Re: [Qemu-devel] [ RFC Patch v5 0/2] Support Receive-Segment-Offload(RSC) for WHQL

[Wei Xu <wexu@redhat.com>]

On 2016年05月24日 16:26, Michael S. Tsirkin wrote:
> On Tue, May 24, 2016 at 04:03:04PM +0800, Jason Wang wrote:
>>
>>
>> On 2016年05月24日 04:14, wexu@redhat.com wrote:
>>> From: Wei Xu <wexu@redhat.com>
>>>
>>> Changes in V5:
>>> - Passed all IPv4/6 test cases
>>> - Add new fields in 'virtio_net_hdr'
>>> - Set 'gso_type' & 'coalesced packets' in new field.
>>> - Bypass all 'tcp option' packet
>>> - Bypass all 'pure ack' packet
>>> - Bypass all 'duplicate ack' packet
>>> - Change 'guest_rsc' feature bit to 'false' by default
>>> - Feedbacks from v4, typo, etc.
>>
>> Patch does not apply on master ...
>>
>>>
>>> Note:
>>> There is still a few pending issues about the feature bit, and need to be
>>> discussed with windows driver maintainer, so linux guests with this patch
>>> won't work at current, haven't figure it out yet, but i'm guessing it's
>>> caused by the 'gso_type' is set to 'VIRTIO_NET_HDR_GSO_TCPV4/6',
>>> will fix it after get the final solution, the below test steps and
>>> performance data is based on v4.
>>
>> Can we split the patches into smaller ones to make review or merging easier?
>> E.g can we send the patches without any feature negotiation and vnet header
>> extension?
>>
>> We can focus on the coalescing (maybe ipv4) without any guest involvement in
>> this series. In this way, the issues were limited and can be converged soon.
>> After this has been merged, we can add patches that co-operate with guests
>> on top (since it needs agreement on virtio specs). Does this sounds a good
>> plan?
>
> True but disabling everything when feature is not negotiated
> reduces the risk somewhat.
Sure.
>
>>>
>>> Another suggestion from Jason is to adjust part of the code to make it
>>> more readable, since there maybe still few change about the flowchart
>>> in the future, such as timestamp, duplicate ack, so i'd like to delay it
>>> temporarily.
>>>
>>> Changes in V4:
>>> - Add new host feature bit
>>> - Replace using fixed header lenght with dynamic header lenght in VirtIONet
>>> - Change ip/ip6 header union in NetRscUnit to void* pointer
>>> - Add macro prefix, adjust code indent, etc.
>>>
>>> Changes in V3:
>>> - Removed big param list, replace it with 'NetRscUnit'
>>> - Different virtio header size
>>> - Modify callback function to direct call.
>>> - Needn't check the failure of g_malloc()
>>> - Other code format adjustment, macro naming, etc
>>>
>>> Changes in V2:
>>> - Add detailed commit log
>>>
>>> This patch is to support WHQL test for Windows guest, while this feature also
>>> benifits other guest works as a kernel 'gro' like feature with userspace
>>> implementation.
>>> Feature information:
>>>    http://msdn.microsoft.com/en-us/library/windows/hardware/jj853324
>>>
>>> Both IPv4 and IPv6 are supported, though performance with userspace virtio
>>> is slow than vhost-net, there is about 1.5x to 2x performance improvement to
>>> userspace virtio, this is done by turning this feature on and disable
>>> 'tso/gso/gro' on corresponding tap interface and guest interface, while get
>>> less improment with all these feature on.
>>>
>>> Linux guest performance data(Netperf):
>>> MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.2.101 () port 0 AF_INET : nodelay
>>> Size   Size    Size     Time     Throughput
>>> bytes  bytes   bytes    secs.    10^6bits/sec
>>>
>>>   87380  16384     64    6.00     1221.20
>>>   87380  16384     64    6.00     1260.30
>>>
>>>   87380  16384    128    6.00     1978.51
>>>   87380  16384    128    6.00     2286.05
>>>
>>>   87380  16384    256    6.00     2677.94
>>>   87380  16384    256    6.00     4615.42
>>>
>>>   87380  16384    512    6.00     2956.54
>>>   87380  16384    512    6.00     5356.39
>>>
>>>   87380  16384   1024    6.00     2798.17
>>>   87380  16384   1024    6.00     4943.30
>>>
>>>   87380  16384   2048    6.00     2681.09
>>>   87380  16384   2048    6.00     4835.81
>>>
>>>   87380  16384   4096    6.00     3390.14
>>>   87380  16384   4096    6.00     5391.54
>>>
>>>   87380  16384   8092    6.00     3008.27
>>>   87380  16384   8092    6.00     5381.68
>>>
>>>   87380  16384  10240    6.00     2999.89
>>>   87380  16384  10240    6.00     5393.11
>>>
>>> Test steps:
>>> Although this feature is mainly used for window guest, i used linux guest to
>>> help test the feature, to make things simple, i used 3 steps to test the patch
>>> as i moved on.
>>>
>>> 1. With a tcp socket client/server pair running on 2 linux guest, thus i can
>>> control
>>> the traffic and debugging the code as i want.
>>> 2. Netperf on linux guest test the throughput.
>>> 3. WHQL test with 2 Windows guests.
>>>
>>> Wei Xu (2):
>>>    virtio-net rsc: support coalescing ipv4 tcp traffic
>>>    virtio-net rsc: support coalescing ipv6 tcp traffic
>>>
>>>   hw/net/virtio-net.c                         | 623 +++++++++++++++++++++++++++-
>>>   include/hw/virtio/virtio-net.h              |   2 +
>>>   include/hw/virtio/virtio.h                  |  75 ++++
>>>   include/standard-headers/linux/virtio_net.h |   2 +
>>>   4 files changed, 699 insertions(+), 3 deletions(-)
>>>