+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: AGPL-3.0
+
package main
import (
+ "bytes"
+ "context"
+ "crypto/sha256"
"encoding/base64"
"encoding/hex"
"flag"
"fmt"
"io"
- "log"
+ "io/ioutil"
"net/http"
"os"
"regexp"
"strings"
"sync"
+ "sync/atomic"
"time"
"git.curoverse.com/arvados.git/sdk/go/arvados"
"github.com/AdRoll/goamz/s3"
)
+const (
+ s3DefaultReadTimeout = arvados.Duration(10 * time.Minute)
+ s3DefaultConnectTimeout = arvados.Duration(time.Minute)
+)
+
var (
// ErrS3TrashDisabled is returned by Trash if that operation
// is impossible with the current config.
s3RaceWindow time.Duration
s3ACL = s3.Private
+
+ zeroTime time.Time
)
const (
&s3UnsafeDelete,
"s3-unsafe-delete",
false,
- "EXPERIMENTAL. Enable deletion (garbage collection), even though there are known race conditions that can cause data loss.")
+ "EXPERIMENTAL. Enable deletion (garbage collection) even when trash lifetime is zero, even though there are known race conditions that can cause data loss.")
}
// S3Volume implements Volume using an S3 bucket.
RaceWindow arvados.Duration
ReadOnly bool
UnsafeDelete bool
+ StorageClasses []string
- bucket *s3.Bucket
+ bucket *s3bucket
startOnce sync.Once
}
// Zero timeouts mean "wait forever", which is a bad
// default. Default to long timeouts instead.
if v.ConnectTimeout == 0 {
- v.ConnectTimeout = arvados.Duration(time.Minute)
+ v.ConnectTimeout = s3DefaultConnectTimeout
}
if v.ReadTimeout == 0 {
- v.ReadTimeout = arvados.Duration(10 * time.Minute)
+ v.ReadTimeout = s3DefaultReadTimeout
}
client := s3.New(auth, region)
+ if region.EC2Endpoint.Signer == aws.V4Signature {
+ // Currently affects only eu-central-1
+ client.Signature = aws.V4Signature
+ }
client.ConnectTimeout = time.Duration(v.ConnectTimeout)
client.ReadTimeout = time.Duration(v.ReadTimeout)
- v.bucket = &s3.Bucket{
- S3: client,
- Name: v.Bucket,
+ v.bucket = &s3bucket{
+ Bucket: &s3.Bucket{
+ S3: client,
+ Name: v.Bucket,
+ },
}
return nil
}
+// DeviceID returns a globally unique ID for the storage bucket.
+func (v *S3Volume) DeviceID() string {
+ return "s3://" + v.Endpoint + "/" + v.Bucket
+}
+
+func (v *S3Volume) getReaderWithContext(ctx context.Context, loc string) (rdr io.ReadCloser, err error) {
+ ready := make(chan bool)
+ go func() {
+ rdr, err = v.getReader(loc)
+ close(ready)
+ }()
+ select {
+ case <-ready:
+ return
+ case <-ctx.Done():
+ theConfig.debugLogf("s3: abandoning getReader(): %s", ctx.Err())
+ go func() {
+ <-ready
+ if err == nil {
+ rdr.Close()
+ }
+ }()
+ return nil, ctx.Err()
+ }
+}
+
// getReader wraps (Bucket)GetReader.
//
// In situations where (Bucket)GetReader would fail because the block
if err == nil || !os.IsNotExist(err) {
return
}
+
_, err = v.bucket.Head("recent/"+loc, nil)
err = v.translateError(err)
if err != nil {
err = os.ErrNotExist
return
}
+
rdr, err = v.bucket.GetReader(loc)
if err != nil {
log.Printf("warning: reading %s after successful fixRace: %s", loc, err)
// Get a block: copy the block data into buf, and return the number of
// bytes copied.
