Donor Migration for Proxy Execution (v21)

Hey All,

I didn't get any feedback on the last iteration, so I wanted to resend this next chunk of the series: Donor Migration

The main change from v20 before it, is that I previously had logic where the ww_mutex paths took the blocked_lock of the task it was waking (either the lock waiter->task or owner), but in a context from __mutex_lock_common() where we already held the current->block_lock. This required using the spin_lock_nested() annotation to keep lockdep happy, and I was leaning on the logic that there is an implied order between running current and the existing not-running lock waiters, which should avoid loops. In the wound case, there is also an order used if the owners context is younger, which sounded likely to avoid loops.

However, after thinking more about the wound case where we are wounding a lock owner, since that owner is not waiting and could be trying to acquire a mutex current owns, I couldn’t quite convince myself we couldn’t get into a ABBA style deadlock with the nested blocked_lock accesses (though, I’ve not been able to contrive it to happen, but that doesn’t prove anything).

So the main difference in v21 is reworking of how we hold the blocked_lock in the mutex_lock_common() code, reducing it so we don’t call into ww_mutex paths while holding it. The lock->waiter_lock still serializes things at top level, but the blocked_lock isn’t held completely in parallel under that, and is focused on its purpose of protecting the blocked_on, blocked_on_state and similar proxy-related values in the task struct.

I also did some cleanups to be more consistent in how the blocked_on_state is handled. I had a few spots previously where I was cheating and just set the value instead of going through the helpers. And sure enough, in fixing those I realized there were a few spots where I wasn’t always holding the right blocked_lock, so some minor rework helped clean that up.

I’m trying to submit this larger work in smallish digestible pieces, so in this portion of the series, I’m only submitting for review and consideration the logic that allows us to do donor(blocked waiter) migration, allowing us to proxy-execute lock owners that might be on other cpu runqueues. This requires some additional changes to locking and extra state tracking to ensure we don’t accidentally run a migrated donor on a cpu it isn’t affined to, as well as some extra handling to deal with balance callback state that needs to be reset when we decide to pick a different task after doing donor migration.

I’d love to get some feedback on any place where these patches are confusing, or could use additional clarifications.

Also you can find the full proxy-exec series here:
https://github.com/johnstultz-work/linux-dev/commits/proxy-exec-v21-6.17-rc4/
https://github.com/johnstultz-work/linux-dev.git proxy-exec-v21-6.17-rc4

Issues still to address with the full series:
* Need to sort out what is needed for sched_ext to be ok with proxy-execution enabled. This is my next priority.

* K Prateek Nayak did some testing about a bit over a year ago with an earlier version of the full series and saw ~3-5% regressions in some cases. Need to re-evaluate this with the proxy-migration avoidance optimization Suleiman suggested having now been implemented.

* The chain migration functionality needs further iterations and better validation to ensure it truly maintains the RT/DL load balancing invariants (despite this being broken in vanilla upstream with RT_PUSH_IPI currently)

Future work:
* Expand to other locking primitives: Suleiman is looking at rw_semaphores, as that is another common source of priority inversion. Figuring out pi-futexes would be good too.
* Eventually: Work to replace rt_mutexes and get things happy with PREEMPT_RT

I’d really appreciate any feedback or review thoughts on the full series as well. I’m trying to keep the chunks small, reviewable and iteratively testable, but if you have any suggestions on how to improve the series, I’m all ears.

Credit/Disclaimer:
—--------------------
As always, this Proxy Execution series has a long history with lots of developers that deserve credit:

First described in a paper[1] by Watkins, Straub, Niehaus, then from patches from Peter Zijlstra, extended with lots of work by Juri Lelli, Valentin Schneider, and Connor O'Brien. (and thank you to Steven Rostedt for providing additional details here!)

So again, many thanks to those above, as all the credit for this series really is due to them - while the mistakes are likely mine.

Thanks so much!
-john

[1] https://static.lwn.net/images/conf/rtlws11/papers/proc/p38.pdf

John Stultz (5):
locking: Add task::blocked_lock to serialize blocked_on state sched/locking: Add blocked_on_state to provide necessary tri-state for
proxy return-migration
sched: Add logic to zap balance callbacks if we pick again
sched: Handle blocked-waiter migration (and return migration)
sched: Migrate whole chain in proxy_migrate_task()

Peter Zijlstra (1):
sched: Add blocked_donor link to task for smarter mutex handoffs

 include/linux/sched.h        | 120 ++++++++-----
 init/init_task.c             |   4 +
 kernel/fork.c                |   4 +
 kernel/locking/mutex-debug.c |   4 +-
 kernel/locking/mutex.c       |  83 +++++++--
 kernel/locking/ww_mutex.h    |  20 +--
 kernel/sched/core.c          | 329 +++++++++++++++++++++++++++++++++--
 kernel/sched/fair.c          |   3 +-
 kernel/sched/sched.h         |   2 +-
 9 files changed, 473 insertions(+), 96 deletions(-)

2.51.0.338.gd7d06c2dae-goog

So far, we have been able to utilize the mutex::wait_lock for serializing the blocked_on state, but when we move to proxying across runqueues, we will need to add more state and a way to serialize changes to this state in contexts where we don't hold the mutex::wait_lock.

So introduce the task::blocked_lock, which nests under the mutex::wait_lock in the locking order, and rework the locking to use it.

---
v15:
* Split back out into later in the series v16:
* Fixups to mark tasks unblocked before sleeping in mutex_optimistic_spin()
* Rework to use guard() as suggested by Peter v19:
* Rework logic for PREEMPT_RT issues reported by K Prateek Nayak
v21:
* After recently thinking more on ww_mutex code, I reworked the blocked_lock usage in mutex lock to avoid having to take nested locks in the ww_mutex paths, as I was concerned the lock ordering constraints weren't as strong as I had previously thought.

---
 include/linux/sched.h        | 52 +++++++++++++++---------------------
 init/init_task.c             |  1 +
 kernel/fork.c                |  1 +
 kernel/locking/mutex-debug.c |  4 +--
 kernel/locking/mutex.c       | 37 +++++++++++++++----------
 kernel/locking/ww_mutex.h    |  4 +--
 kernel/sched/core.c          |  4 ++-
 7 files changed, 54 insertions(+), 49 deletions(-)

index f8188b8333503..3ec0ef0d91603 100644
@@ -1235,6 +1235,7 @@ struct task_struct {
 #endif
 
 	struct mutex			*blocked_on;	/* lock we're blocked on */
+	raw_spinlock_t			blocked_lock;
 
 #ifdef CONFIG_DETECT_HUNG_TASK_BLOCKER
 	/*

@@ -2143,57 +2144,48 @@ extern int __cond_resched_rwlock_write(rwlock_t *lock);
 #ifndef CONFIG_PREEMPT_RT
 static inline struct mutex *__get_task_blocked_on(struct task_struct *p)
 {
-	struct mutex *m = p->blocked_on;
+	lockdep_assert_held_once(&p->blocked_lock);
+	return p->blocked_on;
+}
 
-	if (m)
-		lockdep_assert_held_once(&m->wait_lock);
-	return m;
+static inline struct mutex *get_task_blocked_on(struct task_struct *p)
+{
+	guard(raw_spinlock_irqsave)(&p->blocked_lock);
+	return __get_task_blocked_on(p);
 }
 
 static inline void __set_task_blocked_on(struct task_struct *p, struct mutex *m)
 {
-	struct mutex *blocked_on = READ_ONCE(p->blocked_on);
-
 	WARN_ON_ONCE(!m);
 	/* The task should only be setting itself as blocked */
 	WARN_ON_ONCE(p != current);
-	/* Currently we serialize blocked_on under the mutex::wait_lock */
-	lockdep_assert_held_once(&m->wait_lock);
+	/* Currently we serialize blocked_on under the task::blocked_lock */
+	lockdep_assert_held_once(&p->blocked_lock);
 	/*
 	 * Check ensure we don't overwrite existing mutex value
 	 * with a different mutex. Note, setting it to the same
 	 * lock repeatedly is ok.
 	 */
-	WARN_ON_ONCE(blocked_on && blocked_on != m);
-	WRITE_ONCE(p->blocked_on, m);
-}
-
-static inline void set_task_blocked_on(struct task_struct *p, struct mutex *m)
-{
-	guard(raw_spinlock_irqsave)(&m->wait_lock);
-	__set_task_blocked_on(p, m);
+	WARN_ON_ONCE(p->blocked_on && p->blocked_on != m);
+	p->blocked_on = m;
 }
 
 static inline void __clear_task_blocked_on(struct task_struct *p, struct mutex *m)
 {
-	if (m) {
-		struct mutex *blocked_on = READ_ONCE(p->blocked_on);
-
-		/* Currently we serialize blocked_on under the mutex::wait_lock */
-		lockdep_assert_held_once(&m->wait_lock);
-		/*
-		 * There may be cases where we re-clear already cleared
-		 * blocked_on relationships, but make sure we are not
-		 * clearing the relationship with a different lock.
-		 */
-		WARN_ON_ONCE(blocked_on && blocked_on != m);
-	}
-	WRITE_ONCE(p->blocked_on, NULL);
+	/* Currently we serialize blocked_on under the task::blocked_lock */
+	lockdep_assert_held_once(&p->blocked_lock);
+	/*
+	 * There may be cases where we re-clear already cleared
+	 * blocked_on relationships, but make sure we are not
+	 * clearing the relationship with a different lock.
+	 */
+	WARN_ON_ONCE(m && p->blocked_on && p->blocked_on != m);
+	p->blocked_on = NULL;
 }
 
 static inline void clear_task_blocked_on(struct task_struct *p, struct mutex *m)
 {
-	guard(raw_spinlock_irqsave)(&m->wait_lock);
+	guard(raw_spinlock_irqsave)(&p->blocked_lock);
 	__clear_task_blocked_on(p, m);
 }
 #else

index e557f622bd906..7e29d86153d9f 100644
@@ -140,6 +140,7 @@ struct task_struct init_task __aligned(L1_CACHE_BYTES) = {
 	.journal_info	= NULL,
 	INIT_CPU_TIMERS(init_task)
 	.pi_lock	= __RAW_SPIN_LOCK_UNLOCKED(init_task.pi_lock),
+	.blocked_lock	= __RAW_SPIN_LOCK_UNLOCKED(init_task.blocked_lock),
 	.timer_slack_ns = 50000, /* 50 usec default slack */
 	.thread_pid	= &init_struct_pid,
 	.thread_node	= LIST_HEAD_INIT(init_signals.thread_head),

index af673856499dc..db6d08946ec11 100644
@@ -2030,6 +2030,7 @@ __latent_entropy struct task_struct *copy_process(
 	ftrace_graph_init_task(p);
 
 	rt_mutex_init_task(p);
+	raw_spin_lock_init(&p->blocked_lock);
 
 	lockdep_assert_irqs_enabled();
 #ifdef CONFIG_PROVE_LOCKING

index 949103fd8e9b5..1d8cff71f65e1 100644
@@ -54,13 +54,13 @@ void debug_mutex_add_waiter(struct mutex *lock, struct mutex_waiter *waiter,
 	lockdep_assert_held(&lock->wait_lock);
 
