This change introduces a new subcommand, `exit-node`, along with a
subsubcommand of `list` and a `--filter` flag.
Exit nodes without location data will continue to be displayed when
`status` is used. Exit nodes with location data will only be displayed
behind `exit-node list`, and in status if they are the active exit node.
The `filter` flag can be used to filter exit nodes with location data by
country.
Exit nodes with Location.Priority data will have only the highest
priority option for each country and city listed. For countries with
multiple cities, a <Country> <Any> option will be displayed, indicating
the highest priority node within that country.
Updates tailscale/corp#13025
Signed-off-by: Charlotte Brandhorst-Satzkorn <charlotte@tailscale.com>
Define PeerCapabilty and PeerCapMap as the new way of sending down
inter-peer capability information.
Previously, this was unstructured and you could only send down strings
which got too limiting for certain usecases. Instead add the ability
to send down raw JSON messages that are opaque to Tailscale but provide
the applications to define them however they wish.
Also update accessors to use the new values.
Updates #4217
Signed-off-by: Maisem Ali <maisem@tailscale.com>
The server hasn't sent it in ages.
Updates #cleanup
Change-Id: I9695ab0f074ec6fb006e11faf3cdfc5ca049fbf8
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Trying to SSH when SELinux is enforced results in errors like:
```
➜ ~ ssh ec2-user@<ip>
Last login: Thu Jun 1 22:51:44 from <ip2>
ec2-user: no shell: Permission denied
Connection to <ip> closed.
```
while the `/var/log/audit/audit.log` has
```
type=AVC msg=audit(1685661291.067:465): avc: denied { transition } for pid=5296 comm="login" path="/usr/bin/bash" dev="nvme0n1p1" ino=2564 scontext=system_u:system_r:unconfined_service_t:s0 tcontext=unconfined_u:unconfined_r:unconfined_t:s0 tclass=process permissive=0
```
The right fix here would be to somehow install the appropriate context when
tailscale is installed on host, but until we figure out a way to do that
stop using the `login` cmd in these situations.
Updates #4908
Signed-off-by: Maisem Ali <maisem@tailscale.com>
The netstack code had a bunch of logic to figure out if the LocalBackend should handle an
incoming connection and then would call the function directly on LocalBackend. Move that
logic to LocalBackend and refactor the methods to return conn handlers.
Updates #cleanup
Signed-off-by: Maisem Ali <maisem@tailscale.com>
This is a follow up on PR #8172 that adds a synchronous Poll method
which allows for the Poller to be used as a zero value without needing
the constructor. The local backend is also changed to use the new API.
A follow up PR will remove the async functionality from the portlist package.
Updates #8171
Signed-off-by: Marwan Sulaiman <marwan@tailscale.com>
This change introduces a NodeKey func on localbackend that returns the
public node key.
Updates tailscale/corp#9967
Signed-off-by: Charlotte Brandhorst-Satzkorn <charlotte@tailscale.com>
This is part of an effort to clean up tailscaled initialization between
tailscaled, tailscaled Windows service, tsnet, and the mac GUI.
Updates #8036
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
On some platforms (notably macOS and iOS) we look up the default
interface to bind outgoing connections to. This is both duplicated
work and results in logspam when the default interface is not available
(i.e. when a phone has no connectivity, we log an error and thus cause
more things that we will try to upload and fail).
Fixed by passing around a netmon.Monitor to more places, so that we can
use its cached interface state.
Fixes#7850
Updates #7621
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
We're using it in more and more places, and it's not really specific to
our use of Wireguard (and does more just link/interface monitoring).
Also removes the separate interface we had for it in sockstats -- it's
a small enough package (we already pull in all of its dependencies
via other paths) that it's not worth the extra complexity.
Updates #7621
Updates #7850
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
This is a follow-up to #7905 that adds two more linters and fixes the corresponding findings. As per the previous PR, this only flags things that are "obviously" wrong, and fixes the issues found.
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: I8739bdb7bc4f75666a7385a7a26d56ec13741b7c
Redoes the approach from #5550 and #7539 to explicitly pass in the logf
function, instead of having global state that can be overridden.
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
This makes the sockstat logger available on all builds, but only enables
it by default for unstable. For stable builds, the logger must be
explicitly enabled via C2N component logger.
Updates tailscale/corp#9230
Updates #3363
Signed-off-by: Will Norris <will@tailscale.com>
For stores like k8s secrets we need to dial out to the k8s API as though Tailscale
wasn't running. The issue currently only manifests when you try to use an exit node
while running inside a k8s cluster and are trying to use Kubernetes secrets as the
backing store.