-func (v *S3Volume) Get(loc string, buf []byte) (int, error) {
- rdr, err := v.getReader(loc)
+func (v *S3Volume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
+ rdr, err := v.getReaderWithContext(ctx, loc)
if err != nil {
return 0, err
}
- defer rdr.Close()
- n, err := io.ReadFull(rdr, buf)
- switch err {
- case nil, io.EOF, io.ErrUnexpectedEOF:
- return n, nil
- default:
- return 0, v.translateError(err)
+
+ var n int
+ ready := make(chan bool)
+ go func() {
+ defer close(ready)
+
+ defer rdr.Close()
+ n, err = io.ReadFull(rdr, buf)
+
+ switch err {
+ case nil, io.EOF, io.ErrUnexpectedEOF:
+ err = nil
+ default:
+ err = v.translateError(err)
+ }
+ }()
+ select {
+ case <-ctx.Done():
+ theConfig.debugLogf("s3: interrupting ReadFull() with Close() because %s", ctx.Err())
+ rdr.Close()
+ // Must wait for ReadFull to return, to ensure it
+ // doesn't write to buf after we return.
+ theConfig.debugLogf("s3: waiting for ReadFull() to fail")
+ <-ready
+ return 0, ctx.Err()
+ case <-ready:
+ return n, err
}
}
// Compare the given data with the stored data.
-func (v *S3Volume) Compare(loc string, expect []byte) error {
- rdr, err := v.getReader(loc)
+func (v *S3Volume) Compare(ctx context.Context, loc string, expect []byte) error {
+ errChan := make(chan error, 1)
+ go func() {
+ _, err := v.bucket.Head("recent/"+loc, nil)
+ errChan <- err
+ }()
+ var err error
+ select {
+ case <-ctx.Done():
+ return ctx.Err()
+ case err = <-errChan:
+ }
+ if err != nil {
+ // Checking for "loc" itself here would interfere with
+ // future GET requests.
+ //
+ // On AWS, if X doesn't exist, a HEAD or GET request
+ // for X causes X's non-existence to be cached. Thus,
+ // if we test for X, then create X and return a
+ // signature to our client, the client might still get
+ // 404 from all keepstores when trying to read it.
+ //
+ // To avoid this, we avoid doing HEAD X or GET X until
+ // we know X has been written.
+ //
+ // Note that X might exist even though recent/X
+ // doesn't: for example, the response to HEAD recent/X
+ // might itself come from a stale cache. In such
+ // cases, we will return a false negative and
+ // PutHandler might needlessly create another replica
+ // on a different volume. That's not ideal, but it's
+ // better than passing the eventually-consistent
+ // problem on to our clients.
+ return v.translateError(err)
+ }
+ rdr, err := v.getReaderWithContext(ctx, loc)
if err != nil {
return err
}
defer rdr.Close()
- return v.translateError(compareReaderWithBuf(rdr, expect, loc[:32]))
+ return v.translateError(compareReaderWithBuf(ctx, rdr, expect, loc[:32]))
}
// Put writes a block.
-func (v *S3Volume) Put(loc string, block []byte) error {
+func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error {
if v.ReadOnly {
return MethodDisabledError
}
var opts s3.Options
- if len(block) > 0 {
+ size := len(block)
+ if size > 0 {
md5, err := hex.DecodeString(loc)
if err != nil {
return err
}
opts.ContentMD5 = base64.StdEncoding.EncodeToString(md5)
+ // In AWS regions that use V4 signatures, we need to
+ // provide ContentSHA256 up front. Otherwise, the S3
+ // library reads the request body (from our buffer)
+ // into another new buffer in order to compute the
+ // SHA256 before sending the request -- which would
+ // mean consuming 128 MiB of memory for the duration
+ // of a 64 MiB write.
