X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/1416b0952adc0bfee85e15d9c86a51c32fcfd003..022107bd52092c658208e74161581c6bedda4a5f:/services/keepstore/s3_volume.go diff --git a/services/keepstore/s3_volume.go b/services/keepstore/s3_volume.go index f78084dbcf..d4b90540ea 100644 --- a/services/keepstore/s3_volume.go +++ b/services/keepstore/s3_volume.go @@ -2,18 +2,17 @@ // // SPDX-License-Identifier: AGPL-3.0 -package main +package keepstore import ( "bytes" "context" "encoding/base64" "encoding/hex" - "flag" + "encoding/json" + "errors" "fmt" "io" - "io/ioutil" - "net/http" "os" "regexp" "strings" @@ -21,843 +20,309 @@ import ( "sync/atomic" "time" - "git.curoverse.com/arvados.git/sdk/go/arvados" - "github.com/AdRoll/goamz/aws" - "github.com/AdRoll/goamz/s3" + "git.arvados.org/arvados.git/sdk/go/arvados" + "github.com/aws/aws-sdk-go-v2/aws" + "github.com/aws/aws-sdk-go-v2/aws/awserr" + "github.com/aws/aws-sdk-go-v2/aws/defaults" + "github.com/aws/aws-sdk-go-v2/aws/ec2metadata" + "github.com/aws/aws-sdk-go-v2/aws/ec2rolecreds" + "github.com/aws/aws-sdk-go-v2/aws/endpoints" + "github.com/aws/aws-sdk-go-v2/service/s3" + "github.com/aws/aws-sdk-go-v2/service/s3/s3manager" + "github.com/prometheus/client_golang/prometheus" + "github.com/sirupsen/logrus" ) +func init() { + driver["S3"] = news3Volume +} + const ( - s3DefaultReadTimeout = arvados.Duration(10 * time.Minute) - s3DefaultConnectTimeout = arvados.Duration(time.Minute) + s3DefaultReadTimeout = arvados.Duration(10 * time.Minute) + s3DefaultConnectTimeout = arvados.Duration(time.Minute) + maxClockSkew = 600 * time.Second + nearlyRFC1123 = "Mon, 2 Jan 2006 15:04:05 GMT" + s3downloaderPartSize = 5 * 1024 * 1024 + s3downloaderReadConcurrency = 13 + s3uploaderPartSize = 5 * 1024 * 1024 + s3uploaderWriteConcurrency = 5 ) var ( - // ErrS3TrashDisabled is returned by Trash if that operation - // is impossible with the current config. - ErrS3TrashDisabled = fmt.Errorf("trash function is disabled because -trash-lifetime=0 and -s3-unsafe-delete=false") - - s3AccessKeyFile string - s3SecretKeyFile string - s3RegionName string - s3Endpoint string - s3Replication int - s3UnsafeDelete bool - s3RaceWindow time.Duration - - s3ACL = s3.Private - - zeroTime time.Time + errS3TrashDisabled = fmt.Errorf("trash function is disabled because Collections.BlobTrashLifetime=0 and DriverParameters.UnsafeDelete=false") + s3AWSKeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) + s3AWSZeroTime time.Time ) -const ( - maxClockSkew = 600 * time.Second - nearlyRFC1123 = "Mon, 2 Jan 2006 15:04:05 GMT" -) - -type s3VolumeAdder struct { - *Config -} - -// String implements flag.Value -func (s *s3VolumeAdder) String() string { - return "-" -} - -func (s *s3VolumeAdder) Set(bucketName string) error { - if bucketName == "" { - return fmt.Errorf("no container name given") - } - if s3AccessKeyFile == "" || s3SecretKeyFile == "" { - return fmt.Errorf("-s3-access-key-file and -s3-secret-key-file arguments must given before -s3-bucket-volume") - } - if deprecated.flagSerializeIO { - log.Print("Notice: -serialize is not supported by s3-bucket volumes.") - } - s.Config.Volumes = append(s.Config.Volumes, &S3Volume{ - Bucket: bucketName, - AccessKeyFile: s3AccessKeyFile, - SecretKeyFile: s3SecretKeyFile, - Endpoint: s3Endpoint, - Region: s3RegionName, - RaceWindow: arvados.Duration(s3RaceWindow), - S3Replication: s3Replication, - UnsafeDelete: s3UnsafeDelete, - ReadOnly: deprecated.flagReadonly, - IndexPageSize: 1000, - }) - return nil -} - -func s3regions() (okList []string) { - for r := range aws.Regions { - okList = append(okList, r) - } - return -} - -func init() { - VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &S3Volume{} }) - - flag.Var(&s3VolumeAdder{theConfig}, - "s3-bucket-volume", - "Use the given bucket as a storage volume. Can be given multiple times.") - flag.StringVar( - &s3RegionName, - "s3-region", - "", - fmt.Sprintf("AWS region used for subsequent -s3-bucket-volume arguments. Allowed values are %+q.", s3regions())) - flag.StringVar( - &s3Endpoint, - "s3-endpoint", - "", - "Endpoint URL used for subsequent -s3-bucket-volume arguments. If blank, use the AWS endpoint corresponding to the -s3-region argument. For Google Storage, use \"https://storage.googleapis.com\".") - flag.StringVar( - &s3AccessKeyFile, - "s3-access-key-file", - "", - "`File` containing the access key used for subsequent -s3-bucket-volume arguments.") - flag.StringVar( - &s3SecretKeyFile, - "s3-secret-key-file", - "", - "`File` containing the secret key used for subsequent -s3-bucket-volume arguments.") - flag.DurationVar( - &s3RaceWindow, - "s3-race-window", - 24*time.Hour, - "Maximum eventual consistency latency for subsequent -s3-bucket-volume arguments.") - flag.IntVar( - &s3Replication, - "s3-replication", - 2, - "Replication level reported to clients for subsequent -s3-bucket-volume arguments.") - flag.BoolVar( - &s3UnsafeDelete, - "s3-unsafe-delete", - false, - "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. -type S3Volume struct { - AccessKeyFile string - SecretKeyFile string - Endpoint string - Region string - Bucket string - LocationConstraint bool - IndexPageSize int - S3Replication int - ConnectTimeout arvados.Duration - ReadTimeout arvados.Duration - RaceWindow arvados.Duration - ReadOnly bool - UnsafeDelete bool - StorageClasses []string - - bucket *s3bucket - - startOnce sync.Once -} - -// Examples implements VolumeWithExamples. -func (*S3Volume) Examples() []Volume { - return []Volume{ - &S3Volume{ - AccessKeyFile: "/etc/aws_s3_access_key.txt", - SecretKeyFile: "/etc/aws_s3_secret_key.txt", - Endpoint: "", - Region: "us-east-1", - Bucket: "example-bucket-name", - IndexPageSize: 1000, - S3Replication: 2, - RaceWindow: arvados.Duration(24 * time.Hour), - ConnectTimeout: arvados.Duration(time.Minute), - ReadTimeout: arvados.Duration(5 * time.Minute), - }, - &S3Volume{ - AccessKeyFile: "/etc/gce_s3_access_key.txt", - SecretKeyFile: "/etc/gce_s3_secret_key.txt", - Endpoint: "https://storage.googleapis.com", - Region: "", - Bucket: "example-bucket-name", - IndexPageSize: 1000, - S3Replication: 2, - RaceWindow: arvados.Duration(24 * time.Hour), - ConnectTimeout: arvados.Duration(time.Minute), - ReadTimeout: arvados.Duration(5 * time.Minute), - }, - } -} - -// Type implements Volume. -func (*S3Volume) Type() string { - return "S3" -} - -// Start populates private fields and verifies the configuration is -// valid. -func (v *S3Volume) Start() error { - region, ok := aws.Regions[v.Region] - if v.Endpoint == "" { - if !ok { - return fmt.Errorf("unrecognized region %+q; try specifying -s3-endpoint instead", v.Region) - } - } else if ok { - return fmt.Errorf("refusing to use AWS region name %+q with endpoint %+q; "+ - "specify empty endpoint (\"-s3-endpoint=\") or use a different region name", v.Region, v.Endpoint) +// s3Volume implements Volume using an S3 bucket. +type s3Volume struct { + arvados.S3VolumeDriverParameters + AuthToken string // populated automatically when IAMRole is used + AuthExpiration time.Time // populated automatically when IAMRole is used + + cluster *arvados.Cluster + volume arvados.Volume + logger logrus.FieldLogger + metrics *volumeMetricsVecs + bufferPool *bufferPool + bucket *s3Bucket + region string + startOnce sync.Once +} + +// s3bucket wraps s3.bucket and counts I/O and API usage stats. The +// wrapped bucket can be replaced atomically with SetBucket in order +// to update credentials. +type s3Bucket struct { + bucket string + svc *s3.Client + stats s3awsbucketStats + mu sync.Mutex +} + +func (v *s3Volume) isKeepBlock(s string) (string, bool) { + if v.PrefixLength > 0 && len(s) == v.PrefixLength+33 && s[:v.PrefixLength] == s[v.PrefixLength+1:v.PrefixLength*2+1] { + s = s[v.PrefixLength+1:] + } + return s, s3AWSKeepBlockRegexp.MatchString(s) +} + +// Return the key used for a given loc. If PrefixLength==0 then +// key("abcdef0123") is "abcdef0123", if PrefixLength==3 then key is +// "abc/abcdef0123", etc. +func (v *s3Volume) key(loc string) string { + if v.PrefixLength > 0 && v.PrefixLength < len(loc)-1 { + return loc[:v.PrefixLength] + "/" + loc } else { - region = aws.Region{ - Name: v.Region, - S3Endpoint: v.Endpoint, - S3LocationConstraint: v.LocationConstraint, - } + return loc } +} - var err error - var auth aws.Auth - auth.AccessKey, err = readKeyFromFile(v.AccessKeyFile) - if err != nil { - return err +func news3Volume(params newVolumeParams) (volume, error) { + v := &s3Volume{ + cluster: params.Cluster, + volume: params.ConfigVolume, + metrics: params.MetricsVecs, + bufferPool: params.BufferPool, } - auth.SecretKey, err = readKeyFromFile(v.SecretKeyFile) + err := json.Unmarshal(params.ConfigVolume.DriverParameters, v) if err != nil { - return err - } - - // Zero timeouts mean "wait forever", which is a bad - // default. Default to long timeouts instead. - if v.ConnectTimeout == 0 { - v.ConnectTimeout = s3DefaultConnectTimeout + return nil, err } - if v.ReadTimeout == 0 { - 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 = &s3bucket{ - Bucket: &s3.Bucket{ - S3: client, - Name: v.Bucket, - }, - } - return nil + v.logger = params.Logger.WithField("Volume", v.DeviceID()) + return v, v.check("") } -// 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() +func (v *s3Volume) translateError(err error) error { + if _, ok := err.(*aws.RequestCanceledError); ok { + return context.Canceled + } else if aerr, ok := err.(awserr.Error); ok { + if aerr.Code() == "NotFound" { + return os.ErrNotExist + } else if aerr.Code() == "NoSuchKey" { + return os.ErrNotExist + } } + return err } -// getReader wraps (Bucket)GetReader. +// safeCopy calls CopyObjectRequest, and checks the response to make +// sure the copy succeeded and updated the timestamp on the +// destination object // -// 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) { - rdr, err = v.bucket.GetReader(loc) - err = v.translateError(err) - if err == nil || !os.IsNotExist(err) { - return +// (If something goes wrong during the copy, the error will be +// embedded in the 200 OK response) +func (v *s3Volume) safeCopy(dst, src string) error { + input := &s3.CopyObjectInput{ + Bucket: aws.String(v.bucket.bucket), + ContentType: aws.String("application/octet-stream"), + CopySource: aws.String(v.bucket.bucket + "/" + src), + Key: aws.String(dst), } - _, err = v.bucket.Head("recent/"+loc, nil) + req := v.bucket.svc.CopyObjectRequest(input) + resp, err := req.Send(context.Background()) + err = v.translateError(err) - if err != nil { - // If we can't read recent/X, there's no point in - // trying fixRace. Give up. - return - } - if !v.fixRace(loc) { - err = os.ErrNotExist - return + if os.IsNotExist(err) { + return err + } else if err != nil { + return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.bucket+"/"+src, err) } - rdr, err = v.bucket.GetReader(loc) - if err != nil { - log.Printf("warning: reading %s after successful fixRace: %s", loc, err) - err = v.translateError(err) + if