On 25.02.22 09:50, John Hubbard wrote:

Hi,

Summary:

This puts some prerequisites in place, including a CONFIG parameter,
making it possible to start converting and testing the Direct IO part of
each filesystem, from get_user_pages_fast(), to pin_user_pages_fast().

It will take "a few" kernel releases to get the whole thing done.

Details:

As part of fixing the "get_user_pages() + file-backed memory" problem
[1], and to support various COW-related fixes as well [2], we need to
convert the Direct IO code from get_user_pages_fast(), to
pin_user_pages_fast(). Because pin_user_pages*() calls require a
corresponding call to unpin_user_page(), the conversion is more
elaborate than just substitution.

Further complicating the conversion, the block/bio layers get their
Direct IO pages via iov_iter_get_pages() and iov_iter_get_pages_alloc(),
each of which has a large number of callers. All of those callers need
to be audited and changed so that they call unpin_user_page(), rather
than put_page().

vmsplice is another candidate that uses iov_iter_get_pages() and should
be converted to FOLL_PIN. For that particular user, we have to also pass
FOLL_LONGTERM -- vmsplice as it stands can block memory hotunplug / CMA
/ ... for all eternity.

-- 
Thanks,

David / dhildenb