netns.SetEnabled(false) is for processes like userspace-only clients
whose traffic never routes back into themselves, so they don't need
the namespace setup. But FromDialer also skipped the ALL_PROXY SOCKS
wrapping when disabled, so such processes couldn't reach the internet
in environments that only allow egress via a proxy, such as ChatGPT
Work's sandbox.
Only skip the namespace setup when disabled, and still wrap the dialer
for the proxy.
Updates #22412
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: Ic2023a329c640a1e8de32b157ed608ba385000a1
tcpproxy propagates a backend's close to the client by calling CloseWrite
on the source connection. connHalf did not implement it, so the call was a
no-op and the client never saw EOF. TestTCPRoundRobinHandler only unblocked
the proxy's remaining copy goroutine by writing more data after the first
echo. When the backend closed with that data still buffered in the pipe,
the write was discarded and the test hung until the 10 minute timeout.
Implement both halves of the half-close contract on connHalf. A new test
covers buffered data surviving CloseWrite, EOF on the read after it, and
writes failing after CloseRead.
Fixes#17533
Signed-off-by: Yunus Emre Karaman <emre884k@gmail.com>
The test asserts that a successful probe records a positive transfer
time, but the transfer is shorter than one clock tick on the Windows CI
runner: four attempts at 1 MiB plus their connection setups took 2.3ms
in total, yet every measurement read exactly 0, so the retry loop added
in #20214 just re-measured zero and then failed with
"http_bw_transfer_time_seconds_total = 0 after 4 attempts".
A 1 MiB loopback transfer takes ~0.4ms here, so no payload small enough
for a unit test is reliably measurable against a millisecond-coarse
clock. Make the transfer span more than one tick instead: /download now
writes half the payload, flushes, sleeps 50ms, then writes the rest.
The probe measures ~52ms, so the >0 assertion holds deterministically
and the retry loop can be deleted.
Fixes#21620
Signed-off-by: Yunus Emre Karaman <emre884k@gmail.com>
* golang.org/x/sys: v0.49.0 to v0.50.0
Triggered by @bradfitz via the bumpdep workflow.
Updates #22390
Change-Id: I402717314704a7f7b164a15e073608ef9dcc31fb
Signed-off-by: Dep Updater <noreply+dep-updater@tailscale.com>
A burst of PR pushes on 2026-11-13 left several GitHub-hosted test
shards sitting in "test all" for hours after their disks filled up,
holding hosted runner slots until GitHub's 6 hour job limit. Nothing
bounded them: go test's -timeout only limits a test binary while it
runs, not compiling and linking it, and test.yml set no job timeouts.
Give every test.yml job a timeout-minutes of roughly twice the slowest
successful run seen across ~1200 recent jobs.
In testwrapper, kill go test if it goes quiet for longer than any
single test binary may run (the go test -timeout plus a 5 minute
margin, overridable via TS_TESTWRAPPER_IDLE_MARGIN) and log the
process tree under testwrapper so the stuck process is visible.
Also bound the go list and listpkgs invocations and each retry
attempt, including its rebuild.
While here, fix testwrapper exiting 0 when go test was killed by a
signal (exit code -1) with no failed tests recorded.
Updates tailscale/corp#51500
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: I52bd41c247fca157d8e56381319fe80ae2f7cf02
Control via nodecap.LogUploadAuth may notify the node
that there is an authorization token to use when uploading logs.
If specified, then the logtail.Logger will use that token
when uploading.
Unfortunately, there is a time-of-use and time-of-check discrepancy
that complicates logging. When logs are written under the context
of being for a particular node, we cannot store the token,
as it may be invalid by the time that the logs are uploaded.
Instead, we add Logger.SetAuthID to annotate subsequent log entries
with an identifier for the token to later use.
Only when we upload do we strip out the auth ID, lookup the
auth token to use, and then perform the upload.
We assume that control will notify the client of new tokens
well before they expire. Fetching a refreshed token
is not the job of the client.
Updates tailscale/corp#39092
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
The Windows installers no longer come in a separate -winui variant.
Remove the temporary qualifier.
Updates tailscale/corp#50987
Signed-off-by: Nick Khyl <nickk@tailscale.com>
run assigned keepAliveTimer after AfterFunc had already armed it, and
the timer's func uses the field (writePending resets it), so nothing
ordered the assignment before the first keepalive. Assign it first and
arm it with Reset, which does.
