drbd: allow parallel flushes for multi-volume resources
authorLars Ellenberg <lars.ellenberg@linbit.com>
Mon, 13 Jun 2016 22:26:19 +0000 (00:26 +0200)
committerJens Axboe <axboe@fb.com>
Tue, 14 Jun 2016 03:43:05 +0000 (21:43 -0600)
To maintain write-order fidelity accros all volumes in a DRBD resource,
the receiver of a P_BARRIER needs to issue flushes to all volumes.
We used to do this by calling blkdev_issue_flush(), synchronously,
one volume at a time.

We now submit all flushes to all volumes in parallel, then wait for all
completions, to reduce worst-case latencies on multi-volume resources.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/block/drbd/drbd_receiver.c

index a50cc99..15b2a0d 100644 (file)
@@ -1204,13 +1204,84 @@ static int drbd_recv_header(struct drbd_connection *connection, struct packet_in
        return err;
 }
 
-static void drbd_flush(struct drbd_connection *connection)
+/* This is blkdev_issue_flush, but asynchronous.
+ * We want to submit to all component volumes in parallel,
+ * then wait for all completions.
+ */
+struct issue_flush_context {
+       atomic_t pending;
+       int error;
+       struct completion done;
+};
+struct one_flush_context {
+       struct drbd_device *device;
+       struct issue_flush_context *ctx;
+};
+
+void one_flush_endio(struct bio *bio)
 {
-       int rv;
-       struct drbd_peer_device *peer_device;
-       int vnr;
+       struct one_flush_context *octx = bio->bi_private;
+       struct drbd_device *device = octx->device;
+       struct issue_flush_context *ctx = octx->ctx;
+
+       if (bio->bi_error) {
+               ctx->error = bio->bi_error;
+               drbd_info(device, "local disk FLUSH FAILED with status %d\n", bio->bi_error);
+       }
+       kfree(octx);
+       bio_put(bio);
+
+       clear_bit(FLUSH_PENDING, &device->flags);
+       put_ldev(device);
+       kref_put(&device->kref, drbd_destroy_device);
+
+       if (atomic_dec_and_test(&ctx->pending))
+               complete(&ctx->done);
+}
+
+static void submit_one_flush(struct drbd_device *device, struct issue_flush_context *ctx)
+{
+       struct bio *bio = bio_alloc(GFP_NOIO, 0);
+       struct one_flush_context *octx = kmalloc(sizeof(*octx), GFP_NOIO);
+       if (!bio || !octx) {
+               drbd_warn(device, "Could not allocate a bio, CANNOT ISSUE FLUSH\n");
+               /* FIXME: what else can I do now?  disconnecting or detaching
+                * really does not help to improve the state of the world, either.
+                */
+               kfree(octx);
+               if (bio)
+                       bio_put(bio);
 
+               ctx->error = -ENOMEM;
+               put_ldev(device);
+               kref_put(&device->kref, drbd_destroy_device);
+               return;
+       }
+
+       octx->device = device;
+       octx->ctx = ctx;
+       bio->bi_bdev = device->ldev->backing_bdev;
+       bio->bi_private = octx;
+       bio->bi_end_io = one_flush_endio;
+       bio_set_op_attrs(bio, REQ_OP_FLUSH, WRITE_FLUSH);
+
+       device->flush_jif = jiffies;
+       set_bit(FLUSH_PENDING, &device->flags);
+       atomic_inc(&ctx->pending);
+       submit_bio(bio);
+}
+
+static void drbd_flush(struct drbd_connection *connection)
+{
        if (connection->resource->write_ordering >= WO_BDEV_FLUSH) {
+               struct drbd_peer_device *peer_device;
+               struct issue_flush_context ctx;
+               int vnr;
+
+               atomic_set(&ctx.pending, 1);
+               ctx.error = 0;
+               init_completion(&ctx.done);
+
                rcu_read_lock();
                idr_for_each_entry(&connection->peer_devices, peer_device, vnr) {
                        struct drbd_device *device = peer_device->device;
@@ -1220,31 +1291,24 @@ static void drbd_flush(struct drbd_connection *connection)
                        kref_get(&device->kref);
                        rcu_read_unlock();
 
-                       /* Right now, we have only this one synchronous code path
-                        * for flushes between request epochs.
-                        * We may want to make those asynchronous,
-                        * or at least parallelize the flushes to the volume devices.
-                        */
-                       device->flush_jif = jiffies;
-                       set_bit(FLUSH_PENDING, &device->flags);
-                       rv = blkdev_issue_flush(device->ldev->backing_bdev,
-                                       GFP_NOIO, NULL);
-                       clear_bit(FLUSH_PENDING, &device->flags);
-                       if (rv) {
-                               drbd_info(device, "local disk flush failed with status %d\n", rv);
-                               /* would rather check on EOPNOTSUPP, but that is not reliable.
-                                * don't try again for ANY return value != 0
-                                * if (rv == -EOPNOTSUPP) */
-                               drbd_bump_write_ordering(connection->resource, NULL, WO_DRAIN_IO);
-                       }
-                       put_ldev(device);
-                       kref_put(&device->kref, drbd_destroy_device);
+                       submit_one_flush(device, &ctx);
 
                        rcu_read_lock();
-                       if (rv)
-                               break;
                }
                rcu_read_unlock();
+
+               /* Do we want to add a timeout,
+                * if disk-timeout is set? */
+               if (!atomic_dec_and_test(&ctx.pending))
+                       wait_for_completion(&ctx.done);
+
+               if (ctx.error) {
+                       /* would rather check on EOPNOTSUPP, but that is not reliable.
+                        * don't try again for ANY return value != 0
+                        * if (rv == -EOPNOTSUPP) */
+                       /* Any error is already reported by bio_endio callback. */
+                       drbd_bump_write_ordering(connection->resource, NULL, WO_DRAIN_IO);
+               }
        }
 }