On Wed, Jul 27, 2016 at 11:17:23AM +0200, Jan Kara wrote:

On Tue 26-07-16 22:12:12, Kirill A. Shutemov wrote:

quoted

On Tue, Jul 26, 2016 at 01:29:38PM -0400, Theodore Ts'o wrote:

quoted

On Tue, Jul 26, 2016 at 03:35:02AM +0300, Kirill A. Shutemov wrote:

quoted

Here's the first version of my patchset which intended to bring huge pages
to ext4. It's not yet ready for applying or serious use, but good enough
to show the approach.

Thanks.  The major issues I noticed when doing a quick scan of the
patches you've already mentioned here.  I'll try to take a closer look
in the next week or so when I have time.

Thanks.

quoted

One random question --- in the huge=always approach, how much
additional work would be needed to support file systems with a 64k
block size on a system with 4k pages?

I think it's totally different story.

Here I have block size smaller than page size and it's not new to the
filesystem -- similar to 1k block size with 4k page size. So I was able to
re-use most of infrastructure to handle the situation.

Block size bigger than page size is backward task. I don't think I know
enough to understand how hard it would be. I guess not easy. :)

I think Ted wanted to ask: When you always have huge pages in page cache,
block size of 64k is smaller than the page size of the page cache so there
are chances it could work. Or is there anything which still exposes the
fact that actual pages are 4k even in huge=always case?

As usual with THP, if we failed to allocate huge page, we fallback to 4k
pages. It's normal situation to have both huge and small pages in the same
radix tree.

I guess you can get work 64k blocks with 4k pages if you *always* allocate
order-4 pages for page cache of the filesystem. But I don't think it's
sustainable. It's significant pressure on buddy allocator and compaction.

I guess the right approach would a mechanism to scatter one block to
multiple order-0 pages. At least for fallback.

-- 
 Kirill A. Shutemov