Skip to contentSkip to conversation

locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set

PATCH In review Eduardo Valentin 6 participants 27 messages Nov 6, 2017 – Nov 27, 2017

Subsystems (5)
Documentation x86 KVM KVM x86 +1 more subsystems
lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;
 	/*
 	 * On hypervisors without PARAVIRT_SPINLOCKS support we fall
 	 * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME		5
 #define KVM_FEATURE_PV_EOI		6
 #define KVM_FEATURE_PV_UNHALT		7
+#define KVM_FEATURE_PV_DEDICATED	8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -0
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
 	if (!kvm_para_available())
 		return;
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return;
 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 		return;
Signature · 1 line
2.7.4
maintainer
#1 lore raw
Quoted 10 lines · Eduardo Valentin

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

Quoted 7 lines

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Hi Eduardo,

besides the suggestion to use a separate word than the one for features, is this still needed after Waiman's patch to adaptively switch between tas and pvqspinlock?

Paolo

Quoted 95 lines
Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;
 	/*
 	 * On hypervisors without PARAVIRT_SPINLOCKS support we fall
 	 * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME		5
 #define KVM_FEATURE_PV_EOI		6
 #define KVM_FEATURE_PV_UNHALT		7
+#define KVM_FEATURE_PV_DEDICATED	8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -0
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
 	if (!kvm_para_available())
 		return;
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return;
 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 		return;
#2 lore raw
Quoted 15 lines · Paolo Bonzini
Quoted 14 lines · Eduardo Valentin

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

Quoted 7 lines
Quoted 3 lines

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Hi Eduardo,

besides the suggestion to use a separate word than the one for features,

Ok. I will take a look.

Quoted 2 lines

is this still needed after Waiman's patch to adaptively switch between tas and pvqspinlock?

Can you please point me to it ? Is it already in tip/master?

Quoted 100 lines

Paolo

Quoted 95 lines
Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;
 	/*
 	 * On hypervisors without PARAVIRT_SPINLOCKS support we fall
 	 * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME		5
 #define KVM_FEATURE_PV_EOI		6
 #define KVM_FEATURE_PV_UNHALT		7
+#define KVM_FEATURE_PV_DEDICATED	8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -0
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
 	if (!kvm_para_available())
 		return;
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return;
 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 		return;
Signature · 2 lines
All the best,
Eduardo Valentin
#4 lore raw

Paolo,

Quoted 9 lines · Paolo Bonzini
Quoted 4 lines · Eduardo Valentin
Quoted 2 lines

is this still needed after Waiman's patch to adaptively switch between tas and pvqspinlock?

Can you please point me to it ? Is it already in tip/master?

No, he just posted it:

https://marc.info/?l=linux-kernel&m=150972337909996&w=2

OK, Thanks. I've reviewed his V2. I think the patch to have PV_DEDICATED is still interesting to have, for the case the guest is in autoselect mode and honors what the host gives as hints.

Quoted 2 lines

Paolo

Signature · 2 lines
All the best,
Eduardo Valentin
#5 lore raw
In reply to Eduardo Valentin

2017-11-06 12:26-0800, Eduardo Valentin:

Quoted 79 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +2 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>

arch/x86/include/asm/qspinlock.h static inline bool virt_spin_lock(struct qspinlock *lock) · line 64

Quoted 5 lines
arch/x86/include/asm/qspinlock.h +2 -0
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;

Hm, every spinlock slowpath calls cpuid, which causes a VM exit, so I wouldn't expect it to be faster than the existing implementations. (Using the static key would be better.)

How does this patch perform compared to user-forced qspinlock and hybrid pvqspinlock?

Thanks.

#6 lore raw

Hello,

Quoted 90 lines · Radim Krčmář

2017-11-06 12:26-0800, Eduardo Valentin:

Quoted 79 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +2 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>

arch/x86/include/asm/qspinlock.h static inline bool virt_spin_lock(struct qspinlock *lock) · line 64

Quoted 5 lines
arch/x86/include/asm/qspinlock.h +2 -0
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;

Hm, every spinlock slowpath calls cpuid, which causes a VM exit, so I wouldn't expect it to be faster than the existing implementations. (Using the static key would be better.)

