runtime: GC leak on arm

[rsc]

The arm builders occasionally fail with complaints about finalizers not being run in the
sync.Pool test. Here is one failure:
http://build.golang.org/log/0d19587ed9638d98d6bcc517b0847c2b0649af63

I was unable to reproduce 0 of 100, but using

GOMAXPROCS=4 ./sync.test -test.short -test.run=TestPoolGC
-test.cpu=10,10,10,10,10,10,10,10,10,10,10,10

I was able to reproduce 69 of 100 quite reliably. That turned out to be 31 elements in a
pool cache line being retained due to timerproc holding arg longer than necessary. I
will send CL 64920044 for that.

With that out of the way, I was able to reproduce 0 of 100 fairly reliably as well, with
the same command line but many more 10s.

I do not yet know what is causing the 0 of 100 leak.

A bug to find before Go 1.3.

[rsc]

Comment 1:

This seems to have stopped. I believe the 0 of 100 leak was the test not waiting long
enough for the finalizers to run. And in a few cases it was a runtime race where the
finalizers were not running. We made the timeout longer and Dmitriy fixed the races.

Status changed to Fixed.