]>
git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/txg.c
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.
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.
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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
25 #include <sys/zfs_context.h>
26 #include <sys/txg_impl.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_scan.h>
31 #include <sys/callb.h>
34 * Pool-wide transaction groups.
37 static void txg_sync_thread(dsl_pool_t
*dp
);
38 static void txg_quiesce_thread(dsl_pool_t
*dp
);
40 int zfs_txg_timeout
= 5; /* max seconds worth of delta per txg */
43 * Prepare the txg subsystem.
46 txg_init(dsl_pool_t
*dp
, uint64_t txg
)
48 tx_state_t
*tx
= &dp
->dp_tx
;
50 bzero(tx
, sizeof (tx_state_t
));
52 tx
->tx_cpu
= kmem_zalloc(max_ncpus
* sizeof (tx_cpu_t
), KM_SLEEP
);
54 for (c
= 0; c
< max_ncpus
; c
++) {
57 mutex_init(&tx
->tx_cpu
[c
].tc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
58 for (i
= 0; i
< TXG_SIZE
; i
++) {
59 cv_init(&tx
->tx_cpu
[c
].tc_cv
[i
], NULL
, CV_DEFAULT
,
61 list_create(&tx
->tx_cpu
[c
].tc_callbacks
[i
],
62 sizeof (dmu_tx_callback_t
),
63 offsetof(dmu_tx_callback_t
, dcb_node
));
67 mutex_init(&tx
->tx_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
69 cv_init(&tx
->tx_sync_more_cv
, NULL
, CV_DEFAULT
, NULL
);
70 cv_init(&tx
->tx_sync_done_cv
, NULL
, CV_DEFAULT
, NULL
);
71 cv_init(&tx
->tx_quiesce_more_cv
, NULL
, CV_DEFAULT
, NULL
);
72 cv_init(&tx
->tx_quiesce_done_cv
, NULL
, CV_DEFAULT
, NULL
);
73 cv_init(&tx
->tx_exit_cv
, NULL
, CV_DEFAULT
, NULL
);
75 tx
->tx_open_txg
= txg
;
79 * Close down the txg subsystem.
82 txg_fini(dsl_pool_t
*dp
)
84 tx_state_t
*tx
= &dp
->dp_tx
;
87 ASSERT(tx
->tx_threads
== 0);
89 mutex_destroy(&tx
->tx_sync_lock
);
91 cv_destroy(&tx
->tx_sync_more_cv
);
92 cv_destroy(&tx
->tx_sync_done_cv
);
93 cv_destroy(&tx
->tx_quiesce_more_cv
);
94 cv_destroy(&tx
->tx_quiesce_done_cv
);
95 cv_destroy(&tx
->tx_exit_cv
);
97 for (c
= 0; c
< max_ncpus
; c
++) {
100 mutex_destroy(&tx
->tx_cpu
[c
].tc_lock
);
101 for (i
= 0; i
< TXG_SIZE
; i
++) {
102 cv_destroy(&tx
->tx_cpu
[c
].tc_cv
[i
]);
103 list_destroy(&tx
->tx_cpu
[c
].tc_callbacks
[i
]);
107 if (tx
->tx_commit_cb_taskq
!= NULL
)
108 taskq_destroy(tx
->tx_commit_cb_taskq
);
110 kmem_free(tx
->tx_cpu
, max_ncpus
* sizeof (tx_cpu_t
));
112 bzero(tx
, sizeof (tx_state_t
));
116 * Start syncing transaction groups.
119 txg_sync_start(dsl_pool_t
*dp
)
121 tx_state_t
*tx
= &dp
->dp_tx
;
123 mutex_enter(&tx
->tx_sync_lock
);
125 dprintf("pool %p\n", dp
);
127 ASSERT(tx
->tx_threads
== 0);
131 tx
->tx_quiesce_thread
= thread_create(NULL
, 0, txg_quiesce_thread
,
132 dp
, 0, &p0
, TS_RUN
, minclsyspri
);
135 * The sync thread can need a larger-than-default stack size on
136 * 32-bit x86. This is due in part to nested pools and
137 * scrub_visitbp() recursion.
