KVM: PPC: divide the ics lock into smaller ones for each irq

RFC PATCH In review Li Zhong 2 participants 6 messages May 16, 2016 – Jun 21, 2016

Subsystems (2)

PowerPC KVM PowerPC

lore

Share & tools

lore

Li Zhong

author

2016-05-16 06:58 UTC #0 lore raw

RFC PATCH 1/2 KVM: PPC: divide the ics lock into smaller ones for each irq

This patch tries to use smaller locks for each irq in the ics, instead of a lock at the ics level, to provide better scalability.

Signed-off-by Li Zhong <[email protected]>

---
 arch/powerpc/kvm/book3s_hv_rm_xics.c | 24 ++++++++++-----------
 arch/powerpc/kvm/book3s_xics.c       | 41 +++++++++++++++++++-----------------
 arch/powerpc/kvm/book3s_xics.h       |  2 +-
 3 files changed, 34 insertions(+), 33 deletions(-)

arch/powerpc/kvm/book3s_hv_rm_xics.c +11 ~~-13~~

index 980d8a6..c3f604e 100644
@@ -36,20 +36,19 @@ static void ics_rm_check_resend(struct kvmppc_xics *xics,
 {
 	int i;
 
-	arch_spin_lock(&ics->lock);
-
 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 		struct ics_irq_state *state = &ics->irq_state[i];
 
-		if (!state->resend)
+		arch_spin_lock(&state->lock);
+
+		if (!state->resend) {
+			arch_spin_unlock(&state->lock);
 			continue;
+		}
 
-		arch_spin_unlock(&ics->lock);
+		arch_spin_unlock(&state->lock);
 		icp_rm_deliver_irq(xics, icp, state->number);
-		arch_spin_lock(&ics->lock);
 	}
-
-	arch_spin_unlock(&ics->lock);
 }
 
 /* -- ICP routines -- */

@@ -310,8 +309,7 @@ static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 	}
 	state = &ics->irq_state[src];
 
-	/* Get a lock on the ICS */
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&state->lock);
 
 	/* Get our server */
 	if (!icp || state->server != icp->server_num) {

@@ -367,7 +365,7 @@ static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 * Delivery was successful, did we reject somebody else ?
 		 */
 		if (reject && reject != XICS_IPI) {
-			arch_spin_unlock(&ics->lock);
+			arch_spin_unlock(&state->lock);
 			new_irq = reject;
 			goto again;
 		}

@@ -387,12 +385,12 @@ static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 */
 		smp_mb();
 		if (!icp->state.need_resend) {
-			arch_spin_unlock(&ics->lock);
+			arch_spin_unlock(&state->lock);
 			goto again;
 		}
 	}
- out:
-	arch_spin_unlock(&ics->lock);
+out:
+	arch_spin_unlock(&state->lock);
 }
 
 static void icp_rm_down_cppr(struct kvmppc_xics *xics, struct kvmppc_icp *icp,

arch/powerpc/kvm/book3s_xics.c +21 ~~-19~~

index 46871d5..b9dbf75 100644
@@ -113,25 +113,26 @@ static void ics_check_resend(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 	unsigned long flags;
 
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
 
 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 		struct ics_irq_state *state = &ics->irq_state[i];
 
-		if (!state->resend)
+		arch_spin_lock(&state->lock);
+
+		if (!state->resend) {
+			arch_spin_unlock(&state->lock);
 			continue;
+		}
 
 		XICS_DBG("resend %#x prio %#x\n", state->number,
 			      state->priority);
 
-		arch_spin_unlock(&ics->lock);
+		arch_spin_unlock(&state->lock);
 		local_irq_restore(flags);
 		icp_deliver_irq(xics, icp, state->number);
 		local_irq_save(flags);
-		arch_spin_lock(&ics->lock);
 	}
 
-	arch_spin_unlock(&ics->lock);
 	local_irq_restore(flags);
 }

@@ -143,7 +144,7 @@ static bool write_xive(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 	unsigned long flags;
 
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&state->lock);
 
 	state->server = server;
 	state->priority = priority;

@@ -154,7 +155,7 @@ static bool write_xive(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 		deliver = true;
 	}
 