No existing test ran long enough for a keepalive to fire. The new
TestKeepAlive waits for three under synctest; with the race detector it
fails without this fix.
Updates #13038
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: I8f2c4e6a1b3d5f7091a2c3e4d5f6a7b8c9d0e1f2
onHomeDERPUpdateLocked wrote the netmap cache without checking the
CacheNetworkMaps capability, unlike the full netmap and peer delta
paths. On a node without it, each home DERP change recreated a cache
with no peers, and the next start went Running from it before control
sent any peers.
Check the capability before that write. A cache written before this
fix can still be loaded once; the first netmap from control without
the capability then discards it.
Fixes#22334
Change-Id: I2f49a6617e6756e36316605995b171f6d7b3ff28
Signed-off-by: Andy Tian <me@xiaotian.dev>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The filestore was the only implemented store but we want to be able to
add other stores depending on the platform.
Make space for compiling on more stores, though keep the file store as
the default option.
Updates #22265
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
This is needed for authenticated log uploads (see #20482).
Add a noop declaration now, so that we can call it from
client implementations on other platforms (e.g., Android or iOS).
Updates tailscale/corp#39092
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
For kube-apiserver ProxyGroups with static endpoints, the operator sets
PORT on the k8s-proxy container to the targetPort of the NodePort
Services behind the advertised endpoints. k8s-proxy never read PORT, so
tsnet listened on a random port and traffic to the static endpoints was
forwarded to a port with nothing listening. Peers stayed on DERP.
This fixes this by parsing PORT and setting it as tsnet.Server.Port.
Thanks to @icheishvili (#18841) and @rquitales (#21716), whose changes
this is based on.
Updates #21714
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
policies
This adds support for reading string-list policies represented as
DWORD-backed registry subkeys. Each nonzero DWORD contributes its value
name to the resulting string list, allowing group policy settings to
expose predefined options as checkboxes instead of requiring users to
enter strings manually.
This also adds definitions for HiddenNetworkDevices.
Updates tailscale/corp#50029
Signed-off-by: kari <kari@tailscale.com>
PCP requests with an incorrect nonce are ignored by PCP servers
which follow the simple threat model defined in RFC 6887. Doing
it in natlab will ensure we detect client regressions.
Updates #21127
Change-Id: Ifefcda278800f642c161df90e76babeb089da5a3
Signed-off-by: Francois Marier <francois@tailscale.com>
This uses the same notification delivery logic for policy changes and regular ipn bus updates. Previously, sysPolicyChangedForSession silnetly dropped policy notifications when a watcher's channel was full, potentially leaving clients with stale policy and gui visibility state. Now lagging watchers follow the existing block, suspend, or close behavior, allowing them to recover the latest policy when they resume or reconnect.
Updates tailscale/corp#44104
Signed-off-by: kari-ts <kari@tailscale.com>
tailscale/wireguard-go#116 changed the receive loop to retry any errors
other than net.ErrClosed without a delay. magicsock entered a spin loop
if it encountered a UDP read error.
magicsock's receive funcs now return only net.ErrClosed when the socket
actually closes. Every other read error is counted, logged, and retried
after a short delay. IsUDPReadFeedback and IsUDPReadUnknown are removed
because they can be handled like any other error.
Retrying errors may hide failures which used to stop the receive func.
- There is a new metric magicsock_recv_udp_error which counts all read
errors.
- After 30s of continuous errors, the receive function marks itself as
broken. A warning is raised until a read succeeds.
Updates #22097
Change-Id: Ib9ade0ed001e413205eb7d452aa108aae574e0d1
Signed-off-by: Nick Beirne <nickb@tailscale.com>
Adds a test that issueACMECert, against tstest/acmetest, retries an
order naming "replaces" once without it when the CA rejects it, and
stores the cert from the retried order. It also checks that rate limits
and server errors are not retried, and that the error from the plain
order is returned if that is rejected too.
Updates #22042
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
When a previous cert exists, the renewal order names it in the ARI
"replaces" field. Let's Encrypt rejects that order if the current ACME
account did not issue the cert, which is the case once the account key
has been recreated but the cert kept (for example after
acme-account.key.pem is deleted), so renewal fails until the cert
expires or is deleted. This fixes this by retrying the order once
without "replaces" when the CA rejects it, unless it was rate limited.
