kmap() while holding spinlock

[Anton Blanchard]

Hi,

I was debugging a lockup on sparc32 where we took a page_table_lock
in handle_mm_fault and then tried to take the same one again in
swap_out_mm. So it looks like we are scheduling somewhere inside
handle_mm_fault while holding the page_table_lock.

This machine might have highmem (Im awaiting more information). If
so copy_user_highpage and clear_user_highpage could sleep if we
run out of kmap entries. There are two problems:

do_anonymous_page calls clear_user_highpage with the page_table_lock held.

do_wp_page calls break_cow with the page_table_lock held.

Since I dont think we can drop the lock, do we need a kmap_atomic for
these?

Anton