Re: grab_cache_page deadlock | was Re: set_buffer_dirty_uptodate

[Rajagopal Ananthanarayanan <ananth@xxxxxxx>]

Marcelo Tosatti wrote:
> 

> 
> I think the only clean solution for this case is to not hold any FS lock
> while we're before calling copy_from_user().

Copy from/to user is only one situation. There
are other cases of kmalloc while holding some FS lock.
It will be nearly impossible to weed out those cases from XFS.

IMHO, it is too much of a constraint on the FS to exepect
not to hold any lock while allocating memory. Then there is
that GFP_IO comment in shrink_dcache memory which looks inelegant.
kswapd and/or the memory allocator looks like it can use some improvement.

ananth.