----- Original Message ----- > From: "Catalin Marinas" <catalin.marinas@xxxxxxx> > To: "Chunyu Hu" <chuhu@xxxxxxxxxx> > Cc: "Michal Hocko" <mhocko@xxxxxxxxxx>, "Chunyu Hu" <chuhu.ncepu@xxxxxxxxx>, "Dmitry Vyukov" <dvyukov@xxxxxxxxxx>, > "LKML" <linux-kernel@xxxxxxxxxxxxxxx>, "Linux-MM" <linux-mm@xxxxxxxxx> > Sent: Thursday, April 26, 2018 8:56:35 PM > Subject: Re: [RFC] mm: kmemleak: replace __GFP_NOFAIL to GFP_NOWAIT in gfp_kmemleak_mask > > On Thu, Apr 26, 2018 at 08:23:19AM -0400, Chunyu Hu wrote: > > kmemleak is using kmem_cache to record every pointers returned from kernel > > mem > > allocation activities such as kmem_cache_alloc(). every time an object from > > slab allocator is returned, a following new kmemleak object is allocated. > > > > And when a slab object is freed, then the kmemleak object which contains > > the ptr will also be freed. > > > > and kmemleak scan thread will run in period to scan the kernel data, stack, > > and per cpu areas to check that every pointers recorded by kmemleak has at > > least > > one reference in those areas beside the one recorded by kmemleak. If there > > is no place in the memory acreas recording the ptr, then it's possible a > > leak. > > > > so once a kmemleak object allocation failed, it has to disable itself, > > otherwise > > it would lose track of some object pointers, and become less meaningful to > > continue record and scan the kernel memory for the pointers. So disable > > it forever. so this is why kmemleak can't tolerate a slab alloc fail (from > > fault injection) > > > > @Catalin, > > > > Is this right? If something not so correct or precise, please correct me. > > That's a good description, thanks. > > > I'm thinking about, is it possible that make kmemleak don't disable itself > > when fail_page_alloc is enabled? I can't think clearly what would happen > > if several memory allocation missed by kmelkeak trace, what's the bad > > result? > > Take for example a long linked list. If kmemleak doesn't track an object > in such list (because the metadata allocation failed), such list_head is > never scanned and the subsequent objects in the list (pointed at by > 'next') will be reported as leaks. Kmemleak pretty much becomes unusable > with a high number of false positives. Thanks for the example, one object may contain many pointers, so loose one, means many false reports. I'm clear now. > > -- > Catalin > -- Regards, Chunyu Hu
