x/net/internal/quic: deadlock in `Endpoint.Close` via `connsMu`

[gopherbot]

#!watchflakes
post <- pkg == "golang.org/x/net/internal/quic" && log ~ `panic: test timed out` && log ~ `^\s+TestStreamsCreateConcurrency \(\d+m.*\)`

Issue created automatically to collect these failures.

Example (log):

panic: test timed out after 10m0s
running tests:
	TestStreamsCreateConcurrency (9m57s)

goroutine 382 [running]:
testing.(*M).startAlarm.func1()
	/tmp/workdir/go/src/testing/testing.go:2366 +0x386
created by time.goFunc
	/tmp/workdir/go/src/time/sleep.go:177 +0x29

— watchflakes

[gopherbot]

Found new dashboard test flakes for:

#!watchflakes
default <- pkg == "golang.org/x/net/internal/quic" && test == ""

2024-01-04 19:50 netbsd-386-9_3 net@689bbc70 go@8088b6db x/net/internal/quic (log)

panic: test timed out after 10m0s
running tests:
	TestStreamsCreateConcurrency (9m57s)

goroutine 382 [running]:
testing.(*M).startAlarm.func1()
	/tmp/workdir/go/src/testing/testing.go:2366 +0x386
created by time.goFunc
	/tmp/workdir/go/src/time/sleep.go:177 +0x29

— watchflakes

[bcmills]

Looks like a classic lock-ordering bug:

• Endpoint.Close calls Conn.Abort with e.connsMu locked.
• Conn.Abort calls Conn.sendMsg, which selects on c.msgc and c.donec.
• c.donec is closed by Conn.Cleanup after a call to c.endpoint.connDrained.
• Endpoint.connDrained acquires e.connsMu.

(attn @neild @jba)

[gopherbot]

Change https://go.dev/cl/554695 mentions this issue: quic: avoid deadlock in Endpoint.Close