On 2023-06-23 13:15:11 [-0400], Martin Hicks wrote: > Hi, Hi, > > I'm looking at using a POSIX timer to cause an action to be run every > 500us, but I'm seeing that when the hrtimer fires while the CPU is > busy running another SCHED_RR/SCHED_FIFO process that the hrtimer is > deferred until the process releases the CPU. I'm using SIGEV_SIGNAL > to deliver the timer expiry to a thread that's running at a higher RT > Priority than the other tasks, but no preemption seems to occur. > > If the "other" tasks are running with SCHED_OTHER then the signal is > delivered immediately. > > I've tried using 'cset shield -c 1' and without (just pinning use > taskset -c 1), and modified priorities and scheduler with chrt. > The test is running on x86_64 with 6.3.9-rt15. PREEMPT_RT is enabled. > > A simple test that illustrates the idea is attached. > Is this supposed to work? Well, "work". If you have a SCHED_RR/FIFO task and setup a timer (clock_nanosleep()) then the wakeup is precise/ direct (hw timer -> wakeup of the task). POSIX timer are a bit special due to how they have to be processed. As a consequence, the wake up path is (hw timer -> ktimers/$CPU thread -> wakeup of the task via a signal). That ktimers/$CPU thread runs at SCHED_FIFO 1 priority. If there a task with higher priority then timer handling get deferred. This sounds like what you observe. That ktimers/$CPU is also in charge of waking all non-RT threads. > mh Sebastian
