Hi Maxime,
First off, apologies for the lateness of this reply - I realize that this patch
has already been upstreamed.
Unfortunately, during OvS-DPDK regression testing for DPDK v17.11-rc2 just
today, a regression involving vHost multiq was detected, and pinpointed to this
patch.
Version info for the components involved during the aforementioned testing is
as follows:
DPDK: v17.11-rc2
OvS: af2e40c ("sparse: eliminate "duplicate initialization") + DPDK v17.11
upgrade patch
QEMU: v2.7.0
The regression may be reproduced as follows:
- Set up OvS-DPDK as normal, and add one vHostUser client port:
$OVS_DIR/utilities/ovs-vsctl add-port br0 vhost-user0 -- set Interface
vhost-user0 type=dpdkvhostuserclient
- Start QEMU, specifying 2 ports for the guest interface"
$QEMU_DIR/x86_64-softmmu/qemu-system-x86_64 \
-cpu host -enable-kvm -m 4096M \
-object
memory-backend-file,id=mem,size=4096M,mem-path=/dev/hugepages,share=on \
-numa node,memdev=mem -mem-prealloc \
-drive file=$VM_IMAGE \
-chardev socket,id=char0,path=/tmp/sock0,server \
-netdev type=vhost-user,id=mynet1,chardev=char0,vhostforce,queues=2 \
-device
virtio-net-pci,mac=00:00:00:00:00:01,netdev=mynet1,mrg_rxbuf=off,mq=on,vectors=6
\
-nographic"
- The guest subsequently starts as normal, but then hangs completely, rendering
it completely unusable.
The last lines of ovs-vswitchd.log read as follows:
|00051|dpdk|INFO|VHOST_CONFIG: read message VHOST_USER_SET_VRING_ENABLE
|00052|dpdk|INFO|VHOST_CONFIG: set queue enable: 1 to qp idx: 1
|00053|dpdk|INFO|VHOST_CONFIG: read message VHOST_USER_SET_FEATURES
Moving from QEMU v2.7.0 to v2.10.0 resolves the issue. However, herein lies the
issue: QEMU v2.10.0 was only released in August of this year; anecdotally, we
know that many OvS-DPDK customers use older versions of QEMU (typically,
v2.7.0), and are likely un[able|willing] to move. With this patch, a hard
dependency on QEMU v2.10 is created for users who want to use the vHU multiq
feature in DPDK v17.11 (and subsequently, the upcoming OvS v2.9.0), which IMO
will likely be unacceptable for many.
One potential solution to this problem is to introduce a compile-time option
that would allow the user to [dis|en]able the VHOST_USER_PROTOCOL_F_REPLY_ACK
feature - is that something that would be acceptable to you Maxime?
Thanks in advance,
Mark
>From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Maxime Coquelin
>Sent: Thursday, October 5, 2017 9:36 AM
>To: dev@dpdk.org; Horton, Remy <remy.hor...@intel.com>; Bie, Tiwei
><tiwei....@intel.com>; y...@fridaylinux.org
>Cc: m...@redhat.com; jfrei...@redhat.com; vkapl...@redhat.com;
>jasow...@redhat.com; Maxime Coquelin <maxime.coque...@redhat.com>
>Subject: [dpdk-dev] [PATCH v3 01/19] Revert "vhost: workaround MQ fails to
>startup"
>
>This reverts commit 04d81227960b5c1cf2f11f492100979ead20c526.
>
>As agreed when this workaround was introduced, it can be reverted
>as Qemu v2.10 that fixes the issue is now out.
>
>The reply-ack feature is required for vhost-user IOMMU support.
>
>Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
>---
> lib/librte_vhost/vhost_user.h | 6 +-----
> 1 file changed, 1 insertion(+), 5 deletions(-)
>
>diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
>index 35ebd7190..2ba22dbb0 100644
>--- a/lib/librte_vhost/vhost_user.h
>+++ b/lib/librte_vhost/vhost_user.h
>@@ -49,14 +49,10 @@
> #define VHOST_USER_PROTOCOL_F_REPLY_ACK 3
> #define VHOST_USER_PROTOCOL_F_NET_MTU 4
>
>-/*
>- * disable REPLY_ACK feature to workaround the buggy QEMU implementation.
>- * Proved buggy QEMU includes v2.7 - v2.9.
>- */
> #define VHOST_USER_PROTOCOL_FEATURES ((1ULL << VHOST_USER_PROTOCOL_F_MQ) | \
> (1ULL <<
> VHOST_USER_PROTOCOL_F_LOG_SHMFD) |\
> (1ULL << VHOST_USER_PROTOCOL_F_RARP) |
> \
>- (0ULL <<
>VHOST_USER_PROTOCOL_F_REPLY_ACK) | \
>+ (1ULL <<
>VHOST_USER_PROTOCOL_F_REPLY_ACK) | \
> (1ULL <<
> VHOST_USER_PROTOCOL_F_NET_MTU))
>
> typedef enum VhostUserRequest {
>--
>2.13.6
