On Thu 03-01-19 11:51:57, Qian Cai wrote: > On 1/3/19 4:32 AM, Michal Hocko wrote: > > On Wed 02-01-19 13:06:19, Qian Cai wrote: > > [...] > >> diff --git a/mm/kmemleak.c b/mm/kmemleak.c > >> index f9d9dc250428..9e1aa3b7df75 100644 > >> --- a/mm/kmemleak.c > >> +++ b/mm/kmemleak.c > >> @@ -576,6 +576,16 @@ static struct kmemleak_object *create_object(unsigned > >> long ptr, size_t size, > >> struct rb_node **link, *rb_parent; > >> > >> object = kmem_cache_alloc(object_cache, gfp_kmemleak_mask(gfp)); > >> +#ifdef CONFIG_PREEMPT_COUNT > >> + if (!object) { > >> + /* last-ditch effort in a low-memory situation */ > >> + if (irqs_disabled() || is_idle_task(current) || in_atomic()) > >> + gfp = GFP_ATOMIC; > >> + else > >> + gfp = gfp_kmemleak_mask(gfp) | __GFP_DIRECT_RECLAIM; > >> + object = kmem_cache_alloc(object_cache, gfp); > >> + } > >> +#endif > > > > I do not get it. How can this possibly help when gfp_kmemleak_mask() > > adds __GFP_NOFAIL modifier to the given gfp mask? Or is this not the > > case anymore in some tree? > > Well, __GFP_NOFAIL can still fail easily without __GFP_DIRECT_RECLAIM in a > low-memory situation.
OK, I guess I understand now. So the issue is that a (general) atomic allocation will provide its gfp mask down to kmemleak and you are trying/hoping that if the allocation is no from an atomic context then you can fortify it by using a sleepable allocation for the kmemleak metadata or giving it access to memory reserves for atomic allocations. I think this is still fragile because most atomic allocations are for a good reason. As I've said earlier the current implementation which abuses __GFP_NOFAIL is fra from great and we have discussed some alternatives. Not sure whan came out of it. I will not object to this workaround but I strongly believe that kmemleak should rethink the metadata allocation strategy to be really robust. -- Michal Hocko SUSE Labs