No-CBs grace-period kthread updates for v5.3-rc2
Subsystems (3)
cover No-CBs grace-period kthread updates for v5.3-rc2
Hello!
This series reduces memory footprint (RCU callbacks posted by no-CBs CPUs) by providing separate rcuog kthreads that wait for grace periods, thus avoiding the current situation in which callbacks are delayed while the leader rcuo kthreads invokes callbacks:
1. Rename rcu_data fields to prepare for forward-progress work.
2. Update comments to prepare for forward-progress work.
3. Provide separate no-CBs grace-period kthreads.
4. Rename nocb_follower_wait() to nocb_cb_wait().
5. Rename wake_nocb_leader() to wake_nocb_gp().
6. Rename __wake_nocb_leader() to __wake_nocb_gp().
7. Rename wake_nocb_leader_defer() to wake_nocb_gp_defer().
8. Rename rcu_organize_nocb_kthreads() local variable.
9. Rename and document no-CB CB kthread sleep trace event.
10. Rename rcu_nocb_leader_stride kernel boot parameter.
11. Print gp/cb kthread hierarchy if dump_tree.
Thanx, Paul
------------------------------------------------------------------------
Documentation/admin-guide/kernel-parameters.txt | 13 - include/trace/events/rcu.h | 3 kernel/rcu/tree.h | 28 +- kernel/rcu/tree_plugin.h | 312 ++++++++++++------------ 4 files changed, 183 insertions(+), 173 deletions(-)
PATCH 1/11 rcu/nocb: Rename rcu_data fields to prepare for forward-progress work
This commit simply renames rcu_data fields to prepare for leader nocb kthreads doing only grace-period work and callback shuffling. This will mean the addition of replacement kthreads to invoke callbacks. The "leader" and "follower" thus become less meaningful, so the commit changes no-CB fields with these strings to "gp" and "cb", respectively.
--- kernel/rcu/tree.h | 14 ++++---- kernel/rcu/tree_plugin.h | 78 ++++++++++++++++++++-------------------- 2 files changed, 46 insertions(+), 46 deletions(-)
kernel/rcu/tree.h +7 -7
@@ -198,10 +198,10 @@ struct rcu_data {struct rcu_head **nocb_tail;atomic_long_t nocb_q_count;/* # CBs waiting for nocb */ atomic_long_t nocb_q_count_lazy;/* invocation (all stages). */ -struct rcu_head *nocb_follower_head;/* CBs ready to invoke. */ -struct rcu_head **nocb_follower_tail; +struct rcu_head *nocb_cb_head;/* CBs ready to invoke. */ +struct rcu_head **nocb_cb_tail;struct swait_queue_head nocb_wq;/* For nocb kthreads to sleep on. */ -struct task_struct *nocb_kthread; +struct task_struct *nocb_cb_kthread;raw_spinlock_t nocb_lock;/* Guard following pair of fields. */ int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */
@@ -210,12 +210,12 @@ struct rcu_data {struct rcu_head *nocb_gp_head____cacheline_internodealigned_in_smp ;/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail; -bool nocb_leader_sleep;/* Is the nocb leader thread asleep? */ -struct rcu_data *nocb_next_follower; -/* Next follower in wakeup chain. */ +bool nocb_gp_sleep;/* Is the nocb leader thread asleep? */ +struct rcu_data *nocb_next_cb_rdp; +/* Next rcu_data in wakeup chain. */ /* The following fields are used by the follower, hence new cachline. */ -struct rcu_data *nocb_leader____cacheline_internodealigned_in_smp ; +struct rcu_data *nocb_gp_rdp____cacheline_internodealigned_in_smp ;/* Leader CPU takes GP-end wakeups. */ #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
kernel/rcu/tree_plugin.h +39 -39
199 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH 2/11 rcu/nocb: Update comments to prepare for forward-progress work
In reply to Paul E. McKenneyThis commit simply rewords comments to prepare for leader nocb kthreads doing only grace-period work and callback shuffling. This will mean the addition of replacement kthreads to invoke callbacks. The "leader" and "follower" thus become less meaningful, so the commit changes no-CB comments with these strings to "GP" and "CB", respectively. (Give or take the usual grammatical transformations.)
--- kernel/rcu/tree.h | 8 +++--- kernel/rcu/tree_plugin.h | 57 ++++++++++++++++++++-------------------- 2 files changed, 33 insertions(+), 32 deletions(-)
kernel/rcu/tree.h +4 -4
@@ -206,17 +206,17 @@ struct rcu_data {int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */ -/* The following fields are used by the leader, hence own cacheline. */ +/* The following fields are used by GP kthread, hence own cacheline. */ struct rcu_head *nocb_gp_head____cacheline_internodealigned_in_smp ;/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail; -bool nocb_gp_sleep;/* Is the nocb leader thread asleep? */ +bool nocb_gp_sleep;/* Is the nocb GP thread asleep? */ struct rcu_data *nocb_next_cb_rdp;/* Next rcu_data in wakeup chain. */ -/* The following fields are used by the follower, hence new cachline. */ +/* The following fields are used by CB kthread, hence new cachline. */ struct rcu_data *nocb_gp_rdp____cacheline_internodealigned_in_smp ; -/* Leader CPU takes GP-end wakeups. */ +/* GP rdp takes GP-end wakeups. */ #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ /* 6) RCU priority boosting. */
kernel/rcu/tree_plugin.h +29 -28
170 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH 3/11 rcu/nocb: Provide separate no-CBs grace-period kthreads
In reply to Paul E. McKenneyCurrently, there is one no-CBs rcuo kthread per CPU, and these kthreads are divided into groups. The first rcuo kthread to come online in a given group is that group's leader, and the leader both waits for grace periods and invokes its CPU's callbacks. The non-leader rcuo kthreads only invoke callbacks.
This works well in the real-time/embedded environments for which it was intended because such environments tend not to generate all that many callbacks. However, given huge floods of callbacks, it is possible for the leader kthread to be stuck invoking callbacks while its followers wait helplessly while their callbacks pile up. This is a good recipe for an OOM, and rcutorture's new callback-flood capability does generate such OOMs.
One strategy would be to wait until such OOMs start happening in production, but similar OOMs have in fact happened starting in 2018. It would therefore be wise to take a more proactive approach.
This commit therefore features per-CPU rcuo kthreads that do nothing but invoke callbacks. Instead of having one of these kthreads act as leader, each group has a separate rcog kthread that handles grace periods for its group. Because these rcuog kthreads do not invoke callbacks, callback floods on one CPU no longer block callbacks from reaching the rcuc callback-invocation kthreads on other CPUs.
This change does introduce additional kthreads, however:
1. The number of additional kthreads is about the square root of
the number of CPUs, so that a 4096-CPU system would have only about 64 additional kthreads. Note that recent changes decreased the number of rcuo kthreads by a factor of two (CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so this still represents a significant improvement on most systems.
2. The leading "rcuo" of the rcuog kthreads should allow existing
scripting to affinity these additional kthreads as needed, the same as for the rcuop and rcuos kthreads. (There are no longer any rcuob kthreads.)
3. A state-machine approach was considered and rejected. Although
this would allow the rcuo kthreads to continue their dual leader/follower roles, it complicates callback invocation and makes it more difficult to consolidate rcuo callback invocation with existing softirq callback invocation.
The introduction of rcuog kthreads should thus be acceptable.
--- kernel/rcu/tree.h | 6 +- kernel/rcu/tree_plugin.h | 115 +++++++++++++++++++-------------------- 2 files changed, 61 insertions(+), 60 deletions(-)
kernel/rcu/tree.h +4 -2
@@ -200,8 +200,8 @@ struct rcu_data {atomic_long_t nocb_q_count_lazy;/* invocation (all stages). */ struct rcu_head *nocb_cb_head;/* CBs ready to invoke. */ struct rcu_head **nocb_cb_tail; -struct swait_queue_head nocb_wq;/* For nocb kthreads to sleep on. */ -struct task_struct *nocb_cb_kthread; +struct swait_queue_head nocb_cb_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_gp_kthread;raw_spinlock_t nocb_lock;/* Guard following pair of fields. */ int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */
@@ -211,6 +211,8 @@ struct rcu_data {/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail;bool nocb_gp_sleep;/* Is the nocb GP thread asleep? */ +struct swait_queue_head nocb_gp_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_cb_kthread;struct rcu_data *nocb_next_cb_rdp;/* Next rcu_data in wakeup chain. */
kernel/rcu/tree_plugin.h +57 -58
212 lines not on this page · raw message
Signature · 1 line
2.17.1
Re: [PATCH tip/core/rcu 03/11] rcu/nocb: Provide separate no-CBs grace-period kthreads
Quoted 14 lines · Paul E. McKenney
Currently, there is one no-CBs rcuo kthread per CPU, and these kthreads are divided into groups. The first rcuo kthread to come online in a given group is that group's leader, and the leader both waits for grace periods and invokes its CPU's callbacks. The non-leader rcuo kthreads only invoke callbacks.
This works well in the real-time/embedded environments for which it was intended because such environments tend not to generate all that many callbacks. However, given huge floods of callbacks, it is possible for the leader kthread to be stuck invoking callbacks while its followers wait helplessly while their callbacks pile up. This is a good recipe for an OOM, and rcutorture's new callback-flood capability does generate such OOMs.
Quoted 3 lines
One strategy would be to wait until such OOMs start happening in production, but similar OOMs have in fact happened starting in 2018. It would therefore be wise to take a more proactive approach.
I haven't looked much into nocbs/nohz_full stuff (yet). In particular, I did not even know that the rcuo threads do grace period life-cycle management and waiting, I thought only the RCU GP threads did :-/. however, it seems this is a completely separate grace-period management state machine outside of the RCU GP thread right?
I was wondering for this patch, could we also just have the rcuo leader handle both callback execution and waking other non-leader threads at the same time? So like, execute few callbacks, then do the wake up of the non-leaders to execute their callbacks, the get back to executing their own callbacks, etc. That way we don't need a separate rcuog thread to wait for grace period, would that not work?
If you don't mind could you share with me a kvm.sh command (which has config, boot parameters etc) that can produce the OOM without this patch? I'd like to take a closer look at it.
Is there also a short answer for my the RCU GP thread cannot do the job of these new rcuog threads?
thanks a lot,
- Joel
Quoted 289 lines
This commit therefore features per-CPU rcuo kthreads that do nothing but invoke callbacks. Instead of having one of these kthreads act as leader, each group has a separate rcog kthread that handles grace periods for its group. Because these rcuog kthreads do not invoke callbacks, callback floods on one CPU no longer block callbacks from reaching the rcuc callback-invocation kthreads on other CPUs.
This change does introduce additional kthreads, however:
1. The number of additional kthreads is about the square root of
the number of CPUs, so that a 4096-CPU system would have only about 64 additional kthreads. Note that recent changes decreased the number of rcuo kthreads by a factor of two (CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so this still represents a significant improvement on most systems.2. The leading "rcuo" of the rcuog kthreads should allow existing
scripting to affinity these additional kthreads as needed, the same as for the rcuop and rcuos kthreads. (There are no longer any rcuob kthreads.)3. A state-machine approach was considered and rejected. Although
this would allow the rcuo kthreads to continue their dual leader/follower roles, it complicates callback invocation and makes it more difficult to consolidate rcuo callback invocation with existing softirq callback invocation.The introduction of rcuog kthreads should thus be acceptable.
Signed-off-by Paul E. McKenney <[email protected]>--- kernel/rcu/tree.h | 6 +- kernel/rcu/tree_plugin.h | 115 +++++++++++++++++++-------------------- 2 files changed, 61 insertions(+), 60 deletions(-)kernel/rcu/tree.h +4
-2@@ -200,8 +200,8 @@ struct rcu_data {atomic_long_t nocb_q_count_lazy;/* invocation (all stages). */ struct rcu_head *nocb_cb_head;/* CBs ready to invoke. */ struct rcu_head **nocb_cb_tail; -struct swait_queue_head nocb_wq;/* For nocb kthreads to sleep on. */ -struct task_struct *nocb_cb_kthread; +struct swait_queue_head nocb_cb_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_gp_kthread;raw_spinlock_t nocb_lock;/* Guard following pair of fields. */ int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */ @@ -211,6 +211,8 @@ struct rcu_data {/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail;bool nocb_gp_sleep;/* Is the nocb GP thread asleep? */ +struct swait_queue_head nocb_gp_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_cb_kthread;struct rcu_data *nocb_next_cb_rdp;/* Next rcu_data in wakeup chain. */ kernel/rcu/tree_plugin.h +57
-58@@ -1531,7 +1531,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,struct rcu_data *rdp_leader = rdp->nocb_gp_rdp;lockdep_assert_held (&rdp->nocb_lock); -if (!READ_ONCE (rdp_leader->nocb_cb_kthread)) { +if (!READ_ONCE (rdp_leader->nocb_gp_kthread)) {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);return ; }@@ -1541,7 +1541,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,del_timer (&rdp->nocb_timer);raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);smp_mb ();/* ->nocb_gp_sleep before swake_up_one(). */ -swake_up_one (&rdp_leader->nocb_wq); +swake_up_one (&rdp_leader->nocb_gp_wq); }else {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); }@@ -1646,7 +1646,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,smp_mb__after_atomic ();/* Store *old_rhpp before _wake test. */ /* If we are not being polled and there is a kthread, awaken it ... */ - t =READ_ONCE (rdp->nocb_cb_kthread); + t =READ_ONCE (rdp->nocb_gp_kthread);if (rcu_nocb_poll || !t) {trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeNotPoll" ));@@ -1786,7 +1786,7 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)* No-CBs GP kthreads come here to wait for additional callbacks to show up. * This function does not return until callbacks appear. */ -static void nocb_leader_wait (struct rcu_data *my_rdp) +static void nocb_gp_wait (struct rcu_data *my_rdp) {bool firsttime = true;unsigned long flags;@@ -1794,12 +1794,10 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)struct rcu_data *rdp;struct rcu_head **tail; -wait_again: -/* Wait for callbacks to appear. */ if (!rcu_nocb_poll) {trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("Sleep" )); -swait_event_interruptible_exclusive (my_rdp->nocb_wq, +swait_event_interruptible_exclusive (my_rdp->nocb_gp_wq, !READ_ONCE (my_rdp->nocb_gp_sleep));raw_spin_lock_irqsave (&my_rdp->nocb_lock, flags); my_rdp->nocb_gp_sleep = true;@@ -1838,7 +1836,7 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("WokeEmpty" )); } - goto wait_again; +return ; }/* Wait for one grace period. */ @@ -1862,34 +1860,47 @@ static void nocb_leader_wait(struct rcu_data *my_rdp) rdp->nocb_cb_tail = rdp->nocb_gp_tail; *tail = rdp->nocb_gp_head;raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); -if (rdp != my_rdp && tail == &rdp->nocb_cb_head) { +if (tail == &rdp->nocb_cb_head) {/* List was empty, so wake up the kthread. */ -swake_up_one (&rdp->nocb_wq); +swake_up_one (&rdp->nocb_cb_wq); } } +} -/* If we (the GP kthreads) don't have CBs, go wait some more. */ -if (!my_rdp->nocb_cb_head) - goto wait_again; +/* +* No-CBs grace-period-wait kthread. There is one of these per group +* of CPUs, but only once at least one CPU in that group has come online +* at least once since boot. This kthread checks for newly posted +* callbacks from any of the CPUs it is responsible for, waits for a +* grace period, then awakens all of the rcu_nocb_cb_kthread() instances +* that then have callback-invocation work to do. +*/ +static int rcu_nocb_gp_kthread (void *arg) +{ +struct rcu_data *rdp = arg; + +for (;;) +nocb_gp_wait (rdp); +return 0 ; }/* * No-CBs CB kthreads come here to wait for additional callbacks to show up. -* This function does not return until callbacks appear. +* This function returns true ("keep waiting") until callbacks appear and +* then false ("stop waiting") when callbacks finally do appear. */ -static void nocb_follower_wait (struct rcu_data *rdp) +static bool nocb_follower_wait (struct rcu_data *rdp) { -for (;;) { -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); -swait_event_interruptible_exclusive (rdp->nocb_wq, -READ_ONCE (rdp->nocb_cb_head)); -if (smp_load_acquire (&rdp->nocb_cb_head)) { -/* ^^^ Ensure CB invocation follows _head test. */ -return ; - } -WARN_ON (signal_pending (current)); -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); +swait_event_interruptible_exclusive (rdp->nocb_cb_wq, +READ_ONCE (rdp->nocb_cb_head)); +if (smp_load_acquire (&rdp->nocb_cb_head)) {/* VVV */ +/* ^^^ Ensure CB invocation follows _head test. */ +return false; } +WARN_ON (signal_pending (current)); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +return true; }/* @@ -1899,7 +1910,7 @@ static void nocb_follower_wait(struct rcu_data *rdp)* have to do quite so many wakeups (as in they only need to wake the * no-CBs GP kthreads, not the CB kthreads). */ -static int rcu_nocb_kthread (void *arg) +static int rcu_nocb_cb_kthread (void *arg) {int c, cl;unsigned long flags;@@ -1911,10 +1922,8 @@ static int rcu_nocb_kthread(void *arg)/* Each pass through this loop invokes one batch of callbacks */ for (;;) {/* Wait for callbacks. */ -if (rdp->nocb_gp_rdp == rdp) -nocb_leader_wait (rdp); -else -nocb_follower_wait (rdp); +while (nocb_follower_wait (rdp)) +continue ;/* Pull the ready-to-invoke callbacks onto local list. */ raw_spin_lock_irqsave (&rdp->nocb_lock, flags);@@ -2048,7 +2057,8 @@ void __init rcu_init_nohz(void)static void __init rcu_boot_init_nocb_percpu_data (struct rcu_data *rdp) { rdp->nocb_tail = &rdp->nocb_head; -init_swait_queue_head (&rdp->nocb_wq); +init_swait_queue_head (&rdp->nocb_cb_wq); +init_swait_queue_head (&rdp->nocb_gp_wq); rdp->nocb_cb_tail = &rdp->nocb_cb_head;raw_spin_lock_init (&rdp->nocb_lock);timer_setup (&rdp->nocb_timer, do_nocb_deferred_wakeup_timer,0 );@@ -2056,50 +2066,39 @@ static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)/* * If the specified CPU is a no-CBs CPU that does not already have its -* rcuo kthread, spawn it. If the CPUs are brought online out of order, -* this can require re-organizing the GP-CB relationships. +* rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread +* for this CPU's group has not yet been created, spawn it as well. */ static void rcu_spawn_one_nocb_kthread (int cpu) { -struct rcu_data *rdp; -struct rcu_data *rdp_last; -struct rcu_data *rdp_old_leader; -struct rcu_data *rdp_spawn =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp_gp;struct task_struct *t;/* * If this isn't a no-CBs CPU or if it already has an rcuo kthread, * then nothing to do. */ -if (!rcu_is_nocb_cpu (cpu) || rdp_spawn->nocb_cb_kthread) +if (!rcu_is_nocb_cpu (cpu) || rdp->nocb_cb_kthread)return ;/* If we didn't spawn the GP kthread first, reorganize! */ - rdp_old_leader = rdp_spawn->nocb_gp_rdp; -if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_cb_kthread) { - rdp_last =NULL ; - rdp = rdp_old_leader; -do { - rdp->nocb_gp_rdp = rdp_spawn; -if (rdp_last && rdp != rdp_spawn) - rdp_last->nocb_next_cb_rdp = rdp; -if (rdp == rdp_spawn) { - rdp = rdp->nocb_next_cb_rdp; - }else { - rdp_last = rdp; - rdp = rdp->nocb_next_cb_rdp; - rdp_last->nocb_next_cb_rdp =NULL ; - } - }while (rdp); - rdp_spawn->nocb_next_cb_rdp = rdp_old_leader; + rdp_gp = rdp->nocb_gp_rdp; +if (!rdp_gp->nocb_gp_kthread) { + t =kthread_run (rcu_nocb_gp_kthread, rdp_gp, +"rcuog/%d" , rdp_gp->cpu); +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo GP kthread, OOM is now expected behavior\n" , __func__)) +return ; +WRITE_ONCE (rdp_gp->nocb_gp_kthread, t); }/* Spawn the kthread for this CPU. */ - t =kthread_run (rcu_nocb_kthread, rdp_spawn, + t =kthread_run (rcu_nocb_cb_kthread, rdp,"rcuo%c/%d" , rcu_state.abbr, cpu); -if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo kthread, OOM is now expected behavior\n" , __func__)) +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo CB kthread, OOM is now expected behavior\n" , __func__))return ; -WRITE_ONCE (rdp_spawn->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread); }/* Signature · 1 line
2.17.1
Re: [PATCH tip/core/rcu 03/11] rcu/nocb: Provide separate no-CBs grace-period kthreads
Quoted 19 lines · Joel Fernandes
Quoted 14 lines · Paul E. McKenney
Currently, there is one no-CBs rcuo kthread per CPU, and these kthreads are divided into groups. The first rcuo kthread to come online in a given group is that group's leader, and the leader both waits for grace periods and invokes its CPU's callbacks. The non-leader rcuo kthreads only invoke callbacks.
This works well in the real-time/embedded environments for which it was intended because such environments tend not to generate all that many callbacks. However, given huge floods of callbacks, it is possible for the leader kthread to be stuck invoking callbacks while its followers wait helplessly while their callbacks pile up. This is a good recipe for an OOM, and rcutorture's new callback-flood capability does generate such OOMs.
Quoted 3 lines
One strategy would be to wait until such OOMs start happening in production, but similar OOMs have in fact happened starting in 2018. It would therefore be wise to take a more proactive approach.
Quoted 5 lines
I haven't looked much into nocbs/nohz_full stuff (yet). In particular, I did not even know that the rcuo threads do grace period life-cycle management and waiting, I thought only the RCU GP threads did :-/. however, it seems this is a completely separate grace-period management state machine outside of the RCU GP thread right?
No, the rcuo kthreads interact with the main RCU GP kthread, initiating new grace periods when needed and being awakened as needed by the RCU GP kthread.
Quoted 6 lines
I was wondering for this patch, could we also just have the rcuo leader handle both callback execution and waking other non-leader threads at the same time? So like, execute few callbacks, then do the wake up of the non-leaders to execute their callbacks, the get back to executing their own callbacks, etc. That way we don't need a separate rcuog thread to wait for grace period, would that not work?
I did look into that, but it was more complex and also didn't foster sharing of rcu_do_batch(), which used to only be for non-offloaded callbacks but now does both. Besides which, invoking callbacks would degrade the rcuog kthread's response to new callbacks and the like.
Quoted 3 lines
If you don't mind could you share with me a
kvm.shcommand (which has config, boot parameters etc) that can produce the OOM without this patch? I'd like to take a closer look at it.
Here you go:
tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 8 --duration 120 --configs TREE04
If you add "--memory 1G" in mainline, the OOMs go away. Or at least decrease substantially in probability.
Quoted 2 lines
Is there also a short answer for my the RCU GP thread cannot do the job of these new rcuog threads?
First, the code is more complicated when you do it that way (and yes, I did actually write it out in pen on paper). Second, if the CPU corresponding to the combined grace-period/callback kthread is doing the call_rcu() flooding, you are between a rock and a hard place. On the one hand, you want that kthread to do nothing but invoke callbacks so as to have half a chance of keeping up, and on the other hand you need it to check state frequently so as to react in a timely fashion to a CPU corresponding to one of its callback kthreads starting a second callback flood.