This doesn't address cmd/containerboot, which I'll do in a follow up.
Updates #7695
Signed-off-by: Maisem Ali <maisem@tailscale.com>
Since users can run tailscaled in a variety of ways (root, non-root,
non-root with process capabilities on Linux), this check will print the
current process permissions to the log to aid in debugging.
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: Ida93a206123f98271a0c664775d0baba98b330c7
We were checking against the wrong directory, instead if we
have a custom store configured just use that.
Fixes#7588Fixes#7665
Signed-off-by: Maisem Ali <maisem@tailscale.com>
This change focuses on the backend log ID, which is the mostly commonly
used in the client. Tests which don't seem to make use of the log ID
just use the zero value.
Signed-off-by: Will Norris <will@tailscale.com>
Switch to using logtail for logging sockstat logs. Always log locally
(on supported platforms), but disable automatic uploading. Change
existing c2n sockstats request to trigger upload to log server and
return log ID.
Signed-off-by: Will Norris <will@tailscale.com>
Previously the part that handled Funnel connections was not
aware of any listeners that tsnet.Servers might have had open
so it would check against the ServeConfig and fail.
Adding a ServeConfig for a TCP proxy was also not suitable in this
scenario as that would mean creating two different listeners and have
one forward to the other, which really meant that you could not have
funnel and tailnet-only listeners on the same port.
This also introduces the ipn.FunnelConn as a way for users to identify
whether the call is coming over funnel or not. Currently it only holds
the underlying conn and the target as presented in the "Tailscale-Ingress-Target"
header.
Signed-off-by: Maisem Ali <maisem@tailscale.com>
This change adds a ringbuffer to each magicsock endpoint that keeps a
fixed set of "changes"–debug information about what updates have been
made to that endpoint.
Additionally, this adds a LocalAPI endpoint and associated
"debug peer-status" CLI subcommand to fetch the set of changes for a given
IP or hostname.
Updates tailscale/corp#9364
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: I34f726a71bddd0dfa36ec05ebafffb24f6e0516a
This prevents a panic where we synthesize a new netmap in
setClientStatus after we've shut down and nil'd out the controlclient,
since that function expects to be called while connected to control.
Fixes#7392
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: Ib631eb90f34f6afa008d69bbb386f70da145e102
With #6566 we added an external mechanism for getting the default
interface, and used it on macOS and iOS (see tailscale/corp#8201).
The goal was to be able to get the default physical interface even when
using an exit node (in which case the routing table would say that the
Tailscale utun* interface is the default).
However, the external mechanism turns out to be unreliable in some
cases, e.g. when multiple cellular interfaces are present/toggled (I
have occasionally gotten my phone into a state where it reports the pdp_ip1
interface as the default, even though it can't actually route traffic).
It was observed that `ifconfig -v` on macOS reports an "effective interface"
for the Tailscale utn* interface, which seems promising. By examining
the ifconfig source code, it turns out that this is done via a
SIOCGIFDELEGATE ioctl syscall. Though this is a private API, it appears
to have been around for a long time (e.g. it's in the 10.13 xnu release
at https://opensource.apple.com/source/xnu/xnu-4570.41.2/bsd/net/if_types.h.auto.html)
and thus is unlikely to go away.
We can thus use this ioctl if the routing table says that a utun*
interface is the default, and go back to the simpler mechanism that
we had before #6566.
Updates #7184
Updates #7188
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
With #6566 we started to more aggressively bind to the default interface
on Darwin. We are seeing some reports of the wrong cellular interface
being chosen on iOS. To help with the investigation, this adds to knobs
to control the behavior changes:
- CapabilityDebugDisableAlternateDefaultRouteInterface disables the
alternate function that we use to get the default interface on macOS
and iOS (implemented in tailscale/corp#8201). We still log what it
would have returned so we can see if it gets things wrong.
- CapabilityDebugDisableBindConnToInterface is a bigger hammer that
disables binding of connections to the default interface altogether.
Updates #7184
Updates #7188
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
We now handle the case where the NetworkMap.SelfNode has already expired
and do not return an expiry time in the past (which causes an ~infinite
loop of timers to fire).
Additionally, we now add an explicit check to ensure that the next
expiry time is never before the current local-to-the-system time, to
ensure that we don't end up in a similar situation due to clock skew.
Finally, we add more tests for this logic to ensure that we don't
regress on these edge cases.
Fixes#7193
Change-Id: Iaf8e3d83be1d133a7aab7f8d62939e508cc53f9c
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
If the user passes the --diagnose flag, print a warning if any of the
default or fallback DNS resolvers are Tailscale IPs. This can interfere
with the ability to connect to the controlplane, and is typically
something to pay attention to if there's a connectivity issue.
