Re: [RFC PATCH] mm/vmalloc: make sure to traverse from the beginning when overflow occur

Re: [RFC PATCH] mm/vmalloc: make sure to traverse from the beginning when overflow occur

[Dave Hansen]

On 4/12/20 7:38 AM, Liu Song wrote:
> If overflow, should ensure that "free_vmap_cache" is set to NULL,
> so as to ensure that it can be traversed from the beginning.

This changelog is a bit sparse.

Does this fix a demonstrated problem?  Or was it just discovered via
code review and assumed to help a theoretical problem?

Which tree is it against?  'free_vmap_cache' doesn't show up in Linus's
tree.
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index d8e877365f9f..2638a20d36ce 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -441,7 +441,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
>  	if (!free_vmap_cache ||
>  			size < cached_hole_size ||
>  			vstart < cached_vstart ||
> -			align < cached_align) {
> +			align < cached_align || purged) {
>  nocache:
>  		cached_hole_size = 0;
>  		free_vmap_cache = NULL;

Is 'purged' the right way to do this?

There is a path through the overflow code that also sets purged=0.  Does
'free_vmap_cache' *not* need to get reset in that case?

It also seems more natural to me that the code mucking with the rbtree
would be the one to reset the cache.

Re: [RFC PATCH] mm/vmalloc: make sure to traverse from the beginning when overflow occur

[Uladzislau Rezki]

On Mon, Apr 13, 2020 at 08:02:49AM -0700, Dave Hansen wrote:
> On 4/12/20 7:38 AM, Liu Song wrote:
> > If overflow, should ensure that "free_vmap_cache" is set to NULL,
> > so as to ensure that it can be traversed from the beginning.
> 
> This changelog is a bit sparse.
> 
> Does this fix a demonstrated problem?  Or was it just discovered via
> code review and assumed to help a theoretical problem?
> 
> Which tree is it against?  'free_vmap_cache' doesn't show up in Linus's
> tree.
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index d8e877365f9f..2638a20d36ce 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -441,7 +441,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
> >  	if (!free_vmap_cache ||
> >  			size < cached_hole_size ||
> >  			vstart < cached_vstart ||
> > -			align < cached_align) {
> > +			align < cached_align || purged) {
> >  nocache:
> >  		cached_hole_size = 0;
> >  		free_vmap_cache = NULL;
> >
We do not have such code since 5.2 kernel or so, nor things like
cached_hole_size, free_vmap_cache and so on. It was totally reworked.

--
Vlad Rezki