https://go.dev/cl/432038 changes the frequency with which the HTTP/2 server sends window size updates to clients. Previously it would send updates to extend the stream window immediately upon receiving data; now it sends updates only when at least half the current window has been consumed. This reduces the frequency of window updates and should improve performance overall.

Your heap profiles show that your heap is pretty much entirely consumed by HTTP/2 stream buffers, used to hold data being received or sent. Buffers are allocated as needed from a set of pools of varying size classes; the largest class is 16384 bytes.

I'm not sure why this change would lead to more buffers being consumed. I don't see any obvious way for a new memory leak to have been introduced here; perhaps I'm missing something, but I suspect any change is a result of the pattern of data arriving having changed. This change might result in clients sending data in fewer, larger chunks; perhaps this results in more bursty memory usage? Or perhaps clients are sending data faster, and running ahead of the ability of the server to forward it?

Do you configure the HTTP/2 server flow control limits at all? (Server. MaxUploadBufferPerConnection and Server. MaxUploadBufferPerStream in http2.Server.)

Thanks for taking a look.

I'm not sure why this change would lead to more buffers being consumed. I don't see any obvious way for a new memory leak to have been introduced here; perhaps I'm missing something, but I suspect any change is a result of the pattern of data arriving having changed.

I am also not clear on how this change ultimately results in unbounded memory usage. I set up a closed-environment test workload comparing the behavior at both the good and bad commits of golang.org/x/net (with the same pattern of data sent by the client), and can reliably reproduce unbounded memory growth with the commit in question.

This change might result in clients sending data in fewer, larger chunks; perhaps this results in more bursty memory usage?

This hypothesis makes sense, though in practice I haven't been observing bursty memory usage. This graph illustrates the unbounded heap usage growth over time from ~15 MB to ~80 GB as measured by /memory/classes/heap/objects:bytes from Go's runtime metrics system:

(The sudden drop was due to OOM kill)

Do you configure the HTTP/2 server flow control limits at all?

MaxUploadBufferPerConnection and MaxUploadBufferPerStream are not set explicitly, so the defaults are being used.

I have created a minimally complete and verifiable example workload that demonstrates the issue.

Source tree:

.
├── client
│   ├── client
│   └── main.go
├── go.mod
├── go.sum
├── proto
│   ├── service_grpc.pb.go
│   ├── service.pb.go
│   └── service.proto
├── proxy
│   ├── main.go
│   └── proxy
└── server
    ├── main.go
    └── server

go.mod:

module bug

go 1.19

require (
	golang.org/x/net v0.1.0
	google.golang.org/grpc v1.50.1
	google.golang.org/protobuf v1.28.1
)

require (
	github.com/golang/protobuf v1.5.2 // indirect
	golang.org/x/sys v0.1.0 // indirect
	golang.org/x/text v0.4.0 // indirect
	google.golang.org/genproto v0.0.0-20200526211855-cb27e3aa2013 // indirect
)

proto/service.proto:

syntax = "proto3";

option go_package = "bug/proto";

message Request {
  bytes data = 1;
}

message Response {}

service Example {
    rpc Method(stream Request) returns (Response);
}

proxy/main.go:

package main

import (
	"crypto/tls"
	"net"
	"net/http"
	"net/http/httputil"

	"golang.org/x/net/http2"
	"golang.org/x/net/http2/h2c"
)

func main() {
	tr := &http2.Transport{
		AllowHTTP: true,
		DialTLS: func(network string, addr string, cfg *tls.Config) (net.Conn, error) {
			return net.Dial("tcp", "localhost:7001")
		},
	}

	rp := &httputil.ReverseProxy{
		Transport:     tr,
		FlushInterval: -1,
		Director: func(req *http.Request) {
			req.URL.Scheme = "http"
		},
	}

	handler := h2c.NewHandler(rp, &http2.Server{})

	srv := &http.Server{
		Addr:    "localhost:7000",
		Handler: handler,
	}

	srv.ListenAndServe()
}

server/main.go:

package main

import (
	"fmt"
	"net"

	"google.golang.org/grpc"

	"bug/proto"
)

type example struct {
	proto.UnimplementedExampleServer
}

func (e *example) Method(stream proto.Example_MethodServer) error {
	for {
		request, err := stream.Recv()
		if err != nil {
			break
		}

		fmt.Println("request", request)
	}

	return stream.SendAndClose(&proto.Response{})
}

func main() {
	srv := grpc.NewServer()
	e := &example{}

	ln, _ := net.Listen("tcp", ":7001")

	proto.RegisterExampleServer(srv, e)

	srv.Serve(ln)
}

client/main.go:

package main

import (
	"context"

	"google.golang.org/grpc"
	"google.golang.org/grpc/credentials/insecure"

	"bug/proto"
)

func main() {
	conn, err := grpc.DialContext(context.Background(), "localhost:7000", grpc.WithTransportCredentials(insecure.NewCredentials()))
	if err != nil {
		panic(err)
	}

	e := proto.NewExampleClient(conn)
	stream, err := e.Method(context.Background())
	if err != nil {
		panic(err)
	}

	for {
		err := stream.Send(&proto.Request{Data: []byte{1, 2, 3, 4, 5}})
		if err != nil {
			panic(err)
		}
	}
}

In words, this is:

• A HTTP/2-capable reverse proxy listening on localhost:7000 proxying to localhost:7001.
• A streaming gRPC service definition.
• A gRPC server implementation that simply prints all incoming messages over the stream, listening on localhost:7001.
• A gRPC client that sends the same message over the stream continuously forever, connecting to localhost:7000.

