RE: [RFC PATCH v1 2/3] virtio/vsock: use 'target' in notify_poll_in, callback.

[Dexuan Cui]

> From: Stefano Garzarella <sgarzare@xxxxxxxxxx>
> Sent: Wednesday, July 20, 2022 1:23 AM
> ...
> On Wed, Jul 20, 2022 at 05:38:03AM +0000, Arseniy Krasnov wrote:
> >On 19.07.2022 15:48, Stefano Garzarella wrote:
> >> On Mon, Jul 18, 2022 at 08:17:31AM +0000, Arseniy Krasnov wrote:
> >>> This callback controls setting of POLLIN,POLLRDNORM output bits
> >>> of poll() syscall,but in some cases,it is incorrectly to set it,
> >>> when socket has at least 1 bytes of available data. Use 'target'
> >>> which is already exists and equal to sk_rcvlowat in this case.
> >>>
> >>> Signed-off-by: Arseniy Krasnov <AVKrasnov@xxxxxxxxxxxxxx>
> >>> ---
> >>> net/vmw_vsock/virtio_transport_common.c | 2 +-
> >>> 1 file changed, 1 insertion(+), 1 deletion(-)
> >>>
> >>> diff --git a/net/vmw_vsock/virtio_transport_common.c
> b/net/vmw_vsock/virtio_transport_common.c
> >>> index ec2c2afbf0d0..591908740992 100644
> >>> --- a/net/vmw_vsock/virtio_transport_common.c
> >>> +++ b/net/vmw_vsock/virtio_transport_common.c
> >>> @@ -634,7 +634,7 @@ virtio_transport_notify_poll_in(struct vsock_sock
> *vsk,
> >>>                 size_t target,
> >>>                 bool *data_ready_now)
> >>> {
> >>> -    if (vsock_stream_has_data(vsk))
> >>> +    if (vsock_stream_has_data(vsk) >= target)
> >>>         *data_ready_now = true;
> >>>     else
> >>>         *data_ready_now = false;
> >>
> >> Perhaps we can take the opportunity to clean up the code in this way:
> >>
> >>     *data_ready_now = vsock_stream_has_data(vsk) >= target;
> >Ack
> >>
> >> Anyway, I think we also need to fix the other transports (vmci and hyperv),
> >> what do you think?
> >For vmci it is look clear to fix it. For hyperv i need to check it more, because it
> > already uses some internal target value.
> 
> Yep, I see. Maybe you can pass `target` to hvs_channel_readable() and
> use it as parameter of HVS_PKT_LEN().
> 
> @Dexuan what do you think?
> 
> Thanks,
> Stefano

Can we return "not supported" to set_rcvlowat for Hyper-V vsock? :-)

For Hyper-V vsock, it's easy to tell if there is at least 1 byte to read: 
please refer to hvs_channel_readable(), but it's difficult to figure out
exactly how many bytes can be read. 

In hvs_channel_readable(), hv_get_bytes_to_read() returns the total 
bytes of 0, 1 or multiple Hyper-V vsock packets: each packet has a
24-byte header (see HVS_HEADER_LEN), the payload, some padding
bytes (if the payload length is not a multiple of 8), and 8 trailing
useless bytes.

It's hard to get the total payload length because there is no API in
include/linux/hyperv.h, drivers/hv/channel.c and 
drivers/hv/ring_buffer.c that allows us to peek at the data in the
VMBus channel's ringbuffer. 

We could add such a "peek" API in drivers/hv/channel.c (see the
non-peek version of the APIs in hvs_stream_dequeue(): 
hv_pkt_iter_first() and hv_pkt_iter_next()), and examine the whole
ringbuffe to figure out the exact total payload length, but I feel it may 
not be worth the non-trivial complexity just to be POSIX-compliant --
nobody ever complained about this for the past 5 years :-) So I'm
wondering if we should allow Hyper-V vsock to not support the 
set_rcvlowat op?

Thanks,
-- Dexuan