On Mon, 2018-10-29 at 09:17 +0100, Michal Hocko wrote: > On Mon 29-10-18 09:07:08, Michal Hocko wrote: > [...] > > Besides that, the following doesn't make much sense to me. It simply > > makes no sense to use vmalloc for sub page allocation regardless of > > HIGHMEM. > > OK, it is still early morning here. Now I get the point of the patch. > You just want to (ab)use highmeme for smaller requests. I do not like > this, to be honest. It causes an internal fragmentation and more > importantly the VMALLOC space on 32b where HIGHMEM is enabled (do we > have any 64b with HIGHMEM btw?) is quite small to be wasted like that. > thanks for your comment. It looks like that using vmalloc fallback for sub page allocation is not good here. Your comment gave another idea: 1. force kbuf to PAGE_SIZE 2. allocate a page by alloc_page(GFP_KERNEL | __GFP_HIGHMEM); so we can get a highmem page if possible 3. use kmap/kunmap pair to create mapping for this page. No vmalloc space is used. 4. do not change kvmalloc logic. > In any case such a changes should come with some numbers and as a > separate patch for sure. > > > > diff --git a/mm/util.c b/mm/util.c > > > index 8bf08b5b5760..7b1c59b9bfbf 100644 > > > --- a/mm/util.c > > > +++ b/mm/util.c > > > @@ -416,10 +416,10 @@ void *kvmalloc_node(size_t size, gfp_t flags, int node) > > > ret = kmalloc_node(size, kmalloc_flags, node); > > > > > > /* > > > - * It doesn't really make sense to fallback to vmalloc for sub page > > > - * requests > > > + * It only makes sense to fallback to vmalloc for sub page > > > + * requests if we might be able to allocate highmem pages. > > > */ > > > - if (ret || size <= PAGE_SIZE) > > > + if (ret || (!IS_ENABLED(CONFIG_HIGHMEM) && size <= PAGE_SIZE)) > > > return ret; > > > > > > return __vmalloc_node_flags_caller(size, node, flags, > > > -- > > > 2.18.0 > > > > > > > -- > > Michal Hocko > > SUSE Labs >