Quoted 2 lines

How does this patch perform compared to user-forced qspinlock and hybrid pvqspinlock?

This patch should have same effect as user-forced qspinlock. However, the key aspect here is this patch gives a way for the host to instruct the guest to use qspinlock. Even with Longman's patch which allows guest to select the spinlock implementation, there should still be the auto-select mode. In such mode, PV_DEDICATED should allow the host to get the guest to use qspinlock, without, the guest will fallback to tas when PV_UNHALT == 0.

Quoted 3 lines

Thanks.

Signature · 2 lines
All the best,
Eduardo Valentin
#7 lore raw

2017-11-09 00:55-0800, Eduardo Valentin:

Quoted 93 lines

Hello,

Quoted 90 lines · Radim Krčmář

2017-11-06 12:26-0800, Eduardo Valentin:

Quoted 79 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +2 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>

arch/x86/include/asm/qspinlock.h static inline bool virt_spin_lock(struct qspinlock *lock) · line 64

Quoted 5 lines
arch/x86/include/asm/qspinlock.h +2 -0
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+		return false;

Hm, every spinlock slowpath calls cpuid, which causes a VM exit, so I wouldn't expect it to be faster than the existing implementations. (Using the static key would be better.)

Quoted 4 lines
Quoted 2 lines

How does this patch perform compared to user-forced qspinlock and hybrid pvqspinlock?

This patch should have same effect as user-forced qspinlock.

This is what I'm doubting, because the patch is adding about two thousand cycles to every spinlock-taken path.
Doesn't this patch yield better results?

arch/x86/kernel/kvm.c +6 -0
index 3df743b60c80..d9225e48c11a 100644
@@ -676,6 +676,12 @@ void __init kvm_spinlock_init(void)
 {
 	if (!kvm_para_available())
 		return;
+
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED)) {
+		static_branch_disable(&virt_spin_lock_key);
+		return;
+	}
+
 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 		return;
Quoted 6 lines

However, the key aspect
here is this patch gives a way for the host to instruct the guest to use qspinlock. Even with Longman's patch which allows guest to select the spinlock implementation, there should still be the auto-select mode. In such mode, PV_DEDICATED should allow the host to get the guest to use qspinlock, without, the guest will fallback to tas when PV_UNHALT == 0.

I agree that a flag can be useful for certains setups.

#8 lore raw

Hey Radim,

On Thu, Nov 09, 2017 at 03:17:33PM +0100, Radim Krčmář wrote:

<cut>

Quoted 10 lines

This is what I'm doubting, because the patch is adding about two thousand cycles to every spinlock-taken path.
Doesn't this patch yield better results?

arch/x86/kernel/kvm.c +0 -0
index 3df743b60c80..d9225e48c11a 100644

arch/x86/kernel/kvm.c void __init kvm_spinlock_init(void) · line 684

Quoted 9 lines
arch/x86/kernel/kvm.c +6 -0
@@ -676,6 +676,12 @@ void __init kvm_spinlock_init(void)
 {
 	if (!kvm_para_available())
 		return;
+
+	if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED)) {
+		static_branch_disable(&virt_spin_lock_key);
+		return;
+	}
+

Yes, the above suggestion is a much better approach. The code has probably changed from the time I wrote the first version. I will refresh with the above suggestion.

Quoted 4 lines
arch/x86/kernel/kvm.c +0 -0
 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 		return;
Quoted 8 lines
Quoted 6 lines

However, the key aspect
here is this patch gives a way for the host to instruct the guest to use qspinlock. Even with Longman's patch which allows guest to select the spinlock implementation, there should still be the auto-select mode. In such mode, PV_DEDICATED should allow the host to get the guest to use qspinlock, without, the guest will fallback to tas when PV_UNHALT == 0.

I agree that a flag can be useful for certains setups.

Cool!

Quoted 1 line
Signature · 2 lines
All the best,
Eduardo Valentin
#9 lore raw

Hi Eduardo,
2017-11-16 12:54 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 11 lines

Hey Radim,

On Thu, Nov 09, 2017 at 03:17:33PM +0100, Radim Krčmář wrote:

<cut>

Quoted 5 lines

This is what I'm doubting, because the patch is adding about two thousand cycles to every spinlock-taken path.
Doesn't this patch yield better results?

