1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Portions Copyright 2011 Martin Matuska <mm@FreeBSD.org>
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/txg_impl.h>
29#include <sys/dmu_impl.h>
30#include <sys/dmu_tx.h>
31#include <sys/dsl_pool.h>
32#include <sys/dsl_scan.h>
33#include <sys/callb.h>
34
35/*
36 * ZFS Transaction Groups
37 * ----------------------
38 *
39 * ZFS transaction groups are, as the name implies, groups of transactions
40 * that act on persistent state. ZFS asserts consistency at the granularity of
41 * these transaction groups. Each successive transaction group (txg) is
42 * assigned a 64-bit consecutive identifier. There are three active
43 * transaction group states: open, quiescing, or syncing. At any given time,
44 * there may be an active txg associated with each state; each active txg may
45 * either be processing, or blocked waiting to enter the next state. There may
46 * be up to three active txgs, and there is always a txg in the open state
47 * (though it may be blocked waiting to enter the quiescing state). In broad
48 * strokes, transactions ��� operations that change in-memory structures ��� are
49 * accepted into the txg in the open state, and are completed while the txg is
50 * in the open or quiescing states. The accumulated changes are written to
51 * disk in the syncing state.
52 *
53 * Open
54 *
55 * When a new txg becomes active, it first enters the open state. New
56 * transactions ��� updates to in-memory structures ��� are assigned to the
57 * currently open txg. There is always a txg in the open state so that ZFS can
58 * accept new changes (though the txg may refuse new changes if it has hit
59 * some limit). ZFS advances the open txg to the next state for a variety of
60 * reasons such as it hitting a time or size threshold, or the execution of an
61 * administrative action that must be completed in the syncing state.
62 *
63 * Quiescing
64 *
65 * After a txg exits the open state, it enters the quiescing state. The
66 * quiescing state is intended to provide a buffer between accepting new
67 * transactions in the open state and writing them out to stable storage in
68 * the syncing state. While quiescing, transactions can continue their
69 * operation without delaying either of the other states. Typically, a txg is
70 * in the quiescing state very briefly since the operations are bounded by
71 * software latencies rather than, say, slower I/O latencies. After all
72 * transactions complete, the txg is ready to enter the next state.
73 *
74 * Syncing
75 *
76 * In the syncing state, the in-memory state built up during the open and (to
77 * a lesser degree) the quiescing states is written to stable storage. The
78 * process of writing out modified data can, in turn modify more data. For
79 * example when we write new blocks, we need to allocate space for them; those
80 * allocations modify metadata (space maps)... which themselves must be
81 * written to stable storage. During the sync state, ZFS iterates, writing out
82 * data until it converges and all in-memory changes have been written out.
83 * The first such pass is the largest as it encompasses all the modified user
84 * data (as opposed to filesystem metadata). Subsequent passes typically have
85 * far less data to write as they consist exclusively of filesystem metadata.
86 *
87 * To ensure convergence, after a certain number of passes ZFS begins
88 * overwriting locations on stable storage that had been allocated earlier in
89 * the syncing state (and subsequently freed). ZFS usually allocates new
90 * blocks to optimize for large, continuous, writes. For the syncing state to
91 * converge however it must complete a pass where no new blocks are allocated
92 * since each allocation requires a modification of persistent metadata.
93 * Further, to hasten convergence, after a prescribed number of passes, ZFS
94 * also defers frees, and stops compressing.
95 *
96 * In addition to writing out user data, we must also execute synctasks during
97 * the syncing context. A synctask is the mechanism by which some
98 * administrative activities work such as creating and destroying snapshots or
99 * datasets. Note that when a synctask is initiated it enters the open txg,
100 * and ZFS then pushes that txg as quickly as possible to completion of the
101 * syncing state in order to reduce the latency of the administrative
102 * activity. To complete the syncing state, ZFS writes out a new uberblock,
103 * the root of the tree of blocks that comprise all state stored on the ZFS
104 * pool. Finally, if there is a quiesced txg waiting, we signal that it can
105 * now transition to the syncing state.
106 */
107
108static void txg_sync_thread(void *arg);
109static void txg_quiesce_thread(void *arg);
110
111int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
112
113SYSCTL_DECL(_vfs_zfs);
114SYSCTL_NODE(_vfs_zfs, OID_AUTO, txg, CTLFLAG_RW, 0, "ZFS TXG");
115TUNABLE_INT("vfs.zfs.txg.timeout", &zfs_txg_timeout);
116SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, timeout, CTLFLAG_RW, &zfs_txg_timeout, 0,
117 "Maximum seconds worth of delta per txg");
118
119/*
120 * Prepare the txg subsystem.
121 */
122void
123txg_init(dsl_pool_t *dp, uint64_t txg)
124{
125 tx_state_t *tx = &dp->dp_tx;
126 int c;
127 bzero(tx, sizeof (tx_state_t));
128
129 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
130
131 for (c = 0; c < max_ncpus; c++) {
132 int i;
133
134 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
135 mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT,
136 NULL);
137 for (i = 0; i < TXG_SIZE; i++) {
138 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
139 NULL);
140 list_create(&tx->tx_cpu[c].tc_callbacks[i],
141 sizeof (dmu_tx_callback_t),
142 offsetof(dmu_tx_callback_t, dcb_node));
143 }
144 }
145
146 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
147
148 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
149 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
150 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
151 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
152 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
153
154 tx->tx_open_txg = txg;
155}
156
157/*
158 * Close down the txg subsystem.
159 */
160void
161txg_fini(dsl_pool_t *dp)
162{
163 tx_state_t *tx = &dp->dp_tx;
164 int c;
165
166 ASSERT(tx->tx_threads == 0);
167
168 mutex_destroy(&tx->tx_sync_lock);
169
170 cv_destroy(&tx->tx_sync_more_cv);
171 cv_destroy(&tx->tx_sync_done_cv);
172 cv_destroy(&tx->tx_quiesce_more_cv);
173 cv_destroy(&tx->tx_quiesce_done_cv);
174 cv_destroy(&tx->tx_exit_cv);
175
176 for (c = 0; c < max_ncpus; c++) {
177 int i;
178
179 mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
180 mutex_destroy(&tx->tx_cpu[c].tc_lock);
181 for (i = 0; i < TXG_SIZE; i++) {
182 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
183 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
184 }
185 }
186
187 if (tx->tx_commit_cb_taskq != NULL)
188 taskq_destroy(tx->tx_commit_cb_taskq);
189
190 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
191
192 bzero(tx, sizeof (tx_state_t));
193}
194
195/*
196 * Start syncing transaction groups.
197 */
198void
199txg_sync_start(dsl_pool_t *dp)
200{
201 tx_state_t *tx = &dp->dp_tx;
202
203 mutex_enter(&tx->tx_sync_lock);
204
205 dprintf("pool %p\n", dp);
206
207 ASSERT(tx->tx_threads == 0);
208
209 tx->tx_threads = 2;
210
211 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
212 dp, 0, &p0, TS_RUN, minclsyspri);
213
214 /*
215 * The sync thread can need a larger-than-default stack size on
216 * 32-bit x86. This is due in part to nested pools and
217 * scrub_visitbp() recursion.
218 */
219 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
220 dp, 0, &p0, TS_RUN, minclsyspri);
221
222 mutex_exit(&tx->tx_sync_lock);
223}
224
225static void
226txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
227{
228 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
229 mutex_enter(&tx->tx_sync_lock);
230}
231
232static void
233txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
234{
235 ASSERT(*tpp != NULL);
236 *tpp = NULL;
237 tx->tx_threads--;
238 cv_broadcast(&tx->tx_exit_cv);
239 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
240 thread_exit();
241}
242
243static void
244txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
245{
246 CALLB_CPR_SAFE_BEGIN(cpr);
247
248 if (time)
249 (void) cv_timedwait(cv, &tx->tx_sync_lock, time);
250 else
251 cv_wait(cv, &tx->tx_sync_lock);
252
253 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
254}
255
256/*
257 * Stop syncing transaction groups.
258 */
259void
260txg_sync_stop(dsl_pool_t *dp)
261{
262 tx_state_t *tx = &dp->dp_tx;
263
264 dprintf("pool %p\n", dp);
265 /*
266 * Finish off any work in progress.
267 */
268 ASSERT(tx->tx_threads == 2);
269
270 /*
271 * We need to ensure that we've vacated the deferred space_maps.
272 */
273 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
274
275 /*
276 * Wake all sync threads and wait for them to die.
277 */
278 mutex_enter(&tx->tx_sync_lock);
279
280 ASSERT(tx->tx_threads == 2);
281
282 tx->tx_exiting = 1;
283
284 cv_broadcast(&tx->tx_quiesce_more_cv);
285 cv_broadcast(&tx->tx_quiesce_done_cv);
286 cv_broadcast(&tx->tx_sync_more_cv);
287
288 while (tx->tx_threads != 0)
289 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
290
291 tx->tx_exiting = 0;
292
293 mutex_exit(&tx->tx_sync_lock);
294}
295
296uint64_t
297txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
298{
299 tx_state_t *tx = &dp->dp_tx;
300 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
301 uint64_t txg;
302
303 mutex_enter(&tc->tc_open_lock);
304 txg = tx->tx_open_txg;
305
306 mutex_enter(&tc->tc_lock);
307 tc->tc_count[txg & TXG_MASK]++;
308 mutex_exit(&tc->tc_lock);
309
310 th->th_cpu = tc;
311 th->th_txg = txg;
312
313 return (txg);
314}
315
316void
317txg_rele_to_quiesce(txg_handle_t *th)
318{
319 tx_cpu_t *tc = th->th_cpu;
320
321 ASSERT(!MUTEX_HELD(&tc->tc_lock));
322 mutex_exit(&tc->tc_open_lock);
323}
324
325void
326txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
327{
328 tx_cpu_t *tc = th->th_cpu;
329 int g = th->th_txg & TXG_MASK;
330
331 mutex_enter(&tc->tc_lock);
332 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
333 mutex_exit(&tc->tc_lock);
334}
335
336void
337txg_rele_to_sync(txg_handle_t *th)
338{
339 tx_cpu_t *tc = th->th_cpu;
340 int g = th->th_txg & TXG_MASK;
341
342 mutex_enter(&tc->tc_lock);
343 ASSERT(tc->tc_count[g] != 0);
344 if (--tc->tc_count[g] == 0)
345 cv_broadcast(&tc->tc_cv[g]);
346 mutex_exit(&tc->tc_lock);
347
348 th->th_cpu = NULL; /* defensive */
349}
350
| 1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Portions Copyright 2011 Martin Matuska <mm@FreeBSD.org>
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/txg_impl.h>
29#include <sys/dmu_impl.h>
30#include <sys/dmu_tx.h>
31#include <sys/dsl_pool.h>
32#include <sys/dsl_scan.h>
33#include <sys/callb.h>
34
35/*
36 * ZFS Transaction Groups
37 * ----------------------
38 *
39 * ZFS transaction groups are, as the name implies, groups of transactions
40 * that act on persistent state. ZFS asserts consistency at the granularity of
41 * these transaction groups. Each successive transaction group (txg) is
42 * assigned a 64-bit consecutive identifier. There are three active
43 * transaction group states: open, quiescing, or syncing. At any given time,
44 * there may be an active txg associated with each state; each active txg may
45 * either be processing, or blocked waiting to enter the next state. There may
46 * be up to three active txgs, and there is always a txg in the open state
47 * (though it may be blocked waiting to enter the quiescing state). In broad
48 * strokes, transactions ��� operations that change in-memory structures ��� are
49 * accepted into the txg in the open state, and are completed while the txg is
50 * in the open or quiescing states. The accumulated changes are written to
51 * disk in the syncing state.
