On 04/28/2017 09:35 AM, Yuanhan Liu wrote:
On Fri, Apr 28, 2017 at 09:23:54AM +0200, Maxime Coquelin wrote:
On 04/28/2017 04:25 AM, Yuanhan Liu wrote:
On Thu, Apr 27, 2017 at 10:52:20AM +0200, Maxime Coquelin wrote:
On 04/27/2017 10:20 AM, Yuanhan Liu wrote:
On Thu, Apr 27, 2017 at 09:56:47AM +0200, Maxime Coquelin wrote:
Hi Zhiyong,
+Marc-André
On 04/27/2017 08:34 AM, Zhiyong Yang wrote:
vhost since dpdk17.02 + qemu2.7 and above will cause failures of
new connection when negotiating to set MQ. (one queue pair works
well).Because there exist some bugs in qemu code when introducing
VHOST_USER_PROTOCOL_F_REPLY_ACK to qemu. when dealing with the vhost
message VHOST_USER_SET_MEM_TABLE for the second time, qemu indeed
doesn't send the messge (The message needs to be sent only once)but
still will be waiting for dpdk's reply ack, then, qemu is always
freezing. DPDK code works in the right way.
I'm looking at Qemu's vhost_user_set_mem_table() function, but fail to
see how it could wait for the reply-ack if it didn't send the
VHOST_USER_SET_MEM_TABLE request before.
But the feature
VHOST_USER_PROTOCOL_F_REPLY_ACK has to be disabled by default at the
dpdk side in order to avoid the feature support of DPDK + qemu at
the same time. if doing like that, MQ can works well. Once Qemu bugs
have been fixed and upstreamed, we can enable it.
The problem is for DPDK to detect whether bug is fixed in Qemu.
Maybe only way would be to have a new protocol feature flag, which is
not really its role.
Wouldn't that be an overkill, judging that REPLY_ACK is not a must
feature?
Yes, maybe. But it was introduced to fix (possible) race conditions:
https://lists.gnu.org/archive/html/qemu-devel/2016-07/msg06173.html
But AFAIK, that commit has been reverted:
commit 94c9cb31c04737f86be29afefbff401cd23bc24d
Author: Michael S. Tsirkin <m...@redhat.com>
Date: Mon Aug 15 16:35:24 2016 +0300
Revert "vhost-user: Attempt to fix a race with set_mem_table."
This reverts commit 28ed5ef16384f12500abd3647973ee21b03cbe23.
I still think it's the right thing to do, but
tests have been failing sporadically.
Revert for now, and hope to fix it before the release.
No, what has been reverted is a workaround when REPLY_ACK protocol
feature has not been negotiated.
Good to know.
Instead of waiting for the backend to send the ack, the workaround
consisted in sending a GET_FEATURES request after having sent the
SET_MEM_TABLE request, in order to ensure SET_MEM_TABLE request handling
was done before.
The problem is that it sometimes created a deadlock when when running
QEMU's vhost-user-test in TCG mode.
Note that I planned to use this feature for the device IOTLB
implementation to let the backend decide whether it wants the IOTLB
misses synchronous or asynchronous. But I can still change the protocol
spec to make this behavior specific to this request.
Maybe we could introduce a version message? With that, we could tell
whether the frontend has fixed the known bug or not.
That's a possibility, but this is not really the role of a protocol
version. As in this case, the protocol does not change, just an
implementation.
Maybe. Well, you might could think this way: we do increase the version
when we make a new release (with bugs being fixed).
Or, we could also make the version two parts: major and minor. We increase
major for major updates (say, new features, etc). We increase minor for
bug fixes.
The only thing that doesn't make too much sense is the bug is actually
from the QEMU implementation but not from the vhost-user spec.
Yes, I was maybe not clear, but that's what I meant when saying that was
not the role of the protocol version.
Talking
about that, it may make more sense to introduce a new message to carry
the frontend version, something like a string "QEMU v2.8".
I don't think this is a good idea as it would create more problems that
it would solve. Indeed, you would need also the distro version, as for
example, Red Hat could backport the fix in its QEMU v2.6 package, Ubuntu
in its v2.7, etc...
--yliu
Note that we already has the "version" info in current vhost-user spec.
It's just 2 bits in the message "flag" field though, which is not quite
enough.
Indeed, it does not let room for lots of bugs :)
Thanks,
Maxime
--yliu