resp.CopyObjectResult.LastModified == nil { + return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.CopyObjectResult.LastModified, err) + } else if time.Now().Sub(*resp.CopyObjectResult.LastModified) > maxClockSkew { + return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.CopyObjectResult.LastModified, resp.CopyObjectResult.LastModified) } - return + return nil } -// Get a block: copy the block data into buf, and return the number of -// bytes copied. -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 +func (v *s3Volume) check(ec2metadataHostname string) error { + if v.Bucket == "" { + return errors.New("DriverParameters: Bucket must be provided") } - - 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(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 + if v.IndexPageSize == 0 { + v.IndexPageSize = 1000 } - defer rdr.Close() - return v.translateError(compareReaderWithBuf(ctx, rdr, expect, loc[:32])) -} - -// Put writes a block. -func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error { - if v.ReadOnly { - return MethodDisabledError - } - var opts s3.Options - size := len(block) - if size > 0 { - md5, err := hex.DecodeString(loc) - if err != nil { - return err - } - opts.ContentMD5 = base64.StdEncoding.EncodeToString(md5) - } - - // 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) - } -} - -// Touch sets the timestamp for the given locator to the current time. -func (v *S3Volume) Touch(loc string) error { - if v.ReadOnly { - return MethodDisabledError - } - _, err := v.bucket.Head(loc, nil) - err = v.translateError(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 + if v.RaceWindow < 0 { + return errors.New("DriverParameters: RaceWindow must not be negative") } - err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) - return v.translateError(err) -} -// Mtime returns the stored timestamp for the given locator. -func (v *S3Volume) Mtime(loc string) (time.Time, error) { - _, err := v.bucket.Head(loc, nil) - if err != nil { - return zeroTime, v.translateError(err) + if v.V2Signature { + return errors.New("DriverParameters: V2Signature is not supported") } - resp, err := v.bucket.Head("recent/"+loc, nil) - err = v.translateError(err) - if os.IsNotExist(err) { - // The data object X exists, but recent/X is missing. - 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) - } - log.Printf("info: created %q to migrate existing block to new storage scheme", "recent/"+loc) - 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) - } - } else if err != nil { - // HEAD recent/X failed for some other reason. - return zeroTime, err - } - return v.lastModified(resp) -} -// IndexTo writes a complete list of locators with the given prefix -// for which Get() can retrieve data. -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, - Prefix: prefix, - PageSize: v.IndexPageSize, - } - recentL := s3Lister{ - Bucket: v.bucket.Bucket, - Prefix: "recent/" + prefix, - PageSize: v.IndexPageSize, - } - v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps) - v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps) - for data, recent := dataL.First(), recentL.First(); data != nil; data = dataL.Next() { - v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps) - if data.Key >= "g" { - // Conveniently, "recent/*" and "trash/*" are - // lexically greater than all hex-encoded data - // hashes, so stopping here avoids iterating - // over all of them needlessly with dataL. - break - } - if !v.isKeepBlock(data.Key) { - continue - } + defaultResolver := endpoints.NewDefaultResolver() - // stamp is the list entry we should use to report the - // last-modified time for this data block: it will be - // the recent/X entry if one exists, otherwise the - // entry for the data block itself. - stamp := data + cfg := defaults.Config() - // Advance to the corresponding recent/X marker, if any - for recent != nil { - if cmp := strings.Compare(recent.Key[7:], data.Key); cmp < 0 { - recent = recentL.Next() - v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps) - continue - } else if cmp == 0 { - stamp = recent - recent = recentL.Next() - v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps) - break + if v.Endpoint == "" && v.Region == "" { + return fmt.Errorf("AWS region or endpoint must be specified") + } else if v.Endpoint != "" || ec2metadataHostname != "" { + myCustomResolver := func(service, region string) (aws.Endpoint, error) { + if v.Endpoint != "" && service == "s3" { + return aws.Endpoint{ + URL: v.Endpoint, + SigningRegion: region, + }, nil + } else if service == "ec2metadata" && ec2metadataHostname != "" { + return aws.Endpoint{ + URL: ec2metadataHostname, + }, nil } else { - // recent/X marker is missing: we'll - // use the timestamp on the data - // object. - break + return defaultResolver.ResolveEndpoint(service, region) } } - 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 -} - -// Trash a Keep block. -func (v *S3Volume) Trash(loc string) error { - if v.ReadOnly { - return MethodDisabledError - } - if t, err := v.Mtime(loc); err != nil { - return err - } else if time.Since(t) < theConfig.BlobSignatureTTL.Duration() { - return nil - } - if theConfig.TrashLifetime == 0 { - if !s3UnsafeDelete { - return ErrS3TrashDisabled - } - return v.translateError(v.bucket.Del(loc)) - } - err := v.checkRaceWindow(loc) - if err != nil { - return err - } - err = v.safeCopy("trash/"+loc, loc) - if err != nil { - return err - } - return v.translateError(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 { - resp, err := v.bucket.Head("trash/"+loc, nil) - err = v.translateError(err) - if os.IsNotExist(err) { - // OK, trash/X doesn't exist so we're not in the race - // window - return nil - } else if err != nil { - // Error looking up trash/X. We don't know whether - // we're in the race window - return err + cfg.EndpointResolver = aws.EndpointResolverFunc(myCustomResolver) } - t, err := v.lastModified(resp) - if err != nil { - // Can't parse timestamp - return err + if v.Region == "" { + // Endpoint is already specified (otherwise we would + // have errored out above), but Region is also + // required by the aws sdk, in order to determine + // SignatureVersions. + v.Region = "us-east-1" } - safeWindow := t.Add(theConfig.TrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow))) - if safeWindow <= 0 { - // We can't count on "touch trash/X" to prolong - // trash/X's lifetime. The new timestamp might not - // become visible until now+raceWindow, and EmptyTrash - // is allowed to delete trash/X before then. - return fmt.Errorf("same block is already in trash, and safe window ended %s ago", -safeWindow) - } - // trash/X exists, but it won't be eligible for deletion until - // after now+raceWindow, so it's safe to overwrite it. - return nil -} + cfg.Region = v.Region -// safeCopy calls PutCopy, and checks the response to make sure the -// copy succeeded and updated the timestamp on the destination object -// (PutCopy returns 200 OK if the request was received, even if the -// copy failed). -func (v *S3Volume) safeCopy(dst, src string) error { - resp, err := v.bucket.PutCopy(dst, s3ACL, s3.CopyOptions{ - ContentType: "application/octet-stream", - MetadataDirective: "REPLACE", - }, v.bucket.Name+"/"+src) - err = v.translateError(err) - if err != nil { - return 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) - } else if time.Now().Sub(t) > maxClockSkew { - return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.LastModified, t) - } - return nil -} - -// Get the LastModified header from resp, and parse it as RFC1123 or -// -- if it isn't valid RFC1123 -- as Amazon's variant of RFC1123. -func (v *S3Volume) lastModified(resp *http.Response) (t time.Time, err error) { - s := resp.Header.Get("Last-Modified") - t, err = time.Parse(time.RFC1123, s) - if err != nil && s != "" { - // AWS example is "Sun, 1 Jan 2006 12:00:00 GMT", - // which isn't quite "Sun, 01 Jan 2006 12:00:00 GMT" - // as required by HTTP spec. If it's not a valid HTTP - // header value, it's probably AWS (or s3test) giving - // us a nearly-RFC1123 timestamp. - t, err = time.Parse(nearlyRFC1123, s) - } - return -} - -// Untrash moves block from trash back into store -func (v *S3Volume) Untrash(loc string) error { - err := v.safeCopy(loc, "trash/"+loc) - if err != nil { - return err + // Zero timeouts mean "wait forever", which is a bad + // default. Default to long timeouts instead. + if v.ConnectTimeout == 0 { + v.ConnectTimeout = s3DefaultConnectTimeout } - err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) - return v.translateError(err) -} - -// Status returns a *VolumeStatus representing the current in-use -// storage capacity and a fake available capacity that doesn't make -// the volume seem full or nearly-full. -func (v *S3Volume) Status() *VolumeStatus { - return &VolumeStatus{ - DeviceNum: 1, - BytesFree: BlockSize * 1000, - BytesUsed: 1, + if v.ReadTimeout == 0 { + v.ReadTimeout = s3DefaultReadTimeout } -} -// 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) -} + creds := aws.NewChainProvider( + []aws.CredentialsProvider{ + aws.NewStaticCredentialsProvider(v.AccessKeyID, v.SecretAccessKey, v.AuthToken), + ec2rolecreds.New(ec2metadata.New(cfg)), + }) -// Writable returns false if all future Put, Mtime, and Delete calls -// are expected to fail. -func (v *S3Volume) Writable() bool { - return !v.ReadOnly -} - -// Replication returns the storage redundancy of the underlying -// device. Configured via command line flag. -func (v *S3Volume) Replication() int { - 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 { - return s3KeepBlockRegexp.MatchString(s) -} - -// fixRace(X) is called when "recent/X" exists but "X" doesn't -// exist. If the timestamps on "recent/"+loc and "trash/"+loc indicate -// there was a race between Put and Trash, fixRace recovers from the -// race by Untrashing the block. -func (v *S3Volume) fixRace(loc string) bool { - trash, err := v.bucket.Head("trash/"+loc, nil) - if err != nil { - if !os.IsNotExist(v.translateError(err)) { - log.Printf("error: fixRace: HEAD %q: %s", "trash/"+loc, err) - } - return false - } - trashTime, err := v.lastModified(trash) - if err != nil { - log.Printf("error: fixRace: parse %q: %s", trash.Header.Get("Last-Modified"), err) - return false - } + cfg.Credentials = creds - recent, err := v.bucket.Head("recent/"+loc, nil) - if err != nil { - log.Printf("error: fixRace: HEAD %q: %s", "recent/"+loc, err) - return false - } - recentTime, err := v.lastModified(recent) - if err != nil { - log.Printf("error: fixRace: parse %q: %s", recent.Header.Get("Last-Modified"), err) - return false + v.bucket = &s3Bucket{ + bucket: v.Bucket, + svc: s3.New(cfg), } - ageWhenTrashed := trashTime.Sub(recentTime) - if ageWhenTrashed >= theConfig.BlobSignatureTTL.Duration() { - // No evidence of a race: block hasn't been written - // since it became eligible for Trash. No fix needed. - return false - } + // Set up prometheus metrics + lbls := prometheus.Labels{"device_id": v.DeviceID()} + v.bucket.stats.opsCounters, v.bucket.stats.errCounters, v.bucket.stats.ioBytes = v.metrics.getCounterVecsFor(lbls) - log.Printf("notice: fixRace: %q: trashed at %s but touched at %s (age when trashed = %s < %s)", loc, trashTime, recentTime, ageWhenTrashed, theConfig.BlobSignatureTTL) - log.Printf("notice: fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+loc, loc) - err = v.safeCopy(loc, "trash/"+loc) - if err != nil { - log.Printf("error: fixRace: %s", err) - return false - } - return true + return nil } -func (v *S3Volume) translateError(err error) error { - switch err := err.