On Tue, Apr 22, 2025 at 09:57:12AM +0200, Clément Léger wrote:
On 21/04/2025 09:06, Alexandre Ghiti wrote:
Hi Clément,
On 14/04/2025 14:34, Clément Léger wrote:
misaligned accesses traps are not nmi and should be treated as normal one using irqentry_enter()/exit().
All the traps that come from kernel mode are treated as nmi as it was suggested by Peter here: https://lore.kernel.org/linux-riscv/ Yyhv4UUXuSfvMOw+@hirez.programming.kicks-ass.net/
I don't know the differences between irq_nmi_entry/exit() and irq_entry/ exit(), so is that still correct to now treat the kernel traps as non-nmi?
Hi Alex,
Actually, this discussion was raised on a previous series [1] by Maciej which replied that we should actually reenable interrupt depending on the state that was interrupted. Looking at other architecture/code, it seems like treating misaligned accesses as NMI is probably not the right way. For instance, loongarch treats them as normal IRQ using a irqentry_enter()/exit() and reenabling IRQS if possible.
So, a trap that happens in kernel space while IRQs are disabled, SHOULD really be NMI-like.
You then have a choice, make all such traps from kernel space NMI-like; this makes it easy on the trap handler, since the context is always the same. Mistakes are 'easy' to find.
Or,.. do funny stuff and only make it NMI like if IRQs were disabled. Which gives inconsistent context for the handler and you'll find yourself scratching your head at some point in the future wondering why this one rare occasion goes BOOM.
x86 mostly does the first, any trap that can happen with IRQs disabled is treated unconditionally as NMI like. The obvious exception is page-fault, but that already has a from-non-preemptible-context branch that is 'careful'.
As to unaligned traps from kernel space, I would imagine they mostly BUG the kernel, except when there's an exception entry for that location, in which case it might do a fixup?
Anyway, the reason these exceptions should be NMI like, is because interrupts are not allowed to nest. Notably something like:
raw_spin_lock_irqsave(&foo); <IRQ> raw_spin_lock_irqsave(&foo); ...
Is an obvious problem. Exceptions that can run while IRQs are disabled, must not use locks -- treating them as NMI-like (they are non-maskable after all), ensures this.