rclone/fs/march/march.go
Nick Craig-Wood 88bd80c1fa march: Fix excessive parallelism when using --no-traverse
When using `--no-traverse` the march routines call NewObject on each
potential object in the destination.

The concurrency limiter was accidentally arranged so that there were
`--checkers` * `--checkers` NewObject calls going on at once.

This became obvious when using the sftp backend which used too many
connections.

Fixes #5824
2023-11-20 17:36:31 +00:00

503 lines
14 KiB
Go

// Package march traverses two directories in lock step
package march
import (
"context"
"fmt"
"path"
"sort"
"strings"
"sync"
"github.com/rclone/rclone/fs"
"github.com/rclone/rclone/fs/dirtree"
"github.com/rclone/rclone/fs/filter"
"github.com/rclone/rclone/fs/list"
"github.com/rclone/rclone/fs/walk"
"golang.org/x/text/unicode/norm"
)
// March holds the data used to traverse two Fs simultaneously,
// calling Callback for each match
type March struct {
// parameters
Ctx context.Context // context for background goroutines
Fdst fs.Fs // source Fs
Fsrc fs.Fs // dest Fs
Dir string // directory
NoTraverse bool // don't traverse the destination
SrcIncludeAll bool // don't include all files in the src
DstIncludeAll bool // don't include all files in the destination
Callback Marcher // object to call with results
NoCheckDest bool // transfer all objects regardless without checking dst
NoUnicodeNormalization bool // don't normalize unicode characters in filenames
// internal state
srcListDir listDirFn // function to call to list a directory in the src
dstListDir listDirFn // function to call to list a directory in the dst
transforms []matchTransformFn
limiter chan struct{} // make sure we don't do too many operations at once
}
// Marcher is called on each match
type Marcher interface {
// SrcOnly is called for a DirEntry found only in the source
SrcOnly(src fs.DirEntry) (recurse bool)
// DstOnly is called for a DirEntry found only in the destination
DstOnly(dst fs.DirEntry) (recurse bool)
// Match is called for a DirEntry found both in the source and destination
Match(ctx context.Context, dst, src fs.DirEntry) (recurse bool)
}
// init sets up a march over opt.Fsrc, and opt.Fdst calling back callback for each match
// Note: this will flag filter-aware backends on the source side
func (m *March) init(ctx context.Context) {
ci := fs.GetConfig(ctx)
m.srcListDir = m.makeListDir(ctx, m.Fsrc, m.SrcIncludeAll)
if !m.NoTraverse {
m.dstListDir = m.makeListDir(ctx, m.Fdst, m.DstIncludeAll)
}
// Now create the matching transform
// ..normalise the UTF8 first
if !m.NoUnicodeNormalization {
m.transforms = append(m.transforms, norm.NFC.String)
}
// ..if destination is caseInsensitive then make it lower case
// case Insensitive | src | dst | lower case compare |
// | No | No | No |
// | Yes | No | No |
// | No | Yes | Yes |
// | Yes | Yes | Yes |
if m.Fdst.Features().CaseInsensitive || ci.IgnoreCaseSync {
m.transforms = append(m.transforms, strings.ToLower)
}
// Limit parallelism for operations
m.limiter = make(chan struct{}, ci.Checkers)
}
// list a directory into entries, err
type listDirFn func(dir string) (entries fs.DirEntries, err error)
// makeListDir makes constructs a listing function for the given fs
// and includeAll flags for marching through the file system.
