virtio-blk: Fix hot-unplug race in remove method
authorAsias He <asias@redhat.com>
Fri, 4 May 2012 12:22:04 +0000 (20:22 +0800)
committerRusty Russell <rusty@rustcorp.com.au>
Tue, 22 May 2012 02:46:13 +0000 (12:16 +0930)
If we reset the virtio-blk device before the requests already dispatched
to the virtio-blk driver from the block layer are finised, we will stuck
in blk_cleanup_queue() and the remove will fail.

blk_cleanup_queue() calls blk_drain_queue() to drain all requests queued
before DEAD marking. However it will never success if the device is
already stopped. We'll have q->in_flight[] > 0, so the drain will not
finish.

How to reproduce the race:
1. hot-plug a virtio-blk device
2. keep reading/writing the device in guest
3. hot-unplug while the device is busy serving I/O

Test:
~1000 rounds of hot-plug/hot-unplug test passed with this patch.

Changes in v3:
- Drop blk_abort_queue and blk_abort_request
- Use __blk_end_request_all to complete request dispatched to driver

Changes in v2:
- Drop req_in_flight
- Use virtqueue_detach_unused_buf to get request dispatched to driver

Signed-off-by: Asias He <asias@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
drivers/block/virtio_blk.c

index 0d39f2f4294a3c2587e5c019e39904b3f8440fc2..9d2223bba90c31e477c0ab31832272f7c7e2b23a 100644 (file)
@@ -583,6 +583,8 @@ static void __devexit virtblk_remove(struct virtio_device *vdev)
 {
        struct virtio_blk *vblk = vdev->priv;
        int index = vblk->index;
+       struct virtblk_req *vbr;
+       unsigned long flags;
 
        /* Prevent config work handler from accessing the device. */
        mutex_lock(&vblk->config_lock);
@@ -598,6 +600,15 @@ static void __devexit virtblk_remove(struct virtio_device *vdev)
        flush_work(&vblk->config_work);
 
        del_gendisk(vblk->disk);
+
+       /* Abort requests dispatched to driver. */
+       spin_lock_irqsave(&vblk->lock, flags);
+       while ((vbr = virtqueue_detach_unused_buf(vblk->vq))) {
+               __blk_end_request_all(vbr->req, -EIO);
+               mempool_free(vbr, vblk->pool);
+       }
+       spin_unlock_irqrestore(&vblk->lock, flags);
+
        blk_cleanup_queue(vblk->disk->queue);
        put_disk(vblk->disk);
        mempool_destroy(vblk->pool);