On Wed, Nov 08, 2023 at 05:45:19PM +0530, Kanchan Joshi wrote:
> On 11/7/2023 8:38 PM, Keith Busch wrote:
> > On Tue, Nov 07, 2023 at 03:55:14PM +0530, Kanchan Joshi wrote:
> >> On 11/6/2023 8:32 PM, Keith Busch wrote:
> >>> On Mon, Nov 06, 2023 at 11:18:03AM +0530, Kanchan Joshi wrote:
> >>>> On 10/27/2023 11:49 PM, Keith Busch wrote:
> >>>>> + for (i = 0; i < nr_vecs; i = j) {
> >>>>> + size_t size = min_t(size_t, bytes, PAGE_SIZE - offs);
> >>>>> + struct folio *folio = page_folio(pages[i]);
> >>>>> +
> >>>>> + bytes -= size;
> >>>>> + for (j = i + 1; j < nr_vecs; j++) {
> >>>>> + size_t next = min_t(size_t, PAGE_SIZE, bytes);
> >>>>> +
> >>>>> + if (page_folio(pages[j]) != folio ||
> >>>>> + pages[j] != pages[j - 1] + 1)
> >>>>> + break;
> >>>>> + unpin_user_page(pages[j]);
> >>>>
> >>>> Is this unpin correct here?
> >>>
> >>> Should be. The pages are bound to the folio, so this doesn't really
> >>> unpin the user page. It just drops a reference, and the folio holds the
> >>> final reference to the contiguous pages, which is released on
> >>> completion.
> >>
> >> But the completion is still going to see multiple pages and not one
> >> (folio). The bip_for_each_vec loop is going to drop the reference again.
> >> I suspect it is not folio-aware.
> >
> > The completion unpins once per bvec, not individual pages. The setup
> > creates multipage bvecs with only one pin remaining per bvec for all of
> > the bvec's pages. If a page can't be merged into the current bvec, then
> > that page is not unpinned and becomes the first page of to the next
> > bvec.
> >
>
> Here is a test program [2] that creates this scenario.
> Single 8KB+16b read on a 4KB+8b formatted namespace. It prepares
> meta-buffer out of a huge-page in a way that it spans two regular 4K pages.
> With this, I see more unpins than expected.
I understand now. The bip_for_each_bvec is using single page vector
iterators. Will fix it up for next time.
