Re: [PATCH 2/2] Catching errors on probes-based dynamic linker interface

[Sergio Durigan Junior <sergiodj at redhat dot com>]

On Tuesday, September 01 2015, I wrote:

>> I am ok with doing this:
>>
>>   TRY
>>     {
>>       probe_argc = get_probe_argument_count (pa->probe, frame);
>>     }
>>   CATCH (ex, RETURN_MASK_ERROR)
>>     {
>>       exception_print (gdb_stderr, ex);
>>       probe_argc = 0;
>>     }
>>   END_CATCH
>>
>> If you put a big fat comment above the following block, e.g.:
>>
>>   /* Note that failure of get_probe_argument_count will
>>      set probe_argc == 0.  This must result in returning
>>      action = PROBES_INTERFACE_FAILED.  */
>>   if (probe_argc == 2)
>>     action = FULL_RELOAD;
>>   else if (probe_argc < 2)
>>     action = PROBES_INTERFACE_FAILED;
>
> Great, that works for me as well.  I will update the patch here to
> address this.

I took the liberty to modify and expand the comment; I hope you still
find it OK.  Here's what I pushed.

Thanks,

-- 
Sergio
GPG key ID: 237A 54B1 0287 28BF 00EF  31F4 D0EB 7628 65FC 5E36
Please send encrypted e-mail if possible
http://sergiodj.net/

>From eb7f0ec56321645e1999ddbac9db4e4cd14d03e4 Mon Sep 17 00:00:00 2001
From: Sergio Durigan Junior <sergiodj@redhat.com>
Date: Fri, 21 Aug 2015 18:28:07 -0400
Subject: [PATCH 2/2] Catching errors on probes-based dynamic linker interface

This patch is intended to make the interaction between the
probes-based dynamic linker interface and the SystemTap SDT probe code
on GDB more robust.  It does that by wrapping the calls to the probe
API with TRY...CATCH'es, so that any exception thrown will be caught
and handled properly.

The idea for this patch came from
<https://bugzilla.redhat.com/show_bug.cgi?id=1196181>, which is a bug
initially filed against Fedora GDB (but now under Fedora GLIBC).  This
bug happens on armhfp (although it could happen on other targets as
well), and is triggered because GCC generates a strange argument for
one of the probes used by GDB in the dynamic linker interface.  As can
be seen in the bug, this argument is "-4@.L1052".

I don't want to discuss the reasons for this argument to be there
(this discussion belongs to the bug, or to another thread), but GDB
could definitely do a better error handling here.  Currently, one sees
the following message when there is an error in the probes-based
dynamic linker interface:

  (gdb) run
  Starting program: /bin/inferior
  warning: Probes-based dynamic linker interface failed.
  Reverting to original interface.

  Cannot parse expression `.L976 4@r4'.
  (gdb)

Which means that one needs to explicitly issue a "continue" command to
make GDB continue running the inferior, even though this error is not
fatal and GDB will fallback to the old interface automatically.

This is where this patch helps: it makes GDB still print the necessary
warnings or error messages, but it *also* does not stop the inferior
unnecessarily.

I have tested this patch on the systems where this error happens, but
I could not come up with a way to create a testcase for it.
Nevertheless, it should be straightforward to see that this patch does
improve the current situation.

OK to apply?

gdb/ChangeLog:
2015-08-21  Sergio Durigan Junior  <sergiodj@redhat.com>

	* solib-svr4.c (solib_event_probe_action): Call
	get_probe_argument_count using TRY...CATCH.
	(svr4_handle_solib_event): Likewise, for evaluate_probe_argument.
---
 gdb/solib-svr4.c | 43 ++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)

diff --git a/gdb/solib-svr4.c b/gdb/solib-svr4.c
index 36b6c59..5d2b9dd 100644
--- a/gdb/solib-svr4.c
+++ b/gdb/solib-svr4.c
@@ -1786,7 +1786,23 @@ solib_event_probe_action (struct probe_and_action *pa)
        arg0: Lmid_t lmid (mandatory)
        arg1: struct r_debug *debug_base (mandatory)
        arg2: struct link_map *new (optional, for incremental updates)  */
-  probe_argc = get_probe_argument_count (pa->probe, frame);
+  TRY
+    {
+      probe_argc = get_probe_argument_count (pa->probe, frame);
+    }
+  CATCH (ex, RETURN_MASK_ERROR)
+    {
+      exception_print (gdb_stderr, ex);
+      probe_argc = 0;
+    }
+  END_CATCH
+
+  /* If get_probe_argument_count throws an exception, probe_argc will
+     be set to zero.  However, if pa->probe does not have arguments,
+     then get_probe_argument_count will succeed but probe_argc will
+     also be zero.  Both cases happen because of different things, but
+     they are treated equally here: action will be set to
+     PROBES_INTERFACE_FAILED.  */
   if (probe_argc == 2)
     action = FULL_RELOAD;
   else if (probe_argc < 2)
@@ -1940,7 +1956,17 @@ svr4_handle_solib_event (void)
   usm_chain = make_cleanup (resume_section_map_updates_cleanup,
 			    current_program_space);
 
-  val = evaluate_probe_argument (pa->probe, 1, frame);
+  TRY
+    {
+      val = evaluate_probe_argument (pa->probe, 1, frame);
+    }
+  CATCH (ex, RETURN_MASK_ERROR)
+    {
+      exception_print (gdb_stderr, ex);
+      val = NULL;
+    }
+  END_CATCH
+
   if (val == NULL)
     {
       do_cleanups (old_chain);
@@ -1971,7 +1997,18 @@ svr4_handle_solib_event (void)
 
   if (action == UPDATE_OR_RELOAD)
     {
-      val = evaluate_probe_argument (pa->probe, 2, frame);
+      TRY
+	{
+	  val = evaluate_probe_argument (pa->probe, 2, frame);
+	}
+      CATCH (ex, RETURN_MASK_ERROR)
+	{
+	  exception_print (gdb_stderr, ex);
+	  do_cleanups (old_chain);
+	  return;
+	}
+      END_CATCH
+
       if (val != NULL)
 	lm = value_as_address (val);
 
-- 
2.4.3