Comment 1:

Owner changed to @ianlancetaylor.

Status changed to Started.

Comment 2:

https://golang.org/cl/14526048

Comment 3:

This issue was closed by revision 649fc25.

Status changed to Fixed.

Comment 4:

This test still seems flaky, as reported here:
https://golang.org/cl/14526048#msg6

Comment 5:

Status changed to Accepted.

Comment 6:

Labels changed: added go1.3maybe.

Comment 7:

Labels changed: added release-none, removed go1.3maybe.

Comment 8:

Labels changed: added repo-main.

Comment 9:

https://golang.org/cl/14742043/

Comment 10:

CL https://golang.org/cl/14742043 references this issue.

Comment 11:

Applying https://golang.org/cl/90370043 and running with -test.cpu=2 makes the
test fail reliably on my system, instead of intermittently.
I don't see how to make the test as written work. Consider the following (after applying
CL 90370043):
$ go test -c net
$ GODEBUG=allocfreetrace=1 ./net.test -test.cpu=2 -test.run=TestDialFailPDLeak -test.v
2> aft.log
$ grep "single object" aft.log | awk -F "," '{print $3}' | sort | uniq -c | sort -n -r |
head
2523  single object of net.TCPAddr)
2503  single object of net.OpError)
2500  single object of syscall.SockaddrInet4)
2500  single object of struct { F uintptr; A0 *sync.WaitGroup; A1 **net.Dialer; A2
**testing.T })
2500  single object of net.netFD)
 689  single object)
  43  single object of errors.errorString)
  26  single object of net.Interface)
  22  single object of syscall.SockaddrDatalink)
  22  single object of syscall.InterfaceMessage)
Every pass through the loop allocates a net.TCPAddr, a net.OpError, a
syscall.SockaddrInet4, a closure, and a net.netFD. The test as written assumes that
nothing in the loop allocates, and thus any allocs must be due to allocPollDesc. I'm not
sure how to alter the test to fix this, aside perhaps from instrumenting allocPollDesc
directly.
Suggestions?

Comment 12:

I abandoned CL 14742043 because it's been ignored for a few months and this is a subtle
issue. I think that checking runtime.MemStats.OtherSys might be better; see changeset:
18019:cbacca9a8c67; poll descriptor pool stays at the non-GC region.

Comment 13:

Naively switching to runtime.MemStats.OtherSys didn't do the trick. Surprisingly, there
were no OtherSys allocs at all during the first few loops, which makes me wonder whether
runtime.MemStats.OtherSys is working correctly.
I'd be curious to see CL 14742043, if it is still floating around somewhere on your
machine. (I'm just tired of ignoring this test failure when running other net tests.)

Comment 14:

CL https://golang.org/cl/98080043 mentions this issue.

Comment 15:

This issue was closed by revision f40f0b2.

Status changed to Fixed.