The new cert belongs to the current account, so later renewals use
"replaces" again.
Fixes#22042
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
Adds a test that checks whoIs gets the identity from the request's remote
address, and returns an error instead of using the X-Forwarded-For header
when the remote address is not a tailnet node.
Updates tailscale/corp#51000
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
whoIs fell back to the X-Forwarded-For header when WhoIs on the request's
remote address failed (only when serving over plain HTTP, as cmd/k8s-proxy
does). That header comes from the client, so the identity could come from
the request instead of the connection. This fixes this by removing the
fallback, so whoIs only uses the connection's remote address. It also asks
for TCP only, as the proxy only serves TCP, so a miss does not also try UDP.
Updates tailscale/corp#51000
Thank you Emilio Mendoza @a3l6 for reporting
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
wgengine/magicsock: retire stale TSMP disco key after peer restart
When a nonzero control key replaces an old control key equal to the
active TSMP key, updateDiscoKey clears the TSMP-active flag and
subsequent sends use the new control key, while the retired TSMP
material stays stored. A zero control key falls back to the TSMP key as
active, and a different TSMP key that is already active stays preferred.
Each compare-and-swap attempt allocates a fresh disco snapshot before
choosing the active key.
Fixes#22201
Signed-off-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
ParseDuration rewrites the digits before a d or w unit as hours, but only
looked at the digits after the decimal point, so 0.5d quietly returned
7m12s. Return an error for fractional day and week values instead.
Fixes#21552
Signed-off-by: Aniket <aniketkomailkro@gmail.com>
Lets an embedder pick the dialer for non-tailnet connections, such
as to control. When control runs in the same process, the node can
then reach it in memory (e.g. memnet) instead of via loopback.
Updates tailscale/corp#27636
Signed-off-by: Kristoffer Dalby <kristoffer@dalby.cc>
Prepare for shared socket policy for DERP, control, and log uploads.
Callers need to configure options such as TCP user timeout on the same
socket that netns configures to avoid routing loops.
Use ControlContext for dialer namespace setup and run any caller-supplied
ControlContext afterward, only if namespace setup succeeds. Reject a
non-nil Control unconditionally with a panic directing callers to
ControlContext, including when netns is disabled, rather than support
both hook types and their precedence rules. Listener Control hooks
remain unchanged.
Existing in-tree FromDialer callers supply neither hook, so no caller
migration is required. Keep socket defaults and proxy-wrapping behavior
unchanged. Test hook ordering, context and argument forwarding, errors,
legacy-hook rejection, disabled mode, and proxy delegation.
Updates tailscale/corp#50948
Change-Id: I9d778273f3257fe8a298463ef29ec2dd5d5fb7b0
Signed-off-by: James Tucker <james@tailscale.com>
The eventbus routes events by exact type, so a subscriber of *T never
sees a T. Add a test that type-checks the module for linux, windows and
darwin and fails if any event type is used with eventbus or
eventbustest both by value and by pointer. To keep memory use down, it
drops function bodies outside the packages that import those. It skips
in race builds because of a go/types data race, golang/go#81122.
Updates #22307
Change-Id: Ia7486c43936f367f8767ad9c8750faa9b6bdf8f0
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Both tests passed filter funcs taking a pointer to an event type that
is published by value, so the filters never matched. The health test
passed only because Expect timed out, so make it look for the event it
means to forbid.
Updates #22307
Change-Id: I7daf64656cd9b0e9e98372cd803a6e824db96594
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Count DERP keepalive and ping/pong traffic, terminal read-deadline
expirations, and control map-request watchdog expirations. Count
successful pre-compression log upload bytes alongside existing upload
totals to support batch-size comparisons.
Keep timeout attribution narrow: expirations do not imply a successful
reconnect, and completed writes do not imply peer receipt. Log upload
metrics remain on the existing per-logger ExpVar surface.
Add deterministic coverage without changing heartbeat timing, socket
settings, reconnect behavior, or upload policy.
Updates tailscale/corp#50948
Change-Id: I95f7bab62c000172a4825283bb4c055dd29d8043
Signed-off-by: James Tucker <james@tailscale.com>
Use ControlContext for production net.Dialer socket hooks:
PeerAPI and debug peer dialing, captive-portal probes, stunstamp TCP
measurements, and the test agent connection.
