open-nomad/client/agent_endpoint.go

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

234 lines
5.4 KiB
Go
Raw Normal View History

package client
import (
"bytes"
"context"
"errors"
"io"
"time"
"github.com/hashicorp/go-msgpack/codec"
"github.com/hashicorp/nomad/command/agent/host"
"github.com/hashicorp/nomad/command/agent/monitor"
"github.com/hashicorp/nomad/command/agent/pprof"
"github.com/hashicorp/nomad/helper/pointer"
"github.com/hashicorp/nomad/nomad/structs"
metrics "github.com/armon/go-metrics"
log "github.com/hashicorp/go-hclog"
sframer "github.com/hashicorp/nomad/client/lib/streamframer"
cstructs "github.com/hashicorp/nomad/client/structs"
)
type Agent struct {
c *Client
}
func NewAgentEndpoint(c *Client) *Agent {
a := &Agent{c: c}
a.c.streamingRpcs.Register("Agent.Monitor", a.monitor)
return a
}
func (a *Agent) Profile(args *structs.AgentPprofRequest, reply *structs.AgentPprofResponse) error {
// Check ACL for agent write
aclObj, err := a.c.ResolveToken(args.AuthToken)
if err != nil {
return err
} else if aclObj != nil && !aclObj.AllowAgentWrite() {
return structs.ErrPermissionDenied
}
// If ACLs are disabled, EnableDebug must be enabled
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
if aclObj == nil && !a.c.GetConfig().EnableDebug {
return structs.ErrPermissionDenied
}
var resp []byte
var headers map[string]string
// Determine which profile to run and generate profile.
// Blocks for args.Seconds
// Our RPC endpoints currently don't support context
// or request cancellation so stubbing with TODO
switch args.ReqType {
case pprof.CPUReq:
resp, headers, err = pprof.CPUProfile(context.TODO(), args.Seconds)
case pprof.CmdReq:
resp, headers, err = pprof.Cmdline()
case pprof.LookupReq:
resp, headers, err = pprof.Profile(args.Profile, args.Debug, args.GC)
case pprof.TraceReq:
resp, headers, err = pprof.Trace(context.TODO(), args.Seconds)
}
if err != nil {
if pprof.IsErrProfileNotFound(err) {
return structs.NewErrRPCCoded(404, err.Error())
}
return structs.NewErrRPCCoded(500, err.Error())
}
// Copy profile response to reply
reply.Payload = resp
reply.AgentID = a.c.NodeID()
reply.HTTPHeaders = headers
return nil
}
func (a *Agent) monitor(conn io.ReadWriteCloser) {
defer metrics.MeasureSince([]string{"client", "agent", "monitor"}, time.Now())
2019-10-15 14:33:07 +00:00
defer conn.Close()
// Decode arguments
var args cstructs.MonitorRequest
decoder := codec.NewDecoder(conn, structs.MsgpackHandle)
encoder := codec.NewEncoder(conn, structs.MsgpackHandle)
if err := decoder.Decode(&args); err != nil {
handleStreamResultError(err, pointer.Of(int64(500)), encoder)
return
}
// Check acl
if aclObj, err := a.c.ResolveToken(args.AuthToken); err != nil {
handleStreamResultError(err, pointer.Of(int64(403)), encoder)
return
2019-10-25 18:25:19 +00:00
} else if aclObj != nil && !aclObj.AllowAgentRead() {
handleStreamResultError(structs.ErrPermissionDenied, pointer.Of(int64(403)), encoder)
return
}
logLevel := log.LevelFromString(args.LogLevel)
if args.LogLevel == "" {
logLevel = log.LevelFromString("INFO")
}
if logLevel == log.NoLevel {
handleStreamResultError(errors.New("Unknown log level"), pointer.Of(int64(400)), encoder)
return
}
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
monitor := monitor.New(512, a.c.logger, &log.LoggerOptions{
JSONFormat: args.LogJSON,
Level: logLevel,
})
frames := make(chan *sframer.StreamFrame, streamFramesBuffer)
errCh := make(chan error)
var buf bytes.Buffer
frameCodec := codec.NewEncoder(&buf, structs.JsonHandle)
framer := sframer.NewStreamFramer(frames, 1*time.Second, 200*time.Millisecond, 1024)
framer.Run()
defer framer.Destroy()
// goroutine to detect remote side closing
go func() {
if _, err := conn.Read(nil); err != nil {
