blob: cfe97973ba36d1d586c3704b536aebce2e391af1 [file] [log] [blame]
Dave Chinner71e330b2010-05-21 14:37:18 +10001/*
2 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it would be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write the Free Software Foundation,
15 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
16 */
17
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
Dave Chinner71e330b2010-05-21 14:37:18 +100021#include "xfs_log.h"
Dave Chinner71e330b2010-05-21 14:37:18 +100022#include "xfs_trans.h"
23#include "xfs_trans_priv.h"
24#include "xfs_log_priv.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
Dave Chinner71e330b2010-05-21 14:37:18 +100027#include "xfs_mount.h"
28#include "xfs_error.h"
29#include "xfs_alloc.h"
Dave Chinnerefc27b52012-04-29 10:39:43 +000030#include "xfs_extent_busy.h"
Christoph Hellwige84661a2011-05-20 13:45:32 +000031#include "xfs_discard.h"
Dave Chinner71e330b2010-05-21 14:37:18 +100032
33/*
Dave Chinner71e330b2010-05-21 14:37:18 +100034 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
35 * recover, so we don't allow failure here. Also, we allocate in a context that
36 * we don't want to be issuing transactions from, so we need to tell the
37 * allocation code this as well.
38 *
39 * We don't reserve any space for the ticket - we are going to steal whatever
40 * space we require from transactions as they commit. To ensure we reserve all
41 * the space required, we need to set the current reservation of the ticket to
42 * zero so that we know to steal the initial transaction overhead from the
43 * first transaction commit.
44 */
45static struct xlog_ticket *
46xlog_cil_ticket_alloc(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -050047 struct xlog *log)
Dave Chinner71e330b2010-05-21 14:37:18 +100048{
49 struct xlog_ticket *tic;
50
51 tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
52 KM_SLEEP|KM_NOFS);
53 tic->t_trans_type = XFS_TRANS_CHECKPOINT;
54
55 /*
56 * set the current reservation to zero so we know to steal the basic
57 * transaction overhead reservation from the first transaction commit.
58 */
59 tic->t_curr_res = 0;
60 return tic;
61}
62
63/*
64 * After the first stage of log recovery is done, we know where the head and
65 * tail of the log are. We need this log initialisation done before we can
66 * initialise the first CIL checkpoint context.
67 *
68 * Here we allocate a log ticket to track space usage during a CIL push. This
69 * ticket is passed to xlog_write() directly so that we don't slowly leak log
70 * space by failing to account for space used by log headers and additional
71 * region headers for split regions.
72 */
73void
74xlog_cil_init_post_recovery(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -050075 struct xlog *log)
Dave Chinner71e330b2010-05-21 14:37:18 +100076{
Dave Chinner71e330b2010-05-21 14:37:18 +100077 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
78 log->l_cilp->xc_ctx->sequence = 1;
79 log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
80 log->l_curr_block);
81}
82
Dave Chinner7492c5b2013-08-12 20:50:05 +100083STATIC int
84xlog_cil_lv_item_format(
85 struct xfs_log_item *lip,
86 struct xfs_log_vec *lv)
87{
88 int index;
89 char *ptr;
90
91 /* format new vectors into array */
92 lip->li_ops->iop_format(lip, lv->lv_iovecp);
93
94 /* copy data into existing array */
95 ptr = lv->lv_buf;
96 for (index = 0; index < lv->lv_niovecs; index++) {
97 struct xfs_log_iovec *vec = &lv->lv_iovecp[index];
98
99 memcpy(ptr, vec->i_addr, vec->i_len);
100 vec->i_addr = ptr;
101 ptr += vec->i_len;
102 }
103
104 /*
105 * some size calculations for log vectors over-estimate, so the caller
106 * doesn't know the amount of space actually used by the item. Return
107 * the byte count to the caller so they can check and store it
108 * appropriately.
109 */
110 return ptr - lv->lv_buf;
111}
112
Dave Chinner71e330b2010-05-21 14:37:18 +1000113/*
Dave Chinner991aaf62013-08-12 20:50:07 +1000114 * Prepare the log item for insertion into the CIL. Calculate the difference in
115 * log space and vectors it will consume, and if it is a new item pin it as
116 * well.
117 */
118STATIC void
119xfs_cil_prepare_item(
120 struct xlog *log,
121 struct xfs_log_vec *lv,
122 struct xfs_log_vec *old_lv,
123 int *diff_len,
124 int *diff_iovecs)
125{
126 /* Account for the new LV being passed in */
127 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
128 *diff_len += lv->lv_buf_len;
129 *diff_iovecs += lv->lv_niovecs;
130 }
131
132 /*
133 * If there is no old LV, this is the first time we've seen the item in
134 * this CIL context and so we need to pin it. If we are replacing the
135 * old_lv, then remove the space it accounts for and free it.
136 */
137 if (!old_lv)
138 lv->lv_item->li_ops->iop_pin(lv->lv_item);
139 else if (old_lv != lv) {
140 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
141
142 *diff_len -= old_lv->lv_buf_len;
143 *diff_iovecs -= old_lv->lv_niovecs;
144 kmem_free(old_lv);
145 }
146
147 /* attach new log vector to log item */
148 lv->lv_item->li_lv = lv;
149
150 /*
151 * If this is the first time the item is being committed to the
152 * CIL, store the sequence number on the log item so we can
153 * tell in future commits whether this is the first checkpoint
154 * the item is being committed into.
155 */
156 if (!lv->lv_item->li_seq)
157 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
158}
159
160/*
Dave Chinner71e330b2010-05-21 14:37:18 +1000161 * Format log item into a flat buffers
162 *
163 * For delayed logging, we need to hold a formatted buffer containing all the
164 * changes on the log item. This enables us to relog the item in memory and
165 * write it out asynchronously without needing to relock the object that was
166 * modified at the time it gets written into the iclog.
167 *
168 * This function builds a vector for the changes in each log item in the
169 * transaction. It then works out the length of the buffer needed for each log
170 * item, allocates them and formats the vector for the item into the buffer.
171 * The buffer is then attached to the log item are then inserted into the
172 * Committed Item List for tracking until the next checkpoint is written out.
173 *
174 * We don't set up region headers during this process; we simply copy the
175 * regions into the flat buffer. We can do this because we still have to do a
176 * formatting step to write the regions into the iclog buffer. Writing the
177 * ophdrs during the iclog write means that we can support splitting large
178 * regions across iclog boundares without needing a change in the format of the
179 * item/region encapsulation.
180 *
181 * Hence what we need to do now is change the rewrite the vector array to point
182 * to the copied region inside the buffer we just allocated. This allows us to
183 * format the regions into the iclog as though they are being formatted
184 * directly out of the objects themselves.
185 */
Dave Chinner991aaf62013-08-12 20:50:07 +1000186static void
187xlog_cil_insert_format_items(
188 struct xlog *log,
189 struct xfs_trans *tp,
190 int *diff_len,
191 int *diff_iovecs)
Dave Chinner71e330b2010-05-21 14:37:18 +1000192{
Christoph Hellwig0244b962011-12-06 21:58:08 +0000193 struct xfs_log_item_desc *lidp;
Dave Chinner71e330b2010-05-21 14:37:18 +1000194
Christoph Hellwig0244b962011-12-06 21:58:08 +0000195
196 /* Bail out if we didn't find a log item. */
197 if (list_empty(&tp->t_items)) {
198 ASSERT(0);
Dave Chinner991aaf62013-08-12 20:50:07 +1000199 return;
Christoph Hellwig0244b962011-12-06 21:58:08 +0000200 }
201
202 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
Dave Chinner166d1362013-08-12 20:50:04 +1000203 struct xfs_log_item *lip = lidp->lid_item;
Dave Chinner7492c5b2013-08-12 20:50:05 +1000204 struct xfs_log_vec *lv;
Dave Chinner991aaf62013-08-12 20:50:07 +1000205 struct xfs_log_vec *old_lv;
Dave Chinner7492c5b2013-08-12 20:50:05 +1000206 int niovecs = 0;
207 int nbytes = 0;
208 int buf_size;
Dave Chinnerfd638752013-06-27 16:04:51 +1000209 bool ordered = false;
Dave Chinner71e330b2010-05-21 14:37:18 +1000210
Christoph Hellwig0244b962011-12-06 21:58:08 +0000211 /* Skip items which aren't dirty in this transaction. */
212 if (!(lidp->lid_flags & XFS_LID_DIRTY))
213 continue;
214
Dave Chinner166d1362013-08-12 20:50:04 +1000215 /* get number of vecs and size of data to be stored */
216 lip->li_ops->iop_size(lip, &niovecs, &nbytes);
217
Christoph Hellwig0244b962011-12-06 21:58:08 +0000218 /* Skip items that do not have any vectors for writing */
Christoph Hellwigb3934212011-12-06 21:58:09 +0000219 if (!niovecs)
Christoph Hellwig0244b962011-12-06 21:58:08 +0000220 continue;
221
Dave Chinnerfd638752013-06-27 16:04:51 +1000222 /*
223 * Ordered items need to be tracked but we do not wish to write
224 * them. We need a logvec to track the object, but we do not
225 * need an iovec or buffer to be allocated for copying data.
226 */
227 if (niovecs == XFS_LOG_VEC_ORDERED) {
228 ordered = true;
229 niovecs = 0;
Dave Chinner7492c5b2013-08-12 20:50:05 +1000230 nbytes = 0;
Dave Chinnerfd638752013-06-27 16:04:51 +1000231 }
232
Dave Chinner991aaf62013-08-12 20:50:07 +1000233 /* grab the old item if it exists for reservation accounting */
234 old_lv = lip->li_lv;
235
Dave Chinner7492c5b2013-08-12 20:50:05 +1000236 /* calc buffer size */
237 buf_size = sizeof(struct xfs_log_vec) + nbytes +
238 niovecs * sizeof(struct xfs_log_iovec);
Christoph Hellwig0244b962011-12-06 21:58:08 +0000239
Dave Chinnerf5baac32013-08-12 20:50:06 +1000240 /* compare to existing item size */
241 if (lip->li_lv && buf_size <= lip->li_lv->lv_size) {
242 /* same or smaller, optimise common overwrite case */
243 lv = lip->li_lv;
244 lv->lv_next = NULL;
245
246 if (ordered)
247 goto insert;
248
Dave Chinner991aaf62013-08-12 20:50:07 +1000249 /*
250 * set the item up as though it is a new insertion so
251 * that the space reservation accounting is correct.
