Initialize mmp_last_write when the mmp thread starts

A great deal of time may go by between when mmp_init() is called and
the MMP thread starts, particularly if there are bad devices, because
there is I/O checking configs etc.  If this time is too long,

    (gethrtime() - mmp_last_write) > mmp_fail_ns

at the time the MMP thread starts.  If MMP is configured to suspend
the pool, the pool will be suspended immediately.

This can be seen in issue #10838

The value of mmp_last_write doesn't matter before the mmp thread
starts.  To give the MMP thread time to issue and land MMP writes,
initialize mmp_last_write when the MMP thread starts.

Reviewed-by: Giuseppe Di Natale <guss80@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Olaf Faaland <faaland1@llnl.gov>
Closes #10873

Author: ofaaland

module/zfs/mmp.c

@@ -198,14 +198,6 @@ 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
@@ -557,6 +549,18 @@ mmp_thread(void *arg)
 
 	mmp_thread_enter(mmp, &cpr);
 
+	/*
+	 * There have been no MMP writes yet.  Setting mmp_last_write here gives
+	 * us one mmp_fail_ns period, which is consistent with the activity
+	 * check duration, to try to land an MMP write before MMP suspends the
+	 * pool (if so configured).
+	 */
+
+	mutex_enter(&mmp->mmp_io_lock);
+	mmp->mmp_last_write = gethrtime();
+	mmp->mmp_delay = MSEC2NSEC(MMP_INTERVAL_OK(zfs_multihost_interval));
+	mutex_exit(&mmp->mmp_io_lock);
+
 	while (!mmp->mmp_thread_exiting) {
 		hrtime_t next_time = gethrtime() +
 		    MSEC2NSEC(MMP_DEFAULT_INTERVAL);