New an improved taskq implementation for the SPL. It allows a

configurable number of threads like the Solaris version and almost
all of the options are supported.  Unfortunately, it appears to have
made absolutely no difference to our performance numbers.  I need
to keep looking for where we are bottle necking.



git-svn-id: https://outreach.scidac.gov/svn/spl/trunk@93 7e1ea52c-4ff2-0310-8f11-9dd32ca42a1c

1 files changed, 388 insertions, 70 deletions

diff --git a/modules/spl/spl-taskq.c b/modules/spl/spl-taskq.c
index d26b40db8..ad9be695b 100644
--- a/modules/spl/spl-taskq.c
+++ b/modules/spl/spl-taskq.c
@@ -6,109 +6,427 @@
 
 #define DEBUG_SUBSYSTEM S_TASKQ
 
-/*
- * Task queue interface
- *
- * The taskq_work_wrapper functions are used to manage the work_structs
- * which must be submitted to linux.  The shim layer allocates a wrapper
- * structure for all items which contains a pointer to itself as well as
- * the real work to be performed.  When the work item run the generic
- * handle is called which calls the real work function and then using
- * the self pointer frees the work_struct.
+/* 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.
  */
-typedef struct taskq_work_wrapper {
-        struct work_struct tww_work;
-        task_func_t        tww_func;
-        void *             tww_priv;
-} taskq_work_wrapper_t;
+static task_t *
+task_alloc(taskq_t *tq, uint_t flags)
+{
+        task_t *t;
+        int count = 0;
+        ENTRY;
+
+        ASSERT(tq);
+        ASSERT(flags & (TQ_SLEEP | TQ_NOSLEEP));               /* One set */
+        ASSERT(!((flags & TQ_SLEEP) && (flags & TQ_NOSLEEP))); /* Not both */
+	ASSERT(spin_is_locked(&tq->tq_lock));
+retry:
+        /* Aquire task_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, task_t, t_list);
+	        list_del_init(&t->t_list);
+		RETURN(t);
+        }
+
+        /* Free list is empty and memory allocs are prohibited */
+        if (flags & TQ_NOALLOC)
+                RETURN(NULL);
+
+        /* Hit maximum task_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
+                 * task_t.  If a full second passes and we have not found
+                 * one gives up and return a NULL to the caller. */
+                if (flags & TQ_SLEEP) {
+                        spin_unlock_irq(&tq->tq_lock);
+                        schedule_timeout(HZ / 100);
+                        spin_lock_irq(&tq->tq_lock);
+                        if (count < 100)
+                                GOTO(retry, count++);
+
+                        RETURN(NULL);
+                }
+
+                /* Unreachable, TQ_SLEEP xor TQ_NOSLEEP */
+                SBUG();
+        }
+
+	spin_unlock_irq(&tq->tq_lock);
+        t = kmem_alloc(sizeof(task_t), flags & (TQ_SLEEP | TQ_NOSLEEP));
+        spin_lock_irq(&tq->tq_lock);
+
+	if (t) {
+		spin_lock_init(&t->t_lock);
+                INIT_LIST_HEAD(&t->t_list);
+		t->t_id = 0;
+		t->t_func = NULL;
+		t->t_arg = NULL;
+		tq->tq_nalloc++;
+	}
+
+        RETURN(t);
+}
+
+/* NOTE: Must be called with tq->tq_lock held, expectes the task_t
+ * to already be removed from the free, work, or pending taskq lists.
+ */
+static void
+task_free(taskq_t *tq, task_t *t)
+{
+        ENTRY;
+
+        ASSERT(tq);
+        ASSERT(t);
+	ASSERT(spin_is_locked(&tq->tq_lock));
+	ASSERT(list_empty(&t->t_list));
+
+        kmem_free(t, sizeof(task_t));
+        tq->tq_nalloc--;
 
+	EXIT;
+}
+
+/* NOTE: Must be called with tq->tq_lock held, either destroyes the
+ * task_t if too many exist or moves it to the free list for later use.
+ */
 static void
-taskq_work_handler(void *priv)
+task_done(taskq_t *tq, task_t *t)
 {
-        taskq_work_wrapper_t *tww = priv;
+	ENTRY;
+	ASSERT(tq);
+	ASSERT(t);
+	ASSERT(spin_is_locked(&tq->tq_lock));
+
+	list_del_init(&t->t_list);
 
