[PATCH v2] arm64: hibernate: Fix level3 translation fault in swsusp_save()

[Catalin Marinas]

For some reason I missed the updated patch.

On Fri, Mar 01, 2024 at 10:19:24AM +0800, Yaxiong Tian wrote:
> From: Yaxiong Tian <tianyaxiong at kylinos.cn>
> 
> On ARM64 machines using UEFI, if can_set_direct_map() return false by
> setting some CONFIGS in kernel build or grub,such as
> NO CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT、NO CONFIG_KFENCE
> NO CONFIG_RODATA_FULL_DEFAULT_ENABLED.Also with setting rodata=off、
> debug_pagealloc=off in grub and NO CONFIG_KFENCE.
> swsusp_save() will fail due to can't finding the map table under the 
> nomap memory.such as:
[...]
> [   48.532162] Call trace:
> [   48.532162]  swsusp_save+0x280/0x538
> [   48.532162]  swsusp_arch_suspend+0x148/0x190
> [   48.532162]  hibernation_snapshot+0x240/0x39c
> [   48.532162]  hibernate+0xc4/0x378
> [   48.532162]  state_store+0xf0/0x10c
> [   48.532162]  kobj_attr_store+0x14/0x24
> 
> This issue can be reproduced in QEMU using UEFI when booting with 
> rodata=off、debug_pagealloc=off in grub and NO CONFIG_KFENCE.
> 
> This is because in swsusp_save()->copy_data_pages()->page_is_saveable(),
> kernel_page_present() presumes that a page is present when can_set_direct_map()
> returns false even for NOMAP ranges.So NOMAP pages will saved in after,and then
> cause level3 translation fault in this pages.

I can see how kernel_page_present() ended up returning true if
!can_set_direct_map(), though based on the function naming only, it
feels a bit unintuitive. Is arm64 the only architecture making use of
MEMBLOCK_NOMAP? Or is it the only one where kernel_page_present() also
returns true if !can_set_direct_map()?

> diff --git a/arch/arm64/kernel/hibernate.c b/arch/arm64/kernel/hibernate.c
> index 02870beb271e..d90005de1d26 100644
> --- a/arch/arm64/kernel/hibernate.c
> +++ b/arch/arm64/kernel/hibernate.c
> @@ -94,7 +94,7 @@ int pfn_is_nosave(unsigned long pfn)
>  	unsigned long nosave_end_pfn = sym_to_pfn(&__nosave_end - 1);
>  
>  	return ((pfn >= nosave_begin_pfn) && (pfn <= nosave_end_pfn)) ||
> -		crash_is_nosave(pfn);
> +		crash_is_nosave(pfn) || !pfn_is_map_memory(pfn);
>  }

This indeed fixes the problem but it looks like an arm64-specific
workaround. I can see at least arm, loongarch and riscv making use of
memblock_is_map_memory() (which is what pfn_is_map_memory() calls). Do
they not have the same problem? On riscv, for example,
kernel_page_present() does not depend on any ARCH_HAS_SET_DIRECT_MAP
related options/conditions (neither does x86 though not sure it cares
about MEMBLOCK_NOMAP). Should we do the same for arm64 and drop the
!can_set_direct_map() condition in kernel_page_present()?

-- 
Catalin