Quoted 16 lines
Quoted 5 lines
arch/x86/kernel/kvm.c +0 -0
index 3df743b60c80..d9225e48c11a 100644

arch/x86/kernel/kvm.c void __init kvm_spinlock_init(void) · line 684

Quoted 9 lines
arch/x86/kernel/kvm.c +6 -0
@@ -676,6 +676,12 @@ void __init kvm_spinlock_init(void)
 {
      if (!kvm_para_available())
              return;
+
+     if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED)) {
+             static_branch_disable(&virt_spin_lock_key);
+             return;
+     }
+

Yes, the above suggestion is a much better approach. The code has probably changed from the time I wrote the first version. I will refresh with the above suggestion.

Do you mind to send a new version since the merge window is closed?

Regards,
Wanpeng Li

Quoted 22 lines
Quoted 6 lines
arch/x86/kernel/kvm.c +0 -0
      /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
      if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
              return;
Quoted 8 lines
Quoted 6 lines

However, the key aspect
here is this patch gives a way for the host to instruct the guest to use qspinlock. Even with Longman's patch which allows guest to select the spinlock implementation, there should still be the auto-select mode. In such mode, PV_DEDICATED should allow the host to get the guest to use qspinlock, without, the guest will fallback to tas when PV_UNHALT == 0.

I agree that a flag can be useful for certains setups.

Cool!

Quoted 1 line

--
All the best,
Eduardo Valentin

#10 lore raw
In reply to Eduardo Valentin

2017-11-07 4:26 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 113 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +6 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest checks this feature bit
                                    ||       || before enabling paravirtualized
                                    ||       || spinlock support.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED           ||     8 || guest checks this feature bit
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing opti-
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
        if (!static_branch_likely(&virt_spin_lock_key))
                return false;
 
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return false;
        /*
         * On hypervisors without PARAVIRT_SPINLOCKS support we fall
         * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME         5
 #define KVM_FEATURE_PV_EOI             6
 #define KVM_FEATURE_PV_UNHALT          7
+#define KVM_FEATURE_PV_DEDICATED       8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -1
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
        if (!kvm_para_available())
                return;
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return;
        /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
        if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
                return;
--
Quoted 2 lines

2.7.4

You should also add a cpuid flag in kvm part.

Regards,
Wanpeng Li

#11 lore raw
Quoted 121 lines

2017-11-07 4:26 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 118 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +0 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest

checks this feature bit

                                    ||       || before enabling
                                    ||       || paravirtualized
                                    ||       || spinlock support.

------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED || 8 || guest checks this feature
bit

Documentation/virtual/kvm/cpuid.txt +2 -0
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing

opti-

Documentation/virtual/kvm/cpuid.txt +3 -0
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no
 guest-side
                                    ||       || per-cpu warps are expected
                                    ||       || in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
        if (!static_branch_likely(&virt_spin_lock_key))
                return false;
 
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return false;
        /*
         * On hypervisors without PARAVIRT_SPINLOCKS support we fall
         * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME         5
 #define KVM_FEATURE_PV_EOI             6
 #define KVM_FEATURE_PV_UNHALT          7
+#define KVM_FEATURE_PV_DEDICATED       8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -0
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
        if (!kvm_para_available())
                return;
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return;
Quoted 8 lines
Quoted 6 lines
arch/x86/kernel/kvm.c +0 -1
        /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
        if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
                return;
--

2.7.4

You should also add a cpuid flag in kvm part.

Also, I am thinking if PV_DEDICATED helps in performance and with conjunction with PV TLB patch in other thread. For use-case e.g KVM-RT where we don't overcommit vCPU's and pin vCPU:pCPU 1:1 we need a way from host side with which user can decide to enable PV_DEDICATED option. Such that if vCPU's are unlikely going to preempt or sleep we should avoid traversing the cpulist in PV TLB code.

So, two things:

1] A way to configure PV_DEDICATED from host.
2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Or I am missing any context here?

