For various reasons (mostly during rollouts or config changes on our
side), nodes may end up connecting to a fallback DERP node in a
region, rather than the primary one we tell them about in the DERP
map.
Connecting to the "wrong" node is fine, but it's in our best interest
for all nodes in a domain to connect to the same node, to reduce
intra-region packet forwarding.
This adds a privileged frame type used by the control system that can
kick off a client connection when they're connected to the wrong node
in a region. Then they hopefully reconnect immediately to the correct
location. (If not, we can leave them alone and stop closing them.)
Updates tailscale/corp#372
The remove hook implementation was copy/pasted from the line above and
I didn't change the body, resulting in packet forwarding routes never
being removed.
Fortunately we weren't using this path yet, but it led to stats being
off, and (very) slow memory growth.
Later we'll want to use the presence of a discovery key as a signal
that the node knows how to participate in discovery. Currently the
code generates keys and sends them to the control server but doesn't
do anything with them, which is a bad state to stay in lest we release
this code and end up with nodes in the future that look like they're
functional with the new discovery protocol but aren't.
So for now, make this opt-in as a debug option for now, until the rest
of it is in.
Updates #483
The magicsock derpReader was holding onto 65KB for each DERP
connection forever, just in case.
Make the derp{,http}.Client be in charge of memory instead. It can
reuse its bufio.Reader buffer space.
... it was crashing for some reason, running out of stack while
loading a DLL in goversion. I don't understand Windows (or the Go
runtime for Windows) enough to know why that'd be problematic in that
context.
In any case, don't call it, as tryFixLogStateLocation does nothing on
Windows anyway.
tryFixLogStateLocation should probably just call version.CmdName
itself if/when it needs to, after the GOOS check.
(The NewMeshClient constructor I added recently was gross in
retrospect at call sites, especially when it wasn't obvious that a
meshKey empty string meant a regular client)
This lets a trusted DERP client that knows a pre-shared key subscribe
to the connection list. Upon subscribing, they get the current set
of connected public keys, and then all changes over time.
This lets a set of DERP server peers within a region all stay connected to
each other and know which clients are connected to which nodes.
Updates #388
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
It'll be called a bunch, so worth a bit of effort. Could go further, but not yet.
(really, should hook into wgengine/monitor and only re-read on netlink changes?)
name old time/op new time/op delta
DefaultRouteInterface-8 60.8µs ±11% 44.6µs ± 5% -26.65% (p=0.000 n=20+19)
name old alloc/op new alloc/op delta
DefaultRouteInterface-8 3.29kB ± 0% 0.55kB ± 0% -83.21% (p=0.000 n=20+20)
name old allocs/op new allocs/op delta
DefaultRouteInterface-8 9.00 ± 0% 6.00 ± 0% -33.33% (p=0.000 n=20+20)
My earlier 3fa58303d0 tried to implement
the net/http.Tranhsport.DialTLSContext hook, but I didn't return a
*tls.Conn, so we ended up sending a plaintext HTTP request to an HTTPS
port. The response ended up being Go telling as such, not the
/derp/latency-check handler's response (which is currently still a
404). But we didn't even get the 404.
This happened to work well enough because Go's built-in error response
was still a valid HTTP response that we can measure for timing
purposes, but it's not a great answer. Notably, it means we wouldn't
be able to get a future handler to run server-side and count those
latency requests.
tailscale netcheck was broken otherwise.
We can fix this a better way later; I'm just fixing a regression in
some way because I'm trying to work on netcheck at the moment.
Also:
* add -verbose flag to cmd/tailscale netcheck
* remove some API from the interfaces package
* convert some of the interfaces package to netaddr.IP
* don't even send IPv4 probes on machines with no IPv4 (or only v4
loopback)
* and once three regions have replied, stop waiting for other probes
at 2x the slowest duration.
Updates #376
... and thus does not need to worry about when it escapes into
unprovable fmt interface{} land.
Also, add some convenience methods for efficiently writing integers.
Instead of hard-coding the DERP map (except for cmd/tailscale netcheck
for now), get it from the control server at runtime.
And make the DERP map support multiple nodes per region with clients
picking the first one that's available. (The server will balance the
order presented to clients for load balancing)
This deletes the stunner package, merging it into the netcheck package
instead, to minimize all the config hooks that would've been
required.
Also fix some test flakes & races.
Fixes#387 (Don't hard-code the DERP map)
Updates #388 (Add DERP region support)
Fixes#399 (wgengine: flaky tests)
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We don't want those extra dependencies on iOS, at least yet.
Especially since there's no way to set the relevant environment
variables so it's just bloat with no benefits. Perhaps we'll need to
do SOCKS on iOS later, but probably differently if/when so.
Updates #227
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
For "tailscale status" on macOS (from separately downloaded
cmd/tailscale binary against App Store IPNExtension).
(This isn't all of it, but I've had this sitting around uncommitted.)
If Australia's far away and not going to be used, it's still going to
be far away a minute later. No need to send backup
just-in-case-UDP-gets-lost STUN packets to the known far away
destinations. Those are the ones most likely to trigger retries due to
delay anyway (in random 50-250ms, currently). But we'll keep sending 1
packet to them, just in case our airplane landed.
Likewise, be less aggressive with IPv6. The main point is just to see
whether IPv6 works. No need to send up to 10 packets every round. Max
two is enough (except for the first round). This does mean our STUN
traffic graphs for IPv4-vs-IPv6 will change shape. Oh well. It was a
weird eyeball metric for IPv6 connectivity anyway and we have better
metrics.
We can tweak this policy over time. It's factored out and has tests
now.
This was the last of the three places that do TLS from clients (logs,
control, derp). With this, iOS should be able to use the
memory-efficient x509 root CertPool.
I removed the HTTPC field in b6fa5a69be but it was apparently still
used in [oss-skipped] tests.
Restore it, but name it so it's more obvious that it's only for
tests. (It currently is, and I'd like to keep it like that for now.)
Brad Fitzpatrick
