xref: /openbmc/linux/fs/xfs/xfs_log_cil.c (revision 8a10bc9d)
1 /*
2  * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
11  * GNU General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public License
14  * along with this program; if not, write the Free Software Foundation,
15  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
16  */
17 
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_log_format.h"
21 #include "xfs_shared.h"
22 #include "xfs_trans_resv.h"
23 #include "xfs_sb.h"
24 #include "xfs_ag.h"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_alloc.h"
28 #include "xfs_extent_busy.h"
29 #include "xfs_discard.h"
30 #include "xfs_trans.h"
31 #include "xfs_trans_priv.h"
32 #include "xfs_log.h"
33 #include "xfs_log_priv.h"
34 
35 /*
36  * Allocate a new ticket. Failing to get a new ticket makes it really hard to
37  * recover, so we don't allow failure here. Also, we allocate in a context that
38  * we don't want to be issuing transactions from, so we need to tell the
39  * allocation code this as well.
40  *
41  * We don't reserve any space for the ticket - we are going to steal whatever
42  * space we require from transactions as they commit. To ensure we reserve all
43  * the space required, we need to set the current reservation of the ticket to
44  * zero so that we know to steal the initial transaction overhead from the
45  * first transaction commit.
46  */
47 static struct xlog_ticket *
48 xlog_cil_ticket_alloc(
49 	struct xlog	*log)
50 {
51 	struct xlog_ticket *tic;
52 
53 	tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
54 				KM_SLEEP|KM_NOFS);
55 	tic->t_trans_type = XFS_TRANS_CHECKPOINT;
56 
57 	/*
58 	 * set the current reservation to zero so we know to steal the basic
59 	 * transaction overhead reservation from the first transaction commit.
60 	 */
61 	tic->t_curr_res = 0;
62 	return tic;
63 }
64 
65 /*
66  * After the first stage of log recovery is done, we know where the head and
67  * tail of the log are. We need this log initialisation done before we can
68  * initialise the first CIL checkpoint context.
69  *
70  * Here we allocate a log ticket to track space usage during a CIL push.  This
71  * ticket is passed to xlog_write() directly so that we don't slowly leak log
72  * space by failing to account for space used by log headers and additional
73  * region headers for split regions.
74  */
75 void
76 xlog_cil_init_post_recovery(
77 	struct xlog	*log)
78 {
79 	log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
80 	log->l_cilp->xc_ctx->sequence = 1;
81 	log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
82 								log->l_curr_block);
83 }
84 
85 /*
86  * Prepare the log item for insertion into the CIL. Calculate the difference in
87  * log space and vectors it will consume, and if it is a new item pin it as
88  * well.
89  */
90 STATIC void
91 xfs_cil_prepare_item(
92 	struct xlog		*log,
93 	struct xfs_log_vec	*lv,
94 	struct xfs_log_vec	*old_lv,
95 	int			*diff_len,
96 	int			*diff_iovecs)
97 {
98 	/* Account for the new LV being passed in */
99 	if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
100 		*diff_len += lv->lv_buf_len;
101 		*diff_iovecs += lv->lv_niovecs;
102 	}
103 
104 	/*
105 	 * If there is no old LV, this is the first time we've seen the item in
106 	 * this CIL context and so we need to pin it. If we are replacing the
107 	 * old_lv, then remove the space it accounts for and free it.
108 	 */
109 	if (!old_lv)
110 		lv->lv_item->li_ops->iop_pin(lv->lv_item);
111 	else if (old_lv != lv) {
112 		ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
113 
114 		*diff_len -= old_lv->lv_buf_len;
115 		*diff_iovecs -= old_lv->lv_niovecs;
116 		kmem_free(old_lv);
117 	}
118 
119 	/* attach new log vector to log item */
120 	lv->lv_item->li_lv = lv;
121 
122 	/*
123 	 * If this is the first time the item is being committed to the
124 	 * CIL, store the sequence number on the log item so we can
125 	 * tell in future commits whether this is the first checkpoint
126 	 * the item is being committed into.
