]> git.proxmox.com Git - mirror_spl.git/blobdiff - module/spl/spl-taskq.c
Add defclsyspri macro
[mirror_spl.git] / module / spl / spl-taskq.c
index b58eb8e947cfc79274a13548443acde1f3148b40..b4282333e7e6c1e82f53c06133d4292ce7f74f25 100644 (file)
@@ -6,7 +6,7 @@
  *  UCRL-CODE-235197
  *
  *  This file is part of the SPL, Solaris Porting Layer.
- *  For details, see <http://github.com/behlendorf/spl/>.
+ *  For details, see <http://zfsonlinux.org/>.
  *
  *  The SPL is free software; you can redistribute it and/or modify it
  *  under the terms of the GNU General Public License as published by the
 
 #include <sys/taskq.h>
 #include <sys/kmem.h>
-#include <spl-debug.h>
 
-#ifdef SS_DEBUG_SUBSYS
-#undef SS_DEBUG_SUBSYS
-#endif
+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");
 
-#define SS_DEBUG_SUBSYS SS_TASKQ
+
+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)
+{
+       if (flags & TQ_NOSLEEP)
+               return KM_NOSLEEP;
+
+       if (flags & TQ_PUSHPAGE)
+               return KM_PUSHPAGE;
+
+       return KM_SLEEP;
+}
+
 /*
  * NOTE: Must be called with tq->tq_lock held, returns a list_t which
  * is not attached to the free, work, or pending taskq lists.
@@ -45,60 +73,63 @@ EXPORT_SYMBOL(system_taskq);
 static taskq_ent_t *
 task_alloc(taskq_t *tq, uint_t flags)
 {
-        taskq_ent_t *t;
-        int count = 0;
-        SENTRY;
+       taskq_ent_t *t;
+       int count = 0;
 
-        ASSERT(tq);
-        ASSERT(spin_is_locked(&tq->tq_lock));
+       ASSERT(tq);
+       ASSERT(spin_is_locked(&tq->tq_lock));
 retry:
-        /* Acquire taskq_ent_t's from free list if available */
-        if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
-                t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
-
-                ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
-
-                list_del_init(&t->tqent_list);
-                SRETURN(t);
-        }
-
-        /* Free list is empty and memory allocations are prohibited */
-        if (flags & TQ_NOALLOC)
-                SRETURN(NULL);
-
-        /* Hit maximum taskq_ent_t pool size */
-        if (tq->tq_nalloc >= tq->tq_maxalloc) {
-                if (flags & TQ_NOSLEEP)
-                        SRETURN(NULL);
-
-                /*
-                 * Sleep periodically polling the free list for an available
-                 * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
-                 * but we cannot block forever waiting for an taskq_entq_t to
-                 * show up in the free list, otherwise a deadlock can happen.
-                 *
-                 * Therefore, we need to allocate a new task even if the number
-                 * of allocated tasks is above tq->tq_maxalloc, but we still
-                 * 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);
-                 schedule_timeout(HZ / 100);
-                 spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-                 if (count < 100)
-                        SGOTO(retry, count++);
-        }
-
-        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
-        t = kmem_alloc(sizeof(taskq_ent_t), flags & (TQ_SLEEP | TQ_NOSLEEP));
-        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-
-        if (t) {
-                taskq_init_ent(t);
-                tq->tq_nalloc++;
-        }
-
-        SRETURN(t);
+       /* Acquire taskq_ent_t's from free list if available */
+       if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
+               t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
+
+               ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
+               ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL));
+               ASSERT(!timer_pending(&t->tqent_timer));
+
+               list_del_init(&t->tqent_list);
+               return (t);
+       }
+
+       /* Free list is empty and memory allocations are prohibited */
+       if (flags & TQ_NOALLOC)
+               return (NULL);
+
+       /* Hit maximum taskq_ent_t pool size */
+       if (tq->tq_nalloc >= tq->tq_maxalloc) {
+               if (flags & TQ_NOSLEEP)
+                       return (NULL);
+
+               /*
+                * Sleep periodically polling the free list for an available
+                * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
+                * but we cannot block forever waiting for an taskq_ent_t to
+                * show up in the free list, otherwise a deadlock can happen.
+                *
+                * Therefore, we need to allocate a new task even if the number
+                * of allocated tasks is above tq->tq_maxalloc, but we still
+                * 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);
+               schedule_timeout(HZ / 100);
+               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               if (count < 100) {
+                       count++;
+                       goto retry;
+               }
+       }
+
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       t = kmem_alloc(sizeof(taskq_ent_t), task_km_flags(flags));
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+
+       if (t) {
+               taskq_init_ent(t);
+               tq->tq_nalloc++;
+       }
+
+       return (t);
 }
 
 /*
@@ -108,17 +139,14 @@ retry:
 static void
 task_free(taskq_t *tq, taskq_ent_t *t)
 {
-        SENTRY;
-
-        ASSERT(tq);
-        ASSERT(t);
+       ASSERT(tq);
+       ASSERT(t);
        ASSERT(spin_is_locked(&tq->tq_lock));
        ASSERT(list_empty(&t->tqent_list));
+       ASSERT(!timer_pending(&t->tqent_timer));
 
-        kmem_free(t, sizeof(taskq_ent_t));
-        tq->tq_nalloc--;
-
-       SEXIT;
+       kmem_free(t, sizeof(taskq_ent_t));
+       tq->tq_nalloc--;
 }
 
 /*
@@ -128,148 +156,400 @@ task_free(taskq_t *tq, taskq_ent_t *t)
 static void
 task_done(taskq_t *tq, taskq_ent_t *t)
 {
-       SENTRY;
        ASSERT(tq);
        ASSERT(t);
        ASSERT(spin_is_locked(&tq->tq_lock));
 
+       /* Wake tasks blocked in taskq_wait_id() */
+       wake_up_all(&t->tqent_waitq);
+
        list_del_init(&t->tqent_list);
 
-        if (tq->tq_nalloc <= tq->tq_minalloc) {
+       if (tq->tq_nalloc <= tq->tq_minalloc) {
                t->tqent_id = 0;
                t->tqent_func = NULL;
                t->tqent_arg = NULL;
                t->tqent_flags = 0;
 
-                list_add_tail(&t->tqent_list, &tq->tq_free_list);
+               list_add_tail(&t->tqent_list, &tq->tq_free_list);
        } else {
                task_free(tq, t);
        }
+}
+
+/*
+ * When a delayed task timer expires remove it from the delay list and
+ * add it to the priority list in order for immediate processing.
+ */
+static void
+task_expire(unsigned long data)
+{
+       taskq_ent_t *w, *t = (taskq_ent_t *)data;
+       taskq_t *tq = t->tqent_taskq;
+       struct list_head *l;
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+
+       if (t->tqent_flags & TQENT_FLAG_CANCEL) {
+               ASSERT(list_empty(&t->tqent_list));
+               spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+               return;
+       }
+
+       /*
+        * The priority list must be maintained in strict task id order
+        * from lowest to highest for lowest_id to be easily calculable.
+        */
+       list_del(&t->tqent_list);
+       list_for_each_prev(l, &tq->tq_prio_list) {
+               w = list_entry(l, taskq_ent_t, tqent_list);
+               if (w->tqent_id < t->tqent_id) {
+                       list_add(&t->tqent_list, l);
+                       break;
+               }
+       }
+       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);
+
+       wake_up(&tq->tq_work_waitq);
+}
+
+/*
+ * Returns the lowest incomplete taskqid_t.  The taskqid_t may
+ * be queued on the pending list, on the priority list, on the
+ * delay list, or on the work list currently being handled, but
+ * it is not 100% complete yet.
+ */
+static taskqid_t
+taskq_lowest_id(taskq_t *tq)
+{
+       taskqid_t lowest_id = tq->tq_next_id;
+       taskq_ent_t *t;
+       taskq_thread_t *tqt;
+
+       ASSERT(tq);
+       ASSERT(spin_is_locked(&tq->tq_lock));
+
+       if (!list_empty(&tq->tq_pend_list)) {
+               t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
+               lowest_id = MIN(lowest_id, t->tqent_id);
+       }
+
+       if (!list_empty(&tq->tq_prio_list)) {
+               t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
+               lowest_id = MIN(lowest_id, t->tqent_id);
+       }
+
+       if (!list_empty(&tq->tq_delay_list)) {
+               t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list);
+               lowest_id = MIN(lowest_id, t->tqent_id);
+       }
+
+       if (!list_empty(&tq->tq_active_list)) {
+               tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
+                   tqt_active_list);
+               ASSERT(tqt->tqt_id != 0);
+               lowest_id = MIN(lowest_id, tqt->tqt_id);
+       }
+
+       return (lowest_id);
+}
+
+/*
+ * Insert a task into a list keeping the list sorted by increasing taskqid.
+ */
+static void
+taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
+{
+       taskq_thread_t *w;
+       struct list_head *l;
+
+       ASSERT(tq);
+       ASSERT(tqt);
+       ASSERT(spin_is_locked(&tq->tq_lock));
+
+       list_for_each_prev(l, &tq->tq_active_list) {
+               w = list_entry(l, taskq_thread_t, tqt_active_list);
+               if (w->tqt_id < tqt->tqt_id) {
+                       list_add(&tqt->tqt_active_list, l);
+                       break;
+               }
+       }
+       if (l == &tq->tq_active_list)
+               list_add(&tqt->tqt_active_list, &tq->tq_active_list);
+}
+
+/*
+ * Find and return a task from the given list if it exists.  The list
+ * must be in lowest to highest task id order.
+ */
+static taskq_ent_t *
+taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id)
+{
+       struct list_head *l;
+       taskq_ent_t *t;
+
+       ASSERT(spin_is_locked(&tq->tq_lock));
+
+       list_for_each(l, lh) {
+               t = list_entry(l, taskq_ent_t, tqent_list);
+
+               if (t->tqent_id == id)
+                       return (t);
+
+               if (t->tqent_id > id)
+                       break;
+       }
+
+       return (NULL);
+}
+
+/*
+ * Find an already dispatched task given the task id regardless of what
+ * state it is in.  If a task is still pending or executing it will be
+ * returned and 'active' set appropriately.  If the task has already
+ * been run then NULL is returned.
+ */
+static taskq_ent_t *
+taskq_find(taskq_t *tq, taskqid_t id, int *active)
+{
+       taskq_thread_t *tqt;
+       struct list_head *l;
+       taskq_ent_t *t;
+
+       ASSERT(spin_is_locked(&tq->tq_lock));
+       *active = 0;
+
+       t = taskq_find_list(tq, &tq->tq_delay_list, id);
+       if (t)
+               return (t);
+
+       t = taskq_find_list(tq, &tq->tq_prio_list, id);
+       if (t)
+               return (t);
+
+       t = taskq_find_list(tq, &tq->tq_pend_list, id);
+       if (t)
+               return (t);
+
+       list_for_each(l, &tq->tq_active_list) {
+               tqt = list_entry(l, taskq_thread_t, tqt_active_list);
+               if (tqt->tqt_id == id) {
+                       t = tqt->tqt_task;
+                       *active = 1;
+                       return (t);
+               }
+       }
 
