Linux Virtualization - Feb 2014 - [PATCH net] vhost: net: switch to use data copy if pending DMAs exceed the limit

On 02/25/2014 09:57 PM, Michael S. Tsirkin wrote:
> On Tue, Feb 25, 2014 at 02:53:58PM +0800, Jason Wang wrote:
>> We used to stop the handling of tx when the number of pending DMAs
>> exceeds VHOST_MAX_PEND. This is used to reduce the memory occupation
>> of both host and guest. But it was too aggressive in some cases, since
>> any delay or blocking of a single packet may delay or block the guest
>> transmission. Consider the following setup:
>>
>>      +-----+        +-----+
>>      | VM1 |        | VM2 |
>>      +--+--+        +--+--+
>>         |              |
>>      +--+--+        +--+--+
>>      | tap0|        | tap1|
>>      +--+--+        +--+--+
>>         |              |
>>      pfifo_fast   htb(10Mbit/s)
>>         |              |
>>      +--+--------------+---+
>>      |     bridge          |
>>      +--+------------------+
>>         |
>>      pfifo_fast
>>         |
>>      +-----+
>>      | eth0|(100Mbit/s)
>>      +-----+
>>
>> - start two VMs and connect them to a bridge
>> - add an physical card (100Mbit/s) to that bridge
>> - setup htb on tap1 and limit its throughput to 10Mbit/s
>> - run two netperfs in the same time, one is from VM1 to VM2. Another is
>>    from VM1 to an external host through eth0.
>> - result shows that not only the VM1 to VM2 traffic were throttled but
>>    also the VM1 to external host through eth0 is also throttled
somehow.
>>
>> This is because the delay added by htb may lead the delay the finish
>> of DMAs and cause the pending DMAs for tap0 exceeds the limit
>> (VHOST_MAX_PEND). In this case vhost stop handling tx request until
>> htb send some packets. The problem here is all of the packets
>> transmission were blocked even if it does not go to VM2.
>>
>> We can solve this issue by relaxing it a little bit: switching to use
>> data copy instead of stopping tx when the number of pending DMAs
>> exceed the VHOST_MAX_PEND. This is safe because:
>>
>> - The number of pending DMAs were still limited by VHOST_MAX_PEND
>> - The out of order completion during mode switch can make sure that
>>    most of the tx buffers were freed in time in guest.
>>
>> So even if about 50% packets were delayed in zero-copy case, vhost
>> could continue to do the transmission through data copy in this case.
>>
>> Test result:
>>
>> Before this patch:
>> VM1 to VM2 throughput is 9.3Mbit/s
>> VM1 to External throughput is 40Mbit/s
>>
>> After this patch:
>> VM1 to VM2 throughput is 9.3Mbit/s
>> Vm1 to External throughput is 93Mbit/s
> Would like to see CPU utilization #s as well.
>
Will measure this.
>> Simple performance test on 40gbe shows no obvious changes in
>> throughput after this patch.
>>
>> The patch only solve this issue when unlimited sndbuf. We still need a
>> solution for limited sndbuf.
>>
>> Cc: Michael S. Tsirkin<mst at redhat.com>
>> Cc: Qin Chuanyu<qinchuanyu at huawei.com>
>> Signed-off-by: Jason Wang<jasowang at redhat.com>
> I think this needs some thought.
>
> In particular I think this works because VHOST_MAX_PEND
> is much smaller than the ring size.
> Shouldn't max_pend then be tied to the ring size if it's small?
>
Yes it should. I just reuse the VHOST_MAX_PEND since it was there for a 
long time.
> Another question is about stopping vhost:
> ATM it's waiting for skbs to complete.
> Should we maybe hunt down skbs queued and destroy them
> instead?
> I think this happens when a device is removed.
>
> Thoughts?
>
Agree, vhost net removal should not be blocked by a skb. But since the 
skbs could be queued may places, just destroy them may need extra locks.

