Re: [PATCH v2] mm/hugetlb: fix a addressing exception caused by huge_pte_offset()

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 27 Feb 2020 13:41:46 -0800 Mike Kravetz <mike.kravetz@oracle.com> wrote:

> > Secondly, huge_pte_offset in mm/hugetlb.c is for ARCH_WANT_GENERAL_HUGETLB, many
> > architectures use it, can you make sure there is no issue on all the
> > architectures using it with all the version of gcc ?
> > 
> > Thirdly, there are several places use READ_ONCE to access the page table in mm/*
> > (e.g. gup_pmd_range), they're also generical for all architectures, and they're
> > much more like unnecessary than here, so why there can use but not here? What's
> > more, you can read this commit 688272809.
> 
> Apologies for the late reply.
> 
> In commit 20a004e7 the message says that "Whilst there are some scenarios
> where this cannot happen ... the overhead of using READ_ONCE/WRITE_ONCE
> everywhere is minimal and makes the code an awful lot easier to reason about."
> Therefore, a decision was made to ALWAYS use READ_ONCE in the arm64 code
> whether or not it was absolutely necessary.  Therefore, I do not think
> we can assume all the READ_ONCE additions made in 20a004e7 are necessary.
> Then the question remains, it it necessary in two statements above?
> I do not believe it is necessary.  Why?  In the statements,
> 	if (!pgd_present(*pgd))
> and
> 	if (!p4d_present(*p4d))
> the variables are only accessed and dereferenced once.  I can not imagine
> any way in which the compiler could perform multiple accesses of the variable.
> 
> I do believe the READ_ONCE in code accessing the pud and pmd is necessary.
> This is because the variables (pud_entry or pmd_entry) are accessed more than
> once.  And, I could imagine some strange compiler optimization where it would
> dereference the pud or pmd pointer more than once.  For this same reason
> (multiple accesses), I believe the READ_ONCE was added in commit 688272809.
> 
> I am no expert in this area, so corrections/comments appreciated.
> 
> BTW, I still think there may be races present in lookup_address_in_pgd().
> Multiple dereferences of a p4d, pud and pmd are done.

Based on Mike's observations I shall drop this patch.  If we still
believe it is needed, please enhance the changelog, resend and let's
take another look.