Add unit tests

This commit is contained in:
Mayuresh Gaitonde
2020-12-17 17:25:53 -08:00
parent 3702339f44
commit 6be4d69d02
705 changed files with 120529 additions and 150051 deletions

View File

@@ -25,44 +25,39 @@ package backoff
import (
"time"
grpcbackoff "google.golang.org/grpc/backoff"
"google.golang.org/grpc/internal/grpcrand"
)
// Strategy defines the methodology for backing off after a grpc connection
// failure.
//
type Strategy interface {
// Backoff returns the amount of time to wait before the next retry given
// the number of consecutive failures.
Backoff(retries int) time.Duration
}
const (
// baseDelay is the amount of time to wait before retrying after the first
// failure.
baseDelay = 1.0 * time.Second
// factor is applied to the backoff after each retry.
factor = 1.6
// jitter provides a range to randomize backoff delays.
jitter = 0.2
)
// DefaultExponential is an exponential backoff implementation using the
// default values for all the configurable knobs defined in
// https://github.com/grpc/grpc/blob/master/doc/connection-backoff.md.
var DefaultExponential = Exponential{Config: grpcbackoff.DefaultConfig}
// Exponential implements exponential backoff algorithm as defined in
// https://github.com/grpc/grpc/blob/master/doc/connection-backoff.md.
type Exponential struct {
// MaxDelay is the upper bound of backoff delay.
MaxDelay time.Duration
// Config contains all options to configure the backoff algorithm.
Config grpcbackoff.Config
}
// Backoff returns the amount of time to wait before the next retry given the
// number of retries.
func (bc Exponential) Backoff(retries int) time.Duration {
if retries == 0 {
return baseDelay
return bc.Config.BaseDelay
}
backoff, max := float64(baseDelay), float64(bc.MaxDelay)
backoff, max := float64(bc.Config.BaseDelay), float64(bc.Config.MaxDelay)
for backoff < max && retries > 0 {
backoff *= factor
backoff *= bc.Config.Multiplier
retries--
}
if backoff > max {
@@ -70,7 +65,7 @@ func (bc Exponential) Backoff(retries int) time.Duration {
}
// Randomize backoff delays so that if a cluster of requests start at
// the same time, they won't operate in lockstep.
backoff *= 1 + jitter*(grpcrand.Float64()*2-1)
backoff *= 1 + bc.Config.Jitter*(grpcrand.Float64()*2-1)
if backoff < 0 {
return 0
}

View File

@@ -0,0 +1,46 @@
/*
* Copyright 2019 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
// Package balancerload defines APIs to parse server loads in trailers. The
// parsed loads are sent to balancers in DoneInfo.
package balancerload
import (
"google.golang.org/grpc/metadata"
)
// Parser converts loads from metadata into a concrete type.
type Parser interface {
// Parse parses loads from metadata.
Parse(md metadata.MD) interface{}
}
var parser Parser
// SetParser sets the load parser.
//
// Not mutex-protected, should be called before any gRPC functions.
func SetParser(lr Parser) {
parser = lr
}
// Parse calls parser.Read().
func Parse(md metadata.MD) interface{} {
if parser == nil {
return nil
}
return parser.Parse(md)
}

View File

@@ -0,0 +1,167 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package binarylog implementation binary logging as defined in
// https://github.com/grpc/proposal/blob/master/A16-binary-logging.md.
package binarylog
import (
"fmt"
"os"
"google.golang.org/grpc/grpclog"
)
// Logger is the global binary logger. It can be used to get binary logger for
// each method.
type Logger interface {
getMethodLogger(methodName string) *MethodLogger
}
// binLogger is the global binary logger for the binary. One of this should be
// built at init time from the configuration (environment variable or flags).
//
// It is used to get a methodLogger for each individual method.
var binLogger Logger
// SetLogger sets the binarg logger.
//
// Only call this at init time.
func SetLogger(l Logger) {
binLogger = l
}
// GetMethodLogger returns the methodLogger for the given methodName.
//
// methodName should be in the format of "/service/method".
//
// Each methodLogger returned by this method is a new instance. This is to
// generate sequence id within the call.
func GetMethodLogger(methodName string) *MethodLogger {
if binLogger == nil {
return nil
}
return binLogger.getMethodLogger(methodName)
}
func init() {
const envStr = "GRPC_BINARY_LOG_FILTER"
configStr := os.Getenv(envStr)
binLogger = NewLoggerFromConfigString(configStr)
}
type methodLoggerConfig struct {
// Max length of header and message.
hdr, msg uint64
}
type logger struct {
all *methodLoggerConfig
services map[string]*methodLoggerConfig
methods map[string]*methodLoggerConfig
blacklist map[string]struct{}
}
// newEmptyLogger creates an empty logger. The map fields need to be filled in
// using the set* functions.
func newEmptyLogger() *logger {
return &logger{}
}
// Set method logger for "*".
func (l *logger) setDefaultMethodLogger(ml *methodLoggerConfig) error {
if l.all != nil {
return fmt.Errorf("conflicting global rules found")
}
l.all = ml
return nil
}
// Set method logger for "service/*".
//
// New methodLogger with same service overrides the old one.
func (l *logger) setServiceMethodLogger(service string, ml *methodLoggerConfig) error {
if _, ok := l.services[service]; ok {
return fmt.Errorf("conflicting service rules for service %v found", service)
}
if l.services == nil {
l.services = make(map[string]*methodLoggerConfig)
}
l.services[service] = ml
return nil
}
// Set method logger for "service/method".
//
// New methodLogger with same method overrides the old one.
func (l *logger) setMethodMethodLogger(method string, ml *methodLoggerConfig) error {
if _, ok := l.blacklist[method]; ok {
return fmt.Errorf("conflicting blacklist rules for method %v found", method)
}
if _, ok := l.methods[method]; ok {
return fmt.Errorf("conflicting method rules for method %v found", method)
}
if l.methods == nil {
l.methods = make(map[string]*methodLoggerConfig)
}
l.methods[method] = ml
return nil
}
// Set blacklist method for "-service/method".
func (l *logger) setBlacklist(method string) error {
if _, ok := l.blacklist[method]; ok {
return fmt.Errorf("conflicting blacklist rules for method %v found", method)
}
if _, ok := l.methods[method]; ok {
return fmt.Errorf("conflicting method rules for method %v found", method)
}
if l.blacklist == nil {
l.blacklist = make(map[string]struct{})
}
l.blacklist[method] = struct{}{}
return nil
}
// getMethodLogger returns the methodLogger for the given methodName.
//
// methodName should be in the format of "/service/method".
//
// Each methodLogger returned by this method is a new instance. This is to
// generate sequence id within the call.
func (l *logger) getMethodLogger(methodName string) *MethodLogger {
s, m, err := parseMethodName(methodName)
if err != nil {
grpclog.Infof("binarylogging: failed to parse %q: %v", methodName, err)
return nil
}
if ml, ok := l.methods[s+"/"+m]; ok {
return newMethodLogger(ml.hdr, ml.msg)
}
if _, ok := l.blacklist[s+"/"+m]; ok {
return nil
}
if ml, ok := l.services[s]; ok {
return newMethodLogger(ml.hdr, ml.msg)
}
if l.all == nil {
return nil
}
return newMethodLogger(l.all.hdr, l.all.msg)
}

View File

@@ -0,0 +1,42 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// This file contains exported variables/functions that are exported for testing
// only.
//
// An ideal way for this would be to put those in a *_test.go but in binarylog
// package. But this doesn't work with staticcheck with go module. Error was:
// "MdToMetadataProto not declared by package binarylog". This could be caused
// by the way staticcheck looks for files for a certain package, which doesn't
// support *_test.go files.
//
// Move those to binary_test.go when staticcheck is fixed.
package binarylog
var (
// AllLogger is a logger that logs all headers/messages for all RPCs. It's
// for testing only.
AllLogger = NewLoggerFromConfigString("*")
// MdToMetadataProto converts metadata to a binary logging proto message.
// It's for testing only.
MdToMetadataProto = mdToMetadataProto
// AddrToProto converts an address to a binary logging proto message. It's
// for testing only.
AddrToProto = addrToProto
)

View File

@@ -0,0 +1,210 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package binarylog
import (
"errors"
"fmt"
"regexp"
"strconv"
"strings"
"google.golang.org/grpc/grpclog"
)
// NewLoggerFromConfigString reads the string and build a logger. It can be used
// to build a new logger and assign it to binarylog.Logger.
//
// Example filter config strings:
// - "" Nothing will be logged
// - "*" All headers and messages will be fully logged.
// - "*{h}" Only headers will be logged.
// - "*{m:256}" Only the first 256 bytes of each message will be logged.
// - "Foo/*" Logs every method in service Foo
// - "Foo/*,-Foo/Bar" Logs every method in service Foo except method /Foo/Bar
// - "Foo/*,Foo/Bar{m:256}" Logs the first 256 bytes of each message in method
// /Foo/Bar, logs all headers and messages in every other method in service
// Foo.
//
// If two configs exist for one certain method or service, the one specified
// later overrides the previous config.
func NewLoggerFromConfigString(s string) Logger {
if s == "" {
return nil
}
l := newEmptyLogger()
methods := strings.Split(s, ",")
for _, method := range methods {
if err := l.fillMethodLoggerWithConfigString(method); err != nil {
grpclog.Warningf("failed to parse binary log config: %v", err)
return nil
}
}
return l
}
// fillMethodLoggerWithConfigString parses config, creates methodLogger and adds
// it to the right map in the logger.
func (l *logger) fillMethodLoggerWithConfigString(config string) error {
// "" is invalid.
if config == "" {
return errors.New("empty string is not a valid method binary logging config")
}
// "-service/method", blacklist, no * or {} allowed.
if config[0] == '-' {
s, m, suffix, err := parseMethodConfigAndSuffix(config[1:])
if err != nil {
return fmt.Errorf("invalid config: %q, %v", config, err)
}
if m == "*" {
return fmt.Errorf("invalid config: %q, %v", config, "* not allowed in blacklist config")
}
if suffix != "" {
return fmt.Errorf("invalid config: %q, %v", config, "header/message limit not allowed in blacklist config")
}
if err := l.setBlacklist(s + "/" + m); err != nil {
return fmt.Errorf("invalid config: %v", err)
}
return nil
}
// "*{h:256;m:256}"
if config[0] == '*' {
hdr, msg, err := parseHeaderMessageLengthConfig(config[1:])
if err != nil {
return fmt.Errorf("invalid config: %q, %v", config, err)
}
if err := l.setDefaultMethodLogger(&methodLoggerConfig{hdr: hdr, msg: msg}); err != nil {
return fmt.Errorf("invalid config: %v", err)
}
return nil
}
s, m, suffix, err := parseMethodConfigAndSuffix(config)
if err != nil {
return fmt.Errorf("invalid config: %q, %v", config, err)
}
hdr, msg, err := parseHeaderMessageLengthConfig(suffix)
if err != nil {
return fmt.Errorf("invalid header/message length config: %q, %v", suffix, err)
}
if m == "*" {
if err := l.setServiceMethodLogger(s, &methodLoggerConfig{hdr: hdr, msg: msg}); err != nil {
return fmt.Errorf("invalid config: %v", err)
}
} else {
if err := l.setMethodMethodLogger(s+"/"+m, &methodLoggerConfig{hdr: hdr, msg: msg}); err != nil {
return fmt.Errorf("invalid config: %v", err)
}
}
return nil
}
const (
// TODO: this const is only used by env_config now. But could be useful for
// other config. Move to binarylog.go if necessary.
maxUInt = ^uint64(0)
// For "p.s/m" plus any suffix. Suffix will be parsed again. See test for
// expected output.
longMethodConfigRegexpStr = `^([\w./]+)/((?:\w+)|[*])(.+)?$`
// For suffix from above, "{h:123,m:123}". See test for expected output.
optionalLengthRegexpStr = `(?::(\d+))?` // Optional ":123".
headerConfigRegexpStr = `^{h` + optionalLengthRegexpStr + `}$`
messageConfigRegexpStr = `^{m` + optionalLengthRegexpStr + `}$`
headerMessageConfigRegexpStr = `^{h` + optionalLengthRegexpStr + `;m` + optionalLengthRegexpStr + `}$`
)
var (
longMethodConfigRegexp = regexp.MustCompile(longMethodConfigRegexpStr)
headerConfigRegexp = regexp.MustCompile(headerConfigRegexpStr)
messageConfigRegexp = regexp.MustCompile(messageConfigRegexpStr)
headerMessageConfigRegexp = regexp.MustCompile(headerMessageConfigRegexpStr)
)
// Turn "service/method{h;m}" into "service", "method", "{h;m}".
func parseMethodConfigAndSuffix(c string) (service, method, suffix string, _ error) {
// Regexp result:
//
// in: "p.s/m{h:123,m:123}",
// out: []string{"p.s/m{h:123,m:123}", "p.s", "m", "{h:123,m:123}"},
match := longMethodConfigRegexp.FindStringSubmatch(c)
if match == nil {
return "", "", "", fmt.Errorf("%q contains invalid substring", c)
}
service = match[1]
method = match[2]
suffix = match[3]
return
}
// Turn "{h:123;m:345}" into 123, 345.
//
// Return maxUInt if length is unspecified.
func parseHeaderMessageLengthConfig(c string) (hdrLenStr, msgLenStr uint64, err error) {
if c == "" {
return maxUInt, maxUInt, nil
}
// Header config only.
if match := headerConfigRegexp.FindStringSubmatch(c); match != nil {
if s := match[1]; s != "" {
hdrLenStr, err = strconv.ParseUint(s, 10, 64)
if err != nil {
return 0, 0, fmt.Errorf("failed to convert %q to uint", s)
}
return hdrLenStr, 0, nil
}
return maxUInt, 0, nil
}
// Message config only.
if match := messageConfigRegexp.FindStringSubmatch(c); match != nil {
if s := match[1]; s != "" {
msgLenStr, err = strconv.ParseUint(s, 10, 64)
if err != nil {
return 0, 0, fmt.Errorf("failed to convert %q to uint", s)
}
return 0, msgLenStr, nil
}
return 0, maxUInt, nil
}
// Header and message config both.
if match := headerMessageConfigRegexp.FindStringSubmatch(c); match != nil {
// Both hdr and msg are specified, but one or two of them might be empty.
hdrLenStr = maxUInt
msgLenStr = maxUInt
if s := match[1]; s != "" {
hdrLenStr, err = strconv.ParseUint(s, 10, 64)
if err != nil {
return 0, 0, fmt.Errorf("failed to convert %q to uint", s)
}
}
if s := match[2]; s != "" {
msgLenStr, err = strconv.ParseUint(s, 10, 64)
if err != nil {
return 0, 0, fmt.Errorf("failed to convert %q to uint", s)
}
}
return hdrLenStr, msgLenStr, nil
}
return 0, 0, fmt.Errorf("%q contains invalid substring", c)
}

View File

@@ -0,0 +1,423 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package binarylog
import (
"net"
"strings"
"sync/atomic"
"time"
"github.com/golang/protobuf/proto"
"github.com/golang/protobuf/ptypes"
pb "google.golang.org/grpc/binarylog/grpc_binarylog_v1"
"google.golang.org/grpc/grpclog"
"google.golang.org/grpc/metadata"
"google.golang.org/grpc/status"
)
type callIDGenerator struct {
id uint64
}
func (g *callIDGenerator) next() uint64 {
id := atomic.AddUint64(&g.id, 1)
return id
}
// reset is for testing only, and doesn't need to be thread safe.
func (g *callIDGenerator) reset() {
g.id = 0
}
var idGen callIDGenerator
// MethodLogger is the sub-logger for each method.
type MethodLogger struct {
headerMaxLen, messageMaxLen uint64
callID uint64
idWithinCallGen *callIDGenerator
sink Sink // TODO(blog): make this plugable.
}
func newMethodLogger(h, m uint64) *MethodLogger {
return &MethodLogger{
headerMaxLen: h,
messageMaxLen: m,
callID: idGen.next(),
idWithinCallGen: &callIDGenerator{},
sink: defaultSink, // TODO(blog): make it plugable.
}
}
// Log creates a proto binary log entry, and logs it to the sink.
func (ml *MethodLogger) Log(c LogEntryConfig) {
m := c.toProto()
timestamp, _ := ptypes.TimestampProto(time.Now())
m.Timestamp = timestamp
m.CallId = ml.callID
m.SequenceIdWithinCall = ml.idWithinCallGen.next()
switch pay := m.Payload.(type) {
case *pb.GrpcLogEntry_ClientHeader:
m.PayloadTruncated = ml.truncateMetadata(pay.ClientHeader.GetMetadata())
case *pb.GrpcLogEntry_ServerHeader:
m.PayloadTruncated = ml.truncateMetadata(pay.ServerHeader.GetMetadata())
case *pb.GrpcLogEntry_Message:
m.PayloadTruncated = ml.truncateMessage(pay.Message)
}
ml.sink.Write(m)
}
func (ml *MethodLogger) truncateMetadata(mdPb *pb.Metadata) (truncated bool) {
if ml.headerMaxLen == maxUInt {
return false
}
var (
bytesLimit = ml.headerMaxLen
index int
)
// At the end of the loop, index will be the first entry where the total
// size is greater than the limit:
//
// len(entry[:index]) <= ml.hdr && len(entry[:index+1]) > ml.hdr.
for ; index < len(mdPb.Entry); index++ {
entry := mdPb.Entry[index]
if entry.Key == "grpc-trace-bin" {
// "grpc-trace-bin" is a special key. It's kept in the log entry,
// but not counted towards the size limit.
continue
}
currentEntryLen := uint64(len(entry.Value))
if currentEntryLen > bytesLimit {
break
}
bytesLimit -= currentEntryLen
}
truncated = index < len(mdPb.Entry)
mdPb.Entry = mdPb.Entry[:index]
return truncated
}
func (ml *MethodLogger) truncateMessage(msgPb *pb.Message) (truncated bool) {
if ml.messageMaxLen == maxUInt {
return false
}
if ml.messageMaxLen >= uint64(len(msgPb.Data)) {
return false
}
msgPb.Data = msgPb.Data[:ml.messageMaxLen]
return true
}
// LogEntryConfig represents the configuration for binary log entry.
type LogEntryConfig interface {
toProto() *pb.GrpcLogEntry
}
// ClientHeader configs the binary log entry to be a ClientHeader entry.
type ClientHeader struct {
OnClientSide bool
Header metadata.MD
MethodName string
Authority string
Timeout time.Duration
// PeerAddr is required only when it's on server side.
PeerAddr net.Addr
}
func (c *ClientHeader) toProto() *pb.GrpcLogEntry {
// This function doesn't need to set all the fields (e.g. seq ID). The Log
// function will set the fields when necessary.
clientHeader := &pb.ClientHeader{
Metadata: mdToMetadataProto(c.Header),
MethodName: c.MethodName,
Authority: c.Authority,
}
if c.Timeout > 0 {
clientHeader.Timeout = ptypes.DurationProto(c.Timeout)
}
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_CLIENT_HEADER,
Payload: &pb.GrpcLogEntry_ClientHeader{
ClientHeader: clientHeader,
},
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
if c.PeerAddr != nil {
ret.Peer = addrToProto(c.PeerAddr)
}
return ret
}
// ServerHeader configs the binary log entry to be a ServerHeader entry.
type ServerHeader struct {
OnClientSide bool
Header metadata.MD
// PeerAddr is required only when it's on client side.
PeerAddr net.Addr
}
func (c *ServerHeader) toProto() *pb.GrpcLogEntry {
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_SERVER_HEADER,
Payload: &pb.GrpcLogEntry_ServerHeader{
ServerHeader: &pb.ServerHeader{
Metadata: mdToMetadataProto(c.Header),
},
},
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
if c.PeerAddr != nil {
ret.Peer = addrToProto(c.PeerAddr)
}
return ret
}
// ClientMessage configs the binary log entry to be a ClientMessage entry.
type ClientMessage struct {
OnClientSide bool
// Message can be a proto.Message or []byte. Other messages formats are not
// supported.
Message interface{}
}
func (c *ClientMessage) toProto() *pb.GrpcLogEntry {
var (
data []byte
err error
)
if m, ok := c.Message.(proto.Message); ok {
data, err = proto.Marshal(m)
if err != nil {
grpclog.Infof("binarylogging: failed to marshal proto message: %v", err)
}
} else if b, ok := c.Message.([]byte); ok {
data = b
} else {
grpclog.Infof("binarylogging: message to log is neither proto.message nor []byte")
}
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_CLIENT_MESSAGE,
Payload: &pb.GrpcLogEntry_Message{
Message: &pb.Message{
Length: uint32(len(data)),
Data: data,
},
},
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
return ret
}
// ServerMessage configs the binary log entry to be a ServerMessage entry.
type ServerMessage struct {
OnClientSide bool
// Message can be a proto.Message or []byte. Other messages formats are not
// supported.
Message interface{}
}
func (c *ServerMessage) toProto() *pb.GrpcLogEntry {
var (
data []byte
err error
)
if m, ok := c.Message.(proto.Message); ok {
data, err = proto.Marshal(m)
if err != nil {
grpclog.Infof("binarylogging: failed to marshal proto message: %v", err)
}
} else if b, ok := c.Message.([]byte); ok {
data = b
} else {
grpclog.Infof("binarylogging: message to log is neither proto.message nor []byte")
}
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_SERVER_MESSAGE,
Payload: &pb.GrpcLogEntry_Message{
Message: &pb.Message{
Length: uint32(len(data)),
Data: data,
},
},
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
return ret
}
// ClientHalfClose configs the binary log entry to be a ClientHalfClose entry.
type ClientHalfClose struct {
OnClientSide bool
}
func (c *ClientHalfClose) toProto() *pb.GrpcLogEntry {
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_CLIENT_HALF_CLOSE,
Payload: nil, // No payload here.
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
return ret
}
// ServerTrailer configs the binary log entry to be a ServerTrailer entry.
type ServerTrailer struct {
OnClientSide bool
Trailer metadata.MD
// Err is the status error.
Err error
// PeerAddr is required only when it's on client side and the RPC is trailer
// only.
PeerAddr net.Addr
}
func (c *ServerTrailer) toProto() *pb.GrpcLogEntry {
st, ok := status.FromError(c.Err)
if !ok {
grpclog.Info("binarylogging: error in trailer is not a status error")
}
var (
detailsBytes []byte
err error
)
stProto := st.Proto()
if stProto != nil && len(stProto.Details) != 0 {
detailsBytes, err = proto.Marshal(stProto)
if err != nil {
grpclog.Infof("binarylogging: failed to marshal status proto: %v", err)
}
}
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_SERVER_TRAILER,
Payload: &pb.GrpcLogEntry_Trailer{
Trailer: &pb.Trailer{
Metadata: mdToMetadataProto(c.Trailer),
StatusCode: uint32(st.Code()),
StatusMessage: st.Message(),
StatusDetails: detailsBytes,
},
},
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
if c.PeerAddr != nil {
ret.Peer = addrToProto(c.PeerAddr)
}
return ret
}
// Cancel configs the binary log entry to be a Cancel entry.
type Cancel struct {
OnClientSide bool
}
func (c *Cancel) toProto() *pb.GrpcLogEntry {
ret := &pb.GrpcLogEntry{
Type: pb.GrpcLogEntry_EVENT_TYPE_CANCEL,
Payload: nil,
}
if c.OnClientSide {
ret.Logger = pb.GrpcLogEntry_LOGGER_CLIENT
} else {
ret.Logger = pb.GrpcLogEntry_LOGGER_SERVER
}
return ret
}
// metadataKeyOmit returns whether the metadata entry with this key should be
// omitted.
func metadataKeyOmit(key string) bool {
switch key {
case "lb-token", ":path", ":authority", "content-encoding", "content-type", "user-agent", "te":
return true
case "grpc-trace-bin": // grpc-trace-bin is special because it's visiable to users.
return false
}
return strings.HasPrefix(key, "grpc-")
}
func mdToMetadataProto(md metadata.MD) *pb.Metadata {
ret := &pb.Metadata{}
for k, vv := range md {
if metadataKeyOmit(k) {
continue
}
for _, v := range vv {
ret.Entry = append(ret.Entry,
&pb.MetadataEntry{
Key: k,
Value: []byte(v),
},
)
}
}
return ret
}
func addrToProto(addr net.Addr) *pb.Address {
ret := &pb.Address{}
switch a := addr.(type) {
case *net.TCPAddr:
if a.IP.To4() != nil {
ret.Type = pb.Address_TYPE_IPV4
} else if a.IP.To16() != nil {
ret.Type = pb.Address_TYPE_IPV6
} else {
ret.Type = pb.Address_TYPE_UNKNOWN
// Do not set address and port fields.
break
}
ret.Address = a.IP.String()
ret.IpPort = uint32(a.Port)
case *net.UnixAddr:
ret.Type = pb.Address_TYPE_UNIX
ret.Address = a.String()
default:
ret.Type = pb.Address_TYPE_UNKNOWN
}
return ret
}

