[PATCH 2/2] arm64: Allocate crashkernel always in ZONE_DMA

[Bhupesh Sharma]

Hi Will,

On Thu, Jul 2, 2020 at 1:20 PM Will Deacon <will at kernel.org> wrote:
>
> On Thu, Jul 02, 2020 at 03:44:20AM +0530, Bhupesh Sharma wrote:
> > commit bff3b04460a8 ("arm64: mm: reserve CMA and crashkernel in
> > ZONE_DMA32") allocates crashkernel for arm64 in the ZONE_DMA32.
> >
> > However as reported by Prabhakar, this breaks kdump kernel booting in
> > ThunderX2 like arm64 systems. I have noticed this on another ampere
> > arm64 machine. The OOM log in the kdump kernel looks like this:
> >
> >   [    0.240552] DMA: preallocated 128 KiB GFP_KERNEL pool for atomic allocations
> >   [    0.247713] swapper/0: page allocation failure: order:1, mode:0xcc1(GFP_KERNEL|GFP_DMA), nodemask=(null),cpuset=/,mems_allowed=0
> >   <..snip..>
> >   [    0.274706] Call trace:
> >   [    0.277170]  dump_backtrace+0x0/0x208
> >   [    0.280863]  show_stack+0x1c/0x28
> >   [    0.284207]  dump_stack+0xc4/0x10c
> >   [    0.287638]  warn_alloc+0x104/0x170
> >   [    0.291156]  __alloc_pages_slowpath.constprop.106+0xb08/0xb48
> >   [    0.296958]  __alloc_pages_nodemask+0x2ac/0x2f8
> >   [    0.301530]  alloc_page_interleave+0x20/0x90
> >   [    0.305839]  alloc_pages_current+0xdc/0xf8
> >   [    0.309972]  atomic_pool_expand+0x60/0x210
> >   [    0.314108]  __dma_atomic_pool_init+0x50/0xa4
> >   [    0.318504]  dma_atomic_pool_init+0xac/0x158
> >   [    0.322813]  do_one_initcall+0x50/0x218
> >   [    0.326684]  kernel_init_freeable+0x22c/0x2d0
> >   [    0.331083]  kernel_init+0x18/0x110
> >   [    0.334600]  ret_from_fork+0x10/0x18
> >
> > This patch limits the crashkernel allocation to the first 1GB of
> > the RAM accessible (ZONE_DMA), as otherwise we might run into OOM
> > issues when crashkernel is executed, as it might have been originally
> > allocated from either a ZONE_DMA32 memory or mixture of memory chunks
> > belonging to both ZONE_DMA and ZONE_DMA32.
>
> How does this interact with this ongoing series:
>
> https://lore.kernel.org/r/20200628083458.40066-1-chenzhou10@huawei.com
>
> (patch 4, in particular)

Many thanks for having a look at this patchset. I was not aware that
Chen had sent out a new version.
I had noted in the v9 review of the high/low range allocation
<https://lists.gt.net/linux/kernel/3726052#3726052> that I was working
on a generic solution (irrespective of the crashkernel, low and high
range allocation) which resulted in this patchset.

The issue is two-fold: OOPs in memcfg layer (PATCH 1/2, which has been
Acked-by memcfg maintainer) and OOM in the kdump kernel due to
crashkernel allocation in ZONE_DMA32 regions(s) which is addressed by
this PATCH.

I will have a closer look at the v10 patchset Chen shared, but seems
it needs some rework as per Dave's review comments which he shared
today.
IMO, in the meanwhile this patchset  can be used to fix the existing
kdump issue with upstream kernel.

> > Fixes: bff3b04460a8 ("arm64: mm: reserve CMA and crashkernel in ZONE_DMA32")
> > Cc: Johannes Weiner <hannes at cmpxchg.org>
> > Cc: Michal Hocko <mhocko at kernel.org>
> > Cc: Vladimir Davydov <vdavydov.dev at gmail.com>
> > Cc: James Morse <james.morse at arm.com>
> > Cc: Mark Rutland <mark.rutland at arm.com>
> > Cc: Will Deacon <will at kernel.org>
> > Cc: Catalin Marinas <catalin.marinas at arm.com>
> > Cc: cgroups at vger.kernel.org
> > Cc: linux-mm at kvack.org
> > Cc: linux-arm-kernel at lists.infradead.org
> > Cc: linux-kernel at vger.kernel.org
> > Cc: kexec at lists.infradead.org
> > Reported-by: Prabhakar Kushwaha <pkushwaha at marvell.com>
> > Signed-off-by: Bhupesh Sharma <bhsharma at redhat.com>
> > ---
> >  arch/arm64/mm/init.c | 16 ++++++++++++++--
> >  1 file changed, 14 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > index 1e93cfc7c47a..02ae4d623802 100644
> > --- a/arch/arm64/mm/init.c
> > +++ b/arch/arm64/mm/init.c
> > @@ -91,8 +91,15 @@ static void __init reserve_crashkernel(void)
> >       crash_size = PAGE_ALIGN(crash_size);
> >
> >       if (crash_base == 0) {
> > -             /* Current arm64 boot protocol requires 2MB alignment */
> > -             crash_base = memblock_find_in_range(0, arm64_dma32_phys_limit,
> > +             /* Current arm64 boot protocol requires 2MB alignment.
> > +              * Also limit the crashkernel allocation to the first
> > +              * 1GB of the RAM accessible (ZONE_DMA), as otherwise we
> > +              * might run into OOM issues when crashkernel is executed,
> > +              * as it might have been originally allocated from
> > +              * either a ZONE_DMA32 memory or mixture of memory
> > +              * chunks belonging to both ZONE_DMA and ZONE_DMA32.
> > +              */
>
> This comment needs help. Why does putting the crashkernel in ZONE_DMA
> prevent "OOM issues"?

Sure, I can work on adding more details in the comment so that it
explains the potential OOM issue(s) better.

Thanks,
Bhupesh