Disable TCP & UDP GRO if the probe fails.
torvalds/linux@e269d79c7d broke virtio_net
TCP & UDP GRO causing GRO writes to return EINVAL. The bug was then
resolved later in
torvalds/linux@89add40066. The offending
commit was pulled into various LTS releases.
Updates #13041
Signed-off-by: Jordan Whited <jordan@tailscale.com>
We've added more probe targets recently which has resulted in more
timeouts behind restrictive NATs in localized testing that don't
like how many flows we are creating at once. Not so much an issue
for datacenter or cloud-hosted deployments.
Updates tailscale/corp#22114
Signed-off-by: Jordan Whited <jordan@tailscale.com>
When the TS_DEBUG_NETSTACK_LOOPBACK_PORT environment variable is set,
netstack will loop back (dnat to addressFamilyLoopback:loopbackPort)
TCP & UDP flows originally destined to localServicesIP:loopbackPort.
localServicesIP is quad-100 or the IPv6 equivalent.
Updates tailscale/corp#22713
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This is the equivalent of quad-100, but for IPv6. This is technically
already contained in the Tailscale IPv6 ULA prefix, but that is only
installed when remote peers are visible via control with contained
addrs. The service addr should always be reachable.
Updates #1152
Signed-off-by: Jordan Whited <jordan@tailscale.com>
net/tstun.Wrapper.InjectInboundPacketBuffer is not GSO-aware, which can
break quad-100 TCP streams as a result. Linux is the only platform where
gVisor GSO was previously enabled.
Updates tailscale/corp#22511
Updates #13211
Signed-off-by: Jordan Whited <jordan@tailscale.com>
In df6014f1d7 we removed build tag
gating preventing importation, which tripped a NetworkExtension limit
test in corp. This was a reversal of
25f0a3fc8f which actually made the
situation worse, hence the simplification.
This commit goes back to the strategy in
25f0a3fc8f, and gets us back under the
limit in my local testing. Admittedly, we don't fully understand
the effects of importing or excluding importation of this package,
and have seen mixed results, but this commit allows us to move forward
again.
Updates tailscale/corp#22125
Signed-off-by: Jordan Whited <jordan@tailscale.com>
In 2f27319baf we disabled GRO due to a
data race around concurrent calls to tstun.Wrapper.Write(). This commit
refactors GRO to be thread-safe, and re-enables it on Linux.
This refactor now carries a GRO type across tstun and netstack APIs
with a lifetime that is scoped to a single tstun.Wrapper.Write() call.
In 25f0a3fc8f we used build tags to
prevent importation of gVisor's GRO package on iOS as at the time we
believed it was contributing to additional memory usage on that
platform. It wasn't, so this commit simplifies and removes those
build tags.
Updates tailscale/corp#22353
Updates tailscale/corp#22125
Updates #6816
Signed-off-by: Jordan Whited <jordan@tailscale.com>
In a93dc6cdb1 tryUpgradeToBatchingConn()
moved to build tag gated files, but the runtime.GOOS condition excluding
Android was removed unintentionally from batching_conn_linux.go. Add it
back.
Updates tailscale/corp#22348
Signed-off-by: Jordan Whited <jordan@tailscale.com>
getConns() is now responsible for returning both stable and unstable
conns. conn and measureFn are now passed together via connAndMeasureFn.
newConnAndMeasureFn() is responsible for constructing them.
TCP measurement timeouts are adjusted to more closely match netcheck.
Updates tailscale/corp#22114
Signed-off-by: Jordan Whited <jordan@tailscale.com>
'stun' has been removed from metric names and replaced with a protocol
label. This refactor is preparation work for HTTPS & ICMP support.
Updates tailscale/corp#22114
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit adds a batchingConn interface, and renames batchingUDPConn
to linuxBatchingConn. tryUpgradeToBatchingConn() may return a platform-
specific implementation of batchingConn. So far only a Linux
implementation of this interface exists, but this refactor is being
done in anticipation of a Windows implementation.
Updates tailscale/corp#21874
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit increases gVisor's TCP max send (4->6MiB) and receive
(4->8MiB) buffer sizes on all platforms except iOS. These values are
biased towards higher throughput on high bandwidth-delay product paths.
The iperf3 results below demonstrate the effect of this commit between
two Linux computers with i5-12400 CPUs. 100ms of RTT latency is
introduced via Linux's traffic control network emulator queue
discipline.