139 tx
->tx_sync_thread
= thread_create(NULL
, 32<<10, txg_sync_thread
,
140 dp
, 0, &p0
, TS_RUN
, minclsyspri
);
142 mutex_exit(&tx
->tx_sync_lock
);
146 txg_thread_enter(tx_state_t
*tx
, callb_cpr_t
*cpr
)
148 CALLB_CPR_INIT(cpr
, &tx
->tx_sync_lock
, callb_generic_cpr
, FTAG
);
149 mutex_enter(&tx
->tx_sync_lock
);
153 txg_thread_exit(tx_state_t
*tx
, callb_cpr_t
*cpr
, kthread_t
**tpp
)
155 ASSERT(*tpp
!= NULL
);
158 cv_broadcast(&tx
->tx_exit_cv
);
159 CALLB_CPR_EXIT(cpr
); /* drops &tx->tx_sync_lock */
164 txg_thread_wait(tx_state_t
*tx
, callb_cpr_t
*cpr
, kcondvar_t
*cv
, uint64_t time
)
166 CALLB_CPR_SAFE_BEGIN(cpr
);
169 (void) cv_timedwait(cv
, &tx
->tx_sync_lock
,
170 ddi_get_lbolt() + time
);
172 cv_wait(cv
, &tx
->tx_sync_lock
);
174 CALLB_CPR_SAFE_END(cpr
, &tx
->tx_sync_lock
);
178 * Stop syncing transaction groups.
181 txg_sync_stop(dsl_pool_t
*dp
)
183 tx_state_t
*tx
= &dp
->dp_tx
;
185 dprintf("pool %p\n", dp
);
187 * Finish off any work in progress.
189 ASSERT(tx
->tx_threads
== 2);
192 * We need to ensure that we've vacated the deferred space_maps.
194 txg_wait_synced(dp
, tx
->tx_open_txg
+ TXG_DEFER_SIZE
);
197 * Wake all sync threads and wait for them to die.
199 mutex_enter(&tx
->tx_sync_lock
);
201 ASSERT(tx
->tx_threads
== 2);
205 cv_broadcast(&tx
->tx_quiesce_more_cv
);
206 cv_broadcast(&tx
->tx_quiesce_done_cv
);
207 cv_broadcast(&tx
->tx_sync_more_cv
);
209 while (tx
->tx_threads
!= 0)
210 cv_wait(&tx
->tx_exit_cv
, &tx
->tx_sync_lock
);
214 mutex_exit(&tx
->tx_sync_lock
);
218 txg_hold_open(dsl_pool_t
*dp
, txg_handle_t
*th
)
220 tx_state_t
*tx
= &dp
->dp_tx
;
221 tx_cpu_t
*tc
= &tx
->tx_cpu
[CPU_SEQID
];
224 mutex_enter(&tc
->tc_lock
);
226 txg
= tx
->tx_open_txg
;
227 tc
->tc_count
[txg
& TXG_MASK
]++;
236 txg_rele_to_quiesce(txg_handle_t
*th
)
238 tx_cpu_t
*tc
= th
->th_cpu
;
240 mutex_exit(&tc
->tc_lock
);
244 txg_register_callbacks(txg_handle_t
*th
, list_t
*tx_callbacks
)
246 tx_cpu_t
*tc
= th
->th_cpu
;
247 int g
= th
->th_txg
& TXG_MASK
;
249 mutex_enter(&tc
->tc_lock
);
250 list_move_tail(&tc
->tc_callbacks
[g
], tx_callbacks
);
251 mutex_exit(&tc
->tc_lock
);
255 txg_rele_to_sync(txg_handle_t
*th
)
257 tx_cpu_t
*tc
= th
->th_cpu
;
258 int g
= th
->th_txg
& TXG_MASK
;
260 mutex_enter(&tc
->tc_lock
);
261 ASSERT(tc
->tc_count
[g
] != 0);
262 if (--tc
->tc_count
[g
] == 0)
263 cv_broadcast(&tc
->tc_cv
[g
]);
264 mutex_exit(&tc
->tc_lock
);
266 th
->th_cpu
= NULL
; /* defensive */
270 txg_quiesce(dsl_pool_t
*dp
, uint64_t txg
)
272 tx_state_t
*tx
= &dp
->dp_tx
;
273 int g
= txg
& TXG_MASK
;
277 * Grab all tx_cpu locks so nobody else can get into this txg.
