xref: /freebsd/sys/contrib/openzfs/module/zfs/mmp.c (revision 8a62a2a5659d1839d8799b4274c04469d7f17c78)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2017 by Lawrence Livermore National Security, LLC.
24  */
25 
26 #include <sys/abd.h>
27 #include <sys/mmp.h>
28 #include <sys/spa.h>
29 #include <sys/spa_impl.h>
30 #include <sys/time.h>
31 #include <sys/vdev.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zfs_context.h>
34 #include <sys/callb.h>
35 
36 /*
37  * Multi-Modifier Protection (MMP) attempts to prevent a user from importing
38  * or opening a pool on more than one host at a time.  In particular, it
39  * prevents "zpool import -f" on a host from succeeding while the pool is
40  * already imported on another host.  There are many other ways in which a
41  * device could be used by two hosts for different purposes at the same time
42  * resulting in pool damage.  This implementation does not attempt to detect
43  * those cases.
44  *
45  * MMP operates by ensuring there are frequent visible changes on disk (a
46  * "heartbeat") at all times.  And by altering the import process to check
47  * for these changes and failing the import when they are detected.  This
48  * functionality is enabled by setting the 'multihost' pool property to on.
49  *
50  * Uberblocks written by the txg_sync thread always go into the first
51  * (N-MMP_BLOCKS_PER_LABEL) slots, the remaining slots are reserved for MMP.
52  * They are used to hold uberblocks which are exactly the same as the last
53  * synced uberblock except that the ub_timestamp and mmp_config are frequently
54  * updated.  Like all other uberblocks, the slot is written with an embedded
55  * checksum, and slots with invalid checksums are ignored.  This provides the
56  * "heartbeat", with no risk of overwriting good uberblocks that must be
57  * preserved, e.g. previous txgs and associated block pointers.
58  *
59  * Three optional fields are added to uberblock structure; ub_mmp_magic,
60  * ub_mmp_config, and ub_mmp_delay.  The ub_mmp_magic value allows zfs to tell
61  * whether the other ub_mmp_* fields are valid.  The ub_mmp_config field tells
62  * the importing host the settings of zfs_multihost_interval and
63  * zfs_multihost_fail_intervals on the host which last had (or currently has)
64  * the pool imported.  These determine how long a host must wait to detect
65  * activity in the pool, before concluding the pool is not in use.  The
66  * mmp_delay field is a decaying average of the amount of time between
67  * completion of successive MMP writes, in nanoseconds.  It indicates whether
68  * MMP is enabled.
69  *
70  * During import an activity test may now be performed to determine if
71  * the pool is in use.  The activity test is typically required if the
72  * ZPOOL_CONFIG_HOSTID does not match the system hostid, the pool state is
73  * POOL_STATE_ACTIVE, and the pool is not a root pool.
74  *
75  * The activity test finds the "best" uberblock (highest txg, timestamp, and, if
76  * ub_mmp_magic is valid, sequence number from ub_mmp_config).  It then waits
77  * some time, and finds the "best" uberblock again.  If any of the mentioned
78  * fields have different values in the newly read uberblock, the pool is in use
79  * by another host and the import fails.  In order to assure the accuracy of the
80  * activity test, the default values result in an activity test duration of 20x
81  * the mmp write interval.
82  *
83  * The duration of the "zpool import" activity test depends on the information
84  * available in the "best" uberblock:
85  *
86  * 1) If uberblock was written by zfs-0.8 or newer and fail_intervals > 0:
87  *    ub_mmp_config.fail_intervals * ub_mmp_config.multihost_interval * 2
88  *
89  *    In this case, a weak guarantee is provided.  Since the host which last had
90  *    the pool imported will suspend the pool if no mmp writes land within
91  *    fail_intervals * multihost_interval ms, the absence of writes during that
92  *    time means either the pool is not imported, or it is imported but the pool
93  *    is suspended and no further writes will occur.
94  *
95  *    Note that resuming the suspended pool on the remote host would invalidate
96  *    this guarantee, and so it is not allowed.
97  *
98  *    The factor of 2 provides a conservative safety factor and derives from
99  *    MMP_IMPORT_SAFETY_FACTOR;
100  *
101  * 2) If uberblock was written by zfs-0.8 or newer and fail_intervals == 0:
102  *    (ub_mmp_config.multihost_interval + ub_mmp_delay) *
103  *        zfs_multihost_import_intervals
104  *
105  *    In this case no guarantee can provided.  However, as long as some devices
106  *    are healthy and connected, it is likely that at least one write will land
107  *    within (multihost_interval + mmp_delay) because multihost_interval is
108  *    enough time for a write to be attempted to each leaf vdev, and mmp_delay
109  *    is enough for one to land, based on past delays.  Multiplying by
110  *    zfs_multihost_import_intervals provides a conservative safety factor.
111  *
112  * 3) If uberblock was written by zfs-0.7:
113  *    (zfs_multihost_interval + ub_mmp_delay) * zfs_multihost_import_intervals
114  *
115  *    The same logic as case #2 applies, but we do not know remote tunables.
