Our prod code doesn't eagerly handshake, because our disco layer enables
on-demand handshaking. Configuring both peers to eagerly handshake leads
to WireGuard handshake races that make TestTwoDevicePing flaky.
Signed-off-by: David Anderson <danderson@tailscale.com>
It only existed to override one test-only behavior with a
different test-only behavior, in both cases working around
an annoying feature of our CI environments. Instead, handle
that weirdness entirely in the test code, with a tweaked
TestOnlyPacketListener that gets injected.
Signed-off-by: David Anderson <danderson@tailscale.com>
The docstring said it was meant for use in tests, but it's specifically a
special codepath that is _only_ used in tests, so make the claim stronger.
Signed-off-by: David Anderson <danderson@tailscale.com>
Instead of using the legacy codepath, teach discoEndpoint to handle
peers that have a home DERP, but no disco key. We can still communicate
with them, but only over DERP.
Signed-off-by: David Anderson <danderson@tailscale.com>
magicsock makes multiple calls to Now per packet.
Move to mono.Now. Changing some of the calls to
use package mono has a cascading effect,
causing non-per-packet call sites to also switch.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
logBufWriter had no serialization.
It just so happens that none of its users currently ever log concurrently.
Make it safe for concurrent use.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
The DERPTestPort int meant two things before: which port to use, and
whether to disable TLS verification. Users would like to set the port
without disabling TLS, so break it into two options.
Updates #1264
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
After allowing for custom DERP maps, it's convenient to be able to see their latency in
netcheck. This adds a query to the local tailscaled for the current DERPMap.
Updates #1264
Signed-off-by: julianknodt <julianknodt@gmail.com>
Pull in the latest version of wireguard-windows.
Switch to upstream wireguard-go.
This requires reverting all of our import paths.
Unfortunately, this has to happen at the same time.
The wireguard-go change is very low risk,
as that commit matches our fork almost exactly.
(The only changes are import paths, CI files, and a go.mod entry.)
So if there are issues as a result of this commit,
the first place to look is wireguard-windows changes.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
magicsock.Conn.ParseEndpoint requires a peer's public key,
disco key, and legacy ip/ports in order to do its job.
We currently accomplish that by:
* adding the public key in our wireguard-go fork
* encoding the disco key as magic hostname
* using a bespoke comma-separated encoding
It's a bit messy.
Instead, switch to something simpler: use a json-encoded struct
containing exactly the information we need, in the form we use it.
Our wireguard-go fork still adds the public key to the
address when it passes it to ParseEndpoint, but now the code
compensating for that is just a couple of simple, well-commented lines.
Once this commit is in, we can remove that part of the fork
and remove the compensating code.
Signed-off-by: Josh Bleecher Snyder <josharian@gmail.com>
For historical reasons, we ended up with two near-duplicate
copies of curve25519 key types, one in the wireguard-go module
(wgcfg) and one in the tailscale module (types/wgkey).
Then we moved wgcfg to the tailscale module.
We can now remove the wgcfg key type in favor of wgkey.
Signed-off-by: Josh Bleecher Snyder <josharian@gmail.com>
Track endpoints internally with a new tailcfg.Endpoint type that
includes a typed netaddr.IPPort (instead of just a string) and
includes a type for how that endpoint was discovered (STUN, local,
etc).
Use []tailcfg.Endpoint instead of []string internally.
At the last second, send it to the control server as the existing
[]string for endpoints, but also include a new parallel
MapRequest.EndpointType []tailcfg.EndpointType, so the control server
can start filtering out less-important endpoint changes from
new-enough clients. Notably, STUN-discovered endpoints can be filtered
out from 1.6+ clients, as they can discover them amongst each other
via CallMeMaybe disco exchanges started over DERP. And STUN endpoints
change a lot, causing a lot of MapResposne updates. But portmapped
endpoints are worth keeping for now, as they they work right away
without requiring the firewall traversal extra RTT dance.
End result will be less control->client bandwidth. (despite negligible
increase in client->control bandwidth)
Updates tailscale/corp#1543
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
We don't use the port that wireguard-go passes to us (via magicsock.connBind.Open).
We ignore it entirely and use the port we selected.
When we tell wireguard-go that we're changing the listen_port,
it calls connBind.Close and then connBind.Open.