View File

@@ -0,0 +1,33 @@
#!/bin/bash
# Copyright 2018 gRPC authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
set -eux -o pipefail
TMP=$(mktemp -d)
function finish {
rm -rf "$TMP"
}
trap finish EXIT
pushd "$TMP"
mkdir -p grpc/binarylog/grpc_binarylog_v1
curl https://raw.githubusercontent.com/grpc/grpc-proto/master/grpc/binlog/v1/binarylog.proto > grpc/binarylog/grpc_binarylog_v1/binarylog.proto
protoc --go_out=plugins=grpc,paths=source_relative:. -I. grpc/binarylog/grpc_binarylog_v1/*.proto
popd
rm -f ./grpc_binarylog_v1/*.pb.go
cp "$TMP"/grpc/binarylog/grpc_binarylog_v1/*.pb.go ../../binarylog/grpc_binarylog_v1/

View File

@@ -0,0 +1,162 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package binarylog
import (
"bufio"
"encoding/binary"
"fmt"
"io"
"io/ioutil"
"sync"
"time"
"github.com/golang/protobuf/proto"
pb "google.golang.org/grpc/binarylog/grpc_binarylog_v1"
"google.golang.org/grpc/grpclog"
)
var (
defaultSink Sink = &noopSink{} // TODO(blog): change this default (file in /tmp).
)
// SetDefaultSink sets the sink where binary logs will be written to.
//
// Not thread safe. Only set during initialization.
func SetDefaultSink(s Sink) {
if defaultSink != nil {
defaultSink.Close()
}
defaultSink = s
}
// Sink writes log entry into the binary log sink.
type Sink interface {
// Write will be called to write the log entry into the sink.
//
// It should be thread-safe so it can be called in parallel.
Write(*pb.GrpcLogEntry) error
// Close will be called when the Sink is replaced by a new Sink.
Close() error
}
type noopSink struct{}
func (ns *noopSink) Write(*pb.GrpcLogEntry) error { return nil }
func (ns *noopSink) Close() error { return nil }
// newWriterSink creates a binary log sink with the given writer.
//
// Write() marshals the proto message and writes it to the given writer. Each
// message is prefixed with a 4 byte big endian unsigned integer as the length.
//
// No buffer is done, Close() doesn't try to close the writer.
func newWriterSink(w io.Writer) *writerSink {
return &writerSink{out: w}
}
type writerSink struct {
out io.Writer
}
func (ws *writerSink) Write(e *pb.GrpcLogEntry) error {
b, err := proto.Marshal(e)
if err != nil {
grpclog.Infof("binary logging: failed to marshal proto message: %v", err)
}
hdr := make([]byte, 4)
binary.BigEndian.PutUint32(hdr, uint32(len(b)))
if _, err := ws.out.Write(hdr); err != nil {
return err
}
if _, err := ws.out.Write(b); err != nil {
return err
}
return nil
}
func (ws *writerSink) Close() error { return nil }
type bufWriteCloserSink struct {
mu sync.Mutex
closer io.Closer
out *writerSink // out is built on buf.
buf *bufio.Writer // buf is kept for flush.
writeStartOnce sync.Once
writeTicker *time.Ticker
}
func (fs *bufWriteCloserSink) Write(e *pb.GrpcLogEntry) error {
// Start the write loop when Write is called.
fs.writeStartOnce.Do(fs.startFlushGoroutine)
fs.mu.Lock()
if err := fs.out.Write(e); err != nil {
fs.mu.Unlock()
return err
}
fs.mu.Unlock()
return nil
}
const (
bufFlushDuration = 60 * time.Second
)
func (fs *bufWriteCloserSink) startFlushGoroutine() {
fs.writeTicker = time.NewTicker(bufFlushDuration)
go func() {
for range fs.writeTicker.C {
fs.mu.Lock()
fs.buf.Flush()
fs.mu.Unlock()
}
}()
}
func (fs *bufWriteCloserSink) Close() error {
if fs.writeTicker != nil {
fs.writeTicker.Stop()
}
fs.mu.Lock()
fs.buf.Flush()
fs.closer.Close()
fs.out.Close()
fs.mu.Unlock()
return nil
}
func newBufWriteCloserSink(o io.WriteCloser) Sink {
bufW := bufio.NewWriter(o)
return &bufWriteCloserSink{
closer: o,
out: newWriterSink(bufW),
buf: bufW,
}
}
// NewTempFileSink creates a temp file and returns a Sink that writes to this
// file.
func NewTempFileSink() (Sink, error) {
tempFile, err := ioutil.TempFile("/tmp", "grpcgo_binarylog_*.txt")
if err != nil {
return nil, fmt.Errorf("failed to create temp file: %v", err)
}
return newBufWriteCloserSink(tempFile), nil
}

View File

@@ -0,0 +1,41 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package binarylog
import (
"errors"
"strings"
)
// parseMethodName splits service and method from the input. It expects format
// "/service/method".
//
// TODO: move to internal/grpcutil.
func parseMethodName(methodName string) (service, method string, _ error) {
if !strings.HasPrefix(methodName, "/") {
return "", "", errors.New("invalid method name: should start with /")
}
methodName = methodName[1:]
pos := strings.LastIndex(methodName, "/")
if pos < 0 {
return "", "", errors.New("invalid method name: suffix /method is missing")
}
return methodName[:pos], methodName[pos+1:], nil
}

View File

@@ -0,0 +1,85 @@
/*
* Copyright 2019 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package buffer provides an implementation of an unbounded buffer.
package buffer
import "sync"
// Unbounded is an implementation of an unbounded buffer which does not use
// extra goroutines. This is typically used for passing updates from one entity
// to another within gRPC.
//
// All methods on this type are thread-safe and don't block on anything except
// the underlying mutex used for synchronization.
//
// Unbounded supports values of any type to be stored in it by using a channel
// of `interface{}`. This means that a call to Put() incurs an extra memory
// allocation, and also that users need a type assertion while reading. For
// performance critical code paths, using Unbounded is strongly discouraged and
// defining a new type specific implementation of this buffer is preferred. See
// internal/transport/transport.go for an example of this.
type Unbounded struct {
c chan interface{}
mu sync.Mutex
backlog []interface{}
}
// NewUnbounded returns a new instance of Unbounded.
func NewUnbounded() *Unbounded {
return &Unbounded{c: make(chan interface{}, 1)}
}
// Put adds t to the unbounded buffer.
func (b *Unbounded) Put(t interface{}) {
b.mu.Lock()
if len(b.backlog) == 0 {
select {
case b.c <- t:
b.mu.Unlock()
return
default:
}
}
b.backlog = append(b.backlog, t)
b.mu.Unlock()
}
// Load sends the earliest buffered data, if any, onto the read channel
// returned by Get(). Users are expected to call this every time they read a
// value from the read channel.
func (b *Unbounded) Load() {
b.mu.Lock()
if len(b.backlog) > 0 {
select {
case b.c <- b.backlog[0]:
b.backlog[0] = nil
b.backlog = b.backlog[1:]
default:
}
}
b.mu.Unlock()
}
// Get returns a read channel on which values added to the buffer, via Put(),
// are sent on.
//
// Upon reading a value from this channel, users are expected to call Load() to
// send the next buffered value onto the channel if there is any.
func (b *Unbounded) Get() <-chan interface{} {
return b.c
}

View File

@@ -24,19 +24,26 @@
package channelz
import (
"fmt"
"sort"
"sync"
"sync/atomic"
"time"
"google.golang.org/grpc/grpclog"
)
const (
defaultMaxTraceEntry int32 = 30
)
var (
db dbWrapper
idGen idGenerator
// EntryPerPage defines the number of channelz entries to be shown on a web page.
EntryPerPage = 50
curState int32
EntryPerPage = int64(50)
curState int32
maxTraceEntry = defaultMaxTraceEntry
)
// TurnOn turns on channelz data collection.
@@ -52,6 +59,22 @@ func IsOn() bool {
return atomic.CompareAndSwapInt32(&curState, 1, 1)
}
// SetMaxTraceEntry sets maximum number of trace entry per entity (i.e. channel/subchannel).
// Setting it to 0 will disable channel tracing.
func SetMaxTraceEntry(i int32) {
atomic.StoreInt32(&maxTraceEntry, i)
}
// ResetMaxTraceEntryToDefault resets the maximum number of trace entry per entity to default.
func ResetMaxTraceEntryToDefault() {
atomic.StoreInt32(&maxTraceEntry, defaultMaxTraceEntry)
}
func getMaxTraceEntry() int {
i := atomic.LoadInt32(&maxTraceEntry)
return int(i)
}
// dbWarpper wraps around a reference to internal channelz data storage, and
// provide synchronized functionality to set and get the reference.
type dbWrapper struct {
@@ -73,9 +96,14 @@ func (d *dbWrapper) get() *channelMap {
// NewChannelzStorage initializes channelz data storage and id generator.
//
// This function returns a cleanup function to wait for all channelz state to be reset by the
// grpc goroutines when those entities get closed. By using this cleanup function, we make sure tests
// don't mess up each other, i.e. lingering goroutine from previous test doing entity removal happen
// to remove some entity just register by the new test, since the id space is the same.
//
// Note: This function is exported for testing purpose only. User should not call
// it in most cases.
func NewChannelzStorage() {
func NewChannelzStorage() (cleanup func() error) {
db.set(&channelMap{
topLevelChannels: make(map[int64]struct{}),
channels: make(map[int64]*channel),
@@ -85,26 +113,48 @@ func NewChannelzStorage() {
subChannels: make(map[int64]*subChannel),
})
idGen.reset()
return func() error {
var err error
cm := db.get()
if cm == nil {
return nil
}
for i := 0; i < 1000; i++ {
cm.mu.Lock()
if len(cm.topLevelChannels) == 0 && len(cm.servers) == 0 && len(cm.channels) == 0 && len(cm.subChannels) == 0 && len(cm.listenSockets) == 0 && len(cm.normalSockets) == 0 {
cm.mu.Unlock()
// all things stored in the channelz map have been cleared.
return nil
}
cm.mu.Unlock()
time.Sleep(10 * time.Millisecond)
}
cm.mu.Lock()
err = fmt.Errorf("after 10s the channelz map has not been cleaned up yet, topchannels: %d, servers: %d, channels: %d, subchannels: %d, listen sockets: %d, normal sockets: %d", len(cm.topLevelChannels), len(cm.servers), len(cm.channels), len(cm.subChannels), len(cm.listenSockets), len(cm.normalSockets))
cm.mu.Unlock()
return err
}
}
// GetTopChannels returns a slice of top channel's ChannelMetric, along with a
// boolean indicating whether there's more top channels to be queried for.
//
// The arg id specifies that only top channel with id at or above it will be included
// in the result. The returned slice is up to a length of EntryPerPage, and is
// sorted in ascending id order.
func GetTopChannels(id int64) ([]*ChannelMetric, bool) {
return db.get().GetTopChannels(id)
// in the result. The returned slice is up to a length of the arg maxResults or
// EntryPerPage if maxResults is zero, and is sorted in ascending id order.
func GetTopChannels(id int64, maxResults int64) ([]*ChannelMetric, bool) {
return db.get().GetTopChannels(id, maxResults)
}
// GetServers returns a slice of server's ServerMetric, along with a
// boolean indicating whether there's more servers to be queried for.
//
// The arg id specifies that only server with id at or above it will be included
// in the result. The returned slice is up to a length of EntryPerPage, and is
// sorted in ascending id order.
func GetServers(id int64) ([]*ServerMetric, bool) {
return db.get().GetServers(id)
// in the result. The returned slice is up to a length of the arg maxResults or
// EntryPerPage if maxResults is zero, and is sorted in ascending id order.
func GetServers(id int64, maxResults int64) ([]*ServerMetric, bool) {
return db.get().GetServers(id, maxResults)
}
// GetServerSockets returns a slice of server's (identified by id) normal socket's
@@ -112,10 +162,10 @@ func GetServers(id int64) ([]*ServerMetric, bool) {
// be queried for.
//
// The arg startID specifies that only sockets with id at or above it will be
// included in the result. The returned slice is up to a length of EntryPerPage,
// and is sorted in ascending id order.
func GetServerSockets(id int64, startID int64) ([]*SocketMetric, bool) {
return db.get().GetServerSockets(id, startID)
// included in the result. The returned slice is up to a length of the arg maxResults
// or EntryPerPage if maxResults is zero, and is sorted in ascending id order.
func GetServerSockets(id int64, startID int64, maxResults int64) ([]*SocketMetric, bool) {
return db.get().GetServerSockets(id, startID, maxResults)
}
// GetChannel returns the ChannelMetric for the channel (identified by id).
@@ -133,6 +183,11 @@ func GetSocket(id int64) *SocketMetric {
return db.get().GetSocket(id)
}
// GetServer returns the ServerMetric for the server (identified by id).
func GetServer(id int64) *ServerMetric {
return db.get().GetServer(id)
}
// RegisterChannel registers the given channel c in channelz database with ref
// as its reference name, and add it to the child list of its parent (identified
// by pid). pid = 0 means no parent. It returns the unique channelz tracking id
@@ -146,6 +201,7 @@ func RegisterChannel(c Channel, pid int64, ref string) int64 {
nestedChans: make(map[int64]string),
id: id,
pid: pid,
trace: &channelTrace{createdTime: time.Now(), events: make([]*TraceEvent, 0, getMaxTraceEntry())},
}
if pid == 0 {
db.get().addChannel(id, cn, true, pid, ref)
@@ -170,6 +226,7 @@ func RegisterSubChannel(c Channel, pid int64, ref string) int64 {
sockets: make(map[int64]string),
id: id,
pid: pid,
trace: &channelTrace{createdTime: time.Now(), events: make([]*TraceEvent, 0, getMaxTraceEntry())},
}
db.get().addSubChannel(id, sc, pid, ref)
return id
@@ -226,6 +283,24 @@ func RemoveEntry(id int64) {
db.get().removeEntry(id)
}
// TraceEventDesc is what the caller of AddTraceEvent should provide to describe the event to be added
// to the channel trace.
// The Parent field is optional. It is used for event that will be recorded in the entity's parent
// trace also.
type TraceEventDesc struct {
Desc string
Severity Severity
Parent *TraceEventDesc
}
// AddTraceEvent adds trace related to the entity with specified id, using the provided TraceEventDesc.
func AddTraceEvent(id int64, desc *TraceEventDesc) {
if getMaxTraceEntry() == 0 {
return
}
db.get().traceEvent(id, desc)
}
// channelMap is the storage data structure for channelz.
// Methods of channelMap can be divided in two two categories with respect to locking.
// 1. Methods acquire the global lock.
@@ -251,6 +326,7 @@ func (c *channelMap) addServer(id int64, s *server) {
func (c *channelMap) addChannel(id int64, cn *channel, isTopChannel bool, pid int64, ref string) {
c.mu.Lock()
cn.cm = c
cn.trace.cm = c
c.channels[id] = cn
if isTopChannel {
c.topLevelChannels[id] = struct{}{}
@@ -263,6 +339,7 @@ func (c *channelMap) addChannel(id int64, cn *channel, isTopChannel bool, pid in
func (c *channelMap) addSubChannel(id int64, sc *subChannel, pid int64, ref string) {
c.mu.Lock()
sc.cm = c
sc.trace.cm = c
c.subChannels[id] = sc
c.findEntry(pid).addChild(id, sc)
c.mu.Unlock()
@@ -284,16 +361,25 @@ func (c *channelMap) addNormalSocket(id int64, ns *normalSocket, pid int64, ref
c.mu.Unlock()
}
// removeEntry triggers the removal of an entry, which may not indeed delete the
// entry, if it has to wait on the deletion of its children, or may lead to a chain
// of entry deletion. For example, deleting the last socket of a gracefully shutting
// down server will lead to the server being also deleted.
// removeEntry triggers the removal of an entry, which may not indeed delete the entry, if it has to
// wait on the deletion of its children and until no other entity's channel trace references it.
// It may lead to a chain of entry deletion. For example, deleting the last socket of a gracefully
// shutting down server will lead to the server being also deleted.
func (c *channelMap) removeEntry(id int64) {
c.mu.Lock()
c.findEntry(id).triggerDelete()
c.mu.Unlock()
}
// c.mu must be held by the caller
func (c *channelMap) decrTraceRefCount(id int64) {
e := c.findEntry(id)
if v, ok := e.(tracedChannel); ok {
v.decrTraceRefCount()
e.deleteSelfIfReady()
}
}
// c.mu must be held by the caller.
func (c *channelMap) findEntry(id int64) entry {
var v entry
@@ -347,6 +433,39 @@ func (c *channelMap) deleteEntry(id int64) {
}
}
func (c *channelMap) traceEvent(id int64, desc *TraceEventDesc) {
c.mu.Lock()
child := c.findEntry(id)
childTC, ok := child.(tracedChannel)
if !ok {
c.mu.Unlock()
return
}
childTC.getChannelTrace().append(&TraceEvent{Desc: desc.Desc, Severity: desc.Severity, Timestamp: time.Now()})
if desc.Parent != nil {
parent := c.findEntry(child.getParentID())
var chanType RefChannelType
switch child.(type) {
case *channel:
chanType = RefChannel
case *subChannel:
chanType = RefSubChannel
}
if parentTC, ok := parent.(tracedChannel); ok {
parentTC.getChannelTrace().append(&TraceEvent{
Desc: desc.Parent.Desc,
Severity: desc.Parent.Severity,
Timestamp: time.Now(),
RefID: id,
RefName: childTC.getRefName(),
RefType: chanType,
})
childTC.incrTraceRefCount()
}
}
c.mu.Unlock()
}
type int64Slice []int64
func (s int64Slice) Len() int { return len(s) }
@@ -361,29 +480,32 @@ func copyMap(m map[int64]string) map[int64]string {
return n
}
func min(a, b int) int {
func min(a, b int64) int64 {
if a < b {
return a
}
return b
}
func (c *channelMap) GetTopChannels(id int64) ([]*ChannelMetric, bool) {
func (c *channelMap) GetTopChannels(id int64, maxResults int64) ([]*ChannelMetric, bool) {
if maxResults <= 0 {
maxResults = EntryPerPage
}
c.mu.RLock()
l := len(c.topLevelChannels)
l := int64(len(c.topLevelChannels))
ids := make([]int64, 0, l)
cns := make([]*channel, 0, min(l, EntryPerPage))
cns := make([]*channel, 0, min(l, maxResults))
for k := range c.topLevelChannels {
ids = append(ids, k)
}
sort.Sort(int64Slice(ids))
idx := sort.Search(len(ids), func(i int) bool { return ids[i] >= id })
count := 0
count := int64(0)
var end bool
var t []*ChannelMetric
for i, v := range ids[idx:] {
if count == EntryPerPage {
if count == maxResults {
break
}
if cn, ok := c.channels[v]; ok {
@@ -408,25 +530,29 @@ func (c *channelMap) GetTopChannels(id int64) ([]*ChannelMetric, bool) {
t[i].ChannelData = cn.c.ChannelzMetric()
t[i].ID = cn.id
t[i].RefName = cn.refName
t[i].Trace = cn.trace.dumpData()
}
return t, end
}
func (c *channelMap) GetServers(id int64) ([]*ServerMetric, bool) {
func (c *channelMap) GetServers(id, maxResults int64) ([]*ServerMetric, bool) {
if maxResults <= 0 {
maxResults = EntryPerPage
}
c.mu.RLock()
l := len(c.servers)
l := int64(len(c.servers))
ids := make([]int64, 0, l)
ss := make([]*server, 0, min(l, EntryPerPage))
ss := make([]*server, 0, min(l, maxResults))
for k := range c.servers {
ids = append(ids, k)
}
sort.Sort(int64Slice(ids))
idx := sort.Search(len(ids), func(i int) bool { return ids[i] >= id })
count := 0
count := int64(0)
var end bool
var s []*ServerMetric
for i, v := range ids[idx:] {
if count == EntryPerPage {
if count == maxResults {
break
}
if svr, ok := c.servers[v]; ok {
@@ -454,7 +580,10 @@ func (c *channelMap) GetServers(id int64) ([]*ServerMetric, bool) {
return s, end
}
func (c *channelMap) GetServerSockets(id int64, startID int64) ([]*SocketMetric, bool) {
func (c *channelMap) GetServerSockets(id int64, startID int64, maxResults int64) ([]*SocketMetric, bool) {
if maxResults <= 0 {
maxResults = EntryPerPage
}
var svr *server
var ok bool
c.mu.RLock()
@@ -464,18 +593,18 @@ func (c *channelMap) GetServerSockets(id int64, startID int64) ([]*SocketMetric,
return nil, true
}
svrskts := svr.sockets
l := len(svrskts)
l := int64(len(svrskts))
ids := make([]int64, 0, l)
sks := make([]*normalSocket, 0, min(l, EntryPerPage))
sks := make([]*normalSocket, 0, min(l, maxResults))
for k := range svrskts {
ids = append(ids, k)
}
sort.Sort((int64Slice(ids)))
idx := sort.Search(len(ids), func(i int) bool { return ids[i] >= id })
count := 0
sort.Sort(int64Slice(ids))
idx := sort.Search(len(ids), func(i int) bool { return ids[i] >= startID })
count := int64(0)
var end bool
for i, v := range ids[idx:] {
if count == EntryPerPage {
if count == maxResults {
break
}
if ns, ok := c.normalSockets[v]; ok {
@@ -514,10 +643,14 @@ func (c *channelMap) GetChannel(id int64) *ChannelMetric {
}
cm.NestedChans = copyMap(cn.nestedChans)
cm.SubChans = copyMap(cn.subChans)
// cn.c can be set to &dummyChannel{} when deleteSelfFromMap is called. Save a copy of cn.c when
// holding the lock to prevent potential data race.
chanCopy := cn.c
c.mu.RUnlock()
cm.ChannelData = cn.c.ChannelzMetric()
cm.ChannelData = chanCopy.ChannelzMetric()
cm.ID = cn.id
cm.RefName = cn.refName
cm.Trace = cn.trace.dumpData()
return cm
}
@@ -532,10 +665,14 @@ func (c *channelMap) GetSubChannel(id int64) *SubChannelMetric {
return nil
}
cm.Sockets = copyMap(sc.sockets)
// sc.c can be set to &dummyChannel{} when deleteSelfFromMap is called. Save a copy of sc.c when
// holding the lock to prevent potential data race.
chanCopy := sc.c
c.mu.RUnlock()
cm.ChannelData = sc.c.ChannelzMetric()
cm.ChannelData = chanCopy.ChannelzMetric()
cm.ID = sc.id
cm.RefName = sc.refName
cm.Trace = sc.trace.dumpData()
return cm
}
@@ -560,6 +697,23 @@ func (c *channelMap) GetSocket(id int64) *SocketMetric {
return nil
}
func (c *channelMap) GetServer(id int64) *ServerMetric {
sm := &ServerMetric{}
var svr *server
var ok bool
c.mu.RLock()
if svr, ok = c.servers[id]; !ok {
c.mu.RUnlock()
return nil
}
sm.ListenSockets = copyMap(svr.listenSockets)
c.mu.RUnlock()
sm.ID = svr.id
sm.RefName = svr.refName
sm.ServerData = svr.s.ChannelzMetric()
return sm
}
type idGenerator struct {
id int64
}