116  *
117  *    We use the local value for zfs_multihost_interval because the original MMP
118  *    did not record this value in the uberblock.
119  *
120  *    ub_mmp_delay >= (zfs_multihost_interval / leaves), so if the other host
121  *    has a much larger zfs_multihost_interval set, ub_mmp_delay will reflect
122  *    that.  We will have waited enough time for zfs_multihost_import_intervals
123  *    writes to be issued and all but one to land.
124  *
125  *    single device pool example delays
126  *
127  *    import_delay = (1 + 1) * 20   =  40s #defaults, no I/O delay
128  *    import_delay = (1 + 10) * 20  = 220s #defaults, 10s I/O delay
129  *    import_delay = (10 + 10) * 20 = 400s #10s multihost_interval,
130  *                                          no I/O delay
131  *    100 device pool example delays
132  *
133  *    import_delay = (1 + .01) * 20 =  20s #defaults, no I/O delay
134  *    import_delay = (1 + 10) * 20  = 220s #defaults, 10s I/O delay
135  *    import_delay = (10 + .1) * 20 = 202s #10s multihost_interval,
136  *                                          no I/O delay
137  *
138  * 4) Otherwise, this uberblock was written by a pre-MMP zfs:
139  *    zfs_multihost_import_intervals * zfs_multihost_interval
140  *
141  *    In this case local tunables are used.  By default this product = 10s, long
142  *    enough for a pool with any activity at all to write at least one
143  *    uberblock.  No guarantee can be provided.
144  *
145  * Additionally, the duration is then extended by a random 25% to attempt to to
146  * detect simultaneous imports.  For example, if both partner hosts are rebooted
147  * at the same time and automatically attempt to import the pool.
148  *
149  * Once the read-only activity check completes and the pool is determined to
150  * be inactive a second check is performed to claim the pool.  During this
151  * phase the host writes out MMP uberblocks to each of the devices which are
152  * identical to the best uberblock but with a randomly selected sequence id.
153  * The "best" uberblock is then read back and it must contain this new sequence
154  * number.  This check is performed multiple times to ensure that there is
155  * no window where a concurrently importing system can incorrectly determine
156  * the pool to be inactive.
157  */
158 
159 /*
160  * Used to control the frequency of mmp writes which are performed when the
161  * 'multihost' pool property is on.  This is one factor used to determine the
162  * length of the activity check during import.
163  *
164  * On average an mmp write will be issued for each leaf vdev every
165  * zfs_multihost_interval milliseconds.  In practice, the observed period can
166  * vary with the I/O load and this observed value is the ub_mmp_delay which is
167  * stored in the uberblock.  The minimum allowed value is 100 ms.
168  */
169 uint64_t zfs_multihost_interval = MMP_DEFAULT_INTERVAL;
170 
171 /*
172  * Used to control the duration of the activity test on import.  Smaller values
173  * of zfs_multihost_import_intervals will reduce the import time but increase
174  * the risk of failing to detect an active pool.  The total activity check time
175  * is never allowed to drop below one second.  A value of 0 is ignored and
176  * treated as if it was set to 1.
177  */
178 uint_t zfs_multihost_import_intervals = MMP_DEFAULT_IMPORT_INTERVALS;
179 
180 /*
181  * Controls the behavior of the pool when mmp write failures or delays are
182  * detected.
183  *
184  * When zfs_multihost_fail_intervals = 0, mmp write failures or delays are
185  * ignored.  The failures will still be reported to the ZED which depending on
186  * its configuration may take action such as suspending the pool or taking a
187  * device offline.
188  *
189  * When zfs_multihost_fail_intervals > 0, the pool will be suspended if
190  * zfs_multihost_fail_intervals * zfs_multihost_interval milliseconds pass
191  * without a successful mmp write.  This guarantees the activity test will see
192  * mmp writes if the pool is imported.  A value of 1 is ignored and treated as
193  * if it was set to 2, because a single leaf vdev pool will issue a write once
194  * per multihost_interval and thus any variation in latency would cause the
195  * pool to be suspended.
