Re: [PATCH] block: call elv_bio_merged() when merged

[Vivek Goyal <vgoyal@redhat.com>]

On Fri, May 20, 2011 at 08:50:47PM +0200, Jens Axboe wrote:
> On 2011-05-20 11:51, Namhyung Kim wrote:
> > Hello,
> > 
> > 2011-05-20 (금), 16:31 +0800, Shaohua Li:
> >> 2011/5/20 Namhyung Kim <namhyung@gmail.com>:
> >>> Commit 73c101011926 ("block: initial patch for on-stack per-task plugging")
> >>> removed calls to elv_bio_merged() when @bio merged with @req. Re-add them.
> >>>
> >>> Signed-off-by: Namhyung Kim <namhyung@gmail.com>
> >>> Cc: Divyesh Shah <dpshah@google.com>
> >>> ---
> >>>  block/blk-core.c |    2 ++
> >>>  1 files changed, 2 insertions(+), 0 deletions(-)
> >>>
> >>> diff --git a/block/blk-core.c b/block/blk-core.c
> >>> index 3fe00a14822a..4dc02ef5fc82 100644
> >>> --- a/block/blk-core.c
> >>> +++ b/block/blk-core.c
> >>> @@ -1132,6 +1132,7 @@ static bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
> >>>        req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
> >>>
> >>>        drive_stat_acct(req, 0);
> >>> +       elv_bio_merged(q, req, bio);
> >>>        return true;
> >>>  }
> >>>
> >>> @@ -1173,6 +1174,7 @@ static bool bio_attempt_front_merge(struct request_queue *q,
> >>>        req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
> >>>
> >>>        drive_stat_acct(req, 0);
> >>> +       elv_bio_merged(q, req, bio);
> >>>        return true;
> >>>  }
> >> Looks you should do this in __make_request. when the routine is called
> >> in attempt_plug_merge, the request isn't added to elevator yet.
> >>
> > 
> > Hmm.. anyway it is merged. Is there any reason why we shouldn't collect
> > the stat - or invoke the callback routine - if the @req is not in the
> > elevator? Or we need to add a separate stat item for this case?
> 
> Your patch should be safe, it's essentially only for the cgroup stuff
> that does its own accounting and has appropriate protection for it. We'd
> want to do this for both the plug and not-plugged merge case.
> 
> It's a bit of a shame to add this though, since now we are hitting the
> cgroup lock for each merge.

I think wer can make MERGED per cpu as I have done in my other patch
series for following dispatch stats.

BLKIO_STAT_CPU_SECTORS,
/* Total bytes transferred */
BLKIO_STAT_CPU_SERVICE_BYTES,
/* Total IOs serviced, post merge */
BLKIO_STAT_CPU_SERVICED,

Jens are you planning to merge lockless throttling series? Once that gets
merged, we can convert this MERGED stat also per cpu and get rid of need of
taking blkg->stats_lock.

Thanks
Vivek