Brad FitzpatrickandBrad Fitzpatrick
125fd88c30
FreeBSD guests downloaded their test binaries from vnet's
files.tailscale VIP at roughly 250 kB/s in CI, and transfers sometimes
wedged outright for many minutes, which is why TestSubnetRouterFreeBSD
timed out in about a third of its runs. Locally the same path moves
data at 100+ MB/s, so the problem was never CPU; it was TCP behavior
under two independent constraints, both diagnosed with a new
throughput harness (TestVnetPerfFreeBSDDownload), a VNET_TCP_DEBUG
endpoint sampler, and pcaps:
First, throughput is capped at receive-window/RTT. FreeBSD starts its
receive window at 64 kB and autoscales it in slow 16 kB steps, and on
an oversubscribed CI runner the effective RTT of the userspace vnet
data path reaches hundreds of milliseconds, giving almost exactly the
observed 250 kB/s. Fix: raise the FreeBSD guest's TCP buffer sysctls
in cloud-init before the downloads, and raise netstack's receive
buffer sizing for the reverse (upload) direction.
Second, the outright wedge: when netstack bursts more data than the
QEMU socket plus the guest's virtio RX ring can absorb, a wide swath
of segments is dropped downstream of vnet, and netstack's loss
recovery then crawls, retransmitting one or two segments per 200 ms
RTO for minutes at a time (a 33 MB transfer was observed taking 526
seconds against an otherwise idle receiver). Rather than depending on
recovery from mass loss, make the path effectively lossless by keeping
the maximum in-flight data (the 1 MB netstack send buffer) below the
downstream buffering: grow the guests' virtio RX rings from 256 to
1024 descriptors, enlarge the vnet-QEMU unix socket buffers, and grow
the netstack link endpoint queue from 512 to 4096 packets so a send
burst can't overflow it.
Also fixed along the way, found while chasing the above:
* pcapWriter fsync'd after every packet, serializing all traffic
behind disk writes when a test enables pcap; a pcap-enabled run
was capped at about 290 kB/s. Keep the per-packet Flush but drop
the per-packet fsync.
* Traffic originating from vnet's own netstack (control plane,
DERP, file servers) bypassed conditionedWrite, so SetLatency and
SetPacketLoss silently didn't apply to it.
* writeEthernetFrameToVM held one global mutex (and a shared
scratch buffer) across writes to all VMs, so one guest slow to
drain its socket stalled traffic to every VM on the server. The
write lock is now per-VM-connection.
TestSubnetRouterFreeBSD now passes locally in 31s (down from 4.5
minutes), still passes with the vnet simulating a 100 ms RTT
(downloads at 2-8 MB/s, previously 250-600 kB/s), and passes in 65s
with KVM disabled while pinned to two host CPUs, a harsher environment
than the CI runners. The benchmark test is opt-in via
--run-perf-tests (in addition to --run-vm-tests) so CI doesn't spend
a matrix job re-measuring it on every run. VMTEST_NO_KVM=1
forces TCG for reproducing slow-host behavior.
Fixes tailscale/corp#44805
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: I1a7945a7e9c7d083b0ea2a3530eda0e9757dff18
Tailscale
Private WireGuard® networks made easy
Overview
This repository contains the majority of Tailscale's open source code.
Notably, it includes the tailscaled daemon and
the tailscale CLI tool. The tailscaled daemon runs on Linux, Windows,
macOS, and to varying degrees
on FreeBSD and OpenBSD. The Tailscale iOS and Android apps use this repo's
code, but this repo doesn't contain the mobile GUI code.
Other Tailscale repos of note:
- the Android app is at https://github.com/tailscale/tailscale-android
- the Synology package is at https://github.com/tailscale/tailscale-synology
- the QNAP package is at https://github.com/tailscale/tailscale-qpkg
- the Chocolatey packaging is at https://github.com/tailscale/tailscale-chocolatey
For background on which parts of Tailscale are open source and why, see https://tailscale.com/opensource/.
Using
We serve packages for a variety of distros and platforms at https://pkgs.tailscale.com.
Other clients
The macOS, iOS, and Windows clients use the code in this repository but additionally include small GUI wrappers. The GUI wrappers on non-open source platforms are themselves not open source.
Building
We always require the latest Go release, currently Go 1.26. (While we build releases with our Go fork, its use is not required.)
go install tailscale.com/cmd/tailscale{,d}
If you're packaging Tailscale for distribution, use build_dist.sh
instead, to burn commit IDs and version info into the binaries:
./build_dist.sh tailscale.com/cmd/tailscale
./build_dist.sh tailscale.com/cmd/tailscaled
If your distro has conventions that preclude the use of
build_dist.sh, please do the equivalent of what it does in your
distro's way, so that bug reports contain useful version information.
Bugs
Please file any issues about this code or the hosted service on the issue tracker.
Contributing
PRs welcome! But please file bugs. Commit messages should reference bugs.
We require Developer Certificate of
Origin
Signed-off-by lines in commits.
See commit-messages.md (or skim git log) for our commit message style.
About Us
Tailscale is primarily developed by the people at https://github.com/orgs/tailscale/people. For other contributors, see:
- https://github.com/tailscale/tailscale/graphs/contributors
- https://github.com/tailscale/tailscale-android/graphs/contributors
Legal
WireGuard is a registered trademark of Jason A. Donenfeld.