Re: blk-throttle.c : When limit is changed, must start a new slice

Re: blk-throttle.c : When limit is changed, must start a new slice

[Vivek Goyal]

On Tue, Mar 08, 2011 at 11:03:59PM +0800, lina wrote:

[..]
> >>  Unfortunately, the following patch still has 5~10 seconds latency. I have no
> >>  idea to resolve this problem, it seens hard to find a more suitable func to
> >>  call throtl_start_new_slice().
> >
> >So are you saying that following patch did not solve the latnecy issue?
> >Resetting slice upon limit change did not work for you?
> ></:includetail>
>  </:includetail> 
>  Yes, the following patch did not solve the latency issue. There is still 5~10</:includetail>
>  seconds latency when I change the limit from a very high value to low. From </:includetail>
>  blktrace, I find that the throtl_process_limit_change() is called after work </:includetail>
>  queue </:includetail>delay.</:includetail>
>  </:includetail> 
>  Thanks</:includetail>
>  Lina</:includetail>
>  </:includetail></:includetail> 
>  </:includetail>>Thanks

Ok,

Can you try the attached patch. I think what was happening that after
changing limits, work was not being scheduled as there were no queued
bios hence no slice reset was taking place immediately.

Also I am not sure from where these "</:includetail>" strings are coming.
Looks like your mailer is inserting those. Trying sending mails in text
format.

Thanks
Vivek

---
 block/blk-throttle.c |   24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)

Index: linux-2.6/block/blk-throttle.c
===================================================================
--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
+++ linux-2.6/block/blk-throttle.c	2011-03-08 15:41:19.384654732 -0500
@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
 				" riops=%u wiops=%u", tg->bps[READ],
 				tg->bps[WRITE], tg->iops[READ],
 				tg->iops[WRITE]);
+			/*
+			 * Restart the slices for both READ and WRITES. It
+			 * might happen that a group's limit are dropped
+			 * suddenly and we don't want to account recently
+			 * dispatched IO with new low rate
+			 */
+			throtl_start_new_slice(td, tg, 0);
+			throtl_start_new_slice(td, tg, 1);
 			tg_update_disptime(td, tg);
 			tg->limits_changed = false;
 		}
@@ -825,7 +833,8 @@ throtl_schedule_delayed_work(struct thro
 
 	struct delayed_work *dwork = &td->throtl_work;
 
-	if (total_nr_queued(td) > 0) {
+	/* schedule work if limits changed even if no bio is queued */
+	if (total_nr_queued(td) > 0 || atomic_read(&td->limits_changed)) {
 		/*
 		 * We might have a work scheduled to be executed in future.
 		 * Cancel that and schedule a new one.
@@ -1023,6 +1032,19 @@ int blk_throtl_bio(struct request_queue 
 	/* Bio is with-in rate limit of group */
 	if (tg_may_dispatch(td, tg, bio, NULL)) {
 		throtl_charge_bio(tg, bio);
+
+		/*
+		 * We need to trim slice even when bios are not being queued
+		 * otherwise it might happen that a bio is not queued for
+		 * a long time and slice keeps on extending and trim is not
+		 * called for a long time. Now if limits are reduced suddenly
+		 * we take into account all the IO dispatched so far at new
+		 * low rate and * newly queued IO gets a really long dispatch
+		 * time.
+		 *
+		 * So keep on trimming slice even if bio is not queued.
+		 */
+		throtl_trim_slice(td, tg, rw);
 		goto out;
 	}
 

Re: Re: blk-throttle.c : When limit is changed, must start a new slice

[lulina_nuaa]

>On 2011-03-09 04:54:43, Vivek Goyal wrote:
>
>On Tue, Mar 08, 2011 at 11:03:59PM +0800, lina wrote:

>[..]
>> >>  Unfortunately, the following patch still has 5~10 seconds latency. I have no
>> >>  idea to resolve this problem, it seens hard to find a more suitable func to
>> >>  call throtl_start_new_slice().
>> >
>> >So are you saying that following patch did not solve the latnecy issue?
>> >Resetting slice upon limit change did not work for you?
>> >
>>   
>>  Yes, the following patch did not solve the latency issue. There is still 5~10
>>  seconds latency when I change the limit from a very high value to low. From
>>  blktrace, I find that the throtl_process_limit_change() is called after work 
>>  queue delay.
>>   
>>  Thanks
>>  Lina
>
>Ok,
>
>Can you try the attached patch. I think what was happening that after
>changing limits, work was not being scheduled as there were no queued
>bios hence no slice reset was taking place immediately.
>
>[..]
>
>Thanks
>Vivek
>

I have remove the HTML code, I'm sorry for the mail format!

Thank you very much for the following patch! I think it can solve the problem.
I'll test it as soon as possible, and will inform you once get the result!

Thanks
Lina

>---
> block/blk-throttle.c |   24 +++++++++++++++++++++++-
> 1 file changed, 23 insertions(+), 1 deletion(-)
>
>Index: linux-2.6/block/blk-throttle.c
>===================================================================
>--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
>+++ linux-2.6/block/blk-throttle.c	2011-03-08 15:41:19.384654732 -0500
>@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
> 				" riops=%u wiops=%u", tg->bps[READ],
> 				tg->bps[WRITE], tg->iops[READ],
> 				tg->iops[WRITE]);
>+			/*
>+			 * Restart the slices for both READ and WRITES. It
>+			 * might happen that a group's limit are dropped
>+			 * suddenly and we don't want to account recently
>+			 * dispatched IO with new low rate
>+			 */
>+			throtl_start_new_slice(td, tg, 0);
>+			throtl_start_new_slice(td, tg, 1);
> 			tg_update_disptime(td, tg);
> 			tg->limits_changed = false;
> 		}
>@@ -825,7 +833,8 @@ throtl_schedule_delayed_work(struct thro
> 
> 	struct delayed_work *dwork = &td->throtl_work;
> 
>-	if (total_nr_queued(td) > 0) {
>+	/* schedule work if limits changed even if no bio is queued */
>+	if (total_nr_queued(td) > 0 || atomic_read(&td->limits_changed)) {
> 		/*
> 		 * We might have a work scheduled to be executed in future.
> 		 * Cancel that and schedule a new one.
>@@ -1023,6 +1032,19 @@ int blk_throtl_bio(struct request_queue 
> 	/* Bio is with-in rate limit of group */
> 	if (tg_may_dispatch(td, tg, bio, NULL)) {
> 		throtl_charge_bio(tg, bio);
>+
>+		/*
>+		 * We need to trim slice even when bios are not being queued
>+		 * otherwise it might happen that a bio is not queued for
>+		 * a long time and slice keeps on extending and trim is not
>+		 * called for a long time. Now if limits are reduced suddenly
>+		 * we take into account all the IO dispatched so far at new
>+		 * low rate and * newly queued IO gets a really long dispatch
>+		 * time.
>+		 *
>+		 * So keep on trimming slice even if bio is not queued.
>+		 */
>+		throtl_trim_slice(td, tg, rw);
> 		goto out;
> 	}
 

Re: Re: blk-throttle.c : When limit is changed, must start a new slice

[Lina Lu]

On 2011-03-09 04:54:43, Vivek Goyal wrote:
>
>On Tue, Mar 08, 2011 at 11:03:59PM +0800, lina wrote:

>[..]
>> >>  Unfortunately, the following patch still has 5~10 seconds latency. I have no
>> >>  idea to resolve this problem, it seens hard to find a more suitable func to
>> >>  call throtl_start_new_slice().
>> >
>> >So are you saying that following patch did not solve the latnecy issue?
>> >Resetting slice upon limit change did not work for you?
>> >
>>   
>>  Yes, the following patch did not solve the latency issue. There is still 5~10
>>  seconds latency when I change the limit from a very high value to low. From
>>  blktrace, I find that the throtl_process_limit_change() is called after work 
>>  queue delay.
>>   
>>  Thanks
>>  Lina
>
>Ok,
>
>Can you try the attached patch. I think what was happening that after
>changing limits, work was not being scheduled as there were no queued
>bios hence no slice reset was taking place immediately.
>
>[..]
>
>Thanks
>Vivek
>

Hi Vivek,
I have test the following patch, but the latency still there.

I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
although we call it once dispatch a bio.

Suppose that if the limits change now from 102400000000 to 1024000, the 
tg->slice_start[rw] and tg->slice_end[rw] just like in the following chart. There is two 
throtl_slice in the chart. Here my HZ is 250, so the throtl_slice is 25.

                  jiffies
                  |
   |------------------|------------------|
   |                         |
start                    end

As the jiffies - start < 25(throtl_slice), throtl_trim_slice() will not update the 
tg->slice_start[rw] and tg->bytes_disp[rw]. If the tg->bytes_disp[rw] now is 8M, then
there will be about 7 seconds from jiffies 0 bps as I have set the limits at 1M/s, in 
these seconds no bio can be dispatched.

As the tg->slice_start[rw] must less than or equal to jiffies, and we can not know the 
reason of tg->bytes_disp[rw] > the theoretical value with limits 1M/s, So can not just 
set the tg->slice_start[rw] to jiffies here. If set the start to jiffies, throtl will not work.

I think if we can start a new slice in the next throtl_slice when the limits changed from 
high to low and the tg->bytes_disp[rw] is critical greater than the theoretical value with
now limits, this problem can be solved.

Thanks
Lina

>---
> block/blk-throttle.c |   24 +++++++++++++++++++++++-
> 1 file changed, 23 insertions(+), 1 deletion(-)
>
>Index: linux-2.6/block/blk-throttle.c
>===================================================================
>--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
>+++ linux-2.6/block/blk-throttle.c	2011-03-08 15:41:19.384654732 -0500
>@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
> 				" riops=%u wiops=%u", tg->bps[READ],
> 				tg->bps[WRITE], tg->iops[READ],
> 				tg->iops[WRITE]);
>+			/*
>+			 * Restart the slices for both READ and WRITES. It
>+			 * might happen that a group's limit are dropped
>+			 * suddenly and we don't want to account recently
>+			 * dispatched IO with new low rate
>+			 */
>+			throtl_start_new_slice(td, tg, 0);
>+			throtl_start_new_slice(td, tg, 1);
> 			tg_update_disptime(td, tg);
> 			tg->limits_changed = false;
> 		}
>@@ -825,7 +833,8 @@ throtl_schedule_delayed_work(struct thro
> 
> 	struct delayed_work *dwork = &td->throtl_work;
> 
>-	if (total_nr_queued(td) > 0) {
>+	/* schedule work if limits changed even if no bio is queued */
>+	if (total_nr_queued(td) > 0 || atomic_read(&td->limits_changed)) {
> 		/*
> 		 * We might have a work scheduled to be executed in future.
> 		 * Cancel that and schedule a new one.
>@@ -1023,6 +1032,19 @@ int blk_throtl_bio(struct request_queue 
> 	/* Bio is with-in rate limit of group */
> 	if (tg_may_dispatch(td, tg, bio, NULL)) {
> 		throtl_charge_bio(tg, bio);
>+
>+		/*
>+		 * We need to trim slice even when bios are not being queued
>+		 * otherwise it might happen that a bio is not queued for
>+		 * a long time and slice keeps on extending and trim is not
>+		 * called for a long time. Now if limits are reduced suddenly
>+		 * we take into account all the IO dispatched so far at new
>+		 * low rate and * newly queued IO gets a really long dispatch
>+		 * time.
>+		 *
>+		 * So keep on trimming slice even if bio is not queued.
>+		 */
>+		throtl_trim_slice(td, tg, rw);
> 		goto out;
> 	}
 

Re: Re: blk-throttle.c : When limit is changed, must start a new slice

[Vivek Goyal]

On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
> On 2011-03-09 04:54:43, Vivek Goyal wrote:
> >
> >On Tue, Mar 08, 2011 at 11:03:59PM +0800, lina wrote:
> 
> >[..]
> >> >>  Unfortunately, the following patch still has 5~10 seconds latency. I have no
> >> >>  idea to resolve this problem, it seens hard to find a more suitable func to
> >> >>  call throtl_start_new_slice().
> >> >
> >> >So are you saying that following patch did not solve the latnecy issue?
> >> >Resetting slice upon limit change did not work for you?
> >> >
> >>   
> >>  Yes, the following patch did not solve the latency issue. There is still 5~10
> >>  seconds latency when I change the limit from a very high value to low. From
> >>  blktrace, I find that the throtl_process_limit_change() is called after work 
> >>  queue delay.
> >>   
> >>  Thanks
> >>  Lina
> >
> >Ok,
> >
> >Can you try the attached patch. I think what was happening that after
> >changing limits, work was not being scheduled as there were no queued
> >bios hence no slice reset was taking place immediately.
> >
> >[..]
> >
> >Thanks
> >Vivek
> >
> 
> Hi Vivek,
> I have test the following patch, but the latency still there.
> 
> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
> although we call it once dispatch a bio.

lina,

Trim slice should not even matter now. Upon limit change, this patch
should reset the slice and start a new one irrespective of the fact
where are.

In your traces, do you see limit change message and do you see a new
slice starting.

I did similar test yesterday on my box and this patch worked. Can you
capture some block traces and I can have a look at those. Key thing
to look for is limit change message and whether it started a new
slice or not.

Thanks
Vivek

> 
> Suppose that if the limits change now from 102400000000 to 1024000, the 
> tg->slice_start[rw] and tg->slice_end[rw] just like in the following chart. There is two 
> throtl_slice in the chart. Here my HZ is 250, so the throtl_slice is 25.
> 
>                   jiffies
>                   |
>    |------------------|------------------|
>    |                         |
> start                    end
> 
> As the jiffies - start < 25(throtl_slice), throtl_trim_slice() will not update the 
> tg->slice_start[rw] and tg->bytes_disp[rw]. If the tg->bytes_disp[rw] now is 8M, then
> there will be about 7 seconds from jiffies 0 bps as I have set the limits at 1M/s, in 
> these seconds no bio can be dispatched.
> 
> As the tg->slice_start[rw] must less than or equal to jiffies, and we can not know the 
> reason of tg->bytes_disp[rw] > the theoretical value with limits 1M/s, So can not just 
> set the tg->slice_start[rw] to jiffies here. If set the start to jiffies, throtl will not work.
> 
> I think if we can start a new slice in the next throtl_slice when the limits changed from 
> high to low and the tg->bytes_disp[rw] is critical greater than the theoretical value with
> now limits, this problem can be solved.
> 
> Thanks
> Lina
> 
> >---
> > block/blk-throttle.c |   24 +++++++++++++++++++++++-
> > 1 file changed, 23 insertions(+), 1 deletion(-)
> >
> >Index: linux-2.6/block/blk-throttle.c
> >===================================================================
> >--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
> >+++ linux-2.6/block/blk-throttle.c	2011-03-08 15:41:19.384654732 -0500
> >@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
> > 				" riops=%u wiops=%u", tg->bps[READ],
> > 				tg->bps[WRITE], tg->iops[READ],
> > 				tg->iops[WRITE]);
> >+			/*
> >+			 * Restart the slices for both READ and WRITES. It
> >+			 * might happen that a group's limit are dropped
> >+			 * suddenly and we don't want to account recently
> >+			 * dispatched IO with new low rate
> >+			 */
> >+			throtl_start_new_slice(td, tg, 0);
> >+			throtl_start_new_slice(td, tg, 1);
> > 			tg_update_disptime(td, tg);
> > 			tg->limits_changed = false;
> > 		}
> >@@ -825,7 +833,8 @@ throtl_schedule_delayed_work(struct thro
> > 
> > 	struct delayed_work *dwork = &td->throtl_work;
> > 
> >-	if (total_nr_queued(td) > 0) {
> >+	/* schedule work if limits changed even if no bio is queued */
> >+	if (total_nr_queued(td) > 0 || atomic_read(&td->limits_changed)) {
> > 		/*
> > 		 * We might have a work scheduled to be executed in future.
> > 		 * Cancel that and schedule a new one.
> >@@ -1023,6 +1032,19 @@ int blk_throtl_bio(struct request_queue 
> > 	/* Bio is with-in rate limit of group */
> > 	if (tg_may_dispatch(td, tg, bio, NULL)) {
> > 		throtl_charge_bio(tg, bio);
> >+
> >+		/*
> >+		 * We need to trim slice even when bios are not being queued
> >+		 * otherwise it might happen that a bio is not queued for
> >+		 * a long time and slice keeps on extending and trim is not
> >+		 * called for a long time. Now if limits are reduced suddenly
> >+		 * we take into account all the IO dispatched so far at new
> >+		 * low rate and * newly queued IO gets a really long dispatch
> >+		 * time.
> >+		 *
> >+		 * So keep on trimming slice even if bio is not queued.
> >+		 */
> >+		throtl_trim_slice(td, tg, rw);
> > 		goto out;
> > 	}
>  
> 

Re: Re: Re: blk-throttle.c : When limit is changed, must start a newslice

[Lina Lu]

On 2011-03-11 03:55:55, Vivek Goyal wrote:
>On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
>> [..]
>> Hi Vivek,
>> I have test the following patch, but the latency still there.
>> 
>> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
>> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
>> although we call it once dispatch a bio.
>
>lina,
>
>Trim slice should not even matter now. Upon limit change, this patch
>should reset the slice and start a new one irrespective of the fact
>where are.
>
>In your traces, do you see limit change message and do you see a new
>slice starting.
>
>I did similar test yesterday on my box and this patch worked. Can you
>capture some block traces and I can have a look at those. Key thing
>to look for is limit change message and whether it started a new
>slice or not.
>
>Thanks
>Vivek
>
	
Hi Vivek,
	
Here is the blktrace and iostat results when I change the limit from 1024000000000000
to 1024000. When the limit changed, there is about 3 seconds lantency.

blktrace:	
253,1    0        0     4.177733270     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297788991 end=4297789100 jiffies=4297788992
253,1    0        0     4.187393582     0  m   N throtl / [R] extend slice start=4297788991 end=4297789200 jiffies=4297789002
253,1    0        0     4.276120505     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297789091 end=4297789200 jiffies=4297789091
253,1    0        0     4.285934091     0  m   N throtl / [R] extend slice start=4297789091 end=4297789300 jiffies=4297789101
253,1    1        0     4.348552814     0  m   N throtl schedule work. delay=0 jiffies=4297789163
253,1    1        0     4.348571560     0  m   N throtl limit changed =1
253,1    0        0     4.349839104     0  m   N throtl / [R] extend slice start=4297789091 end=4297793000 jiffies=4297789164
253,1    0        0     4.349844118     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=0/0
253,1    0        0     4.349850121     0  m   N throtl schedule work. delay=3767 jiffies=4297789164
253,1    0        0     4.349912607     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=1/0
253,1    0        0     4.349915880     0  m   N throtl schedule work. delay=3766 jiffies=4297789165
253,1    0        0     4.349921567     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=2/0
...            #queued 63 read bios with no new slice.
253,1    0        0     4.353728869     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=61/0
253,1    0        0     4.353731799     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=62/0
253,1    0        0     4.353735427     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=63/0
253,1    0        0     8.129092326     0  m   N throtl dispatch nr_queued=64 read=64 write=0
253,1    0        0     8.129096924     0  m   N throtl / [R] extend slice start=4297789091 end=4297793100 jiffies=4297792944
253,1    0        0     8.129100584     0  m   N throtl / [R] trim slice nr=38 bytes=3891200 io=16320875721 start=4297792891 end=4297793100 jiffies=4297792944
253,1    0        0     8.129108331     0  m   N throtl bios disp=16
253,1    0        0     8.129111864     0  m   N throtl schedule work. delay=51 jiffies=4297792944
253,1    0        0     8.180899035     0  m   N throtl dispatch nr_queued=48 read=48 write=0
253,1    0        0     8.180905222     0  m   N throtl / [R] trim slice nr=1 bytes=102400 io=429496729 start=4297792991 end=4297793100 jiffies=4297792996
253,1    0        0     8.180915206     0  m   N throtl bios disp=25
253,1    0        0     8.180919011     0  m   N throtl schedule work. delay=99 jiffies=4297792996
253,1    0        0     8.182058927     0  m   N throtl / [R] bio. bdisp=102400 sz=4096 bps=1024000 iodisp=24 iops=4294967295 queued=23/0

iostat:
Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s avgrq-sz avgqu-sz   await  svctm  %util
dm-1              0.00     0.00 12992.00    0.00    50.75     0.00     8.00    59.47    4.57   0.08  99.70
dm-1              0.00     0.00 12598.00    0.00    49.21     0.00     8.00    59.03    4.70   0.08  99.70
dm-1              0.00     0.00 12923.00    0.00    50.48     0.00     8.00    54.43    4.22   0.08  98.50
dm-1              0.00     0.00 13103.00    0.00    51.18     0.00     8.00    57.54    4.38   0.08  99.70
dm-1              0.00     0.00 13024.00    0.00    50.88     0.00     8.00    58.67    4.51   0.08  99.70
dm-1              0.00     0.00 12928.00    0.00    50.50     0.00     8.00    58.50    4.53   0.08  99.60
dm-1              0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00   0.00   0.00
dm-1              0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00   0.00   0.00
dm-1              0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00   0.00   0.00
dm-1              0.00     0.00   66.00    0.00     0.26     0.00     8.00     0.05    0.76   0.03   0.20
dm-1              0.00     0.00  250.00    0.00     0.98     0.00     8.00     0.24    0.98   0.04   1.00

>From the trace we can find 3766 delay(3766/HZ~3.7 seconds), the greater delay value, the
longer latency. Sometimes, the delay is low, so the latency does not appear everytime when I
change limit from high to low.
	
And the latency seens relate to the device's physic capacity.
Here my device has only 50MB/s physic capacity, so there is about 3~5 seconds latency.
If the device has 100MB/s physic capacity, the lantcy will be 5~10 seconds latency. 

There is no new slice trace because the throtl_process_limit_change() is not been called. 
throtl_process_limit_change() function is called only in throtl_dispatch(), and throtl_dispatch()
is called only in blk_throtl_work(). When the limit change from high to low, there is no work
queue, so blk_throtl_work() is never called.

When the limit change from low to high, I find the new slice trace like following. So it only new
slice when there is work queue.

253,1    0        0    60.250888001     0  m   N throtl / [R] bio. bdisp=102400 sz=4096 bps=1024000 iodisp=24 iops=4294967295 queued=49/0
253,1    0        0    60.250890858     0  m   N throtl / [R] bio. bdisp=102400 sz=4096 bps=1024000 iodisp=24 iops=4294967295 queued=50/0
253,1    0        0    60.349455559     0  m   N throtl dispatch nr_queued=51 read=51 write=0
253,1    0        0    60.349460882     0  m   N throtl / [R] extend slice start=4297998658 end=4297998900 jiffies=4297998762
253,1    0        0    60.349464810     0  m   N throtl / [R] trim slice nr=1 bytes=102400 io=429496729 start=4297998758 end=4297998900 jiffies=4297998762
253,1    0        0    60.349473330     0  m   N throtl bios disp=25
253,1    0        0    60.349476631     0  m   N throtl schedule work. delay=100 jiffies=4297998762
253,1    1        0    60.375043834     0  m   N throtl schedule work. delay=0 jiffies=4297998787
253,1    1        0    60.375062998     0  m   N throtl limit changed =1
253,1    1        0    60.375066704     0  m   N throtl / limit change rbps=1024000000000000 wbps=18446744073709551615 riops=4294967295 wiops=4294967295
253,1    1        0    60.375069747     0  m   N throtl / [R] new slice start=4297998787 end=4297998887 jiffies=4297998787
253,1    1        0    60.375070919     0  m   N throtl / [W] new slice start=4297998787 end=4297998887 jiffies=4297998787
253,1    1        0    60.375073946     0  m   N throtl dispatch nr_queued=26 read=26 write=0
253,1    1        0    60.375083440     0  m   N throtl bios disp=26
253,1    1        0    60.430614460     0  m   N throtl / [R] extend slice start=4297998787 end=4297999000 jiffies=4297998843
253,1    1        0    60.476022578     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297998887 end=4297999000 jiffies=4297998888