Introducing rcug grace-period-only kthreads means that you get the best of both worlds. Plus last year's flavor consolidation decreased the number of rcuo kthreads from either 2N or 3N to N, so increasing it to only (N + sqrt(N)) should be just fine. Though I would expect that there will be at least some screaming and shouting. ;-)
Thanx, Paul
Quoted 295 lines
thanks a lot,
- Joel
Quoted 289 lines
This commit therefore features per-CPU rcuo kthreads that do nothing but invoke callbacks. Instead of having one of these kthreads act as leader, each group has a separate rcog kthread that handles grace periods for its group. Because these rcuog kthreads do not invoke callbacks, callback floods on one CPU no longer block callbacks from reaching the rcuc callback-invocation kthreads on other CPUs.
This change does introduce additional kthreads, however:
1. The number of additional kthreads is about the square root of
the number of CPUs, so that a 4096-CPU system would have only about 64 additional kthreads. Note that recent changes decreased the number of rcuo kthreads by a factor of two (CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so this still represents a significant improvement on most systems.2. The leading "rcuo" of the rcuog kthreads should allow existing
scripting to affinity these additional kthreads as needed, the same as for the rcuop and rcuos kthreads. (There are no longer any rcuob kthreads.)3. A state-machine approach was considered and rejected. Although
this would allow the rcuo kthreads to continue their dual leader/follower roles, it complicates callback invocation and makes it more difficult to consolidate rcuo callback invocation with existing softirq callback invocation.The introduction of rcuog kthreads should thus be acceptable.
Signed-off-by Paul E. McKenney <[email protected]>--- kernel/rcu/tree.h | 6 +- kernel/rcu/tree_plugin.h | 115 +++++++++++++++++++-------------------- 2 files changed, 61 insertions(+), 60 deletions(-)kernel/rcu/tree.h +4
-2@@ -200,8 +200,8 @@ struct rcu_data {atomic_long_t nocb_q_count_lazy;/* invocation (all stages). */ struct rcu_head *nocb_cb_head;/* CBs ready to invoke. */ struct rcu_head **nocb_cb_tail; -struct swait_queue_head nocb_wq;/* For nocb kthreads to sleep on. */ -struct task_struct *nocb_cb_kthread; +struct swait_queue_head nocb_cb_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_gp_kthread;raw_spinlock_t nocb_lock;/* Guard following pair of fields. */ int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */ @@ -211,6 +211,8 @@ struct rcu_data {/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail;bool nocb_gp_sleep;/* Is the nocb GP thread asleep? */ +struct swait_queue_head nocb_gp_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_cb_kthread;struct rcu_data *nocb_next_cb_rdp;/* Next rcu_data in wakeup chain. */ kernel/rcu/tree_plugin.h +57
-58@@ -1531,7 +1531,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,struct rcu_data *rdp_leader = rdp->nocb_gp_rdp;lockdep_assert_held (&rdp->nocb_lock); -if (!READ_ONCE (rdp_leader->nocb_cb_kthread)) { +if (!READ_ONCE (rdp_leader->nocb_gp_kthread)) {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);return ; }@@ -1541,7 +1541,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,del_timer (&rdp->nocb_timer);raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);smp_mb ();/* ->nocb_gp_sleep before swake_up_one(). */ -swake_up_one (&rdp_leader->nocb_wq); +swake_up_one (&rdp_leader->nocb_gp_wq); }else {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); }@@ -1646,7 +1646,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,smp_mb__after_atomic ();/* Store *old_rhpp before _wake test. */ /* If we are not being polled and there is a kthread, awaken it ... */ - t =READ_ONCE (rdp->nocb_cb_kthread); + t =READ_ONCE (rdp->nocb_gp_kthread);if (rcu_nocb_poll || !t) {trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeNotPoll" ));@@ -1786,7 +1786,7 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)* No-CBs GP kthreads come here to wait for additional callbacks to show up. * This function does not return until callbacks appear. */ -static void nocb_leader_wait (struct rcu_data *my_rdp) +static void nocb_gp_wait (struct rcu_data *my_rdp) {bool firsttime = true;unsigned long flags;@@ -1794,12 +1794,10 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)struct rcu_data *rdp;struct rcu_head **tail; -wait_again: -/* Wait for callbacks to appear. */ if (!rcu_nocb_poll) {trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("Sleep" )); -swait_event_interruptible_exclusive (my_rdp->nocb_wq, +swait_event_interruptible_exclusive (my_rdp->nocb_gp_wq, !READ_ONCE (my_rdp->nocb_gp_sleep));raw_spin_lock_irqsave (&my_rdp->nocb_lock, flags); my_rdp->nocb_gp_sleep = true;@@ -1838,7 +1836,7 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("WokeEmpty" )); } - goto wait_again; +return ; }/* Wait for one grace period. */ @@ -1862,34 +1860,47 @@ static void nocb_leader_wait(struct rcu_data *my_rdp) rdp->nocb_cb_tail = rdp->nocb_gp_tail; *tail = rdp->nocb_gp_head;raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); -if (rdp != my_rdp && tail == &rdp->nocb_cb_head) { +if (tail == &rdp->nocb_cb_head) {/* List was empty, so wake up the kthread. */ -swake_up_one (&rdp->nocb_wq); +swake_up_one (&rdp->nocb_cb_wq); } } +} -/* If we (the GP kthreads) don't have CBs, go wait some more. */ -if (!my_rdp->nocb_cb_head) - goto wait_again; +/* +* No-CBs grace-period-wait kthread. There is one of these per group +* of CPUs, but only once at least one CPU in that group has come online +* at least once since boot. This kthread checks for newly posted +* callbacks from any of the CPUs it is responsible for, waits for a +* grace period, then awakens all of the rcu_nocb_cb_kthread() instances +* that then have callback-invocation work to do. +*/ +static int rcu_nocb_gp_kthread (void *arg) +{ +struct rcu_data *rdp = arg; + +for (;;) +nocb_gp_wait (rdp); +return 0 ; }/* * No-CBs CB kthreads come here to wait for additional callbacks to show up. -* This function does not return until callbacks appear. +* This function returns true ("keep waiting") until callbacks appear and +* then false ("stop waiting") when callbacks finally do appear. */ -static void nocb_follower_wait (struct rcu_data *rdp) +static bool nocb_follower_wait (struct rcu_data *rdp) { -for (;;) { -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); -swait_event_interruptible_exclusive (rdp->nocb_wq, -READ_ONCE (rdp->nocb_cb_head)); -if (smp_load_acquire (&rdp->nocb_cb_head)) { -/* ^^^ Ensure CB invocation follows _head test. */ -return ; - } -WARN_ON (signal_pending (current)); -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); +swait_event_interruptible_exclusive (rdp->nocb_cb_wq, +READ_ONCE (rdp->nocb_cb_head)); +if (smp_load_acquire (&rdp->nocb_cb_head)) {/* VVV */ +/* ^^^ Ensure CB invocation follows _head test. */ +return false; } +WARN_ON (signal_pending (current)); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +return true; }/* @@ -1899,7 +1910,7 @@ static void nocb_follower_wait(struct rcu_data *rdp)* have to do quite so many wakeups (as in they only need to wake the * no-CBs GP kthreads, not the CB kthreads). */ -static int rcu_nocb_kthread (void *arg) +static int rcu_nocb_cb_kthread (void *arg) {int c, cl;unsigned long flags;@@ -1911,10 +1922,8 @@ static int rcu_nocb_kthread(void *arg)/* Each pass through this loop invokes one batch of callbacks */ for (;;) {/* Wait for callbacks. */ -if (rdp->nocb_gp_rdp == rdp) -nocb_leader_wait (rdp); -else -nocb_follower_wait (rdp); +while (nocb_follower_wait (rdp)) +continue ;/* Pull the ready-to-invoke callbacks onto local list. */ raw_spin_lock_irqsave (&rdp->nocb_lock, flags);@@ -2048,7 +2057,8 @@ void __init rcu_init_nohz(void)static void __init rcu_boot_init_nocb_percpu_data (struct rcu_data *rdp) { rdp->nocb_tail = &rdp->nocb_head; -init_swait_queue_head (&rdp->nocb_wq); +init_swait_queue_head (&rdp->nocb_cb_wq); +init_swait_queue_head (&rdp->nocb_gp_wq); rdp->nocb_cb_tail = &rdp->nocb_cb_head;raw_spin_lock_init (&rdp->nocb_lock);timer_setup (&rdp->nocb_timer, do_nocb_deferred_wakeup_timer,0 );@@ -2056,50 +2066,39 @@ static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)/* * If the specified CPU is a no-CBs CPU that does not already have its -* rcuo kthread, spawn it. If the CPUs are brought online out of order, -* this can require re-organizing the GP-CB relationships. +* rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread +* for this CPU's group has not yet been created, spawn it as well. */ static void rcu_spawn_one_nocb_kthread (int cpu) { -struct rcu_data *rdp; -struct rcu_data *rdp_last; -struct rcu_data *rdp_old_leader; -struct rcu_data *rdp_spawn =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp_gp;struct task_struct *t;/* * If this isn't a no-CBs CPU or if it already has an rcuo kthread, * then nothing to do. */ -if (!rcu_is_nocb_cpu (cpu) || rdp_spawn->nocb_cb_kthread) +if (!rcu_is_nocb_cpu (cpu) || rdp->nocb_cb_kthread)return ;/* If we didn't spawn the GP kthread first, reorganize! */ - rdp_old_leader = rdp_spawn->nocb_gp_rdp; -if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_cb_kthread) { - rdp_last =NULL ; - rdp = rdp_old_leader; -do { - rdp->nocb_gp_rdp = rdp_spawn; -if (rdp_last && rdp != rdp_spawn) - rdp_last->nocb_next_cb_rdp = rdp; -if (rdp == rdp_spawn) { - rdp = rdp->nocb_next_cb_rdp; - }else { - rdp_last = rdp; - rdp = rdp->nocb_next_cb_rdp; - rdp_last->nocb_next_cb_rdp =NULL ; - } - }while (rdp); - rdp_spawn->nocb_next_cb_rdp = rdp_old_leader; + rdp_gp = rdp->nocb_gp_rdp; +if (!rdp_gp->nocb_gp_kthread) { + t =kthread_run (rcu_nocb_gp_kthread, rdp_gp, +"rcuog/%d" , rdp_gp->cpu); +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo GP kthread, OOM is now expected behavior\n" , __func__)) +return ; +WRITE_ONCE (rdp_gp->nocb_gp_kthread, t); }/* Spawn the kthread for this CPU. */ - t =kthread_run (rcu_nocb_kthread, rdp_spawn, + t =kthread_run (rcu_nocb_cb_kthread, rdp,"rcuo%c/%d" , rcu_state.abbr, cpu); -if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo kthread, OOM is now expected behavior\n" , __func__)) +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo CB kthread, OOM is now expected behavior\n" , __func__))return ; -WRITE_ONCE (rdp_spawn->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread); }/* Signature · 1 line
2.17.1
Re: [PATCH tip/core/rcu 03/11] rcu/nocb: Provide separate no-CBs grace-period kthreads
Quoted 26 lines · Paul E. McKenney
Quoted 19 lines · Joel Fernandes
Quoted 14 lines · Paul E. McKenney
Currently, there is one no-CBs rcuo kthread per CPU, and these kthreads are divided into groups. The first rcuo kthread to come online in a given group is that group's leader, and the leader both waits for grace periods and invokes its CPU's callbacks. The non-leader rcuo kthreads only invoke callbacks.
This works well in the real-time/embedded environments for which it was intended because such environments tend not to generate all that many callbacks. However, given huge floods of callbacks, it is possible for the leader kthread to be stuck invoking callbacks while its followers wait helplessly while their callbacks pile up. This is a good recipe for an OOM, and rcutorture's new callback-flood capability does generate such OOMs.
Quoted 3 lines
One strategy would be to wait until such OOMs start happening in production, but similar OOMs have in fact happened starting in 2018. It would therefore be wise to take a more proactive approach.
Quoted 5 lines
I haven't looked much into nocbs/nohz_full stuff (yet). In particular, I did not even know that the rcuo threads do grace period life-cycle management and waiting, I thought only the RCU GP threads did :-/. however, it seems this is a completely separate grace-period management state machine outside of the RCU GP thread right?
Quoted 3 lines
No, the rcuo kthreads interact with the main RCU GP kthread, initiating new grace periods when needed and being awakened as needed by the RCU GP kthread.
Ok, I see the interactions in rcu_nocb_wait_gp(). This what I was thinking too is that there has to be these interactions with the main RCU GP kthread, for anything to work :) Thanks for the explanation!
Quoted 11 lines
Quoted 6 lines
I was wondering for this patch, could we also just have the rcuo leader handle both callback execution and waking other non-leader threads at the same time? So like, execute few callbacks, then do the wake up of the non-leaders to execute their callbacks, the get back to executing their own callbacks, etc. That way we don't need a separate rcuog thread to wait for grace period, would that not work?
I did look into that, but it was more complex and also didn't foster sharing of
rcu_do_batch(), which used to only be for non-offloaded callbacks but now does both. Besides which, invoking callbacks would degrade the rcuog kthread's response to new callbacks and the like.
Makes sense.
Quoted 10 lines
Quoted 3 lines
If you don't mind could you share with me a
kvm.shcommand (which has config, boot parameters etc) that can produce the OOM without this patch? I'd like to take a closer look at it.Here you go:
tools/testing/selftests/rcutorture/bin/kvm.sh--cpus 8 --duration 120 --configs TREE04If you add "--memory 1G" in mainline, the OOMs go away. Or at least decrease substantially in probability.
I could reproduce it, it look around 5-10 minutes for the OOM for 512MB memory. Thanks.
Quoted 13 lines
Quoted 2 lines
Is there also a short answer for my the RCU GP thread cannot do the job of these new rcuog threads?
First, the code is more complicated when you do it that way (and yes, I did actually write it out in pen on paper). Second, if the CPU corresponding to the combined grace-period/callback kthread is doing the
call_rcu()flooding, you are between a rock and a hard place. On the one hand, you want that kthread to do nothing but invoke callbacks so as to have half a chance of keeping up, and on the other hand you need it to check state frequently so as to react in a timely fashion to a CPU corresponding to one of its callback kthreads starting a second callback flood.
Quoted 5 lines
Introducing rcug grace-period-only kthreads means that you get the best of both worlds. Plus last year's flavor consolidation decreased the number of rcuo kthreads from either 2N or 3N to N, so increasing it to only (N + sqrt(N)) should be just fine. Though I would expect that there will be at least some screaming and shouting. ;-)
Ok got it. Yes, fewer newer threads now even with nocb improvements :)
thanks,
- Joel
Quoted 298 lines
Thanx, Paul
Quoted 295 lines
thanks a lot,
- Joel
Quoted 289 lines
This commit therefore features per-CPU rcuo kthreads that do nothing but invoke callbacks. Instead of having one of these kthreads act as leader, each group has a separate rcog kthread that handles grace periods for its group. Because these rcuog kthreads do not invoke callbacks, callback floods on one CPU no longer block callbacks from reaching the rcuc callback-invocation kthreads on other CPUs.
This change does introduce additional kthreads, however:
1. The number of additional kthreads is about the square root of
the number of CPUs, so that a 4096-CPU system would have only about 64 additional kthreads. Note that recent changes decreased the number of rcuo kthreads by a factor of two (CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so this still represents a significant improvement on most systems.2. The leading "rcuo" of the rcuog kthreads should allow existing
scripting to affinity these additional kthreads as needed, the same as for the rcuop and rcuos kthreads. (There are no longer any rcuob kthreads.)3. A state-machine approach was considered and rejected. Although
this would allow the rcuo kthreads to continue their dual leader/follower roles, it complicates callback invocation and makes it more difficult to consolidate rcuo callback invocation with existing softirq callback invocation.The introduction of rcuog kthreads should thus be acceptable.
Signed-off-by Paul E. McKenney <[email protected]>--- kernel/rcu/tree.h | 6 +- kernel/rcu/tree_plugin.h | 115 +++++++++++++++++++-------------------- 2 files changed, 61 insertions(+), 60 deletions(-)kernel/rcu/tree.h +4
-2@@ -200,8 +200,8 @@ struct rcu_data {atomic_long_t nocb_q_count_lazy;/* invocation (all stages). */ struct rcu_head *nocb_cb_head;/* CBs ready to invoke. */ struct rcu_head **nocb_cb_tail; -struct swait_queue_head nocb_wq;/* For nocb kthreads to sleep on. */ -struct task_struct *nocb_cb_kthread; +struct swait_queue_head nocb_cb_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_gp_kthread;raw_spinlock_t nocb_lock;/* Guard following pair of fields. */ int nocb_defer_wakeup;/* Defer wakeup of nocb_kthread. */ struct timer_list nocb_timer;/* Enforce finite deferral. */ @@ -211,6 +211,8 @@ struct rcu_data {/* CBs waiting for GP. */ struct rcu_head **nocb_gp_tail;bool nocb_gp_sleep;/* Is the nocb GP thread asleep? */ +struct swait_queue_head nocb_gp_wq;/* For nocb kthreads to sleep on. */ +struct task_struct *nocb_cb_kthread;struct rcu_data *nocb_next_cb_rdp;/* Next rcu_data in wakeup chain. */ kernel/rcu/tree_plugin.h +57
-58@@ -1531,7 +1531,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,struct rcu_data *rdp_leader = rdp->nocb_gp_rdp;lockdep_assert_held (&rdp->nocb_lock); -if (!READ_ONCE (rdp_leader->nocb_cb_kthread)) { +if (!READ_ONCE (rdp_leader->nocb_gp_kthread)) {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);return ; }@@ -1541,7 +1541,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,del_timer (&rdp->nocb_timer);raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);smp_mb ();/* ->nocb_gp_sleep before swake_up_one(). */ -swake_up_one (&rdp_leader->nocb_wq); +swake_up_one (&rdp_leader->nocb_gp_wq); }else {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); }@@ -1646,7 +1646,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,smp_mb__after_atomic ();/* Store *old_rhpp before _wake test. */ /* If we are not being polled and there is a kthread, awaken it ... */ - t =READ_ONCE (rdp->nocb_cb_kthread); + t =READ_ONCE (rdp->nocb_gp_kthread);if (rcu_nocb_poll || !t) {trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeNotPoll" ));@@ -1786,7 +1786,7 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)* No-CBs GP kthreads come here to wait for additional callbacks to show up. * This function does not return until callbacks appear. */ -static void nocb_leader_wait (struct rcu_data *my_rdp) +static void nocb_gp_wait (struct rcu_data *my_rdp) {bool firsttime = true;unsigned long flags;@@ -1794,12 +1794,10 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)struct rcu_data *rdp;struct rcu_head **tail; -wait_again: -/* Wait for callbacks to appear. */ if (!rcu_nocb_poll) {trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("Sleep" )); -swait_event_interruptible_exclusive (my_rdp->nocb_wq, +swait_event_interruptible_exclusive (my_rdp->nocb_gp_wq, !READ_ONCE (my_rdp->nocb_gp_sleep));raw_spin_lock_irqsave (&my_rdp->nocb_lock, flags); my_rdp->nocb_gp_sleep = true;@@ -1838,7 +1836,7 @@ static void nocb_leader_wait(struct rcu_data *my_rdp)trace_rcu_nocb_wake (rcu_state.name, my_rdp->cpu,TPS ("WokeEmpty" )); } - goto wait_again; +return ; }/* Wait for one grace period. */ @@ -1862,34 +1860,47 @@ static void nocb_leader_wait(struct rcu_data *my_rdp) rdp->nocb_cb_tail = rdp->nocb_gp_tail; *tail = rdp->nocb_gp_head;raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); -if (rdp != my_rdp && tail == &rdp->nocb_cb_head) { +if (tail == &rdp->nocb_cb_head) {/* List was empty, so wake up the kthread. */ -swake_up_one (&rdp->nocb_wq); +swake_up_one (&rdp->nocb_cb_wq); } } +} -/* If we (the GP kthreads) don't have CBs, go wait some more. */ -if (!my_rdp->nocb_cb_head) - goto wait_again; +/* +* No-CBs grace-period-wait kthread. There is one of these per group +* of CPUs, but only once at least one CPU in that group has come online +* at least once since boot. This kthread checks for newly posted +* callbacks from any of the CPUs it is responsible for, waits for a +* grace period, then awakens all of the rcu_nocb_cb_kthread() instances +* that then have callback-invocation work to do. +*/ +static int rcu_nocb_gp_kthread (void *arg) +{ +struct rcu_data *rdp = arg; + +for (;;) +nocb_gp_wait (rdp); +return 0 ; }/* * No-CBs CB kthreads come here to wait for additional callbacks to show up. -* This function does not return until callbacks appear. +* This function returns true ("keep waiting") until callbacks appear and +* then false ("stop waiting") when callbacks finally do appear. */ -static void nocb_follower_wait (struct rcu_data *rdp) +static bool nocb_follower_wait (struct rcu_data *rdp) { -for (;;) { -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); -swait_event_interruptible_exclusive (rdp->nocb_wq, -READ_ONCE (rdp->nocb_cb_head)); -if (smp_load_acquire (&rdp->nocb_cb_head)) { -/* ^^^ Ensure CB invocation follows _head test. */ -return ; - } -WARN_ON (signal_pending (current)); -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); +swait_event_interruptible_exclusive (rdp->nocb_cb_wq, +READ_ONCE (rdp->nocb_cb_head)); +if (smp_load_acquire (&rdp->nocb_cb_head)) {/* VVV */ +/* ^^^ Ensure CB invocation follows _head test. */ +return false; } +WARN_ON (signal_pending (current)); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WokeEmpty" )); +return true; }/* @@ -1899,7 +1910,7 @@ static void nocb_follower_wait(struct rcu_data *rdp)* have to do quite so many wakeups (as in they only need to wake the * no-CBs GP kthreads, not the CB kthreads). */ -static int rcu_nocb_kthread (void *arg) +static int rcu_nocb_cb_kthread (void *arg) {int c, cl;unsigned long flags;@@ -1911,10 +1922,8 @@ static int rcu_nocb_kthread(void *arg)/* Each pass through this loop invokes one batch of callbacks */ for (;;) {/* Wait for callbacks. */ -if (rdp->nocb_gp_rdp == rdp) -nocb_leader_wait (rdp); -else -nocb_follower_wait (rdp); +while (nocb_follower_wait (rdp)) +continue ;/* Pull the ready-to-invoke callbacks onto local list. */ raw_spin_lock_irqsave (&rdp->nocb_lock, flags);@@ -2048,7 +2057,8 @@ void __init rcu_init_nohz(void)static void __init rcu_boot_init_nocb_percpu_data (struct rcu_data *rdp) { rdp->nocb_tail = &rdp->nocb_head; -init_swait_queue_head (&rdp->nocb_wq); +init_swait_queue_head (&rdp->nocb_cb_wq); +init_swait_queue_head (&rdp->nocb_gp_wq); rdp->nocb_cb_tail = &rdp->nocb_cb_head;raw_spin_lock_init (&rdp->nocb_lock);timer_setup (&rdp->nocb_timer, do_nocb_deferred_wakeup_timer,0 );@@ -2056,50 +2066,39 @@ static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)/* * If the specified CPU is a no-CBs CPU that does not already have its -* rcuo kthread, spawn it. If the CPUs are brought online out of order, -* this can require re-organizing the GP-CB relationships. +* rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread +* for this CPU's group has not yet been created, spawn it as well. */ static void rcu_spawn_one_nocb_kthread (int cpu) { -struct rcu_data *rdp; -struct rcu_data *rdp_last; -struct rcu_data *rdp_old_leader; -struct rcu_data *rdp_spawn =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp =per_cpu_ptr (&rcu_data, cpu); +struct rcu_data *rdp_gp;struct task_struct *t;/* * If this isn't a no-CBs CPU or if it already has an rcuo kthread, * then nothing to do. */ -if (!rcu_is_nocb_cpu (cpu) || rdp_spawn->nocb_cb_kthread) +if (!rcu_is_nocb_cpu (cpu) || rdp->nocb_cb_kthread)return ;/* If we didn't spawn the GP kthread first, reorganize! */ - rdp_old_leader = rdp_spawn->nocb_gp_rdp; -if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_cb_kthread) { - rdp_last =NULL ; - rdp = rdp_old_leader; -do { - rdp->nocb_gp_rdp = rdp_spawn; -if (rdp_last && rdp != rdp_spawn) - rdp_last->nocb_next_cb_rdp = rdp; -if (rdp == rdp_spawn) { - rdp = rdp->nocb_next_cb_rdp; - }else { - rdp_last = rdp; - rdp = rdp->nocb_next_cb_rdp; - rdp_last->nocb_next_cb_rdp =NULL ; - } - }while (rdp); - rdp_spawn->nocb_next_cb_rdp = rdp_old_leader; + rdp_gp = rdp->nocb_gp_rdp; +if (!rdp_gp->nocb_gp_kthread) { + t =kthread_run (rcu_nocb_gp_kthread, rdp_gp, +"rcuog/%d" , rdp_gp->cpu); +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo GP kthread, OOM is now expected behavior\n" , __func__)) +return ; +WRITE_ONCE (rdp_gp->nocb_gp_kthread, t); }/* Spawn the kthread for this CPU. */ - t =kthread_run (rcu_nocb_kthread, rdp_spawn, + t =kthread_run (rcu_nocb_cb_kthread, rdp,"rcuo%c/%d" , rcu_state.abbr, cpu); -if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo kthread, OOM is now expected behavior\n" , __func__)) +if (WARN_ONCE (IS_ERR (t),"%s: Could not start rcuo CB kthread, OOM is now expected behavior\n" , __func__))return ; -WRITE_ONCE (rdp_spawn->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_cb_kthread, t); +WRITE_ONCE (rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread); }/* Signature · 1 line
2.17.1
PATCH 4/11 rcu/nocb: Rename nocb_follower_wait() to nocb_cb_wait()
In reply to Paul E. McKenneyThis commit adjusts naming to account for the new distinction between callback and grace-period no-CBs kthreads.
--- kernel/rcu/tree_plugin.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/rcu/tree_plugin.h +2 -2
@@ -1889,7 +1889,7 @@ static int rcu_nocb_gp_kthread(void *arg)* This function returns true ("keep waiting") until callbacks appear and * then false ("stop waiting") when callbacks finally do appear. */ -static bool nocb_follower_wait (struct rcu_data *rdp) +static bool nocb_cb_wait (struct rcu_data *rdp) {trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
@@ -1922,7 +1922,7 @@ static int rcu_nocb_cb_kthread(void *arg)/* Each pass through this loop invokes one batch of callbacks */ for (;;) {/* Wait for callbacks. */ -while (nocb_follower_wait (rdp)) +while (nocb_cb_wait (rdp))continue ;/* Pull the ready-to-invoke callbacks onto local list. */
Signature · 1 line
2.17.1
PATCH 5/11 rcu/nocb: Rename wake_nocb_leader() to wake_nocb_gp()
In reply to Paul E. McKenneyThis commit adjusts naming to account for the new distinction between callback and grace-period no-CBs kthreads.
--- kernel/rcu/tree_plugin.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
kernel/rcu/tree_plugin.h +3 -3
@@ -1551,7 +1551,7 @@ static void __wake_nocb_leader(struct rcu_data *rdp, bool force,* Kick the GP kthread for this NOCB group, but caller has not * acquired locks. */ -static void wake_nocb_leader (struct rcu_data *rdp,bool force) +static void wake_nocb_gp (struct rcu_data *rdp,bool force) {unsigned long flags;
@@ -1656,7 +1656,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,if (old_rhpp == &rdp->nocb_head) {if (!irqs_disabled_flags (flags)) {/* ... if queue was empty ... */ -wake_nocb_leader (rdp, false); +wake_nocb_gp (rdp, false);trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeEmpty" )); }else {
@@ -1667,7 +1667,7 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, }else if (len > rdp->qlen_last_fqs_check + qhimark) {/* ... or if many callbacks queued. */ if (!irqs_disabled_flags (flags)) { -wake_nocb_leader (rdp, true); +wake_nocb_gp (rdp, true);trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeOvf" )); }else {
Signature · 1 line
2.17.1
PATCH 6/11 rcu/nocb: Rename __wake_nocb_leader() to __wake_nocb_gp()
In reply to Paul E. McKenneyThis commit adjusts naming to account for the new distinction between callback and grace-period no-CBs kthreads. While in the area, it also updates local variables.
--- kernel/rcu/tree_plugin.h | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-)
kernel/rcu/tree_plugin.h +9 -9
@@ -1524,24 +1524,24 @@ bool rcu_is_nocb_cpu(int cpu)* Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock * and this function releases it. */ -static void __wake_nocb_leader (struct rcu_data *rdp,bool force, -unsigned long flags) +static void __wake_nocb_gp (struct rcu_data *rdp,bool force, +unsigned long flags) __releases(rdp->nocb_lock) { -struct rcu_data *rdp_leader = rdp->nocb_gp_rdp; +struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;lockdep_assert_held (&rdp->nocb_lock); -if (!READ_ONCE (rdp_leader->nocb_gp_kthread)) { +if (!READ_ONCE (rdp_gp->nocb_gp_kthread)) {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);return ; } -if (rdp_leader->nocb_gp_sleep || force) { +if (rdp_gp->nocb_gp_sleep || force) {/* Prior smp_mb__after_atomic() orders against prior enqueue. */ -WRITE_ONCE (rdp_leader->nocb_gp_sleep, false); +WRITE_ONCE (rdp_gp->nocb_gp_sleep, false);del_timer (&rdp->nocb_timer);raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags);smp_mb ();/* ->nocb_gp_sleep before swake_up_one(). */ -swake_up_one (&rdp_leader->nocb_gp_wq); +swake_up_one (&rdp_gp->nocb_gp_wq); }else {raw_spin_unlock_irqrestore (&rdp->nocb_lock, flags); }
@@ -1556,7 +1556,7 @@ static void wake_nocb_gp(struct rcu_data *rdp, bool force)unsigned long flags;raw_spin_lock_irqsave (&rdp->nocb_lock, flags); -__wake_nocb_leader (rdp, force, flags); +__wake_nocb_gp (rdp, force, flags); }/*
@@ -1988,7 +1988,7 @@ static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp) } ndw =READ_ONCE (rdp->nocb_defer_wakeup);WRITE_ONCE (rdp->nocb_defer_wakeup,RCU_NOCB_WAKE_NOT ); -__wake_nocb_leader (rdp, ndw ==RCU_NOCB_WAKE_FORCE , flags); +__wake_nocb_gp (rdp, ndw ==RCU_NOCB_WAKE_FORCE , flags);trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("DeferredWake" )); }
Signature · 1 line
2.17.1
PATCH 7/11 rcu/nocb: Rename wake_nocb_leader_defer() to wake_nocb_gp_defer()
In reply to Paul E. McKenneyThis commit adjusts naming to account for the new distinction between callback and grace-period no-CBs kthreads.
--- kernel/rcu/tree_plugin.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-)
kernel/rcu/tree_plugin.h +6 -6
@@ -1563,8 +1563,8 @@ static void wake_nocb_gp(struct rcu_data *rdp, bool force)* Arrange to wake the GP kthread for this NOCB group at some future * time when it is safe to do so. */ -static void wake_nocb_leader_defer (struct rcu_data *rdp,int waketype, -const char *reason) +static void wake_nocb_gp_defer (struct rcu_data *rdp,int waketype, +const char *reason) {unsigned long flags;
@@ -1660,8 +1660,8 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeEmpty" )); }else { -wake_nocb_leader_defer (rdp,RCU_NOCB_WAKE , -TPS ("WakeEmptyIsDeferred" )); +wake_nocb_gp_defer (rdp,RCU_NOCB_WAKE , +TPS ("WakeEmptyIsDeferred" )); } rdp->qlen_last_fqs_check =0 ; }else if (len > rdp->qlen_last_fqs_check + qhimark) {
@@ -1671,8 +1671,8 @@ static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("WakeOvf" )); }else { -wake_nocb_leader_defer (rdp,RCU_NOCB_WAKE_FORCE , -TPS ("WakeOvfIsDeferred" )); +wake_nocb_gp_defer (rdp,RCU_NOCB_WAKE_FORCE , +TPS ("WakeOvfIsDeferred" )); } rdp->qlen_last_fqs_check =LONG_MAX /2 ; }else {
Signature · 1 line
2.17.1
PATCH 8/11 rcu/nocb: Rename rcu_organize_nocb_kthreads() local variable
In reply to Paul E. McKenneyThis commit renames rdp_leader to rdp_gp in order to account for the new distinction between callback and grace-period no-CBs kthreads.