 	/* Current thread can't be already blocked (since it's executing!) */
-	DEBUG_LOCKS_WARN_ON(__get_task_blocked_on(task));
+	DEBUG_LOCKS_WARN_ON(get_task_blocked_on(task));
 }
 
 void debug_mutex_remove_waiter(struct mutex *lock, struct mutex_waiter *waiter,
 			 struct task_struct *task)
 {
-	struct mutex *blocked_on = __get_task_blocked_on(task);
+	struct mutex *blocked_on = get_task_blocked_on(task);
 
 	DEBUG_LOCKS_WARN_ON(list_empty(&waiter->list));
 	DEBUG_LOCKS_WARN_ON(waiter->task != task);

index de7d6702cd96c..fac40c456098e 100644
@@ -640,6 +640,7 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 			goto err_early_kill;
 	}
 
+	raw_spin_lock(&current->blocked_lock);
 	__set_task_blocked_on(current, lock);
 	set_current_state(state);
 	trace_contention_begin(lock, LCB_F_MUTEX);

@@ -653,8 +654,9 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 		 * the handoff.
 		 */
 		if (__mutex_trylock(lock))
-			goto acquired;
+			break;
 
+		raw_spin_unlock(&current->blocked_lock);
 		/*
 		 * Check for signals and kill conditions while holding
 		 * wait_lock. This ensures the lock cancellation is ordered

@@ -677,12 +679,14 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 
 		first = __mutex_waiter_is_first(lock, &waiter);
 
+		raw_spin_lock_irqsave(&lock->wait_lock, flags);
+		raw_spin_lock(&current->blocked_lock);
 		/*
 		 * As we likely have been woken up by task
 		 * that has cleared our blocked_on state, re-set
 		 * it to the lock we are trying to acquire.
 		 */
-		set_task_blocked_on(current, lock);
+		__set_task_blocked_on(current, lock);
 		set_current_state(state);
 		/*
 		 * Here we order against unlock; we must either see it change

@@ -693,25 +697,30 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 			break;
 
 		if (first) {
-			trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
+			bool opt_acquired;
+
 			/*
 			 * mutex_optimistic_spin() can call schedule(), so
-			 * clear blocked on so we don't become unselectable
+			 * we need to release these locks before calling it,
+			 * and clear blocked on so we don't become unselectable
 			 * to run.
 			 */
-			clear_task_blocked_on(current, lock);
-			if (mutex_optimistic_spin(lock, ww_ctx, &waiter))
+			__clear_task_blocked_on(current, lock);
+			raw_spin_unlock(&current->blocked_lock);
+			raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
+			trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
+			opt_acquired = mutex_optimistic_spin(lock, ww_ctx, &waiter);
+			raw_spin_lock_irqsave(&lock->wait_lock, flags);
+			raw_spin_lock(&current->blocked_lock);
+			__set_task_blocked_on(current, lock);
+			if (opt_acquired)
 				break;
-			set_task_blocked_on(current, lock);
 			trace_contention_begin(lock, LCB_F_MUTEX);
 		}
-
-		raw_spin_lock_irqsave(&lock->wait_lock, flags);
 	}
-	raw_spin_lock_irqsave(&lock->wait_lock, flags);
-acquired:
 	__clear_task_blocked_on(current, lock);
 	__set_current_state(TASK_RUNNING);
+	raw_spin_unlock(&current->blocked_lock);
 
 	if (ww_ctx) {
 		/*

@@ -740,11 +749,11 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 	return 0;
 
 err:
-	__clear_task_blocked_on(current, lock);
+	clear_task_blocked_on(current, lock);
 	__set_current_state(TASK_RUNNING);
 	__mutex_remove_waiter(lock, &waiter);
 err_early_kill:
-	WARN_ON(__get_task_blocked_on(current));
+	WARN_ON(get_task_blocked_on(current));
 	trace_contention_end(lock, ret);
 	raw_spin_unlock_irqrestore_wake(&lock->wait_lock, flags, &wake_q);
 	debug_mutex_free_waiter(&waiter);

@@ -955,7 +964,7 @@ static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigne
 		next = waiter->task;
 
 		debug_mutex_wake_waiter(lock, waiter);
-		__clear_task_blocked_on(next, lock);
+		clear_task_blocked_on(next, lock);
 		wake_q_add(&wake_q, next);
 	}
 

index 31a785afee6c0..e4a81790ea7dd 100644
@@ -289,7 +289,7 @@ __ww_mutex_die(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
 		 * blocked_on pointer. Otherwise we can see circular
 		 * blocked_on relationships that can't resolve.
 		 */
-		__clear_task_blocked_on(waiter->task, lock);
+		clear_task_blocked_on(waiter->task, lock);
 		wake_q_add(wake_q, waiter->task);
 	}
 

@@ -347,7 +347,7 @@ static bool __ww_mutex_wound(struct MUTEX *lock,
 			 * are waking the mutex owner, who may be currently
 			 * blocked on a different mutex.
 			 */
-			__clear_task_blocked_on(owner, NULL);
+			clear_task_blocked_on(owner, NULL);
 			wake_q_add(wake_q, owner);
 		}
 		return true;

index be00629f0ba4c..0180853dd48c5 100644
@@ -6639,6 +6639,7 @@ static struct task_struct *proxy_deactivate(struct rq *rq, struct task_struct *d
  *   p->pi_lock
  *     rq->lock
  *       mutex->wait_lock
+ *         p->blocked_lock
  *
  * Returns the task that is going to be used as execution context (the one
  * that is actually going to be run on cpu_of(rq)).

@@ -6662,8 +6663,9 @@ find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 		 * and ensure @owner sticks around.
 		 */
 		guard(raw_spinlock)(&mutex->wait_lock);
+		guard(raw_spinlock)(&p->blocked_lock);
 
-		/* Check again that p is blocked with wait_lock held */
+		/* Check again that p is blocked with blocked_lock held */
 		if (mutex != __get_task_blocked_on(p)) {
 			/*
 			 * Something changed in the blocked_on chain and

2.51.0.338.gd7d06c2dae-goog

Hello John,

Thanks and Regards,
Prateek

So, basically yes, it's for symmetry, and trying to avoid subtlety where I've found it's easy to confuse myself.

At this point in the patch series, as we don't prevent double-clears, I guess it could be moved below the break, without much impact. However, after the "Add blocked_on_state" patch, we do enforce that we we don't clear cleared values, and we manage the blocked_on_state separately here, so at that point we'd have to put it back or loosen the constraint checking. So I think leaving it here seems the most sensible.

Sure! Done!

Thanks so much again for your review and feedback! -john

As we add functionality to proxy execution, we may migrate a donor task to a runqueue where it can't run due to cpu affinity. Thus, we must be careful to ensure we return-migrate the task back to a cpu in its cpumask when it becomes unblocked.

Thus we need more then just a binary concept of the task being blocked on a mutex or not.

So add a blocked_on_state value to the task, that allows the task to move through BO_RUNNING -> BO_BLOCKED -> BO_WAKING and back to BO_RUNNING. This provides a guard state in BO_WAKING so we can know the task is no longer blocked but we don't want to run it until we have potentially done return migration, back to a usable cpu.

---
v15:
* Split blocked_on_state into its own patch later in the series, as the tri-state isn't necessary until we deal with proxy/return migrations
v16:
* Handle case where task in the chain is being set as BO_WAKING by another cpu (usually via ww_mutex die code). Make sure we release the rq lock so the wakeup can complete.
* Rework to use guard() in find_proxy_task() as suggested by Peter
v18:
* Add initialization of blocked_on_state for init_task v19:
* PREEMPT_RT build fixups and rework suggested by
K Prateek Nayak
v20:
* Simplify one of the blocked_on_state changes to avoid extra PREMEPT_RT conditionals
v21:
* Slight reworks due to avoiding nested blocked_lock locking
* Be consistent in use of blocked_on_state helper functions
* Rework calls to proxy_deactivate() to do proper locking around blocked_on_state changes that we were cheating in previous versions.
* Minor cleanups, some comment improvements

---
 include/linux/sched.h     | 80 +++++++++++++++++++++++++++++----------
 init/init_task.c          |  1 +
 kernel/fork.c             |  1 +
 kernel/locking/mutex.c    | 15 ++++----
 kernel/locking/ww_mutex.h | 20 ++++------
 kernel/sched/core.c       | 44 +++++++++++++++++++--
 kernel/sched/sched.h      |  2 +-
 7 files changed, 120 insertions(+), 43 deletions(-)

index 3ec0ef0d91603..5801de1a44a79 100644
@@ -815,6 +815,12 @@ struct kmap_ctrl {
 #endif
 };
 
+enum blocked_on_state {
+	BO_RUNNABLE,
+	BO_BLOCKED,
+	BO_WAKING,
+};
+
 struct task_struct {
 #ifdef CONFIG_THREAD_INFO_IN_TASK
 	/*

@@ -1234,6 +1240,7 @@ struct task_struct {
 	struct rt_mutex_waiter		*pi_blocked_on;
 #endif
 
+	enum blocked_on_state		blocked_on_state;
 	struct mutex			*blocked_on;	/* lock we're blocked on */
 	raw_spinlock_t			blocked_lock;
 

@@ -2141,7 +2148,52 @@ extern int __cond_resched_rwlock_write(rwlock_t *lock);
 	__cond_resched_rwlock_write(lock);					\
 })
 