-        SEXIT;
+       return (NULL);
 }
 
 /*
- * As tasks are submitted to the task queue they are assigned a
- * monotonically increasing taskqid and added to the tail of the pending
- * list.  As worker threads become available the tasks are removed from
- * the head of the pending or priority list, giving preference to the
- * priority list.  The tasks are then removed from their respective
- * list, and the taskq_thread servicing the task is added to the active
- * list, preserving the order using the serviced task's taskqid.
- * Finally, as tasks complete the taskq_thread servicing the task is
- * removed from the active list.  This means that the pending task and
- * active taskq_thread lists are always kept sorted by taskqid. Thus the
- * lowest outstanding incomplete taskqid can be determined simply by
- * checking the min taskqid for each head item on the pending, priority,
- * and active taskq_thread list. This value is stored in
- * tq->tq_lowest_id and only updated to the new lowest id when the
- * previous lowest id completes.  All taskqids lower than
- * tq->tq_lowest_id must have completed.  It is also possible larger
- * taskqid's have completed because they may be processed in parallel by
- * several worker threads.  However, this is not a problem because the
- * behavior of taskq_wait_id() is to block until all previously
- * submitted taskqid's have completed.
+ * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and
+ * taskq_wait() functions below.
+ *
+ * Taskq waiting is accomplished by tracking the lowest outstanding task
+ * id and the next available task id.  As tasks are dispatched they are
+ * added to the tail of the pending, priority, or delay lists.  As worker
+ * threads become available the tasks are removed from the heads of these
+ * lists and linked to the worker threads.  This ensures the lists are
+ * kept sorted by lowest to highest task id.
+ *
+ * Therefore the lowest outstanding task id can be quickly determined by
+ * checking the head item from all of these lists.  This value is stored
+ * with the taskq as the lowest id.  It only needs to be recalculated when
+ * either the task with the current lowest id completes or is canceled.
  *
- * XXX: Taskqid_t wrapping is not handled.  However, taskqid_t's are
- * 64-bit values so even if a taskq is processing 2^24 (16,777,216)
- * taskqid_ts per second it will still take 2^40 seconds, 34,865 years,
- * before the wrap occurs.  I can live with that for now.
+ * By blocking until the lowest task id exceeds the passed task id the
+ * taskq_wait_outstanding() function can be easily implemented.  Similarly,
+ * by blocking until the lowest task id matches the next task id taskq_wait()
+ * can be implemented.
+ *
+ * Callers should be aware that when there are multiple worked threads it
+ * is possible for larger task ids to complete before smaller ones.  Also
+ * when the taskq contains delay tasks with small task ids callers may
+ * block for a considerable length of time waiting for them to expire and
+ * execute.
  */
 static int
-taskq_wait_check(taskq_t *tq, taskqid_t id)
+taskq_wait_id_check(taskq_t *tq, taskqid_t id)
 {
+       int active = 0;
        int rc;
 