Change-Id: Ib14bf6228c037877fbdcd22b069212b1a4b2c456
Signed-off-by: Andrew Dunham <andrew@tailscale.com>
We can log too quickly for logtail to catch up, even when we opt out of
log rate-limiting. When the user passes the --diagnose flag to
bugreport, we use a token bucket to control how many logs per second are
printed and sleep until we're able to write more.
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: If27672d66b621b589280bd0fe228de367ffcbd8f
The iOS has a command to reset the persisted state of the app, but it
was doing its own direct keychain manipulation. This proved to be
brittle (since we changed how preferences are stored with #6022), so
we instead add a LocalAPI endpoint to do do this, which can be updated
in tandem.
This clears the same state as the iOS implementation (tailscale/corp#3186),
that is the machine key and preferences (which includes the node key).
Notably this does not clear the logtail ID, so that logs from the device
still end up in the same place.
Updates tailscale/corp#8923
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
This updates all source files to use a new standard header for copyright
and license declaration. Notably, copyright no longer includes a date,
and we now use the standard SPDX-License-Identifier header.
This commit was done almost entirely mechanically with perl, and then
some minimal manual fixes.
Updates #6865
Signed-off-by: Will Norris <will@tailscale.com>
When turned on via environment variable (off by default), this will use
the BSD routing APIs to query what interface index a socket should be
bound to, rather than binding to the default interface in all cases.
Updates #5719
Updates #5940
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: Ib4c919471f377b7a08cd3413f8e8caacb29fee0b
This allows users to temporarily enable/disable dnscache logging via a
new node capability, to aid in debugging strange connectivity issues.
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: I46cf2596a8ae4c1913880a78d0033f8b668edc08
The current node isn't in NetMap.Peers, so without this we would not
have fired this timer on self expiry.
Updates #6932
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: Id57f96985397e372f9226802d63b42ff92c95093
This change delays the first flush in the /watch-ipn-bus/ handler
until after the watcher has been successfully installed on the IPN
bus. It does this by adding a new onWatchAdded callback to
LocalBackend.WatchNotifications().
Without this, the endpoint returns a 200 almost immediatly, and
only then installs a watcher for IPN events. This means there's a
small window where events could be missed by clients after calling
WatchIPNBus().
Fixestailscale/corp#8594.
Signed-off-by: salman <salman@tailscale.com>
In order to be able to synthesize a new NetMap when a node expires, have
LocalBackend start a timer when receiving a new NetMap that fires
slightly after the next node expires. Additionally, move the logic that
updates expired nodes into LocalBackend so it runs on every netmap
(whether received from controlclient or self-triggered).
Updates #6932
Signed-off-by: Andrew Dunham <andrew@du.nham.ca>
Change-Id: I833390e16ad188983eac29eb34cc7574f555f2f3
Needed for clients that get information via the /v0/status LocalAPI
endpoint (e.g. to not offer expired exit nodes as options).
Updates tailscale/corp#8702
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
UI works remains, but data is there now.
Updates #4015
Change-Id: Ib91e94718b655ad60a63596e59468f3b3b102306
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
This makes `tailscale debug watch-ipn` safe to use for troubleshooting
user issues, in addition to local debugging during development.
Signed-off-by: David Anderson <danderson@tailscale.com>
By default, `http.Transport` keeps idle connections open hoping to re-use them in the future. Combined with a separate transport per request in HTTP proxy this results in idle connection leak.
Fixes#6773
Nodes which have both -advertise-exit-node and -exit-node in prefs
should continue have them until the next invocation of `tailscale up`.
Updates #3569.
Signed-off-by: salman <salman@tailscale.com>
x/exp/slices now has ContainsFunc (golang/go#53983) so we can delete
our versions.
Change-Id: I5157a403bfc1b30e243bf31c8b611da25e995078
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
For testing of Windows GUI client.
Updates #6480
Change-Id: I42f7526d95723e14bed7085fb759e371b43aa9da
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
To simplify clients getting the initial state when they subscribe.
Change-Id: I2490a5ab2411253717c74265a46a98012b80db82
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
If user's fn returned false and never canceled their ctx, we never
stopped the NotifyWatchEngineUpdates goroutine.
This was introduced recently (this cycle).
Change-Id: I3453966ac71e00727296ddd237ef845782f4e52e
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Previously, `TAILSCALE_USE_WIP_CODE` was needed to hit a bunch of the TKA paths. With
this change:
- Enablement codepaths (NetworkLockInit) and initialization codepaths (tkaBootstrapFromGenesisLocked via tkaSyncIfNeeded)
require either the WIP envknob or CapabilityTailnetLockAlpha.