196  */
197 uint_t zfs_multihost_fail_intervals = MMP_DEFAULT_FAIL_INTERVALS;
198 
199 static const void *const mmp_tag = "mmp_write_uberblock";
200 static __attribute__((noreturn)) void mmp_thread(void *arg);
201 
202 void
mmp_init(spa_t * spa)203 mmp_init(spa_t *spa)
204 {
205 	mmp_thread_t *mmp = &spa->spa_mmp;
206 
207 	mutex_init(&mmp->mmp_thread_lock, NULL, MUTEX_DEFAULT, NULL);
208 	cv_init(&mmp->mmp_thread_cv, NULL, CV_DEFAULT, NULL);
209 	mutex_init(&mmp->mmp_io_lock, NULL, MUTEX_DEFAULT, NULL);
210 	mmp->mmp_kstat_id = 1;
211 }
212 
213 void
mmp_fini(spa_t * spa)214 mmp_fini(spa_t *spa)
215 {
216 	mmp_thread_t *mmp = &spa->spa_mmp;
217 
218 	mutex_destroy(&mmp->mmp_thread_lock);
219 	cv_destroy(&mmp->mmp_thread_cv);
220 	mutex_destroy(&mmp->mmp_io_lock);
221 }
222 
223 static void
mmp_thread_enter(mmp_thread_t * mmp,callb_cpr_t * cpr)224 mmp_thread_enter(mmp_thread_t *mmp, callb_cpr_t *cpr)
225 {
226 	CALLB_CPR_INIT(cpr, &mmp->mmp_thread_lock, callb_generic_cpr, FTAG);
227 	mutex_enter(&mmp->mmp_thread_lock);
228 }
229 
230 static void
mmp_thread_exit(mmp_thread_t * mmp,kthread_t ** mpp,callb_cpr_t * cpr)231 mmp_thread_exit(mmp_thread_t *mmp, kthread_t **mpp, callb_cpr_t *cpr)
232 {
233 	ASSERT(*mpp != NULL);
234 	*mpp = NULL;
235 	cv_broadcast(&mmp->mmp_thread_cv);
236 	CALLB_CPR_EXIT(cpr);		/* drops &mmp->mmp_thread_lock */
237 }
238 
239 void
mmp_thread_start(spa_t * spa)240 mmp_thread_start(spa_t *spa)
241 {
242 	mmp_thread_t *mmp = &spa->spa_mmp;
243 
244 	if (spa_writeable(spa)) {
245 		mutex_enter(&mmp->mmp_thread_lock);
246 		if (!mmp->mmp_thread) {
247 			mmp->mmp_thread = thread_create(NULL, 0, mmp_thread,
248 			    spa, 0, &p0, TS_RUN, defclsyspri);
249 			zfs_dbgmsg("mmp: mmp thread started spa=%s "
250 			    "gethrtime=%llu", spa_name(spa), gethrtime());
251 		}
252 		mutex_exit(&mmp->mmp_thread_lock);
253 	}
254 }
255 
256 void
mmp_thread_stop(spa_t * spa)257 mmp_thread_stop(spa_t *spa)
258 {
259 	mmp_thread_t *mmp = &spa->spa_mmp;
260 
261 	mutex_enter(&mmp->mmp_thread_lock);
262 	mmp->mmp_thread_exiting = 1;
263 	cv_broadcast(&mmp->mmp_thread_cv);
264 
265 	while (mmp->mmp_thread) {
266 		cv_wait(&mmp->mmp_thread_cv, &mmp->mmp_thread_lock);
267 	}
268 	mutex_exit(&mmp->mmp_thread_lock);
269 	zfs_dbgmsg("mmp: mmp thread stopped spa=%s gethrtime=%llu",
270 	    spa_name(spa), gethrtime());
271 
272 	ASSERT0P(mmp->mmp_thread);
273 	mmp->mmp_thread_exiting = 0;
274 }
275 
276 typedef enum mmp_vdev_state_flag {
277 	MMP_FAIL_NOT_WRITABLE	= (1 << 0),
278 	MMP_FAIL_WRITE_PENDING	= (1 << 1),
279 } mmp_vdev_state_flag_t;
280 
281 /*
282  * Find a leaf vdev to write an MMP block to.  It must not have an outstanding
283  * mmp write (if so a new write will also likely block).  If there is no usable
284  * leaf, a nonzero error value is returned. The error value returned is a bit
285  * field.
286  *
287  * MMP_FAIL_WRITE_PENDING   One or more leaf vdevs are writeable, but have an
288  *                          outstanding MMP write.
289  * MMP_FAIL_NOT_WRITABLE    One or more leaf vdevs are not writeable.
290  */
291 
292 static int
mmp_next_leaf(spa_t * spa)293 mmp_next_leaf(spa_t *spa)
294 {
295 	vdev_t *leaf;
296 	vdev_t *starting_leaf;
297 	int fail_mask = 0;
298 
299 	ASSERT(MUTEX_HELD(&spa->spa_mmp.mmp_io_lock));
300 	ASSERT(spa_config_held(spa, SCL_STATE, RW_READER));
301 	ASSERT(list_link_active(&spa->spa_leaf_list.list_head) == B_TRUE);
302 	ASSERT(!list_is_empty(&spa->spa_leaf_list));
303 
304 	if (spa->spa_mmp.mmp_leaf_last_gen != spa->spa_leaf_list_gen) {
305 		spa->spa_mmp.mmp_last_leaf = list_head(&spa->spa_leaf_list);
306 		spa->spa_mmp.mmp_leaf_last_gen = spa->spa_leaf_list_gen;
307 	}
308 
309 	leaf = spa->spa_mmp.mmp_last_leaf;
310 	if (leaf == NULL)
311 		leaf = list_head(&spa->spa_leaf_list);
312 	starting_leaf = leaf;
313 
314 	do {
315 		leaf = list_next(&spa->spa_leaf_list, leaf);
316 		if (leaf == NULL) {
317 			leaf = list_head(&spa->spa_leaf_list);
318 			ASSERT3P(leaf, !=, NULL);
319 		}
320 
321 		/*
322 		 * We skip unwritable, offline, detached, and dRAID spare
323 		 * devices as they are either not legal targets or the write
324 		 * may fail or not be seen by other hosts.  Skipped dRAID
325 		 * spares can never be written so the fail mask is not set.
