On Tue, 11 Jun 2019, Aaro Koskinen wrote: > However, with your patch the "nobody cared" is never reached so all is > good. I tried 10 boots with the patch and all were successful. Without > the patch 8 out of 10 failed with the "nobody cared" warning.
I wouldn't call it "good", just less obvious or painful. This is still causing wasted CPU cycles that are used for taking the phantom interrupts. There is clearly a completion barrier missing somewhere that causes the interrupt request to linger beyond the point interrupts are reenabled at the CPU. One way to attempt to narrow it down might be taking a backtrace from where IRQ 14 is found to be spurious. This would indicate the offending interrupt unmask action. E.g. I see no explicit completion barrier between the final `outb' in `mask_and_ack_8259A' and the following call to `raw_spin_unlock_irqrestore', which are obviously otherwise unordered WRT each other (because `outb' is I/O or MMIO and `raw_spin_unlock_irqrestore' is contained within the CPU on UP). I can see provisions however for issuing an architecture-specific barrier in `do_raw_spin_unlock', which is the workhorse for `raw_spin_unlock_irqrestore', so maybe this is the place to look into? Also how's IRQ 14 registered as indicated by /proc/interrupts? Maciej