-#ifndef CONFIG_PREEMPT_RT
+static inline void __force_blocked_on_runnable(struct task_struct *p)
+{
+	lockdep_assert_held(&p->blocked_lock);
+	p->blocked_on_state = BO_RUNNABLE;
+}
+
+static inline void force_blocked_on_runnable(struct task_struct *p)
+{
+	guard(raw_spinlock_irqsave)(&p->blocked_lock);
+	__force_blocked_on_runnable(p);
+}
+
+static inline void __force_blocked_on_blocked(struct task_struct *p)
+{
+	lockdep_assert_held(&p->blocked_lock);
+	p->blocked_on_state = BO_BLOCKED;
+}
+
+static inline void __set_blocked_on_runnable(struct task_struct *p)
+{
+	lockdep_assert_held(&p->blocked_lock);
+	if (p->blocked_on_state == BO_WAKING)
+		p->blocked_on_state = BO_RUNNABLE;
+}
+
+static inline void set_blocked_on_runnable(struct task_struct *p)
+{
+	if (!sched_proxy_exec())
+		return;
+	guard(raw_spinlock_irqsave)(&p->blocked_lock);
+	__set_blocked_on_runnable(p);
+}
+
+static inline void __set_blocked_on_waking(struct task_struct *p)
+{
+	lockdep_assert_held(&p->blocked_lock);
+	if (p->blocked_on_state == BO_BLOCKED)
+		p->blocked_on_state = BO_WAKING;
+}
+
+static inline void set_blocked_on_waking(struct task_struct *p)
+{
+	guard(raw_spinlock_irqsave)(&p->blocked_lock);
+	__set_blocked_on_waking(p);
+}
+
 static inline struct mutex *__get_task_blocked_on(struct task_struct *p)
 {
 	lockdep_assert_held_once(&p->blocked_lock);

@@ -2163,24 +2215,23 @@ static inline void __set_task_blocked_on(struct task_struct *p, struct mutex *m)
 	lockdep_assert_held_once(&p->blocked_lock);
 	/*
 	 * Check ensure we don't overwrite existing mutex value
-	 * with a different mutex. Note, setting it to the same
-	 * lock repeatedly is ok.
+	 * with a different mutex.
 	 */
-	WARN_ON_ONCE(p->blocked_on && p->blocked_on != m);
+	WARN_ON_ONCE(p->blocked_on);
 	p->blocked_on = m;
+	p->blocked_on_state = BO_BLOCKED;
 }
 
 static inline void __clear_task_blocked_on(struct task_struct *p, struct mutex *m)
 {
+	/* The task should only be clearing itself */
+	WARN_ON_ONCE(p != current);
 	/* Currently we serialize blocked_on under the task::blocked_lock */
 	lockdep_assert_held_once(&p->blocked_lock);
-	/*
-	 * There may be cases where we re-clear already cleared
-	 * blocked_on relationships, but make sure we are not
-	 * clearing the relationship with a different lock.
-	 */
-	WARN_ON_ONCE(m && p->blocked_on && p->blocked_on != m);
+	/* Make sure we are clearing the relationship with the right lock */
+	WARN_ON_ONCE(m && p->blocked_on != m);
 	p->blocked_on = NULL;
+	p->blocked_on_state = BO_RUNNABLE;
 }
 
 static inline void clear_task_blocked_on(struct task_struct *p, struct mutex *m)

@@ -2188,15 +2239,6 @@ static inline void clear_task_blocked_on(struct task_struct *p, struct mutex *m)
 	guard(raw_spinlock_irqsave)(&p->blocked_lock);
 	__clear_task_blocked_on(p, m);
 }
-#else
-static inline void __clear_task_blocked_on(struct task_struct *p, struct rt_mutex *m)
-{
-}
-
-static inline void clear_task_blocked_on(struct task_struct *p, struct rt_mutex *m)
-{
-}
-#endif /* !CONFIG_PREEMPT_RT */
 
 static __always_inline bool need_resched(void)
 {

index 7e29d86153d9f..6d72ec23410a6 100644
@@ -174,6 +174,7 @@ struct task_struct init_task __aligned(L1_CACHE_BYTES) = {
 	.mems_allowed_seq = SEQCNT_SPINLOCK_ZERO(init_task.mems_allowed_seq,
 						 &init_task.alloc_lock),
 #endif
+	.blocked_on_state = BO_RUNNABLE,
 #ifdef CONFIG_RT_MUTEXES
 	.pi_waiters	= RB_ROOT_CACHED,
 	.pi_top_task	= NULL,

index db6d08946ec11..4bd0731995e86 100644
@@ -2129,6 +2129,7 @@ __latent_entropy struct task_struct *copy_process(
 	lockdep_init_task(p);
 #endif
 
+	p->blocked_on_state = BO_RUNNABLE;
 	p->blocked_on = NULL; /* not blocked yet */
 
 #ifdef CONFIG_BCACHE

index fac40c456098e..42e4d2e6e4ad4 100644
@@ -682,11 +682,9 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 		raw_spin_lock_irqsave(&lock->wait_lock, flags);
 		raw_spin_lock(&current->blocked_lock);
 		/*
-		 * As we likely have been woken up by task
-		 * that has cleared our blocked_on state, re-set
-		 * it to the lock we are trying to acquire.
+		 * Re-set blocked_on_state as unlock path set it to WAKING/RUNNABLE
 		 */
-		__set_task_blocked_on(current, lock);
+		__force_blocked_on_blocked(current);
 		set_current_state(state);
 		/*
 		 * Here we order against unlock; we must either see it change

@@ -705,14 +703,14 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
 			 * and clear blocked on so we don't become unselectable
 			 * to run.
 			 */
-			__clear_task_blocked_on(current, lock);
+			__force_blocked_on_runnable(current);
 			raw_spin_unlock(&current->blocked_lock);
 			raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 			trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
 			opt_acquired = mutex_optimistic_spin(lock, ww_ctx, &waiter);
 			raw_spin_lock_irqsave(&lock->wait_lock, flags);
 			raw_spin_lock(&current->blocked_lock);
-			__set_task_blocked_on(current, lock);
+			__force_blocked_on_blocked(current);
 			if (opt_acquired)
 				break;
 			trace_contention_begin(lock, LCB_F_MUTEX);

@@ -963,8 +961,11 @@ static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigne
 
 		next = waiter->task;
 
+		raw_spin_lock(&next->blocked_lock);
 		debug_mutex_wake_waiter(lock, waiter);
-		clear_task_blocked_on(next, lock);
+		WARN_ON_ONCE(__get_task_blocked_on(next) != lock);
+		__set_blocked_on_waking(next);
+		raw_spin_unlock(&next->blocked_lock);
 		wake_q_add(&wake_q, next);
 	}
 

index e4a81790ea7dd..f34363615eb34 100644
@@ -285,11 +285,11 @@ __ww_mutex_die(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
 		debug_mutex_wake_waiter(lock, waiter);
 #endif
 		/*
-		 * When waking up the task to die, be sure to clear the
-		 * blocked_on pointer. Otherwise we can see circular
-		 * blocked_on relationships that can't resolve.
+		 * When waking up the task to die, be sure to set the
+		 * blocked_on_state to BO_WAKING. Otherwise we can see
+		 * circular blocked_on relationships that can't resolve.
 		 */
-		clear_task_blocked_on(waiter->task, lock);
+		set_blocked_on_waking(waiter->task);
 		wake_q_add(wake_q, waiter->task);
 	}
 

@@ -339,15 +339,11 @@ static bool __ww_mutex_wound(struct MUTEX *lock,
 		 */
 		if (owner != current) {
 			/*
-			 * When waking up the task to wound, be sure to clear the
-			 * blocked_on pointer. Otherwise we can see circular
-			 * blocked_on relationships that can't resolve.
-			 *
-			 * NOTE: We pass NULL here instead of lock, because we
-			 * are waking the mutex owner, who may be currently
-			 * blocked on a different mutex.
+			 * When waking up the task to wound, be sure to set the
+			 * blocked_on_state to BO_WAKING. Otherwise we can see
+			 * circular blocked_on relationships that can't resolve.
 			 */
-			clear_task_blocked_on(owner, NULL);
+			set_blocked_on_waking(owner);
 			wake_q_add(wake_q, owner);
 		}
 		return true;

index be9745d104f75..845454ec81a22 100644
@@ -2264,7 +2264,7 @@ static inline bool task_is_blocked(struct task_struct *p)
 	if (!sched_proxy_exec())
 		return false;
 
-	return !!p->blocked_on;
+	return !!p->blocked_on && p->blocked_on_state != BO_RUNNABLE;
 }
 
 static inline int task_on_cpu(struct rq *rq, struct task_struct *p)

2.51.0.338.gd7d06c2dae-goog

Hello John,

Of the three {set,clear}_task_blcoked_on() usage:

    $ grep -r "\(set\|clear\)_task_blocked_on" include/
    kernel/locking/mutex.c: __set_task_blocked_on(current, lock);
    kernel/locking/mutex.c: __clear_task_blocked_on(current, lock);
    kernel/locking/mutex.c: clear_task_blocked_on(current, lock);

two can be converted directly and perhaps clear_task_blocked_on() can be renamed as clear_task_blocked_on_set_runnable()?

This is just me rambling on so feel free to ignore. I probably have to train my mind enough to see __clear_task_blocked_on() not only clears "blocked_on" but also sets task to runnable :)

[..snip..]

Thanks and Regards,
Prateek

Hrm. So I guess I was thinking of set_task_blocked_on()/clear_task_blocked_on() as the main enter/exit states, with set_blocked_on_waking(), set_blocked_on_runnable() as transition states within.

But, the various force_blocked_on_<state>() cases do add more complexity, so maybe handling it completely separately would be more intuitive? I dunno.

Yeah, the case where we don't already hold the lock and want to do both gets more complex in that case.

Hrm. Maybe just the way the functions are organized in the header make it seem like we're managing two separate bits of state, where really they are ordered.
I'll try to re-arrange that introducing the set_task_blocked_on/clear_task_blocked_on first, then the transition set_blocked_on_<state>() helpers after?
Maybe that and some comments will make that clearer?

Hrrrrmmmm. I hadn't seen that detail. :/ I guess I was just lucky
it worked with my toolchain.

Oof. That may require reworking all this logic back to explicit lock/unlock calls, away from the guard() usage.

Let me think on if there's a better way.

Thanks so much again for pointing this out!
-john

Hello John,

Again that was me rambling. Even a small comment above clear_task_blocked_on() would also be sufficient if the whole rework turns out to be more extensive.

I have been too. Maybe it is okay to use a goto if folks know what they are doing ¯\_(ツ)_/¯

Another idea is to have:

    bool deactivate_donor = false;

    for (p = donor; task_is_blocked(p); p = owner) {
        guard(raw_spinlock)(...);
        ...
        if (<condition> {
            deactivate_donor = true;
            break;
        }
        ...
    }
    if (deactivate_donor)
        return proxy_deactivate(rq, donor);

Can that work?

