summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--cmd/zdb/zdb.c15
-rw-r--r--include/sys/fs/zfs.h1
-rw-r--r--include/sys/mmp.h6
-rw-r--r--include/sys/uberblock_impl.h58
-rw-r--r--man/man5/zfs-module-parameters.556
-rw-r--r--module/zfs/mmp.c321
-rw-r--r--module/zfs/spa.c123
-rw-r--r--module/zfs/uberblock.c12
-rw-r--r--module/zfs/vdev_label.c27
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/cleanup.ksh2
-rw-r--r--tests/zfs-tests/tests/functional/mmp/mmp.cfg6
-rw-r--r--tests/zfs-tests/tests/functional/mmp/mmp.kshlib119
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/mmp_active_import.ksh20
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/mmp_inactive_import.ksh18
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/mmp_on_uberblocks.ksh22
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/mmp_reset_interval.ksh59
-rwxr-xr-xtests/zfs-tests/tests/functional/mmp/setup.ksh2
17 files changed, 692 insertions, 175 deletions
diff --git a/cmd/zdb/zdb.c b/cmd/zdb/zdb.c
index a60f0f255..4b07cdb8e 100644
--- a/cmd/zdb/zdb.c
+++ b/cmd/zdb/zdb.c
@@ -2555,9 +2555,22 @@ dump_uberblock(uberblock_t *ub, const char *header, const char *footer)
(void) printf("\tmmp_magic = %016llx\n",
(u_longlong_t)ub->ub_mmp_magic);
- if (ub->ub_mmp_magic == MMP_MAGIC)
+ if (MMP_VALID(ub)) {
(void) printf("\tmmp_delay = %0llu\n",
(u_longlong_t)ub->ub_mmp_delay);
+ if (MMP_SEQ_VALID(ub))
+ (void) printf("\tmmp_seq = %u\n",
+ (unsigned int) MMP_SEQ(ub));
+ if (MMP_FAIL_INT_VALID(ub))
+ (void) printf("\tmmp_fail = %u\n",
+ (unsigned int) MMP_FAIL_INT(ub));
+ if (MMP_INTERVAL_VALID(ub))
+ (void) printf("\tmmp_write = %u\n",
+ (unsigned int) MMP_INTERVAL(ub));
+ /* After MMP_* to make summarize_uberblock_mmp cleaner */
+ (void) printf("\tmmp_valid = %x\n",
+ (unsigned int) ub->ub_mmp_config & 0xFF);
+ }
if (dump_opt['u'] >= 4) {
char blkbuf[BP_SPRINTF_LEN];
diff --git a/include/sys/fs/zfs.h b/include/sys/fs/zfs.h
index d25e8b5a5..e49a58f43 100644
--- a/include/sys/fs/zfs.h
+++ b/include/sys/fs/zfs.h
@@ -718,6 +718,7 @@ typedef struct zpool_load_policy {
#define ZPOOL_CONFIG_CACHEFILE "cachefile" /* not stored on disk */
#define ZPOOL_CONFIG_MMP_STATE "mmp_state" /* not stored on disk */
#define ZPOOL_CONFIG_MMP_TXG "mmp_txg" /* not stored on disk */
+#define ZPOOL_CONFIG_MMP_SEQ "mmp_seq" /* not stored on disk */
#define ZPOOL_CONFIG_MMP_HOSTNAME "mmp_hostname" /* not stored on disk */
#define ZPOOL_CONFIG_MMP_HOSTID "mmp_hostid" /* not stored on disk */
#define ZPOOL_CONFIG_ALLOCATION_BIAS "alloc_bias" /* not stored on disk */
diff --git a/include/sys/mmp.h b/include/sys/mmp.h
index 6f4672a38..527e3323b 100644
--- a/include/sys/mmp.h
+++ b/include/sys/mmp.h
@@ -31,6 +31,11 @@ extern "C" {
#define MMP_DEFAULT_INTERVAL 1000 /* ms */
#define MMP_DEFAULT_IMPORT_INTERVALS 20
#define MMP_DEFAULT_FAIL_INTERVALS 10
+#define MMP_MIN_FAIL_INTERVALS 2 /* min if != 0 */
+#define MMP_IMPORT_SAFETY_FACTOR 200 /* pct */
+#define MMP_INTERVAL_OK(interval) MAX(interval, MMP_MIN_INTERVAL)
+#define MMP_FAIL_INTVS_OK(fails) (fails == 0 ? 0 : MAX(fails, \
+ MMP_MIN_FAIL_INTERVALS))
typedef struct mmp_thread {
kmutex_t mmp_thread_lock; /* protect thread mgmt fields */
@@ -46,6 +51,7 @@ typedef struct mmp_thread {
int mmp_skip_error; /* reason for last skipped write */
vdev_t *mmp_last_leaf; /* last mmp write sent here */
uint64_t mmp_leaf_last_gen; /* last mmp write sent here */
+ uint32_t mmp_seq; /* intra-second update counter */
} mmp_thread_t;
diff --git a/include/sys/uberblock_impl.h b/include/sys/uberblock_impl.h
index 113df7c61..91699e651 100644
--- a/include/sys/uberblock_impl.h
+++ b/include/sys/uberblock_impl.h
@@ -44,7 +44,36 @@ extern "C" {
*/
#define UBERBLOCK_MAGIC 0x00bab10c /* oo-ba-bloc! */
#define UBERBLOCK_SHIFT 10 /* up to 1K */
-#define MMP_MAGIC 0xa11cea11 /* all-see-all */
+#define MMP_MAGIC 0xa11cea11 /* all-see-all */
+
+#define MMP_INTERVAL_VALID_BIT 0x01
+#define MMP_SEQ_VALID_BIT 0x02
+#define MMP_FAIL_INT_VALID_BIT 0x04
+
+#define MMP_VALID(ubp) (ubp->ub_magic == UBERBLOCK_MAGIC && \
+ ubp->ub_mmp_magic == MMP_MAGIC)
+#define MMP_INTERVAL_VALID(ubp) (MMP_VALID(ubp) && (ubp->ub_mmp_config & \
+ MMP_INTERVAL_VALID_BIT))
+#define MMP_SEQ_VALID(ubp) (MMP_VALID(ubp) && (ubp->ub_mmp_config & \
+ MMP_SEQ_VALID_BIT))
+#define MMP_FAIL_INT_VALID(ubp) (MMP_VALID(ubp) && (ubp->ub_mmp_config & \
+ MMP_FAIL_INT_VALID_BIT))
+
+#define MMP_INTERVAL(ubp) ((ubp->ub_mmp_config & 0x00000000FFFFFF00) \
+ >> 8)
+#define MMP_SEQ(ubp) ((ubp->ub_mmp_config & 0x0000FFFF00000000) \
+ >> 32)
+#define MMP_FAIL_INT(ubp) ((ubp->ub_mmp_config & 0xFFFF000000000000) \
+ >> 48)
+
+#define MMP_INTERVAL_SET(write) \
+ (((uint64_t)(write & 0xFFFFFF) << 8) | MMP_INTERVAL_VALID_BIT)
+
+#define MMP_SEQ_SET(seq) \
+ (((uint64_t)(seq & 0xFFFF) << 32) | MMP_SEQ_VALID_BIT)
+
+#define MMP_FAIL_INT_SET(fail) \
+ (((uint64_t)(fail & 0xFFFF) << 48) | MMP_FAIL_INT_VALID_BIT)
struct uberblock {
uint64_t ub_magic; /* UBERBLOCK_MAGIC */
@@ -59,8 +88,31 @@ struct uberblock {
/* Maybe missing in uberblocks we read, but always written */
uint64_t ub_mmp_magic; /* MMP_MAGIC */
- uint64_t ub_mmp_delay; /* nanosec since last MMP write */
- uint64_t ub_mmp_seq; /* reserved for sequence number */
+ /*
+ * If ub_mmp_delay == 0 and ub_mmp_magic is valid, MMP is off.
+ * Otherwise, nanosec since last MMP write.
+ */
+ uint64_t ub_mmp_delay;
+
+ /*
+ * The ub_mmp_config contains the multihost write interval, multihost
+ * fail intervals, sequence number for sub-second granularity, and
+ * valid bit mask. This layout is as follows:
+ *
+ * 64 56 48 40 32 24 16 8 0
+ * +-------+-------+-------+-------+-------+-------+-------+-------+
+ * 0 | Fail Intervals| Seq | Write Interval (ms) | VALID |
+ * +-------+-------+-------+-------+-------+-------+-------+-------+
+ *
+ * This allows a write_interval of (2^24/1000)s, over 4.5 hours
+ *
+ * VALID Bits:
+ * - 0x01 - Write Interval (ms)
+ * - 0x02 - Sequence number exists
+ * - 0x04 - Fail Intervals
+ * - 0xf8 - Reserved
+ */
+ uint64_t ub_mmp_config;
/*
* ub_checkpoint_txg indicates two things about the current uberblock:
diff --git a/man/man5/zfs-module-parameters.5 b/man/man5/zfs-module-parameters.5
index f75f09917..c1994f340 100644
--- a/man/man5/zfs-module-parameters.5
+++ b/man/man5/zfs-module-parameters.5
@@ -1849,20 +1849,14 @@ Default value: \fB0\fR.
.ad
.RS 12n
Used to control the frequency of multihost writes which are performed when the
-\fBmultihost\fR pool property is on. This is one factor used to determine
-the length of the activity check during import.
+\fBmultihost\fR pool property is on. This is one factor used to determine the
+length of the activity check during import.
.sp
-The multihost write period is \fBzfs_multihost_interval / leaf-vdevs\fR milliseconds.
-This means that on average a multihost write will be issued for each leaf vdev every
-\fBzfs_multihost_interval\fR milliseconds. In practice, the observed period can
-vary with the I/O load and this observed value is the delay which is stored in
-the uberblock.
-.sp
-On import the activity check waits a minimum amount of time determined by
-\fBzfs_multihost_interval * zfs_multihost_import_intervals\fR. The activity
-check time may be further extended if the value of mmp delay found in the best
-uberblock indicates actual multihost updates happened at longer intervals than
-\fBzfs_multihost_interval\fR. A minimum value of \fB100ms\fR is enforced.
+The multihost write period is \fBzfs_multihost_interval / leaf-vdevs\fR
+milliseconds. On average a multihost write will be issued for each leaf vdev
+every \fBzfs_multihost_interval\fR milliseconds. In practice, the observed
+period can vary with the I/O load and this observed value is the delay which is
+stored in the uberblock.
.sp
Default value: \fB1000\fR.
.RE
@@ -1876,8 +1870,17 @@ Default value: \fB1000\fR.
Used to control the duration of the activity test on import. Smaller values of
\fBzfs_multihost_import_intervals\fR will reduce the import time but increase
the risk of failing to detect an active pool. The total activity check time is
-never allowed to drop below one second. A value of 0 is ignored and treated as
-if it was set to 1
+never allowed to drop below one second.
+.sp
+On import the activity check waits a minimum amount of time determined by
+\fBzfs_multihost_interval * zfs_multihost_import_intervals\fR, or the same
+product computed on the host which last had the pool imported (whichever is
+greater). The activity check time may be further extended if the value of mmp
+delay found in the best uberblock indicates actual multihost updates happened
+at longer intervals than \fBzfs_multihost_interval\fR. A minimum value of
+\fB100ms\fR is enforced.
+.sp
+A value of 0 is ignored and treated as if it was set to 1.
.sp
Default value: \fB20\fR.
.RE
@@ -1888,17 +1891,22 @@ Default value: \fB20\fR.
\fBzfs_multihost_fail_intervals\fR (uint)
.ad
.RS 12n
-Controls the behavior of the pool when multihost write failures are detected.
+Controls the behavior of the pool when multihost write failures or delays are
+detected.
.sp
-When \fBzfs_multihost_fail_intervals = 0\fR then multihost write failures are ignored.
-The failures will still be reported to the ZED which depending on its
-configuration may take action such as suspending the pool or offlining a device.
+When \fBzfs_multihost_fail_intervals = 0\fR, multihost write failures or delays
+are ignored. The failures will still be reported to the ZED which depending on
+its configuration may take action such as suspending the pool or offlining a
+device.
+
.sp
-When \fBzfs_multihost_fail_intervals > 0\fR then sequential multihost write failures
-will cause the pool to be suspended. This occurs when
-\fBzfs_multihost_fail_intervals * zfs_multihost_interval\fR milliseconds have
-passed since the last successful multihost write. This guarantees the activity test
-will see multihost writes if the pool is imported.
+When \fBzfs_multihost_fail_intervals > 0\fR, the pool will be suspended if
+\fBzfs_multihost_fail_intervals * zfs_multihost_interval\fR milliseconds pass
+without a successful mmp write. This guarantees the activity test will see
+mmp writes if the pool is imported. A value of 1 is ignored and treated as
+if it was set to 2. This is necessary to prevent the pool from being suspended
+due to normal, small I/O latency variations.
+
.sp
Default value: \fB10\fR.
.RE
diff --git a/module/zfs/mmp.c b/module/zfs/mmp.c
index 16975dd98..cd5603a1a 100644
--- a/module/zfs/mmp.c
+++ b/module/zfs/mmp.c
@@ -49,40 +49,101 @@
* Uberblocks written by the txg_sync thread always go into the first
* (N-MMP_BLOCKS_PER_LABEL) slots, the remaining slots are reserved for MMP.
* They are used to hold uberblocks which are exactly the same as the last
- * synced uberblock except that the ub_timestamp is frequently updated.
- * Like all other uberblocks, the slot is written with an embedded checksum,
- * and slots with invalid checksums are ignored. This provides the
+ * synced uberblock except that the ub_timestamp and mmp_config are frequently
+ * updated. Like all other uberblocks, the slot is written with an embedded
+ * checksum, and slots with invalid checksums are ignored. This provides the
* "heartbeat", with no risk of overwriting good uberblocks that must be
* preserved, e.g. previous txgs and associated block pointers.
*
- * Two optional fields are added to uberblock structure: ub_mmp_magic and
- * ub_mmp_delay. The magic field allows zfs to tell whether ub_mmp_delay is
- * valid. The delay field is a decaying average of the amount of time between
- * completion of successive MMP writes, in nanoseconds. It is used to predict
- * how long the import must wait to detect activity in the pool, before
- * concluding it is not in use.
+ * Three optional fields are added to uberblock structure; ub_mmp_magic,
+ * ub_mmp_config, and ub_mmp_delay. The ub_mmp_magic value allows zfs to tell
+ * whether the other ub_mmp_* fields are valid. The ub_mmp_config field tells
+ * the importing host the settings of zfs_multihost_interval and
+ * zfs_multihost_fail_intervals on the host which last had (or currently has)
+ * the pool imported. These determine how long a host must wait to detect
+ * activity in the pool, before concluding the pool is not in use. The
+ * mmp_delay field is a decaying average of the amount of time between
+ * completion of successive MMP writes, in nanoseconds. It indicates whether
+ * MMP is enabled.
*
* During import an activity test may now be performed to determine if
* the pool is in use. The activity test is typically required if the
* ZPOOL_CONFIG_HOSTID does not match the system hostid, the pool state is
* POOL_STATE_ACTIVE, and the pool is not a root pool.
*
- * The activity test finds the "best" uberblock (highest txg & timestamp),
- * waits some time, and then finds the "best" uberblock again. If the txg
- * and timestamp in both "best" uberblocks do not match, the pool is in use
- * by another host and the import fails. Since the granularity of the
- * timestamp is in seconds this activity test must take a bare minimum of one
- * second. In order to assure the accuracy of the activity test, the default
- * values result in an activity test duration of 10x the mmp write interval.
- *
- * The "zpool import" activity test can be expected to take a minimum time of
- * zfs_multihost_import_intervals * zfs_multihost_interval milliseconds. If the
- * "best" uberblock has a valid ub_mmp_delay field, then the duration of the
- * test may take longer if MMP writes were occurring less frequently than
- * expected. Additionally, the duration is then extended by a random 25% to
- * attempt to to detect simultaneous imports. For example, if both partner
- * hosts are rebooted at the same time and automatically attempt to import the
- * pool.
+ * The activity test finds the "best" uberblock (highest txg, timestamp, and, if
+ * ub_mmp_magic is valid, sequence number from ub_mmp_config). It then waits
+ * some time, and finds the "best" uberblock again. If any of the mentioned
+ * fields have different values in the newly read uberblock, the pool is in use
+ * by another host and the import fails. In order to assure the accuracy of the
+ * activity test, the default values result in an activity test duration of 20x
+ * the mmp write interval.
