RFC: dma-fence: Document recoverable page fault implications

Commit Message

Daniel Vetter Jan. 21, 2021, 7:40 p.m. UTC

Recently there was a fairly long thread about recoreable hardware page
faults, how they can deadlock, and what to do about that.

While the discussion is still fresh I figured good time to try and
document the conclusions a bit.

References: https://lore.kernel.org/dri-devel/[e-mail]/
Cc: Maarten Lankhorst <[e-mail]>
Cc: Thomas Hellström <[e-mail]>
Cc: "Christian König" <[e-mail]>
Cc: Jerome Glisse <[e-mail]>
Cc: Felix Kuehling <[e-mail]>
Signed-off-by: Daniel Vetter <[e-mail]>
Cc: Sumit Semwal <[e-mail]>
Cc: [e-mail]
Cc: [e-mail]
--
I'll be away next week, but figured I'll type this up quickly for some
comments and to check whether I got this all roughly right.

Critique very much wanted on this, so that we can make sure hw which
can't preempt (with pagefaults pending) like gfx10 has a clear path to
support page faults in upstream. So anything I missed, got wrong or
like that would be good.
-Daniel
---
 Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
 1 file changed, 66 insertions(+)

Comments

Christian König Jan. 22, 2021, 1:10 p.m. UTC | #1

Am 21.01.21 um 20:40 schrieb Daniel Vetter:
> Recently there was a fairly long thread about recoreable hardware page
> faults, how they can deadlock, and what to do about that.
>
> While the discussion is still fresh I figured good time to try and
> document the conclusions a bit.
>
> References: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7C94782d99ad7d4e1cc57c08d8be447d74%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637468548672516391%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=AT8QP2r2UczSqCKkPRTJI1cQ0GOGyykgLcMfW8NbD8w%3D&reserved=0
> Cc: Maarten Lankhorst <[e-mail]>
> Cc: Thomas Hellström <[e-mail]>
> Cc: "Christian König" <[e-mail]>
> Cc: Jerome Glisse <[e-mail]>
> Cc: Felix Kuehling <[e-mail]>
> Signed-off-by: Daniel Vetter <[e-mail]>
> Cc: Sumit Semwal <[e-mail]>
> Cc: [e-mail]
> Cc: [e-mail]
> --
> I'll be away next week, but figured I'll type this up quickly for some
> comments and to check whether I got this all roughly right.
>
> Critique very much wanted on this, so that we can make sure hw which
> can't preempt (with pagefaults pending) like gfx10 has a clear path to
> support page faults in upstream. So anything I missed, got wrong or
> like that would be good.
> -Daniel
> ---
>   Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>   1 file changed, 66 insertions(+)
>
> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> index a2133d69872c..e924c1e4f7a3 100644
> --- a/Documentation/driver-api/dma-buf.rst
> +++ b/Documentation/driver-api/dma-buf.rst
> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>     userspace is allowed to use userspace fencing or long running compute
>     workloads. This also means no implicit fencing for shared buffers in these
>     cases.
> +
> +Recoverable Hardware Page Faults Implications
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> +
> +Modern hardware supports recoverable page faults, which has a lot of
> +implications for DMA fences.
> +
> +First, a pending page fault obviously holds up the work that's running on the
> +accelerator and a memory allocation is usually required to resolve the fault.
> +But memory allocations are not allowed to gate completion of DMA fences, which
> +means any workload using recoverable page faults cannot use DMA fences for
> +synchronization. Synchronization fences controlled by userspace must be used
> +instead.
> +
> +On GPUs this poses a problem, because current desktop compositor protocols on
> +Linus rely on DMA fences, which means without an entirely new userspace stack
> +built on top of userspace fences, they cannot benefit from recoverable page
> +faults. The exception is when page faults are only used as migration hints and
> +never to on-demand fill a memory request. For now this means recoverable page
> +faults on GPUs are limited to pure compute workloads.
> +
> +Furthermore GPUs usually have shared resources between the 3D rendering and
> +compute side, like compute units or command submission engines. If both a 3D
> +job with a DMA fence and a compute workload using recoverable page faults are
> +pending they could deadlock:
> +
> +- The 3D workload might need to wait for the compute job to finish and release
> +  hardware resources first.
> +
> +- The compute workload might be stuck in a page fault, because the memory
> +  allocation is waiting for the DMA fence of the 3D workload to complete.
> +
> +There are a few ways to prevent this problem:
> +
> +- Compute workloads can always be preempted, even when a page fault is pending
> +  and not yet repaired. Not all hardware supports this.
> +
> +- DMA fence workloads and workloads which need page fault handling have
> +  independent hardware resources to guarantee forward progress. This could be
> +  achieved through e.g. through dedicated engines and minimal compute unit
> +  reservations for DMA fence workloads.
> +

> +- The reservation approach could be further refined by only reserving the
> +  hardware resources for DMA fence workloads when they are in-flight. This must
> +  cover the time from when the DMA fence is visible to other threads up to
> +  moment when fence is completed through dma_fence_signal().

Up till here it makes perfect sense, but what should this paragraph mean ?

> +
> +- As a last resort, if the hardware provides no useful reservation mechanics,
> +  all workloads must be flushed from the GPU when switching between jobs
> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
> +  fences must complete before a compute job with page fault handling can be
> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
> +  made visible anywhere in the system, all compute workloads must be preempted
> +  to guarantee all pending GPU page faults are flushed.
> +
> +Note that workloads that run on independent hardware like copy engines or other
> +GPUs do not have any impact. This allows us to keep using DMA fences internally
> +in the kernel even for resolving hardware page faults, e.g. by using copy
> +engines to clear or copy memory needed to resolve the page fault.
> +
> +In some ways this page fault problem is a special case of the `Infinite DMA
> +Fences` discussions: Infinite fences from compute workloads are allowed to
> +depend on DMA fences, but not the other way around. And not even the page fault
> +problem is new, because some other CPU thread in userspace might
> +hit a page fault which holds up a userspace fence - supporting page faults on
> +GPUs doesn't anything fundamentally new.

Maybe worth noting that it just doesn't work with the implicit 
synchronization we have for existing userspace.

Regards,
Christian.

Daniel Vetter Jan. 22, 2021, 1:18 p.m. UTC | #2

On Fri, Jan 22, 2021 at 2:10 PM Christian König
<[e-mail]> wrote:
>
> Am 21.01.21 um 20:40 schrieb Daniel Vetter:
> > Recently there was a fairly long thread about recoreable hardware page
> > faults, how they can deadlock, and what to do about that.
> >
> > While the discussion is still fresh I figured good time to try and
> > document the conclusions a bit.
> >
> > References: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&amp;data=04%7C01%7Cchristian.koenig%40amd.com%7C94782d99ad7d4e1cc57c08d8be447d74%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637468548672516391%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=AT8QP2r2UczSqCKkPRTJI1cQ0GOGyykgLcMfW8NbD8w%3D&amp;reserved=0
> > Cc: Maarten Lankhorst <[e-mail]>
> > Cc: Thomas Hellström <[e-mail]>
> > Cc: "Christian König" <[e-mail]>
> > Cc: Jerome Glisse <[e-mail]>
> > Cc: Felix Kuehling <[e-mail]>
> > Signed-off-by: Daniel Vetter <[e-mail]>
> > Cc: Sumit Semwal <[e-mail]>
> > Cc: [e-mail]
> > Cc: [e-mail]
> > --
> > I'll be away next week, but figured I'll type this up quickly for some
> > comments and to check whether I got this all roughly right.
> >
> > Critique very much wanted on this, so that we can make sure hw which
> > can't preempt (with pagefaults pending) like gfx10 has a clear path to
> > support page faults in upstream. So anything I missed, got wrong or
> > like that would be good.
> > -Daniel
> > ---
> >   Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
> >   1 file changed, 66 insertions(+)
> >
> > diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> > index a2133d69872c..e924c1e4f7a3 100644
> > --- a/Documentation/driver-api/dma-buf.rst
> > +++ b/Documentation/driver-api/dma-buf.rst
> > @@ -257,3 +257,69 @@ fences in the kernel. This means:
> >     userspace is allowed to use userspace fencing or long running compute
> >     workloads. This also means no implicit fencing for shared buffers in these
> >     cases.
> > +
> > +Recoverable Hardware Page Faults Implications
> > +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > +
> > +Modern hardware supports recoverable page faults, which has a lot of
> > +implications for DMA fences.
> > +
> > +First, a pending page fault obviously holds up the work that's running on the
> > +accelerator and a memory allocation is usually required to resolve the fault.
> > +But memory allocations are not allowed to gate completion of DMA fences, which
> > +means any workload using recoverable page faults cannot use DMA fences for
> > +synchronization. Synchronization fences controlled by userspace must be used
> > +instead.
> > +
> > +On GPUs this poses a problem, because current desktop compositor protocols on
> > +Linus rely on DMA fences, which means without an entirely new userspace stack
> > +built on top of userspace fences, they cannot benefit from recoverable page
> > +faults. The exception is when page faults are only used as migration hints and
> > +never to on-demand fill a memory request. For now this means recoverable page
> > +faults on GPUs are limited to pure compute workloads.
> > +
> > +Furthermore GPUs usually have shared resources between the 3D rendering and
> > +compute side, like compute units or command submission engines. If both a 3D
> > +job with a DMA fence and a compute workload using recoverable page faults are
> > +pending they could deadlock:
> > +
> > +- The 3D workload might need to wait for the compute job to finish and release
> > +  hardware resources first.
> > +
> > +- The compute workload might be stuck in a page fault, because the memory
> > +  allocation is waiting for the DMA fence of the 3D workload to complete.
> > +
> > +There are a few ways to prevent this problem:
> > +
> > +- Compute workloads can always be preempted, even when a page fault is pending
> > +  and not yet repaired. Not all hardware supports this.
> > +
> > +- DMA fence workloads and workloads which need page fault handling have
> > +  independent hardware resources to guarantee forward progress. This could be
> > +  achieved through e.g. through dedicated engines and minimal compute unit
> > +  reservations for DMA fence workloads.
> > +
>
> > +- The reservation approach could be further refined by only reserving the
> > +  hardware resources for DMA fence workloads when they are in-flight. This must
> > +  cover the time from when the DMA fence is visible to other threads up to
> > +  moment when fence is completed through dma_fence_signal().
>
> Up till here it makes perfect sense, but what should this paragraph mean ?

Instead of reserving a few CU at driver load, to guarantee that
dma-fence workloads can always complete, we only do the reservatation
while a problematic dma_fence is in the system, and note yet
signalled. Of course that approach needs to be very careful, to really
make sure you can't ever deadlock by accident because the dynamic
reservation at runtime was done a notch too late.

This allows us to use all CUs on pure compute workloads (on servers,
and on desktop as long as nothing gets rendered).

If that clicks, got ideas for how to improve the text?

> > +- As a last resort, if the hardware provides no useful reservation mechanics,
> > +  all workloads must be flushed from the GPU when switching between jobs
> > +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
> > +  fences must complete before a compute job with page fault handling can be
> > +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
> > +  made visible anywhere in the system, all compute workloads must be preempted
> > +  to guarantee all pending GPU page faults are flushed.
> > +
> > +Note that workloads that run on independent hardware like copy engines or other
> > +GPUs do not have any impact. This allows us to keep using DMA fences internally
> > +in the kernel even for resolving hardware page faults, e.g. by using copy
> > +engines to clear or copy memory needed to resolve the page fault.
> > +
> > +In some ways this page fault problem is a special case of the `Infinite DMA
> > +Fences` discussions: Infinite fences from compute workloads are allowed to
> > +depend on DMA fences, but not the other way around. And not even the page fault
> > +problem is new, because some other CPU thread in userspace might
> > +hit a page fault which holds up a userspace fence - supporting page faults on
> > +GPUs doesn't anything fundamentally new.
>
> Maybe worth noting that it just doesn't work with the implicit
> synchronization we have for existing userspace.

At the very beginning I covered that already with ".... they cannot
benefit from recoverable page faults." Should I hammer that home more
that with implicit sync, it's really not going to happen? The thing
is, this is just a special case from dma_fence sync, and we also have
some protocols (I think only android) which use sync_file for that,
and that also is not going to work. I guess we could do the same trick
as with timeline fence objects for waiting for the actual dma_fence to
materialize there. But given how much trouble rolling out timeline
syncobj was I don't think that'll happen. So maybe just specifically
include implicitly synced protocols?
-Daniel

Christian König Jan. 22, 2021, 1:24 p.m. UTC | #3

Am 22.01.21 um 14:18 schrieb Daniel Vetter:
> On Fri, Jan 22, 2021 at 2:10 PM Christian König
> <[e-mail]> wrote:
>> Am 21.01.21 um 20:40 schrieb Daniel Vetter:
>>> Recently there was a fairly long thread about recoreable hardware page
>>> faults, how they can deadlock, and what to do about that.
>>>
>>> While the discussion is still fresh I figured good time to try and
>>> document the conclusions a bit.
>>>
>>> References: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&amp;data=04%7C01%7Cchristian.koenig%40amd.com%7C25c2b659bc8f47e0bace08d8bed83728%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637469183153437091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=GlEKsPLRRRO%2BI1JSDpvNeBFbnFacmymxkj8e7QqM5SA%3D&amp;reserved=0
>>> Cc: Maarten Lankhorst <[e-mail]>
>>> Cc: Thomas Hellström <[e-mail]>
>>> Cc: "Christian König" <[e-mail]>
>>> Cc: Jerome Glisse <[e-mail]>
>>> Cc: Felix Kuehling <[e-mail]>
>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>> Cc: Sumit Semwal <[e-mail]>
>>> Cc: [e-mail]
>>> Cc: [e-mail]
>>> --
>>> I'll be away next week, but figured I'll type this up quickly for some
>>> comments and to check whether I got this all roughly right.
>>>
>>> Critique very much wanted on this, so that we can make sure hw which
>>> can't preempt (with pagefaults pending) like gfx10 has a clear path to

One more comment here: You should probably mention that gfx10 is 
referring to AMD GPUs.

>>> support page faults in upstream. So anything I missed, got wrong or
>>> like that would be good.
>>> -Daniel
>>> ---
>>>    Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>>>    1 file changed, 66 insertions(+)
>>>
>>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
>>> index a2133d69872c..e924c1e4f7a3 100644
>>> --- a/Documentation/driver-api/dma-buf.rst
>>> +++ b/Documentation/driver-api/dma-buf.rst
>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>      userspace is allowed to use userspace fencing or long running compute
>>>      workloads. This also means no implicit fencing for shared buffers in these
>>>      cases.
>>> +
>>> +Recoverable Hardware Page Faults Implications
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +Modern hardware supports recoverable page faults, which has a lot of
>>> +implications for DMA fences.
>>> +
>>> +First, a pending page fault obviously holds up the work that's running on the
>>> +accelerator and a memory allocation is usually required to resolve the fault.
>>> +But memory allocations are not allowed to gate completion of DMA fences, which
>>> +means any workload using recoverable page faults cannot use DMA fences for
>>> +synchronization. Synchronization fences controlled by userspace must be used
>>> +instead.
>>> +
>>> +On GPUs this poses a problem, because current desktop compositor protocols on
>>> +Linus rely on DMA fences, which means without an entirely new userspace stack
>>> +built on top of userspace fences, they cannot benefit from recoverable page
>>> +faults. The exception is when page faults are only used as migration hints and
>>> +never to on-demand fill a memory request. For now this means recoverable page
>>> +faults on GPUs are limited to pure compute workloads.
>>> +
>>> +Furthermore GPUs usually have shared resources between the 3D rendering and
>>> +compute side, like compute units or command submission engines. If both a 3D
>>> +job with a DMA fence and a compute workload using recoverable page faults are
>>> +pending they could deadlock:
>>> +
>>> +- The 3D workload might need to wait for the compute job to finish and release
>>> +  hardware resources first.
>>> +
>>> +- The compute workload might be stuck in a page fault, because the memory
>>> +  allocation is waiting for the DMA fence of the 3D workload to complete.
>>> +
>>> +There are a few ways to prevent this problem:
>>> +
>>> +- Compute workloads can always be preempted, even when a page fault is pending
>>> +  and not yet repaired. Not all hardware supports this.
>>> +
>>> +- DMA fence workloads and workloads which need page fault handling have
>>> +  independent hardware resources to guarantee forward progress. This could be
>>> +  achieved through e.g. through dedicated engines and minimal compute unit
>>> +  reservations for DMA fence workloads.
>>> +
>>> +- The reservation approach could be further refined by only reserving the
>>> +  hardware resources for DMA fence workloads when they are in-flight. This must
>>> +  cover the time from when the DMA fence is visible to other threads up to
>>> +  moment when fence is completed through dma_fence_signal().
>> Up till here it makes perfect sense, but what should this paragraph mean ?
> Instead of reserving a few CU at driver load, to guarantee that
> dma-fence workloads can always complete, we only do the reservatation
> while a problematic dma_fence is in the system, and note yet
> signalled. Of course that approach needs to be very careful, to really
> make sure you can't ever deadlock by accident because the dynamic
> reservation at runtime was done a notch too late.
>
> This allows us to use all CUs on pure compute workloads (on servers,
> and on desktop as long as nothing gets rendered).
>
> If that clicks, got ideas for how to improve the text?

Yeah, that makes much more sense.

But if I'm not totally mistaken AMD GPUs can't change the CU config on 
the fly.

I mean you could be in the middle of a compute calculation on the CU 
when a 3d job comes in.

So from my perspective I think we could just drop that.

>>> +- As a last resort, if the hardware provides no useful reservation mechanics,
>>> +  all workloads must be flushed from the GPU when switching between jobs
>>> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
>>> +  fences must complete before a compute job with page fault handling can be
>>> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
>>> +  made visible anywhere in the system, all compute workloads must be preempted
>>> +  to guarantee all pending GPU page faults are flushed.
>>> +
>>> +Note that workloads that run on independent hardware like copy engines or other
>>> +GPUs do not have any impact. This allows us to keep using DMA fences internally
>>> +in the kernel even for resolving hardware page faults, e.g. by using copy
>>> +engines to clear or copy memory needed to resolve the page fault.
>>> +
>>> +In some ways this page fault problem is a special case of the `Infinite DMA
>>> +Fences` discussions: Infinite fences from compute workloads are allowed to
>>> +depend on DMA fences, but not the other way around. And not even the page fault
>>> +problem is new, because some other CPU thread in userspace might
>>> +hit a page fault which holds up a userspace fence - supporting page faults on
>>> +GPUs doesn't anything fundamentally new.
>> Maybe worth noting that it just doesn't work with the implicit
>> synchronization we have for existing userspace.
> At the very beginning I covered that already with ".... they cannot
> benefit from recoverable page faults." Should I hammer that home more
> that with implicit sync, it's really not going to happen? The thing
> is, this is just a special case from dma_fence sync, and we also have
> some protocols (I think only android) which use sync_file for that,
> and that also is not going to work. I guess we could do the same trick
> as with timeline fence objects for waiting for the actual dma_fence to
> materialize there. But given how much trouble rolling out timeline
> syncobj was I don't think that'll happen. So maybe just specifically
> include implicitly synced protocols?

