Re: [PATCH] gdb: relax requirement for the map_failed stap probe to be present

[Simon Marchi <simark@simark.ca>]

On 11/22/22 10:09, Andrew Burgess via Gdb-patches wrote:
> From glibc 2.35 and later, the "map_failed" stap probe is no longer
> included in glibc.  The removal of the probe looks like an accident,
> but it was caused by a glibc commit which meant that the "map_failed"
> probe could no longer be reached; the compiler than helpfully
> optimised out the probe.
> 
> In GDB, in solib-svr4.c, we have a list of probes that we look for
> related to the shared library loading detection.  If any of these
> probes are missing then GDB will fall back to the non-probe based
> mechanism for detecting shared library loading.  The "map_failed"
> probe is include in the list of required probes.
> 
> This means that on glibc 2.35 (or later) systems, GDB is going to
> always fall back to the non-probes based mechanism for detecting
> shared library loading.
> 
> I raised a glibc bug to discuss this issue:
> 
>   https://sourceware.org/bugzilla/show_bug.cgi?id=29818
> 
> But, whatever the ultimate decision from the glibc team, given there
> are version of glibc in the wild without the "map_failed" probe, we
> probably should update GDB to handle this situation.
> 
> The "map_failed" probe is already a little strange, very early
> versions of glibc didn't include this probe, so, in some cases, if
> this probe is missing GDB is happy to ignore it.  In this commit I
> just expand this logic to make the "map_failed" probe fully optional.
> 
> With this commit in place, then, when using a glibc 2.35 or later
> system, GDB will once again use the stap probes for shared library
> detection.
> ---
>  gdb/solib-svr4.c | 13 +++++++++----
>  1 file changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/gdb/solib-svr4.c b/gdb/solib-svr4.c
> index 6acaf87960b..87cd06f251a 100644
> --- a/gdb/solib-svr4.c
> +++ b/gdb/solib-svr4.c
> @@ -2205,10 +2205,15 @@ svr4_find_and_create_probe_breakpoints (svr4_info *info,
>  
>        probes[i] = find_probes_in_objfile (os->objfile, "rtld", name);
>  
> -      /* The "map_failed" probe did not exist in early
> -	 versions of the probes code in which the probes'
> -	 names were prefixed with "rtld_".  */
> -      if (with_prefix && streq (name, "rtld_map_failed"))
> +      /* The "map_failed" probe did not exist in early versions of the
> +	 probes code in which the probes' names were prefixed with
> +	 "rtld_".
> +
> +	 Additionally, the "map_failed" probe was accidentally removed
> +	 from glibc 2.35 and later, when changes in glibc meant the probe
> +	 could no longer be reached.  In this case the probe name doesn't
> +	 have the "rtld_" prefix.  */
> +      if (streq (probe_info[i].name, "map_failed"))
>  	continue;
>  
>        /* Ensure at least one probe for the current name was found.  */
> 
> base-commit: 84f9fbe90e5429adb9dee68f04f44c92fa9e2345
> -- 
> 2.25.4

Hi,

I looked at this separately, and this was one of the fixes I considered.

Another option was to make GDB not give up on the probes interface if
failing to look up a probe whose action is DO_NOTHING.  Probes with that
action are not used by GDB for solib bookkeeping, but can be used to
stop on solib events, with "set stop-on-solib-events".  I was just
worried if there was some cases where a probe would be missing, but the
corresponding event could be caught if using the original interface.  In
that case, using the probes interface would be a regression.  But it's
probably not worth wondering about.  If that happens it's just a bug
that needs to be fixed.  In the case we are looking at, if the
map_failed probe gets optimized out, then surely the corresponding call
to the r_brk function would also be optimized out.

So, the patch LGTM:

Approved-By: Simon Marchi <simon.marchi@efficios.com>

Simon