52 *
53 * Open
54 *
55 * When a new txg becomes active, it first enters the open state. New
56 * transactions ��� updates to in-memory structures ��� are assigned to the
57 * currently open txg. There is always a txg in the open state so that ZFS can
58 * accept new changes (though the txg may refuse new changes if it has hit
59 * some limit). ZFS advances the open txg to the next state for a variety of
60 * reasons such as it hitting a time or size threshold, or the execution of an
61 * administrative action that must be completed in the syncing state.
62 *
63 * Quiescing
64 *
65 * After a txg exits the open state, it enters the quiescing state. The
66 * quiescing state is intended to provide a buffer between accepting new
67 * transactions in the open state and writing them out to stable storage in
68 * the syncing state. While quiescing, transactions can continue their
69 * operation without delaying either of the other states. Typically, a txg is
70 * in the quiescing state very briefly since the operations are bounded by
71 * software latencies rather than, say, slower I/O latencies. After all
72 * transactions complete, the txg is ready to enter the next state.
73 *
74 * Syncing
75 *
76 * In the syncing state, the in-memory state built up during the open and (to
77 * a lesser degree) the quiescing states is written to stable storage. The
78 * process of writing out modified data can, in turn modify more data. For
79 * example when we write new blocks, we need to allocate space for them; those
80 * allocations modify metadata (space maps)... which themselves must be
81 * written to stable storage. During the sync state, ZFS iterates, writing out
82 * data until it converges and all in-memory changes have been written out.
83 * The first such pass is the largest as it encompasses all the modified user
84 * data (as opposed to filesystem metadata). Subsequent passes typically have
85 * far less data to write as they consist exclusively of filesystem metadata.
86 *
87 * To ensure convergence, after a certain number of passes ZFS begins
88 * overwriting locations on stable storage that had been allocated earlier in
89 * the syncing state (and subsequently freed). ZFS usually allocates new
90 * blocks to optimize for large, continuous, writes. For the syncing state to
91 * converge however it must complete a pass where no new blocks are allocated
92 * since each allocation requires a modification of persistent metadata.
93 * Further, to hasten convergence, after a prescribed number of passes, ZFS
94 * also defers frees, and stops compressing.
95 *
96 * In addition to writing out user data, we must also execute synctasks during
97 * the syncing context. A synctask is the mechanism by which some
98 * administrative activities work such as creating and destroying snapshots or
99 * datasets. Note that when a synctask is initiated it enters the open txg,
100 * and ZFS then pushes that txg as quickly as possible to completion of the
101 * syncing state in order to reduce the latency of the administrative
102 * activity. To complete the syncing state, ZFS writes out a new uberblock,
103 * the root of the tree of blocks that comprise all state stored on the ZFS
104 * pool. Finally, if there is a quiesced txg waiting, we signal that it can
105 * now transition to the syncing state.
106 */
107
108static void txg_sync_thread(void *arg);
109static void txg_quiesce_thread(void *arg);
110
111int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
112
113SYSCTL_DECL(_vfs_zfs);
114SYSCTL_NODE(_vfs_zfs, OID_AUTO, txg, CTLFLAG_RW, 0, "ZFS TXG");
115TUNABLE_INT("vfs.zfs.txg.timeout", &zfs_txg_timeout);
116SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, timeout, CTLFLAG_RW, &zfs_txg_timeout, 0,
117 "Maximum seconds worth of delta per txg");
118
119/*
120 * Prepare the txg subsystem.
121 */
122void
123txg_init(dsl_pool_t *dp, uint64_t txg)
124{
125 tx_state_t *tx = &dp->dp_tx;
126 int c;
127 bzero(tx, sizeof (tx_state_t));
128
129 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
130
131 for (c = 0; c < max_ncpus; c++) {
132 int i;
133
134 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
135 mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT,
136 NULL);
137 for (i = 0; i < TXG_SIZE; i++) {
138 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
139 NULL);
140 list_create(&tx->tx_cpu[c].tc_callbacks[i],
141 sizeof (dmu_tx_callback_t),
142 offsetof(dmu_tx_callback_t, dcb_node));
143 }
144 }
145
146 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
147
148 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
149 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
150 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
151 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
152 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
153
154 tx->tx_open_txg = txg;
155}
156
157/*
158 * Close down the txg subsystem.
