Avoid select inside a recursive function to avoid CPU spikes (#8923)

Additionally also allow configurable go-routines

cmd/data-usage.go

@@ -22,6 +22,7 @@ import (
 	"encoding/json"
 	"os"
 	"path/filepath"
+	"sync"
 	"time"
 
 	jsoniter "github.com/json-iterator/go"
@@ -150,77 +151,55 @@ type Item struct {
 }
 
 type getSizeFn func(item Item) (int64, error)
-type activeIOFn func() error
 
-func updateUsage(basePath string, endCh <-chan struct{}, waitForLowActiveIO activeIOFn, getSize getSizeFn) DataUsageInfo {
+func updateUsage(basePath string, doneCh <-chan struct{}, waitForLowActiveIO func(), getSize getSizeFn) DataUsageInfo {
 	var dataUsageInfo = DataUsageInfo{
 		BucketsSizes:          make(map[string]uint64),
 		ObjectsSizesHistogram: make(map[string]uint64),
 	}
 
-	itemCh := make(chan Item)
+	numWorkers := 4
-	skipCh := make(chan error)
-	defer close(skipCh)
 
-	go func() {
+	var mutex sync.Mutex // Mutex to update dataUsageInfo
-		defer close(itemCh)
-		fastWalk(basePath, func(path string, typ os.FileMode) error {
-			if err := waitForLowActiveIO(); err != nil {
-				return filepath.SkipDir
-			}
 
-			select {
+	fastWalk(basePath, numWorkers, doneCh, func(path string, typ os.FileMode) error {
-			case <-endCh:
+		// Wait for I/O to go down.
-				return filepath.SkipDir
+		waitForLowActiveIO()
-			case itemCh <- Item{path, typ}:
-			}
-			return <-skipCh
-		})
-	}()
 
-	for {
+		bucket, entry := path2BucketObjectWithBasePath(basePath, path)
-		select {
-		case <-endCh:
-			return dataUsageInfo
-		case item, ok := <-itemCh:
-			if !ok {
-				return dataUsageInfo
-			}
-
-			bucket, entry := path2BucketObjectWithBasePath(basePath, item.Path)
 		if bucket == "" {
-				skipCh <- nil
+			return nil
-				continue
 		}
 
 		if isReservedOrInvalidBucket(bucket, false) {
-				skipCh <- filepath.SkipDir
+			return filepath.SkipDir
-				continue
 		}
 
-			if entry == "" && item.Typ&os.ModeDir != 0 {
+		if entry == "" && typ&os.ModeDir != 0 {
+			mutex.Lock()
 			dataUsageInfo.BucketsCount++
 			dataUsageInfo.BucketsSizes[bucket] = 0
-				skipCh <- nil
+			mutex.Unlock()
-				continue
+			return nil
 		}
 
-			if item.Typ&os.ModeDir != 0 {
+		if typ&os.ModeDir != 0 {
-				skipCh <- nil
+			return nil
-				continue
 		}
 
-			size, err := getSize(item)
+		size, err := getSize(Item{path, typ})
 		if err != nil {
-				skipCh <- errSkipFile
+			return errSkipFile
-				continue
 		}
 
+		mutex.Lock()
 		dataUsageInfo.ObjectsCount++
 		dataUsageInfo.ObjectsTotalSize += uint64(size)
 		dataUsageInfo.BucketsSizes[bucket] += uint64(size)
 		dataUsageInfo.ObjectsSizesHistogram[objSizeToHistoInterval(uint64(size))]++
-			skipCh <- nil
+		mutex.Unlock()
-		}
+		return nil
-	}
+	})
+
+	return dataUsageInfo
 }

