[Intel-gfx] [PATCH 40/59] drm/vram-helper: Drop drm_gem_prime_export/import

[Daniel Vetter]

On Thu, Jun 27, 2019 at 10:27 AM Thomas Zimmermann <tzimmermann at suse.de> wrote:
>
> Hi
>
> Am 17.06.19 um 15:59 schrieb Daniel Vetter:
> > On Mon, Jun 17, 2019 at 10:24:38AM +0200, Gerd Hoffmann wrote:
> >>   Hi,
> >>
> >>> Aside: Would be really nice to switch the others over to
> >>> drm_gem_object_funcs.
> >>
> >> While most callbacks are pretty straight forward (just hook the same
> >> callbacks into the drm_gem_object_funcs. struct) the mmap bits are a
> >> bit more obscure.
> >>
> >> First, there seem to be two ways to mmap a gem buffer:
> >>
> >>   (1) drm_driver->fops->mmap, and
> >>   (2) drm_driver->gem_prime_mmap.
> >>
> >> drm_gem_object_funcs has just a single vm_ops ...
> >>
> >> Also it is not obvious how one would convert something which basically
> >> calls ttm_bo_mmap() in drm_driver->fops->mmap to the new interface.
> >
> > Yeah the mmap side is still a mess, but my series here was getting a bit
> > too long already. There's a bunch of problems here:
> >
> > drm_driver->gem_prime_mmap could be nuked and instead we use
> > drm_gem_prime_mmap everywhere. Especially the various versions in helpers
> > really don't add much.
> >
> > The trouble is that removing the hook outright isn't quite right, because
> > it also signals "is mmap allowed on this dma-buf". I'm kinda tempted to
> > just make that a hard requirement, and force people who can't support mmap
> > on the dma-buf (or who need begin/end_cpu_access hooks) to supply their
> > own set of dma_buf_ops.
> >
> > Second one is drm_driver->fops->mmap. That one we need to keep, but this
> > isn't mmap on a buffer, but mmap on the entire drm_device. The one which
> > should be replaced by drm_gem_object_funcs.vm_ops is
> > drm_driver->gem_vm_ops.
>
> From what I've seen in fbdev drivers, it's an mmap of the framebuffer
> memory, which typically is the same as the device's memory but doesn't
> have to. And it's only valid for/while the current display mode (e.g.,
> mgafb doesn't set fixes.smem_length until you configure a mode).
>
> I guess it should be legal to just mmap the shadow FB from the fbcon
> emulation.

fbdev mmap is an entirely different beast ... but yeah maybe we should
rework that one too. Problem there is that it intereacts badly with
defio, so for at least some drivers you can't do the straightforward
thing at all.
-Daniel

>
> Best regards
> Thomas
>
> > Hope that explains a bit better what's going on here. Step one here for
> > mmap is definitely to roll out drm_gem_prime_mmap as far as possible, so
> > it's easier to understand where the exceptions are.
> >
> > Cheers, Daniel
> >
>
> --
> Thomas Zimmermann
> Graphics Driver Developer
> SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany
> GF: Felix Imendörffer, Mary Higgins, Sri Rasiah
> HRB 21284 (AG Nürnberg)
>


-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch