Re: [PATCH] mm: convert three more cases to kvmalloc

[Michal Hocko]

On Thu 29-06-17 22:13:26, Mikulas Patocka wrote:
> 
> 
> On Thu, 29 Jun 2017, Michal Hocko wrote:
[...]
> > > Index: linux-2.6/kernel/bpf/syscall.c
> > > ===================================================================
> > > --- linux-2.6.orig/kernel/bpf/syscall.c
> > > +++ linux-2.6/kernel/bpf/syscall.c
> > > @@ -58,16 +58,7 @@ void *bpf_map_area_alloc(size_t size)
> > >  	 * trigger under memory pressure as we really just want to
> > >  	 * fail instead.
> > >  	 */
> > > -	const gfp_t flags = __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO;
> > > -	void *area;
> > > -
> > > -	if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
> > > -		area = kmalloc(size, GFP_USER | flags);
> > > -		if (area != NULL)
> > > -			return area;
> > > -	}
> > > -
> > > -	return __vmalloc(size, GFP_KERNEL | flags, PAGE_KERNEL);
> > > +	return kvmalloc(size, GFP_USER | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO);
> > 
> > kvzalloc without additional flags would be more appropriate.
> > __GFP_NORETRY is explicitly documented as non-supported
> 
> How is __GFP_NORETRY non-supported?

Because its semantic cannot be guaranteed throughout the alloaction
stack. vmalloc will ignore it e.g. for page table allocations.

> > and NOWARN wouldn't be applied everywhere in the vmalloc path.
> 
> __GFP_NORETRY and __GFP_NOWARN wouldn't be applied in the page-table 
> allocation and they would be applied in the page allocation - that seems 
> acceptable.

This is rather muddy semantic to me. Both page table and the page is an
order-0 allocation. Page table allocations are much less likely but I've
explicitly documented that explicit __GFP_NORETRY is unsupported. Slab
allocation is already __GFP_NORETRY (unless you specify
__GFP_RETRY_MAYFAIL in the current mmotm tree).
-- 
Michal Hocko
SUSE Labs