On Thu, 2019-10-31 at 15:51 +0000, Catalin Marinas wrote:
> (sorry, I've been away last week and only now caught up with emails)
> 
> On Tue, Oct 22, 2019 at 01:23:32PM +0200, Nicolas Saenz Julienne wrote:
> > On Mon, 2019-10-21 at 16:36 -0400, Qian Cai wrote:
> > > I managed to get more information here,
> > > 
> > > [    0.000000] cma: dma_contiguous_reserve(limit c0000000)
> > > [    0.000000] cma: dma_contiguous_reserve: reserving 64 MiB for global
> > > area
> > > [    0.000000] cma: cma_declare_contiguous(size 0x0000000004000000, base
> > > 0x0000000000000000, limit 0x00000000c0000000 alignment 0x0000000000000000)
> > > [    0.000000] cma: Failed to reserve 512 MiB
> > > 
> > > Full dmesg:
> > > 
> > > https://cailca.github.io/files/dmesg.txt
> > 
> > OK I got it, reproduced it too.
> > 
> > Here are the relevant logs:
> > 
> > 	[    0.000000]   DMA      [mem 0x00000000802f0000-0x00000000bfffffff]
> > 	[    0.000000]   DMA32    [mem 0x00000000c0000000-0x00000000ffffffff]
> > 	[    0.000000]   Normal   [mem 0x0000000100000000-0x00000097fcffffff]
> > 
> > As you can see ZONE_DMA spans from 0x00000000802f0000-0x00000000bfffffff
> > which
> > is slightly smaller than 1GB.
> > 
> > 	[    0.000000] crashkernel reserved: 0x000000009fe00000 -
> > 0x00000000bfe00000 (512 MB)
> > 
> > Here crashkernel reserved 512M in ZONE_DMA.
> > 
> > 	[    0.000000] cma: Failed to reserve 512 MiB
> > 
> > CMA tried to allocate 512M in ZONE_DMA which fails as there is no enough
> > space.
> > Makes sense.
> > 
> > A fix could be moving crashkernel reservations after CMA and then if unable
> > to
> > fit in ZONE_DMA try ZONE_DMA32 before bailing out. Maybe it's a little over
> > the
> > top, yet although most devices will be fine with ZONE_DMA32, the RPi4 needs
> > crashkernel to be reserved in ZONE_DMA.
> 
> Does RPi4 need CMA in ZONE_DMA? If not, I'd rather reserve the CMA from
> ZONE_DMA32.

Yes, CMA is imperatively to be reserved in ZONE_DMA.

> Even if you moved the crash kernel, someone else might complain that
> they had 2GB of CMA and it no longer works.

I have yet to look into it, but I've been told that on x86/x64 they have a
'high' flag to be set alongside with crashkernel that forces the allocation
into ZONE_DMA32. We could mimic this behavior for big servers that don't depend
on ZONE_DMA but need to reserve big chunks of memory.

Regards,
Nicolas