On Mon, 2021-03-22 at 14:40 -0400, Jon Masters wrote:
> On 3/22/21 2:34 PM, Jon Masters wrote:
> > Hi Nicolas,
> > 
> > On 11/7/19 4:56 AM, Nicolas Saenz Julienne wrote:
> > > With the introduction of ZONE_DMA in arm64 we moved the default CMA and
> > > crashkernel reservation into that area. This caused a regression on big
> > > machines that need big CMA and crashkernel reservations. Note that
> > > ZONE_DMA is only 1GB big.
> > > 
> > > Restore the previous behavior as the wide majority of devices are OK
> > > with reserving these in ZONE_DMA32. The ones that need them in ZONE_DMA
> > > will configure it explicitly.
> > > 
> > > Reported-by: Qian Cai <cai@lca.pw>
> > > Signed-off-by: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
> > > ---
> > >   arch/arm64/mm/init.c | 4 ++--
> > >   1 file changed, 2 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > > index 580d1052ac34..8385d3c0733f 100644
> > > --- a/arch/arm64/mm/init.c
> > > +++ b/arch/arm64/mm/init.c
> > > @@ -88,7 +88,7 @@ static void __init reserve_crashkernel(void)
> > >       if (crash_base == 0) {
> > >           /* Current arm64 boot protocol requires 2MB alignment */
> > > -        crash_base = memblock_find_in_range(0, ARCH_LOW_ADDRESS_LIMIT,
> > > +        crash_base = memblock_find_in_range(0, arm64_dma32_phys_limit,
> > >                   crash_size, SZ_2M);
> > >           if (crash_base == 0) {
> > >               pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
> > > @@ -454,7 +454,7 @@ void __init arm64_memblock_init(void)
> > >       high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
> > > -    dma_contiguous_reserve(arm64_dma_phys_limit ? : 
> > > arm64_dma32_phys_limit);
> > > +    dma_contiguous_reserve(arm64_dma32_phys_limit);
> > >   }
> > >   void __init bootmem_init(void)
> > 
> > Can we get a bit more of a backstory about what the regression was on 
> > larger machines? If the 32-bit DMA region is too small, but the machine 
> > otherwise has plenty of memory, the crashkernel reservation will fail. 
> > Most e.g. enterprise users aren't going to respond to that situation by 
> > determining the placement manually, they'll just not have a crashkernel.
> 
> Nevermind, looks like Catalin already changed this logic in Jan 2021 by 
> removing arm64_dma32_phys_limit and I'm out of date.

Also see this series (already merged):

https://lore.kernel.org/linux-arm-kernel/20201119175400.9995-1-nsaenzjulienne@suse.de/

Regads,
Nicolas