lkml.org 
[lkml]   [2013]   [Jan]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] blkcg: Before starting a new slice, firstly count bps/iops limit in func tg_may_dispatch.
On Fri, Jan 11, 2013 at 03:26:37PM +0100, Jens Axboe wrote:
> On 2013-01-11 10:11, majianpeng wrote:
> > In func tg_may_dispatch,
> >> if (throtl_slice_used(td, tg, rw))
> >> throtl_start_new_slice(td, tg, rw);
> > ...
> >> if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
> >> && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
> >
> > In funcs tg_with_in_(bps/iops)_limit, it used the slice_start to count.
> > So if privious slice expired, it start a new slice.This can cause hung
> > task.
> >
> > The next steps can repeat this bug.
> > 1:echo "8:48 10240" > blkio.throttle.write_bps_devic
> > 2:dd if=/dev/zero of=/dev/sdd bs=1M count=1 oflag=direct
> >
> > Using the blktrace, the messages about throttle:
> > root@kernel:/mnt/programs# blktrace -d /dev/sdd -a notify -o -|blkparse -i -
> > 8,48 1 0 0.000000000 0 m N throtl / [W] new slice start=4294854679 end=4294854779 jiffies=4294854679
> > 8,48 1 0 0.000000966 0 m N throtl / [W] extend slice start=4294854679 end=4294905900 jiffies=4294854679
> > 8,48 1 0 0.000002553 0 m N throtl / [W] bio. bdisp=0 sz=524288 bps=10240 iodisp=0 iops=4294967295 queued=0/0
> > 8,48 1 0 0.000004788 0 m N throtl schedule work. delay=51200 jiffies=4294854679
> > 8,48 1 0 51.304698681 0 m N throtl dispatch nr_queued=1 read=0 write=1
> > 8,48 1 0 51.304701979 0 m N throtl / [W] new slice start=4294905984 end=4294906084 jiffies=4294905984
> > 8,48 1 0 51.304703329 0 m N throtl / [W] extend slice start=4294905984 end=4294957200 jiffies=4294905984
> > 8,48 1 0 51.304705783 0 m N throtl schedule work. delay=51200 jiffies=4294905984
> > 8,48 1 0 102.632697082 0 m N throtl dispatch nr_queued=1 read=0 write=1
> > 8,48 1 0 102.632700544 0 m N throtl / [W] new slice start=4294957312 end=4294957412 jiffies=4294957312
> > 8,48 1 0 102.632701922 0 m N throtl / [W] extend slice start=4294957312 end=4295008600 jiffies=4294957312
> > 8,48 1 0 102.632704016 0 m N throtl schedule work. delay=51200 jiffies=4294957312
> > 8,48 1 0 153.960696503 0 m N throtl dispatch nr_queued=1 read=0 write=1
> > 8,48 1 0 153.960699797 0 m N throtl / [W] new slice start=4295008640 end=4295008740 jiffies=4295008640
> > 8,48 1 0 153.960701153 0 m N throtl / [W] extend slice start=4295008640 end=4295059900 jiffies=4295008640
> > 8,48 1 0 153.960703218 0 m N throtl schedule work. delay=51200 jiffies=4295008640
> > 8,48 1 0 205.288697067 0 m N throtl dispatch nr_queued=1 read=0 write=1
> > 8,48 1 0 205.288700268 0 m N throtl / [W] new slice start=4295059968 end=4295060068 jiffies=4295059968
> > 8,48 1 0 205.288701630 0 m N throtl / [W] extend slice start=4295059968 end=4295111200 jiffies=4295059968
> > 8,48 1 0 205.288703784 0 m N throtl schedule work. delay=51200 jiffies=4295059968
> > 8,48 1 0 256.616696184 0 m N throtl dispatch nr_queued=1 read=0 write=1
> > 8,48 1 0 256.616699266 0 m N throtl / [W] new slice start=4295111296 end=4295111396 jiffies=4295111296
> > 8,48 1 0 256.616700574 0 m N throtl / [W] extend slice start=4295111296 end=4295162500 jiffies=4295111296
> > 8,48 1 0 256.616702701 0 m N throtl schedule work. delay=51200 jiffies=4295111296
> >
> > Signed-off-by: Jianpeng Ma <majianpeng@gmail.com>
> > ---
> > block/blk-throttle.c | 15 ++++++++++-----
> > 1 file changed, 10 insertions(+), 5 deletions(-)
> >
> > diff --git a/block/blk-throttle.c b/block/blk-throttle.c
> > index 3114622..9258789 100644
> > --- a/block/blk-throttle.c
> > +++ b/block/blk-throttle.c
> > @@ -645,6 +645,15 @@ static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
> > }
> >
> > /*
> > + * If privious slice expired,then start new slice.
> > + * But counting bps and iops limit need privious slice info
> > + * which ->slice_start.
> > + */
> > + if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
> > + && tg_with_in_iops_limit(td, tg, bio, &iops_wait))
> > + if (wait)
> > + *wait = 0;
> > + /*
> > * If previous slice expired, start a new one otherwise renew/extend
> > * existing slice to make sure it is at least throtl_slice interval
> > * long since now.
> > @@ -656,12 +665,8 @@ static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
> > throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
> > }
> >
> > - if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
> > - && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
> > - if (wait)
> > - *wait = 0;
> > + if (!(bps_wait || iops_wait))
> > return 1;
> > - }
> >
> > max_wait = max(bps_wait, iops_wait);
>
> Looks pretty sane to me. Vivek?

