[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 2 Feb 2017 02:56:51 -0800
From: Christoph Hellwig <hch@...radead.org>
To: Al Viro <viro@...IV.linux.org.uk>
Cc: Jeff Layton <jlayton@...hat.com>, linux-fsdevel@...r.kernel.org,
linux-kernel@...r.kernel.org, linux-nfs@...r.kernel.org,
ceph-devel@...r.kernel.org, lustre-devel@...ts.lustre.org,
v9fs-developer@...ts.sourceforge.net,
Linus Torvalds <torvalds@...ux-foundation.org>,
Jan Kara <jack@...e.cz>,
Chris Wilson <chris@...is-wilson.co.uk>,
"Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>
Subject: Re: [PATCH v3 0/2] iov_iter: allow iov_iter_get_pages_alloc to
allocate more pages per call
On Thu, Feb 02, 2017 at 09:51:25AM +0000, Al Viro wrote:
> On Wed, Jan 25, 2017 at 08:32:03AM -0500, Jeff Layton wrote:
> > Small respin of the patch that I sent yesterday for the same thing.
> >
> > This moves the maxsize handling into iov_iter_pvec_size, so that we don't
> > end up iterating past the max size we'll use anyway when trying to
> > determine the pagevec length.
> >
> > Also, a respun patch to make ceph use iov_iter_get_pages_alloc instead of
> > trying to do it via its own routine.
> >
> > Al, if these look ok, do you want to pick these up or shall I ask
> > Ilya to merge them via the ceph tree?
>
> I'd rather have that kind of work go through the vfs tree; said that,
> I really wonder if this is the right approach. Most of the users of
> iov_iter_get_pages()/iov_iter_get_pages_alloc() look like they want
> something like
> iov_iter_for_each_page(iter, size, f, data)
> with int (*f)(struct page *page, size_t from, size_t size, void *data)
> passed as callback. Not everything fits that model, but there's a whole
> lot of things that do.
I was planning to do that, mostly because of the iomap dio code that
would not only get a lot cleaner with this, but also support multi-page
bvecs that we hope to have in the block layer soon. The issue with it
is that we need to touch all the arch get_user_pages_fast
implementations, so it's going to be a relatively invasive change that I
didn't want to fix with just introducing the new direct I/O code.
Powered by blists - more mailing lists