----- On Jul 27, 2021, at 7:44 AM, Peter Zijlstra peterz@xxxxxxxxxxxxx wrote: > On Mon, Jul 26, 2021 at 02:49:03PM -0400, Steven Rostedt wrote: >> OK. I see the issue you are saying. And this came from my assumption >> that the tracepoint code did a synchronization when unregistering the >> last callback. But of course it wont because that would make a lot of >> back to back synchronizations of a large number of tracepoints being >> unregistered at once. >> >> And doing it for all 0->1 or 1->0 or even a 1->0->1 can be costly. >> >> One way to handle this is when going from 1->0, set off a worker that >> will do the synchronization asynchronously, and if a 0->1 comes in, >> have that block until the synchronization is complete. This should >> work, and not have too much of an overhead. >> >> If one 1->0 starts the synchronization, and one or more 1->0 >> transitions happen, it will be recorded where the worker will do >> another synchronization, to make sure all 1->0 have went through a full >> synchronization before a 0->1 can happen. >> >> If a 0->1 comes in while a synchronization is happening, it will note >> the current "number" for the synchronizations (if another one is >> queued, it will wait for one more), before it can begin. As locks will >> be held while waiting for synchronizations to finish, we don't need to >> worry about another 1->0 coming in while a 0->1 is waiting. > > Wouldn't get_state_synchronize_rcu() and cond_synchronize_rcu() get you > what you need? Indeed, snapshotting the state and conditionally waiting for a grace period if none happened since the snapshot appears to be the intent here. Using get_state+cond_sync should allow us to do this without any additional worker thread. Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
