On Mon, Sep 23, 2024 at 10:25:28AM +0800, Jinjie Ruan wrote: > > > On 2024/9/20 21:38, Stephan Gerhold wrote: > > On Fri, Sep 20, 2024 at 03:05:13PM +0200, Dmitry Baryshkov wrote: > >> On Fri, 20 Sept 2024 at 14:44, Stephan Gerhold > >> <stephan.gerhold@xxxxxxxxxx> wrote: > >>> > >>> On Fri, Sep 20, 2024 at 01:48:15PM +0300, Dmitry Baryshkov wrote: > >>>> On Fri, Sep 20, 2024 at 06:07:11PM GMT, Jinjie Ruan wrote: > >>>>> It's important to undo pm_runtime_use_autosuspend() with > >>>>> pm_runtime_dont_use_autosuspend() at driver exit time. > >>>>> > >>>>> But the pm_runtime_disable() and pm_runtime_dont_use_autosuspend() > >>>>> is missing in the error path for bam_dmux_probe(). So add it. > >>>> > >>>> Please use devm_pm_runtime_enable(), which handles autosuspend. > >>>> > >>> > >>> This would conflict with the existing cleanup in bam_dmux_remove(), > >>> which probably needs to stay manually managed since the tear down order > >>> is quite important there. > >> > >> Hmm, the setup and teardown code makes me wonder now. > > > > Yeah, you ask the right questions. :-) It's really tricky to get this > > 100% right. I spent quite some time to get close, but there are likely > > still some loopholes. I haven't heard of anyone running into trouble, > > though. This driver has been rock solid for the past few years. > > > >> Are we guaranteed that the IRQs can not be delivered after suspending > >> the device? > > > > I think bam_dmux_remove() should be safe. disable_irq(dmux->pc_irq) > > prevents any further delivery of IRQs before doing the final power off. > > > >> Also is there a race between IRQs being enabled, manual check of the > >> IRQ state and the pc_ack / power_off calls? > > > > Yes, I'm pretty sure this race exists in theory. I'm not sure how to > > avoid it. We would need an atomic "return current state and enable IRQ" > > operation, but I don't think this exists at the moment. Do you have any > > suggestions? > > Maybe use IRQF_NO_AUTOEN flag to reuqest irq and enable_irq() after that? > I thought about that too, but I think that might introduce a small window in between the two calls where we would miss the state change: irq_get_irqchip_state(..., IRQCHIP_STATE_LINE_LEVEL, ...); /* if an interrupt arrives here we will miss the state change */ enable_irq(); Thanks, Stephan