First, generate gRPC code:

protoc --plugin=protoc-gen-go=protoc-gen-go --go_out=paths=source_relative:. --plugin=protoc-gen-go-grpc=protoc-gen-go-grpc --go-grpc_out=paths=source_relative:. service.proto

To reproduce the bug:

1. Install golang.org/x/net@latest
2. Build and run both the proxy and server.
3. Build and run the client, and let it run for 5 minutes.
4. Observe a steady increase in memory usage in the proxy process.

To verify that this is a regression:

1. Install golang.org/x/net@v0.0.0-20220920183852-bf014ff85ad5
2. Build and run both the proxy and sever.
3. Build and run the client, and let it run for 5 minutes.
4. Observe consistent/non-growing memory usage in the proxy process.

I think I have found out what's wrong. There is not memory leak. It's just that the proxy writes to the buffer (data from the client to the proxy) too fast and reads from the buffer (data from the proxy to the server) too slow. As indicated by neild@, you should configure the HTTP/2 server flow control limits.

The proxy writes to the buffer too fast thanks to the improvement of golang/net@2e0b12c. To verify this hypothesis:

1. add log.Println("chunks:", len(b.chunks)) after this line https://github.com/golang/net/blob/master/http2/databuffer.go#L91, and run the test, you should find that the number of chunks keeps growing.
2. simplify the server to make it faster, you should find that the number of b.chunks grows slower or does not grow.

   --- a/server/main.go
   +++ b/server/main.go
   @@ -1,38 +1,35 @@
    package main
    
    import (
   -   "fmt"
       "net"
    
       "google.golang.org/grpc"
    
       "bug/proto"
    )
    
    type example struct {
       proto.UnimplementedExampleServer
    }
    
    func (e *example) Method(stream proto.Example_MethodServer) error {
       for {
   -       request, err := stream.Recv()
   +       _, err := stream.Recv()
           if err != nil {
               break
           }
   -
   -       fmt.Println("request", request)
       }
    
       return stream.SendAndClose(&proto.Response{})
    }
    
    func main() {
       srv := grpc.NewServer()
       e := &example{}
    
       ln, _ := net.Listen("tcp", ":7001")
    
       proto.RegisterExampleServer(srv, e)
    
       srv.Serve(ln)
    }
3. configure the flow control limits, 1<<13 works for me:

   --- a/proxy/main.go
   +++ b/proxy/main.go
   @@ -1,37 +1,41 @@
    package main
    
    import (
       "crypto/tls"
       "net"
       "net/http"
       "net/http/httputil"
    
       "golang.org/x/net/http2"
       "golang.org/x/net/http2/h2c"
    )
    
    func main() {
       tr := &http2.Transport{
           AllowHTTP: true,
           DialTLS: func(network string, addr string, cfg *tls.Config) (net.Conn, error) {
               return net.Dial("tcp", "localhost:7001")
           },
       }
    
       rp := &httputil.ReverseProxy{
           Transport:     tr,
           FlushInterval: -1,
           Director: func(req *http.Request) {
               req.URL.Scheme = "http"
           },
       }
    
   -   handler := h2c.NewHandler(rp, &http2.Server{})
   +   handler := h2c.NewHandler(rp, &http2.Server{
   +       // The default value is 1<<20.
   +       MaxUploadBufferPerConnection: 1 << 13,
   +       MaxUploadBufferPerStream:     1 << 13,
   +   })
    
       srv := &http.Server{
           Addr:    "localhost:7000",
           Handler: handler,
       }
    
       srv.ListenAndServe()
    }

Thanks for looking into this @ZekeLu. I think you are right. I tested a more restrictive MaxUploadBufferPerConnection and MaxUploadBufferPerStream in the proxy at golang.org/x/net@v0.0.0-20221019024206-cb67ada4b0ad and can confirm that the example test workload is no longer causing unbounded memory growth. I guess this was a latent behavior in the proxy that was only exposed by the optimization introduced in 2e0b12c274b70e5e836cf4bb0cda99c44f61a4d1.

That said, I think there's a couple remaining open questions:

(1) The example workload described in the issue seems fairly vanilla. Is there an opportunity to perhaps improve the "out-of-the-box" default http2.Server behavior to avoid the original problem? For example, does it make sense to change the default value of MaxUploadBufferPerConnection and MaxUploadBufferPerStream to 1 << 13?

(2) The original incident described in the memory usage graph above increased total heap allocation size from 15 MB to 80 GB (which exhausted all 80 GB of on-disk swap space on the system in question). It sounds like there should exist some capability to set/tune the maximum http2.Server buffer usage across all connections and streams to avoid this potential failure mode.

Change https://go.dev/cl/444816 mentions this issue: http2: calcualte a correct window increment size for a stream

My two cents:

1. Every application has its own characteristic, I don't think 1 << 13 works for most of the applications. I think it's better to leave the decision to the developers. Maybe it should be documented to remind the developers to provide a proper configuration. I'm not sure.
2. MaxUploadBufferPerConnection and MaxUploadBufferPerStream affect the buffer indirectly. So this question seems essentially the same as question 1 to me.

I found a new issue while investigating this issue. It seems that the increment size for a stream could be wrong in some cases. For example, when the flow control configuration is like this:

handler := h2c.NewHandler(rp, &http2.Server{
	// The default value is 1<<20.
	MaxUploadBufferPerConnection: 1 << 16,
	MaxUploadBufferPerStream:     1 << 20,
})

The proxy will panic with http: panic serving 127.0.0.1:44136: internal error; sent too many window updates without decrements?. I have sent CL 444816 to fix it.

Change https://go.dev/cl/448055 mentions this issue: Revert "http2: Send WindowUpdates when remaining bytes are below a threshold"

I've discovered another issue with https://go.dev/cl/432038, which is effectively disabling server pushback on clients. (Sorry, I really should have detected that in review. Mea culpa.)

Sent rollback.