[patch] Fix run-unstrip-n.sh regression on CentOS-5 ppc

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

Hi,

last patch was a bit more heuristic than needed which was found on RHEL-5 ppc
(32-bit):

FAIL: run-unstrip-n.sh (exit: 1)
================================
--- unstrip.out 2013-07-23 23:23:49.000000000 +0200
+++ -   2013-07-23 23:23:49.434052534 +0200
@@ -4,4 +4,3 @@
 0xfdf0000+0x1c0000 edf3dd232e09d01b90683889bd16b9406c52d4de(a)0xfdf0184 - - libc.so.6
 0xffb0000+0x50000 edec437a85026a1cf8cda94003706202733130c1(a)0xffb0124 - - ld.so.1
 0x10000000+0x20000 979b7a26747cc09bd84a42b311b5288c704baea5(a)0x10000174 . - [exe]
-0xf880000+0x201d4 - /lib/librt.so.1 /usr/lib/debug/lib/librt-2.5.so.debug librt.so.1

Therefore the new code generated this excessive line:
	0xf880000+0x201d4 - /lib/librt.so.1 /usr/lib/debug/lib/librt-2.5.so.debug librt.so.1

The first part of debug dump is from DT_DEBUG, second part is from segments:

start=0xf880000 end=0xf8a01d4 l_ld=0xfd6fe20 name=/lib/librt.so.1
start=0xfc60000 end=0xfe031e4 l_ld=0xff9e270 name=/lib/libc.so.6
start=0xfe10000 end=0xfe421dc l_ld=0xfddfd98 name=/lib/libpthread.so.0
start=0xffb0000 end=0xfff0668 l_ld=0xffef9ac name=/lib/ld.so.1
module_start=0x100000 module_end=0x110000 dyn_vaddr=0x100ee4
module_start=0xfd50000 module_end=0xfd80000 dyn_vaddr=0xfd6fe20 /lib/librt.so.1
module_start=0xfdb0000 module_end=0xfdf0000 dyn_vaddr=0xfddfd98 /lib/libpthread.so.0
module_start=0xfdf0000 module_end=0xffb0000 dyn_vaddr=0xff9e270 /lib/libc.so.6
module_start=0xffb0000 module_end=0x10000000 dyn_vaddr=0xffef9ac /lib/ld.so.1
module_start=0x10000000 module_end=0x10020000 dyn_vaddr=0x10010850

When comparing conflicts for (found in segments)
	module_start=0xfd50000 module_end=0xfd80000 dyn_vaddr=0xfd6fe20 /lib/librt.so.1

the code found this line conflicts (and discarded it):
	start=0xfc60000 end=0xfe031e4 l_ld=0xff9e270 name=/lib/libc.so.6

but it did not discard also conflicting:
	start=0xf880000 end=0xf8a01d4 l_ld=0xfd6fe20 name=/lib/librt.so.1

So I have changed/improved the algorithm - L_LD can be IMO compared exactly
but otherwise the ranges should be compared for every module, not just the
first one.

Again I am not much happy from this code, it should be using NT_FILE instead,
but when we keep compatibility with old OSes elfutils should not regress
there.

This is a code of mine written several days ago so I will check it in if there
is no review soon.


Thanks,
Jan


libdwfl/
2013-07-24  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* dwfl_segment_report_module.c (dwfl_segment_report_module): Check for
	conflicts all the modules, not just the first one.  Compare L_LD if it
	is equal, not if it is in a module address range.

diff --git a/libdwfl/dwfl_segment_report_module.c b/libdwfl/dwfl_segment_report_module.c
index 41487bf..97f4a1a 100644
--- a/libdwfl/dwfl_segment_report_module.c
+++ b/libdwfl/dwfl_segment_report_module.c
@@ -470,28 +470,37 @@ dwfl_segment_report_module (Dwfl *dwfl, int ndx, const char *name,
 	    }
 	}
 
-  /* Ignore this found module if it would conflict in address space with any
-     already existing module of DWFL.  */
   if (r_debug_info != NULL)
-    for (struct r_debug_info_module *module = r_debug_info->module;
-	 module != NULL; module = module->next)
-      if ((module_end > module->start && module_start < module->end)
-          || (module_start <= module->l_ld && module->l_ld < module_end))
-	{
-	  if (! module->disk_file_has_build_id && build_id_len > 0)
-	    {
-	      if (module->elf != NULL)
-		{
-		  elf_end (module->elf);
-		  close (module->fd);
-		  module->elf = NULL;
-		  module->fd = -1;
-		}
-	      break;
-	    }
-	  else if (module->elf != NULL)
-	    return finish ();
-	}
+    {
+      bool skip_this_module = false;
+      for (struct r_debug_info_module *module = r_debug_info->module;
+	   module != NULL; module = module->next)
+	if ((module_end > module->start && module_start < module->end)
+	    || dyn_vaddr == module->l_ld)
+	  {
+	    if (! module->disk_file_has_build_id && build_id_len > 0)
+	      {
+		/* Module found in segments with build-id is more reliable
+		   than a module found via DT_DEBUG on disk without any
+		   build-id.   */
+		if (module->elf != NULL)
+		  {
+		    elf_end (module->elf);
+		    close (module->fd);
+		    module->elf = NULL;
+		    module->fd = -1;
+		  }
+	      }
+	    else if (module->elf != NULL)
+	      {
+		/* Ignore this found module if it would conflict in address
+		   space with any already existing module of DWFL.  */
+		skip_this_module = true;
+	      }
+	  }
+      if (skip_this_module)
+	return finish ();
+    }
 
   /* Our return value now says to skip the segments contained
      within the module.  */