| Message ID | 20230729142738.222208-5-joel@joelfernandes.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | misc RCU fixes and cleanups | expand |
On Sat, Jul 29, 2023 at 02:27:34PM +0000, Joel Fernandes (Google) wrote: > /* > * No-CBs GP kthreads come here to wait for additional callbacks to show up > * or for grace periods to end. > */ > static void nocb_gp_wait(struct rcu_data *my_rdp) > { > - bool bypass = false; > int __maybe_unused cpu = my_rdp->cpu; > unsigned long cur_gp_seq; > unsigned long flags; > bool gotcbs = false; > - unsigned long j = jiffies; > - bool lazy = false; > bool needwait_gp = false; // This prevents actual uninitialized use. > bool needwake; > bool needwake_gp; > + int defer_wake_type = RCU_NOCB_WAKE_NOT; > struct rcu_data *rdp, *rdp_toggling = NULL; > struct rcu_node *rnp; > unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning. > @@ -712,44 +758,24 @@ static void nocb_gp_wait(struct rcu_data *my_rdp) > * won't be ignored for long. > */ > list_for_each_entry(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp) { > - long bypass_ncbs; > - bool flush_bypass = false; > - long lazy_ncbs; > + int defer_wake_type_one = RCU_NOCB_WAKE_NOT; No need to initialize it, nocb_gp_flush_wake() always returns a value, and it will avoid mistakes in the future if nocb_gp_flush_wake() is moved and accidentally not called. > + bool flushed; > + bool empty; > > - trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check")); > rcu_nocb_lock_irqsave(rdp, flags); > - lockdep_assert_held(&rdp->nocb_lock); > - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); > - lazy_ncbs = READ_ONCE(rdp->lazy_len); > + defer_wake_type_one = nocb_gp_flush_wake(rdp, &empty, &flushed); > > - if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) && > - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) || > - bypass_ncbs > 2 * qhimark)) { > - flush_bypass = true; > - } else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) && > - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) || > - bypass_ncbs > 2 * qhimark)) { > - flush_bypass = true; > - } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) { > - rcu_nocb_unlock_irqrestore(rdp, flags); > - continue; /* No callbacks here, try next. */ > - } > + // We may need to do a deferred wakeup later for bypass/lazy > + // So note down what we learnt from the rdp. > + defer_wake_type = max(defer_wake_type_one, defer_wake_type); > > - if (flush_bypass) { > - // Bypass full or old, so flush it. > - (void)rcu_nocb_try_flush_bypass(rdp, j); > - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); > - lazy_ncbs = READ_ONCE(rdp->lazy_len); > + // Did we make any updates to main cblist? If not, no > + // non-deferred wake up to do for this rdp. > + if (!flushed && empty) { Can you ever have (flushed && empty) ? If not you should be able to remove the flushed parameter. Thanks. > + rcu_nocb_unlock_irqrestore(rdp, flags); > + continue; > } > > - if (bypass_ncbs) { > - trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, > - bypass_ncbs == lazy_ncbs ? TPS("Lazy") : TPS("Bypass")); > - if (bypass_ncbs == lazy_ncbs) > - lazy = true; > - else > - bypass = true; > - } > rnp = rdp->mynode; >
On Tue, Aug 29, 2023 at 10:45 AM Frederic Weisbecker <frederic@kernel.org> wrote: > > On Sat, Jul 29, 2023 at 02:27:34PM +0000, Joel Fernandes (Google) wrote: > > /* > > * No-CBs GP kthreads come here to wait for additional callbacks to show up > > * or for grace periods to end. > > */ > > static void nocb_gp_wait(struct rcu_data *my_rdp) > > { > > - bool bypass = false; > > int __maybe_unused cpu = my_rdp->cpu; > > unsigned long cur_gp_seq; > > unsigned long flags; > > bool gotcbs = false; > > - unsigned long j = jiffies; > > - bool lazy = false; > > bool needwait_gp = false; // This prevents actual uninitialized use. > > bool needwake; > > bool needwake_gp; > > + int defer_wake_type = RCU_NOCB_WAKE_NOT; > > struct rcu_data *rdp, *rdp_toggling = NULL; > > struct rcu_node *rnp; > > unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning. > > @@ -712,44 +758,24 @@ static void nocb_gp_wait(struct rcu_data *my_rdp) > > * won't be ignored for long. > > */ > > list_for_each_entry(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp) { > > - long bypass_ncbs; > > - bool flush_bypass = false; > > - long lazy_ncbs; > > + int defer_wake_type_one = RCU_NOCB_WAKE_NOT; > > No need to initialize it, nocb_gp_flush_wake() always returns a value, and > it will avoid mistakes in the future if nocb_gp_flush_wake() is moved and > accidentally not called. Compiler does not like that if it is not initialized: tree_nocb.h:821:51: error: ‘defer_wake_type’ is used uninitialized [-Werror=uninitialized] > > + bool flushed; > > + bool empty; > > > > - trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check")); > > rcu_nocb_lock_irqsave(rdp, flags); > > - lockdep_assert_held(&rdp->nocb_lock); > > - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); > > - lazy_ncbs = READ_ONCE(rdp->lazy_len); > > + defer_wake_type_one = nocb_gp_flush_wake(rdp, &empty, &flushed); > > > > - if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) && > > - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) || > > - bypass_ncbs > 2 * qhimark)) { > > - flush_bypass = true; > > - } else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) && > > - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) || > > - bypass_ncbs > 2 * qhimark)) { > > - flush_bypass = true; > > - } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) { > > - rcu_nocb_unlock_irqrestore(rdp, flags); > > - continue; /* No callbacks here, try next. */ > > - } > > + // We may need to do a deferred wakeup later for bypass/lazy > > + // So note down what we learnt from the rdp. > > + defer_wake_type = max(defer_wake_type_one, defer_wake_type); > > > > - if (flush_bypass) { > > - // Bypass full or old, so flush it. > > - (void)rcu_nocb_try_flush_bypass(rdp, j); > > - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); > > - lazy_ncbs = READ_ONCE(rdp->lazy_len); > > + // Did we make any updates to main cblist? If not, no > > + // non-deferred wake up to do for this rdp. > > + if (!flushed && empty) { > > Can you ever have (flushed && empty) ? If not you should be able to remove the > flushed parameter. We can return 3 different states from the new nocb_gp_flush_wake(): 1. flushed = true and empty = false 2. flushed = false and empty = true 3. flushed = false and empty = false I guess you mean, we don't care about case #1 and #3 ? That's a good point. What I will do then is pass a single bool "wakeup" and and set it for case #2 from within the nocb_gp_flush_wake(). Thanks! - Joel
diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h index 5598212d1f27..c805825c3f00 100644 --- a/kernel/rcu/tree_nocb.h +++ b/kernel/rcu/tree_nocb.h @@ -670,22 +670,68 @@ static void nocb_gp_sleep(struct rcu_data *my_rdp, int cpu) trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep")); } +/* + * Given an rdp, flush its bypass list if needed and return information about + * if a deferred-wakeup needs to be organized depending on whether things are + * still in the bypass list. Also tell caller if the list was flushed and if it + * is still empty after any flushing. + */ +static int nocb_gp_flush_wake(struct rcu_data *rdp, bool *empty, bool *flush) +{ + long bypass_ncbs; + long lazy_ncbs; + unsigned long j = jiffies; + + trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check")); + lockdep_assert_held(&rdp->nocb_lock); + bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); + lazy_ncbs = READ_ONCE(rdp->lazy_len); + + *flush = false; + *empty = false; + if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) && + (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) || + bypass_ncbs > 2 * qhimark)) { + *flush = true; + } else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) && + (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) || + bypass_ncbs > 2 * qhimark)) { + *flush = true; + } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) { + *empty = true; + return RCU_NOCB_WAKE_NOT; + } + + if (*flush) { + // Bypass full or old, so flush it. + (void)rcu_nocb_try_flush_bypass(rdp, j); + bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); + lazy_ncbs = READ_ONCE(rdp->lazy_len); + } + + if (bypass_ncbs) { + trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, + bypass_ncbs == lazy_ncbs ? TPS("Lazy") : TPS("Bypass")); + return (bypass_ncbs == lazy_ncbs ? RCU_NOCB_WAKE_LAZY : + RCU_NOCB_WAKE_BYPASS); + } + return RCU_NOCB_WAKE_NOT; +} + /* * No-CBs GP kthreads come here to wait for additional callbacks to show up * or for grace periods to end. */ static void