linux-kernel - PEBS level 2/3 breaks dwarf unwinding! [WAS: Re: Broken dwarf unwinding

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 01 Nov 2018 23:08:18 +0100
From:   Milian Wolff <milian.wolff@...b.com>
To:     Milian Wolff <milian.wolff@...b.com>
Cc:     Andi Kleen <ak@...ux.intel.com>, linux-kernel@...r.kernel.org,
        Jiri Olsa <jolsa@...nel.org>, namhyung@...nel.org,
        linux-perf-users@...r.kernel.org,
        Arnaldo Carvalho <acme@...nel.org>
Subject: PEBS level 2/3 breaks dwarf unwinding! [WAS: Re: Broken dwarf
 unwinding - wrong stack pointer register value?]

On Dienstag, 30. Oktober 2018 23:34:35 CET Milian Wolff wrote:
> On Mittwoch, 24. Oktober 2018 16:48:18 CET Andi Kleen wrote:
> > > Can someone at least confirm whether unwinding from a function prologue
> > > via
> > > .eh_frame (but without .debug_frame) should actually be possible?
> > 
> > Yes it should be possible. Asynchronous unwind tables should work
> > from any instruction.

<snip>

> We can find `7f91345bdaf8+1 = 7f91345bdaf9" at offset 16 (search for "f9 da
> 5b 34 91 7f"). Using that address makes unwinding work for this sample.
> What could be the reason for this shift?

I believe I have found the culprit: PEBS seems to be at fault here - i.e. the 
RIP/RSP and the ustack dump of the sample simply don't fit together.

Check this out:

```
$ for i in $(seq 10); do perf record -q -e "cycles:" --call-graph dwarf ./cpp-
inlining > /dev/null; perf script | pcre2grep -c -M "hypot_finite.*\n.*\
[unknown\]"; done
0
0
0
0
0
0
0
0
0
0

$ for i in $(seq 10); do perf record -q -e "cycles:p" --call-graph dwarf ./
cpp-inlining > /dev/null; perf script | pcre2grep -c -M "hypot_finite.*\n.*\
[unknown\]"; done
0
0
0
0
0
0
0
0
0
0

$ for i in $(seq 10); do perf record -q -e "cycles:pp" --call-graph dwarf ./
cpp-inlining > /dev/null; perf script | pcre2grep -c -M "hypot_finite.*\n.*\
[unknown\]"; done
37
39
35
28
40
39
29
37
31
26

$ for i in $(seq 10); do perf record -q -e "cycles:ppp" --call-graph dwarf ./
cpp-inlining > /dev/null; perf script | pcre2grep -c -M "hypot_finite.*\n.*\
[unknown\]"; done
79
70
76
77
70
90
64
78
86
74
```

Note how precise levels 0 and 1 do not produce any samples where unwinding 
fails. But precise level 2 produces some, and precise level 3 increases the 
amount (by ca. ~2x).

I can reproduce this pattern on two separate Intel CPUs and kernel versions 
currently:

Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz with 4.18.16-arch1-1-ARCH
Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz with 4.14.78-1-lts

Could someone else try this? What about AMD and IBS - is it also affected? 
What about newer/different Intel CPUs?

Better yet, can someone come up with a fix for this on Intel with maximum 
precise level?

Thanks

-- 
Milian Wolff | milian.wolff@...b.com | Senior Software Engineer
KDAB (Deutschland) GmbH, a KDAB Group company
Tel: +49-30-521325470
KDAB - The Qt, C++ and OpenGL Experts
Download attachment "smime.p7s" of type "application/pkcs7-signature" (3826 bytes)