279 for (c
= 0; c
< max_ncpus
; c
++)
280 mutex_enter(&tx
->tx_cpu
[c
].tc_lock
);
282 ASSERT(txg
== tx
->tx_open_txg
);
286 * Now that we've incremented tx_open_txg, we can let threads
287 * enter the next transaction group.
289 for (c
= 0; c
< max_ncpus
; c
++)
290 mutex_exit(&tx
->tx_cpu
[c
].tc_lock
);
293 * Quiesce the transaction group by waiting for everyone to txg_exit().
295 for (c
= 0; c
< max_ncpus
; c
++) {
296 tx_cpu_t
*tc
= &tx
->tx_cpu
[c
];
297 mutex_enter(&tc
->tc_lock
);
298 while (tc
->tc_count
[g
] != 0)
299 cv_wait(&tc
->tc_cv
[g
], &tc
->tc_lock
);
300 mutex_exit(&tc
->tc_lock
);
305 txg_do_callbacks(list_t
*cb_list
)
307 dmu_tx_do_callbacks(cb_list
, 0);
309 list_destroy(cb_list
);
311 kmem_free(cb_list
, sizeof (list_t
));
315 * Dispatch the commit callbacks registered on this txg to worker threads.
318 txg_dispatch_callbacks(dsl_pool_t
*dp
, uint64_t txg
)
321 tx_state_t
*tx
= &dp
->dp_tx
;
324 for (c
= 0; c
< max_ncpus
; c
++) {
325 tx_cpu_t
*tc
= &tx
->tx_cpu
[c
];
326 /* No need to lock tx_cpu_t at this point */
328 int g
= txg
& TXG_MASK
;
330 if (list_is_empty(&tc
->tc_callbacks
[g
]))
333 if (tx
->tx_commit_cb_taskq
== NULL
) {
335 * Commit callback taskq hasn't been created yet.
337 tx
->tx_commit_cb_taskq
= taskq_create("tx_commit_cb",
338 100, minclsyspri
, max_ncpus
, INT_MAX
,
339 TASKQ_THREADS_CPU_PCT
| TASKQ_PREPOPULATE
);
342 cb_list
= kmem_alloc(sizeof (list_t
), KM_SLEEP
);
343 list_create(cb_list
, sizeof (dmu_tx_callback_t
),
344 offsetof(dmu_tx_callback_t
, dcb_node
));
346 list_move_tail(cb_list
, &tc
->tc_callbacks
[g
]);
348 (void) taskq_dispatch(tx
->tx_commit_cb_taskq
, (task_func_t
*)
349 txg_do_callbacks
, cb_list
, TQ_SLEEP
);
354 txg_sync_thread(dsl_pool_t
*dp
)
356 spa_t
*spa
= dp
->dp_spa
;
357 tx_state_t
*tx
= &dp
->dp_tx
;
359 uint64_t start
, delta
;
361 txg_thread_enter(tx
, &cpr
);
365 uint64_t timer
, timeout
= zfs_txg_timeout
* hz
;
369 * We sync when we're scanning, there's someone waiting
370 * on us, or the quiesce thread has handed off a txg to
371 * us, or we have reached our timeout.