View File

@@ -20,6 +20,8 @@ package channelz
import (
"net"
"sync"
"sync/atomic"
"time"
"google.golang.org/grpc/connectivity"
@@ -40,6 +42,8 @@ type entry interface {
// deleteSelfIfReady check whether triggerDelete() has been called before, and whether child
// list is now empty. If both conditions are met, then delete self from database.
deleteSelfIfReady()
// getParentID returns parent ID of the entry. 0 value parent ID means no parent.
getParentID() int64
}
// dummyEntry is a fake entry to handle entry not found case.
@@ -73,6 +77,10 @@ func (*dummyEntry) deleteSelfIfReady() {
// code should not reach here. deleteSelfIfReady is always called on an existing entry.
}
func (*dummyEntry) getParentID() int64 {
return 0
}
// ChannelMetric defines the info channelz provides for a specific Channel, which
// includes ChannelInternalMetric and channelz-specific data, such as channelz id,
// child list, etc.
@@ -95,6 +103,8 @@ type ChannelMetric struct {
// Note current grpc implementation doesn't allow channel having sockets directly,
// therefore, this is field is unused.
Sockets map[int64]string
// Trace contains the most recent traced events.
Trace *ChannelTrace
}
// SubChannelMetric defines the info channelz provides for a specific SubChannel,
@@ -121,6 +131,8 @@ type SubChannelMetric struct {
// Sockets tracks the socket type children of this subchannel in the format of a map
// from socket channelz id to corresponding reference string.
Sockets map[int64]string
// Trace contains the most recent traced events.
Trace *ChannelTrace
}
// ChannelInternalMetric defines the struct that the implementor of Channel interface
@@ -138,7 +150,35 @@ type ChannelInternalMetric struct {
CallsFailed int64
// The last time a call was started on the channel.
LastCallStartedTimestamp time.Time
//TODO: trace
}
// ChannelTrace stores traced events on a channel/subchannel and related info.
type ChannelTrace struct {
// EventNum is the number of events that ever got traced (i.e. including those that have been deleted)
EventNum int64
// CreationTime is the creation time of the trace.
CreationTime time.Time
// Events stores the most recent trace events (up to $maxTraceEntry, newer event will overwrite the
// oldest one)
Events []*TraceEvent
}
// TraceEvent represent a single trace event
type TraceEvent struct {
// Desc is a simple description of the trace event.
Desc string
// Severity states the severity of this trace event.
Severity Severity
// Timestamp is the event time.
Timestamp time.Time
// RefID is the id of the entity that gets referenced in the event. RefID is 0 if no other entity is
// involved in this event.
// e.g. SubChannel (id: 4[]) Created. --> RefID = 4, RefName = "" (inside [])
RefID int64
// RefName is the reference name for the entity that gets referenced in the event.
RefName string
// RefType indicates the referenced entity type, i.e Channel or SubChannel.
RefType RefChannelType
}
// Channel is the interface that should be satisfied in order to be tracked by
@@ -147,6 +187,12 @@ type Channel interface {
ChannelzMetric() *ChannelInternalMetric
}
type dummyChannel struct{}
func (d *dummyChannel) ChannelzMetric() *ChannelInternalMetric {
return &ChannelInternalMetric{}
}
type channel struct {
refName string
c Channel
@@ -156,6 +202,10 @@ type channel struct {
id int64
pid int64
cm *channelMap
trace *channelTrace
// traceRefCount is the number of trace events that reference this channel.
// Non-zero traceRefCount means the trace of this channel cannot be deleted.
traceRefCount int32
}
func (c *channel) addChild(id int64, e entry) {
@@ -180,25 +230,96 @@ func (c *channel) triggerDelete() {
c.deleteSelfIfReady()
}
func (c *channel) deleteSelfIfReady() {
func (c *channel) getParentID() int64 {
return c.pid
}
// deleteSelfFromTree tries to delete the channel from the channelz entry relation tree, which means
// deleting the channel reference from its parent's child list.
//
// In order for a channel to be deleted from the tree, it must meet the criteria that, removal of the
// corresponding grpc object has been invoked, and the channel does not have any children left.
//
// The returned boolean value indicates whether the channel has been successfully deleted from tree.
func (c *channel) deleteSelfFromTree() (deleted bool) {
if !c.closeCalled || len(c.subChans)+len(c.nestedChans) != 0 {
return
return false
}
c.cm.deleteEntry(c.id)
// not top channel
if c.pid != 0 {
c.cm.findEntry(c.pid).deleteChild(c.id)
}
return true
}
// deleteSelfFromMap checks whether it is valid to delete the channel from the map, which means
// deleting the channel from channelz's tracking entirely. Users can no longer use id to query the
// channel, and its memory will be garbage collected.
//
// The trace reference count of the channel must be 0 in order to be deleted from the map. This is
// specified in the channel tracing gRFC that as long as some other trace has reference to an entity,
// the trace of the referenced entity must not be deleted. In order to release the resource allocated
// by grpc, the reference to the grpc object is reset to a dummy object.
//
// deleteSelfFromMap must be called after deleteSelfFromTree returns true.
//
// It returns a bool to indicate whether the channel can be safely deleted from map.
func (c *channel) deleteSelfFromMap() (delete bool) {
if c.getTraceRefCount() != 0 {
c.c = &dummyChannel{}
return false
}
return true
}
// deleteSelfIfReady tries to delete the channel itself from the channelz database.
// The delete process includes two steps:
// 1. delete the channel from the entry relation tree, i.e. delete the channel reference from its
// parent's child list.
// 2. delete the channel from the map, i.e. delete the channel entirely from channelz. Lookup by id
// will return entry not found error.
func (c *channel) deleteSelfIfReady() {
if !c.deleteSelfFromTree() {
return
}
if !c.deleteSelfFromMap() {
return
}
c.cm.deleteEntry(c.id)
c.trace.clear()
}
func (c *channel) getChannelTrace() *channelTrace {
return c.trace
}
func (c *channel) incrTraceRefCount() {
atomic.AddInt32(&c.traceRefCount, 1)
}
func (c *channel) decrTraceRefCount() {
atomic.AddInt32(&c.traceRefCount, -1)
}
func (c *channel) getTraceRefCount() int {
i := atomic.LoadInt32(&c.traceRefCount)
return int(i)
}
func (c *channel) getRefName() string {
return c.refName
}
type subChannel struct {
refName string
c Channel
closeCalled bool
sockets map[int64]string
id int64
pid int64
cm *channelMap
refName string
c Channel
closeCalled bool
sockets map[int64]string
id int64
pid int64
cm *channelMap
trace *channelTrace
traceRefCount int32
}
func (sc *subChannel) addChild(id int64, e entry) {
@@ -219,12 +340,82 @@ func (sc *subChannel) triggerDelete() {
sc.deleteSelfIfReady()
}
func (sc *subChannel) deleteSelfIfReady() {
func (sc *subChannel) getParentID() int64 {
return sc.pid
}
// deleteSelfFromTree tries to delete the subchannel from the channelz entry relation tree, which
// means deleting the subchannel reference from its parent's child list.
//
// In order for a subchannel to be deleted from the tree, it must meet the criteria that, removal of
// the corresponding grpc object has been invoked, and the subchannel does not have any children left.
//
// The returned boolean value indicates whether the channel has been successfully deleted from tree.
func (sc *subChannel) deleteSelfFromTree() (deleted bool) {
if !sc.closeCalled || len(sc.sockets) != 0 {
return false
}
sc.cm.findEntry(sc.pid).deleteChild(sc.id)
return true
}
// deleteSelfFromMap checks whether it is valid to delete the subchannel from the map, which means
// deleting the subchannel from channelz's tracking entirely. Users can no longer use id to query
// the subchannel, and its memory will be garbage collected.
//
// The trace reference count of the subchannel must be 0 in order to be deleted from the map. This is
// specified in the channel tracing gRFC that as long as some other trace has reference to an entity,
// the trace of the referenced entity must not be deleted. In order to release the resource allocated
// by grpc, the reference to the grpc object is reset to a dummy object.
//
// deleteSelfFromMap must be called after deleteSelfFromTree returns true.
//
// It returns a bool to indicate whether the channel can be safely deleted from map.
func (sc *subChannel) deleteSelfFromMap() (delete bool) {
if sc.getTraceRefCount() != 0 {
// free the grpc struct (i.e. addrConn)
sc.c = &dummyChannel{}
return false
}
return true
}
// deleteSelfIfReady tries to delete the subchannel itself from the channelz database.
// The delete process includes two steps:
// 1. delete the subchannel from the entry relation tree, i.e. delete the subchannel reference from
// its parent's child list.
// 2. delete the subchannel from the map, i.e. delete the subchannel entirely from channelz. Lookup
// by id will return entry not found error.
func (sc *subChannel) deleteSelfIfReady() {
if !sc.deleteSelfFromTree() {
return
}
if !sc.deleteSelfFromMap() {
return
}
sc.cm.deleteEntry(sc.id)
sc.cm.findEntry(sc.pid).deleteChild(sc.id)
sc.trace.clear()
}
func (sc *subChannel) getChannelTrace() *channelTrace {
return sc.trace
}
func (sc *subChannel) incrTraceRefCount() {
atomic.AddInt32(&sc.traceRefCount, 1)
}
func (sc *subChannel) decrTraceRefCount() {
atomic.AddInt32(&sc.traceRefCount, -1)
}
func (sc *subChannel) getTraceRefCount() int {
i := atomic.LoadInt32(&sc.traceRefCount)
return int(i)
}
func (sc *subChannel) getRefName() string {
return sc.refName
}
// SocketMetric defines the info channelz provides for a specific Socket, which
@@ -318,6 +509,10 @@ func (ls *listenSocket) deleteSelfIfReady() {
grpclog.Errorf("cannot call deleteSelfIfReady on a listen socket")
}
func (ls *listenSocket) getParentID() int64 {
return ls.pid
}
type normalSocket struct {
refName string
s Socket
@@ -343,6 +538,10 @@ func (ns *normalSocket) deleteSelfIfReady() {
grpclog.Errorf("cannot call deleteSelfIfReady on a normal socket")
}
func (ns *normalSocket) getParentID() int64 {
return ns.pid
}
// ServerMetric defines the info channelz provides for a specific Server, which
// includes ServerInternalMetric and channelz-specific data, such as channelz id,
// child list, etc.
@@ -370,7 +569,6 @@ type ServerInternalMetric struct {
CallsFailed int64
// The last time a call was started on the server.
LastCallStartedTimestamp time.Time
//TODO: trace
}
// Server is the interface to be satisfied in order to be tracked by channelz as
@@ -417,3 +615,88 @@ func (s *server) deleteSelfIfReady() {
}
s.cm.deleteEntry(s.id)
}
func (s *server) getParentID() int64 {
return 0
}
type tracedChannel interface {
getChannelTrace() *channelTrace
incrTraceRefCount()
decrTraceRefCount()
getRefName() string
}
type channelTrace struct {
cm *channelMap
createdTime time.Time
eventCount int64
mu sync.Mutex
events []*TraceEvent
}
func (c *channelTrace) append(e *TraceEvent) {
c.mu.Lock()
if len(c.events) == getMaxTraceEntry() {
del := c.events[0]
c.events = c.events[1:]
if del.RefID != 0 {
// start recursive cleanup in a goroutine to not block the call originated from grpc.
go func() {
// need to acquire c.cm.mu lock to call the unlocked attemptCleanup func.
c.cm.mu.Lock()
c.cm.decrTraceRefCount(del.RefID)
c.cm.mu.Unlock()
}()
}
}
e.Timestamp = time.Now()
c.events = append(c.events, e)
c.eventCount++
c.mu.Unlock()
}
func (c *channelTrace) clear() {
c.mu.Lock()
for _, e := range c.events {
if e.RefID != 0 {
// caller should have already held the c.cm.mu lock.
c.cm.decrTraceRefCount(e.RefID)
}
}
c.mu.Unlock()
}
// Severity is the severity level of a trace event.
// The canonical enumeration of all valid values is here:
// https://github.com/grpc/grpc-proto/blob/9b13d199cc0d4703c7ea26c9c330ba695866eb23/grpc/channelz/v1/channelz.proto#L126.
type Severity int
const (
// CtUNKNOWN indicates unknown severity of a trace event.
CtUNKNOWN Severity = iota
// CtINFO indicates info level severity of a trace event.
CtINFO
// CtWarning indicates warning level severity of a trace event.
CtWarning
// CtError indicates error level severity of a trace event.
CtError
)
// RefChannelType is the type of the entity being referenced in a trace event.
type RefChannelType int
const (
// RefChannel indicates the referenced entity is a Channel.
RefChannel RefChannelType = iota
// RefSubChannel indicates the referenced entity is a SubChannel.
RefSubChannel
)
func (c *channelTrace) dumpData() *ChannelTrace {
c.mu.Lock()
ct := &ChannelTrace{EventNum: c.eventCount, CreationTime: c.createdTime}
ct.Events = c.events[:len(c.events)]
c.mu.Unlock()
return ct
}

View File

@@ -1,4 +1,4 @@
// +build !appengine,go1.7
// +build !appengine
/*
*

View File

@@ -1,4 +1,4 @@
// +build !linux appengine !go1.7
// +build !linux appengine
/*
*
@@ -20,11 +20,13 @@
package channelz
import "google.golang.org/grpc/grpclog"
import (
"sync"
func init() {
grpclog.Infof("Channelz: socket options are not supported on non-linux os and appengine.")
}
"google.golang.org/grpc/grpclog"
)
var once sync.Once
// SocketOptionData defines the struct to hold socket option data, and related
// getter function to obtain info from fd.
@@ -35,4 +37,8 @@ type SocketOptionData struct {
// Getsockopt defines the function to get socket options requested by channelz.
// It is to be passed to syscall.RawConn.Control().
// Windows OS doesn't support Socket Option
func (s *SocketOptionData) Getsockopt(fd uintptr) {}
func (s *SocketOptionData) Getsockopt(fd uintptr) {
once.Do(func() {
grpclog.Warningln("Channelz: socket options are not supported on non-linux os and appengine.")
})
}

View File

@@ -1,4 +1,4 @@
// +build linux,go1.9,!appengine
// +build linux,!appengine
/*
*

View File

@@ -1,4 +1,4 @@
// +build !linux !go1.9 appengine
// +build !linux appengine
/*
*

View File

@@ -25,11 +25,14 @@ import (
)
const (
prefix = "GRPC_GO_"
retryStr = prefix + "RETRY"
prefix = "GRPC_GO_"
retryStr = prefix + "RETRY"
txtErrIgnoreStr = prefix + "IGNORE_TXT_ERRORS"
)
var (
// Retry is set if retry is explicitly enabled via "GRPC_GO_RETRY=on".
Retry = strings.EqualFold(os.Getenv(retryStr), "on")
// TXTErrIgnore is set if TXT errors should be ignored ("GRPC_GO_IGNORE_TXT_ERRORS" is not "false").
TXTErrIgnore = !strings.EqualFold(os.Getenv(retryStr), "false")
)

View File

@@ -0,0 +1,61 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package grpcsync implements additional synchronization primitives built upon
// the sync package.
package grpcsync
import (
"sync"
"sync/atomic"
)
// Event represents a one-time event that may occur in the future.
type Event struct {
fired int32
c chan struct{}
o sync.Once
}
// Fire causes e to complete. It is safe to call multiple times, and
// concurrently. It returns true iff this call to Fire caused the signaling
// channel returned by Done to close.
func (e *Event) Fire() bool {
ret := false
e.o.Do(func() {
atomic.StoreInt32(&e.fired, 1)
close(e.c)
ret = true
})
return ret
}
// Done returns a channel that will be closed when Fire is called.
func (e *Event) Done() <-chan struct{} {
return e.c
}
// HasFired returns true if Fire has been called.
func (e *Event) HasFired() bool {
return atomic.LoadInt32(&e.fired) == 1
}
// NewEvent returns a new, ready-to-use Event.
func NewEvent() *Event {
return &Event{c: make(chan struct{})}
}

View File

@@ -20,9 +20,53 @@
// symbols to avoid circular dependencies.
package internal
var (
// WithContextDialer is exported by clientconn.go
WithContextDialer interface{} // func(context.Context, string) (net.Conn, error) grpc.DialOption
// WithResolverBuilder is exported by clientconn.go
WithResolverBuilder interface{} // func (resolver.Builder) grpc.DialOption
import (
"context"
"time"
"google.golang.org/grpc/connectivity"
)
var (
// WithHealthCheckFunc is set by dialoptions.go
WithHealthCheckFunc interface{} // func (HealthChecker) DialOption
// HealthCheckFunc is used to provide client-side LB channel health checking
HealthCheckFunc HealthChecker
// BalancerUnregister is exported by package balancer to unregister a balancer.
BalancerUnregister func(name string)
// KeepaliveMinPingTime is the minimum ping interval. This must be 10s by
// default, but tests may wish to set it lower for convenience.
KeepaliveMinPingTime = 10 * time.Second
// StatusRawProto is exported by status/status.go. This func returns a
// pointer to the wrapped Status proto for a given status.Status without a
// call to proto.Clone(). The returned Status proto should not be mutated by
// the caller.
StatusRawProto interface{} // func (*status.Status) *spb.Status
// NewRequestInfoContext creates a new context based on the argument context attaching
// the passed in RequestInfo to the new context.
NewRequestInfoContext interface{} // func(context.Context, credentials.RequestInfo) context.Context
// ParseServiceConfigForTesting is for creating a fake
// ClientConn for resolver testing only
ParseServiceConfigForTesting interface{} // func(string) *serviceconfig.ParseResult
)
// HealthChecker defines the signature of the client-side LB channel health checking function.
//
// The implementation is expected to create a health checking RPC stream by
// calling newStream(), watch for the health status of serviceName, and report
// it's health back by calling setConnectivityState().
//
// The health checking protocol is defined at:
// https://github.com/grpc/grpc/blob/master/doc/health-checking.md
type HealthChecker func(ctx context.Context, newStream func(string) (interface{}, error), setConnectivityState func(connectivity.State, error), serviceName string) error
const (
// CredsBundleModeFallback switches GoogleDefaultCreds to fallback mode.
CredsBundleModeFallback = "fallback"
// CredsBundleModeBalancer switches GoogleDefaultCreds to grpclb balancer
// mode.
CredsBundleModeBalancer = "balancer"
// CredsBundleModeBackendFromBalancer switches GoogleDefaultCreds to mode
// that supports backend returned by grpclb balancer.
CredsBundleModeBackendFromBalancer = "backend-from-balancer"
)

