On Wed, Mar 18, 2020 at 09:58:23PM -0400, Joel Fernandes wrote: > On Wed, Mar 18, 2020 at 9:37 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote: > > > > On Wed, Mar 18, 2020 at 05:10:52PM -0700, paulmck@xxxxxxxxxx wrote: > > [...] > > > +/* Initialize for a new RCU-tasks-trace grace period. */ > > > +static void rcu_tasks_trace_pregp_step(void) > > > +{ > > > + int cpu; > > > + > > > + // Wait for CPU-hotplug paths to complete. > > > + cpus_read_lock(); > > > + cpus_read_unlock(); > > > + > > > + // Allow for fast-acting IPIs. > > > + atomic_set(&trc_n_readers_need_end, 1); > > > + > > > + // There shouldn't be any old IPIs, but... > > > + for_each_possible_cpu(cpu) > > > + WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu)); > > > +} > > > + > > > +/* Do first-round processing for the specified task. */ > > > +static void rcu_tasks_trace_pertask(struct task_struct *t, > > > + struct list_head *hop) > > > +{ > > > + WRITE_ONCE(t->trc_reader_need_end, false); > > > + t->trc_reader_checked = false; > > > + t->trc_ipi_to_cpu = -1; > > > + trc_wait_for_one_reader(t, hop); > > > +} > > > + > > > +/* Do intermediate processing between task and holdout scans. */ > > > +static void rcu_tasks_trace_postscan(void) > > > +{ > > > + // Wait for late-stage exiting tasks to finish exiting. > > > + // These might have passed the call to exit_tasks_rcu_finish(). > > > + synchronize_rcu(); > > > + // Any tasks that exit after this point will set ->trc_reader_checked. > > > +} > > > + > > > +/* Do one scan of the holdout list. */ > > > +static void check_all_holdout_tasks_trace(struct list_head *hop, > > > + bool ndrpt, bool *frptp) > > > +{ > > > + struct task_struct *g, *t; > > > + > > > + list_for_each_entry_safe(t, g, hop, trc_holdout_list) { > > > + // If safe and needed, try to check the current task. > > > + if (READ_ONCE(t->trc_ipi_to_cpu) == -1 && > > > + !READ_ONCE(t->trc_reader_checked)) > > > + trc_wait_for_one_reader(t, hop); > > > > Just some questions: > > > > 1. How are we ensuring on the reader-side that we are executing memory > > barriers that are sufficient to ensure that all update-side memory operations > > Apologies, here I meant "update memory operations". > > > in reader section is visible to code executing after the grace period? The most pithy response is that in many of the cases, we are not executing any additional memory barriers on the read side because it is not necessary to do so. Another pithy response would be that there are very likely still memory-ordering bugs, hence the "RFC". ;-) Perhaps more constructively... There are a number of cases, taking things task by task. 1. The target task is blocked and not in a read-side critical section. In this case, the grace-period kthread's call to try_invoke_on_locked_down_task() will acquire the target task's ->pi_lock. This lock will have been acquired by the target task while sleeping and will be acquired again when it awakened. This set of locks will order prior and subsequent read-side critical sections against that point in the grace-period process. The smp_mb__after_spinlock() at the beginning of rcu_tasks_kthread() in combination with the lock will order this against posting of the callbacks. The smp_mb() at the end of rcu_tasks_trace_postgp() does the same for post-grace-period actions. Note that the ordering from the callbacks to the subsequent readers still holds even if the target task is in a read-side critical section. 2. The target task is runnable and not in a read-side critical section. This proceeds the same as #1 above, except that both the ->pi_lock and the runqueue lock will be held when checking the target tasks's state. Note that the ordering from the callbacks to the subsequent readers still holds even if the target task is in a read-side critical section. 3. The target task is running that we IPI and is not in a read-side critical section. In this case, there is ordering from the smp_call_function_single() to the IPI handler on the one hand (and thus to later readers), and due to the smp_store_release() (from earlier readers) on the other hand. The READ_ONCE(t->trc_ipi_to_cpu) pairs with the smp_store_release() in combination with the aforementioned smp_mb(). Note that the ordering from the callbacks to the subsequent readers still holds even if the target task is in a read-side critical section. 4. If the target task passes through a context switch while not in a read-side critical section, it also passes through rcu_tasks_trace_qs(), which has an smp_mb() and smp_store_release() that work not unlike the IPI case above. If the task is in a read-side critical section, nothing happens, and either this or some other case will apply at some later time. 5. Otherwise, the task's ->trc_reader_special.b.need_qs is set, which will cause the rcu_read_unlock_trace() to invoke rcu_read_unlock_trace_special(). The chain of atomic_dec_and_test() calls will order all of these events, and that chain, when combined with wait_event_idle_exclusive_timeout()'s read of that same variable and its later smp_mb() will force the needed ordering with all prior readers. Hey, you asked! Your turn. Find the bug(s). > > 2. Is it possible that a hold-out task is removed from the hold-out list and is > > not waited on in the updater side, before the reader side got a chance to > > indirectly execute such memory barriers? The only purpose of the hold-out task list is to keep track of tasks that still need one of the above options to be applied. Once a task is removed from that list one of the above options has happened or is in flight, so there is no need for the hold-out task list to be involved in any ordering. > > 3. If a reader sees updates that were done before the grace period started, it > > should not see any updates that happen after