Re: ping: [patch 6/6] PIE: Fix back re-run

[Jan Kratochvil <jan dot kratochvil at redhat dot com>]

On Mon, 05 Jul 2010 19:48:20 +0200, Joel Brobecker wrote:
> Congrats on being able to teach me some about PIE and prelinking ;-).

Thanks, I even discovered that PIEs are intentionally skipped during
prelinking while responding to your review.  Not that it changes too much on
the code IIRC, anyway it is done.


Checked-in.  The whole series is now in.


Thanks,
Jan


http://sourceware.org/ml/gdb-cvs/2010-07/msg00029.html

--- src/gdb/ChangeLog	2010/07/05 18:00:39	1.11966
+++ src/gdb/ChangeLog	2010/07/05 18:04:32	1.11967
@@ -1,6 +1,13 @@
 2010-07-05  Jan Kratochvil  <jan.kratochvil@redhat.com>
 	    Joel Brobecker  <brobecker@adacore.com>
 
+	Fix re-run of PIE executable, PR shlibs/11776.
+	* solib-svr4.c (svr4_relocate_main_executable) <symfile_objfile>: Remove
+	the part of pre-set SYMFILE_OBJFILE->SECTION_OFFSETS.
+
+2010-07-05  Jan Kratochvil  <jan.kratochvil@redhat.com>
+	    Joel Brobecker  <brobecker@adacore.com>
+
 	* auxv.c (memory_xfer_auxv): Update attach comment.
 	* solib-svr4.c (svr4_special_symbol_handling): Remove the call to
 	svr4_relocate_main_executable.
--- src/gdb/solib-svr4.c	2010/07/05 18:00:39	1.136
+++ src/gdb/solib-svr4.c	2010/07/05 18:04:33	1.137
@@ -1989,17 +1989,32 @@
 {
   CORE_ADDR displacement;
 
-  if (symfile_objfile)
-    {
-      int i;
-
-      /* Remote target may have already set specific offsets by `qOffsets'
-	 which should be preferred.  */
-
-      for (i = 0; i < symfile_objfile->num_sections; i++)
-	if (ANOFFSET (symfile_objfile->section_offsets, i) != 0)
-	  return;
-    }
+  /* If we are re-running this executable, SYMFILE_OBJFILE->SECTION_OFFSETS
+     probably contains the offsets computed using the PIE displacement
+     from the previous run, which of course are irrelevant for this run.
+     So we need to determine the new PIE displacement and recompute the
+     section offsets accordingly, even if SYMFILE_OBJFILE->SECTION_OFFSETS
+     already contains pre-computed offsets.
+
+     If we cannot compute the PIE displacement, either:
+
+       - The executable is not PIE.
+
+       - SYMFILE_OBJFILE does not match the executable started in the target.
+	 This can happen for main executable symbols loaded at the host while
+	 `ld.so --ld-args main-executable' is loaded in the target.
+
+     Then we leave the section offsets untouched and use them as is for
+     this run.  Either:
+
+       - These section offsets were properly reset earlier, and thus
+	 already contain the correct values.  This can happen for instance
+	 when reconnecting via the remote protocol to a target that supports
+	 the `qOffsets' packet.
+
+       - The section offsets were not reset earlier, and the best we can
+	 hope is that the old offsets are still applicable to the new run.
+   */
 
   if (! svr4_exec_displacement (&displacement))
     return;
--- src/gdb/testsuite/ChangeLog	2010/07/05 18:02:56	1.2375
+++ src/gdb/testsuite/ChangeLog	2010/07/05 18:04:33	1.2376
@@ -1,6 +1,14 @@
 2010-07-05  Jan Kratochvil  <jan.kratochvil@redhat.com>
 	    Joel Brobecker  <brobecker@adacore.com>
 
+	Fix re-run of PIE executable, PR shlibs/11776.
+	* gdb.base/break-interp.exp (test_ld): Turn off "disable-randomization".
+	Remove $displacement_main to match the solib-svr4.c change.  New "kill"
+	and re-"run" of the inferior.
+
+2010-07-05  Jan Kratochvil  <jan.kratochvil@redhat.com>
+	    Joel Brobecker  <brobecker@adacore.com>
+
 	Cope with missing /usr/sbin/prelink.
 	* lib/prelink-support.exp (prelink_no):
 	<result == 1 && $output is "no such file or directory">: New.
--- src/gdb/testsuite/gdb.base/break-interp.exp	2010/07/05 18:01:53	1.17
+++ src/gdb/testsuite/gdb.base/break-interp.exp	2010/07/05 18:04:33	1.18
@@ -337,6 +337,11 @@
     # Print the "PIE (Position Independent Executable) displacement" message.
     gdb_test_no_output "set verbose on"
 
+    # We want to test the re-run of a PIE in the case where the executable
+    # is loaded with a different displacement, but disable-randomization
+    # prevents that from happening.  So turn it off.
+    gdb_test "set disable-randomization off"
+
     reach "dl_main" "run segv" $displacement
 
     gdb_test "bt" "#0 +\[^\r\n\]*\\mdl_main\\M.*" "dl bt"
@@ -347,7 +352,13 @@
 	reach "libfunc" continue "NONE"
 
 	gdb_test "bt" "#0 +\[^\r\n\]*\\mlibfunc\\M\[^\r\n\]*\r\n#1 +\[^\r\n\]*\\mmain\\M.*" "main bt"
+    }
 
+    # Try re-run if the new PIE displacement takes effect.
+    gdb_test "kill" "" "kill" {Kill the program being debugged\? \(y or n\) } "y"
+    reach "dl_main" "run segv" $displacement
+
+    if $ifmain {
 	test_core $file $displacement
 
 	test_attach $file $displacement