RE: [PATCH RFC 0/3] riscv: Add DMA_COHERENT support

From: Anup Patel
Date: Sun Jun 06 2021 - 23:38:39 EST




> -----Original Message-----
> From: Guo Ren <guoren@xxxxxxxxxx>
> Sent: 06 June 2021 22:42
> To: Anup Patel <Anup.Patel@xxxxxxx>; Atish Patra <atishp@xxxxxxxxxxxxxx>
> Cc: Palmer Dabbelt <palmer@xxxxxxxxxxx>; anup@xxxxxxxxxxxxxx;
> drew@xxxxxxxxxxxxxxx; Christoph Hellwig <hch@xxxxxx>; wefu@xxxxxxxxxx;
> lazyparser@xxxxxxxxx; linux-riscv@xxxxxxxxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; linux-arch@xxxxxxxxxxxxxxx; linux-
> sunxi@xxxxxxxxxxxxxxx; guoren@xxxxxxxxxxxxxxxxx; Paul Walmsley
> <paul.walmsley@xxxxxxxxxx>
> Subject: Re: [PATCH RFC 0/3] riscv: Add DMA_COHERENT support
>
> Hi Anup and Atish,
>
> On Thu, Jun 3, 2021 at 2:00 PM Anup Patel <Anup.Patel@xxxxxxx> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Palmer Dabbelt <palmer@xxxxxxxxxxx>
> > > Sent: 03 June 2021 09:43
> > > To: guoren@xxxxxxxxxx
> > > Cc: anup@xxxxxxxxxxxxxx; drew@xxxxxxxxxxxxxxx; Christoph Hellwig
> > > <hch@xxxxxx>; Anup Patel <Anup.Patel@xxxxxxx>; wefu@xxxxxxxxxx;
> > > lazyparser@xxxxxxxxx; linux-riscv@xxxxxxxxxxxxxxxxxxx; linux-
> > > kernel@xxxxxxxxxxxxxxx; linux-arch@xxxxxxxxxxxxxxx; linux-
> > > sunxi@xxxxxxxxxxxxxxx; guoren@xxxxxxxxxxxxxxxxx; Paul Walmsley
> > > <paul.walmsley@xxxxxxxxxx>
> > > Subject: Re: [PATCH RFC 0/3] riscv: Add DMA_COHERENT support
> > >
> > > On Sat, 29 May 2021 17:30:18 PDT (-0700), Palmer Dabbelt wrote:
> > > > On Fri, 21 May 2021 17:36:08 PDT (-0700), guoren@xxxxxxxxxx wrote:
> > > >> On Wed, May 19, 2021 at 3:15 PM Anup Patel <anup@xxxxxxxxxxxxxx>
> > > wrote:
> > > >>>
> > > >>> On Wed, May 19, 2021 at 12:24 PM Drew Fustini
> > > <drew@xxxxxxxxxxxxxxx> wrote:
> > > >>> >
> > > >>> > On Wed, May 19, 2021 at 08:06:17AM +0200, Christoph Hellwig
> > > wrote:
> > > >>> > > On Wed, May 19, 2021 at 02:05:00PM +0800, Guo Ren wrote:
> > > >>> > > > Since the existing RISC-V ISA cannot solve this problem,
> > > >>> > > > it is better to provide some configuration for the SOC
> > > >>> > > > vendor to
> > > customize.
> > > >>> > >
> > > >>> > > We've been talking about this problem for close to five years.
> > > >>> > > So no, if you don't manage to get the feature into the ISA
> > > >>> > > it can't be supported.
> > > >>> >
> > > >>> > Isn't it a good goal for Linux to support the capabilities
> > > >>> > present in the SoC that a currently being fab'd?
> > > >>> >
> > > >>> > I believe the CMO group only started last year [1] so the
> > > >>> > RV64GC SoCs that are going into mass production this year
> > > >>> > would not have had the opporuntiy of utilizing any RISC-V ISA
> > > >>> > extension for handling cache management.
> > > >>>
> > > >>> The current Linux RISC-V policy is to only accept patches for
> > > >>> frozen or ratified ISA specs.
> > > >>> (Refer, Documentation/riscv/patch-acceptance.rst)
> > > >>>
> > > >>> This means even if emulate CMO instructions in OpenSBI, the
> > > >>> Linux patches won't be taken by Palmer because CMO specification
> > > >>> is still in draft stage.
> > > >> Before CMO specification release, could we use a sbi_ecall to
> > > >> solve the current problem? This is not against the specification,
> > > >> when CMO is ready we could let users choose to use the new CMO in
> Linux.
> > > >>
> > > >> From a tech view, CMO trap emulation is the same as sbi_ecall.
> > > >>
> > > >>>
> > > >>> Also, we all know how much time it takes for RISCV international
> > > >>> to freeze some spec. Judging by that we are looking at another
> > > >>> 3-4 years at minimum.
> > > >
> > > > Sorry for being slow here, this thread got buried.
> > > >
> > > > I've been trying to work with a handful of folks at the RISC-V
> > > > foundation to try and get a subset of the various in-development
> > > > specifications (some simple CMOs, something about non-caching in
> > > > the page tables, and some way to prevent speculative accesse from
> > > > generating coherence traffic that will break non-coherent systems).
> > > > I'm not sure we can get this together quickly, but I'd prefer to
> > > > at least try before we jump to taking vendor-specificed behavior here.
> > > > It's obviously an up-hill battle to try and get specifications
> > > > through the process and I'm certainly not going to promise it will
> > > > work, but I'm hoping that the impending need to avoid forking the
> > > > ISA will be sufficient to get people behind producing some
> > > > specifications in a timely
> > > fashion.
> > > >
> > > > I wasn't aware than this chip had non-coherent devices until I saw
> > > > this thread, so we'd been mostly focused on the Beagle V chip.
> > > > That was in a sense an easier problem because the SiFive IP in it
> > > > was never designed to have non-coherent devices so we'd have to
