Unsigned peers aren't covered by tailnet lock, so they must never hold peer capabilities even if the packet filter grants them. This change extends the check for unsigned-peers to ensure full coverage in capabilities.
Fixestailscale/corp#45116
Change-Id: I918af24f0b9855e55921cbdad109cc68e745e125
Signed-off-by: Mike Jensen <mikej@tailscale.com>
Processing a peer add/remove delta still materialized the full netmap
(an O(n) slicesx.MapValues plus sort over all peers, at 10k+
peers in a large tailnet) twice per delta: once in UpdateNetmapDelta
purely to hand the self node to Engine.SetSelfNode, and once in
authReconfigLocked.
Neither needs peers anymore. SetSelfNode gets the self node from the
existing nodeBackend.Self accessor. authReconfigLocked only reads
self-node fields (SelfNode, NodeKey, GetAddresses, HasCap) now that
WireGuard peers ride the incremental route manager and per-peer config
source, so it can use the peers-free NetMap accessor.
That also makes nmcfg.WGCfg vestigial: since wgcfg.Config lost its
Peers field, its peer walk existed only to emit the [v1] skip logs
(expired peers, unselected exit nodes, unaccepted subnet routes),
duplicating filtering the route manager already does. Delete the
package and construct the two-field wgcfg.Config inline. The skip
logs go away; if they're missed, the route manager can log them
incrementally at upsert time instead of rescanning every peer on
every reconfig.
With this, the runtime.DidRange analysis (see the ts_rangehook test)
shows a delta netmap update performing no O(n) range loops except
updateRouteManagerExtras, and the delta phase of that test drops from
1.09s to 0.14s for 400 deltas at n=10000 (from 4.79s at the
start of this effort, before the incremental route manager work).
Updates #12542
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Change-Id: Ia0e03ef9db0c988790b2c29de1f0505305e93f58
The wireguard-go device now learns its peer set solely from the live
per-peer config source that LocalBackend installs with
Engine.SetPeerConfigFunc, backed by the route manager. Peers are
created lazily on first packet and converged per peer with
Engine.SyncDevicePeer, so the full-peer-list snapshot in wgcfg.Config
and the diff-and-reconfigure machinery around it (wgcfg.Peer,
ReconfigDevice, and the engine's full device sync in
maybeReconfigWireguardLocked) are dead weight: they duplicated state
that the route manager already owns and forced every netmap change to
rebuild and rehash the entire peer list.
Delete the Peers field and the Peer type from wgcfg, along with
ReconfigDevice and maybeReconfigWireguardLocked. Engine.Reconfig no
longer does any device peer work; it only manages the private key,
addresses, and the non-peer subsystems. Full-netmap application converges the device by
syncing exactly the peers whose routes the route manager reports as
changed or removed.
Updates #12542
Change-Id: Ic776e42cfaa5be6b9329b3d381d5cbde17d7078b
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Previously tstun.Wrapper.SetWGConfig walked wgcfg.Config.Peers on every
netmap to rebuild its own IP-to-peer table for masquerade NAT rewrites
and jailed-peer classification. Now the tun layer instead consumes the
route manager's shared immutable outbound snapshot directly, via a new
Engine.SetPeerRoutes method: LocalBackend pushes the snapshot (plus this
node's native Tailscale addresses) after every route manager commit that
can change it, and per-packet lookups read the interned PeerRoute
attributes from that table.
When no current peer is jailed or masqueraded, LocalBackend installs a
nil table (gated on RouteManager.HasDataPlaneAttrs), preserving the
per-packet nil-check fast path. The exitNodeRequiresMasq machinery is
deleted: its purpose was populating the table with all peers so that
more-specific entries shadow an exit node's /0, and the always-full
route manager table gives that shadowing inherently.
This is the last step before removing the Peers field from wgcfg.Config.