Update the PeerDialControlFunc callback signature and Apple extension
implementation together with both consumers. Preserve namespace binding,
socket settings, timeout values, and existing error handling. Listener
Control and RawConn.Control are separate APIs and remain unchanged.
Add coverage for platform hook forwarding and errors, debug and captive
portal dialing, cancellation, SO_REUSEADDR, and test-agent hook selection
and keepalive configuration. Cover the Apple missing-interface error and
non-Linux bypass no-op in their platform-specific tests.
Updates tailscale/corp#50948
Change-Id: I4fb162666724f307f2d3d0beba6ed1b0ed15bb4f
Signed-off-by: James Tucker <james@tailscale.com>
https://github.com/tailscale/wireguard-go/compare/79ead109a81a8f57c75ac9e6ddee61714c82744b...9e10ef9af41578baabde0283237d5cbb6422ff25
Start off with device.WithQueueStagedGlobalLimit inheriting the same
value from the knob it replaced, device.WithPreallocatedBuffersPerPool.
Unlike WithPreallocatedBuffersPerPool, WithQueueStagedGlobalLimit is
exclusive of packet memory allocated for transport messages that are
immediately transmitted. So, one must consider
WithQueueStagedGlobalLimit alongside .Queue{In,Out}boundSize to
understand overall memory ceiling impacts.
An approximation of the configured packet memory ceiling on iOS can be
calculated with the following formula:
device.singlePacketSlabSize * ( // 1788
(2 * device.WithQueueOutboundSize) + // 128
(2 * device.WithQueueInboundSize) + // 128
device.WithQueueStagedGlobalLimit // 64
)
=
~572KB
Where previously it was:
device.singlePacketSlabSize * ( // 1788
device.WithPreallocatedBuffersPerPool // 64
)
=
~114KB
Updates tailscale/tailscale#1971
Updates tailscale/tailscale#21773
Updates tailscale/corp#46396
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Add basic support for the deletion of port mappings in natlab.
NAT-PMP clients are supposed to set the external port to 0 in their
requests, but we test with a bogus port since servers are supposed
to ignore this value.
Note that passing a lifetime of 0 to a UPnP mapping will delete
that mapping (like it's supposed to in PCP/NAT-PMP) instead of
following the spec and turning that into 604800s.
Updates #21713
Change-Id: I487cc0cae1247dff69484658f86a45d278bcdea7
Signed-off-by: Francois Marier <francois@tailscale.com>
The eventbus routes events by exact type, and these three subscribers
took a pointer to an event type that is only ever published by value,
so they never ran. logtail could not detect the network coming back up
after an upload failure, LocalBackend never recorded or announced the
latest client version, and the magicsock portmapper log limiter never
reset after a major link change.
Updates #22307
Signed-off-by: Nick Khyl <nickk@tailscale.com>
Time.LoadAtomic and StoreAtomic only worked on 32-bit platforms if
the field happened to be 64-bit aligned, which tstun and magicsock
each guarded with a test. AtomicTime wraps an atomic.Int64, so the
compiler guarantees its alignment, and the tests can go. Also use it
for tstime/rate's Limiter.next, which was an atomic.Int64 holding a
mono.Time.
Updates #cleanup
Updates #22290 (where this was discovered)
Change-Id: I02807906ed9be59446456853bccece6e32e472ed
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
wgengine/filter asks its accept-logging Limiter about every accepted
packet, so with tailscaled's TUN and UDP queues spread over CPUs
(tailscale/tailscale#22188), that Limiter's mutex took 11% of a subnet
router's CPU on a 128-CPU host, most of it contended. On a denial,
record when the next token is due, and deny before then without taking
the mutex. A denial with a backwards timestamp no longer rewinds the
bucket's last update time, which credited it with tokens it never
earned.
│ old │ new │
│ sec/op │ sec/op vs base │
AllowN 13.550n ± 0% 1.143n ± 2% -91.56% (p=0.000 n=10)
AllowN-8 108.6500n ± 4% 0.1497n ± 4% -99.86% (p=0.000 n=10)
geomean 38.37n 0.4137n -98.92%
Updates #22188
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: I928f64d3df7b51c6cac0da7dafe9ef5e71acaa05
tsnet called the serve LocalAPI client and LocalBackend methods
unconditionally, so it didn't build with ts_omit_serve, which every
minimal feature set includes. Make those methods return
feature.ErrUnavailable when serve is omitted, as ListenFunnel and
ListenService now do too.
