On 2018-10-08 14:09:37 [+0300], Horia Geantă wrote: > CGRs (Congestion Groups) have to be freed by the same CPU that > initialized them. > This is why currently the driver takes special measures; however, using > set_cpus_allowed_ptr() is incorrect - as reported by Sebastian. > > Instead of the generic solution of replacing set_cpus_allowed_ptr() with > work_on_cpu_safe(), we use the qman_delete_cgr_safe() QBMan API instead > of qman_delete_cgr() - which internally takes care of proper CGR > deletion. > > Link: https://lkml.kernel.org/r/20181005125443.dfhd2asqktm22ney@xxxxxxxxxxxxx > Reported-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> > Signed-off-by: Horia Geantă <horia.geanta@xxxxxxx> Oh. No more usage of set_cpus_allowed_ptr(). Wonderful. Thank you. Acked-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> for that. Now that you shifted my attention to qman_delete_cgr_safe(). Could you please use work_on_cpu_safe() here instead smp_call_function_single() with preempt_disable() around it? Now, what is the problem with the CPU limitation? Is this a HW limitation that you can access the registers from a certain CPU? This still fails (silently) if the CPU is missing, right? If you can't get around it, you could block the CPU from going offline. You could register a HP notifier cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, … and the function would return -EINVAL if this is the special CPU. The other thing would be forbid rmmod. This *could* work but if I remember correctly, an explicit unbind can't be stopped. Sebastian
