Hi Michael, This could potentially affect other VIRTIO drivers too. Please see below. On Tue, Jan 23, 2024 at 11:27:40AM +0800, yi sun wrote: > On Mon, Jan 22, 2024 at 11:43 PM Stefan Hajnoczi <stefanha@xxxxxxxxxx> wrote: > > > > On Mon, Jan 22, 2024 at 07:07:22PM +0800, Yi Sun wrote: > > > Ensure no remaining requests in virtqueues before resetting vdev and > > > deleting virtqueues. Otherwise these requests will never be completed. > > > It may cause the system to become unresponsive. So it is better to place > > > blk_mq_quiesce_queue() in front of virtio_reset_device(). > > > > QEMU's virtio-blk device implementation completes all requests during > > device reset. Most device implementations have to do the same to avoid > > leaving dangling requests in flight across device reset. > > > > Which device implementation are you using and why is it safe for the > > device is simply drop requests across device reset? > > > > Stefan > > Virtio-blk device implementation completes all requests during device reset, but > this can only ensure that the device has finished using virtqueue. We should > also consider the driver's use of virtqueue. > > I caught such an example. Before del_vqs, the request had been processed by > the device, but it had not been processed by the driver. Although I am > using kernel5.4, > I think this problem may also occur with the latest version of kernel. > > The debug code I added is as follows: > virtblk_freeze() > { > vdev reset(); > quiesce queue(); > if (virtqueue->num_free != 1024) //1024 is the init value. > BUG_ON(1); > vdev del_vqs(); > } > > BUG_ON triggered the dump, the analysis is as follows: > > There is one request left in request_queue. > crash_arm64> struct request ffffff81f0560000 | grep -e state -e __data_len > __data_len = 20480, > state = MQ_RQ_IN_FLIGHT, > > crash_arm64> vring_virtqueue.packed,last_used_idx,broken,vq 0xffffff8086f92900 | > grep -e num -e used_wrap_counter -e last_used_idx -e broken -e > num_free -e desc_state -e "desc =" > num = 1024, > desc = 0xffffff8085ff8000, > used_wrap_counter = false, > desc_state = 0xffffff8085610000, > last_used_idx = 487, > broken = false, > num_free = 1017, > > Find desc based on last_used_idx. Through flags, we can know that the request > has been processed by the device, but it is still in flight state > because it has not > had time to run virtblk_done(). > crash_arm> vring_packed_desc ffffff8085ff9e70 > struct vring_packed_desc { > addr = 10474619192, > len = 20481, > id = 667, > flags = 2 > } > > I'm using a closed source virtual machine, so I can't see the source > code for it, > but I'm guessing it's similar to qemu. > > After the device completes the request, we must also ensure that the driver can > complete the request in virtblk_done(). > Okay, I think your approach of waiting for requests before virtio_device_reset() makes sense. blk_mq_complete_request() is async (might be deferred to an IPI or softirq) so it's not enough for virtblk_done() to run before virtio_device_reset() returns. There is no guarantee that virtblk_request_done() runs before virtio_device_reset() returns. A separate issue about virtio_device_reset(): Are you using virtio-mmio? virtio-mmio's vm_reset() doesn't offer the same guarantees as virtio-pci's reset functions. virtio-pci guarantees that irqs sent by the device during the reset operation complete and the irq handlers run before virtio_device_reset() returns. virtio-mmio does not. (On top of this, virtio_device_reset() now has CONFIG_VIRTIO_HARDEN_NOTIFICATION which changes the semantics. Drivers cannot expect to complete any in-flight requests in virtio_device_reset() when complied with this config option.) Other drivers may be affected by these inconsistent virtio_device_reset() semantics. I haven't audited the code. Stefan
Attachment:
signature.asc
Description: PGP signature