]> git.proxmox.com Git - mirror_spl.git/blobdiff - module/spl/spl-taskq.c
Don't use tq->tq_lock_flags
[mirror_spl.git] / module / spl / spl-taskq.c
index 49bb40a25123a52eeb03e2062bea3e6574f0a449..ded6d3b80c70114b18ee2521cfe4761aae8805ca 100644 (file)
@@ -31,10 +31,29 @@ int spl_taskq_thread_bind = 0;
 module_param(spl_taskq_thread_bind, int, 0644);
 MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default");
 
+
+int spl_taskq_thread_dynamic = 1;
+module_param(spl_taskq_thread_dynamic, int, 0644);
+MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads");
+
+int spl_taskq_thread_priority = 1;
+module_param(spl_taskq_thread_priority, int, 0644);
+MODULE_PARM_DESC(spl_taskq_thread_priority,
+    "Allow non-default priority for taskq threads");
+
+int spl_taskq_thread_sequential = 4;
+module_param(spl_taskq_thread_sequential, int, 0644);
+MODULE_PARM_DESC(spl_taskq_thread_sequential,
+    "Create new taskq threads after N sequential tasks");
+
 /* Global system-wide dynamic task queue available for all consumers */
 taskq_t *system_taskq;
 EXPORT_SYMBOL(system_taskq);
 
+/* Private dedicated taskq for creating new taskq threads on demand. */
+static taskq_t *dynamic_taskq;
+static taskq_thread_t *taskq_thread_create(taskq_t *);
+
 static int
 task_km_flags(uint_t flags)
 {
@@ -52,7 +71,7 @@ task_km_flags(uint_t flags)
  * is not attached to the free, work, or pending taskq lists.
  */
 static taskq_ent_t *
-task_alloc(taskq_t *tq, uint_t flags)
+task_alloc(taskq_t *tq, uint_t flags, unsigned long *irqflags)
 {
        taskq_ent_t *t;
        int count = 0;
@@ -92,18 +111,19 @@ retry:
                 * end up delaying the task allocation by one second, thereby
                 * throttling the task dispatch rate.
                 */
-               spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+               spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
                schedule_timeout(HZ / 100);
-               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               spin_lock_irqsave_nested(&tq->tq_lock, *irqflags,
+                   tq->tq_lock_class);
                if (count < 100) {
                        count++;
                        goto retry;
                }
        }
 
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
        t = kmem_alloc(sizeof(taskq_ent_t), task_km_flags(flags));
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class);
 
        if (t) {
                taskq_init_ent(t);
@@ -168,12 +188,13 @@ task_expire(unsigned long data)
        taskq_ent_t *w, *t = (taskq_ent_t *)data;
        taskq_t *tq = t->tqent_taskq;
        struct list_head *l;
+       unsigned long flags;
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
 
        if (t->tqent_flags & TQENT_FLAG_CANCEL) {
                ASSERT(list_empty(&t->tqent_list));
-               spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+               spin_unlock_irqrestore(&tq->tq_lock, flags);
                return;
        }
 
@@ -192,7 +213,7 @@ task_expire(unsigned long data)
        if (l == &tq->tq_prio_list)
                list_add(&t->tqent_list, &tq->tq_prio_list);
 
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        wake_up(&tq->tq_work_waitq);
 }
@@ -359,10 +380,11 @@ taskq_wait_id_check(taskq_t *tq, taskqid_t id)
 {
        int active = 0;
        int rc;
+       unsigned long flags;
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
        rc = (taskq_find(tq, id, &active) == NULL);
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        return (rc);
 }
@@ -382,10 +404,11 @@ static int
 taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id)
 {
        int rc;
+       unsigned long flags;
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
        rc = (id < tq->tq_lowest_id);
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        return (rc);
 }
@@ -409,10 +432,11 @@ static int
 taskq_wait_check(taskq_t *tq)
 {
        int rc;
+       unsigned long flags;
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
        rc = (tq->tq_lowest_id == tq->tq_next_id);
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        return (rc);
 }
@@ -429,22 +453,38 @@ taskq_wait(taskq_t *tq)
 }
 EXPORT_SYMBOL(taskq_wait);
 
-int
-taskq_member(taskq_t *tq, void *t)
+static int
+taskq_member_impl(taskq_t *tq, void *t)
 {
        struct list_head *l;
        taskq_thread_t *tqt;
+       int found = 0;
 