+ *
+ * The duration of the "zpool import" activity test depends on the information
+ * available in the "best" uberblock:
+ *
+ * 1) If uberblock was written by zfs-0.8 or newer and fail_intervals > 0:
+ * ub_mmp_config.fail_intervals * ub_mmp_config.multihost_interval * 2
+ *
+ * In this case, a weak guarantee is provided. Since the host which last had
+ * the pool imported will suspend the pool if no mmp writes land within
+ * fail_intervals * multihost_interval ms, the absense of writes during that
+ * time means either the pool is not imported, or it is imported but the pool
+ * is suspended and no further writes will occur.
+ *
+ * Note that resuming the suspended pool on the remote host would invalidate
+ * this gurantee, and so it is not allowed.
+ *
+ * The factor of 2 provides a conservative safety factor and derives from
+ * MMP_IMPORT_SAFETY_FACTOR;
+ *
+ * 2) If uberblock was written by zfs-0.8 or newer and fail_intervals == 0:
+ * (ub_mmp_config.multihost_interval + ub_mmp_delay) *
+ * zfs_multihost_import_intervals
+ *
+ * In this case no guarantee can provided. However, as long as some devices
+ * are healthy and connected, it is likely that at least one write will land
+ * within (multihost_interval + mmp_delay) because multihost_interval is
+ * enough time for a write to be attempted to each leaf vdev, and mmp_delay
+ * is enough for one to land, based on past delays. Multiplying by
+ * zfs_multihost_import_intervals provides a conservative safety factor.
+ *
+ * 3) If uberblock was written by zfs-0.7:
+ * (zfs_multihost_interval + ub_mmp_delay) * zfs_multihost_import_intervals
+ *
+ * The same logic as case #2 applies, but we do not know remote tunables.
+ *
+ * We use the local value for zfs_multihost_interval because the original MMP
+ * did not record this value in the uberblock.
+ *
+ * ub_mmp_delay >= (zfs_multihost_interval / leaves), so if the other host
+ * has a much larger zfs_multihost_interval set, ub_mmp_delay will reflect
+ * that. We will have waited enough time for zfs_multihost_import_intervals
+ * writes to be issued and all but one to land.
+ *
+ * single device pool example delays
+ *
+ * import_delay = (1 + 1) * 20 = 40s #defaults, no I/O delay
+ * import_delay = (1 + 10) * 20 = 220s #defaults, 10s I/O delay
+ * import_delay = (10 + 10) * 20 = 400s #10s multihost_interval,
+ * no I/O delay
+ * 100 device pool example delays
+ *
+ * import_delay = (1 + .01) * 20 = 20s #defaults, no I/O delay
+ * import_delay = (1 + 10) * 20 = 220s #defaults, 10s I/O delay
+ * import_delay = (10 + .1) * 20 = 202s #10s multihost_interval,
+ * no I/O delay
+ *
+ * 4) Otherwise, this uberblock was written by a pre-MMP zfs:
+ * zfs_multihost_import_intervals * zfs_multihost_interval
+ *
+ * In this case local tunables are used. By default this product = 10s, long
+ * enough for a pool with any activity at all to write at least one
+ * uberblock. No guarantee can be provided.
+ *
+ * Additionally, the duration is then extended by a random 25% to attempt to to
+ * detect simultaneous imports. For example, if both partner hosts are rebooted
+ * at the same time and automatically attempt to import the pool.
*/
/*
@@ -90,10 +151,9 @@
* 'multihost' pool property is on. This is one factor used to determine the
* length of the activity check during import.
*
- * The mmp write period is zfs_multihost_interval / leaf-vdevs milliseconds.
- * This means that on average an mmp write will be issued for each leaf vdev
- * every zfs_multihost_interval milliseconds. In practice, the observed period
- * can vary with the I/O load and this observed value is the delay which is
+ * On average an mmp write will be issued for each leaf vdev every
+ * zfs_multihost_interval milliseconds. In practice, the observed period can
+ * vary with the I/O load and this observed value is the ub_mmp_delay which is
* stored in the uberblock. The minimum allowed value is 100 ms.
*/
ulong_t zfs_multihost_interval = MMP_DEFAULT_INTERVAL;
@@ -108,19 +168,21 @@ ulong_t zfs_multihost_interval = MMP_DEFAULT_INTERVAL;
uint_t zfs_multihost_import_intervals = MMP_DEFAULT_IMPORT_INTERVALS;
/*
- * Controls the behavior of the pool when mmp write failures are detected.
+ * Controls the behavior of the pool when mmp write failures or delays are
+ * detected.
*
- * When zfs_multihost_fail_intervals = 0 then mmp write failures are ignored.
- * The failures will still be reported to the ZED which depending on its
- * configuration may take action such as suspending the pool or taking a
+ * When zfs_multihost_fail_intervals = 0, mmp write failures or delays are
+ * ignored. The failures will still be reported to the ZED which depending on
+ * its configuration may take action such as suspending the pool or taking a
* device offline.
*
- * When zfs_multihost_fail_intervals > 0 then sequential mmp write failures will
- * cause the pool to be suspended. This occurs when
- * zfs_multihost_fail_intervals * zfs_multihost_interval milliseconds have
- * passed since the last successful mmp write. This guarantees the activity
- * test will see mmp writes if the
- * pool is imported.
+ * When zfs_multihost_fail_intervals > 0, the pool will be suspended if
+ * zfs_multihost_fail_intervals * zfs_multihost_interval milliseconds pass
+ * without a successful mmp write. This guarantees the activity test will see
+ * mmp writes if the pool is imported. A value of 1 is ignored and treated as
+ * if it was set to 2, because a single leaf vdev pool will issue a write once
+ * per multihost_interval and thus any variation in latency would cause the
+ * pool to be suspended.
*/
uint_t zfs_multihost_fail_intervals = MMP_DEFAULT_FAIL_INTERVALS;
@@ -136,6 +198,14 @@ mmp_init(spa_t *spa)
cv_init(&mmp->mmp_thread_cv, NULL, CV_DEFAULT, NULL);
mutex_init(&mmp->mmp_io_lock, NULL, MUTEX_DEFAULT, NULL);
mmp->mmp_kstat_id = 1;
+
+ /*
+ * mmp_write_done() calculates mmp_delay based on prior mmp_delay and
+ * the elapsed time since the last write. For the first mmp write,
+ * there is no "last write", so we start with fake non-zero values.
+ */
+ mmp->mmp_last_write = gethrtime();
+ mmp->mmp_delay = MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval));
}
void
@@ -173,10 +243,10 @@ mmp_thread_start(spa_t *spa)
if (spa_writeable(spa)) {
mutex_enter(&mmp->mmp_thread_lock);
if (!mmp->mmp_thread) {
- dprintf("mmp_thread_start pool %s\n",
- spa->spa_name);
mmp->mmp_thread = thread_create(NULL, 0, mmp_thread,
spa, 0, &p0, TS_RUN, defclsyspri);
+ zfs_dbgmsg("MMP thread started pool '%s' "
+ "gethrtime %llu", spa_name(spa), gethrtime());
}
mutex_exit(&mmp->mmp_thread_lock);
}
@@ -195,6 +265,8 @@ mmp_thread_stop(spa_t *spa)
cv_wait(&mmp->mmp_thread_cv, &mmp->mmp_thread_lock);
}
mutex_exit(&mmp->mmp_thread_lock);
+ zfs_dbgmsg("MMP thread stopped pool '%s' gethrtime %llu",
+ spa_name(spa), gethrtime());
ASSERT(mmp->mmp_thread == NULL);
mmp->mmp_thread_exiting = 0;
@@ -304,7 +376,8 @@ mmp_delay_update(spa_t *spa, boolean_t write_completed)
* strictly less than, in case delay was changed above.
