Files
tailscale/util/eventbus/subscribe.go
T
James Tucker 0def0f19bd util/eventbus: extract SubscriberFunc.dispatch loop to a non-generic helper
The (*SubscriberFunc[T]).dispatch method body — a ~40-line select
loop with slow-subscriber timer, snapshot handling, ctx-cancel
draining, and a CI stack-dump branch — was previously fully
duplicated by the Go compiler for every distinct GC shape of T.
None of that body actually depends on T except for the type
assertion and the user callback invocation.

This change moves the loop body into a non-generic dispatchFunc()
helper, leaving (*SubscriberFunc[T]).dispatch as a tiny wrapper
that:

  - performs the vals.Peek().Event.(T) type assertion
  - spawns the callback goroutine via `go runFuncCallback(s.read,
    t, callDone)` — a regular generic function call, not a closure,
    so that `go` binds the args to the goroutine's frame instead of
    allocating a closure on the heap. This preserves the
    zero-extra-allocation behavior of the original
    (*SubscriberFunc[T]).runCallback method.
  - resolves T's name via reflect.TypeFor[T]().String() (cached on
    the stack rather than recomputed on each %T formatting)
  - calls dispatchFunc with the callDone channel

The %T formatting in the original logf calls is replaced with %s
on the resolved name string, removing per-T fmt instantiations.

A new BenchmarkBasicFuncThroughput is added alongside the existing
BenchmarkBasicThroughput so per-event allocation behavior on the
SubscribeFunc dispatch path is covered by the benchmark suite.

Measured impact (util/eventbus/sizetest):

  SubscriberFunc per-flow attribution:
    linux/amd64:  912.5 B/flow -> 840.8 B/flow  (-71.7 B/flow)
    linux/arm64:  917.5 B/flow -> 849.9 B/flow  (-67.6 B/flow)

The total per-flow size delta on amd64 dropped from 3,096.6 B to
3,039.2 B (-57 B/flow). The arm64 total stayed at 3,145.7 B
because the linker's page-aligned section sizing absorbed the
improvement on this binary; the symcost-attributed per-receiver
number is the real signal.

Behavior is unchanged: BenchmarkBasicThroughput stays at 0
allocs/op and BenchmarkBasicFuncThroughput holds at the same 2
allocs/op, 144 B/op as the prior eventbus implementation. All
eventbus tests pass.

Updates #12614

Change-Id: I85f933f50f58cd25bbfe5cc46bdda7aab22f0bf7
Signed-off-by: James Tucker <james@tailscale.com>
2026-05-06 18:56:09 -07:00

