On 2020/5/28 下午10:00, Stefan Hajnoczi wrote:
On Thu, May 28, 2020 at 09:06:36PM +0800, Jason Wang wrote:
Hi:
I found ambiguity in the virtio specification:
In PCI part, it describes the queue_enable as:
The driver uses this to selectively prevent the device from executing
requests from this virtqueue. 1 - enabled; 0 - disabled.
In MMIO part, it describes the QueueReady as:
Writing one (0x1) to this register notifies the device that it can execute
requests from this virtual queue. Reading from this register returns the
last value written to it. Both read and write accesses apply to the queue
selected by writing to QueueSel.
If I understand this correctly, they have the same meaning, but the driver
requirements section looks conflict:
PCI said: The driver MUST NOT write a 0 to queue_enable.
MMIO said:
To stop using the queue the driver MUST write zero (0x0) to this QueueReady
and MUST read the value back to ensure synchronization.
So we can't disable a queue via queue_enable but QueueReady. Any reason for
such inconsistency?
I think MMIO is the outlier here. The device emulation code in QEMU
doesn't deal with queue shutdown. That only happens during device reset,
so it's like that a if the guest disables a virtio-mmio queue then
something undefined will happen on the QEMU side.
Qemu's MMIO only implement the fucntion of value write and read. Its PCI
implementation is also buggy which assumes the value is one.
For example, writing
used elements back to the virtqueue after it has been disabled.
Ok, but kernel virtio-mmio driver did the following in vm_del_vq():
/* Select and deactivate the queue */
writel(index, vm_dev->base + VIRTIO_MMIO_QUEUE_SEL);
if (vm_dev->version == 1) {
writel(0, vm_dev->base + VIRTIO_MMIO_QUEUE_PFN);
} else {
writel(0, vm_dev->base + VIRTIO_MMIO_QUEUE_READY);
WARN_ON(readl(vm_dev->base + VIRTIO_MMIO_QUEUE_READY));
}
Which tries to align with the spec ...
If the VIRTIO spec really intends to support virtqueue shutdown then the
semantics need to be spelled out clearly: what happens to in-flight
requests? How do devices behave that rely on multiple virtqueues during
normal operation (e.g. virtio-vsock where you can't really have rx-only
or tx-only)?
If we plan to do this, we need do something similar to status. That
means driver must wait for a read of queue_enable to return 0.
For in-flight request we can do the same thing as device reset but just
for a virtqueue probably.
Thanks
Stefan
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