A one line reference to the stuff above should be enough.

Alternatively merge this into the stuff above as well.

Regards,
Christian.

> -Daniel

Daniel Vetter Jan. 22, 2021, 1:34 p.m. UTC | #4

On Fri, Jan 22, 2021 at 2:24 PM Christian König
<[e-mail]> wrote:
>
> Am 22.01.21 um 14:18 schrieb Daniel Vetter:
> > On Fri, Jan 22, 2021 at 2:10 PM Christian König
> > <[e-mail]> wrote:
> >> Am 21.01.21 um 20:40 schrieb Daniel Vetter:
> >>> Recently there was a fairly long thread about recoreable hardware page
> >>> faults, how they can deadlock, and what to do about that.
> >>>
> >>> While the discussion is still fresh I figured good time to try and
> >>> document the conclusions a bit.
> >>>
> >>> References: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&amp;data=04%7C01%7Cchristian.koenig%40amd.com%7C25c2b659bc8f47e0bace08d8bed83728%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637469183153437091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=GlEKsPLRRRO%2BI1JSDpvNeBFbnFacmymxkj8e7QqM5SA%3D&amp;reserved=0
> >>> Cc: Maarten Lankhorst <[e-mail]>
> >>> Cc: Thomas Hellström <[e-mail]>
> >>> Cc: "Christian König" <[e-mail]>
> >>> Cc: Jerome Glisse <[e-mail]>
> >>> Cc: Felix Kuehling <[e-mail]>
> >>> Signed-off-by: Daniel Vetter <[e-mail]>
> >>> Cc: Sumit Semwal <[e-mail]>
> >>> Cc: [e-mail]
> >>> Cc: [e-mail]
> >>> --
> >>> I'll be away next week, but figured I'll type this up quickly for some
> >>> comments and to check whether I got this all roughly right.
> >>>
> >>> Critique very much wanted on this, so that we can make sure hw which
> >>> can't preempt (with pagefaults pending) like gfx10 has a clear path to
>
> One more comment here: You should probably mention that gfx10 is
> referring to AMD GPUs.

Oh that was just the single-patch cover letter. I'll drop it for the
next round since that's not going to be part of the real patch.

> >>> support page faults in upstream. So anything I missed, got wrong or
> >>> like that would be good.
> >>> -Daniel
> >>> ---
> >>>    Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
> >>>    1 file changed, 66 insertions(+)
> >>>
> >>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> >>> index a2133d69872c..e924c1e4f7a3 100644
> >>> --- a/Documentation/driver-api/dma-buf.rst
> >>> +++ b/Documentation/driver-api/dma-buf.rst
> >>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
> >>>      userspace is allowed to use userspace fencing or long running compute
> >>>      workloads. This also means no implicit fencing for shared buffers in these
> >>>      cases.
> >>> +
> >>> +Recoverable Hardware Page Faults Implications
> >>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>> +
> >>> +Modern hardware supports recoverable page faults, which has a lot of
> >>> +implications for DMA fences.
> >>> +
> >>> +First, a pending page fault obviously holds up the work that's running on the
> >>> +accelerator and a memory allocation is usually required to resolve the fault.
> >>> +But memory allocations are not allowed to gate completion of DMA fences, which
> >>> +means any workload using recoverable page faults cannot use DMA fences for
> >>> +synchronization. Synchronization fences controlled by userspace must be used
> >>> +instead.
> >>> +
> >>> +On GPUs this poses a problem, because current desktop compositor protocols on
> >>> +Linus rely on DMA fences, which means without an entirely new userspace stack
> >>> +built on top of userspace fences, they cannot benefit from recoverable page
> >>> +faults. The exception is when page faults are only used as migration hints and
> >>> +never to on-demand fill a memory request. For now this means recoverable page
> >>> +faults on GPUs are limited to pure compute workloads.
> >>> +
> >>> +Furthermore GPUs usually have shared resources between the 3D rendering and
> >>> +compute side, like compute units or command submission engines. If both a 3D
> >>> +job with a DMA fence and a compute workload using recoverable page faults are
> >>> +pending they could deadlock:
> >>> +
> >>> +- The 3D workload might need to wait for the compute job to finish and release
> >>> +  hardware resources first.
> >>> +
> >>> +- The compute workload might be stuck in a page fault, because the memory
> >>> +  allocation is waiting for the DMA fence of the 3D workload to complete.
> >>> +
> >>> +There are a few ways to prevent this problem:
> >>> +
> >>> +- Compute workloads can always be preempted, even when a page fault is pending
> >>> +  and not yet repaired. Not all hardware supports this.
> >>> +
> >>> +- DMA fence workloads and workloads which need page fault handling have
> >>> +  independent hardware resources to guarantee forward progress. This could be
> >>> +  achieved through e.g. through dedicated engines and minimal compute unit
> >>> +  reservations for DMA fence workloads.
> >>> +
> >>> +- The reservation approach could be further refined by only reserving the
> >>> +  hardware resources for DMA fence workloads when they are in-flight. This must
> >>> +  cover the time from when the DMA fence is visible to other threads up to
> >>> +  moment when fence is completed through dma_fence_signal().
> >> Up till here it makes perfect sense, but what should this paragraph mean ?
> > Instead of reserving a few CU at driver load, to guarantee that
> > dma-fence workloads can always complete, we only do the reservatation
> > while a problematic dma_fence is in the system, and note yet
> > signalled. Of course that approach needs to be very careful, to really
> > make sure you can't ever deadlock by accident because the dynamic
> > reservation at runtime was done a notch too late.
> >
> > This allows us to use all CUs on pure compute workloads (on servers,
> > and on desktop as long as nothing gets rendered).
> >
> > If that clicks, got ideas for how to improve the text?
>
> Yeah, that makes much more sense.
>
> But if I'm not totally mistaken AMD GPUs can't change the CU config on
> the fly.
>
> I mean you could be in the middle of a compute calculation on the CU
> when a 3d job comes in.
>
> So from my perspective I think we could just drop that.

Yeah, I don't expect a single driver to implement everything :-) I
just wanted to go through all the options, because hw is tricky and
tends to change. My idea is also that I can use this here and go to
(our) hw people, if they don't like the consequences, with a list of
options where they can pick the one they like the most. That's why I
want to cover the entire palette of options, down to the pretty big
"full flush" hammer.

I think for the polished version I'll want to give each option some
good name, like "preemptible page-fault" "shared resource
reservation", "dynamic shared resource reservation" and "full flush".
Just so we have some good names to talk about these concepts, instead
of me giving them confusion names like the "gang schedule" thing :-)

> >>> +- As a last resort, if the hardware provides no useful reservation mechanics,
> >>> +  all workloads must be flushed from the GPU when switching between jobs
> >>> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
> >>> +  fences must complete before a compute job with page fault handling can be
> >>> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
> >>> +  made visible anywhere in the system, all compute workloads must be preempted
> >>> +  to guarantee all pending GPU page faults are flushed.
> >>> +
> >>> +Note that workloads that run on independent hardware like copy engines or other
> >>> +GPUs do not have any impact. This allows us to keep using DMA fences internally
> >>> +in the kernel even for resolving hardware page faults, e.g. by using copy
> >>> +engines to clear or copy memory needed to resolve the page fault.
> >>> +
> >>> +In some ways this page fault problem is a special case of the `Infinite DMA
> >>> +Fences` discussions: Infinite fences from compute workloads are allowed to
> >>> +depend on DMA fences, but not the other way around. And not even the page fault
> >>> +problem is new, because some other CPU thread in userspace might
> >>> +hit a page fault which holds up a userspace fence - supporting page faults on
> >>> +GPUs doesn't anything fundamentally new.
> >> Maybe worth noting that it just doesn't work with the implicit
> >> synchronization we have for existing userspace.
> > At the very beginning I covered that already with ".... they cannot
> > benefit from recoverable page faults." Should I hammer that home more
> > that with implicit sync, it's really not going to happen? The thing
> > is, this is just a special case from dma_fence sync, and we also have
> > some protocols (I think only android) which use sync_file for that,
> > and that also is not going to work. I guess we could do the same trick
> > as with timeline fence objects for waiting for the actual dma_fence to
> > materialize there. But given how much trouble rolling out timeline
> > syncobj was I don't think that'll happen. So maybe just specifically
> > include implicitly synced protocols?
>
> A one line reference to the stuff above should be enough.
>
> Alternatively merge this into the stuff above as well.

Yeah I'll try to make this clearer in the paragraph in the beginning.

I'll send out some polished version when I'm back after next week.
-Daniel

Felix Kuehling Jan. 22, 2021, 4:29 p.m. UTC | #5

Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> Recently there was a fairly long thread about recoreable hardware page
> faults, how they can deadlock, and what to do about that.
>
> While the discussion is still fresh I figured good time to try and
> document the conclusions a bit.
Thank you Daniel. This is a good summary of our discussion. It's also an
external reference I can point our HW engineers at when they're
wondering about what "real software" does.

Regards,
  Felix


>
> References: https://lore.kernel.org/dri-devel/[e-mail]/
> Cc: Maarten Lankhorst <[e-mail]>
> Cc: Thomas Hellström <[e-mail]>
> Cc: "Christian König" <[e-mail]>
> Cc: Jerome Glisse <[e-mail]>
> Cc: Felix Kuehling <[e-mail]>
> Signed-off-by: Daniel Vetter <[e-mail]>
> Cc: Sumit Semwal <[e-mail]>
> Cc: [e-mail]
> Cc: [e-mail]
> --
> I'll be away next week, but figured I'll type this up quickly for some
> comments and to check whether I got this all roughly right.
>
> Critique very much wanted on this, so that we can make sure hw which
> can't preempt (with pagefaults pending) like gfx10 has a clear path to
> support page faults in upstream. So anything I missed, got wrong or
> like that would be good.
> -Daniel
> ---
>  Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>  1 file changed, 66 insertions(+)
>
> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> index a2133d69872c..e924c1e4f7a3 100644
> --- a/Documentation/driver-api/dma-buf.rst
> +++ b/Documentation/driver-api/dma-buf.rst
> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>    userspace is allowed to use userspace fencing or long running compute
>    workloads. This also means no implicit fencing for shared buffers in these
>    cases.
> +
> +Recoverable Hardware Page Faults Implications
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> +
> +Modern hardware supports recoverable page faults, which has a lot of
> +implications for DMA fences.
> +
> +First, a pending page fault obviously holds up the work that's running on the
> +accelerator and a memory allocation is usually required to resolve the fault.
> +But memory allocations are not allowed to gate completion of DMA fences, which
> +means any workload using recoverable page faults cannot use DMA fences for
> +synchronization. Synchronization fences controlled by userspace must be used
> +instead.
> +
> +On GPUs this poses a problem, because current desktop compositor protocols on
> +Linus rely on DMA fences, which means without an entirely new userspace stack
> +built on top of userspace fences, they cannot benefit from recoverable page
> +faults. The exception is when page faults are only used as migration hints and
> +never to on-demand fill a memory request. For now this means recoverable page
> +faults on GPUs are limited to pure compute workloads.
> +
> +Furthermore GPUs usually have shared resources between the 3D rendering and
> +compute side, like compute units or command submission engines. If both a 3D
> +job with a DMA fence and a compute workload using recoverable page faults are
> +pending they could deadlock:
> +
> +- The 3D workload might need to wait for the compute job to finish and release
> +  hardware resources first.
> +
> +- The compute workload might be stuck in a page fault, because the memory
> +  allocation is waiting for the DMA fence of the 3D workload to complete.
> +
> +There are a few ways to prevent this problem:
> +
> +- Compute workloads can always be preempted, even when a page fault is pending
> +  and not yet repaired. Not all hardware supports this.
> +
> +- DMA fence workloads and workloads which need page fault handling have
> +  independent hardware resources to guarantee forward progress. This could be
> +  achieved through e.g. through dedicated engines and minimal compute unit
> +  reservations for DMA fence workloads.
> +
> +- The reservation approach could be further refined by only reserving the
> +  hardware resources for DMA fence workloads when they are in-flight. This must
> +  cover the time from when the DMA fence is visible to other threads up to
> +  moment when fence is completed through dma_fence_signal().
> +
> +- As a last resort, if the hardware provides no useful reservation mechanics,
> +  all workloads must be flushed from the GPU when switching between jobs
> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
> +  fences must complete before a compute job with page fault handling can be
> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
> +  made visible anywhere in the system, all compute workloads must be preempted
> +  to guarantee all pending GPU page faults are flushed.
> +
> +Note that workloads that run on independent hardware like copy engines or other
> +GPUs do not have any impact. This allows us to keep using DMA fences internally
> +in the kernel even for resolving hardware page faults, e.g. by using copy
> +engines to clear or copy memory needed to resolve the page fault.
> +
> +In some ways this page fault problem is a special case of the `Infinite DMA
> +Fences` discussions: Infinite fences from compute workloads are allowed to
> +depend on DMA fences, but not the other way around. And not even the page fault
> +problem is new, because some other CPU thread in userspace might
> +hit a page fault which holds up a userspace fence - supporting page faults on
> +GPUs doesn't anything fundamentally new.

Felix Kuehling Jan. 27, 2021, 12:22 a.m. UTC | #6

Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> Recently there was a fairly long thread about recoreable hardware page
> faults, how they can deadlock, and what to do about that.
>
> While the discussion is still fresh I figured good time to try and
> document the conclusions a bit.
>
> References: https://lore.kernel.org/dri-devel/[e-mail]/
> Cc: Maarten Lankhorst <[e-mail]>
> Cc: Thomas Hellström <[e-mail]>
> Cc: "Christian König" <[e-mail]>
> Cc: Jerome Glisse <[e-mail]>
> Cc: Felix Kuehling <[e-mail]>
> Signed-off-by: Daniel Vetter <[e-mail]>
> Cc: Sumit Semwal <[e-mail]>
> Cc: [e-mail]
> Cc: [e-mail]
> --
> I'll be away next week, but figured I'll type this up quickly for some
> comments and to check whether I got this all roughly right.
>
> Critique very much wanted on this, so that we can make sure hw which
> can't preempt (with pagefaults pending) like gfx10 has a clear path to
> support page faults in upstream. So anything I missed, got wrong or
> like that would be good.
> -Daniel
> ---
>  Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>  1 file changed, 66 insertions(+)
>
> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> index a2133d69872c..e924c1e4f7a3 100644
> --- a/Documentation/driver-api/dma-buf.rst
> +++ b/Documentation/driver-api/dma-buf.rst
> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>    userspace is allowed to use userspace fencing or long running compute
>    workloads. This also means no implicit fencing for shared buffers in these
>    cases.
> +
> +Recoverable Hardware Page Faults Implications
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> +
> +Modern hardware supports recoverable page faults, which has a lot of
> +implications for DMA fences.
> +
> +First, a pending page fault obviously holds up the work that's running on the
> +accelerator and a memory allocation is usually required to resolve the fault.
> +But memory allocations are not allowed to gate completion of DMA fences, which
> +means any workload using recoverable page faults cannot use DMA fences for
> +synchronization. Synchronization fences controlled by userspace must be used
> +instead.
> +
> +On GPUs this poses a problem, because current desktop compositor protocols on
> +Linus rely on DMA fences, which means without an entirely new userspace stack
> +built on top of userspace fences, they cannot benefit from recoverable page
> +faults. The exception is when page faults are only used as migration hints and
> +never to on-demand fill a memory request. For now this means recoverable page
> +faults on GPUs are limited to pure compute workloads.
> +
> +Furthermore GPUs usually have shared resources between the 3D rendering and
> +compute side, like compute units or command submission engines. If both a 3D
> +job with a DMA fence and a compute workload using recoverable page faults are
> +pending they could deadlock:
> +
> +- The 3D workload might need to wait for the compute job to finish and release
> +  hardware resources first.
> +
> +- The compute workload might be stuck in a page fault, because the memory
> +  allocation is waiting for the DMA fence of the 3D workload to complete.
> +
> +There are a few ways to prevent this problem:
> +
> +- Compute workloads can always be preempted, even when a page fault is pending
> +  and not yet repaired. Not all hardware supports this.
> +
> +- DMA fence workloads and workloads which need page fault handling have
> +  independent hardware resources to guarantee forward progress. This could be
> +  achieved through e.g. through dedicated engines and minimal compute unit
> +  reservations for DMA fence workloads.
> +
> +- The reservation approach could be further refined by only reserving the
> +  hardware resources for DMA fence workloads when they are in-flight. This must
> +  cover the time from when the DMA fence is visible to other threads up to
> +  moment when fence is completed through dma_fence_signal().
> +
> +- As a last resort, if the hardware provides no useful reservation mechanics,
> +  all workloads must be flushed from the GPU when switching between jobs
> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
> +  fences must complete before a compute job with page fault handling can be
> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
> +  made visible anywhere in the system, all compute workloads must be preempted
> +  to guarantee all pending GPU page faults are flushed.

I thought of another possible workaround:

  * Partition the memory. Servicing of page faults will use a separate
    memory pool that can always be allocated from without waiting for
    fences. This includes memory for page tables and memory for
    migrating data to. You may steal memory from other processes that
    can page fault, so no fence waiting is necessary. Being able to
    steal memory at any time also means there are basically no
    out-of-memory situations you need to worry about. Even page tables
    (except the root page directory of each process) can be stolen in
    the worst case.

Regards,
  Felix

> +
> +Note that workloads that run on independent hardware like copy engines or other
> +GPUs do not have any impact. This allows us to keep using DMA fences internally
> +in the kernel even for resolving hardware page faults, e.g. by using copy
> +engines to clear or copy memory needed to resolve the page fault.
> +
> +In some ways this page fault problem is a special case of the `Infinite DMA
> +Fences` discussions: Infinite fences from compute workloads are allowed to
> +depend on DMA fences, but not the other way around. And not even the page fault
> +problem is new, because some other CPU thread in userspace might
> +hit a page fault which holds up a userspace fence - supporting page faults on
> +GPUs doesn't anything fundamentally new.

Maarten Lankhorst Jan. 27, 2021, 12:11 p.m. UTC | #7

Op 27-01-2021 om 01:22 schreef Felix Kuehling:
> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>> Recently there was a fairly long thread about recoreable hardware page
>> faults, how they can deadlock, and what to do about that.
>>
>> While the discussion is still fresh I figured good time to try and
>> document the conclusions a bit.
>>
>> References: https://lore.kernel.org/dri-devel/[e-mail]/
>> Cc: Maarten Lankhorst <[e-mail]>
>> Cc: Thomas Hellström <[e-mail]>
>> Cc: "Christian König" <[e-mail]>
>> Cc: Jerome Glisse <[e-mail]>
>> Cc: Felix Kuehling <[e-mail]>
>> Signed-off-by: Daniel Vetter <[e-mail]>
>> Cc: Sumit Semwal <[e-mail]>
>> Cc: [e-mail]
>> Cc: [e-mail]
>> --
>> I'll be away next week, but figured I'll type this up quickly for some
>> comments and to check whether I got this all roughly right.
>>
>> Critique very much wanted on this, so that we can make sure hw which
>> can't preempt (with pagefaults pending) like gfx10 has a clear path to
>> support page faults in upstream. So anything I missed, got wrong or
>> like that would be good.
>> -Daniel
>> ---
>>  Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>>  1 file changed, 66 insertions(+)
>>
>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
>> index a2133d69872c..e924c1e4f7a3 100644
>> --- a/Documentation/driver-api/dma-buf.rst
>> +++ b/Documentation/driver-api/dma-buf.rst
>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>    userspace is allowed to use userspace fencing or long running compute
>>    workloads. This also means no implicit fencing for shared buffers in these
>>    cases.
>> +
>> +Recoverable Hardware Page Faults Implications
>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> +
>> +Modern hardware supports recoverable page faults, which has a lot of
>> +implications for DMA fences.
>> +
>> +First, a pending page fault obviously holds up the work that's running on the
>> +accelerator and a memory allocation is usually required to resolve the fault.
>> +But memory allocations are not allowed to gate completion of DMA fences, which
>> +means any workload using recoverable page faults cannot use DMA fences for
>> +synchronization. Synchronization fences controlled by userspace must be used
>> +instead.
>> +
>> +On GPUs this poses a problem, because current desktop compositor protocols on
>> +Linus rely on DMA fences, which means without an entirely new userspace stack
>> +built on top of userspace fences, they cannot benefit from recoverable page
>> +faults. The exception is when page faults are only used as migration hints and
>> +never to on-demand fill a memory request. For now this means recoverable page
>> +faults on GPUs are limited to pure compute workloads.
>> +
>> +Furthermore GPUs usually have shared resources between the 3D rendering and
>> +compute side, like compute units or command submission engines. If both a 3D
>> +job with a DMA fence and a compute workload using recoverable page faults are
>> +pending they could deadlock:
>> +
>> +- The 3D workload might need to wait for the compute job to finish and release
>> +  hardware resources first.
>> +
>> +- The compute workload might be stuck in a page fault, because the memory
>> +  allocation is waiting for the DMA fence of the 3D workload to complete.
>> +
>> +There are a few ways to prevent this problem:
>> +
>> +- Compute workloads can always be preempted, even when a page fault is pending
>> +  and not yet repaired. Not all hardware supports this.
>> +
>> +- DMA fence workloads and workloads which need page fault handling have
>> +  independent hardware resources to guarantee forward progress. This could be
>> +  achieved through e.g. through dedicated engines and minimal compute unit
>> +  reservations for DMA fence workloads.
>> +
>> +- The reservation approach could be further refined by only reserving the
>> +  hardware resources for DMA fence workloads when they are in-flight. This must
>> +  cover the time from when the DMA fence is visible to other threads up to
>> +  moment when fence is completed through dma_fence_signal().
>> +
>> +- As a last resort, if the hardware provides no useful reservation mechanics,
>> +  all workloads must be flushed from the GPU when switching between jobs
>> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
>> +  fences must complete before a compute job with page fault handling can be
>> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
>> +  made visible anywhere in the system, all compute workloads must be preempted
>> +  to guarantee all pending GPU page faults are flushed.
> I thought of another possible workaround:
>
>   * Partition the memory. Servicing of page faults will use a separate
>     memory pool that can always be allocated from without waiting for
>     fences. This includes memory for page tables and memory for
>     migrating data to. You may steal memory from other processes that
>     can page fault, so no fence waiting is necessary. Being able to
>     steal memory at any time also means there are basically no
>     out-of-memory situations you need to worry about. Even page tables
>     (except the root page directory of each process) can be stolen in
>     the worst case.

I think 'overcommit' would be a nice way to describe this. But I'm not
sure how easy this is to implement in practice. You would basically need
to create your own memory manager for this.

But from a design point of view, definitely a valid solution.

But this looks good, those solutions are definitely the valid options we
can choose from.

~Maarten

Christian König Jan. 27, 2021, 12:16 p.m. UTC | #8

Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>> Recently there was a fairly long thread about recoreable hardware page
>>> faults, how they can deadlock, and what to do about that.
>>>
>>> While the discussion is still fresh I figured good time to try and
>>> document the conclusions a bit.
>>>
>>> References: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7Cbee0aeff80f440bcc52108d8c2bcc11f%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637473463245588199%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=ncr%2Fqv5lw0ONrYxFvfdcFAXAZ%2BXcJJa6UY%2BxGfcKGVM%3D&reserved=0
>>> Cc: Maarten Lankhorst <[e-mail]>
>>> Cc: Thomas Hellström <[e-mail]>
>>> Cc: "Christian König" <[e-mail]>
>>> Cc: Jerome Glisse <[e-mail]>
>>> Cc: Felix Kuehling <[e-mail]>
>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>> Cc: Sumit Semwal <[e-mail]>
>>> Cc: [e-mail]
>>> Cc: [e-mail]
>>> --
>>> I'll be away next week, but figured I'll type this up quickly for some
>>> comments and to check whether I got this all roughly right.
>>>
>>> Critique very much wanted on this, so that we can make sure hw which
>>> can't preempt (with pagefaults pending) like gfx10 has a clear path to
>>> support page faults in upstream. So anything I missed, got wrong or
>>> like that would be good.
>>> -Daniel
>>> ---
>>>   Documentation/driver-api/dma-buf.rst | 66 ++++++++++++++++++++++++++++
>>>   1 file changed, 66 insertions(+)
>>>
>>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
>>> index a2133d69872c..e924c1e4f7a3 100644
>>> --- a/Documentation/driver-api/dma-buf.rst
>>> +++ b/Documentation/driver-api/dma-buf.rst
>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>     userspace is allowed to use userspace fencing or long running compute
>>>     workloads. This also means no implicit fencing for shared buffers in these
>>>     cases.
>>> +
>>> +Recoverable Hardware Page Faults Implications
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +Modern hardware supports recoverable page faults, which has a lot of
>>> +implications for DMA fences.
>>> +
>>> +First, a pending page fault obviously holds up the work that's running on the
>>> +accelerator and a memory allocation is usually required to resolve the fault.
>>> +But memory allocations are not allowed to gate completion of DMA fences, which
>>> +means any workload using recoverable page faults cannot use DMA fences for
>>> +synchronization. Synchronization fences controlled by userspace must be used
>>> +instead.
>>> +
>>> +On GPUs this poses a problem, because current desktop compositor protocols on
>>> +Linus rely on DMA fences, which means without an entirely new userspace stack
>>> +built on top of userspace fences, they cannot benefit from recoverable page
>>> +faults. The exception is when page faults are only used as migration hints and
>>> +never to on-demand fill a memory request. For now this means recoverable page
>>> +faults on GPUs are limited to pure compute workloads.
>>> +
>>> +Furthermore GPUs usually have shared resources between the 3D rendering and
>>> +compute side, like compute units or command submission engines. If both a 3D
>>> +job with a DMA fence and a compute workload using recoverable page faults are
>>> +pending they could deadlock:
>>> +
>>> +- The 3D workload might need to wait for the compute job to finish and release
>>> +  hardware resources first.
>>> +
>>> +- The compute workload might be stuck in a page fault, because the memory
>>> +  allocation is waiting for the DMA fence of the 3D workload to complete.
>>> +
>>> +There are a few ways to prevent this problem:
>>> +
>>> +- Compute workloads can always be preempted, even when a page fault is pending
>>> +  and not yet repaired. Not all hardware supports this.
>>> +
>>> +- DMA fence workloads and workloads which need page fault handling have
>>> +  independent hardware resources to guarantee forward progress. This could be
>>> +  achieved through e.g. through dedicated engines and minimal compute unit
>>> +  reservations for DMA fence workloads.
>>> +
>>> +- The reservation approach could be further refined by only reserving the
>>> +  hardware resources for DMA fence workloads when they are in-flight. This must
>>> +  cover the time from when the DMA fence is visible to other threads up to
>>> +  moment when fence is completed through dma_fence_signal().
>>> +
>>> +- As a last resort, if the hardware provides no useful reservation mechanics,
>>> +  all workloads must be flushed from the GPU when switching between jobs
>>> +  requiring DMA fences or jobs requiring page fault handling: This means all DMA
>>> +  fences must complete before a compute job with page fault handling can be
>>> +  inserted into the scheduler queue. And vice versa, before a DMA fence can be
>>> +  made visible anywhere in the system, all compute workloads must be preempted
>>> +  to guarantee all pending GPU page faults are flushed.
>> I thought of another possible workaround:
>>
>>    * Partition the memory. Servicing of page faults will use a separate
>>      memory pool that can always be allocated from without waiting for
>>      fences. This includes memory for page tables and memory for
>>      migrating data to. You may steal memory from other processes that
>>      can page fault, so no fence waiting is necessary. Being able to
>>      steal memory at any time also means there are basically no
>>      out-of-memory situations you need to worry about. Even page tables
>>      (except the root page directory of each process) can be stolen in
>>      the worst case.
> I think 'overcommit' would be a nice way to describe this. But I'm not
> sure how easy this is to implement in practice. You would basically need
> to create your own memory manager for this.

Well you would need a completely separate pool for both device as well 
as system memory.

E.g. on boot we say we steal X GB system memory only for HMM.

> But from a design point of view, definitely a valid solution.

I think the restriction above makes it pretty much unusable.

> But this looks good, those solutions are definitely the valid options we
> can choose from.

It's certainly worth noting, yes. And just to make sure that nobody has 
the idea to reserve only device memory.

Christian.

>
> ~Maarten
>

Felix Kuehling Jan. 27, 2021, 10 p.m. UTC | #9

Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>> Recently there was a fairly long thread about recoreable hardware page
>>>> faults, how they can deadlock, and what to do about that.
>>>>
>>>> While the discussion is still fresh I figured good time to try and
>>>> document the conclusions a bit.
>>>>
>>>> References:
>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7Cbee0aeff80f440bcc52108d8c2bcc11f%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637473463245588199%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=ncr%2Fqv5lw0ONrYxFvfdcFAXAZ%2BXcJJa6UY%2BxGfcKGVM%3D&reserved=0
>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>> Cc: Thomas Hellström <[e-mail]>
>>>> Cc: "Christian König" <[e-mail]>
>>>> Cc: Jerome Glisse <[e-mail]>
>>>> Cc: Felix Kuehling <[e-mail]>
>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>> Cc: Sumit Semwal <[e-mail]>
>>>> Cc: [e-mail]
>>>> Cc: [e-mail]
>>>> -- 
>>>> I'll be away next week, but figured I'll type this up quickly for some
>>>> comments and to check whether I got this all roughly right.
>>>>
>>>> Critique very much wanted on this, so that we can make sure hw which
>>>> can't preempt (with pagefaults pending) like gfx10 has a clear path to
>>>> support page faults in upstream. So anything I missed, got wrong or
>>>> like that would be good.
>>>> -Daniel
>>>> ---
>>>>   Documentation/driver-api/dma-buf.rst | 66
>>>> ++++++++++++++++++++++++++++
>>>>   1 file changed, 66 insertions(+)
>>>>
>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>> b/Documentation/driver-api/dma-buf.rst
>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>     userspace is allowed to use userspace fencing or long running
>>>> compute
>>>>     workloads. This also means no implicit fencing for shared
>>>> buffers in these
>>>>     cases.
>>>> +
>>>> +Recoverable Hardware Page Faults Implications
>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>> +
>>>> +Modern hardware supports recoverable page faults, which has a lot of
>>>> +implications for DMA fences.
>>>> +
>>>> +First, a pending page fault obviously holds up the work that's
>>>> running on the
>>>> +accelerator and a memory allocation is usually required to resolve
>>>> the fault.
>>>> +But memory allocations are not allowed to gate completion of DMA
>>>> fences, which
>>>> +means any workload using recoverable page faults cannot use DMA
>>>> fences for
>>>> +synchronization. Synchronization fences controlled by userspace
>>>> must be used
>>>> +instead.
>>>> +
>>>> +On GPUs this poses a problem, because current desktop compositor
>>>> protocols on
>>>> +Linus rely on DMA fences, which means without an entirely new
>>>> userspace stack
>>>> +built on top of userspace fences, they cannot benefit from
>>>> recoverable page
>>>> +faults. The exception is when page faults are only used as
>>>> migration hints and
>>>> +never to on-demand fill a memory request. For now this means
>>>> recoverable page
>>>> +faults on GPUs are limited to pure compute workloads.
>>>> +
>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>> rendering and
>>>> +compute side, like compute units or command submission engines. If
>>>> both a 3D
>>>> +job with a DMA fence and a compute workload using recoverable page
>>>> faults are
>>>> +pending they could deadlock:
>>>> +
>>>> +- The 3D workload might need to wait for the compute job to finish
>>>> and release
>>>> +  hardware resources first.
>>>> +
>>>> +- The compute workload might be stuck in a page fault, because the
>>>> memory
>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>> complete.
>>>> +
>>>> +There are a few ways to prevent this problem:
>>>> +
>>>> +- Compute workloads can always be preempted, even when a page
>>>> fault is pending
>>>> +  and not yet repaired. Not all hardware supports this.
>>>> +
>>>> +- DMA fence workloads and workloads which need page fault handling
>>>> have
>>>> +  independent hardware resources to guarantee forward progress.
>>>> This could be
>>>> +  achieved through e.g. through dedicated engines and minimal
>>>> compute unit
>>>> +  reservations for DMA fence workloads.
>>>> +
>>>> +- The reservation approach could be further refined by only
>>>> reserving the
>>>> +  hardware resources for DMA fence workloads when they are
>>>> in-flight. This must
>>>> +  cover the time from when the DMA fence is visible to other
>>>> threads up to
>>>> +  moment when fence is completed through dma_fence_signal().
>>>> +
>>>> +- As a last resort, if the hardware provides no useful reservation
>>>> mechanics,
>>>> +  all workloads must be flushed from the GPU when switching
>>>> between jobs
>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>> means all DMA
>>>> +  fences must complete before a compute job with page fault
>>>> handling can be
>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>> fence can be
>>>> +  made visible anywhere in the system, all compute workloads must
>>>> be preempted
>>>> +  to guarantee all pending GPU page faults are flushed.
>>> I thought of another possible workaround:
>>>
>>>    * Partition the memory. Servicing of page faults will use a separate
>>>      memory pool that can always be allocated from without waiting for
>>>      fences. This includes memory for page tables and memory for
>>>      migrating data to. You may steal memory from other processes that
>>>      can page fault, so no fence waiting is necessary. Being able to
>>>      steal memory at any time also means there are basically no
>>>      out-of-memory situations you need to worry about. Even page tables
>>>      (except the root page directory of each process) can be stolen in
>>>      the worst case.
>> I think 'overcommit' would be a nice way to describe this. But I'm not
>> sure how easy this is to implement in practice. You would basically need
>> to create your own memory manager for this.
>
> Well you would need a completely separate pool for both device as well
> as system memory.
>
> E.g. on boot we say we steal X GB system memory only for HMM.

Why? The GPU driver doesn't need to allocate system memory for HMM.
Migrations to system memory are handled by the kernel's handle_mm_fault
and page allocator and swap logic. It doesn't depend on any fences, so
it cannot deadlock with any GPU driver-managed memory. The GPU driver
gets involved in the MMU notifier to invalidate device page tables. But
that also doesn't need to wait for any fences.

And if the kernel runs out of pageable memory, you're in trouble anyway.
The OOM killer will step in, nothing new there.

Regards,
  Felix


>
>> But from a design point of view, definitely a valid solution.
>
> I think the restriction above makes it pretty much unusable.
>
>> But this looks good, those solutions are definitely the valid options we
>> can choose from.
>
> It's certainly worth noting, yes. And just to make sure that nobody
> has the idea to reserve only device memory.
>
> Christian.
>
>>
>> ~Maarten
>>
>

Christian König Jan. 28, 2021, 7:39 a.m. UTC | #10

Am 27.01.21 um 23:00 schrieb Felix Kuehling:
> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>>> Recently there was a fairly long thread about recoreable hardware page
>>>>> faults, how they can deadlock, and what to do about that.
>>>>>
>>>>> While the discussion is still fresh I figured good time to try and
>>>>> document the conclusions a bit.
>>>>>
>>>>> References:
>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7Cbee0aeff80f440bcc52108d8c2bcc11f%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637473463245588199%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=ncr%2Fqv5lw0ONrYxFvfdcFAXAZ%2BXcJJa6UY%2BxGfcKGVM%3D&reserved=0
>>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>>> Cc: Thomas Hellström <[e-mail]>
>>>>> Cc: "Christian König" <[e-mail]>
>>>>> Cc: Jerome Glisse <[e-mail]>
>>>>> Cc: Felix Kuehling <[e-mail]>
>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>>> Cc: Sumit Semwal <[e-mail]>
>>>>> Cc: [e-mail]
>>>>> Cc: [e-mail]
>>>>> -- 
>>>>> I'll be away next week, but figured I'll type this up quickly for some
>>>>> comments and to check whether I got this all roughly right.
>>>>>
>>>>> Critique very much wanted on this, so that we can make sure hw which
>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear path to
>>>>> support page faults in upstream. So anything I missed, got wrong or
>>>>> like that would be good.
>>>>> -Daniel
>>>>> ---
>>>>>    Documentation/driver-api/dma-buf.rst | 66
>>>>> ++++++++++++++++++++++++++++
>>>>>    1 file changed, 66 insertions(+)
>>>>>
>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>>> b/Documentation/driver-api/dma-buf.rst
>>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>>      userspace is allowed to use userspace fencing or long running
>>>>> compute
>>>>>      workloads. This also means no implicit fencing for shared
>>>>> buffers in these
>>>>>      cases.
>>>>> +
>>>>> +Recoverable Hardware Page Faults Implications
>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>> +
>>>>> +Modern hardware supports recoverable page faults, which has a lot of
>>>>> +implications for DMA fences.
>>>>> +
>>>>> +First, a pending page fault obviously holds up the work that's
>>>>> running on the
>>>>> +accelerator and a memory allocation is usually required to resolve
>>>>> the fault.
>>>>> +But memory allocations are not allowed to gate completion of DMA
>>>>> fences, which
>>>>> +means any workload using recoverable page faults cannot use DMA
>>>>> fences for
>>>>> +synchronization. Synchronization fences controlled by userspace
>>>>> must be used
>>>>> +instead.
>>>>> +
>>>>> +On GPUs this poses a problem, because current desktop compositor
>>>>> protocols on
>>>>> +Linus rely on DMA fences, which means without an entirely new
>>>>> userspace stack
>>>>> +built on top of userspace fences, they cannot benefit from
>>>>> recoverable page
>>>>> +faults. The exception is when page faults are only used as
>>>>> migration hints and
>>>>> +never to on-demand fill a memory request. For now this means
>>>>> recoverable page
>>>>> +faults on GPUs are limited to pure compute workloads.
>>>>> +
>>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>>> rendering and
>>>>> +compute side, like compute units or command submission engines. If
>>>>> both a 3D
>>>>> +job with a DMA fence and a compute workload using recoverable page
>>>>> faults are
>>>>> +pending they could deadlock:
>>>>> +
>>>>> +- The 3D workload might need to wait for the compute job to finish
>>>>> and release
>>>>> +  hardware resources first.
>>>>> +
>>>>> +- The compute workload might be stuck in a page fault, because the
>>>>> memory
>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>>> complete.
>>>>> +
>>>>> +There are a few ways to prevent this problem:
>>>>> +
>>>>> +- Compute workloads can always be preempted, even when a page
>>>>> fault is pending
>>>>> +  and not yet repaired. Not all hardware supports this.
>>>>> +
>>>>> +- DMA fence workloads and workloads which need page fault handling
>>>>> have
>>>>> +  independent hardware resources to guarantee forward progress.
>>>>> This could be
>>>>> +  achieved through e.g. through dedicated engines and minimal
>>>>> compute unit
>>>>> +  reservations for DMA fence workloads.
>>>>> +
>>>>> +- The reservation approach could be further refined by only
>>>>> reserving the
>>>>> +  hardware resources for DMA fence workloads when they are
>>>>> in-flight. This must
>>>>> +  cover the time from when the DMA fence is visible to other
>>>>> threads up to
>>>>> +  moment when fence is completed through dma_fence_signal().
>>>>> +
>>>>> +- As a last resort, if the hardware provides no useful reservation
>>>>> mechanics,
>>>>> +  all workloads must be flushed from the GPU when switching
>>>>> between jobs
>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>>> means all DMA
>>>>> +  fences must complete before a compute job with page fault
>>>>> handling can be
>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>>> fence can be
>>>>> +  made visible anywhere in the system, all compute workloads must
>>>>> be preempted
>>>>> +  to guarantee all pending GPU page faults are flushed.
>>>> I thought of another possible workaround:
>>>>
>>>>     * Partition the memory. Servicing of page faults will use a separate
>>>>       memory pool that can always be allocated from without waiting for
>>>>       fences. This includes memory for page tables and memory for
>>>>       migrating data to. You may steal memory from other processes that
>>>>       can page fault, so no fence waiting is necessary. Being able to
>>>>       steal memory at any time also means there are basically no
>>>>       out-of-memory situations you need to worry about. Even page tables
>>>>       (except the root page directory of each process) can be stolen in
>>>>       the worst case.
>>> I think 'overcommit' would be a nice way to describe this. But I'm not
>>> sure how easy this is to implement in practice. You would basically need
>>> to create your own memory manager for this.
>> Well you would need a completely separate pool for both device as well
>> as system memory.
>>
>> E.g. on boot we say we steal X GB system memory only for HMM.
> Why? The GPU driver doesn't need to allocate system memory for HMM.
> Migrations to system memory are handled by the kernel's handle_mm_fault
> and page allocator and swap logic.

