Re: [PATCH] mm/thp: Make ARCH_ENABLE_SPLIT_PMD_PTLOCK dependent on PGTABLE_LEVELS > 2

[Gerald Schaefer]

On Mon, 17 May 2021 16:13:57 +0200
Gerald Schaefer <gerald.schaefer@xxxxxxxxxxxxx> wrote:

> On Mon, 17 May 2021 09:45:31 +0530
> Anshuman Khandual <anshuman.khandual@xxxxxxx> wrote:
> 
> > 
> > 
> > On 5/10/21 10:05 AM, Anshuman Khandual wrote:  
> > > ARCH_ENABLE_SPLIT_PMD_PTLOCK is irrelevant unless there are two page table
> > > levels including PMD (also per Documentation/vm/split_page_table_lock.rst).
> > > Make this dependency explicit on remaining platforms i.e x86 and s390 where
> > > ARCH_ENABLE_SPLIT_PMD_PTLOCK is subscribed.  
> 
> For s390, I don't think this makes a lot of sense. We always have 5 levels
> defined for PGTABLE_LEVELS, and we would not even compile with any other
> value, because of the "#error CONFIG_PGTABLE_LEVELS" in include/linux/pgtable.h.
> 
> Our pagetable folding also works a bit different than it does on other archs,
> and we would actually have pmd level entries for 2-level pagetables, so it should
> all work fine also with PGTABLE_LEVELS == 2 (if it was possible).
> 
> In fact, I do not really see why you would need "more than two levels" on any
> arch, in order to use split PMD locks. Your description also just says
> "irrelevant unless there are two page table levels", and not "more than two
> levels", like in Documentation/vm/split_page_table_lock.rst.
> 
> Yet, your patch adds checks for "more than", so at least the description
> seems a bit misleading. I assume that the "more than" has to do with folded
> PMD on a 2-level system, but the way we fold on s390 I do not see why that
> should be a problem. Could you please elaborate a bit?
> 
> We also have different levels of pagetables for kernel (CONFIG_PGTABLE_LEVELS)
> and user processes on s390. The latter can have dynamic levels, currently
> starting with 3, but previously we also had 2 levels for compat tasks e.g.
> These dynamic levels for user processes are also independent from the
> CONFIG_PGTABLE_LEVELS used for the kernel pagetable, while the split PMD lock
> of course also affects user process pagetables, so that would be another
> reason not to add such a dependency for ARCH_ENABLE_SPLIT_PMD_PTLOCK on s390.

Ouch, I guess I was a bit confused here. I thought the split PMD lock
was part of the struct page for the 4 KB page where the PMD entry is located,
and therefore, with more than one page, it still would make (a little) sense
to use it also for 2 pagetable levels.

However, pmd_to_page() always returns the struct page of the first page,
so there is only one split PMD lock for the whole thing (4 pages for s390).
Of course that means that with 2 pagetable levels, and only one PMD directory,
the split PMD lock would be equivalent to the global pagetable lock, and
therefore not make any sense.

Maybe you could change the description to also mention "more than two"
levels?

I still do not see a real benefit of the patch, e.g. it does not really
fix any possible misconfiguration, at least on s390. But it certainly is not
wrong, and at least it had the benefit of making me aware again of how split
PMD locks work, so I'll happily add this

Acked-by: Gerald Schaefer <gerald.schaefer@xxxxxxxxxxxxx> # s390