Thanks and Regards,
Prateek

Yeah, I've reworked the logic to switch() on an action enum, which will let us do something similar without gotos.

thanks
-john

Hello John,

Thanks and Regards,
Prateek

So yeah, your suggestion is a decent one, though it gets a little messy in a few spots. I'm working on integrating this and propagating it through the full series, and hopefully I can clean it up further. There are a few spots where this and other proxy related values do get checked, so wrapping those up so they can be ifdef'ed out will require some extra logic.

Thanks for the review and suggestion!
-john

Hello John,

Looking at the proxy-exec-6.17-rc4, most of the direct references to "p->blocked_on_state" is already under CONFIG_PROXY_EXEC. Only the functions that modify it needs a stub and a helper for the usage in task_is_blocked()

Thanks and Regards,
Prateek

With proxy-exec, a task is selected to run via pick_next_task(), and then if it is a mutex blocked task, we call find_proxy_task() to find a runnable owner. If the runnable owner is on another cpu, we will need to migrate the selected donor task away, after which we will pick_again can call pick_next_task() to choose something else.

However, in the first call to pick_next_task(), we may have had a balance_callback setup by the class scheduler. After we pick again, its possible pick_next_task_fair() will be called which calls sched_balance_newidle() and sched_balance_rq().

This will throw a warning:

[    8.796467] rq->balance_callback && rq->balance_callback != &balance_push_callback
[    8.796467] WARNING: CPU: 32 PID: 458 at kernel/sched/sched.h:1750 sched_balance_rq+0xe92/0x1250
...
[    8.796467] Call Trace:
[    8.796467]  <TASK>
[    8.796467]  ? __warn.cold+0xb2/0x14e
[    8.796467]  ? sched_balance_rq+0xe92/0x1250
[    8.796467]  ? report_bug+0x107/0x1a0
[    8.796467]  ? handle_bug+0x54/0x90
[    8.796467]  ? exc_invalid_op+0x17/0x70
[    8.796467]  ? asm_exc_invalid_op+0x1a/0x20
[    8.796467]  ? sched_balance_rq+0xe92/0x1250
[    8.796467]  sched_balance_newidle+0x295/0x820
[    8.796467]  pick_next_task_fair+0x51/0x3f0
[    8.796467]  __schedule+0x23a/0x14b0
[    8.796467]  ? lock_release+0x16d/0x2e0
[    8.796467]  schedule+0x3d/0x150
[    8.796467]  worker_thread+0xb5/0x350
[    8.796467]  ? __pfx_worker_thread+0x10/0x10
[    8.796467]  kthread+0xee/0x120
[    8.796467]  ? __pfx_kthread+0x10/0x10
[    8.796467]  ret_from_fork+0x31/0x50
[    8.796467]  ? __pfx_kthread+0x10/0x10
[    8.796467]  ret_from_fork_asm+0x1a/0x30
[    8.796467]  </TASK>

This is because if a RT task was originally picked, it will setup the rq->balance_callback with push_rt_tasks() via set_next_task_rt().

Once the task is migrated away and we pick again, we haven't processed any balance callbacks, so rq->balance_callback is not in the same state as it was the first time pick_next_task was called.

To handle this, add a zap_balance_callbacks() helper function which cleans up the blance callbacks without running them. This should be ok, as we are effectively undoing the state set in the first call to pick_next_task(), and when we pick again, the new callback can be configured for the donor task actually selected.

---
v20:
* Tweaked to avoid build issues with different configs

---
 kernel/sched/core.c | 39 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 38 insertions(+), 1 deletion(-)

index e0007660161fa..01bf5ef8d9fcc 100644
@@ -5001,6 +5001,40 @@ static inline void finish_task(struct task_struct *prev)
 	smp_store_release(&prev->on_cpu, 0);
 }
 
+#if defined(CONFIG_SCHED_PROXY_EXEC)
+/*
+ * Only called from __schedule context
+ *
+ * There are some cases where we are going to re-do the action
+ * that added the balance callbacks. We may not be in a state
+ * where we can run them, so just zap them so they can be
+ * properly re-added on the next time around. This is similar
+ * handling to running the callbacks, except we just don't call
+ * them.
+ */
+static void zap_balance_callbacks(struct rq *rq)
+{
+	struct balance_callback *next, *head;
+	bool found = false;
+
+	lockdep_assert_rq_held(rq);
+
+	head = rq->balance_callback;
+	while (head) {
+		if (head == &balance_push_callback)
+			found = true;
+		next = head->next;
+		head->next = NULL;
+		head = next;
+	}
+	rq->balance_callback = found ? &balance_push_callback : NULL;
+}
+#else
+static inline void zap_balance_callbacks(struct rq *rq)
+{
+}
+#endif
+
 static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
 {
 	void (*func)(struct rq *rq);

@@ -6941,8 +6975,11 @@ static void __sched notrace __schedule(int sched_mode)
 	rq_set_donor(rq, next);
 	if (unlikely(task_is_blocked(next))) {
 		next = find_proxy_task(rq, next, &rf);
-		if (!next)
+		if (!next) {
+			/* zap the balance_callbacks before picking again */
+			zap_balance_callbacks(rq);
 			goto pick_again;
+		}
 		if (next == rq->idle)
 			goto keep_resched;
 	}

2.51.0.338.gd7d06c2dae-goog

Hello John,

s/blance/balance/

Thanks and Regards,
Prateek

Ah. Yes, this is leftover from it previously checking for PROXY_EXEC && CONFIG_SMP. I'll be sure to clean that up.

Ah, if folks are ok with that, I'd prefer it as well! Thanks for the suggestion! I'll try to work that throughout the series.

Hrm. So I don't think it's strictly necessary, because we will run the set callback as part of finish_task_switch() when we switch briefly to idle. So we don't end up with stale callbacks in the next pick_next_task().

But I guess zapping them could help just avoid running it spuriously. I'll give that a shot and see how it affects things.

Thanks again for all the suggestions!
-john

Add logic to handle migrating a blocked waiter to a remote cpu where the lock owner is runnable.

Additionally, as the blocked task may not be able to run on the remote cpu, add logic to handle return migration once the waiting task is given the mutex.

Because tasks may get migrated to where they cannot run, also modify the scheduling classes to avoid sched class migrations on mutex blocked tasks, leaving find_proxy_task() and related logic to do the migrations and return migrations.

This was split out from the larger proxy patch, and significantly reworked.

Credits for the original patch go to:
Peter Zijlstra (Intel) <[email protected]>
Juri Lelli <[email protected]>
Valentin Schneider <[email protected]>
Connor O'Brien <[email protected]>

---
v6:
* Integrated sched_proxy_exec() check in proxy_return_migration()
* Minor cleanups to diff
* Unpin the rq before calling __balance_callbacks()
* Tweak proxy migrate to migrate deeper task in chain, to avoid tasks pingponging between rqs
v7:
* Fixup for unused function arguments
* Switch from that_rq -> target_rq, other minor tweaks, and typo fixes suggested by Metin Kaya
* Switch back to doing return migration in the ttwu path, which avoids nasty lock juggling and performance issues
* Fixes for UP builds
v8:
* More simplifications from Metin Kaya
* Fixes for null owner case, including doing return migration
* Cleanup proxy_needs_return logic
v9:
* Narrow logic in ttwu that sets BO_RUNNABLE, to avoid missed return migrations
* Switch to using zap_balance_callbacks rathern then running them when we are dropping rq locks for proxy_migration.
* Drop task_is_blocked check in sched_submit_work as suggested by Metin (may re-add later if this causes trouble)
* Do return migration when we're not on wake_cpu. This avoids bad task placement caused by proxy migrations raised by Xuewen Yan
* Fix to call set_next_task(rq->curr) prior to dropping rq lock to avoid rq->curr getting migrated before we have actually switched from it
* Cleanup to re-use proxy_resched_idle() instead of open coding it in proxy_migrate_task()
* Fix return migration not to use DEQUEUE_SLEEP, so that we properly see the task as task_on_rq_migrating() after it is dequeued but before set_task_cpu() has been called on it
* Fix to broaden find_proxy_task() checks to avoid race where a task is dequeued off the rq due to return migration, but set_task_cpu() and the enqueue on another rq happened after we checked task_cpu(owner). This ensures we don't proxy using a task that is not actually on our runqueue.
* Cleanup to avoid the locked BO_WAKING->BO_RUNNABLE transition in try_to_wake_up() if proxy execution isn't enabled.
* Cleanup to improve comment in proxy_migrate_task() explaining the set_next_task(rq->curr) logic
* Cleanup deadline.c change to stylistically match rt.c change
* Numerous cleanups suggested by Metin
v10:
* Drop WARN_ON(task_is_blocked(p)) in ttwu current case v11:
* Include proxy_set_task_cpu from later in the series to this change so we can use it, rather then reworking logic later in the series.
* Fix problem with return migration, where affinity was changed and wake_cpu was left outside the affinity mask.
* Avoid reading the owner's cpu twice (as it might change inbetween) to avoid occasional migration-to-same-cpu edge cases
* Add extra WARN_ON checks for wake_cpu and return migration edge cases.
* Typo fix from Metin
v13:
* As we set ret, return it, not just NULL (pulling this change in from later patch)
* Avoid deadlock between try_to_wake_up() and find_proxy_task() when blocked_on cycle with ww_mutex is trying a mid-chain wakeup.
* Tweaks to use new __set_blocked_on_runnable() helper
* Potential fix for incorrectly updated task->dl_server issues
* Minor comment improvements
* Add logic to handle missed wakeups, in that case doing return migration from the find_proxy_task() path
* Minor cleanups
v14:
* Improve edge cases where we wouldn't set the task as BO_RUNNABLE v15:
* Added comment to better describe proxy_needs_return() as suggested by Qais
* Build fixes for !CONFIG_SMP reported by
Maciej Żenczykowski <[email protected]>
* Adds fix for re-evaluating proxy_needs_return when sched_proxy_exec() is disabled, reported and diagnosed by:
kuyo chang <[email protected]>
v16:
* Larger rework of needs_return logic in find_proxy_task, in order to avoid problems with cpuhotplug
* Rework to use guard() as suggested by Peter
v18:
* Integrate optimization suggested by Suleiman to do the checks for sleeping owners before checking if the task_cpu is this_cpu, so that we can avoid needlessly proxy-migrating tasks to only then dequeue them. Also check if migrating last.
* Improve comments around guard locking
* Include tweak to ttwu_runnable() as suggested by
hupu <[email protected]>
* Rework the logic releasing the rq->donor reference before letting go of the rqlock. Just use rq->idle.
* Go back to doing return migration on BO_WAKING owners, as I was hitting some softlockups caused by running tasks not making it out of BO_WAKING.
v19:
* Fixed proxy_force_return() logic for !SMP cases
v21:
* Reworked donor deactivation for unhandled sleeping owners
* Commit message tweaks