        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-       rc = (id < tq->tq_lowest_id);
+       rc = (taskq_find(tq, id, &active) == NULL);
        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-       SRETURN(rc);
+       return (rc);
 }
 
+/*
+ * The taskq_wait_id() function blocks until the passed task id completes.
+ * This does not guarantee that all lower task ids have completed.
+ */
 void
-__taskq_wait_id(taskq_t *tq, taskqid_t id)
+taskq_wait_id(taskq_t *tq, taskqid_t id)
 {
-       SENTRY;
-       ASSERT(tq);
+       wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id));
+}
+EXPORT_SYMBOL(taskq_wait_id);
 
-       wait_event(tq->tq_wait_waitq, taskq_wait_check(tq, id));
+static int
+taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id)
+{
+       int rc;
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       rc = (id < tq->tq_lowest_id);
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-       SEXIT;
+       return (rc);
 }
-EXPORT_SYMBOL(__taskq_wait_id);
 
+/*
+ * The taskq_wait_outstanding() function will block until all tasks with a
+ * lower taskqid than the passed 'id' have been completed.  Note that all
+ * task id's are assigned monotonically at dispatch time.  Zero may be
+ * passed for the id to indicate all tasks dispatch up to this point,
+ * but not after, should be waited for.
+ */
 void
-__taskq_wait(taskq_t *tq)
+taskq_wait_outstanding(taskq_t *tq, taskqid_t id)
 {
-       taskqid_t id;
-       SENTRY;
-       ASSERT(tq);
+       wait_event(tq->tq_wait_waitq,
+           taskq_wait_outstanding_check(tq, id ? id : tq->tq_next_id - 1));
+}
+EXPORT_SYMBOL(taskq_wait_outstanding);
+
+static int
+taskq_wait_check(taskq_t *tq)
+{
+       int rc;
 
-       /* Wait for the largest outstanding taskqid */
        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-       id = tq->tq_next_id - 1;
+       rc = (tq->tq_lowest_id == tq->tq_next_id);
        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-       __taskq_wait_id(tq, id);
-
-       SEXIT;
+       return (rc);
+}
 
+/*
+ * The taskq_wait() function will block until the taskq is empty.
+ * This means that if a taskq re-dispatches work to itself taskq_wait()
+ * callers will block indefinitely.
+ */
+void
+taskq_wait(taskq_t *tq)
+{
+       wait_event(tq->tq_wait_waitq, taskq_wait_check(tq));
 }
-EXPORT_SYMBOL(__taskq_wait);
+EXPORT_SYMBOL(taskq_wait);
 
 int
-__taskq_member(taskq_t *tq, void *t)
+taskq_member(taskq_t *tq, void *t)
 {
        struct list_head *l;
        taskq_thread_t *tqt;
-        SENTRY;
+       int found = 0;
 
        ASSERT(tq);
-        ASSERT(t);
+       ASSERT(t);
 
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
        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)
-                       SRETURN(1);
+               if (tqt->tqt_thread == (struct task_struct *)t) {
+                       found = 1;
+                       break;
+               }
        }
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-        SRETURN(0);
+       return (found);
 }
-EXPORT_SYMBOL(__taskq_member);
+EXPORT_SYMBOL(taskq_member);
 
-taskqid_t
-__taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
+/*
+ * Cancel an already dispatched task given the task id.  Still pending tasks
+ * will be immediately canceled, and if the task is active the function will
+ * block until it completes.  Preallocated tasks which are canceled must be
+ * freed by the caller.
+ */
+int
+taskq_cancel_id(taskq_t *tq, taskqid_t id)
 {
-        taskq_ent_t *t;
-       taskqid_t rc = 0;
-        SENTRY;
+       taskq_ent_t *t;
+       int active = 0;
+       int rc = ENOENT;
+
+       ASSERT(tq);
 