> > > > make anything work via a series of slow workarounds, which would
> > > > make emulating the eventually standardized behavior reasonable in
> > > > terms of performance (ie, everything would be super slow so who really
> cares).
> > > >
> > > > I don't think relying on some sort of SBI call for the CMOs whould
> > > > be such a performance hit that it would prevent these systems from
> > > > being viable, but assuming you have reasonable performance on your
> > > > non-cached accesses then that's probably not going to be viable to
> > > > trap and emulate. At that point it really just becomes silly to
> > > > pretend that we're still making things work by emulating the
> > > > eventually ratified behavior, as anyone who actually tries to use
> > > > this thing to do IO would need out of tree patches. I'm not sure
> > > > exactly what the plan is for the page table bits in the
> > > > specification right now, but if you can give me a pointer to some
> > > > documentation then I'm happy to try and push for something
> compatible.
> > > >
> > > > If we can't make the process work at the foundation then I'd be
> > > > strongly in favor of just biting the bullet and starting to take
> > > > vendor-specific code that's been implemented in hardware and is
> > > > necessarry to make things work acceptably. That's obviously a
> > > > sub-optimal solution as it'll lead to a bunch of ISA
> > > > fragmentation, but at least we'll be able to keep the software stack
> together.
> > > >
> > > > Can you tell us when these will be in the hands of users? That's
> > > > pretty important here, as I don't want to be blocking real users
> > > > from having their hardware work. IIRC there were some plans to
> > > > distribute early boards, but it looks like the foundation got
> > > > involved and I guess I lost the thread at that point.
> > > >
> > > > Sorry this is all such a headache, but hopefully we can get things
> > > > sorted out.
> > >
> > > I talked with some of the RISC-V foundation folks, we're not going
> > > to have an ISA specification for the non-coherent stuff any time
> > > soon. I took a look at this code and I definately don't want to
> > > take it as is, but I'm not opposed to taking something that makes the
> hardware work as long as it's a lot cleaner.
> > > We've already got two of these non-coherent chips, I'm sure more
> > > will come, and I'd rather have the extra headaches than make
> > > everyone fork the software stack.
> >
> > Thanks for confirming. The CMO extension is still in early stages so
> > it will certainly take more time for them. After CMO extension is
> > finalized, it will take some more time to have actual RISC-V platforms with
> CMO implemented.
> >
> > >
> > > After talking to Atish it looks like there's likely to be an SBI
> > > extension to handle the CMOs, which should let us avoid the bulk of
> > > the vendor-specific behavior in the kernel. I know some people are
> > > worried about adding to the SBI surface. I'm worried about that
> > > too, but that's way better than sticking a bunch of vendor-specific
> > > instructions into the kernel. The SBI extension should make for a
> > > straight-forward cache flush implementation in Linux, so let's just plan on
> that getting through quickly (as has been done before).
> >
> > Yes, I agree. We can have just a single SBI call which is meant for
> > DMA sync purpose only which means it will flush/invalidate pages from
> > all cache levels irrespective of the cache hierarchy (i.e.
> > flush/invalidate to RAM). The CMO extension might more generic cache
> > operations which can target any cache level.
> >
> > I am already preparing a write-up for SBI DMA sync call in SBI spec. I
> > will share it with you separately as well.
> >
> > >
> > > Unfortunately we've yet to come up with a way to handle the
> > > non-cacheable mappings without introducing a degree of
> > > vendor-specific behavior or seriously impacting performance (mark
> > > them as not valid and deal with them in the trap handler). I'm not
> > > really sure it counts as supporting the hardware if it's massively
> > > slow, so that really leaves us with vendor-specific mappings as the only
> option to make these chips work.
> >
> > A RISC-V platform can have non-cacheable mappings is following
> > possible
> > ways:
> > 1) Fixed physical address range as non-cacheable using PMAs
> > 2) Custom page table attributes
> > 3) Svpmbt extension being defined by RVI
> >
> > Atish and me both think it is possible to have RISC-V specific DMA ops
> > implementation which can handle all above case. Atish is already
> > working on DMA ops implementation for RISC-V.
> Not only DMA ops, but also icache_sync & __vdso_icache_sync. Please have a
> look at:
> https://lore.kernel.org/linux-riscv/1622970249-50770-12-git-send-email-
> guoren@xxxxxxxxxx/T/#u

