package main
import (
+ "bytes"
+ "context"
"encoding/base64"
"encoding/hex"
"flag"
"fmt"
"io"
+ "io/ioutil"
"log"
"net/http"
"os"
"regexp"
"strings"
"sync"
+ "sync/atomic"
"time"
"git.curoverse.com/arvados.git/sdk/go/arvados"
ReadOnly bool
UnsafeDelete bool
- bucket *s3.Bucket
+ bucket *s3.Bucket
+ bucketStats bucketStats
+ volumeStats ioStats
startOnce sync.Once
}
+type bucketStats struct {
+ Errors uint64
+ Ops uint64
+ GetOps uint64
+ PutOps uint64
+ HeadOps uint64
+ DelOps uint64
+ InBytes uint64
+ OutBytes uint64
+
+ ErrorCodes map[string]uint64 `json:",omitempty"`
+
+ lock sync.Mutex
+}
+
// Examples implements VolumeWithExamples.
func (*S3Volume) Examples() []Volume {
return []Volume{
return nil
}
+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
// disappeared in a Trash race, getReader calls fixRace to recover the
// data, and tries again.
func (v *S3Volume) getReader(loc string) (rdr io.ReadCloser, err error) {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.GetOps)
rdr, err = v.bucket.GetReader(loc)
err = v.translateError(err)
- if err == nil || !os.IsNotExist(err) {
+ if err == nil {
+ rdr = NewCountingReader(rdr, v.tickInBytes)
+ return
+ } else if !os.IsNotExist(v.tickErr(err)) {
return
}
+
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err = v.bucket.Head("recent/"+loc, nil)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
if err != nil {
// If we can't read recent/X, there's no point in
// trying fixRace. Give up.
err = os.ErrNotExist
return
}
+
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.GetOps)
rdr, err = v.bucket.GetReader(loc)
if err != nil {
log.Printf("warning: reading %s after successful fixRace: %s", loc, err)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
}
+ rdr = NewCountingReader(rdr, v.tickInBytes)
return
}
// 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 {
+ 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)
}
- 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)
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
+ rdr := NewCountingReader(bufr, v.tickOutBytes)
+ err = v.bucket.PutReader(loc, rdr, int64(size), "application/octet-stream", s3ACL, opts)
+ if err != nil {
+ v.tickErr(err)
+ return
+ }
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
+ err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
+ v.tickErr(err)
+ }()
+ 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
+ // io.Copy() 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:
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.
if v.ReadOnly {
return MethodDisabledError
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err := v.bucket.Head(loc, nil)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
if os.IsNotExist(err) && v.fixRace(loc) {
// The data object got trashed in a race, but fixRace
// rescued it.
} else if err != nil {
return err
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- return v.translateError(err)
+ return v.translateError(v.tickErr(err))
}
// Mtime returns the stored timestamp for the given locator.
func (v *S3Volume) Mtime(loc string) (time.Time, error) {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err := v.bucket.Head(loc, nil)
if err != nil {
- return zeroTime, v.translateError(err)
+ return zeroTime, v.translateError(v.tickErr(err))
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err := v.bucket.Head("recent/"+loc, nil)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
if os.IsNotExist(err) {
// The data object X exists, but recent/X is missing.
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
if err != nil {
log.Printf("error: creating %q: %s", "recent/"+loc, err)
- return zeroTime, v.translateError(err)
+ return zeroTime, v.translateError(v.tickErr(err))
}
log.Printf("info: created %q to migrate existing block to new storage scheme", "recent/"+loc)
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err = v.bucket.Head("recent/"+loc, nil)
if err != nil {
log.Printf("error: created %q but HEAD failed: %s", "recent/"+loc, err)
- return zeroTime, v.translateError(err)
+ return zeroTime, v.translateError(v.tickErr(err))
}
} else if err != nil {
// HEAD recent/X failed for some other reason.
if !s3UnsafeDelete {
return ErrS3TrashDisabled
}
- return v.bucket.Del(loc)
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
+ return v.translateError(v.tickErr(v.bucket.Del(loc)))
}
err := v.checkRaceWindow(loc)
if err != nil {
if err != nil {
return err
}
- return v.translateError(v.bucket.Del(loc))
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
+ return v.translateError(v.tickErr(v.bucket.Del(loc)))
}
// checkRaceWindow returns a non-nil error if trash/loc is, or might
// be, in the race window (i.e., it's not safe to trash loc).