-        ASSERT(tww);
-        ASSERT(tww->tww_func);
+        if (tq->tq_nalloc <= tq->tq_minalloc) {
+		t->t_id = 0;
+		t->t_func = NULL;
+		t->t_arg = NULL;
+                list_add(&t->t_list, &tq->tq_free_list);
+	} else {
+		task_free(tq, t);
+	}
 
-        /* Call the real function and free the wrapper */
-        tww->tww_func(tww->tww_priv);
-        kfree(tww);
+        EXIT;
 }
 
-/* XXX - All flags currently ignored */
-taskqid_t
-__taskq_dispatch(taskq_t *tq, task_func_t func, void *priv, uint_t flags)
+/* Taskqid's are handed out in a monotonically increasing fashion per
+ * taskq_t.  We don't handle taskqid wrapping yet, but fortuntely it isi
+ * a 64-bit value so this is probably never going to happen.  The lowest
+ * pending taskqid is stored in the taskq_t to make it easy for any
+ * taskq_wait()'ers to know if the tasks they're waiting for have
+ * completed.  Unfortunately, tq_task_lowest is kept up to date is
+ * a pretty brain dead way, something more clever should be done.
+ */
+static int
+taskq_wait_check(taskq_t *tq, taskqid_t id)
+{
+	RETURN(tq->tq_lowest_id >= id);
+}
+
+/* Expected to wait for all previously scheduled tasks to complete.  We do
+ * not need to wait for tasked scheduled after this call to complete.  In
+ * otherwords we do not need to drain the entire taskq. */
+void
+__taskq_wait_id(taskq_t *tq, taskqid_t id)
 {
-        struct workqueue_struct *wq = tq;
-        taskq_work_wrapper_t *tww;
-        int rc;
 	ENTRY;
+	ASSERT(tq);
+
+	wait_event(tq->tq_wait_waitq, taskq_wait_check(tq, id));
+
+	EXIT;
+}
+EXPORT_SYMBOL(__taskq_wait_id);
+
+void
+__taskq_wait(taskq_t *tq)
+{
+	taskqid_t id;
+	ENTRY;
+	ASSERT(tq);
+
+	spin_lock_irq(&tq->tq_lock);
+	id = tq->tq_next_id;
+	spin_unlock_irq(&tq->tq_lock);
+
+	__taskq_wait_id(tq, id);
+
+	EXIT;
+
+}
+EXPORT_SYMBOL(__taskq_wait);
+
+int
+__taskq_member(taskq_t *tq, void *t)
+{
+        int i;
+        ENTRY;
+
+	ASSERT(tq);
+        ASSERT(t);
+
+        for (i = 0; i < tq->tq_nthreads; i++)
+                if (tq->tq_threads[i] == (struct task_struct *)t)
+                        RETURN(1);
+
+        RETURN(0);
+}
+EXPORT_SYMBOL(__taskq_member);
+
+taskqid_t
+__taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
+{
+        task_t *t;
+	taskqid_t rc = 0;
+        ENTRY;
 
         ASSERT(tq);
         ASSERT(func);
+        if (unlikely(in_atomic() && (flags & TQ_SLEEP))) {
+		CERROR("May schedule while atomic: %s/0x%08x/%d\n",
+                       current->comm, preempt_count(), current->pid);
+		SBUG();
+	}
 
-	/* Use GFP_ATOMIC since this may be called in interrupt context */
-        tww = (taskq_work_wrapper_t *)kmalloc(sizeof(*tww), GFP_ATOMIC);
-        if (!tww)
-                RETURN((taskqid_t)0);
+        spin_lock_irq(&tq->tq_lock);
 
-        INIT_WORK(&(tww->tww_work), taskq_work_handler, tww);
-        tww->tww_func = func;
-        tww->tww_priv = priv;
+	/* Taskq being destroyed and all tasks drained */
+	if (!(tq->tq_flags & TQ_ACTIVE))
+		GOTO(out, rc = 0);
 