127 	 */
128 	if (!lv->lv_item->li_seq)
129 		lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
130 }
131 
132 /*
133  * Format log item into a flat buffers
134  *
135  * For delayed logging, we need to hold a formatted buffer containing all the
136  * changes on the log item. This enables us to relog the item in memory and
137  * write it out asynchronously without needing to relock the object that was
138  * modified at the time it gets written into the iclog.
139  *
140  * This function builds a vector for the changes in each log item in the
141  * transaction. It then works out the length of the buffer needed for each log
142  * item, allocates them and formats the vector for the item into the buffer.
143  * The buffer is then attached to the log item are then inserted into the
144  * Committed Item List for tracking until the next checkpoint is written out.
145  *
146  * We don't set up region headers during this process; we simply copy the
147  * regions into the flat buffer. We can do this because we still have to do a
148  * formatting step to write the regions into the iclog buffer.  Writing the
149  * ophdrs during the iclog write means that we can support splitting large
150  * regions across iclog boundares without needing a change in the format of the
151  * item/region encapsulation.
152  *
153  * Hence what we need to do now is change the rewrite the vector array to point
154  * to the copied region inside the buffer we just allocated. This allows us to
155  * format the regions into the iclog as though they are being formatted
156  * directly out of the objects themselves.
157  */
158 static void
159 xlog_cil_insert_format_items(
160 	struct xlog		*log,
161 	struct xfs_trans	*tp,
162 	int			*diff_len,
163 	int			*diff_iovecs)
164 {
165 	struct xfs_log_item_desc *lidp;
166 
167 
168 	/* Bail out if we didn't find a log item.  */
169 	if (list_empty(&tp->t_items)) {
170 		ASSERT(0);
171 		return;
172 	}
173 
174 	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
175 		struct xfs_log_item *lip = lidp->lid_item;
176 		struct xfs_log_vec *lv;
177 		struct xfs_log_vec *old_lv;
178 		int	niovecs = 0;
179 		int	nbytes = 0;
180 		int	buf_size;
181 		bool	ordered = false;
182 
183 		/* Skip items which aren't dirty in this transaction. */
184 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
185 			continue;
186 
187 		/* get number of vecs and size of data to be stored */
188 		lip->li_ops->iop_size(lip, &niovecs, &nbytes);
189 
190 		/* Skip items that do not have any vectors for writing */
191 		if (!niovecs)
192 			continue;
193 
194 		/*
195 		 * Ordered items need to be tracked but we do not wish to write
196 		 * them. We need a logvec to track the object, but we do not
197 		 * need an iovec or buffer to be allocated for copying data.
198 		 */
199 		if (niovecs == XFS_LOG_VEC_ORDERED) {
200 			ordered = true;
201 			niovecs = 0;
202 			nbytes = 0;
203 		}
204 
205 		/*
206 		 * We 64-bit align the length of each iovec so that the start
207 		 * of the next one is naturally aligned.  We'll need to
208 		 * account for that slack space here.
209 		 */
210 		nbytes += niovecs * sizeof(uint64_t);
211 
212 		/* grab the old item if it exists for reservation accounting */
213 		old_lv = lip->li_lv;
214 
215 		/* calc buffer size */
216 		buf_size = sizeof(struct xfs_log_vec) + nbytes +
217 				niovecs * sizeof(struct xfs_log_iovec);
218 
219 		/* compare to existing item size */
220 		if (lip->li_lv && buf_size <= lip->li_lv->lv_size) {
221 			/* same or smaller, optimise common overwrite case */
222 			lv = lip->li_lv;
223 			lv->lv_next = NULL;
224 
225 			if (ordered)
226 				goto insert;
227 
228 			/*
229 			 * set the item up as though it is a new insertion so
230 			 * that the space reservation accounting is correct.
