Re: [PATCH] gdb/breakpoint: remove assertion when printing internal breakpoints without locspecs

[Andrew Burgess <aburgess@redhat.com>]

Hi Sébastien,

Thanks for looking at this.  I think this looks good, I have a couple of
minor nits to feedback, see inline below.

Sébastien Darche <sdarche@efficios.com> writes:

> When debugging gdb with `set debug breakpoint on`, gdb would crash if
> the inferior happened to unload (through dlclose ()) a shared library
> that contains jit debugging symbols:
>
> 	../../gdb/breakpoint.c:6473: internal-error: print_breakpoint_location:
> 	Assertion `b->locspec != nullptr || (!user_breakpoint_p (b) && (b->type
> 	== bp_shlib_event || b->type == bp_thread_event))' failed.
> 	A problem internal to GDB has been detected [...]
>
> This assertion was added in commit 5770f68 ("gdb: handle empty locspec
> when printing breakpoints"). The assumption for this assert is that the
> user may not call any (gdb-)debugging command between the time a shared
> library is marked (bp_location->shlib_disabled is set) and the
> breakpoint is actually removed. This would be true if the user called
> `maint info breakpoints` as explained in the original commit message.
>
> However, if `set debug breakpoint on` is called, gdb prints debugging
> info as the shared library is being unloaded. The ~jiter_objfile_data
> dtor deletes the breakpoint, which calls remove_breakpoint_1 on the
> bp_location in the recently unloaded shared object. Since `debug
> breakpoint` is enabled, we go through print_breakpoint_location with the
> bp_location's shlib_disabled flag set.. which results in the assert
> above.

I think the ".. " should be replace with ", ".

>
> This commit removes the assertion as it is not true in all cases. I've
> also included a minimal test case that loads a shared library with a
> __jit_debug_register_code symbol. The test fails with the assert.


>
> Change-Id: Ia19c84f194a6f0c10315548c7f423bfd86ef0266
> ---
>  gdb/breakpoint.c                          | 17 ---------
>  gdb/testsuite/gdb.base/jit-unload-solib.c | 21 +++++++++++
>  gdb/testsuite/gdb.base/jit-unload.c       | 45 +++++++++++++++++++++++
>  gdb/testsuite/gdb.base/jit-unload.exp     | 44 ++++++++++++++++++++++
>  4 files changed, 110 insertions(+), 17 deletions(-)
>  create mode 100644 gdb/testsuite/gdb.base/jit-unload-solib.c
>  create mode 100644 gdb/testsuite/gdb.base/jit-unload.c
>  create mode 100644 gdb/testsuite/gdb.base/jit-unload.exp
>
> diff --git a/gdb/breakpoint.c b/gdb/breakpoint.c
> index 101dc57ee6b..85c4ffc3171 100644
> --- a/gdb/breakpoint.c
> +++ b/gdb/breakpoint.c
> @@ -6457,23 +6457,6 @@ print_breakpoint_location (const breakpoint *b, const bp_location *loc)
>      }
>    else
>      {
> -      /* Internal breakpoints don't have a locspec string, but can become
> -	 pending if the shared library the breakpoint is in is unloaded.
> -	 For most internal breakpoint types though, after unloading the
> -	 shared library, the breakpoint will be deleted and never recreated
> -	 (see internal_breakpoint::re_set).  But for two internal
> -	 breakpoint types bp_shlib_event and bp_thread_event this is not
> -	 true.  Usually we don't expect the libraries that contain these
> -	 breakpoints to ever be unloaded, but a buggy inferior might do
> -	 such a thing, in which case GDB should be prepared to handle this
> -	 case.
> -
> -	 If these two breakpoint types become pending then there will be no
> -	 locspec string.  */
> -      gdb_assert (b->locspec != nullptr
> -		  || (!user_breakpoint_p (b)
> -		      && (b->type == bp_shlib_event
> -			  || b->type == bp_thread_event)));
>        const char *locspec_str
>  	= (b->locspec != nullptr ? b->locspec->to_string () : "");
>        uiout->field_string ("pending", locspec_str);
> diff --git a/gdb/testsuite/gdb.base/jit-unload-solib.c b/gdb/testsuite/gdb.base/jit-unload-solib.c
> new file mode 100644
> index 00000000000..3cf2c580320
> --- /dev/null
> +++ b/gdb/testsuite/gdb.base/jit-unload-solib.c
> @@ -0,0 +1,21 @@
> +/* This test program is part of GDB, the GNU debugger.
> +
> +   Copyright 2026 Free Software Foundation, Inc.
> +
> +   This program is free software; you can redistribute it and/or modify
> +   it under the terms of the GNU General Public License as published by
> +   the Free Software Foundation; either version 3 of the License, or
> +   (at your option) any later version.
> +
> +   This program is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> +   GNU General Public License for more details.
> +
> +   You should have received a copy of the GNU General Public License
> +   along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
> +
> +/* This simulates a JIT library.  */
> +
> +
> +#include "jit-protocol.h"
> diff --git a/gdb/testsuite/gdb.base/jit-unload.c b/gdb/testsuite/gdb.base/jit-unload.c
> new file mode 100644
> index 00000000000..def66b53861
> --- /dev/null
> +++ b/gdb/testsuite/gdb.base/jit-unload.c
> @@ -0,0 +1,45 @@
> +/* This test program is part of GDB, the GNU debugger.
> +
> +   Copyright 2026 Free Software Foundation, Inc.
> +
> +   This program is free software; you can redistribute it and/or modify
> +   it under the terms of the GNU General Public License as published by
> +   the Free Software Foundation; either version 3 of the License, or
> +   (at your option) any later version.
> +
> +   This program is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> +   GNU General Public License for more details.
> +
> +   You should have received a copy of the GNU General Public License
> +   along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
> +
> +/* Simulate loading of a library JIT code.  */
> +
> +#include <stdio.h>
> +#include <stdlib.h>
> +
> +#ifdef __WIN32__
> +#include <windows.h>
> +#define dlopen(name, mode) LoadLibrary (TEXT (name))
> +#define dlclose(handle) FreeLibrary (handle)
> +#else
> +#include <dlfcn.h>
> +#endif
> +
> +
> +int
> +main()

