Close closes the events channel, resulting in a panic from send on
closed channel.
Reported-By: Brad Fitzpatrick <brad@tailscale.com>
Signed-off-by: James Tucker <james@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Queue{In,Out}boundElement locking can contribute to significant
overhead via sync.Mutex.lockSlow() in some environments. These types
are passed throughout the device package as elements in a slice, so
move the per-element Mutex to a container around the slice.
Reviewed-by: Maisem Ali <maisem@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.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>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
After reducing UDP stack traversal overhead via GSO and GRO,
runtime.chanrecv() began to account for a high percentage (20% in one
environment) of perf samples during a throughput benchmark. The
individual packet channel ops with the crypto goroutines was the primary
contributor to this overhead.
Updating these channels to pass vectors, which the device package
already handles at its ends, reduced this overhead substantially, and
improved throughput.
The iperf3 results below demonstrate the effect of this commit between
two Linux computers with i5-12400 CPUs. There is roughly ~13us of round
trip latency between them.
The first result is with UDP GSO and GRO, and with single element
channels.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 12.3 GBytes 10.6 Gbits/sec 232 3.15 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 12.3 GBytes 10.6 Gbits/sec 232 sender
[ 5] 0.00-10.04 sec 12.3 GBytes 10.6 Gbits/sec receiver
The second result is with channels updated to pass a slice of
elements.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 13.2 GBytes 11.3 Gbits/sec 182 3.15 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 13.2 GBytes 11.3 Gbits/sec 182 sender
[ 5] 0.00-10.04 sec 13.2 GBytes 11.3 Gbits/sec receiver
Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
StdNetBind probes for UDP GSO and GRO support at runtime. UDP GSO is
dependent on checksum offload support on the egress netdev. UDP GSO
will be disabled in the event sendmmsg() returns EIO, which is a strong
signal that the egress netdev does not support checksum offload.
The iperf3 results below demonstrate the effect of this commit between
two Linux computers with i5-12400 CPUs. There is roughly ~13us of round
trip latency between them.
The first result is from commit 052af4a without UDP GSO or GRO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 9.85 GBytes 8.46 Gbits/sec 1139 3.01 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 9.85 GBytes 8.46 Gbits/sec 1139 sender
[ 5] 0.00-10.04 sec 9.85 GBytes 8.42 Gbits/sec receiver
The second result is with UDP GSO and GRO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 12.3 GBytes 10.6 Gbits/sec 232 3.15 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 12.3 GBytes 10.6 Gbits/sec 232 sender
[ 5] 0.00-10.04 sec 12.3 GBytes 10.6 Gbits/sec receiver
Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Replace the src storage inside StdNetEndpoint with a copy of the raw
control message buffer, to reduce allocation and perform less work on a
per-packet basis.
Signed-off-by: James Tucker <james@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
If an IPC operation is in flight while close starts, it is possible for
both processes to deadlock. Prevent this by taking the IPC lock at the
start of close and for the duration.
Signed-off-by: James Tucker <jftucker@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.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>
Looks like a simple copy&paste error.
Fixes: 9e2f386 ("conn, device, tun: implement vectorized I/O on Linux")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
In 9e2f386 ("conn, device, tun: implement vectorized I/O on Linux"), the
Linux-specific Bind implementation was collapsed into StdNetBind. This
introduced a race on StdNetEndpoint from getSrcFromControl() and
setSrcControl().
Remove the sync.Pool involved in the race, and simplify StdNetBind's
receive path to allocate StdNetEndpoint on the heap instead, with the
intent for it to be cleaned up by the GC, later. This essentially
reverts ef5c587 ("conn: remove the final alloc per packet receive"),
adding back that allocation, unfortunately.
This does slightly increase resident memory usage in higher throughput
scenarios. StdNetBind is the only Bind implementation that was using
this Endpoint recycling technique prior to this commit.
This is considered a stop-gap solution, and there are plans to replace
the allocation with a better mechanism.
Reported-by: lsc <lsc@lv6.tw>
Link: https://lore.kernel.org/wireguard/ac87f86f-6837-4e0e-ec34-1df35f52540e@lv6.tw/
Fixes: 9e2f386 ("conn, device, tun: implement vectorized I/O on Linux")
Cc: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: James Tucker <james@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
We can't have the netlink listener socket, so it's not possible to
support it. Plus, android networking stack complexity makes it a bit
tricky anyway, so best to leave it disabled.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
We only expect a single control message in the normal case, but this
would loop infinitely if there were more.
Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
CmsgLen() does not account for data alignment.
Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Cmsghdr uses uint32 and uint64 on 32-bit and 64-bit respectively for the
Len member, which makes assignments and comparisons slightly more
irksome than usual.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
This commit updates StdNetBind.BatchSize() to return 1 instead of
IdealBatchSize for non-Linux platforms. Non-Linux platforms do not
yet benefit from values > 1, which only serves to increase memory
consumption.
Reviewed-by: James Tucker <james@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Enable TCP SACK for the gVisor Stack used in tun/netstack. This can
improve throughput by an order of magnitude in the presence of packet
loss.
