage.
1027	 */
1028	ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA;
1029	xlog_state_release_iclog(log, ctx->commit_iclog, preflush_tail_lsn);
1030
1031	/* Not safe to reference ctx now! */
1032
1033	spin_unlock(&log->l_icloglock);
1034	return;
1035
1036out_skip:
1037	up_write(&cil->xc_ctx_lock);
1038	xfs_log_ticket_put(new_ctx->ticket);
1039	kmem_free(new_ctx);
1040	return;
1041
1042out_abort_free_ticket:
1043	xfs_log_ticket_ungrant(log, tic);
1044	ASSERT(xlog_is_shutdown(log));
1045	if (!ctx->commit_iclog) {
1046		xlog_cil_committed(ctx);
1047		return;
1048	}
1049	spin_lock(&log->l_icloglock);
1050	xlog_state_release_iclog(log, ctx->commit_iclog, 0);
1051	/* Not safe to reference ctx now! */
1052	spin_unlock(&log->l_icloglock);
1053}
1054
1055/*
1056 * We need to push CIL every so often so we don't cache more than we can fit in
1057 * the log. The limit really is that a checkpoint can't be more than half the
1058 * log (the current checkpoint is not allowed to overwrite the previous
1059 * checkpoint), but commit latency and memory usage limit this to a smaller
1060 * size.
1061 */
1062static void
1063xlog_cil_push_background(
1064	struct xlog	*log) __releases(cil->xc_ctx_lock)
1065{
1066	struct xfs_cil	*cil = log->l_cilp;
1067
1068	/*
1069	 * The cil won't be empty because we are called while holding the
1070	 * context lock so whatever we added to the CIL will still be there
1071	 */
1072	ASSERT(!list_empty(&cil->xc_cil));
1073
1074	/*
1075	 * Don't do a background push if we haven't used up all the
1076	 * space available yet.
1077	 */
1078	if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log)) {
1079		up_read(&cil->xc_ctx_lock);
1080		return;
1081	}
1082
1083	spin_lock(&cil->xc_push_lock);
1084	if (cil->xc_push_seq < cil->xc_current_sequence) {
1085		cil->xc_push_seq = cil->xc_current_sequence;
1086		queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
1087	}
1088
1089	/*
1090	 * Drop the context lock now, we can't hold that if we need to sleep
1091	 * because we are over the blocking threshold. The push_lock is still
1092	 * held, so blocking threshold sleep/wakeup is still correctly
1093	 * serialised here.
1094	 */
1095	up_read(&cil->xc_ctx_lock);
1096
1097	/*
1098	 * If we are well over the space limit, throttle the work that is being
1099	 * done until the push work on this context has begun. Enforce the hard
1100	 * throttle on all transaction commits once it has been activated, even
1101	 * if the committing transactions have resulted in the space usage
1102	 * dipping back down under the hard limit.
1103	 *
1104	 * The ctx->xc_push_lock provides the serialisation necessary for safely
1105	 * using the lockless waitqueue_active() check in this context.
1106	 */
1107	if (cil->xc_ctx->space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log) ||
1108	    waitqueue_active(&cil->xc_push_wait)) {
1109		trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket);
1110		ASSERT(cil->xc_ctx->space_used < log->l_logsize);
1111		xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock);
1112		return;
1113	}
1114
1115	spin_unlock(&cil->xc_push_lock);
1116
1117}
1118
1119/*
1120 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
1121 * number that is passed. When it returns, the work will be queued for
1122 * @push_seq, but it won't be completed. The caller is expected to do any
1123 * waiting for push_seq to complete if it is required.
1124 */
1125static void
1126xlog_cil_push_now(
1127	struct xlog	*log,
1128	xfs_lsn_t	push_seq)
1129{
1130	struct xfs_cil	*cil = log->l_cilp;
1131
1132	if (!cil)
1133		return;
1134
1135	ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
1136
1137	/* start on any pending background push to minimise wait time on it */
1138	flush_work(&cil->xc_push_work);
1139
1140	/*
1141	 * If the CIL is empty or we've already pushed the sequence then
1142	 * there's no work we need to do.
1143	 */
1144	spin_lock(&cil->xc_push_lock);
1145	if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
1146		spin_unlock(&cil->xc_push_lock);
1147		return;
1148	}
1149
1150	cil->xc_push_seq = push_seq;
1151	queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
1152	spin_unlock(&cil->xc_push_lock);
1153}
1154
1155bool
1156xlog_cil_empty(
1157	struct xlog	*log)
1158{
1159	struct xfs_cil	*cil = log->l_cilp;
1160	bool		empty = false;
1161
1162	spin_lock(&cil->xc_push_lock);
1163	if (list_empty(&cil->xc_cil))
1164		empty = true;
1165	spin_unlock(&cil->xc_push_lock);
1166	return empty;
1167}
1168
1169/*
1170 * Commit a transaction with the given vector to the Committed Item List.
1171 *
1172 * To do this, we need to format the item, pin it in memory if required and
1173 * account for the space used by the transaction. Once we have done that we
1174 * need to release the unused reservation for the transaction, attach the
1175 * transaction to the checkpoint context so we carry the busy extents through
1176 * to checkpoint completion, and then unlock all the items in the transaction.
1177 *
1178 * Called with the context lock already held in read mode to lock out
1179 * background commit, returns without it held once background commits are
1180 * allowed again.
1181 */
1182void
1183xlog_cil_commit(
1184	struct xlog		*log,
1185	struct xfs_trans	*tp,
1186	xfs_csn_t		*commit_seq,
1187	bool			regrant)
1188{
1189	struct xfs_cil		*cil = log->l_cilp;
1190	struct xfs_log_item	*lip, *next;
1191
1192	/*
1193	 * Do all necessary memory allocation before we lock the CIL.
1194	 * This ensures the allocation does not deadlock with a CIL
1195	 * push in memory reclaim (e.g. from kswapd).
1196	 */
1197	xlog_cil_alloc_shadow_bufs(log, tp);
1198
1199	/* lock out background commit */
1200	down_read(&cil->xc_ctx_lock);
1201
1202	xlog_cil_insert_items(log, tp);
1203
1204	if (regrant && !xlog_is_shutdown(log))
1205		xfs_log_ticket_regrant(log, tp->t_ticket);
1206	else
1207		xfs_log_ticket_ungrant(log, tp->t_ticket);
1208	tp->t_ticket = NULL;
1209	xfs_trans_unreserve_and_mod_sb(tp);
1210
1211	/*
1212	 * Once all the items of the transaction have been copied to the CIL,
1213	 * the items can be unlocked and possibly freed.
1214	 *
1215	 * This needs to be done before we drop the CIL context lock because we
1216	 * have to update state in the log items and unlock them before they go
1217	 * to disk. If we don't, then the CIL checkpoint can race with us and
1218	 * we can run checkpoint completion before we've updated and unlocked
1219	 * the log items. This affects (at least) processing of stale buffers,
1220	 * inodes and EFIs.
1221	 */
1222	trace_xfs_trans_commit_items(tp, _RET_IP_);
1223	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
1224		xfs_trans_del_item(lip);
1225		if (lip->li_ops->iop_committing)
1226			lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence);
1227	}
1228	if (commit_seq)
1229		*commit_seq = cil->xc_ctx->sequence;
1230
1231	/* xlog_cil_push_background() releases cil->xc_ctx_lock */
1232	xlog_cil_push_background(log);
1233}
1234
1235/*
1236 * Conditionally push the CIL based on the sequence passed in.
1237 *
1238 * We only need to push if we haven't already pushed the sequence
1239 * number given. Hence the only time we will trigger a push here is
1240 * if the push sequence is the same as the current context.
1241 *
1242 * We return the current commit lsn to allow the callers to determine if a
1243 * iclog flush is necessary following this call.
1244 */
1245xfs_lsn_t
1246xlog_cil_force_seq(
1247	struct xlog	*log,
1248	xfs_csn_t	sequence)
1249{
1250	struct xfs_cil		*cil = log->l_cilp;
1251	struct xfs_cil_ctx	*ctx;
1252	xfs_lsn_t		commit_lsn = NULLCOMMITLSN;
1253
1254	ASSERT(sequence <= cil->xc_current_sequence);
1255
1256	/*
1257	 * check to see if we need to force out the current context.
1258	 * xlog_cil_push() handles racing pushes for the same sequence,
1259	 * so no need to deal with it here.
1260	 */
1261restart:
1262	xlog_cil_push_now(log, sequence);
1263
1264	/*
1265	 * See if we can find a previous sequence still committing.
1266	 * We need to wait for all previous sequence commits to complete
1267	 * before allowing the force of push_seq to go ahead. Hence block
1268	 * on commits for those as well.
1269	 */
1270	spin_lock(&cil->xc_push_lock);
1271	list_for_each_entry(ctx, &cil->xc_committing, committing) {
1272		/*
1273		 * Avoid getting stuck in this loop because we were woken by the
1274		 * shutdown, but then went back to sleep once already in the
1275		 * shutdown state.
1276		 */
1277		if (xlog_is_shutdown(log))
1278			goto out_shutdown;
1279		if (ctx->sequence > sequence)
1280			continue;
1281		if (!ctx->commit_lsn) {
1282			/*
1283			 * It is still being pushed! Wait for the push to
1284			 * complete, then start again from the beginning.
