Re: [PATCH] slub: Proper kmemleak tracking if CONFIG_SLUB_DEBUGdisabled

[Catalin Marinas]

On Wed, Oct 02, 2013 at 04:33:47PM +0100, Bird, Tim wrote:
> On Wednesday, October 02, 2013 7:41 AM, Christoph Lameter [cl@xxxxxxxxx] wrote:
> >
> >On Fri, 27 Sep 2013, Frank Rowand wrote:
> >
> >> Move the kmemleak code for small block allocation out from
> >> under CONFIG_SLUB_DEBUG.
> >
> >Well in that case it may be better to move the hooks as a whole out of
> >the CONFIG_SLUB_DEBUG section. Do the #ifdeffering for each call from the
> >hooks instead.
> >
> >The point of the hook functions is to separate the hooks out of the
> >functions so taht they do not accumulate in the main code.
> >
> >The patch moves one hook back into the main code. Please keep the checks
> >in the hooks.
> 
> Thanks for the feedback.  Roman's first patch, which we discussed internally
> before sending this one, did exactly that.  I guess Roman gets to say "I told
> you so." :-)  My bad for telling him to change it.
> 
> We'll refactor along the lines that you describe, and send another one.
> 
> The problem child is actually the unconditional call to kmemleak_alloc()
> in kmalloc_large_node() (in slub.c).  The problem comes because that call
> is unconditional on CONFIG_SLUB_DEBUG but the kmemleak
> calls in the hook routines are conditional on CONFIG_SLUB_DEBUG.
> So if you have CONFIG_SLUB_DEBUG=n but CONFIG_DEBUG_KMEMLEAK=y,
> you get the false reports.
> 
> Now, there are kmemleak calls in kmalloc_large_node() and kfree() that don't
> follow the "hook" pattern.  Should these be moved to 'hook' routines, to keep
> all the checks in the hooks?
> 
> Personally, I like the idea of keeping bookeeping/tracing/debug stuff in hook
> routines.  I also like de-coupling CONFIG_SLUB_DEBUG and CONFIG_DEBUG_KMEMLEAK,
> but maybe others have a different opinon.  Unless someone speaks up, we'll
> move the the currently in-function kmemleak calls into hooks, and all of the
> kmemleak stuff out from under CONFIG_SLUB_DEBUG.
> We'll have to see if the ifdefs get a little messy.

Kmemleak doesn't depend on SLUB_DEBUG (at least it didn't originally ;),
so I don't think we should add an artificial dependency (or select). Can
we have kmemleak_*() calls in both debug and !debug hooks?

-- 
Catalin
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/