* Re: GDB 10.1: Backtrace goes into infinite loop
2020-11-13 22:16 GDB 10.1: Backtrace goes into infinite loop Paul Smith via Gdb
@ 2020-11-16 1:04 ` Simon Marchi
2020-11-21 18:48 ` Paul Smith via Gdb
2021-02-02 3:21 ` Repro case! " Paul Smith via Gdb
1 sibling, 1 reply; 7+ messages in thread
From: Simon Marchi @ 2020-11-16 1:04 UTC (permalink / raw)
To: psmith, gdb
On 2020-11-13 5:16 p.m., Paul Smith via Gdb wrote:
> Hi all;
>
> I just upgraded our users from a toolset using GCC 8.1.0, binutils
> 2.30, and GDB 8.2.1, to a new one using GCC 10.2, binutils 2.35.1, and
> GDB 10.1 (on GNU/Linux x86_64).
>
> Now some of my users are running into a problem where they run the "bt"
> command and it shows some subset of the stack frames, then jumps back
> and starts over printing from frame 0, and does this forever until you
> use ^C to stop it.
>
> Apparently this doesn't happen every time, and the number of frames
> that are shown are variable (but usually a smaller number like 2 to 5
> frames). By "not every time" I mean after a breakpoint sometimes we
> get a good bt and sometimes it recurses, but if it recurses for a given
> bt it will always recurse (that is if you use ^C to stop then "bt"
> again it recurses again).
>
> If we do the same thing with the older GDB (keeping the newer
> compiler/binutils) then we don't see this behavior.
>
> FWIW, the code in question is C++ code and was compiled with -ggdb3 and
> no optimization.
>
> Just wondering if anyone has seen something like this, and/or how to
> try to collect more details.
>
Hi Paul,
When you see the problem happening, you could save a core file (just run
the generate-core-file command). Then, try to reproduce the problem
using the core file. If it reproduces, you could upload it to a bug
report.
Simon
^ permalink raw reply [flat|nested] 7+ messages in thread
* Repro case! Re: GDB 10.1: Backtrace goes into infinite loop
2020-11-13 22:16 GDB 10.1: Backtrace goes into infinite loop Paul Smith via Gdb
2020-11-16 1:04 ` Simon Marchi
@ 2021-02-02 3:21 ` Paul Smith via Gdb
2021-02-02 4:27 ` Paul Smith via Gdb
1 sibling, 1 reply; 7+ messages in thread
From: Paul Smith via Gdb @ 2021-02-02 3:21 UTC (permalink / raw)
To: gdb
Hi all. I'm back! I've still not seen this myself but someone
(finally!) provided me with a reproducible test case (see the quoted
message below for all details). After examining it I am able to come
up with a repro case.
I've discovered that there's something wonky with the embedded Python
interpreter and checking threads while pretty-printing stack variables.
If I don't load our custom Python macros (so that none of our pretty-
printers are defined) then GDB works fine. If I load our macros,
Badness Happens.
FYI I'm linking Python 2.7.18 statically with GDB if that matters.
After spelunking my macros I've discovered that I can reproduce the
problem by replacing all our macros with one that does nothing but show
all threads.
Here is the stupid code I'm using:
// ----- foo.cpp
#include <stdlib.h>
#include <iostream>
class Foo
{
public:
const char* s;
};
void foo(Foo& f)
{
std::cout << f.s << "\n";
abort();
}
int main(int argc, const char** argv)
{
Foo f;
f.s = argv[1] ? argv[1] : "hi";
foo(f);
return 0;
}
// -----
Then I do this:
$ g++ -g -ggdb3 -pthread -o foo foo.cpp
(-pthread is required!!)
Then I run it and get a core:
$ ./foo hiya
hiya
Aborted (core dumped)
Then if I run a simple GDB bt with no Python macros it's fine:
$ gdb -q -batch -ex 'bt' -c core.* foo
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Core was generated by `./foo hiya'.
Program terminated with signal SIGABRT, Aborted.
#0 0x00007f28af56a18b in raise () from /lib/x86_64-linux-gnu/libc.so.6
#0 0x00007f28af56a18b in raise () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x00007f28af549859 in abort () from /lib/x86_64-linux-gnu/libc.so.6
#2 0x000055e3b4dc4203 in foo (f=...) at foo.cpp:13
#3 0x000055e3b4dc4256 in main (argc=2, argv=0x7ffe1045c1c8) at foo.cpp:20
Now I add in a pretty-printer for the Foo class, with this file:
# ----- foo.py
import gdb
class SyncPrinter(object):
def __init__(self, val): self.val = val
def to_string(self):
gdb.selected_inferior().threads() # <---- NOTE!!
return 'hello'
def display_hint(self): return "Foo"
def mkpp():
pp = gdb.printing.RegexpCollectionPrettyPrinter("test")
pp.add_printer('Foo', r'^Foo$', SyncPrinter)
return pp
gdb.printing.register_pretty_printer(gdb.current_objfile(),
mkpp(), replace=True)
# ----- foo.py
Then I run GDB with this:
$ gdb -q -x foo.py -batch -ex 'bt' -c core.* foo
Now the backtrace loops quite a few times, then sometimes it finishes,
sometimes it dumps core with:
gdb/frame.c:2467: internal-error: bool get_frame_pc_if_available(frame_info*,
CORE_ADDR*): Assertion `frame->next != NULL' failed.
If I modify the pretty-printer to run gdb.selected_inferior() without
the threads() call, then it works again.
If anyone has any advice or thoughts I'm interested!! Thx.
On Fri, 2020-11-13 at 17:16 -0500, Paul Smith via Gdb wrote:
> Hi all;
>
> I just upgraded our users from a toolset using GCC 8.1.0, binutils
> 2.30, and GDB 8.2.1, to a new one using GCC 10.2, binutils 2.35.1, and
> GDB 10.1 (on GNU/Linux x86_64).
>
> Now some of my users are running into a problem where they run the "bt"
> command and it shows some subset of the stack frames, then jumps back
> and starts over printing from frame 0, and does this forever until you
> use ^C to stop it.
>
> Apparently this doesn't happen every time, and the number of frames
> that are shown are variable (but usually a smaller number like 2 to 5
> frames). By "not every time" I mean after a breakpoint sometimes we
> get a good bt and sometimes it recurses, but if it recurses for a given
> bt it will always recurse (that is if you use ^C to stop then "bt"
> again it recurses again).
>
> If we do the same thing with the older GDB (keeping the newer
> compiler/binutils) then we don't see this behavior.
>
> FWIW, the code in question is C++ code and was compiled with -ggdb3 and
> no optimization.
>
> Just wondering if anyone has seen something like this, and/or how to
> try to collect more details.
>
^ permalink raw reply [flat|nested] 7+ messages in thread