-        ASSERT(tq);
-        ASSERT(func);
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       t = taskq_find(tq, id, &active);
+       if (t && !active) {
+               list_del_init(&t->tqent_list);
+               t->tqent_flags |= TQENT_FLAG_CANCEL;
+
+               /*
+                * When canceling the lowest outstanding task id we
+                * must recalculate the new lowest outstanding id.
+                */
+               if (tq->tq_lowest_id == t->tqent_id) {
+                       tq->tq_lowest_id = taskq_lowest_id(tq);
+                       ASSERT3S(tq->tq_lowest_id, >, t->tqent_id);
+               }
 
-       /* Solaris assumes TQ_SLEEP if not passed explicitly */
-       if (!(flags & (TQ_SLEEP | TQ_NOSLEEP)))
-               flags |= TQ_SLEEP;
+               /*
+                * The task_expire() function takes the tq->tq_lock so drop
+                * drop the lock before synchronously cancelling the timer.
+                */
+               if (timer_pending(&t->tqent_timer)) {
+                       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+                       del_timer_sync(&t->tqent_timer);
+                       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               }
 
-       if (unlikely(in_atomic() && (flags & TQ_SLEEP)))
-               PANIC("May schedule while atomic: %s/0x%08x/%d\n",
-                   current->comm, preempt_count(), current->pid);
+               if (!(t->tqent_flags & TQENT_FLAG_PREALLOC))
+                       task_done(tq, t);
 
-        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               rc = 0;
+       }
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+
+       if (active) {
+               taskq_wait_id(tq, id);
+               rc = EBUSY;
+       }
+
+       return (rc);
+}
+EXPORT_SYMBOL(taskq_cancel_id);
+
+taskqid_t
+taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
+{
+       taskq_ent_t *t;
+       taskqid_t rc = 0;
+
+       ASSERT(tq);
+       ASSERT(func);
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 
        /* Taskq being destroyed and all tasks drained */
-       if (!(tq->tq_flags & TQ_ACTIVE))
-               SGOTO(out, rc = 0);
+       if (!(tq->tq_flags & TASKQ_ACTIVE))
+               goto out;
 
        /* Do not queue the task unless there is idle thread for it */
        ASSERT(tq->tq_nactive <= tq->tq_nthreads);
        if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads))
-               SGOTO(out, rc = 0);
+               goto out;
 
-        if ((t = task_alloc(tq, flags)) == NULL)
-               SGOTO(out, rc = 0);
+       if ((t = task_alloc(tq, flags)) == NULL)
+               goto out;
 
        spin_lock(&t->tqent_lock);
 
@@ -281,8 +561,12 @@ __taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
 
        t->tqent_id = rc = tq->tq_next_id;
        tq->tq_next_id++;
-        t->tqent_func = func;
-        t->tqent_arg = arg;
+       t->tqent_func = func;
+       t->tqent_arg = arg;
+       t->tqent_taskq = tq;
+       t->tqent_timer.data = 0;
+       t->tqent_timer.function = NULL;
+       t->tqent_timer.expires = 0;
 
        ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
 
@@ -291,24 +575,64 @@ __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);
-       SRETURN(rc);
+       return (rc);
+}
+EXPORT_SYMBOL(taskq_dispatch);
+
+taskqid_t
+taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
+    uint_t flags, clock_t expire_time)
+{
+       taskqid_t rc = 0;
+       taskq_ent_t *t;
+
+       ASSERT(tq);
+       ASSERT(func);
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+
+       /* Taskq being destroyed and all tasks drained */
+       if (!(tq->tq_flags & TASKQ_ACTIVE))
+               goto out;
+
+       if ((t = task_alloc(tq, flags)) == NULL)
+               goto out;
+
+       spin_lock(&t->tqent_lock);
+
+       /* Queue to the delay list for subsequent execution */
+       list_add_tail(&t->tqent_list, &tq->tq_delay_list);
+
+       t->tqent_id = rc = tq->tq_next_id;
+       tq->tq_next_id++;
+       t->tqent_func = func;
+       t->tqent_arg = arg;
+       t->tqent_taskq = tq;
+       t->tqent_timer.data = (unsigned long)t;
+       t->tqent_timer.function = task_expire;
+       t->tqent_timer.expires = (unsigned long)expire_time;
+       add_timer(&t->tqent_timer);
+
+       ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
+
+       spin_unlock(&t->tqent_lock);
+out:
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       return (rc);
 }
-EXPORT_SYMBOL(__taskq_dispatch);
+EXPORT_SYMBOL(taskq_dispatch_delay);
 
 void
-__taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
+taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
    taskq_ent_t *t)
 {
-       SENTRY;
-
        ASSERT(tq);
        ASSERT(func);
-       ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
 
