Skip to contentSkip to conversation

swait: Introduce and use swq_has_sleepers

PATCH Superseded Davidlohr Bueso 3 participants 13 messages Sep 5, 2017 – Oct 12, 2017

Subsystems (6)
x86 KVM PowerPC KVM x86 +2 more subsystems

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover swait: Introduce and use swq_has_sleepers

Hi.

Recently[1] Nick mentioned that a lot of swait_active() callers look fishy. This is because it inherited bad habits from regular waitqueues. Other than rcu, kvm is one of the main callers, which I audited.

The following patches fix and/or justify (in baby steps) some of the callers. Note that s390 and mips are other offenders, but I have no idea if it can actually occur -- ie:

CPU0 CPU1
kvm_vcpu_block kvm_mips_comparecount_func

[S] prepare_to_swait(&vcpu->wq)

					  [L] swait_active(&vcpu->wq)
					  [S] queue_timer_int(vcpu)

[L] if (!kvm_vcpu_has_pending_timer(vcpu))
schedule()

[1] swait: add missing barrier to swake_up: https://lkml.org/lkml/2017/9/1/165

Thanks!

Davidlohr Bueso (6):
sched/wait: Add swq_has_sleepers()
kvm,async_pf: Use swq_has_sleepers()
kvm,lapic: Justify use of swait_activate() x86,kvm: Fix apf_task_wake_one() serialization
kvm: Serialize wq active checks in kvm_vcpu_wake_up() kvm,powerpc: Serialize wq active checks in ops->vcpu_kick

 arch/powerpc/kvm/book3s_hv.c |  2 +-
 arch/x86/kernel/kvm.c        |  2 +-
 arch/x86/kvm/lapic.c         |  6 +++++
 include/linux/swait.h        | 57 ++++++++++++++++++++++++++++++++++++++++++--
 virt/kvm/async_pf.c          |  6 +----
 virt/kvm/kvm_main.c          |  2 +-
 6 files changed, 65 insertions(+), 10 deletions(-)
Signature · 1 line
2.12.0
#1 lore raw

PATCH 1/6 sched/wait: Add swq_has_sleepers()

Which is the equivalent of what we have in regular waitqueues. I'm not crazy about the name, but this also helps us get both apis closer -- which iirc comes originally from the -net folks.

We also duplicate the comments for the lockless swait_active(), from wait.h. Future users will make use of this interface.

---
 include/linux/swait.h | 57 +++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 55 insertions(+), 2 deletions(-)
include/linux/swait.h +55 -2
index 4a4e180d0a35..ce4c2582fd42 100644
@@ -79,9 +79,62 @@ extern void __init_swait_queue_head(struct swait_queue_head *q, const char *name
 	DECLARE_SWAIT_QUEUE_HEAD(name)
 #endif
 
-static inline int swait_active(struct swait_queue_head *q)
+/**
+ * swait_active -- locklessly test for waiters on the queue
+ * @wq: the waitqueue to test for waiters
+ *
+ * returns true if the wait list is not empty
+ *
+ * NOTE: this function is lockless and requires care, incorrect usage _will_
+ * lead to sporadic and non-obvious failure.
+ *
+ * NOTE2: this function has the same above implications as regular waitqueues.
+ *
+ * Use either while holding swait_queue_head::lock or when used for wakeups
+ * with an extra smp_mb() like:
+ *
+ *      CPU0 - waker                    CPU1 - waiter
+ *
+ *                                      for (;;) {
+ *      @cond = true;                     prepare_to_wait(&wq_head, &wait, state);
+ *      smp_mb();                         // smp_mb() from set_current_state()
+ *      if (swait_active(wq_head))        if (@cond)
+ *        wake_up(wq_head);                      break;
+ *                                        schedule();
+ *                                      }
+ *                                      finish_wait(&wq_head, &wait);
+ *
+ * Because without the explicit smp_mb() it's possible for the
+ * swait_active() load to get hoisted over the @cond store such that we'll
+ * observe an empty wait list while the waiter might not observe @cond.
+ *
+ * Also note that this 'optimization' trades a spin_lock() for an smp_mb(),
+ * which (when the lock is uncontended) are of roughly equal cost.
+ */
+static inline int swait_active(struct swait_queue_head *wq)
+{
+	return !list_empty(&wq->task_list);
+}
+
+/**
+ * swq_has_sleeper - check if there are any waiting processes
+ * @wq: wait queue head
+ *
+ * Returns true if @wq has waiting processes
+ *
+ * Please refer to the comment for swait_active.
+ */
+static inline bool swq_has_sleeper(struct swait_queue_head *wq)
 {
-	return !list_empty(&q->task_list);
+	/*
+	 * We need to be sure we are in sync with the list_add()
+	 * modifications to the wait queue (task_list).
+	 *
+	 * This memory barrier should be paired with one on the
+	 * waiting side.
+	 */
+	smp_mb();
+	return swait_active(wq);
 }
 
