On Wed, 9 Apr 2014 11:25:47 +0800 Shaohua Li <shli@xxxxxxxxxx> wrote:
> On Wed, Apr 09, 2014 at 12:08:08PM +1000, NeilBrown wrote:
> > On Tue, 8 Apr 2014 12:05:07 +0800 Shaohua Li <shli@xxxxxxxxxx> wrote:
> >
> > >
> > > In NUMA machine, prepare_to_wait/finish_wait in make_request exposes a lot of
> > > contention for sequential workload (or big request size workload). For such
> > > workload, each bio includes several stripes. So we can just do
> > > prepare_to_wait/finish_wait once for the whold bio instead of every stripe.
> > > This reduces the lock contention completely for such workload. Random workload
> > > might have the similar lock contention too, but I didn't see it yet, maybe
> > > because my stroage is still not fast enough.
> > >
> > > Signed-off-by: Shaohua Li <shli@xxxxxxxxxxxx>
> >
> > Thanks,
> > this looks every sensible, except .....
> >
> >
> > > ---
> > > drivers/md/raid5.c | 18 ++++++++++++++----
> > > 1 file changed, 14 insertions(+), 4 deletions(-)
> > >
> > > Index: linux/drivers/md/raid5.c
> > > ===================================================================
> > > --- linux.orig/drivers/md/raid5.c 2014-04-08 09:04:20.000000000 +0800
> > > +++ linux/drivers/md/raid5.c 2014-04-08 09:11:08.201533487 +0800
> > > @@ -4552,6 +4552,8 @@ static void make_request(struct mddev *m
> > > struct stripe_head *sh;
> > > const int rw = bio_data_dir(bi);
> > > int remaining;
> > > + DEFINE_WAIT(w);
> > > + bool do_prepare;
> > >
> > > if (unlikely(bi->bi_rw & REQ_FLUSH)) {
> > > md_flush_request(mddev, bi);
> > > @@ -4575,15 +4577,19 @@ static void make_request(struct mddev *m
> > > bi->bi_next = NULL;
> > > bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
> > >
> > > + prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
> > > for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
> > > DEFINE_WAIT(w);
> > ^^^^^^^^^^^^^^^
> >
> > Shouldn't this be removed? If so, please resubmit with that line deleted and
> > I'll apply the patch.
>
> Ah, that's silly, looks I sent wrong patch, sorry! Below is the correct one and I double
> checked it's the one working for me.
>
>
> Subject: raid5: make_request does less prepare wait
>
> In NUMA machine, prepare_to_wait/finish_wait in make_request exposes a lot of
> contention for sequential workload (or big request size workload). For such
> workload, each bio includes several stripes. So we can just do
> prepare_to_wait/finish_wait once for the whold bio instead of every stripe.
> This reduces the lock contention completely for such workload. Random workload
> might have the similar lock contention too, but I didn't see it yet, maybe
> because my stroage is still not fast enough.
>
> Signed-off-by: Shaohua Li <shli@xxxxxxxxxxxx>
> ---
> drivers/md/raid5.c | 19 ++++++++++++++-----
> 1 file changed, 14 insertions(+), 5 deletions(-)
>
> Index: linux/drivers/md/raid5.c
> ===================================================================
> --- linux.orig/drivers/md/raid5.c 2014-04-08 12:02:54.485630590 +0800
> +++ linux/drivers/md/raid5.c 2014-04-09 11:03:04.276210597 +0800
> @@ -4552,6 +4552,8 @@ static void make_request(struct mddev *m
> struct stripe_head *sh;
> const int rw = bio_data_dir(bi);
> int remaining;
> + DEFINE_WAIT(w);
> + bool do_prepare;
>
> if (unlikely(bi->bi_rw & REQ_FLUSH)) {
> md_flush_request(mddev, bi);
> @@ -4575,15 +4577,18 @@ static void make_request(struct mddev *m
> bi->bi_next = NULL;
> bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
>
> + prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
> for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
> - DEFINE_WAIT(w);
> int previous;
> int seq;
>
> + do_prepare = false;
> retry:
> seq = read_seqcount_begin(&conf->gen_lock);
> previous = 0;
> - prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
> + if (do_prepare)
> + prepare_to_wait(&conf->wait_for_overlap, &w,
> + TASK_UNINTERRUPTIBLE);
> if (unlikely(conf->reshape_progress != MaxSector)) {
> /* spinlock is needed as reshape_progress may be
> * 64bit on a 32bit platform, and so it might be
> @@ -4604,6 +4609,7 @@ static void make_request(struct mddev *m
> : logical_sector >= conf->reshape_safe) {
> spin_unlock_irq(&conf->device_lock);
> schedule();
> + do_prepare = true;
> goto retry;
> }
> }
> @@ -4640,6 +4646,7 @@ static void make_request(struct mddev *m
> if (must_retry) {
> release_stripe(sh);
> schedule();
> + do_prepare = true;
> goto retry;
> }
> }
> @@ -4663,8 +4670,10 @@ static void make_request(struct mddev *m
> prepare_to_wait(&conf->wait_for_overlap,
> &w, TASK_INTERRUPTIBLE);
> if (logical_sector >= mddev->suspend_lo &&
> - logical_sector < mddev->suspend_hi)
> + logical_sector < mddev->suspend_hi) {
> schedule();
> + do_prepare = true;
> + }
> goto retry;
> }
>
> @@ -4677,9 +4686,9 @@ static void make_request(struct mddev *m
> md_wakeup_thread(mddev->thread);
> release_stripe(sh);
> schedule();
> + do_prepare = true;
> goto retry;
> }
> - finish_wait(&conf->wait_for_overlap, &w);
> set_bit(STRIPE_HANDLE, &sh->state);
> clear_bit(STRIPE_DELAYED, &sh->state);
> if ((bi->bi_rw & REQ_SYNC) &&
> @@ -4689,10 +4698,10 @@ static void make_request(struct mddev *m
> } else {
> /* cannot get stripe for read-ahead, just give-up */
> clear_bit(BIO_UPTODATE, &bi->bi_flags);
> - finish_wait(&conf->wait_for_overlap, &w);
> break;
> }
> }
> + finish_wait(&conf->wait_for_overlap, &w);
>
> remaining = raid5_dec_bi_active_stripes(bi);
> if (remaining == 0) {
Applied, thanks.
NeilBrown
Attachment:
signature.asc
Description: PGP signature