        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 
        /* 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;
        }
@@ -331,153 +655,188 @@ __taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
        tq->tq_next_id++;
        t->tqent_func = func;
        t->tqent_arg = arg;
+       t->tqent_taskq = tq;
 
        spin_unlock(&t->tqent_lock);
 
        wake_up(&tq->tq_work_waitq);
 out:
        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
-       SEXIT;
 }
-EXPORT_SYMBOL(__taskq_dispatch_ent);
+EXPORT_SYMBOL(taskq_dispatch_ent);
 
 int
-__taskq_empty_ent(taskq_ent_t *t)
+taskq_empty_ent(taskq_ent_t *t)
 {
        return list_empty(&t->tqent_list);
 }
-EXPORT_SYMBOL(__taskq_empty_ent);
+EXPORT_SYMBOL(taskq_empty_ent);
 
 void
-__taskq_init_ent(taskq_ent_t *t)
+taskq_init_ent(taskq_ent_t *t)
 {
        spin_lock_init(&t->tqent_lock);
+       init_waitqueue_head(&t->tqent_waitq);
+       init_timer(&t->tqent_timer);
        INIT_LIST_HEAD(&t->tqent_list);
        t->tqent_id = 0;
        t->tqent_func = NULL;
        t->tqent_arg = NULL;
        t->tqent_flags = 0;
+       t->tqent_taskq = NULL;
 }
-EXPORT_SYMBOL(__taskq_init_ent);
+EXPORT_SYMBOL(taskq_init_ent);
 
 /*
- * Returns the lowest incomplete taskqid_t.  The taskqid_t may
- * be queued on the pending list, on the priority list,  or on
- * the work list currently being handled, but it is not 100%
- * complete yet.
+ * Return the next pending task, preference is given to tasks on the
+ * priority list which were dispatched with TQ_FRONT.
  */
-static taskqid_t
-taskq_lowest_id(taskq_t *tq)
+static taskq_ent_t *
+taskq_next_ent(taskq_t *tq)
 {
-       taskqid_t lowest_id = tq->tq_next_id;
-        taskq_ent_t *t;
-       taskq_thread_t *tqt;
-       SENTRY;
+       struct list_head *list;
 
-       ASSERT(tq);
        ASSERT(spin_is_locked(&tq->tq_lock));
 
-       if (!list_empty(&tq->tq_pend_list)) {
-               t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
-               lowest_id = MIN(lowest_id, t->tqent_id);
-       }
+       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);
 
-       if (!list_empty(&tq->tq_prio_list)) {
-               t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
-               lowest_id = MIN(lowest_id, t->tqent_id);
-       }
+       return (list_entry(list->next, taskq_ent_t, tqent_list));
+}
 
-       if (!list_empty(&tq->tq_active_list)) {
-               tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
-                                tqt_active_list);
-               ASSERT(tqt->tqt_id != 0);
-               lowest_id = MIN(lowest_id, tqt->tqt_id);
-       }
+/*
+ * Spawns a new thread for the specified taskq.
+ */
+static void
+taskq_thread_spawn_task(void *arg)
+{
+       taskq_t *tq = (taskq_t *)arg;
+
+       (void) taskq_thread_create(tq);
 
-       SRETURN(lowest_id);
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       tq->tq_nspawn--;
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 }
 
 /*
- * Insert a task into a list keeping the list sorted by increasing
- * taskqid.
+ * Spawn addition threads for dynamic taskqs (TASKQ_DYNMAIC) 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 void
-taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
+static int
+taskq_thread_spawn(taskq_t *tq, int seq_tasks)
 {
-       taskq_thread_t *w;
-       struct list_head *l;
+       int spawning = 0;
 
-       SENTRY;
-       ASSERT(tq);
-       ASSERT(tqt);
-       ASSERT(spin_is_locked(&tq->tq_lock));
+       if (!(tq->tq_flags & TASKQ_DYNAMIC))
+               return (0);
 
-       list_for_each_prev(l, &tq->tq_active_list) {
-               w = list_entry(l, taskq_thread_t, tqt_active_list);
-               if (w->tqt_id < tqt->tqt_id) {
-                       list_add(&tqt->tqt_active_list, l);
-                       break;
-               }
+       if ((seq_tasks > spl_taskq_thread_sequential) &&
+           (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);
        }
-       if (l == &tq->tq_active_list)
-               list_add(&tqt->tqt_active_list, &tq->tq_active_list);
 