And in the meantime, it stops calling the receive functions,
which means that we stop receiving and processing UDP and DERP packets.
And that is Very Bad.
That was never a problem prior to b3ceca1dd7,
because we passed the SkipBindUpdate flag to our wireguard-go fork,
which told wireguard-go not to re-bind on listen_port changes.
That commit eliminated the SkipBindUpdate flag.
We could write a bunch of code to work around the gap.
We could add background readers that process UDP and DERP packets when wireguard-go isn't.
But it's simpler to never create the conditions in which wireguard-go rebinds.
The other scenario in which wireguard-go re-binds is device.Down.
Conveniently, we never call device.Down. We go from device.Up to device.Close,
and the latter only when we're shutting down a magicsock.Conn completely.
Rubber-ducked-by: Avery Pennarun <apenwarr@tailscale.com>
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Upstream wireguard-go has changed its receive model.
NewDevice now accepts a conn.Bind interface.
The conn.Bind is stateless; magicsock.Conns are stateful.
To work around this, we add a connBind type that supports
cheap teardown and bring-up, backed by a Conn.
The new conn.Bind allows us to specify a set of receive functions,
rather than having to shoehorn everything into ReceiveIPv4 and ReceiveIPv6.
This lets us plumbing DERP messages directly into wireguard-go,
instead of having to mux them via ReceiveIPv4.
One consequence of the new conn.Bind layer is that
closing the wireguard-go device is now indistinguishable
from the routine bring-up and tear-down normally experienced
by a conn.Bind. We thus have to explicitly close the magicsock.Conn
when the close the wireguard-go device.
One downside of this change is that we are reliant on wireguard-go
to call receiveDERP to process DERP messages. This is fine for now,
but is perhaps something we should fix in the future.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
The code is not obviously better or worse, but this makes the little warning
triangle in my editor go away, and the distraction removal is worth it.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
It can end up executing an a new goroutine,
at which point instead of immediately stopping test execution, it hangs.
Since this is unexpected anyway, panic instead.
As a bonus, it makes call sites nicer and removes a kludge comment.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
The tstun packagen contains both constructors for generic tun
Devices, and a wrapper that provides additional functionality.
Signed-off-by: David Anderson <danderson@tailscale.com>
There was a logical race where Conn.Rebind could acquire the
RebindingUDPConn mutex, close the connection, fail to rebind, release
the mutex, and then because the mutex was no longer held, ReceiveIPv4
wouldn't retry reads that failed with net.ErrClosed, letting that
error back to wireguard-go, which would then stop running that receive
IP goroutine.
Instead, keep the RebindingUDPConn mutex held for the entirety of the
replacement in all cases.
Updates tailscale/corp#1289
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
UIs need to see the full unedited netmap in order to know what exit nodes they
can offer to the user.
Signed-off-by: David Anderson <danderson@tailscale.com>
* move probing out of netcheck into new net/portmapper package
* use PCP ANNOUNCE op codes for PCP discovery, rather than causing
short-lived (sub-second) side effects with a 1-second-expiring map +
delete.
* track when we heard things from the router so we can be less wasteful
in querying the router's port mapping services in the future
* use portmapper from magicsock to map a public port
Fixes#1298Fixes#1080Fixes#1001
Updates #864
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
When a handshake race occurs, a queued data packet can get lost.
TestTwoDevicePing expected that the very first data packet would arrive.
This caused occasional flakes.
Change TestTwoDevicePing to repeatedly re-send packets
and succeed when one of them makes it through.
This is acceptable (vs making WireGuard not drop the packets)
because this only affects communication with extremely old clients.
And those extremely old clients will eventually connect,
because the kernel will retry sends on timeout.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
We modified the standard net package to not allocate a *net.UDPAddr
during a call to (*net.UDPConn).ReadFromUDP if the caller's use
of the *net.UDPAddr does not cause it to escape.
That is https://golang.org/cl/291390.
This is the companion change to magicsock.
There are two changes required.
First, call ReadFromUDP instead of ReadFrom, if possible.
ReadFrom returns a net.Addr, which is an interface, which always allocates.
Second, reduce the lifetime of the returned *net.UDPAddr.
We do this by immediately converting it into a netaddr.IPPort.