---
 kernel/sched/core.c | 247 +++++++++++++++++++++++++++++++++++++++-----
 kernel/sched/fair.c |   3 +-
 2 files changed, 222 insertions(+), 28 deletions(-)

index 01bf5ef8d9fcc..0f824446c6046 100644
@@ -3157,6 +3157,14 @@ static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
 
 	__do_set_cpus_allowed(p, ctx);
 
+	/*
+	 * It might be that the p->wake_cpu is no longer
+	 * allowed, so set it to the dest_cpu so return
+	 * migration doesn't send it to an invalid cpu
+	 */
+	if (!is_cpu_allowed(p, p->wake_cpu))
+		p->wake_cpu = dest_cpu;
+
 	return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
 
 out:

@@ -3717,6 +3725,67 @@ static inline void ttwu_do_wakeup(struct task_struct *p)
 	trace_sched_wakeup(p);
 }
 
+#ifdef CONFIG_SCHED_PROXY_EXEC
+static inline void proxy_set_task_cpu(struct task_struct *p, int cpu)
+{
+	unsigned int wake_cpu;
+
+	/*
+	 * Since we are enqueuing a blocked task on a cpu it may
+	 * not be able to run on, preserve wake_cpu when we
+	 * __set_task_cpu so we can return the task to where it
+	 * was previously runnable.
+	 */
+	wake_cpu = p->wake_cpu;
+	__set_task_cpu(p, cpu);
+	p->wake_cpu = wake_cpu;
+}
+
+static bool proxy_task_runnable_but_waking(struct task_struct *p)
+{
+	if (!sched_proxy_exec())
+		return false;
+	return (READ_ONCE(p->__state) == TASK_RUNNING &&
+		READ_ONCE(p->blocked_on_state) == BO_WAKING);
+}
+#else /* !CONFIG_SCHED_PROXY_EXEC */
+static bool proxy_task_runnable_but_waking(struct task_struct *p)
+{
+	return false;
+}
+#endif /* CONFIG_SCHED_PROXY_EXEC */
+
+/*
+ * Checks to see if task p has been proxy-migrated to another rq
+ * and needs to be returned. If so, we deactivate the task here
+ * so that it can be properly woken up on the p->wake_cpu
+ * (or whichever cpu select_task_rq() picks at the bottom of
+ * try_to_wake_up()
+ */
+static inline bool proxy_needs_return(struct rq *rq, struct task_struct *p)
+{
+	bool ret = false;
+
+	if (!sched_proxy_exec())
+		return false;
+
+	raw_spin_lock(&p->blocked_lock);
+	if (__get_task_blocked_on(p) && p->blocked_on_state == BO_WAKING) {
+		if (!task_current(rq, p) && (p->wake_cpu != cpu_of(rq))) {
+			if (task_current_donor(rq, p)) {
+				put_prev_task(rq, p);
+				rq_set_donor(rq, rq->idle);
+			}
+			deactivate_task(rq, p, DEQUEUE_NOCLOCK);
+			ret = true;
+		}
+		__set_blocked_on_runnable(p);
+		resched_curr(rq);
+	}
+	raw_spin_unlock(&p->blocked_lock);
+	return ret;
+}
+
 static void
 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
 		 struct rq_flags *rf)

@@ -3802,6 +3871,8 @@ static int ttwu_runnable(struct task_struct *p, int wake_flags)
 		update_rq_clock(rq);
 		if (p->se.sched_delayed)
 			enqueue_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_DELAYED);
+		if (proxy_needs_return(rq, p))
+			goto out;
 		if (!task_on_cpu(rq, p)) {
 			/*
 			 * When on_rq && !on_cpu the task is preempted, see if

@@ -3812,6 +3883,7 @@ static int ttwu_runnable(struct task_struct *p, int wake_flags)
 		ttwu_do_wakeup(p);
 		ret = 1;
 	}
+out:
 	__task_rq_unlock(rq, &rf);
 
 	return ret;

@@ -4199,6 +4271,8 @@ int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 		 *    it disabling IRQs (this allows not taking ->pi_lock).
 		 */
 		WARN_ON_ONCE(p->se.sched_delayed);
+		/* If current is waking up, we know we can run here, so set BO_RUNNBLE */
+		set_blocked_on_runnable(p);
 		if (!ttwu_state_match(p, state, &success))
 			goto out;
 

@@ -4215,8 +4289,15 @@ int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 	 */
 	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
 		smp_mb__after_spinlock();
-		if (!ttwu_state_match(p, state, &success))
-			break;
+		if (!ttwu_state_match(p, state, &success)) {
+			/*
+			 * If we're already TASK_RUNNING, and BO_WAKING
+			 * continue on to ttwu_runnable check to force
+			 * proxy_needs_return evaluation
+			 */
+			if (!proxy_task_runnable_but_waking(p))
+				break;
+		}
 
 		trace_sched_waking(p);
 

@@ -4278,6 +4359,7 @@ int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 		 * enqueue, such as ttwu_queue_wakelist().
 		 */
 		WRITE_ONCE(p->__state, TASK_WAKING);
+		set_blocked_on_runnable(p);
 
 		/*
 		 * If the owning (remote) CPU is still in the middle of schedule() with

@@ -4328,12 +4410,6 @@ int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 		ttwu_queue(p, cpu, wake_flags);
 	}
 out:
-	/*
-	 * For now, if we've been woken up, set us as BO_RUNNABLE
-	 * We will need to be more careful later when handling
-	 * proxy migration
-	 */
-	set_blocked_on_runnable(p);
 	if (success)
 		ttwu_stat(p, task_cpu(p), wake_flags);
 

@@ -6635,7 +6711,7 @@ static inline struct task_struct *proxy_resched_idle(struct rq *rq)
 	return rq->idle;
 }
 
-static bool __proxy_deactivate(struct rq *rq, struct task_struct *donor)
+static bool proxy_deactivate(struct rq *rq, struct task_struct *donor)
 {
 	unsigned long state = READ_ONCE(donor->__state);
 

@@ -6655,17 +6731,98 @@ static bool __proxy_deactivate(struct rq *rq, struct task_struct *donor)
 	return try_to_block_task(rq, donor, &state, true);
 }
 
-static struct task_struct *proxy_deactivate(struct rq *rq, struct task_struct *donor)
+/*
+ * If the blocked-on relationship crosses CPUs, migrate @p to the
+ * owner's CPU.
+ *
+ * This is because we must respect the CPU affinity of execution
+ * contexts (owner) but we can ignore affinity for scheduling
+ * contexts (@p). So we have to move scheduling contexts towards
+ * potential execution contexts.
+ *
+ * Note: The owner can disappear, but simply migrate to @target_cpu
+ * and leave that CPU to sort things out.
+ */
+static void proxy_migrate_task(struct rq *rq, struct rq_flags *rf,
+			       struct task_struct *p, int target_cpu)
 {
-	if (!__proxy_deactivate(rq, donor)) {
-		/*
-		 * XXX: For now, if deactivation failed, set donor
-		 * as unblocked, as we aren't doing proxy-migrations
-		 * yet (more logic will be needed then).
-		 */
-		force_blocked_on_runnable(donor);
-	}
-	return NULL;
+	struct rq *target_rq = cpu_rq(target_cpu);
+
+	lockdep_assert_rq_held(rq);
+
+	/*
+	 * Since we're going to drop @rq, we have to put(@rq->donor) first,
+	 * otherwise we have a reference that no longer belongs to us.
+	 *
+	 * Additionally, as we put_prev_task(prev) earlier, its possible that
+	 * prev will migrate away as soon as we drop the rq lock, however we
+	 * still have it marked as rq->curr, as we've not yet switched tasks.
+	 *
+	 * After the migration, we are going to pick_again in the __schedule
+	 * logic, so backtrack a bit before we release the lock:
+	 * Put rq->donor, and set rq->curr as rq->donor and set_next_task,
+	 * so that we're close to the situation we had entering __schedule
+	 * the first time.
+	 *
+	 * Then when we re-aquire the lock, we will re-put rq->curr then
+	 * rq_set_donor(rq->idle) and set_next_task(rq->idle), before
+	 * picking again.
+	 */
+	/* XXX - Added to address problems with changed dl_server semantics - double check */
+	__put_prev_set_next_dl_server(rq, rq->donor, rq->curr);
+	put_prev_task(rq, rq->donor);
+	rq_set_donor(rq, rq->idle);
+	set_next_task(rq, rq->idle);
+
+	WARN_ON(p == rq->curr);
+
+	deactivate_task(rq, p, 0);
+	proxy_set_task_cpu(p, target_cpu);
+
+	zap_balance_callbacks(rq);
+	rq_unpin_lock(rq, rf);
+	raw_spin_rq_unlock(rq);
+	raw_spin_rq_lock(target_rq);
+
+	activate_task(target_rq, p, 0);
+	wakeup_preempt(target_rq, p, 0);
+
+	raw_spin_rq_unlock(target_rq);
+	raw_spin_rq_lock(rq);
+	rq_repin_lock(rq, rf);
+}
+
+static void proxy_force_return(struct rq *rq, struct rq_flags *rf,
+			       struct task_struct *p)
+{
+	lockdep_assert_rq_held(rq);
+
+	put_prev_task(rq, rq->donor);
+	rq_set_donor(rq, rq->idle);
+	set_next_task(rq, rq->idle);
+
+	WARN_ON(p == rq->curr);
+
+	set_blocked_on_waking(p);
+	get_task_struct(p);
+	block_task(rq, p, 0);
+
+	zap_balance_callbacks(rq);
+	rq_unpin_lock(rq, rf);
+	raw_spin_rq_unlock(rq);
+
+	wake_up_process(p);
+	put_task_struct(p);
+
+	raw_spin_rq_lock(rq);
+	rq_repin_lock(rq, rf);
+}
+
+static inline bool proxy_can_run_here(struct rq *rq, struct task_struct *p)
+{
+	if (p == rq->curr || p->wake_cpu == cpu_of(rq))
+		return true;
+	return false;
 }
 
