[lkml]   [2022]   [Aug]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v8 2/2] drm/gem: Don't map imported GEMs
On 8/12/22 18:01, Rob Clark wrote:
> On Fri, Aug 12, 2022 at 7:57 AM Rob Clark <> wrote:
>> On Fri, Aug 12, 2022 at 4:26 AM Dmitry Osipenko
>> <> wrote:
>>> On 8/11/22 02:19, Rob Clark wrote:
>>>> On Wed, Aug 10, 2022 at 3:23 PM Dmitry Osipenko
>>>> <> wrote:
>>>>> On 8/11/22 01:03, Rob Clark wrote:
>>>>>> On Wed, Aug 10, 2022 at 12:26 PM Dmitry Osipenko
>>>>>> <> wrote:
>>>>>>> On 8/10/22 18:08, Rob Clark wrote:
>>>>>>>> On Wed, Aug 10, 2022 at 4:47 AM Daniel Vetter <> wrote:
>>>>>>>>> On Wed, Jul 06, 2022 at 10:02:07AM +0300, Dmitry Osipenko wrote:
>>>>>>>>>> On 7/6/22 00:48, Rob Clark wrote:
>>>>>>>>>>> On Tue, Jul 5, 2022 at 4:51 AM Christian König <> wrote:
>>>>>>>>>>>> Am 01.07.22 um 11:02 schrieb Dmitry Osipenko:
>>>>>>>>>>>>> Drivers that use drm_gem_mmap() and drm_gem_mmap_obj() helpers don't
>>>>>>>>>>>>> handle imported dma-bufs properly, which results in mapping of something
>>>>>>>>>>>>> else than the imported dma-buf. On NVIDIA Tegra we get a hard lockup when
>>>>>>>>>>>>> userspace writes to the memory mapping of a dma-buf that was imported into
>>>>>>>>>>>>> Tegra's DRM GEM.
>>>>>>>>>>>>> Majority of DRM drivers prohibit mapping of the imported GEM objects.
>>>>>>>>>>>>> Mapping of imported GEMs require special care from userspace since it
>>>>>>>>>>>>> should sync dma-buf because mapping coherency of the exporter device may
>>>>>>>>>>>>> not match the DRM device. Let's prohibit the mapping for all DRM drivers
>>>>>>>>>>>>> for consistency.
>>>>>>>>>>>>> Suggested-by: Thomas Hellström <>
>>>>>>>>>>>>> Signed-off-by: Dmitry Osipenko <>
>>>>>>>>>>>> I'm pretty sure that this is the right approach, but it's certainly more
>>>>>>>>>>>> than possible that somebody abused this already.
>>>>>>>>>>> I suspect that this is abused if you run deqp cts on android.. ie. all
>>>>>>>>>>> winsys buffers are dma-buf imports from gralloc. And then when you
>>>>>>>>>>> hit readpix...
>>>>>>>>>>> You might only hit this in scenarios with separate gpu and display (or
>>>>>>>>>>> dGPU+iGPU) because self-imports are handled differently in
>>>>>>>>>>> drm_gem_prime_import_dev().. and maybe not in cases where you end up
>>>>>>>>>>> with a blit from tiled/compressed to linear.. maybe that narrows the
>>>>>>>>>>> scope enough to just fix it in userspace?
>>>>>>>>>> Given that that only drivers which use DRM-SHMEM potentially could've
>>>>>>>>>> map imported dma-bufs (Panfrost, Lima) and they already don't allow to
>>>>>>>>>> do that, I think we're good.
>>>>>>>>> So can I have an ack from Rob here or are there still questions that this
>>>>>>>>> might go boom?
>>>>>>>>> Dmitry, since you have a bunch of patches merged now I think would also be
>>>>>>>>> good to get commit rights so you can drive this more yourself. I've asked
>>>>>>>>> Daniel Stone to help you out with getting that.
>>>>>>>> I *think* we'd be ok with this on msm, mostly just by dumb luck.
>>>>>>>> Because the dma-buf's we import will be self-import. I'm less sure
>>>>>>>> about panfrost (src/panfrost/lib/pan_bo.c doesn't seem to have a
>>>>>>>> special path for imported dma-bufs either, and in that case they won't
>>>>>>>> be self-imports.. but I guess no one has tried to run android cts on
>>>>>>>> panfrost).
>>>>>>> The last time I tried to mmap dma-buf imported to Panfrost didn't work
>>>>>>> because Panfrost didn't implement something needed for that. I'll need
>>>>>>> to take a look again because can't recall what it was.
>>> Upd: I re-checked Panfrost using today's linux-next and mapping of
>>> imported dma-buf works, I mapped imported buf from video decoder.
>>> Apparently previously I had some local kernel change that broke the mapping.
>>>>>>>> What about something less drastic to start, like (apologies for
>>>>>>>> hand-edited patch):
>>>>>>>> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
>>>>>>>> index 86d670c71286..fc9ec42fa0ab 100644
>>>>>>>> --- a/drivers/gpu/drm/drm_gem.c
>>>>>>>> +++ b/drivers/gpu/drm/drm_gem.c
>>>>>>>> @@ -1034,6 +1034,10 @@ int drm_gem_mmap_obj(struct drm_gem_object
>>>>>>>> *obj, unsigned long obj_size,
>>>>>>>> {
>>>>>>>> int ret;
>>>>>>>> + WARN_ON_ONCE(obj->import_attach);
>>>>>>> This will hang NVIDIA Tegra, which is what this patch fixed initially.
>>>>>>> If neither of upstream DRM drivers need to map imported dma-bufs and
>>>>>>> never needed, then why do we need this?
>>>>>> oh, tegra isn't using shmem helpers? I assumed it was. Well my point
>>>>>> was to make a more targeted fail on tegra, and a WARN_ON for everyone
>>>>>> else to make it clear that what they are doing is undefined behavior.
>>>>>> Because so far existing userspace (or well, panfrost and freedreno at
>>>>>> least, those are the two I know or checked) don't make special cases
>>>>>> for mmap'ing against the dmabuf fd against the dmabuf fd instead of
>>>>>> the drm device fd.
>>>>> It's not clear to me what bad Android does form yours comments. Does it
>>>>> export dma-buf from GPU and then import it to GPU? If yes, then DRM core
>>>>> has a check for the self-importing [1].
>>>>> [1]
>>>>> If you're meaning something else, then please explain in a more details.
>>>> So, android/gralloc allocates buffers externally to the driver and
>>>> imports them into driver. (And that seems to not just be window
>>>> surfaces, but in cases random textures, etc)
>>>> In the normal case these should be allocated from drm/msm so it should
>>>> hit [1].. this is the "dumb luck" I mentioned earlier. But I'm not
>>>> confident enough to say that there is no other case.
>>>>>> I *think* it should work out that we don't hit this path with
>>>>>> freedreno but on android I can't really guarantee or prove it. So
>>>>>> your patch would potentially break existing working userspace. Maybe
>>>>>> it is userspace that isn't portable (but OTOH it isn't like you are
>>>>>> going to be using freedreno on tegra). So why don't you go for a more
>>>>>> targeted fix that only returns an error on hw where this is
>>>>>> problematic?
>>>>> That's what the first versions of the patch did and Christian suggested
>>>>> that it's not a good approach. In fact it should be not only Tegra that
>>>>> has a broken dma-buf mapping, but apparently OMAP driver too.
>>>> Hmm, I guess I'm a bit more conservative when it comes to potentially
>>>> breaking userspace.
>>> If such userspace exists, then of course the mapping should continue to
>>> work. Still will be great to know what that userpsace is.
>> Definitely existing mesa does not have a special mmap path for
>> imported dma-bufs (both in the case of panfrost and freedreno, I
>> didn't check any others). The only question is whether there is a
>> case where some app/test/etc imports a foreign dma-buf fd and then
>> does something that would trigger mmap'ing, like readpix.
> The other complication I noticed is that we don't seem to keep around
> the fd after importing to a GEM handle. And I could imagine that
> doing so could cause issues with too many fd's. So I guess the best
> thing is to keep the status quo and let drivers that cannot mmap
> imported buffers just fail mmap?

That actually should be all the drivers excluding those that use
DRM-SHMEM because only DRM-SHMEM uses dma_buf_mmap(), that's why it
works for Panfrost. I'm pretty sure mmaping of imported GEMs doesn't
work for the MSM driver, isn't it?

Intel and AMD drivers don't allow to map the imported dma-bufs. Both
refuse to do the mapping.

Although, AMDGPU "succeeds" to do the mapping using
AMDGPU_GEM_DOMAIN_GTT, but then touching the mapping causes bus fault,
hence mapping actually fails. I think it might be the AMDGPU
driver/libdrm bug, haven't checked yet.

So we're back to the point that neither of DRM drivers need to map
imported dma-bufs and this was never tested. In this case this patch is
valid, IMO.

Best regards,

 \ /
  Last update: 2022-08-16 14:04    [W:0.103 / U:2.396 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site