[linux-dvb] High CPU load in "top" due to budget_av slot polling

[linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

Hello,

on 2.6.24 I run into a small issue with the budget_av driver:

Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module

Error description:
After the budget_av driver module is loaded (even without any DVB 
application), the CPU load average in 'top' rises to ~1, but in top no 
active tasks are found. After unloading the driver, the load decreases 
again to ~0.

Displaying the blocked tasks during high load with Alt-SysRq-W shows 
that the kdvb-ca kernel thread seems to be accounted as blocked when it 
polls for the CI slot status:
---------------------------------------------------
SysRq : Show Blocked State
task                        PC stack   pid father
kdvb-ca-0:0   D 0000000100d78984     0  2046      2
ffff81007ee73c70 0000000000000046 0000000000000000 0000000000000008
0000000000000800 0000000000000000 0000000000000286 ffff81007bebe080
ffff81007f8d47e0 ffff81007ee73c80 ffff81007ed91a00 0000000100d78984
Call Trace:
[<ffffffff80241da6>] __mod_timer+0xb6/0xd0
[<ffffffff804e52ff>] schedule_timeout+0x5f/0xd0
[<ffffffff80241890>] process_timeout+0x0/0x10
[<ffffffff80241dd6>] msleep+0x16/0x30
[<ffffffff880a0189>] :saa7146:saa7146_wait_for_debi_done+0x159/0x260
[<ffffffff880bba3a>] :budget_core:ttpci_budget_debiread+0x6a/0x140
[<ffffffff880d1cfb>] :budget_av:ciintf_poll_slot_status+0xbb/0x1c0
[<ffffffff880aa800>] :dvb_core:dvb_ca_en50221_thread+0x0/0xa10
[<ffffffff880a9339>] :dvb_core:dvb_ca_en50221_check_camstatus+0x59/0x100
[<ffffffff880aa8ba>] :dvb_core:dvb_ca_en50221_thread+0xba/0xa10
[<ffffffff8022f221>] update_curr+0x61/0xb0
[<ffffffff8022f221>] update_curr+0x61/0xb0
[<ffffffff802341e6>] dequeue_task_fair+0x46/0x80
[<ffffffff8022f79d>] __dequeue_entity+0x3d/0x50
[<ffffffff804e4a7e>] thread_return+0x3d/0x52f
[<ffffffff80288da4>] filp_close+0x54/0x90
[<ffffffff880aa800>] :dvb_core:dvb_ca_en50221_thread+0x0/0xa10
[<ffffffff8024d48b>] kthread+0x4b/0x80
[<ffffffff8020ca48>] child_rip+0xa/0x12
[<ffffffff8024d440>] kthread+0x0/0x80
[<ffffffff8020ca3e>] child_rip+0x0/0x12
---------------------------------------------------

Enabling debug traces shows that polling for the PSR in function 
'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
sleeps:

 > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
completion

Increasing the 250ms did not avoid the timeout. And as I understood from 
previous list mails, this timeout is normal when no CI/CAM is connected 
to the DEBI. However, for me the high frequency polling does not make 
sense if someone does not plan to plug in a CI/CAM.

When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
CPU load went down to 0. So this is some kind of workaround for me.

Please note: The high CPU load is not only a statistical issue due to 
the kernel accounting. It would put at least some unnecessary stress on 
PCI and CPU and e.g. also has the effect that cron tasks with idle guard 
condition would not start, so on midterm this should be fixed.

Finally, my questions:
1. Did I understand this right, that the behaviour above is expected 
when no CI is connected?

2. Are all budget_av cards unable to check CAM slot state via interrupt 
  for HW reasons (as budget_ci does)?

3. Would it be possible to substitute the current PSR DEBI_S polling 
with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
seems to do this for its DEBI DMA)? Or was this not considered worthy, 
due to the expected short waiting period and the tricky IER handling? 
The code does not state further thoughts about this.

4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
the 100ms poll timer intended? (I found the recent patch to this code in 
the mailing list end of last year)

5. If we would be restricted to poll with high frequency: Why not at 
least allow users without CI to disable polling for slots or change the 
interval, e.g. via module options?

Regards,
Robert Schedel


_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

Hello all,

meanwhile I also found old threads from August 2007 and earlier, named 
"System load raises when budget_av is loaded" and "System load raises 
when budget_av is loaded", which describe exactly the same issue. 
However, both were written before the latest code changes.

The latest debi_done code change from November 2007 does not use any 
DEBI_E checks anymore. To my understanding this SAA7146 event indicates 
  DEBI errors like timeout via PCI and should be received by the driver 
  much earlier than those 250ms (=end of cycle). Seems like I have to 
play around with this myself.

Time for a bugzilla ticket?

Regards,
Robert Schedel




_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Oliver Endriss]

Robert Schedel wrote:
> Hello all,
> 
> meanwhile I also found old threads from August 2007 and earlier, named 
> "System load raises when budget_av is loaded" and "System load raises 
> when budget_av is loaded", which describe exactly the same issue. 
> However, both were written before the latest code changes.
> 
> The latest debi_done code change from November 2007 does not use any 
> DEBI_E checks anymore. To my understanding this SAA7146 event indicates 
>   DEBI errors like timeout via PCI and should be received by the driver 
>   much earlier than those 250ms (=end of cycle). Seems like I have to 
> play around with this myself.

IIRC there was a problem with the DEBI_E flag but I don't remember what
it was exactly.

> Time for a bugzilla ticket?

You can do this if you prefer (but this will neither fix the bug nor
have any other positive effect).