        ASSERT(tq);
        ASSERT(t);
+       ASSERT(spin_is_locked(&tq->tq_lock));
 
        list_for_each(l, &tq->tq_thread_list) {
                tqt = list_entry(l, taskq_thread_t, tqt_thread_list);
-               if (tqt->tqt_thread == (struct task_struct *)t)
-                       return (1);
+               if (tqt->tqt_thread == (struct task_struct *)t) {
+                       found = 1;
+                       break;
+               }
        }
+       return (found);
+}
 
-       return (0);
+int
+taskq_member(taskq_t *tq, void *t)
+{
+       int found;
+       unsigned long flags;
+
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
+       found = taskq_member_impl(tq, t);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
+
+       return (found);
 }
 EXPORT_SYMBOL(taskq_member);
 
@@ -460,10 +500,11 @@ taskq_cancel_id(taskq_t *tq, taskqid_t id)
        taskq_ent_t *t;
        int active = 0;
        int rc = ENOENT;
+       unsigned long flags;
 
        ASSERT(tq);
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
        t = taskq_find(tq, id, &active);
        if (t && !active) {
                list_del_init(&t->tqent_list);
@@ -483,9 +524,10 @@ taskq_cancel_id(taskq_t *tq, taskqid_t id)
                 * drop the lock before synchronously cancelling the timer.
                 */
                if (timer_pending(&t->tqent_timer)) {
-                       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+                       spin_unlock_irqrestore(&tq->tq_lock, flags);
                        del_timer_sync(&t->tqent_timer);
-                       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+                       spin_lock_irqsave_nested(&tq->tq_lock, flags,
+                           tq->tq_lock_class);
                }
 
                if (!(t->tqent_flags & TQENT_FLAG_PREALLOC))
@@ -493,7 +535,7 @@ taskq_cancel_id(taskq_t *tq, taskqid_t id)
 
                rc = 0;
        }
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        if (active) {
                taskq_wait_id(tq, id);
@@ -504,19 +546,22 @@ taskq_cancel_id(taskq_t *tq, taskqid_t id)
 }
 EXPORT_SYMBOL(taskq_cancel_id);
 
+static int taskq_thread_spawn(taskq_t *tq);
+
 taskqid_t
 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
 {
        taskq_ent_t *t;
        taskqid_t rc = 0;
+       unsigned long irqflags;
 
        ASSERT(tq);
        ASSERT(func);
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
 
        /* Taskq being destroyed and all tasks drained */
-       if (!(tq->tq_flags & TQ_ACTIVE))
+       if (!(tq->tq_flags & TASKQ_ACTIVE))
                goto out;
 
        /* Do not queue the task unless there is idle thread for it */
@@ -524,7 +569,7 @@ taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
        if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads))
                goto out;
 
-       if ((t = task_alloc(tq, flags)) == NULL)
+       if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
                goto out;
 
        spin_lock(&t->tqent_lock);
@@ -550,7 +595,11 @@ taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
 
        wake_up(&tq->tq_work_waitq);
 out:
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       /* Spawn additional taskq threads if required. */
+       if (tq->tq_nactive == tq->tq_nthreads)
+               (void) taskq_thread_spawn(tq);
+
+       spin_unlock_irqrestore(&tq->tq_lock, irqflags);
        return (rc);
 }
 EXPORT_SYMBOL(taskq_dispatch);
@@ -561,17 +610,18 @@ taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
 {
        taskqid_t rc = 0;
        taskq_ent_t *t;
+       unsigned long irqflags;
 
        ASSERT(tq);
        ASSERT(func);
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
 
        /* Taskq being destroyed and all tasks drained */
-       if (!(tq->tq_flags & TQ_ACTIVE))
+       if (!(tq->tq_flags & TASKQ_ACTIVE))
                goto out;
 
-       if ((t = task_alloc(tq, flags)) == NULL)
+       if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
                goto out;
 
        spin_lock(&t->tqent_lock);
@@ -593,7 +643,10 @@ taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
 
        spin_unlock(&t->tqent_lock);
 out:
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       /* Spawn additional taskq threads if required. */
+       if (tq->tq_nactive == tq->tq_nthreads)
+               (void) taskq_thread_spawn(tq);
+       spin_unlock_irqrestore(&tq->tq_lock, irqflags);
        return (rc);
 }
 EXPORT_SYMBOL(taskq_dispatch_delay);
@@ -602,14 +655,15 @@ void
 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
    taskq_ent_t *t)
 {
+       unsigned long irqflags;
        ASSERT(tq);
        ASSERT(func);
-       ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
+           tq->tq_lock_class);
 
