On 2018.12.10 02:52 Peter Zijlstra wrote: >On Mon, Dec 10, 2018 at 10:36:40PM +0100, Rafael J. Wysocki wrote: >> On Mon, Dec 10, 2018 at 1:21 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: >>> Would not a tracepoint be better?; then there is no overhead in the >>> normal case where nobody gives a crap about these here numbers. >> >> There is an existing tracepoint that in principle could be used to >> produce this information, but it is such a major PITA in practice that >> nobody does that. Guess why. :-) > > Sounds like you need to ship a convenient script or something :-) For the histogram plots of idle durations that I sometimes provide, trace is used. It is more work to do it the trace way. Very often, when the rate of idle state entries/ exits is high, turning on trace influences the system under test significantly. Also, even if I allocate the majority of my memory to the trace buffer (i.e. 15 of my 16 gigabytes), I can only acquire a few minutes of trace data before filling the buffer. Some of my tests run for hours, and these new counters provide a way to acquire potentially useful (I don't have enough experience with them yet to know how useful) information, while having no influence on the system under test because I only take samples once per minute, or sometimes 4 times per minute. >> Also, the "usage" and "time" counters are there in sysfs, so why not these two? I agree, how are these two counters any different? In about a week or so, I'll have some test data comparing 4.20-rc5 with teov6 teov7 along with the idle data (graphs) that I usually provide and also these new counters. ... Doug