Updates #12542
Change-Id: Ifce09ca929a3f2511303ca1d6efdd583739494ce
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
[This commit is pulled out of a branch that ultimately removes the
wgcfg.Config.Peers field and removes all O(n peers) processing when
handling deltas]
magicsock.Conn.UpdatePeers existed so wgengine.Reconfig could tell
magicsock the set of WireGuard peers from cfg.Peers, used only to
garbage collect the derpRoute and peerLastDerp maps and to ReSTUN when
the first peers appear. magicsock already learns the full peer list
directly from LocalBackend via SetNetworkMap, UpsertPeer, and
RemovePeer, so do that bookkeeping there and delete the API and its
cfg.Peers use.
Updates #12542
Change-Id: Id07551fc1950239f08a73a9ab02d69ce78d0de0c
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
`TestNetworkSendErrors/network-down` causes a data race because it
tried to `tstest.Replace` the `checkNetworkDownDuringTests` global
while `wgengine.Conn.networkDown` would read from it. This patch moves
this flag into a field within the `wgengine.Conn` struct, so there’s
no chance that two tests could trample on each other.
It also renames this field to `Conn.checkNetworkUpDuringTests`,
because `Conn.networkUp` is the name of the field that gets checked.
Fixes#20260
Signed-off-by: Simon Law <sfllaw@tailscale.com>
Tests in magicsock_test.go would routinely emit this warning:
## WARNING: (non-fatal) nil health.Tracker (being strict in CI):
because they would run NewConn without initializing a health.Tracker.
This patch initializes Conn correctly with a health.Tracker. It also
fixes some missing Close calls that can be handled in t.Cleanup.
Fixes#20263
Signed-off-by: Simon Law <sfllaw@tailscale.com>
This applies the same treatment from 8f210454dd (netlog) to wglog,
ending use of netmap.NetworkMap and instead getting the canonical data
from LocalBackend/nodeBackend.
This is a dependency to removing the netmap.NetworkMap from
upstream callers, like wgengine.Engine in general.
Updates #12542
Change-Id: Icb5af0799322def048a6f594b49f7d11273f025d
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Bumps wireguard-go pin to include the roaming endpoints fix, and
two internal enhancements.
Pulls stock wireguard-go for non-tailscale simulation in tests,
to use its endpoint discovery mechanism.
Updates #20082
Change-Id: I2ff282cb7fe4ab099ce5e780a1d40ae86a6a6964
Signed-off-by: Alex Valiushko <alexvaliushko@tailscale.com>
magicsock de-duplicates NetInfo callbacks against c.netInfoLast, a cache
that lives on the long-lived magicsock.Conn. That cache survives a control
client swap (interactive login or profile switch), where only the control
client (and its own per-client NetInfo dedup) is replaced. As a result, the
first netcheck after the swap produces a structurally-identical NetInfo
(same PreferredDERP, same NAT shape), magicsock suppresses it as unchanged,
and the new control session never learns our home DERP. Peers can't reach
the node over DERP until some unrelated NetInfo field happens to change.
Add Conn.ResetNetInfoLast to clear the dedup cache, and call it from
LocalBackend.setControlClientLocked whenever a control client is installed,
so the next netcheck re-reports the current NetInfo to the new client.
netInfoLast is only a dedup/optimization cache (all readers nil-guard, and
it is recomputed by every netcheck), so clearing it can only add a delivery,
never lose or misroute one; it is scoped to control-client lifecycle events,
not steady-state operation.
Updates #17887Fixes#20024
Signed-off-by: Mike O'Driscoll <mikeo@tailscale.com>
Add four control-plane node attributes that let us disable UDP GSO/GRO
on the magicsock UDP socket and UDP/TCP GRO on the Tailscale TUN
device.
These complement the pre-existing TS_DEBUG_DISABLE_UDP_{GRO,GSO} and
TS_TUN_DISABLE_{UDP,TCP}_GRO envknobs. They exist so we can mitigate
upstream Linux kernel regressions on a deployed fleet without
requiring a client release, after two incidents (#13041, #19777) where
buggy kernel patches landed upstream and the fix took an excessively
long time to reach downstream distros.