Quoted 4 lines

Regards,
Wanpeng Li

#12 lore raw

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 136 lines
Quoted 121 lines

2017-11-07 4:26 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 118 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---
V3:
- When PV_DEDICATED is set (1), qspinlock is selected, regardless of the value of PV_UNHAULT. Suggested by Paolo Bonzini.
- Refreshed on top of tip/master.
V2:
- rebase on top of tip/master

 Documentation/virtual/kvm/cpuid.txt  | 6 ++++++
 arch/x86/include/asm/qspinlock.h     | 4 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 1 +
 arch/x86/kernel/kvm.c                | 2 ++
 4 files changed, 13 insertions(+)
Documentation/virtual/kvm/cpuid.txt +0 -0
index 3c65feb..117066a 100644
@@ -54,6 +54,12 @@ KVM_FEATURE_PV_UNHALT              ||     7 || guest

checks this feature bit

                                    ||       || before enabling
                                    ||       || paravirtualized
                                    ||       || spinlock support.

------------------------------------------------------------------------------
+KVM_FEATURE_PV_DEDICATED || 8 || guest checks this feature
bit

Documentation/virtual/kvm/cpuid.txt +2 -0
+                                   ||       || to determine if they run on
+                                   ||       || dedicated vCPUs, allowing

opti-

Documentation/virtual/kvm/cpuid.txt +3 -0
+                                   ||       || mizations such as usage of
+                                   ||       || qspinlocks.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no
 guest-side
                                    ||       || per-cpu warps are expected
                                    ||       || in
                                    ||       || kvmclock.
arch/x86/include/asm/qspinlock.h +4 -0
index 5e16b5d..de42694 100644
@@ -3,6 +3,8 @@
 #define _ASM_X86_QSPINLOCK_H
 
 #include <linux/jump_label.h>
+#include <linux/kvm_para.h>
+
 #include <asm/cpufeature.h>
 #include <asm-generic/qspinlock_types.h>
 #include <asm/paravirt.h>
@@ -58,6 +60,8 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
        if (!static_branch_likely(&virt_spin_lock_key))
                return false;
 
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return false;
        /*
         * On hypervisors without PARAVIRT_SPINLOCKS support we fall
         * back to a Test-and-Set spinlock, because fair locks have
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 554aa8f..85a9875 100644
@@ -25,6 +25,7 @@
 #define KVM_FEATURE_STEAL_TIME         5
 #define KVM_FEATURE_PV_EOI             6
 #define KVM_FEATURE_PV_UNHALT          7
+#define KVM_FEATURE_PV_DEDICATED       8
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
arch/x86/kernel/kvm.c +2 -0
index 8bb9594..dacd7cf 100644
@@ -642,6 +642,8 @@ void __init kvm_spinlock_init(void)
 {
        if (!kvm_para_available())
                return;
+       if (kvm_para_has_feature(KVM_FEATURE_PV_DEDICATED))
+               return;
Quoted 8 lines
Quoted 6 lines
arch/x86/kernel/kvm.c +0 -1
        /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
        if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
                return;
--

2.7.4

You should also add a cpuid flag in kvm part.

Also, I am thinking if PV_DEDICATED helps in performance and with conjunction with PV TLB patch in other thread. For use-case e.g KVM-RT where we don't overcommit vCPU's and pin vCPU:pCPU 1:1 we need a way from host side with which user can decide to enable PV_DEDICATED option. Such that if vCPU's are unlikely going to preempt or sleep we should avoid traversing the cpulist in PV TLB code.

Quoted 3 lines

So, two things:

1] A way to configure PV_DEDICATED from host.

Userspace can already configure it through KVM_SET_CPUID2 ioctl (regardless of KVM version).

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

thanks.

#13 lore raw
Quoted 4 lines · Radim Krcmar

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

#14 lore raw
Quoted 4 lines

Subject: Re: [PATCHv3 1/1] locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set

On Thu, Nov 09, 2017 at 05:05:36PM +0100, Radim Krcmar wrote:

Quoted 8 lines
Quoted 5 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 2 lines

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

Yes.

If I understand it correctly, It will just avoid traversing all the cpus another time just to check/set 'KVM_VCPU_PREEMPTED'/FLUSH value and clear the cpu mask.

Thanks,
Pankaj

#15 lore raw
In reply to Peter Zijlstra

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

#16 lore raw
Quoted 10 lines · Radim Krcmar

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

Quoted 8 lines

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

Also, you should not put cpumask_t on stack, that's 'broken'.

