net, x/net/internal/socket: ipv6ZoneCache not updated on cache misses

[stapelberg]

Please answer these questions before submitting your issue. Thanks!

What version of Go are you using (go version)?

go version go1.11.1 linux/amd64

What operating system and processor architecture are you using (go env)?

GOARCH="amd64"
GOBIN=""
GOCACHE="/home/michael/.cache/go-build"
GOEXE=""
GOFLAGS=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/michael/go"
GOPROXY=""
GORACE=""
GOROOT="/home/michael/sdk/go1.11.1"
GOTMPDIR=""
GOTOOLDIR="/home/michael/sdk/go1.11.1/pkg/tool/linux_amd64"
GCCGO="/usr/bin/gccgo"
CC="gcc"
CXX="g++"
CGO_ENABLED="1"
GOMOD=""
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build554331710=/tmp/go-build -gno-record-gcc-switches"

What did you do?

Creating a listener on a link-local IPv6 address fails on bootup for ≈ 1 minute longer than necessary. This code:

ln, err := net.Listen("tcp", "[fe80::ba27:ebff:fe3b:11b0%eth0]:80")
if err != nil {
  log.Println(err)
}

results in:

1970/01/01 01:00:05 listen tcp [fe80::ba27:ebff:fe3b:11b0%eth0]:80: bind: cannot assign requested address
2018/11/01 18:19:37 listen tcp [fe80::ba27:ebff:fe3b:11b0%eth0]:80: bind: invalid argument
2018/11/01 18:20:11 listen tcp [fe80::ba27:ebff:fe3b:11b0%eth0]:80: bind: invalid argument
[…]
2018/11/01 18:21:11 now listening on [fe80::ba27:ebff:fe3b:11b0%eth0]:80

Notably, listening fails for about 1 minute, which is the time it takes for zoneCache to be invalidated.

This is on a Raspberry Pi 3B+, where the eth0 interface is connected via USB (as is the case on all Raspberry Pi models).

What’s happening here is the following (in order):

1. System boots, I install a listener on all available IP addresses, including ::1, which fills the zoneCache.
2. The network interface eth0 is discovered once the USB hub is probed.
3. eth0 comes up, assigning itself a link-local IPv6 address.
4. At this point, listening fails with “cannot assign requested address”, because the link-local IPv6 address is marked as tentative while the kernel performs duplicate address detection (DAD).
5. After a second or so, DAD finishes and the address becomes available.
6. At this point, listening still fails, this time with “invalid argument”, because the bind call’s sin6_scope_id parameter is set to 0, because the zoneCache does not contain the correct entry for the eth0 interface. This is because zoneCache was filled before that interface existed. In strace, you’ll see: [pid 1716] bind(9, {sa_family=AF_INET6, sin6_port=htons(80), inet_pton(AF_INET6, "fe80::ba27:ebff:fe3b:11b0", &sin6_addr), sin6_flowinfo=htonl(0), sin6_scope_id=0}, 28) = -1 EINVAL (Invalid argument)
7. After (at most) 1 minute, the zoneCache is invalidated and updated.
8. Only now does listening succeed. In strace, you’ll see: [pid 1715] bind(7, {sa_family=AF_INET6, sin6_port=htons(80), inet_pton(AF_INET6, "fe80::ba27:ebff:fe3b:11b0", &sin6_addr), sin6_flowinfo=htonl(0), sin6_scope_id=if_nametoindex("eth0")}, 28) = 0

I suggest we update the zoneCache upon cache misses, which fixes this issue. I can send a CL to do that.

[mikioh]

It's fine if this proposal focuses on the invalidation using the signal of operation malfunction; otherwise, we should take another approach that working together with the routing information base inside the kernel for accuracy and correctness (upon the fate/nature of IP routing; we cannot avoid any micro routing loop at any time, anywhere.)

Also we need to fix x/net/internal/socket too.

I can send a CL to do that.

Thanks. I miss a cup of Glühwein and Heidelberg Christmas market.

[gopherbot]

Change https://golang.org/cl/146941 mentions this issue: net: update zoneCache on cache misses to cover appearing interfaces

[stapelberg]

Once the change landed, I can update x/net, too.

[mikioh]

Please add Updates golang/go#28535 to the CL description for x/net.

[gopherbot]

Change https://golang.org/cl/147739 mentions this issue: net: update zoneCache on cache misses to cover appearing interfaces