Knob changes are reacted to in setNetworkMapInternal / SetNetworkMap via
a comparison against a cached "last applied" value and only an actual
transition triggers work: magicsock Rebind()+ReSTUN for UDP,
ApplyGROKnobs for TUN. The TUN side is gated by buildfeatures.HasGRO and
is one-way (wireguard-go GRO disablement is sticky); re-enabling
requires a client restart.
Updates #13041
Updates #19777
Change-Id: I802993070afa659cc06809bb0bfbb7f8a0cdb273
Signed-off-by: James Tucker <james@tailscale.com>
Per recent chat with @raggi about all this, I went and looked at this
test again.
Updates #cleanup
Change-Id: Icb7d87b1ed2cebf481ee4e358a3aa603e63fb8a4
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Replace the UAPI text protocol-based wireguard configuration with
wireguard-go's new direct callback API (SetPeerLookupFunc,
SetPeerByIPPacketFunc, RemoveMatchingPeers, SetPrivateKey).
Instead of computing a trimmed wireguard config ahead of time upon
control plane updates and pushing it via UAPI, install callbacks so
wireguard-go creates peers on demand when packets arrive. This removes
all the LazyWG trimming machinery: idle peer tracking, activity maps,
noteRecvActivity callbacks, the KeepFullWGConfig control knob, and the
ts_omit_lazywg build tag.
For incoming packets, PeerLookupFunc answers wireguard-go's questions
about unknown public keys by looking up the peer in the full config.
For outgoing packets, PeerByIPPacketFunc (installed from
LocalBackend.lookupPeerByIP) maps destination IPs to node public keys
using the existing nodeByAddr index.
Updates tailscale/corp#12345
Change-Id: I4cba80979ac49a1231d00a01fdba5f0c2af95dd8
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Commit 78627c132f changed the signature of magicsock.Conn.SetDERPMap to
take an additional bool doReStun parameter. Avoid both the boolean
parameter and the API signature change by restoring SetDERPMap to its
original single-argument form and adding a new SetDERPMapWithoutReSTUN
method for the cache-loading caller that wants to skip the post-set
ReSTUN.
Updates #19490
Change-Id: I97d9e82156bfc546ccf59756d1ea52f039b5de06
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
With netmap caching, the home DERP of the self node was neither saved to
the cache or loaded from it, making nodes not stick to a DERP when
starting without a connection to control.
Instead, make sure that when a cache is available, load that cache,
before looking for DERP servers. This is implemented by allowing a skip
of ReSTUN in setting the DERP map (we must have a DERP map before
setting the home DERP), so the DERP from cache will set itself and be
sticky until a connection to control is established.
Making DERP only change when connected to control is handled by existing
code from f072d017bd.
Updates #19490
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
When there is an active connection between devices, do not send new
disco keys via TSMP.
Updates #12639
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
Replace Conn.peers (sorted views.Slice) with peersByID, a
map[tailcfg.NodeID]tailcfg.NodeView. The only caller that needed
the sorted slice (the disco message receive path's binary search)
becomes a single map lookup. Drop nodesEqual.
Add Conn.UpsertPeer / Conn.RemovePeer for O(1) single-peer endpoint
work. RemovePeer also performs a targeted single-disco-key cleanup
(previously that scan was O(discoInfo)).
Extract the shared per-peer upsert body as upsertPeerLocked; still
used by SetNetworkMap's bulk path. SetNetworkMap is documented as
the bulk / initial / self-change path; UpsertPeer and RemovePeer
are preferred for single-peer changes.
Make the relay server set update O(1) per peer: add serverUpsertCh
/ serverRemoveCh to relayManager with matching run-loop handlers.
UpsertPeer / RemovePeer evaluate the per-peer relay predicate
locally and dispatch upsert or remove. The full-rebuild
updateRelayServersSet stays for the initial netmap, filter
changes, and fallback.
Move the hasPeerRelayServers atomic from Conn onto relayManager,
next to the serversByNodeKey map it summarizes. The run loop is
now the single writer and needs no back-pointer to Conn;
endpoint's two hot-path readers take one extra hop to
de.c.relayManager.hasPeerRelayServers but the cost is the same
atomic load.
No callers use UpsertPeer/RemovePeer yet; a subsequent change will
plumb per-peer add/remove through the incremental map update path.
Updates #12542
Change-Id: If6a3442fe29ccbd77890ea61b754a4d1ad6ef225
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
pickPort would bind a UDP socket on :0 to get a free port, close
the socket, then hope to rebind to the same port in NewConn. This
is a TOCTOU race that can cause flaky test failures when another
process grabs the port in between.
Instead, pass Port: 0 to NewConn and let the OS assign the port
atomically, then read back the assigned port via conn.LocalPort().
Fixes#19409
Change-Id: Ie44b599fb93c361e29a05f2171ad747c46f82b7a
Co-authored-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
The compare-metrics-stats subtest reset two independent counting
systems (physical connection counters and expvar.Int user metrics)
non-atomically. Background WireGuard keepalives arriving between the
resets could increment one system but not the other, causing
off-by-one packet/byte mismatches in either direction.
Replace the reset-then-compare pattern with snapshot-and-delta:
snapshot both systems before pings, snapshot again after, and compare
the deltas. This eliminates the non-atomic reset window entirely.
As a belt-and-suspenders safety net, tolerate a difference of exactly
one packet (and corresponding bytes) from a stray keepalive that
could still arrive in the narrow window between the two snapshots.
flakestress passes with ~5900 runs (~2800 without -race, ~3100 with
-race) but it also passed previously too. This is an annoying one to
repro.
Fixes#11762
Change-Id: I3447ad67e71c8146e85eed38b7a665033ef9e284
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Add a new vet analyzer that checks t.Run subtest names don't contain
characters requiring quoting when re-running via "go test -run". This
enforces the style guide rule: don't use spaces or punctuation in
subtest names.
The analyzer flags:
- Direct t.Run calls with string literal names containing spaces,
regex metacharacters, quotes, or other problematic characters
- Table-driven t.Run(tt.name, ...) calls where tt ranges over a
slice/map literal with bad name field values
Also fix all 978 existing violations across 81 test files, replacing
spaces with hyphens and shortening long sentence-like names to concise
hyphenated forms.
Updates #19242
Change-Id: Ib0ad96a111bd8e764582d1d4902fe2599454ab65
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
Instead of sending out disco keys via TSMP once, send them out in
intervals of 60+ seconds. The trigger is still callmemaaybe and the keys
will not be send if no direct connection needs to be established.
This fixes a case where a node can have stale keys but have communicated
with the other peer before, leading to an infinite DERP state.
Updates #12639
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
Fix three independent flake sources, at least as debugged by Claude,
though empirically no longer flaking as it was before:
1. Poll for connection counter data instead of reading immediately.
The conncount callback fires asynchronously on received WireGuard
traffic, so after counts.Reset() there is no guarantee the counter
has been repopulated before checkStats reads it. Use tstest.WaitFor
with a 5s timeout to retry until a matching connection appears.
2. Replace the *2 symmetry assumption in global metric assertions.
metricSendUDP and friends are AggregateCounters that sum per-conn
expvars from both magicsock instances. The old assertion assumed
both instances had identical packet counts, which breaks under
asymmetric background WireGuard activity (handshake retries, etc).
The new assertGlobalMetricsMatchPerConn computes the actual sum of
both conns' expvars and compares against the AggregateCounter value.
3. Tolerate physical stats being 0 when user metrics are non-zero.
A rebind event replaces the socket mid-measurement, resetting the
physical connection counter while user metrics still reflect packets
processed before the rebind. Log instead of failing in this case.
Also move counts.Reset() after metric reads and reorder the reset
sequence (counts before metrics) to minimize the race window.
Fixestailscale/tailscale#13420
Change-Id: I7b090a4dc229a862c1a52161b3f2547ec1d1f23f
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
I omitted a lot of the min/max modernizers because they didn't
result in more clear code.
Some of it's older "for x := range 123".
Also: errors.AsType, any, fmt.Appendf, etc.
Updates #18682
Change-Id: I83a451577f33877f962766a5b65ce86f7696471c
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
derpActiveFunc was being called immediately as a bare goroutine,
before startGate was resolved. For the firstDerp case, startGate
is c.derpStarted which only closes after dc.Connect() completes,
so derpActiveFunc was firing before the DERP connection existed.
We now block it with the same logic used by runDerpReader and by
runDerpWriter.
Updates: #18810
Signed-off-by: Fernando Serboncini <fserb@tailscale.com>
Restore synchronous method calls from LocalBackend to magicsock.Conn
for node views, filter, and delta mutations. The eventbus delivery
introduced in 8e6f63cf1 was invalid for these updates because
subsequent operations in the same call chain depend on magicsock
already having the current state. The Synchronize/settleEventBus
workaround was fragile and kept requiring more workarounds and
introducing new mystery bugs.
Since eventbus was added, we've since learned more about when to use
eventbus, and this wasn't one of the cases.
We can take another swing at using eventbus for netmap changes in a
future change.
Fixes#16369
Updates #18575 (likely fixes)
Change-Id: I79057cc9259993368bb1e350ff0e073adf6b9a8f
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
This file was never truly necessary and has never actually been used in
the history of Tailscale's open source releases.
A Brief History of AUTHORS files
---
The AUTHORS file was a pattern developed at Google, originally for
Chromium, then adopted by Go and a bunch of other projects. The problem
was that Chromium originally had a copyright line only recognizing
Google as the copyright holder. Because Google (and most open source
projects) do not require copyright assignemnt for contributions, each
contributor maintains their copyright. Some large corporate contributors
then tried to add their own name to the copyright line in the LICENSE
file or in file headers. This quickly becomes unwieldy, and puts a
tremendous burden on anyone building on top of Chromium, since the
license requires that they keep all copyright lines intact.
The compromise was to create an AUTHORS file that would list all of the
copyright holders. The LICENSE file and source file headers would then
include that list by reference, listing the copyright holder as "The
Chromium Authors".
This also become cumbersome to simply keep the file up to date with a
high rate of new contributors. Plus it's not always obvious who the
copyright holder is. Sometimes it is the individual making the
contribution, but many times it may be their employer. There is no way
for the proejct maintainer to know.
Eventually, Google changed their policy to no longer recommend trying to
keep the AUTHORS file up to date proactively, and instead to only add to
it when requested: https://opensource.google/docs/releasing/authors.
They are also clear that:
> Adding contributors to the AUTHORS file is entirely within the
> project's discretion and has no implications for copyright ownership.
It was primarily added to appease a small number of large contributors
that insisted that they be recognized as copyright holders (which was
entirely their right to do). But it's not truly necessary, and not even
the most accurate way of identifying contributors and/or copyright
holders.
In practice, we've never added anyone to our AUTHORS file. It only lists
Tailscale, so it's not really serving any purpose. It also causes
confusion because Tailscalars put the "Tailscale Inc & AUTHORS" header
in other open source repos which don't actually have an AUTHORS file, so
it's ambiguous what that means.
Instead, we just acknowledge that the contributors to Tailscale (whoever
they are) are copyright holders for their individual contributions. We
also have the benefit of using the DCO (developercertificate.org) which
provides some additional certification of their right to make the
contribution.
The source file changes were purely mechanical with:
git ls-files | xargs sed -i -e 's/\(Tailscale Inc &\) AUTHORS/\1 contributors/g'
Updates #cleanup
Change-Id: Ia101a4a3005adb9118051b3416f5a64a4a45987d
Signed-off-by: Will Norris <will@tailscale.com>
Adds a new types of TSMP messages for advertising disco keys keys
to/from a peer, and implements the advertising triggered by a TSMP ping.
Needed as part of the effort to cache the netmap and still let clients
connect without control being reachable.
Updates #12639
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
Co-authored-by: James Tucker <james@tailscale.com>
Adds the ability to rotate discovery keys on running clients, needed for
testing upcoming disco key distribution changes.
Introduces key.DiscoKey, an atomic container for a disco private key,
public key, and the public key's ShortString, replacing the prior
separate atomic fields.
magicsock.Conn has a new RotateDiscoKey method, and access to this is
provided via localapi and a CLI debug command.