252 */
253 *diff_iovecs -= lv->lv_niovecs;
254 *diff_len -= lv->lv_buf_len;
255
Dave Chinnerf5baac32013-08-12 20:50:06 +1000256 /* Ensure the lv is set up according to ->iop_size */
257 lv->lv_niovecs = niovecs;
258 lv->lv_buf = (char *)lv + buf_size - nbytes;
Dave Chinner991aaf62013-08-12 20:50:07 +1000259
Dave Chinnerf5baac32013-08-12 20:50:06 +1000260 lv->lv_buf_len = xlog_cil_lv_item_format(lip, lv);
261 goto insert;
262 }
263
Dave Chinner7492c5b2013-08-12 20:50:05 +1000264 /* allocate new data chunk */
265 lv = kmem_zalloc(buf_size, KM_SLEEP|KM_NOFS);
266 lv->lv_item = lip;
267 lv->lv_size = buf_size;
268 lv->lv_niovecs = niovecs;
Dave Chinnerfd638752013-06-27 16:04:51 +1000269 if (ordered) {
270 /* track as an ordered logvec */
Dave Chinner7492c5b2013-08-12 20:50:05 +1000271 ASSERT(lip->li_lv == NULL);
272 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
273 goto insert;
Dave Chinnerfd638752013-06-27 16:04:51 +1000274 }
275
Christoph Hellwig0244b962011-12-06 21:58:08 +0000276 /* The allocated iovec region lies beyond the log vector. */
Dave Chinner7492c5b2013-08-12 20:50:05 +1000277 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
Christoph Hellwig0244b962011-12-06 21:58:08 +0000278
Dave Chinner7492c5b2013-08-12 20:50:05 +1000279 /* The allocated data region lies beyond the iovec region */
280 lv->lv_buf = (char *)lv + buf_size - nbytes;
Dave Chinner71e330b2010-05-21 14:37:18 +1000281
Dave Chinner7492c5b2013-08-12 20:50:05 +1000282 lv->lv_buf_len = xlog_cil_lv_item_format(lip, lv);
Dave Chinner7492c5b2013-08-12 20:50:05 +1000283insert:
Dave Chinnerf5baac32013-08-12 20:50:06 +1000284 ASSERT(lv->lv_buf_len <= nbytes);
Dave Chinner991aaf62013-08-12 20:50:07 +1000285 xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
Dave Chinner71e330b2010-05-21 14:37:18 +1000286 }
Dave Chinnerd1583a32010-09-24 18:14:13 +1000287}
288
289/*
290 * Insert the log items into the CIL and calculate the difference in space
291 * consumed by the item. Add the space to the checkpoint ticket and calculate
292 * if the change requires additional log metadata. If it does, take that space
Justin P. Mattock42b2aa82011-11-28 20:31:00 -0800293 * as well. Remove the amount of space we added to the checkpoint ticket from
Dave Chinnerd1583a32010-09-24 18:14:13 +1000294 * the current transaction ticket so that the accounting works out correctly.
295 */
Dave Chinner71e330b2010-05-21 14:37:18 +1000296static void
Dave Chinner3b93c7a2010-08-24 11:45:53 +1000297xlog_cil_insert_items(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500298 struct xlog *log,
Dave Chinner991aaf62013-08-12 20:50:07 +1000299 struct xfs_trans *tp)
Dave Chinner3b93c7a2010-08-24 11:45:53 +1000300{
Dave Chinnerd1583a32010-09-24 18:14:13 +1000301 struct xfs_cil *cil = log->l_cilp;
302 struct xfs_cil_ctx *ctx = cil->xc_ctx;
Dave Chinner991aaf62013-08-12 20:50:07 +1000303 struct xfs_log_item_desc *lidp;
Dave Chinnerd1583a32010-09-24 18:14:13 +1000304 int len = 0;
305 int diff_iovecs = 0;
306 int iclog_space;
Dave Chinner3b93c7a2010-08-24 11:45:53 +1000307
Dave Chinner991aaf62013-08-12 20:50:07 +1000308 ASSERT(tp);
Dave Chinnerd1583a32010-09-24 18:14:13 +1000309
310 /*
Dave Chinnerd1583a32010-09-24 18:14:13 +1000311 * We can do this safely because the context can't checkpoint until we
312 * are done so it doesn't matter exactly how we update the CIL.
313 */
Dave Chinner991aaf62013-08-12 20:50:07 +1000314 xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
315
316 /*
317 * Now (re-)position everything modified at the tail of the CIL.
318 * We do this here so we only need to take the CIL lock once during
319 * the transaction commit.
320 */
Dave Chinnerfd638752013-06-27 16:04:51 +1000321 spin_lock(&cil->xc_cil_lock);
Dave Chinner991aaf62013-08-12 20:50:07 +1000322 list_for_each_entry(lidp, &tp->t_items, lid_trans) {
323 struct xfs_log_item *lip = lidp->lid_item;
Dave Chinnerfd638752013-06-27 16:04:51 +1000324
Dave Chinner991aaf62013-08-12 20:50:07 +1000325 /* Skip items which aren't dirty in this transaction. */
326 if (!(lidp->lid_flags & XFS_LID_DIRTY))
327 continue;
Dave Chinnerfd638752013-06-27 16:04:51 +1000328
Dave Chinner991aaf62013-08-12 20:50:07 +1000329 list_move_tail(&lip->li_cil, &cil->xc_cil);
Dave Chinnerfd638752013-06-27 16:04:51 +1000330 }
Dave Chinnerd1583a32010-09-24 18:14:13 +1000331
332 /* account for space used by new iovec headers */
333 len += diff_iovecs * sizeof(xlog_op_header_t);
Dave Chinnerd1583a32010-09-24 18:14:13 +1000334 ctx->nvecs += diff_iovecs;
335
Dave Chinner991aaf62013-08-12 20:50:07 +1000336 /* attach the transaction to the CIL if it has any busy extents */
337 if (!list_empty(&tp->t_busy))
338 list_splice_init(&tp->t_busy, &ctx->busy_extents);
339
Dave Chinnerd1583a32010-09-24 18:14:13 +1000340 /*
341 * Now transfer enough transaction reservation to the context ticket
342 * for the checkpoint. The context ticket is special - the unit
343 * reservation has to grow as well as the current reservation as we
344 * steal from tickets so we can correctly determine the space used
345 * during the transaction commit.
346 */
347 if (ctx->ticket->t_curr_res == 0) {
Dave Chinnerd1583a32010-09-24 18:14:13 +1000348 ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
Dave Chinner991aaf62013-08-12 20:50:07 +1000349 tp->t_ticket->t_curr_res -= ctx->ticket->t_unit_res;
Dave Chinnerd1583a32010-09-24 18:14:13 +1000350 }
351
352 /* do we need space for more log record headers? */
353 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
354 if (len > 0 && (ctx->space_used / iclog_space !=
355 (ctx->space_used + len) / iclog_space)) {
356 int hdrs;
357
358 hdrs = (len + iclog_space - 1) / iclog_space;
359 /* need to take into account split region headers, too */
360 hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
361 ctx->ticket->t_unit_res += hdrs;
362 ctx->ticket->t_curr_res += hdrs;
Dave Chinner991aaf62013-08-12 20:50:07 +1000363 tp->t_ticket->t_curr_res -= hdrs;
364 ASSERT(tp->t_ticket->t_curr_res >= len);
Dave Chinnerd1583a32010-09-24 18:14:13 +1000365 }
Dave Chinner991aaf62013-08-12 20:50:07 +1000366 tp->t_ticket->t_curr_res -= len;
Dave Chinnerd1583a32010-09-24 18:14:13 +1000367 ctx->space_used += len;
368
369 spin_unlock(&cil->xc_cil_lock);
Dave Chinner3b93c7a2010-08-24 11:45:53 +1000370}
371
372static void
Dave Chinner71e330b2010-05-21 14:37:18 +1000373xlog_cil_free_logvec(
374 struct xfs_log_vec *log_vector)
375{
376 struct xfs_log_vec *lv;
377
378 for (lv = log_vector; lv; ) {
379 struct xfs_log_vec *next = lv->lv_next;
Dave Chinner71e330b2010-05-21 14:37:18 +1000380 kmem_free(lv);
381 lv = next;
382 }
383}
384
385/*
Dave Chinner71e330b2010-05-21 14:37:18 +1000386 * Mark all items committed and clear busy extents. We free the log vector
387 * chains in a separate pass so that we unpin the log items as quickly as
388 * possible.
389 */
390static void
391xlog_cil_committed(
392 void *args,
393 int abort)
394{
395 struct xfs_cil_ctx *ctx = args;
Christoph Hellwige84661a2011-05-20 13:45:32 +0000396 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
Dave Chinner71e330b2010-05-21 14:37:18 +1000397
Dave Chinner0e57f6a2010-12-20 12:02:19 +1100398 xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
399 ctx->start_lsn, abort);
Dave Chinner71e330b2010-05-21 14:37:18 +1000400
Dave Chinner4ecbfe62012-04-29 10:41:10 +0000401 xfs_extent_busy_sort(&ctx->busy_extents);
402 xfs_extent_busy_clear(mp, &ctx->busy_extents,
Christoph Hellwige84661a2011-05-20 13:45:32 +0000403 (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
Dave Chinner71e330b2010-05-21 14:37:18 +1000404
Dave Chinner4bb928c2013-08-12 20:50:08 +1000405 spin_lock(&ctx->cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000406 list_del(&ctx->committing);
Dave Chinner4bb928c2013-08-12 20:50:08 +1000407 spin_unlock(&ctx->cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000408
409 xlog_cil_free_logvec(ctx->lv_chain);
Christoph Hellwige84661a2011-05-20 13:45:32 +0000410
411 if (!list_empty(&ctx->busy_extents)) {
412 ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
413
414 xfs_discard_extents(mp, &ctx->busy_extents);
Dave Chinner4ecbfe62012-04-29 10:41:10 +0000415 xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
Christoph Hellwige84661a2011-05-20 13:45:32 +0000416 }
417
Dave Chinner71e330b2010-05-21 14:37:18 +1000418 kmem_free(ctx);
419}
420
421/*
Dave Chinnera44f13e2010-08-24 11:40:03 +1000422 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
423 * is a background flush and so we can chose to ignore it. Otherwise, if the
424 * current sequence is the same as @push_seq we need to do a flush. If
425 * @push_seq is less than the current sequence, then it has already been
426 * flushed and we don't need to do anything - the caller will wait for it to
427 * complete if necessary.
428 *
429 * @push_seq is a value rather than a flag because that allows us to do an
430 * unlocked check of the sequence number for a match. Hence we can allows log
431 * forces to run racily and not issue pushes for the same sequence twice. If we
432 * get a race between multiple pushes for the same sequence they will block on
433 * the first one and then abort, hence avoiding needless pushes.
Dave Chinner71e330b2010-05-21 14:37:18 +1000434 */
Dave Chinnera44f13e2010-08-24 11:40:03 +1000435STATIC int
Dave Chinner71e330b2010-05-21 14:37:18 +1000436xlog_cil_push(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500437 struct xlog *log)
Dave Chinner71e330b2010-05-21 14:37:18 +1000438{
439 struct xfs_cil *cil = log->l_cilp;
440 struct xfs_log_vec *lv;
441 struct xfs_cil_ctx *ctx;
442 struct xfs_cil_ctx *new_ctx;
443 struct xlog_in_core *commit_iclog;
444 struct xlog_ticket *tic;
Dave Chinner71e330b2010-05-21 14:37:18 +1000445 int num_iovecs;
Dave Chinner71e330b2010-05-21 14:37:18 +1000446 int error = 0;
447 struct xfs_trans_header thdr;
448 struct xfs_log_iovec lhdr;
449 struct xfs_log_vec lvhdr = { NULL };
450 xfs_lsn_t commit_lsn;
Dave Chinner4c2d5422012-04-23 17:54:32 +1000451 xfs_lsn_t push_seq;
Dave Chinner71e330b2010-05-21 14:37:18 +1000452
453 if (!cil)
454 return 0;
455
Dave Chinner71e330b2010-05-21 14:37:18 +1000456 new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
457 new_ctx->ticket = xlog_cil_ticket_alloc(log);
458
Dave Chinner4c2d5422012-04-23 17:54:32 +1000459 down_write(&cil->xc_ctx_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000460 ctx = cil->xc_ctx;
461
Dave Chinner4bb928c2013-08-12 20:50:08 +1000462 spin_lock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000463 push_seq = cil->xc_push_seq;
464 ASSERT(push_seq <= ctx->sequence);
Dave Chinner71e330b2010-05-21 14:37:18 +1000465
Dave Chinner4c2d5422012-04-23 17:54:32 +1000466 /*
467 * Check if we've anything to push. If there is nothing, then we don't
468 * move on to a new sequence number and so we have to be able to push
469 * this sequence again later.
470 */
471 if (list_empty(&cil->xc_cil)) {
472 cil->xc_push_seq = 0;
Dave Chinner4bb928c2013-08-12 20:50:08 +1000473 spin_unlock(&cil->xc_push_lock);
Dave Chinnera44f13e2010-08-24 11:40:03 +1000474 goto out_skip;
Dave Chinner4c2d5422012-04-23 17:54:32 +1000475 }
Dave Chinner4bb928c2013-08-12 20:50:08 +1000476 spin_unlock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000477
Dave Chinnera44f13e2010-08-24 11:40:03 +1000478
479 /* check for a previously pushed seqeunce */
Dave Chinner4c2d5422012-04-23 17:54:32 +1000480 if (push_seq < cil->xc_ctx->sequence)
Dave Chinnerdf806152010-05-17 15:52:13 +1000481 goto out_skip;
482
Dave Chinner71e330b2010-05-21 14:37:18 +1000483 /*
484 * pull all the log vectors off the items in the CIL, and
485 * remove the items from the CIL. We don't need the CIL lock
486 * here because it's only needed on the transaction commit
487 * side which is currently locked out by the flush lock.
488 */
489 lv = NULL;
Dave Chinner71e330b2010-05-21 14:37:18 +1000490 num_iovecs = 0;
Dave Chinner71e330b2010-05-21 14:37:18 +1000491 while (!list_empty(&cil->xc_cil)) {
492 struct xfs_log_item *item;
Dave Chinner71e330b2010-05-21 14:37:18 +1000493
494 item = list_first_entry(&cil->xc_cil,
495 struct xfs_log_item, li_cil);
496 list_del_init(&item->li_cil);
497 if (!ctx->lv_chain)
498 ctx->lv_chain = item->li_lv;
499 else
500 lv->lv_next = item->li_lv;
501 lv = item->li_lv;
502 item->li_lv = NULL;
Dave Chinner71e330b2010-05-21 14:37:18 +1000503 num_iovecs += lv->lv_niovecs;
Dave Chinner71e330b2010-05-21 14:37:18 +1000504 }
505
506 /*
507 * initialise the new context and attach it to the CIL. Then attach
508 * the current context to the CIL committing lsit so it can be found
509 * during log forces to extract the commit lsn of the sequence that
510 * needs to be forced.
511 */
512 INIT_LIST_HEAD(&new_ctx->committing);
513 INIT_LIST_HEAD(&new_ctx->busy_extents);
514 new_ctx->sequence = ctx->sequence + 1;
515 new_ctx->cil = cil;
516 cil->xc_ctx = new_ctx;
517
518 /*
Dave Chinnera44f13e2010-08-24 11:40:03 +1000519 * mirror the new sequence into the cil structure so that we can do
520 * unlocked checks against the current sequence in log forces without
521 * risking deferencing a freed context pointer.
522 */
523 cil->xc_current_sequence = new_ctx->sequence;
524
525 /*
Dave Chinner71e330b2010-05-21 14:37:18 +1000526 * The switch is now done, so we can drop the context lock and move out
527 * of a shared context. We can't just go straight to the commit record,
528 * though - we need to synchronise with previous and future commits so
529 * that the commit records are correctly ordered in the log to ensure
530 * that we process items during log IO completion in the correct order.
531 *
532 * For example, if we get an EFI in one checkpoint and the EFD in the
533 * next (e.g. due to log forces), we do not want the checkpoint with
534 * the EFD to be committed before the checkpoint with the EFI. Hence
535 * we must strictly order the commit records of the checkpoints so
536 * that: a) the checkpoint callbacks are attached to the iclogs in the
537 * correct order; and b) the checkpoints are replayed in correct order
538 * in log recovery.
539 *
540 * Hence we need to add this context to the committing context list so
541 * that higher sequences will wait for us to write out a commit record
542 * before they do.
543 */
Dave Chinner4bb928c2013-08-12 20:50:08 +1000544 spin_lock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000545 list_add(&ctx->committing, &cil->xc_committing);
Dave Chinner4bb928c2013-08-12 20:50:08 +1000546 spin_unlock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000547 up_write(&cil->xc_ctx_lock);
548
549 /*
550 * Build a checkpoint transaction header and write it to the log to
551 * begin the transaction. We need to account for the space used by the
552 * transaction header here as it is not accounted for in xlog_write().
553 *
554 * The LSN we need to pass to the log items on transaction commit is
555 * the LSN reported by the first log vector write. If we use the commit
556 * record lsn then we can move the tail beyond the grant write head.
557 */
558 tic = ctx->ticket;
559 thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
560 thdr.th_type = XFS_TRANS_CHECKPOINT;
561 thdr.th_tid = tic->t_tid;
562 thdr.th_num_items = num_iovecs;
Christoph Hellwig4e0d5f92010-06-23 18:11:15 +1000563 lhdr.i_addr = &thdr;
Dave Chinner71e330b2010-05-21 14:37:18 +1000564 lhdr.i_len = sizeof(xfs_trans_header_t);
565 lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
566 tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
567
568 lvhdr.lv_niovecs = 1;
569 lvhdr.lv_iovecp = &lhdr;
570 lvhdr.lv_next = ctx->lv_chain;
571
572 error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
573 if (error)
Dave Chinner7db37c52011-01-27 12:02:00 +1100574 goto out_abort_free_ticket;
Dave Chinner71e330b2010-05-21 14:37:18 +1000575
576 /*
577 * now that we've written the checkpoint into the log, strictly
578 * order the commit records so replay will get them in the right order.
579 */
580restart:
Dave Chinner4bb928c2013-08-12 20:50:08 +1000581 spin_lock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000582 list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
583 /*
584 * Higher sequences will wait for this one so skip them.
585 * Don't wait for own own sequence, either.
586 */
587 if (new_ctx->sequence >= ctx->sequence)
588 continue;
589 if (!new_ctx->commit_lsn) {
590 /*
591 * It is still being pushed! Wait for the push to
592 * complete, then start again from the beginning.
593 */
Dave Chinner4bb928c2013-08-12 20:50:08 +1000594 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000595 goto restart;
596 }
597 }
Dave Chinner4bb928c2013-08-12 20:50:08 +1000598 spin_unlock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000599
Dave Chinner7db37c52011-01-27 12:02:00 +1100600 /* xfs_log_done always frees the ticket on error. */
Dave Chinner71e330b2010-05-21 14:37:18 +1000601 commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
Dave Chinner7db37c52011-01-27 12:02:00 +1100602 if (commit_lsn == -1)
Dave Chinner71e330b2010-05-21 14:37:18 +1000603 goto out_abort;
604
605 /* attach all the transactions w/ busy extents to iclog */
606 ctx->log_cb.cb_func = xlog_cil_committed;
607 ctx->log_cb.cb_arg = ctx;
608 error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
609 if (error)
610 goto out_abort;
611
612 /*
613 * now the checkpoint commit is complete and we've attached the
614 * callbacks to the iclog we can assign the commit LSN to the context
615 * and wake up anyone who is waiting for the commit to complete.
616 */
Dave Chinner4bb928c2013-08-12 20:50:08 +1000617 spin_lock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000618 ctx->commit_lsn = commit_lsn;
Dave Chinnereb40a872010-12-21 12:09:01 +1100619 wake_up_all(&cil->xc_commit_wait);
Dave Chinner4bb928c2013-08-12 20:50:08 +1000620 spin_unlock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000621
622 /* release the hounds! */
623 return xfs_log_release_iclog(log->l_mp, commit_iclog);
624
625out_skip:
626 up_write(&cil->xc_ctx_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000627 xfs_log_ticket_put(new_ctx->ticket);
628 kmem_free(new_ctx);
629 return 0;
630
Dave Chinner7db37c52011-01-27 12:02:00 +1100631out_abort_free_ticket:
632 xfs_log_ticket_put(tic);
Dave Chinner71e330b2010-05-21 14:37:18 +1000633out_abort:
634 xlog_cil_committed(ctx, XFS_LI_ABORTED);
635 return XFS_ERROR(EIO);
636}
637
Dave Chinner4c2d5422012-04-23 17:54:32 +1000638static void
639xlog_cil_push_work(
640 struct work_struct *work)
641{
642 struct xfs_cil *cil = container_of(work, struct xfs_cil,
643 xc_push_work);
644 xlog_cil_push(cil->xc_log);
645}
646
647/*
648 * We need to push CIL every so often so we don't cache more than we can fit in
649 * the log. The limit really is that a checkpoint can't be more than half the
650 * log (the current checkpoint is not allowed to overwrite the previous
651 * checkpoint), but commit latency and memory usage limit this to a smaller
652 * size.
653 */
654static void
655xlog_cil_push_background(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500656 struct xlog *log)
Dave Chinner4c2d5422012-04-23 17:54:32 +1000657{
658 struct xfs_cil *cil = log->l_cilp;
659
660 /*
661 * The cil won't be empty because we are called while holding the
662 * context lock so whatever we added to the CIL will still be there
663 */
664 ASSERT(!list_empty(&cil->xc_cil));
665
666 /*
667 * don't do a background push if we haven't used up all the
668 * space available yet.
669 */
670 if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
671 return;
672
Dave Chinner4bb928c2013-08-12 20:50:08 +1000673 spin_lock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000674 if (cil->xc_push_seq < cil->xc_current_sequence) {
675 cil->xc_push_seq = cil->xc_current_sequence;
676 queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
677 }
Dave Chinner4bb928c2013-08-12 20:50:08 +1000678 spin_unlock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000679
680}
681
682static void
683xlog_cil_push_foreground(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500684 struct xlog *log,
Dave Chinner4c2d5422012-04-23 17:54:32 +1000685 xfs_lsn_t push_seq)
686{
687 struct xfs_cil *cil = log->l_cilp;
688
689 if (!cil)
690 return;
691
692 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
693
694 /* start on any pending background push to minimise wait time on it */
695 flush_work(&cil->xc_push_work);
696
697 /*
698 * If the CIL is empty or we've already pushed the sequence then
699 * there's no work we need to do.
700 */
Dave Chinner4bb928c2013-08-12 20:50:08 +1000701 spin_lock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000702 if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
Dave Chinner4bb928c2013-08-12 20:50:08 +1000703 spin_unlock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000704 return;
705 }
706
707 cil->xc_push_seq = push_seq;
Dave Chinner4bb928c2013-08-12 20:50:08 +1000708 spin_unlock(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000709
710 /* do the push now */
711 xlog_cil_push(log);
712}
713
Dave Chinner71e330b2010-05-21 14:37:18 +1000714/*
Dave Chinnera44f13e2010-08-24 11:40:03 +1000715 * Commit a transaction with the given vector to the Committed Item List.
716 *
717 * To do this, we need to format the item, pin it in memory if required and
718 * account for the space used by the transaction. Once we have done that we
719 * need to release the unused reservation for the transaction, attach the
720 * transaction to the checkpoint context so we carry the busy extents through
721 * to checkpoint completion, and then unlock all the items in the transaction.
722 *
Dave Chinnera44f13e2010-08-24 11:40:03 +1000723 * Called with the context lock already held in read mode to lock out
724 * background commit, returns without it held once background commits are
725 * allowed again.
726 */
Christoph Hellwig0244b962011-12-06 21:58:08 +0000727int
Dave Chinnera44f13e2010-08-24 11:40:03 +1000728xfs_log_commit_cil(
729 struct xfs_mount *mp,
730 struct xfs_trans *tp,
Dave Chinnera44f13e2010-08-24 11:40:03 +1000731 xfs_lsn_t *commit_lsn,
732 int flags)
733{
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500734 struct xlog *log = mp->m_log;
Dave Chinner991aaf62013-08-12 20:50:07 +1000735 struct xfs_cil *cil = log->l_cilp;
Dave Chinnera44f13e2010-08-24 11:40:03 +1000736 int log_flags = 0;
Dave Chinnera44f13e2010-08-24 11:40:03 +1000737
738 if (flags & XFS_TRANS_RELEASE_LOG_RES)
739 log_flags = XFS_LOG_REL_PERM_RESERV;
740
Dave Chinnerf5baac32013-08-12 20:50:06 +1000741 /* lock out background commit */
Dave Chinner991aaf62013-08-12 20:50:07 +1000742 down_read(&cil->xc_ctx_lock);
Dave Chinnerf5baac32013-08-12 20:50:06 +1000743
Dave Chinner991aaf62013-08-12 20:50:07 +1000744 xlog_cil_insert_items(log, tp);
Dave Chinnera44f13e2010-08-24 11:40:03 +1000745
746 /* check we didn't blow the reservation */
747 if (tp->t_ticket->t_curr_res < 0)
Dave Chinner991aaf62013-08-12 20:50:07 +1000748 xlog_print_tic_res(mp, tp->t_ticket);
Dave Chinnera44f13e2010-08-24 11:40:03 +1000749
Dave Chinner991aaf62013-08-12 20:50:07 +1000750 tp->t_commit_lsn = cil->xc_ctx->sequence;
751 if (commit_lsn)
752 *commit_lsn = tp->t_commit_lsn;
Dave Chinnera44f13e2010-08-24 11:40:03 +1000753
Dave Chinnera44f13e2010-08-24 11:40:03 +1000754 xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
755 xfs_trans_unreserve_and_mod_sb(tp);
756
757 /*
758 * Once all the items of the transaction have been copied to the CIL,
759 * the items can be unlocked and freed.
760 *
761 * This needs to be done before we drop the CIL context lock because we
762 * have to update state in the log items and unlock them before they go
763 * to disk. If we don't, then the CIL checkpoint can race with us and
764 * we can run checkpoint completion before we've updated and unlocked
765 * the log items. This affects (at least) processing of stale buffers,
766 * inodes and EFIs.
767 */
Dave Chinner991aaf62013-08-12 20:50:07 +1000768 xfs_trans_free_items(tp, tp->t_commit_lsn, 0);
Dave Chinnera44f13e2010-08-24 11:40:03 +1000769
Dave Chinner4c2d5422012-04-23 17:54:32 +1000770 xlog_cil_push_background(log);
Dave Chinnera44f13e2010-08-24 11:40:03 +1000771
Dave Chinner991aaf62013-08-12 20:50:07 +1000772 up_read(&cil->xc_ctx_lock);
Christoph Hellwig0244b962011-12-06 21:58:08 +0000773 return 0;
Dave Chinnera44f13e2010-08-24 11:40:03 +1000774}
775
776/*
Dave Chinner71e330b2010-05-21 14:37:18 +1000777 * Conditionally push the CIL based on the sequence passed in.
778 *
779 * We only need to push if we haven't already pushed the sequence
780 * number given. Hence the only time we will trigger a push here is
781 * if the push sequence is the same as the current context.
782 *
783 * We return the current commit lsn to allow the callers to determine if a
784 * iclog flush is necessary following this call.
