Re: [BUG -tip] kmemleak and stacktrace cause page faul

From: Cyrill Gorcunov
Date: Tue Oct 22 2019 - 10:23:30 EST


On Sat, Oct 19, 2019 at 02:44:21PM +0300, Cyrill Gorcunov wrote:
...
>
> I nailed it down to the following kmemleak code
>
> create_object
> ...
> object->trace_len = __save_stack_trace(object->trace);
>
> if I drop this line out it boots fine. Just wanted to share the observation,
> probably it is known issue already.
>
> Sidenote: The last -tip kernel which I've been working with is dated Sep 18
> so the changes which cause the problem should be introduced last month.

I've just tried to boot with fresh -tip

commit c2e50c28eeb90c0f3309d43c2ce0bd292a6751b3 (HEAD -> master, origin/master, origin/HEAD)
Merge: aec1ea9d4f48 27a0a90d6301
Author: Ingo Molnar <mingo@xxxxxxxxxx>
Date: Tue Oct 22 01:16:59 2019 +0200

Merge branch 'perf/core'

Conflicts:
tools/perf/check-headers.sh

but got the same issue. So I tried to go deeper, and here is a result: we're failing
in arch/x86/kernel/dumpstack_64.c:in_exception_stack routine, precisely at line

ep = &estack_pages[k];
/* Guard page? */
if (!ep->size)
return false;

so I added a logline here

[ 0.082275] stk 0x1010 k 1 begin 0x0 end 0xd000 estack_pages 0xffffffff82014880 ep 0xffffffff82014888
[ 0.084951] BUG: unable to handle page fault for address: 0000000000001ff0
[ 0.086724] #PF: supervisor read access in kernel mode
[ 0.088506] #PF: error_code(0x0000) - not-present page
[ 0.090265] PGD 0 P4D 0
[ 0.090846] Oops: 0000 [#2] PREEMPT SMP PTI
[ 0.091734] CPU: 0 PID: 0 Comm: swapper Not tainted 5.4.0-rc4-00258-gc2e50c28eeb9-dirty #114
[ 0.093514] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014
[ 0.096993] RIP: 0010:get_stack_info+0xdc/0x173
[ 0.097994] Code: 84 48 01 82 66 85 c0 74 27 42 8b 14 f5 80 48 01 82 49 01 d5 42 0f b7 14 f5 86 48 01 82 4c 01 e8 4c 89 6b 08 89 13 48 89 43 10 <48> 8b 40 f0 eb 2b 65 48 8b 05 16 f4 f9 7e 48 8d 90 00 c0 ff ff 48

I presume the kmemleak tries to save stack trace too early when estack_pages are not
yet filled.