On Thu, Oct 31, 2024 at 10:50:29AM -0700, Paul E. McKenney wrote: > On Thu, Oct 31, 2024 at 08:55:09AM +0100, Sebastian Andrzej Siewior wrote: > > On 2024-10-31 08:35:45 [+0100], Vlastimil Babka wrote: > > > On 10/31/24 08:21, Sebastian Andrzej Siewior wrote: > > > > On 2024-10-30 16:10:58 [-0700], Paul E. McKenney wrote: > > > >> > > > >> So I need to avoid calling kfree() within an smp_call_function() handler? > > > > > > > > Yes. No kmalloc()/ kfree() in IRQ context. > > > > > > However, isn't this the case that the rule is actually about hardirq context > > > on RT, and most of these operations that are in IRQ context on !RT become > > > the threaded interrupt context on RT, so they are actually fine? Or is smp > > > call callback a hardirq context on RT and thus it really can't do those > > > operations? > > > > interrupt handlers as of request_irq() are forced-threaded on RT so you > > can do kmalloc()/ kfree() there. smp_call_function.*() on the other hand > > are not threaded and invoked directly within the IRQ context. > > OK, thank you all for the explanation! I will fix using Boqun's > suggestion of irq work, but avoiding the issue Boqun raises by invoking I've tried fixing this with irq work, however, unlike normal work_struct, irq_work will still touch the work item header after the work function is executed (see irq_work_single()). So it needs more work to build an "one-off free" functionality on it. I think we can just use normal workqueue, because queue_work() uses local_irq_save() + raw_spin_lock(), so it's irq-safe even for non-threaded interrupts. Sending a patch soon. Regards, Boqun > the irq-work handler from the smp_call_function() handler. > > It will be a few days before I get to this, so if there is a better way, > please do not keep it a secret! > > Thanx, Paul