-        rc = queue_work(wq, &(tww->tww_work));
-        if (!rc) {
-                kfree(tww);
-                RETURN((taskqid_t)0);
-        }
+	/* 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))
+		GOTO(out, rc = 0);
+
+        if ((t = task_alloc(tq, flags)) == NULL)
+		GOTO(out, rc = 0);
 
-        RETURN((taskqid_t)wq);
+
+	spin_lock(&t->t_lock);
+	list_add(&t->t_list, &tq->tq_pend_list);
+	t->t_id = rc = tq->tq_next_id;
+	tq->tq_next_id++;
+        t->t_func = func;
+        t->t_arg = arg;
+	spin_unlock(&t->t_lock);
+
+	wake_up(&tq->tq_work_waitq);
+out:
+	spin_unlock_irq(&tq->tq_lock);
+	RETURN(rc);
 }
 EXPORT_SYMBOL(__taskq_dispatch);
 
-/* XXX - We must fully implement dynamic workqueues since they make a
- *       significant impact in terms of performance.  For now I've made
- *       a trivial compromise.  If you ask for one thread you get one
- *       thread, if you ask for more than that you get one per core.
- *       It's unclear if you ever really need/want more than one per-core
- *       anyway.  More analysis is required.
- *
- * name  - Workqueue names are limited to 10 chars
- * pri   - Ignore priority
- * min   - Ignored until this is a dynamic thread pool
- * max   - Ignored until this is a dynamic thread pool
- * flags - Ignored until this is a dynamic thread_pool
- */
+/* NOTE: Must be called with tq->tq_lock held */
+static taskqid_t
+taskq_lowest_id(taskq_t *tq)
+{
+	taskqid_t lowest_id = ~0;
+        task_t *t;
+	ENTRY;
+
+	ASSERT(tq);
+	ASSERT(spin_is_locked(&tq->tq_lock));
+
+	list_for_each_entry(t, &tq->tq_pend_list, t_list)
+		if (t->t_id < lowest_id)
+			lowest_id = t->t_id;
+
+	list_for_each_entry(t, &tq->tq_work_list, t_list)
+		if (t->t_id < lowest_id)
+			lowest_id = t->t_id;
+
+	RETURN(lowest_id);
+}
+
+static int
+taskq_thread(void *args)
+{
+        DECLARE_WAITQUEUE(wait, current);
+        sigset_t blocked;
+	taskqid_t id;
+        taskq_t *tq = args;
+        task_t *t;
+	ENTRY;
+
+        ASSERT(tq);
+        current->flags |= PF_NOFREEZE;
+
+        sigfillset(&blocked);
+        sigprocmask(SIG_BLOCK, &blocked, NULL);
+        flush_signals(current);
+
+        spin_lock_irq(&tq->tq_lock);
+        tq->tq_nthreads++;
+        wake_up(&tq->tq_wait_waitq);
+        set_current_state(TASK_INTERRUPTIBLE);
+
+        while (!kthread_should_stop()) {
+
+		add_wait_queue(&tq->tq_work_waitq, &wait);
+		if (list_empty(&tq->tq_pend_list)) {
+	                spin_unlock_irq(&tq->tq_lock);
+			schedule();
+	                spin_lock_irq(&tq->tq_lock);
+		} else {
+			__set_current_state(TASK_RUNNING);
+		}
+
+		remove_wait_queue(&tq->tq_work_waitq, &wait);
+                if (!list_empty(&tq->tq_pend_list)) {
+                        t = list_entry(tq->tq_pend_list.next, task_t, t_list);
+                        list_del_init(&t->t_list);
+			list_add(&t->t_list, &tq->tq_work_list);
+                        tq->tq_nactive++;
+			spin_unlock_irq(&tq->tq_lock);
+
+			/* Perform the requested task */
+                        t->t_func(t->t_arg);
+
+			spin_lock_irq(&tq->tq_lock);
+                        tq->tq_nactive--;
+			id = t->t_id;
+                        task_done(tq, t);
+
+			/* Update the lowest remaining taskqid yet to run */
+			if (tq->tq_lowest_id == id) {
+				tq->tq_lowest_id = taskq_lowest_id(tq);
+				ASSERT(tq->tq_lowest_id > id);
+			}
+
+                        wake_up_all(&tq->tq_wait_waitq);
+		}
+
+		set_current_state(TASK_INTERRUPTIBLE);
+
+        }
+
+	__set_current_state(TASK_RUNNING);
+        tq->tq_nthreads--;
+        spin_unlock_irq(&tq->tq_lock);
+
+	RETURN(0);
+}
+
 taskq_t *
 __taskq_create(const char *name, int nthreads, pri_t pri,
                int minalloc, int maxalloc, uint_t flags)
 {
-	taskq_t *tq;
-	ENTRY;
+        taskq_t *tq;
+        struct task_struct *t;
+        int rc = 0, i, j = 0;
+        ENTRY;
+
+        ASSERT(name != NULL);
+        ASSERT(pri <= maxclsyspri);
+        ASSERT(minalloc >= 0);
+        ASSERT(maxalloc <= INT_MAX);
+        ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */
+
+        tq = kmem_alloc(sizeof(*tq), KM_SLEEP);
+        if (tq == NULL)
+                RETURN(NULL);
+
+        tq->tq_threads = kmem_alloc(nthreads * sizeof(t), KM_SLEEP);
+        if (tq->tq_threads == NULL) {
+                kmem_free(tq, sizeof(*tq));
+                RETURN(NULL);
+        }
+
+        spin_lock_init(&tq->tq_lock);
+	spin_lock_irq(&tq->tq_lock);
+        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_work_list);
+        INIT_LIST_HEAD(&tq->tq_pend_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_SLEEP | TQ_NEW));
 