Note that this implementation is primarily for testing as it stands, and
regular use should likely introduce an additional mechanism that allows
the old key to be used for some time, to provide a seamless key rotation
rather than one that invalidates all sessions.
Updates tailscale/corp#34037
Signed-off-by: James Tucker <james@tailscale.com>
It's an unnecessary nuisance having it. We go out of our way to redact
it in so many places when we don't even need it there anyway.
Updates #12639
Change-Id: I5fc72e19e9cf36caeb42cf80ba430873f67167c3
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
They distracted me in some refactoring. They're set but never used.
Updates #17858
Change-Id: I6ec7d6841ab684a55bccca7b7cbf7da9c782694f
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
The connstats package was an unnecessary layer of indirection.
It was seperated out of wgengine/netlog so that net/tstun and
wgengine/magicsock wouldn't need a depenedency on the concrete
implementation of network flow logging.
Instead, we simply register a callback for counting connections.
This PR does the bare minimum work to prepare tstun and magicsock
to only care about that callback.
A future PR will delete connstats and merge it into netlog.
Updates tailscale/corp#33352
Signed-off-by: Joe Tsai <joetsai@digital-static.net>
On platforms that are causing EPIPE at a high frequency this is
resulting in non-working connections, for example when Apple decides to
forcefully close UDP sockets due to an unsoliced packet rejection in the
firewall.
Too frequent rebinds cause a failure to solicit the endpoints triggering
the rebinds, that would normally happen via CallMeMaybe.
Updates #14551
Updates tailscale/corp#25648
Signed-off-by: James Tucker <james@tailscale.com>
PR #17258 extracted `derp.Server` into `derp/derpserver.Server`.
This followup patch adds the following cleanups:
1. Rename `derp_server*.go` files to `derpserver*.go` to match
the package name.
2. Rename the `derpserver.NewServer` constructor to `derpserver.New`
to reduce stuttering.
3. Remove the unnecessary `derpserver.Conn` type alias.
Updates #17257
Updates #cleanup
Signed-off-by: Simon Law <sfllaw@tailscale.com>
This exports a number of things from the derp (generic + client) package
to be used by the new derpserver package, as now used by cmd/derper.
And then enough other misc changes to lock in that cmd/tailscaled can
be configured to not bring in tailscale.com/client/local. (The webclient
in particular, even when disabled, was bringing it in, so that's now fixed)
Fixes#17257
Change-Id: I88b6c7958643fb54f386dd900bddf73d2d4d96d5
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>
The Tracker was using direct callbacks to ipnlocal. This PR moves those
to be triggered via the eventbus.
Additionally, the eventbus is now closed on exit from tailscaled
explicitly, and health is now a SubSystem in tsd.
Updates #15160
Signed-off-by: Claus Lensbøl <claus@tailscale.com>
This test ostensibly checks whether we record an error metric if a packet
is dropped because the network is down, but the network connectivity is
irrelevant -- the send error is actually because the arguments to Send()
are invalid:
RebindingUDPConn.WriteWireGuardBatchTo:
[unexpected] offset (0) != Geneve header length (8)
This patch changes the test so we try to send a valid packet, and we
verify this by sending it once before taking the network down. The new
error is:
magicsock: network down
which is what we're trying to test.
We then test sending an invalid payload as a separate test case.
Updates tailscale/corp#22075
Signed-off-by: Alex Chan <alexc@tailscale.com>
DERP writes go via TCP and the host OS will have plenty of buffer space.
We've observed in the wild with a backed up TCP socket kernel side
buffers of >2.4MB. The DERP internal queue being larger causes an
increase in the probability that the contents of the backbuffer are
"dead letters" - packets that were assumed to be lost.
A first step to improvement is to size this queue only large enough to
avoid some of the initial connect stall problem, but not large enough
that it is contributing in a substantial way to buffer bloat /
dead-letter retention.
Updates tailscale/corp#31762
Signed-off-by: James Tucker <james@tailscale.com>
