Fix for leaking file descriptor issue

Fix for leaking file descriptor issue

When calling a non-existent executable, a pair of file descriptors will be
created with pipe(). One end goes to the exitcode handler thread, the other end
goes to the epoll() thread.

If an error occurs during process creation, the pipe is still created and the
exit handler thread writes the exitcode to the pipe, but the other file
descriptor is not added to the epoll() set and is therefore leaked. (i.e.
Process_Start ignores the other end of the pipe if exec() failed in
Proces::Start()).

R=sgjesse@google.com

Committed: https://code.google.com/p/dart/source/detail?r=23534

[kustermann, 2013-06-03 07:30:27]

https://codereview.chromium.org/15980006/diff/1/runtime/bin/process_linux.cc
File runtime/bin/process_linux.cc (right):

https://codereview.chromium.org/15980006/diff/1/runtime/bin/process_linux.cc#...
runtime/bin/process_linux.cc:521: }
In these two cases, we return a nonzero value. The caller (i.e. Process_Start)
will then ignore the '*exit_event' fd.

[Søren Gjesse, 2013-06-03 07:49:51]

lgtm, thanks

Isn't this also the case on Mac OS?

[kustermann, 2013-06-03 09:31:24]

As discussed offline, we should just close the reading part in Process::Start().
I updated it to include the fix for linux/android/macos.

Turns out that GetProcessExitCodes is already ignoring the 'broken pipe' errors.
All tests are passing now.

[Søren Gjesse, 2013-06-03 09:40:53]

LGTM!

[kustermann, 2013-06-03 09:41:04]

If you're fine with it, I'll land it.

[kustermann, 2013-06-03 09:41:28]

On 2013/06/03 09:40:53, Søren Gjesse wrote:
> LGTM!

Thank you.

[kustermann, 2013-06-03 09:41:49]

Committed patchset #2 manually as r23534 (presubmit successful).