Merge branch 'gue-next'
Tom Herbert says:
====================
gue: Remote checksum offload
This patch set implements remote checksum offload for
GUE, which is a mechanism that provides checksum offload of
encapsulated packets using rudimentary offload capabilities found in
most Network Interface Card (NIC) devices. The outer header checksum
for UDP is enabled in packets and, with some additional meta
information in the GUE header, a receiver is able to deduce the
checksum to be set for an inner encapsulated packet. Effectively this
offloads the computation of the inner checksum. Enabling the outer
checksum in encapsulation has the additional advantage that it covers
more of the packet than the inner checksum including the encapsulation
headers.
Remote checksum offload is described in:
http://tools.ietf.org/html/draft-herbert-remotecsumoffload-01
The GUE transmit and receive paths are modified to support the
remote checksum offload option. The option contains a checksum
offset and checksum start which are directly derived from values
set in stack when doing CHECKSUM_PARTIAL. On receipt of the option, the
operation is to calculate the packet checksum from "start" to end of
the packet (normally derived for checksum complete), and then set
the resultant value at checksum "offset" (the checksum field has
already been primed with the pseudo header). This emulates a NIC
that implements NETIF_F_HW_CSUM.
The primary purpose of this feature is to eliminate cost of performing
checksum calculation over a packet when encpasulating.
In this patch set:
- Move fou_build_header into fou.c and split it into a couple of
functions
- Enable offloading of outer UDP checksum in encapsulation
- Change udp_offload to support remote checksum offload, includes
new GSO type and ensuring encapsulated layers (TCP) doesn't try to
set a checksum covered by RCO
- TX support for RCO with GUE. This is configured through ip_tunnel
and set the option on transmit when packet being encapsulated is
CHECKSUM_PARTIAL
- RX support for RCO with GUE for normal and GRO paths. Includes
resolving the offloaded checksum
v2:
Address comments from davem: Move accounting for private option
field in gue_encap_hlen to patch in which we add the remote checksum
offload option.
Testing:
I ran performance numbers using netperf TCP_STREAM and TCP_RR with 200
streams, comparing GUE with and without remote checksum offload (doing
checksum-unnecessary to complete conversion in both cases). These
were run on mlnx4 and bnx2x. Some mlnx4 results are below.
GRE/GUE
TCP_STREAM
IPv4, with remote checksum offload
9.71% TX CPU utilization
7.42% RX CPU utilization
36380 Mbps
IPv4, without remote checksum offload
12.40% TX CPU utilization
7.36% RX CPU utilization
36591 Mbps
TCP_RR
IPv4, with remote checksum offload
77.79% CPU utilization
91/144/216 90/95/99% latencies
1.95127e+06 tps
IPv4, without remote checksum offload
78.70% CPU utilization
89/152/297 90/95/99% latencies
1.95458e+06 tps
IPIP/GUE
TCP_STREAM
With remote checksum offload
10.30% TX CPU utilization
7.43% RX CPU utilization
36486 Mbps
Without remote checksum offload
12.47% TX CPU utilization
7.49% RX CPU utilization
36694 Mbps
TCP_RR
With remote checksum offload
77.80% CPU utilization
87/153/270 90/95/99% latencies
1.98735e+06 tps
Without remote checksum offload
77.98% CPU utilization
87/150/287 90/95/99% latencies
1.98737e+06 tps
SIT/GUE
TCP_STREAM
With remote checksum offload
9.68% TX CPU utilization
7.36% RX CPU utilization
35971 Mbps
Without remote checksum offload
12.95% TX CPU utilization
8.04% RX CPU utilization
36177 Mbps
TCP_RR
With remote checksum offload
79.32% CPU utilization
94/158/295 90/95/99% latencies
1.88842e+06 tps
Without remote checksum offload
80.23% CPU utilization
94/149/226 90/95/99% latencies
1.90338e+06 tps
VXLAN
TCP_STREAM
35.03% TX CPU utilization
20.85% RX CPU utilization
36230 Mbps
TCP_RR
77.36% CPU utilization
84/146/270 90/95/99% latencies
2.08063e+06 tps
We can also look at CPU time in csum_partial using perf (with bnx2x
setup). For GRE with TCP_STREAM I see:
With remote checksum offload
0.33% TX
1.81% RX
Without remote checksum offload
6.00% TX
0.51% RX
I suspect the fact that time in csum_partial noticably increases
with remote checksum offload for RX is due to taking the cache miss on
the encapsulated header in that function. By similar reasoning, if on
the TX side the packet were not in cache (say we did a splice from a
file whose data was never touched by the CPU) the CPU savings for TX
would probably be more pronounced.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>