And that one depends on dma_fence completion because you can easily need 
to wait for an MMU notifier callback.

As Maarten wrote when you want to go down this route you need a complete 
separate memory management parallel to the one of the kernel.

Regards,
Christian.

>   It doesn't depend on any fences, so
> it cannot deadlock with any GPU driver-managed memory. The GPU driver
> gets involved in the MMU notifier to invalidate device page tables. But
> that also doesn't need to wait for any fences.
>
> And if the kernel runs out of pageable memory, you're in trouble anyway.
> The OOM killer will step in, nothing new there.
>
> Regards,
>    Felix
>
>
>>> But from a design point of view, definitely a valid solution.
>> I think the restriction above makes it pretty much unusable.
>>
>>> But this looks good, those solutions are definitely the valid options we
>>> can choose from.
>> It's certainly worth noting, yes. And just to make sure that nobody
>> has the idea to reserve only device memory.
>>
>> Christian.
>>
>>> ~Maarten
>>>
> _______________________________________________
> Linaro-mm-sig mailing list
> [e-mail]
> https://lists.linaro.org/mailman/listinfo/linaro-mm-sig

Felix Kuehling Jan. 28, 2021, 3:39 p.m. UTC | #11

Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
> Am 27.01.21 um 23:00 schrieb Felix Kuehling:
>> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
>>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>>>> Recently there was a fairly long thread about recoreable hardware
>>>>>> page
>>>>>> faults, how they can deadlock, and what to do about that.
>>>>>>
>>>>>> While the discussion is still fresh I figured good time to try and
>>>>>> document the conclusions a bit.
>>>>>>
>>>>>> References:
>>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=y2VzC4vbfMi0ctyerAHfqODZ6tthz1FUDwpMCp0PIrQ%3D&reserved=0
>>>>>>
>>>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>>>> Cc: Thomas Hellström <[e-mail]>
>>>>>> Cc: "Christian König" <[e-mail]>
>>>>>> Cc: Jerome Glisse <[e-mail]>
>>>>>> Cc: Felix Kuehling <[e-mail]>
>>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>>>> Cc: Sumit Semwal <[e-mail]>
>>>>>> Cc: [e-mail]
>>>>>> Cc: [e-mail]
>>>>>> -- 
>>>>>> I'll be away next week, but figured I'll type this up quickly for
>>>>>> some
>>>>>> comments and to check whether I got this all roughly right.
>>>>>>
>>>>>> Critique very much wanted on this, so that we can make sure hw which
>>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
>>>>>> path to
>>>>>> support page faults in upstream. So anything I missed, got wrong or
>>>>>> like that would be good.
>>>>>> -Daniel
>>>>>> ---
>>>>>>    Documentation/driver-api/dma-buf.rst | 66
>>>>>> ++++++++++++++++++++++++++++
>>>>>>    1 file changed, 66 insertions(+)
>>>>>>
>>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>>>> b/Documentation/driver-api/dma-buf.rst
>>>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>>>      userspace is allowed to use userspace fencing or long running
>>>>>> compute
>>>>>>      workloads. This also means no implicit fencing for shared
>>>>>> buffers in these
>>>>>>      cases.
>>>>>> +
>>>>>> +Recoverable Hardware Page Faults Implications
>>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>> +
>>>>>> +Modern hardware supports recoverable page faults, which has a
>>>>>> lot of
>>>>>> +implications for DMA fences.
>>>>>> +
>>>>>> +First, a pending page fault obviously holds up the work that's
>>>>>> running on the
>>>>>> +accelerator and a memory allocation is usually required to resolve
>>>>>> the fault.
>>>>>> +But memory allocations are not allowed to gate completion of DMA
>>>>>> fences, which
>>>>>> +means any workload using recoverable page faults cannot use DMA
>>>>>> fences for
>>>>>> +synchronization. Synchronization fences controlled by userspace
>>>>>> must be used
>>>>>> +instead.
>>>>>> +
>>>>>> +On GPUs this poses a problem, because current desktop compositor
>>>>>> protocols on
>>>>>> +Linus rely on DMA fences, which means without an entirely new
>>>>>> userspace stack
>>>>>> +built on top of userspace fences, they cannot benefit from
>>>>>> recoverable page
>>>>>> +faults. The exception is when page faults are only used as
>>>>>> migration hints and
>>>>>> +never to on-demand fill a memory request. For now this means
>>>>>> recoverable page
>>>>>> +faults on GPUs are limited to pure compute workloads.
>>>>>> +
>>>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>>>> rendering and
>>>>>> +compute side, like compute units or command submission engines. If
>>>>>> both a 3D
>>>>>> +job with a DMA fence and a compute workload using recoverable page
>>>>>> faults are
>>>>>> +pending they could deadlock:
>>>>>> +
>>>>>> +- The 3D workload might need to wait for the compute job to finish
>>>>>> and release
>>>>>> +  hardware resources first.
>>>>>> +
>>>>>> +- The compute workload might be stuck in a page fault, because the
>>>>>> memory
>>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>>>> complete.
>>>>>> +
>>>>>> +There are a few ways to prevent this problem:
>>>>>> +
>>>>>> +- Compute workloads can always be preempted, even when a page
>>>>>> fault is pending
>>>>>> +  and not yet repaired. Not all hardware supports this.
>>>>>> +
>>>>>> +- DMA fence workloads and workloads which need page fault handling
>>>>>> have
>>>>>> +  independent hardware resources to guarantee forward progress.
>>>>>> This could be
>>>>>> +  achieved through e.g. through dedicated engines and minimal
>>>>>> compute unit
>>>>>> +  reservations for DMA fence workloads.
>>>>>> +
>>>>>> +- The reservation approach could be further refined by only
>>>>>> reserving the
>>>>>> +  hardware resources for DMA fence workloads when they are
>>>>>> in-flight. This must
>>>>>> +  cover the time from when the DMA fence is visible to other
>>>>>> threads up to
>>>>>> +  moment when fence is completed through dma_fence_signal().
>>>>>> +
>>>>>> +- As a last resort, if the hardware provides no useful reservation
>>>>>> mechanics,
>>>>>> +  all workloads must be flushed from the GPU when switching
>>>>>> between jobs
>>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>>>> means all DMA
>>>>>> +  fences must complete before a compute job with page fault
>>>>>> handling can be
>>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>>>> fence can be
>>>>>> +  made visible anywhere in the system, all compute workloads must
>>>>>> be preempted
>>>>>> +  to guarantee all pending GPU page faults are flushed.
>>>>> I thought of another possible workaround:
>>>>>
>>>>>     * Partition the memory. Servicing of page faults will use a
>>>>> separate
>>>>>       memory pool that can always be allocated from without
>>>>> waiting for
>>>>>       fences. This includes memory for page tables and memory for
>>>>>       migrating data to. You may steal memory from other processes
>>>>> that
>>>>>       can page fault, so no fence waiting is necessary. Being able to
>>>>>       steal memory at any time also means there are basically no
>>>>>       out-of-memory situations you need to worry about. Even page
>>>>> tables
>>>>>       (except the root page directory of each process) can be
>>>>> stolen in
>>>>>       the worst case.
>>>> I think 'overcommit' would be a nice way to describe this. But I'm not
>>>> sure how easy this is to implement in practice. You would basically
>>>> need
>>>> to create your own memory manager for this.
>>> Well you would need a completely separate pool for both device as well
>>> as system memory.
>>>
>>> E.g. on boot we say we steal X GB system memory only for HMM.
>> Why? The GPU driver doesn't need to allocate system memory for HMM.
>> Migrations to system memory are handled by the kernel's handle_mm_fault
>> and page allocator and swap logic.
>
> And that one depends on dma_fence completion because you can easily
> need to wait for an MMU notifier callback.

I see, the GFX MMU notifier for userpointers in amdgpu currently waits
for fences. For the KFD MMU notifier I am planning to fix this by
causing GPU page faults instead of preempting the queues. Can we limit
userptrs in amdgpu to engines that can page fault. Basically make it
illegal to attach userptr BOs to graphics CS BO lists, so they can only
be used in user mode command submissions, which can page fault. Then the
GFX MMU notifier could invalidate PTEs and would not have to wait for
fences.


>
> As Maarten wrote when you want to go down this route you need a
> complete separate memory management parallel to the one of the kernel.

Not really. I'm trying to make the GPU memory management more similar to
what the kernel does for system memory.

I understood Maarten's comment as "I'm creating a new memory manager and
not using TTM any more". This is true. The idea is that this portion of
VRAM would be managed more like system memory.

Regards,
  Felix


>
> Regards,
> Christian.
>
>>   It doesn't depend on any fences, so
>> it cannot deadlock with any GPU driver-managed memory. The GPU driver
>> gets involved in the MMU notifier to invalidate device page tables. But
>> that also doesn't need to wait for any fences.
>>
>> And if the kernel runs out of pageable memory, you're in trouble anyway.
>> The OOM killer will step in, nothing new there.
>>
>> Regards,
>>    Felix
>>
>>
>>>> But from a design point of view, definitely a valid solution.
>>> I think the restriction above makes it pretty much unusable.
>>>
>>>> But this looks good, those solutions are definitely the valid
>>>> options we
>>>> can choose from.
>>> It's certainly worth noting, yes. And just to make sure that nobody
>>> has the idea to reserve only device memory.
>>>
>>> Christian.
>>>
>>>> ~Maarten
>>>>
>> _______________________________________________
>> Linaro-mm-sig mailing list
>> [e-mail]
>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.linaro.org%2Fmailman%2Flistinfo%2Flinaro-mm-sig&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=gQj51eDK8OUWoQcbYliY639jOPleRjyLY3Q16nj2PL0%3D&amp;reserved=0
>>
>

Christian König Jan. 28, 2021, 3:46 p.m. UTC | #12

Am 28.01.21 um 16:39 schrieb Felix Kuehling:
> Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
>> Am 27.01.21 um 23:00 schrieb Felix Kuehling:
>>> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
>>>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>>>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>>>>> Recently there was a fairly long thread about recoreable hardware
>>>>>>> page
>>>>>>> faults, how they can deadlock, and what to do about that.
>>>>>>>
>>>>>>> While the discussion is still fresh I figured good time to try and
>>>>>>> document the conclusions a bit.
>>>>>>>
>>>>>>> References:
>>>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=y2VzC4vbfMi0ctyerAHfqODZ6tthz1FUDwpMCp0PIrQ%3D&reserved=0
>>>>>>>
>>>>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>>>>> Cc: Thomas Hellström <[e-mail]>
>>>>>>> Cc: "Christian König" <[e-mail]>
>>>>>>> Cc: Jerome Glisse <[e-mail]>
>>>>>>> Cc: Felix Kuehling <[e-mail]>
>>>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>>>>> Cc: Sumit Semwal <[e-mail]>
>>>>>>> Cc: [e-mail]
>>>>>>> Cc: [e-mail]
>>>>>>> -- 
>>>>>>> I'll be away next week, but figured I'll type this up quickly for
>>>>>>> some
>>>>>>> comments and to check whether I got this all roughly right.
>>>>>>>
>>>>>>> Critique very much wanted on this, so that we can make sure hw which
>>>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
>>>>>>> path to
>>>>>>> support page faults in upstream. So anything I missed, got wrong or
>>>>>>> like that would be good.
>>>>>>> -Daniel
>>>>>>> ---
>>>>>>>     Documentation/driver-api/dma-buf.rst | 66
>>>>>>> ++++++++++++++++++++++++++++
>>>>>>>     1 file changed, 66 insertions(+)
>>>>>>>
>>>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>>>>> b/Documentation/driver-api/dma-buf.rst
>>>>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>>>>       userspace is allowed to use userspace fencing or long running
>>>>>>> compute
>>>>>>>       workloads. This also means no implicit fencing for shared
>>>>>>> buffers in these
>>>>>>>       cases.
>>>>>>> +
>>>>>>> +Recoverable Hardware Page Faults Implications
>>>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>>> +
>>>>>>> +Modern hardware supports recoverable page faults, which has a
>>>>>>> lot of
>>>>>>> +implications for DMA fences.
>>>>>>> +
>>>>>>> +First, a pending page fault obviously holds up the work that's
>>>>>>> running on the
>>>>>>> +accelerator and a memory allocation is usually required to resolve
>>>>>>> the fault.
>>>>>>> +But memory allocations are not allowed to gate completion of DMA
>>>>>>> fences, which
>>>>>>> +means any workload using recoverable page faults cannot use DMA
>>>>>>> fences for
>>>>>>> +synchronization. Synchronization fences controlled by userspace
>>>>>>> must be used
>>>>>>> +instead.
>>>>>>> +
>>>>>>> +On GPUs this poses a problem, because current desktop compositor
>>>>>>> protocols on
>>>>>>> +Linus rely on DMA fences, which means without an entirely new
>>>>>>> userspace stack
>>>>>>> +built on top of userspace fences, they cannot benefit from
>>>>>>> recoverable page
>>>>>>> +faults. The exception is when page faults are only used as
>>>>>>> migration hints and
>>>>>>> +never to on-demand fill a memory request. For now this means
>>>>>>> recoverable page
>>>>>>> +faults on GPUs are limited to pure compute workloads.
>>>>>>> +
>>>>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>>>>> rendering and
>>>>>>> +compute side, like compute units or command submission engines. If
>>>>>>> both a 3D
>>>>>>> +job with a DMA fence and a compute workload using recoverable page
>>>>>>> faults are
>>>>>>> +pending they could deadlock:
>>>>>>> +
>>>>>>> +- The 3D workload might need to wait for the compute job to finish
>>>>>>> and release
>>>>>>> +  hardware resources first.
>>>>>>> +
>>>>>>> +- The compute workload might be stuck in a page fault, because the
>>>>>>> memory
>>>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>>>>> complete.
>>>>>>> +
>>>>>>> +There are a few ways to prevent this problem:
>>>>>>> +
>>>>>>> +- Compute workloads can always be preempted, even when a page
>>>>>>> fault is pending
>>>>>>> +  and not yet repaired. Not all hardware supports this.
>>>>>>> +
>>>>>>> +- DMA fence workloads and workloads which need page fault handling
>>>>>>> have
>>>>>>> +  independent hardware resources to guarantee forward progress.
>>>>>>> This could be
>>>>>>> +  achieved through e.g. through dedicated engines and minimal
>>>>>>> compute unit
>>>>>>> +  reservations for DMA fence workloads.
>>>>>>> +
>>>>>>> +- The reservation approach could be further refined by only
>>>>>>> reserving the
>>>>>>> +  hardware resources for DMA fence workloads when they are
>>>>>>> in-flight. This must
>>>>>>> +  cover the time from when the DMA fence is visible to other
>>>>>>> threads up to
>>>>>>> +  moment when fence is completed through dma_fence_signal().
>>>>>>> +
>>>>>>> +- As a last resort, if the hardware provides no useful reservation
>>>>>>> mechanics,
>>>>>>> +  all workloads must be flushed from the GPU when switching
>>>>>>> between jobs
>>>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>>>>> means all DMA
>>>>>>> +  fences must complete before a compute job with page fault
>>>>>>> handling can be
>>>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>>>>> fence can be
>>>>>>> +  made visible anywhere in the system, all compute workloads must
>>>>>>> be preempted
>>>>>>> +  to guarantee all pending GPU page faults are flushed.
>>>>>> I thought of another possible workaround:
>>>>>>
>>>>>>      * Partition the memory. Servicing of page faults will use a
>>>>>> separate
>>>>>>        memory pool that can always be allocated from without
>>>>>> waiting for
>>>>>>        fences. This includes memory for page tables and memory for
>>>>>>        migrating data to. You may steal memory from other processes
>>>>>> that
>>>>>>        can page fault, so no fence waiting is necessary. Being able to
>>>>>>        steal memory at any time also means there are basically no
>>>>>>        out-of-memory situations you need to worry about. Even page
>>>>>> tables
>>>>>>        (except the root page directory of each process) can be
>>>>>> stolen in
>>>>>>        the worst case.
>>>>> I think 'overcommit' would be a nice way to describe this. But I'm not
>>>>> sure how easy this is to implement in practice. You would basically
>>>>> need
>>>>> to create your own memory manager for this.
>>>> Well you would need a completely separate pool for both device as well
>>>> as system memory.
>>>>
>>>> E.g. on boot we say we steal X GB system memory only for HMM.
>>> Why? The GPU driver doesn't need to allocate system memory for HMM.
>>> Migrations to system memory are handled by the kernel's handle_mm_fault
>>> and page allocator and swap logic.
>> And that one depends on dma_fence completion because you can easily
>> need to wait for an MMU notifier callback.
> I see, the GFX MMU notifier for userpointers in amdgpu currently waits
> for fences. For the KFD MMU notifier I am planning to fix this by
> causing GPU page faults instead of preempting the queues. Can we limit
> userptrs in amdgpu to engines that can page fault. Basically make it
> illegal to attach userptr BOs to graphics CS BO lists, so they can only
> be used in user mode command submissions, which can page fault. Then the
> GFX MMU notifier could invalidate PTEs and would not have to wait for
> fences.