-       SEXIT;
+       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)
 {
-        DECLARE_WAITQUEUE(wait, current);
-        sigset_t blocked;
+       DECLARE_WAITQUEUE(wait, current);
+       sigset_t blocked;
        taskq_thread_t *tqt = args;
-        taskq_t *tq;
-        taskq_ent_t *t;
-       struct list_head *pend_list;
-       SENTRY;
+       taskq_t *tq;
+       taskq_ent_t *t;
+       int seq_tasks = 0;
 
-        ASSERT(tqt);
+       ASSERT(tqt);
        tq = tqt->tqt_tq;
-        current->flags |= PF_NOFREEZE;
+       current->flags |= PF_NOFREEZE;
+
+       sigfillset(&blocked);
+       sigprocmask(SIG_BLOCK, &blocked, NULL);
+       flush_signals(current);
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 
-        sigfillset(&blocked);
-        sigprocmask(SIG_BLOCK, &blocked, NULL);
-        flush_signals(current);
+       /* Immediately exit if more threads than allowed were created. */
+       if (tq->tq_nthreads >= tq->tq_maxthreads)
+               goto error;
 
-        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-        tq->tq_nthreads++;
-        wake_up(&tq->tq_wait_waitq);
-        set_current_state(TASK_INTERRUPTIBLE);
+       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);
 
-        while (!kthread_should_stop()) {
+       while (!kthread_should_stop()) {
 
                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);
+
                        schedule();
+                       seq_tasks = 0;
+
                        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
                        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);
-                        list_del_init(&t->tqent_list);
+               if ((t = taskq_next_ent(tq)) != NULL) {
+                       list_del_init(&t->tqent_list);
 
                        /* In order to support recursively dispatching a
                         * preallocated taskq_ent_t, tqent_id must be
                         * stored prior to executing tqent_func. */
                        tqt->tqt_id = t->tqent_id;
+                       tqt->tqt_task = t;
 
                        /* We must store a copy of the flags prior to
                         * servicing the task (servicing a prealloc'd task
@@ -487,19 +846,19 @@ taskq_thread(void *args)
                        tqt->tqt_flags = t->tqent_flags;
 
                        taskq_insert_in_order(tq, tqt);
-                        tq->tq_nactive++;
+                       tq->tq_nactive++;
                        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
                        /* Perform the requested task */
-                        t->tqent_func(t->tqent_arg);
+                       t->tqent_func(t->tqent_arg);
 
                        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-                        tq->tq_nactive--;
+                       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
@@ -509,39 +868,77 @@ taskq_thread(void *args)
                                ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
                        }
 
+                       /* Spawn additional taskq threads if required. */
+                       if (taskq_thread_spawn(tq, ++seq_tasks))
+                               seq_tasks = 0;
+
                        tqt->tqt_id = 0;
                        tqt->tqt_flags = 0;
-                        wake_up_all(&tq->tq_wait_waitq);
+                       wake_up_all(&tq->tq_wait_waitq);
+               } else {
+                       if (taskq_thread_should_stop(tq, tqt))
+                               break;
                }
 
                set_current_state(TASK_INTERRUPTIBLE);
 
-        }
+       }
 
        __set_current_state(TASK_RUNNING);
-        tq->tq_nthreads--;
+       tq->tq_nthreads--;
        list_del_init(&tqt->tqt_thread_list);
-       kmem_free(tqt, sizeof(taskq_thread_t));
+error:
+       kmem_free(tqt, sizeof (taskq_thread_t));
+       spin_unlock_irqrestore(&tq->tq_lock, 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);
+       }
 
-        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       if (spl_taskq_thread_priority)
+               set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri));
 
-       SRETURN(0);
+       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)
+taskq_create(const char *name, int nthreads, pri_t pri,
+    int minalloc, int maxalloc, uint_t flags)
 {
-        taskq_t *tq;
+       taskq_t *tq;
        taskq_thread_t *tqt;
-        int rc = 0, i, j = 0;
-        SENTRY;
+       int count = 0, rc = 0, i;
 
-        ASSERT(name != NULL);
-        ASSERT(pri <= maxclsyspri);
-        ASSERT(minalloc >= 0);
-        ASSERT(maxalloc <= INT_MAX);
-        ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */
+       ASSERT(name != NULL);
+       ASSERT(pri <= maxclsyspri);
+       ASSERT(minalloc >= 0);
+       ASSERT(maxalloc <= INT_MAX);
+       ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */
 