- Normal operation codepaths (tkaSyncIfNeeded, tkaFilterNetmapLocked) require TKA to be initialized, or either-or the
envknob / capability.
- Auxillary commands (ie: changing tka keys) require TKA to be initialized.
The end result is that it shouldn't be possible to initialize TKA (or subsequently use any of its features) without being
sent the capability or setting the envknob on tailscaled yourself.
I've also pulled out a bunch of unnecessary checks for CanSupportNetworkLock().
Signed-off-by: Tom DNetto <tom@tailscale.com>
The Go style weirds people out so we try to stick to the more
well-known double hyphen style in docs.
Change-Id: Iad6db5c82cda37f6b7687eed7ecd9276f8fd94d6
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We want users to have the freedom to start tailscaled with `-no-logs-no-support`,
but that is obviously in direct conflict with tailnets that have network logging
enabled.
When we detect that condition, we record the issue in health, notify the client,
set WantRunning=false, and bail.
We clear the item in health when a profile switch occurs, since it is a
per-tailnet condition that should not propagate across profiles.
Signed-off-by: Aaron Klotz <aaron@tailscale.com>
We use this pattern in a number of places (in this repo and elsewhere)
and I was about to add a fourth to this repo which was crossing the line.
Add this type instead so they're all the same.
Also, we have another Set type (SliceSet, which tracks its keys in
order) in another repo we can move to this package later.
Change-Id: Ibbdcdba5443fae9b6956f63990bdb9e9443cefa9
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
The "userID is empty everywhere but Windows" docs on lots of places
but not everywhere while using just a string type was getting
confusing. This makes a new type to wrap up those rules, however
weird/historical they might be.
Change-Id: I142e85a8e38760988d6c0c91d0efecedade81b9b
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
So GUI clients don't need to poll for it.
We still poll internally (for now!) but that's still cheaper. And will
get much cheaper later, without having to modify clients once they
start sending this bit.
Change-Id: I36647b701c8d1fe197677e5eb76f6894e8ff79f7
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Follow-up to #6467 and #6506.
LocalBackend knows the server-mode state, so move more auth checking
there, removing some bookkeeping from ipnserver.Server.
Updates #6417
Updates tailscale/corp#8051
Change-Id: Ic5d14a077bf0dccc92a3621bd2646bab2cc5b837
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
While reading the DNS code noticed that we were still using FallbackResolvers
in this code path but the comment was out of date.
Signed-off-by: Maisem Ali <maisem@tailscale.com>
The earlier 5f6d63936f was not complete.
Updates tailscale/corp#7515
Change-Id: I35efca51d1584c48ef6834a7d29cd42d7c943628
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We would end up with duplicate profiles for the node as the UserID
would have chnaged. In order to correctly deduplicate profiles, we
need to look at both the UserID and the NodeID. A single machine can
only ever have 1 profile per NodeID and 1 profile per UserID.
Note: UserID of a Node can change when the node is tagged/untagged,
and the NodeID of a device can change when the node is deleted so we
need to check for both.
Updates #713
Signed-off-by: Maisem Ali <maisem@tailscale.com>
The fix in 4fc8538e2 was sufficient for IPv6. Browsers (can?) send the
IPv6 literal, even without a port number, in brackets.
Updates tailscale/corp#7948
Change-Id: I0e429d3de4df8429152c12f251ab140b0c8f6b77
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
I was too late with review feedback to 513780f4f8.
Updates tailscale/corp#7948
Change-Id: I8fa3b4eba4efaff591a2d0bfe6ab4795638b7c3a
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We were not updating the LoginProfile.UserProfile when a netmap
updated the UserProfile (e.g. when a node was tagged via the admin panel).
Updates #713
Signed-off-by: Maisem Ali <maisem@tailscale.com>
This moves the NetworkLock key from a dedicated StateKey to be part of the persist.Persist struct.
This struct is stored as part for ipn.Prefs and is also the place where we store the NodeKey.
It also moves the ChonkDir from "/tka" to "/tka-profile/<profile-id>". The rename was intentional
to be able to delete the "/tka" dir if it exists.
This means that we will have a unique key per profile, and a unique directory per profile.
Note: `tailscale logout` will delete the entire profile, including any keys. It currently does not
delete the ChonkDir.
Signed-off-by: Maisem Ali <maisem@tailscale.com>
We do not need to wait for it to complete. And we might have to
call Shutdown from callback from the controlclient which might
already be holding a lock that Shutdown requires.