Haven't thought this deeply, but another possible sloution is to rcuify 
destructor_arg and assign it to NULL during vhost_net
removing.
>> ---
>>   drivers/vhost/net.c | 17 +++++++----------
>>   1 file changed, 7 insertions(+), 10 deletions(-)
>>
>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>> index a0fa5de..3e96e47 100644
>> --- a/drivers/vhost/net.c
>> +++ b/drivers/vhost/net.c
>> @@ -345,7 +345,7 @@ static void handle_tx(struct vhost_net *net)
>>   		.msg_flags = MSG_DONTWAIT,
>>   	};
>>   	size_t len, total_len = 0;
>> -	int err;
>> +	int err, num_pends;
>>   	size_t hdr_size;
>>   	struct socket *sock;
>>   	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
>> @@ -366,13 +366,6 @@ static void handle_tx(struct vhost_net *net)
>>   		if (zcopy)
>>   			vhost_zerocopy_signal_used(net, vq);
>>
>> -		/* If more outstanding DMAs, queue the work.
>> -		 * Handle upend_idx wrap around
>> -		 */
>> -		if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
>> -			      % UIO_MAXIOV == nvq->done_idx))
>> -			break;
>> -
>>   		head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
>>   					 ARRAY_SIZE(vq->iov),
>>   					&out,&in,
>> @@ -405,9 +398,13 @@ static void handle_tx(struct vhost_net *net)
>>   			break;
>>   		}
>>
>> +		num_pends = likely(nvq->upend_idx>= nvq->done_idx) ?
>> +			    (nvq->upend_idx - nvq->done_idx) :
>> +			    (nvq->upend_idx + UIO_MAXIOV -
>> +			     nvq->done_idx);
>> +
>>   		zcopy_used = zcopy&&  len>= VHOST_GOODCOPY_LEN
>> -				&&  (nvq->upend_idx + 1) % UIO_MAXIOV !>> -				  
nvq->done_idx
>> +				&&  num_pends<= VHOST_MAX_PEND
>>   				&&  vhost_net_tx_select_zcopy(net);
>>
>>   		/* use msg_control to pass vhost zerocopy ubuf info to skb */
>> -- 
>> 1.8.3.2

On 2014/2/26 13:53, Jason Wang wrote:
> On 02/25/2014 09:57 PM, Michael S. Tsirkin wrote:
>> On Tue, Feb 25, 2014 at 02:53:58PM +0800, Jason Wang wrote:
>>> We used to stop the handling of tx when the number of pending DMAs
>>> exceeds VHOST_MAX_PEND. This is used to reduce the memory
occupation
>>> of both host and guest. But it was too aggressive in some cases,
since
>>> any delay or blocking of a single packet may delay or block the
guest
>>> transmission. Consider the following setup:
>>>
>>>      +-----+        +-----+
>>>      | VM1 |        | VM2 |
>>>      +--+--+        +--+--+
>>>         |              |
>>>      +--+--+        +--+--+
>>>      | tap0|        | tap1|
>>>      +--+--+        +--+--+
>>>         |              |
>>>      pfifo_fast   htb(10Mbit/s)
>>>         |              |
>>>      +--+--------------+---+
>>>      |     bridge          |
>>>      +--+------------------+
>>>         |
>>>      pfifo_fast
>>>         |
>>>      +-----+
>>>      | eth0|(100Mbit/s)
>>>      +-----+
>>>
>>> - start two VMs and connect them to a bridge
>>> - add an physical card (100Mbit/s) to that bridge
>>> - setup htb on tap1 and limit its throughput to 10Mbit/s
>>> - run two netperfs in the same time, one is from VM1 to VM2.
Another is
>>>    from VM1 to an external host through eth0.
>>> - result shows that not only the VM1 to VM2 traffic were throttled
but
>>>    also the VM1 to external host through eth0 is also throttled
somehow.
>>>
>>> This is because the delay added by htb may lead the delay the
finish
>>> of DMAs and cause the pending DMAs for tap0 exceeds the limit
>>> (VHOST_MAX_PEND). In this case vhost stop handling tx request until
>>> htb send some packets. The problem here is all of the packets
>>> transmission were blocked even if it does not go to VM2.
>>>
>>> We can solve this issue by relaxing it a little bit: switching to
use
>>> data copy instead of stopping tx when the number of pending DMAs
>>> exceed the VHOST_MAX_PEND. This is safe because:
>>>
>>> - The number of pending DMAs were still limited by VHOST_MAX_PEND
>>> - The out of order completion during mode switch can make sure that
>>>    most of the tx buffers were freed in time in guest.
>>>
>>> So even if about 50% packets were delayed in zero-copy case, vhost
>>> could continue to do the transmission through data copy in this
case.
>>>
>>> Test result:
>>>
>>> Before this patch:
>>> VM1 to VM2 throughput is 9.3Mbit/s
>>> VM1 to External throughput is 40Mbit/s
>>>
>>> After this patch:
>>> VM1 to VM2 throughput is 9.3Mbit/s
>>> Vm1 to External throughput is 93Mbit/s
>> Would like to see CPU utilization #s as well.
>>
>
> Will measure this.
>>> Simple performance test on 40gbe shows no obvious changes in
>>> throughput after this patch.
>>>
>>> The patch only solve this issue when unlimited sndbuf. We still
need a
>>> solution for limited sndbuf.
>>>
>>> Cc: Michael S. Tsirkin<mst at redhat.com>
>>> Cc: Qin Chuanyu<qinchuanyu at huawei.com>
>>> Signed-off-by: Jason Wang<jasowang at redhat.com>
>> I think this needs some thought.
>>
>> In particular I think this works because VHOST_MAX_PEND
>> is much smaller than the ring size.
>> Shouldn't max_pend then be tied to the ring size if it's small?
>>
>
> Yes it should. I just reuse the VHOST_MAX_PEND since it was there for a
> long time.
>> Another question is about stopping vhost:
>> ATM it's waiting for skbs to complete.
>> Should we maybe hunt down skbs queued and destroy them
>> instead?
>> I think this happens when a device is removed.
>>
>> Thoughts?
>>
>
> Agree, vhost net removal should not be blocked by a skb. But since the
> skbs could be queued may places, just destroy them may need extra locks.
>
> Haven't thought this deeply, but another possible sloution is to rcuify
> destructor_arg and assign it to NULL during vhost_net removing.
Xen treat it by a timer, for those skbs which has been delivered for a
while, netback would exchange page of zero_copy's skb with dom0's page.