cmd/fastwalk.go

@@ -12,7 +12,6 @@ import (
 	"errors"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
 	"sync"
 )
@@ -44,16 +43,7 @@ var errSkipFile = errors.New("fastwalk: skip this file")
 //   * fastWalk can follow symlinks if walkFn returns the TraverseLink
 //     sentinel error. It is the walkFn's responsibility to prevent
 //     fastWalk from going into symlink cycles.
-func fastWalk(root string, walkFn func(path string, typ os.FileMode) error) error {
+func fastWalk(root string, nworkers int, doneCh <-chan struct{}, walkFn func(path string, typ os.FileMode) error) error {
-	// TODO(bradfitz): make numWorkers configurable? We used a
-	// minimum of 4 to give the kernel more info about multiple
-	// things we want, in hopes its I/O scheduling can take
-	// advantage of that. Hopefully most are in cache. Maybe 4 is
-	// even too low of a minimum. Profile more.
-	numWorkers := 4
-	if n := runtime.NumCPU(); n > numWorkers {
-		numWorkers = n
-	}
 
 	// Make sure to wait for all workers to finish, otherwise
 	// walkFn could still be called after returning. This Wait call
@@ -63,19 +53,20 @@ func fastWalk(root string, walkFn func(path string, typ os.FileMode) error) erro
 
 	w := &walker{
 		fn:       walkFn,
-		enqueuec: make(chan walkItem, numWorkers), // buffered for performance
+		enqueuec: make(chan walkItem, nworkers), // buffered for performance
-		workc:    make(chan walkItem, numWorkers), // buffered for performance
+		workc:    make(chan walkItem, nworkers), // buffered for performance
 		donec:    make(chan struct{}),
 
 		// buffered for correctness & not leaking goroutines:
-		resc: make(chan error, numWorkers),
+		resc: make(chan error, nworkers),
 	}
 	defer close(w.donec)
 
-	for i := 0; i < numWorkers; i++ {
+	for i := 0; i < nworkers; i++ {
 		wg.Add(1)
 		go w.doWork(&wg)
 	}
+
 	todo := []walkItem{{dir: root}}
 	out := 0
 	for {
@@ -87,6 +78,8 @@ func fastWalk(root string, walkFn func(path string, typ os.FileMode) error) erro
 			workItem = todo[len(todo)-1]
 		}
 		select {
+		case <-doneCh:
+			return nil
 		case workc <- workItem:
 			todo = todo[:len(todo)-1]
 			out++

cmd/fs-v1.go

@@ -19,7 +19,6 @@ package cmd
 import (
 	"bytes"
 	"context"
-	"errors"
 	"fmt"
 	"io"
 	"io/ioutil"
@@ -226,18 +225,10 @@ func (fs *FSObjects) StorageInfo(ctx context.Context) StorageInfo {
 	return storageInfo
 }
 
-func (fs *FSObjects) waitForLowActiveIO() error {
+func (fs *FSObjects) waitForLowActiveIO() {
 	for atomic.LoadInt64(&fs.activeIOCount) >= fs.maxActiveIOCount {
-		select {
+		time.Sleep(lowActiveIOWaitTick)
-		case <-GlobalServiceDoneCh:
-			return errors.New("forced exit")
-		case <-time.NewTimer(lowActiveIOWaitTick).C:
-			continue
-		}
 	}
-
-	return nil
-
 }
 
 // CrawlAndGetDataUsage returns data usage stats of the current FS deployment

cmd/posix.go

@@ -333,16 +333,10 @@ func isQuitting(endCh chan struct{}) bool {
 	}
 }
 
-func (s *posix) waitForLowActiveIO() error {
+func (s *posix) waitForLowActiveIO() {
 	for atomic.LoadInt32(&s.activeIOCount) >= s.maxActiveIOCount {
-		select {
+		time.Sleep(lowActiveIOWaitTick)
-		case <-GlobalServiceDoneCh:
-			return errors.New("forced exit")
-		case <-time.NewTimer(lowActiveIOWaitTick).C:
-			continue
 	}
-	}
-	return nil
 }
 
 func (s *posix) CrawlAndGetDataUsage(endCh <-chan struct{}) (DataUsageInfo, error) {