Thanks
Lina                            
  	
>> 
>> Suppose that if the limits change now from 102400000000 to 1024000, the 
>> tg->slice_start[rw] and tg->slice_end[rw] just like in the following chart. There is two 
>> throtl_slice in the chart. Here my HZ is 250, so the throtl_slice is 25.
>> 
>>                   jiffies
>>                   |
>>    |------------------|------------------|
>>    |                         |
>> start                    end
>> 
>> As the jiffies - start < 25(throtl_slice), throtl_trim_slice() will not update the 
>> tg->slice_start[rw] and tg->bytes_disp[rw]. If the tg->bytes_disp[rw] now is 8M, then
>> there will be about 7 seconds from jiffies 0 bps as I have set the limits at 1M/s, in 
>> these seconds no bio can be dispatched.
>> 
>> As the tg->slice_start[rw] must less than or equal to jiffies, and we can not know the 
>> reason of tg->bytes_disp[rw] > the theoretical value with limits 1M/s, So can not just 
>> set the tg->slice_start[rw] to jiffies here. If set the start to jiffies, throtl will not work.
>> 
>> I think if we can start a new slice in the next throtl_slice when the limits changed from 
>> high to low and the tg->bytes_disp[rw] is critical greater than the theoretical value with
>> now limits, this problem can be solved.
>> 
>> Thanks
>> Lina
>> 
>> >---
>> > block/blk-throttle.c |   24 +++++++++++++++++++++++-
>> > 1 file changed, 23 insertions(+), 1 deletion(-)
>> >
>> >Index: linux-2.6/block/blk-throttle.c
>> >===================================================================
>> >--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
>> >+++ linux-2.6/block/blk-throttle.c	2011-03-08 15:41:19.384654732 -0500
>> >@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
>> > 				" riops=%u wiops=%u", tg->bps[READ],
>> > 				tg->bps[WRITE], tg->iops[READ],
>> > 				tg->iops[WRITE]);
>> >+			/*
>> >+			 * Restart the slices for both READ and WRITES. It
>> >+			 * might happen that a group's limit are dropped
>> >+			 * suddenly and we don't want to account recently
>> >+			 * dispatched IO with new low rate
>> >+			 */
>> >+			throtl_start_new_slice(td, tg, 0);
>> >+			throtl_start_new_slice(td, tg, 1);
>> > 			tg_update_disptime(td, tg);
>> > 			tg->limits_changed = false;
>> > 		}
>> >@@ -825,7 +833,8 @@ throtl_schedule_delayed_work(struct thro
>> > 
>> > 	struct delayed_work *dwork = &td->throtl_work;
>> > 
>> >-	if (total_nr_queued(td) > 0) {
>> >+	/* schedule work if limits changed even if no bio is queued */
>> >+	if (total_nr_queued(td) > 0 || atomic_read(&td->limits_changed)) {
>> > 		/*
>> > 		 * We might have a work scheduled to be executed in future.
>> > 		 * Cancel that and schedule a new one.
>> >@@ -1023,6 +1032,19 @@ int blk_throtl_bio(struct request_queue 
>> > 	/* Bio is with-in rate limit of group */
>> > 	if (tg_may_dispatch(td, tg, bio, NULL)) {
>> > 		throtl_charge_bio(tg, bio);
>> >+
>> >+		/*
>> >+		 * We need to trim slice even when bios are not being queued
>> >+		 * otherwise it might happen that a bio is not queued for
>> >+		 * a long time and slice keeps on extending and trim is not
>> >+		 * called for a long time. Now if limits are reduced suddenly
>> >+		 * we take into account all the IO dispatched so far at new
>> >+		 * low rate and * newly queued IO gets a really long dispatch
>> >+		 * time.
>> >+		 *
>> >+		 * So keep on trimming slice even if bio is not queued.
>> >+		 */
>> >+		throtl_trim_slice(td, tg, rw);
>> > 		goto out;
>> > 	}
>>  
>> 
 

Re: Re: Re: blk-throttle.c : When limit is changed, must start a newslice

[Vivek Goyal]

On Sat, Mar 12, 2011 at 07:33:07PM +0800, Lina Lu wrote:
> On 2011-03-11 03:55:55, Vivek Goyal wrote:
> >On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
> >> [..]
> >> Hi Vivek,
> >> I have test the following patch, but the latency still there.
> >> 
> >> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
> >> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
> >> although we call it once dispatch a bio.
> >
> >lina,
> >
> >Trim slice should not even matter now. Upon limit change, this patch
> >should reset the slice and start a new one irrespective of the fact
> >where are.
> >
> >In your traces, do you see limit change message and do you see a new
> >slice starting.
> >
> >I did similar test yesterday on my box and this patch worked. Can you
> >capture some block traces and I can have a look at those. Key thing
> >to look for is limit change message and whether it started a new
> >slice or not.
> >
> >Thanks
> >Vivek
> >
> 	
> Hi Vivek,
> 	
> Here is the blktrace and iostat results when I change the limit from 1024000000000000
> to 1024000. When the limit changed, there is about 3 seconds lantency.
> 
> blktrace:	
> 253,1    0        0     4.177733270     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297788991 end=4297789100 jiffies=4297788992
> 253,1    0        0     4.187393582     0  m   N throtl / [R] extend slice start=4297788991 end=4297789200 jiffies=4297789002
> 253,1    0        0     4.276120505     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297789091 end=4297789200 jiffies=4297789091
> 253,1    0        0     4.285934091     0  m   N throtl / [R] extend slice start=4297789091 end=4297789300 jiffies=4297789101
> 253,1    1        0     4.348552814     0  m   N throtl schedule work. delay=0 jiffies=4297789163
> 253,1    1        0     4.348571560     0  m   N throtl limit changed =1
> 253,1    0        0     4.349839104     0  m   N throtl / [R] extend slice start=4297789091 end=4297793000 jiffies=4297789164
> 253,1    0        0     4.349844118     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=0/0

Lina,

Thanks for the traces.

I think we did call process_limit_change() but we did not start the new
slice. I guess this happened because, we seem to be starting slice only
if group on run tree. Because before limit udpates, most likely group
is not on run tree as limits are very high, hence we missed resetting
the slice.

        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
                if (throtl_tg_on_rr(tg) && tg->limits_changed) {
                        throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
                                " riops=%u wiops=%u", tg->bps[READ],
                                tg->bps[WRITE], tg->iops[READ],
                                tg->iops[WRITE]);

Actually many races have been fixed in Jens's block tree. Is it possible to
test origin/for-2.6.39/core branch of Jens's tree with following patch applied
and see if it fixes the issue for you?

Thanks
Vivek

---
 block/blk-throttle.c |   25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

Index: linux-2.6-block/block/blk-throttle.c
===================================================================
--- linux-2.6-block.orig/block/blk-throttle.c	2011-03-14 10:27:57.000000000 -0400
+++ linux-2.6-block/block/blk-throttle.c	2011-03-14 10:30:47.267170956 -0400
@@ -756,6 +756,15 @@ static void throtl_process_limit_change(
 			" riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
 			tg->iops[READ], tg->iops[WRITE]);
 
+		/*
+		 * Restart the slices for both READ and WRITES. It
+		 * might happen that a group's limit are dropped
+		 * suddenly and we don't want to account recently
+		 * dispatched IO with new low rate
+		 */
+		throtl_start_new_slice(td, tg, 0);
+		throtl_start_new_slice(td, tg, 1);
+
 		if (throtl_tg_on_rr(tg))
 			tg_update_disptime(td, tg);
 	}
@@ -821,7 +830,8 @@ throtl_schedule_delayed_work(struct thro
 
 	struct delayed_work *dwork = &td->throtl_work;
 
-	if (total_nr_queued(td) > 0) {
+	/* schedule work if limits changed even if no bio is queued */
+	if (total_nr_queued(td) > 0 || td->limits_changed) {
 		/*
 		 * We might have a work scheduled to be executed in future.
 		 * Cancel that and schedule a new one.
@@ -1002,6 +1012,19 @@ int blk_throtl_bio(struct request_queue 
 	/* Bio is with-in rate limit of group */
 	if (tg_may_dispatch(td, tg, bio, NULL)) {
 		throtl_charge_bio(tg, bio);
+
+		/*
+		 * We need to trim slice even when bios are not being queued
+		 * otherwise it might happen that a bio is not queued for
+		 * a long time and slice keeps on extending and trim is not
+		 * called for a long time. Now if limits are reduced suddenly
+		 * we take into account all the IO dispatched so far at new
+		 * low rate and * newly queued IO gets a really long dispatch
+		 * time.
+		 *
+		 * So keep on trimming slice even if bio is not queued.
+		 */
+		throtl_trim_slice(td, tg, rw);
 		goto out;
 	}
 

Re: Re: Re: blk-throttle.c : When limit is changed, must start anewslice

[Lina Lu]