159 */
160void
161txg_fini(dsl_pool_t *dp)
162{
163 tx_state_t *tx = &dp->dp_tx;
164 int c;
165
166 ASSERT(tx->tx_threads == 0);
167
168 mutex_destroy(&tx->tx_sync_lock);
169
170 cv_destroy(&tx->tx_sync_more_cv);
171 cv_destroy(&tx->tx_sync_done_cv);
172 cv_destroy(&tx->tx_quiesce_more_cv);
173 cv_destroy(&tx->tx_quiesce_done_cv);
174 cv_destroy(&tx->tx_exit_cv);
175
176 for (c = 0; c < max_ncpus; c++) {
177 int i;
178
179 mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
180 mutex_destroy(&tx->tx_cpu[c].tc_lock);
181 for (i = 0; i < TXG_SIZE; i++) {
182 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
183 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
184 }
185 }
186
187 if (tx->tx_commit_cb_taskq != NULL)
188 taskq_destroy(tx->tx_commit_cb_taskq);
189
190 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
191
192 bzero(tx, sizeof (tx_state_t));
193}
194
195/*
196 * Start syncing transaction groups.
197 */
198void
199txg_sync_start(dsl_pool_t *dp)
200{
201 tx_state_t *tx = &dp->dp_tx;
202
203 mutex_enter(&tx->tx_sync_lock);
204
205 dprintf("pool %p\n", dp);
206
207 ASSERT(tx->tx_threads == 0);
208
209 tx->tx_threads = 2;
210
211 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
212 dp, 0, &p0, TS_RUN, minclsyspri);
213
214 /*
215 * The sync thread can need a larger-than-default stack size on
216 * 32-bit x86. This is due in part to nested pools and
217 * scrub_visitbp() recursion.
218 */
219 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
220 dp, 0, &p0, TS_RUN, minclsyspri);
221
222 mutex_exit(&tx->tx_sync_lock);
223}
224
225static void
226txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
227{
228 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
229 mutex_enter(&tx->tx_sync_lock);
230}
231
232static void
233txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
234{
235 ASSERT(*tpp != NULL);
236 *tpp = NULL;
237 tx->tx_threads--;
238 cv_broadcast(&tx->tx_exit_cv);
239 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
240 thread_exit();
241}
242
243static void
244txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
245{
246 CALLB_CPR_SAFE_BEGIN(cpr);
247
248 if (time)
249 (void) cv_timedwait(cv, &tx->tx_sync_lock, time);
250 else
251 cv_wait(cv, &tx->tx_sync_lock);
252
253 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
254}
255
256/*
257 * Stop syncing transaction groups.
258 */
259void
260txg_sync_stop(dsl_pool_t *dp)
261{
262 tx_state_t *tx = &dp->dp_tx;
263
264 dprintf("pool %p\n", dp);
265 /*
266 * Finish off any work in progress.
267 */
268 ASSERT(tx->tx_threads == 2);
269
270 /*
271 * We need to ensure that we've vacated the deferred space_maps.
272 */
273 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
274
275 /*
276 * Wake all sync threads and wait for them to die.
277 */
278 mutex_enter(&tx->tx_sync_lock);
279
280 ASSERT(tx->tx_threads == 2);
281
282 tx->tx_exiting = 1;
283
284 cv_broadcast(&tx->tx_quiesce_more_cv);
285 cv_broadcast(&tx->tx_quiesce_done_cv);
286 cv_broadcast(&tx->tx_sync_more_cv);
287
288 while (tx->tx_threads != 0)
289 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
290
291 tx->tx_exiting = 0;
292
293 mutex_exit(&tx->tx_sync_lock);
294}
295
296uint64_t
297txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
298{
299 tx_state_t *tx = &dp->dp_tx;
300 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
301 uint64_t txg;
302
303 mutex_enter(&tc->tc_open_lock);
304 txg = tx->tx_open_txg;
305
306 mutex_enter(&tc->tc_lock);
307 tc->tc_count[txg & TXG_MASK]++;
308 mutex_exit(&tc->tc_lock);
309
310 th->th_cpu = tc;
311 th->th_txg = txg;
312
313 return (txg);
314}
315
316void
317txg_rele_to_quiesce(txg_handle_t *th)
318{
319 tx_cpu_t *tc = th->th_cpu;
320
321 ASSERT(!MUTEX_HELD(&tc->tc_lock));
322 mutex_exit(&tc->tc_open_lock);
323}
324
325void
326txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
327{
328 tx_cpu_t *tc = th->th_cpu;
329 int g = th->th_txg & TXG_MASK;
330
331 mutex_enter(&tc->tc_lock);
332 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
333 mutex_exit(&tc->tc_lock);
334}
335
336void
337txg_rele_to_sync(txg_handle_t *th)
338{
339 tx_cpu_t *tc = th->th_cpu;
340 int g = th->th_txg & TXG_MASK;
341
342 mutex_enter(&tc->tc_lock);
343 ASSERT(tc->tc_count[g] != 0);
344 if (--tc->tc_count[g] == 0)
345 cv_broadcast(&tc->tc_cv[g]);
346 mutex_exit(&tc->tc_lock);
347
348 th->th_cpu = NULL; /* defensive */
349}
350
|
411
412 int g = txg & TXG_MASK;
413
414 if (list_is_empty(&tc->tc_callbacks[g]))
415 continue;
416
417 if (tx->tx_commit_cb_taskq == NULL) {
418 /*
419 * Commit callback taskq hasn't been created yet.
