Ensure that the pageout daemon runs on schedule

[alc]

Summary

In vm_pageout_worker(), we do not account for the time that it takes to scan the active and inactive queues, or delays in starting vm_pageout_worker(), when computing the sleep time. Consequently, as the amount of time that vm_pageout_worker() spends scanning the queues varies, so does the timing of the calls to pidctrl_daemon(). For example, under a PostgreSQL workload, I see the interval between calls stretch by 10+%. This change computes a sleep time based on the next desired execution tick for the controller.

On my test machine, calls to pidctrl_daemon() often occur one tick earlier than expected, and this triggers the "multiple calls within an interval" case. Suppose that calls occur at time t, t+99, and t+199. The effect of the "multiple calls within an interval" handling at time t+99 is that the integral term has much reduced impact on the output.

Test Plan

I'm really curious what values you see for debug.counter1. After about 2.5 hours of running a PostgreSQL workload, I see:

debug.dt: 89
debug.counter: 7
vm.domain.0.pidctrl.output: 11609
vm.domain.0.pidctrl.error: -38
debug.foo1: 2156480700
debug.foo2: 2156480700
debug.counter1: 7705

[kib]

The change tightly alias the worker wakeup to the period ticks. It might result at least in the cosmetics issues with the load average.

[markj]

Regarding the testing section, after running pgbench in a loop for roughly 12 hours, I see:

debug.dt: 95
debug.counter: 632
vm.domain.0.pidctrl.output: 10481
vm.domain.0.pidctrl.error: 99
debug.foo1: 2205322400
debug.foo2: 2205322400
debug.counter1: 7

[alc]

In D15749#332911, @kib wrote:

The change tightly alias the worker wakeup to the period ticks. It might result at least in the cosmetics issues with the load average.

Isn't this addressed by the P_SYSTEM flag, specifically, that we don't include processes with the P_SYSTEM flag set in the load average?

[kib]

In D15749#355638, @alc wrote:

In D15749#332911, @kib wrote:

The change tightly alias the worker wakeup to the period ticks. It might result at least in the cosmetics issues with the load average.

Isn't this addressed by the P_SYSTEM flag, specifically, that we don't include processes with the P_SYSTEM flag set in the load average?

I more do not like the sync with ticks, and the LA was only (bad) example. I believe a small amount of noise in the sleep period would be beneficial regardless of P_SYSTEM.