326 		 */
327 		if (!vdev_writeable(leaf) || leaf->vdev_offline ||
328 		    leaf->vdev_detached) {
329 			fail_mask |= MMP_FAIL_NOT_WRITABLE;
330 		} else if (leaf->vdev_ops == &vdev_draid_spare_ops) {
331 			continue;
332 		} else if (leaf->vdev_mmp_pending != 0) {
333 			fail_mask |= MMP_FAIL_WRITE_PENDING;
334 		} else {
335 			spa->spa_mmp.mmp_last_leaf = leaf;
336 			return (0);
337 		}
338 	} while (leaf != starting_leaf);
339 
340 	ASSERT(fail_mask);
341 
342 	return (fail_mask);
343 }
344 
345 /*
346  * MMP writes are issued on a fixed schedule, but may complete at variable,
347  * much longer, intervals.  The mmp_delay captures long periods between
348  * successful writes for any reason, including disk latency, scheduling delays,
349  * etc.
350  *
351  * The mmp_delay is usually calculated as a decaying average, but if the latest
352  * delay is higher we do not average it, so that we do not hide sudden spikes
353  * which the importing host must wait for.
354  *
355  * If writes are occurring frequently, such as due to a high rate of txg syncs,
356  * the mmp_delay could become very small.  Since those short delays depend on
357  * activity we cannot count on, we never allow mmp_delay to get lower than rate
358  * expected if only mmp_thread writes occur.
359  *
360  * If an mmp write was skipped or fails, and we have already waited longer than
361  * mmp_delay, we need to update it so the next write reflects the longer delay.
362  *
363  * Do not set mmp_delay if the multihost property is not on, so as not to
364  * trigger an activity check on import.
365  */
366 static void
mmp_delay_update(spa_t * spa,boolean_t write_completed)367 mmp_delay_update(spa_t *spa, boolean_t write_completed)
368 {
369 	mmp_thread_t *mts = &spa->spa_mmp;
370 	hrtime_t delay = gethrtime() - mts->mmp_last_write;
371 
372 	ASSERT(MUTEX_HELD(&mts->mmp_io_lock));
373 
374 	if (spa_multihost(spa) == B_FALSE) {
375 		mts->mmp_delay = 0;
376 		return;
377 	}
378 
379 	if (delay > mts->mmp_delay)
380 		mts->mmp_delay = delay;
381 
382 	if (write_completed == B_FALSE)
383 		return;
384 
385 	mts->mmp_last_write = gethrtime();
386 
387 	/*
388 	 * strictly less than, in case delay was changed above.
389 	 */
390 	if (delay < mts->mmp_delay) {
391 		hrtime_t min_delay =
392 		    MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval)) /
393 		    MAX(1, vdev_count_leaves(spa));
394 		mts->mmp_delay = MAX(((delay + mts->mmp_delay * 127) / 128),
395 		    min_delay);
396 	}
397 }
398 
399 static void
mmp_write_done(zio_t * zio)400 mmp_write_done(zio_t *zio)
401 {
402 	spa_t *spa = zio->io_spa;
403 	vdev_t *vd = zio->io_vd;
404 	mmp_thread_t *mts = zio->io_private;
405 
406 	mutex_enter(&mts->mmp_io_lock);
407 	uint64_t mmp_kstat_id = vd->vdev_mmp_kstat_id;
408 	hrtime_t mmp_write_duration = gethrtime() - vd->vdev_mmp_pending;
409 
410 	mmp_delay_update(spa, (zio->io_error == 0));
411 
412 	vd->vdev_mmp_pending = 0;
413 	vd->vdev_mmp_kstat_id = 0;
414 
415 	mutex_exit(&mts->mmp_io_lock);
416 	spa_config_exit(spa, SCL_STATE, mmp_tag);
417 
418 	spa_mmp_history_set(spa, mmp_kstat_id, zio->io_error,
419 	    mmp_write_duration);
420 
421 	abd_free(zio->io_abd);
422 }
423 
424 /*
425  * When the uberblock on-disk is updated by a spa_sync,
426  * creating a new "best" uberblock, update the one stored
427  * in the mmp thread state, used for mmp writes.
428  */
429 void
mmp_update_uberblock(spa_t * spa,uberblock_t * ub)430 mmp_update_uberblock(spa_t *spa, uberblock_t *ub)
431 {
432 	mmp_thread_t *mmp = &spa->spa_mmp;
433 
434 	mutex_enter(&mmp->mmp_io_lock);
435 	mmp->mmp_ub = *ub;
436 	mmp->mmp_seq = 1;
437 	mmp->mmp_ub.ub_timestamp = gethrestime_sec();
438 	mmp_delay_update(spa, B_TRUE);
439 	mutex_exit(&mmp->mmp_io_lock);
440 }
441 
442 /*
443  * Choose a random vdev, label, and MMP block, and write over it
444  * with a copy of the last-synced uberblock, whose timestamp
445  * has been updated to reflect that the pool is in use.
