[patch/rfc] Fix gdb/1426: osabi initialization race condition

[patch/rfc] Fix gdb/1426: osabi initialization race condition

[Andrew Cagney]

[-- Attachment #1: Type: text/plain, Size: 488 bytes --]

Hello,

ref: http://sources.redhat.com/gdb/bugs/1426 for details.

The attached implements the "hack" suggested in tdep/1426 to work around 
a bogus error message:

A handler for the OS ABI "GNU/Linux" is not built into this 
configuration of GDB.  Attempting to continue with the default 
rs6000:6000 settingsGNU gdb 2003-10-23-cvs
Copyright 2003 Free Software Foundation, Inc.
...

that appears on some platforms during startup.

Baring comment, I'll commit this in a few days.

Andrew

[-- Attachment #2: diffs --]
[-- Type: text/plain, Size: 3778 bytes --]

2003-10-23  Andrew Cagney  <cagney@redhat.com>

	* osabi.c (GDB_OSABI_DEFAULT): Delete macro definition.
	(user_osabi_default): New static variable.
	(_initialize_gdb_osabi): Initialize "user_osabi_default" to
	GDB_OSABI_DEFAULT if defined, GDB_OSABI_UNKNOWN otherwize.
	(gdbarch_lookup_osabi): Replace GDB_OSABI_DEFAULT with
	"user_osabi_default".
	(set_osabi, show_osabi): Ditto.
	Work around GDB PR tdep/1426.

Index: osabi.c
===================================================================
RCS file: /cvs/src/src/gdb/osabi.c,v
retrieving revision 1.17
diff -u -r1.17 osabi.c
--- osabi.c	24 Aug 2003 11:47:18 -0000	1.17
+++ osabi.c	24 Oct 2003 02:30:08 -0000
@@ -30,10 +30,6 @@
 
 #include "elf-bfd.h"
 
-#ifndef GDB_OSABI_DEFAULT
-#define GDB_OSABI_DEFAULT GDB_OSABI_UNKNOWN
-#endif
-
 /* State for the "set osabi" command.  */
 static enum { osabi_auto, osabi_default, osabi_user } user_osabi_state;
 static enum gdb_osabi user_selected_osabi;
@@ -45,6 +41,23 @@
 };
 static const char *set_osabi_string;
 
+/* Value for the "default" osabi.
+
+   It is initialized, during per-file initialization, by the function
+   _initialize_osabi below, to the build-time macro GDB_OSABI_DEFAULT.
+   Due to a race condition with the code trying to create the initial
+   "static" architecture, it is not statically initialized to
+   GDB_OSABI_DEFAULT.
+
+   The problem with static initialization is that the code creating
+   that initial "static" architecture calls gdbarch_osabi_init
+   (requesting the "default" OSABI) before any per-file initialization
+   has occured, but it is only during that per-file initialization
+   that the OSABI target code gets to register that said "default"
+   OSABI.  Starting the user_osabi_default out as "uninitialized", and
+   then initializing it at per-file time works around the problem.  */
+static int user_osabi_default = GDB_OSABI_UNINITIALIZED;
+
 /* This table matches the indices assigned to enum gdb_osabi.  Keep
    them in sync.  */
 static const char * const gdb_osabi_names[] =
@@ -210,8 +223,8 @@
      an inconclusive result (otherwise).  */
   if (abfd == NULL) 
     {
-      if (GDB_OSABI_DEFAULT != GDB_OSABI_UNKNOWN)
-	return GDB_OSABI_DEFAULT;
+      if (user_osabi_default != GDB_OSABI_UNKNOWN)
+	return user_osabi_default;
       else
 	return GDB_OSABI_UNINITIALIZED;
     }
@@ -277,8 +290,8 @@
 
