Re: [PATCH v4] kmemleak: survive in a low-memory situation

[Catalin Marinas]

On Wed, Mar 27, 2019 at 09:44:32AM +0100, Michal Hocko wrote:
> On Tue 26-03-19 20:59:48, Qian Cai wrote:
> [...]
> > Unless there is a brave soul to reimplement the kmemleak to embed it's
> > metadata into the tracked memory itself in a foreseeable future,

Revisiting the kmemleak memory scanning code, that's not actually
possible without some long periods with kmemleak_lock held. The scanner
relies on the kmemleak_object (refcounted) being around even when the
actual memory block has been freed.

> > this
> > provides a good balance between enabling kmemleak in a low-memory
> > situation and not introducing too much hackiness into the existing
> > code for now. Another approach is to fail back the original allocation
> > once kmemleak_alloc() failed, but there are too many call sites to
> > deal with which makes it error-prone.
> 
> As long as there is an implicit __GFP_NOFAIL then kmemleak is simply
> broken no matter what other gfp flags you play with. Has anybody looked
> at some sort of preallocation where gfpflags_allow_blocking context
> allocate objects into a pool that non-sleeping allocations can eat from?

Quick attempt below and it needs some more testing (pretty random pick
of the EMERGENCY_POOL_SIZE value). Also, with __GFP_NOFAIL removed, are
the other flags safe or we should trim them further?

---------------8<-------------------------------