Page migration issue with UBIFS

[Kirill A. Shutemov]

On Tue, Mar 15, 2016 at 04:25:50PM +0100, Richard Weinberger wrote:
> Kirill,
> 
> Am 15.03.2016 um 16:17 schrieb Kirill A. Shutemov:
> > On Tue, Mar 15, 2016 at 03:16:11PM +0100, Richard Weinberger wrote:
> >> Hi!
> >>
> >> We're facing this issue from 2014 on UBIFS:
> >> http://www.spinics.net/lists/linux-fsdevel/msg79941.html
> >>
> >> So sum up:
> >> UBIFS does not allow pages directly marked as dirty. It want's everyone to do it via UBIFS's
> >> ->wirte_end() and ->page_mkwirte() functions.
> >> This assumption *seems* to be violated by CMA which migrates pages.
> > 
> > I don't thing the CMA/migration is the root cause.
> > 
> > How did we end up with writable and dirty pte, but not having
> > ->page_mkwrite() called for the page?
> > 
> > Or if ->page_mkwrite() was called, why the page is not dirty?
> 
> Thanks for your quick response!
> 
> I also don't think that the root cause is CMA or migration but it seems
> to be the messenger.
> 
> Can you confirm that UBIFS's assumptions are valid?
> I'm trying to rule out possible issues and hunt down the root cause...

The assumption looks reasonable for me, but I am not confident enough to
"confirm" it.

-- 
 Kirill A. Shutemov