+ opts.ContentSHA256 = fmt.Sprintf("%x", sha256.Sum256(block))
}
- err := v.bucket.Put(loc, block, "application/octet-stream", s3ACL, opts)
- if err != nil {
+
+ // Send the block data through a pipe, so that (if we need to)
+ // we can close the pipe early and abandon our PutReader()
+ // goroutine, without worrying about PutReader() accessing our
+ // block buffer after we release it.
+ bufr, bufw := io.Pipe()
+ go func() {
+ io.Copy(bufw, bytes.NewReader(block))
+ bufw.Close()
+ }()
+
+ var err error
+ ready := make(chan bool)
+ go func() {
+ defer func() {
+ if ctx.Err() != nil {
+ theConfig.debugLogf("%s: abandoned PutReader goroutine finished with err: %s", v, err)
+ }
+ }()
+ defer close(ready)
+ err = v.bucket.PutReader(loc, bufr, int64(size), "application/octet-stream", s3ACL, opts)
+ if err != nil {
+ return
+ }
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
+ }()
+ select {
+ case <-ctx.Done():
+ theConfig.debugLogf("%s: taking PutReader's input away: %s", v, ctx.Err())
+ // Our pipe might be stuck in Write(), waiting for
+ // PutReader() to read. If so, un-stick it. This means
+ // PutReader will get corrupt data, but that's OK: the
+ // size and MD5 won't match, so the write will fail.
+ go io.Copy(ioutil.Discard, bufr)
+ // CloseWithError() will return once pending I/O is done.
+ bufw.CloseWithError(ctx.Err())
+ theConfig.debugLogf("%s: abandoning PutReader goroutine", v)
+ return ctx.Err()
+ case <-ready:
+ // Unblock pipe in case PutReader did not consume it.
+ io.Copy(ioutil.Discard, bufr)
return v.translateError(err)
}
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- return v.translateError(err)
}
// Touch sets the timestamp for the given locator to the current time.
} else if err != nil {
return err
}
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
return v.translateError(err)
}
err = v.translateError(err)
if os.IsNotExist(err) {
// The data object X exists, but recent/X is missing.
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
if err != nil {
log.Printf("error: creating %q: %s", "recent/"+loc, err)
return zeroTime, v.translateError(err)
func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
// Use a merge sort to find matching sets of X and recent/X.
dataL := s3Lister{
- Bucket: v.bucket,
+ Bucket: v.bucket.Bucket,
Prefix: prefix,
PageSize: v.IndexPageSize,
+ Stats: &v.bucket.stats,
}
recentL := s3Lister{
- Bucket: v.bucket,
+ Bucket: v.bucket.Bucket,
Prefix: "recent/" + prefix,
PageSize: v.IndexPageSize,
+ Stats: &v.bucket.stats,
}
- for data, recent := dataL.First(), recentL.First(); data != nil; data = dataL.Next() {
+ for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() {
if data.Key >= "g" {
// Conveniently, "recent/*" and "trash/*" are
// lexically greater than all hex-encoded data
stamp := data
// Advance to the corresponding recent/X marker, if any
- for recent != nil {
+ for recent != nil && recentL.Error() == nil {
if cmp := strings.Compare(recent.Key[7:], data.Key); cmp < 0 {
recent = recentL.Next()
continue
break
}
}
+ if err := recentL.Error(); err != nil {
+ return err
+ }
t, err := time.Parse(time.RFC3339, stamp.LastModified)
if err != nil {
return err
}
fmt.Fprintf(writer, "%s+%d %d\n", data.Key, data.Size, t.UnixNano())
}
- return nil
+ return dataL.Error()
}
// Trash a Keep block.
if !s3UnsafeDelete {
return ErrS3TrashDisabled
}
- return v.bucket.Del(loc)
+ return v.translateError(v.bucket.Del(loc))
}
err := v.checkRaceWindow(loc)
if err != nil {
MetadataDirective: "REPLACE",
}, v.bucket.Name+"/"+src)
err = v.translateError(err)
- if err != nil {
+ if os.IsNotExist(err) {
return err
+ } else if err != nil {
+ return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.Name+"/"+src, err)
}
if t, err := time.Parse(time.RFC3339Nano, resp.LastModified); err != nil {
return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.LastModified, err)
if err != nil {
return err
}
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
return v.translateError(err)
}
}
}
+// InternalStats returns bucket I/O and API call counters.