*/
if (delay < mts->mmp_delay) {
- hrtime_t min_delay = MSEC2NSEC(zfs_multihost_interval) /
+ hrtime_t min_delay =
+ MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval)) /
MAX(1, vdev_count_leaves(spa));
mts->mmp_delay = MAX(((delay + mts->mmp_delay * 127) / 128),
min_delay);
@@ -348,6 +421,7 @@ mmp_update_uberblock(spa_t *spa, uberblock_t *ub)
mutex_enter(&mmp->mmp_io_lock);
mmp->mmp_ub = *ub;
+ mmp->mmp_seq = 1;
mmp->mmp_ub.ub_timestamp = gethrestime_sec();
mmp_delay_update(spa, B_TRUE);
mutex_exit(&mmp->mmp_io_lock);
@@ -372,8 +446,9 @@ mmp_write_uberblock(spa_t *spa)
spa_config_enter(spa, SCL_STATE, mmp_tag, RW_READER);
lock_acquire_time = gethrtime() - lock_acquire_time;
if (lock_acquire_time > (MSEC2NSEC(MMP_MIN_INTERVAL) / 10))
- zfs_dbgmsg("SCL_STATE acquisition took %llu ns\n",
- (u_longlong_t)lock_acquire_time);
+ zfs_dbgmsg("MMP SCL_STATE acquisition pool '%s' took %llu ns "
+ "gethrtime %llu", spa_name(spa), lock_acquire_time,
+ gethrtime());
mutex_enter(&mmp->mmp_io_lock);
@@ -396,6 +471,9 @@ mmp_write_uberblock(spa_t *spa)
spa_mmp_history_add(spa, mmp->mmp_ub.ub_txg,
gethrestime_sec(), mmp->mmp_delay, NULL, 0,
mmp->mmp_kstat_id++, error);
+ zfs_dbgmsg("MMP error choosing leaf pool '%s' "
+ "gethrtime %llu fail_mask %#x", spa_name(spa),
+ gethrtime(), error);
}
mutex_exit(&mmp->mmp_io_lock);
spa_config_exit(spa, SCL_STATE, mmp_tag);
@@ -403,16 +481,34 @@ mmp_write_uberblock(spa_t *spa)
}
vd = spa->spa_mmp.mmp_last_leaf;
- mmp->mmp_skip_error = 0;
+ if (mmp->mmp_skip_error != 0) {
+ mmp->mmp_skip_error = 0;
+ zfs_dbgmsg("MMP write after skipping due to unavailable "
+ "leaves, pool '%s' gethrtime %llu leaf %#llu",
+ spa_name(spa), gethrtime(), vd->vdev_guid);
+ }
if (mmp->mmp_zio_root == NULL)
mmp->mmp_zio_root = zio_root(spa, NULL, NULL,
flags | ZIO_FLAG_GODFATHER);
+ if (mmp->mmp_ub.ub_timestamp != gethrestime_sec()) {
+ /*
+ * Want to reset mmp_seq when timestamp advances because after
+ * an mmp_seq wrap new values will not be chosen by
+ * uberblock_compare() as the "best".
+ */
+ mmp->mmp_ub.ub_timestamp = gethrestime_sec();
+ mmp->mmp_seq = 1;
+ }
+
ub = &mmp->mmp_ub;
- ub->ub_timestamp = gethrestime_sec();
ub->ub_mmp_magic = MMP_MAGIC;
ub->ub_mmp_delay = mmp->mmp_delay;
+ ub->ub_mmp_config = MMP_SEQ_SET(mmp->mmp_seq) |
+ MMP_INTERVAL_SET(MMP_INTERVAL_OK(zfs_multihost_interval)) |
+ MMP_FAIL_INT_SET(MMP_FAIL_INTVS_OK(
+ zfs_multihost_fail_intervals));
vd->vdev_mmp_pending = gethrtime();
vd->vdev_mmp_kstat_id = mmp->mmp_kstat_id;
@@ -421,6 +517,7 @@ mmp_write_uberblock(spa_t *spa)
abd_zero(ub_abd, VDEV_UBERBLOCK_SIZE(vd));
abd_copy_from_buf(ub_abd, ub, sizeof (uberblock_t));
+ mmp->mmp_seq++;
mmp->mmp_kstat_id++;
mutex_exit(&mmp->mmp_io_lock);
@@ -443,38 +540,75 @@ mmp_thread(void *arg)
{
spa_t *spa = (spa_t *)arg;
mmp_thread_t *mmp = &spa->spa_mmp;
- boolean_t last_spa_suspended = spa_suspended(spa);
- boolean_t last_spa_multihost = spa_multihost(spa);
+ boolean_t suspended = spa_suspended(spa);
+ boolean_t multihost = spa_multihost(spa);
+ uint64_t mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK(
+ zfs_multihost_interval));
+ uint32_t mmp_fail_intervals = MMP_FAIL_INTVS_OK(
+ zfs_multihost_fail_intervals);
+ hrtime_t mmp_fail_ns = mmp_fail_intervals * mmp_interval;
+ boolean_t last_spa_suspended = suspended;
+ boolean_t last_spa_multihost = multihost;
+ uint64_t last_mmp_interval = mmp_interval;
+ uint32_t last_mmp_fail_intervals = mmp_fail_intervals;
+ hrtime_t last_mmp_fail_ns = mmp_fail_ns;
callb_cpr_t cpr;
- hrtime_t max_fail_ns = zfs_multihost_fail_intervals *
- MSEC2NSEC(MAX(zfs_multihost_interval, MMP_MIN_INTERVAL));
+ int skip_wait = 0;
mmp_thread_enter(mmp, &cpr);
- /*
- * The mmp_write_done() function calculates mmp_delay based on the
- * prior value of mmp_delay and the elapsed time since the last write.
- * For the first mmp write, there is no "last write", so we start
- * with fake, but reasonable, default non-zero values.
- */
- mmp->mmp_delay = MSEC2NSEC(MAX(zfs_multihost_interval,
- MMP_MIN_INTERVAL)) / MAX(vdev_count_leaves(spa), 1);
- mmp->mmp_last_write = gethrtime() - mmp->mmp_delay;
-
while (!mmp->mmp_thread_exiting) {
- uint64_t mmp_fail_intervals = zfs_multihost_fail_intervals;
- uint64_t mmp_interval = MSEC2NSEC(
- MAX(zfs_multihost_interval, MMP_MIN_INTERVAL));
- boolean_t suspended = spa_suspended(spa);
- boolean_t multihost = spa_multihost(spa);
- hrtime_t next_time;
+ hrtime_t next_time = gethrtime() +
+ MSEC2NSEC(MMP_DEFAULT_INTERVAL);
+ int leaves = MAX(vdev_count_leaves(spa), 1);
+
+ /* Detect changes in tunables or state */
+
+ last_spa_suspended = suspended;
+ last_spa_multihost = multihost;
+ suspended = spa_suspended(spa);
+ multihost = spa_multihost(spa);
+
+ last_mmp_interval = mmp_interval;
+ last_mmp_fail_intervals = mmp_fail_intervals;
+ last_mmp_fail_ns = mmp_fail_ns;
+ mmp_interval = MSEC2NSEC(MMP_INTERVAL_OK(
+ zfs_multihost_interval));
+ mmp_fail_intervals = MMP_FAIL_INTVS_OK(
+ zfs_multihost_fail_intervals);
+
+ /* Smooth so pool is not suspended when reducing tunables */
+ if (mmp_fail_intervals * mmp_interval < mmp_fail_ns) {
+ mmp_fail_ns = (mmp_fail_ns * 31 +
+ mmp_fail_intervals * mmp_interval) / 32;
+ } else {
+ mmp_fail_ns = mmp_fail_intervals *
+ mmp_interval;
+ }
+
+ if (mmp_interval != last_mmp_interval ||
+ mmp_fail_intervals != last_mmp_fail_intervals) {
+ /*
+ * We want other hosts to see new tunables as quickly as
+ * possible. Write out at higher frequency than usual.
+ */
+ skip_wait += leaves;
+ }
if (multihost)
- next_time = gethrtime() + mmp_interval /
- MAX(vdev_count_leaves(spa), 1);
- else
- next_time = gethrtime() +
- MSEC2NSEC(MMP_DEFAULT_INTERVAL);
+ next_time = gethrtime() + mmp_interval / leaves;
+
+ if (mmp_fail_ns != last_mmp_fail_ns) {
+ zfs_dbgmsg("MMP interval change pool '%s' "
+ "gethrtime %llu last_mmp_interval %llu "
+ "mmp_interval %llu last_mmp_fail_intervals %u "
+ "mmp_fail_intervals %u mmp_fail_ns %llu "
+ "skip_wait %d leaves %d next_time %llu",
+ spa_name(spa), gethrtime(), last_mmp_interval,
+ mmp_interval, last_mmp_fail_intervals,
+ mmp_fail_intervals, mmp_fail_ns, skip_wait, leaves,
+ next_time);
+ }
/*
* MMP off => on, or suspended => !suspended:
@@ -483,8 +617,14 @@ mmp_thread(void *arg)
*/
if ((!last_spa_multihost && multihost) ||
(last_spa_suspended && !suspended)) {
+ zfs_dbgmsg("MMP state change pool '%s': gethrtime %llu "
+ "last_spa_multihost %u multihost %u "
+ "last_spa_suspended %u suspended %u",
+ spa_name(spa), last_spa_multihost, multihost,
+ last_spa_suspended, suspended);
mutex_enter(&mmp->mmp_io_lock);
mmp->mmp_last_write = gethrtime();
+ mmp->mmp_delay = mmp_interval;
mutex_exit(&mmp->mmp_io_lock);
}
@@ -497,38 +637,39 @@ mmp_thread(void *arg)
mmp->mmp_delay = 0;
mutex_exit(&mmp->mmp_io_lock);
}
- last_spa_multihost = multihost;
- last_spa_suspended = suspended;
-
- /*
- * Smooth max_fail_ns when its factors are decreased, because
- * making (max_fail_ns < mmp_interval) results in the pool being
- * immediately suspended before writes can occur at the new
- * higher frequency.
