go-pulse/les/servingqueue.go
ucwong e3f3e01504
les: use new atomic types (#27856)
Co-authored-by: Felix Lange <fjl@twurst.com>
2023-08-23 21:19:04 +02:00

366 lines
10 KiB
Go

// Copyright 2019 The go-ethereum Authors
// This file is part of the go-ethereum library.
//
// The go-ethereum library is free software: you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// The go-ethereum library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
package les
import (
"sync"
"sync/atomic"
"github.com/ethereum/go-ethereum/common/mclock"
"github.com/ethereum/go-ethereum/common/prque"
"golang.org/x/exp/slices"
)
// servingQueue allows running tasks in a limited number of threads and puts the
// waiting tasks in a priority queue
type servingQueue struct {
recentTime, queuedTime uint64
servingTimeDiff atomic.Uint64
burstLimit, burstDropLimit uint64
burstDecRate float64
lastUpdate mclock.AbsTime
queueAddCh, queueBestCh chan *servingTask
stopThreadCh, quit chan struct{}
setThreadsCh chan int
wg sync.WaitGroup
threadCount int // number of currently running threads
queue *prque.Prque[int64, *servingTask] // priority queue for waiting or suspended tasks
best *servingTask // the highest priority task (not included in the queue)
suspendBias int64 // priority bias against suspending an already running task
}
// servingTask represents a request serving task. Tasks can be implemented to
// run in multiple steps, allowing the serving queue to suspend execution between
// steps if higher priority tasks are entered. The creator of the task should
// set the following fields:
//
// - priority: greater value means higher priority; values can wrap around the int64 range
// - run: execute a single step; return true if finished
// - after: executed after run finishes or returns an error, receives the total serving time
type servingTask struct {
sq *servingQueue
servingTime, timeAdded, maxTime, expTime uint64
peer *clientPeer
priority int64
biasAdded bool
token runToken
tokenCh chan runToken
}
// runToken received by servingTask.start allows the task to run. Closing the
// channel by servingTask.stop signals the thread controller to allow a new task
// to start running.
type runToken chan struct{}
// start blocks until the task can start and returns true if it is allowed to run.
// Returning false means that the task should be cancelled.
func (t *servingTask) start() bool {
if t.peer.isFrozen() {
return false
}
t.tokenCh = make(chan runToken, 1)
select {
case t.sq.queueAddCh <- t:
case <-t.sq.quit:
return false
}
select {
case t.token = <-t.tokenCh:
case <-t.sq.quit:
return false
}
if t.token == nil {
return false
}
t.servingTime -= uint64(mclock.Now())
return true
}
// done signals the thread controller about the task being finished and returns
// the total serving time of the task in nanoseconds.
func (t *servingTask) done() uint64 {
t.servingTime += uint64(mclock.Now())
close(t.token)
diff := t.servingTime - t.timeAdded
t.timeAdded = t.servingTime
if t.expTime > diff {
t.expTime -= diff
t.sq.servingTimeDiff.Add(t.expTime)
} else {
t.expTime = 0
}
return t.servingTime
}
// waitOrStop can be called during the execution of the task. It blocks if there
// is a higher priority task waiting (a bias is applied in favor of the currently
// running task). Returning true means that the execution can be resumed. False
// means the task should be cancelled.
func (t *servingTask) waitOrStop() bool {
t.done()
if !t.biasAdded {
t.priority += t.sq.suspendBias
t.biasAdded = true
}
return t.start()
}
// newServingQueue returns a new servingQueue
func newServingQueue(suspendBias int64, utilTarget float64) *servingQueue {
sq := &servingQueue{
queue: prque.New[int64, *servingTask](nil),
suspendBias: suspendBias,
queueAddCh: make(chan *servingTask, 100),
queueBestCh: make(chan *servingTask),
stopThreadCh: make(chan struct{}),
quit: make(chan struct{}),
setThreadsCh: make(chan int, 10),
burstLimit: uint64(utilTarget * bufLimitRatio * 1200000),
burstDropLimit: uint64(utilTarget * bufLimitRatio * 1000000),
burstDecRate: utilTarget,
lastUpdate: mclock.Now(),
}
sq.wg.Add(2)
go sq.queueLoop()
go sq.threadCountLoop()
return sq
}
// newTask creates a new task with the given priority
func (sq *servingQueue) newTask(peer *clientPeer, maxTime uint64, priority int64) *servingTask {
return &servingTask{
sq: sq,
peer: peer,
maxTime: maxTime,
expTime: maxTime,
priority: priority,
}
}
// threadController is started in multiple goroutines and controls the execution
// of tasks. The number of active thread controllers equals the allowed number of
// concurrently running threads. It tries to fetch the highest priority queued
// task first. If there are no queued tasks waiting then it can directly catch
// run tokens from the token channel and allow the corresponding tasks to run
// without entering the priority queue.
