net/http: IdleConnTimeout is invalid for HTTP/2

[ahajunzi]

Go version

go version go1.22.2 linux/amd64

Output of go env in your module/workspace:

GO111MODULE='on'
GOARCH='amd64'
GOBIN=''
GOCACHE='/data/home/sakiluzhang/.cache/go-build'
GOENV='/data/home/sakiluzhang/.config/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFLAGS=''
GOHOSTARCH='amd64'
GOHOSTOS='linux'
GOINSECURE=''
GOMODCACHE='/data/home/sakiluzhang/go/pkg/mod'
GONOPROXY=''
GONOSUMDB=''
GOOS='linux'
GOPATH='/data/home/sakiluzhang/go'
GOPRIVATE=''
GOPROXY='https://sakiluzhang:lFEajkXQ@goproxy.woa.com'
GOROOT='/data/home/sakiluzhang/go'
GOSUMDB='sum.woa.com+643d7a06+Ac5f5VOC4N8NUXdmhbm8pZSXIWfhek5JSmWdWrq7pLX4'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/data/home/sakiluzhang/go/pkg/tool/linux_amd64'
GOVCS=''
GOVERSION='go1.22.2'
GCCGO='gccgo'
GOAMD64='v1'
AR='ar'
CC='gcc'
CXX='g++'
CGO_ENABLED='1'
GOMOD='/data/home/sakiluzhang/DServer/go.mod'
GOWORK=''
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
PKG_CONFIG='pkg-config'
GOGCCFLAGS='-fPIC -m64 -pthread -Wl,--no-gc-sections -fmessage-length=0 -ffile-prefix-map=/tmp/go-build929430334=/tmp/go-build -gno-record-gcc-switches'

What did you do?

func (t *Transport) tryPutIdleConn(pconn *persistConn) error {

    if t.DisableKeepAlives || t.MaxIdleConnsPerHost < 0 {
        return errKeepAlivesDisabled
    }
    if pconn.isBroken() {
        return errConnBroken
    }
    pconn.markReused()

    t.idleMu.Lock()
    defer t.idleMu.Unlock()

    // HTTP/2 (pconn.alt != nil) connections do not come out of the idle list,
    // because multiple goroutines can use them simultaneously.
    // If this is an HTTP/2 connection being “returned,” we're done.
    if pconn.alt != nil && t.idleLRU.m[pconn] != nil {
             **here pconn.idleAt has not been updated and returned.**
        return nil
    }
       ......

What did you see happen?

I hope one connection is checked IdleConnTimeout should start at the latest time I used it rather than the first time

I hope *http.Transport.idleLRU list should remove connection not only by MaxIdleConns (in many cases, you can't estimate proper num for in China there maybe Tens of millions users use our product simultaneously and more http requests will be produced) but also by IdleConnTimeout

What did you expect to see?

read above

[gabyhelp]

Related Issues and Documentation

• net/http: Connection to HTTP/2 site with IdleConnTimeout hangs #34387 (closed)
• net/http: transport closes healthy connections when pool limit is reached #42650
• net/http: dial should not be canceled if an idle connection is reused #66442 (closed)
• x/net/http2: SetKeepAlivesEnabled(false) closes all HTTP/2 connections older than 5 seconds #36946 (closed)
• net/http: setting IdleConnTimeout to a low value on http.Transport is causing non-deterministic tls handshake failures #33891
• x/net/http2: data race in Transport.newClientConn() when connection idle timeout is too low #66763 (closed)
• net/http: zombie connection leaked after Server.IdleTimeout #51614 (closed)
• net/http2: client immediately close most of connections via forward proxy even if idleConn was configured #50000 (closed)
• net/http http2 reverse proxy connection pool #63112 (closed)
• [net/http] Support timeout of idle "keep-alive" connections in http.Transport #15522 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[ianlancetaylor]

CC @neild

[neild]

The HTTP/2 idle timeout is handled in the HTTP/2 transport: https://go.googlesource.com/net/+/refs/tags/v0.28.0/http2/transport.go#862