TLB flushes on s390

TLB flushes on s390

[Kirill A. Shutemov]

Hi,

I'm looking though s390 code around page table handling and I found that
in many places s390 does tlb flush before changing page table entry.

Let's look for instance on pmdp_clear_flush() implementation on s390.
It's implemented with pmdp_get_and_clear() which does pmdp_flush_direct()
*before* pmd_clear(). That's invert order comparing to generic
pmdp_flush_direct().

The question is what prevents tlb from being re-fill between flushing tlb
and clearing page table entry?

-- 
 Kirill A. Shutemov

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: TLB flushes on s390

[Martin Schwidefsky]

On Wed, 6 May 2015 14:29:39 +0300
"Kirill A. Shutemov" <kirill@shutemov.name> wrote:

> I'm looking though s390 code around page table handling and I found that
> in many places s390 does tlb flush before changing page table entry.

Uhh, have fun with that.. it is complicated :-/
 
> Let's look for instance on pmdp_clear_flush() implementation on s390.
> It's implemented with pmdp_get_and_clear() which does pmdp_flush_direct()
> *before* pmd_clear(). That's invert order comparing to generic
> pmdp_flush_direct().
> 
> The question is what prevents tlb from being re-fill between flushing tlb
> and clearing page table entry?
 
Look again at pmdp_flush_direct(), either __pmdp_idte_local or __pmdp_idte is
called. Both functions use the IDTE instruction but in two different flavors.
The mnemonic IDTE stands for invalidate-dat-table-entry, the instruction sets
the invalid bit in the PMD and flushes all TLB entries on all CPUs that are
affected by the now invalid PMD. The pmd_clear after the pmdp_flush_direct is
done to set all the other bits of the PMD to the "empty" state. The invalid
bit is already set prior to pmd_clear.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: TLB flushes on s390

[Kirill A. Shutemov]

On Wed, May 06, 2015 at 02:00:02PM +0200, Martin Schwidefsky wrote:
> On Wed, 6 May 2015 14:29:39 +0300
> "Kirill A. Shutemov" <kirill@shutemov.name> wrote:
> 
> > I'm looking though s390 code around page table handling and I found that
> > in many places s390 does tlb flush before changing page table entry.
> 
> Uhh, have fun with that.. it is complicated :-/
>  
> > Let's look for instance on pmdp_clear_flush() implementation on s390.
> > It's implemented with pmdp_get_and_clear() which does pmdp_flush_direct()
> > *before* pmd_clear(). That's invert order comparing to generic
> > pmdp_flush_direct().
> > 
> > The question is what prevents tlb from being re-fill between flushing tlb
> > and clearing page table entry?
>  
> Look again at pmdp_flush_direct(), either __pmdp_idte_local or __pmdp_idte is
> called. Both functions use the IDTE instruction but in two different flavors.
> The mnemonic IDTE stands for invalidate-dat-table-entry, the instruction sets
> the invalid bit in the PMD and flushes all TLB entries on all CPUs that are
> affected by the now invalid PMD. The pmd_clear after the pmdp_flush_direct is
> done to set all the other bits of the PMD to the "empty" state. The invalid
> bit is already set prior to pmd_clear.

Okay, it makes some sense.

One more question: why does __tlb_flush_full()/__tlb_flush_asce() require
disabling preemption and pmdp_flush_direct() doesn't?

-- 
 Kirill A. Shutemov

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: TLB flushes on s390

[Martin Schwidefsky]

On Wed, 6 May 2015 15:50:00 +0300
"Kirill A. Shutemov" <kirill@shutemov.name> wrote:

> On Wed, May 06, 2015 at 02:00:02PM +0200, Martin Schwidefsky wrote:
> > On Wed, 6 May 2015 14:29:39 +0300
> > "Kirill A. Shutemov" <kirill@shutemov.name> wrote:
> > 
> > > I'm looking though s390 code around page table handling and I found that
> > > in many places s390 does tlb flush before changing page table entry.
> > 
> > Uhh, have fun with that.. it is complicated :-/
> >  
> > > Let's look for instance on pmdp_clear_flush() implementation on s390.
> > > It's implemented with pmdp_get_and_clear() which does pmdp_flush_direct()
> > > *before* pmd_clear(). That's invert order comparing to generic
> > > pmdp_flush_direct().
> > > 
> > > The question is what prevents tlb from being re-fill between flushing tlb
> > > and clearing page table entry?
> >  
> > Look again at pmdp_flush_direct(), either __pmdp_idte_local or __pmdp_idte is
> > called. Both functions use the IDTE instruction but in two different flavors.
> > The mnemonic IDTE stands for invalidate-dat-table-entry, the instruction sets
> > the invalid bit in the PMD and flushes all TLB entries on all CPUs that are
> > affected by the now invalid PMD. The pmd_clear after the pmdp_flush_direct is
> > done to set all the other bits of the PMD to the "empty" state. The invalid
> > bit is already set prior to pmd_clear.
> 
> Okay, it makes some sense.
> 
> One more question: why does __tlb_flush_full()/__tlb_flush_asce() require
> disabling preemption and pmdp_flush_direct() doesn't?
 
Easy: for pmdp_flush_direct you are required to hold the pmd-lock as you
modify the pmd. The spinlock provides the preempt_disable.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>