In very low-latency conditions, a STUN request can complete before
the startup loop has finished firing off goroutines, leading to
a concurrent map mutation.
This avoids the server blocking on misbehaving or heavily contended
clients. We attempt to drop from the head of the queue to keep
overall queueing time lower.
Also:
- fixes server->client keepalives, which weren't happening.
- removes read rate-limiter, deferring instead to kernel-level
global limiter/fair queuer.
Signed-off-by: David Anderson <dave@natulte.net>
Breaks something deep in wireguard or magicsock's brainstem, no packets at all
can flow. All received packets fail decryption with "invalid mac1".
This reverts commit 94024355ed.
Signed-off-by: David Anderson <dave@natulte.net>
Handler is like http.Handler, but returns errors. ErrHandler
converts back to an http.Handler, with added error handling
and logging.
Signed-off-by: David Anderson <dave@natulte.net>
This lets us publish sets of vars that are breakdowns along one
dimension in a format that Prometheus and Grafana natively know
how to do useful things with.
Signed-off-by: David Anderson <dave@natulte.net>
On BSD, /var/db is what linux calls /var/lib.
On modern linux, /run and /var/run are the same directory, but
on BSD the correct path is /var/run, so use that.
Fixes#79
Signed-off-by: David Anderson <dave@natulte.net>
Both RPM and Deb require us to specify both Replaces and Conflicts:
Conflicts tells them that the packages cannot coexist on the system,
Replaces tells them which one to keep.
Two commands for now, `up` and `netcheck`. The commands and the flags they take
will change a bunch in the future, but this is good enough to launch on parity
with relaynode.
Signed-Off-By: David Anderson <dave@natulte.net>
Prefs has become a heavy object with non-memcpy copy
semantics. We should not pass such a thing by value.
Signed-off-by: David Anderson <dave@natulte.net>
staticcheck defaults to running with no tags set, which only
works if redo hasn't run and generated ver.go. If it has,
we end up with a redeclaration conflict.
Signed-off-by: David Anderson <dave@natulte.net>
The autoselection should pick sensible paths for all of:
- Windows (LocalAppData)
- Mac (Library/Caches)
- Unix user (XDG_CACHE_DIR)
- Linux systemd service (CACHE_DIRECTORY)
As a last resort, if cache dir lookup fails, plops sufficiently
uniquely named files into the current working directory.
Signed-off-by: David Anderson <dave@natulte.net>
We can't rely on a frontend to provide a control
server URL, so this naturally belongs in server-persisted
state.
Signed-off-by: David Anderson <dave@natulte.net>
Port number has to be by itself for substitution to work.
Disabling the restart rate-limiting has to be in [Unit] not
[Service].
Signed-off-by: David Anderson <dave@natulte.net>
On unix, we want to provide a full path to the desired unix socket.
On windows, currently we want to provide a TCP port, but someday
we'll also provide a "path-ish" object for a named pipe.
For now, simplify the API down to exactly a path and a TCP port.
Signed-off-by: David Anderson <dave@natulte.net>
We need iptables to make subnet routing work. Without it,
Tailscale mostly works, but subnet routing mysteriously doesn't.
Signed-off-by: David Anderson <dave@natulte.net>
This is a prelude to supporting relaynode's --routes in
tailscaled. The daemon needs to remembers routes to
advertise, and the CLI needs to be able to change the
set of advertised routes. Prefs is the thing used for
both of these.
Signed-off-by: David Anderson <dave@natulte.net>
With this change, tailscaled can be restarted and reconnect
without interaction from `tailscale`, and `tailscale` is merely
there to provide login assistance and adjust preferences.
Signed-off-by: David Anderson <dave@natulte.net>
It was previously used by the MacOS client, but it now does
something different. ipnserver should never obey a client's
request to exit.
Signed-off-by: David Anderson <dave@natulte.net>
In particular, the Dup2 syscall is not defined in the syscall
package on arm64, but is defined in x/sys/unix.
Signed-off-by: David Anderson <dave@natulte.net>
The store is passed-in by callers of NewLocalBackend and
ipnserver.Run, but currently all callers are hardcoded to
an in-memory store. The store is unused.
Signed-Off-By: David Anderson <dave@natulte.net>
This is a prelude to making it truly optional, once state
management has moved into the backend. For now though, it's
still required. This change is just isolating the bubbling-up
of the pointerification into other layers.
Signed-Off-By: David Anderson <dave@natulte.net>
- It was only used in one currently-unused client.
- It's an imperative command, not a configuration setting.
- The LoginFlags stuff in controlclient feels like it needs
a refactor anyway.
I'll put this logic back once ipnd owns its state and Backend
commands reflect that.
Signed-Off-By: David Anderson <dave@natulte.net>
The MTU is currently set when creating the tun device,
elsewhere in the code. Maybe someday we'll want some kind
of per-platform MTU configuration here, but not in the
short-medium term.
Signed-off-by: David Anderson <dave@natulte.net>
It seems changelog generation got broken by moving the code between
corp and OSS repos, because one of the commit SHAs doesn't have an
associated tag.
In the interest of fixing the build, and because we're not yet trying
to upstream the debian package, I fixed this by allowing hash-based
versions to show up in the changelog. This maybe wrong from a
debian standards perspective, but for our current point in life
it'll work until we learn to do it better.
Signed-Off-By: David Anderson <dave@natulte.net>
The linter is strictly correct, but the code is structured
this way to avoid variable shadowing problems in the following
for loop. The context doesn't leak.
Staticcheck is correctly pointing out that this code is hard to
follow. However, this chunk of code is in service of enforcing
one frontend <> one backend, and we want to remove that limitation.
So, we'll just ignore the lint warning until this entire piece of
code goes away.
Signed-off-by: David Anderson <dave@natulte.net>
Looks like Github doesn't understand mailto: links, so
we'll have to create KB pages on tailscale.com and point
to those.
Signed-off-by: David Anderson <dave@natulte.net>
The function is indeed unused on linux, but we want it around
so that unit tests exercise the netstat logic even on linux,
even if it's only used on !linux.
Signed-off-by: David Anderson <dave@natulte.net>
David Anderson