 extern void swake_up(struct swait_queue_head *q);
Signature · 1 line
2.12.0
#2 lore raw

Re: [PATCH 1/6] sched/wait: Add swq_has_sleepers()

Quoted 3 lines · Davidlohr Bueso

Which is the equivalent of what we have in regular waitqueues. I'm not crazy about the name, but this also helps us get both apis closer -- which iirc comes originally from the -net folks.

I'm confused. I don't see anything like this in the original waitqueue.

maintainer
#3 lore raw

Re: [PATCH 1/6] sched/wait: Add swq_has_sleepers()

Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Davidlohr Bueso

Which is the equivalent of what we have in regular waitqueues. I'm not crazy about the name, but this also helps us get both apis closer -- which iirc comes originally from the -net folks.

I'm confused. I don't see anything like this in the original waitqueue.

/**
* wq_has_sleeper - check if there are any waiting processes
* @wq_head: wait queue head
*
* Returns true if wq_head has waiting processes
*
* Please refer to the comment for waitqueue_active.
*/
static inline bool wq_has_sleeper(struct wait_queue_head *wq_head)

{
        /*
         * We need to be sure we are in sync with the
         * add_wait_queue modifications to the wait queue.
         *
         * This memory barrier should be paired with one on the
         * waiting side.
         */
        smp_mb();
        return waitqueue_active(wq_head);
}

Paolo

#4 lore raw

Re: [PATCH 1/6] sched/wait: Add swq_has_sleepers()

Quoted 11 lines · Paolo Bonzini
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Davidlohr Bueso

Which is the equivalent of what we have in regular waitqueues. I'm not crazy about the name, but this also helps us get both apis closer -- which iirc comes originally from the -net folks.

I'm confused. I don't see anything like this in the original waitqueue.

/**
* wq_has_sleeper - check if there are any waiting processes

Bah, clearly my grep-fu isn't strong today :/

#5 lore raw

PATCH 2/6 kvm,async_pf: Use swq_has_sleepers()

In reply to Davidlohr Bueso

... as we've got the new helper now. This caller already does the right thing, hence no changes in semantics.

---
 virt/kvm/async_pf.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
virt/kvm/async_pf.c +1 -5
index bb298a200cd3..57bcb27dcf30 100644
@@ -106,11 +106,7 @@ static void async_pf_execute(struct work_struct *work)
 
 	trace_kvm_async_pf_completed(addr, gva);
 
-	/*
-	 * This memory barrier pairs with prepare_to_wait's set_current_state()
-	 */
-	smp_mb();
-	if (swait_active(&vcpu->wq))
+	if (swq_has_sleeper(&vcpu->wq))
 		swake_up(&vcpu->wq);
 
 	mmput(mm);
Signature · 1 line
2.12.0
#6 lore raw

PATCH 3/6 kvm,lapic: Justify use of swait_activate()

In reply to Davidlohr Bueso

A comment might serve future readers.