- */
- if ((mmp_interval * mmp_fail_intervals) < max_fail_ns) {
- max_fail_ns = ((31 * max_fail_ns) + (mmp_interval *
- mmp_fail_intervals)) / 32;
- } else {
- max_fail_ns = mmp_interval * mmp_fail_intervals;
- }
/*
* Suspend the pool if no MMP write has succeeded in over
* mmp_interval * mmp_fail_intervals nanoseconds.
*/
- if (!suspended && mmp_fail_intervals && multihost &&
- (gethrtime() - mmp->mmp_last_write) > max_fail_ns) {
+ if (multihost && !suspended && mmp_fail_intervals &&
+ (gethrtime() - mmp->mmp_last_write) > mmp_fail_ns) {
+ zfs_dbgmsg("MMP suspending pool '%s': gethrtime %llu "
+ "mmp_last_write %llu mmp_interval %llu "
+ "mmp_fail_intervals %llu mmp_fail_ns %llu",
+ spa_name(spa), (u_longlong_t)gethrtime(),
+ (u_longlong_t)mmp->mmp_last_write,
+ (u_longlong_t)mmp_interval,
+ (u_longlong_t)mmp_fail_intervals,
+ (u_longlong_t)mmp_fail_ns);
cmn_err(CE_WARN, "MMP writes to pool '%s' have not "
- "succeeded in over %llus; suspending pool",
+ "succeeded in over %llu ms; suspending pool. "
+ "Hrtime %llu",
spa_name(spa),
- NSEC2SEC(gethrtime() - mmp->mmp_last_write));
+ NSEC2MSEC(gethrtime() - mmp->mmp_last_write),
+ gethrtime());
zio_suspend(spa, NULL, ZIO_SUSPEND_MMP);
}
if (multihost && !suspended)
mmp_write_uberblock(spa);
+ if (skip_wait > 0) {
+ next_time = gethrtime() + MSEC2NSEC(MMP_MIN_INTERVAL) /
+ leaves;
+ skip_wait--;
+ }
+
CALLB_CPR_SAFE_BEGIN(&cpr);
(void) cv_timedwait_sig_hires(&mmp->mmp_thread_cv,
&mmp->mmp_thread_lock, next_time, USEC2NSEC(1),
diff --git a/module/zfs/spa.c b/module/zfs/spa.c
index d95a43001..9d798ebac 100644
--- a/module/zfs/spa.c
+++ b/module/zfs/spa.c
@@ -2437,6 +2437,7 @@ spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
uint64_t hostid = 0;
uint64_t tryconfig_txg = 0;
uint64_t tryconfig_timestamp = 0;
+ uint16_t tryconfig_mmp_seq = 0;
nvlist_t *nvinfo;
if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
@@ -2445,6 +2446,8 @@ spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
&tryconfig_txg);
(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
&tryconfig_timestamp);
+ (void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ,
+ &tryconfig_mmp_seq);
}
(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state);
@@ -2462,13 +2465,15 @@ spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0)
return (B_FALSE);
/*
- * If the tryconfig_* values are nonzero, they are the results of an
- * earlier tryimport. If they match the uberblock we just found, then
- * the pool has not changed and we return false so we do not test a
- * second time.
+ * If the tryconfig_ values are nonzero, they are the results of an
+ * earlier tryimport. If they all match the uberblock we just found,
+ * then the pool has not changed and we return false so we do not test
+ * a second time.
*/
if (tryconfig_txg && tryconfig_txg == ub->ub_txg &&
- tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp)
+ tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp &&
+ tryconfig_mmp_seq && tryconfig_mmp_seq ==
+ (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0))
return (B_FALSE);
/*
@@ -2492,16 +2497,87 @@ spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
}
/*
+ * Nanoseconds the activity check must watch for changes on-disk.
+ */
+static uint64_t
+spa_activity_check_duration(spa_t *spa, uberblock_t *ub)
+{
+ uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
+ uint64_t multihost_interval = MSEC2NSEC(
+ MMP_INTERVAL_OK(zfs_multihost_interval));
+ uint64_t import_delay = MAX(NANOSEC, import_intervals *
+ multihost_interval);
+
+ /*
+ * Local tunables determine a minimum duration except for the case
+ * where we know when the remote host will suspend the pool if MMP
+ * writes do not land.
+ *
+ * See Big Theory comment at the top of mmp.c for the reasoning behind
+ * these cases and times.
+ */
+
+ ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100);
+
+ if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
+ MMP_FAIL_INT(ub) > 0) {
+
+ /* MMP on remote host will suspend pool after failed writes */
+ import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) *
+ MMP_IMPORT_SAFETY_FACTOR / 100;
+
+ zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp "
+ "mmp_fails=%llu ub_mmp mmp_interval=%llu "
+ "import_intervals=%u", import_delay, MMP_FAIL_INT(ub),
+ MMP_INTERVAL(ub), import_intervals);
+
+ } else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
+ MMP_FAIL_INT(ub) == 0) {
+
+ /* MMP on remote host will never suspend pool */
+ import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) +
+ ub->ub_mmp_delay) * import_intervals);
+
+ zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp "
+ "mmp_interval=%llu ub_mmp_delay=%llu "
+ "import_intervals=%u", import_delay, MMP_INTERVAL(ub),
+ ub->ub_mmp_delay, import_intervals);
+
+ } else if (MMP_VALID(ub)) {
+ /*
+ * zfs-0.7 compatability case
+ */
+
+ import_delay = MAX(import_delay, (multihost_interval +
+ ub->ub_mmp_delay) * import_intervals);
+
+ zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu "
+ "import_intervals=%u leaves=%u", import_delay,
+ ub->ub_mmp_delay, import_intervals,
+ vdev_count_leaves(spa));
+ } else {
+ /* Using local tunings is the only reasonable option */
+ zfs_dbgmsg("pool last imported on non-MMP aware "
+ "host using import_delay=%llu multihost_interval=%llu "
+ "import_intervals=%u", import_delay, multihost_interval,
+ import_intervals);
+ }
+
+ return (import_delay);
+}
+
+/*
* Perform the import activity check. If the user canceled the import or
* we detected activity then fail.
*/
static int
spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
{
- uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
uint64_t txg = ub->ub_txg;
uint64_t timestamp = ub->ub_timestamp;
- uint64_t import_delay = NANOSEC;
+ uint64_t mmp_config = ub->ub_mmp_config;
+ uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0;
+ uint64_t import_delay;
hrtime_t import_expire;
nvlist_t *mmp_label = NULL;
vdev_t *rvd = spa->spa_root_vdev;
@@ -2518,7 +2594,7 @@ spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
* during the earlier tryimport. If the txg recorded there is 0 then
* the pool is known to be active on another host.
*
- * Otherwise, the pool might be in use on another node. Check for
+ * Otherwise, the pool might be in use on another host. Check for
* changes in the uberblocks on disk if necessary.
*/
if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
@@ -2533,23 +2609,7 @@ spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
}
}
- /*
- * Preferentially use the zfs_multihost_interval from the node which
- * last imported the pool. This value is stored in an MMP uberblock as.