-	arch_spin_unlock(&ics->lock);
+	arch_spin_unlock(&state->lock);
 	local_irq_restore(flags);
 
 	return deliver;

@@ -207,10 +208,10 @@ int kvmppc_xics_get_xive(struct kvm *kvm, u32 irq, u32 *server, u32 *priority)
 	state = &ics->irq_state[src];
 
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&state->lock);
 	*server = state->server;
 	*priority = state->priority;
-	arch_spin_unlock(&ics->lock);
+	arch_spin_unlock(&state->lock);
 	local_irq_restore(flags);
 
 	return 0;

@@ -406,7 +407,7 @@ static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 
 	/* Get a lock on the ICS */
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&state->lock);
 
 	/* Get our server */
 	if (!icp || state->server != icp->server_num) {

@@ -463,7 +464,7 @@ static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 * Delivery was successful, did we reject somebody else ?
 		 */
 		if (reject && reject != XICS_IPI) {
-			arch_spin_unlock(&ics->lock);
+			arch_spin_unlock(&state->lock);
 			local_irq_restore(flags);
 			new_irq = reject;
 			goto again;

@@ -484,13 +485,13 @@ static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 */
 		smp_mb();
 		if (!icp->state.need_resend) {
-			arch_spin_unlock(&ics->lock);
+			arch_spin_unlock(&state->lock);
 			local_irq_restore(flags);
 			goto again;
 		}
 	}
  out:
-	arch_spin_unlock(&ics->lock);
+	arch_spin_unlock(&state->lock);
 	local_irq_restore(flags);
 }

@@ -950,18 +951,18 @@ static int xics_debug_show(struct seq_file *m, void *private)
 			   icsid);
 
 		local_irq_save(flags);
-		arch_spin_lock(&ics->lock);
 
 		for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 			struct ics_irq_state *irq = &ics->irq_state[i];
+			arch_spin_lock(&irq->lock);
 
 			seq_printf(m, "irq 0x%06x: server %#x prio %#x save prio %#x asserted %d resend %d masked pending %d\n",
 				   irq->number, irq->server, irq->priority,
 				   irq->saved_priority, irq->asserted,
 				   irq->resend, irq->masked_pending);
+			arch_spin_unlock(&irq->lock);
 
 		}
-		arch_spin_unlock(&ics->lock);
 		local_irq_restore(flags);
 	}
 	return 0;

@@ -1021,6 +1022,8 @@ static struct kvmppc_ics *kvmppc_xics_create_ics(struct kvm *kvm,
 		ics->irq_state[i].number = (icsid << KVMPPC_XICS_ICS_SHIFT) | i;
 		ics->irq_state[i].priority = MASKED;
 		ics->irq_state[i].saved_priority = MASKED;
+		ics->irq_state[i].lock +			(arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
 	}
 	smp_wmb();
 	xics->ics[icsid] = ics;

@@ -1164,7 +1167,7 @@ static int xics_get_source(struct kvmppc_xics *xics, long irq, u64 addr)
 
 	irqp = &ics->irq_state[idx];
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&irqp->lock);
 	ret = -ENOENT;
 	if (irqp->exists) {
 		val = irqp->server;

@@ -1180,7 +1183,7 @@ static int xics_get_source(struct kvmppc_xics *xics, long irq, u64 addr)
 			val |= KVM_XICS_PENDING;
 		ret = 0;
 	}
-	arch_spin_unlock(&ics->lock);
+	arch_spin_unlock(&irqp->lock);
 	local_irq_restore(flags);
 
 	if (!ret && put_user(val, ubufp))

@@ -1220,7 +1223,7 @@ static int xics_set_source(struct kvmppc_xics *xics, long irq, u64 addr)
 		return -EINVAL;
 
 	local_irq_save(flags);
-	arch_spin_lock(&ics->lock);
+	arch_spin_lock(&irqp->lock);
 	irqp->server = server;
 	irqp->saved_priority = prio;
 	if (val & KVM_XICS_MASKED)