On 2011-03-14 23:18:31, Vivek Goyal wrote:
>On Sat, Mar 12, 2011 at 07:33:07PM +0800, Lina Lu wrote:
>> On 2011-03-11 03:55:55, Vivek Goyal wrote:
>> >On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
>> >> [..]
>> >> Hi Vivek,
>> >> I have test the following patch, but the latency still there.
>> >> 
>> >> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
>> >> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
>> >> although we call it once dispatch a bio.
>> >
>> >lina,
>> >
>> >Trim slice should not even matter now. Upon limit change, this patch
>> >should reset the slice and start a new one irrespective of the fact
>> >where are.
>> >
>> >In your traces, do you see limit change message and do you see a new
>> >slice starting.
>> >
>> >I did similar test yesterday on my box and this patch worked. Can you
>> >capture some block traces and I can have a look at those. Key thing
>> >to look for is limit change message and whether it started a new
>> >slice or not.
>> >
>> >Thanks
>> >Vivek
>> >
>> 	
>> Hi Vivek,
>> 	
>> Here is the blktrace and iostat results when I change the limit from 1024000000000000
>> to 1024000. When the limit changed, there is about 3 seconds lantency.
>> 
>> blktrace:	
>> 253,1    0        0     4.177733270     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297788991 end=4297789100 jiffies=4297788992
>> 253,1    0        0     4.187393582     0  m   N throtl / [R] extend slice start=4297788991 end=4297789200 jiffies=4297789002
>> 253,1    0        0     4.276120505     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297789091 end=4297789200 jiffies=4297789091
>> 253,1    0        0     4.285934091     0  m   N throtl / [R] extend slice start=4297789091 end=4297789300 jiffies=4297789101
>> 253,1    1        0     4.348552814     0  m   N throtl schedule work. delay=0 jiffies=4297789163
>> 253,1    1        0     4.348571560     0  m   N throtl limit changed =1
>> 253,1    0        0     4.349839104     0  m   N throtl / [R] extend slice start=4297789091 end=4297793000 jiffies=4297789164
>> 253,1    0        0     4.349844118     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=0/0
>
>Lina,
>
>Thanks for the traces.
>
>I think we did call process_limit_change() but we did not start the new
>slice. I guess this happened because, we seem to be starting slice only
>if group on run tree. Because before limit udpates, most likely group
>is not on run tree as limits are very high, hence we missed resetting
>the slice.
>
>        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
>                if (throtl_tg_on_rr(tg) && tg->limits_changed) {
>                        throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
>                                " riops=%u wiops=%u", tg->bps[READ],
>                                tg->bps[WRITE], tg->iops[READ],
>                                tg->iops[WRITE]);
>
	
Do you mean that throtl_tg_on_rr() function returns 0 when the limits are very
high?
	
>Actually many races have been fixed in Jens's block tree. Is it possible to
>test origin/for-2.6.39/core branch of Jens's tree with following patch applied
>and see if it fixes the issue for you?

I only find 2.6.38 core in gitweb. Do you mean origin/for-2.6.38/core branch? 
I'll test it as soon as possible and keep you know the result.

>Thanks
>Vivek
>
>---
> block/blk-throttle.c |   25 ++++++++++++++++++++++++-
> 1 file changed, 24 insertions(+), 1 deletion(-)
>
>Index: linux-2.6-block/block/blk-throttle.c
>===================================================================
>--- linux-2.6-block.orig/block/blk-throttle.c	2011-03-14 10:27:57.000000000 -0400
>+++ linux-2.6-block/block/blk-throttle.c	2011-03-14 10:30:47.267170956 -0400
>@@ -756,6 +756,15 @@ static void throtl_process_limit_change(
> 			" riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
> 			tg->iops[READ], tg->iops[WRITE]);
> 
>+		/*
>+		 * Restart the slices for both READ and WRITES. It
>+		 * might happen that a group's limit are dropped
>+		 * suddenly and we don't want to account recently
>+		 * dispatched IO with new low rate
>+		 */
>+		throtl_start_new_slice(td, tg, 0);
>+		throtl_start_new_slice(td, tg, 1);
>+
> 		if (throtl_tg_on_rr(tg))
> 			tg_update_disptime(td, tg);
> 	}
>@@ -821,7 +830,8 @@ throtl_schedule_delayed_work(struct thro
> 
> 	struct delayed_work *dwork = &td->throtl_work;
> 
>-	if (total_nr_queued(td) > 0) {
>+	/* schedule work if limits changed even if no bio is queued */
>+	if (total_nr_queued(td) > 0 || td->limits_changed) {
> 		/*
> 		 * We might have a work scheduled to be executed in future.
> 		 * Cancel that and schedule a new one.
>@@ -1002,6 +1012,19 @@ int blk_throtl_bio(struct request_queue 
> 	/* Bio is with-in rate limit of group */
> 	if (tg_may_dispatch(td, tg, bio, NULL)) {
> 		throtl_charge_bio(tg, bio);
>+
>+		/*
>+		 * We need to trim slice even when bios are not being queued
>+		 * otherwise it might happen that a bio is not queued for
>+		 * a long time and slice keeps on extending and trim is not
>+		 * called for a long time. Now if limits are reduced suddenly
>+		 * we take into account all the IO dispatched so far at new
>+		 * low rate and * newly queued IO gets a really long dispatch
>+		 * time.
>+		 *
>+		 * So keep on trimming slice even if bio is not queued.
>+		 */
>+		throtl_trim_slice(td, tg, rw);
> 		goto out;
> 	}
 
t

Re: Re: Re: blk-throttle.c : When limit is changed, must start anewslice

[Vivek Goyal]

On Mon, Mar 14, 2011 at 11:52:36PM +0800, Lina Lu wrote:
> On 2011-03-14 23:18:31, Vivek Goyal wrote:
> >On Sat, Mar 12, 2011 at 07:33:07PM +0800, Lina Lu wrote:
> >> On 2011-03-11 03:55:55, Vivek Goyal wrote:
> >> >On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
> >> >> [..]
> >> >> Hi Vivek,
> >> >> I have test the following patch, but the latency still there.
> >> >> 
> >> >> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
> >> >> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
> >> >> although we call it once dispatch a bio.
> >> >
> >> >lina,
> >> >
> >> >Trim slice should not even matter now. Upon limit change, this patch
> >> >should reset the slice and start a new one irrespective of the fact
> >> >where are.
> >> >
> >> >In your traces, do you see limit change message and do you see a new
> >> >slice starting.
> >> >
> >> >I did similar test yesterday on my box and this patch worked. Can you
> >> >capture some block traces and I can have a look at those. Key thing
> >> >to look for is limit change message and whether it started a new
> >> >slice or not.
> >> >
> >> >Thanks
> >> >Vivek
> >> >
> >> 	
> >> Hi Vivek,
> >> 	
> >> Here is the blktrace and iostat results when I change the limit from 1024000000000000
> >> to 1024000. When the limit changed, there is about 3 seconds lantency.
> >> 
> >> blktrace:	
> >> 253,1    0        0     4.177733270     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297788991 end=4297789100 jiffies=4297788992
> >> 253,1    0        0     4.187393582     0  m   N throtl / [R] extend slice start=4297788991 end=4297789200 jiffies=4297789002
> >> 253,1    0        0     4.276120505     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297789091 end=4297789200 jiffies=4297789091
> >> 253,1    0        0     4.285934091     0  m   N throtl / [R] extend slice start=4297789091 end=4297789300 jiffies=4297789101
> >> 253,1    1        0     4.348552814     0  m   N throtl schedule work. delay=0 jiffies=4297789163
> >> 253,1    1        0     4.348571560     0  m   N throtl limit changed =1
> >> 253,1    0        0     4.349839104     0  m   N throtl / [R] extend slice start=4297789091 end=4297793000 jiffies=4297789164
> >> 253,1    0        0     4.349844118     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=0/0
> >
> >Lina,
> >
> >Thanks for the traces.
> >
> >I think we did call process_limit_change() but we did not start the new
> >slice. I guess this happened because, we seem to be starting slice only
> >if group on run tree. Because before limit udpates, most likely group
> >is not on run tree as limits are very high, hence we missed resetting
> >the slice.
> >
> >        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
> >                if (throtl_tg_on_rr(tg) && tg->limits_changed) {
> >                        throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
> >                                " riops=%u wiops=%u", tg->bps[READ],
> >                                tg->bps[WRITE], tg->iops[READ],
> >                                tg->iops[WRITE]);
> >
> 	
> Do you mean that throtl_tg_on_rr() function returns 0 when the limits are very
> high?

Yes. When limits are very high, you will never enqueue a bio hence a 
group will never be enqueued hence throtl_tg_on_rr=0.

> 	
> >Actually many races have been fixed in Jens's block tree. Is it possible to
> >test origin/for-2.6.39/core branch of Jens's tree with following patch applied
> >and see if it fixes the issue for you?
> 
> I only find 2.6.38 core in gitweb. Do you mean origin/for-2.6.38/core branch? 
> I'll test it as soon as possible and keep you know the result.

