On Tue, Oct 26, 2021 at 12:24 PM NeilBrown <neilb@xxxxxxx> wrote: > > On Tue, 26 Oct 2021, Michal Hocko wrote: > > On Tue 26-10-21 10:50:21, Neil Brown wrote: > > > On Mon, 25 Oct 2021, Uladzislau Rezki wrote: > > > > On Fri, Oct 22, 2021 at 09:49:08AM +1100, NeilBrown wrote: > > > > > However I'm not 100% certain, and the behaviour might change in the > > > > > future. So having one place (the definition of memalloc_retry_wait()) > > > > > where we can change the sleeping behaviour if the alloc_page behavour > > > > > changes, would be ideal. Maybe memalloc_retry_wait() could take a > > > > > gfpflags arg. > > > > > > > > > At sleeping is required for __get_vm_area_node() because in case of lack > > > > of vmap space it will end up in tight loop without sleeping what is > > > > really bad. > > > > > > > So vmalloc() has two failure modes. alloc_page() failure and > > > __alloc_vmap_area() failure. The caller cannot tell which... > > > > > > Actually, they can. If we pass __GFP_NOFAIL to vmalloc(), and it fails, > > > then it must have been __alloc_vmap_area() which failed. > > > What do we do in that case? > > > Can we add a waitq which gets a wakeup when __purge_vmap_area_lazy() > > > finishes? > > > If we use the spinlock from that waitq in place of free_vmap_area_lock, > > > then the wakeup would be nearly free if no-one was waiting, and worth > > > while if someone was waiting. > > > > Is this really required to be part of the initial support? > > No.... I was just thinking out-loud. > alloc_vmap_area() has an retry path, basically if it fails the code will try to "purge" areas and repeat it one more time. So we do not need to purge outside some where else. -- Uladzislau Rezki