CU
Oliver

-- 
----------------------------------------------------------------
VDR Remote Plugin 0.4.0: http://www.escape-edv.de/endriss/vdr/
----------------------------------------------------------------

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Oliver Endriss]

Robert Schedel wrote:
> Hello,
> 
> on 2.6.24 I run into a small issue with the budget_av driver:
> 
> Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
> Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module
> 
> Error description:
> After the budget_av driver module is loaded (even without any DVB 
> application), the CPU load average in 'top' rises to ~1, but in top no 
> active tasks are found. After unloading the driver, the load decreases 
> again to ~0.
> 
> Displaying the blocked tasks during high load with Alt-SysRq-W shows 
> that the kdvb-ca kernel thread seems to be accounted as blocked when it 
> polls for the CI slot status:
> ---------------------------------------------------
> SysRq : Show Blocked State
> task                        PC stack   pid father
> kdvb-ca-0:0   D 0000000100d78984     0  2046      2
> ffff81007ee73c70 0000000000000046 0000000000000000 0000000000000008
> 0000000000000800 0000000000000000 0000000000000286 ffff81007bebe080
> ffff81007f8d47e0 ffff81007ee73c80 ffff81007ed91a00 0000000100d78984
> Call Trace:
> [<ffffffff80241da6>] __mod_timer+0xb6/0xd0
> [<ffffffff804e52ff>] schedule_timeout+0x5f/0xd0
> [<ffffffff80241890>] process_timeout+0x0/0x10
> [<ffffffff80241dd6>] msleep+0x16/0x30
> [<ffffffff880a0189>] :saa7146:saa7146_wait_for_debi_done+0x159/0x260
> [<ffffffff880bba3a>] :budget_core:ttpci_budget_debiread+0x6a/0x140
> [<ffffffff880d1cfb>] :budget_av:ciintf_poll_slot_status+0xbb/0x1c0
> [<ffffffff880aa800>] :dvb_core:dvb_ca_en50221_thread+0x0/0xa10
> [<ffffffff880a9339>] :dvb_core:dvb_ca_en50221_check_camstatus+0x59/0x100
> [<ffffffff880aa8ba>] :dvb_core:dvb_ca_en50221_thread+0xba/0xa10
> [<ffffffff8022f221>] update_curr+0x61/0xb0
> [<ffffffff8022f221>] update_curr+0x61/0xb0
> [<ffffffff802341e6>] dequeue_task_fair+0x46/0x80
> [<ffffffff8022f79d>] __dequeue_entity+0x3d/0x50
> [<ffffffff804e4a7e>] thread_return+0x3d/0x52f
> [<ffffffff80288da4>] filp_close+0x54/0x90
> [<ffffffff880aa800>] :dvb_core:dvb_ca_en50221_thread+0x0/0xa10
> [<ffffffff8024d48b>] kthread+0x4b/0x80
> [<ffffffff8020ca48>] child_rip+0xa/0x12
> [<ffffffff8024d440>] kthread+0x0/0x80
> [<ffffffff8020ca3e>] child_rip+0x0/0x12
> ---------------------------------------------------
> 
> Enabling debug traces shows that polling for the PSR in function 
> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
> sleeps:
> 
>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
> completion
> 
> Increasing the 250ms did not avoid the timeout. And as I understood from 
> previous list mails, this timeout is normal when no CI/CAM is connected 
> to the DEBI. However, for me the high frequency polling does not make 
> sense if someone does not plan to plug in a CI/CAM.
> 
> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
> CPU load went down to 0. So this is some kind of workaround for me.

Afaics the polling interval could be increased to something like 5s or
10s if (and only if) the slot is empty. Could you provide a patch?

> Please note: The high CPU load is not only a statistical issue due to 
> the kernel accounting. It would put at least some unnecessary stress on 
> PCI and CPU and e.g. also has the effect that cron tasks with idle guard 
> condition would not start, so on midterm this should be fixed.

Agreed.

> Finally, my questions:
> 1. Did I understand this right, that the behaviour above is expected 
> when no CI is connected?

Yes, but afaics the polling interval is way too short.

> 2. Are all budget_av cards unable to check CAM slot state via interrupt 
>   for HW reasons (as budget_ci does)?

I don't have any budget-av hardware, so I don't know.
But I think that Andrew(?) had a good reason to implement it this way.
(In contrast the budget-ci driver was always interrupt-driven.)

If someone finds out that a given card can operate in interrupt mode,
it should be changed for this card. Patches are welcome. ;-)

> 3. Would it be possible to substitute the current PSR DEBI_S polling 
> with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
> seems to do this for its DEBI DMA)? Or was this not considered worthy, 
> due to the expected short waiting period and the tricky IER handling? 
> The code does not state further thoughts about this.

The av7110 driver uses interrupt mode for buffer transfers in dma mode.
It does not make much sense to use interrupt mode for single-word
transfers, because the single-word transfers are very fast.
But I understand that the timeout causes a problem in this case.

> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
> the 100ms poll timer intended? (I found the recent patch to this code in 
> the mailing list end of last year)

That patch was applied to reduce the load on the pci bus in busy-wait
mode. Basically it did not change anything for cam polling. (In fact I
was not aware that the CAM was polled every 100ms. Imho this should be
fixed.)

> 5. If we would be restricted to poll with high frequency: Why not at 
> least allow users without CI to disable polling for slots or change the 
> interval, e.g. via module options?

Module options are evil. ;-)
We should do this only if everything else fails.

CU
Oliver

-- 
----------------------------------------------------------------
VDR Remote Plugin 0.4.0: http://www.escape-edv.de/endriss/vdr/
----------------------------------------------------------------

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

[-- Attachment #1: Type: text/plain, Size: 5321 bytes --]

Oliver Endriss wrote:
> Robert Schedel wrote:
>> Hello,
>>
>> on 2.6.24 I run into a small issue with the budget_av driver:
>>
>> Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
>> Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module
>>
>> Error description:
>> After the budget_av driver module is loaded (even without any DVB 
>> application), the CPU load average in 'top' rises to ~1, but in top no 
>> active tasks are found. After unloading the driver, the load decreases 
>> again to ~0.
>>
>> Displaying the blocked tasks during high load with Alt-SysRq-W shows 
>> that the kdvb-ca kernel thread seems to be accounted as blocked when it 
>> polls for the CI slot status:
>> ---------------------------------------------------
>> [...]
>> ---------------------------------------------------
>>
>> Enabling debug traces shows that polling for the PSR in function 
>> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
>> sleeps:
>>
>>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
>> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
>> completion
>>
>> Increasing the 250ms did not avoid the timeout. And as I understood from 
>> previous list mails, this timeout is normal when no CI/CAM is connected 
>> to the DEBI. However, for me the high frequency polling does not make 
>> sense if someone does not plan to plug in a CI/CAM.
>>
>> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
>> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
>> CPU load went down to 0. So this is some kind of workaround for me.
> 
> Afaics the polling interval could be increased to something like 5s or
> 10s if (and only if) the slot is empty. Could you provide a patch?