        /* Taskq being destroyed and all tasks drained */
-       if (!(tq->tq_flags & TQ_ACTIVE)) {
+       if (!(tq->tq_flags & TASKQ_ACTIVE)) {
                t->tqent_id = 0;
                goto out;
        }
@@ -638,7 +692,10 @@ taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
 
        wake_up(&tq->tq_work_waitq);
 out:
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       /* Spawn additional taskq threads if required. */
+       if (tq->tq_nactive == tq->tq_nthreads)
+               (void) taskq_thread_spawn(tq);
+       spin_unlock_irqrestore(&tq->tq_lock, irqflags);
 }
 EXPORT_SYMBOL(taskq_dispatch_ent);
 
@@ -664,6 +721,97 @@ taskq_init_ent(taskq_ent_t *t)
 }
 EXPORT_SYMBOL(taskq_init_ent);
 
+/*
+ * Return the next pending task, preference is given to tasks on the
+ * priority list which were dispatched with TQ_FRONT.
+ */
+static taskq_ent_t *
+taskq_next_ent(taskq_t *tq)
+{
+       struct list_head *list;
+
+       ASSERT(spin_is_locked(&tq->tq_lock));
+
+       if (!list_empty(&tq->tq_prio_list))
+               list = &tq->tq_prio_list;
+       else if (!list_empty(&tq->tq_pend_list))
+               list = &tq->tq_pend_list;
+       else
+               return (NULL);
+
+       return (list_entry(list->next, taskq_ent_t, tqent_list));
+}
+
+/*
+ * Spawns a new thread for the specified taskq.
+ */
+static void
+taskq_thread_spawn_task(void *arg)
+{
+       taskq_t *tq = (taskq_t *)arg;
+       unsigned long flags;
+
+       (void) taskq_thread_create(tq);
+
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
+       tq->tq_nspawn--;
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
+}
+
+/*
+ * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current
+ * number of threads is insufficient to handle the pending tasks.  These
+ * new threads must be created by the dedicated dynamic_taskq to avoid
+ * deadlocks between thread creation and memory reclaim.  The system_taskq
+ * which is also a dynamic taskq cannot be safely used for this.
+ */
+static int
+taskq_thread_spawn(taskq_t *tq)
+{
+       int spawning = 0;
+
+       if (!(tq->tq_flags & TASKQ_DYNAMIC))
+               return (0);
+
+       if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) &&
+           (tq->tq_flags & TASKQ_ACTIVE)) {
+               spawning = (++tq->tq_nspawn);
+               taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task,
+                   tq, TQ_NOSLEEP);
+       }
+
+       return (spawning);
+}
+
+/*
+ * Threads in a dynamic taskq should only exit once it has been completely
+ * drained and no other threads are actively servicing tasks.  This prevents
+ * threads from being created and destroyed more than is required.
+ *
+ * The first thread is the thread list is treated as the primary thread.
+ * There is nothing special about the primary thread but in order to avoid
+ * all the taskq pids from changing we opt to make it long running.
+ */
+static int
+taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt)
+{
+       ASSERT(spin_is_locked(&tq->tq_lock));
+
+       if (!(tq->tq_flags & TASKQ_DYNAMIC))
+               return (0);
+
+       if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t,
+           tqt_thread_list) == tqt)
+               return (0);
+
+       return
+           ((tq->tq_nspawn == 0) &&    /* No threads are being spawned */
+           (tq->tq_nactive == 0) &&    /* No threads are handling tasks */
+           (tq->tq_nthreads > 1) &&    /* More than 1 thread is running */
+           (!taskq_next_ent(tq)) &&    /* There are no pending tasks */
+           (spl_taskq_thread_dynamic));/* Dynamic taskqs are allowed */
+}
+
 static int
 taskq_thread(void *args)
 {
@@ -672,18 +820,30 @@ taskq_thread(void *args)
        taskq_thread_t *tqt = args;
        taskq_t *tq;
        taskq_ent_t *t;
-       struct list_head *pend_list;
+       int seq_tasks = 0;
+       unsigned long flags;
 
        ASSERT(tqt);
+       ASSERT(tqt->tqt_tq);
        tq = tqt->tqt_tq;
        current->flags |= PF_NOFREEZE;
 
