[patch] Fix bfd_elf_bfd_from_remote_memory() loadbase

[Jan Kratochvil]

Hi,

bfd_elf_bfd_from_remote_memory() is currently in use only for vDSOs where it
works fine.  It fails for real memory-mapped ELF files, though.

On the other hand I do not need this fix as my build-id GDB patch no longer
uses bfd_elf_bfd_from_remote_memory() (as it was performance-ineffective).


My memory-mapped dummy test library has:
  Type           Offset   VirtAddr           PhysAddr           FileSiz  MemSiz   Flg Align
  LOAD           0x000000 0x0000000000000000 0x0000000000000000 0x0005dc 0x0005dc R E 0x200000
  LOAD           0x0005e0 0x00000000002005e0 0x00000000002005e0 0x0001e8 0x0001f8 RW  0x200000

which sets the LOADBASE variable by the current BFD CVS HEAD for both for the
first and for the second PT_LOAD segment by

        if (loadbase == ehdr_vma)
                loadbase = ehdr_vma - (i_phdrs[i].p_vaddr & -i_phdrs[i].p_align);

This command has no effect on the first PT_LOAD there (as P_VADDR is zero) and
so even the second PT_LOAD gets processed.

I believe a simple rule "just the first PT_LOAD segment" implemented in my
patch should be enough.

The current PF_R condition there targets ia64 vDSO which has:
  Type           Offset   VirtAddr           PhysAddr           FileSiz  MemSiz   Flg Align
  LOAD           0x000000 0xa000000000000000 0xa000000000000000 0x000610 0x000610 R   0x10000
  LOAD           0x000000 0xa000000000010000 0xa000000000010000 0x0009d0 0x0009d0   E 0x10000
so the (i_phdrs[i].p_flags & PF_R) condition picked the first one (the right
one).  Still this condition is needed there as the memory needs to be readable.


Tested on x86_64 on GDB testsuite and on some IA64 vDSO GDB accesses with no
regressions.


Regards,
Jan
-------------- next part --------------
2007-08-10  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* elfcode.h (NAME(_bfd_elf,bfd_from_remote_memory)): LOADBASE is now
	initialized only on the first PT_LOAD.  New variable LOADBASE_SET.
	Moved the IA-64 vDSO handling comments to the LOADBASE assignment.

--- bfd/elfcode.h	4 Aug 2007 16:31:00 -0000	1.85
+++ bfd/elfcode.h	10 Aug 2007 14:46:27 -0000
@@ -1635,6 +1635,7 @@ NAME(_bfd_elf,bfd_from_remote_memory)
   int err;
   unsigned int i;
   bfd_vma loadbase;
+  bfd_boolean loadbase_set; 
 
   /* Read in the ELF header in external format.  */
   err = target_read_memory (ehdr_vma, (bfd_byte *) &x_ehdr, sizeof x_ehdr);
@@ -1711,12 +1712,11 @@ NAME(_bfd_elf,bfd_from_remote_memory)
   contents_size = 0;
   last_phdr = NULL;
   loadbase = ehdr_vma;
+  loadbase_set = FALSE;
   for (i = 0; i < i_ehdr.e_phnum; ++i)
     {
       elf_swap_phdr_in (templ, &x_phdrs[i], &i_phdrs[i]);
-      /* IA-64 vDSO may have two mappings for one segment, where one mapping
-	 is executable only, and one is read only.  We must not use the
-	 executable one.  */
+      /* PF_R as we need to read the segment memory.  */
       if (i_phdrs[i].p_type == PT_LOAD && (i_phdrs[i].p_flags & PF_R))
 	{
 	  bfd_vma segment_end;
@@ -1725,8 +1725,18 @@ NAME(_bfd_elf,bfd_from_remote_memory)
 	  if (segment_end > (bfd_vma) contents_size)
 	    contents_size = segment_end;
 
-	  if ((i_phdrs[i].p_offset & -i_phdrs[i].p_align) == 0)
-	    loadbase = ehdr_vma - (i_phdrs[i].p_vaddr & -i_phdrs[i].p_align);
+	  /* Only the first PT_LOAD segment indicates the file bias.
+	     Next segments may have P_VADDR arbitrarily higher.
+	     If the first segment has P_VADDR zero any next segment must not
+	     confuse us, the first one sets LOADBASE certainly enough.
+	     IA-64 vDSO may have two mappings for one segment, where the first
+	     mapping is executable only, and the second one is read only.
+	     We must use the first one.  */
+	  if (!loadbase_set && i_phdrs[i].p_offset == 0)
+	    {
+	      loadbase = ehdr_vma - i_phdrs[i].p_vaddr;
+	      loadbase_set = TRUE;
+	    }
 
 	  last_phdr = &i_phdrs[i];
 	}
@@ -1764,9 +1774,7 @@ NAME(_bfd_elf,bfd_from_remote_memory)
     }
 
   for (i = 0; i < i_ehdr.e_phnum; ++i)
-    /* IA-64 vDSO may have two mappings for one segment, where one mapping
-       is executable only, and one is read only.  We must not use the
-       executable one.  */
+    /* PF_R as we need to read the segment memory.  */
     if (i_phdrs[i].p_type == PT_LOAD && (i_phdrs[i].p_flags & PF_R))
       {
 	bfd_vma start = i_phdrs[i].p_offset & -i_phdrs[i].p_align;