When a DNS server claims to be unable or unwilling to handle a request,
instead of passing that refusal along to the client, just treat it as
any other error trying to connect to the DNS server. This prevents DNS
requests from failing based on if a server can respond with a transient
error before another server is able to give an actual response. DNS
requests only failing *sometimes* is really hard to find the cause of
(#1033).
Signed-off-by: Smitty <me@smitop.com>
"skipping portmap; gateway range likely lacks support" is really
spammy on cloud systems, and not very useful in debugging.
Fixes https://github.com/tailscale/tailscale/issues/3034
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
We added the initial handling only for macOS and iOS.
With 1.16.0 now released, suppress forwarding DNS-SD
on all platforms to test it through the 1.17.x cycle.
Updates #2442
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
On iOS (and possibly other platforms), sometimes our UDP socket would
get stuck in a state where it was bound to an invalid interface (or no
interface) after a network reconfiguration. We can detect this by
actually checking the error codes from sending our STUN packets.
If we completely fail to send any STUN packets, we know something is
very broken. So on the next STUN attempt, let's rebind the UDP socket
to try to correct any problems.
This fixes a problem where iOS would sometimes get stuck using DERP
instead of direct connections until the backend was restarted.
Fixes#2994
Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
I forgot to include this file in the earlier
7cf8ec8108 commit.
This exists purely to keep "go mod tidy" happy.
Updates #1609
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We still try the host's x509 roots first, but if that fails (like if
the host is old), we fall back to using LetsEncrypt's root and
retrying with that.
tlsdial was used in the three main places: logs, control, DERP. But it
was missing in dnsfallback. So added it there too, so we can run fine
now on a machine with no DNS config and no root CAs configured.
Also, move SSLKEYLOGFILE support out of DERP. tlsdial is the logical place
for that support.
Fixes#1609
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
DNSSEC is an availability issue, as recently demonstrated by the
Slack issue, with limited security advantage. DoH on the other hand
is a critical security upgrade. This change adds DoH support for the
non-DNSSEC endpoints of Quad9.
https://www.quad9.net/service/service-addresses-and-features#unsec
Signed-off-by: Filippo Valsorda <hi@filippo.io>
It was in the wrong filter direction before, per CPU profiles
we now have.
Updates #1526 (maybe fixes? time will tell)
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Windows has a public dns.Flush used in router_windows.go.
However that won't work for platforms like Linux, where
we need a different flush mechanism for resolved versus
other implementations.
We're instead adding a FlushCaches method to the dns Manager,
which can be made to work on all platforms as needed.
Fixes https://github.com/tailscale/tailscale/issues/2132
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
The earlier 382b349c54 was too late,
as engine creation itself needed to listen on things.
Fixes#2827
Updates #2822
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We currently plumb full URLs for DNS resolvers from the control server
down to the client. But when we pass the values into the net/dns
package, we throw away any URL that isn't a bare IP. This commit
continues the plumbing, and gets the URL all the way to the built in
forwarder. (It stops before plumbing URLs into the OS configurations
that can handle them.)
For #2596
Signed-off-by: David Crawshaw <crawshaw@tailscale.com>
Reported on IRC: in an edge case, you can end up with a directManager DNS
manager and --accept-dns=false, in which case we should do nothing, but
actually end up restarting resolved whenever the netmap changes, even though
the user told us to not manage DNS.
Signed-off-by: David Anderson <danderson@tailscale.com>
Reported on IRC: a resolv.conf that contained two entries for
"nameserver 127.0.0.53", which defeated our "is resolved actually
in charge" check. Relax that check to allow any number of nameservers,
as long as they're all 127.0.0.53.
Signed-off-by: David Anderson <danderson@tailscale.com>
It wasn't using the right metric. Apparently you're supposed to sum the route
metric and interface metric. Whoops.
While here, optimize a few little things too, not that this code
should be too hot.
