On Tue, 2017-08-29 at 17:43 -0300, Daniel Henrique Barboza wrote: > Hi, > > This is a scenario I've been facing when working in early device > hotplugs in QEMU. When a device is added, a IRQ pulse is fired to warn > the guest of the event, then the kernel fetches it by calling > 'check_exception' and handles it. If the hotplug is done too early > (before SLOF, for example), the pulse is ignored and the hotplug event > is left unchecked in the events queue. > > One solution would be to pulse the hotplug queue interrupt after CAS, > when we are sure that the hotplug queue is negotiated. However, this > panics the kernel with sig 11 kernel access of bad area, which suggests > that the kernel wasn't quite ready to handle it.
That's not right. This is a bug that needs fixing. The interrupt should be masked anyway but still. Tell us more about the crash (backtrace etc...) this definitely needs fixing. > In my experiments using upstream 4.13 I saw that there is a 'safe time' > to pulse the queue, sometime after CAS and before mounting the root fs, > but I wasn't able to pinpoint it. From QEMU perspective, the last hcall > done (an h_set_mode) is still too early to pulse it and the kernel > panics. Looking at the kernel source I saw that the IRQ handling is > initiated quite early in the init process. > > So my question (ok, actually 2 questions): > > - Is my analysis correct? Is there an unsafe time to fire a IRQ pulse > before CAS that can break the kernel or am I overlooking/doing something > wrong? > - is there a reliable way to know when can the kernel safely handle the > hotplug interrupt? So I don't think that's the right approach. Virtual interrutps are edge sensitive and we will potentially lose them if they occur early. I think what needs to happen is: - Fix whatever's causing the above crash and - The hotplug code should check for pending events (check_exception ?) at boot time to enqueue whatever's there. It needs to do that after unmasking the interrupt and in a way that is protected from races with said interrupt. Cheers, Ben. > > Thanks, > > > Daniel