On 1/26/23 14:36, Al Viro wrote:
...
+static inline bool iov_iter_extract_will_pin(const struct iov_iter *iter)
+{
+ return user_backed_iter(iter);
+}
+
Wait a sec; why would we want a pin for pages we won't be modifying?
A reference - sure, but...
After having looked through the earlier iterations of the patchset -
sorry, but that won't fly for (at least) vmsplice(). There we can't
pin those suckers; thankfully, we don't need to - they are used only
for fetches, so FOLL_GET is sufficient. With your "we'll just pin them,
source or destination" you won't be able to convert at least that
call of iov_iter_get_pages2(). And there might be other similar cases;
I won't swear there's more, but ISTR running into more than one of
the "pin won't be OK here, but fortunately it's a data source" places.
Assuming that "page is a data source" means that we are writing out from
the page to a block device (so, a WRITE operation, which of course
actually *reads* from the page), then...
...one thing I'm worried about now is whether Jan's original problem
report [1] can be fixed, because that involves page writeback. And it
seems like we need to mark the pages involved as "maybe dma-pinned" via
FOLL_PIN pins, in order to solve it.
Or am I missing a key point (I hope)?
[1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@xxxxxxxxxxxxxx/T/#u
thanks,
--
John Hubbard
NVIDIA
