Hello Peter, thanks for the feedback!
On Thu, 2019-10-03 at 13:51 +0200, Peter Zijlstra wrote:
> On Thu, Oct 03, 2019 at 09:11:45AM +0200, Peter Zijlstra wrote:
> > On Wed, Oct 02, 2019 at 10:33:15PM -0300, Leonardo Bras wrote:
> > > diff --git a/include/asm-generic/pgtable.h b/include/asm-generic/pgtable.h
> > > index 818691846c90..3043ea9812d5 100644
> > > --- a/include/asm-generic/pgtable.h
> > > +++ b/include/asm-generic/pgtable.h
> > > @@ -1171,6 +1171,64 @@ static inline bool arch_has_pfn_modify_check(void)
> > > #endif
> > > #endif
> > >
> > > +#ifndef __HAVE_ARCH_LOCKLESS_PGTBL_WALK_CONTROL
> > > +static inline unsigned long begin_lockless_pgtbl_walk(struct mm_struct *mm)
> > > +{
> > > + unsigned long irq_mask;
> > > +
> > > + if (IS_ENABLED(CONFIG_LOCKLESS_PAGE_TABLE_WALK_TRACKING))
> > > + atomic_inc(&mm->lockless_pgtbl_walkers);
> >
> > This will not work for file backed THP. Also, this is a fairly serious
> > contention point all on its own.
>
> Kiryl says we have tmpfs-thp, this would be broken vs that, as would
> your (PowerPC) use of mm_cpumask() for that IPI.
Could you please explain it?
I mean, why this breaks tmpfs-thp?
Also, why mm_cpumask() is also broken?
>
> > > + /*
> > > + * Interrupts must be disabled during the lockless page table walk.
> > > + * That's because the deleting or splitting involves flushing TLBs,
> > > + * which in turn issues interrupts, that will block when disabled.
> > > + */
> > > + local_irq_save(irq_mask);
> > > +
> > > + /*
> > > + * This memory barrier pairs with any code that is either trying to
> > > + * delete page tables, or split huge pages. Without this barrier,
> > > + * the page tables could be read speculatively outside of interrupt
> > > + * disabling.
> > > + */
> > > + smp_mb();
> >
> > I don't think this is something smp_mb() can guarantee. smp_mb() is
> > defined to order memory accesses, in this case the store of the old
> > flags vs whatever comes after this.
> >
> > It cannot (in generic) order against completion of prior instructions,
> > like clearing the interrupt enabled flags.
> >
> > Possibly you want barrier_nospec().
>
> I'm still really confused about this barrier. It just doesn't make
> sense.
>
> If an interrupt happens before the local_irq_disable()/save(), then it
> will discard any and all speculation that would be in progress to handle
> the exception.
>
> If there isn't an interrupt (or it happens after disable) it is
> irrelevant.
>
> Specifically, that serialize-IPI thing wants to ensure in-progress
> lookups are complete, and I can't find a scenario where
> local_irq_disable/enable() needs additional help vs IPIs. The moment an
> interrupt lands it kills speculation and forces things into
> program-order.
>
> Did you perhaps want something like:
>
> if (IS_ENABLED(CONFIG_LOCKLESS_PAGE_TABLE_WALK_TRACKING)) {
> atomic_inc(&foo);
> smp_mb__after_atomic();
> }
>
> ...
>
> if (IS_ENABLED(CONFIG_LOCKLESS_PAGE_TABLE_WALK_TRACKING)) {
> smp_mb__before_atomic();
> atomic_dec(&foo);
> }
>
> To ensure everything happens inside of the increment?
>
I need to rethink this barrier, but yes. I think that's it.
It's how it was on v4.
I have changed it because I thought it would be better this way. Well,
it was probably a mistake of my part.
> And I still think all that wrong, you really shouldn't need to wait on
> munmap().
That is something I need to better understand. I mean, before coming
with this patch, I thought exactly this: not serialize when on munmap.
But on the way I was convinced it would not work on munmap. I need to
recall why, and if it was false to assume this, re-think the whole
solution.
Best regards,
Leonardo Brás
On Thu, Oct 03, 2019 at 06:24:07PM -0300, Leonardo Bras wrote:
> Hello Peter, thanks for the feedback!
>
> On Thu, 2019-10-03 at 13:51 +0200, Peter Zijlstra wrote:
> > On Thu, Oct 03, 2019 at 09:11:45AM +0200, Peter Zijlstra wrote:
> > > On Wed, Oct 02, 2019 at 10:33:15PM -0300, Leonardo Bras wrote:
> > > > diff --git a/include/asm-generic/pgtable.h b/include/asm-generic/pgtable.h
> > > > index 818691846c90..3043ea9812d5 100644
> > > > --- a/include/asm-generic/pgtable.h
> > > > +++ b/include/asm-generic/pgtable.h
> > > > @@ -1171,6 +1171,64 @@ static inline bool arch_has_pfn_modify_check(void)
> > > > #endif
> > > > #endif
> > > >
> > > > +#ifndef __HAVE_ARCH_LOCKLESS_PGTBL_WALK_CONTROL
> > > > +static inline unsigned long begin_lockless_pgtbl_walk(struct mm_struct *mm)
> > > > +{
> > > > + unsigned long irq_mask;
> > > > +
> > > > + if (IS_ENABLED(CONFIG_LOCKLESS_PAGE_TABLE_WALK_TRACKING))
> > > > + atomic_inc(&mm->lockless_pgtbl_walkers);
> > >
> > > This will not work for file backed THP. Also, this is a fairly serious
> > > contention point all on its own.
> >
> > Kiryl says we have tmpfs-thp, this would be broken vs that, as would
> > your (PowerPC) use of mm_cpumask() for that IPI.
>
> Could you please explain it?
> I mean, why this breaks tmpfs-thp?
> Also, why mm_cpumask() is also broken?
Because shared pages are not bound by a mm; or does it not share the thp
state between mappings?
> > And I still think all that wrong, you really shouldn't need to wait on
> > munmap().
>
> That is something I need to better understand. I mean, before coming
> with this patch, I thought exactly this: not serialize when on munmap.
>
> But on the way I was convinced it would not work on munmap. I need to
> recall why, and if it was false to assume this, re-think the whole
> solution.
And once you (re)figure it out, please write it down. It is a crucial
bit of the puzzle and needs to be part of the Changelogs.
On Fri, 2019-10-04 at 13:28 +0200, Peter Zijlstra wrote:
> > Could you please explain it?
> > I mean, why this breaks tmpfs-thp?
> > Also, why mm_cpumask() is also broken?
>
> Because shared pages are not bound by a mm; or does it not share the thp
> state between mappings?
By what I could understand, even though the memory is shared, the
mapping may differ for different processes (i.e. the same physical
memory that is mapped as a hugepage in process A can be mapped as a lot
of smallpages in process B).
Did I miss something here?
> And once you (re)figure it out, please write it down. It is a crucial
> bit of the puzzle and needs to be part of the Changelogs.
I am still investing time studying this. More on this later :)
Thanks!