Here is Jens's block tree. It is separate from linus's tree.

http://git.kernel.org/?p=linux/kernel/git/axboe/linux-2.6-block.git;a=summary

Thanks
Vivek

> 
> >Thanks
> >Vivek
> >
> >---
> > block/blk-throttle.c |   25 ++++++++++++++++++++++++-
> > 1 file changed, 24 insertions(+), 1 deletion(-)
> >
> >Index: linux-2.6-block/block/blk-throttle.c
> >===================================================================
> >--- linux-2.6-block.orig/block/blk-throttle.c	2011-03-14 10:27:57.000000000 -0400
> >+++ linux-2.6-block/block/blk-throttle.c	2011-03-14 10:30:47.267170956 -0400
> >@@ -756,6 +756,15 @@ static void throtl_process_limit_change(
> > 			" riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
> > 			tg->iops[READ], tg->iops[WRITE]);
> > 
> >+		/*
> >+		 * Restart the slices for both READ and WRITES. It
> >+		 * might happen that a group's limit are dropped
> >+		 * suddenly and we don't want to account recently
> >+		 * dispatched IO with new low rate
> >+		 */
> >+		throtl_start_new_slice(td, tg, 0);
> >+		throtl_start_new_slice(td, tg, 1);
> >+
> > 		if (throtl_tg_on_rr(tg))
> > 			tg_update_disptime(td, tg);
> > 	}
> >@@ -821,7 +830,8 @@ throtl_schedule_delayed_work(struct thro
> > 
> > 	struct delayed_work *dwork = &td->throtl_work;
> > 
> >-	if (total_nr_queued(td) > 0) {
> >+	/* schedule work if limits changed even if no bio is queued */
> >+	if (total_nr_queued(td) > 0 || td->limits_changed) {
> > 		/*
> > 		 * We might have a work scheduled to be executed in future.
> > 		 * Cancel that and schedule a new one.
> >@@ -1002,6 +1012,19 @@ int blk_throtl_bio(struct request_queue 
> > 	/* Bio is with-in rate limit of group */
> > 	if (tg_may_dispatch(td, tg, bio, NULL)) {
> > 		throtl_charge_bio(tg, bio);
> >+
> >+		/*
> >+		 * We need to trim slice even when bios are not being queued
> >+		 * otherwise it might happen that a bio is not queued for
> >+		 * a long time and slice keeps on extending and trim is not
> >+		 * called for a long time. Now if limits are reduced suddenly
> >+		 * we take into account all the IO dispatched so far at new
> >+		 * low rate and * newly queued IO gets a really long dispatch
> >+		 * time.
> >+		 *
> >+		 * So keep on trimming slice even if bio is not queued.
> >+		 */
> >+		throtl_trim_slice(td, tg, rw);
> > 		goto out;
> > 	}
>  
> t

Re: Re: Re: blk-throttle.c : When limit is changed, must startanewslice

[Lina Lu]

On 2011-03-14 23:52:31, Vivek Goyal wrote:
>On Mon, Mar 14, 2011 at 11:52:36PM +0800, Lina Lu wrote:
>> On 2011-03-14 23:18:31, Vivek Goyal wrote:
>> >On Sat, Mar 12, 2011 at 07:33:07PM +0800, Lina Lu wrote:
>> >> On 2011-03-11 03:55:55, Vivek Goyal wrote:
>> >> >On Fri, Mar 11, 2011 at 12:38:18AM +0800, Lina Lu wrote:
>> >> >> [..]
>> >> >> Hi Vivek,
>> >> >> I have test the following patch, but the latency still there.
>> >> >> 
>> >> >> I try to find why there are 5~10 seconds latency today. After collect the blktrace, I 
>> >> >> think the reason is that throtl_trim_slice() don't aways update the tg->slice_start[rw], 
>> >> >> although we call it once dispatch a bio.
>> >> >
>> >> >lina,
>> >> >
>> >> >Trim slice should not even matter now. Upon limit change, this patch
>> >> >should reset the slice and start a new one irrespective of the fact
>> >> >where are.
>> >> >
>> >> >In your traces, do you see limit change message and do you see a new
>> >> >slice starting.
>> >> >
>> >> >I did similar test yesterday on my box and this patch worked. Can you
>> >> >capture some block traces and I can have a look at those. Key thing
>> >> >to look for is limit change message and whether it started a new
>> >> >slice or not.
>> >> >
>> >> >Thanks
>> >> >Vivek
>> >> >
>> >> 	
>> >> Hi Vivek,
>> >> 	
>> >> Here is the blktrace and iostat results when I change the limit from 1024000000000000
>> >> to 1024000. When the limit changed, there is about 3 seconds lantency.
>> >> 
>> >> blktrace:	
>> >> 253,1    0        0     4.177733270     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297788991 end=4297789100 jiffies=4297788992
>> >> 253,1    0        0     4.187393582     0  m   N throtl / [R] extend slice start=4297788991 end=4297789200 jiffies=4297789002
>> >> 253,1    0        0     4.276120505     0  m   N throtl / [R] trim slice nr=1 bytes=102400000000000 io=429496729 start=4297789091 end=4297789200 jiffies=4297789091
>> >> 253,1    0        0     4.285934091     0  m   N throtl / [R] extend slice start=4297789091 end=4297789300 jiffies=4297789101
>> >> 253,1    1        0     4.348552814     0  m   N throtl schedule work. delay=0 jiffies=4297789163
>> >> 253,1    1        0     4.348571560     0  m   N throtl limit changed =1
>> >> 253,1    0        0     4.349839104     0  m   N throtl / [R] extend slice start=4297789091 end=4297793000 jiffies=4297789164
>> >> 253,1    0        0     4.349844118     0  m   N throtl / [R] bio. bdisp=3928064 sz=4096 bps=1024000 iodisp=959 iops=4294967295 queued=0/0
>> >
>> >Lina,
>> >
>> >Thanks for the traces.
>> >
>> >I think we did call process_limit_change() but we did not start the new
>> >slice. I guess this happened because, we seem to be starting slice only
>> >if group on run tree. Because before limit udpates, most likely group
>> >is not on run tree as limits are very high, hence we missed resetting
>> >the slice.
>> >
>> >        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
>> >                if (throtl_tg_on_rr(tg) && tg->limits_changed) {
>> >                        throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
>> >                                " riops=%u wiops=%u", tg->bps[READ],
>> >                                tg->bps[WRITE], tg->iops[READ],
>> >                                tg->iops[WRITE]);
>> >
>> 	
>> Do you mean that throtl_tg_on_rr() function returns 0 when the limits are very
>> high?
>
>Yes. When limits are very high, you will never enqueue a bio hence a 
>group will never be enqueued hence throtl_tg_on_rr=0.
>
>> 	
>> >Actually many races have been fixed in Jens's block tree. Is it possible to
>> >test origin/for-2.6.39/core branch of Jens's tree with following patch applied
>> >and see if it fixes the issue for you?
>> 
>> I only find 2.6.38 core in gitweb. Do you mean origin/for-2.6.38/core branch? 
>> I'll test it as soon as possible and keep you know the result.
>
>Here is Jens's block tree. It is separate from linus's tree.
>
>http://git.kernel.org/?p=linux/kernel/git/axboe/linux-2.6-block.git;a=summary
>
>Thanks
>Vivek
>
Hi Vivek,
I have test the following patch on 2.6.39 core branch of Jens's tree, the bug has
been fixed. 

Can you tell me which patch makes the throtl_tg_on_rr() return 1 even if the limits
are very high?