373 timer
= (delta
>= timeout
? 0 : timeout
- delta
);
374 while (!dsl_scan_active(dp
->dp_scan
) &&
375 !tx
->tx_exiting
&& timer
> 0 &&
376 tx
->tx_synced_txg
>= tx
->tx_sync_txg_waiting
&&
377 tx
->tx_quiesced_txg
== 0) {
378 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
379 tx
->tx_synced_txg
, tx
->tx_sync_txg_waiting
, dp
);
380 txg_thread_wait(tx
, &cpr
, &tx
->tx_sync_more_cv
, timer
);
381 delta
= ddi_get_lbolt() - start
;
382 timer
= (delta
> timeout
? 0 : timeout
- delta
);
386 * Wait until the quiesce thread hands off a txg to us,
387 * prompting it to do so if necessary.
389 while (!tx
->tx_exiting
&& tx
->tx_quiesced_txg
== 0) {
390 if (tx
->tx_quiesce_txg_waiting
< tx
->tx_open_txg
+1)
391 tx
->tx_quiesce_txg_waiting
= tx
->tx_open_txg
+1;
392 cv_broadcast(&tx
->tx_quiesce_more_cv
);
393 txg_thread_wait(tx
, &cpr
, &tx
->tx_quiesce_done_cv
, 0);
397 txg_thread_exit(tx
, &cpr
, &tx
->tx_sync_thread
);
400 * Consume the quiesced txg which has been handed off to
401 * us. This may cause the quiescing thread to now be
402 * able to quiesce another txg, so we must signal it.
404 txg
= tx
->tx_quiesced_txg
;
405 tx
->tx_quiesced_txg
= 0;
406 tx
->tx_syncing_txg
= txg
;
407 cv_broadcast(&tx
->tx_quiesce_more_cv
);
409 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
410 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
411 mutex_exit(&tx
->tx_sync_lock
);
413 start
= ddi_get_lbolt();
415 delta
= ddi_get_lbolt() - start
;
417 mutex_enter(&tx
->tx_sync_lock
);
418 tx
->tx_synced_txg
= txg
;
419 tx
->tx_syncing_txg
= 0;
420 cv_broadcast(&tx
->tx_sync_done_cv
);
423 * Dispatch commit callbacks to worker threads.
425 txg_dispatch_callbacks(dp
, txg
);
430 txg_quiesce_thread(dsl_pool_t
*dp
)
432 tx_state_t
*tx
= &dp
->dp_tx
;
435 txg_thread_enter(tx
, &cpr
);
441 * We quiesce when there's someone waiting on us.
442 * However, we can only have one txg in "quiescing" or
443 * "quiesced, waiting to sync" state. So we wait until
444 * the "quiesced, waiting to sync" txg has been consumed
445 * by the sync thread.
447 while (!tx
->tx_exiting
&&
448 (tx
->tx_open_txg
>= tx
->tx_quiesce_txg_waiting
||
449 tx
->tx_quiesced_txg
!= 0))
450 txg_thread_wait(tx
, &cpr
, &tx
->tx_quiesce_more_cv
, 0);
453 txg_thread_exit(tx
, &cpr
, &tx
->tx_quiesce_thread
);
455 txg
= tx
->tx_open_txg
;
456 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
457 txg
, tx
->tx_quiesce_txg_waiting
,
458 tx
->tx_sync_txg_waiting
);
459 mutex_exit(&tx
->tx_sync_lock
);
460 txg_quiesce(dp
, txg
);
461 mutex_enter(&tx
->tx_sync_lock
);
464 * Hand this txg off to the sync thread.
466 dprintf("quiesce done, handing off txg %llu\n", txg
);
467 tx
->tx_quiesced_txg
= txg
;
468 cv_broadcast(&tx
->tx_sync_more_cv
);
469 cv_broadcast(&tx
->tx_quiesce_done_cv
);
474 * Delay this thread by 'ticks' if we are still in the open transaction
475 * group and there is already a waiting txg quiesing or quiesced. Abort
476 * the delay if this txg stalls or enters the quiesing state.