411 lines
12 KiB
Go

// Copyright (c) Tailscale Inc & contributors
// SPDX-License-Identifier: BSD-3-Clause
package eventbus
import (
"context"
"fmt"
"reflect"
"runtime"
"time"
"tailscale.com/syncs"
"tailscale.com/types/logger"
"tailscale.com/util/cibuild"
)
type DeliveredEvent struct {
Event any
From *Client
To *Client
}
// subscriber is a uniformly typed wrapper around Subscriber[T], so
// that debugging facilities can look at active subscribers.
type subscriber interface {
subscribeType() reflect.Type
// dispatch is a function that dispatches the head value in vals to
// a subscriber, while also handling stop and incoming queue write
// events.
//
// dispatch exists because of the strongly typed Subscriber[T]
// wrapper around subscriptions: within the bus events are boxed in an
// 'any', and need to be unpacked to their full type before delivery
// to the subscriber. This involves writing to a strongly-typed
// channel, so subscribeState cannot handle that dispatch by itself -
// but if that strongly typed send blocks, we also need to keep
// processing other potential sources of wakeups, which is how we end
// up at this awkward type signature and sharing of internal state
// through dispatch.
dispatch(ctx context.Context, vals *queue[DeliveredEvent], acceptCh func() chan DeliveredEvent, snapshot chan chan []DeliveredEvent) bool
Close()
}
// subscribeState handles dispatching of events received from a Bus.
type subscribeState struct {
client *Client
dispatcher *worker
write chan DeliveredEvent
snapshot chan chan []DeliveredEvent
debug hook[DeliveredEvent]
outputsMu syncs.Mutex
outputs map[reflect.Type]subscriber
}
func newSubscribeState(c *Client) *subscribeState {
ret := &subscribeState{
client: c,
write: make(chan DeliveredEvent),
snapshot: make(chan chan []DeliveredEvent),
outputs: map[reflect.Type]subscriber{},
}
ret.dispatcher = runWorker(ret.pump)
return ret
}
func (s *subscribeState) pump(ctx context.Context) {
var vals queue[DeliveredEvent]
acceptCh := func() chan DeliveredEvent {
if vals.Full() {
return nil
}
return s.write
}
for {
if !vals.Empty() {
val := vals.Peek()
sub := s.subscriberFor(val.Event)
if sub == nil {
// Raced with unsubscribe.
vals.Drop()
continue
}
if !sub.dispatch(ctx, &vals, acceptCh, s.snapshot) {
return
}
if s.debug.active() {
s.debug.run(DeliveredEvent{
Event: val.Event,
From: val.From,
To: s.client,
})
}
} else {
// Keep the cases in this select in sync with
// Subscriber.dispatch and SubscriberFunc.dispatch below.
// The only difference should be that this select doesn't deliver
// queued values to anyone, and unconditionally accepts new values.
select {
case val := <-s.write:
vals.Add(val)
case <-ctx.Done():
return
case ch := <-s.snapshot:
ch <- vals.Snapshot()
}
}
}
}
func (s *subscribeState) snapshotQueue() []DeliveredEvent {
if s == nil {
return nil
}
resp := make(chan []DeliveredEvent)
select {
case s.snapshot <- resp:
return <-resp
case <-s.dispatcher.Done():
return nil
}
}
func (s *subscribeState) subscribeTypes() []reflect.Type {
if s == nil {
return nil
}
s.outputsMu.Lock()
defer s.outputsMu.Unlock()
ret := make([]reflect.Type, 0, len(s.outputs))
for t := range s.outputs {
ret = append(ret, t)
}
return ret
}
func (s *subscribeState) addSubscriber(sub subscriber) {
s.outputsMu.Lock()
defer s.outputsMu.Unlock()
t := sub.subscribeType()
if s.outputs[t] != nil {
panic(fmt.Errorf("double subscription for event %s", t))
}
s.outputs[t] = sub
s.client.addSubscriber(t, s)
}
func (s *subscribeState) deleteSubscriber(t reflect.Type) {
s.outputsMu.Lock()
defer s.outputsMu.Unlock()
delete(s.outputs, t)
s.client.deleteSubscriber(t, s)
}
func (s *subscribeState) subscriberFor(val any) subscriber {
s.outputsMu.Lock()
defer s.outputsMu.Unlock()
return s.outputs[reflect.TypeOf(val)]
}
// Close closes the subscribeState. It implicitly closes all Subscribers
// linked to this state, and any pending events are discarded.
func (s *subscribeState) close() {
s.dispatcher.StopAndWait()
var subs map[reflect.Type]subscriber
s.outputsMu.Lock()
subs, s.outputs = s.outputs, nil
s.outputsMu.Unlock()
for _, sub := range subs {
sub.Close()
}
}
func (s *subscribeState) closed() <-chan struct{} {
return s.dispatcher.Done()
}
// A Subscriber delivers one type of event from a [Client].
// Events are sent to the [Subscriber.Events] channel.
type Subscriber[T any] struct {
stop stopFlag
read chan T
unregister func()
logf logger.Logf
slow *time.Timer // used to detect slow subscriber service
}
func newSubscriber[T any](r *subscribeState, logf logger.Logf) *Subscriber[T] {
slow := time.NewTimer(0)
slow.Stop() // reset in dispatch
return &Subscriber[T]{
read: make(chan T),
unregister: func() { r.deleteSubscriber(reflect.TypeFor[T]()) },
logf: logf,
slow: slow,
}
}
func newMonitor[T any](attach func(fn func(T)) (cancel func())) *Subscriber[T] {
ret := &Subscriber[T]{
read: make(chan T, 100), // arbitrary, large
}
ret.unregister = attach(ret.monitor)
return ret
}
func (s *Subscriber[T]) subscribeType() reflect.Type {
return reflect.TypeFor[T]()
}
func (s *Subscriber[T]) monitor(debugEvent T) {
select {
case s.read <- debugEvent:
case <-s.stop.Done():
}
}
func (s *Subscriber[T]) dispatch(ctx context.Context, vals *queue[DeliveredEvent], acceptCh func() chan DeliveredEvent, snapshot chan chan []DeliveredEvent) bool {
t := vals.Peek().Event.(T)
start := time.Now()
s.slow.Reset(slowSubscriberTimeout)
defer s.slow.Stop()
for {
// Keep the cases in this select in sync with subscribeState.pump
// above. The only difference should be that this select
// delivers a value on s.read.
select {
case s.read <- t:
vals.Drop()
return true
case val := <-acceptCh():
vals.Add(val)
case <-ctx.Done():
return false
case ch := <-snapshot:
ch <- vals.Snapshot()
case <-s.slow.C:
s.logf("subscriber for %T is slow (%v elapsed)", t, time.Since(start))
s.slow.Reset(slowSubscriberTimeout)
}
}
}
// Events returns a channel on which the subscriber's events are
// delivered.
func (s *Subscriber[T]) Events() <-chan T {
return s.read
}
// Done returns a channel that is closed when the subscriber is
// closed.
func (s *Subscriber[T]) Done() <-chan struct{} {
return s.stop.Done()
}
// Close closes the Subscriber, indicating the caller no longer wishes
// to receive this event type. After Close, receives on
// [Subscriber.Events] block for ever.
//
// If the Bus from which the Subscriber was created is closed,
// the Subscriber is implicitly closed and does not need to be closed
// separately.
func (s *Subscriber[T]) Close() {
s.stop.Stop() // unblock receivers
s.unregister()
}
// A SubscriberFunc delivers one type of event from a [Client].
// Events are forwarded synchronously to a function provided at construction.
type SubscriberFunc[T any] struct {
stop stopFlag
read func(T)
unregister func()
logf logger.Logf
slow *time.Timer // used to detect slow subscriber service
}
func newSubscriberFunc[T any](r *subscribeState, f func(T), logf logger.Logf) *SubscriberFunc[T] {
slow := time.NewTimer(0)
slow.Stop() // reset in dispatch
return &SubscriberFunc[T]{
read: f,
unregister: func() { r.deleteSubscriber(reflect.TypeFor[T]()) },
logf: logf,
slow: slow,
}
}
// Close closes the SubscriberFunc, indicating the caller no longer wishes to
// receive this event type. After Close, no further events will be passed to
// the callback.
//
// If the [Bus] from which s was created is closed, s is implicitly closed and
// does not need to be closed separately.
func (s *SubscriberFunc[T]) Close() { s.stop.Stop(); s.unregister() }
// subscribeType implements part of the subscriber interface.
func (s *SubscriberFunc[T]) subscribeType() reflect.Type { return reflect.TypeFor[T]() }
// dispatch implements part of the subscriber interface.
//
// We deliberately keep this method body small and delegate the
// dispatch loop to dispatchFunc, a non-generic helper. Each
// instantiation of SubscriberFunc[T] otherwise produces a fresh
// stencil of the full ~40-line select loop (including the slow-
// subscriber timer, snapshot handling, and CI stack-dump branch),
// which is responsible for hundreds of bytes of binary size per
// distinct T. By isolating the per-T work to the type assertion
// and the callback closure, only a small fixed-size wrapper is
// emitted per T.
func (s *SubscriberFunc[T]) dispatch(ctx context.Context, vals *queue[DeliveredEvent], acceptCh func() chan DeliveredEvent, snapshot chan chan []DeliveredEvent) bool {
t := vals.Peek().Event.(T)
callDone := make(chan struct{})
// Launch the user callback on a goroutine via a generic worker
// rather than via a closure (`go func() { s.read(t) }()` would
// allocate the closure object on the heap on every dispatched
// event). `go runFuncCallback(s.read, t, callDone)` binds its
// arguments directly to the goroutine's frame, leaving allocation
// behavior identical to the original (*SubscriberFunc[T]).runCallback
// method.
go runFuncCallback(s.read, t, callDone)
return dispatchFunc(ctx, dispatchFuncState{
slow: s.slow,
logf: s.logf,
typeName: reflect.TypeFor[T]().String(),
}, vals, acceptCh, snapshot, callDone)
}
// dispatchFuncState is the non-generic state needed by dispatchFunc.
// Bundling these lets us pass them as one argument and keeps the
// per-T wrapper at the dispatch call site small.
type dispatchFuncState struct {
slow *time.Timer
logf logger.Logf
typeName string // cached reflect.TypeFor[T]().String()
}
// dispatchFunc is the non-generic body of SubscriberFunc[T].dispatch.
// It is identical in observable behavior to the original loop; the
// only differences are that the dispatched value has already been
// unboxed by the caller (and the user callback is already running
// on its own goroutine, signaling completion via callDone) and the
// type name has already been resolved (and is passed as a string).
//
// callDone is closed by runFuncCallback when the user callback returns.
func dispatchFunc(
ctx context.Context,
st dispatchFuncState,
vals *queue[DeliveredEvent],
acceptCh func() chan DeliveredEvent,
snapshot chan chan []DeliveredEvent,
callDone chan struct{},
) bool {
start := time.Now()
st.slow.Reset(slowSubscriberTimeout)
defer st.slow.Stop()
// Keep the cases in this select in sync with subscribeState.pump
// above. The only difference should be that this select
// delivers a value by calling the user callback (via the
// goroutine spawned by the typed wrapper).
for {
select {
case <-callDone:
vals.Drop()
return true
case val := <-acceptCh():
vals.Add(val)
case <-ctx.Done():
// Wait for the callback to be complete, but not forever.
st.slow.Reset(5 * slowSubscriberTimeout)
select {
case <-st.slow.C:
st.logf("giving up on subscriber for %s after %v at close", st.typeName, time.Since(start))
if cibuild.On() {
all := make([]byte, 2<<20)
n := runtime.Stack(all, true)
st.logf("goroutine stacks:\n%s", all[:n])
}
case <-callDone:
}
return false
case ch := <-snapshot:
ch <- vals.Snapshot()
case <-st.slow.C:
st.logf("subscriber for %s is slow (%v elapsed)", st.typeName, time.Since(start))
st.slow.Reset(slowSubscriberTimeout)
}
}
}
// runFuncCallback runs f(t) and closes done when it returns. It is
// the per-T worker spawned as a goroutine for each dispatched
// event. Keeping it as a regular generic function (rather than a
// closure) means `go runFuncCallback(f, t, done)` binds its
// arguments to the goroutine's frame directly, with no per-event
// closure allocation. The body is small (defer + one indirect
// call), so the per-shape stencil cost is minimal.
func runFuncCallback[T any](f func(T), t T, done chan struct{}) {
defer close(done)
f(t)
}