On 8/27/22 17:12, Andrew Morton wrote:

On Sat, 27 Aug 2022 16:59:32 -0700 John Hubbard [off-list ref] wrote:

quoted

Anyway, I'll change my patch locally for now, to this:

static inline void dio_w_unpin_user_pages(struct page **pages,
					  unsigned long npages)
{
	/* Careful, release_pages() uses a smaller integer type for npages: */
	if (WARN_ON_ONCE(npages > (unsigned long)INT_MAX))
		return;

	release_pages(pages, (int)npages);
}

Well, it might be slower.  release_pages() has a ton of fluff.

As mentioned, the above might be faster if the pages tend
to have page_count()==1??

I don't think we can know the answer to that. This code is called
in all kinds of situations, and it seems to me that whatever
tradeoffs are best for release_pages(), are probably also reasonable
for this code.

Even with all the fluff in release_pages(), it at least batches
the pages, as opposed to the simple put_page() in a loop. Most of 
the callers do have more than one page to release in non-error cases,
so release_pages() does seem better.


thanks,

-- 
John Hubbard
NVIDIA