231 			 */
232 			*diff_iovecs -= lv->lv_niovecs;
233 			*diff_len -= lv->lv_buf_len;
234 		} else {
235 			/* allocate new data chunk */
236 			lv = kmem_zalloc(buf_size, KM_SLEEP|KM_NOFS);
237 			lv->lv_item = lip;
238 			lv->lv_size = buf_size;
239 			if (ordered) {
240 				/* track as an ordered logvec */
241 				ASSERT(lip->li_lv == NULL);
242 				lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
243 				goto insert;
244 			}
245 			lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
246 		}
247 
248 		/* Ensure the lv is set up according to ->iop_size */
249 		lv->lv_niovecs = niovecs;
250 
251 		/* The allocated data region lies beyond the iovec region */
252 		lv->lv_buf_len = 0;
253 		lv->lv_buf = (char *)lv + buf_size - nbytes;
254 		lip->li_ops->iop_format(lip, lv);
255 insert:
256 		ASSERT(lv->lv_buf_len <= nbytes);
257 		xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
258 	}
259 }
260 
261 /*
262  * Insert the log items into the CIL and calculate the difference in space
263  * consumed by the item. Add the space to the checkpoint ticket and calculate
264  * if the change requires additional log metadata. If it does, take that space
265  * as well. Remove the amount of space we added to the checkpoint ticket from
266  * the current transaction ticket so that the accounting works out correctly.
267  */
268 static void
269 xlog_cil_insert_items(
270 	struct xlog		*log,
271 	struct xfs_trans	*tp)
272 {
273 	struct xfs_cil		*cil = log->l_cilp;
274 	struct xfs_cil_ctx	*ctx = cil->xc_ctx;
275 	struct xfs_log_item_desc *lidp;
276 	int			len = 0;
277 	int			diff_iovecs = 0;
278 	int			iclog_space;
279 
280 	ASSERT(tp);
281 
282 	/*
283 	 * We can do this safely because the context can't checkpoint until we
284 	 * are done so it doesn't matter exactly how we update the CIL.
285 	 */
286 	xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
287 
288 	/*
289 	 * Now (re-)position everything modified at the tail of the CIL.
290 	 * We do this here so we only need to take the CIL lock once during
291 	 * the transaction commit.
292 	 */
293 	spin_lock(&cil->xc_cil_lock);
294 	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
295 		struct xfs_log_item	*lip = lidp->lid_item;
296 
297 		/* Skip items which aren't dirty in this transaction. */
298 		if (!(lidp->lid_flags & XFS_LID_DIRTY))
299 			continue;
300 
301 		list_move_tail(&lip->li_cil, &cil->xc_cil);
302 	}
303 
304 	/* account for space used by new iovec headers  */
305 	len += diff_iovecs * sizeof(xlog_op_header_t);
306 	ctx->nvecs += diff_iovecs;
307 
308 	/* attach the transaction to the CIL if it has any busy extents */
309 	if (!list_empty(&tp->t_busy))
310 		list_splice_init(&tp->t_busy, &ctx->busy_extents);
311 
312 	/*
313 	 * Now transfer enough transaction reservation to the context ticket
314 	 * for the checkpoint. The context ticket is special - the unit
315 	 * reservation has to grow as well as the current reservation as we
316 	 * steal from tickets so we can correctly determine the space used
317 	 * during the transaction commit.
318 	 */
319 	if (ctx->ticket->t_curr_res == 0) {
320 		ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
321 		tp->t_ticket->t_curr_res -= ctx->ticket->t_unit_res;
322 	}
323 
324 	/* do we need space for more log record headers? */
325 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
326 	if (len > 0 && (ctx->space_used / iclog_space !=
327 				(ctx->space_used + len) / iclog_space)) {
328 		int hdrs;
329 
330 		hdrs = (len + iclog_space - 1) / iclog_space;
331 		/* need to take into account split region headers, too */
332 		hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
333 		ctx->ticket->t_unit_res += hdrs;
334 		ctx->ticket->t_curr_res += hdrs;
335 		tp->t_ticket->t_curr_res -= hdrs;
336 		ASSERT(tp->t_ticket->t_curr_res >= len);
337 	}
338 	tp->t_ticket->t_curr_res -= len;
339 	ctx->space_used += len;
340 
341 	spin_unlock(&cil->xc_cil_lock);
342 }
343 
344 static void
345 xlog_cil_free_logvec(
346 	struct xfs_log_vec	*log_vector)
347 {
348 	struct xfs_log_vec	*lv;
349 
350 	for (lv = log_vector; lv; ) {
351 		struct xfs_log_vec *next = lv->lv_next;
352 		kmem_free(lv);
353 		lv = next;
354 	}
355 }
356 
357 /*
358  * Mark all items committed and clear busy extents. We free the log vector
359  * chains in a separate pass so that we unpin the log items as quickly as
360  * possible.