Thanks
Lina
	
>> 
>> >Thanks
>> >Vivek
>> >
>> >---
>> > block/blk-throttle.c |   25 ++++++++++++++++++++++++-
>> > 1 file changed, 24 insertions(+), 1 deletion(-)
>> >
>> >Index: linux-2.6-block/block/blk-throttle.c
>> >===================================================================
>> >--- linux-2.6-block.orig/block/blk-throttle.c	2011-03-14 10:27:57.000000000 -0400
>> >+++ linux-2.6-block/block/blk-throttle.c	2011-03-14 10:30:47.267170956 -0400
>> >@@ -756,6 +756,15 @@ static void throtl_process_limit_change(
>> > 			" riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
>> > 			tg->iops[READ], tg->iops[WRITE]);
>> > 
>> >+		/*
>> >+		 * Restart the slices for both READ and WRITES. It
>> >+		 * might happen that a group's limit are dropped
>> >+		 * suddenly and we don't want to account recently
>> >+		 * dispatched IO with new low rate
>> >+		 */
>> >+		throtl_start_new_slice(td, tg, 0);
>> >+		throtl_start_new_slice(td, tg, 1);
>> >+
>> > 		if (throtl_tg_on_rr(tg))
>> > 			tg_update_disptime(td, tg);
>> > 	}
>> >@@ -821,7 +830,8 @@ throtl_schedule_delayed_work(struct thro
>> > 
>> > 	struct delayed_work *dwork = &td->throtl_work;
>> > 
>> >-	if (total_nr_queued(td) > 0) {
>> >+	/* schedule work if limits changed even if no bio is queued */
>> >+	if (total_nr_queued(td) > 0 || td->limits_changed) {
>> > 		/*
>> > 		 * We might have a work scheduled to be executed in future.
>> > 		 * Cancel that and schedule a new one.
>> >@@ -1002,6 +1012,19 @@ int blk_throtl_bio(struct request_queue 
>> > 	/* Bio is with-in rate limit of group */
>> > 	if (tg_may_dispatch(td, tg, bio, NULL)) {
>> > 		throtl_charge_bio(tg, bio);
>> >+
>> >+		/*
>> >+		 * We need to trim slice even when bios are not being queued
>> >+		 * otherwise it might happen that a bio is not queued for
>> >+		 * a long time and slice keeps on extending and trim is not
>> >+		 * called for a long time. Now if limits are reduced suddenly
>> >+		 * we take into account all the IO dispatched so far at new
>> >+		 * low rate and * newly queued IO gets a really long dispatch
>> >+		 * time.
>> >+		 *
>> >+		 * So keep on trimming slice even if bio is not queued.
>> >+		 */
>> >+		throtl_trim_slice(td, tg, rw);
>> > 		goto out;
>> > 	}
0

Re: Re: Re: blk-throttle.c : When limit is changed, must startanewslice

[Vivek Goyal]

On Tue, Mar 15, 2011 at 11:00:25PM +0800, Lina Lu wrote:

[..]
> Hi Vivek,
> I have test the following patch on 2.6.39 core branch of Jens's tree, the bug has
> been fixed. 

Thanks Lina. I will cleanup this patch and post for inclusion.

> 
> Can you tell me which patch makes the throtl_tg_on_rr() return 1 even if the limits
> are very high?

Following patches fixes the issue in Jens's tree.

commit de701c74a34005e637e1ca2634fbf28fd1debba2
Author: Vivek Goyal <vgoyal@redhat.com>
Date:   Mon Mar 7 21:09:32 2011 +0100

    blk-throttle: Some cleanups and race fixes in limit  update code

Thanks
Vivek

Re: blk-throttle.c : When limit is changed, must start a new slice

[Vivek Goyal]

On Tue, Mar 08, 2011 at 10:33:47PM +0800, lina wrote:
> >On Mon, Mar 07, 2011 at 11:05:49PM +0800, lina wrote:
> >> >On Sat, Mar 05, 2011 at 12:30:55AM +0800, lina wrote:
> >> >> Hi Vivek,    
> >> >>      Take you a litter time to read this letter, I think there seens  to 
> >> >>  be a small bug in blk-throttle.c.
> >> >>      This might happen that initially cgroup limit was greater than the 
> >> >>  device's physic ability, but later limit was under physic ability. In this
> >> >>  case, all of the bio in the device was throttled for serveral minutes. 
> >> >>   
> >> >>      I did some analysis as the following:
> >> >>      First, setting a very large bps on device lead all of the bio go
> >> >>  through. The slice is begin when test begin, and only extend but 
> >> >>  can not start new one. 
> >> >>      Second, change the limit under physic ability to make one bio 
> >> >>  queued. Once one bio queued, blk_throtl_bio func will call 
> >> >>  tg_update_disptime to estimate the delay time for the throtl_work.
> >> >>      As the slice is very old, there is a very large value in 
> >> >>  tg->bytes_disp[rw], and the tg->disptime is a long time after jiffies. 
> >> >>  During this time, all of the bio is queued. And the work_queue can 
> >> >>  not start, so tg->slice_start[rw] still can not be reset.
> >> >>   
> >> >>      Although after serveral minutes everything will be ok, but it still
> >> >>  seens no-good for users.
> >> >>      
> >> >>      I think it should start a new slice when the limit is changed. 
> >> >>      Here is my patch, please conrrect it if there is something wrong
> >> >>  follow.
> >>  >
> >> >
> >> >CCing to lkml. Lets keep all the testing and bug reports regarding
> >> >blkio throttling on mailing list.
> >> >
> >> >thanks for the bug report lina. I think this is a bug. I am not too keen
> >> >on restarting slice all the time upon limit change as somebody can exploit
> >> >that to get higher BW by doing it frequently. Can you try attached patch
> >> >and see if it solves your problem.
> >> 
> >>  Thank you for this patch, it can solve the problem. But there still has 5~10
> >>  seconds 0 bps in the test. I think this is because we first tirm the end of
> >>  slice, then new one, there is some latency. Do you have any idea to let the 
> >>  limit change work immediately or make less latency(maybe 1 or 2 seconds
> >>   
> >
> >Ok, if you are concerned about those few seconds, can you please try
> >following patch. I think starting a new slice is better when processing
> >limit change instead of in blk_throtl_bio().
> >
>   
>  Yes. It's better starting a new slice in processing limit change function. I
>  have test starting a new slice in blk_throtl_bio(), if I change the limit very
>  quickly, the bps go out of my control.

I thought about changing limits too quickly. Is this a problem? In many
setups a non-priviliged user will not be able to write to limit files so
he can not exploit it. root should not anyway do that. To get higher
bandwidth root can simply write to files and there is no reason to try
to write to reset limit frequently.

So resetting the slice upon limit change probably is not a huge problem then.

>   
>  Unfortunately, the following patch still has 5~10 seconds latency. I have no
>  idea to resolve this problem, it seens hard to find a more suitable func to
>  call throtl_start_new_slice().

So are you saying that following patch did not solve the latnecy issue?
Resetting slice upon limit change did not work for you?

Thanks
Vivek

> >
> >---
> > block/blk-throttle.c |   21 +++++++++++++++++++++
> > 1 file changed, 21 insertions(+)
> >
> >Index: linux-2.6/block/blk-throttle.c
> >===================================================================
> >--- linux-2.6.orig/block/blk-throttle.c 2011-03-04 13:59:45.000000000 -0500
> >+++ linux-2.6/block/blk-throttle.c 2011-03-07 10:54:30.639467538 -0500
> >@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
> >  " riops=%u wiops=%u", tg->bps[READ],
> >  tg->bps[WRITE], tg->iops[READ],
> >  tg->iops[WRITE]);
> >+ /*
> >+ * Restart the slices for both READ and WRITES. It
> >+ * might happen that a group's limit are dropped
> >+ * suddenly and we don't want to account recently
> >+ * dispatched IO with new low rate
> >+ */
> >+ throtl_start_new_slice(td, tg, 0);
> >+ throtl_start_new_slice(td, tg, 1);
> >  tg_update_disptime(td, tg);
> >  tg->limits_changed = false;
> >  }
> >@@ -1023,6 +1031,19 @@ int blk_throtl_bio(struct request_queue 
> >  /* Bio is with-in rate limit of group */
> >  if (tg_may_dispatch(td, tg, bio, NULL)) {
> >  throtl_charge_bio(tg, bio);
> >+
> >+ /*
> >+ * We need to trim slice even when bios are not being queued
> >+ * otherwise it might happen that a bio is not queued for
> >+ * a long time and slice keeps on extending and trim is not
> >+ * called for a long time. Now if limits are reduced suddenly
> >+ * we take into account all the IO dispatched so far at new
> >+ * low rate and * newly queued IO gets a really long dispatch
> >+ * time.
> >+ *
> >+ * So keep on trimming slice even if bio is not queued.
> >+ */
> >+ throtl_trim_slice(td, tg, rw);
> >  goto out;
> >  }
>  
> Thanks
>  Lina

Re: blk-throttle.c : When limit is changed, must start a new slice

[Vivek Goyal]

