From: Mark Fortescue <redacted>
Date: Tue, 31 Jul 2007 08:55:20 +0100 (BST)

I have formulated a patch that prevents the update_mmu_cache from doing 
enything if there is no context available. This apears to have no 
immediate, undesirable side effects.

This worked better than the alternative of setting up a context to work with.

Can you for see any issues in doing this?

If not, can you check+apply the attached (un-mangled) patch.

Thanks for tracking this down Mark.

The issue is that, when exec()'ing to userspace from a kernel thread,
we need activate_context() to be invoked before we try to touch
userspace at all.  This new argument handling is invoking
get_user_pages() before that happens.

activate_context() happens via flush_old_exec(), but that occurs via
load_elf_binary() et al. which is long after the argument fetching
code runs in fs/exec.c that is using get_user_pages().

(Mark, hint: activate_context() is defined to switch_mm() on
 sparc32, which is sun4c_switch_mm() which you thought was only
 invoked from context switches :-))

Touching userspace before activate_context() is questionable at best,
in my opinion.  But I can't come up with a good way to fix this right
now other than Mark's sparc patch, so I will apply it.

Thanks again Mark!