479 txg_delay(dsl_pool_t
*dp
, uint64_t txg
, int ticks
)
481 tx_state_t
*tx
= &dp
->dp_tx
;
482 int timeout
= ddi_get_lbolt() + ticks
;
484 /* don't delay if this txg could transition to quiesing immediately */
485 if (tx
->tx_open_txg
> txg
||
486 tx
->tx_syncing_txg
== txg
-1 || tx
->tx_synced_txg
== txg
-1)
489 mutex_enter(&tx
->tx_sync_lock
);
490 if (tx
->tx_open_txg
> txg
|| tx
->tx_synced_txg
== txg
-1) {
491 mutex_exit(&tx
->tx_sync_lock
);
495 while (ddi_get_lbolt() < timeout
&&
496 tx
->tx_syncing_txg
< txg
-1 && !txg_stalled(dp
))
497 (void) cv_timedwait(&tx
->tx_quiesce_more_cv
, &tx
->tx_sync_lock
,
500 mutex_exit(&tx
->tx_sync_lock
);
504 txg_wait_synced(dsl_pool_t
*dp
, uint64_t txg
)
506 tx_state_t
*tx
= &dp
->dp_tx
;
508 mutex_enter(&tx
->tx_sync_lock
);
509 ASSERT(tx
->tx_threads
== 2);
511 txg
= tx
->tx_open_txg
+ TXG_DEFER_SIZE
;
512 if (tx
->tx_sync_txg_waiting
< txg
)
513 tx
->tx_sync_txg_waiting
= txg
;
514 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
515 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
516 while (tx
->tx_synced_txg
< txg
) {
517 dprintf("broadcasting sync more "
518 "tx_synced=%llu waiting=%llu dp=%p\n",
519 tx
->tx_synced_txg
, tx
->tx_sync_txg_waiting
, dp
);
520 cv_broadcast(&tx
->tx_sync_more_cv
);
521 cv_wait(&tx
->tx_sync_done_cv
, &tx
->tx_sync_lock
);
523 mutex_exit(&tx
->tx_sync_lock
);
527 txg_wait_open(dsl_pool_t
*dp
, uint64_t txg
)
529 tx_state_t
*tx
= &dp
->dp_tx
;
531 mutex_enter(&tx
->tx_sync_lock
);
532 ASSERT(tx
->tx_threads
== 2);
534 txg
= tx
->tx_open_txg
+ 1;
535 if (tx
->tx_quiesce_txg_waiting
< txg
)
536 tx
->tx_quiesce_txg_waiting
= txg
;
537 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
538 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
539 while (tx
->tx_open_txg
< txg
) {
540 cv_broadcast(&tx
->tx_quiesce_more_cv
);
541 cv_wait(&tx
->tx_quiesce_done_cv
, &tx
->tx_sync_lock
);
543 mutex_exit(&tx
->tx_sync_lock
);
547 txg_stalled(dsl_pool_t
*dp
)
549 tx_state_t
*tx
= &dp
->dp_tx
;
550 return (tx
->tx_quiesce_txg_waiting
> tx
->tx_open_txg
);
554 txg_sync_waiting(dsl_pool_t
*dp
)
556 tx_state_t
*tx
= &dp
->dp_tx
;
558 return (tx
->tx_syncing_txg
<= tx
->tx_sync_txg_waiting
||
559 tx
->tx_quiesced_txg
!= 0);
563 * Per-txg object lists.
566 txg_list_create(txg_list_t
*tl
, size_t offset
)
570 mutex_init(&tl
->tl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
572 tl
->tl_offset
= offset
;
574 for (t
= 0; t
< TXG_SIZE
; t
++)
575 tl
->tl_head
[t
] = NULL
;
579 txg_list_destroy(txg_list_t
*tl
)
583 for (t
= 0; t
< TXG_SIZE
; t
++)
584 ASSERT(txg_list_empty(tl
, t
));
586 mutex_destroy(&tl
->tl_lock
);
590 txg_list_empty(txg_list_t
*tl
, uint64_t txg
)
592 return (tl
->tl_head
[txg
& TXG_MASK
] == NULL
);
596 * Add an entry to the list.
597 * Returns 0 if it's a new entry, 1 if it's already there.
600 txg_list_add(txg_list_t
*tl
, void *p
, uint64_t txg
)
602 int t
= txg
& TXG_MASK
;
603 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
606 mutex_enter(&tl
->tl_lock
);
607 already_on_list
= tn
->tn_member
[t
];
608 if (!already_on_list
) {
609 tn
->tn_member
[t
] = 1;
610 tn
->tn_next
[t
] = tl
->tl_head
[t
];
613 mutex_exit(&tl
->tl_lock
);
615 return (already_on_list
);
619 * Add an entry to the end of the list (walks list to find end).
620 * Returns 0 if it's a new entry, 1 if it's already there.
623 txg_list_add_tail(txg_list_t
*tl
, void *p
, uint64_t txg
)
625 int t
= txg
& TXG_MASK
;
626 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
629 mutex_enter(&tl
->tl_lock
);
630 already_on_list
= tn
->tn_member
[t
];
631 if (!already_on_list
) {
634 for (tp
= &tl
->tl_head
[t
]; *tp
!= NULL
; tp
= &(*tp
)->tn_next
[t
])
637 tn
->tn_member
[t
] = 1;
638 tn
->tn_next
[t
] = NULL
;
641 mutex_exit(&tl
->tl_lock
);
643 return (already_on_list
);
647 * Remove the head of the list and return it.
650 txg_list_remove(txg_list_t
*tl
, uint64_t txg
)
652 int t
= txg
& TXG_MASK
;
656 mutex_enter(&tl
->tl_lock
);
657 if ((tn
= tl
->tl_head
[t
]) != NULL
) {
658 p
= (char *)tn
- tl
->tl_offset
;
659 tl
->tl_head
[t
] = tn
->tn_next
[t
];
660 tn
->tn_next
[t
] = NULL
;
661 tn
->tn_member
[t
] = 0;
663 mutex_exit(&tl
->tl_lock
);
669 * Remove a specific item from the list and return it.
672 txg_list_remove_this(txg_list_t
*tl
, void *p
, uint64_t txg
)
674 int t
= txg
& TXG_MASK
;
675 txg_node_t
*tn
, **tp
;
677 mutex_enter(&tl
->tl_lock
);
679 for (tp
= &tl
->tl_head
[t
]; (tn
= *tp
) != NULL
; tp
= &tn
->tn_next
[t
]) {
680 if ((char *)tn
- tl
->tl_offset
== p
) {
681 *tp
= tn
->tn_next
[t
];
682 tn
->tn_next
[t
] = NULL
;
683 tn
->tn_member
[t
] = 0;
684 mutex_exit(&tl
->tl_lock
);
689 mutex_exit(&tl
->tl_lock
);
695 txg_list_member(txg_list_t
*tl
, void *p
, uint64_t txg
)
697 int t
= txg
& TXG_MASK
;
698 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
700 return (tn
->tn_member
[t
]);
704 * Walk a txg list -- only safe if you know it's not changing.
707 txg_list_head(txg_list_t
*tl
, uint64_t txg
)
709 int t
= txg
& TXG_MASK
;
710 txg_node_t
*tn
= tl
->tl_head
[t
];
712 return (tn
== NULL
? NULL
: (char *)tn
- tl
->tl_offset
);
716 txg_list_next(txg_list_t
*tl
, void *p
, uint64_t txg
)
718 int t
= txg
& TXG_MASK
;
719 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
723 return (tn
== NULL
? NULL
: (char *)tn
- tl
->tl_offset
);