---
 arch/x86/kvm/lapic.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/lapic.c +6 -0
index 589dcc117086..fb8be28c7093 100644
@@ -1326,6 +1326,12 @@ static void apic_timer_expired(struct kvm_lapic *apic)
 	atomic_inc(&apic->lapic_timer.pending);
 	kvm_set_pending_timer(vcpu);
 
+	/*
+	 * The above kvm_set_pending_timer implies a wmb
+	 * which pairs with the swaiter side. Either way,
+	 * the atomic_inc() is also serialized so using
+	 * swait_active() is safe.
+	 */
 	if (swait_active(q))
 		swake_up(q);
 
Signature · 1 line
2.12.0
#7 lore raw

PATCH 4/6 x86,kvm: Fix apf_task_wake_one() serialization

In reply to Davidlohr Bueso

During code inspection, the following potential race was seen:

CPU0 CPU1
kvm_async_pf_task_wait apf_task_wake_one
[S] prepare_to_swait(&n.wq)

					  [L] swait_active(&n->wq)
					  [S] hlist_del_init(&n->link);

[L] if (!hlist_unhahed(&n.link))
schedule()

Properly serialize swait_active() checks such that a wakeup is not missed.

---
 arch/x86/kernel/kvm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/kvm.c +1 -1
index 874827b0d7ca..aa60a08b65b1 100644
@@ -180,7 +180,7 @@ static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 	hlist_del_init(&n->link);
 	if (n->halted)
 		smp_send_reschedule(n->cpu);
-	else if (swait_active(&n->wq))
+	else if (swq_has_sleeper(&n->wq))
 		swake_up(&n->wq);
 }
 
Signature · 1 line
2.12.0
maintainer
#8 lore raw

Re: [PATCH 4/6] x86,kvm: Fix apf_task_wake_one() serialization

Quoted 23 lines · Davidlohr Bueso

During code inspection, the following potential race was seen:

CPU0 CPU1
kvm_async_pf_task_wait apf_task_wake_one
[S] prepare_to_swait(&n.wq)

					  [L] swait_active(&n->wq)
					  [S] hlist_del_init(&n->link);

[L] if (!hlist_unhahed(&n.link))
schedule()

Properly serialize swait_active() checks such that a wakeup is not missed.

---
 arch/x86/kernel/kvm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/kvm.c +0 -0
index 874827b0d7ca..aa60a08b65b1 100644

arch/x86/kernel/kvm.c static void apf_task_wake_one(struct kvm_task_sleep_node *n) · line 185

Quoted 7 lines
arch/x86/kernel/kvm.c +1 -1
@@ -180,7 +180,7 @@ static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 	hlist_del_init(&n->link);
 	if (n->halted)
 		smp_send_reschedule(n->cpu);
-	else if (swait_active(&n->wq))
+	else if (swq_has_sleeper(&n->wq))
 		swake_up(&n->wq);
 }

After Nick's patch, swake_up starts with:

	smp_mb();
        if (!swait_active(q))
                return;

so we can just remove the test here (and in patch 2).

The other patches could also use a better swait API, for example:

1) add a public __swake_up routine that omits the memory barrier, and which can be used in patch 3. Perhaps better: omit the out-of-lock check in __swake_up: then the caller can use it if it knows there is a waiter. In those cases the memory barrier is expensive.

2) change swake_up and __swake_up to return true if they woke up a process (or alternatively 0/-EAGAIN). Patches 5 and 6 now need not call anymore either swq_has_sleepers or swait_active, and that saves a memory barrier too.

What do you think?

Thanks,

Paolo

maintainer
#9 lore raw

Re: [PATCH 4/6] x86,kvm: Fix apf_task_wake_one() serialization

Quoted 47 lines · Paolo Bonzini
Quoted 23 lines · Davidlohr Bueso

During code inspection, the following potential race was seen:

CPU0 CPU1
kvm_async_pf_task_wait apf_task_wake_one
[S] prepare_to_swait(&n.wq)

					  [L] swait_active(&n->wq)
					  [S] hlist_del_init(&n->link);

[L] if (!hlist_unhahed(&n.link))
schedule()

Properly serialize swait_active() checks such that a wakeup is not missed.