View File

@@ -0,0 +1,441 @@
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package dns implements a dns resolver to be installed as the default resolver
// in grpc.
package dns
import (
"context"
"encoding/json"
"errors"
"fmt"
"net"
"os"
"strconv"
"strings"
"sync"
"time"
"google.golang.org/grpc/grpclog"
"google.golang.org/grpc/internal/envconfig"
"google.golang.org/grpc/internal/grpcrand"
"google.golang.org/grpc/resolver"
"google.golang.org/grpc/serviceconfig"
)
// EnableSRVLookups controls whether the DNS resolver attempts to fetch gRPCLB
// addresses from SRV records. Must not be changed after init time.
var EnableSRVLookups = false
func init() {
resolver.Register(NewBuilder())
}
const (
defaultPort = "443"
defaultDNSSvrPort = "53"
golang = "GO"
// txtPrefix is the prefix string to be prepended to the host name for txt record lookup.
txtPrefix = "_grpc_config."
// In DNS, service config is encoded in a TXT record via the mechanism
// described in RFC-1464 using the attribute name grpc_config.
txtAttribute = "grpc_config="
)
var (
errMissingAddr = errors.New("dns resolver: missing address")
// Addresses ending with a colon that is supposed to be the separator
// between host and port is not allowed. E.g. "::" is a valid address as
// it is an IPv6 address (host only) and "[::]:" is invalid as it ends with
// a colon as the host and port separator
errEndsWithColon = errors.New("dns resolver: missing port after port-separator colon")
)
var (
defaultResolver netResolver = net.DefaultResolver
// To prevent excessive re-resolution, we enforce a rate limit on DNS
// resolution requests.
minDNSResRate = 30 * time.Second
)
var customAuthorityDialler = func(authority string) func(ctx context.Context, network, address string) (net.Conn, error) {
return func(ctx context.Context, network, address string) (net.Conn, error) {
var dialer net.Dialer
return dialer.DialContext(ctx, network, authority)
}
}
var customAuthorityResolver = func(authority string) (netResolver, error) {
host, port, err := parseTarget(authority, defaultDNSSvrPort)
if err != nil {
return nil, err
}
authorityWithPort := net.JoinHostPort(host, port)
return &net.Resolver{
PreferGo: true,
Dial: customAuthorityDialler(authorityWithPort),
}, nil
}
// NewBuilder creates a dnsBuilder which is used to factory DNS resolvers.
func NewBuilder() resolver.Builder {
return &dnsBuilder{}
}
type dnsBuilder struct{}
// Build creates and starts a DNS resolver that watches the name resolution of the target.
func (b *dnsBuilder) Build(target resolver.Target, cc resolver.ClientConn, opts resolver.BuildOptions) (resolver.Resolver, error) {
host, port, err := parseTarget(target.Endpoint, defaultPort)
if err != nil {
return nil, err
}
// IP address.
if ipAddr, ok := formatIP(host); ok {
addr := []resolver.Address{{Addr: ipAddr + ":" + port}}
cc.UpdateState(resolver.State{Addresses: addr})
return deadResolver{}, nil
}
// DNS address (non-IP).
ctx, cancel := context.WithCancel(context.Background())
d := &dnsResolver{
host: host,
port: port,
ctx: ctx,
cancel: cancel,
cc: cc,
rn: make(chan struct{}, 1),
disableServiceConfig: opts.DisableServiceConfig,
}
if target.Authority == "" {
d.resolver = defaultResolver
} else {
d.resolver, err = customAuthorityResolver(target.Authority)
if err != nil {
return nil, err
}
}
d.wg.Add(1)
go d.watcher()
d.ResolveNow(resolver.ResolveNowOptions{})
return d, nil
}
// Scheme returns the naming scheme of this resolver builder, which is "dns".
func (b *dnsBuilder) Scheme() string {
return "dns"
}
type netResolver interface {
LookupHost(ctx context.Context, host string) (addrs []string, err error)
LookupSRV(ctx context.Context, service, proto, name string) (cname string, addrs []*net.SRV, err error)
LookupTXT(ctx context.Context, name string) (txts []string, err error)
}
// deadResolver is a resolver that does nothing.
type deadResolver struct{}
func (deadResolver) ResolveNow(resolver.ResolveNowOptions) {}
func (deadResolver) Close() {}
// dnsResolver watches for the name resolution update for a non-IP target.
type dnsResolver struct {
host string
port string
resolver netResolver
ctx context.Context
cancel context.CancelFunc
cc resolver.ClientConn
// rn channel is used by ResolveNow() to force an immediate resolution of the target.
rn chan struct{}
// wg is used to enforce Close() to return after the watcher() goroutine has finished.
// Otherwise, data race will be possible. [Race Example] in dns_resolver_test we
// replace the real lookup functions with mocked ones to facilitate testing.
// If Close() doesn't wait for watcher() goroutine finishes, race detector sometimes
// will warns lookup (READ the lookup function pointers) inside watcher() goroutine
// has data race with replaceNetFunc (WRITE the lookup function pointers).
wg sync.WaitGroup
disableServiceConfig bool
}
// ResolveNow invoke an immediate resolution of the target that this dnsResolver watches.
func (d *dnsResolver) ResolveNow(resolver.ResolveNowOptions) {
select {
case d.rn <- struct{}{}:
default:
}
}
// Close closes the dnsResolver.
func (d *dnsResolver) Close() {
d.cancel()
d.wg.Wait()
}
func (d *dnsResolver) watcher() {
defer d.wg.Done()
for {
select {
case <-d.ctx.Done():
return
case <-d.rn:
}
state, err := d.lookup()
if err != nil {
d.cc.ReportError(err)
} else {
d.cc.UpdateState(*state)
}
// Sleep to prevent excessive re-resolutions. Incoming resolution requests
// will be queued in d.rn.
t := time.NewTimer(minDNSResRate)
select {
case <-t.C:
case <-d.ctx.Done():
t.Stop()
return
}
}
}
func (d *dnsResolver) lookupSRV() ([]resolver.Address, error) {
if !EnableSRVLookups {
return nil, nil
}
var newAddrs []resolver.Address
_, srvs, err := d.resolver.LookupSRV(d.ctx, "grpclb", "tcp", d.host)
if err != nil {
err = handleDNSError(err, "SRV") // may become nil
return nil, err
}
for _, s := range srvs {
lbAddrs, err := d.resolver.LookupHost(d.ctx, s.Target)
if err != nil {
err = handleDNSError(err, "A") // may become nil
if err == nil {
// If there are other SRV records, look them up and ignore this
// one that does not exist.
continue
}
return nil, err
}
for _, a := range lbAddrs {
ip, ok := formatIP(a)
if !ok {
return nil, fmt.Errorf("dns: error parsing A record IP address %v", a)
}
addr := ip + ":" + strconv.Itoa(int(s.Port))
newAddrs = append(newAddrs, resolver.Address{Addr: addr, Type: resolver.GRPCLB, ServerName: s.Target})
}
}
return newAddrs, nil
}
var filterError = func(err error) error {
if dnsErr, ok := err.(*net.DNSError); ok && !dnsErr.IsTimeout && !dnsErr.IsTemporary {
// Timeouts and temporary errors should be communicated to gRPC to
// attempt another DNS query (with backoff). Other errors should be
// suppressed (they may represent the absence of a TXT record).
return nil
}
return err
}
func handleDNSError(err error, lookupType string) error {
err = filterError(err)
if err != nil {
err = fmt.Errorf("dns: %v record lookup error: %v", lookupType, err)
grpclog.Infoln(err)
}
return err
}
func (d *dnsResolver) lookupTXT() *serviceconfig.ParseResult {
ss, err := d.resolver.LookupTXT(d.ctx, txtPrefix+d.host)
if err != nil {
if envconfig.TXTErrIgnore {
return nil
}
if err = handleDNSError(err, "TXT"); err != nil {
return &serviceconfig.ParseResult{Err: err}
}
return nil
}
var res string
for _, s := range ss {
res += s
}
// TXT record must have "grpc_config=" attribute in order to be used as service config.
if !strings.HasPrefix(res, txtAttribute) {
grpclog.Warningf("dns: TXT record %v missing %v attribute", res, txtAttribute)
// This is not an error; it is the equivalent of not having a service config.
return nil
}
sc := canaryingSC(strings.TrimPrefix(res, txtAttribute))
return d.cc.ParseServiceConfig(sc)
}
func (d *dnsResolver) lookupHost() ([]resolver.Address, error) {
var newAddrs []resolver.Address
addrs, err := d.resolver.LookupHost(d.ctx, d.host)
if err != nil {
err = handleDNSError(err, "A")
return nil, err
}
for _, a := range addrs {
ip, ok := formatIP(a)
if !ok {
return nil, fmt.Errorf("dns: error parsing A record IP address %v", a)
}
addr := ip + ":" + d.port
newAddrs = append(newAddrs, resolver.Address{Addr: addr})
}
return newAddrs, nil
}
func (d *dnsResolver) lookup() (*resolver.State, error) {
srv, srvErr := d.lookupSRV()
addrs, hostErr := d.lookupHost()
if hostErr != nil && (srvErr != nil || len(srv) == 0) {
return nil, hostErr
}
state := &resolver.State{
Addresses: append(addrs, srv...),
}
if !d.disableServiceConfig {
state.ServiceConfig = d.lookupTXT()
}
return state, nil
}
// formatIP returns ok = false if addr is not a valid textual representation of an IP address.
// If addr is an IPv4 address, return the addr and ok = true.
// If addr is an IPv6 address, return the addr enclosed in square brackets and ok = true.
func formatIP(addr string) (addrIP string, ok bool) {
ip := net.ParseIP(addr)
if ip == nil {
return "", false
}
if ip.To4() != nil {
return addr, true
}
return "[" + addr + "]", true
}
// parseTarget takes the user input target string and default port, returns formatted host and port info.
// If target doesn't specify a port, set the port to be the defaultPort.
// If target is in IPv6 format and host-name is enclosed in square brackets, brackets
// are stripped when setting the host.
// examples:
// target: "www.google.com" defaultPort: "443" returns host: "www.google.com", port: "443"
// target: "ipv4-host:80" defaultPort: "443" returns host: "ipv4-host", port: "80"
// target: "[ipv6-host]" defaultPort: "443" returns host: "ipv6-host", port: "443"
// target: ":80" defaultPort: "443" returns host: "localhost", port: "80"
func parseTarget(target, defaultPort string) (host, port string, err error) {
if target == "" {
return "", "", errMissingAddr
}
if ip := net.ParseIP(target); ip != nil {
// target is an IPv4 or IPv6(without brackets) address
return target, defaultPort, nil
}
if host, port, err = net.SplitHostPort(target); err == nil {
if port == "" {
// If the port field is empty (target ends with colon), e.g. "[::1]:", this is an error.
return "", "", errEndsWithColon
}
// target has port, i.e ipv4-host:port, [ipv6-host]:port, host-name:port
if host == "" {
// Keep consistent with net.Dial(): If the host is empty, as in ":80", the local system is assumed.
host = "localhost"
}
return host, port, nil
}
if host, port, err = net.SplitHostPort(target + ":" + defaultPort); err == nil {
// target doesn't have port
return host, port, nil
}
return "", "", fmt.Errorf("invalid target address %v, error info: %v", target, err)
}
type rawChoice struct {
ClientLanguage *[]string `json:"clientLanguage,omitempty"`
Percentage *int `json:"percentage,omitempty"`
ClientHostName *[]string `json:"clientHostName,omitempty"`
ServiceConfig *json.RawMessage `json:"serviceConfig,omitempty"`
}
func containsString(a *[]string, b string) bool {
if a == nil {
return true
}
for _, c := range *a {
if c == b {
return true
}
}
return false
}
func chosenByPercentage(a *int) bool {
if a == nil {
return true
}
return grpcrand.Intn(100)+1 <= *a
}
func canaryingSC(js string) string {
if js == "" {
return ""
}
var rcs []rawChoice
err := json.Unmarshal([]byte(js), &rcs)
if err != nil {
grpclog.Warningf("dns: error parsing service config json: %v", err)
return ""
}
cliHostname, err := os.Hostname()
if err != nil {
grpclog.Warningf("dns: error getting client hostname: %v", err)
return ""
}
var sc string
for _, c := range rcs {
if !containsString(c.ClientLanguage, golang) ||
!chosenByPercentage(c.Percentage) ||
!containsString(c.ClientHostName, cliHostname) ||
c.ServiceConfig == nil {
continue
}
sc = string(*c.ServiceConfig)
break
}
return sc
}

View File

@@ -0,0 +1,33 @@
// +build go1.13
/*
*
* Copyright 2019 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package dns
import "net"
func init() {
filterError = func(err error) error {
if dnsErr, ok := err.(*net.DNSError); ok && dnsErr.IsNotFound {
// The name does not exist; not an error.
return nil
}
return err
}
}

View File

@@ -0,0 +1,57 @@
/*
*
* Copyright 2017 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package passthrough implements a pass-through resolver. It sends the target
// name without scheme back to gRPC as resolved address.
package passthrough
import "google.golang.org/grpc/resolver"
const scheme = "passthrough"
type passthroughBuilder struct{}
func (*passthroughBuilder) Build(target resolver.Target, cc resolver.ClientConn, opts resolver.BuildOptions) (resolver.Resolver, error) {
r := &passthroughResolver{
target: target,
cc: cc,
}
r.start()
return r, nil
}
func (*passthroughBuilder) Scheme() string {
return scheme
}
type passthroughResolver struct {
target resolver.Target
cc resolver.ClientConn
}
func (r *passthroughResolver) start() {
r.cc.UpdateState(resolver.State{Addresses: []resolver.Address{{Addr: r.target.Endpoint}}})
}
func (*passthroughResolver) ResolveNow(o resolver.ResolveNowOptions) {}
func (*passthroughResolver) Close() {}
func init() {
resolver.Register(&passthroughBuilder{})
}

View File

@@ -0,0 +1,114 @@
// +build !appengine
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
// Package syscall provides functionalities that grpc uses to get low-level operating system
// stats/info.
package syscall
import (
"fmt"
"net"
"syscall"
"time"
"golang.org/x/sys/unix"
"google.golang.org/grpc/grpclog"
)
// GetCPUTime returns the how much CPU time has passed since the start of this process.
func GetCPUTime() int64 {
var ts unix.Timespec
if err := unix.ClockGettime(unix.CLOCK_PROCESS_CPUTIME_ID, &ts); err != nil {
grpclog.Fatal(err)
}
return ts.Nano()
}
// Rusage is an alias for syscall.Rusage under linux non-appengine environment.
type Rusage syscall.Rusage
// GetRusage returns the resource usage of current process.
func GetRusage() (rusage *Rusage) {
rusage = new(Rusage)
syscall.Getrusage(syscall.RUSAGE_SELF, (*syscall.Rusage)(rusage))
return
}
// CPUTimeDiff returns the differences of user CPU time and system CPU time used
// between two Rusage structs.
func CPUTimeDiff(first *Rusage, latest *Rusage) (float64, float64) {
f := (*syscall.Rusage)(first)
l := (*syscall.Rusage)(latest)
var (
utimeDiffs = l.Utime.Sec - f.Utime.Sec
utimeDiffus = l.Utime.Usec - f.Utime.Usec
stimeDiffs = l.Stime.Sec - f.Stime.Sec
stimeDiffus = l.Stime.Usec - f.Stime.Usec
)
uTimeElapsed := float64(utimeDiffs) + float64(utimeDiffus)*1.0e-6
sTimeElapsed := float64(stimeDiffs) + float64(stimeDiffus)*1.0e-6
return uTimeElapsed, sTimeElapsed
}
// SetTCPUserTimeout sets the TCP user timeout on a connection's socket
func SetTCPUserTimeout(conn net.Conn, timeout time.Duration) error {
tcpconn, ok := conn.(*net.TCPConn)
if !ok {
// not a TCP connection. exit early
return nil
}
rawConn, err := tcpconn.SyscallConn()
if err != nil {
return fmt.Errorf("error getting raw connection: %v", err)
}
err = rawConn.Control(func(fd uintptr) {
err = syscall.SetsockoptInt(int(fd), syscall.IPPROTO_TCP, unix.TCP_USER_TIMEOUT, int(timeout/time.Millisecond))
})
if err != nil {
return fmt.Errorf("error setting option on socket: %v", err)
}
return nil
}
// GetTCPUserTimeout gets the TCP user timeout on a connection's socket
func GetTCPUserTimeout(conn net.Conn) (opt int, err error) {
tcpconn, ok := conn.(*net.TCPConn)
if !ok {
err = fmt.Errorf("conn is not *net.TCPConn. got %T", conn)
return
}
rawConn, err := tcpconn.SyscallConn()
if err != nil {
err = fmt.Errorf("error getting raw connection: %v", err)
return
}
err = rawConn.Control(func(fd uintptr) {
opt, err = syscall.GetsockoptInt(int(fd), syscall.IPPROTO_TCP, unix.TCP_USER_TIMEOUT)
})
if err != nil {
err = fmt.Errorf("error getting option on socket: %v", err)
return
}
return
}

View File

@@ -0,0 +1,73 @@
// +build !linux appengine
/*
*
* Copyright 2018 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package syscall
import (
"net"
"sync"
"time"
"google.golang.org/grpc/grpclog"
)
var once sync.Once
func log() {
once.Do(func() {
grpclog.Info("CPU time info is unavailable on non-linux or appengine environment.")
})
}
// GetCPUTime returns the how much CPU time has passed since the start of this process.
// It always returns 0 under non-linux or appengine environment.
func GetCPUTime() int64 {
log()
return 0
}
// Rusage is an empty struct under non-linux or appengine environment.
type Rusage struct{}
// GetRusage is a no-op function under non-linux or appengine environment.
func GetRusage() (rusage *Rusage) {
log()
return nil
}
// CPUTimeDiff returns the differences of user CPU time and system CPU time used
// between two Rusage structs. It a no-op function for non-linux or appengine environment.
func CPUTimeDiff(first *Rusage, latest *Rusage) (float64, float64) {
log()
return 0, 0
}
// SetTCPUserTimeout is a no-op function under non-linux or appengine environments
func SetTCPUserTimeout(conn net.Conn, timeout time.Duration) error {
log()
return nil
}
// GetTCPUserTimeout is a no-op function under non-linux or appengine environments
// a negative return value indicates the operation is not supported
func GetTCPUserTimeout(conn net.Conn) (int, error) {
log()
return -1, nil
}

View File

@@ -24,9 +24,10 @@ import (
)
const (
// bdpLimit is the maximum value the flow control windows
// will be increased to.
bdpLimit = (1 << 20) * 4
// bdpLimit is the maximum value the flow control windows will be increased
// to. TCP typically limits this to 4MB, but some systems go up to 16MB.
// Since this is only a limit, it is safe to make it optimistic.
bdpLimit = (1 << 20) * 16
// alpha is a constant factor used to keep a moving average
// of RTTs.
alpha = 0.9

View File

@@ -23,6 +23,7 @@ import (
"fmt"
"runtime"
"sync"
"sync/atomic"
"golang.org/x/net/http2"
"golang.org/x/net/http2/hpack"
@@ -84,24 +85,40 @@ func (il *itemList) isEmpty() bool {
// the control buffer of transport. They represent different aspects of
// control tasks, e.g., flow control, settings, streaming resetting, etc.
// maxQueuedTransportResponseFrames is the most queued "transport response"
// frames we will buffer before preventing new reads from occurring on the
// transport. These are control frames sent in response to client requests,
// such as RST_STREAM due to bad headers or settings acks.
const maxQueuedTransportResponseFrames = 50
type cbItem interface {
isTransportResponseFrame() bool
}
// registerStream is used to register an incoming stream with loopy writer.
type registerStream struct {
streamID uint32
wq *writeQuota
}
func (*registerStream) isTransportResponseFrame() bool { return false }
// headerFrame is also used to register stream on the client-side.
type headerFrame struct {
streamID uint32
hf []hpack.HeaderField
endStream bool // Valid on server side.
initStream func(uint32) (bool, error) // Used only on the client side.
endStream bool // Valid on server side.
initStream func(uint32) error // Used only on the client side.
onWrite func()
wq *writeQuota // write quota for the stream created.
cleanup *cleanupStream // Valid on the server side.
onOrphaned func(error) // Valid on client-side
}
func (h *headerFrame) isTransportResponseFrame() bool {
return h.cleanup != nil && h.cleanup.rst // Results in a RST_STREAM
}
type cleanupStream struct {
streamID uint32
rst bool
@@ -109,6 +126,8 @@ type cleanupStream struct {
onWrite func()
}
func (c *cleanupStream) isTransportResponseFrame() bool { return c.rst } // Results in a RST_STREAM
type dataFrame struct {
streamID uint32
endStream bool
@@ -119,27 +138,41 @@ type dataFrame struct {
onEachWrite func()
}
func (*dataFrame) isTransportResponseFrame() bool { return false }
type incomingWindowUpdate struct {
streamID uint32
increment uint32
}
func (*incomingWindowUpdate) isTransportResponseFrame() bool { return false }
type outgoingWindowUpdate struct {
streamID uint32
increment uint32
}
func (*outgoingWindowUpdate) isTransportResponseFrame() bool {
return false // window updates are throttled by thresholds
}
type incomingSettings struct {
ss []http2.Setting
}
func (*incomingSettings) isTransportResponseFrame() bool { return true } // Results in a settings ACK
type outgoingSettings struct {
ss []http2.Setting
}
func (*outgoingSettings) isTransportResponseFrame() bool { return false }
type incomingGoAway struct {
}
func (*incomingGoAway) isTransportResponseFrame() bool { return false }
type goAway struct {
code http2.ErrCode
debugData []byte
@@ -147,15 +180,21 @@ type goAway struct {
closeConn bool
}
func (*goAway) isTransportResponseFrame() bool { return false }
type ping struct {
ack bool
data [8]byte
}
func (*ping) isTransportResponseFrame() bool { return true }
type outFlowControlSizeRequest struct {
resp chan uint32
}
func (*outFlowControlSizeRequest) isTransportResponseFrame() bool { return false }
type outStreamState int
const (
@@ -238,6 +277,14 @@ type controlBuffer struct {
consumerWaiting bool
list *itemList
err error
// transportResponseFrames counts the number of queued items that represent
// the response of an action initiated by the peer. trfChan is created
// when transportResponseFrames >= maxQueuedTransportResponseFrames and is
// closed and nilled when transportResponseFrames drops below the
// threshold. Both fields are protected by mu.
transportResponseFrames int
trfChan atomic.Value // *chan struct{}
}
func newControlBuffer(done <-chan struct{}) *controlBuffer {
@@ -248,12 +295,24 @@ func newControlBuffer(done <-chan struct{}) *controlBuffer {
}
}
func (c *controlBuffer) put(it interface{}) error {
// throttle blocks if there are too many incomingSettings/cleanupStreams in the
// controlbuf.
func (c *controlBuffer) throttle() {
ch, _ := c.trfChan.Load().(*chan struct{})
if ch != nil {
select {
case <-*ch:
case <-c.done:
}
}
}
func (c *controlBuffer) put(it cbItem) error {
_, err := c.executeAndPut(nil, it)
return err
}
func (c *controlBuffer) executeAndPut(f func(it interface{}) bool, it interface{}) (bool, error) {
func (c *controlBuffer) executeAndPut(f func(it interface{}) bool, it cbItem) (bool, error) {
var wakeUp bool
c.mu.Lock()
if c.err != nil {
@@ -271,6 +330,15 @@ func (c *controlBuffer) executeAndPut(f func(it interface{}) bool, it interface{
c.consumerWaiting = false
}
c.list.enqueue(it)
if it.isTransportResponseFrame() {
c.transportResponseFrames++
if c.transportResponseFrames == maxQueuedTransportResponseFrames {
// We are adding the frame that puts us over the threshold; create
// a throttling channel.
ch := make(chan struct{})
c.trfChan.Store(&ch)
}
}
c.mu.Unlock()
if wakeUp {
select {
@@ -304,7 +372,17 @@ func (c *controlBuffer) get(block bool) (interface{}, error) {
return nil, c.err
}
if !c.list.isEmpty() {
h := c.list.dequeue()
h := c.list.dequeue().(cbItem)
if h.isTransportResponseFrame() {
if c.transportResponseFrames == maxQueuedTransportResponseFrames {
// We are removing the frame that put us over the
// threshold; close and clear the throttling channel.
ch := c.trfChan.Load().(*chan struct{})
close(*ch)
c.trfChan.Store((*chan struct{})(nil))
}
c.transportResponseFrames--
}
c.mu.Unlock()
return h, nil
}
@@ -559,21 +637,17 @@ func (l *loopyWriter) headerHandler(h *headerFrame) error {
func (l *loopyWriter) originateStream(str *outStream) error {
hdr := str.itl.dequeue().(*headerFrame)
sendPing, err := hdr.initStream(str.id)
if err != nil {
if err := hdr.initStream(str.id); err != nil {
if err == ErrConnClosing {
return err
}
// Other errors(errStreamDrain) need not close transport.
return nil
}
if err = l.writeHeader(str.id, hdr.endStream, hdr.hf, hdr.onWrite); err != nil {
if err := l.writeHeader(str.id, hdr.endStream, hdr.hf, hdr.onWrite); err != nil {
return err
}
l.estdStreams[str.id] = str
if sendPing {
return l.pingHandler(&ping{data: [8]byte{}})
}
return nil
}

