Re: [PATCH] reiserfs: use __GFP_NOFAIL instead of yield and retry loop for allocation

[Andrew Morton <akpm@osdl.org>]

Pekka J Enberg <penberg@cs.Helsinki.FI> wrote:
>
> Hi,
> 
> Pekka J Enberg <penberg@cs.Helsinki.FI> wrote:
> > >
> > >  -      retry:
> > >  -	jl = kzalloc(sizeof(struct reiserfs_journal_list), GFP_NOFS);
> > >  -	if (!jl) {
> > >  -		yield();
> > >  -		goto retry;
> > >  -	}
> > >  +	jl = kzalloc(sizeof(struct reiserfs_journal_list),
> > >  +		     GFP_NOFS | __GFP_NOFAIL);
> 
> On Thu, 12 Jan 2006, Andrew Morton wrote:
> > yup, that's what __GFP_NOFAIL is for: to consolidate and identify all those
> > places which want to lock up when we're short of memory...  They all need
> > fixing, really.
> 
> Out of curiosity, are there any potential problems with combining GFP_NOFS 
> and __GFP_NOFAIL? Can we really guarantee to give out memory if we're not 
> allowed to page out?
> 

GFP_NOFS increases the risk (relative to GFP_KERNEL) because page reclaim
can do less things than GFP_KERNEL to free memory.

GFP_NOFS allocations can still perform swapspace writes, however.  GFP_NOIO
cannot even do that.