A question about PROT_EXEC-only pages on IA64

[Jamie Lokier <jamie@shareable.org>]

Hi IA64 folks,

I'm doing a survey of the different architectural implementations of
PROT_* flags for mmap() and mprotect().  I'm looking at linux-2.6.5.

According to my reading of the source, from include/asm-ia64/pgtable.h
and arch/ia64/mm/fault.c, IA64 on Linux implements the following:

Requested PROT flags | ---    R--    -W-    RW-    --X    R-X    -WX    RWX
=====================+======================================================
ia64 MAP_SHARED      | ---    r--    !w-    rw-    *-x    r-x    !wx    rwx
ia64 MAP_PRIVATE     | ---    r--    !w-    rw-    *-x    r-x    !wx    rwx

"!" means that a read access raises a signal *sometimes*.  This is
because the page protection when the page is installed allows reading,
but if there isn't a page installed, then a read fault will raise a
signal.

(Several architectures have "!" entries, but their places in the table
vary.  On x86_64, and i386 with NX, for example, the sequence is "---
r-- !w- rw- r-x r-x rwx rwx": notice no "!" in the -WX case.  That's
because of different logic in the fault handler.)

I have a question about about the column with "*".  You have
implemented PROT_EXEC-only pages using these flags:

#define __P100	__pgprot(__ACCESS_BITS | _PAGE_PL_3 | _PAGE_AR_X_RX)

I.e. _PAGE_AR_X_RX.  My question is: does this mean that reading those
pages in *kernel* mode will succeed, i.e. so that write() would
succeed in reading from those pages?

If that's the behaviour, and the intention is to create exec-only
pages, then it's a bug.  A similar bug has been found on the Sparc for
PROT_NONE pages, which is more serious.

It's not necessarily important on IA64, because IA64 is the _only_
Linux platform which does exec-only pages.  (This despite other
hardware being able to have exec-only permissions.  See another mail
about that in a few minutes).

I would simply like to know if this is the IA64 behaviour, where a
PROT_EXEC-only area is readable by get_user(), to document it.  (It's
possible that I don't understand the technical implications of
_PAGE_AR_X_RX, and that in fact get_user() in kernel won't be able to
read exec-only pages.)

Thanks,
-- Jamie