[PATCH] md: improve raid10 "IO Barrier" concept
authorNeilBrown <neilb@suse.de>
Fri, 6 Jan 2006 08:20:13 +0000 (00:20 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 6 Jan 2006 16:34:02 +0000 (08:34 -0800)
raid10 needs to put up a barrier to new requests while it does resync or other
background recovery.  The code for this is currently open-coded, slighty
obscure by its use of two waitqueues, and not documented.

This patch gathers all the related code into 4 functions, and includes a
comment which (hopefully) explains what is happening.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/md/raid10.c
include/linux/raid/raid10.h

index 713dc9c2c73093a67348e473f2bfc47f888d6e3c..50bd7b152f28e22c2bb9a5a36696f36f5dc390ab 100644 (file)
@@ -47,6 +47,9 @@
 
 static void unplug_slaves(mddev_t *mddev);
 
+static void allow_barrier(conf_t *conf);
+static void lower_barrier(conf_t *conf);
+
 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 {
        conf_t *conf = data;
@@ -175,20 +178,13 @@ static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 
 static inline void free_r10bio(r10bio_t *r10_bio)
 {
-       unsigned long flags;
-
        conf_t *conf = mddev_to_conf(r10_bio->mddev);
 
        /*
         * Wake up any possible resync thread that waits for the device
         * to go idle.
         */
-       spin_lock_irqsave(&conf->resync_lock, flags);
-       if (!--conf->nr_pending) {
-               wake_up(&conf->wait_idle);
-               wake_up(&conf->wait_resume);
-       }
-       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       allow_barrier(conf);
 
        put_all_bios(conf, r10_bio);
        mempool_free(r10_bio, conf->r10bio_pool);
@@ -197,22 +193,10 @@ static inline void free_r10bio(r10bio_t *r10_bio)
 static inline void put_buf(r10bio_t *r10_bio)
 {
        conf_t *conf = mddev_to_conf(r10_bio->mddev);
-       unsigned long flags;
 
        mempool_free(r10_bio, conf->r10buf_pool);
 
-       spin_lock_irqsave(&conf->resync_lock, flags);
-       if (!conf->barrier)
-               BUG();
-       --conf->barrier;
-       wake_up(&conf->wait_resume);
-       wake_up(&conf->wait_idle);
-
-       if (!--conf->nr_pending) {
-               wake_up(&conf->wait_idle);
-               wake_up(&conf->wait_resume);
-       }
-       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       lower_barrier(conf);
 }
 
 static void reschedule_retry(r10bio_t *r10_bio)
@@ -640,30 +624,82 @@ static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
        return ret;
 }
 
-/*
- * Throttle resync depth, so that we can both get proper overlapping of
- * requests, but are still able to handle normal requests quickly.
+/* Barriers....
+ * Sometimes we need to suspend IO while we do something else,
+ * either some resync/recovery, or reconfigure the array.
+ * To do this we raise a 'barrier'.
+ * The 'barrier' is a counter that can be raised multiple times
+ * to count how many activities are happening which preclude
+ * normal IO.
+ * We can only raise the barrier if there is no pending IO.
+ * i.e. if nr_pending == 0.
+ * We choose only to raise the barrier if no-one is waiting for the
+ * barrier to go down.  This means that as soon as an IO request
+ * is ready, no other operations which require a barrier will start
+ * until the IO request has had a chance.
+ *
+ * So: regular IO calls 'wait_barrier'.  When that returns there
+ *    is no backgroup IO happening,  It must arrange to call
+ *    allow_barrier when it has finished its IO.
+ * backgroup IO calls must call raise_barrier.  Once that returns
+ *    there is no normal IO happeing.  It must arrange to call
+ *    lower_barrier when the particular background IO completes.
  */
 #define RESYNC_DEPTH 32
 
-static void device_barrier(conf_t *conf, sector_t sect)
+static void raise_barrier(conf_t *conf)
 {
        spin_lock_irq(&conf->resync_lock);
-       wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
-                           conf->resync_lock, unplug_slaves(conf->mddev));
-
-       if (!conf->barrier++) {
-               wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
-                                   conf->resync_lock, unplug_slaves(conf->mddev));
-               if (conf->nr_pending)
-                       BUG();
+
+       /* Wait until no block IO is waiting */
+       wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
+                           conf->resync_lock,
+                           raid10_unplug(conf->mddev->queue));
+
+       /* block any new IO from starting */
+       conf->barrier++;
+
+       /* No wait for all pending IO to complete */
+       wait_event_lock_irq(conf->wait_barrier,
+                           !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
+                           conf->resync_lock,
+                           raid10_unplug(conf->mddev->queue));
+
+       spin_unlock_irq(&conf->resync_lock);
+}
+
+static void lower_barrier(conf_t *conf)
+{
+       unsigned long flags;
+       spin_lock_irqsave(&conf->resync_lock, flags);
+       conf->barrier--;
+       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       wake_up(&conf->wait_barrier);
+}
+
+static void wait_barrier(conf_t *conf)
+{
+       spin_lock_irq(&conf->resync_lock);
+       if (conf->barrier) {
+               conf->nr_waiting++;
+               wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
+                                   conf->resync_lock,
+                                   raid10_unplug(conf->mddev->queue));
+               conf->nr_waiting--;
        }
-       wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
-                           conf->resync_lock, unplug_slaves(conf->mddev));
-       conf->next_resync = sect;
+       conf->nr_pending++;
        spin_unlock_irq(&conf->resync_lock);
 }
 
