Re: [PATCH V15] mm/debug: Add tests validating architecture page table helpers

On 03/07/2020 12:35 PM, Christophe Leroy wrote:

Le 07/03/2020 à 01:56, Anshuman Khandual a écrit :

quoted

On 03/07/2020 06:04 AM, Qian Cai wrote:

quoted

quoted

On Mar 6, 2020, at 7:03 PM, Anshuman Khandual [off-list ref] wrote:

Hmm, set_pte_at() function is not preferred here for these tests. The idea
is to avoid or atleast minimize TLB/cache flushes triggered from these sort
of 'static' tests. set_pte_at() is platform provided and could/might trigger
these flushes or some other platform specific synchronization stuff. Just

Why is that important for this debugging option?

Primarily reason is to avoid TLB/cache flush instructions on the system
during these tests that only involve transforming different page table
level entries through helpers. Unless really necessary, why should it
emit any TLB/cache flush instructions ?

What's the problem with thoses flushes ?

quoted

quoted

quoted

wondering is there specific reason with respect to the soft lock up problem
making it necessary to use set_pte_at() rather than a simple WRITE_ONCE() ?

Looks at the s390 version of set_pte_at(), it has this comment,
vmaddr);

/*
  * Certain architectures need to do special things when PTEs
  * within a page table are directly modified.  Thus, the following
  * hook is made available.
  */

I can only guess that powerpc  could be the same here.

This comment is present in multiple platforms while defining set_pte_at().
Is not 'barrier()' here alone good enough ? Else what exactly set_pte_at()
does as compared to WRITE_ONCE() that avoids the soft lock up, just trying
to understand.

Argh ! I didn't realise that you were writing directly into the page tables. When it works, that's only by chance I guess.

To properly set the page table entries, set_pte_at() has to be used:
- On powerpc 8xx, with 16k pages, the page table entry must be copied four times. set_pte_at() does it, WRITE_ONCE() doesn't.
- On powerpc book3s/32 (hash MMU), the flag _PAGE_HASHPTE must be preserved among writes. set_pte_at() preserves it, WRITE_ONCE() doesn't.

set_pte_at() also does a few other mandatory things, like calling pte_mkpte()

So, the WRITE_ONCE() must definitely become a set_pte_at()

Sure, will do. These are part of the clear tests that populates a given
entry with a non zero value before clearing and testing it with pxx_none().
In that context, WRITE_ONCE() seemed sufficient. But pte_clear() might be
closely tied with proper page table entry update and hence a preceding
set_pte_at() will be better.

There are still more WRITE_ONCE() for other page table levels during these
clear tests. set_pmd_at() and set_pud_at() are defined on platforms that
support (and enable) THP and PUD based THP respectively. Hence they could
not be used for clear tests as remaining helpers pmd_clear(), pud_clear(),
p4d_clear() and pgd_clear() still need to be validated with or without
THP support and enablement. We should just leave all other WRITE_ONCE()
instances unchanged. Please correct me if I am missing something here.

Christophe