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.
