[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20180522190933.GA25904@redhat.com>
Date: Tue, 22 May 2018 15:09:34 -0400
From: Mike Snitzer <snitzer@...hat.com>
To: Christoph Hellwig <hch@...radead.org>
Cc: Kent Overstreet <kent.overstreet@...il.com>,
Jens Axboe <axboe@...nel.dk>, linux-kernel@...r.kernel.org,
linux-block@...r.kernel.org, colyli@...e.de,
darrick.wong@...cle.com, clm@...com, bacik@...com,
linux-xfs@...r.kernel.org, drbd-dev@...ts.linbit.com,
linux-btrfs@...r.kernel.org, linux-raid@...r.kernel.org,
neilb@...e.com
Subject: Re: [PATCH 00/13] convert block layer to bioset_init()/mempool_init()
On Tue, May 22 2018 at 2:41am -0400,
Christoph Hellwig <hch@...radead.org> wrote:
> On Mon, May 21, 2018 at 07:38:55PM -0400, Kent Overstreet wrote:
> > On Mon, May 21, 2018 at 02:24:32PM -0400, Mike Snitzer wrote:
> > > Every single data structure change in this series should be reviewed for
> > > unforeseen alignment consequences. Jens seemed to say that is
> > > worthwhile. Not sure if he'll do it or we divide it up. If we divide
> > > it up a temp topic branch should be published for others to inspect.
> > >
> > > Could be alignment hasn't been a historic concern for a bunch of the
> > > data structures changed in this series.. if so then all we can do is fix
> > > up any obvious potential for false sharing.
> >
> > Honestly, I almost never worry about alignment... the very few times I do care,
> > I use __cacheline_aligned_in_smp.
>
> And that generally is the right stratgey. If Mike really doesn't want
> a change we can just open code the kmalloc for the bio set there, the
> important point is that we should not keep the old API around for no
> good reason.
Again, I never said I didn't want these changes. I just thought it
worthwhile to mention the potential for alignment quirks arising.
Reality is false sharing is quite uncommon. So my concern was/is pretty
niche and unlikely to be applicable.
That said, I did take the opportunity to look at all the DM structures
that were modified. The bio_set and mempool_t structs are so large that
they span multiple cachelines as is. So I focused on eliminating
unnecessary spanning of cachelines (by non-bio_set and non-mempool_t
members) and eliminated most holes in DM structs. This is the result:
http://people.redhat.com/msnitzer/dm-4.18-struct-reorg/
Compare before.txt vs after_kent.txt vs after_mike.txt
Nothing staggering or special. Just something I'll add once Kent's
latest changes land.
But, I also eliminated 2 4-byte holes in struct bio_set, Jens please
feel free to pick this up (if you think it useful):
From: Mike Snitzer <snitzer@...hat.com>
Subject: [PATCH] block: eliminate 2 4-byte holes in bio_set structure
Signed-off-by: Mike Snitzer <snitzer@...hat.com>
---
include/linux/bio.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/bio.h b/include/linux/bio.h
index 5e472fc..e6fc692 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -735,7 +735,6 @@ static inline void bio_inc_remaining(struct bio *bio)
struct bio_set {
struct kmem_cache *bio_slab;
- unsigned int front_pad;
mempool_t bio_pool;
mempool_t bvec_pool;
@@ -743,6 +742,7 @@ struct bio_set {
mempool_t bio_integrity_pool;
mempool_t bvec_integrity_pool;
#endif
+ unsigned int front_pad;
/*
* Deadlock avoidance for stacking block drivers: see comments in
Powered by blists - more mailing lists