[PATCH] program header instability

[Nathan Sidwell]

This patch fixes a problem with instability in calculating program headers.  The 
original testcase came from a vxworks build, but the problem can manifest on any 
elf target.  The problem occurs when a section boundary happens to lie on a page 
boundary, and the previous and next sections differ in the WRITABLE attribute.

This bit of code in elf.c:
	  else if (! writable
		   && (hdr->flags & SEC_READONLY) == 0
		   && (((last_hdr->lma + last_size - 1)
			& ~(maxpagesize - 1))
		       != (hdr->lma & ~(maxpagesize - 1))))
	    {
	      /* We don't want to put a writable section in a read only
		 segment, unless they are on the same page in memory
		 anyhow.  We already know that the last section does not
		 bring us past the current section on the page, so the
		 only case in which the new section is not on the same
		 page as the previous section is when the previous section
		 ends precisely on a page boundary.  */
	      new_segment = TRUE;

places those two sections in different segments, necessitating an additional 
program header.  Now, when the new program header is included, the section 
boundary is no longer page aligned, and we decide we only need 1 program header 
for the two sections.  So the section boundary shifts down and we're back where 
we started.

The linker bails out with
	./ld-new: looping in map_segments

That loop fails because, even though bfd caches the program_header size during 
the linking layout phase, we keep changing our minds as to whether the first 
segment contains the pheaders or not, depending on whether we think we need N or 
N+1 of them.
	  if ((abfd->flags & D_PAGED) == 0
	      || sections[0]->lma < phdr_size
	      || sections[0]->lma % maxpagesize < phdr_size % maxpagesize)
	    phdr_in_segment = FALSE;

This patch changes elf.c to initialize the program_header_size to zero (rather 
than the 'unknown' value of -1), and then always recalculate it.  furthermore we 
never reduce the size of the program headers, to avoid the above instability.

Tested in i686-pc-linux-gnu, ok?

nathan
-- 
Nathan Sidwell    ::   http://www.codesourcery.com   ::         CodeSourcery
nathan@codesourcery.com    ::     http://www.planetfall.pwp.blueyonder.co.uk

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: all.diff
URL: <https://sourceware.org/pipermail/binutils/attachments/20061219/ae1f7490/attachment.ksh>