(type) { - case *s3.Error: - if (err.StatusCode == http.StatusNotFound && err.Code == "NoSuchKey") || - strings.Contains(err.Error(), "Not Found") { - return os.ErrNotExist - } - // Other 404 errors like NoSuchVersion and - // NoSuchBucket are different problems which should - // get called out downstream, so we don't convert them - // to os.ErrNotExist. - } - return err +// DeviceID returns a globally unique ID for the storage bucket. +func (v *s3Volume) DeviceID() string { + return "s3://" + v.Endpoint + "/" + v.Bucket } -// EmptyTrash looks for trashed blocks that exceeded TrashLifetime +// EmptyTrash looks for trashed blocks that exceeded BlobTrashLifetime // and deletes them from the volume. -func (v *S3Volume) EmptyTrash() { +func (v *s3Volume) EmptyTrash() { var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64 // Define "ready to delete" as "...when EmptyTrash started". startT := time.Now() - emptyOneKey := func(trash *s3.Key) { - loc := trash.Key[6:] - if !v.isKeepBlock(loc) { + emptyOneKey := func(trash *s3.Object) { + key := strings.TrimPrefix(*trash.Key, "trash/") + loc, isblk := v.isKeepBlock(key) + if !isblk { return } - atomic.AddInt64(&bytesInTrash, trash.Size) + 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) - return - } - recent, err := v.bucket.Head("recent/"+loc, nil) + trashT := *trash.LastModified + recent, err := v.head("recent/" + key) if err != nil && os.IsNotExist(v.translateError(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) + v.logger.Warnf("EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", *trash.Key, "recent/"+key, err) + err = v.BlockUntrash(loc) if err != nil { - log.Printf("error: %s: EmptyTrash: Untrash(%q): %s", v, loc, err) + v.logger.WithError(err).Errorf("EmptyTrash: Untrash(%q) failed", loc) } return } else if err != nil { - log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err) - 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) + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", "recent/"+key) return } - if trashT.Sub(recentT) < theConfig.BlobSignatureTTL.Duration() { - if age := startT.Sub(recentT); age >= theConfig.BlobSignatureTTL.Duration()-time.Duration(v.RaceWindow) { - // recent/loc is too old to protect + if trashT.Sub(*recent.LastModified) < v.cluster.Collections.BlobSigningTTL.Duration() { + if age := startT.Sub(*recent.LastModified); age >= v.cluster.Collections.BlobSigningTTL.Duration()-time.Duration(v.RaceWindow) { + // recent/key is too old to protect // loc from being Trashed again during // the raceWindow that starts if we // delete trash/X now. // - // Note this means (TrashCheckInterval - // < BlobSignatureTTL - raceWindow) is + // Note this means (TrashSweepInterval + // < BlobSigningTTL - raceWindow) is // necessary to avoid starvation. - log.Printf("notice: %s: EmptyTrash: detected old race for %q, calling fixRace + Touch", v, loc) - v.fixRace(loc) - v.Touch(loc) + v.logger.Infof("EmptyTrash: detected old race for %q, calling fixRace + Touch", loc) + v.fixRace(key) + v.BlockTouch(loc) return } - _, err := v.bucket.Head(loc, nil) + _, err := v.head(key) if os.IsNotExist(err) { - log.Printf("notice: %s: EmptyTrash: detected recent race for %q, calling fixRace", v, loc) - v.fixRace(loc) + v.logger.Infof("EmptyTrash: detected recent race for %q, calling fixRace", loc) + v.fixRace(key) return } else if err != nil { - log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err) + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc) return } } - if startT.Sub(trashT) < theConfig.TrashLifetime.Duration() { + if startT.Sub(trashT) < v.cluster.Collections.BlobTrashLifetime.Duration() { return } - err = v.bucket.Del(trash.Key) + err = v.bucket.Del(*trash.Key) if err != nil { - log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, trash.Key, err) + v.logger.WithError(err).Errorf("EmptyTrash: error deleting %q", *trash.Key) return } - atomic.AddInt64(&bytesDeleted, trash.Size) + atomic.AddInt64(&bytesDeleted, *trash.Size) atomic.AddInt64(&blocksDeleted, 1) - _, err = v.bucket.Head(loc, nil) + _, err = v.head(*trash.Key) if err == nil { - log.Printf("warning: %s: EmptyTrash: HEAD %q succeeded immediately after deleting %q", v, loc, loc) + v.logger.Warnf("EmptyTrash: HEAD %q succeeded immediately after deleting %q", loc, loc) return } if !os.IsNotExist(v.translateError(err)) { - log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err) + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", key) return } - err = v.bucket.Del("recent/" + loc) + err = v.bucket.Del("recent/" + key) if err != nil { - log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err) + v.logger.WithError(err).Warnf("EmptyTrash: error deleting %q", "recent/"+key) } } var wg sync.WaitGroup - todo := make(chan *s3.Key, theConfig.EmptyTrashWorkers) - for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ { + todo := make(chan *s3.Object, v.cluster.Collections.BlobDeleteConcurrency) + for