+func (v *S3Volume) InternalStats() interface{} {
+ return &v.bucket.stats
+}
+
// String implements fmt.Stringer.
func (v *S3Volume) String() string {
- return fmt.Sprintf("s3-bucket:%+q", v.bucket.Name)
+ return fmt.Sprintf("s3-bucket:%+q", v.Bucket)
}
// Writable returns false if all future Put, Mtime, and Delete calls
return v.S3Replication
}
+// GetStorageClasses implements Volume
+func (v *S3Volume) GetStorageClasses() []string {
+ return v.StorageClasses
+}
+
var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
func (v *S3Volume) isKeepBlock(s string) bool {
func (v *S3Volume) EmptyTrash() {
var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64
- // Use a merge sort to find matching sets of trash/X and recent/X.
- trashL := s3Lister{
- Bucket: v.bucket,
- Prefix: "trash/",
- PageSize: v.IndexPageSize,
- }
// Define "ready to delete" as "...when EmptyTrash started".
startT := time.Now()
- for trash := trashL.First(); trash != nil; trash = trashL.Next() {
+
+ emptyOneKey := func(trash *s3.Key) {
loc := trash.Key[6:]
if !v.isKeepBlock(loc) {
- continue
+ return
}
- bytesInTrash += trash.Size
- blocksInTrash++
+ atomic.AddInt64(&bytesInTrash, trash.Size)
+ atomic.AddInt64(&blocksInTrash, 1)
trashT, err := time.Parse(time.RFC3339, trash.LastModified)
if err != nil {
log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, trash.Key, trash.LastModified, err)
- continue
+ return
}
recent, err := v.bucket.Head("recent/"+loc, nil)
if err != nil && os.IsNotExist(v.translateError(err)) {
if err != nil {
log.Printf("error: %s: EmptyTrash: Untrash(%q): %s", v, loc, err)
}
- continue
+ return
} else if err != nil {
log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err)
- continue
+ return
}
recentT, err := v.lastModified(recent)
if err != nil {
log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, "recent/"+loc, recent.Header.Get("Last-Modified"), err)
- continue
+ return
}
if trashT.Sub(recentT) < theConfig.BlobSignatureTTL.Duration() {
if age := startT.Sub(recentT); age >= theConfig.BlobSignatureTTL.Duration()-time.Duration(v.RaceWindow) {
log.Printf("notice: %s: EmptyTrash: detected old race for %q, calling fixRace + Touch", v, loc)
v.fixRace(loc)
v.Touch(loc)
- continue
- } else if _, err := v.bucket.Head(loc, nil); os.IsNotExist(err) {
+ return
+ }
+ _, err := v.bucket.Head(loc, nil)
+ if os.IsNotExist(err) {
log.Printf("notice: %s: EmptyTrash: detected recent race for %q, calling fixRace", v, loc)
v.fixRace(loc)
- continue
+ return
} else if err != nil {
log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err)
- continue
+ return
}
}
if startT.Sub(trashT) < theConfig.TrashLifetime.Duration() {
- continue
+ return
}
err = v.bucket.Del(trash.Key)
if err != nil {
log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, trash.Key, err)
- continue
+ return
}
- bytesDeleted += trash.Size
- blocksDeleted++
+ atomic.AddInt64(&bytesDeleted, trash.Size)
+ atomic.AddInt64(&blocksDeleted, 1)
_, err = v.bucket.Head(loc, nil)
- if os.IsNotExist(err) {
- err = v.bucket.Del("recent/" + loc)
- if err != nil {
- log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err)
- }