// Note: this will optionally flag filter-aware backends!
func (m *March) makeListDir(ctx context.Context, f fs.Fs, includeAll bool) listDirFn {
ci := fs.GetConfig(ctx)
fi := filter.GetConfig(ctx)
if !(ci.UseListR && f.Features().ListR != nil) && // !--fast-list active and
!(ci.NoTraverse && fi.HaveFilesFrom()) { // !(--files-from and --no-traverse)
return func(dir string) (entries fs.DirEntries, err error) {
dirCtx := filter.SetUseFilter(m.Ctx, f.Features().FilterAware && !includeAll) // make filter-aware backends constrain List
return list.DirSorted(dirCtx, f, includeAll, dir)
}
}
// This returns a closure for use when --fast-list is active or for when
// --files-from and --no-traverse is set
var (
mu sync.Mutex
started bool
dirs dirtree.DirTree
dirsErr error
)
return func(dir string) (entries fs.DirEntries, err error) {
mu.Lock()
defer mu.Unlock()
if !started {
dirCtx := filter.SetUseFilter(m.Ctx, f.Features().FilterAware && !includeAll) // make filter-aware backends constrain List
dirs, dirsErr = walk.NewDirTree(dirCtx, f, m.Dir, includeAll, ci.MaxDepth)
started = true
}
if dirsErr != nil {
return nil, dirsErr
}
entries, ok := dirs[dir]
if !ok {
err = fs.ErrorDirNotFound
} else {
delete(dirs, dir)
}
return entries, err
}
}
// listDirJob describe a directory listing that needs to be done
type listDirJob struct {
srcRemote string
dstRemote string
srcDepth int
dstDepth int
noSrc bool
noDst bool
}
// Run starts the matching process off
func (m *March) Run(ctx context.Context) error {
ci := fs.GetConfig(ctx)
fi := filter.GetConfig(ctx)
m.init(ctx)
srcDepth := ci.MaxDepth
if srcDepth < 0 {
srcDepth = fs.MaxLevel
}
dstDepth := srcDepth
if fi.Opt.DeleteExcluded {
dstDepth = fs.MaxLevel
}
var mu sync.Mutex // Protects vars below
var jobError error
var errCount int
// Start some directory listing go routines
var wg sync.WaitGroup // sync closing of go routines
var traversing sync.WaitGroup // running directory traversals
checkers := ci.Checkers
in := make(chan listDirJob, checkers)
for i := 0; i < checkers; i++ {
wg.Add(1)
go func() {
defer wg.Done()
for {
select {
case <-m.Ctx.Done():
return
case job, ok := <-in:
if !ok {
return
}
jobs, err := m.processJob(job)
if err != nil {
mu.Lock()
// Keep reference only to the first encountered error
if jobError == nil {
jobError = err
}
errCount++
mu.Unlock()
}
if len(jobs) > 0 {
traversing.Add(len(jobs))
go func() {
// Now we have traversed this directory, send these
// jobs off for traversal in the background
for _, newJob := range jobs {
select {
case <-m.Ctx.Done():
// discard job if finishing
traversing.Done()
case in <- newJob:
}
}
}()
}
traversing.Done()
}
}
}()
}
// Start the process
traversing.Add(1)
in <- listDirJob{
srcRemote: m.Dir,
srcDepth: srcDepth - 1,
dstRemote: m.Dir,
dstDepth: dstDepth - 1,
noDst: m.NoCheckDest,
}
go func() {
// when the context is cancelled discard the remaining jobs
<-m.Ctx.Done()
for range in {
traversing.Done()
}
}()
traversing.Wait()
close(in)
wg.Wait()
if errCount > 1 {
return fmt.Errorf("march failed with %d error(s): first error: %w", errCount, jobError)
}
return jobError
}
// Check to see if the context has been cancelled
func (m *March) aborting() bool {
select {
case <-m.Ctx.Done():
return true
default:
}
return false
}
// matchEntry is an entry plus transformed name
type matchEntry struct {
entry fs.DirEntry
leaf string
name string
}
// matchEntries contains many matchEntry~s
type matchEntries []matchEntry
// Len is part of sort.Interface.
func (es matchEntries) Len() int { return len(es) }
// Swap is part of sort.Interface.
func (es matchEntries) Swap(i, j int) { es[i], es[j] = es[j], es[i] }
// Less is part of sort.Interface.