It's not only the MMU notifier, the TTM shrinker I'm adding needs to 
wait for dma_fences as well.

And apart from that we can't limit userptrs since they are part of the 
UAPI and Vulkan/OpenGL.

>> As Maarten wrote when you want to go down this route you need a
>> complete separate memory management parallel to the one of the kernel.
> Not really. I'm trying to make the GPU memory management more similar to
> what the kernel does for system memory.
>
> I understood Maarten's comment as "I'm creating a new memory manager and
> not using TTM any more". This is true. The idea is that this portion of
> VRAM would be managed more like system memory.

I don't think that will fly. We can have the backing store which TTM 
uses for allocation shared with HMM.

But essentially both TTM allocations needs to be able to put pressure on 
HMM allocations as well as the other way around.

Regards,
Christian.

>
> Regards,
>    Felix
>
>
>> Regards,
>> Christian.
>>
>>>    It doesn't depend on any fences, so
>>> it cannot deadlock with any GPU driver-managed memory. The GPU driver
>>> gets involved in the MMU notifier to invalidate device page tables. But
>>> that also doesn't need to wait for any fences.
>>>
>>> And if the kernel runs out of pageable memory, you're in trouble anyway.
>>> The OOM killer will step in, nothing new there.
>>>
>>> Regards,
>>>     Felix
>>>
>>>
>>>>> But from a design point of view, definitely a valid solution.
>>>> I think the restriction above makes it pretty much unusable.
>>>>
>>>>> But this looks good, those solutions are definitely the valid
>>>>> options we
>>>>> can choose from.
>>>> It's certainly worth noting, yes. And just to make sure that nobody
>>>> has the idea to reserve only device memory.
>>>>
>>>> Christian.
>>>>
>>>>> ~Maarten
>>>>>
>>> _______________________________________________
>>> Linaro-mm-sig mailing list
>>> [e-mail]
>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.linaro.org%2Fmailman%2Flistinfo%2Flinaro-mm-sig&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=gQj51eDK8OUWoQcbYliY639jOPleRjyLY3Q16nj2PL0%3D&amp;reserved=0
>>>

Daniel Vetter Feb. 2, 2021, 3:38 p.m. UTC | #13

Back from vacations.

On Thu, Jan 28, 2021 at 04:46:55PM +0100, Christian König wrote:
> Am 28.01.21 um 16:39 schrieb Felix Kuehling:
> > Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
> > > Am 27.01.21 um 23:00 schrieb Felix Kuehling:
> > > > Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
> > > > > Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
> > > > > > Op 27-01-2021 om 01:22 schreef Felix Kuehling:
> > > > > > > Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> > > > > > > > Recently there was a fairly long thread about recoreable hardware
> > > > > > > > page
> > > > > > > > faults, how they can deadlock, and what to do about that.
> > > > > > > > 
> > > > > > > > While the discussion is still fresh I figured good time to try and
> > > > > > > > document the conclusions a bit.
> > > > > > > > 
> > > > > > > > References:
> > > > > > > > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=y2VzC4vbfMi0ctyerAHfqODZ6tthz1FUDwpMCp0PIrQ%3D&reserved=0
> > > > > > > > 
> > > > > > > > Cc: Maarten Lankhorst <[e-mail]>
> > > > > > > > Cc: Thomas Hellström <[e-mail]>
> > > > > > > > Cc: "Christian König" <[e-mail]>
> > > > > > > > Cc: Jerome Glisse <[e-mail]>
> > > > > > > > Cc: Felix Kuehling <[e-mail]>
> > > > > > > > Signed-off-by: Daniel Vetter <[e-mail]>
> > > > > > > > Cc: Sumit Semwal <[e-mail]>
> > > > > > > > Cc: [e-mail]
> > > > > > > > Cc: [e-mail]
> > > > > > > > -- 
> > > > > > > > I'll be away next week, but figured I'll type this up quickly for
> > > > > > > > some
> > > > > > > > comments and to check whether I got this all roughly right.
> > > > > > > > 
> > > > > > > > Critique very much wanted on this, so that we can make sure hw which
> > > > > > > > can't preempt (with pagefaults pending) like gfx10 has a clear
> > > > > > > > path to
> > > > > > > > support page faults in upstream. So anything I missed, got wrong or
> > > > > > > > like that would be good.
> > > > > > > > -Daniel
> > > > > > > > ---
> > > > > > > >     Documentation/driver-api/dma-buf.rst | 66
> > > > > > > > ++++++++++++++++++++++++++++
> > > > > > > >     1 file changed, 66 insertions(+)
> > > > > > > > 
> > > > > > > > diff --git a/Documentation/driver-api/dma-buf.rst
> > > > > > > > b/Documentation/driver-api/dma-buf.rst
> > > > > > > > index a2133d69872c..e924c1e4f7a3 100644
> > > > > > > > --- a/Documentation/driver-api/dma-buf.rst
> > > > > > > > +++ b/Documentation/driver-api/dma-buf.rst
> > > > > > > > @@ -257,3 +257,69 @@ fences in the kernel. This means:
> > > > > > > >       userspace is allowed to use userspace fencing or long running
> > > > > > > > compute
> > > > > > > >       workloads. This also means no implicit fencing for shared
> > > > > > > > buffers in these
> > > > > > > >       cases.
> > > > > > > > +
> > > > > > > > +Recoverable Hardware Page Faults Implications
> > > > > > > > +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > > > > > > +
> > > > > > > > +Modern hardware supports recoverable page faults, which has a
> > > > > > > > lot of
> > > > > > > > +implications for DMA fences.
> > > > > > > > +
> > > > > > > > +First, a pending page fault obviously holds up the work that's
> > > > > > > > running on the
> > > > > > > > +accelerator and a memory allocation is usually required to resolve
> > > > > > > > the fault.
> > > > > > > > +But memory allocations are not allowed to gate completion of DMA
> > > > > > > > fences, which
> > > > > > > > +means any workload using recoverable page faults cannot use DMA
> > > > > > > > fences for
> > > > > > > > +synchronization. Synchronization fences controlled by userspace
> > > > > > > > must be used
> > > > > > > > +instead.
> > > > > > > > +
> > > > > > > > +On GPUs this poses a problem, because current desktop compositor
> > > > > > > > protocols on
> > > > > > > > +Linus rely on DMA fences, which means without an entirely new
> > > > > > > > userspace stack
> > > > > > > > +built on top of userspace fences, they cannot benefit from
> > > > > > > > recoverable page
> > > > > > > > +faults. The exception is when page faults are only used as
> > > > > > > > migration hints and
> > > > > > > > +never to on-demand fill a memory request. For now this means
> > > > > > > > recoverable page
> > > > > > > > +faults on GPUs are limited to pure compute workloads.
> > > > > > > > +
> > > > > > > > +Furthermore GPUs usually have shared resources between the 3D
> > > > > > > > rendering and
> > > > > > > > +compute side, like compute units or command submission engines. If
> > > > > > > > both a 3D
> > > > > > > > +job with a DMA fence and a compute workload using recoverable page
> > > > > > > > faults are
> > > > > > > > +pending they could deadlock:
> > > > > > > > +
> > > > > > > > +- The 3D workload might need to wait for the compute job to finish
> > > > > > > > and release
> > > > > > > > +  hardware resources first.
> > > > > > > > +
> > > > > > > > +- The compute workload might be stuck in a page fault, because the
> > > > > > > > memory
> > > > > > > > +  allocation is waiting for the DMA fence of the 3D workload to
> > > > > > > > complete.
> > > > > > > > +
> > > > > > > > +There are a few ways to prevent this problem:
> > > > > > > > +
> > > > > > > > +- Compute workloads can always be preempted, even when a page
> > > > > > > > fault is pending
> > > > > > > > +  and not yet repaired. Not all hardware supports this.
> > > > > > > > +
> > > > > > > > +- DMA fence workloads and workloads which need page fault handling
> > > > > > > > have
> > > > > > > > +  independent hardware resources to guarantee forward progress.
> > > > > > > > This could be
> > > > > > > > +  achieved through e.g. through dedicated engines and minimal
> > > > > > > > compute unit
> > > > > > > > +  reservations for DMA fence workloads.
> > > > > > > > +
> > > > > > > > +- The reservation approach could be further refined by only
> > > > > > > > reserving the
> > > > > > > > +  hardware resources for DMA fence workloads when they are
> > > > > > > > in-flight. This must
> > > > > > > > +  cover the time from when the DMA fence is visible to other
> > > > > > > > threads up to
> > > > > > > > +  moment when fence is completed through dma_fence_signal().
> > > > > > > > +
> > > > > > > > +- As a last resort, if the hardware provides no useful reservation
> > > > > > > > mechanics,
> > > > > > > > +  all workloads must be flushed from the GPU when switching
> > > > > > > > between jobs
> > > > > > > > +  requiring DMA fences or jobs requiring page fault handling: This
> > > > > > > > means all DMA
> > > > > > > > +  fences must complete before a compute job with page fault
> > > > > > > > handling can be
> > > > > > > > +  inserted into the scheduler queue. And vice versa, before a DMA
> > > > > > > > fence can be
> > > > > > > > +  made visible anywhere in the system, all compute workloads must
> > > > > > > > be preempted
> > > > > > > > +  to guarantee all pending GPU page faults are flushed.
> > > > > > > I thought of another possible workaround:
> > > > > > > 
> > > > > > >      * Partition the memory. Servicing of page faults will use a
> > > > > > > separate
> > > > > > >        memory pool that can always be allocated from without
> > > > > > > waiting for
> > > > > > >        fences. This includes memory for page tables and memory for
> > > > > > >        migrating data to. You may steal memory from other processes
> > > > > > > that
> > > > > > >        can page fault, so no fence waiting is necessary. Being able to
> > > > > > >        steal memory at any time also means there are basically no
> > > > > > >        out-of-memory situations you need to worry about. Even page
> > > > > > > tables
> > > > > > >        (except the root page directory of each process) can be
> > > > > > > stolen in
> > > > > > >        the worst case.
> > > > > > I think 'overcommit' would be a nice way to describe this. But I'm not
> > > > > > sure how easy this is to implement in practice. You would basically
> > > > > > need
> > > > > > to create your own memory manager for this.

Yeah when we discussed this at intel we've come across this one too, but
for the practical reasons laid out below this one is going to be very
hard.

Some more of the things I've dug out when looking into whether this is
feasible below.

> > > > > Well you would need a completely separate pool for both device as well
> > > > > as system memory.
> > > > > 
> > > > > E.g. on boot we say we steal X GB system memory only for HMM.
> > > > Why? The GPU driver doesn't need to allocate system memory for HMM.
> > > > Migrations to system memory are handled by the kernel's handle_mm_fault
> > > > and page allocator and swap logic.
> > > And that one depends on dma_fence completion because you can easily
> > > need to wait for an MMU notifier callback.
> > I see, the GFX MMU notifier for userpointers in amdgpu currently waits
> > for fences. For the KFD MMU notifier I am planning to fix this by
> > causing GPU page faults instead of preempting the queues. Can we limit
> > userptrs in amdgpu to engines that can page fault. Basically make it
> > illegal to attach userptr BOs to graphics CS BO lists, so they can only
> > be used in user mode command submissions, which can page fault. Then the
> > GFX MMU notifier could invalidate PTEs and would not have to wait for
> > fences.
> 
> It's not only the MMU notifier, the TTM shrinker I'm adding needs to wait
> for dma_fences as well.
> 
> And apart from that we can't limit userptrs since they are part of the UAPI
> and Vulkan/OpenGL.

So when I looked I noticed that ->mmap has already a GFP flag, but it
seems largely defunct. It's in struct vm_fault.gfp_mask.

We could also set a PF thread flag somehow to limit this.

But the real risk I'm seeing is that this means we're running the entire
page faulthandler from any fs/driver/whatever under a more limited memory
allocation policy, and experience from other areas says that's very
fragile and prone to blow up real bad. Other examples are loopback block
device (running file i/o under GFP_NOIO because it's a block device) or
nfs, which runs the network stack under GFP_NOFS. I've chatted with some
fs people, and they strongly recommend against these kind of magic
"everything I call here has a limited memory allocation scope" tricks.

That's why I didn't bring it up, but I think for completeness I can
mention this and explain why it's very hard to implement and probably not
going to happen.

> > > As Maarten wrote when you want to go down this route you need a
> > > complete separate memory management parallel to the one of the kernel.
> > Not really. I'm trying to make the GPU memory management more similar to
> > what the kernel does for system memory.
> > 
> > I understood Maarten's comment as "I'm creating a new memory manager and
> > not using TTM any more". This is true. The idea is that this portion of
> > VRAM would be managed more like system memory.
> 
> I don't think that will fly. We can have the backing store which TTM uses
> for allocation shared with HMM.
> 
> But essentially both TTM allocations needs to be able to put pressure on HMM
> allocations as well as the other way around.

Yeah that's another reason why I think full split isn't good, as soon as
you run desktop stuff with mixed workload we want the 2 worlds to press
against each another and figure out a fair memory split. Also when we go
into stuff like cgroups I don't think users want to manage these 2 worlds
explicitly, especially if we want to keep the road open to transition vk
(and maybe also gl/libva) over to the explicit userspace fencing world.

I'll try and respin the patch with the suggestion from Christian and this
thread here address and then resend the patch.

Cheers, Daniel

> 
> Regards,
> Christian.
> 
> > 
> > Regards,
> >    Felix
> > 
> > 
> > > Regards,
> > > Christian.
> > > 
> > > >    It doesn't depend on any fences, so
> > > > it cannot deadlock with any GPU driver-managed memory. The GPU driver
> > > > gets involved in the MMU notifier to invalidate device page tables. But
> > > > that also doesn't need to wait for any fences.
> > > > 
> > > > And if the kernel runs out of pageable memory, you're in trouble anyway.
> > > > The OOM killer will step in, nothing new there.
> > > > 
> > > > Regards,
> > > >     Felix
> > > > 
> > > > 
> > > > > > But from a design point of view, definitely a valid solution.
> > > > > I think the restriction above makes it pretty much unusable.
> > > > > 
> > > > > > But this looks good, those solutions are definitely the valid
> > > > > > options we
> > > > > > can choose from.
> > > > > It's certainly worth noting, yes. And just to make sure that nobody
> > > > > has the idea to reserve only device memory.
> > > > > 
> > > > > Christian.
> > > > > 
> > > > > > ~Maarten
> > > > > > 
> > > > _______________________________________________
> > > > Linaro-mm-sig mailing list
> > > > [e-mail]
> > > > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.linaro.org%2Fmailman%2Flistinfo%2Flinaro-mm-sig&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=gQj51eDK8OUWoQcbYliY639jOPleRjyLY3Q16nj2PL0%3D&amp;reserved=0
> > > > 
>

Bas Nieuwenhuizen Feb. 9, 2021, 3:13 a.m. UTC | #14

On Thu, Jan 28, 2021 at 4:40 PM Felix Kuehling <[e-mail]> wrote:
>
> Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
> > Am 27.01.21 um 23:00 schrieb Felix Kuehling:
> >> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
> >>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
> >>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
> >>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> >>>>>> Recently there was a fairly long thread about recoreable hardware
> >>>>>> page
> >>>>>> faults, how they can deadlock, and what to do about that.
> >>>>>>
> >>>>>> While the discussion is still fresh I figured good time to try and
> >>>>>> document the conclusions a bit.
> >>>>>>
> >>>>>> References:
> >>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=y2VzC4vbfMi0ctyerAHfqODZ6tthz1FUDwpMCp0PIrQ%3D&amp;reserved=0
> >>>>>>
> >>>>>> Cc: Maarten Lankhorst <[e-mail]>
> >>>>>> Cc: Thomas Hellström <[e-mail]>
> >>>>>> Cc: "Christian König" <[e-mail]>
> >>>>>> Cc: Jerome Glisse <[e-mail]>
> >>>>>> Cc: Felix Kuehling <[e-mail]>
> >>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
> >>>>>> Cc: Sumit Semwal <[e-mail]>
> >>>>>> Cc: [e-mail]
> >>>>>> Cc: [e-mail]
> >>>>>> --
> >>>>>> I'll be away next week, but figured I'll type this up quickly for
> >>>>>> some
> >>>>>> comments and to check whether I got this all roughly right.
> >>>>>>
> >>>>>> Critique very much wanted on this, so that we can make sure hw which
> >>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
> >>>>>> path to
> >>>>>> support page faults in upstream. So anything I missed, got wrong or
> >>>>>> like that would be good.
> >>>>>> -Daniel
> >>>>>> ---
> >>>>>>    Documentation/driver-api/dma-buf.rst | 66
> >>>>>> ++++++++++++++++++++++++++++
> >>>>>>    1 file changed, 66 insertions(+)
> >>>>>>
> >>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
> >>>>>> b/Documentation/driver-api/dma-buf.rst
> >>>>>> index a2133d69872c..e924c1e4f7a3 100644
> >>>>>> --- a/Documentation/driver-api/dma-buf.rst
> >>>>>> +++ b/Documentation/driver-api/dma-buf.rst
> >>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
> >>>>>>      userspace is allowed to use userspace fencing or long running
> >>>>>> compute
> >>>>>>      workloads. This also means no implicit fencing for shared
> >>>>>> buffers in these
> >>>>>>      cases.
> >>>>>> +
> >>>>>> +Recoverable Hardware Page Faults Implications
> >>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>>> +
> >>>>>> +Modern hardware supports recoverable page faults, which has a
> >>>>>> lot of
> >>>>>> +implications for DMA fences.
> >>>>>> +
> >>>>>> +First, a pending page fault obviously holds up the work that's
> >>>>>> running on the
> >>>>>> +accelerator and a memory allocation is usually required to resolve
> >>>>>> the fault.
> >>>>>> +But memory allocations are not allowed to gate completion of DMA
> >>>>>> fences, which
> >>>>>> +means any workload using recoverable page faults cannot use DMA
> >>>>>> fences for
> >>>>>> +synchronization. Synchronization fences controlled by userspace
> >>>>>> must be used
> >>>>>> +instead.
> >>>>>> +
> >>>>>> +On GPUs this poses a problem, because current desktop compositor
> >>>>>> protocols on
> >>>>>> +Linus rely on DMA fences, which means without an entirely new
> >>>>>> userspace stack
> >>>>>> +built on top of userspace fences, they cannot benefit from
> >>>>>> recoverable page
> >>>>>> +faults. The exception is when page faults are only used as
> >>>>>> migration hints and
> >>>>>> +never to on-demand fill a memory request. For now this means
> >>>>>> recoverable page
> >>>>>> +faults on GPUs are limited to pure compute workloads.
> >>>>>> +
> >>>>>> +Furthermore GPUs usually have shared resources between the 3D
> >>>>>> rendering and
> >>>>>> +compute side, like compute units or command submission engines. If
> >>>>>> both a 3D
> >>>>>> +job with a DMA fence and a compute workload using recoverable page
> >>>>>> faults are
> >>>>>> +pending they could deadlock:
> >>>>>> +
> >>>>>> +- The 3D workload might need to wait for the compute job to finish
> >>>>>> and release
> >>>>>> +  hardware resources first.
> >>>>>> +
> >>>>>> +- The compute workload might be stuck in a page fault, because the
> >>>>>> memory
> >>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
> >>>>>> complete.
> >>>>>> +
> >>>>>> +There are a few ways to prevent this problem:
> >>>>>> +
> >>>>>> +- Compute workloads can always be preempted, even when a page
> >>>>>> fault is pending
> >>>>>> +  and not yet repaired. Not all hardware supports this.
> >>>>>> +
> >>>>>> +- DMA fence workloads and workloads which need page fault handling
> >>>>>> have
> >>>>>> +  independent hardware resources to guarantee forward progress.
> >>>>>> This could be
> >>>>>> +  achieved through e.g. through dedicated engines and minimal
> >>>>>> compute unit
> >>>>>> +  reservations for DMA fence workloads.
> >>>>>> +
> >>>>>> +- The reservation approach could be further refined by only
> >>>>>> reserving the
> >>>>>> +  hardware resources for DMA fence workloads when they are
> >>>>>> in-flight. This must
> >>>>>> +  cover the time from when the DMA fence is visible to other
> >>>>>> threads up to
> >>>>>> +  moment when fence is completed through dma_fence_signal().
> >>>>>> +
> >>>>>> +- As a last resort, if the hardware provides no useful reservation
> >>>>>> mechanics,
> >>>>>> +  all workloads must be flushed from the GPU when switching
> >>>>>> between jobs
> >>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
> >>>>>> means all DMA
> >>>>>> +  fences must complete before a compute job with page fault
> >>>>>> handling can be
> >>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
> >>>>>> fence can be
> >>>>>> +  made visible anywhere in the system, all compute workloads must
> >>>>>> be preempted
> >>>>>> +  to guarantee all pending GPU page faults are flushed.
> >>>>> I thought of another possible workaround:
> >>>>>
> >>>>>     * Partition the memory. Servicing of page faults will use a
> >>>>> separate
> >>>>>       memory pool that can always be allocated from without
> >>>>> waiting for
> >>>>>       fences. This includes memory for page tables and memory for
> >>>>>       migrating data to. You may steal memory from other processes
> >>>>> that
> >>>>>       can page fault, so no fence waiting is necessary. Being able to
> >>>>>       steal memory at any time also means there are basically no
> >>>>>       out-of-memory situations you need to worry about. Even page
> >>>>> tables
> >>>>>       (except the root page directory of each process) can be
> >>>>> stolen in
> >>>>>       the worst case.
> >>>> I think 'overcommit' would be a nice way to describe this. But I'm not
> >>>> sure how easy this is to implement in practice. You would basically
> >>>> need
> >>>> to create your own memory manager for this.
> >>> Well you would need a completely separate pool for both device as well
> >>> as system memory.
> >>>
> >>> E.g. on boot we say we steal X GB system memory only for HMM.
> >> Why? The GPU driver doesn't need to allocate system memory for HMM.
> >> Migrations to system memory are handled by the kernel's handle_mm_fault
> >> and page allocator and swap logic.
> >
> > And that one depends on dma_fence completion because you can easily
> > need to wait for an MMU notifier callback.
>
> I see, the GFX MMU notifier for userpointers in amdgpu currently waits
> for fences. For the KFD MMU notifier I am planning to fix this by
> causing GPU page faults instead of preempting the queues. Can we limit
> userptrs in amdgpu to engines that can page fault. Basically make it
> illegal to attach userptr BOs to graphics CS BO lists, so they can only
> be used in user mode command submissions, which can page fault. Then the
> GFX MMU notifier could invalidate PTEs and would not have to wait for
> fences.

