Re: [PATCH v2 1/8] drm: Add dummy page per device or GEM object

["Michel Dänzer" <michel@daenzer.net>]

On 2020-11-14 10:57 a.m., Daniel Vetter wrote:
> On Sat, Nov 14, 2020 at 10:51 AM Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
>>
>> On Sat, Nov 14, 2020 at 9:41 AM Christian König
>> <ckoenig.leichtzumerken@gmail.com> wrote:
>>>
>>> Am 13.11.20 um 21:52 schrieb Andrey Grodzovsky:
>>>>
>>>> On 6/22/20 1:50 PM, Daniel Vetter wrote:
>>>>> On Mon, Jun 22, 2020 at 7:45 PM Christian König
>>>>> <christian.koenig@amd.com> wrote:
>>>>>> Am 22.06.20 um 16:32 schrieb Andrey Grodzovsky:
>>>>>>> On 6/22/20 9:18 AM, Christian König wrote:
>>>>>>>> Am 21.06.20 um 08:03 schrieb Andrey Grodzovsky:
>>>>>>>>> Will be used to reroute CPU mapped BO's page faults once
>>>>>>>>> device is removed.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
>>>>>>>>> ---
>>>>>>>>>     drivers/gpu/drm/drm_file.c  |  8 ++++++++
>>>>>>>>>     drivers/gpu/drm/drm_prime.c | 10 ++++++++++
>>>>>>>>>     include/drm/drm_file.h      |  2 ++
>>>>>>>>>     include/drm/drm_gem.h       |  2 ++
>>>>>>>>>     4 files changed, 22 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c
>>>>>>>>> index c4c704e..67c0770 100644
>>>>>>>>> --- a/drivers/gpu/drm/drm_file.c
>>>>>>>>> +++ b/drivers/gpu/drm/drm_file.c
>>>>>>>>> @@ -188,6 +188,12 @@ struct drm_file *drm_file_alloc(struct
>>>>>>>>> drm_minor *minor)
>>>>>>>>>                 goto out_prime_destroy;
>>>>>>>>>         }
>>>>>>>>>     +    file->dummy_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
>>>>>>>>> +    if (!file->dummy_page) {
>>>>>>>>> +        ret = -ENOMEM;
>>>>>>>>> +        goto out_prime_destroy;
>>>>>>>>> +    }
>>>>>>>>> +
>>>>>>>>>         return file;
>>>>>>>>>       out_prime_destroy:
>>>>>>>>> @@ -284,6 +290,8 @@ void drm_file_free(struct drm_file *file)
>>>>>>>>>         if (dev->driver->postclose)
>>>>>>>>>             dev->driver->postclose(dev, file);
>>>>>>>>>     +    __free_page(file->dummy_page);
>>>>>>>>> +
>>>>>>>>> drm_prime_destroy_file_private(&file->prime);
>>>>>>>>>           WARN_ON(!list_empty(&file->event_list));
>>>>>>>>> diff --git a/drivers/gpu/drm/drm_prime.c
>>>>>>>>> b/drivers/gpu/drm/drm_prime.c
>>>>>>>>> index 1de2cde..c482e9c 100644
>>>>>>>>> --- a/drivers/gpu/drm/drm_prime.c
>>>>>>>>> +++ b/drivers/gpu/drm/drm_prime.c
>>>>>>>>> @@ -335,6 +335,13 @@ int drm_gem_prime_fd_to_handle(struct
>>>>>>>>> drm_device *dev,
>>>>>>>>>           ret = drm_prime_add_buf_handle(&file_priv->prime,
>>>>>>>>>                 dma_buf, *handle);
>>>>>>>>> +
>>>>>>>>> +    if (!ret) {
>>>>>>>>> +        obj->dummy_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
>>>>>>>>> +        if (!obj->dummy_page)
>>>>>>>>> +            ret = -ENOMEM;
>>>>>>>>> +    }
>>>>>>>>> +
>>>>>>>> While the per file case still looks acceptable this is a clear NAK
>>>>>>>> since it will massively increase the memory needed for a prime
>>>>>>>> exported object.
>>>>>>>>
>>>>>>>> I think that this is quite overkill in the first place and for the
>>>>>>>> hot unplug case we can just use the global dummy page as well.
>>>>>>>>
>>>>>>>> Christian.
>>>>>>>
>>>>>>> Global dummy page is good for read access, what do you do on write
>>>>>>> access ? My first approach was indeed to map at first global dummy
>>>>>>> page as read only and mark the vma->vm_flags as !VM_SHARED assuming
>>>>>>> that this would trigger Copy On Write flow in core mm
>>>>>>> (https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.bootlin.com%2Flinux%2Fv5.7-rc7%2Fsource%2Fmm%2Fmemory.c%23L3977&amp;data=02%7C01%7CAndrey.Grodzovsky%40amd.com%7C3753451d037544e7495408d816d4c4ee%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637284450384586120&amp;sdata=ZpRaQgqA5K4jRfidOiedey0AleeYQ97WNUkGA29ERA0%3D&amp;reserved=0)
>>>>>>>
>>>>>>> on the next page fault to same address triggered by a write access but
>>>>>>> then i realized a new COW page will be allocated for each such mapping
>>>>>>> and this is much more wasteful then having a dedicated page per GEM
>>>>>>> object.
>>>>>> Yeah, but this is only for a very very small corner cases. What we need
>>>>>> to prevent is increasing the memory usage during normal operation to
>>>>>> much.
>>>>>>
>>>>>> Using memory during the unplug is completely unproblematic because we
>>>>>> just released quite a bunch of it by releasing all those system memory
>>>>>> buffers.
>>>>>>
>>>>>> And I'm pretty sure that COWed pages are correctly accounted towards
>>>>>> the
>>>>>> used memory of a process.
>>>>>>
>>>>>> So I think if that approach works as intended and the COW pages are
>>>>>> released again on unmapping it would be the perfect solution to the
>>>>>> problem.
>>>>>>
>>>>>> Daniel what do you think?
>>>>> If COW works, sure sounds reasonable. And if we can make sure we
>>>>> managed to drop all the system allocations (otherwise suddenly 2x
>>>>> memory usage, worst case). But I have no idea whether we can
>>>>> retroshoehorn that into an established vma, you might have fun stuff
>>>>> like a mkwrite handler there (which I thought is the COW handler
>>>>> thing, but really no idea).
>>>>>
>>>>> If we need to massively change stuff then I think rw dummy page,
>>>>> allocated on first fault after hotunplug (maybe just make it one per
>>>>> object, that's simplest) seems like the much safer option. Much less
>>>>> code that can go wrong.
>>>>> -Daniel
>>>>
>>>>
>>>> Regarding COW, i was looking into how to properly implement it from
>>>> within the fault handler (i.e. ttm_bo_vm_fault)
>>>> and the main obstacle I hit is that of exclusive access to the
>>>> vm_area_struct, i need to be able to modify
>>>> vma->vm_flags (and vm_page_prot)  to remove VM_SHARED bit so COW can
>>>> be triggered on subsequent write access
>>>> fault (here
>>>> https://elixir.bootlin.com/linux/latest/source/mm/memory.c#L4128)
>>>> but core mm takes only read side mm_sem (here for example
>>>> https://elixir.bootlin.com/linux/latest/source/drivers/iommu/amd/iommu_v2.c#L488)
>>>> and so I am not supposed to modify vm_area_struct in this case. I am
>>>> not sure if it's legit to write lock tthe mm_sem from this point.
>>>> I found some discussions about this here
>>>> http://lkml.iu.edu/hypermail/linux/kernel/1909.1/02754.html but it
>>>> wasn't really clear to me
>>>> what's the solution.
>>>>
>>>> In any case, seems to me that easier and more memory saving solution
>>>> would be to just switch to per ttm bo dumy rw page that
>>>> would be allocated on demand as you suggested here.  This should also
>>>> take care of imported BOs and flink cases.
>>>> Then i can drop the per device FD and per GEM object FD dummy BO and
>>>> the ugly loop i am using in patch 2 to match faulting BO to the right
>>>> dummy page.
>>>>
>>>> Does this makes sense ?
>>>
>>> I still don't see the information leak as much of a problem, but if
>>> Daniel insists we should probably do this.
>>
>> Well amdgpu doesn't clear buffers by default, so indeed you guys are a
>> lot more laissez-faire here. But in general we really don't do that
>> kind of leaking. Iirc there's even radeonsi bugs because else clears,
>> and radeonsi happily displays gunk :-)
> 
> btw I think not clearing at alloc breaks the render node model a bit.
> Without that this was all fine, since system pages still got cleared
> by alloc_page(), and we only leaked vram. And for the legacy node
> model with authentication of clients against the X server, leaking
> that all around was ok. With render nodes no leaking should happen,
> with no knob for userspace to opt out of the forced clearing.

Seconded.


-- 
Earthling Michel Dänzer               |               https://redhat.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx