net: Dial with specified source address selects wrong scoped destination address

[mikioh]

When running TestDialRoutingToBlackhole in golang.org/cl/12561, we see the following on various platforms:

# sudo GODEBUG=netdns=go go test -v -run=Blackhole
=== RUN   TestDialRoutingToBlackhole
installing [route add -host -inet6 2404:6800:4004:807::1011/128 ::1 -blackhole]
installing [route add -host 173.194.117.241/32 127.0.0.1 -blackhole]
installing [route add -host 173.194.117.244/32 127.0.0.1 -blackhole]
installing [route add -host 173.194.117.242/32 127.0.0.1 -blackhole]
installing [route add -host 173.194.117.243/32 127.0.0.1 -blackhole]
installing [route add -host 173.194.117.240/32 127.0.0.1 -blackhole]
uninstalling [route add -host -inet6 2404:6800:4004:807::1011/128 ::1 -blackhole]
uninstalling [route add -host 173.194.117.241/32 127.0.0.1 -blackhole]
uninstalling [route add -host 173.194.117.244/32 127.0.0.1 -blackhole]
uninstalling [route add -host 173.194.117.242/32 127.0.0.1 -blackhole]
uninstalling [route add -host 173.194.117.243/32 127.0.0.1 -blackhole]
uninstalling [route add -host 173.194.117.240/32 127.0.0.1 -blackhole]
--- FAIL: TestDialRoutingToBlackhole (9.25s)
    dial_unix_test.go:58: #0/0: elapsed 504.732193ms for {500ms 0001-01-01 00:00:00 +0000 UTC [240f:6d:1ced:1:815b:276a:fe81:7727]:0 false 0 0}
    dial_unix_test.go:58: #0/1: elapsed 504.187842ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:66: #0/1: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:58: #0/2: elapsed 505.047238ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:66: #0/2: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:58: #0/3: elapsed 502.103169ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:66: #0/3: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:58: #0/4: elapsed 503.668158ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:66: #0/4: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:58: #0/5: elapsed 502.566501ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:66: #0/5: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 false 0 0}
    dial_unix_test.go:58: #1/0: elapsed 502.806944ms for {500ms 0001-01-01 00:00:00 +0000 UTC [240f:6d:1ced:1:815b:276a:fe81:7727]:0 true 0 0}
    dial_unix_test.go:58: #1/1: elapsed 503.798456ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:66: #1/1: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:58: #1/2: elapsed 504.844686ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:66: #1/2: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:58: #1/3: elapsed 505.069618ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:66: #1/3: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:58: #1/4: elapsed 503.374228ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:66: #1/4: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:58: #1/5: elapsed 501.865917ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:66: #1/5: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 0 0}
    dial_unix_test.go:58: #2/0: elapsed 502.118066ms for {500ms 0001-01-01 00:00:00 +0000 UTC [240f:6d:1ced:1:815b:276a:fe81:7727]:0 true 1ns 0}
    dial_unix_test.go:58: #2/1: elapsed 502.060709ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:66: #2/1: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:58: #2/2: elapsed 501.901003ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:66: #2/2: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:58: #2/3: elapsed 504.841677ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:66: #2/3: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:58: #2/4: elapsed 504.515092ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:66: #2/4: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:58: #2/5: elapsed 500.103138ms for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
    dial_unix_test.go:66: #2/5: dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80: connect: no route to host for {500ms 0001-01-01 00:00:00 +0000 UTC 192.168.0.2:0 true 1ns 0}
FAIL

"dial tcp 192.168.0.2:0->[2404:6800:4004:807::1011]:80" looks clearly wrong.

[mikioh]

/cc @pmarks-net

[mikioh]

This issue would not be critical for go1.5 because it implements new Dial using connect-by-name method that tries each IP address until one succeeds.

[pmarks-net]

Looks like there are two problems here:

• LocalAddr is a single value, when it should be a mapping from destination prefix to source address. But taking a step back, isn't source address selection supposed to be the kernel's job?
• If you specify an IPv4 source but connect to an IPv6 destination (or vice versa), the failure requires a timeout to elapse. Instead, Dial should either filter out incompatible destinations, or ignore LocalAddr when using an incompatible destination.

[mikioh]

But taking a step back, isn't source address selection supposed to be the kernel's job?

I don't think so. If the net package abandons taking that responsibility, users could not implement an application which needs to be run on multi-homed nodes. I think that source and/or destination address selection is part of IP routing, and we've introduced the feature for fancy connect-by-name method in go1.5. Seems like we need to make the feature work correctly for both [dst:name, src:literal] and [dst:name, src:wildcard] cases.

[gopherbot]

CL https://golang.org/cl/20586 mentions this issue.