420 */
421 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
422 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
423 TASKQ_PREPOPULATE);
424 }
425
426 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
427 list_create(cb_list, sizeof (dmu_tx_callback_t),
428 offsetof(dmu_tx_callback_t, dcb_node));
429
430 list_move_tail(&tc->tc_callbacks[g], cb_list);
431
432 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
433 txg_do_callbacks, cb_list, TQ_SLEEP);
434 }
435}
436
437static void
438txg_sync_thread(void *arg)
439{
440 dsl_pool_t *dp = arg;
441 spa_t *spa = dp->dp_spa;
442 tx_state_t *tx = &dp->dp_tx;
443 callb_cpr_t cpr;
444 uint64_t start, delta;
445
446 txg_thread_enter(tx, &cpr);
447
448 start = delta = 0;
449 for (;;) {
450 uint64_t timer, timeout = zfs_txg_timeout * hz;
451 uint64_t txg;
452
453 /*
454 * We sync when we're scanning, there's someone waiting
455 * on us, or the quiesce thread has handed off a txg to
456 * us, or we have reached our timeout.
457 */
458 timer = (delta >= timeout ? 0 : timeout - delta);
459 while (!dsl_scan_active(dp->dp_scan) &&
460 !tx->tx_exiting && timer > 0 &&
461 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
462 tx->tx_quiesced_txg == 0) {
463 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
464 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
465 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
466 delta = ddi_get_lbolt() - start;
467 timer = (delta > timeout ? 0 : timeout - delta);
468 }
469
470 /*
471 * Wait until the quiesce thread hands off a txg to us,
472 * prompting it to do so if necessary.
473 */
474 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
475 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
476 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
477 cv_broadcast(&tx->tx_quiesce_more_cv);
478 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
479 }
480
481 if (tx->tx_exiting)
482 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
483
484 /*
485 * Consume the quiesced txg which has been handed off to
486 * us. This may cause the quiescing thread to now be
487 * able to quiesce another txg, so we must signal it.
488 */
489 txg = tx->tx_quiesced_txg;
490 tx->tx_quiesced_txg = 0;
491 tx->tx_syncing_txg = txg;
492 cv_broadcast(&tx->tx_quiesce_more_cv);
493
494 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
495 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
496 mutex_exit(&tx->tx_sync_lock);
497
498 start = ddi_get_lbolt();
499 spa_sync(spa, txg);
500 delta = ddi_get_lbolt() - start;
501
502 mutex_enter(&tx->tx_sync_lock);
503 tx->tx_synced_txg = txg;
504 tx->tx_syncing_txg = 0;
505 cv_broadcast(&tx->tx_sync_done_cv);
506
507 /*
508 * Dispatch commit callbacks to worker threads.
509 */
510 txg_dispatch_callbacks(dp, txg);
511 }
512}
513
514static void
515txg_quiesce_thread(void *arg)
516{
517 dsl_pool_t *dp = arg;
518 tx_state_t *tx = &dp->dp_tx;
519 callb_cpr_t cpr;
520
521 txg_thread_enter(tx, &cpr);
522
523 for (;;) {
524 uint64_t txg;
525
526 /*
527 * We quiesce when there's someone waiting on us.
528 * However, we can only have one txg in "quiescing" or
529 * "quiesced, waiting to sync" state. So we wait until
530 * the "quiesced, waiting to sync" txg has been consumed
531 * by the sync thread.
532 */
533 while (!tx->tx_exiting &&
534 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
535 tx->tx_quiesced_txg != 0))
536 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
537
538 if (tx->tx_exiting)
539 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
540
541 txg = tx->tx_open_txg;
542 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
543 txg, tx->tx_quiesce_txg_waiting,
544 tx->tx_sync_txg_waiting);
545 mutex_exit(&tx->tx_sync_lock);
546 txg_quiesce(dp, txg);
547 mutex_enter(&tx->tx_sync_lock);
548
549 /*
550 * Hand this txg off to the sync thread.
551 */
552 dprintf("quiesce done, handing off txg %llu\n", txg);
553 tx->tx_quiesced_txg = txg;
554 cv_broadcast(&tx->tx_sync_more_cv);
555 cv_broadcast(&tx->tx_quiesce_done_cv);
556 }
557}
558
559/*
560 * Delay this thread by 'ticks' if we are still in the open transaction
561 * group and there is already a waiting txg quiesing or quiesced. Abort
562 * the delay if this txg stalls or enters the quiesing state.
563 */
564void
565txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
566{
567 tx_state_t *tx = &dp->dp_tx;
568 clock_t timeout = ddi_get_lbolt() + ticks;
569
570 /* don't delay if this txg could transition to quiesing immediately */
571 if (tx->tx_open_txg > txg ||
572 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
573 return;
574
575 mutex_enter(&tx->tx_sync_lock);
576 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
577 mutex_exit(&tx->tx_sync_lock);
578 return;
579 }
580
581 while (ddi_get_lbolt() < timeout &&
582 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
583 (void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
584 timeout - ddi_get_lbolt());
585
586 mutex_exit(&tx->tx_sync_lock);
587}
588
589void
590txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
591{
592 tx_state_t *tx = &dp->dp_tx;
593
594 ASSERT(!dsl_pool_config_held(dp));
595
596 mutex_enter(&tx->tx_sync_lock);
597 ASSERT(tx->tx_threads == 2);
598 if (txg == 0)
599 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
600 if (tx->tx_sync_txg_waiting < txg)
601 tx->tx_sync_txg_waiting = txg;
602 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
603 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
604 while (tx->tx_synced_txg < txg) {
605 dprintf("broadcasting sync more "
606 "tx_synced=%llu waiting=%llu dp=%p\n",
607 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
608 cv_broadcast(&tx->tx_sync_more_cv);
609 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
610 }
611 mutex_exit(&tx->tx_sync_lock);
612}
613
614void
615txg_wait_open(dsl_pool_t *dp, uint64_t txg)
616{
617 tx_state_t *tx = &dp->dp_tx;
618
619 ASSERT(!dsl_pool_config_held(dp));
620
621 mutex_enter(&tx->tx_sync_lock);
622 ASSERT(tx->tx_threads == 2);
623 if (txg == 0)
624 txg = tx->tx_open_txg + 1;
625 if (tx->tx_quiesce_txg_waiting < txg)
626 tx->tx_quiesce_txg_waiting = txg;
627 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
628 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
629 while (tx->tx_open_txg < txg) {
630 cv_broadcast(&tx->tx_quiesce_more_cv);
631 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
632 }
633 mutex_exit(&tx->tx_sync_lock);
634}
635
636boolean_t
637txg_stalled(dsl_pool_t *dp)
638{
639 tx_state_t *tx = &dp->dp_tx;
640 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
641}
642
643boolean_t
644txg_sync_waiting(dsl_pool_t *dp)
645{
646 tx_state_t *tx = &dp->dp_tx;
647
648 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
649 tx->tx_quiesced_txg != 0);
650}
651
652/*
653 * Per-txg object lists.
