Re: [PATCH v1 3/5] mm: ptdump: Provide page size to notepage()

[Christophe Leroy <christophe.leroy@csgroup.eu>]

Le 16/04/2021 à 12:51, Steven Price a écrit :
> On 16/04/2021 11:38, Christophe Leroy wrote:
>>
>>
>> Le 16/04/2021 à 11:28, Steven Price a écrit :
>>> On 15/04/2021 18:18, Christophe Leroy wrote:
>>>> In order to support large pages on powerpc, notepage()
>>>> needs to know the page size of the page.
>>>>
>>>> Add a page_size argument to notepage().
>>>>
>>>> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
>>>> ---
>>>>   arch/arm64/mm/ptdump.c         |  2 +-
>>>>   arch/riscv/mm/ptdump.c         |  2 +-
>>>>   arch/s390/mm/dump_pagetables.c |  3 ++-
>>>>   arch/x86/mm/dump_pagetables.c  |  2 +-
>>>>   include/linux/ptdump.h         |  2 +-
>>>>   mm/ptdump.c                    | 16 ++++++++--------
>>>>   6 files changed, 14 insertions(+), 13 deletions(-)
>>>>
>>> [...]
>>>> diff --git a/mm/ptdump.c b/mm/ptdump.c
>>>> index da751448d0e4..61cd16afb1c8 100644
>>>> --- a/mm/ptdump.c
>>>> +++ b/mm/ptdump.c
>>>> @@ -17,7 +17,7 @@ static inline int note_kasan_page_table(struct mm_walk *walk,
>>>>   {
>>>>       struct ptdump_state *st = walk->private;
>>>> -    st->note_page(st, addr, 4, pte_val(kasan_early_shadow_pte[0]));
>>>> +    st->note_page(st, addr, 4, pte_val(kasan_early_shadow_pte[0]), PAGE_SIZE);
>>>
>>> I'm not completely sure what the page_size is going to be used for, but note that KASAN presents 
>>> an interesting case here. We short-cut by detecting it's a KASAN region at a high level 
>>> (PGD/P4D/PUD/PMD) and instead of walking the tree down just call note_page() *once* but with 
>>> level==4 because we know KASAN sets up the page table like that.
>>>
>>> However the one call actually covers a much larger region - so while PAGE_SIZE matches the level 
>>> it doesn't match the region covered. AFAICT this will lead to odd results if you enable KASAN on 
>>> powerpc.
>>
>> Hum .... I successfully tested it with KASAN, I now realise that I tested it with 
>> CONFIG_KASAN_VMALLOC selected. In this situation, since 
>> https://github.com/torvalds/linux/commit/af3d0a686 we don't have any common shadow page table 
>> anymore.
>>
>> I'll test again without CONFIG_KASAN_VMALLOC.
>>
>>>
>>> To be honest I don't fully understand why powerpc requires the page_size - it appears to be using 
>>> it purely to find "holes" in the calls to note_page(), but I haven't worked out why such holes 
>>> would occur.
>>
>> I was indeed introduced for KASAN. We have a first commit 
>> https://github.com/torvalds/linux/commit/cabe8138 which uses page size to detect whether it is a 
>> KASAN like stuff.
>>
>> Then came https://github.com/torvalds/linux/commit/b00ff6d8c as a fix. I can't remember what the 
>> problem was exactly, something around the use of hugepages for kernel memory, came as part of the 
>> series 
>> https://patchwork.ozlabs.org/project/linuxppc-dev/cover/cover.1589866984.git.christophe.leroy@csgroup.eu/ 
> 
> 
> Ah, that's useful context. So it looks like powerpc took a different route to reducing the KASAN 
> output to x86.
> 
> Given the generic ptdump code has handling for KASAN already it should be possible to drop that from 
> the powerpc arch code, which I think means we don't actually need to provide page size to 
> notepage(). Hopefully that means more code to delete ;)
> 

Yes ... and no.

It looks like the generic ptdump handles the case when several pgdir entries points to the same 
kasan_early_shadow_pte. But it doesn't take into account the powerpc case where we have regular page 
tables where several (if not all) PTEs are pointing to the kasan_early_shadow_page .

Christophe