Re: [PATCH] arm64: mm: make CONFIG_ZONE_DMA configurable without EXPERT

[Catalin Marinas]

On Wed, Mar 25, 2020 at 12:34:15AM +0000, Peng Fan wrote:
> > On Tue, Mar 10, 2020 at 08:48:46PM +0800, peng.fan@xxxxxxx wrote:
> > > From: Peng Fan <peng.fan@xxxxxxx>
> > >
> > > commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32")
> > > enables both ZONE_DMA and ZONE_DMA32. The lower 1GB memory will be
> > > occupied by ZONE_DMA, this will cause CMA allocation fail on some
> > > platforms, because CMA area could not across different type of memory
> > > zones.
> > >
> > > Make CONFIG_ZONE_DMA configurable without EXPERT option could let
> > > people build non debug kernel image with CONFIG_ZONE_DMA disabled.
> > 
> > While I see why you need to toggle this feature, I'd rather try to figure out
> > whether there is a better solution that does not break the single kernel image
> > aim (i.e. the same config works for all supported SoCs).
> > 
> > When we decided to go ahead with a static 1GB ZONE_DMA for Raspberry Pi
> > 4, we thought that other platforms would be fine, ZONE_DMA32 allocations
> > fall back to ZONE_DMA. We missed the large CMA case.
> > 
> > I see a few potential options:
> > 
> > a) Ensure the CMA is contained within a single zone. 
> 
> This will break legacy dts with new version kernel.
> 
> > How large is it in your case? 
> 
> It is 1GB
> 
> > Is it allocated by the kernel dynamically or a fixed start set by
> > the boot loader?
> 
> We use alloc-ranges and size in kernel dts.
> 
> But there is only 2GB DRAM in the board.

So I guess without changing the dts, option (a) doesn't really work.

> > b) Change the CMA allocator to allow spanning multiple zones (last time
> >    I looked it wasn't trivial since it relied on some per-zone lock).
> > 
> > c) Make ZONE_DMA dynamic on arm64 and only enable it if RPi4.
> 
> Option c seems a bit easier to me :)
> 
> I will try to explore both, but if you have time to help, that would be
> appreciated.

I don't have time but option (c) was already discussed and there are
patches from Nicolas on the list:

https://lore.kernel.org/linux-arm-kernel/20190820145821.27214-5-nsaenzjulienne@xxxxxxx/

The above series was checking whether the platform is RPi4 and limiting
the ZONE_DMA size to 1GB (otherwise 4GB with ZONE_DMA32 empty). We ended
up with a static 1GB for ZONE_DMA but we missed the fact that it may
break existing platforms.

So I don't think it would be too hard to revive the above series (most
of it was already merged).

-- 
Catalin