Re: kmemleak reports in kernel 3.9.5+

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

On Tue, Jun 11, 2013 at 08:52:41PM +0100, Ben Greear wrote:
> On 06/10/2013 03:32 PM, Catalin Marinas wrote:
> > On 10 June 2013 19:22, Ben Greear <greearb@candelatech.com> wrote:
> >> We had a system go OOM while doing lots of wireless
> >> stations.  (System had 8GB of RAM, so I suspect a leak).
> >>
> >> I enabled kmemleak in a 3.9.5 (plus some local patches) and
> >> I see the entries below.  Any idea if these are real or not?
> >>
> >> unreferenced object 0xffff880212281c80 (size 128):
> >>    comm "systemd", pid 1, jiffies 4294682684 (age 1159.517s)
> >>    hex dump (first 32 bytes):
> >>      60 39 27 12 02 88 ff ff 00 02 20 00 00 00 ad de  `9'....... .....
> >>      10 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
> >>    backtrace:
> >>      [<ffffffff815de7bf>] kmemleak_alloc+0x73/0x98
> >>      [<ffffffff8118b4d4>] slab_post_alloc_hook+0x28/0x2a
> >>      [<ffffffff8118d605>] __kmalloc+0xf9/0x122
> >>      [<ffffffff8154946d>] kzalloc.clone.0+0xe/0x10
> >>      [<ffffffff81549494>] fib_default_rule_add+0x25/0x7a
> >>      [<ffffffffa014f5a9>] ip6mr_net_init+0x7e/0x118 [ipv6]
> >>      [<ffffffff8152c992>] ops_init+0xd6/0xf7
> >>      [<ffffffff8152cb51>] register_pernet_operations+0xc2/0x16b
> >>      [<ffffffff8152cc87>] register_pernet_subsys+0x2e/0x47
> >>      [<ffffffffa016db69>] 0xffffffffa016db69
> >>      [<ffffffffa016d109>] 0xffffffffa016d109
> >>      [<ffffffff8100207f>] do_one_initcall+0x7f/0x13e
> >>      [<ffffffff810f3985>] do_init_module+0x44/0x18f
> >>      [<ffffffff810f5da7>] load_module+0x14d1/0x168e
> >>      [<ffffffff810f6114>] sys_init_module+0xfd/0x101
> >>      [<ffffffff815f6599>] system_call_fastpath+0x16/0x1b
> >
> > No idea yet. You can try:
> >
> > echo clear > /sys/kernel/debug/kmemleak
> >
> > and see if there are more appearing after. All seem to have a common
> > allocation path via debug_object_activate -> ... ->
> > rcuhead_fixup_activate -> ... -> __debug_object_init.
> 
> I tried the command below, and it printed out quite a few things.

Can you send me your .config file? I can't reproduce this.

> Also, I read the kmemleak.txt documentation, but a question remains:
> 
> If I enable kmemleak at compile time, but disable it at boot
> time using kmemleak=off, is there any significant runtime overhead?

You still get the callback into kmemleak for each allocation/freeing but
it doesn't walk the stack for the backtrace and it doesn't scan the
memory either. I would say the overhead is very small, probably
unnoticeable if you already have other debug options enabled.

-- 
Catalin