View File

@@ -149,6 +149,7 @@ func (f *inFlow) maybeAdjust(n uint32) uint32 {
n = uint32(math.MaxInt32)
}
f.mu.Lock()
defer f.mu.Unlock()
// estSenderQuota is the receiver's view of the maximum number of bytes the sender
// can send without a window update.
estSenderQuota := int32(f.limit - (f.pendingData + f.pendingUpdate))
@@ -169,10 +170,8 @@ func (f *inFlow) maybeAdjust(n uint32) uint32 {
// is padded; We will fallback on the current available window(at least a 1/4th of the limit).
f.delta = n
}
f.mu.Unlock()
return f.delta
}
f.mu.Unlock()
return 0
}

View File

@@ -1,52 +0,0 @@
// +build go1.6,!go1.7
/*
*
* Copyright 2016 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package transport
import (
"net"
"net/http"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
"golang.org/x/net/context"
)
// dialContext connects to the address on the named network.
func dialContext(ctx context.Context, network, address string) (net.Conn, error) {
return (&net.Dialer{Cancel: ctx.Done()}).Dial(network, address)
}
// ContextErr converts the error from context package into a status error.
func ContextErr(err error) error {
switch err {
case context.DeadlineExceeded:
return status.Error(codes.DeadlineExceeded, err.Error())
case context.Canceled:
return status.Error(codes.Canceled, err.Error())
}
return status.Errorf(codes.Internal, "Unexpected error from context packet: %v", err)
}
// contextFromRequest returns a background context.
func contextFromRequest(r *http.Request) context.Context {
return context.Background()
}

View File

@@ -1,53 +0,0 @@
// +build go1.7
/*
*
* Copyright 2016 gRPC authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*/
package transport
import (
"context"
"net"
"net/http"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
netctx "golang.org/x/net/context"
)
// dialContext connects to the address on the named network.
func dialContext(ctx context.Context, network, address string) (net.Conn, error) {
return (&net.Dialer{}).DialContext(ctx, network, address)
}
// ContextErr converts the error from context package into a status error.
func ContextErr(err error) error {
switch err {
case context.DeadlineExceeded, netctx.DeadlineExceeded:
return status.Error(codes.DeadlineExceeded, err.Error())
case context.Canceled, netctx.Canceled:
return status.Error(codes.Canceled, err.Error())
}
return status.Errorf(codes.Internal, "Unexpected error from context packet: %v", err)
}
// contextFromRequest returns a context from the HTTP Request.
func contextFromRequest(r *http.Request) context.Context {
return r.Context()
}

View File

@@ -24,6 +24,8 @@
package transport
import (
"bytes"
"context"
"errors"
"fmt"
"io"
@@ -34,7 +36,6 @@ import (
"time"
"github.com/golang/protobuf/proto"
"golang.org/x/net/context"
"golang.org/x/net/http2"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/credentials"
@@ -63,9 +64,6 @@ func NewServerHandlerTransport(w http.ResponseWriter, r *http.Request, stats sta
if _, ok := w.(http.Flusher); !ok {
return nil, errors.New("gRPC requires a ResponseWriter supporting http.Flusher")
}
if _, ok := w.(http.CloseNotifier); !ok {
return nil, errors.New("gRPC requires a ResponseWriter supporting http.CloseNotifier")
}
st := &serverHandlerTransport{
rw: w,
@@ -176,17 +174,11 @@ func (a strAddr) String() string { return string(a) }
// do runs fn in the ServeHTTP goroutine.
func (ht *serverHandlerTransport) do(fn func()) error {
// Avoid a panic writing to closed channel. Imperfect but maybe good enough.
select {
case <-ht.closedCh:
return ErrConnClosing
default:
select {
case ht.writes <- fn:
return nil
case <-ht.closedCh:
return ErrConnClosing
}
case ht.writes <- fn:
return nil
}
}
@@ -235,9 +227,10 @@ func (ht *serverHandlerTransport) WriteStatus(s *Stream, st *status.Status) erro
if err == nil { // transport has not been closed
if ht.stats != nil {
ht.stats.HandleRPC(s.Context(), &stats.OutTrailer{})
ht.stats.HandleRPC(s.Context(), &stats.OutTrailer{
Trailer: s.trailer.Copy(),
})
}
close(ht.writes)
}
ht.Close()
return err
@@ -298,7 +291,9 @@ func (ht *serverHandlerTransport) WriteHeader(s *Stream, md metadata.MD) error {
if err == nil {
if ht.stats != nil {
ht.stats.HandleRPC(s.Context(), &stats.OutHeader{})
ht.stats.HandleRPC(s.Context(), &stats.OutHeader{
Header: md.Copy(),
})
}
}
return err
@@ -307,7 +302,7 @@ func (ht *serverHandlerTransport) WriteHeader(s *Stream, md metadata.MD) error {
func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), traceCtx func(context.Context, string) context.Context) {
// With this transport type there will be exactly 1 stream: this HTTP request.
ctx := contextFromRequest(ht.req)
ctx := ht.req.Context()
var cancel context.CancelFunc
if ht.timeoutSet {
ctx, cancel = context.WithTimeout(ctx, ht.timeout)
@@ -315,19 +310,13 @@ func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), trace
ctx, cancel = context.WithCancel(ctx)
}
// requestOver is closed when either the request's context is done
// or the status has been written via WriteStatus.
// requestOver is closed when the status has been written via WriteStatus.
requestOver := make(chan struct{})
// clientGone receives a single value if peer is gone, either
// because the underlying connection is dead or because the
// peer sends an http2 RST_STREAM.
clientGone := ht.rw.(http.CloseNotifier).CloseNotify()
go func() {
select {
case <-requestOver:
case <-ht.closedCh:
case <-clientGone:
case <-ht.req.Context().Done():
}
cancel()
ht.Close()
@@ -349,7 +338,7 @@ func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), trace
Addr: ht.RemoteAddr(),
}
if req.TLS != nil {
pr.AuthInfo = credentials.TLSInfo{State: *req.TLS}
pr.AuthInfo = credentials.TLSInfo{State: *req.TLS, CommonAuthInfo: credentials.CommonAuthInfo{credentials.PrivacyAndIntegrity}}
}
ctx = metadata.NewIncomingContext(ctx, ht.headerMD)
s.ctx = peer.NewContext(ctx, pr)
@@ -363,7 +352,7 @@ func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), trace
ht.stats.HandleRPC(s.ctx, inHeader)
}
s.trReader = &transportReader{
reader: &recvBufferReader{ctx: s.ctx, ctxDone: s.ctx.Done(), recv: s.buf},
reader: &recvBufferReader{ctx: s.ctx, ctxDone: s.ctx.Done(), recv: s.buf, freeBuffer: func(*bytes.Buffer) {}},
windowHandler: func(int) {},
}
@@ -377,7 +366,7 @@ func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), trace
for buf := make([]byte, readSize); ; {
n, err := req.Body.Read(buf)
if n > 0 {
s.buf.put(recvMsg{data: buf[:n:n]})
s.buf.put(recvMsg{buffer: bytes.NewBuffer(buf[:n:n])})
buf = buf[n:]
}
if err != nil {
@@ -407,10 +396,7 @@ func (ht *serverHandlerTransport) HandleStreams(startStream func(*Stream), trace
func (ht *serverHandlerTransport) runStream() {
for {
select {
case fn, ok := <-ht.writes:
if !ok {
return
}
case fn := <-ht.writes:
fn()
case <-ht.closedCh:
return