446  */
447 static void
mmp_write_uberblock(spa_t * spa)448 mmp_write_uberblock(spa_t *spa)
449 {
450 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
451 	mmp_thread_t *mmp = &spa->spa_mmp;
452 	uberblock_t *ub;
453 	vdev_t *vd = NULL;
454 	int label, error;
455 	uint64_t offset;
456 
457 	hrtime_t lock_acquire_time = gethrtime();
458 	spa_config_enter_priority(spa, SCL_STATE, mmp_tag, RW_READER);
459 	lock_acquire_time = gethrtime() - lock_acquire_time;
460 	if (lock_acquire_time > (MSEC2NSEC(MMP_MIN_INTERVAL) / 10))
461 		zfs_dbgmsg("mmp: long SCL_STATE acquisition, spa=%s "
462 		    "acquire_time=%llu gethrtime=%llu", spa_name(spa),
463 		    lock_acquire_time, gethrtime());
464 
465 	mutex_enter(&mmp->mmp_io_lock);
466 
467 	error = mmp_next_leaf(spa);
468 
469 	/*
470 	 * spa_mmp_history has two types of entries:
471 	 * Issued MMP write: records time issued, error status, etc.
472 	 * Skipped MMP write: an MMP write could not be issued because no
473 	 * suitable leaf vdev was available.  See comment above struct
474 	 * spa_mmp_history for details.
475 	 */
476 
477 	if (error) {
478 		mmp_delay_update(spa, B_FALSE);
479 		if (mmp->mmp_skip_error == error) {
480 			spa_mmp_history_set_skip(spa, mmp->mmp_kstat_id - 1);
481 		} else {
482 			mmp->mmp_skip_error = error;
483 			spa_mmp_history_add(spa, mmp->mmp_ub.ub_txg,
484 			    gethrestime_sec(), mmp->mmp_delay, NULL, 0,
485 			    mmp->mmp_kstat_id++, error);
486 			zfs_dbgmsg("mmp: error choosing leaf, spa=%s "
487 			    "gethrtime=%llu fail_mask=%#x", spa_name(spa),
488 			    gethrtime(), error);
489 		}
490 		mutex_exit(&mmp->mmp_io_lock);
491 		spa_config_exit(spa, SCL_STATE, mmp_tag);
492 		return;
493 	}
494 
495 	vd = spa->spa_mmp.mmp_last_leaf;
496 	if (mmp->mmp_skip_error != 0) {
497 		zfs_dbgmsg("mmp: write after skipping due to unavailable "
498 		    "leaves, spa=%s gethrtime=%llu vdev=%llu error=%d",
499 		    spa_name(spa), (u_longlong_t)gethrtime(),
500 		    (u_longlong_t)vd->vdev_guid, mmp->mmp_skip_error);
501 		mmp->mmp_skip_error = 0;
502 	}
503 
504 	if (mmp->mmp_zio_root == NULL)
505 		mmp->mmp_zio_root = zio_root(spa, NULL, NULL,
506 		    flags | ZIO_FLAG_GODFATHER);
507 
508 	if (mmp->mmp_ub.ub_timestamp != gethrestime_sec()) {
509 		/*
510 		 * Want to reset mmp_seq when timestamp advances because after
511 		 * an mmp_seq wrap new values will not be chosen by
512 		 * uberblock_compare() as the "best".
513 		 */
514 		mmp->mmp_ub.ub_timestamp = gethrestime_sec();
515 		mmp->mmp_seq = 1;
516 	}
517 
518 	ub = &mmp->mmp_ub;
519 	ub->ub_mmp_magic = MMP_MAGIC;
520 	ub->ub_mmp_delay = mmp->mmp_delay;
521 	ub->ub_mmp_config = MMP_SEQ_SET(mmp->mmp_seq) |
522 	    MMP_INTERVAL_SET(MMP_INTERVAL_OK(zfs_multihost_interval)) |
523 	    MMP_FAIL_INT_SET(MMP_FAIL_INTVS_OK(
524 	    zfs_multihost_fail_intervals));
525 	vd->vdev_mmp_pending = gethrtime();
526 	vd->vdev_mmp_kstat_id = mmp->mmp_kstat_id;
527 
528 	zio_t *zio  = zio_null(mmp->mmp_zio_root, spa, NULL, NULL, NULL, flags);
529 	abd_t *ub_abd = abd_alloc_for_io(VDEV_UBERBLOCK_SIZE(vd), B_TRUE);
530 	abd_copy_from_buf(ub_abd, ub, sizeof (uberblock_t));
531 	abd_zero_off(ub_abd, sizeof (uberblock_t),
532 	    VDEV_UBERBLOCK_SIZE(vd) - sizeof (uberblock_t));
533 
534 	mmp->mmp_seq++;
535 	mmp->mmp_kstat_id++;
536 	mutex_exit(&mmp->mmp_io_lock);
537 
538 	offset = VDEV_UBERBLOCK_OFFSET(vd, VDEV_UBERBLOCK_COUNT(vd) -
539 	    MMP_BLOCKS_PER_LABEL + random_in_range(MMP_BLOCKS_PER_LABEL));
540 
541 	label = random_in_range(VDEV_LABELS);
542 	vdev_label_write(zio, vd, label, ub_abd, offset,
543 	    VDEV_UBERBLOCK_SIZE(vd), mmp_write_done, mmp,
544 	    flags | ZIO_FLAG_DONT_PROPAGATE);
545 
546 	(void) spa_mmp_history_add(spa, ub->ub_txg, ub->ub_timestamp,
547 	    ub->ub_mmp_delay, vd, label, vd->vdev_mmp_kstat_id, 0);
548 
549 	zio_nowait(zio);
550 }
551 
552 static void
mmp_claim_uberblock_sync_done(zio_t * zio)553 mmp_claim_uberblock_sync_done(zio_t *zio)
554 {
555 	uint64_t *good_writes = zio->io_private;
556 
557 	if (zio->io_error == 0 && zio->io_vd->vdev_top->vdev_ms_array != 0)
558 		atomic_inc_64(good_writes);
559 }
560 
561 /*
562  * Write the uberblock to the first label of all leaves of the specified vdev.
