Re: [PATCH] runtime/unwind: Allow to increase MAX_CFI

[Mark Wielaard <mark@klomp.org>]

Hi Torsten,
[Sorry for the duplicate, I said I would CC the list, then forgot...]

On Fri, Nov 30, 2018 at 10:31:11PM +0100, Torsten Polle wrote:
> > Am 29.11.2018 um 19:06 schrieb Mark Wielaard <mark@klomp.org>:
> > I am still curious what program/library defines more than 512 CFI
> > instructions. If you could post some example of eu-readelf --debug-dump=frame
> > that would be interesting. I assume it must be some really big
> > functions that haven't been split up?
> 
> PFA the dump. As it is even compressed more than 2 MB, I just send it to you and not to the mailing list.

Wow, I had no idea. That is indeed huge. I inspected it and include the
mailinglist again so there is a bit of a record.

> This is the output of unwind.c with slightly modified messages.
> 
> unwind_frame:1291: processCFI for CIE: /usr/lib/libgtk-3.so.0.2000.9
> unwind_frame:1299: processCFI for FDE
> processCFI:312: Too many CFI instuctions: 5547
> 
> This is the first part backtrace.
> 0x4a8aa:libglib-2.0.so.0.4800.2:0xca2c:libgobject-2.0.so.0.4800.2:0xdf68:libgobject-2.0.so.0.4800.2:0x230c9:libgobject-2.0.so.0.4800.2:0x359d25:libgtk-3.so.0.2000.9

So 5547 is more than a factor 10 bigger than the current 512 limit.
Looking through the dump I see this must be either gtk_widget_class_init
or gtk_settings_class_intern_init. Both have a really big CFI descriptions.

I see the same in my local /usr/lib64/libgtk-3.so.0.2400.1 library.
Both functions seem to be just very long initialization sequnces.

Maybe we should increase the MAX_CFI by default.
But processing so many CFI instructions is resource intensive.
Maybe we should have a better way to just skip frames with
such complicated CFI.

Thanks,

Mark