Message ID | cover.1500373914.git.viresh.kumar@linaro.org |
---|---|
Headers | show |
Series | cpufreq: transition-latency cleanups | expand |
On Wednesday, July 19, 2017 03:42:40 PM Viresh Kumar wrote: > Hi Rafael, > > This series tries to cleanup the code around transition-latency and its > users. Some of the old legacy code, which may not make much sense now, > is dropped as well. And some code consolidation is also done across > governors. > > Based of: v4.13-rc1 > Tested on: ARM64 Hikey board. From the first quick look this version is fine by me. Unless I find anything of concern later, this will be queued up for 4.14. Thanks, Rafael
On Wed, 2017-07-26 at 11:36 +0530, Viresh Kumar wrote: > On 25-07-17, 14:54, Leonard Crestez wrote: > > This patch made it's way into linux-next and it seems to cause imx socs > > to almost always hang around their max frequency with the ondemand > > governor, even when almost completely idle. The lowest frequency is > > never reached. This seems wrong? > > This driver calculates transition_latency at probe time, the value is > > not terribly accurate but it reaches values like latency = 109 us, so > So this is the value that is stored in the global variable > "transition_latency" in the imx6q-cpufreq.c file? i.e. > transition_latency = 109000 (ns) to be exact ? Yes. > - Don't use this patch and try to change ondemand's sampling rate from > sysfs. Try setting it to 10000 and see if the behavior is identical > to after this patch. Yes, it seems to be. Also setting 100000 explicitly fixes this. I also tried to switch from HZ=100 to HZ=1000 but that did not make a difference. > - Find how much time does it really take to change the frequency of > the CPU. I don't really thing 109 us is the right transition > latency. Use attached patch for that and look for the print message. Your patch measures latencies of around 2.5ms, but it can vary between 1.6 ms to 3ms from boot-to-boot. This is a lot more than what the driver reports. Most transitions seem to be faster. I did a little digging and it seems that a majority of time is always spent inside clk_pllv3_wait_lock which spins on a HW bit while doing usleep_range(50, 500). I originally thought it was because of regulators but the delays involved in that are smaller. Measuring wall time on a process that can sleep seems dubious, isn't this vulnerable to random delays because of other tasks? > Without this patch the sampling rate of ondemand governor will be 109 > ms. And after this patch it would be capped at 10 ms. Why would that > screw up anyone's setup ? I don't have an answer to that right now. On a closer look it seems that most of the time is actually spent at low cpufreq though (90%+). Your change makes it so that even something like "sleep 1; cat scaling_cur_freq" raises the frequency to the maximum. This happens enough that even if you do it in a loop you will never see the minimum frequency. It seems there is enough internal bookkeeping on such a wakeup that it takes more than 10ms and enough for a reevaluation of cpufreq until cat returns the value?! I found this by enabling the power:cpu_frequency tracepoint event and checking for deltas with a script. Enabling CPU_FREQ_STAT show this: time_in_state: 396000 1609 792000 71 996000 54 trans_table: From : To : 396000 792000 996000 396000: 0 10 7 792000: 16 0 12 996000: 1 18 0 This is very unexpected but not necessarily wrong. -- Regards, Leonard
+ IMX maintainers. On 27-07-17, 19:54, Leonard Crestez wrote: > On Wed, 2017-07-26 at 11:36 +0530, Viresh Kumar wrote: > > - Find how much time does it really take to change the frequency of > > the CPU. I don't really thing 109 us is the right transition > > latency. Use attached patch for that and look for the print message. > > Your patch measures latencies of around 2.5ms, but it can vary between > 1.6 ms to 3ms from boot-to-boot. This is a lot more than what the > driver reports. Most transitions seem to be faster. Wow !! I was pretty sure all these figures are just made up by some coder :) > I did a little digging and it seems that a majority of time is always > spent inside clk_pllv3_wait_lock which spins on a HW bit while doing > usleep_range(50, 500). I originally thought it was because of > regulators but the delays involved in that are smaller. > > Measuring wall time on a process that can sleep seems dubious, isn't > this vulnerable to random delays because of other tasks? I am not sure I understood that, sorry. > > Without this patch the sampling rate of ondemand governor will be 109 > > ms. And after this patch it would be capped at 10 ms. Why would that > > screw up anyone's setup ? I don't have an answer to that right now. > > On a closer look it seems that most of the time is actually