Re: [RFA] Only try to load libthread_db when we load libpthread.

[Doug Evans <dje@google.com>]

[-- Attachment #1: Type: text/plain, Size: 2785 bytes --]

On Thu, Oct 6, 2011 at 1:25 PM, Pedro Alves <pedro@codesourcery.com> wrote:
>
> On Thursday 06 October 2011 21:08:08, Doug Evans wrote:
> > On Thu, Oct 6, 2011 at 12:56 PM, Pedro Alves <pedro@codesourcery.com> wrote:
> > >
> > > On Thursday 06 October 2011 12:22:24, Pedro Alves wrote:
> > > > On Wednesday 05 October 2011 19:27:05, Doug Evans wrote:
> > > > > 2011-10-05  Doug Evans  <dje@google.com>
> > > > >
> > > > >         * linux-thread-db.c (thread_db_new_objfile): Only try to load
> > > > >         libthread_db when we load libpthread.
> > > >
> > > > Makes sense to me.
> > > >
> > > > > No regressions in amd64-linux,
> > > > > but I can imagine it misses some cases.
> > > >
> > > > Yeah.  I think we'll no longer activate thread_db when debugging core
> > > > files of static executables (e.g., a core of gdb.threads/staticthreads).
> > > > It works with live debugging since we call check_for_thread_db
> > > > from linux_child_post_attach/linux_child_post_startup_inferior.
> > > > Maybe moving that to an inferior_created observer in
> > > > linux-thread-db.c would work.
> > >
> > > And all the talk about executables made me realize something else.  :-)
> > >
> > > For static threaded executables, we'll want to check for thread
> > > db when the symbols of the main executable are (re)loaded too.
> > > I don't recall off hand if there's a flag in the objfile to
> > > know that it's from the main executable though.
> >
> > In what scenario?
> > [what would the user type?]
>
> (gdb) file wrong_executable
> (gdb) attach PID or core-file core.1234
> whooops!
> (gdb) file right_executable

What would the user expect to be able to do next?
I ask because I played with it, and things don't necessarily work,
e.g. if wrong_executable didn't have libpthread.
Perhaps more smarts could be added to file to make this work, or maybe
a new command could be added to (effectively) reattach so that the
initialization that attach does was redone (in case one is
uncomfortable with having file do that.  Though I think commands
shouldn't try to do too much.  Otherwise one could say "Why doesn't
attach automagically redo the "file" command since if the first "file"
wasn't done it would anyway?"  Maybe attach should at least warn the
user though).

Anyways, I think checking for the main symbol file is reasonable (even
if more work is needed), so how about this?

2011-10-06  Doug Evans  <dje@google.com>

        * linux-thread-db.c (thread_db_new_objfile): Only try to load
        libthread_db when we load libpthread or the main symbol file.
        * objfiles.h (OBJF_MAINLINE): Define.
        * symfile.c (symbol_file_add_with_addrs_or_offsets): Pass it to
        allocate_objfile when appropriate.

[-- Attachment #2: gdb-111006-libthread-db-2.patch.txt --]
[-- Type: text/plain, Size: 3073 bytes --]

2011-10-06  Doug Evans  <dje@google.com>

	* linux-thread-db.c (thread_db_new_objfile): Only try to load
	libthread_db when we load libpthread or the main symbol file.
	* objfiles.h (OBJF_MAINLINE): Define.
	* symfile.c (symbol_file_add_with_addrs_or_offsets): Pass it to
	allocate_objfile when appropriate.

Index: linux-thread-db.c
===================================================================
RCS file: /cvs/src/src/gdb/linux-thread-db.c,v
retrieving revision 1.91
diff -u -p -r1.91 linux-thread-db.c
--- linux-thread-db.c	13 Sep 2011 19:27:01 -0000	1.91
+++ linux-thread-db.c	6 Oct 2011 21:39:32 -0000
@@ -1083,7 +1083,16 @@ thread_db_new_objfile (struct objfile *o
   /* This observer must always be called with inferior_ptid set
      correctly.  */
 
-  if (objfile != NULL)
+  if (objfile != NULL
+      /* Only check for thread_db if we loaded libpthread,
+	 or if this is the main symbol file.
+	 We need to check OBJF_MAINLINE in case this is a statically
+	 linked executable.
+	 For dynamically linked executables, libpthread can be near the end
+	 of the list of shared libraries to load, and in an app of several
+	 thousand shared libraries, this can otherwise be painful.  */
+      && ((objfile->flags & OBJF_MAINLINE) != 0
+	  || libpthread_name_p (objfile->name)))
     check_for_thread_db ();
 }
 
Index: objfiles.h
===================================================================
RCS file: /cvs/src/src/gdb/objfiles.h,v
retrieving revision 1.85
diff -u -p -r1.85 objfiles.h
--- objfiles.h	14 Jun 2011 16:49:41 -0000	1.85
+++ objfiles.h	6 Oct 2011 21:39:32 -0000
@@ -196,7 +196,8 @@ struct objfile
 
     CORE_ADDR addr_low;
 
-    /* Some flag bits for this objfile.  */
+    /* Some flag bits for this objfile.
+       The values are defined by OBJF_*.  */
 
     unsigned short flags;
 
@@ -434,6 +435,11 @@ struct objfile
 
 #define OBJF_PSYMTABS_READ (1 << 4)
 
+/* Set if this is the main symbol file
+   (as opposed to symbol file for dynamically loaded code).  */
+
+#define OBJF_MAINLINE (1 << 5)
+
 /* The object file that contains the runtime common minimal symbols
    for SunOS4.  Note that this objfile has no associated BFD.  */
 
Index: symfile.c
===================================================================
RCS file: /cvs/src/src/gdb/symfile.c,v
retrieving revision 1.316
diff -u -p -r1.316 symfile.c
--- symfile.c	29 Sep 2011 02:04:25 -0000	1.316
+++ symfile.c	6 Oct 2011 21:39:34 -0000
@@ -1084,6 +1084,7 @@ symbol_file_add_with_addrs_or_offsets (b
   const int should_print = ((from_tty || info_verbose)
 			    && (readnow_symbol_files
 				|| (add_flags & SYMFILE_NO_READ) == 0));
+  const int mainline = add_flags & SYMFILE_MAINLINE;
 
   if (readnow_symbol_files)
     {
@@ -1102,7 +1103,7 @@ symbol_file_add_with_addrs_or_offsets (b
       && !query (_("Load new symbol table from \"%s\"? "), name))
     error (_("Not confirmed."));
 
-  objfile = allocate_objfile (abfd, flags);
+  objfile = allocate_objfile (abfd, flags | (mainline ? OBJF_MAINLINE : 0));
   discard_cleanups (my_cleanups);
 
   if (parent)