We left the existing RebindingUDPConn.ReadFrom method in place,
as it is required to satisfy the net.PacketConn interface.
With the upstream change and both of these fixes in place,
we have removed one large allocation per packet received.
name old time/op new time/op delta
ReceiveFrom-8 16.7µs ± 5% 16.4µs ± 8% ~ (p=0.310 n=5+5)
name old alloc/op new alloc/op delta
ReceiveFrom-8 112B ± 0% 64B ± 0% -42.86% (p=0.008 n=5+5)
name old allocs/op new allocs/op delta
ReceiveFrom-8 3.00 ± 0% 2.00 ± 0% -33.33% (p=0.008 n=5+5)
Co-authored-by: Sonia Appasamy <sonia@tailscale.com>
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
addrSet maintained duplicate lists of netaddr.IPPorts and net.UDPAddrs.
Unify to use the netaddr type only.
This makes (*Conn).ReceiveIPvN a bit uglier,
but that'll be cleaned up in a subsequent commit.
This is preparatory work to remove an allocation from ReceiveIPv4.
Co-authored-by: Sonia Appasamy <sonia@tailscale.com>
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
I based my estimation of the required timeout based on locally
observed behavior. But CI machines are worse than my local machine.
16s was enough to reduce flakiness but not eliminate it. Bump it up again.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
We removed the "fast retry" code from our wireguard-go fork.
As a result, pings can take longer to transit when retries are required.
Allow that.
Fixes#1277
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
The fix can make this test run unconditionally.
This moves code from 5c619882bc for
testability but doesn't fix it yet. The #1282 problem remains (when I
wrote its wake-up mechanism, I forgot there were N DERP readers
funneling into 1 UDP reader, and the code just isn't correct at all
for that case).
Also factor out some test helper code from BenchmarkReceiveFrom.
The refactoring in magicsock.go for testability should have no
behavior change.
Magicsock started dropping all traffic internally when Tailscale is
shut down, to avoid spurious wireguard logspam. This made the benchmark
not receive anything. Setting a dummy private key is sufficient to get
magicsock to pass traffic for benchmarking purposes.
Fixes#1270.
Signed-off-by: David Anderson <danderson@tailscale.com>
Use tb.Cleanup to simplify both the API and the implementation.
One behavior change: When the number of goroutines shrinks, don't log.
I've never found these logs to be useful, and they frequently add noise.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
This is mostly code movement from the wireguard-go repo.
Most of the new wgcfg package corresponds to the wireguard-go wgcfg package.
wgengine/wgcfg/device{_test}.go was device/config{_test}.go.
There were substantive but simple changes to device_test.go to remove
internal package device references.
The API of device.Config (now wgcfg.DeviceConfig) grew an error return;
we previously logged the error and threw it away.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
Rewrite log lines on the fly, based on the set of known peers.
This enables us to use upstream wireguard-go logging,
but maintain the Tailscale-style peer public key identifiers
that the rest of our systems (and people) expect.
Fixes#1183
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
22507adf54 stopped relying on
our fork of wireguard-go's UpdateDst callback.
As a result, we can unwind that code,
and the extra return value of ReceiveIPv{4,6}.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
TwoDevicePing is explicitly testing the behavior of the legacy codepath, everything
else is happy to assume that code no longer exists.
Signed-off-by: David Anderson <danderson@tailscale.com>
Previously, this benchmark relied on behavior of the legacy
receive codepath, which I changed in 22507adf. With this
change, the benchmark instead relies on the new active discovery
path.
Signed-off-by: David Anderson <danderson@tailscale.com>
This eliminates a dependency on wgcfg.Endpoint,
as part of the effort to eliminate our wireguard-go fork.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
This makes connectivity between ancient and new tailscale nodes slightly
worse in some cases, but only in cases where the ancient version would
likely have failed to get connectivity anyway.
Signed-off-by: David Anderson <danderson@tailscale.com>
In sendDiscoMessage there is a check of whether the connection is
closed, which is not being reliably exercised by other tests.
This shows up in code coverage reports, the lines of code in
sendDiscoMessage are alternately added and subtracted from
code coverage.
Add a test to specifically exercise and verify this code path.