--- kernel/rcu/tree_plugin.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
kernel/rcu/tree_plugin.h +3 -3
@@ -2138,7 +2138,7 @@ static void __init rcu_organize_nocb_kthreads(void)int ls = rcu_nocb_leader_stride;int nl =0 ;/* Next GP kthread. */ struct rcu_data *rdp; -struct rcu_data *rdp_leader =NULL ;/* Suppress misguided gcc warn. */ +struct rcu_data *rdp_gp =NULL ;/* Suppress misguided gcc warn. */ struct rcu_data *rdp_prev =NULL ;if (!cpumask_available (rcu_nocb_mask))
@@ -2159,10 +2159,10 @@ static void __init rcu_organize_nocb_kthreads(void)/* New GP kthread, set up for CBs & next GP. */ nl =DIV_ROUND_UP (rdp->cpu +1 , ls) * ls; rdp->nocb_gp_rdp = rdp; - rdp_leader = rdp; + rdp_gp = rdp; }else {/* Another CB kthread, link to previous GP kthread. */ - rdp->nocb_gp_rdp = rdp_leader; + rdp->nocb_gp_rdp = rdp_gp; rdp_prev->nocb_next_cb_rdp = rdp; } rdp_prev = rdp;
Signature · 1 line
2.17.1
PATCH 9/11 rcu/nocb: Rename and document no-CB CB kthread sleep trace event
In reply to Paul E. McKenneyThe nocb_cb_wait() function traces a "FollowerSleep" trace_rcu_nocb_wake() event, which never was documented and is now misleading. This commit therefore changes "FollowerSleep" to "CBSleep", documents this, and updates the documentation for "Sleep" as well.
--- include/trace/events/rcu.h | 3 ++- kernel/rcu/tree_plugin.h | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-)
include/trace/events/rcu.h +2 -1
@@ -267,7 +267,8 @@ TRACE_EVENT_RCU(rcu_exp_funnel_lock,* "WakeNotPoll": Don't wake rcuo kthread because it is polling. * "DeferredWake": Carried out the "IsDeferred" wakeup. * "Poll": Start of new polling cycle for rcu_nocb_poll. -* "Sleep": Sleep waiting for CBs for !rcu_nocb_poll. +* "Sleep": Sleep waiting for GP for !rcu_nocb_poll. +* "CBSleep": Sleep waiting for CBs for !rcu_nocb_poll. * "WokeEmpty": rcuo kthread woke to find empty list. * "WokeNonEmpty": rcuo kthread woke to find non-empty list. * "WaitQueue": Enqueue partially done, timed wait for it to complete.
kernel/rcu/tree_plugin.h +1 -1
@@ -1891,7 +1891,7 @@ static int rcu_nocb_gp_kthread(void *arg)*/ static bool nocb_cb_wait (struct rcu_data *rdp) { -trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("FollowerSleep" )); +trace_rcu_nocb_wake (rcu_state.name, rdp->cpu,TPS ("CBSleep" ));swait_event_interruptible_exclusive (rdp->nocb_cb_wq,READ_ONCE (rdp->nocb_cb_head));if (smp_load_acquire (&rdp->nocb_cb_head)) {/* VVV */
Signature · 1 line
2.17.1
PATCH 10/11 rcu/nocb: Rename rcu_nocb_leader_stride kernel boot parameter
In reply to Paul E. McKenneyThis commit changes the name of the rcu_nocb_leader_stride kernel boot parameter to rcu_nocb_gp_stride in order to account for the new distinction between callback and grace-period no-CBs kthreads.
--- Documentation/admin-guide/kernel-parameters.txt | 13 +++++++------ kernel/rcu/tree_plugin.h | 8 ++++---- 2 files changed, 11 insertions(+), 10 deletions(-)
Documentation/admin-guide/kernel-parameters.txt +7 -6
@@ -3837,12 +3837,13 @@ RCU_BOOST is not set, valid values are 0-99 and the default is zero (non-realtime operation). - rcutree.rcu_nocb_leader_stride= [KNL] - Set the number of NOCB kthread groups, which - defaults to the square root of the number of - CPUs. Larger numbers reduces the wakeup overhead - on the per-CPU grace-period kthreads, but increases - that same overhead on each group's leader. + rcutree.rcu_nocb_gp_stride= [KNL] + Set the number of NOCB callback kthreads in + each group, which defaults to the square root + of the number of CPUs. Larger numbers reduce + the wakeup overhead on the global grace-period + kthread, but increases that same overhead on + each group's NOCB grace-period kthread. rcutree.qhimark= [KNL] Set threshold of queued RCU callbacks beyond which
kernel/rcu/tree_plugin.h +4 -4
@@ -2126,8 +2126,8 @@ static void __init rcu_spawn_nocb_kthreads(void) }/* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */ -static int rcu_nocb_leader_stride =-1 ; -module_param (rcu_nocb_leader_stride, int,0444 ); +static int rcu_nocb_gp_stride =-1 ; +module_param (rcu_nocb_gp_stride, int,0444 );/* * Initialize GP-CB relationships for all no-CBs CPU.
@@ -2135,7 +2135,7 @@ module_param(rcu_nocb_leader_stride, int, 0444);static void __init rcu_organize_nocb_kthreads (void ) {int cpu; -int ls = rcu_nocb_leader_stride; +int ls = rcu_nocb_gp_stride;int nl =0 ;/* Next GP kthread. */ struct rcu_data *rdp;struct rcu_data *rdp_gp =NULL ;/* Suppress misguided gcc warn. */
@@ -2145,7 +2145,7 @@ static void __init rcu_organize_nocb_kthreads(void)return ;if (ls ==-1 ) { ls =int_sqrt (nr_cpu_ids); - rcu_nocb_leader_stride = ls; + rcu_nocb_gp_stride = ls; }/*
Signature · 1 line
2.17.1
PATCH 11/11 rcu/nocb: Print gp/cb kthread hierarchy if dump_tree
In reply to Paul E. McKenneyThis commit causes the no-CBs grace-period/callback hierarchy to be printed to the console when the dump_tree kernel boot parameter is set.
--- kernel/rcu/tree_plugin.h | 6 ++++++ 1 file changed, 6 insertions(+)
kernel/rcu/tree_plugin.h +6 -0
@@ -2135,6 +2135,7 @@ module_param(rcu_nocb_gp_stride, int, 0444);static void __init rcu_organize_nocb_kthreads (void ) {int cpu; +bool firsttime = true;int ls = rcu_nocb_gp_stride;int nl =0 ;/* Next GP kthread. */ struct rcu_data *rdp;
@@ -2160,10 +2161,15 @@ static void __init rcu_organize_nocb_kthreads(void) nl =DIV_ROUND_UP (rdp->cpu +1 , ls) * ls; rdp->nocb_gp_rdp = rdp; rdp_gp = rdp; +if (!firsttime && dump_tree) +pr_cont ("\n" ); + firsttime = false; +pr_alert ("%s: No-CB GP kthread CPU %d:" , __func__, cpu); }else {/* Another CB kthread, link to previous GP kthread. */ rdp->nocb_gp_rdp = rdp_gp; rdp_prev->nocb_next_cb_rdp = rdp; +pr_alert (" %d" , cpu); } rdp_prev = rdp; }
Signature · 1 line
2.17.1