View File

@@ -19,6 +19,8 @@
package transport
import (
"context"
"fmt"
"io"
"math"
"net"
@@ -28,13 +30,14 @@ import (
"sync/atomic"
"time"
"golang.org/x/net/context"
"golang.org/x/net/http2"
"golang.org/x/net/http2/hpack"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/credentials"
"google.golang.org/grpc/internal"
"google.golang.org/grpc/internal/channelz"
"google.golang.org/grpc/internal/syscall"
"google.golang.org/grpc/keepalive"
"google.golang.org/grpc/metadata"
"google.golang.org/grpc/peer"
@@ -42,8 +45,14 @@ import (
"google.golang.org/grpc/status"
)
// clientConnectionCounter counts the number of connections a client has
// initiated (equal to the number of http2Clients created). Must be accessed
// atomically.
var clientConnectionCounter uint64
// http2Client implements the ClientTransport interface with HTTP2.
type http2Client struct {
lastRead int64 // Keep this field 64-bit aligned. Accessed atomically.
ctx context.Context
cancel context.CancelFunc
ctxDone <-chan struct{} // Cache the ctx.Done() chan.
@@ -60,8 +69,6 @@ type http2Client struct {
// goAway is closed to notify the upper layer (i.e., addrConn.transportMonitor)
// that the server sent GoAway on this transport.
goAway chan struct{}
// awakenKeepalive is used to wake up keepalive when after it has gone dormant.
awakenKeepalive chan struct{}
framer *framer
// controlBuf delivers all the control related tasks (e.g., window
@@ -73,11 +80,8 @@ type http2Client struct {
isSecure bool
creds []credentials.PerRPCCredentials
perRPCCreds []credentials.PerRPCCredentials
// Boolean to keep track of reading activity on transport.
// 1 is true and 0 is false.
activity uint32 // Accessed atomically.
kp keepalive.ClientParameters
keepaliveEnabled bool
@@ -89,10 +93,10 @@ type http2Client struct {
maxSendHeaderListSize *uint32
bdpEst *bdpEstimator
// onSuccess is a callback that client transport calls upon
// onPrefaceReceipt is a callback that client transport calls upon
// receiving server preface to signal that a succefull HTTP2
// connection was established.
onSuccess func()
onPrefaceReceipt func()
maxConcurrentStreams uint32
streamQuota int64
@@ -108,17 +112,34 @@ type http2Client struct {
// goAwayReason records the http2.ErrCode and debug data received with the
// GoAway frame.
goAwayReason GoAwayReason
// A condition variable used to signal when the keepalive goroutine should
// go dormant. The condition for dormancy is based on the number of active
// streams and the `PermitWithoutStream` keepalive client parameter. And
// since the number of active streams is guarded by the above mutex, we use
// the same for this condition variable as well.
kpDormancyCond *sync.Cond
// A boolean to track whether the keepalive goroutine is dormant or not.
// This is checked before attempting to signal the above condition
// variable.
kpDormant bool
// Fields below are for channelz metric collection.
channelzID int64 // channelz unique identification number
czData *channelzData
onGoAway func(GoAwayReason)
onClose func()
bufferPool *bufferPool
connectionID uint64
}
func dial(ctx context.Context, fn func(context.Context, string) (net.Conn, error), addr string) (net.Conn, error) {
if fn != nil {
return fn(ctx, addr)
}
return dialContext(ctx, "tcp", addr)
return (&net.Dialer{}).DialContext(ctx, "tcp", addr)
}
func isTemporary(err error) bool {
@@ -140,7 +161,7 @@ func isTemporary(err error) bool {
// newHTTP2Client constructs a connected ClientTransport to addr based on HTTP2
// and starts to receive messages on it. Non-nil error returns if construction
// fails.
func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts ConnectOptions, onSuccess func()) (_ *http2Client, err error) {
func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts ConnectOptions, onPrefaceReceipt func(), onGoAway func(GoAwayReason), onClose func()) (_ *http2Client, err error) {
scheme := "http"
ctx, cancel := context.WithCancel(ctx)
defer func() {
@@ -162,18 +183,6 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
conn.Close()
}
}(conn)
var (
isSecure bool
authInfo credentials.AuthInfo
)
if creds := opts.TransportCredentials; creds != nil {
scheme = "https"
conn, authInfo, err = creds.ClientHandshake(connectCtx, addr.Authority, conn)
if err != nil {
return nil, connectionErrorf(isTemporary(err), err, "transport: authentication handshake failed: %v", err)
}
isSecure = true
}
kp := opts.KeepaliveParams
// Validate keepalive parameters.
if kp.Time == 0 {
@@ -182,6 +191,36 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
if kp.Timeout == 0 {
kp.Timeout = defaultClientKeepaliveTimeout
}
keepaliveEnabled := false
if kp.Time != infinity {
if err = syscall.SetTCPUserTimeout(conn, kp.Timeout); err != nil {
return nil, connectionErrorf(false, err, "transport: failed to set TCP_USER_TIMEOUT: %v", err)
}
keepaliveEnabled = true
}
var (
isSecure bool
authInfo credentials.AuthInfo
)
transportCreds := opts.TransportCredentials
perRPCCreds := opts.PerRPCCredentials
if b := opts.CredsBundle; b != nil {
if t := b.TransportCredentials(); t != nil {
transportCreds = t
}
if t := b.PerRPCCredentials(); t != nil {
perRPCCreds = append(perRPCCreds, t)
}
}
if transportCreds != nil {
scheme = "https"
conn, authInfo, err = transportCreds.ClientHandshake(connectCtx, addr.Authority, conn)
if err != nil {
return nil, connectionErrorf(isTemporary(err), err, "transport: authentication handshake failed: %v", err)
}
isSecure = true
}
dynamicWindow := true
icwz := int32(initialWindowSize)
if opts.InitialConnWindowSize >= defaultWindowSize {
@@ -207,22 +246,25 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
readerDone: make(chan struct{}),
writerDone: make(chan struct{}),
goAway: make(chan struct{}),
awakenKeepalive: make(chan struct{}, 1),
framer: newFramer(conn, writeBufSize, readBufSize, maxHeaderListSize),
fc: &trInFlow{limit: uint32(icwz)},
scheme: scheme,
activeStreams: make(map[uint32]*Stream),
isSecure: isSecure,
creds: opts.PerRPCCredentials,
perRPCCreds: perRPCCreds,
kp: kp,
statsHandler: opts.StatsHandler,
initialWindowSize: initialWindowSize,
onSuccess: onSuccess,
onPrefaceReceipt: onPrefaceReceipt,
nextID: 1,
maxConcurrentStreams: defaultMaxStreamsClient,
streamQuota: defaultMaxStreamsClient,
streamsQuotaAvailable: make(chan struct{}, 1),
czData: new(channelzData),
onGoAway: onGoAway,
onClose: onClose,
keepaliveEnabled: keepaliveEnabled,
bufferPool: newBufferPool(),
}
t.controlBuf = newControlBuffer(t.ctxDone)
if opts.InitialWindowSize >= defaultWindowSize {
@@ -235,9 +277,6 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
updateFlowControl: t.updateFlowControl,
}
}
// Make sure awakenKeepalive can't be written upon.
// keepalive routine will make it writable, if need be.
t.awakenKeepalive <- struct{}{}
if t.statsHandler != nil {
t.ctx = t.statsHandler.TagConn(t.ctx, &stats.ConnTagInfo{
RemoteAddr: t.remoteAddr,
@@ -249,16 +288,17 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
t.statsHandler.HandleConn(t.ctx, connBegin)
}
if channelz.IsOn() {
t.channelzID = channelz.RegisterNormalSocket(t, opts.ChannelzParentID, "")
t.channelzID = channelz.RegisterNormalSocket(t, opts.ChannelzParentID, fmt.Sprintf("%s -> %s", t.localAddr, t.remoteAddr))
}
if t.kp.Time != infinity {
t.keepaliveEnabled = true
if t.keepaliveEnabled {
t.kpDormancyCond = sync.NewCond(&t.mu)
go t.keepalive()
}
// Start the reader goroutine for incoming message. Each transport has
// a dedicated goroutine which reads HTTP2 frame from network. Then it
// dispatches the frame to the corresponding stream entity.
go t.reader()
// Send connection preface to server.
n, err := t.conn.Write(clientPreface)
if err != nil {
@@ -295,7 +335,12 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
return nil, connectionErrorf(true, err, "transport: failed to write window update: %v", err)
}
}
t.framer.writer.Flush()
t.connectionID = atomic.AddUint64(&clientConnectionCounter, 1)
if err := t.framer.writer.Flush(); err != nil {
return nil, err
}
go func() {
t.loopy = newLoopyWriter(clientSide, t.framer, t.controlBuf, t.bdpEst)
err := t.loopy.run()
@@ -315,6 +360,7 @@ func newHTTP2Client(connectCtx, ctx context.Context, addr TargetInfo, opts Conne
func (t *http2Client) newStream(ctx context.Context, callHdr *CallHdr) *Stream {
// TODO(zhaoq): Handle uint32 overflow of Stream.id.
s := &Stream{
ct: t,
done: make(chan struct{}),
method: callHdr.Method,
sendCompress: callHdr.SendCompress,
@@ -335,6 +381,10 @@ func (t *http2Client) newStream(ctx context.Context, callHdr *CallHdr) *Stream {
ctx: s.ctx,
ctxDone: s.ctx.Done(),
recv: s.buf,
closeStream: func(err error) {
t.CloseStream(s, err)
},
freeBuffer: t.bufferPool.put,
},
windowHandler: func(n int) {
t.updateWindow(s, uint32(n))
@@ -344,23 +394,24 @@ func (t *http2Client) newStream(ctx context.Context, callHdr *CallHdr) *Stream {
}
func (t *http2Client) getPeer() *peer.Peer {
pr := &peer.Peer{
Addr: t.remoteAddr,
return &peer.Peer{
Addr: t.remoteAddr,
AuthInfo: t.authInfo,
}
// Attach Auth info if there is any.
if t.authInfo != nil {
pr.AuthInfo = t.authInfo
}
return pr
}
func (t *http2Client) createHeaderFields(ctx context.Context, callHdr *CallHdr) ([]hpack.HeaderField, error) {
aud := t.createAudience(callHdr)
authData, err := t.getTrAuthData(ctx, aud)
ri := credentials.RequestInfo{
Method: callHdr.Method,
AuthInfo: t.authInfo,
}
ctxWithRequestInfo := internal.NewRequestInfoContext.(func(context.Context, credentials.RequestInfo) context.Context)(ctx, ri)
authData, err := t.getTrAuthData(ctxWithRequestInfo, aud)
if err != nil {
return nil, err
}
callAuthData, err := t.getCallAuthData(ctx, aud, callHdr)
callAuthData, err := t.getCallAuthData(ctxWithRequestInfo, aud, callHdr)
if err != nil {
return nil, err
}
@@ -383,11 +434,12 @@ func (t *http2Client) createHeaderFields(ctx context.Context, callHdr *CallHdr)
if callHdr.SendCompress != "" {
headerFields = append(headerFields, hpack.HeaderField{Name: "grpc-encoding", Value: callHdr.SendCompress})
headerFields = append(headerFields, hpack.HeaderField{Name: "grpc-accept-encoding", Value: callHdr.SendCompress})
}
if dl, ok := ctx.Deadline(); ok {
// Send out timeout regardless its value. The server can detect timeout context by itself.
// TODO(mmukhi): Perhaps this field should be updated when actually writing out to the wire.
timeout := dl.Sub(time.Now())
timeout := time.Until(dl)
headerFields = append(headerFields, hpack.HeaderField{Name: "grpc-timeout", Value: encodeTimeout(timeout)})
}
for k, v := range authData {
@@ -405,6 +457,15 @@ func (t *http2Client) createHeaderFields(ctx context.Context, callHdr *CallHdr)
if md, added, ok := metadata.FromOutgoingContextRaw(ctx); ok {
var k string
for k, vv := range md {
// HTTP doesn't allow you to set pseudoheaders after non pseudoheaders were set.
if isReservedHeader(k) {
continue
}
for _, v := range vv {
headerFields = append(headerFields, hpack.HeaderField{Name: k, Value: encodeMetadataHeader(k, v)})
}
}
for _, vv := range added {
for i, v := range vv {
if i%2 == 0 {
@@ -418,15 +479,6 @@ func (t *http2Client) createHeaderFields(ctx context.Context, callHdr *CallHdr)
headerFields = append(headerFields, hpack.HeaderField{Name: strings.ToLower(k), Value: encodeMetadataHeader(k, v)})
}
}
for k, vv := range md {
// HTTP doesn't allow you to set pseudoheaders after non pseudoheaders were set.
if isReservedHeader(k) {
continue
}
for _, v := range vv {
headerFields = append(headerFields, hpack.HeaderField{Name: k, Value: encodeMetadataHeader(k, v)})
}
}
}
if md, ok := t.md.(*metadata.MD); ok {
for k, vv := range *md {
@@ -443,7 +495,7 @@ func (t *http2Client) createHeaderFields(ctx context.Context, callHdr *CallHdr)
func (t *http2Client) createAudience(callHdr *CallHdr) string {
// Create an audience string only if needed.
if len(t.creds) == 0 && callHdr.Creds == nil {
if len(t.perRPCCreds) == 0 && callHdr.Creds == nil {
return ""
}
// Construct URI required to get auth request metadata.
@@ -457,8 +509,11 @@ func (t *http2Client) createAudience(callHdr *CallHdr) string {
}
func (t *http2Client) getTrAuthData(ctx context.Context, audience string) (map[string]string, error) {
if len(t.perRPCCreds) == 0 {
return nil, nil
}
authData := map[string]string{}
for _, c := range t.creds {
for _, c := range t.perRPCCreds {
data, err := c.GetRequestMetadata(ctx, audience)
if err != nil {
if _, ok := status.FromError(err); ok {
@@ -477,7 +532,7 @@ func (t *http2Client) getTrAuthData(ctx context.Context, audience string) (map[s
}
func (t *http2Client) getCallAuthData(ctx context.Context, audience string, callHdr *CallHdr) (map[string]string, error) {
callAuthData := map[string]string{}
var callAuthData map[string]string
// Check if credentials.PerRPCCredentials were provided via call options.
// Note: if these credentials are provided both via dial options and call
// options, then both sets of credentials will be applied.
@@ -489,6 +544,7 @@ func (t *http2Client) getCallAuthData(ctx context.Context, audience string, call
if err != nil {
return nil, status.Errorf(codes.Internal, "transport: %v", err)
}
callAuthData = make(map[string]string, len(data))
for k, v := range data {
// Capital header names are illegal in HTTP/2
k = strings.ToLower(k)
@@ -517,15 +573,14 @@ func (t *http2Client) NewStream(ctx context.Context, callHdr *CallHdr) (_ *Strea
s.write(recvMsg{err: err})
close(s.done)
// If headerChan isn't closed, then close it.
if atomic.SwapUint32(&s.headerDone, 1) == 0 {
if atomic.CompareAndSwapUint32(&s.headerChanClosed, 0, 1) {
close(s.headerChan)
}
}
hdr := &headerFrame{
hf: headerFields,
endStream: false,
initStream: func(id uint32) (bool, error) {
initStream: func(id uint32) error {
t.mu.Lock()
if state := t.state; state != reachable {
t.mu.Unlock()
@@ -535,29 +590,19 @@ func (t *http2Client) NewStream(ctx context.Context, callHdr *CallHdr) (_ *Strea
err = ErrConnClosing
}
cleanup(err)
return false, err
return err
}
t.activeStreams[id] = s
if channelz.IsOn() {
atomic.AddInt64(&t.czData.streamsStarted, 1)
atomic.StoreInt64(&t.czData.lastStreamCreatedTime, time.Now().UnixNano())
}
var sendPing bool
// If the number of active streams change from 0 to 1, then check if keepalive
// has gone dormant. If so, wake it up.
if len(t.activeStreams) == 1 && t.keepaliveEnabled {
select {
case t.awakenKeepalive <- struct{}{}:
sendPing = true
// Fill the awakenKeepalive channel again as this channel must be
// kept non-writable except at the point that the keepalive()
// goroutine is waiting either to be awaken or shutdown.
t.awakenKeepalive <- struct{}{}
default:
}
// If the keepalive goroutine has gone dormant, wake it up.
if t.kpDormant {
t.kpDormancyCond.Signal()
}
t.mu.Unlock()
return sendPing, nil
return nil
},
onOrphaned: cleanup,
wq: s.wq,
@@ -635,12 +680,14 @@ func (t *http2Client) NewStream(ctx context.Context, callHdr *CallHdr) (_ *Strea
}
}
if t.statsHandler != nil {
header, _, _ := metadata.FromOutgoingContextRaw(ctx)
outHeader := &stats.OutHeader{
Client: true,
FullMethod: callHdr.Method,
RemoteAddr: t.remoteAddr,
LocalAddr: t.localAddr,
Compression: callHdr.SendCompress,
Header: header.Copy(),
}
t.statsHandler.HandleRPC(s.ctx, outHeader)
}
@@ -664,7 +711,9 @@ func (t *http2Client) CloseStream(s *Stream, err error) {
func (t *http2Client) closeStream(s *Stream, err error, rst bool, rstCode http2.ErrCode, st *status.Status, mdata map[string][]string, eosReceived bool) {
// Set stream status to done.
if s.swapState(streamDone) == streamDone {
// If it was already done, return.
// If it was already done, return. If multiple closeStream calls
// happen simultaneously, wait for the first to finish.
<-s.done
return
}
// status and trailers can be updated here without any synchronization because the stream goroutine will
@@ -678,10 +727,8 @@ func (t *http2Client) closeStream(s *Stream, err error, rst bool, rstCode http2.
// This will unblock reads eventually.
s.write(recvMsg{err: err})
}
// This will unblock write.
close(s.done)
// If headerChan isn't closed, then close it.
if atomic.SwapUint32(&s.headerDone, 1) == 0 {
if atomic.CompareAndSwapUint32(&s.headerChanClosed, 0, 1) {
s.noHeaders = true
close(s.headerChan)
}
@@ -715,11 +762,17 @@ func (t *http2Client) closeStream(s *Stream, err error, rst bool, rstCode http2.
return true
}
t.controlBuf.executeAndPut(addBackStreamQuota, cleanup)
// This will unblock write.
close(s.done)
}
// Close kicks off the shutdown process of the transport. This should be called
// only once on a transport. Once it is called, the transport should not be
// accessed any more.
//
// This method blocks until the addrConn that initiated this transport is
// re-connected. This happens because t.onClose() begins reconnect logic at the
// addrConn level and blocks until the addrConn is successfully connected.
func (t *http2Client) Close() error {
t.mu.Lock()
// Make sure we only Close once.
@@ -727,9 +780,17 @@ func (t *http2Client) Close() error {
t.mu.Unlock()
return nil
}
// Call t.onClose before setting the state to closing to prevent the client
// from attempting to create new streams ASAP.
t.onClose()
t.state = closing
streams := t.activeStreams
t.activeStreams = nil
if t.kpDormant {
// If the keepalive goroutine is blocked on this condition variable, we
// should unblock it so that the goroutine eventually exits.
t.kpDormancyCond.Signal()
}
t.mu.Unlock()
t.controlBuf.finish()
t.cancel()
@@ -755,21 +816,21 @@ func (t *http2Client) Close() error {
// stream is closed. If there are no active streams, the transport is closed
// immediately. This does nothing if the transport is already draining or
// closing.
func (t *http2Client) GracefulClose() error {
func (t *http2Client) GracefulClose() {
t.mu.Lock()
// Make sure we move to draining only from active.
if t.state == draining || t.state == closing {
t.mu.Unlock()
return nil
return
}
t.state = draining
active := len(t.activeStreams)
t.mu.Unlock()
if active == 0 {
return t.Close()
t.Close()
return
}
t.controlBuf.put(&incomingGoAway{})
return nil
}
// Write formats the data into HTTP2 data frame(s) and sends it out. The caller
@@ -805,11 +866,11 @@ func (t *http2Client) Write(s *Stream, hdr []byte, data []byte, opts *Options) e
return t.controlBuf.put(df)
}
func (t *http2Client) getStream(f http2.Frame) (*Stream, bool) {
func (t *http2Client) getStream(f http2.Frame) *Stream {
t.mu.Lock()
defer t.mu.Unlock()
s, ok := t.activeStreams[f.Header().StreamID]
return s, ok
s := t.activeStreams[f.Header().StreamID]
t.mu.Unlock()
return s
}
// adjustWindow sends out extra window update over the initial window size
@@ -889,8 +950,8 @@ func (t *http2Client) handleData(f *http2.DataFrame) {
t.controlBuf.put(bdpPing)
}
// Select the right stream to dispatch.
s, ok := t.getStream(f)
if !ok {
s := t.getStream(f)
if s == nil {
return
}
if size > 0 {
@@ -907,9 +968,10 @@ func (t *http2Client) handleData(f *http2.DataFrame) {
// guarantee f.Data() is consumed before the arrival of next frame.
// Can this copy be eliminated?
if len(f.Data()) > 0 {
data := make([]byte, len(f.Data()))
copy(data, f.Data())
s.write(recvMsg{data: data})
buffer := t.bufferPool.get()
buffer.Reset()
buffer.Write(f.Data())
s.write(recvMsg{buffer: buffer})
}
}
// The server has closed the stream without sending trailers. Record that
@@ -920,8 +982,8 @@ func (t *http2Client) handleData(f *http2.DataFrame) {
}
func (t *http2Client) handleRSTStream(f *http2.RSTStreamFrame) {
s, ok := t.getStream(f)
if !ok {
s := t.getStream(f)
if s == nil {
return
}
if f.ErrCode == http2.ErrCodeRefusedStream {
@@ -934,9 +996,9 @@ func (t *http2Client) handleRSTStream(f *http2.RSTStreamFrame) {
statusCode = codes.Unknown
}
if statusCode == codes.Canceled {
// Our deadline was already exceeded, and that was likely the cause of
// this cancelation. Alter the status code accordingly.
if d, ok := s.ctx.Deadline(); ok && d.After(time.Now()) {
if d, ok := s.ctx.Deadline(); ok && !d.After(time.Now()) {
// Our deadline was already exceeded, and that was likely the cause
// of this cancelation. Alter the status code accordingly.
statusCode = codes.DeadlineExceeded
}
}
@@ -1041,8 +1103,12 @@ func (t *http2Client) handleGoAway(f *http2.GoAwayFrame) {
default:
t.setGoAwayReason(f)
close(t.goAway)
t.state = draining
t.controlBuf.put(&incomingGoAway{})
// Notify the clientconn about the GOAWAY before we set the state to
// draining, to allow the client to stop attempting to create streams
// before disallowing new streams on this connection.
t.onGoAway(t.goAwayReason)
t.state = draining
}
// All streams with IDs greater than the GoAwayId
// and smaller than the previous GoAway ID should be killed.
@@ -1094,58 +1160,77 @@ func (t *http2Client) handleWindowUpdate(f *http2.WindowUpdateFrame) {
// operateHeaders takes action on the decoded headers.
func (t *http2Client) operateHeaders(frame *http2.MetaHeadersFrame) {
s, ok := t.getStream(frame)
if !ok {
s := t.getStream(frame)
if s == nil {
return
}
endStream := frame.StreamEnded()
atomic.StoreUint32(&s.bytesReceived, 1)
var state decodeState
if err := state.decodeHeader(frame); err != nil {
t.closeStream(s, err, true, http2.ErrCodeProtocol, status.New(codes.Internal, err.Error()), nil, false)
// Something wrong. Stops reading even when there is remaining.
initialHeader := atomic.LoadUint32(&s.headerChanClosed) == 0
if !initialHeader && !endStream {
// As specified by gRPC over HTTP2, a HEADERS frame (and associated CONTINUATION frames) can only appear at the start or end of a stream. Therefore, second HEADERS frame must have EOS bit set.
st := status.New(codes.Internal, "a HEADERS frame cannot appear in the middle of a stream")
t.closeStream(s, st.Err(), true, http2.ErrCodeProtocol, st, nil, false)
return
}
endStream := frame.StreamEnded()
var isHeader bool
state := &decodeState{}
// Initialize isGRPC value to be !initialHeader, since if a gRPC Response-Headers has already been received, then it means that the peer is speaking gRPC and we are in gRPC mode.
state.data.isGRPC = !initialHeader
if err := state.decodeHeader(frame); err != nil {
t.closeStream(s, err, true, http2.ErrCodeProtocol, status.Convert(err), nil, endStream)
return
}
isHeader := false
defer func() {
if t.statsHandler != nil {
if isHeader {
inHeader := &stats.InHeader{
Client: true,
WireLength: int(frame.Header().Length),
Header: s.header.Copy(),
}
t.statsHandler.HandleRPC(s.ctx, inHeader)
} else {
inTrailer := &stats.InTrailer{
Client: true,
WireLength: int(frame.Header().Length),
Trailer: s.trailer.Copy(),
}
t.statsHandler.HandleRPC(s.ctx, inTrailer)
}
}
}()
// If headers haven't been received yet.
if atomic.SwapUint32(&s.headerDone, 1) == 0 {
// If headerChan hasn't been closed yet
if atomic.CompareAndSwapUint32(&s.headerChanClosed, 0, 1) {
s.headerValid = true
if !endStream {
// Headers frame is not actually a trailers-only frame.
// HEADERS frame block carries a Response-Headers.
isHeader = true
// These values can be set without any synchronization because
// stream goroutine will read it only after seeing a closed
// headerChan which we'll close after setting this.
s.recvCompress = state.encoding
if len(state.mdata) > 0 {
s.header = state.mdata
s.recvCompress = state.data.encoding
if len(state.data.mdata) > 0 {
s.header = state.data.mdata
}
} else {
// HEADERS frame block carries a Trailers-Only.
s.noHeaders = true
}
close(s.headerChan)
}
if !endStream {
return
}
t.closeStream(s, io.EOF, false, http2.ErrCodeNo, state.status(), state.mdata, true)
// if client received END_STREAM from server while stream was still active, send RST_STREAM
rst := s.getState() == streamActive
t.closeStream(s, io.EOF, rst, http2.ErrCodeNo, state.status(), state.data.mdata, true)
}
// reader runs as a separate goroutine in charge of reading data from network
@@ -1159,25 +1244,27 @@ func (t *http2Client) reader() {
// Check the validity of server preface.
frame, err := t.framer.fr.ReadFrame()
if err != nil {
t.Close()
t.Close() // this kicks off resetTransport, so must be last before return
return
}
t.conn.SetReadDeadline(time.Time{}) // reset deadline once we get the settings frame (we didn't time out, yay!)
if t.keepaliveEnabled {
atomic.CompareAndSwapUint32(&t.activity, 0, 1)
atomic.StoreInt64(&t.lastRead, time.Now().UnixNano())
}
sf, ok := frame.(*http2.SettingsFrame)
if !ok {
t.Close()
t.Close() // this kicks off resetTransport, so must be last before return
return
}
t.onSuccess()
t.onPrefaceReceipt()
t.handleSettings(sf, true)
// loop to keep reading incoming messages on this transport.
for {
t.controlBuf.throttle()
frame, err := t.framer.fr.ReadFrame()
if t.keepaliveEnabled {
atomic.CompareAndSwapUint32(&t.activity, 0, 1)
atomic.StoreInt64(&t.lastRead, time.Now().UnixNano())
}
if err != nil {
// Abort an active stream if the http2.Framer returns a
@@ -1221,55 +1308,83 @@ func (t *http2Client) reader() {
}
}
func minTime(a, b time.Duration) time.Duration {
if a < b {
return a
}
return b
}
// keepalive running in a separate goroutune makes sure the connection is alive by sending pings.
func (t *http2Client) keepalive() {
p := &ping{data: [8]byte{}}
// True iff a ping has been sent, and no data has been received since then.
outstandingPing := false
// Amount of time remaining before which we should receive an ACK for the
// last sent ping.
timeoutLeft := time.Duration(0)
// Records the last value of t.lastRead before we go block on the timer.
// This is required to check for read activity since then.
prevNano := time.Now().UnixNano()
timer := time.NewTimer(t.kp.Time)
for {
select {
case <-timer.C:
if atomic.CompareAndSwapUint32(&t.activity, 1, 0) {
timer.Reset(t.kp.Time)
lastRead := atomic.LoadInt64(&t.lastRead)
if lastRead > prevNano {
// There has been read activity since the last time we were here.
outstandingPing = false
// Next timer should fire at kp.Time seconds from lastRead time.
timer.Reset(time.Duration(lastRead) + t.kp.Time - time.Duration(time.Now().UnixNano()))
prevNano = lastRead
continue
}
// Check if keepalive should go dormant.
if outstandingPing && timeoutLeft <= 0 {
t.Close()
return
}
t.mu.Lock()
if t.state == closing {
// If the transport is closing, we should exit from the
// keepalive goroutine here. If not, we could have a race
// between the call to Signal() from Close() and the call to
// Wait() here, whereby the keepalive goroutine ends up
// blocking on the condition variable which will never be
// signalled again.
t.mu.Unlock()
return
}
if len(t.activeStreams) < 1 && !t.kp.PermitWithoutStream {
// Make awakenKeepalive writable.
<-t.awakenKeepalive
t.mu.Unlock()
select {
case <-t.awakenKeepalive:
// If the control gets here a ping has been sent
// need to reset the timer with keepalive.Timeout.
case <-t.ctx.Done():
return
}
} else {
t.mu.Unlock()
// If a ping was sent out previously (because there were active
// streams at that point) which wasn't acked and its timeout
// hadn't fired, but we got here and are about to go dormant,
// we should make sure that we unconditionally send a ping once
// we awaken.
outstandingPing = false
t.kpDormant = true
t.kpDormancyCond.Wait()
}
t.kpDormant = false
t.mu.Unlock()
// We get here either because we were dormant and a new stream was
// created which unblocked the Wait() call, or because the
// keepalive timer expired. In both cases, we need to send a ping.
if !outstandingPing {
if channelz.IsOn() {
atomic.AddInt64(&t.czData.kpCount, 1)
}
// Send ping.
t.controlBuf.put(p)
timeoutLeft = t.kp.Timeout
outstandingPing = true
}
// By the time control gets here a ping has been sent one way or the other.
timer.Reset(t.kp.Timeout)
select {
case <-timer.C:
if atomic.CompareAndSwapUint32(&t.activity, 1, 0) {
timer.Reset(t.kp.Time)
continue
}
t.Close()
return
case <-t.ctx.Done():
if !timer.Stop() {
<-timer.C
}
return
}
// The amount of time to sleep here is the minimum of kp.Time and
// timeoutLeft. This will ensure that we wait only for kp.Time
// before sending out the next ping (for cases where the ping is
// acked).
sleepDuration := minTime(t.kp.Time, timeoutLeft)
timeoutLeft -= sleepDuration
timer.Reset(sleepDuration)
case <-t.ctx.Done():
if !timer.Stop() {
<-timer.C
@@ -1311,6 +1426,8 @@ func (t *http2Client) ChannelzMetric() *channelz.SocketInternalMetric {
return &s
}
func (t *http2Client) RemoteAddr() net.Addr { return t.remoteAddr }
func (t *http2Client) IncrMsgSent() {
atomic.AddInt64(&t.czData.msgSent, 1)
atomic.StoreInt64(&t.czData.lastMsgSentTime, time.Now().UnixNano())

