Fix cpu hotplug atomic sleep issue

We move the spinlock unlock before the thread creation. This should be safe because the thread creation code doesn't actually manipulate any taskq data structures; that's done by the thread once it's created. We also remove the assertion that the maxthreads is the current threads plus one; that assertion could fail if multiple hotplug events come in quick succession, and the first new taskq thread hasn't had a chance to start processing yet. Reviewed-by: Brian Behlendorf <[email protected]> Reviewed-by: Matthew Ahrens <[email protected]> eviewed-by: Tony Nguyen <[email protected]> Signed-off-by: Paul Dagnelie <[email protected]> Closes openzfs#12714
tonyhutter · Feb 16, 2022 · 7bd292e · 7bd292e
1 parent 5c80a25
commit 7bd292e
Showing 1 changed file with 6 additions and 5 deletions.
diff --git a/module/os/linux/spl/spl-taskq.c b/module/os/linux/spl/spl-taskq.c
@@ -1298,8 +1298,10 @@ spl_taskq_expand(unsigned int cpu, struct hlist_node *node)
 	ASSERT(tq);
 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
 
-	if (!(tq->tq_flags & TASKQ_ACTIVE))
-		goto out;
+	if (!(tq->tq_flags & TASKQ_ACTIVE)) {
+		spin_unlock_irqrestore(&tq->tq_lock, flags);
+		return (err);
+	}
 
 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
 	int nthreads = MIN(tq->tq_cpu_pct, 100);
@@ -1308,13 +1310,12 @@ spl_taskq_expand(unsigned int cpu, struct hlist_node *node)
 
 	if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
 	    tq->tq_maxthreads > tq->tq_nthreads) {
-		ASSERT3U(tq->tq_maxthreads, ==, tq->tq_nthreads + 1);
+		spin_unlock_irqrestore(&tq->tq_lock, flags);
 		taskq_thread_t *tqt = taskq_thread_create(tq);
 		if (tqt == NULL)
 			err = -1;
+		return (err);
 	}
-
-out:
 	spin_unlock_irqrestore(&tq->tq_lock, flags);
 	return (err);
 }