Andrea Arcangeli wrote:
>
> Here the latest patch. Now that I consider the 'production' trouble
> closed I'll be porting it to mainline while addressing gup-fast too
> which is an additional complication I didn't have to take care of
> yet. So expect a patch that works for you in the next few days, either
> that or complains about an unfixable gup-fast ;). But frankly I've
> been thinking it should be possible in a much simpler way that I
> ever thought before, by entirely relaying on the tlb flush.
>
> In short if I simply do the page-count vs page-mapcount check _after_
> ptep_set_wrprotect (which implies a tlb flush and after that any
> gup-fast running in other cpus should be forced through the slow path
> and block) I think I'm done. The code now does:
>
>     check mapcount
>     ptep_set_wrprotect
>
> I think to make the thing working with gup-fast I've only to
> ptep_set_wrprotect before the mapcount check.
>
> The reason why the normal pagetable walking of the cpu is ok with
> current code is that ptep_set_wrprotect done later will block any
> access to the page from the other cpus. Not so if it was gup-fast
> taking reference on the page. So we need to stop with
> ptep_set_wrprotect any subsequential gup-fast _before_ we check count
> vs mapcount and the fact the get_page is run inside the critical
> section with local irq disabled in gup-fast should close the race for
> gup-fast too. Will try that ASAP...
>
> 
Hi Andrea,
Good approach, but I think that it isn't enough to just change the order 
of check mapcount and ptep_set_wrprotect(),
the reason is that gup_fast do the get_page() AFTER it fetched the pte, 
so 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()

Anyway sound like a great idea to fix this issue!
good work.