Re: [PATCH] mm/gup: fix omission of check on FOLL_LONGTERM in get_user_pages_fast()

From: Pingfan Liu
Date: Fri May 31 2019 - 07:09:22 EST


On Fri, May 31, 2019 at 7:21 AM John Hubbard <jhubbard@xxxxxxxxxx> wrote:
>
> On 5/30/19 2:47 PM, Ira Weiny wrote:
> > On Thu, May 30, 2019 at 06:54:04AM +0800, Pingfan Liu wrote:
> [...]
> >> + for (j = i; j < nr; j++)
> >> + put_page(pages[j]);
> >
> > Should be put_user_page() now. For now that just calls put_page() but it is
> > slated to change soon.
> >
> > I also wonder if this would be more efficient as a check as we are walking the
> > page tables and bail early.
> >
> > Perhaps the code complexity is not worth it?
>
> Good point, it might be worth it. Because now we've got two loops that
> we run, after the interrupts-off page walk, and it's starting to look like
> a potential performance concern.
>
> >
> >> + nr = i;
> >
> > Why not just break from the loop here?
> >
> > Or better yet just use 'i' in the inner loop...
> >
>
> ...but if you do end up putting in the after-the-fact check, then we can
> go one or two steps further in cleaning it up, by:
>
> * hiding the visible #ifdef that was slicing up gup_fast,
>
> * using put_user_pages() instead of either put_page or put_user_page,
> thus getting rid of j entirely, and
>
> * renaming an ancient minor confusion: nr --> nr_pinned),
>
> we could have this, which is looks cleaner and still does the same thing:
>
> diff --git a/mm/gup.c b/mm/gup.c
> index f173fcbaf1b2..0c1f36be1863 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1486,6 +1486,33 @@ static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
> }
> #endif /* CONFIG_FS_DAX || CONFIG_CMA */
>
> +#ifdef CONFIG_CMA
> +/*
> + * Returns the number of pages that were *not* rejected. This makes it
> + * exactly compatible with its callers.
> + */
> +static int reject_cma_pages(int nr_pinned, unsigned gup_flags,
> + struct page **pages)
> +{
> + int i = 0;
> + if (unlikely(gup_flags & FOLL_LONGTERM)) {
> +
> + for (i = 0; i < nr_pinned; i++)
> + if (is_migrate_cma_page(pages[i])) {
> + put_user_pages(&pages[i], nr_pinned - i);
> + break;
> + }
> + }
> + return i;
> +}
> +#else
> +static int reject_cma_pages(int nr_pinned, unsigned gup_flags,
> + struct page **pages)
> +{
> + return nr_pinned;
> +}
> +#endif
> +
> /*
> * This is the same as get_user_pages_remote(), just with a
> * less-flexible calling convention where we assume that the task
> @@ -2216,7 +2243,7 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
> unsigned int gup_flags, struct page **pages)
> {
> unsigned long addr, len, end;
> - int nr = 0, ret = 0;
> + int nr_pinned = 0, ret = 0;
>
> start &= PAGE_MASK;
> addr = start;
> @@ -2231,25 +2258,27 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
>
> if (gup_fast_permitted(start, nr_pages)) {
> local_irq_disable();
> - gup_pgd_range(addr, end, gup_flags, pages, &nr);
> + gup_pgd_range(addr, end, gup_flags, pages, &nr_pinned);
> local_irq_enable();
> - ret = nr;
> + ret = nr_pinned;
> }
>
> - if (nr < nr_pages) {
> + nr_pinned = reject_cma_pages(nr_pinned, gup_flags, pages);
> +
> + if (nr_pinned < nr_pages) {
> /* Try to get the remaining pages with get_user_pages */
> - start += nr << PAGE_SHIFT;
> - pages += nr;
> + start += nr_pinned << PAGE_SHIFT;
> + pages += nr_pinned;
>
> - ret = __gup_longterm_unlocked(start, nr_pages - nr,
> + ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
> gup_flags, pages);
>
> /* Have to be a bit careful with return values */
> - if (nr > 0) {
> + if (nr_pinned > 0) {
> if (ret < 0)
> - ret = nr;
> + ret = nr_pinned;
> else
> - ret += nr;
> + ret += nr_pinned;
> }
> }
>
>
> Rather lightly tested...I've compile-tested with CONFIG_CMA and !CONFIG_CMA,
> and boot tested with CONFIG_CMA, but could use a second set of eyes on whether
> I've added any off-by-one errors, or worse. :)
>
Do you mind I send V2 based on your above patch? Anyway, it is a simple bug fix.

Thanks,
Pingfan