// One end of the pipe explicitly closed, exit
cancel()
return
}
2020-12-09 19:05:18 +00:00
<-ctx.Done()
}()
logCh := monitor.Start()
defer monitor.Stop()
initialOffset := int64(0)
// receive logs and build frames
go func() {
defer framer.Destroy()
LOOP:
for {
select {
case log := <-logCh:
if err := framer.Send("", "log", log, initialOffset); err != nil {
select {
case errCh <- err:
case <-ctx.Done():
}
break LOOP
}
case <-ctx.Done():
break LOOP
}
}
}()
var streamErr error
OUTER:
for {
select {
case frame, ok := <-frames:
if !ok {
// frame may have been closed when an error
// occurred. Check once more for an error.
select {
case streamErr = <-errCh:
// There was a pending error!
default:
// No error, continue on
}
break OUTER
}
var resp cstructs.StreamErrWrapper
if args.PlainText {
resp.Payload = frame.Data
} else {
if err := frameCodec.Encode(frame); err != nil {
streamErr = err
break OUTER
}
resp.Payload = buf.Bytes()
buf.Reset()
}
if err := encoder.Encode(resp); err != nil {
streamErr = err
break OUTER
}
encoder.Reset(conn)
case <-ctx.Done():
break OUTER
}
}
if streamErr != nil {
handleStreamResultError(streamErr, pointer.Of(int64(500)), encoder)
return
}
}
// Host collects data about the host evironment running the agent
func (a *Agent) Host(args *structs.HostDataRequest, reply *structs.HostDataResponse) error {
aclObj, err := a.c.ResolveToken(args.AuthToken)
if err != nil {
return err
}
if (aclObj != nil && !aclObj.AllowAgentRead()) ||
client: fix data races in config handling (#14139) Before this change, Client had 2 copies of the config object: config and configCopy. There was no guidance around which to use where (other than configCopy's comment to pass it to alloc runners), both are shared among goroutines and mutated in data racy ways. At least at one point I think the idea was to have `config` be mutable and then grab a lock to overwrite `configCopy`'s pointer atomically. This would have allowed alloc runners to read their config copies in data race safe ways, but this isn't how the current implementation worked. This change takes the following approach to safely handling configs in the client: 1. `Client.config` is the only copy of the config and all access must go through the `Client.configLock` mutex 2. Since the mutex *only protects the config pointer itself and not fields inside the Config struct:* all config mutation must be done on a *copy* of the config, and then Client's config pointer is overwritten while the mutex is acquired. Alloc runners and other goroutines with the old config pointer will not see config updates. 3. Deep copying is implemented on the Config struct to satisfy the previous approach. The TLS Keyloader is an exception because it has its own internal locking to support mutating in place. An unfortunate complication but one I couldn't find a way to untangle in a timely fashion. 4. To facilitate deep copying I made an *internally backward incompatible API change:* our `helper/funcs` used to turn containers (slices and maps) with 0 elements into nils. This probably saves a few memory allocations but makes it very easy to cause panics. Since my new config handling approach uses more copying, it became very difficult to ensure all code that used containers on configs could handle nils properly. Since this code has caused panics in the past, I fixed it: nil containers are copied as nil, but 0-element containers properly return a new 0-element container. No more "downgrading to nil!"
2022-08-18 23:32:04 +00:00
(aclObj == nil && !a.c.GetConfig().EnableDebug) {
return structs.ErrPermissionDenied
}
data, err := host.MakeHostData()
if err != nil {
return err
}
reply.AgentID = a.c.NodeID()
reply.HostData = data
return nil
}