On 1/25/21 4:42 AM, Pavel Begunkov wrote: > We don't call io_submit_flush_completions() in interrupt context, no > need to use irq* versions of spinlock. _irq() isn't safe to use from IRQ context, it would have to be _irqsave() for that. So the point here isn't to ensure that we work from IRQ context, it's to protect against a deadlock when the lock is _also_ grabbed from IRQ context. And for that, we need to make sure that IRQs are disabled, if this lock is already grabbed from IRQ context. So this patch is wrong, the code is correct as-is. -- Jens Axboe
