On 07.07.2021 21:44, Michael S. Tsirkin wrote:
On Wed, Jul 07, 2021 at 05:58:50PM +0300, Denis Plotnikov wrote:
On 07.07.2021 17:39, Michael S. Tsirkin wrote:
On Wed, Jul 07, 2021 at 03:19:20PM +0300, Denis Plotnikov wrote:
On 07.07.2021 13:10, Michael S. Tsirkin wrote:
On Fri, Jun 25, 2021 at 11:52:10AM +0300, Denis Plotnikov wrote:
On vhost-user-blk migration, qemu normally sends a number of commands
to enable logging if VHOST_USER_PROTOCOL_F_LOG_SHMFD is negotiated.
Qemu sends VHOST_USER_SET_FEATURES to enable buffers logging and
VHOST_USER_SET_FEATURES per each started ring to enable "used ring"
data logging.
The issue is that qemu doesn't wait for reply from the vhost daemon
for these commands which may result in races between qemu expectation
of logging starting and actual login starting in vhost daemon.
Could you be more explicit please? What kind of race have you
observed? Getting a reply slows down the setup considerably and
should not be done lightly.
I'm talking about the vhost-user-blk case. On migration setup, we enable
logging by sending VHOST_USER_SET_FEATURES. The command doesn't arrive to a
vhost-user-blk daemon immediately and the daemon needs some time turn the
logging on internally. If qemu doesn't wait for reply, after sending the
command qemu may start migrate memory pages. At this time the logging may
not be actually turned on in the daemon but some guest pages, which the
daemon is about to write to, may be already transferred without logging to a
destination. Since the logging wasn't turned on, those pages won't be
transferred again as dirty. So we may end up with corrupted data on the
destination.
Have I managed to explain the case clearly?
Thanks!
Denis
OK so this is just about enabling logging. It would be cleaner to
defer migrating memory until response ... if that is too hard,
at least document why we are doing this please.
And, let's wait for an ack just in that case then - why not?
And what about VHOST_USER_SET_PROTOCOL_FEATURES?
The code uses the same path for both VHOST_USER_SET_PROTOCOL_FEATURES and
VHOST_USER_SET_FEATURES via vhost_user_set_u64(). So, I decided to suggest
adding reply to both of them, so both feature setting commands work
similarly as it doesn't contradicts with vhost-user spec.
I'm not sure that it worth doing that, so if you think it's not I'll just
remove them.
Denis
I'm inclined to say let's not add to the latency of setting up the
device unnecessarily.
ok
I'll remove reply for VHOST_USER_SET_FEATURES and amend the commit
message in v2
Thanks!
Denis
Thanks!
To resolve this issue, this patch makes qemu wait for the commands result
explicilty if VHOST_USER_PROTOCOL_F_REPLY_ACK is negotiated.
Also, this patch adds the reply waiting for VHOST_USER_SET_PROTOCOL_FEATURES
command to make the features setting functions work similary.
Signed-off-by: Denis Plotnikov <den-plotni...@yandex-team.ru>
---
hw/virtio/vhost-user.c | 20 ++++++++++++++++++++
1 file changed, 20 insertions(+)
diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
index ee57abe04526..e47b82adab00 100644
--- a/hw/virtio/vhost-user.c
+++ b/hw/virtio/vhost-user.c
@@ -1105,10 +1105,20 @@ static int vhost_user_set_vring_addr(struct vhost_dev
*dev,
.hdr.size = sizeof(msg.payload.addr),
};
+ bool reply_supported = virtio_has_feature(dev->protocol_features,
+ VHOST_USER_PROTOCOL_F_REPLY_ACK);
+ if (reply_supported) {
+ msg.hdr.flags |= VHOST_USER_NEED_REPLY_MASK;
+ }
+
if (vhost_user_write(dev, &msg, NULL, 0) < 0) {
return -1;
}
+ if (reply_supported) {
+ return process_message_reply(dev, &msg);
+ }
+
return 0;
}
@@ -1297,10 +1307,20 @@ static int vhost_user_set_u64(struct vhost_dev *dev,
int request, uint64_t u64)
.hdr.size = sizeof(msg.payload.u64),
};
+ bool reply_supported = virtio_has_feature(dev->protocol_features,
+ VHOST_USER_PROTOCOL_F_REPLY_ACK);
+ if (reply_supported) {
+ msg.hdr.flags |= VHOST_USER_NEED_REPLY_MASK;
+ }
+
if (vhost_user_write(dev, &msg, NULL, 0) < 0) {
return -1;
}
+ if (reply_supported) {
+ return process_message_reply(dev, &msg);
+ }
+
return 0;
}
--
2.25.1
