Re: [PATCH 2.6.17-rc6 7/9] Remove some of the kmemleak false positives

[Catalin Marinas]

On 13/06/06, Pekka J Enberg <penberg@xxxxxxxxxxxxxx> wrote:
> Hi Catalin,
>
> On Tue, 13 Jun 2006, Catalin Marinas wrote:
> > The gc roots are the data and bss sections (and maybe task kernel
> > stacks) and all the slab-allocated blocks are scanned if a link to
> > them is found from the roots (and all of them are usually scanned). If
> > no link is found, they would be reported as memory leaks (and not
> > scanned). You can't really avoid the scanning of allocated blocks
> > since they may contain pointers to other blocks.
>
> I am not sure you're agreeing or disagreeing :-).

I'm not sure either :-). Doing some quick statistics on an ARM
platform shows that there are plenty of values (almost 15%) in the
scanned memory that look like pointers (i.e. between 3GB and
3GB+128MB) but do not point to any allocated block:

total scanned values = 932102
total detected pointers = 6270
detected by alias (i.e. in-block address) = 2096
looking like pointer = 135675

I'll add a test to see how many of the look-like pointers actually
point to a valid in-block address. If this is considerable larger than
the detected by alias above, it shouldn't be implemented.

> As far as I understood, Ingo is worried about:
>
>         struct s { /* some fields */; char *buf; };
>
>         struct s *p = kmalloc(sizeof(struct s) + BUF_SIZE);
>         p->buf = p + sizeof(struct s);
>
> Which could lead to false negative due to p->buf pointing to p.  However,
> for us to even _find_ p->buf, we would need an incoming pointer _to_ p
> which makes me think this is not a problem in practice.  Hmm?

Not exactly. In the above case, Ingo (and me) is worried about having
a incoming pointer (from other block) equal to p->buf and therefore
inside the block allocated with kmalloc. It doesn't matter whether any
value in a block point to locations inside the block since the actual
block first needs to be accessible on the directed graph starting from
the root.

--
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/