Hello,
during my experiments with zerocopy receive, i found, that in some
cases, poll() implementation violates POSIX: when socket has non-
default SO_RCVLOWAT(e.g. not 1), poll() will always set POLLIN and
POLLRDNORM bits in 'revents' even number of bytes available to read
on socket is smaller than SO_RCVLOWAT value. In this case,user sees
POLLIN flag and then tries to read data(for example using 'read()'
call), but read call will be blocked, because SO_RCVLOWAT logic is
supported in dequeue loop in af_vsock.c. But the same time, POSIX
requires that:
"POLLIN Data other than high-priority data may be read without
blocking.
POLLRDNORM Normal data may be read without blocking."
See https://www.open-std.org/jtc1/sc22/open/n4217.pdf, page 293.
So, we have, that poll() syscall returns POLLIN, but read call will
be blocked.
Also in man page socket(7) i found that:
"Since Linux 2.6.28, select(2), poll(2), and epoll(7) indicate a
socket as readable only if at least SO_RCVLOWAT bytes are available."
I checked TCP callback for poll()(net/ipv4/tcp.c, tcp_poll()), it
uses SO_RCVLOWAT value to set POLLIN bit, also i've tested TCP with
this case for TCP socket, it works as POSIX required.
I've added some fixes to af_vsock.c and virtio_transport_common.c,
test is also implemented.
What do You think guys?
Thank You
Arseniy Krasnov(3):
vsock_test: POLLIN + SO_RCVLOWAT test.
virtio/vsock: use 'target' in notify_poll_in callback.
vsock: use sk_skrcvlowat to set POLLIN,POLLRDNORM bits.
net/vmw_vsock/af_vsock.c | 2 +-
net/vmw_vsock/virtio_transport_common.c | 2 +-
tools/testing/vsock/vsock_test.c | 90 +++++++++++++++++++++++++++++++++
3 files changed, 92 insertions(+), 2 deletions(-)
--
2.25.1