Reviewed-by: James Tucker <james@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
This commit re-slices received control messages in StdNetBind to the
value the OS reports on a successful read. Previously, the len of this
slice would always be srcControlSize, which could result in control
message values leaking through a sync.Pool round trip. This is
unlikely with the IP_PKTINFO socket option set successfully, but
should be guarded against.
Reviewed-by: James Tucker <james@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
If RIO is unavailable, NewWinRingBind() falls back to StdNetBind.
StdNetBind uses x/net/ipv{4,6}.PacketConn for sending and receiving
datagrams, specifically via the {Read,Write}Batch methods.
These methods are unimplemented on Windows and will return runtime
errors as a result. Additionally, only Linux benefits from these
x/net types for reading and writing, so we update StdNetBind to fall
back to the standard library net package for all platforms other than
Linux.
Reviewed-by: James Tucker <james@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>
The conn.Bind UDP sockets' send and receive buffers are now being sized
to 7MB, whereas they were previously inheriting the system defaults.
The system defaults are considerably small and can result in dropped
packets on high speed links. By increasing the size of these buffers we
are able to achieve higher throughput in the aforementioned case.
The iperf3 results below demonstrate the effect of this commit between
two Linux computers with 32-core Xeon Platinum CPUs @ 2.9Ghz. There is
roughly ~125us of round trip latency between them.
The first result is from commit 792b49c which uses the system defaults,
e.g. net.core.{r,w}mem_max = 212992. The TCP retransmits are correlated
with buffer full drops on both sides.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 4.74 GBytes 4.08 Gbits/sec 2742 285 KBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 4.74 GBytes 4.08 Gbits/sec 2742 sender
[ 5] 0.00-10.04 sec 4.74 GBytes 4.06 Gbits/sec receiver
The second result is after increasing SO_{SND,RCV}BUF to 7MB, i.e.
applying this commit.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval Transfer Bitrate Retr Cwnd
[ 5] 0.00-10.00 sec 6.14 GBytes 5.27 Gbits/sec 0 3.15 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 6.14 GBytes 5.27 Gbits/sec 0 sender
[ 5] 0.00-10.04 sec 6.14 GBytes 5.25 Gbits/sec receiver
The specific value of 7MB is chosen as it is the max supported by a
default configuration of macOS. A value greater than 7MB may further
benefit throughput for environments with higher network latency and
lower CPU clocks, but will also increase latency under load
(bufferbloat). Some platforms will silently clamp the value to other
maximums. On Linux, we use SO_{SND,RCV}BUFFORCE in case 7MB is beyond
net.core.{r,w}mem_max.
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>
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>
Accept packet vectors for reading and writing in the tun.Device and
conn.Bind interfaces, so that the internal plumbing between these
interfaces now passes a vector of packets. Vectors move untouched
between these interfaces, i.e. if 128 packets are received from
conn.Bind.Read(), 128 packets are passed to tun.Device.Write(). There is
no internal buffering.
Currently, existing implementations are only adjusted to have vectors
of length one. Subsequent patches will improve that.
Also, as a related fixup, use the unix and windows packages rather than
the syscall package when possible.
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>
For some reason, this was omitted for response messages.
Reported-by: z <dzm@unexpl0.red>
Fixes: 8c34c4c ("First set of code review patches")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
This seems like a much better demonstration as it removes the need for
external components.
Signed-off-by: Søren L. Hansen <sorenisanerd@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
To build with go1.19, gvisor needs
99325baf ("Bump gVisor build tags to go1.19").
However gvisor.dev/gvisor/pkg/tcpip/buffer is no longer available,
so refactor to use gvisor.dev/gvisor/pkg/tcpip/link/channel directly.
Signed-off-by: Shengjing Zhu <i@zhsj.me>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Use unix.ByteSliceToString in (*NativeTun).nameSlice to convert the
TUNGETIFF ioctl result []byte to a string.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
This does bind_std only; other platforms remain.
The remaining alloc per iteration in the Throughput benchmark
comes from the tuntest package, and should not appear in regular use.
name old time/op new time/op delta
Latency-10 25.2µs ± 1% 25.0µs ± 0% -0.58% (p=0.006 n=10+10)
Throughput-10 2.44µs ± 3% 2.41µs ± 2% ~ (p=0.140 n=10+8)
name old alloc/op new alloc/op delta
Latency-10 854B ± 5% 741B ± 3% -13.22% (p=0.000 n=10+10)
Throughput-10 265B ±34% 267B ±39% ~ (p=0.670 n=10+10)
name old allocs/op new allocs/op delta
Latency-10 16.0 ± 0% 14.0 ± 0% -12.50% (p=0.000 n=10+10)
Throughput-10 2.00 ± 0% 1.00 ± 0% -50.00% (p=0.000 n=10+10)
name old packet-loss new packet-loss delta
Throughput-10 0.01 ±82% 0.01 ±282% ~ (p=0.321 n=9+8)
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
