On 1/23/23 05:24, David Hildenbrand wrote:

On 23.01.23 14:19, David Howells wrote:

quoted

David Hildenbrand [off-list ref] wrote:

quoted

Switching from FOLL_GET to FOLL_PIN was in the works by John H. Not sure what
the status is. Interestingly, Documentation/core-api/pin_user_pages.rst
already documents that "CASE 1: Direct IO (DIO)" uses FOLL_PIN ... which does,
unfortunately, no reflect reality yet.

Yes, that part of the documentation is...aspirational. :) But this
series is taking us there, so good. Let me go review v8 of the series in
actual detail to verify but it sounds very promising.

quoted

Yeah - I just came across that.

Should iov_iter.c then switch entirely to using pin_user_pages(), rather than
get_user_pages()?  In which case my patches only need keep track of
pinned/not-pinned and never "got".

That would be the ideal case: whenever intending to access page content, use FOLL_PIN instead of FOLL_GET.

The issue that John was trying to sort out was that there are plenty of callsites that do a simple put_page() instead of calling unpin_user_page(). IIRC, handling that correctly in existing code -- what was pinned must be released via unpin_user_page() -- was the biggest workitem.

Not sure how that relates to your work here (that's why I was asking): if you could avoid FOLL_GET, that would be great :)

The largest part of this problem has been: __iov_iter_get_pages_alloc()
calls get_user_pages_fast() (so, FOLL_GET), as part of the Direct IO
path. And that __iov_iter_get_pages_alloc() is also called by a wide
variety of things that are not Direct IO: networking, crytpo, RDS.

So splitting out a variant that only Direct IO uses is a great move and
should allow conversion of Direct IO to FOLL_PIN. Again, let me go do an
actual review to check on that.

thanks,
-- 
John Hubbard
NVIDIA