i := 0; i < v.cluster.Collections.BlobDeleteConcurrency; i++ { wg.Add(1) go func() { defer wg.Done() @@ -867,10 +332,12 @@ func (v *S3Volume) EmptyTrash() { }() } - trashL := s3Lister{ - Bucket: v.bucket.Bucket, + trashL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, Prefix: "trash/", PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, } for trash := trashL.First(); trash != nil; trash = trashL.Next() { todo <- trash @@ -879,23 +346,195 @@ func (v *S3Volume) EmptyTrash() { wg.Wait() if err := trashL.Error(); err != nil { - log.Printf("error: %s: EmptyTrash: lister: %s", v, err) + v.logger.WithError(err).Error("EmptyTrash: lister failed") + } + v.logger.Infof("EmptyTrash: stats for %v: Deleted %v bytes in %v blocks. Remaining in trash: %v bytes in %v blocks.", v.DeviceID(), bytesDeleted, blocksDeleted, bytesInTrash-bytesDeleted, blocksInTrash-blocksDeleted) +} + +// fixRace(X) is called when "recent/X" exists but "X" doesn't +// exist. If the timestamps on "recent/X" and "trash/X" indicate there +// was a race between Put and Trash, fixRace recovers from the race by +// Untrashing the block. +func (v *s3Volume) fixRace(key string) bool { + trash, err := v.head("trash/" + key) + if err != nil { + if !os.IsNotExist(v.translateError(err)) { + v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "trash/"+key) + } + return false + } + + recent, err := v.head("recent/" + key) + if err != nil { + v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "recent/"+key) + return false + } + + recentTime := *recent.LastModified + trashTime := *trash.LastModified + ageWhenTrashed := trashTime.Sub(recentTime) + if ageWhenTrashed >= v.cluster.Collections.BlobSigningTTL.Duration() { + // No evidence of a race: block hasn't been written + // since it became eligible for Trash. No fix needed. + return false + } + + v.logger.Infof("fixRace: %q: trashed at %s but touched at %s (age when trashed = %s < %s)", key, trashTime, recentTime, ageWhenTrashed, v.cluster.Collections.BlobSigningTTL) + v.logger.Infof("fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+key, key) + err = v.safeCopy(key, "trash/"+key) + if err != nil { + v.logger.WithError(err).Error("fixRace: copy failed") + return false + } + return true +} + +func (v *s3Volume) head(key string) (result *s3.HeadObjectOutput, err error) { + input := &s3.HeadObjectInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), + } + + req := v.bucket.svc.HeadObjectRequest(input) + res, err := req.Send(context.TODO()) + + v.bucket.stats.TickOps("head") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.HeadOps) + v.bucket.stats.TickErr(err) + + if err != nil { + return nil, v.translateError(err) + } + result = res.HeadObjectOutput + return +} + +// BlockRead reads a Keep block that has been stored as a block blob +// in the S3 bucket. +func (v *s3Volume) BlockRead(ctx context.Context, hash string, w io.WriterAt) error { + key := v.key(hash) + err := v.readWorker(ctx, key, w) + if err != nil { + err = v.translateError(err) + if !os.IsNotExist(err) { + return err + } + + _, err = v.head("recent/" + key) + err = v.translateError(err) + if err != nil { + // If we can't read recent/X, there's no point in + // trying fixRace. Give up. + return err + } + if !v.fixRace(key) { + err = os.ErrNotExist + return err + } + + err = v.readWorker(ctx, key, w) + if err != nil { + v.logger.Warnf("reading %s after successful fixRace: %s", hash, err) + err = v.translateError(err) + return err + } + } + return nil +} + +func (v *s3Volume) readWorker(ctx context.Context, key string, dst io.WriterAt) error { + downloader := s3manager.NewDownloaderWithClient(v.bucket.svc, func(u *s3manager.Downloader) { + u.PartSize = s3downloaderPartSize + u.Concurrency = s3downloaderReadConcurrency + }) + count, err := downloader.DownloadWithContext(ctx, dst, &s3.GetObjectInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), + }) + v.bucket.stats.TickOps("get") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.GetOps) + v.bucket.stats.TickErr(err) + v.bucket.stats.TickInBytes(uint64(count)) + return v.translateError(err) +} + +func (v *s3Volume) writeObject(ctx context.Context, key string, r io.Reader) error { + if r == nil { + // r == nil leads to a memory violation in func readFillBuf in + // aws-sdk-go-v2@v0.23.0/service/s3/s3manager/upload.go + r = bytes.NewReader(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) + + uploadInput := s3manager.UploadInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), + Body: r, + } + + if loc, ok := v.isKeepBlock(key); ok { + var contentMD5 string + md5, err := hex.DecodeString(loc) + if err != nil { + return v.translateError(err) + } + contentMD5 = base64.StdEncoding.EncodeToString(md5) + uploadInput.ContentMD5 = &contentMD5 + } + + // Experimentation indicated that using concurrency 5 yields the best + // throughput, better than higher concurrency (10 or 13) by ~5%. + // Defining u.BufferProvider = s3manager.NewBufferedReadSeekerWriteToPool(64 * 1024 * 1024) + // is detrimental to throughput (minus ~15%). + uploader := s3manager.NewUploaderWithClient(v.bucket.svc, func(u *s3manager.Uploader) { + u.PartSize = s3uploaderPartSize + u.Concurrency = s3uploaderWriteConcurrency + }) + + // Unlike the goamz S3 driver, we don't need to precompute ContentSHA256: + // the aws-sdk-go v2 SDK uses a ReadSeeker to avoid having to copy the + // block, so there is no extra memory use to be concerned about. See + // makeSha256Reader in