Bug#993760: libexiv2-27:i386: non-baseline instructions cause crash on AMD Geode

[Simon McVittie]

Control: retitle -1 libexiv2-27:i386: non-baseline instructions cause SIGILL on AMD Geode
Control: reassign -1 libexiv2-27
Control: affects -1 + tracker-extract
Control: tags -1 + moreinfo

On Mon, 06 Sep 2021 at 09:48:00 +0300, Martin-Éric Racine wrote:
>        Message: Process 838 (tracker-extract) of user 1000 dumped core.
>                 
>                 Stack trace of thread 838:
>                 #0  0x00000000b1009f90 n/a (libexiv2.so.27 + 0x74f90)

This looks to me like it's libexiv2.so.27 that is crashing, with an
illegal instruction (SIGILL). You're using an AMD Geode, which is
very close to the minimum for Debian's i386 baseline, so non-baseline
instructions are likely to be found there first.

What version of libexiv2-27 is installed? Unfortunately reportbug
only lists version numbers for direct dependencies, and tracker-extract
uses libexiv2-27 via libgexiv2-2 rather than directly.

> (gdb) x/16i $pc
> => 0xb1009f90:	endbr32 

If I remember correctly, endbr32 is a CET opcode (a repurposed "long NOP")
that is supported on most i686-class CPUs, but notably is considered an
illegal instruction on Geode. This seems most likely to be something
to do with how libexiv2-27 was compiled, rather than the libexiv2-27
source code.

If libexiv2-27 had been rebuilt in unstable recently then I would be saying
this is probably fallout from gcc-11 briefly having enabled CET on i386
(#993162, #993172), but it hasn't - unless you have rebuilt it locally?

It look as though tracker-extract is inspecting some file in "raw" format.
Can you find out (perhaps from /proc/$PID/fd) which file, and try running
the exiv2 tool against it? If that crashes too, then this is certainly not
a tracker bug.

    smcv