The first set of results are from commit f0230ce prior to TCP buffer
resizing.
gVisor write direction:
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 180 MBytes 151 Mbits/sec 0 sender
[ 5] 0.00-10.10 sec 179 MBytes 149 Mbits/sec receiver
gVisor read direction:
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.10 sec 337 MBytes 280 Mbits/sec 20 sender
[ 5] 0.00-10.00 sec 323 MBytes 271 Mbits/sec receiver
The second set of results are from this commit with increased TCP
buffer sizes.
gVisor write direction:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 297 MBytes 249 Mbits/sec 0 sender
[ 5] 0.00-10.10 sec 297 MBytes 247 Mbits/sec receiver
gVisor read direction:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.10 sec 501 MBytes 416 Mbits/sec 17 sender
[ 5] 0.00-10.00 sec 485 MBytes 407 Mbits/sec receiver
Updates #9707
Updates tailscale/corp#22119
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit implements TCP GRO for packets being written to gVisor on
Linux. Windows support will follow later. The wireguard-go dependency is
updated in order to make use of newly exported IP checksum functions.
gVisor is updated in order to make use of newly exported
stack.PacketBuffer GRO logic.
TCP throughput towards gVisor, i.e. TUN write direction, is dramatically
improved as a result of this commit. Benchmarks show substantial
improvement, sometimes as high as 2x. High bandwidth-delay product
paths remain receive window limited, bottlenecked by gVisor's default
TCP receive socket buffer size. This will be addressed in a follow-on
commit.
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 57856fc without TCP GRO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 4.77 GBytes 4.10 Gbits/sec 20 sender
[ 5] 0.00-10.00 sec 4.77 GBytes 4.10 Gbits/sec receiver
The second result is from this commit with TCP GRO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 10.6 GBytes 9.14 Gbits/sec 20 sender
[ 5] 0.00-10.00 sec 10.6 GBytes 9.14 Gbits/sec receiver
Updates #6816
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit implements TCP GSO for packets being read from gVisor on
Linux. Windows support will follow later. The wireguard-go dependency is
updated in order to make use of newly exported GSO logic from its tun
package.
A new gVisor stack.LinkEndpoint implementation has been established
(linkEndpoint) that is loosely modeled after its predecessor
(channel.Endpoint). This new implementation supports GSO of monster TCP
segments up to 64K in size, whereas channel.Endpoint only supports up to
32K. linkEndpoint will also be required for GRO, which will be
implemented in a follow-on commit.
TCP throughput from gVisor, i.e. TUN read direction, is dramatically
improved as a result of this commit. Benchmarks show substantial
improvement through a wide range of RTT and loss conditions, sometimes
as high as 5x.
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 57856fc without TCP GSO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 2.51 GBytes 2.15 Gbits/sec 154 sender
[ 5] 0.00-10.00 sec 2.49 GBytes 2.14 Gbits/sec receiver
The second result is from this commit with TCP GSO.
Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 12.6 GBytes 10.8 Gbits/sec 6 sender
[ 5] 0.00-10.00 sec 12.6 GBytes 10.8 Gbits/sec receiver
Updates #6816
Signed-off-by: Jordan Whited <jordan@tailscale.com>
stunstamp now sends data to Prometheus via remote write, and Prometheus
can serve the same data. Retaining and cleaning up old data in sqlite
leads to long probing pauses, and it's not worth investing more effort
to optimize the schema and/or concurrency model.
Updates tailscale/corp#20344
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Timeouts could already be identified as NaN values on
stunstamp_derp_stun_rtt_ns, but we can't use NaN effectively with
promql to visualize them. So, this commit adds a timeouts metric that
we can use with rate/delta/etc promql functions.
Updates tailscale/corp#20689
Signed-off-by: Jordan Whited <jordan@tailscale.com>
validate_udp_checksum was previously indeterminate (not zero) at
declaration, and IPv4 zero value UDP checksum packets were being passed
to the kernel.
Updates tailscale/corp#20689
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit introduces a userspace program for managing an experimental
eBPF XDP STUN server program. derp/xdp contains the eBPF pseudo-C along
with a Go pkg for loading it and exporting its metrics.
cmd/xdpderper is a package main user of derp/xdp.
Updates tailscale/corp#20689
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Evaluation of remote write errors was using errors.Is() where it should
have been using errors.As().
