On 19 July 2017 at 23:27, H. Peter Anvin [off-list ref] wrote:

On 07/19/17 08:40, Thomas Garnier wrote:

quoted

quoted

This doesn't look right.  It's accessing a per-cpu variable.  The
per-cpu section is an absolute, zero-based section and not subject to
relocation.

PIE does not respect the zero-based section, it tries to have
everything relative. Patch 16/22 also adapt per-cpu to work with PIE
(while keeping the zero absolute design by default).

This is silly.  The right thing is for PIE is to be explicitly absolute,
without (%rip).  The use of (%rip) memory references for percpu is just
an optimization.

Sadly, there is an issue in binutils that may prevent us from doing
this as cleanly as we would want.

For historical reasons, bfd.ld emits special symbols like
__GLOBAL_OFFSET_TABLE__ as absolute symbols with a section index of
SHN_ABS, even though it is quite obvious that they are relative like
any other symbol that points into the image. Unfortunately, this means
that binutils needs to emit R_X86_64_RELATIVE relocations even for
SHN_ABS symbols, which means we lose the ability to use both absolute
and relocatable symbols in the same PIE image (unless the reloc tool
can filter them out)

More info here:
https://sourceware.org/bugzilla/show_bug.cgi?id=19818