 /*

@@ -6688,9 +6845,11 @@ static struct task_struct *
 find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 {
 	struct task_struct *owner = NULL;
+	bool curr_in_chain = false;
 	int this_cpu = cpu_of(rq);
 	struct task_struct *p;
 	struct mutex *mutex;
+	int owner_cpu;
 
 	/* Follow blocked_on chain. */
 	for (p = donor; task_is_blocked(p); p = owner) {

@@ -6716,6 +6875,10 @@ find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 			return NULL;
 		}
 
+		/* Double check blocked_on_state now we're holding the lock */
+		if (p->blocked_on_state == BO_RUNNABLE)
+			return p;
+
 		/*
 		 * If a ww_mutex hits the die/wound case, it marks the task as
 		 * BO_WAKING and calls try_to_wake_up(), so that the mutex

@@ -6731,26 +6894,46 @@ find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 		 * try_to_wake_up from completing and doing the return
 		 * migration.
 		 *
-		 * So when we hit a !BO_BLOCKED task briefly schedule idle
-		 * so we release the rq and let the wakeup complete.
+		 * So when we hit a BO_WAKING task try to wake it up ourselves.
 		 */
-		if (p->blocked_on_state != BO_BLOCKED)
-			return proxy_resched_idle(rq);
+		if (p->blocked_on_state == BO_WAKING) {
+			if (task_current(rq, p)) {
+				/* If its current just set it runnable */
+				__force_blocked_on_runnable(p);
+				return p;
+			}
+			goto needs_return;
+		}
+
+		if (task_current(rq, p))
+			curr_in_chain = true;
 
 		owner = __mutex_owner(mutex);
 		if (!owner) {
+			/* If the owner is null, we may have some work to do */
+			if (!proxy_can_run_here(rq, p))
+				goto needs_return;
+
 			__force_blocked_on_runnable(p);
 			return p;
 		}
 
 		if (!READ_ONCE(owner->on_rq) || owner->se.sched_delayed) {
 			/* XXX Don't handle blocked owners/delayed dequeue yet */
+			if (curr_in_chain)
+				return proxy_resched_idle(rq);
 			goto deactivate_donor;
 		}
 
-		if (task_cpu(owner) != this_cpu) {
-			/* XXX Don't handle migrations yet */
-			goto deactivate_donor;
+		owner_cpu = task_cpu(owner);
+		if (owner_cpu != this_cpu) {
+			/*
+			 * @owner can disappear, simply migrate to @owner_cpu
+			 * and leave that CPU to sort things out.
+			 */
+			if (curr_in_chain)
+				return proxy_resched_idle(rq);
+			goto migrate;
 		}
 
 		if (task_on_rq_migrating(owner)) {

@@ -6817,8 +7000,18 @@ find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 	 * blocked_lock released, so we have to get out of the
 	 * guard() scope.
 	 */
+migrate:
+	proxy_migrate_task(rq, rf, p, owner_cpu);
+	return NULL;
+needs_return:
+	proxy_force_return(rq, rf, p);
+	return NULL;
 deactivate_donor:
-	return proxy_deactivate(rq, donor);
+	if (!proxy_deactivate(rq, donor)) {
+		p = donor;
+		goto needs_return;
+	}
+	return NULL;
 }
 #else /* SCHED_PROXY_EXEC */
 static struct task_struct *

index b173a059315c2..cc531eb939831 100644
@@ -8781,7 +8781,8 @@ pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf
 	se = &p->se;
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
-	if (prev->sched_class != &fair_sched_class)
+	if (prev->sched_class != &fair_sched_class ||
+	    rq->curr != rq->donor)
 		goto simple;
 
 	__put_prev_set_next_dl_server(rq, prev, p);

2.51.0.338.gd7d06c2dae-goog

Hello John,

    __put_prev_set_next_dl_server(rq, rq->donor, rq->idle);

    put_prev_set_next_task(rq, rq->donor, rq->idle);
    rq_set_donor(rq, rq->idle);

Thanks and Regards,
Prateek

Nice! That is simpler. I think I can also simplify those two lines to proxy_resched_idle() as well.
And looking, the big comment above that section needs an update as well as its out of date.

So unfortunately the issue is that when we're letting go of the runqueue other CPUs can jump in via sched_balance_domains() -> sched_balance_rq() -> rq_lock_irqsave() -> rq_pin_lock() which will trip the
WARN_ON_ONCE(rq->balance_callback && rq->balance_callback != &balance_push_callback); case if we leave callbacks.

So we need to zap the callbacks before we drop the rq lock. I'll add a comment to make that clear.

And yeah, that means we do call zap_balance_callbacks() again after we re-take the lock and return null.

I guess we could remove the zap_balance_callbacks() call in __schedule() and instead call it in every case where find_proxy_task() returns NULL? Or we call it twice in the migration paths, as it should just have one entry the second time.

Ack.

Hrmm.. that's a good question, as I haven't thought about that in awhile, and I'm unfortunately forgetful. My instinct is that its a concern that by checking the prev sched class is fair, is assuming the prev task was selected by the fair scheduler's pick_next_task last time around. But since we may have picked a blocked RT task as donor, if we are proxying, we can't assume prev was previously chosen by pick_next_task_fair(). So we skip the optimization just to be careful.

But this may be overly cautious and looking it over I don't immediately see it as necessary. So I'll drop it and do some thorough testing without it. If I do find an issue I'll re-add it with a much better comment so I don't forget again. I do really appreciate you checking me here!

thanks
-john

Hello John,

Ah! That makes sense. I overlooked the fact that we have *other CPUs* that can take the rq lock once dropped.

Thank you!

I think a comment would be good enough with the current flow. I'll let you decide which is the best option based on your understanding of the complete flow (I'm still slowly wrapping my head around all this :)

[..snip..]

My interpretation was - since pick_next_task_fair() selected the scheduling context, it shouldn't matter what the execution context was.

Thank you! I too will do some testing and let you know if I see something go awry.

Thanks and Regards,
Prateek

From: Peter Zijlstra <[email protected]>

Add link to the task this task is proxying for, and use it so the mutex owner can do an intelligent hand-off of the mutex to the task that the owner is running on behalf.

---
v5:
* Split out from larger proxy patch
v6:
* Moved proxied value from earlier patch to this one where it is actually used
* Rework logic to check sched_proxy_exec() instead of using ifdefs
* Moved comment change to this patch where it makes sense v7:
* Use more descriptive term then "us" in comments, as suggested by Metin Kaya.
* Minor typo fixup from Metin Kaya
* Reworked proxied variable to prev_not_proxied to simplify usage v8:
* Use helper for donor blocked_on_state transition
v9:
* Re-add mutex lock handoff in the unlock path, but only when we have a blocked donor
* Slight reword of commit message suggested by Metin v18:
* Add task_init initialization for blocked_donor, suggested by Suleiman

---
 include/linux/sched.h  |  1 +
 init/init_task.c       |  1 +
 kernel/fork.c          |  1 +
 kernel/locking/mutex.c | 41 ++++++++++++++++++++++++++++++++++++++---
 kernel/sched/core.c    | 18 ++++++++++++++++--
 5 files changed, 57 insertions(+), 5 deletions(-)

index 5801de1a44a79..ab12eb738c440 100644
@@ -1242,6 +1242,7 @@ struct task_struct {
 
 	enum blocked_on_state		blocked_on_state;
 	struct mutex			*blocked_on;	/* lock we're blocked on */
+	struct task_struct		*blocked_donor;	/* task that is boosting this task */
 	raw_spinlock_t			blocked_lock;
 
 #ifdef CONFIG_DETECT_HUNG_TASK_BLOCKER

index 6d72ec23410a6..627bbd8953e88 100644
@@ -175,6 +175,7 @@ struct task_struct init_task __aligned(L1_CACHE_BYTES) = {
 						 &init_task.alloc_lock),
 #endif
 	.blocked_on_state = BO_RUNNABLE,
+	.blocked_donor = NULL,
 #ifdef CONFIG_RT_MUTEXES
 	.pi_waiters	= RB_ROOT_CACHED,
 	.pi_top_task	= NULL,

index 4bd0731995e86..86fe43ee35952 100644
@@ -2131,6 +2131,7 @@ __latent_entropy struct task_struct *copy_process(
 
 	p->blocked_on_state = BO_RUNNABLE;
 	p->blocked_on = NULL; /* not blocked yet */
+	p->blocked_donor = NULL; /* nobody is boosting p yet */
 
 #ifdef CONFIG_BCACHE
 	p->sequential_io	= 0;

index 0f824446c6046..cac03f68cbcce 100644
@@ -6829,7 +6829,17 @@ static inline bool proxy_can_run_here(struct rq *rq, struct task_struct *p)
  * Find runnable lock owner to proxy for mutex blocked donor
  *
  * Follow the blocked-on relation:
- *   task->blocked_on -> mutex->owner -> task...
+ *
+ *                ,-> task
+ *                |     | blocked-on
+ *                |     v
+ *  blocked_donor |   mutex
+ *                |     | owner
+ *                |     v
+ *                `-- task
+ *
+ * and set the blocked_donor relation, this latter is used by the mutex
+ * code to find which (blocked) task to hand-off to.
  *
  * Lock order:
  *

@@ -6989,6 +6999,7 @@ find_proxy_task(struct rq *rq, struct task_struct *donor, struct rq_flags *rf)
 		 * rq, therefore holding @rq->lock is sufficient to
 		 * guarantee its existence, as per ttwu_remote().
 		 */
+		owner->blocked_donor = p;
 	}
 
 	WARN_ON_ONCE(owner && !owner->on_rq);

@@ -7091,6 +7102,7 @@ static void __sched notrace __schedule(int sched_mode)
 	unsigned long prev_state;
 	struct rq_flags rf;
 	struct rq *rq;
+	bool prev_not_proxied;
 	int cpu;
 
 	/* Trace preemptions consistently with task switches */

@@ -7163,9 +7175,11 @@ static void __sched notrace __schedule(int sched_mode)
 		switch_count = &prev->nvcsw;
 	}
 
+	prev_not_proxied = !prev->blocked_donor;
 pick_again:
 	next = pick_next_task(rq, rq->donor, &rf);
 	rq_set_donor(rq, next);
+	next->blocked_donor = NULL;
 	if (unlikely(task_is_blocked(next))) {
 		next = find_proxy_task(rq, next, &rf);
 		if (!next) {

@@ -7229,7 +7243,7 @@ static void __sched notrace __schedule(int sched_mode)
 		rq = context_switch(rq, prev, next, &rf);
 	} else {
 		/* In case next was already curr but just got blocked_donor */
-		if (!task_current_donor(rq, next))
+		if (prev_not_proxied && next->blocked_donor)
 			proxy_tag_curr(rq, next);
 
 		rq_unpin_lock(rq, &rf);

2.51.0.338.gd7d06c2dae-goog

Instead of migrating one task each time through find_proxy_task(), we can walk up the blocked_donor ptrs and migrate the entire current chain in one go.