+static void allow_barrier(conf_t *conf)
+{
+       unsigned long flags;
+       spin_lock_irqsave(&conf->resync_lock, flags);
+       conf->nr_pending--;
+       spin_unlock_irqrestore(&conf->resync_lock, flags);
+       wake_up(&conf->wait_barrier);
+}
+
 static int make_request(request_queue_t *q, struct bio * bio)
 {
        mddev_t *mddev = q->queuedata;
@@ -719,10 +755,7 @@ static int make_request(request_queue_t *q, struct bio * bio)
         * thread has put up a bar for new requests.
         * Continue immediately if no resync is active currently.
         */
-       spin_lock_irq(&conf->resync_lock);
-       wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
-       conf->nr_pending++;
-       spin_unlock_irq(&conf->resync_lock);
+       wait_barrier(conf);
 
        disk_stat_inc(mddev->gendisk, ios[rw]);
        disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
@@ -897,13 +930,8 @@ static void print_conf(conf_t *conf)
 
 static void close_sync(conf_t *conf)
 {
-       spin_lock_irq(&conf->resync_lock);
-       wait_event_lock_irq(conf->wait_resume, !conf->barrier,
-                           conf->resync_lock,  unplug_slaves(conf->mddev));
-       spin_unlock_irq(&conf->resync_lock);
-
-       if (conf->barrier) BUG();
-       if (waitqueue_active(&conf->wait_idle)) BUG();
+       wait_barrier(conf);
+       allow_barrier(conf);
 
        mempool_destroy(conf->r10buf_pool);
        conf->r10buf_pool = NULL;
@@ -1395,9 +1423,10 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
         * If there is non-resync activity waiting for us then
         * put in a delay to throttle resync.
         */
-       if (!go_faster && waitqueue_active(&conf->wait_resume))
+       if (!go_faster && conf->nr_waiting)
                msleep_interruptible(1000);
-       device_barrier(conf, sector_nr + RESYNC_SECTORS);
+       raise_barrier(conf);
+       conf->next_resync = sector_nr;
 
        /* Again, very different code for resync and recovery.
         * Both must result in an r10bio with a list of bios that
@@ -1427,7 +1456,6 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
 
                                r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
                                spin_lock_irq(&conf->resync_lock);
-                               conf->nr_pending++;
                                if (rb2) conf->barrier++;
                                spin_unlock_irq(&conf->resync_lock);
                                atomic_set(&r10_bio->remaining, 0);
@@ -1500,10 +1528,6 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
                int count = 0;
                r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
 
-               spin_lock_irq(&conf->resync_lock);
-               conf->nr_pending++;
-               spin_unlock_irq(&conf->resync_lock);
-
                r10_bio->mddev = mddev;
                atomic_set(&r10_bio->remaining, 0);
 
@@ -1713,8 +1737,7 @@ static int run(mddev_t *mddev)
        INIT_LIST_HEAD(&conf->retry_list);
 
        spin_lock_init(&conf->resync_lock);
-       init_waitqueue_head(&conf->wait_idle);
-       init_waitqueue_head(&conf->wait_resume);
+       init_waitqueue_head(&conf->wait_barrier);
 
        /* need to check that every block has at least one working mirror */
        if (!enough(conf)) {
index 60708789c8f978711fe2cf43ce008ef425c1faaa..08317b77802bf6a0b4b6c634dbaab1e5a5496c05 100644 (file)
@@ -39,11 +39,11 @@ struct r10_private_data_s {
 
        spinlock_t              resync_lock;
        int nr_pending;
+       int nr_waiting;
        int barrier;
        sector_t                next_resync;
 
-       wait_queue_head_t       wait_idle;
-       wait_queue_head_t       wait_resume;
+       wait_queue_head_t       wait_barrier;
 
        mempool_t *r10bio_pool;
        mempool_t *r10buf_pool;