Attached a patch for 2.6.24.4. Opinions?

Unfortunately, a 5s poll interval for empty slots still results in a
load average of about 0,06 (~ 250ms/5s).

Increasing to 10s would decrease the load and be fine for people without
CAM, but increase the delay for those users inserting CAMs during
runtime. 5s sounds like a reasonable tradeoff.

Is the 250ms timeout an approved limit? Decreasing it would push the
load further down. Probably it still has to cover slow CAMs as well as a
stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
statements myself.

>> Finally, my questions:
>> 1. Did I understand this right, that the behaviour above is expected 
>> when no CI is connected?
> 
> Yes, but afaics the polling interval is way too short.
> 
>> 2. Are all budget_av cards unable to check CAM slot state via interrupt 
>>   for HW reasons (as budget_ci does)?
> 
> I don't have any budget-av hardware, so I don't know.
> But I think that Andrew(?) had a good reason to implement it this way.
> (In contrast the budget-ci driver was always interrupt-driven.)
> 
> If someone finds out that a given card can operate in interrupt mode,
> it should be changed for this card. Patches are welcome. ;-)

My impression is, due to the different GPIO layout there is no way to
get a CAM change IRQ. But it seems difficult to get information about
the HW architecture of cards at all.

Regarding DEBI_E: Just found a av7110 code comment which also reflects
what my recent tests showed:
          /* Note: Don't try to handle the DEBI error irq (MASK_18), in
           * intel mode the timeout is asserted all the time...
           */

So really only DEBI_S would be left, see below.

>> 3. Would it be possible to substitute the current PSR DEBI_S polling 
>> with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
>> seems to do this for its DEBI DMA)? Or was this not considered worthy, 
>> due to the expected short waiting period and the tricky IER handling? 
>> The code does not state further thoughts about this.
> 
> The av7110 driver uses interrupt mode for buffer transfers in dma mode.
> It does not make much sense to use interrupt mode for single-word
> transfers, because the single-word transfers are very fast.
> But I understand that the timeout causes a problem in this case.

OK, interrupts of course decrease performance in the "sunny day" cases
(=communication with inserted card in slot state READY). Having both
approaches (interrupt when slotstate empty, later polling) would combine
all benefits but also be somewhat crazy.

>> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
>> the 100ms poll timer intended? (I found the recent patch to this code in 
>> the mailing list end of last year)
> 
> That patch was applied to reduce the load on the pci bus in busy-wait
> mode. Basically it did not change anything for cam polling. (In fact I
> was not aware that the CAM was polled every 100ms. Imho this should be
> fixed.)

Only wondered whether the 250ms might have been smaller in former driver
versions.

> 
>> 5. If we would be restricted to poll with high frequency: Why not at 
>> least allow users without CI to disable polling for slots or change the 
>> interval, e.g. via module options?
> 
> Module options are evil. ;-)
> We should do this only if everything else fails.

Correct.

P.S. Had already created bugzilla #10459 for tracking.
Sorry for late response, previous mail was stuck in spam filter.

Regards,
Robert


[-- Attachment #2: incr-empty-ca-slot-poll-2.6.24.4.patch --]
[-- Type: text/plain, Size: 2485 bytes --]

From: Robert Schedel <r.schedel@yahoo.de>

This change addresses kernel bug #10459: In kernel 2.6.24 the
budget_av driver polls for an CI slot in 100ms intervals (because no
interrupt solution for budget_av cards is feasible due to HW reasons).
If no CI/CAM is connected to the DVB card, polling times out only after 250ms.
This periodic polling leads to high CPU load.

The change increases the polling interval for empty slots from 100ms to 5s.
Intervals for remaining slot states (invalid, in progress, ready) are unchanged,
as they are either temporary conditions or no timeout should occur.

Signed-off-by: Robert Schedel <r.schedel@yahoo.de>

---
 drivers/media/dvb/dvb-core/dvb_ca_en50221.c |   20 ++++++++++++------
  1 file changed, 14 insertions(+), 6 deletions(-)

diff -rupN linux-2.6.24.4-orig/drivers/media/dvb/dvb-core/dvb_ca_en50221.c linux-2.6.24.4/drivers/media/dvb/dvb-core/dvb_ca_en50221.c
--- linux-2.6.24.4-orig/drivers/media/dvb/dvb-core/dvb_ca_en50221.c	2008-01-24 23:58:37.000000000 +0100
+++ linux-2.6.24.4/drivers/media/dvb/dvb-core/dvb_ca_en50221.c	2008-04-16 21:41:17.000000000 +0200
@@ -910,14 +910,22 @@ static void dvb_ca_en50221_thread_update
 	int curdelay = 100000000;
 	int slot;
 
+	/* Beware of too high polling frequency, because one polling
+	 * call might take several hundred milliseconds until timeout!
+	 */
 	for (slot = 0; slot < ca->slot_count; slot++) {
 		switch (ca->slot_info[slot].slot_state) {
 		default:
 		case DVB_CA_SLOTSTATE_NONE:
+			delay = HZ * 60;  // 60s
+			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE)) {
+				delay = HZ * 5;  // 5s
+			}
+			break;
 		case DVB_CA_SLOTSTATE_INVALID:
-			delay = HZ * 60;
+			delay = HZ * 60;  // 60s
 			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE)) {
-				delay = HZ / 10;
+				delay = HZ / 10;  // 100ms
 			}
 			break;
 
@@ -926,18 +934,18 @@ static void dvb_ca_en50221_thread_update
 		case DVB_CA_SLOTSTATE_VALIDATE:
 		case DVB_CA_SLOTSTATE_WAITFR:
 		case DVB_CA_SLOTSTATE_LINKINIT:
-			delay = HZ / 10;
+			delay = HZ / 10;  // 100ms
 			break;
 
 		case DVB_CA_SLOTSTATE_RUNNING:
-			delay = HZ * 60;
+			delay = HZ * 60;  // 60s
 			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE)) {
-				delay = HZ / 10;
+				delay = HZ / 10;  // 100ms
 			}
 			if (ca->open) {
 				if ((!ca->slot_info[slot].da_irq_supported) ||
 				    (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_DA))) {
-					delay = HZ / 10;
+					delay = HZ / 10;  // 100ms
 				}
 			}
 			break;


[-- Attachment #3: Type: text/plain, Size: 150 bytes --]

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Guy Martin]

Hi Robert,

This patch looks great. I tested it with Easywatch DVB-C (1011:0019) without CI/CAM and load is 0 now while it was ~0.6 constantly before.

uname -a : Linux devbox 2.6.25 #1 Thu Apr 17 12:18:32 CEST 2008 i686 AMD Athlon(tm) Processor AuthenticAMD GNU/Linux

Cheers,
  Guy

On Wed, 16 Apr 2008 23:28:02 +0200
Robert Schedel <r.schedel@yahoo.de> wrote:

> 
> Attached a patch for 2.6.24.4. Opinions?
> 

-- 
Guy Martin
Gentoo Linux - HPPA port lead


_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

Robert Schedel wrote:

> Is the 250ms timeout an approved limit? Decreasing it would push the
> load further down. Probably it still has to cover slow CAMs as well as a
> stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
> statements myself.

Just got another idea to improve the code: Function 
"saa7146_wait_for_debi_done_sleep" could be reworked to use what is 
known as "truncated binary exponential backoff" algorithm. IOW, on each 
sleep duplicate the period from 1ms until a fixed maximum, e.g. 32ms. 
This way polling ends fast for those users with fast bus/CAM, and those 
requiring 200ms due to slow bus/CAM should not worry about e.g. 216ms 
response time.

My first tests look promising (load goes down to 0). However, is not the 
simple BEB algorithm already patented?

Regards,
Robert


_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Oliver Endriss]

[-- Attachment #1: Type: text/plain, Size: 1229 bytes --]

Robert Schedel wrote:
> Robert Schedel wrote:
> 
> > Is the 250ms timeout an approved limit? Decreasing it would push the
> > load further down. Probably it still has to cover slow CAMs as well as a
> > stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
> > statements myself.
> 
> Just got another idea to improve the code: Function 
> "saa7146_wait_for_debi_done_sleep" could be reworked to use what is 
> known as "truncated binary exponential backoff" algorithm. IOW, on each 
> sleep duplicate the period from 1ms until a fixed maximum, e.g. 32ms. 
> This way polling ends fast for those users with fast bus/CAM, and those 
> requiring 200ms due to slow bus/CAM should not worry about e.g. 216ms 
> response time.
> 
> My first tests look promising (load goes down to 0). However, is not the 
> simple BEB algorithm already patented?

Load should go down to 0 if the sleep call does not busy-wait.

Please test whether the attached code fixes the problem.
Btw, I will not claim a patent for that. :D

CU
Oliver

-- 
----------------------------------------------------------------
VDR Remote Plugin 0.4.0: http://www.escape-edv.de/endriss/vdr/
----------------------------------------------------------------

[-- Attachment #2: saa7146_sleep.diff --]
[-- Type: text/x-diff, Size: 1029 bytes --]

diff -r f500c56d9064 linux/drivers/media/common/saa7146_core.c
--- a/linux/drivers/media/common/saa7146_core.c	Fri Apr 11 11:22:15 2008 +0200
+++ b/linux/drivers/media/common/saa7146_core.c	Fri Apr 18 02:25:39 2008 +0200
@@ -64,8 +64,10 @@ static inline int saa7146_wait_for_debi_
 {
 	unsigned long timeout;
 	int err;
+	bool slow;
 
 	/* wait for registers to be programmed */
+	slow = false;
 	timeout = jiffies + usecs_to_jiffies(us1);
 	while (1) {
 		err = time_after(jiffies, timeout);
@@ -77,10 +79,12 @@ static inline int saa7146_wait_for_debi_
 					dev->name, __func__);
 			return -ETIMEDOUT;
 		}
-		msleep(1);
+		msleep(slow ? 10 : 1);
+		slow = true;
 	}
 
 	/* wait for transfer to complete */
+	slow = false;
 	timeout = jiffies + usecs_to_jiffies(us2);
 	while (1) {
 		err = time_after(jiffies, timeout);
@@ -92,7 +96,8 @@ static inline int saa7146_wait_for_debi_
 				"completion\n",	dev->name, __func__));
 			return -ETIMEDOUT;
 		}
-		msleep(1);
+		msleep(slow ? 10 : 1);
+		slow = true;
 	}
 
 	return 0;

[-- Attachment #3: Type: text/plain, Size: 150 bytes --]

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

Oliver Endriss wrote:
> Robert Schedel wrote:
>> Robert Schedel wrote:
>>
>>> Is the 250ms timeout an approved limit? Decreasing it would push the
>>> load further down. Probably it still has to cover slow CAMs as well as a
>>> stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
>>> statements myself.
>> Just got another idea to improve the code: Function 
>> "saa7146_wait_for_debi_done_sleep" could be reworked to use what is 
>> known as "truncated binary exponential backoff" algorithm. IOW, on each 
>> sleep duplicate the period from 1ms until a fixed maximum, e.g. 32ms. 
>> This way polling ends fast for those users with fast bus/CAM, and those 
>> requiring 200ms due to slow bus/CAM should not worry about e.g. 216ms 
>> response time.
>>
>> My first tests look promising (load goes down to 0). However, is not the 
>> simple BEB algorithm already patented?
> 
> Load should go down to 0 if the sleep call does not busy-wait.
> 
> Please test whether the attached code fixes the problem.
> Btw, I will not claim a patent for that. :D

OK, I just took the time to make a more reliable test series (because 
load measurements varied). All nonessential system processes and modules 
were terminated before the test. Basically, only the login shell and the 
budget_av module were left. 1 minute uptime was used for measurements.

