Re: Table of mmap PROT_* implementations by architecture

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Thu, 1 Jul 2004 04:36:20 +0100, Jamie Lokier <jamie@shareable.org> said:

  >> From a study of Linux 2.6.5 source code, and some patches.
  Jamie> This is based on studying the source, not running tests, so
  Jamie> there may be errors.

  Jamie> This table shows expected page protections, for different
  Jamie> values of PROT_READ, PROT_WRITE and PROT_EXEC passed to
  Jamie> mmap() and mprotect().

  Jamie> (As noted in a recent mail from me, real behaviour isn't
  Jamie> quite this simple.  Reading from a write-only page will
  Jamie> *sometimes* raise a signal, and sometimes not, possibly
  Jamie> dependent on background paging decisions.  Therefore some of
  Jamie> these entries should say "!w!" instead of "rwx", and "!w-"
  Jamie> instead of "rw-".  Perhaps there are other combinations too,
  Jamie> depending on architecture-specific fault handlers).


  Jamie> ==============================================================
  Jamie> Requested PROT flags | --- R-- -W- RW- --X    R-X -WX RWX
  Jamie> ==============================================================
  Jamie> ia64                 | --- r-- rw- rw- --x(1) r-x rwx rwx
  Jamie> --------------------------------------------------------------

  Jamie> (1) - In kernel, maybe these pages are readable using
  Jamie> "write()"?

That's correct for ia64.  The architecture does not support an
"execute-only at all privilege levels" protection per se, so this
behavior is the easiest (most efficient) to implement.  If we really
cared about the "read execute-only at kernel-level", we could use
"probe" instructions in the __access_ok() macro to verify (user-level)
access permission.

	--david