   /* If we didn't find a match, but a default was specified at configure
      time, return the default.  */
-  if (GDB_OSABI_DEFAULT != GDB_OSABI_UNKNOWN && match == GDB_OSABI_UNKNOWN)
-    return GDB_OSABI_DEFAULT;
+  if (user_osabi_default != GDB_OSABI_UNKNOWN && match == GDB_OSABI_UNKNOWN)
+    return user_osabi_default;
   else
     return match;
 }
@@ -504,7 +517,7 @@
     user_osabi_state = osabi_auto;
   else if (strcmp (set_osabi_string, "default") == 0)
     {
-      user_selected_osabi = GDB_OSABI_DEFAULT;
+      user_selected_osabi = user_osabi_default;
       user_osabi_state = osabi_user;
     }
   else if (strcmp (set_osabi_string, "none") == 0)
@@ -545,9 +558,9 @@
     printf_filtered ("The current OS ABI is \"%s\".\n",
 		     gdbarch_osabi_name (user_selected_osabi));
 
-  if (GDB_OSABI_DEFAULT != GDB_OSABI_UNKNOWN)
+  if (user_osabi_default != GDB_OSABI_UNKNOWN)
     printf_filtered ("The default OS ABI is \"%s\".\n",
-		     gdbarch_osabi_name (GDB_OSABI_DEFAULT));
+		     gdbarch_osabi_name (user_osabi_default));
 }
 \f
 extern initialize_file_ftype _initialize_gdb_osabi; /* -Wmissing-prototype */
@@ -577,4 +590,9 @@
   add_cmd ("osabi", class_support, show_osabi, "Show OS/ABI of target.",
 	   &showlist);
   user_osabi_state = osabi_auto;
+#ifdef GDB_OSABI_DEFAULT
+  user_osabi_default = GDB_OSABI_DEFAULT;
+#else
+  user_osabi_default = GDB_OSABI_UNKNOWN;
+#endif
 }

Re: [patch/rfc] Fix gdb/1426: osabi initialization race condition

[Daniel Jacobowitz]

On Thu, Oct 23, 2003 at 10:44:42PM -0400, Andrew Cagney wrote:
> Hello,
> 
> ref: http://sources.redhat.com/gdb/bugs/1426 for details.
> 
> The attached implements the "hack" suggested in tdep/1426 to work around 
> a bogus error message:
> 
> A handler for the OS ABI "GNU/Linux" is not built into this 
> configuration of GDB.  Attempting to continue with the default 
> rs6000:6000 settingsGNU gdb 2003-10-23-cvs
> Copyright 2003 Free Software Foundation, Inc.
> ...
> 
> that appears on some platforms during startup.
> 
> Baring comment, I'll commit this in a few days.
> 
> Andrew

This is a rather unfortunate hack.

This is the way it is now, I think:
  - first architecture is created
  - per-file init:
    - then the target's tdep file creates the osabi

But you've changed it to:
  - application startup
    - osabi is set to uninitialized
  - first architecture is created
  - per-file init
    - then osabi.c sets user_osabi_default
    - then the target's tdep file creates the osabi
    (or these two switched, no difference)

This means that if per-file initialization doesn't happen then the
default osabi is never set.  But one of the points of setting a
compile-time default OS ABI was so that even if not given a file, GDB
would get the right default.  This patch will break using gdbserver on
an arm-linux target without loading a file into GDB (unless set osabi
is used, of course).

Also, one thing I couldn't figure out from the code - why does this
happen on some targets and not others?

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [patch/rfc] Fix gdb/1426: osabi initialization race condition

[Andrew Cagney]

> On Thu, Oct 23, 2003 at 10:44:42PM -0400, Andrew Cagney wrote:
> 
>> Hello,
>> 
>> ref: http://sources.redhat.com/gdb/bugs/1426 for details.
>> 
>> The attached implements the "hack" suggested in tdep/1426 to work around 
>> a bogus error message:
>> 
>> A handler for the OS ABI "GNU/Linux" is not built into this 
>> configuration of GDB.  Attempting to continue with the default 
>> rs6000:6000 settingsGNU gdb 2003-10-23-cvs
>> Copyright 2003 Free Software Foundation, Inc.
>> ...
>> 
>> that appears on some platforms during startup.
>> 
>> Baring comment, I'll commit this in a few days.
>> 
>> Andrew
> 
> 
> This is a rather unfortunate hack.

Hmm, patch withdrawn.

> This is the way it is now, I think:
>   - first architecture is created
>   - per-file init:
>     - then the target's tdep file creates the osabi

The sequence was:

- create fake architecture
- per-file init
- create real architecture

but a more careful check reveals that the first call has been eliminated 
- something else is causing that warning :-/

Andrew