This was broken out of earlier patches and held back while the series was being stabilized, but I wanted to re-introduce it.

---
v12:
* Earlier this was re-using blocked_node, but I hit a race with activating blocked entities, and to avoid it introduced a new migration_node listhead
v18:
* Add init_task initialization of migration_node as suggested by Suleiman

---
 include/linux/sched.h |  1 +
 init/init_task.c      |  1 +
 kernel/fork.c         |  1 +
 kernel/sched/core.c   | 25 +++++++++++++++++--------
 4 files changed, 20 insertions(+), 8 deletions(-)

index ab12eb738c440..176ec117f4041 100644
@@ -1243,6 +1243,7 @@ struct task_struct {
 	enum blocked_on_state		blocked_on_state;
 	struct mutex			*blocked_on;	/* lock we're blocked on */
 	struct task_struct		*blocked_donor;	/* task that is boosting this task */
+	struct list_head		migration_node;
 	raw_spinlock_t			blocked_lock;
 
 #ifdef CONFIG_DETECT_HUNG_TASK_BLOCKER

index 627bbd8953e88..65e0f90285966 100644
@@ -176,6 +176,7 @@ struct task_struct init_task __aligned(L1_CACHE_BYTES) = {
 #endif
 	.blocked_on_state = BO_RUNNABLE,
 	.blocked_donor = NULL,
+	.migration_node = LIST_HEAD_INIT(init_task.migration_node),
 #ifdef CONFIG_RT_MUTEXES
 	.pi_waiters	= RB_ROOT_CACHED,
 	.pi_top_task	= NULL,

index 86fe43ee35952..01fd08c463871 100644
@@ -2132,6 +2132,7 @@ __latent_entropy struct task_struct *copy_process(
 	p->blocked_on_state = BO_RUNNABLE;
 	p->blocked_on = NULL; /* not blocked yet */
 	p->blocked_donor = NULL; /* nobody is boosting p yet */
+	INIT_LIST_HEAD(&p->migration_node);
 
 #ifdef CONFIG_BCACHE
 	p->sequential_io	= 0;

index cac03f68cbcce..26f7a11a39e0e 100644
@@ -6747,6 +6747,7 @@ static void proxy_migrate_task(struct rq *rq, struct rq_flags *rf,
 			       struct task_struct *p, int target_cpu)
 {
 	struct rq *target_rq = cpu_rq(target_cpu);
+	LIST_HEAD(migrate_list);
 
 	lockdep_assert_rq_held(rq);
 

@@ -6774,19 +6775,27 @@ static void proxy_migrate_task(struct rq *rq, struct rq_flags *rf,
 	rq_set_donor(rq, rq->idle);
 	set_next_task(rq, rq->idle);
 
-	WARN_ON(p == rq->curr);
-
-	deactivate_task(rq, p, 0);
-	proxy_set_task_cpu(p, target_cpu);
-
+	for (; p; p = p->blocked_donor) {
+		WARN_ON(p == rq->curr);
+		deactivate_task(rq, p, 0);
+		proxy_set_task_cpu(p, target_cpu);
+		/*
+		 * We can abuse blocked_node to migrate the thing,
+		 * because @p was still on the rq.
+		 */
+		list_add(&p->migration_node, &migrate_list);
+	}
 	zap_balance_callbacks(rq);
 	rq_unpin_lock(rq, rf);
 	raw_spin_rq_unlock(rq);
 	raw_spin_rq_lock(target_rq);
+	while (!list_empty(&migrate_list)) {
+		p = list_first_entry(&migrate_list, struct task_struct, migration_node);
+		list_del_init(&p->migration_node);
 
-	activate_task(target_rq, p, 0);
-	wakeup_preempt(target_rq, p, 0);
-
+		activate_task(target_rq, p, 0);
+		wakeup_preempt(target_rq, p, 0);
+	}
 	raw_spin_rq_unlock(target_rq);
 	raw_spin_rq_lock(rq);
 	rq_repin_lock(rq, rf);

2.51.0.338.gd7d06c2dae-goog

Hi John,

Not ignoring you, but I had to spend some time putting together some testing infra I am not trying to use to see how DEADLINE behaves with the series, as it's somewhat difficult for me to think in abstract about all this. :)

I current have the following on top of your complete series

https://github.com/jlelli/linux/commits/experimental/eval-mbwi/
https://github.com/jlelli/linux experimental/eval-mbwi

of which

https://github.com/jlelli/linux/commit/9d4bbb1aca624e76e5b34938d848dc9a418c6146

introduces the testing (M-BWI is Multiprocessor Bandwidth Inheritance) infra and the rest some additional tracepoints (based on Gabriele's patch) to get more DEADLINE info out of tests (in conjuction with sched_tp [1]).

Nothing bit to report just yet, mainly spent time getting this working.

One thing I noticed thouh (and probably forgot from previous discussions) is that spin_on_owner might be 'confusing' from an RT/DEADLINE perspective as it deviates from what one expects from the ideal theoretical world (as tasks don't immediately block and potentially donate). Not sure what to do about it. Maybe special case it for RT/DEADLINE, but just started playing with it.

Anyway, I will keep playing with all this. Just wanted to give you/others a quick update. Also adding Luca, Tommaso and Yuri to the thread so that they are aware of the testing framework. :)

Thanks!
Juri

1 - https://github.com/jlelli/sched_tp/tree/deadline-tp
will open an MR against Qais' mainline repo as soon as TPs are hopefully merged upstream as well.

Very cool to see! I'll have to pull those and take a look at it!

And I'm of course very interested to hear if you find anything with the proxy set that I need to revise.

Can you refresh me a bit on why blocking to donate is preferred? If the lock owner is running, there's not much that blocking to donate would help with. Does this concern not apply to the current mutex logic without proxy? With proxy-exec, I'm trying to preserve the existing mutex behavior of spin_on_owner, with the main tweak is just the lock handoff to the current donor when we are proxying, otherwise the intent is it should be the same.

Now, I do recognize that rt_mutexes and mutexes do have different lock handoff requirements for RT tasks (needs to strictly go to the highest priority waiter, and we can't let a lower priority task steal it), which is why I've not yet enabled proxy-exec on rt_mutexes.

Thanks so much for sharing!
-john

Yeah, I think we want to preserve that behavior for non-RT mutexes for throughput, but for RT I fear we might risk priority inversion if tasks spin (for a bit) before blocking. My understanding is that with PI enabled futexes (apart from some initial tries to get the lock with atomic ops) we then call into __rt_mutex_start_proxy_lock() which enqueues the blocked tasks onto the pi chain (so that PI rules are respected etc.). Guess maybe we could end-up reintroducing this behavior when we eventually kill rt-mutexes, so don't worry to much about it yet I think, just something to keep in mind. :)

Right. I will probably hack something in to test the DEADLINE scenarios, but again don't worry about it.

Thanks,
Juri

Hello John,

tl;dr

This series seems fine from performance standpoint but the above branch may have some performance issues but take them with a grain of salt since this is not all apples to apples comparison.

For this series things are alright - my harness for longer running benchmarks gave up for some reason so I'll rerun those tests again and report back later but either tip has some improvements for netperf / tbench or "proxy-exec-v21-6.17-rc4" may have some issues around it. I'll take a deeper look later in the week.

o System Details

- 3rd Generation EPYC System
- 2 x 64C/128T
- NPS1 mode

o Kernels

- tip tip:sched/core at commit 5b726e9bf954 ("sched/fair: Get
rid of throttled_lb_pair()") (CONFIG_SCHED_PROXY_EXEC disabled)

- proxy-v21 tip + this series as is
(CONFIG_SCHED_PROXY_EXEC=y)

- proxy-full proxy-exec-v21-6.17-rc4 as is
(CONFIG_SCHED_PROXY_EXEC=y)

o Benchmark results

    ==================================================================
    Test          : hackbench
    Units         : Normalized time in seconds
    Interpretation: Lower is better
    Statistic     : AMean
    ==================================================================
    Case:           tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
     1-groups     1.00 [ -0.00](10.57)     0.94 [  6.24]( 7.88)     0.91 [  9.46](10.11)
     2-groups     1.00 [ -0.00]( 3.33)     1.02 [ -1.75]( 3.16)     1.04 [ -4.17]( 2.51)
     4-groups     1.00 [ -0.00]( 2.41)     1.01 [ -0.87]( 2.29)     1.03 [ -3.03]( 1.27)
     8-groups     1.00 [ -0.00]( 2.67)     1.02 [ -1.66]( 2.10)     1.01 [ -0.55]( 1.45)
    16-groups     1.00 [ -0.00]( 1.83)     1.01 [ -0.82]( 2.30)     1.00 [ -0.25]( 1.72)


    ==================================================================
    Test          : tbench
    Units         : Normalized throughput
    Interpretation: Higher is better
    Statistic     : AMean
    ==================================================================
    Clients:    tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
        1     1.00 [  0.00]( 0.81)     1.00 [ -0.13]( 0.16)     0.92 [ -8.06]( 0.39)
        2     1.00 [  0.00]( 0.32)     0.99 [ -0.84]( 0.66)     0.91 [ -8.85]( 0.54)
        4     1.00 [  0.00]( 0.32)     0.98 [ -2.37]( 1.40)     0.92 [ -8.28]( 0.28)
        8     1.00 [  0.00]( 0.69)     0.98 [ -2.47]( 0.53)     0.90 [ -9.58]( 0.36)
       16     1.00 [  0.00]( 1.24)     0.96 [ -3.94]( 1.51)     0.90 [ -9.83]( 0.69)
       32     1.00 [  0.00]( 0.60)     0.99 [ -1.47]( 3.38)     0.89 [-11.43]( 5.60)
       64     1.00 [  0.00]( 1.22)     0.99 [ -1.33]( 0.88)     0.91 [ -8.52]( 2.67)
      128     1.00 [  0.00]( 0.34)     0.99 [ -1.48]( 0.99)     0.92 [ -7.51]( 0.13)
      256     1.00 [  0.00]( 1.32)     0.98 [ -1.75]( 0.96)     0.97 [ -3.35]( 1.22)
      512     1.00 [  0.00]( 0.25)     0.99 [ -1.29]( 0.41)     0.97 [ -2.90]( 0.17)
     1024     1.00 [  0.00]( 0.24)     0.99 [ -0.59]( 0.14)     0.98 [ -2.36]( 0.33)