Kernel: Linux 2.6.25
HW: Athlon 64 X2 3800+, Satelco EasyWatch DVB-C (as before)

1. Original module budget_av is loaded:
Load: ~0,6-0,8

2. Module + Patch "saa7146_sleep.diff" (1ms/10ms polling intervals in 
debi_done function):
Load: ~0,6-0,8 (same as in 1., no difference visible)

3. Module + Patch "incr-empty-ca-slot-poll-2.6.24.4.patch" (5s polling 
timer on slot state EMPTY):
Load: Decays to 0,02, but after about 105s always a spike to 0,10, then 
again decays to 0,02, and so on

4. Module + Patch "incr-empty-ca-slot-poll" + "saa7146_sleep.diff":
Same as 3.

5. Module + Patch "incr-empty-ca-slot-poll" + "binary exponential backoff":
Same as 3.

6. Module budget_av is unloaded:
Load constantly stays at 0, no spikes

Bottomline for me:
- Increasing the poll timer from 100ms, e.g. to 5s, makes sense. 
Changing the polling intervals in the debi_done function, however, makes 
no difference (unlike my previous assumption which was caused by the 
ugly variations).
- There seems to be a spike in the CPU load, each ~105s, but only when 
budget_av is loaded. I cannot explain it (maybe some frontend background 
functions), but it is no issue for me.

Regards,
Robert


_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Oliver Endriss]

Robert Schedel wrote:
> Oliver Endriss wrote:
> > Robert Schedel wrote:
> >> Hello,
> >>
> >> on 2.6.24 I run into a small issue with the budget_av driver:
> >>
> >> Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
> >> Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module
> >>
> >> Error description:
> >> After the budget_av driver module is loaded (even without any DVB 
> >> application), the CPU load average in 'top' rises to ~1, but in top no 
> >> active tasks are found. After unloading the driver, the load decreases 
> >> again to ~0.
> >>
> >> Displaying the blocked tasks during high load with Alt-SysRq-W shows 
> >> that the kdvb-ca kernel thread seems to be accounted as blocked when it 
> >> polls for the CI slot status:
> >> ---------------------------------------------------
> >> [...]
> >> ---------------------------------------------------
> >>
> >> Enabling debug traces shows that polling for the PSR in function 
> >> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
> >> sleeps:
> >>
> >>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
> >> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
> >> completion
> >>
> >> Increasing the 250ms did not avoid the timeout. And as I understood from 
> >> previous list mails, this timeout is normal when no CI/CAM is connected 
> >> to the DEBI. However, for me the high frequency polling does not make 
> >> sense if someone does not plan to plug in a CI/CAM.
> >>
> >> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
> >> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
> >> CPU load went down to 0. So this is some kind of workaround for me.
> > 
> > Afaics the polling interval could be increased to something like 5s or
> > 10s if (and only if) the slot is empty. Could you provide a patch?
> 
> Attached a patch for 2.6.24.4. Opinions?

Basically it should work but it has to be tested with CI/CAM, too.
Furthermore it is not sufficient to test with budget-av because many
other drivers will be affected.

So I would prefer a patch which does not touch behaviour for other card
drivers (if possible).

Please note for 'final' patches:
Always run 'checkpatch.pl' and fix the errors.
Sorry for that. :-(

> Unfortunately, a 5s poll interval for empty slots still results in a
> load average of about 0,06 (~ 250ms/5s).
> 
> Increasing to 10s would decrease the load and be fine for people without
> CAM, but increase the delay for those users inserting CAMs during
> runtime. 5s sounds like a reasonable tradeoff.
> 
> Is the 250ms timeout an approved limit? Decreasing it would push the
> load further down. Probably it still has to cover slow CAMs as well as a
> stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
> statements myself.
> 
> >> Finally, my questions:
> >> 1. Did I understand this right, that the behaviour above is expected 
> >> when no CI is connected?
> > 
> > Yes, but afaics the polling interval is way too short.
> > 
> >> 2. Are all budget_av cards unable to check CAM slot state via interrupt 
> >>   for HW reasons (as budget_ci does)?
> > 
> > I don't have any budget-av hardware, so I don't know.
> > But I think that Andrew(?) had a good reason to implement it this way.
> > (In contrast the budget-ci driver was always interrupt-driven.)
> > 
> > If someone finds out that a given card can operate in interrupt mode,
> > it should be changed for this card. Patches are welcome. ;-)
> 
> My impression is, due to the different GPIO layout there is no way to
> get a CAM change IRQ. But it seems difficult to get information about
> the HW architecture of cards at all.

For most cards we have no hw info. :-(

> Regarding DEBI_E: Just found a av7110 code comment which also reflects
> what my recent tests showed:
>           /* Note: Don't try to handle the DEBI error irq (MASK_18), in
>            * intel mode the timeout is asserted all the time...
>            */
> 
> So really only DEBI_S would be left, see below.

Did you check whether DEBI_S and/or DEBI_E are ever asserted with your
setup? If not, an interrupt would never occur anyway...

> >> 3. Would it be possible to substitute the current PSR DEBI_S polling 
> >> with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
> >> seems to do this for its DEBI DMA)? Or was this not considered worthy, 
> >> due to the expected short waiting period and the tricky IER handling? 
> >> The code does not state further thoughts about this.
> > 
> > The av7110 driver uses interrupt mode for buffer transfers in dma mode.
> > It does not make much sense to use interrupt mode for single-word
> > transfers, because the single-word transfers are very fast.
> > But I understand that the timeout causes a problem in this case.
> 
> OK, interrupts of course decrease performance in the "sunny day" cases
> (=communication with inserted card in slot state READY). Having both
> approaches (interrupt when slotstate empty, later polling) would combine
> all benefits but also be somewhat crazy.
> 
> >> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
> >> the 100ms poll timer intended? (I found the recent patch to this code in 
> >> the mailing list end of last year)
> > 
> > That patch was applied to reduce the load on the pci bus in busy-wait
> > mode. Basically it did not change anything for cam polling. (In fact I
> > was not aware that the CAM was polled every 100ms. Imho this should be
> > fixed.)
> 
> Only wondered whether the 250ms might have been smaller in former driver
> versions.