aws/signer/v4/v4.go. In fact, we explicitly disable + // calculating the Sha-256 because we don't need it; we already use md5sum + // hashes that match the name of the block. + _, err := uploader.UploadWithContext(ctx, &uploadInput, s3manager.WithUploaderRequestOptions(func(r *aws.Request) { + r.HTTPRequest.Header.Set("X-Amz-Content-Sha256", "UNSIGNED-PAYLOAD") + })) + + v.bucket.stats.TickOps("put") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.PutOps) + v.bucket.stats.TickErr(err) + + return v.translateError(err) +} + +// Put writes a block. +func (v *s3Volume) BlockWrite(ctx context.Context, hash string, data []byte) error { + // Do not use putWithPipe here; we want to pass an io.ReadSeeker to the S3 + // sdk to avoid memory allocation there. See #17339 for more information. + rdr := bytes.NewReader(data) + r := newCountingReaderAtSeeker(rdr, v.bucket.stats.TickOutBytes) + key := v.key(hash) + err := v.writeObject(ctx, key, r) + if err != nil { + return err + } + return v.writeObject(ctx, "recent/"+key, nil) } -type s3Lister struct { - Bucket *s3.Bucket - Prefix string - PageSize int - nextMarker string - buf []s3.Key - err error +type s3awsLister struct { + Logger logrus.FieldLogger + Bucket *s3Bucket + Prefix string + PageSize int + Stats *s3awsbucketStats + ContinuationToken string + buf []s3.Object + err error } // First fetches the first page and returns the first item. It returns // nil if the response is the empty set or an error occurs. -func (lister *s3Lister) First() *s3.Key { +func (lister *s3awsLister) First() *s3.Object { lister.getPage() return lister.pop() } @@ -903,39 +542,65 @@ func (lister *s3Lister) First() *s3.Key { // Next returns the next item, fetching the next page if necessary. It // returns nil if the last available item has already been fetched, or // an error occurs. -func (lister *s3Lister) Next() *s3.Key { - if len(lister.buf) == 0 && lister.nextMarker != "" { +func (lister *s3awsLister) Next() *s3.Object { + if len(lister.buf) == 0 && lister.ContinuationToken != "" { lister.getPage() } return lister.pop() } // Return the most recent error encountered by First or Next. -func (lister *s3Lister) Error() error { +func (lister *s3awsLister) Error() error { return lister.err } -func (lister *s3Lister) getPage() { - resp, err := lister.Bucket.List(lister.Prefix, "", lister.nextMarker, lister.PageSize) - lister.nextMarker = "" +func (lister *s3awsLister) getPage() { + lister.Stats.TickOps("list") + lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps) + + var input *s3.ListObjectsV2Input + if lister.ContinuationToken == "" { + input = &s3.ListObjectsV2Input{ + Bucket: aws.String(lister.Bucket.bucket), + MaxKeys: aws.Int64(int64(lister.PageSize)), + Prefix: aws.String(lister.Prefix), + } + } else { + input = &s3.ListObjectsV2Input{ + Bucket: aws.String(lister.Bucket.bucket), + MaxKeys: aws.Int64(int64(lister.PageSize)), + Prefix: aws.String(lister.Prefix), + ContinuationToken: &lister.ContinuationToken, + } + } + + req := lister.Bucket.svc.ListObjectsV2Request(input) + resp, err := req.Send(context.Background()) if err != nil { - lister.err = err + if aerr, ok := err.(awserr.Error); ok { + lister.err = aerr + } else { + lister.err = err + } return } - if resp.IsTruncated { - lister.nextMarker = resp.NextMarker + + if *resp.IsTruncated { + lister.ContinuationToken = *resp.NextContinuationToken + } else { + lister.ContinuationToken = "" } - lister.buf = make([]s3.Key, 0, len(resp.Contents)) + lister.buf = make([]s3.Object, 0, len(resp.Contents)) for _, key := range resp.Contents { - if !strings.HasPrefix(key.Key, lister.Prefix) { - log.Printf("warning: s3Lister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, key.Key) + if !strings.HasPrefix(*key.Key, lister.Prefix) { + lister.Logger.Warnf("s3awsLister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, *key.Key) continue } lister.buf = append(lister.buf, key) } } -func (lister *s3Lister) pop() (k *s3.Key) { +func (lister *s3awsLister) pop() (k *s3.Object) { if len(lister.buf) > 0 { k = &lister.buf[0] lister.buf = lister.buf[1:] @@ -943,52 +608,201 @@ func (lister *s3Lister) pop() (k *s3.Key) { return } -// s3bucket wraps s3.bucket and counts I/O and API usage stats. -type s3bucket struct { - *s3.Bucket - stats s3bucketStats +// Index writes a complete list of locators with the given prefix +// for which Get() can retrieve data. +func (v *s3Volume) Index(ctx context.Context, prefix string, writer io.Writer) error { + prefix = v.key(prefix) + // Use a merge sort to find matching sets of X and recent/X. + dataL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, + Prefix: prefix, + PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, + } + recentL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, + Prefix: "recent/" + prefix, + PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, + } + for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() { + if ctx.Err() != nil { + return ctx.Err() + } + if *data.Key >= "g" { + // Conveniently, "recent/*" and "trash/*" are + // lexically greater than all hex-encoded data + // hashes, so stopping here avoids iterating + // over all of them needlessly with dataL. + break + } + loc, isblk := v.isKeepBlock(*data.Key) + if !isblk { + continue + } + + // stamp is the list entry we should use to report the + // last-modified time for this data block: it will be + // the recent/X entry if one exists, otherwise the + // entry for the data block itself. + stamp := data + + // Advance