On Tue, Apr 09, 2019 at 11:04:18AM +0800, Huang Shijie wrote:
> On Mon, Apr 08, 2019 at 07:49:29PM -0700, Matthew Wilcox wrote:
> > On Tue, Apr 09, 2019 at 09:08:33AM +0800, Huang Shijie wrote:
> > > On Mon, Apr 08, 2019 at 07:13:13AM -0700, Matthew Wilcox wrote:
> > > > On Mon, Apr 08, 2019 at 10:37:45AM +0800, Huang Shijie wrote:
> > > > > The root cause is that sg_alloc_table_from_pages() requires the
> > > > > page order to keep the same as it used in the user space, but
> > > > > get_user_pages_fast() will mess it up.
> > > >
> > > > I don't understand how get_user_pages_fast() can return the pages in a
> > > > different order in the array from the order they appear in userspace.
> > > > Can you explain?
> > > Please see the code in gup.c:
> > >
> > > int get_user_pages_fast(unsigned long start, int nr_pages,
> > > unsigned int gup_flags, struct page **pages)
> > > {
> > > .......
> > > if (gup_fast_permitted(start, nr_pages)) {
> > > local_irq_disable();
> > > gup_pgd_range(addr, end, gup_flags, pages, &nr); // The @pages array maybe filled at the first time.
> >
> > Right ... but if it's not filled entirely, it will be filled part-way,
> > and then we stop.
> >
> > > local_irq_enable();
> > > ret = nr;
> > > }
> > > .......
> > > if (nr < nr_pages) {
> > > /* Try to get the remaining pages with get_user_pages */
> > > start += nr << PAGE_SHIFT;
> > > pages += nr; // The @pages is moved forward.
> >
> > Yes, to the point where gup_pgd_range() stopped.
> >
> > > if (gup_flags & FOLL_LONGTERM) {
> > > down_read(¤t->mm->mmap_sem);
> > > ret = __gup_longterm_locked(current, current->mm, // The @pages maybe filled at the second time
> >
> > Right.
> >
> > > /*
> > > * retain FAULT_FOLL_ALLOW_RETRY optimization if
> > > * possible
> > > */
> > > ret = get_user_pages_unlocked(start, nr_pages - nr, // The @pages maybe filled at the second time.
> > > pages, gup_flags);
> >
> > Yes. But they'll be in the same order.
> >
> > > BTW, I do not know why we mess up the page order. It maybe used in some special case.
> >
> > I'm not discounting the possibility that you've found a bug.
> > But documenting that a bug exists is not the solution; the solution is
> > fixing the bug.
> I do not think it is a bug :)
>
> If we use the get_user_pages_unlocked(), DMA is okay, such as:
> ....
> get_user_pages_unlocked()
> sg_alloc_table_from_pages()
> .....
>
> I think the comment is not accurate enough. So just add more comments, and tell the driver
> users how to use the GUPs.
gup_fast() and gup_unlocked() should return the pages in the same order.
If they do not, then it is a bug.
