Re: Question about hwpoison handling of 1GB hugepage

[HORIGUCHI NAOYA(堀口 直也)]

On Thu, Mar 31, 2022 at 06:56:25PM +0800, Liu Shixin wrote:
> Hi,
> 
> Recently, I found a problem with hwpoison 1GB hugepage.
> I created a process and mapped 1GB hugepage. This process will then fork a
> child process and write/read this 1GB hugepage. Then I inject hwpoison into
> this 1GB hugepage. The child process triggers the memory failure and is
> being killed as expected. After this, the parent process will try to fork a
> new child process and do the same thing. It is killed again and finally it
> goes into such an infinite loop. I found this was caused by
> commit 31286a8484a8 ("mm: hwpoison: disable memory error handling on 1GB hugepage")

Hello Shixin,

It's little unclear to me about what behavior you are expecting and
what the infinite loop repeats, could you explain little more about them?
(I briefly tried to reproduce it, but didn't make it...)

> 
> It looks like there is a bug for hwpoison 1GB hugepage so I try to reproduce
> the bug described. After trying to revert the patch in an earlier version of
> the kernel, I reproduce the bug described. Then I try to revert the patch in
> latest version, and find the bug is no longer reproduced.
> 
> I compare the code paths of 1 GB hugepage and 2 MB hugepage for second madvise(MADV_HWPOISON),
> and find that the problem is caused because in gup_pud_range(), pud_none() and
> pud_huge() both return false and then trigger the bug. But in gup_pmd_range(),
> the pmd_none() is modified to pmd_present() which will make code return directly.
> The I find that it is commit 15494520b776 ("mm: fix gup_pud_range") which
> cause latest version not reproduced. I backport commit 15494520b776 in
> earlier version and find the bug is no longer reproduced either.

Thank you for the analysis.
So this patch might make 31286a8484a8 unnecessary, that's a good news.

> 
> So I'd like to consult that is it the time to revert commit 31286a8484a8?
> Or if we modify pud_huge to be similar with pmd_huge, is it sufficient?
> 
> I also noticed there is a TODO comment in memory_failure_hugetlb():
>     - conversion of a pud that maps an error hugetlb into hwpoison
>       entry properly works, and
>     - other mm code walking over page table is aware of pud-aligned
>       hwpoison entries. 

These are simply minimum requirements, but might not be sufficient.
We need testing (with removing 31286a8484a8) to make sure that
there's no issues on some corner cases.
(I start to extend existing hugetlb-related testcases to 1GB ones.)

Thanks,
Naoya Horiguchi

> 
> I'm not sure whether the above fix are sufficient, so is there anything else need
> to analysis that I haven't considered?