654 */
655void
656txg_list_create(txg_list_t *tl, size_t offset)
657{
658 int t;
659
660 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
661
662 tl->tl_offset = offset;
663
664 for (t = 0; t < TXG_SIZE; t++)
665 tl->tl_head[t] = NULL;
666}
667
668void
669txg_list_destroy(txg_list_t *tl)
670{
671 int t;
672
673 for (t = 0; t < TXG_SIZE; t++)
674 ASSERT(txg_list_empty(tl, t));
675
676 mutex_destroy(&tl->tl_lock);
677}
678
679boolean_t
680txg_list_empty(txg_list_t *tl, uint64_t txg)
681{
682 return (tl->tl_head[txg & TXG_MASK] == NULL);
683}
684
685/*
686 * Add an entry to the list (unless it's already on the list).
687 * Returns B_TRUE if it was actually added.
688 */
689boolean_t
690txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
691{
692 int t = txg & TXG_MASK;
693 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
694 boolean_t add;
695
696 mutex_enter(&tl->tl_lock);
697 add = (tn->tn_member[t] == 0);
698 if (add) {
699 tn->tn_member[t] = 1;
700 tn->tn_next[t] = tl->tl_head[t];
701 tl->tl_head[t] = tn;
702 }
703 mutex_exit(&tl->tl_lock);
704
705 return (add);
706}
707
708/*
709 * Add an entry to the end of the list, unless it's already on the list.
710 * (walks list to find end)
711 * Returns B_TRUE if it was actually added.
712 */
713boolean_t
714txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
715{
716 int t = txg & TXG_MASK;
717 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
718 boolean_t add;
719
720 mutex_enter(&tl->tl_lock);
721 add = (tn->tn_member[t] == 0);
722 if (add) {
723 txg_node_t **tp;
724
725 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
726 continue;
727
728 tn->tn_member[t] = 1;
729 tn->tn_next[t] = NULL;
730 *tp = tn;
731 }
732 mutex_exit(&tl->tl_lock);
733
734 return (add);
735}
736
737/*
738 * Remove the head of the list and return it.
739 */
740void *
741txg_list_remove(txg_list_t *tl, uint64_t txg)
742{
743 int t = txg & TXG_MASK;
744 txg_node_t *tn;
745 void *p = NULL;
746
747 mutex_enter(&tl->tl_lock);
748 if ((tn = tl->tl_head[t]) != NULL) {
749 p = (char *)tn - tl->tl_offset;
750 tl->tl_head[t] = tn->tn_next[t];
751 tn->tn_next[t] = NULL;
752 tn->tn_member[t] = 0;
753 }
754 mutex_exit(&tl->tl_lock);
755
756 return (p);
757}
758
759/*
760 * Remove a specific item from the list and return it.
761 */
762void *
763txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
764{
765 int t = txg & TXG_MASK;
766 txg_node_t *tn, **tp;
767
768 mutex_enter(&tl->tl_lock);
769
770 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
771 if ((char *)tn - tl->tl_offset == p) {
772 *tp = tn->tn_next[t];
773 tn->tn_next[t] = NULL;
774 tn->tn_member[t] = 0;
775 mutex_exit(&tl->tl_lock);
776 return (p);
777 }
778 }
779
780 mutex_exit(&tl->tl_lock);
781
782 return (NULL);
783}
784
785boolean_t
786txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
787{
788 int t = txg & TXG_MASK;
789 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
790
791 return (tn->tn_member[t] != 0);
792}
793
794/*
795 * Walk a txg list -- only safe if you know it's not changing.
796 */
797void *
798txg_list_head(txg_list_t *tl, uint64_t txg)
799{
800 int t = txg & TXG_MASK;
801 txg_node_t *tn = tl->tl_head[t];
802
803 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
804}
805
806void *
807txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
808{
809 int t = txg & TXG_MASK;
810 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
811
812 tn = tn->tn_next[t];
813
814 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
815}
| 423
424 int g = txg & TXG_MASK;
425
426 if (list_is_empty(&tc->tc_callbacks[g]))
427 continue;
428
429 if (tx->tx_commit_cb_taskq == NULL) {
430 /*
431 * Commit callback taskq hasn't been created yet.
