On Thu, Dec 3, 2020 at 11:59 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Thu, Dec 03, 2020 at 11:40:15AM -0500, Pavel Tatashin wrote:
> > > Looking at this code some more.. How is it even correct?
> > >
> > > 1633 if (!isolate_lru_page(head)) {
> > > 1634 list_add_tail(&head->lru, &cma_page_list);
> > >
> > > Here we are only running under the read side of the mmap sem so multiple
> > > GUPs can be calling that sequence in parallel. I don't see any
> > > obvious exclusion that will prevent corruption of head->lru. The first
> > > GUP thread to do isolate_lru_page() will ClearPageLRU() and the second
> > > GUP thread will be a NOP for isolate_lru_page().
> > >
> > > They will both race list_add_tail and other list ops. That is not OK.
> >
> > Good question. I studied it, and I do not see how this is OK. Worse,
> > this race is also exposable as a syscall instead of via driver: two
> > move_pages() run simultaneously. Perhaps in other places?
> >
> > move_pages()
> > kernel_move_pages()
> > mmget()
> > do_pages_move()
> > add_page_for_migratio()
> > mmap_read_lock(mm);
> > list_add_tail(&head->lru, pagelist); <- Not protected
>
> When this was CMA only it might have been rarer to trigger, but this
> move stuff sounds like it makes it much more broadly, eg on typical
> servers with RDMA exposed/etc
>
> Seems like it needs fixing as part of this too :\
Just to clarify the stack that I showed above is outside of gup, it is
the same issue that you pointed out that happens elsewhere. I suspect
there might be more. All of them should be addressed together.
Pasha
>
> Page at a time inside the gup loop could address both concerns, unsure
> about batching performance here though..
>
> Jason