On Mon, Mar 07, 2011 at 11:05:49PM +0800, lina wrote:
> >On Sat, Mar 05, 2011 at 12:30:55AM +0800, lina wrote:
> >> Hi Vivek,    
> >>      Take you a litter time to read this letter, I think there seens  to 
> >>  be a small bug in blk-throttle.c.
> >>      This might happen that initially cgroup limit was greater than the 
> >>  device's physic ability, but later limit was under physic ability. In this
> >>  case, all of the bio in the device was throttled for serveral minutes. 
> >>   
> >>      I did some analysis as the following:
> >>      First, setting a very large bps on device lead all of the bio go
> >>  through. The slice is begin when test begin, and only extend but 
> >>  can not start new one. 
> >>      Second, change the limit under physic ability to make one bio 
> >>  queued. Once one bio queued, blk_throtl_bio func will call 
> >>  tg_update_disptime to estimate the delay time for the throtl_work.
> >>      As the slice is very old, there is a very large value in 
> >>  tg->bytes_disp[rw], and the tg->disptime is a long time after jiffies. 
> >>  During this time, all of the bio is queued. And the work_queue can 
> >>  not start, so tg->slice_start[rw] still can not be reset.
> >>   
> >>      Although after serveral minutes everything will be ok, but it still
> >>  seens no-good for users.
> >>      
> >>      I think it should start a new slice when the limit is changed. 
> >>      Here is my patch, please conrrect it if there is something wrong
> >>  follow.</:includetail>
>  >
> >
> >CCing to lkml. Lets keep all the testing and bug reports regarding
> >blkio throttling on mailing list.
> >
> >thanks for the bug report lina. I think this is a bug. I am not too keen
> >on restarting slice all the time upon limit change as somebody can exploit
> >that to get higher BW by doing it frequently. Can you try attached patch
> >and see if it solves your problem.
> </:includetail></:includetail>
>  Thank you for this patch, it can solve the problem. But there still has 5~10</:includetail>
>  seconds 0 bps in the test. I think this is because we first tirm the end of </:includetail>
>  slice, then new one, there is some latency. Do you have any idea to let </:includetail>the </:includetail>
>  limit change work immediately or make less latency(maybe 1 or 2 seconds)?</:includetail>
>  </:includetail> 

Ok, if you are concerned about those few seconds, can you please try
following patch. I think starting a new slice is better when processing
limit change instead of in blk_throtl_bio().

Thanks
Vivek


---
 block/blk-throttle.c |   21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

Index: linux-2.6/block/blk-throttle.c
===================================================================
--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
+++ linux-2.6/block/blk-throttle.c	2011-03-07 10:54:30.639467538 -0500
@@ -757,6 +757,14 @@ static void throtl_process_limit_change(
 				" riops=%u wiops=%u", tg->bps[READ],
 				tg->bps[WRITE], tg->iops[READ],
 				tg->iops[WRITE]);
+			/*
+			 * Restart the slices for both READ and WRITES. It
+			 * might happen that a group's limit are dropped
+			 * suddenly and we don't want to account recently
+			 * dispatched IO with new low rate
+			 */
+			throtl_start_new_slice(td, tg, 0);
+			throtl_start_new_slice(td, tg, 1);
 			tg_update_disptime(td, tg);
 			tg->limits_changed = false;
 		}
@@ -1023,6 +1031,19 @@ int blk_throtl_bio(struct request_queue 
 	/* Bio is with-in rate limit of group */
 	if (tg_may_dispatch(td, tg, bio, NULL)) {
 		throtl_charge_bio(tg, bio);
+
+		/*
+		 * We need to trim slice even when bios are not being queued
+		 * otherwise it might happen that a bio is not queued for
+		 * a long time and slice keeps on extending and trim is not
+		 * called for a long time. Now if limits are reduced suddenly
+		 * we take into account all the IO dispatched so far at new
+		 * low rate and * newly queued IO gets a really long dispatch
+		 * time.
+		 *
+		 * So keep on trimming slice even if bio is not queued.
+		 */
+		throtl_trim_slice(td, tg, rw);
 		goto out;
 	}
 

Re: blk-throttle.c : When limit is changed, must start a new slice

[Vivek Goyal]

On Sat, Mar 05, 2011 at 12:30:55AM +0800, lina wrote:
> Hi Vivek,    
>      Take you a litter time to read this letter, I think there seens  to 
>  be a small bug in blk-throttle.c.
>      This might happen that initially cgroup limit was greater than the 
>  device's physic ability, but later limit was under physic ability. In this
>  case, all of the bio in the device was throttled for serveral minutes. 
>   
>      I did some analysis as the following:
>      First, setting a very large bps on device lead all of the bio go
>  through. The slice is begin when test begin, and only extend but 
>  can not start new one. 
>      Second, change the limit under physic ability to make one bio 
>  queued. Once one bio queued, blk_throtl_bio func will call 
>  tg_update_disptime to estimate the delay time for the throtl_work.
>      As the slice is very old, there is a very large value in 
>  tg->bytes_disp[rw], and the tg->disptime is a long time after jiffies. 
>  During this time, all of the bio is queued. And the work_queue can 
>  not start, so tg->slice_start[rw] still can not be reset.
>   
>      Although after serveral minutes everything will be ok, but it still
>  seens no-good for users.
>      
>      I think it should start a new slice when the limit is changed. 
>      Here is my patch, please conrrect it if there is something wrong
>  follow.

CCing to lkml. Lets keep all the testing and bug reports regarding
blkio throttling on mailing list.

thanks for the bug report lina. I think this is a bug. I am not too keen
on restarting slice all the time upon limit change as somebody can exploit
that to get higher BW by doing it frequently. Can you try attached patch
and see if it solves your problem.

>   
>  --- block/blk-throttle.c
>  +++ block/blk-throttle.c
>  @@ -1027,6 +1027,9 @@
>           *biop = NULL;  
>           if (update_disptime) {
>  +                 if (tg->limits_changed)
>  +                 {
>  +                         throtl_start_new_slice(td, tg, rw);
>  +                 }
>                   tg_update_disptime(td, tg);
>                   throtl_schedule_next_dispatch(td);
>           }
>   
>   
>       There is two other problems, can you explain them to me? 
>   
>  First：
>       The blkio dir always like 
>   
>       blkio //patent dir
>         ---blkio.weight
>         ---blkio.throtl_write_bps_device
>         ...
>         ---tst1  //child dir
>               ---blkio.weight              ---blkio.throtl_write_bps_device
>               ...
>   
>       The child dir's throtl files like tst1/blkio.throttl_write_bps_device 
>  seens never be use. It maybe better that do not create these files.
>  If they are useful, please tell me. Thank you!

Throttle creates a flag hierarchy of groups internally. So even if you
create child of a child group, child will be treated as if children 
of root and will be throttled as per throttling rules.
>   
>  Second:
>       When apply weight policy to some pid, I must mkdir tst1 like 
>  above. If all of pid in tst1 is end, the pid in tasks will disappear, but
>  the tst1 dir is still here. I think here will be better if tst1 can be 
>  removed automatically. 
>       Throttle policy has no auto-clean capacity. If I create many 
>  linear dm devices, and apply throttle policy on them, then use 
>  'dmsetup remove--all'. Now I must clean the entries one by one in blkio.throtl_write_bps_device manually. If the throttle policy can 
>  clean the entry automatically when the device is no longger in the 
>  system, It seens to be perfect!

i think it is hard to remove rules also when devie is going away. One
easy way is to remove whole cgroup. Other would be that provide a 
simple command to delete all the rules from a file. 

say echo "0" > blkio.throttle.read_bps_device will remove all the rules
from the file. If that helps you, you can write one patch for that.

Thanks
Vivek

---
 block/blk-throttle.c |   13 +++++++++++++
 1 file changed, 13 insertions(+)

Index: linux-2.6/block/blk-throttle.c
===================================================================
--- linux-2.6.orig/block/blk-throttle.c	2011-03-04 13:59:45.000000000 -0500
+++ linux-2.6/block/blk-throttle.c	2011-03-04 14:05:31.542332242 -0500
@@ -1023,6 +1023,19 @@ int blk_throtl_bio(struct request_queue 
 	/* Bio is with-in rate limit of group */
 	if (tg_may_dispatch(td, tg, bio, NULL)) {
 		throtl_charge_bio(tg, bio);
+
+		/*
+		 * We need to trim slice even when bios are not being queued
+		 * otherwise it might happen that a bio is not queued for
+		 * a long time and slice keeps on extending and trim is not
+		 * called for a long time. Now if limits are reduced suddenly
+		 * we take into account all the IO dispatched so far at new
+		 * low rate and * newly queued IO gets a really long dispatch
+		 * time.
+		 *
+		 * So keep on trimming slice even if bio is not queued.
+		 */
+		throtl_trim_slice(td, tg, rw);
 		goto out;
 	}