func (v *S3Volume) checkRaceWindow(loc string) error {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err := v.bucket.Head("trash/"+loc, nil)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
if os.IsNotExist(err) {
// OK, trash/X doesn't exist so we're not in the race
// window
// (PutCopy returns 200 OK if the request was received, even if the
// copy failed).
func (v *S3Volume) safeCopy(dst, src string) error {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
resp, err := v.bucket.PutCopy(dst, s3ACL, s3.CopyOptions{
ContentType: "application/octet-stream",
MetadataDirective: "REPLACE",
}, v.bucket.Name+"/"+src)
- err = v.translateError(err)
+ err = v.translateError(v.tickErr(err))
if err != nil {
return err
}
if err != nil {
return err
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- return v.translateError(err)
+ return v.translateError(v.tickErr(err))
}
// Status returns a *VolumeStatus representing the current in-use
}
}
+// IOStatus implements InternalStatser.
+func (v *S3Volume) InternalStats() interface{} {
+ return &v.bucketStats
+}
+
// 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
// there was a race between Put and Trash, fixRace recovers from the
// race by Untrashing the block.
func (v *S3Volume) fixRace(loc string) bool {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
trash, err := v.bucket.Head("trash/"+loc, nil)
if err != nil {
- if !os.IsNotExist(v.translateError(err)) {
+ if !os.IsNotExist(v.translateError(v.tickErr(err))) {
log.Printf("error: fixRace: HEAD %q: %s", "trash/"+loc, err)
}
return false
return false
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
recent, err := v.bucket.Head("recent/"+loc, nil)
if err != nil {
+ v.tickErr(err)
log.Printf("error: fixRace: HEAD %q: %s", "recent/"+loc, err)
return false
}
log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, trash.Key, trash.LastModified, err)
continue
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
recent, err := v.bucket.Head("recent/"+loc, nil)
- if err != nil && os.IsNotExist(v.translateError(err)) {
+ if err != nil && os.IsNotExist(v.translateError(v.tickErr(err))) {
log.Printf("warning: %s: EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", v, trash.Key, "recent/"+loc, err)
err = v.Untrash(loc)
if err != nil {
v.fixRace(loc)
v.Touch(loc)
continue
- } else if _, err := v.bucket.Head(loc, nil); os.IsNotExist(err) {
+ }
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
+ _, err := v.bucket.Head(loc, nil)
+ if os.IsNotExist(v.tickErr(err)) {
log.Printf("notice: %s: EmptyTrash: detected recent race for %q, calling fixRace", v, loc)
v.fixRace(loc)
continue
if startT.Sub(trashT) < theConfig.TrashLifetime.Duration() {
continue
}
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
err = v.bucket.Del(trash.Key)
if err != nil {
+ v.tickErr(err)
log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, trash.Key, err)
continue
}
bytesDeleted += trash.Size
blocksDeleted++
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err = v.bucket.Head(loc, nil)
- if os.IsNotExist(err) {
+ if os.IsNotExist(v.tickErr(err)) {
+ v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
err = v.bucket.Del("recent/" + loc)
if err != nil {
+ v.tickErr(err)
log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err)
}
} else if err != nil {
log.Printf("EmptyTrash stats for %v: Deleted %v bytes in %v blocks. Remaining in trash: %v bytes in %v blocks.", v.String(), bytesDeleted, blocksDeleted, bytesInTrash-bytesDeleted, blocksInTrash-blocksDeleted)
}
+func (v *S3Volume) tick(counters ...*uint64) {
+ for _, counter := range counters {
+ atomic.AddUint64(counter, 1)
+ }
+}
+
+func (v *S3Volume) tickErr(err error) error {
+ if err == nil {
+ return nil
+ }
+ atomic.AddUint64(&v.bucketStats.Errors, 1)
+ if err, ok := err.(*s3.Error); ok {
+ errStr := fmt.Sprintf("%d %s", err.StatusCode, err.Code)
+ v.bucketStats.lock.Lock()
+ if v.bucketStats.ErrorCodes == nil {
+ v.bucketStats.ErrorCodes = make(map[string]uint64)
+ }
+ v.bucketStats.ErrorCodes[errStr]++
+ v.bucketStats.lock.Unlock()
+ }
+ return err
+}
+
+func (v *S3Volume) tickInBytes(n uint64) {
+ atomic.AddUint64(&v.bucketStats.InBytes, n)
+}
+
+func (v *S3Volume) tickOutBytes(n uint64) {
+ atomic.AddUint64(&v.bucketStats.OutBytes, n)
+}
+
type s3Lister struct {
Bucket *s3.Bucket
Prefix string
projects / arvados.git / blobdiff