Btrfs: Worker thread optimizations
authorChris Mason <chris.mason@oracle.com>
Thu, 12 Jun 2008 00:21:24 +0000 (20:21 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 25 Sep 2008 15:04:03 +0000 (11:04 -0400)
This changes the worker thread pool to maintain a list of idle threads,
avoiding a complex search for a good thread to wake up.

Threads have two states:

idle - we try to reuse the last thread used in hopes of improving the batching
ratios

busy - each time a new work item is added to a busy task, the task is
rotated to the end of the line.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/async-thread.c
fs/btrfs/async-thread.h

index 8b9e2cf..8d4cc46 100644 (file)
@@ -34,6 +34,9 @@
  * One of these is allocated per thread.
  */
 struct btrfs_worker_thread {
+       /* pool we belong to */
+       struct btrfs_workers *workers;
+
        /* list of struct btrfs_work that are waiting for service */
        struct list_head pending;
 
@@ -51,9 +54,45 @@ struct btrfs_worker_thread {
 
        /* set to non-zero when this thread is already awake and kicking */
        int working;
+
+       /* are we currently idle */
+       int idle;
 };
 
 /*
+ * helper function to move a thread onto the idle list after it
+ * has finished some requests.
+ */
+static void check_idle_worker(struct btrfs_worker_thread *worker)
+{
+       if (!worker->idle && atomic_read(&worker->num_pending) <
+           worker->workers->idle_thresh / 2) {
+               unsigned long flags;
+               spin_lock_irqsave(&worker->workers->lock, flags);
+               worker->idle = 1;
+               list_move(&worker->worker_list, &worker->workers->idle_list);
+               spin_unlock_irqrestore(&worker->workers->lock, flags);
+       }
+}
+
+/*
+ * helper function to move a thread off the idle list after new
+ * pending work is added.
+ */
+static void check_busy_worker(struct btrfs_worker_thread *worker)
+{
+       if (worker->idle && atomic_read(&worker->num_pending) >=
+           worker->workers->idle_thresh) {
+               unsigned long flags;
+               spin_lock_irqsave(&worker->workers->lock, flags);
+               worker->idle = 0;
+               list_move_tail(&worker->worker_list,
+                              &worker->workers->worker_list);
+               spin_unlock_irqrestore(&worker->workers->lock, flags);
+       }
+}
+
+/*
  * main loop for servicing work items
  */
 static int worker_loop(void *arg)
@@ -76,6 +115,7 @@ static int worker_loop(void *arg)
 
                        atomic_dec(&worker->num_pending);
                        spin_lock_irq(&worker->lock);
+                       check_idle_worker(worker);
                }
                worker->working = 0;
                if (freezing(current)) {
@@ -98,6 +138,7 @@ int btrfs_stop_workers(struct btrfs_workers *workers)
        struct list_head *cur;
        struct btrfs_worker_thread *worker;
 
+       list_splice_init(&workers->idle_list, &workers->worker_list);
        while(!list_empty(&workers->worker_list)) {
                cur = workers->worker_list.next;
                worker = list_entry(cur, struct btrfs_worker_thread,
@@ -116,9 +157,10 @@ void btrfs_init_workers(struct btrfs_workers *workers, int max)
 {
        workers->num_workers = 0;
        INIT_LIST_HEAD(&workers->worker_list);
-       workers->last = NULL;
+       INIT_LIST_HEAD(&workers->idle_list);
        spin_lock_init(&workers->lock);
        workers->max_workers = max;
+       workers->idle_thresh = 64;
 }
 
 /*
@@ -143,14 +185,14 @@ int btrfs_start_workers(struct btrfs_workers *workers, int num_workers)
                spin_lock_init(&worker->lock);
                atomic_set(&worker->num_pending, 0);
                worker->task = kthread_run(worker_loop, worker, "btrfs");
+               worker->workers = workers;
                if (IS_ERR(worker->task)) {
                        ret = PTR_ERR(worker->task);
                        goto fail;
                }
 
                spin_lock_irq(&workers->lock);
-               list_add_tail(&worker->worker_list, &workers->worker_list);
-               workers->last = worker;
+               list_add_tail(&worker->worker_list, &workers->idle_list);
                workers->num_workers++;
                spin_unlock_irq(&workers->lock);
        }
@@ -169,42 +211,30 @@ static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
 {
        struct btrfs_worker_thread *worker;
        struct list_head *next;
-       struct list_head *start;
        int enforce_min = workers->num_workers < workers->max_workers;
 
