On Thu, Oct 24, 2024 at 2:46 PM Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx> wrote: > > On (24/10/24 14:18), Tomasz Figa wrote: > > > @@ -1538,9 +1538,25 @@ static int venc_close(struct file *file) > > > > > > venc_pm_get(inst); > > > > > > + /* > > > + * First, remove the inst from the ->instances list, so that > > > + * to_instance() will return NULL. > > > + */ > > > + hfi_session_destroy(inst); > > > + /* > > > + * Second, make sure we don't have IRQ/IRQ-thread currently running or > > > + * pending execution (disable_irq() calls synchronize_irq()), which > > > + * can race with the inst destruction. > > > + */ > > > + disable_irq(inst->core->irq); > > > + /* > > > + * Lastly, inst is gone from the core->instances list and we don't > > > + * have running/pending IRQ/IRQ-thread, proceed with the destruction > > > + */ > > > + enable_irq(inst->core->irq); > > > + > > > > Thanks a lot for looking into this. Wouldn't it be enough to just call > > synchronize_irq() at this point, since the instance was removed from > > the list already? I guess the question is if that's the only way the > > interrupt handler can get hold of the instance. > > Good question. > > synchronize_irq() waits for IRQ-threads, so if inst is accessed only from > IRQ-thread then we are fine. If, however, inst is also accessed from hard > IRQ, then synchronize_irq() won't work, I guess, because it doesn't wait > for "in flight hard IRQs". disable_irq() OTOH "waits for completion", so > we cover in-flight hard IRQs too. Looking at the code, synchronize_irq() internally also calls synchronize_hardirq() and that in turn waits for the IRQD_IRQ_INPROGESS flag to be cleared before returning [1]. The flag is set by handle_irq_event() before most of the IRQ handling is run and cleared at the end of the function [2], which makes me believe that it would actually ensure all the hardirq and threaded IRQ handlers would be waited for. [1] https://elixir.bootlin.com/linux/v6.11.5/source/kernel/irq/manage.c#L38 [2] https://elixir.bootlin.com/linux/v6.11.5/source/kernel/irq/handle.c#L202 Although I guess it would be the best if someone confirmed that, because with all the IRQ handling complexities of SMP, nothing can be certain today. :P Best, Tomasz