spent at > low cpufreq though (90%+). > > Your change makes it so that even something like "sleep 1; cat > scaling_cur_freq" raises the frequency to the maximum. Why? > This happens > enough that even if you do it in a loop you will never see the minimum > frequency. It seems there is enough internal bookkeeping on such a > wakeup that it takes more than 10ms and enough for a reevaluation of > cpufreq until cat returns the value?! At this point I really feel that this is a hardware specific problem and it was working by chance until now. And I am not sure if we shouldn't be stopping this patch from getting merged just because of that. At least you can teach your distribution to go increase the sampling rate from userspace to make it all work. Can you try one more thing? Try using schedutil governor and see how it behaves ? > I found this by enabling the power:cpu_frequency tracepoint event and > checking for deltas with a script. Enabling CPU_FREQ_STAT show this: > > time_in_state: > > 396000 1609 So we still stay at the lowest frequency most of the time. > 792000 71 > 996000 54 > > trans_table: > > From : To > : 396000 792000 996000 > 396000: 0 10 7 > 792000: 16 0 12 > 996000: 1 18 0 What is it that you are trying to point out here? I still see that we are coming back to 396 MHz quite often. Maybe can you compare these values with and without this patch to let us know? > This is very unexpected but not necessarily wrong. I haven't understood the problem completely yet :( -- viresh
On Fri, 2017-07-28 at 10:58 +0530, Viresh Kumar wrote: > On 27-07-17, 19:54, Leonard Crestez wrote: > > On Wed, 2017-07-26 at 11:36 +0530, Viresh Kumar wrote: > > > Without this patch the sampling rate of ondemand governor will be 109 > > > ms. And after this patch it would be capped at 10 ms. Why would that > > > screw up anyone's setup ? I don't have an answer to that right now. > > On a closer look it seems that most of the time is actually spent at > > low cpufreq though (90%+). > > > > Your change makes it so that even something like "sleep 1; cat > > scaling_cur_freq" raises the frequency to the maximum. > Why? > > > > > This happens > > enough that even if you do it in a loop you will never see the minimum > > frequency. It seems there is enough internal bookkeeping on such a > > wakeup that it takes more than 10ms and enough for a reevaluation of > > cpufreq until cat returns the value?! > At this point I really feel that this is a hardware specific problem > and it was working by chance until now. And I am not sure if we > shouldn't be stopping this patch from getting merged just because of > that. Yes, I agree. Something is fishy here but most likely your patch just expose the problem. > At least you can teach your distribution to go increase the sampling > rate from userspace to make it all work. > > Can you try one more thing? Try using schedutil governor and see how > it behaves ? I don't have the time to investigate this properly right now. > > I found this by enabling the power:cpu_frequency tracepoint event and > > checking for deltas with a script. Enabling CPU_FREQ_STAT show this: > > > > time_in_state: > > > > 396000 1609 > So we still stay at the lowest frequency most of the time. Yes > Maybe can you compare these values with and without this patch to let > us know? Without the patch it is always at low freq. Sampling at a lower frequency means spikes get ignored.
On Wed, 2017-08-16 at 12:04 +0530, Viresh Kumar wrote: > On 28-07-17, 10:58, Viresh Kumar wrote: > > > > At this point I really feel that this is a hardware specific problem > > and it was working by chance until now. And I am not sure if we > > shouldn't be stopping this patch from getting merged just because of > > that. > > > > At least you can teach your distribution to go increase the sampling > > rate from userspace to make it all work. > Its been 3 weeks since my last email on this thread and no reply yet > from any of the IMX maintainers. Can someone please help here ? > > @Shawn: Can you help debugging a bit here, to see what's get screwed > up due to this commit ? Its just that your platform isn't able to > change freq at 10 ms rate. > > @Rafael: I am not sure, but should we be stopping this patch because > some hardware isn't able to change freq at 10ms interval and is just > faking the transition delay to start with ? > > Maybe we get this merged again and the IMX guys can figure out what's > wrong on their platform and how to fix it ? I reported the initial issue but did not have the time to do a more thorough investigation, this is more complicated than it seems. I said this before but maybe it got lost: I don't think the odd behavior I noticed justifies keeping the patch from merging. -- Regards, Leonrd