| Message ID | 20201216225648.48037-1-v@nametag.social |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [net-next,v5] udp:allow UDP cmsghdrs through io_uring | expand |
On 12/16/20 3:56 PM, Victor Stewart wrote: > This patch adds PROTO_CMSG_DATA_ONLY to inet_dgram_ops and inet6_dgram_ops so that UDP_SEGMENT (GSO) and UDP_GRO can be used through io_uring. > > GSO and GRO are vital to bring QUIC servers on par with TCP throughputs, and together offer a higher > throughput gain than io_uring alone (rate of data transit > considering), thus io_uring is presently the lesser performance choice. > > RE http://vger.kernel.org/lpc_net2018_talks/willemdebruijn-lpc2018-udpgso-paper-DRAFT-1.pdf, > GSO is about +~63% and GRO +~82%. > > this patch closes that loophole. LGTM Acked-by: Jens Axboe <axboe@kernel.dk> -- Jens Axboe
might this still make it into 5.11? On Thu, Dec 17, 2020 at 3:49 PM Jens Axboe <axboe@kernel.dk> wrote: > > On 12/16/20 3:56 PM, Victor Stewart wrote: > > This patch adds PROTO_CMSG_DATA_ONLY to inet_dgram_ops and inet6_dgram_ops so that UDP_SEGMENT (GSO) and UDP_GRO can be used through io_uring. > > > > GSO and GRO are vital to bring QUIC servers on par with TCP throughputs, and together offer a higher > > throughput gain than io_uring alone (rate of data transit > > considering), thus io_uring is presently the lesser performance choice. > > > > RE http://vger.kernel.org/lpc_net2018_talks/willemdebruijn-lpc2018-udpgso-paper-DRAFT-1.pdf, > > GSO is about +~63% and GRO +~82%. > > > > this patch closes that loophole. > > LGTM > > Acked-by: Jens Axboe <axboe@kernel.dk> > > -- > Jens Axboe >
On 12/17/20 11:30 AM, Victor Stewart wrote:
> might this still make it into 5.11?
Doesn't meet the criteria to go in at this point. I sometimes
make exceptions, but generally speaking, something going into
5.11 should have been completed at least a week ago.
So I'd feel more comfortable pushing this to 5.12.
--
Jens Axboe
On Thu, Dec 17, 2020 at 1:45 PM Jens Axboe <axboe@kernel.dk> wrote: > > On 12/17/20 11:30 AM, Victor Stewart wrote: > > might this still make it into 5.11? > > Doesn't meet the criteria to go in at this point. I sometimes > make exceptions, but generally speaking, something going into > 5.11 should have been completed at least a week ago. > > So I'd feel more comfortable pushing this to 5.12. ping. are we still looking at 5.12 for this? > > -- > Jens Axboe >
diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c index b7260c8cef2e..c9fd5e7cfd6e 100644 --- a/net/ipv4/af_inet.c +++ b/net/ipv4/af_inet.c @@ -1052,6 +1052,7 @@ EXPORT_SYMBOL(inet_stream_ops); const struct proto_ops inet_dgram_ops = { .family = PF_INET, + .flags = PROTO_CMSG_DATA_ONLY, .owner = THIS_MODULE, .release = inet_release, .bind = inet_bind, diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c index e648fbebb167..560f45009d06 100644 --- a/net/ipv6/af_inet6.c +++ b/net/ipv6/af_inet6.c @@ -695,6 +695,7 @@ const struct proto_ops inet6_stream_ops = { const struct proto_ops inet6_dgram_ops = { .family = PF_INET6, + .flags = PROTO_CMSG_DATA_ONLY, .owner = THIS_MODULE, .release = inet6_release, .bind = inet6_bind, diff --git a/net/socket.c b/net/socket.c index 6e6cccc2104f..6995835d6355 100644 --- a/net/socket.c +++ b/net/socket.c @@ -2416,9 +2416,11 @@ static int ___sys_sendmsg(struct socket *sock, struct user_msghdr __user *msg, long __sys_sendmsg_sock(struct socket *sock, struct msghdr *msg, unsigned int flags) { - /* disallow ancillary data requests from this path */ - if (msg->msg_control || msg->msg_controllen) - return -EINVAL; + if (msg->msg_control || msg->msg_controllen) { + /* disallow ancillary data reqs unless cmsg is plain data */ + if (!(sock->ops->flags & PROTO_CMSG_DATA_ONLY)) + return -EINVAL; + } return ____sys_sendmsg(sock, msg, flags, NULL, 0); }
This patch adds PROTO_CMSG_DATA_ONLY to inet_dgram_ops and inet6_dgram_ops so that UDP_SEGMENT (GSO) and UDP_GRO can be used through io_uring. GSO and GRO are vital to bring QUIC servers on par with TCP throughputs, and together offer a higher throughput gain than io_uring alone (rate of data transit considering), thus io_uring is presently the lesser performance choice. RE http://vger.kernel.org/lpc_net2018_talks/willemdebruijn-lpc2018-udpgso-paper-DRAFT-1.pdf, GSO is about +~63% and GRO +~82%. this patch closes that loophole. Signed-off-by: Victor Stewart <v@nametag.social> --- net/ipv4/af_inet.c | 1 + net/ipv6/af_inet6.c | 1 + net/socket.c | 8 +++++--- 3 files changed, 7 insertions(+), 3 deletions(-)