361  */
362 static void
363 xlog_cil_committed(
364 	void	*args,
365 	int	abort)
366 {
367 	struct xfs_cil_ctx	*ctx = args;
368 	struct xfs_mount	*mp = ctx->cil->xc_log->l_mp;
369 
370 	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
371 					ctx->start_lsn, abort);
372 
373 	xfs_extent_busy_sort(&ctx->busy_extents);
374 	xfs_extent_busy_clear(mp, &ctx->busy_extents,
375 			     (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
376 
377 	spin_lock(&ctx->cil->xc_push_lock);
378 	list_del(&ctx->committing);
379 	spin_unlock(&ctx->cil->xc_push_lock);
380 
381 	xlog_cil_free_logvec(ctx->lv_chain);
382 
383 	if (!list_empty(&ctx->busy_extents)) {
384 		ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
385 
386 		xfs_discard_extents(mp, &ctx->busy_extents);
387 		xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
388 	}
389 
390 	kmem_free(ctx);
391 }
392 
393 /*
394  * Push the Committed Item List to the log. If @push_seq flag is zero, then it
395  * is a background flush and so we can chose to ignore it. Otherwise, if the
396  * current sequence is the same as @push_seq we need to do a flush. If
397  * @push_seq is less than the current sequence, then it has already been
398  * flushed and we don't need to do anything - the caller will wait for it to
399  * complete if necessary.
400  *
401  * @push_seq is a value rather than a flag because that allows us to do an
402  * unlocked check of the sequence number for a match. Hence we can allows log
403  * forces to run racily and not issue pushes for the same sequence twice. If we
404  * get a race between multiple pushes for the same sequence they will block on
405  * the first one and then abort, hence avoiding needless pushes.
406  */
407 STATIC int
408 xlog_cil_push(
409 	struct xlog		*log)
410 {
411 	struct xfs_cil		*cil = log->l_cilp;
412 	struct xfs_log_vec	*lv;
413 	struct xfs_cil_ctx	*ctx;
414 	struct xfs_cil_ctx	*new_ctx;
415 	struct xlog_in_core	*commit_iclog;
416 	struct xlog_ticket	*tic;
417 	int			num_iovecs;
418 	int			error = 0;
419 	struct xfs_trans_header thdr;
420 	struct xfs_log_iovec	lhdr;
421 	struct xfs_log_vec	lvhdr = { NULL };
422 	xfs_lsn_t		commit_lsn;
423 	xfs_lsn_t		push_seq;
424 
425 	if (!cil)
426 		return 0;
427 
428 	new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
429 	new_ctx->ticket = xlog_cil_ticket_alloc(log);
430 
431 	down_write(&cil->xc_ctx_lock);
432 	ctx = cil->xc_ctx;
433 
434 	spin_lock(&cil->xc_push_lock);
435 	push_seq = cil->xc_push_seq;
436 	ASSERT(push_seq <= ctx->sequence);
437 
438 	/*
439 	 * Check if we've anything to push. If there is nothing, then we don't
440 	 * move on to a new sequence number and so we have to be able to push
441 	 * this sequence again later.
442 	 */
443 	if (list_empty(&cil->xc_cil)) {
444 		cil->xc_push_seq = 0;
445 		spin_unlock(&cil->xc_push_lock);
446 		goto out_skip;
447 	}
448 	spin_unlock(&cil->xc_push_lock);
449 
450 
451 	/* check for a previously pushed seqeunce */
452 	if (push_seq < cil->xc_ctx->sequence)
453 		goto out_skip;
454 
455 	/*
456 	 * pull all the log vectors off the items in the CIL, and
457 	 * remove the items from the CIL. We don't need the CIL lock
458 	 * here because it's only needed on the transaction commit
459 	 * side which is currently locked out by the flush lock.
460 	 */
461 	lv = NULL;
462 	num_iovecs = 0;
463 	while (!list_empty(&cil->xc_cil)) {
464 		struct xfs_log_item	*item;
465 
466 		item = list_first_entry(&cil->xc_cil,
467 					struct xfs_log_item, li_cil);
468 		list_del_init(&item->li_cil);
469 		if (!ctx->lv_chain)
470 			ctx->lv_chain = item->li_lv;
471 		else
472 			lv->lv_next = item->li_lv;
473 		lv = item->li_lv;
474 		item->li_lv = NULL;
475 		num_iovecs += lv->lv_niovecs;
476 	}
477 
478 	/*
479 	 * initialise the new context and attach it to the CIL. Then attach
480 	 * the current context to the CIL committing lsit so it can be found
481 	 * during log forces to extract the commit lsn of the sequence that
482 	 * needs to be forced.
483 	 */
484 	INIT_LIST_HEAD(&new_ctx->committing);
485 	INIT_LIST_HEAD(&new_ctx->busy_extents);
486 	new_ctx->sequence = ctx->sequence + 1;
487 	new_ctx->cil = cil;
488 	cil->xc_ctx = new_ctx;
489 
490 	/*
491 	 * mirror the new sequence into the cil structure so that we can do
492 	 * unlocked checks against the current sequence in log forces without
493 	 * risking deferencing a freed context pointer.
494 	 */
495 	cil->xc_current_sequence = new_ctx->sequence;
496 
497 	/*
498 	 * The switch is now done, so we can drop the context lock and move out
499 	 * of a shared context. We can't just go straight to the commit record,
500 	 * though - we need to synchronise with previous and future commits so
501 	 * that the commit records are correctly ordered in the log to ensure
502 	 * that we process items during log IO completion in the correct order.
503 	 *
504 	 * For example, if we get an EFI in one checkpoint and the EFD in the
505 	 * next (e.g. due to log forces), we do not want the checkpoint with
506 	 * the EFD to be committed before the checkpoint with the EFI.  Hence
507 	 * we must strictly order the commit records of the checkpoints so
508 	 * that: a) the checkpoint callbacks are attached to the iclogs in the
509 	 * correct order; and b) the checkpoints are replayed in correct order
510 	 * in log recovery.
511 	 *
512 	 * Hence we need to add this context to the committing context list so
513 	 * that higher sequences will wait for us to write out a commit record
514 	 * before they do.
515 	 */
516 	spin_lock(&cil->xc_push_lock);
517 	list_add(&ctx->committing, &cil->xc_committing);
518 	spin_unlock(&cil->xc_push_lock);
519 	up_write(&cil->xc_ctx_lock);
520 
521 	/*
522 	 * Build a checkpoint transaction header and write it to the log to
523 	 * begin the transaction. We need to account for the space used by the
524 	 * transaction header here as it is not accounted for in xlog_write().
525 	 *
526 	 * The LSN we need to pass to the log items on transaction commit is
527 	 * the LSN reported by the first log vector write. If we use the commit
528 	 * record lsn then we can move the tail beyond the grant write head.
529 	 */
530 	tic = ctx->ticket;
531 	thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
532 	thdr.th_type = XFS_TRANS_CHECKPOINT;
533 	thdr.th_tid = tic->t_tid;
534 	thdr.th_num_items = num_iovecs;
535 	lhdr.i_addr = &thdr;
536 	lhdr.i_len = sizeof(xfs_trans_header_t);
537 	lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
538 	tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
539 
540 	lvhdr.lv_niovecs = 1;
541 	lvhdr.lv_iovecp = &lhdr;
542 	lvhdr.lv_next = ctx->lv_chain;
543 
544 	error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
545 	if (error)
546 		goto out_abort_free_ticket;
547 
548 	/*
549 	 * now that we've written the checkpoint into the log, strictly
550 	 * order the commit records so replay will get them in the right order.
551 	 */
552 restart:
553 	spin_lock(&cil->xc_push_lock);
554 	list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
555 		/*
556 		 * Higher sequences will wait for this one so skip them.
557 		 * Don't wait for own own sequence, either.
558 		 */
559 		if (new_ctx->sequence >= ctx->sequence)
560 			continue;
561 		if (!new_ctx->commit_lsn) {
562 			/*
563 			 * It is still being pushed! Wait for the push to
564 			 * complete, then start again from the beginning.
565 			 */
566 			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
567 			goto restart;
568 		}
569 	}
570 	spin_unlock(&cil->xc_push_lock);
571 
572 	/* xfs_log_done always frees the ticket on error. */
573 	commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
574 	if (commit_lsn == -1)
575 		goto out_abort;
576 
577 	/* attach all the transactions w/ busy extents to iclog */
578 	ctx->log_cb.cb_func = xlog_cil_committed;
579 	ctx->log_cb.cb_arg = ctx;
580 	error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
581 	if (error)
582 		goto out_abort;
583 
584 	/*
585 	 * now the checkpoint commit is complete and we've attached the
586 	 * callbacks to the iclog we can assign the commit LSN to the context
587 	 * and wake up anyone who is waiting for the commit to complete.
588 	 */
589 	spin_lock(&cil->xc_push_lock);
590 	ctx->commit_lsn = commit_lsn;
591 	wake_up_all(&cil->xc_commit_wait);
592 	spin_unlock(&cil->xc_push_lock);
593 
594 	/* release the hounds! */
595 	return xfs_log_release_iclog(log->l_mp, commit_iclog);
596 
597 out_skip:
598 	up_write(&cil->xc_ctx_lock);
599 	xfs_log_ticket_put(new_ctx->ticket);
600 	kmem_free(new_ctx);
601 	return 0;
602 
603 out_abort_free_ticket:
604 	xfs_log_ticket_put(tic);
605 out_abort:
606 	xlog_cil_committed(ctx, XFS_LI_ABORTED);
607 	return XFS_ERROR(EIO);
608 }
609 
610 static void
611 xlog_cil_push_work(
612 	struct work_struct	*work)
613 {
614 	struct xfs_cil		*cil = container_of(work, struct xfs_cil,
615 							xc_push_work);
616 	xlog_cil_push(cil->xc_log);
617 }
618 
619 /*
620  * We need to push CIL every so often so we don't cache more than we can fit in
621  * the log. The limit really is that a checkpoint can't be more than half the
622  * log (the current checkpoint is not allowed to overwrite the previous
623  * checkpoint), but commit latency and memory usage limit this to a smaller
624  * size.
625  */
626 static void
627 xlog_cil_push_background(
628 	struct xlog	*log)
629 {
630 	struct xfs_cil	*cil = log->l_cilp;
631 
632 	/*
633 	 * The cil won't be empty because we are called while holding the
634 	 * context lock so whatever we added to the CIL will still be there
635 	 */
636 	ASSERT(!list_empty(&cil->xc_cil));
637 
638 	/*
639 	 * don't do a background push if we haven't used up all the
640 	 * space available yet.
641 	 */
642 	if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
643 		return;
644 
645 	spin_lock(&cil->xc_push_lock);
646 	if (cil->xc_push_seq < cil->xc_current_sequence) {
647 		cil->xc_push_seq = cil->xc_current_sequence;
648 		queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
649 	}
650 	spin_unlock(&cil->xc_push_lock);
651 
652 }
653 
654 static void
655 xlog_cil_push_foreground(
656 	struct xlog	*log,
657 	xfs_lsn_t	push_seq)
658 {
659 	struct xfs_cil	*cil = log->l_cilp;
660 
661 	if (!cil)
662 		return;
663 
664 	ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
665 
666 	/* start on any pending background push to minimise wait time on it */
667 	flush_work(&cil->xc_push_work);
668 
669 	/*
670 	 * If the CIL is empty or we've already pushed the sequence then
671 	 * there's no work we need to do.
