Messages in this thread | | | Date | Wed, 1 Oct 2014 11:39:30 +0100 | From | Catalin Marinas <> | Subject | Re: [PATCH v3 11/13] kmemleak: disable kasan instrumentation for kmemleak |
| |
On Mon, Sep 29, 2014 at 03:10:01PM +0100, Dmitry Vyukov wrote: > On Fri, Sep 26, 2014 at 9:36 PM, Andrey Ryabinin <ryabinin.a.a@gmail.com> wrote: > > 2014-09-26 21:10 GMT+04:00 Dmitry Vyukov <dvyukov@google.com>: > >> Looks good to me. > >> > >> We can disable kasan instrumentation of this file as well. > > > > Yes, but why? I don't think we need that. > > Just gut feeling. Such tools usually don't play well together. For > example, due to asan quarantine lots of leaks will be missed (if we > pretend that tools work together, end users will use them together and > miss bugs). I won't be surprised if leak detector touches freed > objects under some circumstances as well. > We can do this if/when discover actual compatibility issues, of course.
I think it's worth testing them together first.
One issue, as mentioned in the patch log, is that the size information that kmemleak gets is the one from the kmem_cache object rather than the original allocation size, so this would be rounded up.
Kmemleak should not touch freed objects (if an object is freed during a scan, it is protected by some lock until the scan completes). There is a bug however which I haven't got to fixing it yet, if kmemleak fails for some reason (cannot allocate memory) and disables itself, it may access some freed object (though usually hard to trigger).
-- Catalin
| |