@@ -1232,7 +1235,7 @@ static int xics_set_source(struct kvmppc_xics *xics, long irq, u64 addr)
 	if ((val & KVM_XICS_PENDING) && (val & KVM_XICS_LEVEL_SENSITIVE))
 		irqp->asserted = 1;
 	irqp->exists = 1;
-	arch_spin_unlock(&ics->lock);
+	arch_spin_unlock(&irqp->lock);
 	local_irq_restore(flags);
 
 	if (val & KVM_XICS_PENDING)

arch/powerpc/kvm/book3s_xics.h +1 -1

index 56ea44f..d30564b 100644
@@ -33,6 +33,7 @@
 
 /* State for one irq source */
 struct ics_irq_state {
+	arch_spinlock_t lock;
 	u32 number;
 	u32 server;
 	u8  priority;

@@ -93,7 +94,6 @@ struct kvmppc_icp {
 };
 
 struct kvmppc_ics {
-	arch_spinlock_t lock;
 	u16 icsid;
 	struct ics_irq_state irq_state[KVMPPC_XICS_IRQ_PER_ICS];
 };

Signature · 1 line

1.8.3.1

Li Zhong

author

2016-05-16 07:02 UTC #1 lore raw

RFC PATCH 2/2 KVM: PPC: Don't take lock when check irq's resend flag

It seems that we don't need to take the lock before evaluating irq's resend flag. What needed is to make sure when we clear the ics's bit in the icp's resend_map, we don't miss the resend flag of the irqs that set the bit.

And seems this could be ordered through the barrier in test_and_clear_bit(), and an newly added wmb when setting irq's resend flag, and icp's resend_map.

Signed-off-by Li Zhong <[email protected]>

---
 arch/powerpc/kvm/book3s_hv_rm_xics.c | 14 +++++++-------
 arch/powerpc/kvm/book3s_xics.c       | 22 +++++++---------------
 2 files changed, 14 insertions(+), 22 deletions(-)

arch/powerpc/kvm/book3s_hv_rm_xics.c +7 -7

index c3f604e..c3fa386 100644
@@ -39,14 +39,9 @@ static void ics_rm_check_resend(struct kvmppc_xics *xics,
 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 		struct ics_irq_state *state = &ics->irq_state[i];
 
-		arch_spin_lock(&state->lock);
-
-		if (!state->resend) {
-			arch_spin_unlock(&state->lock);
+		if (!state->resend)
 			continue;
-		}
 
-		arch_spin_unlock(&state->lock);
 		icp_rm_deliver_irq(xics, icp, state->number);
 	}
 }

@@ -374,8 +369,13 @@ static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 * We failed to deliver the interrupt we need to set the
 		 * resend map bit and mark the ICS state as needing a resend
 		 */
-		set_bit(ics->icsid, icp->resend_map);
 		state->resend = 1;
+		/*
+		 * Make sure when checking resend, we don't miss the resend if
+		 * resend_map bit is seen and cleared.
+		 */
+		smp_wmb();
+		set_bit(ics->icsid, icp->resend_map);
 
 		/*
 		 * If the need_resend flag got cleared in the ICP some time

arch/powerpc/kvm/book3s_xics.c +7 ~~-15~~

index b9dbf75..420c4fc 100644
@@ -110,30 +110,17 @@ static void ics_check_resend(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 {
 	int i;
 
-	unsigned long flags;
-
-	local_irq_save(flags);
-
 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 		struct ics_irq_state *state = &ics->irq_state[i];
 
-		arch_spin_lock(&state->lock);
-
-		if (!state->resend) {
-			arch_spin_unlock(&state->lock);
+		if (!state->resend)
 			continue;
-		}
 
 		XICS_DBG("resend %#x prio %#x\n", state->number,
 			      state->priority);
 
-		arch_spin_unlock(&state->lock);
-		local_irq_restore(flags);
 		icp_deliver_irq(xics, icp, state->number);
-		local_irq_save(flags);
 	}
-
-	local_irq_restore(flags);
 }
 
 static bool write_xive(struct kvmppc_xics *xics, struct kvmppc_ics *ics,

@@ -474,8 +461,13 @@ static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 		 * We failed to deliver the interrupt we need to set the
 		 * resend map bit and mark the ICS state as needing a resend
 		 */
-		set_bit(ics->icsid, icp->resend_map);
 		state->resend = 1;
+		/*
+		 * Make sure when checking resend, we don't miss the resend if
+		 * resend_map bit is seen and cleared.
+		 */
+		smp_wmb();
+		set_bit(ics->icsid, icp->resend_map);
 