+       #if defined(PF_MEMALLOC_NOIO)
+       (void) memalloc_noio_save();
+       #endif
+
        sigfillset(&blocked);
        sigprocmask(SIG_BLOCK, &blocked, NULL);
        flush_signals(current);
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
+
+       /* Immediately exit if more threads than allowed were created. */
+       if (tq->tq_nthreads >= tq->tq_maxthreads)
+               goto error;
+
        tq->tq_nthreads++;
+       list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list);
        wake_up(&tq->tq_wait_waitq);
        set_current_state(TASK_INTERRUPTIBLE);
 
@@ -691,25 +851,26 @@ taskq_thread(void *args)
 
                if (list_empty(&tq->tq_pend_list) &&
                    list_empty(&tq->tq_prio_list)) {
+
+                       if (taskq_thread_should_stop(tq, tqt)) {
+                               wake_up_all(&tq->tq_wait_waitq);
+                               break;
+                       }
+
                        add_wait_queue_exclusive(&tq->tq_work_waitq, &wait);
-                       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+                       spin_unlock_irqrestore(&tq->tq_lock, flags);
+
                        schedule();
-                       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+                       seq_tasks = 0;
+
+                       spin_lock_irqsave_nested(&tq->tq_lock, flags,
+                           tq->tq_lock_class);
                        remove_wait_queue(&tq->tq_work_waitq, &wait);
                } else {
                        __set_current_state(TASK_RUNNING);
                }
 
-
-               if (!list_empty(&tq->tq_prio_list))
-                       pend_list = &tq->tq_prio_list;
-               else if (!list_empty(&tq->tq_pend_list))
-                       pend_list = &tq->tq_pend_list;
-               else
-                       pend_list = NULL;
-
-               if (pend_list) {
-                       t = list_entry(pend_list->next,taskq_ent_t,tqent_list);
+               if ((t = taskq_next_ent(tq)) != NULL) {
                        list_del_init(&t->tqent_list);
 
                        /* In order to support recursively dispatching a
@@ -727,19 +888,19 @@ taskq_thread(void *args)
 
                        taskq_insert_in_order(tq, tqt);
                        tq->tq_nactive++;
-                       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+                       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
                        /* Perform the requested task */
                        t->tqent_func(t->tqent_arg);
 
-                       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+                       spin_lock_irqsave_nested(&tq->tq_lock, flags,
+                           tq->tq_lock_class);
                        tq->tq_nactive--;
                        list_del_init(&tqt->tqt_active_list);
                        tqt->tqt_task = NULL;
 
                        /* For prealloc'd tasks, we don't free anything. */
-                       if ((tq->tq_flags & TASKQ_DYNAMIC) ||
-                           !(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
+                       if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
                                task_done(tq, t);
 
                        /* When the current lowest outstanding taskqid is
@@ -749,9 +910,17 @@ taskq_thread(void *args)
                                ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
                        }
 
+                       /* Spawn additional taskq threads if required. */
+                       if ((++seq_tasks) > spl_taskq_thread_sequential &&
+                           taskq_thread_spawn(tq))
+                               seq_tasks = 0;
+
                        tqt->tqt_id = 0;
                        tqt->tqt_flags = 0;
                        wake_up_all(&tq->tq_wait_waitq);
+               } else {
+                       if (taskq_thread_should_stop(tq, tqt))
+                               break;
                }
 
                set_current_state(TASK_INTERRUPTIBLE);
@@ -761,27 +930,58 @@ taskq_thread(void *args)
        __set_current_state(TASK_RUNNING);
        tq->tq_nthreads--;
        list_del_init(&tqt->tqt_thread_list);
-       kmem_free(tqt, sizeof(taskq_thread_t));
-
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+error:
+       kmem_free(tqt, sizeof (taskq_thread_t));
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
        return (0);
 }
 
