Le Sat, Mar 22, 2025 at 06:00:13PM +0100, Joel Fernandes a écrit : > > > On 3/18/2025 2:56 PM, Frederic Weisbecker wrote: > > RCU relies on the context tracking nesting counter in order to determine > > if it is running in extended quiescent state. > > > > However the context tracking nesting counter is not completely > > synchronized with the actual context tracking state: > > > > * The nesting counter is set to 1 or incremented further _after_ the > > actual state is set to RCU not watching. > > I agree with patch, but this line is a bit confusing ->nesting is set to 1 > *after* the RCU state is set to "watching". Did you mean "watching" ? > > But I think you meant "After RCU transitions from a state of not-watching to > watching' instead of 'actual state is set to RCU not watching'.. > > ct_kernel_entry(): > > // RCU is not watching here ... > ct_kernel_enter_state(offset); > // ... but is watching here. > WRITE_ONCE(ct->nesting, 1);
Oh I completely inverted the thing in the changelog! > > > (then we know for sure we interrupted RCU not watching) > > > > * The nesting counter is set to 0 or decremented further _before_ the > > actual state is set to RCU watching. > > > > Therefore it is safe to assume that if ct_nesting() > 0, RCU is not > > watching. But if ct_nesting() <= 0, RCU is watching except for a tiny > > window. > > > > This hasn't been a problem so far because rcu_is_cpu_rrupt_from_idle() > > has only been called from interrupts. However the code is confusing > > Agreed, and I could also see the existing code's snippet: > WARN_ON_ONCE(!nesting && !is_idle_task(current)); > > .. not working if this function were to be called from non-interrupt kernel > context. Right. I'll reissue that one. Thanks!