Dave Chinner71e330b2010-05-21 14:37:18 +1000785 */
786xfs_lsn_t
Dave Chinnera44f13e2010-08-24 11:40:03 +1000787xlog_cil_force_lsn(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500788 struct xlog *log,
Dave Chinnera44f13e2010-08-24 11:40:03 +1000789 xfs_lsn_t sequence)
Dave Chinner71e330b2010-05-21 14:37:18 +1000790{
791 struct xfs_cil *cil = log->l_cilp;
792 struct xfs_cil_ctx *ctx;
793 xfs_lsn_t commit_lsn = NULLCOMMITLSN;
794
Dave Chinnera44f13e2010-08-24 11:40:03 +1000795 ASSERT(sequence <= cil->xc_current_sequence);
Dave Chinner71e330b2010-05-21 14:37:18 +1000796
Dave Chinnera44f13e2010-08-24 11:40:03 +1000797 /*
798 * check to see if we need to force out the current context.
799 * xlog_cil_push() handles racing pushes for the same sequence,
800 * so no need to deal with it here.
801 */
Dave Chinner4c2d5422012-04-23 17:54:32 +1000802 xlog_cil_push_foreground(log, sequence);
Dave Chinner71e330b2010-05-21 14:37:18 +1000803
804 /*
805 * See if we can find a previous sequence still committing.
Dave Chinner71e330b2010-05-21 14:37:18 +1000806 * We need to wait for all previous sequence commits to complete
807 * before allowing the force of push_seq to go ahead. Hence block
808 * on commits for those as well.
809 */
Dave Chinnera44f13e2010-08-24 11:40:03 +1000810restart:
Dave Chinner4bb928c2013-08-12 20:50:08 +1000811 spin_lock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000812 list_for_each_entry(ctx, &cil->xc_committing, committing) {
Dave Chinnera44f13e2010-08-24 11:40:03 +1000813 if (ctx->sequence > sequence)
Dave Chinner71e330b2010-05-21 14:37:18 +1000814 continue;
815 if (!ctx->commit_lsn) {
816 /*
817 * It is still being pushed! Wait for the push to
818 * complete, then start again from the beginning.
819 */
Dave Chinner4bb928c2013-08-12 20:50:08 +1000820 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000821 goto restart;
822 }
Dave Chinnera44f13e2010-08-24 11:40:03 +1000823 if (ctx->sequence != sequence)
Dave Chinner71e330b2010-05-21 14:37:18 +1000824 continue;
825 /* found it! */
826 commit_lsn = ctx->commit_lsn;
827 }
Dave Chinner4bb928c2013-08-12 20:50:08 +1000828 spin_unlock(&cil->xc_push_lock);
Dave Chinner71e330b2010-05-21 14:37:18 +1000829 return commit_lsn;
830}
Dave Chinnerccf7c232010-05-20 23:19:42 +1000831
832/*
833 * Check if the current log item was first committed in this sequence.
834 * We can't rely on just the log item being in the CIL, we have to check
835 * the recorded commit sequence number.
836 *
837 * Note: for this to be used in a non-racy manner, it has to be called with
838 * CIL flushing locked out. As a result, it should only be used during the
839 * transaction commit process when deciding what to format into the item.
840 */
841bool
842xfs_log_item_in_current_chkpt(
843 struct xfs_log_item *lip)
844{
845 struct xfs_cil_ctx *ctx;
846
Dave Chinnerccf7c232010-05-20 23:19:42 +1000847 if (list_empty(&lip->li_cil))
848 return false;
849
850 ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
851
852 /*
853 * li_seq is written on the first commit of a log item to record the
854 * first checkpoint it is written to. Hence if it is different to the
855 * current sequence, we're in a new checkpoint.
856 */
857 if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
858 return false;
859 return true;
860}
Dave Chinner4c2d5422012-04-23 17:54:32 +1000861
862/*
863 * Perform initial CIL structure initialisation.
864 */
865int
866xlog_cil_init(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500867 struct xlog *log)
Dave Chinner4c2d5422012-04-23 17:54:32 +1000868{
869 struct xfs_cil *cil;
870 struct xfs_cil_ctx *ctx;
871
872 cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
873 if (!cil)
874 return ENOMEM;
875
876 ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
877 if (!ctx) {
878 kmem_free(cil);
879 return ENOMEM;
880 }
881
882 INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
883 INIT_LIST_HEAD(&cil->xc_cil);
884 INIT_LIST_HEAD(&cil->xc_committing);
885 spin_lock_init(&cil->xc_cil_lock);
Dave Chinner4bb928c2013-08-12 20:50:08 +1000886 spin_lock_init(&cil->xc_push_lock);
Dave Chinner4c2d5422012-04-23 17:54:32 +1000887 init_rwsem(&cil->xc_ctx_lock);
888 init_waitqueue_head(&cil->xc_commit_wait);
889
890 INIT_LIST_HEAD(&ctx->committing);
891 INIT_LIST_HEAD(&ctx->busy_extents);
892 ctx->sequence = 1;
893 ctx->cil = cil;
894 cil->xc_ctx = ctx;
895 cil->xc_current_sequence = ctx->sequence;
896
897 cil->xc_log = log;
898 log->l_cilp = cil;
899 return 0;
900}
901
902void
903xlog_cil_destroy(
Mark Tinguelyf7bdf032012-06-14 09:22:15 -0500904 struct xlog *log)
Dave Chinner4c2d5422012-04-23 17:54:32 +1000905{
906 if (log->l_cilp->xc_ctx) {
907 if (log->l_cilp->xc_ctx->ticket)
908 xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
909 kmem_free(log->l_cilp->xc_ctx);
910 }
911
912 ASSERT(list_empty(&log->l_cilp->xc_cil));
913 kmem_free(log->l_cilp);
914}
915