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
In derpWriteChanOfAddr when we call derphttp.NewRegionClient(),
there is a check of whether the connection is already errored and
if so it returns before grabbing the lock. The lock might already
be held and would be a deadlock.
This corner case is not being reliably exercised by other tests.
This shows up in code coverage reports, the lines of code in
derpWriteChanOfAddr are alternately added and subtracted from
code coverage.
Add a test to specifically exercise this code path, and verify that
it doesn't deadlock.
This is the best tradeoff I could come up with:
+ the moment code calls Err() to check if there is an error, we
grab the lock to make sure it would deadlock if it tries to grab
the lock itself.
+ if a new call to Err() is added in this code path, only the
first one will be covered and the rest will not be tested.
+ this test doesn't verify whether code is checking for Err() in
the right place, which ideally I guess it would.
Signed-off-by: Denton Gentry <dgentry@tailscale.com>
This is a replacement for the key-related parts
of the wireguard-go wgcfg package.
This is almost a straight copy/paste from the wgcfg package.
I have slightly changed some of the exported functions and types
to avoid stutter, added and tweaked some comments,
and removed some now-unused code.
To avoid having wireguard-go depend on this new package,
wgcfg will keep its key types.
We translate into and out of those types at the last minute.
These few remaining uses will be eliminated alongside
the rest of the wgcfg package.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
We still use the packet.* alloc-free types in the data path, but
the compilation from netaddr to packet happens within the filter
package.
Signed-off-by: David Anderson <danderson@tailscale.com>
Updates #654. See that issue for a discussion of why
this timeout reduces flakiness, and what next steps are.
Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>
1) we weren't waking up a discoEndpoint that once existed and
went idle for 5 minutes and then got a disco message again.
2) userspaceEngine.noteReceiveActivity had a buggy check; fixed
and added a test
There is a race in natlab where we might start shutdown while natlab is still running
a goroutine or two to deliver packets. This adds a small grace period to try and receive
it before continuing shutdown.
Signed-off-by: David Anderson <danderson@tailscale.com>
The first packet to transit may take several seconds to do so, because
setup rates in wgengine may result in the initial WireGuard handshake
init to get dropped. So, we have to wait at least long enough for a
retransmit to correct the fault.
Signed-off-by: David Anderson <danderson@tailscale.com>
Active discovery lets us introspect the state of the network stack precisely
enough that it's unnecessary, and dropping the initial DERP packets greatly
slows down tests. Additionally, it's unrealistic since our production network
will never deliver _only_ discovery packets, it'll be all or nothing.
Signed-off-by: David Anderson <danderson@tailscale.com>
Uses natlab only, because the point of this active discovery test is going to be
that it should get through a lot of obstacles.
Signed-off-by: David Anderson <danderson@tailscale.com>
wireguard-go uses 3 goroutines per peer (with reasonably large stacks
& buffers).
Rather than tell wireguard-go about all our peers, only tell it about
peers we're actively communicating with. That means we need hooks into
magicsock's packet receiving path and tstun's packet sending path to
lazily create a wireguard peer on demand from the network map.
This frees up lots of memory for iOS (where we have almost nothing
left for larger domains with many users).
We should ideally do this in wireguard-go itself one day, but that'd
be a pretty big change.
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
The new interface lets implementors more precisely distinguish
local traffic from forwarded traffic, and applies different
forwarding logic within Machines for each type. This allows
Machines to be packet forwarders, which didn't quite work
with the implementation of Inject.
Signed-off-by: David Anderson <danderson@tailscale.com>
The test demonstrates that magicsock can traverse two stateful
firewalls facing each other, that each require localhost to
initiate connections.
Signed-off-by: David Anderson <danderson@tailscale.com>
This adds a new magicsock endpoint type only used when both sides
support discovery (that is, are advertising a discovery
key). Otherwise the old code is used.
So far the new code only communicates over DERP as proof that the new
code paths are wired up. None of the actually discovery messaging is
implemented yet.
Support for discovery (generating and advertising a key) are still
behind an environment variable for now.
Updates #483
And track known peers.
Doesn't yet do anything with the messages. (nor does it send any yet)
Start of docs on the message format. More will come in subsequent changes.
Updates #483
This allows tailscaled's own traffic to bypass Tailscale-managed routes,
so that things like tailscale-provided default routes don't break
tailscaled itself.