672 	 */
673 	spin_lock(&cil->xc_push_lock);
674 	if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
675 		spin_unlock(&cil->xc_push_lock);
676 		return;
677 	}
678 
679 	cil->xc_push_seq = push_seq;
680 	spin_unlock(&cil->xc_push_lock);
681 
682 	/* do the push now */
683 	xlog_cil_push(log);
684 }
685 
686 bool
687 xlog_cil_empty(
688 	struct xlog	*log)
689 {
690 	struct xfs_cil	*cil = log->l_cilp;
691 	bool		empty = false;
692 
693 	spin_lock(&cil->xc_push_lock);
694 	if (list_empty(&cil->xc_cil))
695 		empty = true;
696 	spin_unlock(&cil->xc_push_lock);
697 	return empty;
698 }
699 
700 /*
701  * Commit a transaction with the given vector to the Committed Item List.
702  *
703  * To do this, we need to format the item, pin it in memory if required and
704  * account for the space used by the transaction. Once we have done that we
705  * need to release the unused reservation for the transaction, attach the
706  * transaction to the checkpoint context so we carry the busy extents through
707  * to checkpoint completion, and then unlock all the items in the transaction.
708  *
709  * Called with the context lock already held in read mode to lock out
710  * background commit, returns without it held once background commits are
711  * allowed again.
712  */
713 int
714 xfs_log_commit_cil(
715 	struct xfs_mount	*mp,
716 	struct xfs_trans	*tp,
717 	xfs_lsn_t		*commit_lsn,
718 	int			flags)
719 {
720 	struct xlog		*log = mp->m_log;
721 	struct xfs_cil		*cil = log->l_cilp;
722 	int			log_flags = 0;
723 
724 	if (flags & XFS_TRANS_RELEASE_LOG_RES)
725 		log_flags = XFS_LOG_REL_PERM_RESERV;
726 
727 	/* lock out background commit */
728 	down_read(&cil->xc_ctx_lock);
729 
730 	xlog_cil_insert_items(log, tp);
731 
732 	/* check we didn't blow the reservation */
733 	if (tp->t_ticket->t_curr_res < 0)
734 		xlog_print_tic_res(mp, tp->t_ticket);
735 
736 	tp->t_commit_lsn = cil->xc_ctx->sequence;
737 	if (commit_lsn)
738 		*commit_lsn = tp->t_commit_lsn;
739 
740 	xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
741 	xfs_trans_unreserve_and_mod_sb(tp);
742 
743 	/*
744 	 * Once all the items of the transaction have been copied to the CIL,
745 	 * the items can be unlocked and freed.
746 	 *
747 	 * This needs to be done before we drop the CIL context lock because we
748 	 * have to update state in the log items and unlock them before they go
749 	 * to disk. If we don't, then the CIL checkpoint can race with us and
750 	 * we can run checkpoint completion before we've updated and unlocked
751 	 * the log items. This affects (at least) processing of stale buffers,
752 	 * inodes and EFIs.
753 	 */
754 	xfs_trans_free_items(tp, tp->t_commit_lsn, 0);
755 
756 	xlog_cil_push_background(log);
757 
758 	up_read(&cil->xc_ctx_lock);
759 	return 0;
760 }
761 
762 /*
763  * Conditionally push the CIL based on the sequence passed in.
764  *
765  * We only need to push if we haven't already pushed the sequence
766  * number given. Hence the only time we will trigger a push here is
767  * if the push sequence is the same as the current context.
768  *
769  * We return the current commit lsn to allow the callers to determine if a
770  * iclog flush is necessary following this call.
771  */
772 xfs_lsn_t
773 xlog_cil_force_lsn(
774 	struct xlog	*log,
775 	xfs_lsn_t	sequence)
776 {
777 	struct xfs_cil		*cil = log->l_cilp;
778 	struct xfs_cil_ctx	*ctx;
779 	xfs_lsn_t		commit_lsn = NULLCOMMITLSN;
780 
781 	ASSERT(sequence <= cil->xc_current_sequence);
782 
783 	/*
784 	 * check to see if we need to force out the current context.
