Re: kmemleaks on ac3b43283923 ("module: replace module_layout with module_memory")

[Catalin Marinas <catalin.marinas@arm.com>]

On Tue, Apr 11, 2023 at 04:00:59PM -0700, Luis Chamberlain wrote:
> On Tue, Apr 11, 2023 at 10:07:53AM -0700, Luis Chamberlain wrote:
> > On Tue, Apr 11, 2023 at 04:10:24PM +0100, Catalin Marinas wrote:
> > > On Mon, Apr 03, 2023 at 01:43:58PM -0700, Luis Chamberlain wrote:
> > > > On Fri, Mar 31, 2023 at 05:27:04PM -0700, Song Liu wrote:
> > > > > On Fri, Mar 31, 2023 at 12:00 AM Luis Chamberlain <mcgrof@kernel.org> wrote:
> > > > > > On Thu, Mar 30, 2023 at 04:45:43PM -0600, jim.cromie@gmail.com wrote:
> > > > > > > kmemleak is reporting 19 leaks during boot
> > > > > > >
> > > > > > > because the hexdumps appeared to have module-names,
> > > > > > > and Ive been hacking nearby, and see the same names
> > > > > > > every time I boot my test-vm, I needed a clearer picture
> > > > > > > Jason corroborated and bisected.
> > > > > > >
> > > > > > > the 19 leaks split into 2 groups,
> > > > > > > 9 with names of builtin modules in the hexdump,
> > > > > > > all with the same backtrace
> > > > > > > 9 without module-names (with a shared backtrace)
> > > > > > > +1 wo name-ish and a separate backtrace
> > > > > >
> > > > > > Song, please take a look.
> > > > > 
> > > > > I will look into this next week.
> > > > 
> > > > I'm thinking this may be it, at least this gets us to what we used to do
> > > > as per original Catalinas' 4f2294b6dc88d ("kmemleak: Add modules
> > > > support") and right before Song's patch.
> > > > 
> > > > diff --git a/kernel/module/main.c b/kernel/module/main.c
> > > > index 6b6da80f363f..3b9c71fa6096 100644
> > > > --- a/kernel/module/main.c
> > > > +++ b/kernel/module/main.c
> > > > @@ -2240,7 +2240,10 @@ static int move_module(struct module *mod, struct load_info *info)
> > > >  		 * which is inside the block. Just mark it as not being a
> > > >  		 * leak.
> > > >  		 */
> > > > -		kmemleak_ignore(ptr);
> > > > +		if (type == MOD_INIT_TEXT)
> > > > +			kmemleak_ignore(ptr);
> > > > +		else
> > > > +			kmemleak_not_leak(ptr);
> > > >  		if (!ptr) {
> > > >  			t = type;
> > > >  			goto out_enomem;
> > > > 
> > > > We used to use the grey area for the TEXT but the original commit
> > > > doesn't explain too well why we grey out init but not the others. Ie
> > > > why kmemleak_ignore() on init and kmemleak_not_leak() on the others.
> > > 
> > > It's safe to use the 'grey' colour in all cases. For text sections that
> > > don't need scanning, there's a slight chance of increasing the false
> > > negatives, 
> > 
> > It turns out that there are *tons* of false positives today, unless
> > these are real leaks.
> 
> I should clarify: *if* we leave things as-is, we seem to get tons of
> false positives.

Which makes sense if kmemleak_ignore() is used, such objects would not
be scanned. I'd just replace it with kmemleak_not_leak() irrespective of
the type.

-- 
Catalin