Message-ID: <1484220421.2970.20.camel@redhat.com>
Date:   Thu, 12 Jan 2017 06:27:01 -0500
From:   Jeff Layton <jlayton@...hat.com>
To:     Al Viro <viro@...IV.linux.org.uk>
Cc:     "Yan, Zheng" <zyan@...hat.com>, Sage Weil <sage@...hat.com>,
        Ilya Dryomov <idryomov@...il.com>, ceph-devel@...r.kernel.org,
        linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
        "Zhu, Caifeng" <zhucaifeng@...ssoft-nj.com>
Subject: Re: [PATCH v2] ceph/iov_iter: fix bad iov_iter handling in ceph
 splice codepaths

On Thu, 2017-01-12 at 07:59 +0000, Al Viro wrote:
> On Tue, Jan 10, 2017 at 07:57:31AM -0500, Jeff Layton wrote:
> > 
> > v2: fix bug in offset handling in iov_iter_pvec_size
> > 
> > xfstest generic/095 triggers soft lockups in kcephfs. Basically it uses
> > fio to drive some I/O via vmsplice ane splice. Ceph then ends up trying
> > to access an ITER_BVEC type iov_iter as a ITER_IOVEC one. That causes it
> > to pick up a wrong offset and get stuck in an infinite loop while trying
> > to populate the page array. dio_get_pagev_size has a similar problem.
> > 
> > To fix the first problem, add a new iov_iter helper to determine the
> > offset into the page for the current segment and have ceph call that.
> > I would just replace dio_get_pages_alloc with iov_iter_get_pages_alloc,
> > but that will only return a single page at a time for ITER_BVEC and
> > it's better to make larger requests when possible.
> > 
> > For the second problem, we simply replace it with a new helper that does
> > what it does, but properly for all iov_iter types.
> > 
> > Since we're moving that into generic code, we can also utilize the
> > iterate_all_kinds macro to simplify this. That means that we need to
> > rework the logic a bit since we can't advance to the next vector while
> > checking the current one.
> 
> Yecchhh...  That really looks like exposing way too low-level stuff instead
> of coming up with saner primitive ;-/
> 

Fair point. That said, I'm not terribly thrilled with how
iov_iter_get_pages* works right now.

Note that it only ever touches the first vector. Would it not be better
to keep getting page references if the bvec/iov elements are aligned
properly? It seems quite plausible that they often would be, and being
able to hand back a larger list of pages in most cases would be
advantageous.

IOW, should we have iov_iter_get_pages basically do what
dio_get_pages_alloc does -- try to build as long an array of pages as
possible before returning, provided that the alignment works out?

The NFS DIO code, for instance, could also benefit there. I know we've
had reports there in the past that sending down a bunch of small iovecs
causes a lot of small-sized requests on the wire.

> Is page vector + offset in the first page + number of bytes really what
> ceph wants?  Would e.g. an array of bio_vec be saner?  Because _that_
> would make a lot more natural iov_iter_get_pages_alloc() analogue...
> 
> And yes, I realize that you have ->pages wired into the struct ceph_osd_request;
> how painful would it be to have it switched to struct bio_vec array instead?

Actually...it looks like that might not be too hard. The low-level OSD
handling code can already handle bio_vec arrays in order to service RBD.
It looks like we could switch cephfs to use
osd_req_op_extent_osd_data_bio instead of
osd_req_op_extent_osd_data_pages. That would add a dependency in cephfs
on CONFIG_BLOCK, but I think we could probably live with that.

-- 
Jeff Layton <jlayton@...hat.com>

Powered by blists - more mailing lists