[PATCH 3/5] drm/udl: Add GEM object functions for free(), vmap(), and vunmap()

[Noralf Trønnes]

Den 25.10.2019 12.12, skrev Thomas Zimmermann:
> Hi
> 
> Am 25.10.19 um 11:28 schrieb Daniel Vetter:
>> On Fri, Oct 25, 2019 at 9:59 AM Thomas Zimmermann <tzimmermann at suse.de> wrote
>>>
>>> Hi
>>>
>>> Am 25.10.19 um 09:40 schrieb Daniel Vetter:
>>>> On Thu, Oct 24, 2019 at 04:42:35PM +0200, Thomas Zimmermann wrote:
>>>>> Implementing vmap() and vunmap() of struct drm_gem_object_funcs is
>>>>> required by generic fbdev emulation. Supporting free() is easy as
>>>>> well. More udl-specific interfaces can probably be replaced by GEM
>>>>> object functions in later patches.
>>>>>
>>>>> Signed-off-by: Thomas Zimmermann <tzimmermann at suse.de>
>>>>> ---
>>>>>  drivers/gpu/drm/udl/udl_gem.c | 34 ++++++++++++++++++++++++++++++++++
>>>>>  1 file changed, 34 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/udl/udl_gem.c b/drivers/gpu/drm/udl/udl_gem.c
>>>>> index 3ea0cd9ae2d6..6ca097c270d6 100644
>>>>> --- a/drivers/gpu/drm/udl/udl_gem.c
>>>>> +++ b/drivers/gpu/drm/udl/udl_gem.c
>>>>> @@ -11,6 +11,8 @@
>>>>>
>>>>>  #include "udl_drv.h"
>>>>>
>>>>> +static const struct drm_gem_object_funcs udl_gem_object_funcs;
>>>>> +
>>>>>  struct udl_gem_object *udl_gem_alloc_object(struct drm_device *dev,
>>>>>                                          size_t size)
>>>>>  {
>>>>> @@ -25,6 +27,7 @@ struct udl_gem_object *udl_gem_alloc_object(struct drm_device *dev,
>>>>>              return NULL;
>>>>>      }
>>>>>
>>>>> +    obj->base.funcs = &udl_gem_object_funcs;
>>>>>      obj->flags = UDL_BO_CACHEABLE;
>>>>>      return obj;
>>>>>  }
>>>>> @@ -230,3 +233,34 @@ int udl_gem_mmap(struct drm_file *file, struct drm_device *dev,
>>>>>      mutex_unlock(&udl->gem_lock);
>>>>>      return ret;
>>>>>  }
>>>>> +
>>>>> +/*
>>>>> + * Helpers for struct drm_gem_object_funcs
>>>>> + */
>>>>> +
>>>>> +static void udl_gem_object_free(struct drm_gem_object *obj)
>>>>> +{
>>>>> +    udl_gem_free_object(obj);
>>>>> +}
>>>>> +
>>>>> +static void *udl_gem_object_vmap(struct drm_gem_object *obj)
>>>>> +{
>>>>> +    struct udl_gem_object *uobj = to_udl_bo(obj);
>>>>> +    int ret;
>>>>> +
>>>>> +    ret = udl_gem_vmap(uobj);
>>>>> +    if (ret)
>>>>> +            return ERR_PTR(ret);
>>>>> +    return uobj->vmapping;
>>>>> +}
>>>>> +
>>>>> +static void udl_gem_object_vunmap(struct drm_gem_object *obj, void *vaddr)
>>>>> +{
>>>>> +    return udl_gem_vunmap(to_udl_bo(obj));
>>>>> +}
>>>>> +
>>>>> +static const struct drm_gem_object_funcs udl_gem_object_funcs = {
>>>>> +    .free   = udl_gem_object_free,
>>>>> +    .vmap   = udl_gem_object_vmap,
>>>>> +    .vunmap = udl_gem_object_vunmap,
>>>>
>>>> Yeah this doesn't work, you need to refcount the vmap here I think. I'd
>>>
>>> I see. Should have thought of that...
>>>
>>>> say simpler to first cut over to shmem helpers.
>>>
>>> Right. I was already attempting the conversion and the udl gem is more
>>> or less the same as shmem, except for the flags field. [1] The flag
>>> signals page attributes for mmap-ing [2]. For prime-imported BOs its is
>>> set to writecombining, and for local BOs it is set to cached. Shmem
>>> always maps with writecombining.
>>>
>>> I don't see why udl BOs are special wrt to mmap-ing. It seems to be some
>>> kind of optimization. Do you have an idea?
>>>
>>> Best regards
>>> Thomas
>>>
>>> [1]
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/udl/udl_drv.h?h=v5.3#n78
>>> [2]
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/udl/udl_gem.c?h=v5.3#n57
>>
>> udl doesn't set prefer_shadow = 1, which means compositors will render
>> directly into the buffer. And for that you want caching. For imported
>> dma-buf otoh you want wc to make sure it's coherent. Otherwise I guess
>> udl would need to get the dma_buf_begin/end_cpu_access calls added
>> (and that would probably be faster than going with wc for everything).
>> So we do want cachable, it's going to suck otherwise.
> 
> Thanks for clarifying. Actually, it does have these calls in its dirty
> handler. [1]
> 
>>
>> But that should be fairly easy to do by overwriting the obj->mmap
>> callback and wrapping it around the shmem one.
> 
> Is there a reason why shmem doesn't implement the wc-vs.cached logic?
> Could this be added?
> 

I had a flag to set this in the initial version of the shmem helper
modeled after udl, but Thomas Hellstrom brought up a question and it was
dropped. The issue was beyond my understanding:

[PATCH v3 0/2] drm: Add shmem GEM library
https://lists.freedesktop.org/archives/dri-devel/2018-September/190001.html

In tinydrm I have sped up uncached writes on arm by copying one pixel
line at a time into a temporary buffer before accessing the individual
bytes. See drm_fb_swab16().

Noralf.

> (I'm just trying to understand the available options here before
> attempting to do a conversion.)
> 
> Best regards
> Thomas
> 
>>
>> What surprises me more is that udl supports mmap on imported dma-buf,
>> that's some real quirky stuff. Not sure there's really other drivers
>> doing that. Might be easier to rip that out :-)
>> -Daniel
>>
> 
> [1]
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/udl/udl_fb.c?h=v5.3#n293
> 
> 
> 
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>