On Thu, Dec 14, 2017 at 02:14:00PM -0800, Linus Torvalds wrote: > On Thu, Dec 14, 2017 at 2:02 PM, Peter Zijlstra <pet...@infradead.org> wrote: > > > > _Should_ being the operative word, because I cannot currently see it > > DTRT. But maybe I'm missing the obvious -- I tend to do that at times. > > At least the old get_user_pages_fast() code used to check the USER bit: > > unsigned long need_pte_bits = _PAGE_PRESENT|_PAGE_USER; > > if (write) > need_pte_bits |= _PAGE_RW; > > but that may have been lost when we converted over to the generic code.
The generic gup_pte_range() has pte_access_permitted() (which has the above test) in the right place. > It shouldn't actually _matter_, since we'd need to change access_ok() > anyway (and gup had better check that!) get_user_pages_fast() (both of them) do indeed test access_ok(), but the regular get_user_pages() does not, I suspect because it can operate on a foreign mm. And its the regular old get_user_pages() that's all sorts of broken wrt !PAGE_USER too.