Iirc it should be even worse with older drivers.

Basically the 250ms timeout is just a last resort to escape from the
loop, if the debi transfer hangs for some reason. We might try to reduce
the timeout but I don't know how far we can go. (Touching 'magic' values
might be dangeous.)

CU
Oliver

-- 
----------------------------------------------------------------
VDR Remote Plugin 0.4.0: http://www.escape-edv.de/endriss/vdr/
----------------------------------------------------------------

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[hermann pitton]

Am Freitag, den 18.04.2008, 03:21 +0200 schrieb Oliver Endriss:
> Robert Schedel wrote:
> > Oliver Endriss wrote:
> > > Robert Schedel wrote:
> > >> Hello,
> > >>
> > >> on 2.6.24 I run into a small issue with the budget_av driver:
> > >>
> > >> Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
> > >> Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module
> > >>
> > >> Error description:
> > >> After the budget_av driver module is loaded (even without any DVB 
> > >> application), the CPU load average in 'top' rises to ~1, but in top no 
> > >> active tasks are found. After unloading the driver, the load decreases 
> > >> again to ~0.
> > >>
> > >> Displaying the blocked tasks during high load with Alt-SysRq-W shows 
> > >> that the kdvb-ca kernel thread seems to be accounted as blocked when it 
> > >> polls for the CI slot status:
> > >> ---------------------------------------------------
> > >> [...]
> > >> ---------------------------------------------------
> > >>
> > >> Enabling debug traces shows that polling for the PSR in function 
> > >> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
> > >> sleeps:
> > >>
> > >>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
> > >> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
> > >> completion
> > >>
> > >> Increasing the 250ms did not avoid the timeout. And as I understood from 
> > >> previous list mails, this timeout is normal when no CI/CAM is connected 
> > >> to the DEBI. However, for me the high frequency polling does not make 
> > >> sense if someone does not plan to plug in a CI/CAM.
> > >>
> > >> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
> > >> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
> > >> CPU load went down to 0. So this is some kind of workaround for me.
> > > 
> > > Afaics the polling interval could be increased to something like 5s or
> > > 10s if (and only if) the slot is empty. Could you provide a patch?
> > 
> > Attached a patch for 2.6.24.4. Opinions?
> 
> Basically it should work but it has to be tested with CI/CAM, too.
> Furthermore it is not sufficient to test with budget-av because many
> other drivers will be affected.
> 
> So I would prefer a patch which does not touch behaviour for other card
> drivers (if possible).
> 
> Please note for 'final' patches:
> Always run 'checkpatch.pl' and fix the errors.
> Sorry for that. :-(
> 
> > Unfortunately, a 5s poll interval for empty slots still results in a
> > load average of about 0,06 (~ 250ms/5s).
> > 
> > Increasing to 10s would decrease the load and be fine for people without
> > CAM, but increase the delay for those users inserting CAMs during
> > runtime. 5s sounds like a reasonable tradeoff.
> > 
> > Is the 250ms timeout an approved limit? Decreasing it would push the
> > load further down. Probably it still has to cover slow CAMs as well as a
> > stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
> > statements myself.
> > 
> > >> Finally, my questions:
> > >> 1. Did I understand this right, that the behaviour above is expected 
> > >> when no CI is connected?
> > > 
> > > Yes, but afaics the polling interval is way too short.
> > > 
> > >> 2. Are all budget_av cards unable to check CAM slot state via interrupt 
> > >>   for HW reasons (as budget_ci does)?
> > > 
> > > I don't have any budget-av hardware, so I don't know.
> > > But I think that Andrew(?) had a good reason to implement it this way.
> > > (In contrast the budget-ci driver was always interrupt-driven.)
> > > 
> > > If someone finds out that a given card can operate in interrupt mode,
> > > it should be changed for this card. Patches are welcome. ;-)
> > 
> > My impression is, due to the different GPIO layout there is no way to
> > get a CAM change IRQ. But it seems difficult to get information about
> > the HW architecture of cards at all.
> 
> For most cards we have no hw info. :-(
> 
> > Regarding DEBI_E: Just found a av7110 code comment which also reflects
> > what my recent tests showed:
> >           /* Note: Don't try to handle the DEBI error irq (MASK_18), in
> >            * intel mode the timeout is asserted all the time...
> >            */
> > 
> > So really only DEBI_S would be left, see below.
> 
> Did you check whether DEBI_S and/or DEBI_E are ever asserted with your
> setup? If not, an interrupt would never occur anyway...
> 
> > >> 3. Would it be possible to substitute the current PSR DEBI_S polling 
> > >> with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
> > >> seems to do this for its DEBI DMA)? Or was this not considered worthy, 
> > >> due to the expected short waiting period and the tricky IER handling? 
> > >> The code does not state further thoughts about this.
> > > 
> > > The av7110 driver uses interrupt mode for buffer transfers in dma mode.
> > > It does not make much sense to use interrupt mode for single-word
> > > transfers, because the single-word transfers are very fast.
> > > But I understand that the timeout causes a problem in this case.
> > 
> > OK, interrupts of course decrease performance in the "sunny day" cases
> > (=communication with inserted card in slot state READY). Having both
> > approaches (interrupt when slotstate empty, later polling) would combine
> > all benefits but also be somewhat crazy.
> > 
> > >> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
> > >> the 100ms poll timer intended? (I found the recent patch to this code in 
> > >> the mailing list end of last year)
> > > 
> > > That patch was applied to reduce the load on the pci bus in busy-wait
> > > mode. Basically it did not change anything for cam polling. (In fact I
> > > was not aware that the CAM was polled every 100ms. Imho this should be
> > > fixed.)
> > 
> > Only wondered whether the 250ms might have been smaller in former driver
> > versions.
> 
> Iirc it should be even worse with older drivers.
> 
> Basically the 250ms timeout is just a last resort to escape from the
> loop, if the debi transfer hangs for some reason. We might try to reduce
> the timeout but I don't know how far we can go. (Touching 'magic' values
> might be dangeous.)
> 
> CU
> Oliver
> 

Hi,

what are magic values?

