On 2024-07-18, Andrew Halaney <ahalaney@xxxxxxxxxx> wrote: > I'll be honest, I only considered this from a printk() point of view > so far and pretty much ignored the original splat (as someone else is > looking at that). Here's the original splat from a centos kernel, I > have yet to look into it other than "we took an exception prior to RCU > coming online for this CPU" Thanks for posting the splat. This is indeed a printk() issue. vprintk_emit() or nbcon_wake_threads() need to be aware of the RCU state and act appropriately. I will look into this. > and me naively thinking that printk() should work everywhere. Your thinking is not naive. It is correct (for PREEMPT_RT). printk() stores messages locklessly into its ringbuffer and must work in all cases (including NMI and scheduler). And AFAIK it does. However, printk() is also responsible for reliably triggering or performing console printing. This separate responsibility is tricky. For PREEMPT_RT the problem is supposed to be solved. Obviously it still has some issues, so thank you for the report. For PREEMPT_RT, if you can put a printk() anywhere after the Linux banner and cause a problem, the printk-folks need to know about it. John Ogness