nocb_gp_wait(struct rcu_data *my_rdp) { - bool bypass = false; int __maybe_unused cpu = my_rdp->cpu; unsigned long cur_gp_seq; unsigned long flags; bool gotcbs = false; - unsigned long j = jiffies; - bool lazy = false; bool needwait_gp = false; // This prevents actual uninitialized use. bool needwake; bool needwake_gp; + int defer_wake_type = RCU_NOCB_WAKE_NOT; struct rcu_data *rdp, *rdp_toggling = NULL; struct rcu_node *rnp; unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning. @@ -712,44 +758,24 @@ static void nocb_gp_wait(struct rcu_data *my_rdp) * won't be ignored for long. */ list_for_each_entry(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp) { - long bypass_ncbs; - bool flush_bypass = false; - long lazy_ncbs; + int defer_wake_type_one = RCU_NOCB_WAKE_NOT; + bool flushed; + bool empty; - trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check")); rcu_nocb_lock_irqsave(rdp, flags); - lockdep_assert_held(&rdp->nocb_lock); - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); - lazy_ncbs = READ_ONCE(rdp->lazy_len); + defer_wake_type_one = nocb_gp_flush_wake(rdp, &empty, &flushed); - if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) && - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) || - bypass_ncbs > 2 * qhimark)) { - flush_bypass = true; - } else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) && - (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) || - bypass_ncbs > 2 * qhimark)) { - flush_bypass = true; - } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) { - rcu_nocb_unlock_irqrestore(rdp, flags); - continue; /* No callbacks here, try next. */ - } + // We may need to do a deferred wakeup later for bypass/lazy + // So note down what we learnt from the rdp. + defer_wake_type = max(defer_wake_type_one, defer_wake_type); - if (flush_bypass) { - // Bypass full or old, so flush it. - (void)rcu_nocb_try_flush_bypass(rdp, j); - bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); - lazy_ncbs = READ_ONCE(rdp->lazy_len); + // Did we make any updates to main cblist? If not, no + // non-deferred wake up to do for this rdp. + if (!flushed && empty) { + rcu_nocb_unlock_irqrestore(rdp, flags); + continue; } - if (bypass_ncbs) { - trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, - bypass_ncbs == lazy_ncbs ? TPS("Lazy") : TPS("Bypass")); - if (bypass_ncbs == lazy_ncbs) - lazy = true; - else - bypass = true; - } rnp = rdp->mynode; // Advance callbacks if helpful and low contention. @@ -792,23 +818,16 @@ static void nocb_gp_wait(struct rcu_data *my_rdp) rcu_gp_kthread_wake(); } - my_rdp->nocb_gp_bypass = bypass; + my_rdp->nocb_gp_bypass = (defer_wake_type == RCU_NOCB_WAKE_BYPASS); my_rdp->nocb_gp_gp = needwait_gp; my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0; // At least one child with non-empty ->nocb_bypass, so set // timer in order to avoid stranding its callbacks. - if (!rcu_nocb_poll) { - // If bypass list only has lazy CBs. Add a deferred lazy wake up. - if (lazy && !bypass) { - wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_LAZY, - TPS("WakeLazyIsDeferred")); - // Otherwise add a deferred bypass wake up. - } else if (bypass) { - wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS, - TPS("WakeBypassIsDeferred")); - } - } + if (!rcu_nocb_poll && defer_wake_type != RCU_NOCB_WAKE_NOT) + wake_nocb_gp_defer(my_rdp, defer_wake_type, + defer_wake_type == RCU_NOCB_WAKE_LAZY ? + TPS("WakeLazyIsDeferred") : TPS("WakeBypassIsDeferred")); if (rcu_nocb_poll) { /* Polling, so trace if first poll in the series. */
The nocb_gp_wait() function contains logic to check each rdp's bypass list, flush if needed, and decide on wakeups. This makes the function hard to follow. Split out the bypass checking and flushing into a separate helper nocb_gp_flush_wake(). The new function encapsulates the logic to: - Check if the bypass needs to be flushed - Flush if needed - Return info on wakeups (lazy, bypass or none) nocb_gp_wait() now becomes simpler by calling the helper to handle the bypass flushing and waking logic. This splitting improves the readability and maintainability of the code by encapsulating related logic into a function with a clear purpose. Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> --- kernel/rcu/tree_nocb.h | 113 ++++++++++++++++++++++++----------------- 1 file changed, 66 insertions(+), 47 deletions(-)