- *
- * ub_mmp_delay * vdev_count_leaves() == zfs_multihost_interval
- */
- if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay)
- import_delay = MAX(import_delay, import_intervals *
- ub->ub_mmp_delay * MAX(vdev_count_leaves(spa), 1));
-
- /* Apply a floor using the local default values. */
- import_delay = MAX(import_delay, import_intervals *
- MSEC2NSEC(MAX(zfs_multihost_interval, MMP_MIN_INTERVAL)));
-
- zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu import_intervals=%u "
- "leaves=%u", import_delay, ub->ub_mmp_delay, import_intervals,
- vdev_count_leaves(spa));
+ import_delay = spa_activity_check_duration(spa, ub);
/* Add a small random factor in case of simultaneous imports (0-25%) */
import_expire = gethrtime() + import_delay +
@@ -2558,7 +2618,15 @@ spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
while (gethrtime() < import_expire) {
vdev_uberblock_load(rvd, ub, &mmp_label);
- if (txg != ub->ub_txg || timestamp != ub->ub_timestamp) {
+ if (txg != ub->ub_txg || timestamp != ub->ub_timestamp ||
+ mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) {
+ zfs_dbgmsg("multihost activity detected "
+ "txg %llu ub_txg %llu "
+ "timestamp %llu ub_timestamp %llu "
+ "mmp_config %#llx ub_mmp_config %#llx",
+ txg, ub->ub_txg, timestamp, ub->ub_timestamp,
+ mmp_config, ub->ub_mmp_config);
+
error = SET_ERROR(EREMOTEIO);
break;
}
@@ -2944,6 +3012,9 @@ spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE);
fnvlist_add_uint64(spa->spa_load_info,
ZPOOL_CONFIG_MMP_TXG, ub->ub_txg);
+ fnvlist_add_uint16(spa->spa_load_info,
+ ZPOOL_CONFIG_MMP_SEQ,
+ (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0));
}
/*
diff --git a/module/zfs/uberblock.c b/module/zfs/uberblock.c
index 3b8526076..b8857d74d 100644
--- a/module/zfs/uberblock.c
+++ b/module/zfs/uberblock.c
@@ -26,6 +26,7 @@
#include <sys/zfs_context.h>
#include <sys/uberblock_impl.h>
#include <sys/vdev_impl.h>
+#include <sys/mmp.h>
int
uberblock_verify(uberblock_t *ub)
@@ -58,8 +59,15 @@ uberblock_update(uberblock_t *ub, vdev_t *rvd, uint64_t txg, uint64_t mmp_delay)
ub->ub_timestamp = gethrestime_sec();
ub->ub_software_version = SPA_VERSION;
ub->ub_mmp_magic = MMP_MAGIC;
- ub->ub_mmp_delay = spa_multihost(rvd->vdev_spa) ? mmp_delay : 0;
- ub->ub_mmp_seq = 0;
+ if (spa_multihost(rvd->vdev_spa)) {
+ ub->ub_mmp_delay = mmp_delay;
+ ub->ub_mmp_config = MMP_SEQ_SET(0) |
+ MMP_INTERVAL_SET(zfs_multihost_interval) |
+ MMP_FAIL_INT_SET(zfs_multihost_fail_intervals);
+ } else {
+ ub->ub_mmp_delay = 0;
+ ub->ub_mmp_config = 0;
+ }
ub->ub_checkpoint_txg = 0;
return (ub->ub_rootbp.blk_birth == txg);
diff --git a/module/zfs/vdev_label.c b/module/zfs/vdev_label.c
index 65b847d66..a03722d05 100644
--- a/module/zfs/vdev_label.c
+++ b/module/zfs/vdev_label.c
@@ -1181,10 +1181,35 @@ static int
vdev_uberblock_compare(const uberblock_t *ub1, const uberblock_t *ub2)
{
int cmp = AVL_CMP(ub1->ub_txg, ub2->ub_txg);
+
+ if (likely(cmp))
+ return (cmp);
+
+ cmp = AVL_CMP(ub1->ub_timestamp, ub2->ub_timestamp);
if (likely(cmp))
return (cmp);
- return (AVL_CMP(ub1->ub_timestamp, ub2->ub_timestamp));
+ /*
+ * If MMP_VALID(ub) && MMP_SEQ_VALID(ub) then the host has an MMP-aware
+ * ZFS, e.g. zfsonlinux >= 0.7.
+ *
+ * If one ub has MMP and the other does not, they were written by
+ * different hosts, which matters for MMP. So we treat no MMP/no SEQ as
+ * a 0 value.
+ *
+ * Since timestamp and txg are the same if we get this far, either is
+ * acceptable for importing the pool.
+ */
+ unsigned int seq1 = 0;
+ unsigned int seq2 = 0;
+
+ if (MMP_VALID(ub1) && MMP_SEQ_VALID(ub1))
+ seq1 = MMP_SEQ(ub1);
+
+ if (MMP_VALID(ub2) && MMP_SEQ_VALID(ub2))
+ seq2 = MMP_SEQ(ub2);
+
+ return (AVL_CMP(seq1, seq2));
}
struct ubl_cbdata {
diff --git a/tests/zfs-tests/tests/functional/mmp/cleanup.ksh b/tests/zfs-tests/tests/functional/mmp/cleanup.ksh
index 6e438d88d..8146f773a 100755
--- a/tests/zfs-tests/tests/functional/mmp/cleanup.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/cleanup.ksh
@@ -23,6 +23,6 @@
verify_runnable "global"
-log_must set_tunable64 zfs_multihost_history 0
+log_must set_tunable64 zfs_multihost_history $MMP_HISTORY_OFF
log_pass "mmp cleanup passed"
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp.cfg b/tests/zfs-tests/tests/functional/mmp/mmp.cfg
index 52680c275..9f7e76e27 100644
--- a/tests/zfs-tests/tests/functional/mmp/mmp.cfg
+++ b/tests/zfs-tests/tests/functional/mmp/mmp.cfg
@@ -38,3 +38,9 @@ export MMP_HISTORY_OFF=0
export MMP_INTERVAL_HOUR=$((60*60*1000))
export MMP_INTERVAL_DEFAULT=1000
export MMP_INTERVAL_MIN=100
+
+export MMP_IMPORT_INTERVALS=20
+export MMP_FAIL_INTERVALS_DEFAULT=10
+export MMP_FAIL_INTERVALS_MIN=2
+
+export MMP_TEST_DURATION_DEFAULT=$((MMP_IMPORT_INTERVALS*MMP_INTERVAL_DEFAULT/1000))
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp.kshlib b/tests/zfs-tests/tests/functional/mmp/mmp.kshlib
index e74f04a5b..fda57c002 100644
--- a/tests/zfs-tests/tests/functional/mmp/mmp.kshlib
+++ b/tests/zfs-tests/tests/functional/mmp/mmp.kshlib
@@ -162,15 +162,42 @@ function mmp_pool_set_hostid # pool hostid
return 0
}
-
# Return the number of seconds the activity check portion of the import process
-# will take. Does not include the time to find devices and assemble the
-# preliminary pool configuration passed into the kernel.
+# will take. Does not include the time to find devices and assemble a config.
+# Note that the activity check may be skipped, e.g. if the pool and host
+# hostid's match, but this will return non-zero because mmp_* are populated.
function seconds_mmp_waits_for_activity
{
+ typeset pool=$1
+ typeset devpath=$2
+
+ typeset seconds=0
+ typeset devices=${#DISK[@]}
typeset import_intervals=$(get_tunable zfs_multihost_import_intervals)
- typeset interval=$(get_tunable zfs_multihost_interval)
- typeset seconds=$((interval*import_intervals/1000))
+ typeset import_interval=$(get_tunable zfs_multihost_interval)
+ typeset tmpfile=$(mktemp)
+ typeset mmp_fail
+ typeset mmp_write
+ typeset mmp_delay
+
+ log_must zdb -e -p $devpath $pool >$tmpfile 2>/dev/null
+ mmp_fail=$(awk '/mmp_fail/ {print $NF}' $tmpfile)
+ mmp_write=$(awk '/mmp_write/ {print $NF}' $tmpfile)
+ mmp_delay=$(awk '/mmp_delay/ {print $NF}' $tmpfile)
+ if [ -f $tmpfile ]; then
+ rm $tmpfile
+ fi
+
+ # In order of preference:
+ if [ -n $mmp_fail -a -n $mmp_write ]; then
+ seconds=$((2*mmp_fail*mmp_write/1000))
+ elif [ -n $mmp_delay ]; then
+ # MMP V0: Based on mmp_delay from the best Uberblock
+ seconds=$((import_intervals*devices*mmp_delay/1000000000))
+ else
+ # Non-MMP aware: Based on zfs_multihost_interval and import_intervals
+ seconds=$((import_intervals*import_interval/1000))
+ fi
echo $seconds
}
@@ -180,34 +207,33 @@ function import_no_activity_check # pool opts
typeset pool=$1
typeset opts=$2
- typeset max_duration=$(seconds_mmp_waits_for_activity)
+ typeset max_duration=$((MMP_TEST_DURATION_DEFAULT-1))
SECONDS=0
zpool import $opts $pool
typeset rc=$?
