Things you definitely want to know

Generic issues

isolcpus is our friend too

Isolating some CPUs on the kernel command line using the isolcpus= option, in order to prevent the load balancer from offloading in-band work to them is not only a good idea with PREEMPT_RT, but for any dual kernel configuration too.

By doing so, having some random in-band work evicting cache lines on a CPU where real-time threads briefly sleep is less likely, increasing the odds of costly cache misses, which translates positively into the latency numbers you can get. Even if EVL’s small footprint core has a limited exposure to such kind of disturbance, saving a handful of microseconds is worth it when the worst case figure is already within tenths of microseconds.

CONFIG_DEBUG_HARD_LOCKS is cool but ruins real-time guarantees

When CONFIG_DEBUG_HARD_LOCKS is enabled, the lock dependency engine (CONFIG_LOCKDEP) which helps in tracking down deadlocks and other locking-related issues is also enabled for Dovetail’s hard locks, which underpins most of the serialization mechanisms the EVL core uses.

This is nice as it has the lock validator monitor the hard spinlocks EVL uses too. However, this comes with a high price latency-wise: seeing hundreds of microseconds spent in the validator with hard interrupts off from time to time is not uncommon. Running the latency monitoring utility (aka latmus) which is part of libevl in this configuration should give you pretty ugly numbers.

In short, it is fine enabling CONFIG_DEBUG_HARD_LOCKS for debugging some locking pattern in EVL, but you won’t be able to meet real-time requirements at the same time in such configuration.

CPU frequency scaling (usually) has a negative impact on latency

Enabling the ondemand CPUFreq governor - or any governor performing dynamic adjustment of the CPU frequency - may induce significant latency for EVL on your system, from ten microseconds to more than a hundred depending on the hardware. Selecting the so-called performance governor is the safe option, which guarantees that no frequency transition ever happens, keeping the CPUs at their maximum processing speed.

In other words, if CONFIG_CPU_FREQ has to be enabled in your configuration, enabling CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE and CONFIG_CPU_FREQ_GOV_PERFORMANCE exclusively is most often the best way to prevent unexpectedly high latency peaks.

Disable CONFIG_SMP for best latency on single-core systems

On single-core hardware, some out-of-line code may still be executed for dealing with various types of spinlock with a SMP build, which translates into additional CPU branches and cache misses. On low end hardware, this overhead may be noticeable.

Therefore, if you neither need SMP support nor kernel debug options which depend on instrumenting the spinlock constructs (e.g. CONFIG_DEBUG_PREEMPT), you may want to disable all the related kernel options, starting with CONFIG_SMP.

Architecture-specific issues

x86

  • CONFIG_ACPI_PROCESSOR_IDLE may increase the latency upon wakeup on IRQ from idle on some SoC (up to 30 us observed) on x86. This option is implicitly selected by the following configuration chain: CONFIG_SCHED_MC_PRIO → CONFIG_INTEL_PSTATE → CONFIG_ACPI_PROCESSOR. If out-of-range latency figures are observed on your x86 hardware, turning off this chain may help.

  • When the HPET is disabled, the watchdog which monitors the sanity of the current clocksource for the kernel may use refined-jiffies as the reference clocksource to compare with. Unfortunately, such clocksource is fairly imprecise for timekeeping since timer interrupts might be missed. This could in turn trigger false positives with the watchdog, which would end up declaring the TSC clocksource as ‘unstable’. For instance, it has been observed that enabling CONFIG_FUNCTION_GRAPH_TRACER on some legacy hardware would systematically cause such behavior at boot. The following warning splat appearing in the kernel log is symptomatic of this problem:

    clocksource: timekeeping watchdog on CPU0: Marking clocksource
               'tsc-early' as unstable because the skew is too large:
    clocksource: 'refined-jiffies' wd_now: fffb7018 wd_last: fffb6e9d 
               mask: ffffffff
    clocksource: 'tsc-early' cs_now: 68a6a7070f6a0 cs_last: 68a69ab6f74d6 
               mask: ffffffffffffffff
    tsc: Marking TSC unstable due to clocksource watchdog
    

    This is a problem because the TSC is the best-rated clocksource and directly accessible from the vDSO, which speeds up timestamping operations. If the TSC on your hardware is known to be fine and face this issue nevertheless, you may want to pass tsc=nowatchdog to the kernel to prevent it, or even tsc=reliable if all TSCs are reliable enough to be synchronized across CPUs. If the TSC is really unstable on some legacy hardware and you cannot ignore the watchdog alert, you can still leave it to other clocksources such as acpi_pm. Calls to evl_read_clock() would be significantly slower compared to a direct readout from the vDSO, but the EVL core would manage to get timestamps from its built-in clocks from the out-of-band stage at the expense of a system call, without involving the in-band stage.

    You can retrieve the current clocksource used by the kernel as follows:

    # cat /sys/devices/system/clocksource/clocksource0/current_clocksource
    tsc
    
  • NMI-based perf data collection may cause the kernel to execute utterly sluggish ACPI driver code at each event. Since disabling CONFIG_PERF is not an option, passing nmi_watchodg=0 on the kernel command line at boot may help.

Passing nmi_watchodg=0 turns off the hard lockup detection for the in-band kernel. However, by enabling CONFIG_EVL_WATCHDOG, EVL will still detect runaway EVL threads stuck in out-of-band execution.


Last modified: Mon, 20 Jan 2020 20:23:03 CET