Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

[Catalin Marinas]

On Fri, Aug 11, 2023 at 10:09:08AM +0200, Vlastimil Babka wrote:
> On 8/11/23 04:03, wang xiaolei wrote:
> > On 8/10/23 9:16 PM, Vlastimil Babka wrote:
> >> Looking closer, I think what you want could be achieved by kmemleak_init()
> >> setting a variable that is checked in kmemleak_initialized() instead of the
> >> kmemleak_initialized that's set too late.
> >>
> >> I think this should work because:
> >> - I assume kmemleak can't record anything before kmemleak_init()
> >> - stack depot early init is requested one way or the other
> >> - mm_core_init() calls stack_depot_early_init() before kmemleak_init()
> >>
> >> But I also wonder how kmemleak can even reach set_track_prepare() before
> >> kmemleak_init(), maybe that's the issue?
> > 
> > Before kmemleak_init, many places also need to allocate kmemleak_object,
> > 
> > and also need to save stack in advance, but kmemleak_object is allocated
> > 
> > in the form of an array, after kmemleak_init 'object_cache = 
> > KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);'
> 
> Hm I see, kmemleak has this static mempool so it really can record some
> objects very early.

Indeed, otherwise we'd get a lot of false positives.

> > I think there is still some memory not recorded on the backtrace before
> > 
> > stack_depot_early_init(), does anyone have a better suggestion?
> 
> No we can't record the backtrace earlier. But I don't think it's a problem
> in practice. AFAIU kmemleak needs to record these very early allocations so
> if they point to further objects, those are not suspected as orphans. But
> the early allocations themselves also are very unlikely to be leaks, so does
> it really matter that we don't have a backtrace for their allocation?
> Because the backtrace is the only thing that's missing - the object is
> otherwise recorded even if set_track_prepare() returns 0.

It's not a functional problem, just a reporting one. There are
rare early leaks (usually false positives) so identifying them would
help. That said, I think set_track_prepare() is too conservative in
waiting for kmemleak_initialized to be set in kmemleak_late_init().
That's a late_initcall() meant for the scanning thread etc. not the core
kmemleak functionality (which is on from early boot).

We can instead use a different variable to check in set_track_prepare(),
e.g. object_cache. stack_depot_early_init() is called prior to
kmemleak_init(), so it should be fine.

If "kmemleak_initialized" is confusing, we could rename it to
"kmemleak_late_initialized" or "kmemleak_fully_initialized". I'm not too
fussed about this as long as we add some comment on why we check
object_cache instead of kmemleak_initialized.

-- 
Catalin