+static taskq_thread_t *
+taskq_thread_create(taskq_t *tq)
+{
+       static int last_used_cpu = 0;
+       taskq_thread_t *tqt;
+
+       tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE);
+       INIT_LIST_HEAD(&tqt->tqt_thread_list);
+       INIT_LIST_HEAD(&tqt->tqt_active_list);
+       tqt->tqt_tq = tq;
+       tqt->tqt_id = 0;
+
+       tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt,
+           "%s", tq->tq_name);
+       if (tqt->tqt_thread == NULL) {
+               kmem_free(tqt, sizeof (taskq_thread_t));
+               return (NULL);
+       }
+
+       if (spl_taskq_thread_bind) {
+               last_used_cpu = (last_used_cpu + 1) % num_online_cpus();
+               kthread_bind(tqt->tqt_thread, last_used_cpu);
+       }
+
+       if (spl_taskq_thread_priority)
+               set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri));
+
+       wake_up_process(tqt->tqt_thread);
+
+       return (tqt);
+}
+
 taskq_t *
 taskq_create(const char *name, int nthreads, pri_t pri,
     int minalloc, int maxalloc, uint_t flags)
 {
-       static int last_used_cpu = 0;
        taskq_t *tq;
        taskq_thread_t *tqt;
-       int rc = 0, i, j = 0;
+       int count = 0, rc = 0, i;
+       unsigned long irqflags;
 
        ASSERT(name != NULL);
-       ASSERT(pri <= maxclsyspri);
        ASSERT(minalloc >= 0);
        ASSERT(maxalloc <= INT_MAX);
-       ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */
+       ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */
 
        /* Scale the number of threads using nthreads as a percentage */
        if (flags & TASKQ_THREADS_CPU_PCT) {
@@ -792,63 +992,57 @@ taskq_create(const char *name, int nthreads, pri_t pri,
                nthreads = MAX((num_online_cpus() * nthreads) / 100, 1);
        }
 
-       tq = kmem_alloc(sizeof(*tq), KM_PUSHPAGE);
+       tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE);
        if (tq == NULL)
                return (NULL);
 
        spin_lock_init(&tq->tq_lock);
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
        INIT_LIST_HEAD(&tq->tq_thread_list);
        INIT_LIST_HEAD(&tq->tq_active_list);
-       tq->tq_name      = name;
-       tq->tq_nactive   = 0;
-       tq->tq_nthreads  = 0;
-       tq->tq_pri       = pri;
-       tq->tq_minalloc  = minalloc;
-       tq->tq_maxalloc  = maxalloc;
-       tq->tq_nalloc    = 0;
-       tq->tq_flags     = (flags | TQ_ACTIVE);
-       tq->tq_next_id   = 1;
-       tq->tq_lowest_id = 1;
+       tq->tq_name       = strdup(name);
+       tq->tq_nactive    = 0;
+       tq->tq_nthreads   = 0;
+       tq->tq_nspawn     = 0;
+       tq->tq_maxthreads = nthreads;
+       tq->tq_pri        = pri;
+       tq->tq_minalloc   = minalloc;
+       tq->tq_maxalloc   = maxalloc;
+       tq->tq_nalloc     = 0;
+       tq->tq_flags      = (flags | TASKQ_ACTIVE);
+       tq->tq_next_id    = 1;
+       tq->tq_lowest_id  = 1;
        INIT_LIST_HEAD(&tq->tq_free_list);
        INIT_LIST_HEAD(&tq->tq_pend_list);
        INIT_LIST_HEAD(&tq->tq_prio_list);
        INIT_LIST_HEAD(&tq->tq_delay_list);
        init_waitqueue_head(&tq->tq_work_waitq);
        init_waitqueue_head(&tq->tq_wait_waitq);
+       tq->tq_lock_class = TQ_LOCK_GENERAL;
+
+       if (flags & TASKQ_PREPOPULATE) {
+               spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
+                   tq->tq_lock_class);
 
-       if (flags & TASKQ_PREPOPULATE)
                for (i = 0; i < minalloc; i++)
-                       task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW));
+                       task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW,
+                           &irqflags));
 
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+               spin_unlock_irqrestore(&tq->tq_lock, irqflags);
+       }
+
+       if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic)
+               nthreads = 1;
 
        for (i = 0; i < nthreads; i++) {
-               tqt = kmem_alloc(sizeof(*tqt), KM_PUSHPAGE);
-               INIT_LIST_HEAD(&tqt->tqt_thread_list);
-               INIT_LIST_HEAD(&tqt->tqt_active_list);
-               tqt->tqt_tq = tq;
-               tqt->tqt_id = 0;
-
-               tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt,
-                   "%s/%d", name, i);
-               if (tqt->tqt_thread) {
-                       list_add(&tqt->tqt_thread_list, &tq->tq_thread_list);
-                       if (spl_taskq_thread_bind) {
-                               last_used_cpu = (last_used_cpu + 1) % num_online_cpus();
-                               kthread_bind(tqt->tqt_thread, last_used_cpu);
-                       }
-                       set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(pri));
-                       wake_up_process(tqt->tqt_thread);
-                       j++;
-               } else {
-                       kmem_free(tqt, sizeof(taskq_thread_t));
+               tqt = taskq_thread_create(tq);
+               if (tqt == NULL)
                        rc = 1;
-               }
+               else
+                       count++;
        }
 
