The gfp_mask that is passed to __page_symlink() is being completely
dropped on the floor. Historically this mask was at least used by
ext3 and ext4 to avoid recursing back into the FS from within a
journal transaction; Kirill fixed that issue with this commit:
0adb25d2e71ab047423d6fc63d5d184590d0a66f
I'm quite naive when it comes to Nick's relatively new (>= 2.6.24) AOP
pagecache_write_{begin,end} code that motivated __page_symlink to
change with this commit:
afddba49d18f346e5cc2938b6ed7c512db18ca68
Nick's change clearly did away with using the explicitly passed
gfp_mask in __page_symlink().
So at a minimum it would seem __page_symlink() now has an unused
parameter that should be removed.
But a more serious concern would be: have ext[34]_symlink() regressed
to being susceptible to the bug that Kirill fixed some time ago?
Please advise, thanks.
Mike
On Tue, Oct 28, 2008 at 08:11:48PM -0400, Mike Snitzer wrote:
> The gfp_mask that is passed to __page_symlink() is being completely
> dropped on the floor. Historically this mask was at least used by
> ext3 and ext4 to avoid recursing back into the FS from within a
> journal transaction; Kirill fixed that issue with this commit:
> 0adb25d2e71ab047423d6fc63d5d184590d0a66f
>
> I'm quite naive when it comes to Nick's relatively new (>= 2.6.24) AOP
> pagecache_write_{begin,end} code that motivated __page_symlink to
> change with this commit:
> afddba49d18f346e5cc2938b6ed7c512db18ca68
>
> Nick's change clearly did away with using the explicitly passed
> gfp_mask in __page_symlink().
> So at a minimum it would seem __page_symlink() now has an unused
> parameter that should be removed.
>
> But a more serious concern would be: have ext[34]_symlink() regressed
> to being susceptible to the bug that Kirill fixed some time ago?
Yeah, I think this would be a potential problem for ext3/4. Looks
like pagemap_write_begin() should take a gfp_mask argument, and then
pass it down through to __grab_cache_page(), which should then call
__page_cache_alloc() instead of _page_cache_alloc(). Then
__page_symlink() can actually pass in its gfp_mask to
pagemap_write_begin().
Nick, do you agree?
- Ted
On Tue, Oct 28, 2008 at 10:40:48PM -0400, Theodore Tso wrote:
> On Tue, Oct 28, 2008 at 08:11:48PM -0400, Mike Snitzer wrote:
> > The gfp_mask that is passed to __page_symlink() is being completely
> > dropped on the floor. Historically this mask was at least used by
> > ext3 and ext4 to avoid recursing back into the FS from within a
> > journal transaction; Kirill fixed that issue with this commit:
> > 0adb25d2e71ab047423d6fc63d5d184590d0a66f
> >
> > I'm quite naive when it comes to Nick's relatively new (>= 2.6.24) AOP
> > pagecache_write_{begin,end} code that motivated __page_symlink to
> > change with this commit:
> > afddba49d18f346e5cc2938b6ed7c512db18ca68
> >
> > Nick's change clearly did away with using the explicitly passed
> > gfp_mask in __page_symlink().
> > So at a minimum it would seem __page_symlink() now has an unused
> > parameter that should be removed.
> >
> > But a more serious concern would be: have ext[34]_symlink() regressed
> > to being susceptible to the bug that Kirill fixed some time ago?
>
> Yeah, I think this would be a potential problem for ext3/4. Looks
> like pagemap_write_begin() should take a gfp_mask argument, and then
> pass it down through to __grab_cache_page(), which should then call
> __page_cache_alloc() instead of _page_cache_alloc(). Then
> __page_symlink() can actually pass in its gfp_mask to
> pagemap_write_begin().
>
> Nick, do you agree?
I agree it is a problem. It's a bit hard to pass down a gfp_mask
(because the caller would normally expect _all_ operations in the
called code to obey the mask, basically impossible to do for
GFP_NOFS because by definition we're calling into ->write_begin).
I was leaning towards adding a new AOP_FLAG_ there, usable just by
filesystem code, and just to tell any helper code to clear __GFP_FS.
That way callers won't get confused into thinking they can do
GFP_ATOMIC writes from interrupt context or something ;) (which,
trust me, somebody will attempt to do if it looks remotely feasible!)
On Wed, Oct 29, 2008 at 04:25:57AM +0100, Nick Piggin wrote:
>
> I was leaning towards adding a new AOP_FLAG_ there, usable just by
> filesystem code, and just to tell any helper code to clear __GFP_FS.
> That way callers won't get confused into thinking they can do
> GFP_ATOMIC writes from interrupt context or something ;) (which,
> trust me, somebody will attempt to do if it looks remotely feasible!)
Good point!!
- Ted