[PATCH 07/10] drm/tests: Add test for drm_framebuffer_init()

Maira Canal mairacanal at riseup.net
Fri Sep 8 20:22:32 UTC 2023


Hi Carlos,

On 9/4/23 14:41, Carlos wrote:
> Hi Maíra,
> 
> On 8/26/23 11:16, Maíra Canal wrote:
>> Hi Carlos,
>>
>> On 8/25/23 13:11, Carlos Eduardo Gallo Filho wrote:
>>> Add a single KUnit test case for the drm_framebuffer_init function.
>>>
>>> Signed-off-by: Carlos Eduardo Gallo Filho <gcarlos at disroot.org>
>>> ---
>>>   drivers/gpu/drm/tests/drm_framebuffer_test.c | 52 ++++++++++++++++++++
>>>   1 file changed, 52 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/tests/drm_framebuffer_test.c 
>>> b/drivers/gpu/drm/tests/drm_framebuffer_test.c
>>> index 3d14d35b4c4d..50d88bf3fa65 100644
>>> --- a/drivers/gpu/drm/tests/drm_framebuffer_test.c
>>> +++ b/drivers/gpu/drm/tests/drm_framebuffer_test.c
>>> @@ -557,8 +557,60 @@ static void drm_test_framebuffer_lookup(struct 
>>> kunit *test)
>>>       KUNIT_EXPECT_NULL(test, fb2);
>>>   }
>>>   +static void drm_test_framebuffer_init(struct kunit *test)
>>> +{
>>> +    struct drm_mock *mock = test->priv;
>>> +    struct drm_device *dev = &mock->dev;
>>> +    struct drm_device wrong_drm = { };
>>> +    struct drm_format_info format = { };
>>> +    struct drm_framebuffer fb1 = { .dev = dev, .format = &format };
>>> +    struct drm_framebuffer *fb2;
>>> +    struct drm_framebuffer_funcs funcs = { };
>>> +    int ret;
>>> +
>>> +    /* Fails if fb->dev doesn't point to the drm_device passed on 
>>> first arg */
>>> +    fb1.dev = &wrong_drm;
>>> +    ret = drm_framebuffer_init(dev, &fb1, &funcs);
>>> +    KUNIT_EXPECT_EQ(test, ret, -EINVAL);
>>> +    fb1.dev = dev;
>>> +
>>> +    /* Fails if fb.format isn't set */
>>> +    fb1.format = NULL;
>>> +    ret = drm_framebuffer_init(dev, &fb1, &funcs);
>>> +    KUNIT_EXPECT_EQ(test, ret, -EINVAL);
>>> +    fb1.format = &format;
>>> +
>>> +    ret = drm_framebuffer_init(dev, &fb1, &funcs);
>>> +    KUNIT_EXPECT_EQ(test, ret, 0);
>>> +
>>> +    /*
>>> +     * Check if fb->funcs is actually set to the drm_framebuffer_funcs
>>> +     * passed to it
>>> +     */
>>> +    KUNIT_EXPECT_PTR_EQ(test, fb1.funcs, &funcs);
>>> +
>>> +    /* The fb->comm must be set to the current running process */
>>> +    KUNIT_EXPECT_STREQ(test, fb1.comm, current->comm);
>>> +
>>> +    /* The fb->base must be successfully initialized */
>>> +    KUNIT_EXPECT_EQ(test, fb1.base.id, 1);
>>> +    KUNIT_EXPECT_EQ(test, fb1.base.type, DRM_MODE_OBJECT_FB);
>>> +    KUNIT_EXPECT_EQ(test, kref_read(&fb1.base.refcount), 1);
>>> +    KUNIT_EXPECT_PTR_EQ(test, fb1.base.free_cb, &drm_framebuffer_free);

BTW I believe we should also make sure that dev->mode_config.num_fb was
incremented by 1.

>>> +
>>> +    /* Checks if the fb is really published and findable */
>>> +    fb2 = drm_framebuffer_lookup(dev, NULL, fb1.base.id);
>>> +    KUNIT_EXPECT_PTR_EQ(test, fb2, &fb1);
>>> +
>>> +    /* There must be just that one fb initialized */
>>> +    KUNIT_EXPECT_EQ(test, dev->mode_config.num_fb, 1);
>>> +    KUNIT_EXPECT_PTR_EQ(test, dev->mode_config.fb_list.prev, 
>>> &fb1.head);
>>> +    KUNIT_EXPECT_PTR_EQ(test, dev->mode_config.fb_list.next, 
>>> &fb1.head);
>>
>> Shouldn't we clean the framebuffer object?
> What did you mean by "clean"? Firstly I supposed that it would be about
> freeing some dynamically allocated frambuffer, but it's statically
> allocated, so I believe it isn't what you are meaning. Is there some
> collateral effect I'm not taking into account?

I was talking about calling the function `drm_framebuffer_cleanup()`.

Best Regards,
- Maíra

> 
> Thanks,
> Carlos
> 
>> Best Regards,
>> - Maíra
>>
>>> +}
>>> +
>>>   static struct kunit_case drm_framebuffer_tests[] = {
>>>       KUNIT_CASE(drm_test_framebuffer_cleanup),
>>> +    KUNIT_CASE(drm_test_framebuffer_init),
>>>       KUNIT_CASE(drm_test_framebuffer_lookup),
>>> KUNIT_CASE(drm_test_framebuffer_modifiers_not_supported),
>>>       KUNIT_CASE_PARAM(drm_test_framebuffer_check_src_coords, 
>>> check_src_coords_gen_params),


More information about the dri-devel mailing list