        /* Scale the number of threads using nthreads as a percentage */
        if (flags & TASKQ_THREADS_CPU_PCT) {
@@ -552,86 +949,86 @@ __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);
-        if (tq == NULL)
-                SRETURN(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;
-        INIT_LIST_HEAD(&tq->tq_free_list);
-        INIT_LIST_HEAD(&tq->tq_pend_list);
-        INIT_LIST_HEAD(&tq->tq_prio_list);
-        init_waitqueue_head(&tq->tq_work_waitq);
-        init_waitqueue_head(&tq->tq_wait_waitq);
-
-        if (flags & TASKQ_PREPOPULATE)
-                for (i = 0; i < minalloc; i++)
-                        task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW));
-
-        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE);
+       if (tq == NULL)
+               return (NULL);
+
+       spin_lock_init(&tq->tq_lock);
+       INIT_LIST_HEAD(&tq->tq_thread_list);
+       INIT_LIST_HEAD(&tq->tq_active_list);
+       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);
+
+       if (flags & TASKQ_PREPOPULATE) {
+               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+
+               for (i = 0; i < minalloc; i++)
+                       task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW));
+
+               spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       }
+
+       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 = kthread_create(taskq_thread, tqt,
-                                                "%s/%d", name, i);
-               if (tqt->tqt_thread) {
-                       list_add(&tqt->tqt_thread_list, &tq->tq_thread_list);
-                       kthread_bind(tqt->tqt_thread, i % num_online_cpus());
-                       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 for all threads to be started before potential destroy */
+       wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count);
 
-        if (rc) {
-                __taskq_destroy(tq);
-                tq = NULL;
-        }
+       if (rc) {
+               taskq_destroy(tq);
+               tq = NULL;
+       }
 
-        SRETURN(tq);
+       return (tq);
 }
-EXPORT_SYMBOL(__taskq_create);
+EXPORT_SYMBOL(taskq_create);
 
 void
-__taskq_destroy(taskq_t *tq)
+taskq_destroy(taskq_t *tq)
 {
        struct task_struct *thread;
        taskq_thread_t *tqt;
        taskq_ent_t *t;
-       SENTRY;
 
        ASSERT(tq);
        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
-        tq->tq_flags &= ~TQ_ACTIVE;
+       tq->tq_flags &= ~TASKQ_ACTIVE;
        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-       /* TQ_ACTIVE cleared prevents new tasks being added to pending */
-        __taskq_wait(tq);
+       /*
+        * 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);
 
-        spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+       taskq_wait(tq);
+
+       spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 
        /*
         * Signal each thread to exit and block until it does.  Each thread
@@ -641,59 +1038,65 @@ __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);
 
                kthread_stop(thread);
 
-               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
+               spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
        }
 
-        while (!list_empty(&tq->tq_free_list)) {
+       while (!list_empty(&tq->tq_free_list)) {
                t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
 
                ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
 
-               list_del_init(&t->tqent_list);
-                task_free(tq, t);
-        }
-
-        ASSERT(tq->tq_nthreads == 0);
-        ASSERT(tq->tq_nalloc == 0);
-        ASSERT(list_empty(&tq->tq_thread_list));
-        ASSERT(list_empty(&tq->tq_active_list));
-        ASSERT(list_empty(&tq->tq_free_list));
-        ASSERT(list_empty(&tq->tq_pend_list));
-        ASSERT(list_empty(&tq->tq_prio_list));
+               list_del_init(&t->tqent_list);
+               task_free(tq, t);
+       }
 
-        spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
+       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));
+       ASSERT(list_empty(&tq->tq_pend_list));
+       ASSERT(list_empty(&tq->tq_prio_list));
+       ASSERT(list_empty(&tq->tq_delay_list));
 
-        kmem_free(tq, sizeof(taskq_t));
+       spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 
-       SEXIT;
+       strfree(tq->tq_name);
+       kmem_free(tq, sizeof (taskq_t));
 }
-EXPORT_SYMBOL(__taskq_destroy);
+EXPORT_SYMBOL(taskq_destroy);
 
 int
 spl_taskq_init(void)
 {
-        SENTRY;
-
-       /* 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),
+           defclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
        if (system_taskq == NULL)
-               SRETURN(1);
+               return (1);
 
-        SRETURN(0);
+       dynamic_taskq = taskq_create("spl_dynamic_taskq", 1,
+           defclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE);
+       if (dynamic_taskq == NULL) {
+               taskq_destroy(system_taskq);
+               return (1);
+       }
+
+       return (0);
 }
 
 void
 spl_taskq_fini(void)
 {
-        SENTRY;
+       taskq_destroy(dynamic_taskq);
+       dynamic_taskq = NULL;
+
        taskq_destroy(system_taskq);
-        SEXIT;
+       system_taskq = NULL;
 }