[Piglit] [PATCH] framework: avoid blocking joins, allowing ^C to work

[Ilia Mirkin]

On Sun, Feb 16, 2014 at 6:14 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> On 02/15/2014 10:30 AM, Ilia Mirkin wrote:
>> On Sat, Feb 15, 2014 at 7:37 AM, Dylan Baker <baker.dylan.c at gmail.com> wrote:
>>> On Saturday, February 15, 2014 03:27:44 AM Ilia Mirkin wrote:
>>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>>> ---
>>>>  framework/core.py | 20 ++++++++++++++------
>>>>  1 file changed, 14 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/framework/core.py b/framework/core.py
>>>> index 4bcaa82..7b2083b 100644
>>>> --- a/framework/core.py
>>>> +++ b/framework/core.py
>>>> @@ -31,6 +31,7 @@ import sys
>>>>  import time
>>>>  import traceback
>>>>  from cStringIO import StringIO
>>>> +import itertools
>>>>  import multiprocessing
>>>>  import multiprocessing.dummy
>>>>  import importlib
>>>> @@ -578,22 +579,29 @@ class TestProfile(object):
>>>>          chunksize = 1
>>>>
>>>>          if env.concurrent == "all":
>>>> -            multi.imap(test, self.test_list.iteritems(), chunksize)
>>>> +            testiter = multi.imap(test, self.test_list.iteritems(),
>>>> chunksize) elif env.concurrent == "none":
>>>> -            single.imap(test, self.test_list.iteritems(), chunksize)
>>>> +            testiter = single.imap(test, self.test_list.iteritems(),
>>>> chunksize) else:
>>>>              # Filter and return only thread safe tests to the threaded pool
>>>> -            multi.imap(test, (x for x in self.test_list.iteritems() if -
>>>>                            x[1].runConcurrent), chunksize)
>>>> +            iter1 = multi.imap(test, (x for x in self.test_list.iteritems()
>>>> +                                      if x[1].runConcurrent), chunksize) #
>>>> Filter and return the non thread safe tests to the single pool -
>>>> single.imap(test, (x for x in self.test_list.iteritems() if not -
>>>>                     x[1].runConcurrent), chunksize)
>>>> +            iter2 = single.imap(test, (x for x in
>>>> self.test_list.iteritems() +                                       if not
>>>> x[1].runConcurrent), chunksize) +            testiter =
>>>> itertools.chain(iter1, iter2)
>>>>
>>>>          # Close and join the pools
>>>>          # If we don't close and the join the pools the script will exit
>>>> before # the pools finish running
>>>>          multi.close()
>>>>          single.close()
>>>> +
>>>> +        # This waits for all the results to arrive in a non-blocking
>>>> +        # fashion. This in turn allows ^C to work to stop a piglit run.
>>>> +        for test in testiter:
>>>> +            pass
>>>> +
>>>>          multi.join()
>>>>          single.join()
>>>
>>> The code is valid and works as advertised. I actually like what's happening
>>> now since ctrl-c kills the currently running test, and can be used to kill
>>> stuck tests, but if other like the previous behavior better, it isn't a big
>>> deal to me either way.
>>>
>>> Reviewed-by: Dylan Baker <baker.dylan.c at gmail.com>
>>
>> Hmmmm... I hadn't thought about that. It's just really hard to kill
>> piglit right now -- you have to ^Z + kill -9 %%. How does using ^C in
>> a parallel run (which I understand is the majority use-case) possible
>> to kill the 'current' test -- there is no current test.
>>
>> Anyone else with opinions on what ^C should do while running piglit?
>>
>>   -ilia
>
> In the past, people have objected to Ctrl+C killing whole runs.  That
> might be different if resume support was more reliable...I don't know.
>
> However, with concurrency, it's really not obvious what it does.  Some
> test gets victimized...or sometimes the runner.  It's been discussed,
> and there were no good answers.
>
> Ctrl+\ sends SIGQUIT and that always kills the whole thing.  It's just
> not as well known.

Ah, yes, I did not know that.

>
> I suspect if you change the behavior of Ctrl+C people will come
> complain.  I don't know that it's justified, really, but...

OK. I had assumed that it was accidental that ^C didn't work. If it's
well-thought-out (or at least well-understood/expected) behaviour,
happy to drop the patch.

  -ilia