On 2024-12-20 14:18:43 [+0100], Florian Paul Schmidt wrote: > Hi! Hi, > I'm playing around with a Raspberry Pi 5 and a PREEMPT_RT kernel (raspberry > pi 6.13.y branch) and wonder about the latencies I see. First I tried > cyclictest and saw latencies in the upper 50s up to 80s or so under a > kernel compile load. > > So I compiled the kernel with timerlat and osnoise tracers and ran a bit of > > rtla timerlat top -a 60 --dma-latency 0 > > and I get traces like the following: … > ## CPU 3 hit stop tracing, analyzing it ## > IRQ handler delay: 0.00 us (0.00 %) > IRQ latency: 5.16 us > Timerlat IRQ duration: 47.41 us (42.53 %) > Blocking thread: 50.50 us (45.30 %) > stress:278226 50.50 us > Blocking thread stack trace > -> timerlat_irq > -> __hrtimer_run_queues > -> hrtimer_interrupt > -> arch_timer_handler_phys > -> handle_percpu_devid_irq > -> generic_handle_domain_irq > -> gic_handle_irq > -> call_on_irq_stack > -> do_interrupt_handler > -> el1_interrupt > -> el1h_64_irq_handler > -> el1h_64_irq > -> folios_put_refs … > Thread latency: 111.47 us (100%) The backtrace appears long but everything up to folios_put_refs() appears preemtible. Then there is the interrupt handler. This looks okay. > Max timerlat IRQ latency from idle: 1.98 us in cpu 3 > Saving trace to timerlat_trace.txt > > > OK, here at least something related to virtual memory is happening. > I still wonder though: How come the latencies are so large? Is the > raspberry pi platform just not suited to reach lower latencies? Is > everything working as expected? Are maybe the raspberry pi > patches responsible for latency sources which are invisible in the > traces? I there any hope for getting these latencies a little more > under control? Based on this trace, I don't see a section which disables preemption/ interrupts and be the reason for the interrupt happening. One thing that could delay your doing are small caches. Then there is looking within the system: An inaccurate/ low resolution clocksource and/ or timer device. > Kind regards, > FPS Sebastian