//
// Compare in order (name, leaf, remote)
func (es matchEntries) Less(i, j int) bool {
ei, ej := &es[i], &es[j]
if ei.name == ej.name {
if ei.leaf == ej.leaf {
return fs.CompareDirEntries(ei.entry, ej.entry) < 0
}
return ei.leaf < ej.leaf
}
return ei.name < ej.name
}
// Sort the directory entries by (name, leaf, remote)
//
// We use a stable sort here just in case there are
// duplicates. Assuming the remote delivers the entries in a
// consistent order, this will give the best user experience
// in syncing as it will use the first entry for the sync
// comparison.
func (es matchEntries) sort() {
sort.Stable(es)
}
// make a matchEntries from a newMatch entries
func newMatchEntries(entries fs.DirEntries, transforms []matchTransformFn) matchEntries {
es := make(matchEntries, len(entries))
for i := range es {
es[i].entry = entries[i]
name := path.Base(entries[i].Remote())
es[i].leaf = name
for _, transform := range transforms {
name = transform(name)
}
es[i].name = name
}
es.sort()
return es
}
// matchPair is a matched pair of direntries returned by matchListings
type matchPair struct {
src, dst fs.DirEntry
}
// matchTransformFn converts a name into a form which is used for
// comparison in matchListings.
type matchTransformFn func(name string) string
// Process the two listings, matching up the items in the two slices
// using the transform function on each name first.
//
// Into srcOnly go Entries which only exist in the srcList
// Into dstOnly go Entries which only exist in the dstList
// Into matches go matchPair's of src and dst which have the same name
//
// This checks for duplicates and checks the list is sorted.
func matchListings(srcListEntries, dstListEntries fs.DirEntries, transforms []matchTransformFn) (srcOnly fs.DirEntries, dstOnly fs.DirEntries, matches []matchPair) {
srcList := newMatchEntries(srcListEntries, transforms)
dstList := newMatchEntries(dstListEntries, transforms)
for iSrc, iDst := 0, 0; ; iSrc, iDst = iSrc+1, iDst+1 {
var src, dst fs.DirEntry
var srcName, dstName string
if iSrc < len(srcList) {
src = srcList[iSrc].entry
srcName = srcList[iSrc].name
}
if iDst < len(dstList) {
dst = dstList[iDst].entry
dstName = dstList[iDst].name
}
if src == nil && dst == nil {
break
}
if src != nil && iSrc > 0 {
prev := srcList[iSrc-1].entry
prevName := srcList[iSrc-1].name
if srcName == prevName && fs.DirEntryType(prev) == fs.DirEntryType(src) {
fs.Logf(src, "Duplicate %s found in source - ignoring", fs.DirEntryType(src))
iDst-- // ignore the src and retry the dst
continue
} else if srcName < prevName {
// this should never happen since we sort the listings
panic("Out of order listing in source")
}
}
if dst != nil && iDst > 0 {
prev := dstList[iDst-1].entry
prevName := dstList[iDst-1].name
if dstName == prevName && fs.DirEntryType(dst) == fs.DirEntryType(prev) {
fs.Logf(dst, "Duplicate %s found in destination - ignoring", fs.DirEntryType(dst))
iSrc-- // ignore the dst and retry the src
continue
} else if dstName < prevName {
// this should never happen since we sort the listings
panic("Out of order listing in destination")
}
}
if src != nil && dst != nil {
// we can't use CompareDirEntries because srcName, dstName could
// be different then src.Remote() or dst.Remote()
srcType := fs.DirEntryType(src)
dstType := fs.DirEntryType(dst)