Simply something retrieved under NDA with even that hardware not in
established/stable conditions yet, most likely.

If the commercial research facilities are not established for that,
because too expensive, it likely goes out to some university or major
research project around the globe, which recently collected some
external money for something related, to stay alive, and then please
could solve that minor issue on the run for free.

We had lots of high quality contributions in the past, seemingly coming
out of nothing ...

If it really counts, Open Source is a must and everybody knows it.

Cheers,
Hermann






 


_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Robert Schedel]

[-- Attachment #1: Type: text/plain, Size: 4095 bytes --]

Oliver Endriss wrote:
> Robert Schedel wrote:
>> Oliver Endriss wrote:
>>> Robert Schedel wrote:
>>>>
>>>> Enabling debug traces shows that polling for the PSR in function 
>>>> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
>>>> sleeps:
>>>>
>>>>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
>>>> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
>>>> completion
>>>>
>>>> Increasing the 250ms did not avoid the timeout. And as I understood from 
>>>> previous list mails, this timeout is normal when no CI/CAM is connected 
>>>> to the DEBI. However, for me the high frequency polling does not make 
>>>> sense if someone does not plan to plug in a CI/CAM.
>>>>
>>>> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
>>>> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
>>>> CPU load went down to 0. So this is some kind of workaround for me.
>>> Afaics the polling interval could be increased to something like 5s or
>>> 10s if (and only if) the slot is empty. Could you provide a patch?
>> Attached a patch for 2.6.24.4. Opinions?
> 
> Basically it should work but it has to be tested with CI/CAM, too.

Correct, unfortunately I cannot test it against a CI.

> Furthermore it is not sufficient to test with budget-av because many
> other drivers will be affected.
> 
> So I would prefer a patch which does not touch behaviour for other card
> drivers (if possible).

To my understanding of the DVB code dvb_ca_en50221 is only referenced by 
budget_av and budget_ci, at least in the vanilla kernel 2.6.25. The 
patch only changes the timer for slot state EMPTY if 
DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE is not set, which is for
1) budget_av
2) budget_ci if the CI firmware version is 0xa2 (because IRQs for CAM 
change are not supported in this version)

And those two cases are probably affected by the load issue and should 
be fixed.

Of course, we could add another DVB_CA_EN50221 flag solely for budget_av 
to exclude case 2), but does this make sense? Who is available to test 
against a budget_ci with FW=0xa2 whether it is affected by the load issue?

> Please note for 'final' patches:
> Always run 'checkpatch.pl' and fix the errors.
> Sorry for that. :-(

Sorry, forgot to call the check script. Attached a revised patch 
(against 2.6.25).

>> Regarding DEBI_E: Just found a av7110 code comment which also reflects
>> what my recent tests showed:
>>           /* Note: Don't try to handle the DEBI error irq (MASK_18), in
>>            * intel mode the timeout is asserted all the time...
>>            */
>>
>> So really only DEBI_S would be left, see below.
> 
> Did you check whether DEBI_S and/or DEBI_E are ever asserted with your
> setup? If not, an interrupt would never occur anyway...

DEBI_E was always asserted (as described in the av7110 code comment), so 
it was worthless. DEBI_S was never asserted without CI (therefore the 
250ms timeout), so it would probably only be received when a CI is used. 
But as described in my other email with measurements, it seems that 
there is no need to optimize the debi_done function further.

>>>> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
>>>> the 100ms poll timer intended? (I found the recent patch to this code in 
>>>> the mailing list end of last year)
>>> That patch was applied to reduce the load on the pci bus in busy-wait
>>> mode. Basically it did not change anything for cam polling. (In fact I
>>> was not aware that the CAM was polled every 100ms. Imho this should be
>>> fixed.)
>> Only wondered whether the 250ms might have been smaller in former driver
>> versions.
> 
> Iirc it should be even worse with older drivers.
> 
> Basically the 250ms timeout is just a last resort to escape from the
> loop, if the debi transfer hangs for some reason. We might try to reduce
> the timeout but I don't know how far we can go. (Touching 'magic' values
> might be dangeous.)

As above, according to my measurements we would not need to change the 
250ms timeout.

Regards,
Robert

[-- Attachment #2: incr-empty-ca-slot-poll-2.6.25.patch --]
[-- Type: text/plain, Size: 2708 bytes --]

From: Robert Schedel <r.schedel@yahoo.de>

This change addresses kernel bug #10459: In kernel 2.6.25 the
budget_av driver polls for an CI slot in 100ms intervals (because no
interrupt solution for budget_av cards is feasible due to HW reasons).
If no CI/CAM is connected to the DVB card, polling times out only after 250ms.
This periodic polling leads to high CPU load.

The change increases the polling interval for empty slots from 100ms to 5s.
Intervals for remaining slot states (invalid, in progress, ready) are unchanged,
as they are either temporary conditions or no timeout should occur.

Signed-off-by: Robert Schedel <r.schedel@yahoo.de>

---
 drivers/media/dvb/dvb-core/dvb_ca_en50221.c |   28 ++++++++++--------
  1 file changed, 16 insertions(+), 12 deletions(-)

diff -rupN linux-2.6.25-orig/drivers/media/dvb/dvb-core/dvb_ca_en50221.c linux-2.6.25/drivers/media/dvb/dvb-core/dvb_ca_en50221.c
--- linux-2.6.25-orig/drivers/media/dvb/dvb-core/dvb_ca_en50221.c	2008-04-17 04:49:44.000000000 +0200
+++ linux-2.6.25/drivers/media/dvb/dvb-core/dvb_ca_en50221.c	2008-04-20 21:23:12.000000000 +0200
@@ -910,15 +910,21 @@ static void dvb_ca_en50221_thread_update
 	int curdelay = 100000000;
 	int slot;
 
