This adds AMD64 assembly implementations of IP checksum computation, one
for baseline AMD64 and the other for v3 AMD64 (AVX2 and BMI2).
All performance numbers reported are from a Ryzen 7 4750U but similar
improvements are expected for a wide range of processors.
The generic IP checksum implementation has also been further improved to
be significantly faster using bits.AddUint64 (for a 64KiB buffer the
throughput improves from 15,000MiB/s to 27,600MiB/s; similar gains are
also reported on ARM64 but I do not have specific numbers).
The baseline AMD64 implementation for a 64KiB buffer reports 32,700MiB/s
and the AVX2 implementation is slightly over 107,000MiB/s.
Unfortunately, for very small sizes (e.g. the expected size for an IPv4
header) setting up SIMD computation involves some overhead that makes
computing a checksum for small buffers slower than a non-SIMD
implementation. Even more unfortunately, testing for this at runtimen in
Go and calling a func optimized for small buffers mitigates most of the
improvement due to call overhead. The break even point is around 256
byte buffers; IPv4 headers are no more than 60 bytes including
extensions. IPv6 headers do not have a checksum but are a fixed size of
40 bytes. As a result, the generated assembly code uses an alternate
approach for buffers of less than 256 bytes. Additionally, buffers of
less than 32 bytes need to be handled specially because the strategy for
reading buffers that are not a multiple of 8 bytes fails when the buffer
is too small.
As suggested by additional benchmarking, pseudo header computation has
been rewritten to be faster (benchmark time reduced by 1/2 to 1/4).
Updates tailscale/corp#9755
Signed-off-by: Adrian Dewhurst <adrian@tailscale.com>
IPv4 header and pseudo header checksums were being computed on every
merge operation. Additionally, virtioNetHdr was being written at the
same time. This delays those operations until after all coalescing has
occurred.
Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
tcpGRO() was using an incorrect IPv4 more fragments bit mask.
tcpPacketsCanCoalesce() was not distinguishing tcp6 from tcp4, and TTL
values were not compared. TTL values should be equal at the IP layer,
otherwise the packets should not coalesce. This tracks with the kernel.
Reviewed-by: Denton Gentry <dgentry@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
IP options were not being compared prior to coalescing. They are not
commonly used. Disqualification due to nonzero options is in line with
the kernel.
Reviewed-by: Denton Gentry <dgentry@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
There's not really a use at the moment for making this configurable, and
once bind_windows.go behaves like bind_std.go, we'll be able to use
constants everywhere. So begin that simplification now.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Implement TCP offloading via TSO and GRO for the Linux tun.Device, which
is made possible by virtio extensions in the kernel's TUN driver.
Delete conn.LinuxSocketEndpoint in favor of a collapsed conn.StdNetBind.
conn.StdNetBind makes use of recvmmsg() and sendmmsg() on Linux. All
platforms now fall under conn.StdNetBind, except for Windows, which
remains in conn.WinRingBind, which still needs to be adjusted to handle
multiple packets.
Also refactor sticky sockets support to eventually be applicable on
platforms other than just Linux. However Linux remains the sole platform
that fully implements it for now.
Co-authored-by: James Tucker <james@tailscale.com>
Signed-off-by: James Tucker <james@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
