Re: [PATCH v2 0/8] arm64: Unmap FF-A lent memory from direct map
From: Sumit Garg
Date: Fri Oct 30 2026 - 02:31:04 EST
On Wed, 28 Oct 2026 at 14:50:36 +0000, Vincent Donnefort wrote:
On Tue, Oct 13, 2026 at 03:26:30PM +0530, Sumit Garg wrote:
On Tue, 13 Oct 2026 at 10:37:59 +0100, Vincent Donnefort wrote:
> On Mon, Oct 12, 2026 at 06:46:12PM +0530, Sumit Garg wrote:
> > On Mon, 05 Oct 2026 at 12:02:55 +0100, Vincent Donnefort wrote:
> > > Sorry, not sure why Sumit's email was dropped...
> >
> > The problem was with my mutt config, it should be fixed now.
> >
> > >
> > > On Mon, Oct 05, 2026 at 11:48:00AM +0100, Vincent Donnefort wrote:
> > > > On Thu, Oct 01, 2026 at 06:04:06PM +0530, Sumit Garg wrote:
> > > > > On Tue, 29 Sep 2026 at 12:33:07 +0100, Will Deacon wrote:
> > > > > > On Sat, Sep 26, 2026 at 05:43:08PM +0530, Sumit Garg wrote:
> > > > > > > Looks like you dropped me from your reply.
> > > > > >
> > > > > > Huh, that's weird. If I hit reply-all ('g') in mutt, it moves the entire
> > > > > > CC list to To: and drops you. I suspect the same thing happened to Thierry
> > > > > > when he replied here:
> > > > > >
> > > > > > https://lore.kernel.org/all/arJcN60nagFKXj2Q@orome/
> > > > > >
> > > > > > I can't figure out why that's happening :/
> > > > > > I manually tried to fix the To: and Cc: lines in this message.
> > > > >
> > > > > Thanks.
> > > > >
> > > > > >
> > > > > > > On Tue, 22 Sep 2026 at 15:55:59 +0100, Will Deacon wrote:
> > > > > > > > I spoke to Brendan at LPC (?) last year but this doesn't really work
> > > > > > > > for arm64 because it relies on being able to unmap arbitrary parts of
> > > > > > > > the linear map, which isn't generally possible unless you force pte-level
> > > > > > > > mappings for everything, which is prohibitive for perf/power.
> > > > > > >
> > > > > > > As per the cover letter it's about allocating pages that are not present
> > > > > > > in the direct map. This essentially fits the protected DMABufs use-case
> > > > > > > where we don't want any kernel mapping to exist at any time. Bufers
> > > > > > > allocated from protected DMABufs are only meant to be accessed by the
> > > > > > > TEE implementation or HW accelerators like in the secure media pipeline.
> > > > > >
> > > > > > That sounds like it should probably build on top of this series, then.
> > > > >
> > > > > The main problem I see with this series is it tries to move from the generic
> > > > > CMA allocator to a reserved FF-A CMA pool which again comes from the
> > > > > DT. We already support reserved pool allocations via "no-map" which are
> > > > > discovered dynamically from OP-TEE.
> > > >
> > > > I can't think of a way around a DT declaration. We need to know what region is
> > > > PTE-level before the direct map is installed and of_reserved_mem is very handy
> > > > for that purpose. And as this PTE-level mapping is costly we certainly only want
> > > > to enable it for platforms that absolutely need it.
> > > >
> >
> > I am rather thinking of a way to turn the default CMA allocator to be
> > PTE-level mapped. Since the overall idea for this dynamic protected
> > buffers was to reuse memory efficiently, otherwise the reserved buffer
> > approach is already there to use as "no-map" via DT.
>
> When the default CMA allocator is setup, it is already too late
> (dma_contiguous_reserve() is called after map_mem()). I am not sure this can be
> reordered easily due to the dependency on dma_limits_init().
Yeah I saw that constraint too, let's leave it for future extension
then although having this reordering dependency fixed will be ideal.
>
> So the only case it would work would be when a reserved memory has set
> "linux,cma-default".
>
Let's rather just turn DT based default CMA pool to support PTE-level
mapping and keep it configurable via Kconfig or cmdline or DT.
If we need a DT declaration anyway, a Kconfig and a cmdline option aren't really
necessary.
Agree, it rather makes sense for DT independent generic CMA pool.
Also, Rob wasn't very keen on having a DT option for the reserved memory and
suggested to have a compatible string instead [1]. So perhaps we could have a
a separate driver that handles the unmapping.
linux,cma {
compatible = "shared-dma-pool", "unmap-on-alloc";
reusable;
linux,cma-default;
...
}
That "unmap-on-alloc" driver sets the memblock flags and register some cma hooks
(to be defined) for cma_alloc/cma_free where it unmaps/re-maps into the linear
map.
The nice property is optee won't need to be modified anymore. The downside is
the entire pool is now mapped at PTE-level.
Sounds reasonable to me. If folks are really worried about performance
with generic CMA PTE-level map then OP-TEE already supports no-map based
reserved carveouts for protected DMABufs too.
-Sumit
[1] http://lore.kernel.org/all/20260902172436.GB1440252-robh@xxxxxxxxxx/
--
Vincent
-Sumit
> >
> > If we can have "cma_unmap" cmdline option alongside "cma" then it can
> > turn the existing default CMA allocator to support unmapping buffers.
> > Not requiring a DT change at all, although we can have the DT property
> > for the default pool too if platforms want to opt for.
> >
> > > > >
> > > > > If we can rather support a generic CMA like allocator for arm64 which
> > > > > can support unmapped allocations then I am all for it. These platform
> > > > > specific reserved pools isn't a scalable solution.
> > > >
> > > > What do you mean by generic CMA allocator here? An extension to
> > > > shared-dma-pool? Or having the CMA automatically doing the unmap/map on
> > > > cma_alloc() cma_release()?
> > > >
> > > > Right now I have created an API for the unmapping/remapping part. It adds a
> > > > burden on the caller, but the alternative solution is introducing post-alloc
> > > > pre-dealloc callbacks in the CMA, which looked a bit overkilled to me when
> > > > the users of that "ffa-pool" are only optee and the nvidia video dma-buf heap.
> > > >
> >
> > I am not too worried about where the actual implementation lives but
> > looks like CMA driver to be the better place. I think use-cases like
> > guest_memfd proposed as a user for __GFP_UNMAPPED on arm64 will also be
> > looking for this feature.
> >
> > -Sumit
> >
> > > >
> > > > >
> > > > > > You could allocate the DMABufs from a CMA region that has no linear alias.
> > > > >
> > > > > Allocation from generic CMA pool is already the case here:
> > > > >
> > > > > tee_shm_alloc_dma_mem() ->
> > > > > dma_alloc_pages()
> > > > >
> > > > > Do you mean it's just fine to unmap buffers allocated from generic CMA
> > > > > region since it doesn't have any linear alias?
> > > > >
> > > > > >
> > > > > > > > Vincent's series tackles that by using a pool so that only that part of
> > > > > > > > memory requires the pte-level mappings in the linear map. That's the
> > > > > > > > whole point of it, so I don't think it makes sense to drop it in favour
> > > > > > > > of Brendan's approach (which doesn't work).
> > > > > > > >
> > > > > > >
> > > > > > > For protected DMABufs, we even don't require any pte-level mappings in
> > > > > > > linear map. The whole idea of mapping and unmapping is just a bottleneck
> > > > > > > for performance sensitive secure media pipeline use-cases. That's why if
> > > > > > > we can support __GFP_UNMAPPED with the memory allocator on arm64 would
> > > > > > > be the best fit for this use-case.
> > > > > >
> > > > > > The point is that you will require pte-level mappings for the linear map
> > > > > > if you want to unmap from it at runtime on arm64. If you don't, then you
> > > > > > can end up needing to split a block mapping (e.g. pmd-level) when you
> > > > > > decide to unmap only part of it and there isn't a safe way to do that on
> > > > > > arm64 without transiently unmapping the entire block, which can lead to
> > > > > > fatal translation faults during that window.
> > > > >
> > > > > I agree with you here, what I am rather trying to find if on arm64 there
> > > > > is any possibilty to have a generic allocator for unmapped pages. That's
> > > > > the use-case here as well as what Brendan't patch-set was trying to address.
> > > > >
> > > > > >
> > > > > > > Now the question is why arm64 can't support __GFP_UNMAPPED similar to
> > > > > > > how Brendan is doing it for x86?
> > > > > >
> > > > > > Because arm64 != x86? Presumably x86 doesn't have the strict
> > > > > > break-before-make requirements we have on arm64.
> > > > > >
> > > > >
> > > > > Sure, sounds like you are eluding to that the generic MM flag
> > > > > __GFP_UNMAPPED being proposed can't be supported on arm64, right?
> > > > >
> > > > > -Sumit
> > >
> > > --
> > > Vincent
>