The icache_sync and __vdso_icache_sync will have to be addressed
differently. The SBI DMA sync extension cannot address this.

It seems Allwinner D1 have more non-standard stuff:
1) Custom PTE bits for IO-coherent access
2) Custom data cache flush/invalidate for DMA sync
3) Custom icache flush/invalidate

Other hand, BeagleV has only two problems:
1) Custom physical address range for IO-coherent access
2) Custom L2 cache flush/invalidate for DMA sync

From above #2, can be solved by SBI DMA sync call and
Linux DMA ops for both BeagleV and Allwinner D1

On BeagleV, issue #1 can be solved using "dma-ranges".

On Allwinner D1, issues #1 and #3 need to be addressed
separately.

I think supporting BeagleV in upstream Linux is relatively
easy compared to Allwinner D1.

@Guo, please check if you can reserve dedicated
physical address range for IO-coherent access (just like
BeagleV). If yes, then we can tackle issue #1 for Allwinner
D1 using "dma-ranges" DT property.

Regards,
Anup

>
>
> >
> > >
> > > This implementation, which adds some Kconfig entries that control
> > > page table bits, definately isn't suitable for upstream. Allowing
> > > users to set arbitrary page table bits will eventually conflict with
> > > the standard, and is just going to be a mess. It'll also lead to
> > > kernels that are only compatible with specific designs, which we're
> > > trying very hard to avoid. At a bare minimum we'll need some way to
> > > detect systems with these page table bits before setting them, and
> > > some description of what the bits actually do so we can reason about
> them.
> >
> > Yes, vendor specific Kconfig options are strict NO NO. We can't
> > give-up the goal of unified kernel image for all platforms.
> >
> > Regards,
> > Anup
>
>
>
> --
> Best Regards
> Guo Ren
>
> ML: https://lore.kernel.org/linux-csky/