if srcName > dstName || (srcName == dstName && srcType > dstType) {
src = nil
iSrc--
} else if srcName < dstName || (srcName == dstName && srcType < dstType) {
dst = nil
iDst--
}
}
// Debugf(nil, "src = %v, dst = %v", src, dst)
switch {
case src == nil && dst == nil:
// do nothing
case src == nil:
dstOnly = append(dstOnly, dst)
case dst == nil:
srcOnly = append(srcOnly, src)
default:
matches = append(matches, matchPair{src: src, dst: dst})
}
}
return
}
// processJob processes a listDirJob listing the source and
// destination directories, comparing them and returning a slice of
// more jobs
//
// returns errors using processError
func (m *March) processJob(job listDirJob) ([]listDirJob, error) {
var (
jobs []listDirJob
srcList, dstList fs.DirEntries
srcListErr, dstListErr error
wg sync.WaitGroup
mu sync.Mutex
)
// List the src and dst directories
if !job.noSrc {
wg.Add(1)
go func() {
defer wg.Done()
srcList, srcListErr = m.srcListDir(job.srcRemote)
}()
}
if !m.NoTraverse && !job.noDst {
wg.Add(1)
go func() {
defer wg.Done()
dstList, dstListErr = m.dstListDir(job.dstRemote)
}()
}
// Wait for listings to complete and report errors
wg.Wait()
if srcListErr != nil {
if job.srcRemote != "" {
fs.Errorf(job.srcRemote, "error reading source directory: %v", srcListErr)
} else {
fs.Errorf(m.Fsrc, "error reading source root directory: %v", srcListErr)
}
srcListErr = fs.CountError(srcListErr)
return nil, srcListErr
}
if dstListErr == fs.ErrorDirNotFound {
// Copy the stuff anyway
} else if dstListErr != nil {
if job.dstRemote != "" {
fs.Errorf(job.dstRemote, "error reading destination directory: %v", dstListErr)
} else {
fs.Errorf(m.Fdst, "error reading destination root directory: %v", dstListErr)
}
dstListErr = fs.CountError(dstListErr)
return nil, dstListErr
}
// If NoTraverse is set, then try to find a matching object
// for each item in the srcList to head dst object
if m.NoTraverse && !m.NoCheckDest {
for _, src := range srcList {
wg.Add(1)
m.limiter <- struct{}{}
go func(src fs.DirEntry) {
defer wg.Done()
if srcObj, ok := src.(fs.Object); ok {
leaf := path.Base(srcObj.Remote())
dstObj, err := m.Fdst.NewObject(m.Ctx, path.Join(job.dstRemote, leaf))
if err == nil {
mu.Lock()
dstList = append(dstList, dstObj)
mu.Unlock()
}
}
<-m.limiter
}(src)
}
wg.Wait()
}
// Work out what to do and do it
srcOnly, dstOnly, matches := matchListings(srcList, dstList, m.transforms)
for _, src := range srcOnly {
if m.aborting() {
return nil, m.Ctx.Err()
}
recurse := m.Callback.SrcOnly(src)
if recurse && job.srcDepth > 0 {
jobs = append(jobs, listDirJob{
srcRemote: src.Remote(),
dstRemote: src.Remote(),
srcDepth: job.srcDepth - 1,
noDst: true,
})
}
}
for _, dst := range dstOnly {
if m.aborting() {
return nil, m.Ctx.Err()
}
recurse := m.Callback.DstOnly(dst)
if recurse && job.dstDepth > 0 {
jobs = append(jobs, listDirJob{
srcRemote: dst.Remote(),
dstRemote: dst.Remote(),
dstDepth: job.dstDepth - 1,
noSrc: true,
})
}
}
for _, match := range matches {
if m.aborting() {
return nil, m.Ctx.Err()
}
recurse := m.Callback.Match(m.Ctx, match.dst, match.src)
if recurse && job.srcDepth > 0 && job.dstDepth > 0 {
jobs = append(jobs, listDirJob{
srcRemote: match.src.Remote(),
dstRemote: match.dst.Remote(),
srcDepth: job.srcDepth - 1,
dstDepth: job.dstDepth - 1,
})
}
}
return jobs, nil
}