[PATCH, libgomp] Fix deadlock in acc_set_device_type (ping x2)

[Chung-Lin Tang]

Ping x2.

Hi Jakub,
This patch is fairly straightforward, and solves a easily encountered
deadlock. Please approve for trunk and gcc-6-branch.

Thanks,
Chung-Lin

On 2016/4/16 03:39 PM, Chung-Lin Tang wrote:
> Ping.
> 
> On 2016/3/28 05:45 PM, Chung-Lin Tang wrote:
>> Hi Jakub, there's a path for deadlock on acc_device_lock when going
>> through the acc_set_device_type() OpenACC library function.
>> Basically, the gomp_init_targets_once() function should not be
>> called with that held. The attached patch moves it appropriately.
>>
>> Also in this patch, there are several cases in acc_* functions
>> where gomp_init_targets_once() is guarded by a test of
>> !cached_base_dev. Since that function already uses pthread_once() to
>> call gomp_target_init(), and technically cached_base_dev
>> is protected by acc_device_lock, the cleanest way should be to
>> simply drop those "if(!cached_base_dev)" tests.
>>
>> Tested libgomp without regressions on an nvptx offloaded system,
>> is this okay for trunk?
>>
>> Thanks,
>> Chung-Lin
>>
>> 2016-03-28  Chung-Lin Tang  <cltang@codesourcery.com>
>>
>>         * oacc-init.c (acc_init): Remove !cached_base_dev condition on call to
>>         gomp_init_targets_once().
>>         (acc_set_device_type): Remove !cached_base_dev condition on call to
>>         gomp_init_targets_once(), move call to before acc_device_lock acquire,
>>         to avoid deadlock.
>>         (acc_get_device_num): Remove !cached_base_dev condition on call to
>>         gomp_init_targets_once().
>>         (acc_set_device_num): Likewise.
>>
>