On Thu, May 26, 2016 at 04:04:55PM +0900, Minchan Kim wrote:
> On Wed, May 25, 2016 at 01:30:11PM +0300, Vladimir Davydov wrote:
> > On Tue, May 24, 2016 at 01:04:33PM -0700, Eric Dumazet wrote:
> > > On Tue, 2016-05-24 at 19:13 +0300, Vladimir Davydov wrote:
> > > > On Tue, May 24, 2016 at 05:59:02AM -0700, Eric Dumazet wrote:
> > > > ...
> > > > > > +static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
> > > > > > + struct pipe_buffer *buf)
> > > > > > +{
> > > > > > + struct page *page = buf->page;
> > > > > > +
> > > > > > + if (page_count(page) == 1) {
> > > > >
> > > > > This looks racy : some cpu could have temporarily elevated page count.
> > > >
> > > > All pipe operations (pipe_buf_operations->get, ->release, ->steal) are
> > > > supposed to be called under pipe_lock. So, if we see a pipe_buffer->page
> > > > with refcount of 1 in ->steal, that means that we are the only its user
> > > > and it can't be spliced to another pipe.
> > > >
> > > > In fact, I just copied the code from generic_pipe_buf_steal, adding
> > > > kmemcg related checks along the way, so it should be fine.
> > >
> > > So you guarantee that no other cpu might have done
> > > get_page_unless_zero() right before this test ?
> >
> > Each pipe_buffer holds a reference to its page. If we find page's
> > refcount to be 1 here, then it can be referenced only by our
> > pipe_buffer. And the refcount cannot be increased by a parallel thread,
> > because we hold pipe_lock, which rules out splice, and otherwise it's
> > impossible to reach the page as it is not on lru. That said, I think I
> > guarantee that this should be safe.
>
> I don't know kmemcg internal and pipe stuff so my comment might be
> totally crap.
>
> No one cannot guarantee any CPU cannot held a reference of a page.
> Look at get_page_unless_zero usecases.
>
> 1. balloon_page_isolate
>
> It can hold a reference in random page and then verify the page
> is balloon page. Otherwise, just put.
>
> 2. page_idle_get_page
>
> It has PageLRU check but it's racy so it can hold a reference
> of randome page and then verify within zone->lru_lock. If it's
> not LRU page, just put.
Well, I see your concern now - even if a page is not on lru and we
locked all structs pointing to it, it can always get accessed by pfn in
a completely unrelated thread, like in examples you gave above. That's a
fair point.
However, I still think that it's OK in case of pipe buffers. What can
happen if somebody takes a transient reference to a pipe buffer page? At
worst, we'll see page_count > 1 due to temporary ref and abort stealing,
falling back on copying instead. That's OK, because stealing is not
guaranteed. Can a function that takes a transient ref to page by pfn
mistakenly assume that this is a page it's interested in? I don't think
so, because this page has no marks on it except special _mapcount value,
which should only be set on kmemcg pages.
Thanks,
Vladimir
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