Updates tailscale/corp#20344
Signed-off-by: Jordan Whited <jordan@tailscale.com>
stunstamp timestamping includes userspace and SO_TIMESTAMPING kernel
timestamping where available. Measurements are written locally to a
sqlite DB, exposed over an HTTP API, and written to prometheus
via remote-write protocol.
Updates tailscale/corp#20344
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Certain device drivers (e.g. vxlan, geneve) do not properly handle
coalesced UDP packets later in the stack, resulting in packet loss.
Updates #11026
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit implements probing of UDP path lifetime on the tail end of
an active direct connection. Probing configuration has two parts -
Cliffs, which are various timeout cliffs of interest, and
CycleCanStartEvery, which limits how often a probing cycle can start,
per-endpoint. Initially a statically defined default configuration will
be used. The default configuration has cliffs of 10s, 30s, and 60s,
with a CycleCanStartEvery of 24h. Probing results are communicated via
clientmetric counters. Probing is off by default, and can be enabled
via control knob. Probing is purely informational and does not yet
drive any magicsock behaviors.
Updates #540
Signed-off-by: Jordan Whited <jordan@tailscale.com>
The switch in Conn.runDerpReader() on the derp.ReceivedMessage type
contained cases other than derp.ReceivedPacket that fell through to
writing to c.derpRecvCh, which should only be reached for
derp.ReceivedPacket. This can result in the last/previous
derp.ReceivedPacket to be re-handled, effectively creating a duplicate
packet. If the last derp.ReceivedPacket happens to be a
disco.CallMeMaybe it may result in a disco ping scan towards the
originating peer on the endpoints contained.
The change in this commit moves the channel write on c.derpRecvCh and
subsequent select awaiting the result into the derp.ReceivedMessage
case, preventing it from being reached from any other case. Explicit
continue statements are also added to non-derp.ReceivedPacket cases
where they were missing, in order to signal intent to the reader.
Fixes#10586
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This is gated behind the silent disco control knob, which is still in
its infancy. Prior to this change disco pong reception was the only
event that could move trustBestAddrUntil forward, so even though we
weren't heartbeating, we would kick off discovery pings every
trustUDPAddrDuration and mirror to DERP.
Updates #540
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This change exposes SilentDisco as a control knob, and plumbs it down to
magicsock.endpoint. No changes are being made to magicsock.endpoint
disco behavior, yet.
Updates #540
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Co-authored-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Don't assume Linux lacks UDP_GRO support if it lacks UDP_SEGMENT
support. This mirrors a similar change in wireguard/wireguard-go@177caa7
for consistency sake. We haven't found any issues here, just being
overly paranoid.
Updates #cleanup
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit implements UDP offloading for Linux. GSO size is passed to
and from the kernel via socket control messages. Support is probed 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.
Updates tailscale/corp#8734
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit updates the wireguard-go dependency to pull in fixes for
the tun package, specifically 052af4a and aad7fca.
Signed-off-by: Jordan Whited <jordan@tailscale.com>
The single packet WriteTo() through RebindingUDPConn.WriteBatch() was
not checking for a rebind between loading the PacketConn and writing to
it. Same with ReadFrom()/ReadBatch().
Fixes#6989
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Gateway devices operating as an HA pair w/VRRP or CARP may send UPnP
replies from static addresses rather than the floating gateway address.
This commit relaxes our source address verification such that we parse
responses from non-gateway IPs, and re-point the UPnP root desc
URL to the gateway IP. This ensures we are still interfacing with the
gateway device (assuming L2 security intact), even though we got a
root desc from a non-gateway address.
This relaxed handling is required for ANY port mapping to work on certain
OPNsense/pfsense distributions using CARP at the time of writing, as
miniupnpd may only listen on the static, non-gateway interface address
for PCP and PMP.
Fixes#5502
Signed-off-by: Jordan Whited <jordan@tailscale.com>
This commit updates the wireguard-go dependency and implements the
necessary changes to the tun.Device and conn.Bind implementations to
support passing vectors of packets in tailscaled. This significantly
improves throughput performance on Linux.
Updates #414
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: James Tucker <james@tailscale.com>
Co-authored-by: James Tucker <james@tailscale.com>
TCP selective acknowledgement can improve throughput by an order
of magnitude in the presence of loss.
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Jordan WhitedandGitHub
