9  * or https://opensource.org/licenses/CDDL-1.0.
29 #include <sys/time.h>
36  * Multi-Modifier Protection (MMP) attempts to prevent a user from importing
37  * or opening a pool on more than one host at a time.  In particular, it
38  * prevents "zpool import -f" on a host from succeeding while the pool is
40  * device could be used by two hosts for different purposes at the same time
50  * (N-MMP_BLOCKS_PER_LABEL) slots, the remaining slots are reserved for MMP.
65  * mmp_delay field is a decaying average of the amount of time between
76  * some time, and finds the "best" uberblock again.  If any of the mentioned
85  * 1) If uberblock was written by zfs-0.8 or newer and fail_intervals > 0:
91  *    time means either the pool is not imported, or it is imported but the pool
100  * 2) If uberblock was written by zfs-0.8 or newer and fail_intervals == 0:
107  *    enough time for a write to be attempted to each leaf vdev, and mmp_delay
111  * 3) If uberblock was written by zfs-0.7:
121  *    that.  We will have waited enough time for zfs_multihost_import_intervals
137  * 4) Otherwise, this uberblock was written by a pre-MMP zfs:
146  * at the same time and automatically attempt to import the pool.
163  * of zfs_multihost_import_intervals will reduce the import time but increase
164  * the risk of failing to detect an active pool.  The total activity check time
195 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_init()
197 	mutex_init(&mmp->mmp_thread_lock, NULL, MUTEX_DEFAULT, NULL);  in mmp_init()
198 	cv_init(&mmp->mmp_thread_cv, NULL, CV_DEFAULT, NULL);  in mmp_init()
199 	mutex_init(&mmp->mmp_io_lock, NULL, MUTEX_DEFAULT, NULL);  in mmp_init()
200 	mmp->mmp_kstat_id = 1;  in mmp_init()
206 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_fini()
208 	mutex_destroy(&mmp->mmp_thread_lock);  in mmp_fini()
209 	cv_destroy(&mmp->mmp_thread_cv);  in mmp_fini()
210 	mutex_destroy(&mmp->mmp_io_lock);  in mmp_fini()
216 	CALLB_CPR_INIT(cpr, &mmp->mmp_thread_lock, callb_generic_cpr, FTAG);  in mmp_thread_enter()
217 	mutex_enter(&mmp->mmp_thread_lock);  in mmp_thread_enter()
225 	cv_broadcast(&mmp->mmp_thread_cv);  in mmp_thread_exit()
226 	CALLB_CPR_EXIT(cpr);		/* drops &mmp->mmp_thread_lock */  in mmp_thread_exit()
232 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_thread_start()
235 		mutex_enter(&mmp->mmp_thread_lock);  in mmp_thread_start()
236 		if (!mmp->mmp_thread) {  in mmp_thread_start()
237 			mmp->mmp_thread = thread_create(NULL, 0, mmp_thread,  in mmp_thread_start()
242 		mutex_exit(&mmp->mmp_thread_lock);  in mmp_thread_start()
249 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_thread_stop()
251 	mutex_enter(&mmp->mmp_thread_lock);  in mmp_thread_stop()
252 	mmp->mmp_thread_exiting = 1;  in mmp_thread_stop()
253 	cv_broadcast(&mmp->mmp_thread_cv);  in mmp_thread_stop()
255 	while (mmp->mmp_thread) {  in mmp_thread_stop()
256 		cv_wait(&mmp->mmp_thread_cv, &mmp->mmp_thread_lock);  in mmp_thread_stop()
258 	mutex_exit(&mmp->mmp_thread_lock);  in mmp_thread_stop()
262 	ASSERT(mmp->mmp_thread == NULL);  in mmp_thread_stop()
263 	mmp->mmp_thread_exiting = 0;  in mmp_thread_stop()
289 	ASSERT(MUTEX_HELD(&spa->spa_mmp.mmp_io_lock));  in mmp_next_leaf()
291 	ASSERT(list_link_active(&spa->spa_leaf_list.list_head) == B_TRUE);  in mmp_next_leaf()
292 	ASSERT(!list_is_empty(&spa->spa_leaf_list));  in mmp_next_leaf()
294 	if (spa->spa_mmp.mmp_leaf_last_gen != spa->spa_leaf_list_gen) {  in mmp_next_leaf()
