Re: [patch v2 1/5] mm: add nofail variants of kmalloc kcalloc and kzalloc

[Neil Brown <neilb@suse.de>]

On Thu, 2 Sep 2010 16:51:41 +0200
Jan Kara <jack@suse.cz> wrote:

> On Thu 02-09-10 09:59:13, Jiri Slaby wrote:
> > On 09/02/2010 03:02 AM, David Rientjes wrote:
> > > --- a/include/linux/slab.h +++ b/include/linux/slab.h @@ -334,6 +334,57
> > > @@ static inline void *kzalloc_node(size_t size, gfp_t flags, int node)
> > > return kmalloc_node(size, flags | __GFP_ZERO, node); }
> > >  
> > > +/** + * kmalloc_nofail - infinitely loop until kmalloc() succeeds.  +
> > > * @size: how many bytes of memory are required.  + * @flags: the type
> > > of memory to allocate (see kmalloc).  + * + * NOTE: no new callers of
> > > this function should be implemented!  + * All memory allocations should
> > > be failable whenever possible.  + */ +static inline void
> > > *kmalloc_nofail(size_t size, gfp_t flags) +{ +	void *ret; + +	for
> > > (;;) { +		ret = kmalloc(size, flags); +		if (ret) +
> > > return ret; +		WARN_ON_ONCE(get_order(size) >
> > > PAGE_ALLOC_COSTLY_ORDER);
> > 
> > This doesn't work as you expect. kmalloc will warn every time it fails.
> > __GFP_NOFAIL used to disable the warning. Actually what's wrong with
> > __GFP_NOFAIL? I cannot find a reason in the changelogs why the patches
> > are needed.
>   David should probably add the reasoning to the changelogs so that he
> doesn't have to explain again and again ;). But if I understood it
> correctly, the concern is that the looping checks slightly impact fast path
> of the callers which do not need it. Generally, also looping for a long
> time inside allocator isn't a nice thing but some callers aren't able to do
> better for now to the patch is imperfect in this sence...
>

I'm actually a bit confused about this too.
I thought David said he was removing a branch on the *slow* path - which make
sense as you wouldn't even test NOFAIL until you had a failure.
Why are branches on the slow-path an issue??
This is an important question to me because I still hope to see the
swap-over-nfs patches merged eventually and they add a branch on the slow
path (if I remember correctly).

NeilBrown