if [[ $SECONDS -gt $max_duration ]]; then
- log_fail "unexpected activity check (${SECONDS}s gt \
-$max_duration)"
+ log_fail "ERROR: import_no_activity_check unexpected activity \
+check (${SECONDS}s gt $max_duration)"
fi
return $rc
}
-function import_activity_check # pool opts
+function import_activity_check # pool opts act_test_duration
{
typeset pool=$1
typeset opts=$2
-
- typeset min_duration=$(seconds_mmp_waits_for_activity)
+ typeset min_duration=${3:-$MMP_TEST_DURATION_DEFAULT}
SECONDS=0
zpool import $opts $pool
typeset rc=$?
if [[ $SECONDS -le $min_duration ]]; then
- log_fail "expected activity check (${SECONDS}s le \
-$min_duration)"
+ log_fail "ERROR: import_activity_check expected activity check \
+(${SECONDS}s le min_duration $min_duration)"
fi
return $rc
@@ -238,3 +264,70 @@ function count_mmp_writes # pool duration
sleep $duration
awk 'BEGIN {count=0}; $NF != "-" {count++}; END {print count};' "$hist_path"
}
+
+function summarize_uberblock_mmp # device
+{
+ typeset device=$1
+
+ zdb -luuuu $device | awk '
+ BEGIN {write_fail_present=0; write_fail_missing=0; uber_invalid=0;}
+ /Uberblock\[[0-9][0-9]*\]/ {delay=-99; write=-99; fail=-99; total++; if (/invalid/) {uber_invalid++};};
+ /mmp_fail/ {fail=$3};
+ /mmp_seq/ {seq=$3};
+ /mmp_write/ {write=$3};
+ /mmp_delay/ {delay=$3; if (delay==0) {delay_zero++};};
+ /mmp_valid/ && delay>0 && write>0 && fail>0 {write_fail_present++};
+ /mmp_valid/ && delay>0 && (write<=0 || fail<=0) {write_fail_missing++};
+ /mmp_valid/ && delay>0 && write<=0 {write_missing++};
+ /mmp_valid/ && delay>0 && fail<=0 {fail_missing++};
+ /mmp_valid/ && delay>0 && seq>0 {seq_nonzero++};
+ END {
+ print "total_uberblocks " total;
+ print "delay_zero " delay_zero;
+ print "write_fail_present " write_fail_present;
+ print "write_fail_missing " write_fail_missing;
+ print "write_missing " write_missing;
+ print "fail_missing " fail_missing;
+ print "seq_nonzero " seq_nonzero;
+ print "uberblock_invalid " uber_invalid;
+ }'
+}
+
+function count_mmp_write_fail_present # device
+{
+ typeset device=$1
+
+ summarize_uberblock_mmp $device | awk '/write_fail_present/ {print $NF}'
+}
+
+function count_mmp_write_fail_missing # device
+{
+ typeset device=$1
+
+ summarize_uberblock_mmp $device | awk '/write_fail_missing/ {print $NF}'
+}
+
+function verify_mmp_write_fail_present # device
+{
+ typeset device=$1
+
+ count=$(count_mmp_write_fail_present $device)
+ log_note "present count: $count"
+ if [ $count -eq 0 ]; then
+ summarize_uberblock_mmp $device
+ log_note "----- snip -----"
+ zdb -luuuu $device
+ log_note "----- snip -----"
+ log_fail "No Uberblocks contain valid mmp_write and fail values"
+ fi
+
+ count=$(count_mmp_write_fail_missing $device)
+ log_note "missing count: $count"
+ if [ $count -gt 0 ]; then
+ summarize_uberblock_mmp $device
+ log_note "----- snip -----"
+ zdb -luuuu $device
+ log_note "----- snip -----"
+ log_fail "Uberblocks missing mmp_write or mmp_fail"
+ fi
+}
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp_active_import.ksh b/tests/zfs-tests/tests/functional/mmp/mmp_active_import.ksh
index e39c5ab30..c4ed89482 100755
--- a/tests/zfs-tests/tests/functional/mmp/mmp_active_import.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/mmp_active_import.ksh
@@ -42,8 +42,19 @@ verify_runnable "both"
function cleanup
{
mmp_pool_destroy $MMP_POOL $MMP_DIR
- log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
log_must mmp_clear_hostid
+ ZTESTPID=$(pgrep ztest)
+ if [ -n "$ZTESTPID" ]; then
+ for pid in $ZTESTPID; do
+ log_must kill -9 $pid
+ done
+ else
+ # if ztest not running and log present, ztest crashed
+ if [ -f $MMP_ZTEST_LOG ]; then
+ log_note "ztest appears to have crashed. Tail of log:"
+ tail -n 50 $MMP_ZTEST_LOG
+ fi
+ fi
}
log_assert "multihost=on|off active pool activity checks"
@@ -55,7 +66,6 @@ mmp_pool_create $MMP_POOL $MMP_DIR
# 2. Verify 'zpool import' reports an active pool.
log_must mmp_set_hostid $HOSTID2
-log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_MIN
log_must is_pool_imported $MMP_POOL "-d $MMP_DIR"
# 3. Verify 'zpool import [-f] $MMP_POOL' cannot import the pool.
@@ -79,6 +89,9 @@ if [ -n "$ZTESTPID" ]; then
log_must kill -9 $ZTESTPID
fi
log_must wait_pool_imported $MMP_POOL "-d $MMP_DIR"
+if [ -f $MMP_ZTEST_LOG ]; then
+ log_must rm $MMP_ZTEST_LOG
+fi
# 5. Verify 'zpool import' fails with the expected error message, when
# - hostid=0: - configuration error
@@ -103,9 +116,6 @@ MMP_IMPORTED_MSG="pool was previously in use from another system."
log_must try_pool_import $MMP_POOL "-d $MMP_DIR" "$MMP_IMPORTED_MSG"
# 7. Verify 'zpool import -f $MMP_POOL' can now import the pool.
-# Default interval results in minimum activity test 10s which
-# makes detection of the activity test reliable.
-log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
log_must import_activity_check $MMP_POOL "-f -d $MMP_DIR"
# 8 Verify pool may be exported/imported without -f argument.
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp_inactive_import.ksh b/tests/zfs-tests/tests/functional/mmp/mmp_inactive_import.ksh
index c5c66373e..64ed9bf97 100755
--- a/tests/zfs-tests/tests/functional/mmp/mmp_inactive_import.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/mmp_inactive_import.ksh
@@ -28,7 +28,9 @@
# 4. Verify multihost=off and hostid allowed (no activity check)
# 5. Verify multihost=on and hostids match (no activity check)
# 6. Verify multihost=on and hostids differ (activity check)
-# 7. Verify multihost=on and hostid zero fails (no activity check)
+# 7. Verify mmp_write and mmp_fail are set correctly
+# 8. Verify multihost=on and hostid zero fails (no activity check)
+# 9. Verify activity check duration based on mmp_write and mmp_fail
#
. $STF_SUITE/include/libtest.shlib
@@ -41,6 +43,7 @@ function cleanup
{
default_cleanup_noexit
log_must mmp_clear_hostid
+ log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
}
log_assert "multihost=on|off inactive pool activity checks"
@@ -87,11 +90,22 @@ log_must mmp_set_hostid $HOSTID2
log_mustnot import_activity_check $TESTPOOL ""
log_must import_activity_check $TESTPOOL "-f"
-# 7. Verify multihost=on and hostid zero fails (no activity check)
+# 7. Verify mmp_write and mmp_fail are set correctly
log_must zpool export -F $TESTPOOL
+log_must verify_mmp_write_fail_present ${DISK[0]}
+
+# 8. Verify multihost=on and hostid zero fails (no activity check)
log_must mmp_clear_hostid
MMP_IMPORTED_MSG="Set a unique system hostid"
log_must check_pool_import $TESTPOOL "-f" "action" "$MMP_IMPORTED_MSG"
log_mustnot import_no_activity_check $TESTPOOL "-f"
+# 9. Verify activity check duration based on mmp_write and mmp_fail
+# Specify a short test via tunables but import pool imported while
+# tunables set to default duration.
