[RFC PATCH 1/1] fuse: disable local irqs when processing vq completions

[Stefan Hajnoczi]

The virtqueue completion handler function runs on a work queue and local
irqs are still enabled. There is a race where the completion handler
function grabs the next completed request just before vring_interrupt()
runs. vring_interrupt() sees an empty virtqueue and returns IRQ_NONE,
falsely declaring this interrupt unhandled.

The unhandled irq causes the kernel to disable the irq:

  irq 77: nobody cared (try booting with the "irqpoll" option)
  ...
  handlers:
  [<00000000d33eeed7>] vring_interrupt
  Disabling IRQ #77

The driver hangs afterwards since virtqueue irqs are now ignored.

Disable local irqs before calling virtqueue_get_buf() and re-enable them
afterwards so that vring_interrupt() doesn't run during the race window.

Reported-by: Xiaoling Gao <xiagao@xxxxxxxxxx>
Cc: Michael Tsirkin <mst@xxxxxxxxxx>
Cc: Jason Wang <jasowang@xxxxxxxxxx>
Signed-off-by: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
---
I'm not 100% convinced this fixes everything because vring_interrupt()
can still run after our critical section and find the virtqueue empty.
virtqueue_disable_cb() should minimize that but it's only a hint and
there is a small window when the race condition can happen before it's
called.
---
 fs/fuse/virtio_fs.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c
index 8f52cdaa8445..57e1f264b0a8 100644
--- a/fs/fuse/virtio_fs.c
+++ b/fs/fuse/virtio_fs.c
@@ -319,9 +319,10 @@ static void virtio_fs_hiprio_done_work(struct work_struct *work)
 	struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
 						 done_work);
 	struct virtqueue *vq = fsvq->vq;
+	unsigned long flags;
 
 	/* Free completed FUSE_FORGET requests */
-	spin_lock(&fsvq->lock);
+	spin_lock_irqsave(&fsvq->lock, flags);
 	do {
 		unsigned int len;
 		void *req;
@@ -333,7 +334,7 @@ static void virtio_fs_hiprio_done_work(struct work_struct *work)
 			dec_in_flight_req(fsvq);
 		}
 	} while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
-	spin_unlock(&fsvq->lock);
+	spin_unlock_irqrestore(&fsvq->lock, flags);
 }
 
 static void virtio_fs_request_dispatch_work(struct work_struct *work)
@@ -601,11 +602,15 @@ static void virtio_fs_requests_done_work(struct work_struct *work)
 	struct virtqueue *vq = fsvq->vq;
 	struct fuse_req *req;
 	struct fuse_req *next;
+	unsigned long flags;
 	unsigned int len;
 	LIST_HEAD(reqs);
 
-	/* Collect completed requests off the virtqueue */
-	spin_lock(&fsvq->lock);
+	/*
+	 * Collect completed requests off the virtqueue with irqs disabled to
+	 * prevent races with vring_interrupt().
+	 */
+	spin_lock_irqsave(&fsvq->lock, flags);
 	do {
 		virtqueue_disable_cb(vq);
 
@@ -615,7 +620,7 @@ static void virtio_fs_requests_done_work(struct work_struct *work)
 			spin_unlock(&fpq->lock);
 		}
 	} while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
-	spin_unlock(&fsvq->lock);
+	spin_unlock_irqrestore(&fsvq->lock, flags);
 
 	/* End requests */
 	list_for_each_entry_safe(req, next, &reqs, list) {
-- 
2.31.1