-	if (nthreads == 1)
-		tq = create_singlethread_workqueue(name);
-	else
-		tq = create_workqueue(name);
+	spin_unlock_irq(&tq->tq_lock);
 
-	return tq;
+        for (i = 0; i < nthreads; i++) {
+                t = kthread_create(taskq_thread, tq, "%s/%d", name, i);
+                if (t) {
+                        tq->tq_threads[i] = t;
+                        kthread_bind(t, i % num_online_cpus());
+                        set_user_nice(t, PRIO_TO_NICE(pri));
+                        wake_up_process(t);
+			j++;
+                } else {
+                        tq->tq_threads[i] = NULL;
+                        rc = 1;
+                }
+        }
+
+        /* Wait for all threads to be started before potential destroy */
+	wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j);
+
+        if (rc) {
+                __taskq_destroy(tq);
+                tq = NULL;
+        }
+
+        RETURN(tq);
 }
 EXPORT_SYMBOL(__taskq_create);
 
 void
 __taskq_destroy(taskq_t *tq)
 {
+	task_t *t;
+	int i, nthreads;
 	ENTRY;
-	destroy_workqueue(tq);
-	EXIT;
-}
-EXPORT_SYMBOL(__taskq_destroy);
 
-void
-__taskq_wait(taskq_t *tq)
-{
-	ENTRY;
-	flush_workqueue(tq);
+	ASSERT(tq);
+        spin_lock_irq(&tq->tq_lock);
+        tq->tq_flags &= ~TQ_ACTIVE;
+	spin_unlock_irq(&tq->tq_lock);
+
+	/* TQ_ACTIVE cleared prevents new tasks being added to pending */
+        __taskq_wait(tq);
+
+	nthreads = tq->tq_nthreads;
+	for (i = 0; i < nthreads; i++)
+		if (tq->tq_threads[i])
+			kthread_stop(tq->tq_threads[i]);
+
+        spin_lock_irq(&tq->tq_lock);
+
+        while (!list_empty(&tq->tq_free_list)) {
+		t = list_entry(tq->tq_free_list.next, task_t, t_list);
+	        list_del_init(&t->t_list);
+                task_free(tq, t);
+        }
+
+        ASSERT(tq->tq_nthreads == 0);
+        ASSERT(tq->tq_nalloc == 0);
+        ASSERT(list_empty(&tq->tq_free_list));
+        ASSERT(list_empty(&tq->tq_work_list));
+        ASSERT(list_empty(&tq->tq_pend_list));
+
+        spin_unlock_irq(&tq->tq_lock);
+        kmem_free(tq->tq_threads, nthreads * sizeof(task_t *));
+        kmem_free(tq, sizeof(taskq_t));
+
 	EXIT;
 }
-EXPORT_SYMBOL(__taskq_wait);
+EXPORT_SYMBOL(__taskq_destroy);