432 */
433 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
434 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
435 TASKQ_PREPOPULATE);
436 }
437
438 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
439 list_create(cb_list, sizeof (dmu_tx_callback_t),
440 offsetof(dmu_tx_callback_t, dcb_node));
441
442 list_move_tail(&tc->tc_callbacks[g], cb_list);
443
444 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
445 txg_do_callbacks, cb_list, TQ_SLEEP);
446 }
447}
448
449static void
450txg_sync_thread(void *arg)
451{
452 dsl_pool_t *dp = arg;
453 spa_t *spa = dp->dp_spa;
454 tx_state_t *tx = &dp->dp_tx;
455 callb_cpr_t cpr;
456 uint64_t start, delta;
457
458 txg_thread_enter(tx, &cpr);
459
460 start = delta = 0;
461 for (;;) {
462 uint64_t timer, timeout = zfs_txg_timeout * hz;
463 uint64_t txg;
464
465 /*
466 * We sync when we're scanning, there's someone waiting
467 * on us, or the quiesce thread has handed off a txg to
468 * us, or we have reached our timeout.
469 */
470 timer = (delta >= timeout ? 0 : timeout - delta);
471 while (!dsl_scan_active(dp->dp_scan) &&
472 !tx->tx_exiting && timer > 0 &&
473 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
474 tx->tx_quiesced_txg == 0) {
475 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
476 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
477 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
478 delta = ddi_get_lbolt() - start;
479 timer = (delta > timeout ? 0 : timeout - delta);
480 }
481
482 /*
483 * Wait until the quiesce thread hands off a txg to us,
484 * prompting it to do so if necessary.
485 */
486 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
487 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
488 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
489 cv_broadcast(&tx->tx_quiesce_more_cv);
490 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
491 }
492
493 if (tx->tx_exiting)
494 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
495
496 /*
497 * Consume the quiesced txg which has been handed off to
498 * us. This may cause the quiescing thread to now be
499 * able to quiesce another txg, so we must signal it.
500 */
501 txg = tx->tx_quiesced_txg;
502 tx->tx_quiesced_txg = 0;
503 tx->tx_syncing_txg = txg;
504 cv_broadcast(&tx->tx_quiesce_more_cv);
505
506 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
507 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
508 mutex_exit(&tx->tx_sync_lock);
509
510 start = ddi_get_lbolt();
511 spa_sync(spa, txg);
512 delta = ddi_get_lbolt() - start;
513
514 mutex_enter(&tx->tx_sync_lock);
515 tx->tx_synced_txg = txg;
516 tx->tx_syncing_txg = 0;
517 cv_broadcast(&tx->tx_sync_done_cv);
518
519 /*
520 * Dispatch commit callbacks to worker threads.
521 */
522 txg_dispatch_callbacks(dp, txg);
523 }
524}
525
526static void
527txg_quiesce_thread(void *arg)
528{
529 dsl_pool_t *dp = arg;
530 tx_state_t *tx = &dp->dp_tx;
531 callb_cpr_t cpr;
532
533 txg_thread_enter(tx, &cpr);
534
535 for (;;) {
536 uint64_t txg;
537
538 /*
539 * We quiesce when there's someone waiting on us.
540 * However, we can only have one txg in "quiescing" or
541 * "quiesced, waiting to sync" state. So we wait until
542 * the "quiesced, waiting to sync" txg has been consumed
543 * by the sync thread.
544 */
545 while (!tx->tx_exiting &&
546 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
547 tx->tx_quiesced_txg != 0))
548 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
549
550 if (tx->tx_exiting)
551 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
552
553 txg = tx->tx_open_txg;
554 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
555 txg, tx->tx_quiesce_txg_waiting,
556 tx->tx_sync_txg_waiting);
557 mutex_exit(&tx->tx_sync_lock);
558 txg_quiesce(dp, txg);
559 mutex_enter(&tx->tx_sync_lock);
560
561 /*
562 * Hand this txg off to the sync thread.
563 */
564 dprintf("quiesce done, handing off txg %llu\n", txg);
565 tx->tx_quiesced_txg = txg;
566 cv_broadcast(&tx->tx_sync_more_cv);
567 cv_broadcast(&tx->tx_quiesce_done_cv);
568 }
569}
570
571/*
572 * Delay this thread by 'ticks' if we are still in the open transaction
573 * group and there is already a waiting txg quiesing or quiesced. Abort
574 * the delay if this txg stalls or enters the quiesing state.