+	/* Beware of too high polling frequency, because one polling
+	 * call might take several hundred milliseconds until timeout!
+	 */
 	for (slot = 0; slot < ca->slot_count; slot++) {
 		switch (ca->slot_info[slot].slot_state) {
 		default:
 		case DVB_CA_SLOTSTATE_NONE:
+			delay = HZ * 60;  /* 60s */
+			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE))
+				delay = HZ * 5;  /* 5s */
+			break;
 		case DVB_CA_SLOTSTATE_INVALID:
-			delay = HZ * 60;
-			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE)) {
-				delay = HZ / 10;
-			}
+			delay = HZ * 60;  /* 60s */
+			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE))
+				delay = HZ / 10;  /* 100ms */
 			break;
 
 		case DVB_CA_SLOTSTATE_UNINITIALISED:
@@ -926,19 +932,17 @@ static void dvb_ca_en50221_thread_update
 		case DVB_CA_SLOTSTATE_VALIDATE:
 		case DVB_CA_SLOTSTATE_WAITFR:
 		case DVB_CA_SLOTSTATE_LINKINIT:
-			delay = HZ / 10;
+			delay = HZ / 10;  /* 100ms */
 			break;
 
 		case DVB_CA_SLOTSTATE_RUNNING:
-			delay = HZ * 60;
-			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE)) {
-				delay = HZ / 10;
-			}
+			delay = HZ * 60;  /* 60s */
+			if (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE))
+				delay = HZ / 10;  /* 100ms */
 			if (ca->open) {
 				if ((!ca->slot_info[slot].da_irq_supported) ||
-				    (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_DA))) {
-					delay = HZ / 10;
-				}
+				    (!(ca->flags & DVB_CA_EN50221_FLAG_IRQ_DA)))
+					delay = HZ / 10;  /* 100ms */
 			}
 			break;
 		}

[-- Attachment #3: Type: text/plain, Size: 150 bytes --]

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb

Re: [linux-dvb] High CPU load in "top" due to budget_av slot polling

[Oliver Endriss]

Robert Schedel wrote:
> Oliver Endriss wrote:
> > Robert Schedel wrote:
> >> Oliver Endriss wrote:
> >>> Robert Schedel wrote:
> >>>>
> >>>> Enabling debug traces shows that polling for the PSR in function 
> >>>> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
> >>>> sleeps:
> >>>>
> >>>>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
> >>>> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
> >>>> completion
> >>>>
> >>>> Increasing the 250ms did not avoid the timeout. And as I understood from 
> >>>> previous list mails, this timeout is normal when no CI/CAM is connected 
> >>>> to the DEBI. However, for me the high frequency polling does not make 
> >>>> sense if someone does not plan to plug in a CI/CAM.
> >>>>
> >>>> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
> >>>> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
> >>>> CPU load went down to 0. So this is some kind of workaround for me.
> >>> Afaics the polling interval could be increased to something like 5s or
> >>> 10s if (and only if) the slot is empty. Could you provide a patch?
> >> Attached a patch for 2.6.24.4. Opinions?
> > 
> > Basically it should work but it has to be tested with CI/CAM, too.
> 
> Correct, unfortunately I cannot test it against a CI.

@all:
Could someone who runs budget-av with a CAM please test Robert's patch?
Thanks!

> > Furthermore it is not sufficient to test with budget-av because many
> > other drivers will be affected.
> > 
> > So I would prefer a patch which does not touch behaviour for other card
> > drivers (if possible).
> 
> To my understanding of the DVB code dvb_ca_en50221 is only referenced by 
> budget_av and budget_ci, at least in the vanilla kernel 2.6.25.

You are right (and I am a bit surprised).

> The  patch only changes the timer for slot state EMPTY if 
> DVB_CA_EN50221_FLAG_IRQ_CAMCHANGE is not set, which is for
> 1) budget_av
> 2) budget_ci if the CI firmware version is 0xa2 (because IRQs for CAM 
> change are not supported in this version)
> 
> And those two cases are probably affected by the load issue and should 
> be fixed.
> 
> Of course, we could add another DVB_CA_EN50221 flag solely for budget_av 
> to exclude case 2), but does this make sense? Who is available to test 
> against a budget_ci with FW=0xa2 whether it is affected by the load issue?

After looking closer at your patch I saw that most changes are comments
or coding-style related. Afaics the patch should be safe.

So I will commit your patch next weekend, unless someone spots a
problem.

> >> Regarding DEBI_E: Just found a av7110 code comment which also reflects
> >> what my recent tests showed:
> >>           /* Note: Don't try to handle the DEBI error irq (MASK_18), in
> >>            * intel mode the timeout is asserted all the time...
> >>            */
> >>
> >> So really only DEBI_S would be left, see below.
> > 
> > Did you check whether DEBI_S and/or DEBI_E are ever asserted with your
> > setup? If not, an interrupt would never occur anyway...
> 
> DEBI_E was always asserted (as described in the av7110 code comment), so 
> it was worthless. DEBI_S was never asserted without CI (therefore the 
> 250ms timeout), so it would probably only be received when a CI is used. 
> But as described in my other email with measurements, it seems that 
> there is no need to optimize the debi_done function further.

Ok, thanks for checking.

> >>>> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
> >>>> the 100ms poll timer intended? (I found the recent patch to this code in 
> >>>> the mailing list end of last year)
> >>> That patch was applied to reduce the load on the pci bus in busy-wait
> >>> mode. Basically it did not change anything for cam polling. (In fact I
> >>> was not aware that the CAM was polled every 100ms. Imho this should be
> >>> fixed.)
> >> Only wondered whether the 250ms might have been smaller in former driver
> >> versions.
> > 
> > Iirc it should be even worse with older drivers.
> > 
> > Basically the 250ms timeout is just a last resort to escape from the
> > loop, if the debi transfer hangs for some reason. We might try to reduce
> > the timeout but I don't know how far we can go. (Touching 'magic' values
> > might be dangeous.)
> 
> As above, according to my measurements we would not need to change the 
> 250ms timeout.

Agreed.

CU
Oliver

-- 
----------------------------------------------------------------
VDR Remote Plugin 0.4.0: http://www.escape-edv.de/endriss/vdr/
----------------------------------------------------------------

_______________________________________________
linux-dvb mailing list
linux-dvb@linuxtv.org
http://www.linuxtv.org/cgi-bin/mailman/listinfo/linux-dvb