The 9p protocol mostly relies on a request/reply dialog between the
client and the server. A notable exception to this rule is request
cancellation (ie, flush in 9p wording): when the client requests an
in-flight request to be flushed, the server should only reply to the
flush request and not to the flushed in-flight request (otherwise it
considers the in-flight request to have completed just like it has
never been flushed).
It is up to the client to inform the transport that the in-flight request
has been flushed and won't receive a reply. This is achieved with the
'cancelled' operation of the struct p9_trans_module.
This operation isn't currently implemented with the virtio transport.
As a consequence, flushed requests leave buffers in the used list
forever and the virtqueue ends up in being able to process only one
request at a time.
This issue never popped up because the 9p server in QEMU had a bug
and would always reply to flushed requests. But this will be fixed
soon, so it is time to implement the 'cancelled' operation in the
9p virtio transport.
--
Greg
---
Greg Kurz (2):
virtio: allow to detach a buffer from the virtqueue
9p/trans_virtio: implement cancelled callback
drivers/virtio/virtio_ring.c | 28 ++++++++++++++++++++++++++++
include/linux/virtio.h | 1 +
net/9p/trans_virtio.c | 18 ++++++++++++++++++
3 files changed, 47 insertions(+)