sadly graphics + userptr is already exposed via Mesa.

>
>
> >
> > As Maarten wrote when you want to go down this route you need a
> > complete separate memory management parallel to the one of the kernel.
>
> Not really. I'm trying to make the GPU memory management more similar to
> what the kernel does for system memory.
>
> I understood Maarten's comment as "I'm creating a new memory manager and
> not using TTM any more". This is true. The idea is that this portion of
> VRAM would be managed more like system memory.
>
> Regards,
>   Felix
>
>
> >
> > Regards,
> > Christian.
> >
> >>   It doesn't depend on any fences, so
> >> it cannot deadlock with any GPU driver-managed memory. The GPU driver
> >> gets involved in the MMU notifier to invalidate device page tables. But
> >> that also doesn't need to wait for any fences.
> >>
> >> And if the kernel runs out of pageable memory, you're in trouble anyway.
> >> The OOM killer will step in, nothing new there.
> >>
> >> Regards,
> >>    Felix
> >>
> >>
> >>>> But from a design point of view, definitely a valid solution.
> >>> I think the restriction above makes it pretty much unusable.
> >>>
> >>>> But this looks good, those solutions are definitely the valid
> >>>> options we
> >>>> can choose from.
> >>> It's certainly worth noting, yes. And just to make sure that nobody
> >>> has the idea to reserve only device memory.
> >>>
> >>> Christian.
> >>>
> >>>> ~Maarten
> >>>>
> >> _______________________________________________
> >> Linaro-mm-sig mailing list
> >> [e-mail]
> >> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.linaro.org%2Fmailman%2Flistinfo%2Flinaro-mm-sig&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=gQj51eDK8OUWoQcbYliY639jOPleRjyLY3Q16nj2PL0%3D&amp;reserved=0
> >>
> >
> _______________________________________________
> dri-devel mailing list
> [e-mail]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

Daniel Vetter Feb. 9, 2021, 6:37 a.m. UTC | #15

On Tue, Feb 9, 2021 at 4:13 AM Bas Nieuwenhuizen
<[e-mail]> wrote:
>
> On Thu, Jan 28, 2021 at 4:40 PM Felix Kuehling <[e-mail]> wrote:
> >
> > Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
> > > Am 27.01.21 um 23:00 schrieb Felix Kuehling:
> > >> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
> > >>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
> > >>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
> > >>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> > >>>>>> Recently there was a fairly long thread about recoreable hardware
> > >>>>>> page
> > >>>>>> faults, how they can deadlock, and what to do about that.
> > >>>>>>
> > >>>>>> While the discussion is still fresh I figured good time to try and
> > >>>>>> document the conclusions a bit.
> > >>>>>>
> > >>>>>> References:
> > >>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F20210107030127.20393-1-Felix.Kuehling%40amd.com%2F&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=y2VzC4vbfMi0ctyerAHfqODZ6tthz1FUDwpMCp0PIrQ%3D&amp;reserved=0
> > >>>>>>
> > >>>>>> Cc: Maarten Lankhorst <[e-mail]>
> > >>>>>> Cc: Thomas Hellström <[e-mail]>
> > >>>>>> Cc: "Christian König" <[e-mail]>
> > >>>>>> Cc: Jerome Glisse <[e-mail]>
> > >>>>>> Cc: Felix Kuehling <[e-mail]>
> > >>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
> > >>>>>> Cc: Sumit Semwal <[e-mail]>
> > >>>>>> Cc: [e-mail]
> > >>>>>> Cc: [e-mail]
> > >>>>>> --
> > >>>>>> I'll be away next week, but figured I'll type this up quickly for
> > >>>>>> some
> > >>>>>> comments and to check whether I got this all roughly right.
> > >>>>>>
> > >>>>>> Critique very much wanted on this, so that we can make sure hw which
> > >>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
> > >>>>>> path to
> > >>>>>> support page faults in upstream. So anything I missed, got wrong or
> > >>>>>> like that would be good.
> > >>>>>> -Daniel
> > >>>>>> ---
> > >>>>>>    Documentation/driver-api/dma-buf.rst | 66
> > >>>>>> ++++++++++++++++++++++++++++
> > >>>>>>    1 file changed, 66 insertions(+)
> > >>>>>>
> > >>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
> > >>>>>> b/Documentation/driver-api/dma-buf.rst
> > >>>>>> index a2133d69872c..e924c1e4f7a3 100644
> > >>>>>> --- a/Documentation/driver-api/dma-buf.rst
> > >>>>>> +++ b/Documentation/driver-api/dma-buf.rst
> > >>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
> > >>>>>>      userspace is allowed to use userspace fencing or long running
> > >>>>>> compute
> > >>>>>>      workloads. This also means no implicit fencing for shared
> > >>>>>> buffers in these
> > >>>>>>      cases.
> > >>>>>> +
> > >>>>>> +Recoverable Hardware Page Faults Implications
> > >>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > >>>>>> +
> > >>>>>> +Modern hardware supports recoverable page faults, which has a
> > >>>>>> lot of
> > >>>>>> +implications for DMA fences.
> > >>>>>> +
> > >>>>>> +First, a pending page fault obviously holds up the work that's
> > >>>>>> running on the
> > >>>>>> +accelerator and a memory allocation is usually required to resolve
> > >>>>>> the fault.
> > >>>>>> +But memory allocations are not allowed to gate completion of DMA
> > >>>>>> fences, which
> > >>>>>> +means any workload using recoverable page faults cannot use DMA
> > >>>>>> fences for
> > >>>>>> +synchronization. Synchronization fences controlled by userspace
> > >>>>>> must be used
> > >>>>>> +instead.
> > >>>>>> +
> > >>>>>> +On GPUs this poses a problem, because current desktop compositor
> > >>>>>> protocols on
> > >>>>>> +Linus rely on DMA fences, which means without an entirely new
> > >>>>>> userspace stack
> > >>>>>> +built on top of userspace fences, they cannot benefit from
> > >>>>>> recoverable page
> > >>>>>> +faults. The exception is when page faults are only used as
> > >>>>>> migration hints and
> > >>>>>> +never to on-demand fill a memory request. For now this means
> > >>>>>> recoverable page
> > >>>>>> +faults on GPUs are limited to pure compute workloads.
> > >>>>>> +
> > >>>>>> +Furthermore GPUs usually have shared resources between the 3D
> > >>>>>> rendering and
> > >>>>>> +compute side, like compute units or command submission engines. If
> > >>>>>> both a 3D
> > >>>>>> +job with a DMA fence and a compute workload using recoverable page
> > >>>>>> faults are
> > >>>>>> +pending they could deadlock:
> > >>>>>> +
> > >>>>>> +- The 3D workload might need to wait for the compute job to finish
> > >>>>>> and release
> > >>>>>> +  hardware resources first.
> > >>>>>> +
> > >>>>>> +- The compute workload might be stuck in a page fault, because the
> > >>>>>> memory
> > >>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
> > >>>>>> complete.
> > >>>>>> +
> > >>>>>> +There are a few ways to prevent this problem:
> > >>>>>> +
> > >>>>>> +- Compute workloads can always be preempted, even when a page
> > >>>>>> fault is pending
> > >>>>>> +  and not yet repaired. Not all hardware supports this.
> > >>>>>> +
> > >>>>>> +- DMA fence workloads and workloads which need page fault handling
> > >>>>>> have
> > >>>>>> +  independent hardware resources to guarantee forward progress.
> > >>>>>> This could be
> > >>>>>> +  achieved through e.g. through dedicated engines and minimal
> > >>>>>> compute unit
> > >>>>>> +  reservations for DMA fence workloads.
> > >>>>>> +
> > >>>>>> +- The reservation approach could be further refined by only
> > >>>>>> reserving the
> > >>>>>> +  hardware resources for DMA fence workloads when they are
> > >>>>>> in-flight. This must
> > >>>>>> +  cover the time from when the DMA fence is visible to other
> > >>>>>> threads up to
> > >>>>>> +  moment when fence is completed through dma_fence_signal().
> > >>>>>> +
> > >>>>>> +- As a last resort, if the hardware provides no useful reservation
> > >>>>>> mechanics,
> > >>>>>> +  all workloads must be flushed from the GPU when switching
> > >>>>>> between jobs
> > >>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
> > >>>>>> means all DMA
> > >>>>>> +  fences must complete before a compute job with page fault
> > >>>>>> handling can be
> > >>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
> > >>>>>> fence can be
> > >>>>>> +  made visible anywhere in the system, all compute workloads must
> > >>>>>> be preempted
> > >>>>>> +  to guarantee all pending GPU page faults are flushed.
> > >>>>> I thought of another possible workaround:
> > >>>>>
> > >>>>>     * Partition the memory. Servicing of page faults will use a
> > >>>>> separate
> > >>>>>       memory pool that can always be allocated from without
> > >>>>> waiting for
> > >>>>>       fences. This includes memory for page tables and memory for
> > >>>>>       migrating data to. You may steal memory from other processes
> > >>>>> that
> > >>>>>       can page fault, so no fence waiting is necessary. Being able to
> > >>>>>       steal memory at any time also means there are basically no
> > >>>>>       out-of-memory situations you need to worry about. Even page
> > >>>>> tables
> > >>>>>       (except the root page directory of each process) can be
> > >>>>> stolen in
> > >>>>>       the worst case.
> > >>>> I think 'overcommit' would be a nice way to describe this. But I'm not
> > >>>> sure how easy this is to implement in practice. You would basically
> > >>>> need
> > >>>> to create your own memory manager for this.
> > >>> Well you would need a completely separate pool for both device as well
> > >>> as system memory.
> > >>>
> > >>> E.g. on boot we say we steal X GB system memory only for HMM.
> > >> Why? The GPU driver doesn't need to allocate system memory for HMM.
> > >> Migrations to system memory are handled by the kernel's handle_mm_fault
> > >> and page allocator and swap logic.
> > >
> > > And that one depends on dma_fence completion because you can easily
> > > need to wait for an MMU notifier callback.
> >
> > I see, the GFX MMU notifier for userpointers in amdgpu currently waits
> > for fences. For the KFD MMU notifier I am planning to fix this by
> > causing GPU page faults instead of preempting the queues. Can we limit
> > userptrs in amdgpu to engines that can page fault. Basically make it
> > illegal to attach userptr BOs to graphics CS BO lists, so they can only
> > be used in user mode command submissions, which can page fault. Then the
> > GFX MMU notifier could invalidate PTEs and would not have to wait for
> > fences.
>
> sadly graphics + userptr is already exposed via Mesa.

This is not about userptr, we fake userptr entirely in software. It's
about exposing recoverable gpu page faults (which would make userptr
maybe more efficient since we could do on-demand paging). userptr
itself isn't a problem, but it is part of the reasons why this is
tricky.

Christian/Felix, I think for kernel folks this is clear enough that I
don't need to clarify this in the text?
-Daniel

>
> >
> >
> > >
> > > As Maarten wrote when you want to go down this route you need a
> > > complete separate memory management parallel to the one of the kernel.
> >
> > Not really. I'm trying to make the GPU memory management more similar to
> > what the kernel does for system memory.
> >
> > I understood Maarten's comment as "I'm creating a new memory manager and
> > not using TTM any more". This is true. The idea is that this portion of
> > VRAM would be managed more like system memory.
> >
> > Regards,
> >   Felix
> >
> >
> > >
> > > Regards,
> > > Christian.
> > >
> > >>   It doesn't depend on any fences, so
> > >> it cannot deadlock with any GPU driver-managed memory. The GPU driver
> > >> gets involved in the MMU notifier to invalidate device page tables. But
> > >> that also doesn't need to wait for any fences.
> > >>
> > >> And if the kernel runs out of pageable memory, you're in trouble anyway.
> > >> The OOM killer will step in, nothing new there.
> > >>
> > >> Regards,
> > >>    Felix
> > >>
> > >>
> > >>>> But from a design point of view, definitely a valid solution.
> > >>> I think the restriction above makes it pretty much unusable.
> > >>>
> > >>>> But this looks good, those solutions are definitely the valid
> > >>>> options we
> > >>>> can choose from.
> > >>> It's certainly worth noting, yes. And just to make sure that nobody
> > >>> has the idea to reserve only device memory.
> > >>>
> > >>> Christian.
> > >>>
> > >>>> ~Maarten
> > >>>>
> > >> _______________________________________________
> > >> Linaro-mm-sig mailing list
> > >> [e-mail]
> > >> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.linaro.org%2Fmailman%2Flistinfo%2Flinaro-mm-sig&amp;data=04%7C01%7Cfelix.kuehling%40amd.com%7C4e4884be55d74c4dda1408d8c35fd0ab%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637474163592260552%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=gQj51eDK8OUWoQcbYliY639jOPleRjyLY3Q16nj2PL0%3D&amp;reserved=0
> > >>
> > >
> > _______________________________________________
> > dri-devel mailing list
> > [e-mail]
> > https://lists.freedesktop.org/mailman/listinfo/dri-devel

Felix Kuehling Feb. 9, 2021, 11:15 a.m. UTC | #16

Am 2021-02-09 um 1:37 a.m. schrieb Daniel Vetter:
> On Tue, Feb 9, 2021 at 4:13 AM Bas Nieuwenhuizen
> <[e-mail]> wrote:
>> On Thu, Jan 28, 2021 at 4:40 PM Felix Kuehling <[e-mail]> wrote:
>>> Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
>>>> Am 27.01.21 um 23:00 schrieb Felix Kuehling:
>>>>> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
>>>>>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>>>>>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>>>>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>>>>>>> Recently there was a fairly long thread about recoreable hardware
>>>>>>>>> page
>>>>>>>>> faults, how they can deadlock, and what to do about that.
>>>>>>>>>
>>>>>>>>> While the discussion is still fresh I figured good time to try and
>>>>>>>>> document the conclusions a bit.
>>>>>>>>>
>>>>>>>>> References:
>>>>>>>>> https://lore.kernel.org/dri-devel/[e-mail]/
>>>>>>>>>
>>>>>>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>>>>>>> Cc: Thomas Hellström <[e-mail]>
>>>>>>>>> Cc: "Christian König" <[e-mail]>
>>>>>>>>> Cc: Jerome Glisse <[e-mail]>
>>>>>>>>> Cc: Felix Kuehling <[e-mail]>
>>>>>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>>>>>>> Cc: Sumit Semwal <[e-mail]>
>>>>>>>>> Cc: [e-mail]
>>>>>>>>> Cc: [e-mail]
>>>>>>>>> --
>>>>>>>>> I'll be away next week, but figured I'll type this up quickly for
>>>>>>>>> some
>>>>>>>>> comments and to check whether I got this all roughly right.
>>>>>>>>>
>>>>>>>>> Critique very much wanted on this, so that we can make sure hw which
>>>>>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
>>>>>>>>> path to
>>>>>>>>> support page faults in upstream. So anything I missed, got wrong or
>>>>>>>>> like that would be good.
>>>>>>>>> -Daniel
>>>>>>>>> ---
>>>>>>>>>    Documentation/driver-api/dma-buf.rst | 66
>>>>>>>>> ++++++++++++++++++++++++++++
>>>>>>>>>    1 file changed, 66 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>>>>>>> b/Documentation/driver-api/dma-buf.rst
>>>>>>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>>>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>>>>>>      userspace is allowed to use userspace fencing or long running
>>>>>>>>> compute
>>>>>>>>>      workloads. This also means no implicit fencing for shared
>>>>>>>>> buffers in these
>>>>>>>>>      cases.
>>>>>>>>> +
>>>>>>>>> +Recoverable Hardware Page Faults Implications
>>>>>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>>>>> +
>>>>>>>>> +Modern hardware supports recoverable page faults, which has a
>>>>>>>>> lot of
>>>>>>>>> +implications for DMA fences.
>>>>>>>>> +
>>>>>>>>> +First, a pending page fault obviously holds up the work that's
>>>>>>>>> running on the
>>>>>>>>> +accelerator and a memory allocation is usually required to resolve
>>>>>>>>> the fault.
>>>>>>>>> +But memory allocations are not allowed to gate completion of DMA
>>>>>>>>> fences, which
>>>>>>>>> +means any workload using recoverable page faults cannot use DMA
>>>>>>>>> fences for
>>>>>>>>> +synchronization. Synchronization fences controlled by userspace
>>>>>>>>> must be used
>>>>>>>>> +instead.
>>>>>>>>> +
>>>>>>>>> +On GPUs this poses a problem, because current desktop compositor
>>>>>>>>> protocols on
>>>>>>>>> +Linus rely on DMA fences, which means without an entirely new
>>>>>>>>> userspace stack
>>>>>>>>> +built on top of userspace fences, they cannot benefit from
>>>>>>>>> recoverable page
>>>>>>>>> +faults. The exception is when page faults are only used as
>>>>>>>>> migration hints and
>>>>>>>>> +never to on-demand fill a memory request. For now this means
>>>>>>>>> recoverable page
>>>>>>>>> +faults on GPUs are limited to pure compute workloads.
>>>>>>>>> +
>>>>>>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>>>>>>> rendering and
>>>>>>>>> +compute side, like compute units or command submission engines. If
>>>>>>>>> both a 3D
>>>>>>>>> +job with a DMA fence and a compute workload using recoverable page
>>>>>>>>> faults are
>>>>>>>>> +pending they could deadlock:
>>>>>>>>> +
>>>>>>>>> +- The 3D workload might need to wait for the compute job to finish
>>>>>>>>> and release
>>>>>>>>> +  hardware resources first.
>>>>>>>>> +
>>>>>>>>> +- The compute workload might be stuck in a page fault, because the
>>>>>>>>> memory
>>>>>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>>>>>>> complete.
>>>>>>>>> +
>>>>>>>>> +There are a few ways to prevent this problem:
>>>>>>>>> +
>>>>>>>>> +- Compute workloads can always be preempted, even when a page
>>>>>>>>> fault is pending
>>>>>>>>> +  and not yet repaired. Not all hardware supports this.
>>>>>>>>> +
>>>>>>>>> +- DMA fence workloads and workloads which need page fault handling
>>>>>>>>> have
>>>>>>>>> +  independent hardware resources to guarantee forward progress.
>>>>>>>>> This could be
>>>>>>>>> +  achieved through e.g. through dedicated engines and minimal
>>>>>>>>> compute unit
>>>>>>>>> +  reservations for DMA fence workloads.
>>>>>>>>> +
>>>>>>>>> +- The reservation approach could be further refined by only
>>>>>>>>> reserving the
>>>>>>>>> +  hardware resources for DMA fence workloads when they are
>>>>>>>>> in-flight. This must
>>>>>>>>> +  cover the time from when the DMA fence is visible to other
>>>>>>>>> threads up to
>>>>>>>>> +  moment when fence is completed through dma_fence_signal().
>>>>>>>>> +
>>>>>>>>> +- As a last resort, if the hardware provides no useful reservation
>>>>>>>>> mechanics,
>>>>>>>>> +  all workloads must be flushed from the GPU when switching
>>>>>>>>> between jobs
>>>>>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>>>>>>> means all DMA
>>>>>>>>> +  fences must complete before a compute job with page fault
>>>>>>>>> handling can be
>>>>>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>>>>>>> fence can be
>>>>>>>>> +  made visible anywhere in the system, all compute workloads must
>>>>>>>>> be preempted
>>>>>>>>> +  to guarantee all pending GPU page faults are flushed.
>>>>>>>> I thought of another possible workaround:
>>>>>>>>
>>>>>>>>     * Partition the memory. Servicing of page faults will use a
>>>>>>>> separate
>>>>>>>>       memory pool that can always be allocated from without
>>>>>>>> waiting for
>>>>>>>>       fences. This includes memory for page tables and memory for
>>>>>>>>       migrating data to. You may steal memory from other processes
>>>>>>>> that
>>>>>>>>       can page fault, so no fence waiting is necessary. Being able to
>>>>>>>>       steal memory at any time also means there are basically no
>>>>>>>>       out-of-memory situations you need to worry about. Even page
>>>>>>>> tables
>>>>>>>>       (except the root page directory of each process) can be
>>>>>>>> stolen in
>>>>>>>>       the worst case.
>>>>>>> I think 'overcommit' would be a nice way to describe this. But I'm not
>>>>>>> sure how easy this is to implement in practice. You would basically
>>>>>>> need
>>>>>>> to create your own memory manager for this.
>>>>>> Well you would need a completely separate pool for both device as well
>>>>>> as system memory.
>>>>>>
>>>>>> E.g. on boot we say we steal X GB system memory only for HMM.
>>>>> Why? The GPU driver doesn't need to allocate system memory for HMM.
>>>>> Migrations to system memory are handled by the kernel's handle_mm_fault
>>>>> and page allocator and swap logic.
>>>> And that one depends on dma_fence completion because you can easily
>>>> need to wait for an MMU notifier callback.
>>> I see, the GFX MMU notifier for userpointers in amdgpu currently waits
>>> for fences. For the KFD MMU notifier I am planning to fix this by
>>> causing GPU page faults instead of preempting the queues. Can we limit
>>> userptrs in amdgpu to engines that can page fault. Basically make it
>>> illegal to attach userptr BOs to graphics CS BO lists, so they can only
>>> be used in user mode command submissions, which can page fault. Then the
>>> GFX MMU notifier could invalidate PTEs and would not have to wait for
>>> fences.
>> sadly graphics + userptr is already exposed via Mesa.
> This is not about userptr, we fake userptr entirely in software. It's
> about exposing recoverable gpu page faults (which would make userptr
> maybe more efficient since we could do on-demand paging). userptr
> itself isn't a problem, but it is part of the reasons why this is
> tricky.
>
> Christian/Felix, I think for kernel folks this is clear enough that I
> don't need to clarify this in the text?

Yeah, it's clear to me. Anyway, your latest text doesn't reference
userptr directly and keeps the discussion at a fairly abstract level. So
I think it's fine. It's the practical details of the proposed
workarounds where it feel like walking through a mirror cabinet, bumping
into unexpected obstacles with every other step.

Regards,
  Felix


> -Daniel
>
>>>
>>>> As Maarten wrote when you want to go down this route you need a
>>>> complete separate memory management parallel to the one of the kernel.
>>> Not really. I'm trying to make the GPU memory management more similar to
>>> what the kernel does for system memory.
>>>
>>> I understood Maarten's comment as "I'm creating a new memory manager and
>>> not using TTM any more". This is true. The idea is that this portion of
>>> VRAM would be managed more like system memory.
>>>
>>> Regards,
>>>   Felix
>>>
>>>
>>>> Regards,
>>>> Christian.
>>>>
>>>>>   It doesn't depend on any fences, so
>>>>> it cannot deadlock with any GPU driver-managed memory. The GPU driver
>>>>> gets involved in the MMU notifier to invalidate device page tables. But
>>>>> that also doesn't need to wait for any fences.
>>>>>
>>>>> And if the kernel runs out of pageable memory, you're in trouble anyway.
>>>>> The OOM killer will step in, nothing new there.
>>>>>
>>>>> Regards,
>>>>>    Felix
>>>>>
>>>>>
>>>>>>> But from a design point of view, definitely a valid solution.
>>>>>> I think the restriction above makes it pretty much unusable.
>>>>>>
>>>>>>> But this looks good, those solutions are definitely the valid
>>>>>>> options we
>>>>>>> can choose from.
>>>>>> It's certainly worth noting, yes. And just to make sure that nobody
>>>>>> has the idea to reserve only device memory.
>>>>>>
>>>>>> Christian.
>>>>>>
>>>>>>> ~Maarten
>>>>>>>
>>>>> _______________________________________________
>>>>> Linaro-mm-sig mailing list
>>>>> [e-mail]
>>>>> https://lists.linaro.org/mailman/listinfo/linaro-mm-sig
>>>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> [e-mail]
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>

Daniel Vetter Feb. 9, 2021, 2:08 p.m. UTC | #17

On Tue, Feb 9, 2021 at 12:15 PM Felix Kuehling <[e-mail]> wrote:
> Am 2021-02-09 um 1:37 a.m. schrieb Daniel Vetter:
> > On Tue, Feb 9, 2021 at 4:13 AM Bas Nieuwenhuizen
> > <[e-mail]> wrote:
> >> On Thu, Jan 28, 2021 at 4:40 PM Felix Kuehling <[e-mail]> wrote:
> >>> Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
> >>>> Am 27.01.21 um 23:00 schrieb Felix Kuehling:
> >>>>> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
> >>>>>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
> >>>>>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
> >>>>>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
> >>>>>>>>> Recently there was a fairly long thread about recoreable hardware
> >>>>>>>>> page
> >>>>>>>>> faults, how they can deadlock, and what to do about that.
> >>>>>>>>>
> >>>>>>>>> While the discussion is still fresh I figured good time to try and
> >>>>>>>>> document the conclusions a bit.
> >>>>>>>>>
> >>>>>>>>> References:
> >>>>>>>>> https://lore.kernel.org/dri-devel/[e-mail]/
> >>>>>>>>>
> >>>>>>>>> Cc: Maarten Lankhorst <[e-mail]>
> >>>>>>>>> Cc: Thomas Hellström <[e-mail]>
> >>>>>>>>> Cc: "Christian König" <[e-mail]>
> >>>>>>>>> Cc: Jerome Glisse <[e-mail]>
> >>>>>>>>> Cc: Felix Kuehling <[e-mail]>
> >>>>>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
> >>>>>>>>> Cc: Sumit Semwal <[e-mail]>
> >>>>>>>>> Cc: [e-mail]
> >>>>>>>>> Cc: [e-mail]
> >>>>>>>>> --
> >>>>>>>>> I'll be away next week, but figured I'll type this up quickly for
> >>>>>>>>> some
> >>>>>>>>> comments and to check whether I got this all roughly right.
> >>>>>>>>>
> >>>>>>>>> Critique very much wanted on this, so that we can make sure hw which
> >>>>>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
> >>>>>>>>> path to
> >>>>>>>>> support page faults in upstream. So anything I missed, got wrong or
> >>>>>>>>> like that would be good.
> >>>>>>>>> -Daniel
> >>>>>>>>> ---
> >>>>>>>>>    Documentation/driver-api/dma-buf.rst | 66
> >>>>>>>>> ++++++++++++++++++++++++++++
> >>>>>>>>>    1 file changed, 66 insertions(+)
> >>>>>>>>>
> >>>>>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
> >>>>>>>>> b/Documentation/driver-api/dma-buf.rst
> >>>>>>>>> index a2133d69872c..e924c1e4f7a3 100644
> >>>>>>>>> --- a/Documentation/driver-api/dma-buf.rst
> >>>>>>>>> +++ b/Documentation/driver-api/dma-buf.rst
> >>>>>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
> >>>>>>>>>      userspace is allowed to use userspace fencing or long running
> >>>>>>>>> compute
> >>>>>>>>>      workloads. This also means no implicit fencing for shared
> >>>>>>>>> buffers in these
> >>>>>>>>>      cases.
> >>>>>>>>> +
> >>>>>>>>> +Recoverable Hardware Page Faults Implications
> >>>>>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>>>>>> +
> >>>>>>>>> +Modern hardware supports recoverable page faults, which has a
> >>>>>>>>> lot of
> >>>>>>>>> +implications for DMA fences.
> >>>>>>>>> +
> >>>>>>>>> +First, a pending page fault obviously holds up the work that's
> >>>>>>>>> running on the
> >>>>>>>>> +accelerator and a memory allocation is usually required to resolve
> >>>>>>>>> the fault.
> >>>>>>>>> +But memory allocations are not allowed to gate completion of DMA
> >>>>>>>>> fences, which
> >>>>>>>>> +means any workload using recoverable page faults cannot use DMA
> >>>>>>>>> fences for
> >>>>>>>>> +synchronization. Synchronization fences controlled by userspace
> >>>>>>>>> must be used
> >>>>>>>>> +instead.
> >>>>>>>>> +
> >>>>>>>>> +On GPUs this poses a problem, because current desktop compositor
> >>>>>>>>> protocols on
> >>>>>>>>> +Linus rely on DMA fences, which means without an entirely new
> >>>>>>>>> userspace stack
> >>>>>>>>> +built on top of userspace fences, they cannot benefit from
> >>>>>>>>> recoverable page
> >>>>>>>>> +faults. The exception is when page faults are only used as
> >>>>>>>>> migration hints and
> >>>>>>>>> +never to on-demand fill a memory request. For now this means
> >>>>>>>>> recoverable page
> >>>>>>>>> +faults on GPUs are limited to pure compute workloads.
> >>>>>>>>> +
> >>>>>>>>> +Furthermore GPUs usually have shared resources between the 3D
> >>>>>>>>> rendering and
> >>>>>>>>> +compute side, like compute units or command submission engines. If
> >>>>>>>>> both a 3D
> >>>>>>>>> +job with a DMA fence and a compute workload using recoverable page
> >>>>>>>>> faults are
> >>>>>>>>> +pending they could deadlock:
> >>>>>>>>> +
> >>>>>>>>> +- The 3D workload might need to wait for the compute job to finish
> >>>>>>>>> and release
> >>>>>>>>> +  hardware resources first.
> >>>>>>>>> +
> >>>>>>>>> +- The compute workload might be stuck in a page fault, because the
> >>>>>>>>> memory
> >>>>>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
> >>>>>>>>> complete.
> >>>>>>>>> +
> >>>>>>>>> +There are a few ways to prevent this problem:
> >>>>>>>>> +
> >>>>>>>>> +- Compute workloads can always be preempted, even when a page
> >>>>>>>>> fault is pending
> >>>>>>>>> +  and not yet repaired. Not all hardware supports this.
> >>>>>>>>> +
> >>>>>>>>> +- DMA fence workloads and workloads which need page fault handling
> >>>>>>>>> have
> >>>>>>>>> +  independent hardware resources to guarantee forward progress.
> >>>>>>>>> This could be
> >>>>>>>>> +  achieved through e.g. through dedicated engines and minimal
> >>>>>>>>> compute unit
> >>>>>>>>> +  reservations for DMA fence workloads.
> >>>>>>>>> +
> >>>>>>>>> +- The reservation approach could be further refined by only
> >>>>>>>>> reserving the
> >>>>>>>>> +  hardware resources for DMA fence workloads when they are
> >>>>>>>>> in-flight. This must
> >>>>>>>>> +  cover the time from when the DMA fence is visible to other
> >>>>>>>>> threads up to
> >>>>>>>>> +  moment when fence is completed through dma_fence_signal().
> >>>>>>>>> +
> >>>>>>>>> +- As a last resort, if the hardware provides no useful reservation
> >>>>>>>>> mechanics,
> >>>>>>>>> +  all workloads must be flushed from the GPU when switching
> >>>>>>>>> between jobs
> >>>>>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
> >>>>>>>>> means all DMA
> >>>>>>>>> +  fences must complete before a compute job with page fault
> >>>>>>>>> handling can be
> >>>>>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
> >>>>>>>>> fence can be
> >>>>>>>>> +  made visible anywhere in the system, all compute workloads must
> >>>>>>>>> be preempted
> >>>>>>>>> +  to guarantee all pending GPU page faults are flushed.
> >>>>>>>> I thought of another possible workaround:
> >>>>>>>>
> >>>>>>>>     * Partition the memory. Servicing of page faults will use a
> >>>>>>>> separate
> >>>>>>>>       memory pool that can always be allocated from without
> >>>>>>>> waiting for
> >>>>>>>>       fences. This includes memory for page tables and memory for
> >>>>>>>>       migrating data to. You may steal memory from other processes
> >>>>>>>> that
> >>>>>>>>       can page fault, so no fence waiting is necessary. Being able to
> >>>>>>>>       steal memory at any time also means there are basically no
> >>>>>>>>       out-of-memory situations you need to worry about. Even page
> >>>>>>>> tables
> >>>>>>>>       (except the root page directory of each process) can be
> >>>>>>>> stolen in
> >>>>>>>>       the worst case.
> >>>>>>> I think 'overcommit' would be a nice way to describe this. But I'm not
> >>>>>>> sure how easy this is to implement in practice. You would basically
> >>>>>>> need
> >>>>>>> to create your own memory manager for this.
> >>>>>> Well you would need a completely separate pool for both device as well
> >>>>>> as system memory.
> >>>>>>
> >>>>>> E.g. on boot we say we steal X GB system memory only for HMM.
> >>>>> Why? The GPU driver doesn't need to allocate system memory for HMM.
> >>>>> Migrations to system memory are handled by the kernel's handle_mm_fault
> >>>>> and page allocator and swap logic.
> >>>> And that one depends on dma_fence completion because you can easily
> >>>> need to wait for an MMU notifier callback.
> >>> I see, the GFX MMU notifier for userpointers in amdgpu currently waits
> >>> for fences. For the KFD MMU notifier I am planning to fix this by
> >>> causing GPU page faults instead of preempting the queues. Can we limit
> >>> userptrs in amdgpu to engines that can page fault. Basically make it
> >>> illegal to attach userptr BOs to graphics CS BO lists, so they can only
> >>> be used in user mode command submissions, which can page fault. Then the
> >>> GFX MMU notifier could invalidate PTEs and would not have to wait for
> >>> fences.
> >> sadly graphics + userptr is already exposed via Mesa.
> > This is not about userptr, we fake userptr entirely in software. It's
> > about exposing recoverable gpu page faults (which would make userptr
> > maybe more efficient since we could do on-demand paging). userptr
> > itself isn't a problem, but it is part of the reasons why this is
> > tricky.
> >
> > Christian/Felix, I think for kernel folks this is clear enough that I
> > don't need to clarify this in the text?
>
> Yeah, it's clear to me. Anyway, your latest text doesn't reference
> userptr directly and keeps the discussion at a fairly abstract level. So
> I think it's fine. It's the practical details of the proposed
> workarounds where it feel like walking through a mirror cabinet, bumping
> into unexpected obstacles with every other step.

Oh yes, this is very high-level. The implementation is going to be
very trick, no matter which one we're picking. And tbh I expect
surprises and things we'll learn. But I'm still hoping that this high
level doc patch will help a lot with avoiding the worst problems.

Of course once we have some of these hacks landed we should look at it
again and maybe update where it's wrong/unclear/...

btw r-b: from you too on the patch?

