On Thu, May 30, 2019 at 04:21:19PM -0700, John Hubbard 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.
FWIW I don't see this being a huge issue at the moment. Perhaps those more
familiar with CMA can weigh in here. How was this issue found? If it was
found by running some test perhaps that indicates a performance preference?
>
> >
> >> + 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);
Yes this is cleaner.
> + 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;
To be absolutely pedantic I would have split the nr_pinned change to a separate
patch.
Ira
>
> 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. :)
>
> thanks,
> --
> John Hubbard
> NVIDIA
