On Fri, Nov 03, 2023 at 04:46:24PM +0100, Sebastian Andrzej Siewior wrote: > On 2023-10-20 11:19:49 [-0400], Alan Stern wrote: > > Hmmm... This turns out not to be as easy as one might think. > > > > Sebastian, if you can instead suggest a way to call drivers' interrupt > > handlers (i.e., simulate an interrupt) without causing problems for RT > > kernels, I think that would be a better approach. > > So there is generic_handle_irq_safe(). It should get all the details > right like incrementing the counter in /proc/interrupts, doing nothing > if the interrupt has been masked or waking the interrupt thread if the > interrupt has happen to be threaded. > It triggers the interrupt so for a shared handler it will invoke _all_ > registered interrupt handler and for threaded interrupts it will return > before the thread had a chance to run (free_irq() will handle it > properly and wait for the interrupt thread/handler to complete). Good. Meng Li, can you test a patch that replaces the local_irq_disable() - usb_hcd_irq() - local_irq_enable() lines with a single call to generic_handle_irq_safe()? > > The fundamental problem here is that the uhci-hcd driver was not written > > with unexpected hardware removal in mind. It doesn't have timeouts to > > handle situations where the device doesn't generate an IRQ to indicate > > completion of an I/O operation. And since it's been ten years since > > I've done any significant work on the driver, I'd really like to avoid > > the need for such a far-reaching change (not least because I don't have > > any way to test it). > > I see. Don't over complicate or "correct" things here. What should work > is that the removal callback can be called at any time and things > continue work. That means it will purge all queues, cancel all requests, > timers, whatever and free all resources associated with the driver/ > device. The driver _does_ work under those circumstances -- provided the hardware is still present and accessible. > If it comes to PCI-hotplug you have to have a so called PCI-hotplug > slot. This "slot" will let the OS know if the hardware has been removed > or added. If you don't have such a thing you have to maintain the state > yourself by using the "remove" and "rescan" sysfs files of the PCI slot. > > I'm not aware of any requirement for a PCI-driver to check if its device > has been removed. That's the problem: The driver doesn't really support PCI-hotplug. The code that Meng Li wants to change was sort of a half-baked way to add such support. Alan Stern