Progress on #144.
Signed-off-by: David Anderson <danderson@tailscale.com>
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>
This didn't catch anything yet, but it's good practice for detecting
goroutine leaks that we might not find otherwise.
Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
If a test calls log.Printf, 'go test' horrifyingly rearranges the
output to no longer be in chronological order, which makes debugging
virtually impossible. Let's stop that from happening by making
log.Printf panic if called from any module, no matter how deep, during
tests.
This required us to change the default error handler in at least one
http.Server, as well as plumbing a bunch of logf functions around,
especially in magicsock and wgengine, but also in logtail and backoff.
To add insult to injury, 'go test' also rearranges the output when a
parent test has multiple sub-tests (all the sub-test's t.Logf is always
printed after all the parent tests t.Logf), so we need to screw around
with a special Logf that can point at the "current" t (current_t.Logf)
in some places. Probably our entire way of using subtests is wrong,
since 'go test' would probably like to run them all in parallel if you
called t.Parallel(), but it definitely can't because the're all
manipulating the shared state created by the parent test. They should
probably all be separate toplevel tests instead, with common
setup/teardown logic. But that's a job for another time.
Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
Right now, filtering and packet injection in wgengine depend
on a patch to wireguard-go that probably isn't suitable for upstreaming.
This need not be the case: wireguard-go/tun.Device is an interface.
For example, faketun.go implements it to mock a TUN device for testing.
This patch implements the same interface to provide filtering
and packet injection at the tunnel device level,
at which point the wireguard-go patch should no longer be necessary.
This patch has the following performance impact on i7-7500U @ 2.70GHz,
tested in the following namespace configuration:
┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐
│ $ns1 │ │ $ns0 │ │ $ns2 │
│ client0 │ │ tailcontrol, logcatcher │ │ client1 │
│ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │
│ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │
│ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │
│ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │
│ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │
└────────────────┘ └─────────────────────────────────┘ └────────────────┘
Before:
---------------------------------------------------
| TCP send | UDP send |
|------------------------|------------------------|
| 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec |
---------------------------------------------------
After:
---------------------------------------------------
| TCP send | UDP send |
|------------------------|------------------------|
| 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec |
---------------------------------------------------
The impact on receive performance is similar.
Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
It used to make assumptions based on having Anycast IPs that are super
near. Now we're intentionally going to a bunch of different distant
IPs to measure latency.
Also, optimize how the hairpin detection works. No need to STUN on
that socket. Just use that separate socket for sending, once we know
the other UDP4 socket's endpoint. The trick is: make our test probe
also a STUN packet, so it fits through magicsock's existing STUN
routing.
This drops netcheck from ~5 seconds to ~250-500ms.
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Failure to do this leads to fd exhaustion at -count=10000,
and increasingly poor execution north of -count=100.
Signed-off-by: David Anderson <danderson@tailscale.com>
Failure to do so triggers either a data race or a panic
in the testing package, due to racey use of t.Logf.
Signed-off-by: David Anderson <danderson@tailscale.com>
wireguard-go closes magicsock, and expects this to unblock reads
so that its internal goroutines can wind down. We were incorrectly
blocking the read indefinitey and breaking this contract.
Signed-off-by: David Anderson <danderson@tailscale.com>
It's extremely flaky in several dimensions, as well as very slow.
It's making the CI completely red all the time without telling us
useful information.
Set RUN_CURSED_TESTS=1 to run locally.
This change just alters the semantics of the one flaky test, without
trying to speed up timeouts on the others. Empirically, speeding up
the timeouts causes _more_ flakes right now :(
The remaining flake occurs due to a mysterious packet loss. This
doesn't affect normal tailscaled operations, so until I track down
where the loss occurs and fix it, the flaky test is going to be
lenient about packet loss (but not about whether the spray logic
worked).
Signed-off-by: David Anderson <danderson@tailscale.com>
It previously passed incorrectly due to bugs. With those fixed,
it becomes flaky for 2 reasons. One of them is the wireguard handshake
race, which can eat the 1st sprayed packet and prevent roamAddr
discovery. This change fixes that failure, by spreading the test
traffic out enough that additional spraying occurs.
Signed-Off-By: David Anderson <danderson@tailscale.com>