View File

@@ -20,6 +20,7 @@ package transport
import (
"bytes"
"context"
"errors"
"fmt"
"io"
@@ -31,13 +32,14 @@ import (
"time"
"github.com/golang/protobuf/proto"
"golang.org/x/net/context"
"golang.org/x/net/http2"
"golang.org/x/net/http2/hpack"
spb "google.golang.org/genproto/googleapis/rpc/status"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/credentials"
"google.golang.org/grpc/grpclog"
"google.golang.org/grpc/internal"
"google.golang.org/grpc/internal/channelz"
"google.golang.org/grpc/internal/grpcrand"
"google.golang.org/grpc/keepalive"
@@ -55,13 +57,20 @@ var (
// ErrHeaderListSizeLimitViolation indicates that the header list size is larger
// than the limit set by peer.
ErrHeaderListSizeLimitViolation = errors.New("transport: trying to send header list size larger than the limit set by peer")
// statusRawProto is a function to get to the raw status proto wrapped in a
// status.Status without a proto.Clone().
statusRawProto = internal.StatusRawProto.(func(*status.Status) *spb.Status)
)
// serverConnectionCounter counts the number of connections a server has seen
// (equal to the number of http2Servers created). Must be accessed atomically.
var serverConnectionCounter uint64
// http2Server implements the ServerTransport interface with HTTP2.
type http2Server struct {
lastRead int64 // Keep this field 64-bit aligned. Accessed atomically.
ctx context.Context
ctxDone <-chan struct{} // Cache the context.Done() chan
cancel context.CancelFunc
done chan struct{}
conn net.Conn
loopy *loopyWriter
readerDone chan struct{} // sync point to enable testing.
@@ -79,12 +88,8 @@ type http2Server struct {
controlBuf *controlBuffer
fc *trInFlow
stats stats.Handler
// Flag to keep track of reading activity on transport.
// 1 is true and 0 is false.
activity uint32 // Accessed atomically.
// Keepalive and max-age parameters for the server.
kp keepalive.ServerParameters
// Keepalive enforcement policy.
kep keepalive.EnforcementPolicy
// The time instance last ping was received.
@@ -119,6 +124,9 @@ type http2Server struct {
// Fields below are for channelz metric collection.
channelzID int64 // channelz unique identification number
czData *channelzData
bufferPool *bufferPool
connectionID uint64
}
// newHTTP2Server constructs a ServerTransport based on HTTP2. ConnectionError is
@@ -132,7 +140,10 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
}
framer := newFramer(conn, writeBufSize, readBufSize, maxHeaderListSize)
// Send initial settings as connection preface to client.
var isettings []http2.Setting
isettings := []http2.Setting{{
ID: http2.SettingMaxFrameSize,
Val: http2MaxFrameLen,
}}
// TODO(zhaoq): Have a better way to signal "no limit" because 0 is
// permitted in the HTTP2 spec.
maxStreams := config.MaxStreams
@@ -166,6 +177,12 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
Val: *config.MaxHeaderListSize,
})
}
if config.HeaderTableSize != nil {
isettings = append(isettings, http2.Setting{
ID: http2.SettingHeaderTableSize,
Val: *config.HeaderTableSize,
})
}
if err := framer.fr.WriteSettings(isettings...); err != nil {
return nil, connectionErrorf(false, err, "transport: %v", err)
}
@@ -197,11 +214,10 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
if kep.MinTime == 0 {
kep.MinTime = defaultKeepalivePolicyMinTime
}
ctx, cancel := context.WithCancel(context.Background())
done := make(chan struct{})
t := &http2Server{
ctx: ctx,
cancel: cancel,
ctxDone: ctx.Done(),
ctx: context.Background(),
done: done,
conn: conn,
remoteAddr: conn.RemoteAddr(),
localAddr: conn.LocalAddr(),
@@ -220,8 +236,9 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
kep: kep,
initialWindowSize: iwz,
czData: new(channelzData),
bufferPool: newBufferPool(),
}
t.controlBuf = newControlBuffer(t.ctxDone)
t.controlBuf = newControlBuffer(t.done)
if dynamicWindow {
t.bdpEst = &bdpEstimator{
bdp: initialWindowSize,
@@ -237,8 +254,11 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
t.stats.HandleConn(t.ctx, connBegin)
}
if channelz.IsOn() {
t.channelzID = channelz.RegisterNormalSocket(t, config.ChannelzParentID, "")
t.channelzID = channelz.RegisterNormalSocket(t, config.ChannelzParentID, fmt.Sprintf("%s -> %s", t.remoteAddr, t.localAddr))
}
t.connectionID = atomic.AddUint64(&serverConnectionCounter, 1)
t.framer.writer.Flush()
defer func() {
@@ -263,7 +283,7 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
if err != nil {
return nil, connectionErrorf(false, err, "transport: http2Server.HandleStreams failed to read initial settings frame: %v", err)
}
atomic.StoreUint32(&t.activity, 1)
atomic.StoreInt64(&t.lastRead, time.Now().UnixNano())
sf, ok := frame.(*http2.SettingsFrame)
if !ok {
return nil, connectionErrorf(false, nil, "transport: http2Server.HandleStreams saw invalid preface type %T from client", frame)
@@ -286,7 +306,9 @@ func newHTTP2Server(conn net.Conn, config *ServerConfig) (_ ServerTransport, err
// operateHeader takes action on the decoded headers.
func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(*Stream), traceCtx func(context.Context, string) context.Context) (fatal bool) {
streamID := frame.Header().StreamID
state := decodeState{serverSide: true}
state := &decodeState{
serverSide: true,
}
if err := state.decodeHeader(frame); err != nil {
if se, ok := status.FromError(err); ok {
t.controlBuf.put(&cleanupStream{
@@ -305,16 +327,16 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
st: t,
buf: buf,
fc: &inFlow{limit: uint32(t.initialWindowSize)},
recvCompress: state.encoding,
method: state.method,
contentSubtype: state.contentSubtype,
recvCompress: state.data.encoding,
method: state.data.method,
contentSubtype: state.data.contentSubtype,
}
if frame.StreamEnded() {
// s is just created by the caller. No lock needed.
s.state = streamReadDone
}
if state.timeoutSet {
s.ctx, s.cancel = context.WithTimeout(t.ctx, state.timeout)
if state.data.timeoutSet {
s.ctx, s.cancel = context.WithTimeout(t.ctx, state.data.timeout)
} else {
s.ctx, s.cancel = context.WithCancel(t.ctx)
}
@@ -327,19 +349,19 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
}
s.ctx = peer.NewContext(s.ctx, pr)
// Attach the received metadata to the context.
if len(state.mdata) > 0 {
s.ctx = metadata.NewIncomingContext(s.ctx, state.mdata)
if len(state.data.mdata) > 0 {
s.ctx = metadata.NewIncomingContext(s.ctx, state.data.mdata)
}
if state.statsTags != nil {
s.ctx = stats.SetIncomingTags(s.ctx, state.statsTags)
if state.data.statsTags != nil {
s.ctx = stats.SetIncomingTags(s.ctx, state.data.statsTags)
}
if state.statsTrace != nil {
s.ctx = stats.SetIncomingTrace(s.ctx, state.statsTrace)
if state.data.statsTrace != nil {
s.ctx = stats.SetIncomingTrace(s.ctx, state.data.statsTrace)
}
if t.inTapHandle != nil {
var err error
info := &tap.Info{
FullMethodName: state.method,
FullMethodName: state.data.method,
}
s.ctx, err = t.inTapHandle(s.ctx, info)
if err != nil {
@@ -350,12 +372,14 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
rstCode: http2.ErrCodeRefusedStream,
onWrite: func() {},
})
s.cancel()
return false
}
}
t.mu.Lock()
if t.state != reachable {
t.mu.Unlock()
s.cancel()
return false
}
if uint32(len(t.activeStreams)) >= t.maxStreams {
@@ -366,12 +390,14 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
rstCode: http2.ErrCodeRefusedStream,
onWrite: func() {},
})
s.cancel()
return false
}
if streamID%2 != 1 || streamID <= t.maxStreamID {
t.mu.Unlock()
// illegal gRPC stream id.
errorf("transport: http2Server.HandleStreams received an illegal stream id: %v", streamID)
s.cancel()
return true
}
t.maxStreamID = streamID
@@ -396,6 +422,7 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
LocalAddr: t.localAddr,
Compression: s.recvCompress,
WireLength: int(frame.Header().Length),
Header: metadata.MD(state.data.mdata).Copy(),
}
t.stats.HandleRPC(s.ctx, inHeader)
}
@@ -403,9 +430,10 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
s.wq = newWriteQuota(defaultWriteQuota, s.ctxDone)
s.trReader = &transportReader{
reader: &recvBufferReader{
ctx: s.ctx,
ctxDone: s.ctxDone,
recv: s.buf,
ctx: s.ctx,
ctxDone: s.ctxDone,
recv: s.buf,
freeBuffer: t.bufferPool.put,
},
windowHandler: func(n int) {
t.updateWindow(s, uint32(n))
@@ -426,8 +454,9 @@ func (t *http2Server) operateHeaders(frame *http2.MetaHeadersFrame, handle func(
func (t *http2Server) HandleStreams(handle func(*Stream), traceCtx func(context.Context, string) context.Context) {
defer close(t.readerDone)
for {
t.controlBuf.throttle()
frame, err := t.framer.fr.ReadFrame()
atomic.StoreUint32(&t.activity, 1)
atomic.StoreInt64(&t.lastRead, time.Now().UnixNano())
if err != nil {
if se, ok := err.(http2.StreamError); ok {
warningf("transport: http2Server.HandleStreams encountered http2.StreamError: %v", se)
@@ -435,7 +464,7 @@ func (t *http2Server) HandleStreams(handle func(*Stream), traceCtx func(context.
s := t.activeStreams[se.StreamID]
t.mu.Unlock()
if s != nil {
t.closeStream(s, true, se.Code, nil, false)
t.closeStream(s, true, se.Code, false)
} else {
t.controlBuf.put(&cleanupStream{
streamID: se.StreamID,
@@ -577,7 +606,7 @@ func (t *http2Server) handleData(f *http2.DataFrame) {
}
if size > 0 {
if err := s.fc.onData(size); err != nil {
t.closeStream(s, true, http2.ErrCodeFlowControl, nil, false)
t.closeStream(s, true, http2.ErrCodeFlowControl, false)
return
}
if f.Header().Flags.Has(http2.FlagDataPadded) {
@@ -589,9 +618,10 @@ func (t *http2Server) handleData(f *http2.DataFrame) {
// guarantee f.Data() is consumed before the arrival of next frame.
// Can this copy be eliminated?
if len(f.Data()) > 0 {
data := make([]byte, len(f.Data()))
copy(data, f.Data())
s.write(recvMsg{data: data})
buffer := t.bufferPool.get()
buffer.Reset()
buffer.Write(f.Data())
s.write(recvMsg{buffer: buffer})
}
}
if f.Header().Flags.Has(http2.FlagDataEndStream) {
@@ -602,11 +632,18 @@ func (t *http2Server) handleData(f *http2.DataFrame) {
}
func (t *http2Server) handleRSTStream(f *http2.RSTStreamFrame) {
s, ok := t.getStream(f)
if !ok {
// If the stream is not deleted from the transport's active streams map, then do a regular close stream.
if s, ok := t.getStream(f); ok {
t.closeStream(s, false, 0, false)
return
}
t.closeStream(s, false, 0, nil, false)
// If the stream is already deleted from the active streams map, then put a cleanupStream item into controlbuf to delete the stream from loopy writer's established streams map.
t.controlBuf.put(&cleanupStream{
streamID: f.Header().StreamID,
rst: false,
rstCode: 0,
onWrite: func() {},
})
}
func (t *http2Server) handleSettings(f *http2.SettingsFrame) {
@@ -727,7 +764,7 @@ func (t *http2Server) checkForHeaderListSize(it interface{}) bool {
return true
}
// WriteHeader sends the header metedata md back to the client.
// WriteHeader sends the header metadata md back to the client.
func (t *http2Server) WriteHeader(s *Stream, md metadata.MD) error {
if s.updateHeaderSent() || s.getState() == streamDone {
return ErrIllegalHeaderWrite
@@ -748,6 +785,10 @@ func (t *http2Server) WriteHeader(s *Stream, md metadata.MD) error {
return nil
}
func (t *http2Server) setResetPingStrikes() {
atomic.StoreUint32(&t.resetPingStrikes, 1)
}
func (t *http2Server) writeHeaderLocked(s *Stream) error {
// TODO(mmukhi): Benchmark if the performance gets better if count the metadata and other header fields
// first and create a slice of that exact size.
@@ -762,21 +803,21 @@ func (t *http2Server) writeHeaderLocked(s *Stream) error {
streamID: s.id,
hf: headerFields,
endStream: false,
onWrite: func() {
atomic.StoreUint32(&t.resetPingStrikes, 1)
},
onWrite: t.setResetPingStrikes,
})
if !success {
if err != nil {
return err
}
t.closeStream(s, true, http2.ErrCodeInternal, nil, false)
t.closeStream(s, true, http2.ErrCodeInternal, false)
return ErrHeaderListSizeLimitViolation
}
if t.stats != nil {
// Note: WireLength is not set in outHeader.
// TODO(mmukhi): Revisit this later, if needed.
outHeader := &stats.OutHeader{}
outHeader := &stats.OutHeader{
Header: s.header.Copy(),
}
t.stats.HandleRPC(s.Context(), outHeader)
}
return nil
@@ -808,7 +849,7 @@ func (t *http2Server) WriteStatus(s *Stream, st *status.Status) error {
headerFields = append(headerFields, hpack.HeaderField{Name: "grpc-status", Value: strconv.Itoa(int(st.Code()))})
headerFields = append(headerFields, hpack.HeaderField{Name: "grpc-message", Value: encodeGrpcMessage(st.Message())})
if p := st.Proto(); p != nil && len(p.Details) > 0 {
if p := statusRawProto(st); p != nil && len(p.Details) > 0 {
stBytes, err := proto.Marshal(p)
if err != nil {
// TODO: return error instead, when callers are able to handle it.
@@ -824,9 +865,7 @@ func (t *http2Server) WriteStatus(s *Stream, st *status.Status) error {
streamID: s.id,
hf: headerFields,
endStream: true,
onWrite: func() {
atomic.StoreUint32(&t.resetPingStrikes, 1)
},
onWrite: t.setResetPingStrikes,
}
s.hdrMu.Unlock()
success, err := t.controlBuf.execute(t.checkForHeaderListSize, trailingHeader)
@@ -834,12 +873,16 @@ func (t *http2Server) WriteStatus(s *Stream, st *status.Status) error {
if err != nil {
return err
}
t.closeStream(s, true, http2.ErrCodeInternal, nil, false)
t.closeStream(s, true, http2.ErrCodeInternal, false)
return ErrHeaderListSizeLimitViolation
}
t.closeStream(s, false, 0, trailingHeader, true)
// Send a RST_STREAM after the trailers if the client has not already half-closed.
rst := s.getState() == streamActive
t.finishStream(s, rst, http2.ErrCodeNo, trailingHeader, true)
if t.stats != nil {
t.stats.HandleRPC(s.Context(), &stats.OutTrailer{})
t.stats.HandleRPC(s.Context(), &stats.OutTrailer{
Trailer: s.trailer.Copy(),
})
}
return nil
}
@@ -849,6 +892,9 @@ func (t *http2Server) WriteStatus(s *Stream, st *status.Status) error {
func (t *http2Server) Write(s *Stream, hdr []byte, data []byte, opts *Options) error {
if !s.isHeaderSent() { // Headers haven't been written yet.
if err := t.WriteHeader(s, nil); err != nil {
if _, ok := err.(ConnectionError); ok {
return err
}
// TODO(mmukhi, dfawley): Make sure this is the right code to return.
return status.Errorf(codes.Internal, "transport: %v", err)
}
@@ -858,7 +904,7 @@ func (t *http2Server) Write(s *Stream, hdr []byte, data []byte, opts *Options) e
// TODO(mmukhi, dfawley): Should the server write also return io.EOF?
s.cancel()
select {
case <-t.ctx.Done():
case <-t.done:
return ErrConnClosing
default:
}
@@ -873,16 +919,14 @@ func (t *http2Server) Write(s *Stream, hdr []byte, data []byte, opts *Options) e
hdr = append(hdr, data[:emptyLen]...)
data = data[emptyLen:]
df := &dataFrame{
streamID: s.id,
h: hdr,
d: data,
onEachWrite: func() {
atomic.StoreUint32(&t.resetPingStrikes, 1)
},
streamID: s.id,
h: hdr,
d: data,
onEachWrite: t.setResetPingStrikes,
}
if err := s.wq.get(int32(len(hdr) + len(data))); err != nil {
select {
case <-t.ctx.Done():
case <-t.done:
return ErrConnClosing
default:
}
@@ -899,32 +943,35 @@ func (t *http2Server) Write(s *Stream, hdr []byte, data []byte, opts *Options) e
// after an additional duration of keepalive.Timeout.
func (t *http2Server) keepalive() {
p := &ping{}
var pingSent bool
maxIdle := time.NewTimer(t.kp.MaxConnectionIdle)
maxAge := time.NewTimer(t.kp.MaxConnectionAge)
keepalive := time.NewTimer(t.kp.Time)
// NOTE: All exit paths of this function should reset their
// respective timers. A failure to do so will cause the
// following clean-up to deadlock and eventually leak.
// True iff a ping has been sent, and no data has been received since then.
outstandingPing := false
// Amount of time remaining before which we should receive an ACK for the
// last sent ping.
kpTimeoutLeft := time.Duration(0)
// Records the last value of t.lastRead before we go block on the timer.
// This is required to check for read activity since then.
prevNano := time.Now().UnixNano()
// Initialize the different timers to their default values.
idleTimer := time.NewTimer(t.kp.MaxConnectionIdle)
ageTimer := time.NewTimer(t.kp.MaxConnectionAge)
kpTimer := time.NewTimer(t.kp.Time)
defer func() {
if !maxIdle.Stop() {
<-maxIdle.C
}
if !maxAge.Stop() {
<-maxAge.C
}
if !keepalive.Stop() {
<-keepalive.C
}
// We need to drain the underlying channel in these timers after a call
// to Stop(), only if we are interested in resetting them. Clearly we
// are not interested in resetting them here.
idleTimer.Stop()
ageTimer.Stop()
kpTimer.Stop()
}()
for {
select {
case <-maxIdle.C:
case <-idleTimer.C:
t.mu.Lock()
idle := t.idle
if idle.IsZero() { // The connection is non-idle.
t.mu.Unlock()
maxIdle.Reset(t.kp.MaxConnectionIdle)
idleTimer.Reset(t.kp.MaxConnectionIdle)
continue
}
val := t.kp.MaxConnectionIdle - time.Since(idle)
@@ -933,42 +980,52 @@ func (t *http2Server) keepalive() {
// The connection has been idle for a duration of keepalive.MaxConnectionIdle or more.
// Gracefully close the connection.
t.drain(http2.ErrCodeNo, []byte{})
// Resetting the timer so that the clean-up doesn't deadlock.
maxIdle.Reset(infinity)
return
}
maxIdle.Reset(val)
case <-maxAge.C:
idleTimer.Reset(val)
case <-ageTimer.C:
t.drain(http2.ErrCodeNo, []byte{})
maxAge.Reset(t.kp.MaxConnectionAgeGrace)
ageTimer.Reset(t.kp.MaxConnectionAgeGrace)
select {
case <-maxAge.C:
case <-ageTimer.C:
// Close the connection after grace period.
infof("transport: closing server transport due to maximum connection age.")
t.Close()
// Resetting the timer so that the clean-up doesn't deadlock.
maxAge.Reset(infinity)
case <-t.ctx.Done():
case <-t.done:
}
return
case <-keepalive.C:
if atomic.CompareAndSwapUint32(&t.activity, 1, 0) {
pingSent = false
keepalive.Reset(t.kp.Time)
case <-kpTimer.C:
lastRead := atomic.LoadInt64(&t.lastRead)
if lastRead > prevNano {
// There has been read activity since the last time we were
// here. Setup the timer to fire at kp.Time seconds from
// lastRead time and continue.
outstandingPing = false
kpTimer.Reset(time.Duration(lastRead) + t.kp.Time - time.Duration(time.Now().UnixNano()))
prevNano = lastRead
continue
}
if pingSent {
if outstandingPing && kpTimeoutLeft <= 0 {
infof("transport: closing server transport due to idleness.")
t.Close()
// Resetting the timer so that the clean-up doesn't deadlock.
keepalive.Reset(infinity)
return
}
pingSent = true
if channelz.IsOn() {
atomic.AddInt64(&t.czData.kpCount, 1)
if !outstandingPing {
if channelz.IsOn() {
atomic.AddInt64(&t.czData.kpCount, 1)
}
t.controlBuf.put(p)
kpTimeoutLeft = t.kp.Timeout
outstandingPing = true
}
t.controlBuf.put(p)
keepalive.Reset(t.kp.Timeout)
case <-t.ctx.Done():
// The amount of time to sleep here is the minimum of kp.Time and
// timeoutLeft. This will ensure that we wait only for kp.Time
// before sending out the next ping (for cases where the ping is
// acked).
sleepDuration := minTime(t.kp.Time, kpTimeoutLeft)
kpTimeoutLeft -= sleepDuration
kpTimer.Reset(sleepDuration)
case <-t.done:
return
}
}
@@ -988,7 +1045,7 @@ func (t *http2Server) Close() error {
t.activeStreams = nil
t.mu.Unlock()
t.controlBuf.finish()
t.cancel()
close(t.done)
err := t.conn.Close()
if channelz.IsOn() {
channelz.RemoveEntry(t.channelzID)
@@ -1004,45 +1061,61 @@ func (t *http2Server) Close() error {
return err
}
// closeStream clears the footprint of a stream when the stream is not needed
// any more.
func (t *http2Server) closeStream(s *Stream, rst bool, rstCode http2.ErrCode, hdr *headerFrame, eosReceived bool) {
if s.swapState(streamDone) == streamDone {
// If the stream was already done, return.
return
}
// deleteStream deletes the stream s from transport's active streams.
func (t *http2Server) deleteStream(s *Stream, eosReceived bool) {
// In case stream sending and receiving are invoked in separate
// goroutines (e.g., bi-directional streaming), cancel needs to be
// called to interrupt the potential blocking on other goroutines.
s.cancel()
cleanup := &cleanupStream{
t.mu.Lock()
if _, ok := t.activeStreams[s.id]; ok {
delete(t.activeStreams, s.id)
if len(t.activeStreams) == 0 {
t.idle = time.Now()
}
}
t.mu.Unlock()
if channelz.IsOn() {
if eosReceived {
atomic.AddInt64(&t.czData.streamsSucceeded, 1)
} else {
atomic.AddInt64(&t.czData.streamsFailed, 1)
}
}
}
// finishStream closes the stream and puts the trailing headerFrame into controlbuf.
func (t *http2Server) finishStream(s *Stream, rst bool, rstCode http2.ErrCode, hdr *headerFrame, eosReceived bool) {
oldState := s.swapState(streamDone)
if oldState == streamDone {
// If the stream was already done, return.
return
}
hdr.cleanup = &cleanupStream{
streamID: s.id,
rst: rst,
rstCode: rstCode,
onWrite: func() {
t.mu.Lock()
if t.activeStreams != nil {
delete(t.activeStreams, s.id)
if len(t.activeStreams) == 0 {
t.idle = time.Now()
}
}
t.mu.Unlock()
if channelz.IsOn() {
if eosReceived {
atomic.AddInt64(&t.czData.streamsSucceeded, 1)
} else {
atomic.AddInt64(&t.czData.streamsFailed, 1)
}
}
t.deleteStream(s, eosReceived)
},
}
if hdr != nil {
hdr.cleanup = cleanup
t.controlBuf.put(hdr)
} else {
t.controlBuf.put(cleanup)
}
t.controlBuf.put(hdr)
}
// closeStream clears the footprint of a stream when the stream is not needed any more.
func (t *http2Server) closeStream(s *Stream, rst bool, rstCode http2.ErrCode, eosReceived bool) {
s.swapState(streamDone)
t.deleteStream(s, eosReceived)
t.controlBuf.put(&cleanupStream{
streamID: s.id,
rst: rst,
rstCode: rstCode,
onWrite: func() {},
})
}
func (t *http2Server) RemoteAddr() net.Addr {
@@ -1112,7 +1185,7 @@ func (t *http2Server) outgoingGoAwayHandler(g *goAway) (bool, error) {
select {
case <-t.drainChan:
case <-timer.C:
case <-t.ctx.Done():
case <-t.done:
return
}
t.controlBuf.put(&goAway{code: g.code, debugData: g.debugData})
@@ -1155,14 +1228,14 @@ func (t *http2Server) IncrMsgRecv() {
}
func (t *http2Server) getOutFlowWindow() int64 {
resp := make(chan uint32)
resp := make(chan uint32, 1)
timer := time.NewTimer(time.Second)
defer timer.Stop()
t.controlBuf.put(&outFlowControlSizeRequest{resp})
select {
case sz := <-resp:
return int64(sz)
case <-t.ctxDone:
case <-t.done:
return -1
case <-timer.C:
return -2