295 		spa->spa_mmp.mmp_last_leaf = list_head(&spa->spa_leaf_list);  in mmp_next_leaf()
296 		spa->spa_mmp.mmp_leaf_last_gen = spa->spa_leaf_list_gen;  in mmp_next_leaf()
299 	leaf = spa->spa_mmp.mmp_last_leaf;  in mmp_next_leaf()
301 		leaf = list_head(&spa->spa_leaf_list);  in mmp_next_leaf()
305 		leaf = list_next(&spa->spa_leaf_list, leaf);  in mmp_next_leaf()
307 			leaf = list_head(&spa->spa_leaf_list);  in mmp_next_leaf()
317 		if (!vdev_writeable(leaf) || leaf->vdev_offline ||  in mmp_next_leaf()
318 		    leaf->vdev_detached) {  in mmp_next_leaf()
320 		} else if (leaf->vdev_ops == &vdev_draid_spare_ops) {  in mmp_next_leaf()
322 		} else if (leaf->vdev_mmp_pending != 0) {  in mmp_next_leaf()
325 			spa->spa_mmp.mmp_last_leaf = leaf;  in mmp_next_leaf()
359 	mmp_thread_t *mts = &spa->spa_mmp;  in mmp_delay_update()
360 	hrtime_t delay = gethrtime() - mts->mmp_last_write;  in mmp_delay_update()
362 	ASSERT(MUTEX_HELD(&mts->mmp_io_lock));  in mmp_delay_update()
365 		mts->mmp_delay = 0;  in mmp_delay_update()
369 	if (delay > mts->mmp_delay)  in mmp_delay_update()
370 		mts->mmp_delay = delay;  in mmp_delay_update()
375 	mts->mmp_last_write = gethrtime();  in mmp_delay_update()
380 	if (delay < mts->mmp_delay) {  in mmp_delay_update()
384 		mts->mmp_delay = MAX(((delay + mts->mmp_delay * 127) / 128),  in mmp_delay_update()
392 	spa_t *spa = zio->io_spa;  in mmp_write_done()
393 	vdev_t *vd = zio->io_vd;  in mmp_write_done()
394 	mmp_thread_t *mts = zio->io_private;  in mmp_write_done()
396 	mutex_enter(&mts->mmp_io_lock);  in mmp_write_done()
397 	uint64_t mmp_kstat_id = vd->vdev_mmp_kstat_id;  in mmp_write_done()
398 	hrtime_t mmp_write_duration = gethrtime() - vd->vdev_mmp_pending;  in mmp_write_done()
400 	mmp_delay_update(spa, (zio->io_error == 0));  in mmp_write_done()
402 	vd->vdev_mmp_pending = 0;  in mmp_write_done()
403 	vd->vdev_mmp_kstat_id = 0;  in mmp_write_done()
405 	mutex_exit(&mts->mmp_io_lock);  in mmp_write_done()
408 	spa_mmp_history_set(spa, mmp_kstat_id, zio->io_error,  in mmp_write_done()
411 	abd_free(zio->io_abd);  in mmp_write_done()
415  * When the uberblock on-disk is updated by a spa_sync,
422 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_update_uberblock()
424 	mutex_enter(&mmp->mmp_io_lock);  in mmp_update_uberblock()
425 	mmp->mmp_ub = *ub;  in mmp_update_uberblock()
426 	mmp->mmp_seq = 1;  in mmp_update_uberblock()
427 	mmp->mmp_ub.ub_timestamp = gethrestime_sec();  in mmp_update_uberblock()
429 	mutex_exit(&mmp->mmp_io_lock);  in mmp_update_uberblock()
434  * with a copy of the last-synced uberblock, whose timestamp
441 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_write_uberblock()
449 	lock_acquire_time = gethrtime() - lock_acquire_time;  in mmp_write_uberblock()
451 		zfs_dbgmsg("MMP SCL_STATE acquisition pool '%s' took %llu ns "  in mmp_write_uberblock()
455 	mutex_enter(&mmp->mmp_io_lock);  in mmp_write_uberblock()
461 	 * Issued MMP write: records time issued, error status, etc.  in mmp_write_uberblock()
469 		if (mmp->mmp_skip_error == error) {  in mmp_write_uberblock()
470 			spa_mmp_history_set_skip(spa, mmp->mmp_kstat_id - 1);  in mmp_write_uberblock()
472 			mmp->mmp_skip_error = error;  in mmp_write_uberblock()
473 			spa_mmp_history_add(spa, mmp->mmp_ub.ub_txg,  in mmp_write_uberblock()
474 			    gethrestime_sec(), mmp->mmp_delay, NULL, 0,  in mmp_write_uberblock()
475 			    mmp->mmp_kstat_id++, error);  in mmp_write_uberblock()
480 		mutex_exit(&mmp->mmp_io_lock);  in mmp_write_uberblock()