---
 arch/x86/kernel/kvm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/kvm.c +0 -0
index 874827b0d7ca..aa60a08b65b1 100644

arch/x86/kernel/kvm.c static void apf_task_wake_one(struct kvm_task_sleep_node *n) · line 185

Quoted 7 lines
arch/x86/kernel/kvm.c +1 -1
@@ -180,7 +180,7 @@ static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 	hlist_del_init(&n->link);
 	if (n->halted)
 		smp_send_reschedule(n->cpu);
-	else if (swait_active(&n->wq))
+	else if (swq_has_sleeper(&n->wq))
 		swake_up(&n->wq);
 }

After Nick's patch, swake_up starts with:

	smp_mb();
        if (!swait_active(q))
                return;

so we can just remove the test here (and in patch 2).

The other patches could also use a better swait API, for example:

1) add a public __swake_up routine that omits the memory barrier, and which can be used in patch 3. Perhaps better: omit the out-of-lock check in __swake_up: then the caller can use it if it knows there is a waiter. In those cases the memory barrier is expensive.

Quoted 6 lines

2) change swake_up and __swake_up to return true if they woke up a process (or alternatively 0/-EAGAIN). Patches 5 and 6 now need not call anymore either swq_has_sleepers or swait_active, and that saves a memory barrier too.

What do you think?

... doh, I missed PeterZ's remark that the early test is gone in tip. Then the series makes total sense. Peter, if you ack patch 1 I can push it through the KVM tree.

Paolo

#10 lore raw

Re: [PATCH 4/6] x86,kvm: Fix apf_task_wake_one() serialization

Quoted 3 lines · Paolo Bonzini

... doh, I missed PeterZ's remark that the early test is gone in tip. Then the series makes total sense. Peter, if you ack patch 1 I can push it through the KVM tree.

I see its in already, but Ack ;-)

#11 lore raw

PATCH 5/6 kvm: Serialize wq active checks in kvm_vcpu_wake_up()

In reply to Davidlohr Bueso

This is a generic call and can be suceptible to races in reading the wq task_list while another task is adding itself to the list. Add a full barrier by using the swq_has_sleeper() helper.

---
 virt/kvm/kvm_main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
virt/kvm/kvm_main.c +1 -1
index 4d81f6ded88e..4e76e6d15ce0 100644
@@ -2185,7 +2185,7 @@ bool kvm_vcpu_wake_up(struct kvm_vcpu *vcpu)
 	struct swait_queue_head *wqp;
 
 	wqp = kvm_arch_vcpu_wq(vcpu);
-	if (swait_active(wqp)) {
+	if (swq_has_sleeper(wqp)) {
 		swake_up(wqp);
 		++vcpu->stat.halt_wakeup;
 		return true;
Signature · 1 line
2.12.0
#12 lore raw

PATCH 6/6 kvm,powerpc: Serialize wq active checks in ops->vcpu_kick

In reply to Davidlohr Bueso

Particularly because kvmppc_fast_vcpu_kick_hv() is a callback, ensure that we properly serialize wq active checks in order to avoid potentially missing a wakeup due to racing with the waiter side.

---
 arch/powerpc/kvm/book3s_hv.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/kvm/book3s_hv.c +1 -1
index 359c79cdf0cc..7ffc757ef978 100644
@@ -181,7 +181,7 @@ static void kvmppc_fast_vcpu_kick_hv(struct kvm_vcpu *vcpu)
 	struct swait_queue_head *wqp;
 
 	wqp = kvm_arch_vcpu_wq(vcpu);
-	if (swait_active(wqp)) {
+	if (swq_has_sleeper(wqp)) {
 		swake_up(wqp);
 		++vcpu->stat.halt_wakeup;
 	}
Signature · 1 line
2.12.0