This series of patches adds vhost-user support to passt
and then allows passt to connect to QEMU network backend using
virtqueue rather than a socket.
We have with iperf3 a 10x speed improvement:
$ iperf3 -c localhost -p 10001 -t 60 -6 -u -b 50G
socket:
[ 5] 0.00-60.04 sec 30.5 GBytes 4.36 Gbits/sec 0.065 ms 9127377/10125415 (90%)
receiver
vhost-user:
[ 5] 0.00-60.05 sec 292 GBytes 41.8 Gbits/sec 0.007 ms 259805/9832736 (2.6%)
receiver
$ iperf3 -c localhost -p 10001 -t 60 -4 -u -b 50G
socket:
[ 5] 0.00-60.04 sec 36.4 GBytes 5.21 Gbits/sec 0.048 ms 7535735/8728101 (86%)
receiver
vhost-user:
[ 5] 0.00-60.05 sec 259 GBytes 37.0 Gbits/sec 0.003 ms 142594/8616705 (1.7%)
receiver
$ iperf3 -c localhost -p 10001 -t 60 -6
socket:
[ 5] 0.00-60.00 sec 16.3 GBytes 2.33 Gbits/sec 0 sender
[ 5] 0.00-60.06 sec 16.3 GBytes 2.32 Gbits/sec receiver
vhost-user:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-60.00 sec 205 GBytes 29.3 Gbits/sec 0 sender
[ 5] 0.00-60.04 sec 205 GBytes 29.3 Gbits/sec receiver
$ iperf3 -c localhost -p 10001 -t 60 -4
socket:
[ 5] 0.00-60.00 sec 16.1 GBytes 2.31 Gbits/sec 0 sender
[ 5] 0.00-60.07 sec 16.1 GBytes 2.31 Gbits/sec receiver
vhost-user:
[ 5] 0.00-60.00 sec 201 GBytes 28.7 Gbits/sec 0 sender
[ 5] 0.00-60.04 sec 201 GBytes 28.7 Gbits/sec receiver
With QEMU, rather than using to connect:
-netdev stream,id=s,server=off,addr.type=unix,addr.path=/tmp/passt_1.socket
we will use:
-chardev socket,id=chr0,path=/tmp/passt_1.socket
-netdev vhost-user,id=netdev0,chardev=chr0
-device virtio-net,netdev=netdev0
-object memory-backend-memfd,id=memfd0,share=on,size=$RAMSIZE
-numa node,memdev=memfd0
The memory backend is needed to share data between passt and QEMU.
The series starts to introduce new functions to manage iovec and
to do checksum on unaligned memory (we cannot align buffers provided
by the guest with the same value we use with internal passt buffers):
iov: add some functions to manage iovec
pcap: add pcap_iov()
checksum: align buffers
checksum: add csum_iov()
We introduce new files ip.c and ip.h to provide IP generic functions:
util: move IP stuff from util.[ch] to ip.[ch]
ip: move duplicate IPv4 checksum function to ip.h
Then we extract from existing TCP and UDP function the internal
passt buffer management to be able to use them with the guest
provided buffers:
tcp: extract buffer management from tcp_send_flag()
tcp: extract buffer management from tcp_conn_tap_mss()
tcp: rename functions that manage buffers
tcp: move buffers management functions to their own file
tap: make tap_update_mac() generic
tap: export pool_flush()/tapX_handler()/packet_add()
udp: move udpX_l2_buf_t and udpX_l2_mh_sock out of udp_update_hdrX()
udp: rename udp_sock_handler() to udp_buf_sock_handler()
packet: replace struct desc by struct iovec
As vhost-user is a variant of passt mode, modify the code to compare
to (!MODE_PASTA) rather than (MODE_PASST || MODE_VU)
vhost-user: compare mode MODE_PASTA and not MODE_PASST
We introduce virtio and vhost-user management functions:
vhost-user: introduce virtio API
vhost-user: introduce vhost-user API
And then a first version of vhost-user that copies data from the passt buffers
to guest memory, and vice-versa, as it's done with the socket algorithm:
vhost-user: add vhost-user
And finaly remove the buffers copy on TX and RX (TCP/UDP):
vhost-user: use guest buffer directly in vu_handle_tx()
tcp: vhost-user RX nocopy
udp: vhost-user RX nocopy
vhost-user: remove tap_send_frames_vu()
Thanks,
Laurent
Laurent Vivier (24):
iov: add some functions to manage iovec
pcap: add pcap_iov()
checksum: align buffers
checksum: add csum_iov()
util: move IP stuff from util.[ch] to ip.[ch]
ip: move duplicate IPv4 checksum function to ip.h
ip: introduce functions to compute the header part checksum for
TCP/UDP
tcp: extract buffer management from tcp_send_flag()
tcp: extract buffer management from tcp_conn_tap_mss()
tcp: rename functions that manage buffers
tcp: move buffers management functions to their own file
tap: make tap_update_mac() generic
tap: export pool_flush()/tapX_handler()/packet_add()
udp: move udpX_l2_buf_t and udpX_l2_mh_sock out of udp_update_hdrX()
udp: rename udp_sock_handler() to udp_buf_sock_handler()
packet: replace struct desc by struct iovec
vhost-user: compare mode MODE_PASTA and not MODE_PASST
vhost-user: introduce virtio API
vhost-user: introduce vhost-user API
vhost-user: add vhost-user
vhost-user: use guest buffer directly in vu_handle_tx()
tcp: vhost-user RX nocopy
udp: vhost-user RX nocopy
vhost-user: remove tap_send_frames_vu()
Makefile | 7 +-
checksum.c | 51 ++-
checksum.h | 1 +
conf.c | 33 +-
dhcp.c | 1 +
flow.c | 1 +
icmp.c | 1 +
iov.c | 78 ++++
iov.h | 46 +++
ip.c | 72 ++++
ip.h | 124 ++++++
isolation.c | 10 +-
ndp.c | 1 +
packet.c | 81 ++--
packet.h | 16 +-
passt.c | 18 +-
passt.h | 10 +
pcap.c | 32 ++
pcap.h | 1 +
port_fwd.c | 1 +
qrap.c | 1 +
tap.c | 226 +++++++----
tap.h | 13 +-
tcp.c | 789 ++++++------------------------------
tcp.h | 2 +-
tcp_buf.c | 569 ++++++++++++++++++++++++++
tcp_buf.h | 17 +
tcp_internal.h | 78 ++++
tcp_splice.c | 1 +
tcp_vu.c | 447 +++++++++++++++++++++
tcp_vu.h | 10 +
udp.c | 171 ++++----
udp.h | 4 +-
udp_internal.h | 21 +
udp_vu.c | 215 ++++++++++
udp_vu.h | 8 +
util.c | 55 ---
util.h | 83 +---
vhost_user.c | 1050 ++++++++++++++++++++++++++++++++++++++++++++++++
vhost_user.h | 137 +++++++
virtio.c | 484 ++++++++++++++++++++++
virtio.h | 121 ++++++
42 files changed, 4041 insertions(+), 1046 deletions(-)
create mode 100644 iov.c
create mode 100644 iov.h
create mode 100644 ip.c
create mode 100644 ip.h
create mode 100644 tcp_buf.c
create mode 100644 tcp_buf.h
create mode 100644 tcp_internal.h
create mode 100644 tcp_vu.c
create mode 100644 tcp_vu.h
create mode 100644 udp_internal.h
create mode 100644 udp_vu.c
create mode 100644 udp_vu.h
create mode 100644 vhost_user.c
create mode 100644 vhost_user.h
create mode 100644 virtio.c
create mode 100644 virtio.h
--
2.42.0