On Tue, Mar 24, 2020 at 10:20:26AM -0700, Paul E. McKenney wrote: > On Tue, Mar 24, 2020 at 12:52:55PM -0400, Joel Fernandes wrote: > > On Tue, Mar 24, 2020 at 08:48:22AM -0700, Paul E. McKenney wrote: > > [..] > > > > > > > > > diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h > > > > > index 44edd0a..43991a4 100644 > > > > > --- a/kernel/rcu/tree.h > > > > > +++ b/kernel/rcu/tree.h > > > > > @@ -455,6 +455,8 @@ static void rcu_bind_gp_kthread(void); > > > > > static bool rcu_nohz_full_cpu(void); > > > > > static void rcu_dynticks_task_enter(void); > > > > > static void rcu_dynticks_task_exit(void); > > > > > +static void rcu_dynticks_task_trace_enter(void); > > > > > +static void rcu_dynticks_task_trace_exit(void); > > > > > > > > > > /* Forward declarations for tree_stall.h */ > > > > > static void record_gp_stall_check_time(void); > > > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > > > > > index 9355536..f4a344e 100644 > > > > > --- a/kernel/rcu/tree_plugin.h > > > > > +++ b/kernel/rcu/tree_plugin.h > > > > > @@ -2553,3 +2553,21 @@ static void rcu_dynticks_task_exit(void) > > > > > WRITE_ONCE(current->rcu_tasks_idle_cpu, -1); > > > > > #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ > > > > > } > > > > > + > > > > > +/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */ > > > > > +static void rcu_dynticks_task_trace_enter(void) > > > > > +{ > > > > > +#ifdef CONFIG_TASKS_RCU_TRACE > > > > > + if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) > > > > > + current->trc_reader_special.b.need_mb = true; > > > > > > > > If this is every called from middle of a reader section (that is we > > > > transition from IPI-mode to using heavier reader-sections), then is a memory > > > > barrier needed here just to protect the reader section that already started? > > > > > > That memory barrier is provided by the memory ordering in the callers > > > of rcu_dynticks_task_trace_enter() and rcu_dynticks_task_trace_exit(), > > > namely, those callers' atomic_add_return() invocations. These barriers > > > pair with the pair of smp_rmb() calls in rcu_dynticks_zero_in_eqs(), > > > which is in turn invoked from the function formerly known as > > > trc_inspect_reader_notrunning(), AKA trc_inspect_reader(). > > > > > > This same pair of smp_rmb() calls also pair with the conditional smp_mb() > > > calls in rcu_read_lock_trace() and rcu_read_unlock_trace(). > > > > > > In your scenario, the calls in rcu_read_lock_trace() and > > > rcu_read_unlock_trace() wouldn't happen, but in that case the ordering > > > from atomic_add_return() would suffice. > > > > > > Does that work? Or is there an ordering bug in there somewhere? > > > > Thanks for explaining. Could the following scenario cause a problem? > > > > If we consider the litmus test: > > > > { > > int x = 1; > > int *y = &x; > > int z = 1; > > } > > > > P0(int *x, int *z, int **y) > > { > > int *r0; > > int r1; > > > > dynticks_eqs_trace_enter(); > > > > rcu_read_lock(); > > r0 = rcu_dereference(*y); > > > > dynticks_eqs_trace_exit(); // cut-off reader's mb wings :) > > RCU Tasks Trace currently assumes that a reader will not start within > idle and end outside of idle. However, keep in mind that eqs exit > implies a full memory barrier and changes the ->dynticks counter. > The call to rcu_dynticks_task_trace_exit() is not standalone. Instead, > the atomic_add_return() immediately preceding that call is critically > important. And ditto for rcu_dynticks_task_trace_enter() and the > atomic_add_return() immediately following it. > > The overall effect is similar to that of sequence locks. Yes, sounds good. My corner case did consider the full memory barrier aspect. > > r1 = READ_ONCE(*r0); // Reordering of this beyond the unlock() is bad. > > rcu_read_unlock(); > > } > > > > P1(int *x, int *z, int **y) > > { > > rcu_assign_pointer(*y, z); > > synchronize_rcu(); > > WRITE_ONCE(*x, 0); > > } > > > > exists (0:r0=x /\ 0:r1=0) > > > > Then the following situation can happen? > > > > READER UPDATER > > > > y = &z; > > > > eqs_enter(); // full-mb > > > > rcu_read_lock(); // full-mb > > // r0 = x; > > // GP-start > > // ..zero_in_eqs() notices eqs, no IPI > > eqs_exit(); // full-mb > > > > // actual r1 = *x but will reorder > > > > rcu_read_unlock(); // no-mb > > // GP-finish as notices nesting = 0 > > x = 0; > > Followed by an smp_rmb() followed the second read of ->dynticks, which > will see a non-zero bottom bit for ->dynticks, and thus return false. > This in turn will cause the possible zero nesting counter to be ignored. Ah, I see. You are re-reading dynticks to confirm that the case I brought up does not occur. That sounds good to me :) I drew out all possible (similar) scenarios and could not break it and found the GP ordering guarantees holds :) thanks, - Joel > > // reordered r1 = *x = 0; > > > > > > Basically r0=x /\ r1=0 happened because r1=0. Or did I miss something that > > prevents it? > > Yes, the change in value of ->dynticks and the full ordering associated > with the atomic_add_return() that makes this change. > > Thanx, Paul > > > thanks, > > > > - Joel > > > > > > > > > > > > thanks, > > > > > > > > - Joel > > > > > > > > > > > > > +#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */ > > > > > +} > > > > > + > > > > > +/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */ > > > > > +static void rcu_dynticks_task_trace_exit(void) > > > > > +{ > > > > > +#ifdef CONFIG_TASKS_RCU_TRACE > > > > > + if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) > > > > > + current->trc_reader_special.b.need_mb = false; > > > > > +#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */ > > > > > +}