On Sun, Apr 19, 2020 at 06:44:50PM -0700, Paul E. McKenney wrote: > On Sun, Apr 19, 2020 at 09:17:49PM -0400, Joel Fernandes wrote: > > On Sun, Apr 19, 2020 at 08:27:13PM -0400, Joel Fernandes wrote: > > > On Sun, Apr 19, 2020 at 07:58:36AM -0700, Paul E. McKenney wrote: > > > > On Sat, Apr 18, 2020 at 02:37:48PM +0200, Uladzislau Rezki wrote: > > > > > On Fri, Apr 17, 2020 at 11:54:49AM -0700, Paul E. McKenney wrote: > > > > > > On Fri, Apr 17, 2020 at 02:26:41PM -0400, Joel Fernandes wrote: > > > > > > > On Fri, Apr 17, 2020 at 05:04:42PM +0200, Sebastian Andrzej Siewior wrote: > > > > > > > > On 2020-04-16 23:05:15 [-0400], Joel Fernandes wrote: > > > > > > > > > On Thu, Apr 16, 2020 at 11:34:44PM +0200, Sebastian Andrzej Siewior wrote: > > > > > > > > > > On 2020-04-16 14:00:57 [-0700], Paul E. McKenney wrote: > > > > > > > > > > > > > > > > > > > > > > We might need different calling-context restrictions for the two variants > > > > > > > > > > > of kfree_rcu(). And we might need to come up with some sort of lockdep > > > > > > > > > > > check for "safe to use normal spinlock in -rt". > > > > > > > > > > > > > > > > > > > > Oh. We do have this already, it is called CONFIG_PROVE_RAW_LOCK_NESTING. > > > > > > > > > > This one will scream if you do > > > > > > > > > > raw_spin_lock(); > > > > > > > > > > spin_lock(); > > > > > > > > > > > > > > > > > > > > Sadly, as of today, there is code triggering this which needs to be > > > > > > > > > > addressed first (but it is one list of things to do). > > > > > > > > > > > > > > > > > > > > Given the thread so far, is it okay if I repost the series with > > > > > > > > > > migrate_disable() instead of accepting a possible migration before > > > > > > > > > > grabbing the lock? I would prefer to avoid the extra RT case (avoiding > > > > > > > > > > memory allocations in a possible atomic context) until we get there. > > > > > > > > > > > > > > > > > > I prefer something like the following to make it possible to invoke > > > > > > > > > kfree_rcu() from atomic context considering call_rcu() is already callable > > > > > > > > > from such contexts. Thoughts? > > > > > > > > > > > > > > > > So it looks like it would work. However, could we please delay this > > > > > > > > until we have an actual case on RT? I just added > > > > > > > > WARN_ON(!preemptible()); > > > > > > > > > > > > > > I am not sure if waiting for it to break in the future is a good idea. I'd > > > > > > > rather design it in a forward thinking way. There could be folks replacing > > > > > > > "call_rcu() + kfree in a callback" with kfree_rcu() for example. If they were > > > > > > > in !preemptible(), we'd break on page allocation. > > > > > > > > > > > > > > Also as a sidenote, the additional pre-allocation of pages that Vlad is > > > > > > > planning on adding would further reduce the need for pages from the page > > > > > > > allocator. > > > > > > > > > > > > > > Paul, what is your opinion on this? > > > > > > > > > > > > My experience with call_rcu(), of which kfree_rcu() is a specialization, > > > > > > is that it gets invoked with preemption disabled, with interrupts > > > > > > disabled, and during early boot, as in even before rcu_init() has been > > > > > > invoked. This experience does make me lean towards raw spinlocks. > > > > > > > > > > > > But to Sebastian's point, if we are going to use raw spinlocks, we need > > > > > > to keep the code paths holding those spinlocks as short as possible. > > > > > > I suppose that the inability to allocate memory with raw spinlocks held > > > > > > helps, but it is worth checking. > > > > > > > > > > > How about reducing the lock contention even further? > > > > > > > > Can we do even better by moving the work-scheduling out from under the > > > > spinlock? This of course means that it is necessary to handle the > > > > occasional spurious call to the work handler, but that should be rare > > > > and should be in the noise compared to the reduction in contention. > > > > > > Yes I think that will be required since -rt will sleep on workqueue locks as > > > well :-(. I'm looking into it right now. > > > > > > /* > > > * If @work was previously on a different pool, it might still be > > > * running there, in which case the work needs to be queued on that > > > * pool to guarantee non-reentrancy. > > > */ > > > last_pool = get_work_pool(work); > > > if (last_pool && last_pool != pwq->pool) { > > > struct worker *worker; > > > > > > spin_lock(&last_pool->lock); > > > > Hmm, I think moving schedule_delayed_work() outside lock will work. Just took > > a good look and that's not an issue. However calling schedule_delayed_work() > > itself is an issue if the caller of kfree_rcu() is !preemptible() on > > PREEMPT_RT. Because the schedule_delayed_work() calls spin_lock on pool->lock > > which can sleep on PREEMPT_RT :-(. Which means we have to do either of: > > > > 1. Implement a new mechanism for scheduling delayed work that does not > > acquire sleeping locks. > > > > 2. Allow kfree_rcu() only from preemptible context (That is Sebastian's > > initial patch to replace local_irq_save() + spin_lock() with > > spin_lock_irqsave()). > > > > 3. Queue the work through irq_work or another bottom-half mechanism. > > I use irq_work elsewhere in RCU, but the queue_delayed_work() might > go well with a timer. This can of course be done conditionally. > We can schedule_delayed_work() inside and outside of the spinlock, i.e. it is not an issue for RT kernel, because as it was noted in last message a workqueue system uses raw spinlicks internally. I checked the latest linux-5.6.y-rt also. If we do it inside, we will place the work on current CPU, at least as i see it, even if it is "unbound". If we do it outside, we will reduce a critical section, from the other hand we can introduce a potential delay in placing the context into CPUs run-queuye. As a result we could end up on another CPU, thus placing the work on new CPU, plus memory foot-print might be higher. It would be good to test and have a look at it actually. But it can be negligible :) > > Any other thoughts? > > I did forget to ask you guys your opinions about the downsides (if any) > of moving from unbound to per-CPU workqueues. Thoughts? > If we do it outside of spinlock, there is at least one drawback that i see, i described it above. We can use schedule_delayed_work_on() but we as a caller have to guarantee that a CPU we about to place a work is alive :) -- Vlad Rezki