Re: [PATCH 1/4] of/fdt: Update zone_dma_bits when running in bcm2711

[Catalin Marinas <catalin.marinas@arm.com>]

On Sat, Oct 10, 2020 at 12:53:19PM +0200, Nicolas Saenz Julienne wrote:
> On Sat, 2020-10-10 at 12:36 +0200, Ard Biesheuvel wrote:
> > On Fri, 9 Oct 2020 at 19:10, Catalin Marinas <catalin.marinas@arm.com> wrote:
> > > On Fri, Oct 09, 2020 at 06:23:06PM +0200, Ard Biesheuvel wrote:
> > > > On Fri, 9 Oct 2020 at 17:24, Lorenzo Pieralisi
> > > > <lorenzo.pieralisi@arm.com> wrote:
> > > > > We can move this check to IORT code and call it from arm64 if it
> > > > > can be made to work.
> > > > 
> > > > Finding the smallest value in the IORT, and assigning it to
> > > > zone_dma_bits if it is < 32 should be easy. But as I understand it,
> > > > having these separate DMA and DMA32 zones is what breaks kdump, no? So
> > > > how is this going to fix the underlying issue?
> > > 
> > > If zone_dma_bits is 32, ZONE_DMA32 disappears into ZONE_DMA (GFP_DMA32
> > > allocations fall back to ZONE_DMA).
> > > 
> > > kdump wants DMA-able memory and, without a 30-bit ZONE_DMA, that would
> > > be the bottom 32-bit. With the introduction of ZONE_DMA, this suddenly
> > > became 1GB. We could change kdump to allocate ZONE_DMA32 but this one
> > > may also be small as it lost 1GB to ZONE_DMA. However, the kdump kernel
> > > would need to be rebuilt without ZONE_DMA since it won't have any. IIRC
> > > (it's been a while since I looked), the kdump allocation couldn't span
> > > multiple zones.
> > > 
> > > In a separate thread, we try to fix kdump to use allocations above 4G as
> > > a fallback but this only fixes platforms with enough RAM (and maybe it's
> > > only those platforms that care about kdump).
> > > 
> > 
> > One thing that strikes me as odd is that we are applying the same
> > shifting logic to ZONE_DMA as we are applying to ZONE_DMA32, i.e., if
> > DRAM starts outside of the zone, it is shifted upwards.
> > 
> > On a typical ARM box, this gives me
> > 
> > [    0.000000] Zone ranges:
> > [    0.000000]   DMA      [mem 0x0000000080000000-0x00000000bfffffff]
> > [    0.000000]   DMA32    [mem 0x00000000c0000000-0x00000000ffffffff]
> > [    0.000000]   Normal   [mem 0x0000000100000000-0x0000000fffffffff]
> > 
> > i.e., the 30-bit addressable range has bit 31 set, which is weird.
> 
> Yes I agree it's weird, and IMO plain useless. I sent a series this summer to
> address this[1], which ultimately triggered the discussion we're having right
> now.

I don't mind assuming that ZONE_DMA is always from pfn 0 (i.e. no
dma_offset for such constrained devices). But if ZONE_DMA32 is squeezed
out with ZONE_DMA extended to 4GB, it should allow non-zero upper 32
bits. IIRC we do have SoCs with RAM starting above 4GB.

However, your patch didn't completely solve the problem for non-RPi4
platforms as there's hardware with RAM starting at 0 that doesn't need
the 1GB ZONE_DMA. We may end up with a combination of the two
approaches.

> Although with with your latest patch and the DT counterpart, we should be OK.
> It would be weird for a HW description to define DMA constraints that are
> impossible to reach on that system.

I don't remember the difficulties with parsing a DT early for inferring
the ZONE_DMA requirements. Could we not check the dma-ranges property in
the soc node? I can see bcm2711.dtsi defines a 30-bit address range. We
are not interested in the absolute physical/bus addresses, just the
size to check whether it's smaller than 32-bit.

> > I wonder if it wouldn't be better (and less problematic in the general
> > case) to drop this logic for ZONE_DMA, and simply let it remain empty
> > unless there is really some memory there.
> 
> From my experience, you can't have empty ZONE_DMA when enabled. Empty
> ZONE_DMA32 is OK though. Although I'm sure it's something that can be changed.

Indeed, because we still have GFP_DMA around which can't fall back to
ZONE_DMA32 (well, unless CONFIG_ZONE_DMA is disabled).

-- 
Catalin