but there is still a race between host's another process handle the skb
and netback exchange its page. (This problem has been proved by testing)

and Xen hasn't solved this problem yet, because if anyone want to solve
this problem completely, a page lock is necessary, but it would be
complex and expensive.

rcuify destructor arg and assign it to NULL couldn't solve the problem
of page release that has been reserved by host's another process.

The key problem is how to release the memory of zero_copy's skb while
been reserved.
>>> ---
>>>   drivers/vhost/net.c | 17 +++++++----------
>>>   1 file changed, 7 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>>> index a0fa5de..3e96e47 100644
>>> --- a/drivers/vhost/net.c
>>> +++ b/drivers/vhost/net.c
>>> @@ -345,7 +345,7 @@ static void handle_tx(struct vhost_net *net)
>>>           .msg_flags = MSG_DONTWAIT,
>>>       };
>>>       size_t len, total_len = 0;
>>> -    int err;
>>> +    int err, num_pends;
>>>       size_t hdr_size;
>>>       struct socket *sock;
>>>       struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
>>> @@ -366,13 +366,6 @@ static void handle_tx(struct vhost_net *net)
>>>           if (zcopy)
>>>               vhost_zerocopy_signal_used(net, vq);
>>>
>>> -        /* If more outstanding DMAs, queue the work.
>>> -         * Handle upend_idx wrap around
>>> -         */
>>> -        if (unlikely((nvq->upend_idx + vq->num -
VHOST_MAX_PEND)
>>> -                  % UIO_MAXIOV == nvq->done_idx))
>>> -            break;
>>> -
>>>           head = vhost_get_vq_desc(&net->dev, vq,
vq->iov,
>>>                        ARRAY_SIZE(vq->iov),
>>>                       &out,&in,
>>> @@ -405,9 +398,13 @@ static void handle_tx(struct vhost_net *net)
>>>               break;
>>>           }
>>>
>>> +        num_pends = likely(nvq->upend_idx>=
nvq->done_idx) ?
>>> +                (nvq->upend_idx - nvq->done_idx) :
>>> +                (nvq->upend_idx + UIO_MAXIOV -
>>> +                 nvq->done_idx);
>>> +
>>>           zcopy_used = zcopy&&  len>= VHOST_GOODCOPY_LEN
>>> -                &&  (nvq->upend_idx + 1) % UIO_MAXIOV
!>>> -                      nvq->done_idx
>>> +                &&  num_pends<= VHOST_MAX_PEND
>>>                   &&  vhost_net_tx_select_zcopy(net);
>>>
>>>           /* use msg_control to pass vhost zerocopy ubuf info to
skb */
>>> --
>>> 1.8.3.2
>
>
> .
>