563  * Two writes required for each mirror, one for a singleton, and parity+1 for
564  * raidz or draid vdevs.
565  */
566 static void
mmp_claim_uberblock_sync(zio_t * zio,uint64_t * good_writes,uint64_t * req_writes,uberblock_t * ub,vdev_t * vd,int flags)567 mmp_claim_uberblock_sync(zio_t *zio, uint64_t *good_writes,
568     uint64_t *req_writes, uberblock_t *ub, vdev_t *vd, int flags)
569 {
570 	for (uint64_t c = 0; c < vd->vdev_children; c++) {
571 		vdev_t *cvd = vd->vdev_child[c];
572 
573 		if (cvd->vdev_islog || cvd->vdev_isspare || cvd->vdev_isl2cache)
574 			continue;
575 
576 		if (cvd->vdev_top == cvd) {
577 			uint64_t nparity = vdev_get_nparity(cvd);
578 			if (nparity) {
579 				*req_writes += nparity + 1;
580 			} else {
581 				*req_writes +=
582 				    MIN(MAX(cvd->vdev_children, 1), 2);
583 			}
584 		}
585 
586 		mmp_claim_uberblock_sync(zio, good_writes, req_writes,
587 		    ub, cvd, flags);
588 	}
589 
590 	if (!vd->vdev_ops->vdev_op_leaf)
591 		return;
592 
593 	if (!vdev_writeable(vd))
594 		return;
595 
596 	if (vd->vdev_ops == &vdev_draid_spare_ops)
597 		return;
598 
599 	abd_t *ub_abd = abd_alloc_for_io(VDEV_UBERBLOCK_SIZE(vd), B_TRUE);
600 	abd_copy_from_buf(ub_abd, ub, sizeof (uberblock_t));
601 	abd_zero_off(ub_abd, sizeof (uberblock_t),
602 	    VDEV_UBERBLOCK_SIZE(vd) - sizeof (uberblock_t));
603 
604 	vdev_label_write(zio, vd, 0, ub_abd,
605 	    VDEV_UBERBLOCK_OFFSET(vd, VDEV_UBERBLOCK_COUNT(vd) -
606 	    MMP_BLOCKS_PER_LABEL), VDEV_UBERBLOCK_SIZE(vd),
607 	    mmp_claim_uberblock_sync_done, good_writes,
608 	    flags | ZIO_FLAG_DONT_PROPAGATE);
609 
610 	abd_free(ub_abd);
611 }
612 
613 int
mmp_claim_uberblock(spa_t * spa,vdev_t * vd,uberblock_t * ub)614 mmp_claim_uberblock(spa_t *spa, vdev_t *vd, uberblock_t *ub)
615 {
616 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
617 	uint64_t good_writes = 0;
618 	uint64_t req_writes = 0;
619 	zio_t *zio;
620 
621 	ASSERT(MMP_VALID(ub));
622 	ASSERT(MMP_SEQ_VALID(ub));
623 
624 	spa_config_enter(spa, SCL_ALL, mmp_tag, RW_WRITER);
625 
626 	/* Sync the uberblock to all writeable leaves */
627 	zio = zio_root(spa, NULL, NULL, flags);
628 	mmp_claim_uberblock_sync(zio, &good_writes, &req_writes, ub, vd, flags);
629 	(void) zio_wait(zio);
630 
631 	/* Flush the new uberblocks so they're immediately visible */
632 	zio = zio_root(spa, NULL, NULL, flags);
633 	zio_flush(zio, vd);
634 	(void) zio_wait(zio);
635 
636 	spa_config_exit(spa, SCL_ALL, mmp_tag);
637 
638 	zfs_dbgmsg("mmp: claiming uberblock, spa=%s txg=%llu seq=%llu "
639 	    "req_writes=%llu good_writes=%llu", spa_load_name(spa),
640 	    (u_longlong_t)ub->ub_txg, (u_longlong_t)MMP_SEQ(ub),
641 	    (u_longlong_t)req_writes, (u_longlong_t)good_writes);
642 
643 	/*
644 	 * To guarantee visibility from a remote host we require a minimum
645 	 * number of good writes. For raidz/draid vdevs parity+1 writes, for
646 	 * mirrors 2 writes, and for singletons 1 write.