485 	vd = spa->spa_mmp.mmp_last_leaf;  in mmp_write_uberblock()
486 	if (mmp->mmp_skip_error != 0) {  in mmp_write_uberblock()
487 		mmp->mmp_skip_error = 0;  in mmp_write_uberblock()
491 		    (u_longlong_t)vd->vdev_guid);  in mmp_write_uberblock()
494 	if (mmp->mmp_zio_root == NULL)  in mmp_write_uberblock()
495 		mmp->mmp_zio_root = zio_root(spa, NULL, NULL,  in mmp_write_uberblock()
498 	if (mmp->mmp_ub.ub_timestamp != gethrestime_sec()) {  in mmp_write_uberblock()
504 		mmp->mmp_ub.ub_timestamp = gethrestime_sec();  in mmp_write_uberblock()
505 		mmp->mmp_seq = 1;  in mmp_write_uberblock()
508 	ub = &mmp->mmp_ub;  in mmp_write_uberblock()
509 	ub->ub_mmp_magic = MMP_MAGIC;  in mmp_write_uberblock()
510 	ub->ub_mmp_delay = mmp->mmp_delay;  in mmp_write_uberblock()
511 	ub->ub_mmp_config = MMP_SEQ_SET(mmp->mmp_seq) |  in mmp_write_uberblock()
515 	vd->vdev_mmp_pending = gethrtime();  in mmp_write_uberblock()
516 	vd->vdev_mmp_kstat_id = mmp->mmp_kstat_id;  in mmp_write_uberblock()
518 	zio_t *zio  = zio_null(mmp->mmp_zio_root, spa, NULL, NULL, NULL, flags);  in mmp_write_uberblock()
522 	    VDEV_UBERBLOCK_SIZE(vd) - sizeof (uberblock_t));  in mmp_write_uberblock()
524 	mmp->mmp_seq++;  in mmp_write_uberblock()
525 	mmp->mmp_kstat_id++;  in mmp_write_uberblock()
526 	mutex_exit(&mmp->mmp_io_lock);  in mmp_write_uberblock()
528 	offset = VDEV_UBERBLOCK_OFFSET(vd, VDEV_UBERBLOCK_COUNT(vd) -  in mmp_write_uberblock()
536 	(void) spa_mmp_history_add(spa, ub->ub_txg, ub->ub_timestamp,  in mmp_write_uberblock()
537 	    ub->ub_mmp_delay, vd, label, vd->vdev_mmp_kstat_id, 0);  in mmp_write_uberblock()
546 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_thread()
571 	mutex_enter(&mmp->mmp_io_lock);  in mmp_thread()
572 	mmp->mmp_last_write = gethrtime();  in mmp_thread()
573 	mmp->mmp_delay = MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval));  in mmp_thread()
574 	mutex_exit(&mmp->mmp_io_lock);  in mmp_thread()
576 	while (!mmp->mmp_thread_exiting) {  in mmp_thread()
633 		 * us some time to try.  in mmp_thread()
643 			mutex_enter(&mmp->mmp_io_lock);  in mmp_thread()
644 			mmp->mmp_last_write = gethrtime();  in mmp_thread()
645 			mmp->mmp_delay = mmp_interval;  in mmp_thread()
646 			mutex_exit(&mmp->mmp_io_lock);  in mmp_thread()
654 			mutex_enter(&mmp->mmp_io_lock);  in mmp_thread()
655 			mmp->mmp_delay = 0;  in mmp_thread()
656 			mutex_exit(&mmp->mmp_io_lock);  in mmp_thread()
664 		    (gethrtime() - mmp->mmp_last_write) > mmp_fail_ns) {  in mmp_thread()
669 			    (u_longlong_t)mmp->mmp_last_write,  in mmp_thread()
673 			    (u_longlong_t)spa->spa_uberblock.ub_txg);  in mmp_thread()
678 			    NSEC2MSEC(gethrtime() - mmp->mmp_last_write),  in mmp_thread()
689 			skip_wait--;  in mmp_thread()
693 		(void) cv_timedwait_idle_hires(&mmp->mmp_thread_cv,  in mmp_thread()
694 		    &mmp->mmp_thread_lock, next_time, USEC2NSEC(100),  in mmp_thread()
696 		CALLB_CPR_SAFE_END(&cpr, &mmp->mmp_thread_lock);  in mmp_thread()
700 	zio_wait(mmp->mmp_zio_root);  in mmp_thread()
702 	mmp->mmp_zio_root = NULL;  in mmp_thread()
703 	mmp_thread_exit(mmp, &mmp->mmp_thread, &cpr);  in mmp_thread()
718 	mmp_thread_t *mmp = &spa->spa_mmp;  in mmp_signal_thread()
720 	mutex_enter(&mmp->mmp_thread_lock);  in mmp_signal_thread()
721 	if (mmp->mmp_thread)  in mmp_signal_thread()
722 		cv_broadcast(&mmp->mmp_thread_cv);  in mmp_signal_thread()
723 	mutex_exit(&mmp->mmp_thread_lock);  in mmp_signal_thread()
733 		if (spa->spa_state == POOL_STATE_ACTIVE)  in mmp_signal_all_threads()