+log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_MIN
+log_must mmp_clear_hostid
+log_must mmp_set_hostid $HOSTID1
+log_must import_activity_check $TESTPOOL "-f" $MMP_TEST_DURATION_DEFAULT
+
log_pass "multihost=on|off inactive pool activity checks passed"
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp_on_uberblocks.ksh b/tests/zfs-tests/tests/functional/mmp/mmp_on_uberblocks.ksh
index 0cb38f889..bf1eb54a7 100755
--- a/tests/zfs-tests/tests/functional/mmp/mmp_on_uberblocks.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/mmp_on_uberblocks.ksh
@@ -19,7 +19,7 @@
#
# DESCRIPTION:
-# Ensure that MMP updates uberblocks at the expected intervals.
+# Ensure that MMP updates uberblocks with MMP info at expected intervals.
#
# STRATEGY:
# 1. Set zfs_txg_timeout to large value
@@ -28,6 +28,7 @@
# 4. Sleep, then collect count of uberblocks written
# 5. If number of changes seen is less than min threshold, then fail
# 6. If number of changes seen is more than max threshold, then fail
+# 7. Sequence number increments when no TXGs are syncing
#
. $STF_SUITE/include/libtest.shlib
@@ -39,12 +40,14 @@ verify_runnable "both"
UBER_CHANGES=0
EXPECTED=$(($(echo $DISKS | wc -w) * 10))
FUDGE=$((EXPECTED * 20 / 100))
-MIN=$((EXPECTED - FUDGE))
-MAX=$((EXPECTED + FUDGE))
+MIN_UB_WRITES=$((EXPECTED - FUDGE))
+MAX_UB_WRITES=$((EXPECTED + FUDGE))
+MIN_SEQ_VALUES=7
function cleanup
{
default_cleanup_noexit
+ log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
set_tunable64 zfs_txg_timeout $TXG_TIMEOUT_DEFAULT
log_must mmp_clear_hostid
}
@@ -62,12 +65,21 @@ UBER_CHANGES=$(count_mmp_writes $TESTPOOL 10)
log_note "Uberblock changed $UBER_CHANGES times"
-if [ $UBER_CHANGES -lt $MIN ]; then
+if [ $UBER_CHANGES -lt $MIN_UB_WRITES ]; then
log_fail "Fewer uberblock writes occured than expected ($EXPECTED)"
fi
-if [ $UBER_CHANGES -gt $MAX ]; then
+if [ $UBER_CHANGES -gt $MAX_UB_WRITES ]; then
log_fail "More uberblock writes occured than expected ($EXPECTED)"
fi
+log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_MIN
+SEQ_BEFORE=$(zdb -luuuu ${DISK[0]} | awk '/mmp_seq/ {if ($NF>max) max=$NF}; END {print max}')
+sleep 1
+SEQ_AFTER=$(zdb -luuuu ${DISK[0]} | awk '/mmp_seq/ {if ($NF>max) max=$NF}; END {print max}')
+if [ $((SEQ_AFTER - SEQ_BEFORE)) -lt $MIN_SEQ_VALUES ]; then
+ zdb -luuuu ${DISK[0]}
+ log_fail "ERROR: mmp_seq did not increase by $MIN_SEQ_VALUES; before $SEQ_BEFORE after $SEQ_AFTER"
+fi
+
log_pass "Ensure MMP uberblocks update at the correct interval passed"
diff --git a/tests/zfs-tests/tests/functional/mmp/mmp_reset_interval.ksh b/tests/zfs-tests/tests/functional/mmp/mmp_reset_interval.ksh
index 3c8f00cde..842df284b 100755
--- a/tests/zfs-tests/tests/functional/mmp/mmp_reset_interval.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/mmp_reset_interval.ksh
@@ -20,7 +20,8 @@
# DESCRIPTION:
# Ensure that the MMP thread is notified when zfs_multihost_interval is
-# reduced.
+# reduced, and that changes to zfs_multihost_interval and
+# zfs_multihost_fail_intervals do not trigger pool suspensions.
#
# STRATEGY:
# 1. Set zfs_multihost_interval to much longer than the test duration
@@ -29,6 +30,8 @@
# 4. Set zfs_multihost_interval to 1 second
# 5. Sleep briefly
# 6. Verify MMP writes began
+# 7. Verify mmp_fail and mmp_write in uberblock reflect tunables
+# 8. Repeatedly change tunables relating to pool suspension
#
. $STF_SUITE/include/libtest.shlib
@@ -41,6 +44,8 @@ function cleanup
{
default_cleanup_noexit
log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
+ log_must set_tunable64 zfs_multihost_fail_intervals \
+ $MMP_FAIL_INTERVALS_DEFAULT
log_must mmp_clear_hostid
}
@@ -58,7 +63,57 @@ log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
uber_count=$(count_mmp_writes $TESTPOOL 1)
if [ $uber_count -eq 0 ]; then
- log_fail "mmp writes did not start when zfs_multihost_interval reduced"
+ log_fail "ERROR: mmp writes did not start when zfs_multihost_interval reduced"
fi
+# 7. Verify mmp_write and mmp_fail are written
+for fails in $(seq $MMP_FAIL_INTERVALS_MIN $((MMP_FAIL_INTERVALS_MIN*2))); do
+ for interval in $(seq $MMP_INTERVAL_MIN 200 $MMP_INTERVAL_DEFAULT); do
+ log_must set_tunable64 zfs_multihost_fail_intervals $fails
+ log_must set_tunable64 zfs_multihost_interval $interval
+ log_must sync_pool $TESTPOOL
+ typeset mmp_fail=$(zdb $TESTPOOL 2>/dev/null |
+ awk '/mmp_fail/ {print $NF}')
+ if [ $fails -ne $mmp_fail ]; then
+ log_fail "ERROR: mmp_fail $mmp_fail != $fails"
+ fi
+ typeset mmp_write=$(zdb $TESTPOOL 2>/dev/null |
+ awk '/mmp_write/ {print $NF}')
+ if [ $interval -ne $mmp_write ]; then
+ log_fail "ERROR: mmp_write $mmp_write != $interval"
+ fi
+ done
+done
+
+
+# 8. Repeatedly change zfs_multihost_interval and fail_intervals
+for x in $(seq 10); do
+ typeset new_interval=$(( (RANDOM % 20 + 1) * $MMP_INTERVAL_MIN ))
+ log_must set_tunable64 zfs_multihost_interval $new_interval
+ typeset action=$((RANDOM %10))
+ if [ $action -eq 0 ]; then
+ log_must zpool export -a
+ log_must mmp_clear_hostid
+ log_must mmp_set_hostid $HOSTID1
+ log_must zpool import $TESTPOOL
+ elif [ $action -eq 1 ]; then
+ log_must zpool export -F $TESTPOOL
+ log_must zpool import $TESTPOOL
+ elif [ $action -eq 2 ]; then
+ log_must zpool export -F $TESTPOOL
+ log_must mmp_clear_hostid
+ log_must mmp_set_hostid $HOSTID2
+ log_must zpool import -f $TESTPOOL
+ elif [ $action -eq 3 ]; then
+ log_must zpool export -F $TESTPOOL
+ log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_MIN
+ log_must zpool import $TESTPOOL
+ elif [ $action -eq 4 ]; then
+ log_must set_tunable64 zfs_multihost_fail_intervals \
+ $((RANDOM % MMP_FAIL_INTERVALS_DEFAULT))
+ fi
+ sleep 5
+done
+
+
log_pass "mmp threads notified when zfs_multihost_interval reduced"
diff --git a/tests/zfs-tests/tests/functional/mmp/setup.ksh b/tests/zfs-tests/tests/functional/mmp/setup.ksh
index fde5e3bb7..c91f61979 100755
--- a/tests/zfs-tests/tests/functional/mmp/setup.ksh
+++ b/tests/zfs-tests/tests/functional/mmp/setup.ksh
@@ -28,5 +28,7 @@ if [ -e $HOSTID_FILE ]; then
fi
log_must set_tunable64 zfs_multihost_history $MMP_HISTORY
+log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
+log_must set_tunable64 zfs_multihost_fail_intervals $MMP_FAIL_INTERVALS_DEFAULT
log_pass "mmp setup pass"
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-17 03:11:29 +0000