Fixes#2707 (at least; probably dups but I'm failing to find)
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Now that we have the easier-to-parse go:build build tags,
it is straightforward to simplify them. Yay.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Mostly so the Linux one can use Linux-specific stuff in package
syscall and not use os/exec for uname for portability.
But also it helps deps a tiny bit on iOS.
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
This logs some basic statistics for UPnP, so that tailscale can better understand what routers
are being used and how to connect to them.
Signed-off-by: julianknodt <julianknodt@gmail.com>
This adds a PCP test to the IGD test server, by hardcoding in a few observed packets from
Denton's box.
Signed-off-by: julianknodt <julianknodt@gmail.com>
And use dynamic port numbers in tests, as Linux on GitHub Actions and
Windows in general have things running on these ports.
Co-Author: Julian Knodt <julianknodt@gmail.com>
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Previously, we hashed the question and combined it with the original
txid which was useful when concurrent queries were multiplexed on a
single local source port. We encountered some situations where the DNS
server canonicalizes the question in the response (uppercase converted
to lowercase in this case), which resulted in responses that we couldn't
match to the original request due to hash mismatches. This includes a
new test to cover that situation.
Fixes#2597
Signed-off-by: Adrian Dewhurst <adrian@tailscale.com>
PCP handles external IPs by allowing the client to specify them in the packet, which is more
explicit than requiring 2 packets from PMP, so allow for future changes to add it in easily.
Signed-off-by: julianknodt <julianknodt@gmail.com>
Still very much a prototype (hard-coded IPs, etc) but should be
non-invasive enough to submit at this point and iterate from here.
Updates #2589
Co-Author: David Crawshaw <crawshaw@tailscale.com>
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Prior to Tailscale 1.12 it detected UPnP on any port.
Starting with Tailscale 1.11.x, it stopped detecting UPnP on all ports.
Then start plumbing its discovered Location header port number to the
code that was assuming port 5000.
Fixes#2109
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
There's a call to Now once per packet.
Move to mono.Now.
Though the current implementation provides high precision,
we document it to be coarse, to preserve the ability
to switch to a coarse monotonic time later.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Go 1.17 switches to a register ABI on amd64 platforms.
Part of that switch is that go and defer calls use an argument-less
closure, which allocates. This means that we have an extra
alloc in some DNS work. That's unfortunate but not a showstopper,
and I don't see a clear path to fixing it.
The other performance benefits from the register ABI will all
but certainly outweigh this extra alloc.
Fixes#2545
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
I don't know how to get access to a real packet. Basing this commit
entirely off:
+------------+--------------+------------------------------+
| Field Name | Field Type | Description |
+------------+--------------+------------------------------+
| NAME | domain name | MUST be 0 (root domain) |
| TYPE | u_int16_t | OPT (41) |
| CLASS | u_int16_t | requestor's UDP payload size |
| TTL | u_int32_t | extended RCODE and flags |
| RDLEN | u_int16_t | length of all RDATA |
| RDATA | octet stream | {attribute,value} pairs |
+------------+--------------+------------------------------+
From https://datatracker.ietf.org/doc/html/rfc6891#section-6.1.2
Signed-off-by: David Crawshaw <crawshaw@tailscale.com>
The handoff between tstun.Wrap's Read and poll methods
is one of the per-packet hotspots. It shows up in pprof.
Making outbound buffered increases throughput.
It is hard to measure exactly how much, because the numbers
are highly variable, but I'd estimate it at about 1%,
using the best observed max throughput across three runs.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
The handoff between tstun.Wrap's Read and poll methods
is one of the per-packet hotspots. It shows up in pprof.
Making outbound buffered increases throughput.
It is hard to measure exactly how much, because the numbers
are highly variable, but I'd estimate it at about 1%,
using the best observed max throughput across three runs.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Tested manually with:
$ go test -v ./net/dnscache/ -dial-test=bogusplane.dev.tailscale.com:80
Where bogusplane has three A records, only one of which works.