-       /* start with the last thread if it isn't busy */
-       worker = workers->last;
-       if (atomic_read(&worker->num_pending) < 64)
-               goto done;
-
-       next = worker->worker_list.next;
-       start = &worker->worker_list;
-
        /*
-        * check all the workers for someone that is bored.  FIXME, do
-        * something smart here
+        * if we find an idle thread, don't move it to the end of the
+        * idle list.  This improves the chance that the next submission
+        * will reuse the same thread, and maybe catch it while it is still
+        * working
         */
-       while(next != start) {
-               if (next == &workers->worker_list) {
-                       next = workers->worker_list.next;
-                       continue;
-               }
+       if (!list_empty(&workers->idle_list)) {
+               next = workers->idle_list.next;
                worker = list_entry(next, struct btrfs_worker_thread,
                                    worker_list);
-               if (atomic_read(&worker->num_pending) < 64 || !enforce_min)
-                       goto done;
-               next = next->next;
+               return worker;
        }
+       if (enforce_min || list_empty(&workers->worker_list))
+               return NULL;
+
        /*
-        * nobody was bored, if we're already at the max thread count,
-        * use the last thread
+        * if we pick a busy task, move the task to the end of the list.
+        * hopefully this will keep things somewhat evenly balanced
         */
-       if (!enforce_min || atomic_read(&workers->last->num_pending) < 64) {
-               return workers->last;
-       }
-       return NULL;
-done:
-       workers->last = worker;
+       next = workers->worker_list.next;
+       worker = list_entry(next, struct btrfs_worker_thread, worker_list);
+       list_move_tail(next, &workers->worker_list);
        return worker;
 }
 
@@ -221,11 +251,17 @@ again:
        if (!worker) {
                spin_lock_irqsave(&workers->lock, flags);
                if (workers->num_workers >= workers->max_workers) {
+                       struct list_head *fallback = NULL;
                        /*
                         * we have failed to find any workers, just
                         * return the force one
                         */
-                       worker = list_entry(workers->worker_list.next,
+                       if (!list_empty(&workers->worker_list))
+                               fallback = workers->worker_list.next;
+                       if (!list_empty(&workers->idle_list))
+                               fallback = workers->idle_list.next;
+                       BUG_ON(!fallback);
+                       worker = list_entry(fallback,
                                  struct btrfs_worker_thread, worker_list);
                        spin_unlock_irqrestore(&workers->lock, flags);
                } else {
@@ -254,6 +290,7 @@ int btrfs_requeue_work(struct btrfs_work *work)
        spin_lock_irqsave(&worker->lock, flags);
        atomic_inc(&worker->num_pending);
        list_add_tail(&work->list, &worker->pending);
+       check_busy_worker(worker);
        spin_unlock_irqrestore(&worker->lock, flags);
 out:
        return 0;
@@ -276,6 +313,7 @@ int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
 
        spin_lock_irqsave(&worker->lock, flags);
        atomic_inc(&worker->num_pending);
+       check_busy_worker(worker);
        list_add_tail(&work->list, &worker->pending);
 
        /*
index 52fc9da..3436ff8 100644 (file)
@@ -60,11 +60,12 @@ struct btrfs_workers {
        /* max number of workers allowed.  changed by btrfs_start_workers */
        int max_workers;
 
+       /* once a worker has this many requests or fewer, it is idle */
+       int idle_thresh;
+
        /* list with all the work threads */
        struct list_head worker_list;
-
-       /* the last worker thread to have something queued */
-       struct btrfs_worker_thread *last;
+       struct list_head idle_list;
 
        /* lock for finding the next worker thread to queue on */
        spinlock_t lock;