Re: [PATCH v3 17/19] vhost-user: iommu: postpone device creation until ring are mapped

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 11/02/2017 05:02 PM, Maxime Coquelin wrote:
> 
> 
> On 11/02/2017 09:21 AM, Maxime Coquelin wrote:
>> Hi Lei,
>>
>> On 11/02/2017 08:21 AM, Yao, Lei A wrote:
>>>
>> ...
>>> Hi, Maxime > I met one issue with your patch set during the v17.11 test.
>>
>> Is it with v17.11-rc2 or -rc1?
>>
>>> The test scenario is following,
>>> 1.    Bind one NIC, use test-pmd set vhost-user with 2 queue
>>> usertools/dpdk-devbind.py --bind=igb_uio 0000:05:00.0
>>> ./x86_64-native-linuxapp-gcc/app/testpmd -c 0xe -n 4 --socket-mem 
>>> 1024,1024 \
>>> --vdev 'net_vhost0,iface=vhost-net,queues=2' - -i --rxq=2 --txq=2 
>>> --nb-cores=2 --rss-ip
>>> 2.    Launch qemu with  virtio device which has 2 queue
>>> 3.    In VM, launch testpmd with virtio-pmd using only 1 queue.
>>> x86_64-native-linuxapp-gcc/app/testpmd -c 0x07 -n 3 - -i 
>>> --txqflags=0xf01 \
>>> --rxq=1 --txq=1 --rss-ip --nb-cores=1
>>>
>>> First,
>>> commit 09927b5249694bad1c094d3068124673722e6b8f
>>> vhost: translate ring addresses when IOMMU enabled
>>> The patch causes no traffic in PVP test. but link status is still up 
>>> in vhost-user.
>>>
>>> Second,
>>> eefac9536a901a1f0bb52aa3b6fec8f375f09190
>>> vhost: postpone device creation until rings are mapped
>>> The patch causes link status "down" in vhost-user.
> 
> I reproduced this one, and understand why link status remains down.
> My series did fixed a potential issue Michael raised, that the vring
> addresses should only interpreted once the ring is enabled.
> When VHOST_USER_F_PROTOCOL_FEATURES is negotiated, the rings addrs are
> translated when ring is enabled via VHOST_USER_SET_VRING_ENABLE.
> When not negotiated, the ring is considered started enabled, so
> translation is done at VHOST_USER_SET_VRING_KICK time.
> 
> In your case, protocol features are negotiated, so the ring addresses
> are translated at enable time. The problem is that the code considers
> the device is ready once addresses for all the rings are translated.
> But since only the first pair of rings is used, it never happens, and
> the link remains down.
> 
> One of the reason this check is done is to avoid starting the PMD
> threads before the addresses are translated in case of NUMA
> reallocation, as virtqueues and virtio-net device structs can be
> reallocated on a different node.
> 
> I think the right fix would be to only perform NUMA reallocation for
> vring 0, as today we would end-up reallocating virtio-net struct
> mulitple time if VQs are on different NUMA nodes.
> 
> Doing that, we could then consider the device is ready if vring 0 is
> enabled and its ring addresses are translated, and if other vrings have
> been kicked.
> 
> I'll post a patch shortly implementing this idea.

The proposed solution doesn't work, because disabled queues get accessed 
at device start time:

int
rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
{
..
	dev->virtqueue[queue_id]->used->flags = VRING_USED_F_NO_NOTIFY;
	return 0;
}

The above function being called in Vhost PMD for every queues, enabled
or not. While we could fix the PMD, it could break other applications
using the Vhost lib API directly, so we cannot translate at enable
time reliably.

I think we may be a bit less conservative, and postpone addresses
translation at kick time, whatever VHOST_USER_F_PROTOCOL_FEATURES is
negotiated or not.

Regards,
Maxime

> Thanks,
> Maxime