On Fri, Jul 19, 2024 at 7:02 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Fri 19-07-24 12:35:55, Barry Song wrote: > > On Thu, Jul 18, 2024 at 8:50 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > [...] > > > Yes, those shouldn't really fail. NOWAIT|NOFAIL was something that > > > should never happen and I really hope it doesn't. Others should really > > > retry but it's been some time since I've checked the last time. > > > > > > I assume allocations directly using alloc_pages() might not respect GFP_NOFAIL > > and violate the semantics of GFP_NOFAIL. > > What do you mean? I mean, if we are using wrappers like vmalloc (GFP_NOFAIL | GFP_NOWAIT), though alloc_pages might return NULL, vmalloc for itself will retry. but if you call alloc_pages() directly with GFP_NOFAIL | GFP_NOWAIT, alloc_pages() may return NULL without retry at all. I believe alloc_pages() is also wrong. > > > Additionally, at least drivers/vdpa/vdpa_user/iova_domain.c is > > incorrect with GFP_ATOMIC > > | __GFP_NOFAIL. > > This is broken! Please bring this up with the maintainer of the code. > > [...] > > > GFP_NOFAIL things need to be fixed. Let me investigate further. > > Thanks! > > -- > Michal Hocko > SUSE Labs