- } else if err != nil {
- log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err)
+ if err == nil {
+ log.Printf("warning: %s: EmptyTrash: HEAD %q succeeded immediately after deleting %q", v, loc, loc)
+ return
+ }
+ if !os.IsNotExist(v.translateError(err)) {
+ log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err)
+ return
+ }
+ err = v.bucket.Del("recent/" + loc)
+ if err != nil {
+ log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err)
}
}
+
+ var wg sync.WaitGroup
+ todo := make(chan *s3.Key, theConfig.EmptyTrashWorkers)
+ for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
+ wg.Add(1)
+ go func() {
+ defer wg.Done()
+ for key := range todo {
+ emptyOneKey(key)
+ }
+ }()
+ }
+
+ trashL := s3Lister{
+ Bucket: v.bucket.Bucket,
+ Prefix: "trash/",
+ PageSize: v.IndexPageSize,
+ Stats: &v.bucket.stats,
+ }
+ for trash := trashL.First(); trash != nil; trash = trashL.Next() {
+ todo <- trash
+ }
+ close(todo)
+ wg.Wait()
+
if err := trashL.Error(); err != nil {
log.Printf("error: %s: EmptyTrash: lister: %s", v, err)
}
Bucket *s3.Bucket
Prefix string
PageSize int
+ Stats *s3bucketStats
nextMarker string
buf []s3.Key
err error
}
func (lister *s3Lister) getPage() {
+ lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps)
resp, err := lister.Bucket.List(lister.Prefix, "", lister.nextMarker, lister.PageSize)
lister.nextMarker = ""
if err != nil {
}
return
}
+
+// s3bucket wraps s3.bucket and counts I/O and API usage stats.
+type s3bucket struct {
+ *s3.Bucket
+ stats s3bucketStats
+}
+
+func (b *s3bucket) GetReader(path string) (io.ReadCloser, error) {
+ rdr, err := b.Bucket.GetReader(path)
+ b.stats.Tick(&b.stats.Ops, &b.stats.GetOps)
+ b.stats.TickErr(err)
+ return NewCountingReader(rdr, b.stats.TickInBytes), err
+}
+
+func (b *s3bucket) Head(path string, headers map[string][]string) (*http.Response, error) {
+ resp, err := b.Bucket.Head(path, headers)
+ b.stats.Tick(&b.stats.Ops, &b.stats.HeadOps)
+ b.stats.TickErr(err)
+ return resp, err
+}
+
+func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error {
+ if length == 0 {
+ // goamz will only send Content-Length: 0 when reader
+ // is nil due to net.http.Request.ContentLength
+ // behavior. Otherwise, Content-Length header is
+ // omitted which will cause some S3 services
+ // (including AWS and Ceph RadosGW) to fail to create
+ // empty objects.
+ r = nil
+ } else {
+ r = NewCountingReader(r, b.stats.TickOutBytes)
+ }
+ err := b.Bucket.PutReader(path, r, length, contType, perm, options)
+ b.stats.Tick(&b.stats.Ops, &b.stats.PutOps)
+ b.stats.TickErr(err)
+ return err
+}
+
+func (b *s3bucket) Del(path string) error {
+ err := b.Bucket.Del(path)
+ b.stats.Tick(&b.stats.Ops, &b.stats.DelOps)
+ b.stats.TickErr(err)
+ return err
+}
+
+type s3bucketStats struct {
+ statsTicker
+ Ops uint64
+ GetOps uint64
+ PutOps uint64
+ HeadOps uint64
+ DelOps uint64
+ ListOps uint64
+}
+
+func (s *s3bucketStats) TickErr(err error) {
+ if err == nil {
+ return
+ }
+ errType := fmt.Sprintf("%T", err)
+ if err, ok := err.(*s3.Error); ok {
+ errType = errType + fmt.Sprintf(" %d %s", err.StatusCode, err.Code)
+ }
+ s.statsTicker.TickErr(err, errType)
+}