Re: [PATCH] Fix vmtruncate race and distributed filesystem race

[Andrew Morton]

"Paul E. McKenney" <[e-mail]> wrote:
>
> > but you can't trap this with a single counter increment in do_truncate:
>  > 
>  > 	CPU 0			CPU 1
>  > 	----------		-----------
>  > 				do_no_page
>  > 	truncate

        i_size = new_i_size;

>  > 	increment counter
>  > 				read counter
>  > 				->nopage

                                check i_size

>  > 	vmtruncate
>  > 				read counter again -> different so retry
>  > 
>  > thanks to the second counter increment after vmtruncate in my fix, the
>  > above race couldn't happen.
> 
>  The trick is that CPU 0 is expected to have updated the filesystem's
>  idea of what pages are available before calling vmtruncate,
>  invalidate_mmap_range() or whichever.

i_size has been updated, and filemap_nopage() will return NULL.