Re: [RFC PATCH v3 00/13] virtio/vsock: introduce SOCK_SEQPACKET support

[Stefano Garzarella <sgarzare@redhat.com>]

On Mon, Feb 01, 2021 at 04:57:18PM +0300, Arseny Krasnov wrote:
>
>On 01.02.2021 14:02, Stefano Garzarella wrote:
>> On Fri, Jan 29, 2021 at 06:52:23PM +0300, Arseny Krasnov wrote:
>>> On 29.01.2021 12:26, Stefano Garzarella wrote:
>>>> On Fri, Jan 29, 2021 at 09:41:50AM +0300, Arseny Krasnov wrote:
>>>>> On 28.01.2021 20:19, Stefano Garzarella wrote:
>>>>>> Hi Arseny,
>>>>>> I reviewed a part, tomorrow I hope to finish the other patches.
>>>>>>
>>>>>> Just a couple of comments in the TODOs below.
>>>>>>
>>>>>> On Mon, Jan 25, 2021 at 02:09:00PM +0300, Arseny Krasnov wrote:
>>>>>>> 	This patchset impelements support of SOCK_SEQPACKET for virtio
>>>>>>> transport.
>>>>>>> 	As SOCK_SEQPACKET guarantees to save record boundaries, so to
>>>>>>> do it, new packet operation was added: it marks start of record (with
>>>>>>> record length in header), such packet doesn't carry any data.  To send
>>>>>>> record, packet with start marker is sent first, then all data is sent
>>>>>>> as usual 'RW' packets. On receiver's side, length of record is known
>>>>>> >from packet with start record marker. Now as  packets of one socket
>>>>>>> are not reordered neither on vsock nor on vhost transport layers, such
>>>>>>> marker allows to restore original record on receiver's side. If user's
>>>>>>> buffer is smaller that record length, when all out of size data is
>>>>>>> dropped.
>>>>>>> 	Maximum length of datagram is not limited as in stream socket,
>>>>>>> because same credit logic is used. Difference with stream socket is
>>>>>>> that user is not woken up until whole record is received or error
>>>>>>> occurred. Implementation also supports 'MSG_EOR' and 'MSG_TRUNC' flags.
>>>>>>> 	Tests also implemented.
>>>>>>>
>>>>>>> Arseny Krasnov (13):
>>>>>>>  af_vsock: prepare for SOCK_SEQPACKET support
>>>>>>>  af_vsock: prepare 'vsock_connectible_recvmsg()'
>>>>>>>  af_vsock: implement SEQPACKET rx loop
>>>>>>>  af_vsock: implement send logic for SOCK_SEQPACKET
>>>>>>>  af_vsock: rest of SEQPACKET support
>>>>>>>  af_vsock: update comments for stream sockets
>>>>>>>  virtio/vsock: dequeue callback for SOCK_SEQPACKET
>>>>>>>  virtio/vsock: fetch length for SEQPACKET record
>>>>>>>  virtio/vsock: add SEQPACKET receive logic
>>>>>>>  virtio/vsock: rest of SOCK_SEQPACKET support
>>>>>>>  virtio/vsock: setup SEQPACKET ops for transport
>>>>>>>  vhost/vsock: setup SEQPACKET ops for transport
>>>>>>>  vsock_test: add SOCK_SEQPACKET tests
>>>>>>>
>>>>>>> drivers/vhost/vsock.c                   |   7 +-
>>>>>>> include/linux/virtio_vsock.h            |  12 +
>>>>>>> include/net/af_vsock.h                  |   6 +
>>>>>>> include/uapi/linux/virtio_vsock.h       |   9 +
>>>>>>> net/vmw_vsock/af_vsock.c                | 543 ++++++++++++++++------
>>>>>>> net/vmw_vsock/virtio_transport.c        |   4 +
>>>>>>> net/vmw_vsock/virtio_transport_common.c | 295 ++++++++++--
>>>>>>> tools/testing/vsock/util.c              |  32 +-
>>>>>>> tools/testing/vsock/util.h              |   3 +
>>>>>>> tools/testing/vsock/vsock_test.c        | 126 +++++
>>>>>>> 10 files changed, 862 insertions(+), 175 deletions(-)
>>>>>>>
>>>>>>> TODO:
>>>>>>> - Support for record integrity control. As transport could drop some
>>>>>>>   packets, something like "record-id" and record end marker need to
>>>>>>>   be implemented. Idea is that SEQ_BEGIN packet carries both record
>>>>>>>   length and record id, end marker(let it be SEQ_END) carries only
>>>>>>>   record id. To be sure that no one packet was lost, receiver checks
>>>>>>>   length of data between SEQ_BEGIN and SEQ_END(it must be same with
>>>>>>>   value in SEQ_BEGIN) and record ids of SEQ_BEGIN and SEQ_END(this
>>>>>>>   means that both markers were not dropped. I think that easiest way
>>>>>>>   to implement record id for SEQ_BEGIN is to reuse another field of
>>>>>>>   packet header(SEQ_BEGIN already uses 'flags' as record length).For
>>>>>>>   SEQ_END record id could be stored in 'flags'.
>>>>>> I don't really like the idea of reusing the 'flags' field for this