        /* Wait for all threads to be started before potential destroy */
-       wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j);
+       wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count);
 
        if (rc) {
                taskq_destroy(tq);
@@ -865,16 +1059,23 @@ taskq_destroy(taskq_t *tq)
        struct task_struct *thread;
        taskq_thread_t *tqt;
        taskq_ent_t *t;
+       unsigned long flags;
 
        ASSERT(tq);
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-       tq->tq_flags &= ~TQ_ACTIVE;
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
+       tq->tq_flags &= ~TASKQ_ACTIVE;
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
+
+       /*
+        * When TASKQ_ACTIVE is clear new tasks may not be added nor may
+        * new worker threads be spawned for dynamic taskq.
+        */
+       if (dynamic_taskq != NULL)
+               taskq_wait_outstanding(dynamic_taskq, 0);
 
-       /* TQ_ACTIVE cleared prevents new tasks being added to pending */
        taskq_wait(tq);
 
-       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
 
        /*
         * Signal each thread to exit and block until it does.  Each thread
@@ -884,13 +1085,14 @@ taskq_destroy(taskq_t *tq)
         */
        while (!list_empty(&tq->tq_thread_list)) {
                tqt = list_entry(tq->tq_thread_list.next,
-                                taskq_thread_t, tqt_thread_list);
+                   taskq_thread_t, tqt_thread_list);
                thread = tqt->tqt_thread;
-               spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+               spin_unlock_irqrestore(&tq->tq_lock, flags);
 
                kthread_stop(thread);
 
-               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               spin_lock_irqsave_nested(&tq->tq_lock, flags,
+                   tq->tq_lock_class);
        }
 
        while (!list_empty(&tq->tq_free_list)) {
@@ -902,8 +1104,9 @@ taskq_destroy(taskq_t *tq)
                task_free(tq, t);
        }
 
-       ASSERT(tq->tq_nthreads == 0);
-       ASSERT(tq->tq_nalloc == 0);
+       ASSERT0(tq->tq_nthreads);
+       ASSERT0(tq->tq_nalloc);
+       ASSERT0(tq->tq_nspawn);
        ASSERT(list_empty(&tq->tq_thread_list));
        ASSERT(list_empty(&tq->tq_active_list));
        ASSERT(list_empty(&tq->tq_free_list));
@@ -911,27 +1114,43 @@ taskq_destroy(taskq_t *tq)
        ASSERT(list_empty(&tq->tq_prio_list));
        ASSERT(list_empty(&tq->tq_delay_list));
 
-       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       spin_unlock_irqrestore(&tq->tq_lock, flags);
 
-       kmem_free(tq, sizeof(taskq_t));
+       strfree(tq->tq_name);
+       kmem_free(tq, sizeof (taskq_t));
 }
 EXPORT_SYMBOL(taskq_destroy);
 
 int
 spl_taskq_init(void)
 {
-       /* Solaris creates a dynamic taskq of up to 64 threads, however in
-        * a Linux environment 1 thread per-core is usually about right */
-       system_taskq = taskq_create("spl_system_taskq", num_online_cpus(),
-                                   minclsyspri, 4, 512, TASKQ_PREPOPULATE);
+       system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64),
+           maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
        if (system_taskq == NULL)
                return (1);
 
+       dynamic_taskq = taskq_create("spl_dynamic_taskq", 1,
+           maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE);
+       if (dynamic_taskq == NULL) {
+               taskq_destroy(system_taskq);
+               return (1);
+       }
+
+       /* This is used to annotate tq_lock, so
+        *      taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch
+        * does not trigger a lockdep warning re: possible recursive locking
+        */
+       dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC;
+
        return (0);
 }
 
 void
 spl_taskq_fini(void)
 {
+       taskq_destroy(dynamic_taskq);
+       dynamic_taskq = NULL;
+
        taskq_destroy(system_taskq);
+       system_taskq = NULL;
 }