575 */
576void
577txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
578{
579 tx_state_t *tx = &dp->dp_tx;
580 clock_t timeout = ddi_get_lbolt() + ticks;
581
582 /* don't delay if this txg could transition to quiesing immediately */
583 if (tx->tx_open_txg > txg ||
584 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
585 return;
586
587 mutex_enter(&tx->tx_sync_lock);
588 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
589 mutex_exit(&tx->tx_sync_lock);
590 return;
591 }
592
593 while (ddi_get_lbolt() < timeout &&
594 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
595 (void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
596 timeout - ddi_get_lbolt());
597
598 mutex_exit(&tx->tx_sync_lock);
599}
600
601void
602txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
603{
604 tx_state_t *tx = &dp->dp_tx;
605
606 ASSERT(!dsl_pool_config_held(dp));
607
608 mutex_enter(&tx->tx_sync_lock);
609 ASSERT(tx->tx_threads == 2);
610 if (txg == 0)
611 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
612 if (tx->tx_sync_txg_waiting < txg)
613 tx->tx_sync_txg_waiting = txg;
614 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
615 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
616 while (tx->tx_synced_txg < txg) {
617 dprintf("broadcasting sync more "
618 "tx_synced=%llu waiting=%llu dp=%p\n",
619 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
620 cv_broadcast(&tx->tx_sync_more_cv);
621 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
622 }
623 mutex_exit(&tx->tx_sync_lock);
624}
625
626void
627txg_wait_open(dsl_pool_t *dp, uint64_t txg)
628{
629 tx_state_t *tx = &dp->dp_tx;
630
631 ASSERT(!dsl_pool_config_held(dp));
632
633 mutex_enter(&tx->tx_sync_lock);
634 ASSERT(tx->tx_threads == 2);
635 if (txg == 0)
636 txg = tx->tx_open_txg + 1;
637 if (tx->tx_quiesce_txg_waiting < txg)
638 tx->tx_quiesce_txg_waiting = txg;
639 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
640 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
641 while (tx->tx_open_txg < txg) {
642 cv_broadcast(&tx->tx_quiesce_more_cv);
643 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
644 }
645 mutex_exit(&tx->tx_sync_lock);
646}
647
648boolean_t
649txg_stalled(dsl_pool_t *dp)
650{
651 tx_state_t *tx = &dp->dp_tx;
652 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
653}
654
655boolean_t
656txg_sync_waiting(dsl_pool_t *dp)
657{
658 tx_state_t *tx = &dp->dp_tx;
659
660 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
661 tx->tx_quiesced_txg != 0);
662}
663
664/*
665 * Per-txg object lists.
666 */
667void
668txg_list_create(txg_list_t *tl, size_t offset)
669{
670 int t;
671
672 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
673
674 tl->tl_offset = offset;
675
676 for (t = 0; t < TXG_SIZE; t++)
677 tl->tl_head[t] = NULL;
678}
679
680void
681txg_list_destroy(txg_list_t *tl)
682{
683 int t;
684
685 for (t = 0; t < TXG_SIZE; t++)
686 ASSERT(txg_list_empty(tl, t));
687
688 mutex_destroy(&tl->tl_lock);
689}
690
691boolean_t
692txg_list_empty(txg_list_t *tl, uint64_t txg)
693{
694 return (tl->tl_head[txg & TXG_MASK] == NULL);
695}
696
697/*
698 * Add an entry to the list (unless it's already on the list).
699 * Returns B_TRUE if it was actually added.
700 */
701boolean_t
702txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
703{
704 int t = txg & TXG_MASK;
705 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
706 boolean_t add;
707
708 mutex_enter(&tl->tl_lock);
709 add = (tn->tn_member[t] == 0);
710 if (add) {
711 tn->tn_member[t] = 1;
712 tn->tn_next[t] = tl->tl_head[t];
713 tl->tl_head[t] = tn;
714 }
715 mutex_exit(&tl->tl_lock);
716
717 return (add);
718}
719
720/*
721 * Add an entry to the end of the list, unless it's already on the list.
722 * (walks list to find end)
723 * Returns B_TRUE if it was actually added.
724 */
725boolean_t
726txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
727{
728 int t = txg & TXG_MASK;
729 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
730 boolean_t add;
731
732 mutex_enter(&tl->tl_lock);
733 add = (tn->tn_member[t] == 0);
734 if (add) {
735 txg_node_t **tp;
736
737 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
738 continue;
739
740 tn->tn_member[t] = 1;
741 tn->tn_next[t] = NULL;
742 *tp = tn;
743 }
744 mutex_exit(&tl->tl_lock);
745
746 return (add);
747}
748
749/*
750 * Remove the head of the list and return it.
751 */
752void *
753txg_list_remove(txg_list_t *tl, uint64_t txg)
754{
755 int t = txg & TXG_MASK;
756 txg_node_t *tn;
757 void *p = NULL;
758
759 mutex_enter(&tl->tl_lock);
760 if ((tn = tl->tl_head[t]) != NULL) {
761 p = (char *)tn - tl->tl_offset;
762 tl->tl_head[t] = tn->tn_next[t];
763 tn->tn_next[t] = NULL;
764 tn->tn_member[t] = 0;
765 }
766 mutex_exit(&tl->tl_lock);
767
768 return (p);
769}
770
771/*
772 * Remove a specific item from the list and return it.
773 */
774void *
775txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
776{
777 int t = txg & TXG_MASK;
778 txg_node_t *tn, **tp;
779
780 mutex_enter(&tl->tl_lock);
781
782 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
783 if ((char *)tn - tl->tl_offset == p) {
784 *tp = tn->tn_next[t];
785 tn->tn_next[t] = NULL;
786 tn->tn_member[t] = 0;
787 mutex_exit(&tl->tl_lock);
788 return (p);
789 }
790 }
791
792 mutex_exit(&tl->tl_lock);
793
794 return (NULL);
795}
796
797boolean_t
798txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
799{
800 int t = txg & TXG_MASK;
801 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
802
803 return (tn->tn_member[t] != 0);
804}
805
806/*
807 * Walk a txg list -- only safe if you know it's not changing.
808 */
809void *
810txg_list_head(txg_list_t *tl, uint64_t txg)
811{
812 int t = txg & TXG_MASK;
813 txg_node_t *tn = tl->tl_head[t];
814
815 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
816}
817
818void *
819txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
820{
821 int t = txg & TXG_MASK;
822 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
823
824 tn = tn->tn_next[t];
825
826 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
827}
|