Re: [RFC 2/3] mm/vmalloc: add support for __GFP_NOFAIL

[Uladzislau Rezki <urezki@gmail.com>]

On Tue, Oct 19, 2021 at 01:52:07PM +0200, Michal Hocko wrote:
> On Tue 19-10-21 13:06:49, Uladzislau Rezki wrote:
> > > From: Michal Hocko <mhocko@suse.com>
> > > 
> > > Dave Chinner has mentioned that some of the xfs code would benefit from
> > > kvmalloc support for __GFP_NOFAIL because they have allocations that
> > > cannot fail and they do not fit into a single page.
> > > 
> > > The larg part of the vmalloc implementation already complies with the
> > > given gfp flags so there is no work for those to be done. The area
> > > and page table allocations are an exception to that. Implement a retry
> > > loop for those.
> > > 
> > > Signed-off-by: Michal Hocko <mhocko@suse.com>
> > > ---
> > >  mm/vmalloc.c | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > index 7455c89598d3..3a5a178295d1 100644
> > > --- a/mm/vmalloc.c
> > > +++ b/mm/vmalloc.c
> > > @@ -2941,8 +2941,10 @@ static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
> > >  	else if (!(gfp_mask & (__GFP_FS | __GFP_IO)))
> > >  		flags = memalloc_noio_save();
> > >  
> > > -	ret = vmap_pages_range(addr, addr + size, prot, area->pages,
> > > +	do {
> > > +		ret = vmap_pages_range(addr, addr + size, prot, area->pages,
> > >  			page_shift);
> > > +	} while ((gfp_mask & __GFP_NOFAIL) && (ret < 0));
> > >  
> > >  	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
> > >  		memalloc_nofs_restore(flags);
> > > @@ -3032,6 +3034,8 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
> > >  		warn_alloc(gfp_mask, NULL,
> > >  			"vmalloc error: size %lu, vm_struct allocation failed",
> > >  			real_size);
> > > +		if (gfp_mask && __GFP_NOFAIL)
> > > +			goto again;
> > >  		goto fail;
> > >  	}
> > >  
> > > -- 
> > > 2.30.2
> > > 
> > I have checked the vmap code how it aligns with the __GFP_NOFAIL flag.
> > To me it looks correct from functional point of view.
> > 
> > There is one place though it is kasan_populate_vmalloc_pte(). It does
> > not use gfp_mask, instead it directly deals with GFP_KERNEL for its
> > internal purpose. If it fails the code will end up in loping in the
> > __vmalloc_node_range().
> > 
> > I am not sure how it is important to pass __GFP_NOFAIL into KASAN code.
> > 
> > Any thoughts about it?
> 
> The flag itself is not really necessary down there as long as we
> guarantee that the high level logic doesn't fail. In this case we keep
> retrying at __vmalloc_node_range level which should be possible to cover
> all callers that can control gfp mask. I was thinking to put it into
> __get_vm_area_node but that was slightly more hairy and we would be
> losing the warning which might turn out being helpful in cases where the
> failure is due to lack of vmalloc space or similar constrain. Btw. do we
> want some throttling on a retry?
> 
I think adding kind of schedule() will not make things worse and in corner
cases could prevent a power drain by CPU. It is important for mobile devices. 

As for vmap space, it can be that a user specifies a short range that does
not contain any free area. In that case we might never return back to a caller.

Maybe add a good comment something like: think what you do when deal with the
__vmalloc_node_range() and __GFP_NOFAIL?

--
Vlad Rezki