 		/*
 		 * If the need_resend flag got cleared in the ICP some time

Signature · 1 line

1.8.3.1

Paul Mackerras

2016-06-20 05:27 UTC #2 lore raw

Re: [RFC PATCH 2/2] KVM: PPC: Don't take lock when check irq's resend flag

Quoted 7 lines · Li Zhong

It seems that we don't need to take the lock before evaluating irq's resend flag. What needed is to make sure when we clear the ics's bit in the icp's resend_map, we don't miss the resend flag of the irqs that set the bit.
And seems this could be ordered through the barrier in test_and_clear_bit(), and an newly added wmb when setting irq's resend flag, and icp's resend_map.

This looks fine to me. Is there a measurable performance improvement from this? I understand it could be hard to measure.

Also, you could make the patch description more definite - just say that we don't need to take the lock, there's no need for "seems".

Paul.

Li Zhong

author

2016-06-21 02:57 UTC #3 lore raw

Re: [RFC PATCH 2/2] KVM: PPC: Don't take lock when check irq's resend flag

Quoted 9 lines · Paul Mackerras

Quoted 5 lines · Li Zhong
It seems that we don't need to take the lock before evaluating irq's resend flag. What needed is to make sure when we clear the ics's bit in the icp's resend_map, we don't miss the resend flag of the irqs that set the bit.

Quoted 8 lines

Quoted 2 lines
And seems this could be ordered through the barrier in test_and_clear_bit(), and an newly added wmb when setting irq's resend flag, and icp's resend_map.
This looks fine to me. Is there a measurable performance improvement from this? I understand it could be hard to measure.
Also, you could make the patch description more definite - just say that we don't need to take the lock, there's no need for "seems”.

OK :)

However, we may need to ignore this one for now. To implement the P/Q stuff, we probably need make sure the resend irqs to be resent only once. It’s easier to make sure that with the lock here, and the resend flag can be cleared inside the lock.

Thanks, Zhong

Quoted 6 lines

Paul.
_______________________________________________
Linuxppc-dev mailing list
[email protected]
https://lists.ozlabs.org/listinfo/linuxppc-dev

Paul Mackerras

2016-06-20 05:25 UTC #4 lore raw

In reply to Li Zhong

Quoted 2 lines · Li Zhong

This patch tries to use smaller locks for each irq in the ics, instead of a lock at the ics level, to provide better scalability.

This looks like a worth-while thing to do. Do you have any performance measurements to justify the change? This will increase the size of struct kvmppc_ics by 4kB, so it would be useful to show the performance increase that justifies it.

Also, when you resend the patch, please make the patch description more definite - say "With this patch, we use" rather than "this patch tries to use", for instance.

Regards,
Paul.

Li Zhong

author

2016-06-21 02:47 UTC #5 lore raw

On Jun 20, 2016, at = 13:25, Paul Mackerras <[email protected]> wrote:

On Mon, = May 16, 2016 at 02:58:18PM +0800, Li Zhong wrote:
This patch tries to use smaller locks for each irq in the = ics, instead of a lock at the ics level, to provide better = scalability.

This looks like a worth-while thing to = do. Do you have any performance measurements to justify the = change? This will increase the size of struct kvmppc_ics by = 4kB, so it would be useful to show the performance increase that = justifies it.

Actually, I saw some =E2=80=9Cimprovemen= t=E2=80=9D because of the vcpus were not binded, io jobs and irqs on the = guest were not binded. After I fixed those random factors, the result = became stable, but I couldn=E2=80=99t see any obvious improvements from = the patches...

Maybe I need find some other test cases that = could support this change.

Also, = when you resend the patch, please make the patch description more = definite - say "With this patch, we use" rather than "this = patch tries to use", for instance.

OK, I will = change that when doing a resend, if I can find some workload that could = benefit from this change.

Thanks, Zhong

Regards,
Paul.
___________________________________= ____________
Linuxppc-dev mailing = list
[email protected]
https://lists.ozlabs.org/listinfo= /linuxppc-dev