to the corresponding recent/X marker, if any + for recent != nil && recentL.Error() == nil { + if cmp := strings.Compare((*recent.Key)[7:], *data.Key); cmp < 0 { + recent = recentL.Next() + continue + } else if cmp == 0 { + stamp = recent + recent = recentL.Next() + break + } else { + // recent/X marker is missing: we'll + // use the timestamp on the data + // object. + break + } + } + if err := recentL.Error(); err != nil { + return err + } + // We truncate sub-second precision here. Otherwise + // timestamps will never match the RFC1123-formatted + // Last-Modified values parsed by Mtime(). + fmt.Fprintf(writer, "%s+%d %d\n", loc, *data.Size, stamp.LastModified.Unix()*1000000000) + } + return dataL.Error() } -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 +// Mtime returns the stored timestamp for the given locator. +func (v *s3Volume) Mtime(loc string) (time.Time, error) { + key := v.key(loc) + _, err := v.head(key) + if err != nil { + return s3AWSZeroTime, v.translateError(err) + } + resp, err := v.head("recent/" + key) + err = v.translateError(err) + if os.IsNotExist(err) { + // The data object X exists, but recent/X is missing. + err = v.writeObject(context.Background(), "recent/"+key, nil) + if err != nil { + v.logger.WithError(err).Errorf("error creating %q", "recent/"+key) + return s3AWSZeroTime, v.translateError(err) + } + v.logger.Infof("Mtime: created %q to migrate existing block to new storage scheme", "recent/"+key) + resp, err = v.head("recent/" + key) + if err != nil { + v.logger.WithError(err).Errorf("HEAD failed after creating %q", "recent/"+key) + return s3AWSZeroTime, v.translateError(err) + } + } else if err != nil { + // HEAD recent/X failed for some other reason. + return s3AWSZeroTime, err + } + return *resp.LastModified, 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) +// InternalStats returns bucket I/O and API call counters. +func (v *s3Volume) InternalStats() interface{} { + return &v.bucket.stats +} + +// BlockTouch sets the timestamp for the given locator to the current time. +func (v *s3Volume) BlockTouch(hash string) error { + key := v.key(hash) + _, err := v.head(key) + err = v.translateError(err) + if os.IsNotExist(err) && v.fixRace(key) { + // The data object got trashed in a race, but fixRace + // rescued it. + } else if err != nil { + return err } - 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 + err = v.writeObject(context.Background(), "recent/"+key, nil) + return v.translateError(err) +} + +// checkRaceWindow returns a non-nil error if trash/key is, or might +// be, in the race window (i.e., it's not safe to trash key). +func (v *s3Volume) checkRaceWindow(key string) error { + resp, err := v.head("trash/" + key) + err = v.translateError(err) + if os.IsNotExist(err) { + // OK, trash/X doesn't exist so we're not in the race + // window + return nil + } else if err != nil { + // Error looking up trash/X. We don't know whether + // we're in the race window + return err + } + t := resp.LastModified + safeWindow := t.Add(v.cluster.Collections.BlobTrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow))) + if safeWindow <= 0 { + // We can't count on "touch trash/X" to prolong + // trash/X's lifetime. The new timestamp might not + // become visible until now+raceWindow, and EmptyTrash + // is allowed to delete trash/X before then. + return fmt.Errorf("%s: same block is already in trash, and safe window ended %s ago", key, -safeWindow) + } + // trash/X exists, but it won't be eligible for deletion until + // after now+raceWindow, so it's safe to overwrite it. + return nil } -func (b *s3bucket) Del(path string) error { - err := b.Bucket.Del(path) +func (b *s3Bucket) Del(path string) error { + input := &s3.DeleteObjectInput{ + Bucket: aws.String(b.bucket), + Key: aws.String(path), + } + req := b.svc.DeleteObjectRequest(input) + _, err := req.Send(context.Background()) + b.stats.TickOps("delete") b.stats.Tick(&b.stats.Ops, &b.stats.DelOps) b.stats.TickErr(err) return err } -type s3bucketStats struct { +// Trash a Keep block. +func (v *s3Volume) BlockTrash(loc string) error { + if t, err := v.Mtime(loc); err != nil { + return err + } else if time.Since(t) < v.cluster.Collections.BlobSigningTTL.Duration() { + return nil + } + key := v.key(loc) + if v.cluster.Collections.BlobTrashLifetime == 0 { + if !v.UnsafeDelete { + return errS3TrashDisabled + } + return v.translateError(v.bucket.Del(key)) + } + err := v.checkRaceWindow(key) + if err != nil { + return err + } + err = v.safeCopy("trash/"+key, key) + if err != nil { + return err + } + return v.translateError(v.bucket.Del(key)) +} + +// BlockUntrash moves block from trash back into store +func (v *s3Volume) BlockUntrash(hash string) error { + key := v.key(hash) + err := v.safeCopy(key, "trash/"+key) + if err != nil { + return err + } + err = v.writeObject(context.Background(), "recent/"+key, nil) + return v.translateError(err) +} + +type s3awsbucketStats struct { statsTicker Ops uint64 GetOps uint64 @@ -998,13 +812,18 @@ type s3bucketStats struct { ListOps uint64 } -func (s *s3bucketStats) TickErr(err error) { +func (s *s3awsbucketStats) 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) + if aerr, ok := err.(awserr.Error); ok { + if reqErr, ok := err.(awserr.RequestFailure); ok { + // A service error occurred + errType = errType + fmt.Sprintf(" %d %s", reqErr.StatusCode(), aerr.Code()) + } else { + errType = errType + fmt.Sprintf(" 000 %s", aerr.Code()) + } } s.statsTicker.TickErr(err, errType) }