Missing space after 'main'.  Also, for C code I think 'main (void)' is
what we should be writing, though we do have a lot of 'main ()' already.

> +{
> +  void *handle = dlopen (SHLIB_NAME, RTLD_NOW);
> +
> +  if (handle == nullptr)

The 'nullptr' is a C23 feature.  To allow this test to compile with
older compilers it would be better to use 'NULL'.

> +    {
> +      fprintf (stderr, "%s\n", dlerror ());
> +      exit (1);
> +    }
> +
> +  dlclose (handle);
> +  return 0;
> +}
> diff --git a/gdb/testsuite/gdb.base/jit-unload.exp b/gdb/testsuite/gdb.base/jit-unload.exp
> new file mode 100644
> index 00000000000..0bf9c59385d
> --- /dev/null
> +++ b/gdb/testsuite/gdb.base/jit-unload.exp
> @@ -0,0 +1,44 @@
> +# Copyright 2026 Free Software Foundation, Inc.
> +
> +# This program is free software; you can redistribute it and/or modify
> +# it under the terms of the GNU General Public License as published by
> +# the Free Software Foundation; either version 3 of the License, or
> +# (at your option) any later version.
> +#
> +# This program is distributed in the hope that it will be useful,
> +# but WITHOUT ANY WARRANTY; without even the implied warranty of
> +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> +# GNU General Public License for more details.
> +#
> +# You should have received a copy of the GNU General Public License
> +# along with this program.  If not, see <http://www.gnu.org/licenses/>.
> +
> +# This test checks that gdb does not assert when unloading a shared library
> +# that defines a __jit_debug_register_code if `set debug breakpoint on` is set.
> +
> +require allow_shlib_tests
> +
> +standard_testfile .c -solib.c
> +set libfile ${testfile}-lib
> +
> +set options [list \
> +		 debug \
> +		 shlib_load \
> +		 additional_flags=-DSHLIB_NAME=\"${libfile}\"]
> +
> +if { [build_executable "build main file" $testfile $srcfile $options] == -1 } {
> +    return
> +}
> +
> +if { [build_executable "build shlib" $libfile $srcfile2 {debug shlib}] == -1 } {
> +    return
> +}

This isn't going to work for remote target boards as the shlib will
still be on the build machine.  If you rewrite like this to make use of
gdb_download_shlib then things will be better.  I tested this using
'make check-all-boards', which is described in gdb/testsuite/README:

  require allow_shlib_tests
  
  standard_testfile .c -solib.c
  
  set lib_testfile ${testfile}-lib
  set lib_binfile [standard_output_file $lib_testfile]
  
  if { [build_executable "build shlib" $lib_testfile $srcfile2 \
  	  {debug shlib}] == -1 } {
      return
  }
  
  set lib_testfile_on_target [gdb_download_shlib $lib_binfile]
  
  set options [list \
  		 debug \
  		 shlib_load \
  		 additional_flags=-DSHLIB_NAME=\"${lib_testfile_on_target}\"]
  
  if { [prepare_for_testing "build main file" $testfile $srcfile \
  	  $options] == -1 } {
      return
  }

> +
> +clean_restart $::testfile

I propose using prepare_for_testing above, so this line will no longer
be needed.

> +
> +if { ![runto_main] } {
> +    return
> +}
> +
> +gdb_test_no_output "set debug breakpoint on"
> +gdb_continue_to_end "unload" continue 1

I worry about trying to handle all of the possible debug breakpoint
output using gdb_continue_to_end, which matches all the output with a
".*" pattern; there's a risk of buffer overflow in the future.

I'd prefer to see this handled with something like:

  set saw_exit false
  gdb_test_multiple "continue" unload {
      -re "^\\\[Inferior $decimal \[^\r\n\]+ exited normally\\\]\r\n" {
  	set saw_exit true
  	exp_continue
      }
  
      -re "^$gdb_prompt $" {
  	gdb_assert ${saw_exit} $gdb_test_name
      }
  
      -re "^\[^\r\n\]*\r\n" {
  	exp_continue
      }
  }

which handles the output line by line, and so will be happy with any
volume of debug output.

Thanks,
Andrew