Cheers, Daniel

>
> Regards,
>   Felix
>
>
> > -Daniel
> >
> >>>
> >>>> As Maarten wrote when you want to go down this route you need a
> >>>> complete separate memory management parallel to the one of the kernel.
> >>> Not really. I'm trying to make the GPU memory management more similar to
> >>> what the kernel does for system memory.
> >>>
> >>> I understood Maarten's comment as "I'm creating a new memory manager and
> >>> not using TTM any more". This is true. The idea is that this portion of
> >>> VRAM would be managed more like system memory.
> >>>
> >>> Regards,
> >>>   Felix
> >>>
> >>>
> >>>> Regards,
> >>>> Christian.
> >>>>
> >>>>>   It doesn't depend on any fences, so
> >>>>> it cannot deadlock with any GPU driver-managed memory. The GPU driver
> >>>>> gets involved in the MMU notifier to invalidate device page tables. But
> >>>>> that also doesn't need to wait for any fences.
> >>>>>
> >>>>> And if the kernel runs out of pageable memory, you're in trouble anyway.
> >>>>> The OOM killer will step in, nothing new there.
> >>>>>
> >>>>> Regards,
> >>>>>    Felix
> >>>>>
> >>>>>
> >>>>>>> But from a design point of view, definitely a valid solution.
> >>>>>> I think the restriction above makes it pretty much unusable.
> >>>>>>
> >>>>>>> But this looks good, those solutions are definitely the valid
> >>>>>>> options we
> >>>>>>> can choose from.
> >>>>>> It's certainly worth noting, yes. And just to make sure that nobody
> >>>>>> has the idea to reserve only device memory.
> >>>>>>
> >>>>>> Christian.
> >>>>>>
> >>>>>>> ~Maarten
> >>>>>>>
> >>>>> _______________________________________________
> >>>>> Linaro-mm-sig mailing list
> >>>>> [e-mail]
> >>>>> https://lists.linaro.org/mailman/listinfo/linaro-mm-sig
> >>>>>
> >>> _______________________________________________
> >>> dri-devel mailing list
> >>> [e-mail]
> >>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
> >
> >

Felix Kuehling Feb. 9, 2021, 2:25 p.m. UTC | #18

Am 2021-02-09 um 9:08 a.m. schrieb Daniel Vetter:
> On Tue, Feb 9, 2021 at 12:15 PM Felix Kuehling <[e-mail]> wrote:
>> Am 2021-02-09 um 1:37 a.m. schrieb Daniel Vetter:
>>> On Tue, Feb 9, 2021 at 4:13 AM Bas Nieuwenhuizen
>>> <[e-mail]> wrote:
>>>> On Thu, Jan 28, 2021 at 4:40 PM Felix Kuehling <[e-mail]> wrote:
>>>>> Am 2021-01-28 um 2:39 a.m. schrieb Christian König:
>>>>>> Am 27.01.21 um 23:00 schrieb Felix Kuehling:
>>>>>>> Am 2021-01-27 um 7:16 a.m. schrieb Christian König:
>>>>>>>> Am 27.01.21 um 13:11 schrieb Maarten Lankhorst:
>>>>>>>>> Op 27-01-2021 om 01:22 schreef Felix Kuehling:
>>>>>>>>>> Am 2021-01-21 um 2:40 p.m. schrieb Daniel Vetter:
>>>>>>>>>>> Recently there was a fairly long thread about recoreable hardware
>>>>>>>>>>> page
>>>>>>>>>>> faults, how they can deadlock, and what to do about that.
>>>>>>>>>>>
>>>>>>>>>>> While the discussion is still fresh I figured good time to try and
>>>>>>>>>>> document the conclusions a bit.
>>>>>>>>>>>
>>>>>>>>>>> References:
>>>>>>>>>>> https://lore.kernel.org/dri-devel/[e-mail]/
>>>>>>>>>>>
>>>>>>>>>>> Cc: Maarten Lankhorst <[e-mail]>
>>>>>>>>>>> Cc: Thomas Hellström <[e-mail]>
>>>>>>>>>>> Cc: "Christian König" <[e-mail]>
>>>>>>>>>>> Cc: Jerome Glisse <[e-mail]>
>>>>>>>>>>> Cc: Felix Kuehling <[e-mail]>
>>>>>>>>>>> Signed-off-by: Daniel Vetter <[e-mail]>
>>>>>>>>>>> Cc: Sumit Semwal <[e-mail]>
>>>>>>>>>>> Cc: [e-mail]
>>>>>>>>>>> Cc: [e-mail]
>>>>>>>>>>> --
>>>>>>>>>>> I'll be away next week, but figured I'll type this up quickly for
>>>>>>>>>>> some
>>>>>>>>>>> comments and to check whether I got this all roughly right.
>>>>>>>>>>>
>>>>>>>>>>> Critique very much wanted on this, so that we can make sure hw which
>>>>>>>>>>> can't preempt (with pagefaults pending) like gfx10 has a clear
>>>>>>>>>>> path to
>>>>>>>>>>> support page faults in upstream. So anything I missed, got wrong or
>>>>>>>>>>> like that would be good.
>>>>>>>>>>> -Daniel
>>>>>>>>>>> ---
>>>>>>>>>>>    Documentation/driver-api/dma-buf.rst | 66
>>>>>>>>>>> ++++++++++++++++++++++++++++
>>>>>>>>>>>    1 file changed, 66 insertions(+)
>>>>>>>>>>>
>>>>>>>>>>> diff --git a/Documentation/driver-api/dma-buf.rst
>>>>>>>>>>> b/Documentation/driver-api/dma-buf.rst
>>>>>>>>>>> index a2133d69872c..e924c1e4f7a3 100644
>>>>>>>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>>>>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>>>>>>>> @@ -257,3 +257,69 @@ fences in the kernel. This means:
>>>>>>>>>>>      userspace is allowed to use userspace fencing or long running
>>>>>>>>>>> compute
>>>>>>>>>>>      workloads. This also means no implicit fencing for shared
>>>>>>>>>>> buffers in these
>>>>>>>>>>>      cases.
>>>>>>>>>>> +
>>>>>>>>>>> +Recoverable Hardware Page Faults Implications
>>>>>>>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>>>>>>> +
>>>>>>>>>>> +Modern hardware supports recoverable page faults, which has a
>>>>>>>>>>> lot of
>>>>>>>>>>> +implications for DMA fences.
>>>>>>>>>>> +
>>>>>>>>>>> +First, a pending page fault obviously holds up the work that's
>>>>>>>>>>> running on the
>>>>>>>>>>> +accelerator and a memory allocation is usually required to resolve
>>>>>>>>>>> the fault.
>>>>>>>>>>> +But memory allocations are not allowed to gate completion of DMA
>>>>>>>>>>> fences, which
>>>>>>>>>>> +means any workload using recoverable page faults cannot use DMA
>>>>>>>>>>> fences for
>>>>>>>>>>> +synchronization. Synchronization fences controlled by userspace
>>>>>>>>>>> must be used
>>>>>>>>>>> +instead.
>>>>>>>>>>> +
>>>>>>>>>>> +On GPUs this poses a problem, because current desktop compositor
>>>>>>>>>>> protocols on
>>>>>>>>>>> +Linus rely on DMA fences, which means without an entirely new
>>>>>>>>>>> userspace stack
>>>>>>>>>>> +built on top of userspace fences, they cannot benefit from
>>>>>>>>>>> recoverable page
>>>>>>>>>>> +faults. The exception is when page faults are only used as
>>>>>>>>>>> migration hints and
>>>>>>>>>>> +never to on-demand fill a memory request. For now this means
>>>>>>>>>>> recoverable page
>>>>>>>>>>> +faults on GPUs are limited to pure compute workloads.
>>>>>>>>>>> +
>>>>>>>>>>> +Furthermore GPUs usually have shared resources between the 3D
>>>>>>>>>>> rendering and
>>>>>>>>>>> +compute side, like compute units or command submission engines. If
>>>>>>>>>>> both a 3D
>>>>>>>>>>> +job with a DMA fence and a compute workload using recoverable page
>>>>>>>>>>> faults are
>>>>>>>>>>> +pending they could deadlock:
>>>>>>>>>>> +
>>>>>>>>>>> +- The 3D workload might need to wait for the compute job to finish
>>>>>>>>>>> and release
>>>>>>>>>>> +  hardware resources first.
>>>>>>>>>>> +
>>>>>>>>>>> +- The compute workload might be stuck in a page fault, because the
>>>>>>>>>>> memory
>>>>>>>>>>> +  allocation is waiting for the DMA fence of the 3D workload to
>>>>>>>>>>> complete.
>>>>>>>>>>> +
>>>>>>>>>>> +There are a few ways to prevent this problem:
>>>>>>>>>>> +
>>>>>>>>>>> +- Compute workloads can always be preempted, even when a page
>>>>>>>>>>> fault is pending
>>>>>>>>>>> +  and not yet repaired. Not all hardware supports this.
>>>>>>>>>>> +
>>>>>>>>>>> +- DMA fence workloads and workloads which need page fault handling
>>>>>>>>>>> have
>>>>>>>>>>> +  independent hardware resources to guarantee forward progress.
>>>>>>>>>>> This could be
>>>>>>>>>>> +  achieved through e.g. through dedicated engines and minimal
>>>>>>>>>>> compute unit
>>>>>>>>>>> +  reservations for DMA fence workloads.
>>>>>>>>>>> +
>>>>>>>>>>> +- The reservation approach could be further refined by only
>>>>>>>>>>> reserving the
>>>>>>>>>>> +  hardware resources for DMA fence workloads when they are
>>>>>>>>>>> in-flight. This must
>>>>>>>>>>> +  cover the time from when the DMA fence is visible to other
>>>>>>>>>>> threads up to
>>>>>>>>>>> +  moment when fence is completed through dma_fence_signal().
>>>>>>>>>>> +
>>>>>>>>>>> +- As a last resort, if the hardware provides no useful reservation
>>>>>>>>>>> mechanics,
>>>>>>>>>>> +  all workloads must be flushed from the GPU when switching
>>>>>>>>>>> between jobs
>>>>>>>>>>> +  requiring DMA fences or jobs requiring page fault handling: This
>>>>>>>>>>> means all DMA
>>>>>>>>>>> +  fences must complete before a compute job with page fault
>>>>>>>>>>> handling can be
>>>>>>>>>>> +  inserted into the scheduler queue. And vice versa, before a DMA
>>>>>>>>>>> fence can be
>>>>>>>>>>> +  made visible anywhere in the system, all compute workloads must
>>>>>>>>>>> be preempted
>>>>>>>>>>> +  to guarantee all pending GPU page faults are flushed.
>>>>>>>>>> I thought of another possible workaround:
>>>>>>>>>>
>>>>>>>>>>     * Partition the memory. Servicing of page faults will use a
>>>>>>>>>> separate
>>>>>>>>>>       memory pool that can always be allocated from without
>>>>>>>>>> waiting for
>>>>>>>>>>       fences. This includes memory for page tables and memory for
>>>>>>>>>>       migrating data to. You may steal memory from other processes
>>>>>>>>>> that
>>>>>>>>>>       can page fault, so no fence waiting is necessary. Being able to
>>>>>>>>>>       steal memory at any time also means there are basically no
>>>>>>>>>>       out-of-memory situations you need to worry about. Even page
>>>>>>>>>> tables
>>>>>>>>>>       (except the root page directory of each process) can be
>>>>>>>>>> stolen in
>>>>>>>>>>       the worst case.
>>>>>>>>> I think 'overcommit' would be a nice way to describe this. But I'm not
>>>>>>>>> sure how easy this is to implement in practice. You would basically
>>>>>>>>> need
>>>>>>>>> to create your own memory manager for this.
>>>>>>>> Well you would need a completely separate pool for both device as well
>>>>>>>> as system memory.
>>>>>>>>
>>>>>>>> E.g. on boot we say we steal X GB system memory only for HMM.
>>>>>>> Why? The GPU driver doesn't need to allocate system memory for HMM.
>>>>>>> Migrations to system memory are handled by the kernel's handle_mm_fault
>>>>>>> and page allocator and swap logic.
>>>>>> And that one depends on dma_fence completion because you can easily
>>>>>> need to wait for an MMU notifier callback.
>>>>> I see, the GFX MMU notifier for userpointers in amdgpu currently waits
>>>>> for fences. For the KFD MMU notifier I am planning to fix this by
>>>>> causing GPU page faults instead of preempting the queues. Can we limit
>>>>> userptrs in amdgpu to engines that can page fault. Basically make it
>>>>> illegal to attach userptr BOs to graphics CS BO lists, so they can only
>>>>> be used in user mode command submissions, which can page fault. Then the
>>>>> GFX MMU notifier could invalidate PTEs and would not have to wait for
>>>>> fences.
>>>> sadly graphics + userptr is already exposed via Mesa.
>>> This is not about userptr, we fake userptr entirely in software. It's
>>> about exposing recoverable gpu page faults (which would make userptr
>>> maybe more efficient since we could do on-demand paging). userptr
>>> itself isn't a problem, but it is part of the reasons why this is
>>> tricky.
>>>
>>> Christian/Felix, I think for kernel folks this is clear enough that I
>>> don't need to clarify this in the text?
>> Yeah, it's clear to me. Anyway, your latest text doesn't reference
>> userptr directly and keeps the discussion at a fairly abstract level. So
>> I think it's fine. It's the practical details of the proposed
>> workarounds where it feel like walking through a mirror cabinet, bumping
>> into unexpected obstacles with every other step.
> Oh yes, this is very high-level. The implementation is going to be
> very trick, no matter which one we're picking. And tbh I expect
> surprises and things we'll learn. But I'm still hoping that this high
> level doc patch will help a lot with avoiding the worst problems.
>
> Of course once we have some of these hacks landed we should look at it
> again and maybe update where it's wrong/unclear/...
>
> btw r-b: from you too on the patch?

Yes.

Reviewed-by: Felix Kuehling <[e-mail]>

Thanks,
  Felix


>
> Cheers, Daniel
>
>> Regards,
>>   Felix
>>
>>
>>> -Daniel
>>>
>>>>>> As Maarten wrote when you want to go down this route you need a
>>>>>> complete separate memory management parallel to the one of the kernel.
>>>>> Not really. I'm trying to make the GPU memory management more similar to
>>>>> what the kernel does for system memory.
>>>>>
>>>>> I understood Maarten's comment as "I'm creating a new memory manager and
>>>>> not using TTM any more". This is true. The idea is that this portion of
>>>>> VRAM would be managed more like system memory.
>>>>>
>>>>> Regards,
>>>>>   Felix
>>>>>
>>>>>
>>>>>> Regards,
>>>>>> Christian.
>>>>>>
>>>>>>>   It doesn't depend on any fences, so
>>>>>>> it cannot deadlock with any GPU driver-managed memory. The GPU driver
>>>>>>> gets involved in the MMU notifier to invalidate device page tables. But
>>>>>>> that also doesn't need to wait for any fences.
>>>>>>>
>>>>>>> And if the kernel runs out of pageable memory, you're in trouble anyway.
>>>>>>> The OOM killer will step in, nothing new there.
>>>>>>>
>>>>>>> Regards,
>>>>>>>    Felix
>>>>>>>
>>>>>>>
>>>>>>>>> But from a design point of view, definitely a valid solution.
>>>>>>>> I think the restriction above makes it pretty much unusable.
>>>>>>>>
>>>>>>>>> But this looks good, those solutions are definitely the valid
>>>>>>>>> options we
>>>>>>>>> can choose from.
>>>>>>>> It's certainly worth noting, yes. And just to make sure that nobody
>>>>>>>> has the idea to reserve only device memory.
>>>>>>>>
>>>>>>>> Christian.
>>>>>>>>
>>>>>>>>> ~Maarten
>>>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> Linaro-mm-sig mailing list
>>>>>>> [e-mail]
>>>>>>> https://lists.linaro.org/mailman/listinfo/linaro-mm-sig
>>>>>>>
>>>>> _______________________________________________
>>>>> dri-devel mailing list
>>>>> [e-mail]
>>>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>>>
>
>

Patch

diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
index a2133d69872c..e924c1e4f7a3 100644
--- a/Documentation/driver-api/dma-buf.rst
+++ b/Documentation/driver-api/dma-buf.rst
@@ -257,3 +257,69 @@  fences in the kernel. This means:
   userspace is allowed to use userspace fencing or long running compute
   workloads. This also means no implicit fencing for shared buffers in these
   cases.
+
+Recoverable Hardware Page Faults Implications
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Modern hardware supports recoverable page faults, which has a lot of
+implications for DMA fences.
+
+First, a pending page fault obviously holds up the work that's running on the
+accelerator and a memory allocation is usually required to resolve the fault.
+But memory allocations are not allowed to gate completion of DMA fences, which
+means any workload using recoverable page faults cannot use DMA fences for
+synchronization. Synchronization fences controlled by userspace must be used
+instead.
+
+On GPUs this poses a problem, because current desktop compositor protocols on
+Linus rely on DMA fences, which means without an entirely new userspace stack
+built on top of userspace fences, they cannot benefit from recoverable page
+faults. The exception is when page faults are only used as migration hints and
+never to on-demand fill a memory request. For now this means recoverable page
+faults on GPUs are limited to pure compute workloads.
+
+Furthermore GPUs usually have shared resources between the 3D rendering and
+compute side, like compute units or command submission engines. If both a 3D
+job with a DMA fence and a compute workload using recoverable page faults are
+pending they could deadlock:
+
+- The 3D workload might need to wait for the compute job to finish and release
+  hardware resources first.
+
+- The compute workload might be stuck in a page fault, because the memory
+  allocation is waiting for the DMA fence of the 3D workload to complete.
+
+There are a few ways to prevent this problem:
+
+- Compute workloads can always be preempted, even when a page fault is pending
+  and not yet repaired. Not all hardware supports this.
+
+- DMA fence workloads and workloads which need page fault handling have
+  independent hardware resources to guarantee forward progress. This could be
+  achieved through e.g. through dedicated engines and minimal compute unit
+  reservations for DMA fence workloads.
+
+- The reservation approach could be further refined by only reserving the
+  hardware resources for DMA fence workloads when they are in-flight. This must
+  cover the time from when the DMA fence is visible to other threads up to
+  moment when fence is completed through dma_fence_signal().
+
+- As a last resort, if the hardware provides no useful reservation mechanics,
+  all workloads must be flushed from the GPU when switching between jobs
+  requiring DMA fences or jobs requiring page fault handling: This means all DMA
+  fences must complete before a compute job with page fault handling can be
+  inserted into the scheduler queue. And vice versa, before a DMA fence can be
+  made visible anywhere in the system, all compute workloads must be preempted
+  to guarantee all pending GPU page faults are flushed.
+
+Note that workloads that run on independent hardware like copy engines or other
+GPUs do not have any impact. This allows us to keep using DMA fences internally
+in the kernel even for resolving hardware page faults, e.g. by using copy
+engines to clear or copy memory needed to resolve the page fault.
+
+In some ways this page fault problem is a special case of the `Infinite DMA
+Fences` discussions: Infinite fences from compute workloads are allowed to
+depend on DMA fences, but not the other way around. And not even the page fault
+problem is new, because some other CPU thread in userspace might
+hit a page fault which holds up a userspace fence - supporting page faults on
+GPUs doesn't anything fundamentally new.