[Intel-gfx] [PATCH 1/2] drm/vgem: Do not allocate backing shmemfs file for an import dmabuf object

[Daniel Vetter]

On Wed, Jul 8, 2020 at 4:37 PM Christian König <christian.koenig at amd.com> wrote:
>
> Am 08.07.20 um 11:54 schrieb Daniel Vetter:
> > On Wed, Jul 08, 2020 at 11:22:00AM +0200, Christian König wrote:
> >> Am 07.07.20 um 20:35 schrieb Chris Wilson:
> >>> Quoting lepton (2020-07-07 19:17:51)
> >>>> On Tue, Jul 7, 2020 at 10:20 AM Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >>>>> Quoting lepton (2020-07-07 18:05:21)
> >>>>>> On Tue, Jul 7, 2020 at 9:00 AM Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >>>>>>> If we assign obj->filp, we believe that the create vgem bo is native and
> >>>>>>> allow direct operations like mmap() assuming it behaves as backed by a
> >>>>>>> shmemfs inode. When imported from a dmabuf, the obj->pages are
> >>>>>>> not always meaningful and the shmemfs backing store misleading.
> >>>>>>>
> >>>>>>> Note, that regular mmap access to a vgem bo is via the dumb buffer API,
> >>>>>>> and that rejects attempts to mmap an imported dmabuf,
> >>>>>> What do you mean by "regular mmap access" here?  It looks like vgem is
> >>>>>> using vgem_gem_dumb_map as .dumb_map_offset callback then it doesn't call
> >>>>>> drm_gem_dumb_map_offset
> >>>>> As I too found out, and so had to correct my story telling.
> >>>>>
> >>>>> By regular mmap() access I mean mmap on the vgem bo [via the dumb buffer
> >>>>> API] as opposed to mmap() via an exported dma-buf fd. I had to look at
> >>>>> igt to see how it was being used.
> >>>> Now it seems your fix is to disable "regular mmap" on imported dma buf
> >>>> for vgem. I am not really a graphic guy, but then the api looks like:
> >>>> for a gem handle, user space has to guess to find out the way to mmap
> >>>> it. If user space guess wrong, then it will fail to mmap. Is this the
> >>>> expected way
> >>>> for people to handle gpu buffer?
> >>> You either have a dumb buffer handle, or a dma-buf fd. If you have the
> >>> handle, you have to use the dumb buffer API, there's no other way to
> >>> mmap it. If you have the dma-buf fd, you should mmap it directly. Those
> >>> two are clear.
> >>>
> >>> It's when you import the dma-buf into vgem and create a handle out of
> >>> it, that's when the handle is no longer first class and certain uAPI
> >>> [the dumb buffer API in particular] fail.
> >>>
> >>> It's not brilliant, as you say, it requires the user to remember the
> >>> difference between the handles, but at the same time it does prevent
> >>> them falling into coherency traps by forcing them to use the right
> >>> driver to handle the object, and have to consider the additional ioctls
> >>> that go along with that access.
> >> Yes, Chris is right. Mapping DMA-buf through the mmap() APIs of an importer
> >> is illegal.
> >>
> >> What we could maybe try to do is to redirect this mmap() API call on the
> >> importer to the exporter, but I'm pretty sure that the fs layer wouldn't
> >> like that without changes.
> > We already do that, there's a full helper-ified path from I think shmem
> > helpers through prime helpers to forward this all. Including handling
> > buffer offsets and all the other lolz back&forth.
>
> Oh, that most likely won't work correctly with unpinned DMA-bufs and
> needs to be avoided.
>
> Each file descriptor is associated with an struct address_space. And
> when you mmap() through the importer by redirecting the system call to
> the exporter you end up with the wrong struct address_space in your VMA.
>
> That in turn can go up easily in flames when the exporter tries to
> invalidate the CPU mappings for its DMA-buf while moving it.
>
> Where are we doing this? My last status was that this is forbidden.

Hm I thought we're doing all that already, but looking at the code
again we're only doing this when opening a new drm fd or dma-buf fd.
So the right file->f_mapping is always set at file creation time.

And we indeed don't frob this more when going another indirection ...
Maybe we screwed up something somewhere :-/

Also I thought the mapping is only taken after the vma is instatiated,
otherwise the tricks we're playing with dma-buf already wouldn't work:
dma-buf has the buffer always at offset 0, whereas gem drm_fd mmap has
it somewhere else. We already adjust vma->vm_pgoff, so I'm wondering
whether we could adjust vm_file too. Or is that the thing that's
forbidden?
-Daniel

> Christian.
>
> > Of course there's still the problem that many drivers don't forward the
> > cache coherency calls for begin/end cpu access, so in a bunch of cases
> > you'll cache cacheline dirt soup. But that's kinda standard procedure for
> > dma-buf :-P
> >
> > But yeah trying to handle the mmap as an importer, bypassing the export:
> > nope. The one exception is if you have some kind of fancy gart with
> > cpu-visible pci bar (like at least integrated intel gpus have). But in
> > that case the mmap very much looks&acts like device access in every way.
> >
> > Cheers, Daniel
> >
> >> Regards,
> >> Christian.
> >>
> >>
> >>> -Chris
>


-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch