| Message ID | 20170717023957.7783-1-sysugaozhenyu@gmail.com |
|---|---|
| State | Superseded |
| Headers | show |
Ping.... Thanks Zhenyu Gao 2017-07-17 10:39 GMT+08:00 Zhenyu Gao <sysugaozhenyu@gmail.com>: > Sendmmsg can reduce cpu cycles in sending packets to kernel. > Replace sendmsg with sendmmsg in function netdev_linux_send to send > batch packets if sendmmsg is available. > > If kernel side doesn't support sendmmsg, will fallback to sendmsg. > > netserver > |------------| > | | > | container | > |----veth----| > | > | |------------| > |---veth-| dpdk-ovs | netperf > | | |--------------| > |----dpdk----| | bare-metal | > | |--------------| > | | > | | > pnic-----------pnic > > Netperf was consumed to test the performance: > > 1)cmd:netperf -H remote-container -t UDP_STREAM -l 60 -- -m 1400 > result: netserver received 2383.21Mb(sendmsg)/2551.64Mb(sendmmsg) > > 2)cmd:netperf -H remote-container -t UDP_STREAM -l 60 -- -m 60 > result: netserver received 109.72Mb(sendmsg)/115.18Mb(sendmmsg) > > Sendmmsg show about 6% improvement in netperf UDP testing. > > Signed-off-by: Zhenyu Gao <sysugaozhenyu@gmail.com> > --- > lib/netdev-linux.c | 85 ++++++++++++++++++++++++++++++ > +++++++++--------------- > 1 file changed, 62 insertions(+), 23 deletions(-) > > diff --git a/lib/netdev-linux.c b/lib/netdev-linux.c > index e1d8701..d991d05 100644 > --- a/lib/netdev-linux.c > +++ b/lib/netdev-linux.c > @@ -1182,6 +1182,54 @@ netdev_linux_rxq_drain(struct netdev_rxq *rxq_) > } > } > > +static inline int > +netdev_linux_sock_batch_send(int sock, struct msghdr *msg, > + struct dp_packet_batch *batch) > +{ > + int error = 0; > + ssize_t retval; > + uint32_t resend_idx = 0; > + struct mmsghdr *mmsg; > + struct iovec *iov; > + > + mmsg = xmalloc(sizeof(*mmsg) * batch->count); > + iov = xmalloc(sizeof(*iov) * batch->count); > + > + for (int i = 0; i < batch->count; i++) { > + const void *data = dp_packet_data(batch->packets[i]); > + size_t size = dp_packet_size(batch->packets[i]); > + > + /* Truncate the packet if it is configured. */ > + size -= dp_packet_get_cutlen(batch->packets[i]); > + > + iov[i].iov_base = CONST_CAST(void *, data); > + iov[i].iov_len = size; > + mmsg[i].msg_hdr = *msg; > + mmsg[i].msg_hdr.msg_iov = &iov[i]; > + } > + > +resend_batch: > + retval = sendmmsg(sock, mmsg + resend_idx, > + batch->count - resend_idx, 0); > + if (retval < 0) { > + if (errno == EINTR) { > + goto resend_batch; > + } > + /* The Linux AF_PACKET implementation never blocks waiting for > + * room for packets, instead returning ENOBUFS. Translate this > + * into EAGAIN for the caller. */ > + error = errno == ENOBUFS ? EAGAIN : errno; > + } else if (retval != batch->count - resend_idx) { > + /* Send remain packets again. */ > + resend_idx += retval; > + goto resend_batch; > + } > + > + free(mmsg); > + free(iov); > + return error; > +} > + > /* Sends 'buffer' on 'netdev'. Returns 0 if successful, otherwise a > positive > * errno value. Returns EAGAIN without blocking if the packet cannot be > queued > * immediately. Returns EMSGSIZE if a partial packet was transmitted or > if > @@ -1226,6 +1274,9 @@ netdev_linux_send(struct netdev *netdev_, int qid > OVS_UNUSED, > msg.msg_control = NULL; > msg.msg_controllen = 0; > msg.msg_flags = 0; > + > + error = netdev_linux_sock_batch_send(sock, &msg, batch); > + goto check_error; > } > > /* 'i' is incremented only if there's no error */ > @@ -1236,34 +1287,21 @@ netdev_linux_send(struct netdev *netdev_, int qid > OVS_UNUSED, > > /* Truncate the packet if it is configured. */ > size -= dp_packet_get_cutlen(batch->packets[i]); > + /* Use the tap fd to send to this device. This is essential for > + * tap devices, because packets sent to a tap device with an > + * AF_PACKET socket will loop back to be *received* again on the > + * tap device. This doesn't occur on other interface types > + * because we attach a socket filter to the rx socket. */ > + struct netdev_linux *netdev = netdev_linux_cast(netdev_); > > - if (!is_tap_netdev(netdev_)) { > - /* Use our AF_PACKET socket to send to this device. */ > - struct iovec iov; > - > - iov.iov_base = CONST_CAST(void *, data); > - iov.iov_len = size; > - > - msg.msg_iov = &iov; > - > - retval = sendmsg(sock, &msg, 0); > - } else { > - /* Use the tap fd to send to this device. This is essential > for > - * tap devices, because packets sent to a tap device with an > - * AF_PACKET socket will loop back to be *received* again on > the > - * tap device. This doesn't occur on other interface types > - * because we attach a socket filter to the rx socket. */ > - struct netdev_linux *netdev = netdev_linux_cast(netdev_); > - > - retval = write(netdev->tap_fd, data, size); > - } > + retval = write(netdev->tap_fd, data, size); > > if (retval < 0) { > if (errno == EINTR) { > /* The send was interrupted by a signal. Retry the > packet by > * continuing without incrementing 'i'.*/ > continue; > - } else if (errno == EIO && is_tap_netdev(netdev_)) { > + } else if (errno == EIO) { > /* The Linux tap driver returns EIO if the device is not > up. > * From the OVS side this is not an error, so ignore it. > */ > } else { > @@ -1285,9 +1323,10 @@ netdev_linux_send(struct netdev *netdev_, int qid > OVS_UNUSED, > i++; > } > > +check_error: > if (error && error != EAGAIN) { > - VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s", > - netdev_get_name(netdev_), ovs_strerror(error)); > + VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s", > + netdev_get_name(netdev_), ovs_strerror(error)); > } > > free_batch: > -- > 1.8.3.1 > >
On Mon, Jul 17, 2017 at 02:39:57AM +0000, Zhenyu Gao wrote: > Sendmmsg can reduce cpu cycles in sending packets to kernel. > Replace sendmsg with sendmmsg in function netdev_linux_send to send > batch packets if sendmmsg is available. Thanks for the patch. I like the idea but I found the details to be a little hard to follow, especially given the multiple goto statements. How about the following instead? https://patchwork.ozlabs.org/patch/796898/ https://patchwork.ozlabs.org/patch/796899/ Thanks, Ben.
It looks better to this independent function netdev_linux_tap_batch_send. Thanks for working on it. Thanks Zhenyu Gao 2017-08-03 5:59 GMT+08:00 Ben Pfaff <blp@ovn.org>: > On Mon, Jul 17, 2017 at 02:39:57AM +0000, Zhenyu Gao wrote: > > Sendmmsg can reduce cpu cycles in sending packets to kernel. > > Replace sendmsg with sendmmsg in function netdev_linux_send to send > > batch packets if sendmmsg is available. > > Thanks for the patch. > > I like the idea but I found the details to be a little hard to follow, > especially given the multiple goto statements. > > How about the following instead? > https://patchwork.ozlabs.org/patch/796898/ > https://patchwork.ozlabs.org/patch/796899/ > > Thanks, > > Ben. >
diff --git a/lib/netdev-linux.c b/lib/netdev-linux.c index e1d8701..d991d05 100644 --- a/lib/netdev-linux.c +++ b/lib/netdev-linux.c @@ -1182,6 +1182,54 @@ netdev_linux_rxq_drain(struct netdev_rxq *rxq_) } } +static inline int +netdev_linux_sock_batch_send(int sock, struct msghdr *msg, + struct dp_packet_batch *batch) +{ + int error = 0; + ssize_t retval; + uint32_t resend_idx = 0; + struct mmsghdr *mmsg; + struct iovec *iov; + + mmsg = xmalloc(sizeof(*mmsg) * batch->count); + iov = xmalloc(sizeof(*iov) * batch->count); + + for (int i = 0; i < batch->count; i++) { + const void *data = dp_packet_data(batch->packets[i]); + size_t size = dp_packet_size(batch->packets[i]); + + /* Truncate the packet if it is configured. */ + size -= dp_packet_get_cutlen(batch->packets[i]); + + iov[i].iov_base = CONST_CAST(void *, data); + iov[i].iov_len = size; + mmsg[i].msg_hdr = *msg; + mmsg[i].msg_hdr.msg_iov = &iov[i]; + } + +resend_batch: + retval = sendmmsg(sock, mmsg + resend_idx, + batch->count - resend_idx, 0); + if (retval < 0) { + if (errno == EINTR) { + goto resend_batch; + } + /* The Linux AF_PACKET implementation never blocks waiting for + * room for packets, instead returning ENOBUFS. Translate this + * into EAGAIN for the caller. */ + error = errno == ENOBUFS ? EAGAIN : errno; + } else if (retval != batch->count - resend_idx) { + /* Send remain packets again. */ + resend_idx += retval; + goto resend_batch; + } + + free(mmsg); + free(iov); + return error; +} + /* Sends 'buffer' on 'netdev'. Returns 0 if successful, otherwise a positive * errno value. Returns EAGAIN without blocking if the packet cannot be queued * immediately. Returns EMSGSIZE if a partial packet was transmitted or if @@ -1226,6 +1274,9 @@ netdev_linux_send(struct netdev *netdev_, int qid OVS_UNUSED, msg.msg_control = NULL; msg.msg_controllen = 0; msg.msg_flags = 0; + + error = netdev_linux_sock_batch_send(sock, &msg, batch); + goto check_error; } /* 'i' is incremented only if there's no error */ @@ -1236,34 +1287,21 @@ netdev_linux_send(struct netdev *netdev_, int qid OVS_UNUSED, /* Truncate the packet if it is configured. */ size -= dp_packet_get_cutlen(batch->packets[i]); + /* Use the tap fd to send to this device. This is essential for + * tap devices, because packets sent to a tap device with an + * AF_PACKET socket will loop back to be *received* again on the + * tap device. This doesn't occur on other interface types + * because we attach a socket filter to the rx socket. */ + struct netdev_linux *netdev = netdev_linux_cast(netdev_); - if (!is_tap_netdev(netdev_)) { - /* Use our AF_PACKET socket to send to this device. */ - struct iovec iov; - - iov.iov_base = CONST_CAST(void *, data); - iov.iov_len = size; - - msg.msg_iov = &iov; - - retval = sendmsg(sock, &msg, 0); - } else { - /* Use the tap fd to send to this device. This is essential for - * tap devices, because packets sent to a tap device with an - * AF_PACKET socket will loop back to be *received* again on the - * tap device. This doesn't occur on other interface types - * because we attach a socket filter to the rx socket. */ - struct netdev_linux *netdev = netdev_linux_cast(netdev_); - - retval = write(netdev->tap_fd, data, size); - } + retval = write(netdev->tap_fd, data, size); if (retval < 0) { if (errno == EINTR) { /* The send was interrupted by a signal. Retry the packet by * continuing without incrementing 'i'.*/ continue; - } else if (errno == EIO && is_tap_netdev(netdev_)) { + } else if (errno == EIO) { /* The Linux tap driver returns EIO if the device is not up. * From the OVS side this is not an error, so ignore it. */ } else { @@ -1285,9 +1323,10 @@ netdev_linux_send(struct netdev *netdev_, int qid OVS_UNUSED, i++; } +check_error: if (error && error != EAGAIN) { - VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s", - netdev_get_name(netdev_), ovs_strerror(error)); + VLOG_WARN_RL(&rl, "error sending Ethernet packet on %s: %s", + netdev_get_name(netdev_), ovs_strerror(error)); } free_batch:
Sendmmsg can reduce cpu cycles in sending packets to kernel. Replace sendmsg with sendmmsg in function netdev_linux_send to send batch packets if sendmmsg is available. If kernel side doesn't support sendmmsg, will fallback to sendmsg. netserver |------------| | | | container | |----veth----| | | |------------| |---veth-| dpdk-ovs | netperf | | |--------------| |----dpdk----| | bare-metal | | |--------------| | | | | pnic-----------pnic Netperf was consumed to test the performance: 1)cmd:netperf -H remote-container -t UDP_STREAM -l 60 -- -m 1400 result: netserver received 2383.21Mb(sendmsg)/2551.64Mb(sendmmsg) 2)cmd:netperf -H remote-container -t UDP_STREAM -l 60 -- -m 60 result: netserver received 109.72Mb(sendmsg)/115.18Mb(sendmmsg) Sendmmsg show about 6% improvement in netperf UDP testing. Signed-off-by: Zhenyu Gao <sysugaozhenyu@gmail.com> --- lib/netdev-linux.c | 85 +++++++++++++++++++++++++++++++++++++++--------------- 1 file changed, 62 insertions(+), 23 deletions(-)