Updates #713
Signed-off-by: Maisem Ali <maisem@tailscale.com>
The health package was turning into a rando dumping ground. Make a new
Warnable type instead that callers can request an instance of, and
then Set it locally in their code without the health package being
aware of all the things that are warnable. (For plenty of things the
health package will want to know details of how Tailscale works so it
can better prioritize/suppress errors, but lots of the warnings are
pretty leaf-y and unrelated)
This just moves two of the health warnings. Can probably move more
later.
Change-Id: I51e50e46eb633f4e96ced503d3b18a1891de1452
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
This makes it so that the backend also restarts when users change,
otherwise an extra call to Start was required.
Updates #713
Signed-off-by: Maisem Ali <maisem@tailscale.com>
Noticed this while debugging something else, we would reset all routes if
either `--advertise-exit-node` or `--advertise-routes` were set. This handles
correctly updating them.
Also added tests.
Signed-off-by: Maisem Ali <maisem@tailscale.com>
Example output:
# Health check:
# - Some peers are advertising routes but --accept-routes is false
Also, move "tailscale status" health checks to the bottom, where they
won't be lost in large netmaps.
Updates #2053
Updates #6266
Change-Id: I5ae76a0cd69a452ce70063875cd7d974bfeb8f1a
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Numerous issues have been filed concerning an inability to install and run
Tailscale headlessly in unattended mode, particularly after rebooting. The
server mode `Prefs` stored in `server-state.conf` were not being updated with
`Persist` state once the node had been succesfully logged in.
Users have been working around this by finagling with the GUI to make it force
a state rewrite. This patch makes that unnecessary by ensuring the required
server mode state is updated when prefs are updated by the control client.
Fixes https://github.com/tailscale/tailscale/issues/3186
Signed-off-by: Aaron Klotz <aaron@tailscale.com>
cmd/viewer couldn't deal with that map-of-map. Add a wrapper type
instead, which also gives us a place to add future stuff.
Updates tailscale/corp#7515
Change-Id: I44a4ca1915300ea8678e5b0385056f0642ccb155
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Not for end users (unless directed by support). Mostly for ease of
development for some upcoming webserver work.
Change-Id: I43acfed217514567acb3312367b24d620e739f88
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
It is currently a `ipn.PrefsView` which means when we do a JSON roundtrip,
we go from an invalid Prefs to a valid one.
This makes it a pointer, which fixes the JSON roundtrip.
This was introduced in 0957bc5af2.
Signed-off-by: Maisem Ali <maisem@tailscale.com>
* Plumb disablement values through some of the internals of TKA enablement.
* Transmit the node's TKA hash at the end of sync so the control plane understands each node's head.
* Implement /machine/tka/disable RPC to actuate disablement on the control plane.
There is a partner PR for the control server I'll send shortly.
Signed-off-by: Tom DNetto <tom@tailscale.com>
This makes tags, creation time, exit node option and primary routes
for the current node exposed via `tailscale status --json`
Signed-off-by: Anton Tolchanov <anton@tailscale.com>
Poller.C and Poller.c were duplicated for one caller. Add an accessor
returning the receive-only version instead. It'll inline.
Poller.Err was unused. Remove.
Then Poller is opaque.
The channel usage and shutdown was a bit sketchy. Clean it up.
And document some things.
Change-Id: I5669e54f51a6a13492cf5485c83133bda7ea3ce9
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Running corp/ipn#TestNetworkLockE2E has a 1/300 chance of failing, and
deskchecking suggests thats whats happening are two netmaps are racing each
other to be processed through tkaSyncIfNeededLocked. This happens in the
first place because we release b.mu during network RPCs.
To fix this, we make the tka sync logic an exclusive section, so two
netmaps will need to wait for tka sync to complete serially (which is what
we would want anyway, as the second run through probably wont need to
sync).
Signed-off-by: Tom DNetto <tom@tailscale.com>
The macOS and iOS apps that used the /localapi/v0/file-targets handler
were getting too many candidate targets. They wouldn't actually accept
the file. This is effectively just a UI glitch in the wrong hosts
being listed as valid targets from the source side.
Change-Id: I6907a5a1c3c66920e5ec71601c044e722e7cb888
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
And add a CLI/localapi and c2n mechanism to enable it for a fixed
amount of time.
Updates #1548
Change-Id: I71674aaf959a9c6761ff33bbf4a417ffd42195a7
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Most visible when using tsnet.Server, but could have resulted in dropped
messages in a few other places too.
Fixes#5743
Signed-off-by: Mihai Parparita <mihai@tailscale.com>