647 	 */
648 	if (req_writes == 0 || good_writes < req_writes)
649 		return (SET_ERROR(EIO));
650 
651 	return (0);
652 }
653 
654 static __attribute__((noreturn)) void
mmp_thread(void * arg)655 mmp_thread(void *arg)
656 {
657 	spa_t *spa = (spa_t *)arg;
658 	mmp_thread_t *mmp = &spa->spa_mmp;
659 	boolean_t suspended = spa_suspended(spa);
660 	boolean_t multihost = spa_multihost(spa);
661 	uint64_t mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK(
662 	    zfs_multihost_interval));
663 	uint32_t mmp_fail_intervals = MMP_FAIL_INTVS_OK(
664 	    zfs_multihost_fail_intervals);
665 	hrtime_t mmp_fail_ns = mmp_fail_intervals * mmp_interval;
666 	boolean_t last_spa_suspended;
667 	boolean_t last_spa_multihost;
668 	uint64_t last_mmp_interval;
669 	uint32_t last_mmp_fail_intervals;
670 	hrtime_t last_mmp_fail_ns;
671 	callb_cpr_t cpr;
672 	int skip_wait = 0;
673 
674 	mmp_thread_enter(mmp, &cpr);
675 
676 	/*
677 	 * There have been no MMP writes yet.  Setting mmp_last_write here gives
678 	 * us one mmp_fail_ns period, which is consistent with the activity
679 	 * check duration, to try to land an MMP write before MMP suspends the
680 	 * pool (if so configured).
681 	 */
682 
683 	mutex_enter(&mmp->mmp_io_lock);
684 	mmp->mmp_last_write = gethrtime();
685 	mmp->mmp_delay = MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval));
686 	mutex_exit(&mmp->mmp_io_lock);
687 
688 	while (!mmp->mmp_thread_exiting) {
689 		hrtime_t next_time = gethrtime() +
690 		    MSEC2NSEC(MMP_DEFAULT_INTERVAL);
691 		int leaves = MAX(vdev_count_leaves(spa), 1);
692 
693 		/* Detect changes in tunables or state */
694 
695 		last_spa_suspended = suspended;
696 		last_spa_multihost = multihost;
697 		suspended = spa_suspended(spa);
698 		multihost = spa_multihost(spa);
699 
700 		last_mmp_interval = mmp_interval;
701 		last_mmp_fail_intervals = mmp_fail_intervals;
702 		last_mmp_fail_ns = mmp_fail_ns;
703 		mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK(
704 		    zfs_multihost_interval));
705 		mmp_fail_intervals = MMP_FAIL_INTVS_OK(
706 		    zfs_multihost_fail_intervals);
707 
708 		/* Smooth so pool is not suspended when reducing tunables */
709 		if (mmp_fail_intervals * mmp_interval < mmp_fail_ns) {
710 			mmp_fail_ns = (mmp_fail_ns * 31 +
711 			    mmp_fail_intervals * mmp_interval) / 32;
712 		} else {
713 			mmp_fail_ns = mmp_fail_intervals *
714 			    mmp_interval;
715 		}
716 
717 		if (mmp_interval != last_mmp_interval ||
718 		    mmp_fail_intervals != last_mmp_fail_intervals) {
719 			/*
720 			 * We want other hosts to see new tunables as quickly as
721 			 * possible.  Write out at higher frequency than usual.
722 			 */
723 			skip_wait += leaves;
724 		}
725 
726 		if (multihost)
727 			next_time = gethrtime() + mmp_interval / leaves;
728 
729 		if (mmp_fail_ns != last_mmp_fail_ns) {
730 			zfs_dbgmsg("mmp: interval change, spa=%s "
731 			    "gethrtime=%llu last_mmp_interval=%llu "
732 			    "mmp_interval=%llu last_mmp_fail_intervals=%u "
733 			    "mmp_fail_intervals=%u mmp_fail_ns=%llu "
734 			    "skip_wait=%d leaves=%d next_time=%llu",
735 			    spa_name(spa), (u_longlong_t)gethrtime(),
736 			    (u_longlong_t)last_mmp_interval,
737 			    (u_longlong_t)mmp_interval, last_mmp_fail_intervals,
738 			    mmp_fail_intervals, (u_longlong_t)mmp_fail_ns,
739 			    skip_wait, leaves, (u_longlong_t)next_time);
740 		}
741 
742 		/*
743 		 * MMP off => on, or suspended => !suspended:
744 		 * No writes occurred recently.  Update mmp_last_write to give
745 		 * us some time to try.