On 02/26/2014 02:32 PM, Qin Chuanyu wrote:
> On 2014/2/26 13:53, Jason Wang wrote:
>> On 02/25/2014 09:57 PM, Michael S. Tsirkin wrote:
>>> On Tue, Feb 25, 2014 at 02:53:58PM +0800, Jason Wang wrote:
>>>> We used to stop the handling of tx when the number of pending
DMAs
>>>> exceeds VHOST_MAX_PEND. This is used to reduce the memory
occupation
>>>> of both host and guest. But it was too aggressive in some
cases, since
>>>> any delay or blocking of a single packet may delay or block the
guest
>>>> transmission. Consider the following setup:
>>>>
>>>>      +-----+        +-----+
>>>>      | VM1 |        | VM2 |
>>>>      +--+--+        +--+--+
>>>>         |              |
>>>>      +--+--+        +--+--+
>>>>      | tap0|        | tap1|
>>>>      +--+--+        +--+--+
>>>>         |              |
>>>>      pfifo_fast   htb(10Mbit/s)
>>>>         |              |
>>>>      +--+--------------+---+
>>>>      |     bridge          |
>>>>      +--+------------------+
>>>>         |
>>>>      pfifo_fast
>>>>         |
>>>>      +-----+
>>>>      | eth0|(100Mbit/s)
>>>>      +-----+
>>>>
>>>> - start two VMs and connect them to a bridge
>>>> - add an physical card (100Mbit/s) to that bridge
>>>> - setup htb on tap1 and limit its throughput to 10Mbit/s
>>>> - run two netperfs in the same time, one is from VM1 to VM2. 
>>>> Another is
>>>>    from VM1 to an external host through eth0.
>>>> - result shows that not only the VM1 to VM2 traffic were
throttled but
>>>>    also the VM1 to external host through eth0 is also throttled
>>>> somehow.
>>>>
>>>> This is because the delay added by htb may lead the delay the
finish
>>>> of DMAs and cause the pending DMAs for tap0 exceeds the limit
>>>> (VHOST_MAX_PEND). In this case vhost stop handling tx request
until
>>>> htb send some packets. The problem here is all of the packets
>>>> transmission were blocked even if it does not go to VM2.
>>>>
>>>> We can solve this issue by relaxing it a little bit: switching
to use
>>>> data copy instead of stopping tx when the number of pending
DMAs
>>>> exceed the VHOST_MAX_PEND. This is safe because:
>>>>
>>>> - The number of pending DMAs were still limited by
VHOST_MAX_PEND
>>>> - The out of order completion during mode switch can make sure
that
>>>>    most of the tx buffers were freed in time in guest.
>>>>
>>>> So even if about 50% packets were delayed in zero-copy case,
vhost
>>>> could continue to do the transmission through data copy in this
case.
>>>>
>>>> Test result:
>>>>
>>>> Before this patch:
>>>> VM1 to VM2 throughput is 9.3Mbit/s
>>>> VM1 to External throughput is 40Mbit/s
>>>>
>>>> After this patch:
>>>> VM1 to VM2 throughput is 9.3Mbit/s
>>>> Vm1 to External throughput is 93Mbit/s
>>> Would like to see CPU utilization #s as well.
>>>
>>
>> Will measure this.
>>>> Simple performance test on 40gbe shows no obvious changes in
>>>> throughput after this patch.
>>>>
>>>> The patch only solve this issue when unlimited sndbuf. We still
need a
>>>> solution for limited sndbuf.
>>>>
>>>> Cc: Michael S. Tsirkin<mst at redhat.com>
>>>> Cc: Qin Chuanyu<qinchuanyu at huawei.com>
>>>> Signed-off-by: Jason Wang<jasowang at redhat.com>
>>> I think this needs some thought.
>>>
>>> In particular I think this works because VHOST_MAX_PEND
>>> is much smaller than the ring size.
>>> Shouldn't max_pend then be tied to the ring size if it's
small?
>>>
>>
>> Yes it should. I just reuse the VHOST_MAX_PEND since it was there for a
>> long time.
>>> Another question is about stopping vhost:
>>> ATM it's waiting for skbs to complete.
>>> Should we maybe hunt down skbs queued and destroy them
>>> instead?
>>> I think this happens when a device is removed.
>>>
>>> Thoughts?
>>>
>>
>> Agree, vhost net removal should not be blocked by a skb. But since the
>> skbs could be queued may places, just destroy them may need extra
locks.
>>
>> Haven't thought this deeply, but another possible sloution is to
rcuify
>> destructor_arg and assign it to NULL during vhost_net removing.
>
> Xen treat it by a timer, for those skbs which has been delivered for a
> while, netback would exchange page of zero_copy's skb with dom0's
page.
>
> but there is still a race between host's another process handle the skb
> and netback exchange its page. (This problem has been proved by testing)
>
> and Xen hasn't solved this problem yet, because if anyone want to solve
> this problem completely, a page lock is necessary, but it would be
> complex and expensive.
>
> rcuify destructor arg and assign it to NULL couldn't solve the problem
> of page release that has been reserved by host's another process.
>
There're two issues:

1) if a zerocopy skb won't be freed or frags orphaned in time, vhost_net 
removal will be blocked since it was waiting for the refcnt of ubuf to zero.
2) whether or not we should free all pending skbs during vhost_net removing.

My proposal is for issue 1. Another idea is not wait for the refcnt to 
be zero and then we can defer the freeing of vhost_net during the 
release method of kref_put().

For issue 2, I'm still not sure we should do this or not. Looks like 
there's a similar issue for the packets sent by tcp_sendpage() was 
blocked or delayed.
> The key problem is how to release the memory of zero_copy's skb while
> been reserved.