785 	 * xlog_cil_push() handles racing pushes for the same sequence,
786 	 * so no need to deal with it here.
787 	 */
788 	xlog_cil_push_foreground(log, sequence);
789 
790 	/*
791 	 * See if we can find a previous sequence still committing.
792 	 * We need to wait for all previous sequence commits to complete
793 	 * before allowing the force of push_seq to go ahead. Hence block
794 	 * on commits for those as well.
795 	 */
796 restart:
797 	spin_lock(&cil->xc_push_lock);
798 	list_for_each_entry(ctx, &cil->xc_committing, committing) {
799 		if (ctx->sequence > sequence)
800 			continue;
801 		if (!ctx->commit_lsn) {
802 			/*
803 			 * It is still being pushed! Wait for the push to
804 			 * complete, then start again from the beginning.
805 			 */
806 			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
807 			goto restart;
808 		}
809 		if (ctx->sequence != sequence)
810 			continue;
811 		/* found it! */
812 		commit_lsn = ctx->commit_lsn;
813 	}
814 	spin_unlock(&cil->xc_push_lock);
815 	return commit_lsn;
816 }
817 
818 /*
819  * Check if the current log item was first committed in this sequence.
820  * We can't rely on just the log item being in the CIL, we have to check
821  * the recorded commit sequence number.
822  *
823  * Note: for this to be used in a non-racy manner, it has to be called with
824  * CIL flushing locked out. As a result, it should only be used during the
825  * transaction commit process when deciding what to format into the item.
826  */
827 bool
828 xfs_log_item_in_current_chkpt(
829 	struct xfs_log_item *lip)
830 {
831 	struct xfs_cil_ctx *ctx;
832 
833 	if (list_empty(&lip->li_cil))
834 		return false;
835 
836 	ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
837 
838 	/*
839 	 * li_seq is written on the first commit of a log item to record the
840 	 * first checkpoint it is written to. Hence if it is different to the
841 	 * current sequence, we're in a new checkpoint.
842 	 */
843 	if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
844 		return false;
845 	return true;
846 }
847 
848 /*
849  * Perform initial CIL structure initialisation.
850  */
851 int
852 xlog_cil_init(
853 	struct xlog	*log)
854 {
855 	struct xfs_cil	*cil;
856 	struct xfs_cil_ctx *ctx;
857 
858 	cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
859 	if (!cil)
860 		return ENOMEM;
861 
862 	ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
863 	if (!ctx) {
864 		kmem_free(cil);
865 		return ENOMEM;
866 	}
867 
868 	INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
869 	INIT_LIST_HEAD(&cil->xc_cil);
870 	INIT_LIST_HEAD(&cil->xc_committing);
871 	spin_lock_init(&cil->xc_cil_lock);
872 	spin_lock_init(&cil->xc_push_lock);
873 	init_rwsem(&cil->xc_ctx_lock);
874 	init_waitqueue_head(&cil->xc_commit_wait);
875 
876 	INIT_LIST_HEAD(&ctx->committing);
877 	INIT_LIST_HEAD(&ctx->busy_extents);
878 	ctx->sequence = 1;
879 	ctx->cil = cil;
880 	cil->xc_ctx = ctx;
881 	cil->xc_current_sequence = ctx->sequence;
882 
883 	cil->xc_log = log;
884 	log->l_cilp = cil;
885 	return 0;
886 }
887 
888 void
889 xlog_cil_destroy(
890 	struct xlog	*log)
891 {
892 	if (log->l_cilp->xc_ctx) {
893 		if (log->l_cilp->xc_ctx->ticket)
894 			xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
895 		kmem_free(log->l_cilp->xc_ctx);
896 	}
897 
898 	ASSERT(list_empty(&log->l_cilp->xc_cil));
899 	kmem_free(log->l_cilp);
900 }
901 
902