kswapd0: page allocation failure: order:0, mode:0x820(GFP_ATOMIC), nodemask=(null),cpuset=/,mems_allowed=0 (Kernel v6.5.9, 32bit ppc)

[Michael Ellerman]

Yu Zhao <yuzhao at google.com> writes:
> On Wed, Jun 5, 2024 at 9:12 PM Michael Ellerman <mpe at ellerman.id.au> wrote:
>>
>> David Hildenbrand <david at redhat.com> writes:
>> > On 01.06.24 08:01, Yu Zhao wrote:
>> >> On Wed, May 15, 2024 at 4:06 PM Yu Zhao <yuzhao at google.com> wrote:
>> ...
>> >>
>> >> Your system has 2GB memory and it uses zswap with zsmalloc (which is
>> >> good since it can allocate from the highmem zone) and zstd/lzo (which
>> >> doesn't matter much). Somehow -- I couldn't figure out why -- it
>> >> splits the 2GB into a 0.25GB DMA zone and a 1.75GB highmem zone:
>> >>
>> >> [    0.000000] Zone ranges:
>> >> [    0.000000]   DMA      [mem 0x0000000000000000-0x000000002fffffff]
>> >> [    0.000000]   Normal   empty
>> >> [    0.000000]   HighMem  [mem 0x0000000030000000-0x000000007fffffff]
>> >
>> > That's really odd. But we are messing with "PowerMac3,6", so I don't
>> > really know what's right or wrong ...
>>
>> The DMA zone exists because 9739ab7eda45 ("powerpc: enable a 30-bit
>> ZONE_DMA for 32-bit pmac") selects it.
>>
>> It's 768MB (not 0.25GB) because it's clamped at max_low_pfn:
>
> Right. (I meant 0.75GB.)
>
>> #ifdef CONFIG_ZONE_DMA
>>         max_zone_pfns[ZONE_DMA] = min(max_low_pfn,
>>                                       1UL << (zone_dma_bits - PAGE_SHIFT));
>> #endif
>>
>> Which comes eventually from CONFIG_LOWMEM_SIZE, which defaults to 768MB.
>
> I see. I grep'ed VMSPLIT which is used on x86 and arm but apparently
> not on powerpc.

Those VMSPLIT configs are nice, on powerpc it's all done manually :}

>> I think it's 768MB because the user:kernel split is 3G:1G, and then the
>> kernel needs some of that 1G virtual space for vmalloc/ioremap/highmem,
>> so it splits it 768M:256M.
>>
>> Then ZONE_NORMAL is empty because it is also limited to max_low_pfn:
>>
>>         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
>>
>> The rest of RAM is highmem.
>>
>> So I think that's all behaving as expected, but I don't know 32-bit /
>> highmem stuff that well so I could be wrong.
>
> Yes, the three zones work as intended.
>
> Erhard,
>
> Since your system only has 2GB memory, I'd try the 2G:2G split, which
> would in theory allow both the kernel and userspace to all memory.
>
> CONFIG_LOWMEM_SIZE_BOOL=y
> CONFIG_LOWMEM_SIZE=0x7000000
>
> (Michael, please correct me if the above wouldn't work.)

It's a bit more complicated, in order to increase LOWMEM_SIZE you need
to adjust all the other variables to make space.

To get 2G of user virtual space I think you need:

CONFIG_ADVANCED_OPTIONS=y
CONFIG_LOWMEM_SIZE_BOOL=y
CONFIG_LOWMEM_SIZE=0x60000000
CONFIG_PAGE_OFFSET_BOOL=y
CONFIG_PAGE_OFFSET=0x90000000
CONFIG_KERNEL_START_BOOL=y
CONFIG_KERNEL_START=0x90000000
CONFIG_PHYSICAL_START=0x00000000
CONFIG_TASK_SIZE_BOOL=y
CONFIG_TASK_SIZE=0x80000000

Which results in 1.5GB of lowmem.

Or if you want to map all 2G of RAM directly in the kernel without
highmem, but limit user virtual space to 1.5G:

CONFIG_ADVANCED_OPTIONS=y
CONFIG_LOWMEM_SIZE_BOOL=y
CONFIG_LOWMEM_SIZE=0x80000000
CONFIG_PAGE_OFFSET_BOOL=y
CONFIG_PAGE_OFFSET=0x70000000
CONFIG_KERNEL_START_BOOL=y
CONFIG_KERNEL_START=0x70000000
CONFIG_PHYSICAL_START=0x00000000
CONFIG_TASK_SIZE_BOOL=y
CONFIG_TASK_SIZE=0x60000000

You can also reclaim another 256MB of virtual space if you disable
CONFIG_MODULES.

Those configs do boot on qemu. But I don't have easy access to my 32-bit
machine to test if they boot on actual hardware.

cheers