Re: [patch 3/3 -mmotm] oom: invoke oom killer for __GFP_NOFAIL

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 3 Jun 2009 15:10:38 -0700 (PDT)
David Rientjes <rientjes@google.com> wrote:

> On Tue, 2 Jun 2009, David Rientjes wrote:
> 
> > With my patch, we kill a memory hogging task that will free some memory so 
> > the allocation will succeed (or multiple tasks if insufficient contiguous 
> > memory is available).  Kernel allocations use __GFP_NOFAIL, so the fault 
> > of this memory freeing is entirely on the caller, not the page allocator.
> > 
> > My preference for handling this is to merge my patch (obviously :), and 
> > then hopefully deprecate __GFP_NOFAIL as much as possible although I don't 
> > suspect it could be eradicated forever.
> > 
> 
> I really hope this patch isn't getting dropped because it fixes the 
> possibility that a __GFP_NOFAIL allocation will fail when its definition 
> is to the contrary.  Depending on the size of the allocation, that can 
> cause a panic in at least the reiserfs, ntfs, cxgb3, and gfs2 cases.
> 
> As I mentioned before, it's a noble goal to deprecate __GFP_NOFAIL as much 
> as possible and (at the least) prevent it from trying high-order 
> allocation attempts.  The current implementation of the flag is 
> problematic, however, and this patch addresses it by attempting to free 
> some memory when direct reclaim fails.
> 

Sigh, all right, but we suck.

Divy, could we please at least remove __GFP_NOFAIL from
drivers/net/cxgb?  It's really quite inappropriate for a driver to
assume that core VM can do magic.  Drivers should test the return value
and handle the -ENOMEM in the old-fashioned way, please.

Ditto-in-spades cfq-iosched.c.  We discussed that recently but I forgot the
upshot.  The code and its comment are still in flagrant disagreement?