Re: Race between vmtruncate and mapped areas?

[Andrea Arcangeli <andrea@suse.de>]

On Wed, May 14, 2003 at 11:53:19AM -0700, Andrew Morton wrote:
> converting i_sem to an rwsem and taking it in the pagefault would certainly
> stitch it up.  Unpopular, very messy.

and very slow, down_read on every page fault wouldn't scale

> Could "truncate file" return some code to say pages were left behind, so
> truncate re-runs zap_page_range()?  Sounds unpleasant.
> 
> 
> Yes, re-checking the page against i_size from do_no_page() would fix it up.

think if there are two truncates, one zapping the entere file, and
another restoring the previous i_size, in such case the new_page will be
wrong, as it won't be zeroed out. I mean if we do anything about it, we
should close all races and make it 100% correct.

My fix has no scalability cost, no indirect calls, touches mostly just
hot cachelines anyways, and addresses the multiple truncate case too.


>  But damn, that's another indirect call, 64-bit math, etc on _every_
> file-backed pagefault.
> 
> 
> Remind me again what problem this whole thing is currently causing?

the only thing I can imagine is an app trapping SIGBUS to garbage
collect the end of a file. So for example you truncate the file and you
wait the SIGBUS to know you've to re-extend it. it would be a legitimate
use, and this is a bug, it's not read against write that has no way to
synchronize anyways, however I doubt any application is being bitten by
this race.

Andrea