1285			 */
1286			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1287			goto restart;
1288		}
1289		if (ctx->sequence != sequence)
1290			continue;
1291		/* found it! */
1292		commit_lsn = ctx->commit_lsn;
1293	}
1294
1295	/*
1296	 * The call to xlog_cil_push_now() executes the push in the background.
1297	 * Hence by the time we have got here it our sequence may not have been
1298	 * pushed yet. This is true if the current sequence still matches the
1299	 * push sequence after the above wait loop and the CIL still contains
1300	 * dirty objects. This is guaranteed by the push code first adding the
1301	 * context to the committing list before emptying the CIL.
1302	 *
1303	 * Hence if we don't find the context in the committing list and the
1304	 * current sequence number is unchanged then the CIL contents are
1305	 * significant.  If the CIL is empty, if means there was nothing to push
1306	 * and that means there is nothing to wait for. If the CIL is not empty,
1307	 * it means we haven't yet started the push, because if it had started
1308	 * we would have found the context on the committing list.
1309	 */
1310	if (sequence == cil->xc_current_sequence &&
1311	    !list_empty(&cil->xc_cil)) {
1312		spin_unlock(&cil->xc_push_lock);
1313		goto restart;
1314	}
1315
1316	spin_unlock(&cil->xc_push_lock);
1317	return commit_lsn;
1318
1319	/*
1320	 * We detected a shutdown in progress. We need to trigger the log force
1321	 * to pass through it's iclog state machine error handling, even though
1322	 * we are already in a shutdown state. Hence we can't return
1323	 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
1324	 * LSN is already stable), so we return a zero LSN instead.
1325	 */
1326out_shutdown:
1327	spin_unlock(&cil->xc_push_lock);
1328	return 0;
1329}
1330
1331/*
1332 * Check if the current log item was first committed in this sequence.
1333 * We can't rely on just the log item being in the CIL, we have to check
1334 * the recorded commit sequence number.
1335 *
1336 * Note: for this to be used in a non-racy manner, it has to be called with
1337 * CIL flushing locked out. As a result, it should only be used during the
1338 * transaction commit process when deciding what to format into the item.
1339 */
1340bool
1341xfs_log_item_in_current_chkpt(
1342	struct xfs_log_item *lip)
1343{
1344	struct xfs_cil_ctx *ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
1345
1346	if (list_empty(&lip->li_cil))
1347		return false;
1348
1349	/*
1350	 * li_seq is written on the first commit of a log item to record the
1351	 * first checkpoint it is written to. Hence if it is different to the
1352	 * current sequence, we're in a new checkpoint.
1353	 */
1354	return lip->li_seq == ctx->sequence;
1355}
1356
1357/*
1358 * Perform initial CIL structure initialisation.
1359 */
1360int
1361xlog_cil_init(
1362	struct xlog	*log)
1363{
1364	struct xfs_cil	*cil;
1365	struct xfs_cil_ctx *ctx;
1366
1367	cil = kmem_zalloc(sizeof(*cil), KM_MAYFAIL);
1368	if (!cil)
1369		return -ENOMEM;
1370
1371	ctx = kmem_zalloc(sizeof(*ctx), KM_MAYFAIL);
1372	if (!ctx) {
1373		kmem_free(cil);
1374		return -ENOMEM;
1375	}
1376
1377	INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
1378	INIT_LIST_HEAD(&cil->xc_cil);
1379	INIT_LIST_HEAD(&cil->xc_committing);
1380	spin_lock_init(&cil->xc_cil_lock);
1381	spin_lock_init(&cil->xc_push_lock);
1382	init_waitqueue_head(&cil->xc_push_wait);
1383	init_rwsem(&cil->xc_ctx_lock);
1384	init_waitqueue_head(&cil->xc_commit_wait);
1385
1386	INIT_LIST_HEAD(&ctx->committing);
1387	INIT_LIST_HEAD(&ctx->busy_extents);
1388	ctx->sequence = 1;
1389	ctx->cil = cil;
1390	cil->xc_ctx = ctx;
1391	cil->xc_current_sequence = ctx->sequence;
1392
1393	cil->xc_log = log;
1394	log->l_cilp = cil;
1395	return 0;
1396}
1397
1398void
1399xlog_cil_destroy(
1400	struct xlog	*log)
1401{
1402	if (log->l_cilp->xc_ctx) {
1403		if (log->l_cilp->xc_ctx->ticket)
1404			xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
1405		kmem_free(log->l_cilp->xc_ctx);
1406	}
1407
1408	ASSERT(list_empty(&log->l_cilp->xc_cil));
1409	kmem_free(log->l_cilp);
1410}
1411
1412