On 03/04/2019 08:25, Guillaume Tucker wrote:
> On 02/04/2019 09:35, Petri Latvala wrote:
>> On Mon, Apr 01, 2019 at 04:39:24PM +0200, Guillaume Tucker wrote:
>>> The MIPS architecture doesn't provide the hardware atomics that are
>>> required for the "create-clear" sub-test such as
>>> __sync_add_and_fetch(). As a simple and pragmatic solution, disable
>>> this sub-test when building for MIPS. A better approach would be to
>>> add a fallback implementation for these operations.
>>>
>>> Fixes: 6727e17c00b2 ("i915/gem_create: Verify that all new objects are clear")
>>> Signed-off-by: Guillaume Tucker <guillaume.tucker@xxxxxxxxxxxxx>
>>> ---
>>> tests/i915/gem_create.c | 4 ++++
>>> 1 file changed, 4 insertions(+)
>>>
>>> diff --git a/tests/i915/gem_create.c b/tests/i915/gem_create.c
>>> index 2a861ca8a7ec..8a48496e6c19 100644
>>> --- a/tests/i915/gem_create.c
>>> +++ b/tests/i915/gem_create.c
>>> @@ -142,6 +142,7 @@ static void invalid_nonaligned_size(int fd)
>>> gem_close(fd, handle);
>>> }
>>>
>>> +#if !defined(__mips__) /* MIPS doesn't provide the required hardware atomics */
>>> static uint64_t get_npages(uint64_t *global, uint64_t npages)
>>> {
>>> uint64_t try, old, max;
>>> @@ -208,6 +209,7 @@ static void always_clear(int i915, int timeout)
>>> for (int i = 0; i < ncpus; i++)
>>> pthread_join(thread[i], NULL);
>>> }
>>> +#endif /* !defined(__mips__) */
>>>
>>> igt_main
>>> {
>>> @@ -231,6 +233,8 @@ igt_main
>>> igt_subtest("create-invalid-nonaligned")
>>> invalid_nonaligned_size(fd);
>>>
>>> +#if !defined(__mips__)
>>> igt_subtest("create-clear")
>>> always_clear(fd, 30);
>>> +#endif
>>> }
>>
>>
>> It's a bit ugly. I wonder how much work a fallback mechanism would be?
>
> Sorry I should have sent this as RFC.
>
>> The test is i915 specific and using those on non-x86 architectures
>> sounds silly. We could limit building tests/i915/* only if
>> host_machine.cpu_family() is x86 or x86_64. But that requires
>> revisiting this issue if ever the day comes when i915 can be used on
>> other architectures *cough*.
>
> Right, I thought it might not be worth implementing fallback
> functions if there is no MIPS hardware on which this test can be
> run. Still it would be a shame to leave i-g-t failing to build
> on MIPS.
>
>> Apropos, compile-testing on MIPS in gitlab-CI?
>
> This issue was actually hit while building the KernelCI root file
> system with i-g-t tests. We're starting to add some MIPS
> platforms, so running the generic DRM/KMS tests on them might
> start to make sense at some point (cc khilman).
>
> And yes I guess it seems worth considering adding MIPS to
> Gitlab-CI as it only appears to be covering x86, armhf and
> arm64 (although I did fix an armhf build issue a few weeks ago).
>
>> A compile-tested-only fallback mechanism suggestion, and a bad spot
>> for placing the fallback functions:
>
> Thanks, I agree that does look like a sustainable way forward.
> We don't quite have a MIPS platform to test that yet in KernelCI,
> so hopefully QEMU can be used to test a fallback implementation.
>
> I guess adding placeholder functions as in your example with
> igt_assert_f() statements would just add some technical debt with
> little added benefit, so I'll work on a v2 with something that
> works. Meanwhile we'll just skip i-g-t KernelCI builds on MIPS.
>
>> diff --git a/meson.build b/meson.build
>> index 557400a5..0552e858 100644
>> --- a/meson.build
>> +++ b/meson.build
>> @@ -246,6 +246,9 @@ endif
>> have = cc.has_function('memfd_create', prefix : '''#include <sys/mman.h>''', args : '-D_GNU_SOURCE')
>> config.set10('HAVE_MEMFD_CREATE', have)
>>
>> +have_atomics = cc.compiles('void f() { int x, y; __sync_add_and_fetch(&x, y); }')
>> +config.set10('HAVE_BUILTIN_ATOMICS', have_atomics)
>> +
>> add_project_arguments('-D_GNU_SOURCE', language : 'c')
>> add_project_arguments('-include', 'config.h', language : 'c')
>>
>> diff --git a/tests/i915/gem_create.c b/tests/i915/gem_create.c
>> index 2a861ca8..615bb475 100644
>> --- a/tests/i915/gem_create.c
>> +++ b/tests/i915/gem_create.c
>> @@ -62,6 +62,18 @@ IGT_TEST_DESCRIPTION("This is a test for the extended & old gem_create ioctl,"
>> " that includes allocation of object from stolen memory"
>> " and shmem.");
>>
>> +#if !HAVE_BUILTIN_ATOMICS
>> +int __sync_add_and_fetch(void *ptr, uint64_t val)
>> +{
>> + igt_assert_f(false, "Don't have builtin atomics\n");
>> +}
>> +
>> +int __sync_val_compare_and_swap(void *ptr, uint64_t old, uint64_t new)
>> +{
>> + igt_assert_f(false, "Don't have builtin atomics\n");
>> +}
>> +#endif
>> +
>> #define CLEAR(s) memset(&s, 0, sizeof(s))
>> #define PAGE_SIZE 4096
Just to close this thread, I sent some patches earlier this week
to fix the build on MIPS using the more recent __atomic_*
functions from libatomic as they do have the 64-bit
implementations for MIPS that were missing for the __sync_*
functions (tested with gcc 8.3). I've also just sent a patch to
enable Gitlab CI builds for MIPS, it's all passing:
https://gitlab.freedesktop.org/gtucker/igt-gpu-tools/pipelines/40827
Guillaume
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