the grace period ends. Is that > > guaranteed with this RCU-Trace? Yes, as above. > > If its Ok, it would be nice to mention more about memory ordering aspect in > > the changelog. I can certainly do that, but in the immortal words of MS-DOS, are you sure? The problem is that RCU memory ordering isn't a single pairing or even a chain. It is instead a net weaving through all the readers before the end of the grace period, all the readers after the beginning of the grace period, the grace period itself, as well as the updates both before and after the grace period. See for example Documentation/RCU/Design/Memory-Ordering/TreeRCU-gp.svg. It would look sort of like that, though perhaps a little bit less elaborate. Thanx, Paul > > thanks! > > > > - Joel > > > > > > > + > > > + // If check succeeded, remove this task from the list. > > > + if (READ_ONCE(t->trc_reader_checked)) > > > + trc_del_holdout(t); > > > + } > > > +} > > > + > > > +/* Wait for grace period to complete and provide ordering. */ > > > +static void rcu_tasks_trace_postgp(void) > > > +{ > > > + // Remove the safety count. > > > + smp_mb__before_atomic(); // Order vs. earlier atomics > > > + atomic_dec(&trc_n_readers_need_end); > > > + smp_mb__after_atomic(); // Order vs. later atomics > > > + > > > + // Wait for readers. > > > + wait_event_idle_exclusive(trc_wait, > > > + atomic_read(&trc_n_readers_need_end) == 0); > > > + > > > + smp_mb(); // Caller's code must be ordered after wakeup. > > > +} > > > + > > > +/* Report any needed quiescent state for this exiting task. */ > > > +void exit_tasks_rcu_finish_trace(struct task_struct *t) > > > +{ > > > + WRITE_ONCE(t->trc_reader_checked, true); > > > + WARN_ON_ONCE(t->trc_reader_nesting); > > > + WRITE_ONCE(t->trc_reader_nesting, 0); > > > + if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_need_end))) > > > + rcu_read_unlock_trace_special(t); > > > +} > > > + > > > +void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func); > > > +DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace, > > > + "RCU Tasks Trace"); > > > + > > > +/** > > > + * call_rcu_tasks_trace() - Queue a callback trace task-based grace period > > > + * @rhp: structure to be used for queueing the RCU updates. > > > + * @func: actual callback function to be invoked after the grace period > > > + * > > > + * The callback function will be invoked some time after a full grace > > > + * period elapses, in other words after all currently executing RCU > > > + * read-side critical sections have completed. call_rcu_tasks_trace() > > > + * assumes that the read-side critical sections end at context switch, > > > + * cond_resched_rcu_qs(), or transition to usermode execution. As such, > > > + * there are no read-side primitives analogous to rcu_read_lock() and > > > + * rcu_read_unlock() because this primitive is intended to determine > > > + * that all tasks have passed through a safe state, not so much for > > > + * data-strcuture synchronization. > > > + * > > > + * See the description of call_rcu() for more detailed information on > > > + * memory ordering guarantees. > > > + */ > > > +void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func) > > > +{ > > > + call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace); > > > +} > > > +EXPORT_SYMBOL_GPL(call_rcu_tasks_trace); > > > + > > > +/** > > > + * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period > > > + * > > > + * Control will return to the caller some time after a trace rcu-tasks > > > + * grace period has elapsed, in other words after all currently > > > + * executing rcu-tasks read-side critical sections have elapsed. These > > > + * read-side critical sections are delimited by calls to schedule(), > > > + * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory, > > > + * anyway) cond_resched(). > > > + * > > > + * This is a very specialized primitive, intended only for a few uses in > > > + * tracing and other situations requiring manipulation of function preambles > > > + * and profiling hooks. The synchronize_rcu_tasks_trace() function is not > > > + * (yet) intended for heavy use from multiple CPUs. > > > + * > > > + * See the description of synchronize_rcu() for more detailed information > > > + * on memory ordering guarantees. > > > + */ > > > +void synchronize_rcu_tasks_trace(void) > > > +{ > > > + RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section"); > > > + synchronize_rcu_tasks_generic(&rcu_tasks_trace); > > > +} > > > +EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace); > > > + > > > +/** > > > + * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks. > > > + * > > > + * Although the current implementation is guaranteed to wait, it is not > > > + * obligated to, for example, if there are no pending callbacks. > > > + */ > > > +void rcu_barrier_tasks_trace(void) > > > +{ > > > + /* There is only one callback queue, so this is easy. ;-) */ > > > + synchronize_rcu_tasks_trace(); > > > +} > > > +EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace); > > > + > > > +static int __init rcu_spawn_tasks_trace_kthread(void) > > > +{ > > > + rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step; > > > + rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask; > > > + rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan; > > > + rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace; > > > + rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp; > > > + rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace); > > > + return 0; > > > +} > > > +core_initcall(rcu_spawn_tasks_trace_kthread); > > > + > > > +#else /* #ifdef CONFIG_TASKS_TRACE_RCU */ > > > +void exit_tasks_rcu_finish_trace(struct task_struct *t) { } > > > +#endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */ > > > -- > > > 2.9.5 > > >