According to the 9P spec [*], when a client wants to cancel a pending I/O
request identified by a given tag (uint16), it must send a Tflush message
and wait for the server to respond with a Rflush message before reusing this
tag for another I/O. The server may still send a completion message for the
I/O if it wasn't actually cancelled but the Rflush message must arrive after
that.
QEMU hence waits for the flushed PDU to complete before sending the Rflush
message back to the client.
If a client sends 'Tflush tag oldtag' and tag == oldtag, QEMU will then
allocate a PDU identified by tag, find it in the PDU list and wait for
this same PDU to complete... i.e. wait for a completion that will never
happen. This causes a tag and ring slot leak in the guest, and a PDU
leak in QEMU, all of them limited by the maximal number of PDUs (128).
But, worse, this causes QEMU to hang on device reset since v9fs_reset()
wants to drain all pending I/O.
This insane behavior is likely to denote a bug in the client, and it would
deserve an Rerror message to be sent back. Unfortunately, the protocol
allows it and requires all flush requests to suceed (only a Tflush response
is expected).
The only option is to detect when we have to handle a self-referencing
flush request and report success to the client right away.
[*] http://man.cat-v.org/plan_9/5/flush
Reported-by: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: Greg Kurz <groug@kaod.org>
d5f2af7b95