Re: [RFC] Aligning segments in gold

[Ian Lance Taylor <iant at google dot com>]

On Fri, Jan 9, 2015 at 2:29 PM, Cary Coutant <ccoutant@google.com> wrote:
>
> In Layout::set_segment_offsets, we have this code:
>
>               // Skip the address forward one page, maintaining the same
>               // position within the page.  This lets us store both segments
>               // overlapping on a single page in the file, but the loader will
>               // put them on different pages in memory. We will revisit this
>               // decision once we know the size of the segment.
>
>               addr = align_address(addr, (*p)->maximum_alignment());
>               aligned_addr = addr;
>
>               if (load_seg == *p)
>                 {
>                   // ...
>                 }
>               else
>                 {
>                   if ((addr & (abi_pagesize - 1)) != 0)
>                     addr = addr + abi_pagesize;
>
>                   off = orig_off + ((addr - orig_addr) & (abi_pagesize - 1));
>                 }
>
> Usually, this works fine -- we'll set addr = addr + abi_pagesize, then
> adjust off so that we share a page in the file. But what happens when
> (*p)->maximum_alignment() returns a large value like 4096? We end up
> aligning addr first to a page boundary, and that's all. If -z relro
> adjustment results in more padding such that we *could* have
> overlapped the segments in the file, we lose the opportunity and end
> up wasting a page. If we subsequently run the output through strip,
> strip will "fix" that for us, the p_offset fields won't match, and GDB
> will complain that the unstripped binary doesn't match the core file.
>
> I'm thinking we don't need to align to (*p)->maximum_alignment() at
> all here, unless it's larger than abi_pagesize. Each output section in
> the segment will be aligned appropriately without that, and we don't
> need the segment's starting address to match that alignment, since the
> loader always preserves the offset modulo page size when relocating
> segments. Indeed, with relro adjustment, the starting address of the
> segment rarely matches the alignment of its most strictly-aligned
> section.

(*p)->maximum_alignment is going to feed into the p_align field.  It
sounds like you are saying that it's OK if a phdr is not aligned
according to what its p_align field says.  Is that right?

Other than that, what you say makes sense to me.

Ian