    ==================================================================
    Test          : stream-10
    Units         : Normalized Bandwidth, MB/s
    Interpretation: Higher is better
    Statistic     : HMean
    ==================================================================
    Test:       tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
     Copy     1.00 [  0.00](10.90)     1.07 [  6.53]( 8.21)     1.07 [  7.26]( 7.22)
    Scale     1.00 [  0.00]( 9.62)     1.04 [  4.00]( 6.99)     1.05 [  4.71]( 5.85)
      Add     1.00 [  0.00](10.17)     1.05 [  5.07]( 6.14)     1.06 [  6.03]( 6.56)
    Triad     1.00 [  0.00]( 8.48)     1.04 [  4.34]( 5.09)     1.04 [  4.07]( 4.40)


    ==================================================================
    Test          : stream-100
    Units         : Normalized Bandwidth, MB/s
    Interpretation: Higher is better
    Statistic     : HMean
    ==================================================================
    Test:       tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
     Copy     1.00 [  0.00]( 1.38)     1.01 [  0.99]( 1.21)     1.02 [  1.68]( 1.50)
    Scale     1.00 [  0.00]( 6.19)     1.02 [  1.94]( 4.34)     1.03 [  3.00]( 1.19)
      Add     1.00 [  0.00]( 4.42)     1.01 [  0.94]( 4.17)     1.02 [  1.58]( 1.54)
    Triad     1.00 [  0.00]( 1.30)     1.01 [  0.61]( 1.37)     1.00 [  0.18]( 2.65)


    ==================================================================
    Test          : netperf
    Units         : Normalized Througput
    Interpretation: Higher is better
    Statistic     : AMean
    ==================================================================
    Clients:         tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
     1-clients     1.00 [  0.00]( 0.41)     0.99 [ -1.03]( 0.34)     0.90 [ -9.96]( 0.46)
     2-clients     1.00 [  0.00]( 0.31)     0.99 [ -1.17]( 0.72)     0.90 [ -9.77]( 0.78)
     4-clients     1.00 [  0.00]( 0.57)     0.99 [ -0.68]( 0.32)     0.90 [-10.21]( 0.89)
     8-clients     1.00 [  0.00]( 0.46)     0.99 [ -0.69]( 0.32)     0.90 [-10.20]( 0.70)
    16-clients     1.00 [  0.00]( 0.57)     0.99 [ -1.39]( 1.28)     0.90 [-10.37]( 1.34)
    32-clients     1.00 [  0.00]( 1.03)     0.97 [ -2.53]( 1.92)     0.90 [-10.00]( 1.23)
    64-clients     1.00 [  0.00]( 1.23)     0.97 [ -3.15]( 2.94)     0.90 [ -9.94]( 1.52)
    128-clients    1.00 [  0.00]( 1.14)     0.99 [ -1.07]( 0.95)     0.90 [ -9.91]( 0.90)
    256-clients    1.00 [  0.00]( 3.73)     0.98 [ -1.80]( 3.66)     0.97 [ -3.41]( 4.47)
    512-clients    1.00 [  0.00](54.79)     0.97 [ -3.03](48.98)     0.95 [ -4.63](51.77)


    ==================================================================
    Test          : schbench
    Units         : Normalized 99th percentile latency in us
    Interpretation: Lower is better
    Statistic     : Median
    ==================================================================
    #workers: tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
      1     1.00 [ -0.00](30.14)     1.11 [-11.11](35.78)     1.31 [-30.56](42.87)
      2     1.00 [ -0.00]( 7.87)     0.93 [  7.14]( 8.45)     0.95 [  4.76]( 7.50)
      4     1.00 [ -0.00]( 7.87)     1.07 [ -7.14]( 7.36)     1.14 [-14.29](12.73)
      8     1.00 [ -0.00]( 4.59)     1.08 [ -8.16]( 5.09)     1.12 [-12.24]( 7.44)
     16     1.00 [ -0.00]( 5.33)     1.05 [ -5.08]( 0.93)     1.05 [ -5.08]( 2.75)
     32     1.00 [ -0.00]( 1.04)     1.00 [ -0.00]( 3.12)     1.07 [ -7.29]( 4.49)
     64     1.00 [ -0.00]( 1.04)     0.96 [  3.50]( 3.78)     1.01 [ -1.00]( 2.24)
    128     1.00 [ -0.00]( 5.11)     1.06 [ -6.11]( 7.56)     1.09 [ -8.60]( 6.26)
    256     1.00 [ -0.00](19.39)     1.29 [-28.73](14.92)     1.15 [-14.71](14.83)
    512     1.00 [ -0.00]( 0.15)     0.98 [  2.02]( 1.85)     0.99 [  1.01]( 1.66)


    ==================================================================
    Test          : new-schbench-requests-per-second
    Units         : Normalized Requests per second
    Interpretation: Higher is better
    Statistic     : Median
    ==================================================================
    #workers: tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
      1     1.00 [  0.00]( 0.26)     1.00 [  0.29]( 0.15)     1.00 [ -0.29]( 0.30)
      2     1.00 [  0.00]( 0.00)     1.00 [  0.00]( 0.15)     1.00 [  0.00]( 0.15)
      4     1.00 [  0.00]( 0.00)     1.00 [  0.00]( 0.15)     1.00 [  0.00]( 0.00)
      8     1.00 [  0.00]( 0.15)     1.00 [  0.29]( 0.15)     1.00 [  0.29]( 0.15)
     16     1.00 [  0.00]( 0.00)     1.00 [  0.00]( 0.00)     1.00 [  0.00]( 0.00)
     32     1.00 [  0.00]( 1.86)     1.00 [ -0.31]( 0.28)     1.00 [ -0.31]( 2.12)
     64     1.00 [  0.00](13.62)     0.99 [ -0.77]( 4.78)     0.81 [-18.52](11.11)
    128     1.00 [  0.00]( 0.00)     1.00 [  0.38]( 0.00)     1.00 [  0.38]( 0.00)
    256     1.00 [  0.00]( 1.49)     1.02 [  1.82]( 1.63)     1.00 [  0.00]( 1.19)
    512     1.00 [  0.00]( 0.75)     1.01 [  0.71]( 1.65)     1.01 [  1.19]( 1.53)


    ==================================================================
    Test          : new-schbench-wakeup-latency
    Units         : Normalized 99th percentile latency in us
    Interpretation: Lower is better
    Statistic     : Median
    ==================================================================
    #workers: tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
      1     1.00 [ -0.00]( 6.74)     1.00 [ -0.00]( 6.74)     1.12 [-12.50](19.26)
      2     1.00 [ -0.00](11.18)     1.00 [ -0.00](17.21)     1.50 [-50.00]( 7.45)
      4     1.00 [ -0.00]( 9.94)     1.00 [ -0.00](19.26)     1.56 [-55.56](15.78)
      8     1.00 [ -0.00](10.68)     1.00 [ -0.00](10.68)     1.44 [-44.44](28.77)
     16     1.00 [ -0.00]( 9.68)     1.00 [ -0.00]( 9.68)     1.20 [-20.00]( 8.15)
     32     1.00 [ -0.00](14.08)     1.00 [ -0.00]( 5.34)     1.20 [-20.00](14.70)
     64     1.00 [ -0.00]( 3.52)     1.13 [-13.33]( 5.26)     1.27 [-26.67]( 2.77)
    128     1.00 [ -0.00]( 1.79)     1.07 [ -6.56]( 2.70)     1.07 [ -6.97]( 7.71)
    256     1.00 [ -0.00]( 9.89)     1.04 [ -4.50]( 3.81)     1.02 [ -2.00]( 7.78)
    512     1.00 [ -0.00]( 0.00)     1.01 [ -0.77]( 0.34)     1.00 [ -0.00]( 0.20)


    ==================================================================
    Test          : new-schbench-request-latency
    Units         : Normalized 99th percentile latency in us
    Interpretation: Lower is better
    Statistic     : Median
    ==================================================================
    #workers: tip[pct imp](CV)      proxy-v21[pct imp](CV)   proxy-full[pct imp](CV)
      1     1.00 [ -0.00]( 1.33)     0.96 [  3.89]( 1.46)     1.02 [ -1.82]( 3.02)
      2     1.00 [ -0.00]( 0.14)     1.01 [ -1.09]( 0.24)     1.02 [ -2.44]( 2.73)
      4     1.00 [ -0.00]( 1.24)     1.00 [ -0.26]( 1.69)     0.97 [  2.65]( 0.14)
      8     1.00 [ -0.00]( 0.54)     1.00 [ -0.00]( 1.02)     0.99 [  1.31]( 2.16)
     16     1.00 [ -0.00]( 0.36)     1.00 [ -0.00]( 1.70)     0.98 [  1.59]( 1.00)
     32     1.00 [ -0.00]( 5.51)     0.99 [  0.73]( 2.09)     1.01 [ -1.45]( 7.52)
     64     1.00 [ -0.00]( 5.38)     1.09 [ -9.27]( 0.88)     1.09 [ -8.56]( 0.11)
    128     1.00 [ -0.00]( 0.32)     1.00 [ -0.36]( 0.32)     1.03 [ -2.54]( 1.15)
    256     1.00 [ -0.00](10.51)     1.14 [-14.23](11.19)     1.00 [  0.24](11.42)
    512     1.00 [ -0.00]( 2.00)     1.03 [ -3.27]( 0.94)     1.02 [ -2.41]( 1.96)

Thanks and Regards,
Prateek

Thank you so much for running these test! I really appreciate it! It does look like I need to spend some more work on the full series, as those regressions with the full proxy patch set does seem problematic. Thank you for raising this!

(Thank you also for your feedback on the patches in this series, I've not gotten a chance to address and reply back individually yet, but I hope to do so soon!)

Let me know if you find anything further!

Thank you so much again for all your efforts here, both in testing and review! As well as the testing you do for the whole community! -john