Hi Jens,

This fix will introduce other side affects. And that is when an group
has been idle for few seconds and a new IO gets queued in, we will
not start a new slice and allow dispatch equivalent of those idle
seconds before we throttle the group. So keeping group idle will
become an incentive.

I have attached a patch which might work better. majianpeng, can
you please give it a try.

Having said that it is strange that workqueue thread is triggering
so late. Looking at timestamp of traces attached.

We scheduled a delayed work. Following is trace.

8,48 1 0 51.304705783 0 m N throtl schedule work. delay=51200 jiffies=4294905984

Now a worker should execute after 51.2 seconds (delay=51200).

8,48 1 0 102.632697082 0 m N throtl dispatch nr_queued=1 read=0 write=1

But worker executed at delay of 51.328 seconds (102.632 - 51.304). That
is a delay of around 128 jiffies (51.328 - 51.2). That kind of seems odd.
AFAIK, workers are fairly quick to execute after expiry. And it is
because of this excessive delay that we think slice has expired.

May be it is something new. CCing Tejun, if he has seen anything like this
and may be it is a known issue.

Even if it is a worker thread issue, I think applying below patch makes
sense.

Thanks
Vivek


blk-throttle: Do not start a new slice if IO is pending

It might happen that we queue an IO (because it overshot the rate)
and then wait for certain jiffies to pass. We will set slice_end
accordingly and wait for that time to elapse and worker thread will
kick in, again evaluate whether group can dispatch or not. Now it
might happen that current time is after slice_end and
tg_may_dispatch() will start a new slice.

This will result in IO not being dispatched and be put back on
wait again. And this will repeat, resulting in hang.

Do not start a new slice if an IO is pending in that group in
the direction being queired. Instead extend the slice. New slice
is supposed to start when a group has not been doing IO for some
time and a new IO shows up. In that case we want do discard
history and start a new slice.

Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
---
block/blk-throttle.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)

Index: linux-2.6/block/blk-throttle.c
===================================================================
--- linux-2.6.orig/block/blk-throttle.c 2012-10-18 01:52:28.000000000 -0400
+++ linux-2.6/block/blk-throttle.c 2013-01-14 03:40:41.355731375 -0500
@@ -648,8 +648,14 @@ static bool tg_may_dispatch(struct throt
* If previous slice expired, start a new one otherwise renew/extend
* existing slice to make sure it is at least throtl_slice interval
* long since now.
+ *
+ * Start a new slice only if there is no bio queued in that direction.
+ * That bio is waiting to be dispatched and slice needs to be
+ * extended. It might happen that bio waited to be dispatched but
+ * workqueue execution got little late it might restart a new slice
+ * instead of taking all the waited time into account.
*/
- if (throtl_slice_used(td, tg, rw))
+ if (throtl_slice_used(td, tg, rw) && !tg->nr_queued[rw])
throtl_start_new_slice(td, tg, rw);
else {
if (time_before(tg->slice_end[rw], jiffies + throtl_slice))

\
 
 \ /
  Last update: 2013-01-11 21:01    [W:0.105 / U:1.692 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site