>>>>>> purpose.
>>>>>>
>>>>>>>     Another way to implement it, is to move metadata of both SEQ_END
>>>>>>>   and SEQ_BEGIN to payload. But this approach has problem, because
>>>>>>>   if we move something to payload, such payload is accounted by
>>>>>>>   credit logic, which fragments payload, while payload with record
>>>>>>>   length and id couldn't be fragmented. One way to overcome it is to
>>>>>>>   ignore credit update for SEQ_BEGIN/SEQ_END packet.Another solution
>>>>>>>   is to update 'stream_has_space()' function: current implementation
>>>>>>>   return non-zero when at least 1 byte is allowed to use,but updated
>>>>>>>   version will have extra argument, which is needed length. For 'RW'
>>>>>>>   packet this argument is 1, for SEQ_BEGIN it is sizeof(record len +
>>>>>>>   record id) and for SEQ_END it is sizeof(record id).
>>>>>> Is the payload accounted by credit logic also if hdr.op is not
>>>>>> VIRTIO_VSOCK_OP_RW?
>>>>> Yes, on send any packet with payload could be fragmented if
>>>>>
>>>>> there is not enough space at receiver. On receive 'fwd_cnt' and
>>>>>
>>>>> 'buf_alloc' are updated with header of every packet. Of course,
>>>>>
>>>>> to every such case i've described i can add check for 'RW'
>>>>>
>>>>> packet, to exclude payload from credit accounting, but this is
>>>>>
>>>>> bunch of dumb checks.
>>>>>
>>>>>> I think that we can define a specific header to put after the
>>>>>> virtio_vsock_hdr when hdr.op is SEQ_BEGIN or SEQ_END, and in this header
>>>>>> we can store the id and the length of the message.
>>>>> I think it is better than use payload and touch credit logic
>>>>>
>>>> Cool, so let's try this option, hoping there aren't a lot of issues.
>>> If i understand, current implementation has 'struct 
>>> virtio_vsock_hdr',
>>>
>>> then i'll add 'struct virtio_vsock_hdr_seq' with message length and id.
>>>
>>> After that, in 'struct virtio_vsock_pkt' which describes packet, field for
>>>
>>> header(which is 'struct virtio_vsock_hdr') must be replaced with new
>>>
>>> structure which  contains both 'struct virtio_vsock_hdr' and 'struct
>>>
>>> virtio_vsock_hdr_seq', because header field of 'struct virtio_vsock_pkt'
>>>
>>> is buffer for virtio layer. After it all accesses to header(for example to
>>>
>>> 'buf_alloc' field will go accross new  structure with both headers:
>>>
>>> pkt->hdr.buf_alloc   ->   pkt->extended_hdr.classic_hdr.buf_alloc
>>>
>>> May be to avoid this, packet's header could be allocated dynamically
>>>
>>> in the same manner as packet's buffer? Size of allocation is always
>>>
>>> sizeof(classic header) + sizeof(seq header). In 'struct virtio_vsock_pkt'
>>>
>>> such header will be implemented as union of two pointers: class header
>>>
>>> and extended header containing classic and seq header. Which pointer
>>>
>>> to use is depends on packet's op.
>> I think that the 'classic header' can stay as is, and the extended
>> header can be dynamically allocated, as we do for the payload.
>>
>> But we have to be careful what happens if the other peer doesn't support
>> SEQPACKET and if it counts this extra header as a payload for the credit
>> mechanism.
>
>You mean put extra header to payload(buffer of second virtio desc),
>
>in this way on send/receive auxiliary 'if's are needed to avoid credit
>
>logic(or set length field in header of such packets to 0). But what
>
>about placing extra header after classic header in buffer of first virtio
>
>desc? In this case extra header is not payload and credit works as is.
>
>Or it is critical, that size of first buffer will be not same as size of
>
>classic header?

We need to think about compatibility with old drivers.

What would happen in this case?

I think it's easier to use the second buffer, usually used for the 
payload, to carry the extra header. Also, we can leave hdr.len = 0, so 
we are sure that it is not counted in credit mechanism.

If the driver supports SEQPACKET, it knows it must fetch extra header 
when it must handle SEQ_BEGIN/SEQ_END.

If it is not clear, I'll try to provide a simple PoC of a patch.

Thanks,
Stefano