func (sq *servingQueue) threadController() {
defer sq.wg.Done()
for {
token := make(runToken)
select {
case best := <-sq.queueBestCh:
best.tokenCh <- token
case <-sq.stopThreadCh:
return
case <-sq.quit:
return
}
select {
case <-sq.stopThreadCh:
return
case <-sq.quit:
return
case <-token:
}
}
}
// peerTasks lists the tasks received from a given peer when selecting peers to freeze
type peerTasks struct {
peer *clientPeer
list []*servingTask
sumTime uint64
priority float64
}
// freezePeers selects the peers with the worst priority queued tasks and freezes
// them until burstTime goes under burstDropLimit or all peers are frozen
func (sq *servingQueue) freezePeers() {
peerMap := make(map[*clientPeer]*peerTasks)
var peerList []*peerTasks
if sq.best != nil {
sq.queue.Push(sq.best, sq.best.priority)
}
sq.best = nil
for sq.queue.Size() > 0 {
task := sq.queue.PopItem()
tasks := peerMap[task.peer]
if tasks == nil {
bufValue, bufLimit := task.peer.fcClient.BufferStatus()
if bufLimit < 1 {
bufLimit = 1
}
tasks = &peerTasks{
peer: task.peer,
priority: float64(bufValue) / float64(bufLimit), // lower value comes first
}
peerMap[task.peer] = tasks
peerList = append(peerList, tasks)
}
tasks.list = append(tasks.list, task)
tasks.sumTime += task.expTime
}
slices.SortFunc(peerList, func(a, b *peerTasks) int {
if a.priority < b.priority {
return -1
}
if a.priority > b.priority {
return 1
}
return 0
})
drop := true
for _, tasks := range peerList {
if drop {
tasks.peer.freeze()
tasks.peer.fcClient.Freeze()
sq.queuedTime -= tasks.sumTime
sqQueuedGauge.Update(int64(sq.queuedTime))
clientFreezeMeter.Mark(1)
drop = sq.recentTime+sq.queuedTime > sq.burstDropLimit
for _, task := range tasks.list {
task.tokenCh <- nil
}
} else {
for _, task := range tasks.list {
sq.queue.Push(task, task.priority)
}
}
}
if sq.queue.Size() > 0 {
sq.best = sq.queue.PopItem()
}
}
// updateRecentTime recalculates the recent serving time value
func (sq *servingQueue) updateRecentTime() {
subTime := sq.servingTimeDiff.Swap(0)
now := mclock.Now()
dt := now - sq.lastUpdate
sq.lastUpdate = now
if dt > 0 {
subTime += uint64(float64(dt) * sq.burstDecRate)
}
if sq.recentTime > subTime {
sq.recentTime -= subTime
} else {
sq.recentTime = 0
}
}
// addTask inserts a task into the priority queue
func (sq *servingQueue) addTask(task *servingTask) {
if sq.best == nil {
sq.best = task
} else if task.priority-sq.best.priority > 0 {
sq.queue.Push(sq.best, sq.best.priority)
sq.best = task
} else {
sq.queue.Push(task, task.priority)
}
sq.updateRecentTime()
sq.queuedTime += task.expTime
sqServedGauge.Update(int64(sq.recentTime))
sqQueuedGauge.Update(int64(sq.queuedTime))
if sq.recentTime+sq.queuedTime > sq.burstLimit {
sq.freezePeers()
}
}
// queueLoop is an event loop running in a goroutine. It receives tasks from queueAddCh
// and always tries to send the highest priority task to queueBestCh. Successfully sent
// tasks are removed from the queue.
func (sq *servingQueue) queueLoop() {
defer sq.wg.Done()
for {
if sq.best != nil {
expTime := sq.best.expTime
select {
case task := <-sq.queueAddCh:
sq.addTask(task)
case sq.queueBestCh <- sq.best:
sq.updateRecentTime()
sq.queuedTime -= expTime
sq.recentTime += expTime
sqServedGauge.Update(int64(sq.recentTime))
sqQueuedGauge.Update(int64(sq.queuedTime))
if sq.queue.Size() == 0 {
sq.best = nil
} else {
sq.best = sq.queue.PopItem()
}
case <-sq.quit:
return
}
} else {
select {
case task := <-sq.queueAddCh:
sq.addTask(task)
case <-sq.quit:
return
}
}
}
}
// threadCountLoop is an event loop running in a goroutine. It adjusts the number
// of active thread controller goroutines.
func (sq *servingQueue) threadCountLoop() {
var threadCountTarget int
defer sq.wg.Done()
for {
for threadCountTarget > sq.threadCount {
sq.wg.Add(1)
go sq.threadController()
sq.threadCount++
}
if threadCountTarget < sq.threadCount {
select {
case threadCountTarget = <-sq.setThreadsCh:
case sq.stopThreadCh <- struct{}{}:
sq.threadCount--
case <-sq.quit:
return
}
} else {
select {
case threadCountTarget = <-sq.setThreadsCh:
case <-sq.quit:
return
}
}
}
}
// setThreads sets the allowed processing thread count, suspending tasks as soon as
// possible if necessary.
func (sq *servingQueue) setThreads(threadCount int) {
select {
case sq.setThreadsCh <- threadCount:
case <-sq.quit:
return
}
}
// stop stops task processing as soon as possible and shuts down the serving queue.
func (sq *servingQueue) stop() {
close(sq.quit)
sq.wg.Wait()
}