Re: [PATCH] bcache: dont reset bio opf in bch_data_insert_start

[Dongsheng Yang]

在 2021/1/26 星期二 下午 12:34, Coly Li 写道:
> On 1/26/21 12:32 PM, Dongsheng Yang wrote:
> > 在 2021/1/25 星期一 下午 12:53, Coly Li 写道:
> > > On 1/25/21 12:29 PM, Dongsheng Yang wrote:
> > > > commit ad0d9e76(bcache: use bio op accessors) makes the bi_opf
> > > > modified by bio_set_op_attrs(). But there is a logical
> > > > problem in this commit:
> > > >
> > > >                   trace_bcache_cache_insert(k);
> > > >                   bch_keylist_push(&op->insert_keys);
> > > >
> > > > -               n->bi_rw |= REQ_WRITE;
> > > > +               bio_set_op_attrs(n, REQ_OP_WRITE, 0);
> > > >                   bch_submit_bbio(n, op->c, k, 0);
> > > >           } while (n != bio);
> > > >
> > > > The old code add REQ_WRITE into bio n and keep other flags; the
> > > > new code set REQ_OP_WRITE to bi_opf, but reset all other flags.
> > > >
> > > > This problem is discoverd in our performance testing:
> > > > (1) start a fio with 1M x 128depth for read in /dev/nvme0n1p1
> > > > (2) start a fio with 1M x 128depth for write in /dev/escache0 (cache
> > > > device is /dev/nvme0n1p2)
> > > >
> > > > We found the BW of reading is 2000+M/s, but the BW of writing is
> > > > 0-100M/s. After some debugging, we found the problem is io submit in
> > > > writting is very slow.
> > > >
> > > > bch_data_insert_start() insert a bio to /dev/nvme0n1p1, but as
> > > > cached_dev submit stack bio will be added into current->bio_list, and
> > > > return.Then __submit_bio_noacct() will submit the new bio in bio_list
> > > > into /dev/nvme0n1p1. This operation would be slow in
> > > > blk_mq_submit_bio() -> rq_qos_throttle(q, bio);
> > > >
> > > > The rq_qos_throttle() will call wbt_should_throttle(),
> > > > static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio
> > > > *bio)
> > > > {
> > > >           switch (bio_op(bio)) {
> > > >           case REQ_OP_WRITE:
> > > >                   /*
> > > >                    * Don't throttle WRITE_ODIRECT
> > > >                    */
> > > >                   if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
> > > >                       (REQ_SYNC | REQ_IDLE))
> > > >                           return false;
> > > > ... ...
> > > > }
> > > >
> > > > As the bio_set_op_attrs() reset the (REQ_SYNC | REQ_IDLE), so this write
> > > > bio will be considered as non-direct write.
> > > >
> > > > After this fix, bio to nvme will flaged as (REQ_SYNC | REQ_IDLE),
> > > > then fio for writing will get about 1000M/s bandwidth.
> > > >
> > > > Fixes: ad0d9e76a4124708dddd00c04fc4b56fc86c02d6
> > > It should be,
> > > Fixes: ad0d9e76a412 ("bcache: use bio op accessors")
> > >
> > > > Signed-off-by: Dongsheng Yang<dongsheng.yang@xxxxxxxxxxxx>
> > > Please CC the fixed patch author  Mike Christie<mchristi@xxxxxxxxxx>.
> >
> > Hi Coly,
> >
> >      Should I send a V2 for commit message update?
> >
> > Or you can help to fix it when you take it from maillist?
> Yes, please fix it in v2 version. And let's wait for response from Mike,
> maybe he has better suggestion to fix.


okey，actually, Mike is in my cc list of first mail (but not note in 
commit message), so he can receive my patch.

But anyway, I will send a v2

> Thanks.
>
> Coly Li