746 		 */
747 		if ((!last_spa_multihost && multihost) ||
748 		    (last_spa_suspended && !suspended)) {
749 			zfs_dbgmsg("mmp: state change spa=%s: gethrtime=%llu "
750 			    "last_spa_multihost=%u multihost=%u "
751 			    "last_spa_suspended=%u suspended=%u",
752 			    spa_name(spa), (u_longlong_t)gethrtime(),
753 			    last_spa_multihost, multihost, last_spa_suspended,
754 			    suspended);
755 			mutex_enter(&mmp->mmp_io_lock);
756 			mmp->mmp_last_write = gethrtime();
757 			mmp->mmp_delay = mmp_interval;
758 			mutex_exit(&mmp->mmp_io_lock);
759 		}
760 
761 		/*
762 		 * MMP on => off:
763 		 * mmp_delay == 0 tells importing node to skip activity check.
764 		 */
765 		if (last_spa_multihost && !multihost) {
766 			mutex_enter(&mmp->mmp_io_lock);
767 			mmp->mmp_delay = 0;
768 			mutex_exit(&mmp->mmp_io_lock);
769 		}
770 
771 		/*
772 		 * Suspend the pool if no MMP write has succeeded in over
773 		 * mmp_interval * mmp_fail_intervals nanoseconds.
774 		 */
775 		if (multihost && !suspended && mmp_fail_intervals &&
776 		    (gethrtime() - mmp->mmp_last_write) > mmp_fail_ns) {
777 			zfs_dbgmsg("mmp: suspending pool, spa=%s "
778 			    "gethrtime=%llu mmp_last_write=%llu "
779 			    "mmp_interval=%llu mmp_fail_intervals=%llu "
780 			    "mmp_fail_ns=%llu txg=%llu",
781 			    spa_name(spa), (u_longlong_t)gethrtime(),
782 			    (u_longlong_t)mmp->mmp_last_write,
783 			    (u_longlong_t)mmp_interval,
784 			    (u_longlong_t)mmp_fail_intervals,
785 			    (u_longlong_t)mmp_fail_ns,
786 			    (u_longlong_t)spa->spa_uberblock.ub_txg);
787 			cmn_err(CE_WARN, "MMP writes to pool '%s' have not "
788 			    "succeeded in over %llu ms; suspending pool. "
789 			    "Hrtime %llu",
790 			    spa_name(spa),
791 			    NSEC2MSEC(gethrtime() - mmp->mmp_last_write),
792 			    gethrtime());
793 			zio_suspend(spa, NULL, ZIO_SUSPEND_MMP);
794 		}
795 
796 		if (multihost && !suspended)
797 			mmp_write_uberblock(spa);
798 
799 		if (skip_wait > 0) {
800 			next_time = gethrtime() + MSEC2NSEC(MMP_MIN_INTERVAL) /
801 			    leaves;
802 			skip_wait--;
803 		}
804 
805 		CALLB_CPR_SAFE_BEGIN(&cpr);
806 		(void) cv_timedwait_idle_hires(&mmp->mmp_thread_cv,
807 		    &mmp->mmp_thread_lock, next_time, USEC2NSEC(100),
808 		    CALLOUT_FLAG_ABSOLUTE);
809 		CALLB_CPR_SAFE_END(&cpr, &mmp->mmp_thread_lock);
810 	}
811 
812 	/* Outstanding writes are allowed to complete. */
813 	zio_wait(mmp->mmp_zio_root);
814 
815 	mmp->mmp_zio_root = NULL;
816 	mmp_thread_exit(mmp, &mmp->mmp_thread, &cpr);
817 
818 	thread_exit();
819 }
820 
821 /*
822  * Signal the MMP thread to wake it, when it is sleeping on
823  * its cv.  Used when some module parameter has changed and
824  * we want the thread to know about it.
825  * Only signal if the pool is active and mmp thread is
826  * running, otherwise there is no thread to wake.
827  */
828 static void
mmp_signal_thread(spa_t * spa)829 mmp_signal_thread(spa_t *spa)
830 {
831 	mmp_thread_t *mmp = &spa->spa_mmp;
832 
833 	mutex_enter(&mmp->mmp_thread_lock);
834 	if (mmp->mmp_thread)
835 		cv_broadcast(&mmp->mmp_thread_cv);
836 	mutex_exit(&mmp->mmp_thread_lock);
837 }
838 
839 void
mmp_signal_all_threads(void)840 mmp_signal_all_threads(void)
841 {
842 	spa_t *spa = NULL;
843 
844 	spa_namespace_enter(FTAG);
845 	while ((spa = spa_next(spa))) {
846 		if (spa->spa_state == POOL_STATE_ACTIVE)
847 			mmp_signal_thread(spa);
848 	}
849 	spa_namespace_exit(FTAG);
850 }
851 
852 ZFS_MODULE_PARAM_CALL(zfs_multihost, zfs_multihost_, interval,
853 	param_set_multihost_interval, spl_param_get_u64, ZMOD_RW,
854 	"Milliseconds between mmp writes to each leaf");
855 
856 ZFS_MODULE_PARAM(zfs_multihost, zfs_multihost_, fail_intervals, UINT, ZMOD_RW,
857 	"Max allowed period without a successful mmp write");
858 
859 ZFS_MODULE_PARAM(zfs_multihost, zfs_multihost_, import_intervals, UINT, ZMOD_RW,
860 	"Number of zfs_multihost_interval periods to wait for activity");
861