Declare tsnet's minimum feature set (netstack) as
featuretags.TSNetMin, add cmd/featuretags --tsnet to keep it, and add
a test that type-checks tsnet with it and a few other small feature
sets.
Fixes#22282
Updates #22283
Change-Id: I866fa34d052bfd437b1c9d3309063999d7ebb02c
Co-Authored-By: Fredrik Fornwall <fredrik@fornwall.net>
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
This moves HiddenNetworkDvice policy out of ipnlocal and into an optional ipnext.
This adds a per-session notification hook that uses each watcher's effective policy snapshot to populate NodeGUIVisibility.
A full visibility snapshot is sent on initialization and policy changes, and on subsequent peer updates, only changed peers are evaluated.
The extension populates Notify.NodeGUIVisibility with the hidden peers
among the nodes the Notify covers: all peers when it carries Policy
(initially and on policy change), or the PeersChanged peers otherwise.
Absent means visible, so a Policy-bearing Notify replaces the client's
visibility state and a peer-change Notify updates only the changed peers.
No NodeGUIVisibilityChanges/Removed fields and no per-session node state
are needed: removed peers simply leave the client's peer set, and full
netmap updates already resend every peer alongside SelfChange.
When the feature isn't linked, the hook slice is empty and the fast path
in notifyForSessionLocked is unchanged. When the policy isn't configured,
the hook costs a snapshot lookup per session per Notify and sends nothing.
Updates tailscale/corp#44104
Signed-off-by: kari <kari@tailscale.com>
Signed-off-by: Nick Khyl <nickk@tailscale.com>
Tests that the localapi cert-status/ handler checks read access and returns
the status as JSON, that LocalBackend.CertStatus goes through the hook, and
that feature/acme's hook reports the same missing, valid and expired results
as the C2N /tls-cert-status handler. Also tests that Client.CertStatus
returns ErrCertStatusUnsupported for a 404 and not for other errors.
Updates tailscale/corp#38175
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
The localapi cert/ endpoint fetches a cert when none is cached, so polling it
triggers issuance. feature/acme can already report a cached cert without
contacting the CA, but only over the C2N /tls-cert-status path. This moves
that check into an extension method and serves it as localapi cert-status/,
with a matching Client.CertStatus, reporting Missing on a cache miss instead
of issuing. Client.CertStatus returns ErrCertStatusUnsupported (wrapping
errors.ErrUnsupported) when tailscaled has no cert-status endpoint.
Updates tailscale/corp#38175
Signed-off-by: chaosinthecrd <tom@tmlabs.co.uk>
The cleanup returned by flowCreated decremented activeFlowCount on every
call, and its documented contract pushed the burden of calling it exactly
once onto the datapath. A caller that lost track and ran a cleanup twice
would undercount the flows still forwarding over the assignment, letting
popExpired reclaim both of its addresses back to the pools mid-connection,
or drive the count to -1, which can never again satisfy popExpired's
zero-flow test and so pins the assignment in the expiry heap and leaks
both addresses for the life of the process.
The cleanup now releases its hold at most once, guarded by a bool read and
written under c.mu alongside the decrement it protects. Repeat calls are
no-ops. The Conn25Datapath docs no longer ask callers to guarantee this.
No such double-call is reachable through FlowTable today: removeFlowLocked
deletes both cache entries, so NewFlow's second lookup misses, and the
capacity eviction takes lru.Back(), never the flow just pushed to the
front. This removes the class rather than a live bug.
Updates https://github.com/tailscale/corp/issues/50760
Change-Id: I1888f36eb8fdec9b219ccfd2d290dc4eb19c55a0
Signed-off-by: Mazdak Nasab <mazdak.nasab@gmail.com>
Funnel can't be on without serve. With the HasServe check, the
compiler reduces the method to a constant and inlines it, so the
linker drops it from those builds.
Updates #22283
Change-Id: Idd5da9c1e4f8f41a828a3774f5c4d44872eb1dda
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>