[Bug c/59850] Support sparse-style pointer address spaces (type attributes)

[pageexec at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59850

--- Comment #23 from PaX Team <pageexec at gmail dot com> ---
some data points based on my experience with the 'checker' gcc plugin in PaX:

1. the C address space infrastructure available since gcc 4.6 can be sort of
coerced into implementing the __user/__kernel/etc address spaces and it works
reasonably well (i'd say even better than sparse as it produces no false
positives in my experience and caught real bugs such as CVE-2014-0038).

2. __force itself presents a problem as its semantics isn't well defined and
only sparse knows how to model it. in gcc it cannot be an attribute as
attributes apply to the outermost variable/etc, e.g., you can't use them on a
pointee in a pointer context. what i did instead is that i introduced new
address spaces (__force_user/__force_kernel so far, __rcu/__iomem/etc will need
more of these) that replace the '__force something' combination with
__force_something (yes, this needs patching on the kernel side, and i haven't
done a thorough job of it but it works on my smaller configs at least). this
way the hijacked targetm.addr_space.legitimate_address_p callback can be taught
to allow/disallow the intended conversions.

3. designated_init is a tricky problem because by the time a plugin can examine
variable initializers, gcc will have lost the information. however with a trick
such unwanted initializers can instead be turned into a compile error (that
existing gcc infrastructure can detect). you can find it in spender's
randomize_layout plugin that's distributed in grsecurity.

4. as for maintaining a plugin for kernel and/or other use: inside the kernel
it'll need some kbuild infrastructure (there's one in PaX already, though it's
probably not 100% complete) and it's worked fine for our users for the past 3+
years now. for more  general use distros can package up plugins as they'd do
with any library (as plugins are really nothing more than that). note also that
keeping a plugin in the kernel tree will raise license problems (gplv2 vs
gplv3) but i guess the kernel list is the better forum for discussing that.