Re: ptrace "fix" breaks ia64

[Roland McGrath <roland@redhat.com>]

> The problem is due to the fact that the gate page on ia64 really does
> live in the kernel-mapped segment (as your original code correctly
> assumed).  Furthermore, pgd_offset_k() is different from pgd_offset()
> since the kernel-mapped segment gets a full page-directory inside a
> single region, whereas user-space regions get only 1/8th of a
> page-directory, so it's not OK to use pgd_offset() in lieu of
> pgd_offset().

Sorry.  I skimmed all the code and comments for pgd_offset_k and thought at
the time that it was strictly an optimized shortcut for pgd_offset.
Clearly I did not understand the ia64 code.

> I suppose we could have a new macro pgd_offset_gate() or something
> along those lines to accommodate platform-differences in where the
> gage page lives.

That seems like the reasonable thing to do.  I considered just putting all
that logic into arch-specific code, joining with the get_gate_vma code.
But that would leave x86_64 requiring duplication of the generic version.
At least with the various arch cases around now, just adding pgd_offset_gate 
is the thing that will allow maximal code sharing.


Thanks,
Roland