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Instead of blasting away at all upstream resolvers at the same time,
make a timing plan upon reconfiguration and have each upstream have an
associated start delay, depending on the overall forwarding config.
So now if you have two or four upstream Google or Cloudflare DNS
servers (e.g. two IPv4 and two IPv6), we now usually only send a
query, not four.
This is especially nice on iOS where we start fewer DoH queries and
thus fewer HTTP/1 requests (because we still disable HTTP/2 on iOS),
fewer sockets, fewer goroutines, and fewer associated HTTP buffers,
etc, saving overall memory burstiness.
Fixes#2436
Updates tailscale/corp#2250
Updates tailscale/corp#2238
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Add a place to hang state in a future change for #2436.
For now this just simplifies the send signature without
any functional change.
Updates #2436
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Previously, this was incorrectly returning the internal port, and using that with the external
exposed IP when it did not use WANIPConnection2. In the case when we must provide a port, we
return it instead.
Noticed this while implementing the integration test for upnp.
Signed-off-by: julianknodt <julianknodt@gmail.com>
It was a huge chunk of the overall log output and made debugging
difficult. Omit and summarize the spammy *.arpa parts instead.
Fixestailscale/corp#2066 (to which nobody had opinions, so)
Add in UPnP portmapping, using goupnp library in order to get the UPnP client and run the
portmapping functions. This rips out anywhere where UPnP used to be in portmapping, and has a
flow separate from PMP and PCP.
RELNOTE=portmapper now supports UPnP mappings
Fixes#682
Updates #2109
Signed-off-by: julianknodt <julianknodt@gmail.com>
Recognize Cloudflare, Google, Quad9 which are by far the
majority of upstream DNS servers that people use.
RELNOTE=MagicDNS now uses DNS-over-HTTPS when querying popular upstream resolvers,
so DNS queries aren't sent in the clear over the Internet.
Updates #915 (might fix it?)
Updates #988 (gets us closer, if it fixes Android)
Updates #74 (not yet configurable, but progress)
Updates #2056 (not yet configurable, dup of #74?)
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Added the net/speedtest package that contains code for starting up a
speedtest server and a client. The speedtest command for starting a
client takes in a duration for the speedtest as well as the host and
port of the speedtest server to connect to. The speedtest command for
starting a server takes in a host:port pair to listen on.
Signed-off-by: Aaditya Chaudhary <32117362+AadityaChaudhary@users.noreply.github.com>
With netns handling localhost now, existing tests no longer
need special handling. The tests set up their connections to
localhost, and the connections work without fuss.
Remove the special handling for tests.
Also remove the hostinfo.TestCase support, since this was
the only use of it. It can be added back later if really
needed, but it would be better to try to make tests work
without special cases.
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
netns_linux checked whether "ip rule" could run to determine
whether to use SO_MARK for network namespacing. However in
Linux environments which lack CAP_NET_ADMIN, such as various
container runtimes, the "ip rule" command succeeds but SO_MARK
fails due to lack of permission. SO_BINDTODEVICE would work in
these environments, but isn't tried.
In addition to running "ip rule" check directly whether SO_MARK
works or not. Among others, this allows Microsoft Azure App
Service and AWS App Runner to work.
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
Connections to a control server or log server on localhost,
used in a number of tests, are working right now because the
calls to SO_MARK in netns fail for non-root but then we ignore
the failure when running in tests.
Unfortunately that failure in SO_MARK also affects container
environments without CAP_NET_ADMIN, breaking Tailscale
connectivity. We're about to fix netns to recognize when SO_MARK
doesn't work and use SO_BINDTODEVICE instead. Doing so makes
tests fail, as their sockets now BINDTODEVICE of the default
route and cannot connect to localhost.
Add support to skip namespacing for localhost connections,
which Darwin and Windows already do. This is not conditional
on running within a test, if you tell tailscaled to connect
to localhost it will automatically use a non-namespaced
socket to do so.
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
