lkml.org 
[lkml]   [2015]   [Apr]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/4] mm: Send a single IPI to TLB flush multiple pages when unmapping
Hello Mel,

On Wed, Apr 15, 2015 at 10:28:55PM +0100, Mel Gorman wrote:
> On Wed, Apr 15, 2015 at 02:16:49PM -0700, Hugh Dickins wrote:
> > On Wed, 15 Apr 2015, Rik van Riel wrote:
> > > On 04/15/2015 06:42 AM, Mel Gorman wrote:
> > > > An IPI is sent to flush remote TLBs when a page is unmapped that was
> > > > recently accessed by other CPUs. There are many circumstances where this
> > > > happens but the obvious one is kswapd reclaiming pages belonging to a
> > > > running process as kswapd and the task are likely running on separate CPUs.
> > > >
> > > > On small machines, this is not a significant problem but as machine
> > > > gets larger with more cores and more memory, the cost of these IPIs can
> > > > be high. This patch uses a structure similar in principle to a pagevec
> > > > to collect a list of PFNs and CPUs that require flushing. It then sends
> > > > one IPI to flush the list of PFNs. A new TLB flush helper is required for
> > > > this and one is added for x86. Other architectures will need to decide if
> > > > batching like this is both safe and worth the memory overhead. Specifically
> > > > the requirement is;
> > > >
> > > > If a clean page is unmapped and not immediately flushed, the
> > > > architecture must guarantee that a write to that page from a CPU
> > > > with a cached TLB entry will trap a page fault.
> > > >
> > > > This is essentially what the kernel already depends on but the window is
> > > > much larger with this patch applied and is worth highlighting.
> > >
> > > This means we already have a (hard to hit?) data corruption
> > > issue in the kernel. We can lose data if we unmap a writable
> > > but not dirty pte from a file page, and the task writes before
> > > we flush the TLB.
> >
> > I don't think so. IIRC, when the CPU needs to set the dirty bit,
> > it doesn't just do that in its TLB entry, but has to fetch and update
> > the actual pte entry - and at that point discovers it's no longer
> > valid so traps, as Mel says.
> >
>
> This is what I'm expecting i.e. clean->dirty transition is write-through
> to the PTE which is now unmapped and it traps. I'm assuming there is an
> architectural guarantee that it happens but could not find an explicit
> statement in the docs. I'm hoping Dave or Andi can check with the relevant
> people on my behalf.

A dumb question. It's not related to your patch but MADV_FREE.

clean->dirty transition is *atomic* as well as write-through?
I'm really confusing.
It seems most arches use xchg for ptep_get_and_clear so it's
atomic but some of arches without defining __HAVE_ARCH_PTEP_GET_AND_CLEAR
will use non-atomic version in include/asm-generic/pgtable.h.

#ifndef __HAVE_ARCH_PTEP_GET_AND_CLEAR
static inline pte_t ptep_get_and_clear(struct mm_struct *mm,
unsigned long address,
pte_t *ptep)
{
pte_t pte = *ptep;
pte_clear(mm, address, ptep);
return pte;
}
#endif

I hope they have own lock or something to protect a race between software
and hardware(ie, CPU set dirty bit by itself).

Anyway, if there is a problem about that, we might see data corruption
but didn't so I guess it's atomic.
Otherwise, MADV_FREE will break, too.
I'd like to confirm that.

Thanks.

--
Kind regards,
Minchan Kim


\
 
 \ /
  Last update: 2015-04-16 09:21    [W:0.082 / U:3.264 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site