Re: open(2) says O_DIRECT works on 512 byte boundries?

[Andrea Arcangeli <aarcange@redhat.com>]

On Sat, Feb 07, 2009 at 03:32:03PM +0200, Izik Eidus wrote:
> we are opening here a tiny race:
>
> cpu#1 do get_user_pages_fast and fetch the pte (it think the pte is 
> writeable)
> cpu#2 do ptep_set_wrprotect()
> cpu#2 check the mapcount against pagecount (it think that everything is 
> fine and continue)
> cpu#1 only now do get_page()
>
> Anyway this is minor issue that can be probably solved by just:
> rechecking if the pte isnt read_only in gup_fast after we do the get_page()

Not needed, if I check page_count vs mapcount after marking the pte
readonly and after sending smp-tlb-flush there is no race.

> Anyway sound like a great idea to fix this issue!

The only problem I'm thinking now is the IPI flood that would be
generated if I send IPIs for every pte wrprotected in fork, that
sounds overkill. So to use the IPI fix I could have gup-fast take the
slow path first time around if PG_gup isn't set, and then only second
time take the lockless fast path when PG_gup is already set (PG_gup
gets set by follow_page under PT lock/mmap_sem read mode at
least). And fork/ksm would only send IPIs for PG_gup pages. However
that would make gup-fast slow the first time it runs on an
anonymous/hugetlb page with pte marked writeable and I'm unsure if
that's ok.

Otherwise we've to return to the plan of the slightly more complicated
fix.