Virtualization - Jun 2022 - [RFC PATCH v2 0/8] virtio/vsock: experimental zerocopy receive

On Thu, Jun 09, 2022 at 12:33:32PM +0000, Arseniy Krasnov
wrote:
>On 09.06.2022 11:54, Stefano Garzarella wrote:
>> Hi Arseniy,
>> I left some comments in the patches, and I'm adding something also
here:
>Thanks for comments
>>
>> On Fri, Jun 03, 2022 at 05:27:56AM +0000, Arseniy Krasnov wrote:
>>> ???????????????????????????? INTRODUCTION
>>>
>>> ????Hello, this is experimental implementation of virtio vsock
zerocopy
>>> receive. It was inspired by TCP zerocopy receive by Eric Dumazet.
This API uses
>>> same idea: call 'mmap()' on socket's descriptor, then
every 'getsockopt()' will
>>> fill provided vma area with pages of virtio RX buffers. After
received data was
>>> processed by user, pages must be freed by 'madvise()'? call
with MADV_DONTNEED
>>> flag set(if user won't call 'madvise()', next
'getsockopt()' will fail).
>>
>> If it is not too time-consuming, can we have a table/list to compare
this and the TCP zerocopy?
>You mean compare API with more details?
Yes, maybe a comparison from the user's point of view to do zero-copy 
with TCP and VSOCK.
>>
>>>
>>> ??????????????????????????????? DETAILS
>>>
>>> ????Here is how mapping with mapped pages looks exactly: first page
mapping
>>> contains array of trimmed virtio vsock packet headers (in contains
only length
>>> of data on the corresponding page and 'flags' field):
>>>
>>> ????struct virtio_vsock_usr_hdr {
>>> ??????? uint32_t length;
>>> ??????? uint32_t flags;
>>> ??????? uint32_t copy_len;
>>> ????};
>>>
>>> Field? 'length' allows user to know exact size of payload
within each sequence
>>> of pages and 'flags' allows user to handle SOCK_SEQPACKET
flags(such as message
>>> bounds or record bounds). Field 'copy_len' is described
below in 'v1->v2' part.
>>> All other pages are data pages from RX queue.
>>>
>>> ??????????? Page 0????? Page 1????? Page N
>>>
>>> ????[ hdr1 .. hdrN ][ data ] .. [ data ]
>>> ????????? |??????? |?????? ^?????????? ^
>>> ????????? |??????? |?????? |?????????? |
>>> ????????? |??????? *-------------------*
>>> ????????? |??????????????? |
>>> ????????? |??????????????? |
>>> ????????? *----------------*
>>>
>>> ????Of course, single header could represent array of pages (when
packet's
>>> buffer is bigger than one page).So here is example of detailed
mapping layout
>>> for some set of packages. Lets consider that we have the following
sequence? of
>>> packages: 56 bytes, 4096 bytes and 8200 bytes. All pages: 0,1,2,3,4
and 5 will
>>> be inserted to user's vma(vma is large enough).
>>
>> In order to have a "userspace polling-friendly approach" and
reduce number of syscall, can we allow for example the userspace to mmap at
least the first header before packets arrive.
>> Then the userspace can poll a flag or other fields in the header to
understand that there are new packets.
>You mean to avoid 'poll()' syscall, user will spin on some flag,
provided by kernel on some mapped page? I think yes. This is ok. Also i think,
that i can avoid 'madvise' call
>to clear memory mapping before each 'getsockopt()' - let
'getsockopt()' do 'madvise()' job by removing pages from
previous data. In this case only one system call is needed -
'getsockopt()'.
Yes, that's right. I mean to support both, poll() for interrupt-based 
applications and the ability to actively poll a variable in the shared 
memory for applications that want to minimize latency.

Thanks,
Stefano