View File

@@ -24,6 +24,7 @@ import (
"encoding/base64"
"fmt"
"io"
"math"
"net"
"net/http"
"strconv"
@@ -77,7 +78,8 @@ var (
codes.ResourceExhausted: http2.ErrCodeEnhanceYourCalm,
codes.PermissionDenied: http2.ErrCodeInadequateSecurity,
}
httpStatusConvTab = map[int]codes.Code{
// HTTPStatusConvTab is the HTTP status code to gRPC error code conversion table.
HTTPStatusConvTab = map[int]codes.Code{
// 400 Bad Request - INTERNAL.
http.StatusBadRequest: codes.Internal,
// 401 Unauthorized - UNAUTHENTICATED.
@@ -97,9 +99,7 @@ var (
}
)
// Records the states during HPACK decoding. Must be reset once the
// decoding of the entire headers are finished.
type decodeState struct {
type parsedHeaderData struct {
encoding string
// statusGen caches the stream status received from the trailer the server
// sent. Client side only. Do not access directly. After all trailers are
@@ -119,8 +119,30 @@ type decodeState struct {
statsTags []byte
statsTrace []byte
contentSubtype string
// isGRPC field indicates whether the peer is speaking gRPC (otherwise HTTP).
//
// We are in gRPC mode (peer speaking gRPC) if:
// * We are client side and have already received a HEADER frame that indicates gRPC peer.
// * The header contains valid a content-type, i.e. a string starts with "application/grpc"
// And we should handle error specific to gRPC.
//
// Otherwise (i.e. a content-type string starts without "application/grpc", or does not exist), we
// are in HTTP fallback mode, and should handle error specific to HTTP.
isGRPC bool
grpcErr error
httpErr error
contentTypeErr string
}
// decodeState configures decoding criteria and records the decoded data.
type decodeState struct {
// whether decoding on server side or not
serverSide bool
// Records the states during HPACK decoding. It will be filled with info parsed from HTTP HEADERS
// frame once decodeHeader function has been invoked and returned.
data parsedHeaderData
}
// isReservedHeader checks whether hdr belongs to HTTP2 headers
@@ -201,11 +223,11 @@ func contentType(contentSubtype string) string {
}
func (d *decodeState) status() *status.Status {
if d.statusGen == nil {
if d.data.statusGen == nil {
// No status-details were provided; generate status using code/msg.
d.statusGen = status.New(codes.Code(int32(*(d.rawStatusCode))), d.rawStatusMsg)
d.data.statusGen = status.New(codes.Code(int32(*(d.data.rawStatusCode))), d.data.rawStatusMsg)
}
return d.statusGen
return d.data.statusGen
}
const binHdrSuffix = "-bin"
@@ -243,113 +265,146 @@ func (d *decodeState) decodeHeader(frame *http2.MetaHeadersFrame) error {
if frame.Truncated {
return status.Error(codes.Internal, "peer header list size exceeded limit")
}
for _, hf := range frame.Fields {
if err := d.processHeaderField(hf); err != nil {
return err
d.processHeaderField(hf)
}
if d.data.isGRPC {
if d.data.grpcErr != nil {
return d.data.grpcErr
}
if d.serverSide {
return nil
}
if d.data.rawStatusCode == nil && d.data.statusGen == nil {
// gRPC status doesn't exist.
// Set rawStatusCode to be unknown and return nil error.
// So that, if the stream has ended this Unknown status
// will be propagated to the user.
// Otherwise, it will be ignored. In which case, status from
// a later trailer, that has StreamEnded flag set, is propagated.
code := int(codes.Unknown)
d.data.rawStatusCode = &code
}
}
if d.serverSide {
return nil
}
// If grpc status exists, no need to check further.
if d.rawStatusCode != nil || d.statusGen != nil {
return nil
// HTTP fallback mode
if d.data.httpErr != nil {
return d.data.httpErr
}
// If grpc status doesn't exist and http status doesn't exist,
// then it's a malformed header.
if d.httpStatus == nil {
return status.Error(codes.Internal, "malformed header: doesn't contain status(gRPC or HTTP)")
}
var (
code = codes.Internal // when header does not include HTTP status, return INTERNAL
ok bool
)
if *(d.httpStatus) != http.StatusOK {
code, ok := httpStatusConvTab[*(d.httpStatus)]
if d.data.httpStatus != nil {
code, ok = HTTPStatusConvTab[*(d.data.httpStatus)]
if !ok {
code = codes.Unknown
}
return status.Error(code, http.StatusText(*(d.httpStatus)))
}
// gRPC status doesn't exist and http status is OK.
// Set rawStatusCode to be unknown and return nil error.
// So that, if the stream has ended this Unknown status
// will be propagated to the user.
// Otherwise, it will be ignored. In which case, status from
// a later trailer, that has StreamEnded flag set, is propagated.
code := int(codes.Unknown)
d.rawStatusCode = &code
return nil
return status.Error(code, d.constructHTTPErrMsg())
}
// constructErrMsg constructs error message to be returned in HTTP fallback mode.
// Format: HTTP status code and its corresponding message + content-type error message.
func (d *decodeState) constructHTTPErrMsg() string {
var errMsgs []string
if d.data.httpStatus == nil {
errMsgs = append(errMsgs, "malformed header: missing HTTP status")
} else {
errMsgs = append(errMsgs, fmt.Sprintf("%s: HTTP status code %d", http.StatusText(*(d.data.httpStatus)), *d.data.httpStatus))
}
if d.data.contentTypeErr == "" {
errMsgs = append(errMsgs, "transport: missing content-type field")
} else {
errMsgs = append(errMsgs, d.data.contentTypeErr)
}
return strings.Join(errMsgs, "; ")
}
func (d *decodeState) addMetadata(k, v string) {
if d.mdata == nil {
d.mdata = make(map[string][]string)
if d.data.mdata == nil {
d.data.mdata = make(map[string][]string)
}
d.mdata[k] = append(d.mdata[k], v)
d.data.mdata[k] = append(d.data.mdata[k], v)
}
func (d *decodeState) processHeaderField(f hpack.HeaderField) error {
func (d *decodeState) processHeaderField(f hpack.HeaderField) {
switch f.Name {
case "content-type":
contentSubtype, validContentType := contentSubtype(f.Value)
if !validContentType {
return status.Errorf(codes.Internal, "transport: received the unexpected content-type %q", f.Value)
d.data.contentTypeErr = fmt.Sprintf("transport: received the unexpected content-type %q", f.Value)
return
}
d.contentSubtype = contentSubtype
d.data.contentSubtype = contentSubtype
// TODO: do we want to propagate the whole content-type in the metadata,
// or come up with a way to just propagate the content-subtype if it was set?
// ie {"content-type": "application/grpc+proto"} or {"content-subtype": "proto"}
// in the metadata?
d.addMetadata(f.Name, f.Value)
d.data.isGRPC = true
case "grpc-encoding":
d.encoding = f.Value
d.data.encoding = f.Value
case "grpc-status":
code, err := strconv.Atoi(f.Value)
if err != nil {
return status.Errorf(codes.Internal, "transport: malformed grpc-status: %v", err)
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed grpc-status: %v", err)
return
}
d.rawStatusCode = &code
d.data.rawStatusCode = &code
case "grpc-message":
d.rawStatusMsg = decodeGrpcMessage(f.Value)
d.data.rawStatusMsg = decodeGrpcMessage(f.Value)
case "grpc-status-details-bin":
v, err := decodeBinHeader(f.Value)
if err != nil {
return status.Errorf(codes.Internal, "transport: malformed grpc-status-details-bin: %v", err)
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed grpc-status-details-bin: %v", err)
return
}
s := &spb.Status{}
if err := proto.Unmarshal(v, s); err != nil {
return status.Errorf(codes.Internal, "transport: malformed grpc-status-details-bin: %v", err)
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed grpc-status-details-bin: %v", err)
return
}
d.statusGen = status.FromProto(s)
d.data.statusGen = status.FromProto(s)
case "grpc-timeout":
d.timeoutSet = true
d.data.timeoutSet = true
var err error
if d.timeout, err = decodeTimeout(f.Value); err != nil {
return status.Errorf(codes.Internal, "transport: malformed time-out: %v", err)
if d.data.timeout, err = decodeTimeout(f.Value); err != nil {
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed time-out: %v", err)
}
case ":path":
d.method = f.Value
d.data.method = f.Value
case ":status":
code, err := strconv.Atoi(f.Value)
if err != nil {
return status.Errorf(codes.Internal, "transport: malformed http-status: %v", err)
d.data.httpErr = status.Errorf(codes.Internal, "transport: malformed http-status: %v", err)
return
}
d.httpStatus = &code
d.data.httpStatus = &code
case "grpc-tags-bin":
v, err := decodeBinHeader(f.Value)
if err != nil {
return status.Errorf(codes.Internal, "transport: malformed grpc-tags-bin: %v", err)
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed grpc-tags-bin: %v", err)
return
}
d.statsTags = v
d.data.statsTags = v
d.addMetadata(f.Name, string(v))
case "grpc-trace-bin":
v, err := decodeBinHeader(f.Value)
if err != nil {
return status.Errorf(codes.Internal, "transport: malformed grpc-trace-bin: %v", err)
d.data.grpcErr = status.Errorf(codes.Internal, "transport: malformed grpc-trace-bin: %v", err)
return
}
d.statsTrace = v
d.data.statsTrace = v
d.addMetadata(f.Name, string(v))
default:
if isReservedHeader(f.Name) && !isWhitelistedHeader(f.Name) {
@@ -358,11 +413,10 @@ func (d *decodeState) processHeaderField(f hpack.HeaderField) error {
v, err := decodeMetadataHeader(f.Name, f.Value)
if err != nil {
errorf("Failed to decode metadata header (%q, %q): %v", f.Name, f.Value, err)
return nil
return
}
d.addMetadata(f.Name, v)
}
return nil
}
type timeoutUnit uint8
@@ -435,6 +489,10 @@ func decodeTimeout(s string) (time.Duration, error) {
if size < 2 {
return 0, fmt.Errorf("transport: timeout string is too short: %q", s)
}
if size > 9 {
// Spec allows for 8 digits plus the unit.
return 0, fmt.Errorf("transport: timeout string is too long: %q", s)
}
unit := timeoutUnit(s[size-1])
d, ok := timeoutUnitToDuration(unit)
if !ok {
@@ -444,6 +502,11 @@ func decodeTimeout(s string) (time.Duration, error) {
if err != nil {
return 0, err
}
const maxHours = math.MaxInt64 / int64(time.Hour)
if d == time.Hour && t > maxHours {
// This timeout would overflow math.MaxInt64; clamp it.
return time.Duration(math.MaxInt64), nil
}
return d * time.Duration(t), nil
}
@@ -604,6 +667,7 @@ func newFramer(conn net.Conn, writeBufferSize, readBufferSize int, maxHeaderList
writer: w,
fr: http2.NewFramer(w, r),
}
f.fr.SetMaxReadFrameSize(http2MaxFrameLen)
// Opt-in to Frame reuse API on framer to reduce garbage.
// Frames aren't safe to read from after a subsequent call to ReadFrame.
f.fr.SetReuseFrames()

View File

@@ -22,6 +22,8 @@
package transport
import (
"bytes"
"context"
"errors"
"fmt"
"io"
@@ -29,7 +31,6 @@ import (
"sync"
"sync/atomic"
"golang.org/x/net/context"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/credentials"
"google.golang.org/grpc/keepalive"
@@ -39,10 +40,32 @@ import (
"google.golang.org/grpc/tap"
)
type bufferPool struct {
pool sync.Pool
}
func newBufferPool() *bufferPool {
return &bufferPool{
pool: sync.Pool{
New: func() interface{} {
return new(bytes.Buffer)
},
},
}
}
func (p *bufferPool) get() *bytes.Buffer {
return p.pool.Get().(*bytes.Buffer)
}
func (p *bufferPool) put(b *bytes.Buffer) {
p.pool.Put(b)
}
// recvMsg represents the received msg from the transport. All transport
// protocol specific info has been removed.
type recvMsg struct {
data []byte
buffer *bytes.Buffer
// nil: received some data
// io.EOF: stream is completed. data is nil.
// other non-nil error: transport failure. data is nil.
@@ -50,10 +73,11 @@ type recvMsg struct {
}
// recvBuffer is an unbounded channel of recvMsg structs.
// Note recvBuffer differs from controlBuffer only in that recvBuffer
// holds a channel of only recvMsg structs instead of objects implementing "item" interface.
// recvBuffer is written to much more often than
// controlBuffer and using strict recvMsg structs helps avoid allocation in "recvBuffer.put"
//
// Note: recvBuffer differs from buffer.Unbounded only in the fact that it
// holds a channel of recvMsg structs instead of objects implementing "item"
// interface. recvBuffer is written to much more often and using strict recvMsg
// structs helps avoid allocation in "recvBuffer.put"
type recvBuffer struct {
c chan recvMsg
mu sync.Mutex
@@ -110,15 +134,16 @@ func (b *recvBuffer) get() <-chan recvMsg {
return b.c
}
//
// recvBufferReader implements io.Reader interface to read the data from
// recvBuffer.
type recvBufferReader struct {
ctx context.Context
ctxDone <-chan struct{} // cache of ctx.Done() (for performance).
recv *recvBuffer
last []byte // Stores the remaining data in the previous calls.
err error
closeStream func(error) // Closes the client transport stream with the given error and nil trailer metadata.
ctx context.Context
ctxDone <-chan struct{} // cache of ctx.Done() (for performance).
recv *recvBuffer
last *bytes.Buffer // Stores the remaining data in the previous calls.
err error
freeBuffer func(*bytes.Buffer)
}
// Read reads the next len(p) bytes from last. If last is drained, it tries to
@@ -128,31 +153,74 @@ func (r *recvBufferReader) Read(p []byte) (n int, err error) {
if r.err != nil {
return 0, r.err
}
n, r.err = r.read(p)
if r.last != nil {
// Read remaining data left in last call.
copied, _ := r.last.Read(p)
if r.last.Len() == 0 {
r.freeBuffer(r.last)
r.last = nil
}
return copied, nil
}
if r.closeStream != nil {
n, r.err = r.readClient(p)
} else {
n, r.err = r.read(p)
}
return n, r.err
}
func (r *recvBufferReader) read(p []byte) (n int, err error) {
if r.last != nil && len(r.last) > 0 {
// Read remaining data left in last call.
copied := copy(p, r.last)
r.last = r.last[copied:]
return copied, nil
}
select {
case <-r.ctxDone:
return 0, ContextErr(r.ctx.Err())
case m := <-r.recv.get():
r.recv.load()
if m.err != nil {
return 0, m.err
}
copied := copy(p, m.data)
r.last = m.data[copied:]
return copied, nil
return r.readAdditional(m, p)
}
}
func (r *recvBufferReader) readClient(p []byte) (n int, err error) {
// If the context is canceled, then closes the stream with nil metadata.
// closeStream writes its error parameter to r.recv as a recvMsg.
// r.readAdditional acts on that message and returns the necessary error.
select {
case <-r.ctxDone:
// Note that this adds the ctx error to the end of recv buffer, and
// reads from the head. This will delay the error until recv buffer is
// empty, thus will delay ctx cancellation in Recv().
//
// It's done this way to fix a race between ctx cancel and trailer. The
// race was, stream.Recv() may return ctx error if ctxDone wins the
// race, but stream.Trailer() may return a non-nil md because the stream
// was not marked as done when trailer is received. This closeStream
// call will mark stream as done, thus fix the race.
//
// TODO: delaying ctx error seems like a unnecessary side effect. What
// we really want is to mark the stream as done, and return ctx error
// faster.
r.closeStream(ContextErr(r.ctx.Err()))
m := <-r.recv.get()
return r.readAdditional(m, p)
case m := <-r.recv.get():
return r.readAdditional(m, p)
}
}
func (r *recvBufferReader) readAdditional(m recvMsg, p []byte) (n int, err error) {
r.recv.load()
if m.err != nil {
return 0, m.err
}
copied, _ := m.buffer.Read(p)
if m.buffer.Len() == 0 {
r.freeBuffer(m.buffer)
r.last = nil
} else {
r.last = m.buffer
}
return copied, nil
}
type streamState uint32
const (
@@ -166,6 +234,7 @@ const (
type Stream struct {
id uint32
st ServerTransport // nil for client side Stream
ct *http2Client // nil for server side Stream
ctx context.Context // the associated context of the stream
cancel context.CancelFunc // always nil for client side Stream
done chan struct{} // closed at the end of stream to unblock writers. On the client side.
@@ -182,12 +251,20 @@ type Stream struct {
// is used to adjust flow control, if needed.
requestRead func(int)
headerChan chan struct{} // closed to indicate the end of header metadata.
headerDone uint32 // set when headerChan is closed. Used to avoid closing headerChan multiple times.
headerChan chan struct{} // closed to indicate the end of header metadata.
headerChanClosed uint32 // set when headerChan is closed. Used to avoid closing headerChan multiple times.
// headerValid indicates whether a valid header was received. Only
// meaningful after headerChan is closed (always call waitOnHeader() before
// reading its value). Not valid on server side.
headerValid bool
// hdrMu protects header and trailer metadata on the server-side.
hdrMu sync.Mutex
header metadata.MD // the received header metadata.
hdrMu sync.Mutex
// On client side, header keeps the received header metadata.
//
// On server side, header keeps the header set by SetHeader(). The complete
// header will merged into this after t.WriteHeader() is called.
header metadata.MD
trailer metadata.MD // the key-value map of trailer metadata.
noHeaders bool // set if the client never received headers (set only after the stream is done).
@@ -232,26 +309,28 @@ func (s *Stream) getState() streamState {
return streamState(atomic.LoadUint32((*uint32)(&s.state)))
}
func (s *Stream) waitOnHeader() error {
func (s *Stream) waitOnHeader() {
if s.headerChan == nil {
// On the server headerChan is always nil since a stream originates
// only after having received headers.
return nil
return
}
select {
case <-s.ctx.Done():
return ContextErr(s.ctx.Err())
// Close the stream to prevent headers/trailers from changing after
// this function returns.
s.ct.CloseStream(s, ContextErr(s.ctx.Err()))
// headerChan could possibly not be closed yet if closeStream raced
// with operateHeaders; wait until it is closed explicitly here.
<-s.headerChan
case <-s.headerChan:
return nil
}
}
// RecvCompress returns the compression algorithm applied to the inbound
// message. It is empty string if there is no compression applied.
func (s *Stream) RecvCompress() string {
if err := s.waitOnHeader(); err != nil {
return ""
}
s.waitOnHeader()
return s.recvCompress
}
@@ -266,34 +345,33 @@ func (s *Stream) Done() <-chan struct{} {
return s.done
}
// Header acquires the key-value pairs of header metadata once it
// is available. It blocks until i) the metadata is ready or ii) there is no
// header metadata or iii) the stream is canceled/expired.
// Header returns the header metadata of the stream.
//
// On client side, it acquires the key-value pairs of header metadata once it is
// available. It blocks until i) the metadata is ready or ii) there is no header
// metadata or iii) the stream is canceled/expired.
//
// On server side, it returns the out header after t.WriteHeader is called. It
// does not block and must not be called until after WriteHeader.
func (s *Stream) Header() (metadata.MD, error) {
err := s.waitOnHeader()
// Even if the stream is closed, header is returned if available.
select {
case <-s.headerChan:
if s.header == nil {
return nil, nil
}
if s.headerChan == nil {
// On server side, return the header in stream. It will be the out
// header after t.WriteHeader is called.
return s.header.Copy(), nil
default:
}
return nil, err
s.waitOnHeader()
if !s.headerValid {
return nil, s.status.Err()
}
return s.header.Copy(), nil
}
// TrailersOnly blocks until a header or trailers-only frame is received and
// then returns true if the stream was trailers-only. If the stream ends
// before headers are received, returns true, nil. If a context error happens
// first, returns it as a status error. Client-side only.
func (s *Stream) TrailersOnly() (bool, error) {
err := s.waitOnHeader()
if err != nil {
return false, err
}
// if !headerDone, some other connection error occurred.
return s.noHeaders && atomic.LoadUint32(&s.headerDone) == 1, nil
// before headers are received, returns true, nil. Client-side only.
func (s *Stream) TrailersOnly() bool {
s.waitOnHeader()
return s.noHeaders
}
// Trailer returns the cached trailer metedata. Note that if it is not called
@@ -447,6 +525,7 @@ type ServerConfig struct {
ReadBufferSize int
ChannelzParentID int64
MaxHeaderListSize *uint32
HeaderTableSize *uint32
}
// NewServerTransport creates a ServerTransport with conn or non-nil error
@@ -465,8 +544,12 @@ type ConnectOptions struct {
FailOnNonTempDialError bool
// PerRPCCredentials stores the PerRPCCredentials required to issue RPCs.
PerRPCCredentials []credentials.PerRPCCredentials
// TransportCredentials stores the Authenticator required to setup a client connection.
// TransportCredentials stores the Authenticator required to setup a client
// connection. Only one of TransportCredentials and CredsBundle is non-nil.
TransportCredentials credentials.TransportCredentials
// CredsBundle is the credentials bundle to be used. Only one of
// TransportCredentials and CredsBundle is non-nil.
CredsBundle credentials.Bundle
// KeepaliveParams stores the keepalive parameters.
KeepaliveParams keepalive.ClientParameters
// StatsHandler stores the handler for stats.
@@ -494,8 +577,8 @@ type TargetInfo struct {
// NewClientTransport establishes the transport with the required ConnectOptions
// and returns it to the caller.
func NewClientTransport(connectCtx, ctx context.Context, target TargetInfo, opts ConnectOptions, onSuccess func()) (ClientTransport, error) {
return newHTTP2Client(connectCtx, ctx, target, opts, onSuccess)
func NewClientTransport(connectCtx, ctx context.Context, target TargetInfo, opts ConnectOptions, onPrefaceReceipt func(), onGoAway func(GoAwayReason), onClose func()) (ClientTransport, error) {
return newHTTP2Client(connectCtx, ctx, target, opts, onPrefaceReceipt, onGoAway, onClose)
}
// Options provides additional hints and information for message
@@ -540,9 +623,12 @@ type ClientTransport interface {
// is called only once.
Close() error
// GracefulClose starts to tear down the transport. It stops accepting
// new RPCs and wait the completion of the pending RPCs.
GracefulClose() error
// GracefulClose starts to tear down the transport: the transport will stop
// accepting new RPCs and NewStream will return error. Once all streams are
// finished, the transport will close.
//
// It does not block.
GracefulClose()
// Write sends the data for the given stream. A nil stream indicates
// the write is to be performed on the transport as a whole.
@@ -572,6 +658,9 @@ type ClientTransport interface {
// GetGoAwayReason returns the reason why GoAway frame was received.
GetGoAwayReason() GoAwayReason
// RemoteAddr returns the remote network address.
RemoteAddr() net.Addr
// IncrMsgSent increments the number of message sent through this transport.
IncrMsgSent()
@@ -706,3 +795,14 @@ type channelzData struct {
lastMsgSentTime int64
lastMsgRecvTime int64
}
// ContextErr converts the error from context package into a status error.
func ContextErr(err error) error {
switch err {
case context.DeadlineExceeded:
return status.Error(codes.DeadlineExceeded, err.Error())
case context.Canceled:
return status.Error(codes.Canceled, err.Error())
}
return status.Errorf(codes.Internal, "Unexpected error from context packet: %v", err)
}