blob: 798fe599be4ce56e63d78a72cb85ecb578e1ab7f [file] [log] [blame]
// Copyright 2016 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package fastwalk provides a faster version of filepath.Walk for file system
// scanning tools.
package fastwalk
import (
"errors"
"os"
"path/filepath"
"runtime"
"sync"
)
// ErrTraverseLink is used as a return value from WalkFuncs to indicate that the
// symlink named in the call may be traversed.
var ErrTraverseLink = errors.New("fastwalk: traverse symlink, assuming target is a directory")
// ErrSkipFiles is a used as a return value from WalkFuncs to indicate that the
// callback should not be called for any other files in the current directory.
// Child directories will still be traversed.
var ErrSkipFiles = errors.New("fastwalk: skip remaining files in directory")
// Walk is a faster implementation of filepath.Walk.
//
// filepath.Walk's design necessarily calls os.Lstat on each file,
// even if the caller needs less info.
// Many tools need only the type of each file.
// On some platforms, this information is provided directly by the readdir
// system call, avoiding the need to stat each file individually.
// fastwalk_unix.go contains a fork of the syscall routines.
//
// See golang.org/issue/16399
//
// Walk walks the file tree rooted at root, calling walkFn for
// each file or directory in the tree, including root.
//
// If fastWalk returns filepath.SkipDir, the directory is skipped.
//
// Unlike filepath.Walk:
// - file stat calls must be done by the user.
// The only provided metadata is the file type, which does not include
// any permission bits.
// - multiple goroutines stat the filesystem concurrently. The provided
// walkFn must be safe for concurrent use.
// - 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 Walk(root string, 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
// runs after close(e.donec) below.
var wg sync.WaitGroup
defer wg.Wait()
w := &walker{
fn: walkFn,
enqueuec: make(chan walkItem, numWorkers), // buffered for performance
workc: make(chan walkItem, numWorkers), // buffered for performance
donec: make(chan struct{}),
// buffered for correctness & not leaking goroutines:
resc: make(chan error, numWorkers),
}
defer close(w.donec)
for i := 0; i < numWorkers; i++ {
wg.Add(1)
go w.doWork(&wg)
}
todo := []walkItem{{dir: root}}
out := 0
for {
workc := w.workc
var workItem walkItem
if len(todo) == 0 {
workc = nil
} else {
workItem = todo[len(todo)-1]
}
select {
case workc <- workItem:
todo = todo[:len(todo)-1]
out++
case it := <-w.enqueuec:
todo = append(todo, it)
case err := <-w.resc:
out--
if err != nil {
return err
}
if out == 0 && len(todo) == 0 {
// It's safe to quit here, as long as the buffered
// enqueue channel isn't also readable, which might
// happen if the worker sends both another unit of
// work and its result before the other select was
// scheduled and both w.resc and w.enqueuec were
// readable.
select {
case it := <-w.enqueuec:
todo = append(todo, it)
default:
return nil
}
}
}
}
}
// doWork reads directories as instructed (via workc) and runs the
// user's callback function.
func (w *walker) doWork(wg *sync.WaitGroup) {
defer wg.Done()
for {
select {
case <-w.donec:
return
case it := <-w.workc:
select {
case <-w.donec:
return
case w.resc <- w.walk(it.dir, !it.callbackDone):
}
}
}
}
type walker struct {
fn func(path string, typ os.FileMode) error
donec chan struct{} // closed on fastWalk's return
workc chan walkItem // to workers
enqueuec chan walkItem // from workers
resc chan error // from workers
}
type walkItem struct {
dir string
callbackDone bool // callback already called; don't do it again
}
func (w *walker) enqueue(it walkItem) {
select {
case w.enqueuec <- it:
case <-w.donec:
}
}
func (w *walker) onDirEnt(dirName, baseName string, typ os.FileMode) error {
joined := dirName + string(os.PathSeparator) + baseName
if typ == os.ModeDir {
w.enqueue(walkItem{dir: joined})
return nil
}
err := w.fn(joined, typ)
if typ == os.ModeSymlink {
if err == ErrTraverseLink {
// Set callbackDone so we don't call it twice for both the
// symlink-as-symlink and the symlink-as-directory later:
w.enqueue(walkItem{dir: joined, callbackDone: true})
return nil
}
if err == filepath.SkipDir {
// Permit SkipDir on symlinks too.
return nil
}
}
return err
}
func (w *walker) walk(root string, runUserCallback bool) error {
if runUserCallback {
err := w.fn(root, os.ModeDir)
if err == filepath.SkipDir {
return nil
}
if err != nil {
return err
}
}
return readDir(root, w.onDirEnt)
}