#17 lore raw
Quoted 18 lines · Peter Zijlstra
Quoted 17 lines · Radim Krcmar

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

Quoted 6 lines
Quoted 1 line

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

Also, you should not put cpumask_t on stack, that's 'broken'.

Also, you'll want to use __cpumask_clear_cpu() there.

#18 lore raw
Quoted 19 lines · Peter Zijlstra
Quoted 18 lines · Peter Zijlstra
Quoted 17 lines · Radim Krcmar

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

Quoted 8 lines
Quoted 6 lines
Quoted 1 line

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

Also, you should not put cpumask_t on stack, that's 'broken'.

Also, you'll want to use __cpumask_clear_cpu() there.

Also^2, that patch split is crazy, after patch 2/3 your machine is broken due to lost TLB flushes. You have to first add the SHOULD_FLUSH handling and then clear CPUs from the native_flush_tlb_other() mask.

#19 lore raw

2017-11-09 18:28+0100, Peter Zijlstra:

Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Peter Zijlstra
Quoted 23 lines · Radim Krcmar
Quoted 10 lines

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

Quoted 8 lines

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

Quoted 7 lines
Quoted 3 lines
Quoted 1 line

Also, you should not put cpumask_t on stack, that's 'broken'.

Also, you'll want to use __cpumask_clear_cpu() there.

Also^2, that patch split is crazy, after patch 2/3 your machine is broken due to lost TLB flushes. You have to first add the SHOULD_FLUSH handling and then clear CPUs from the native_flush_tlb_other() mask.

That should be fixed in v2 -- [2/3] must not enable this feature if the host has not exposed it and [3/3] has to expose it. (The ordering of those two doesn't matter as they are separate kernel.)

#20 lore raw
In reply to Peter Zijlstra

2017-11-10 1:15 GMT+08:00 Peter Zijlstra <[email protected]>:

Quoted 19 lines · Peter Zijlstra
Quoted 18 lines · Radim Krcmar
Quoted 17 lines

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

Quoted 6 lines
Quoted 6 lines
Quoted 1 line

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

Also, you should not put cpumask_t on stack, that's 'broken'.

Thanks pointing out this. I found a useful comments in arch/x86/kernel/irq.c:

/* These two declarations are only used in check_irq_vectors_for_cpu_disable()
* below, which is protected by stop_machine(). Putting them on the stack
* results in a stack frame overflow. Dynamically allocating could result in a
* failure so declare these two cpumasks as global.
*/
static struct cpumask affinity_new, online_new;

Quoted 2 lines

Also, you'll want to use __cpumask_clear_cpu() there.

Will do.

Regards,
Wanpeng Li

#21 lore raw
Quoted 10 lines · Wanpeng Li
Quoted 1 line
Quoted 1 line

Also, you should not put cpumask_t on stack, that's 'broken'.

Thanks pointing out this. I found a useful comments in arch/x86/kernel/irq.c:

/* These two declarations are only used in check_irq_vectors_for_cpu_disable()
* below, which is protected by stop_machine(). Putting them on the stack
* results in a stack frame overflow. Dynamically allocating could result in a
* failure so declare these two cpumasks as global.
*/
static struct cpumask affinity_new, online_new;

That code no longer exists.. Also not entirely sure how it would be helpful.

What you probably want to do is have a per-cpu cpumask, since flush_tlb_others() is called with preemption disabled. But you probably don't want an unconditionally allocated one, since most kernels will not in fact be PV.

So you'll want something like:

	static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);

And then you need something like:

	for_each_possible_cpu(cpu) {
		zalloc_cpumask_var_node(per_cpu_ptr(&__pb_tlb_mask, cpu),
					GFP_KERNEL, cpu_to_node(cpu));
	}

before you set the pv-op or so.

#22 lore raw

2017-11-10 15:59 GMT+08:00 Peter Zijlstra <[email protected]>:

Quoted 25 lines · Wanpeng Li
Quoted 11 lines
Quoted 2 lines
Quoted 2 lines

Also, you should not put cpumask_t on stack, that's 'broken'.

Thanks pointing out this. I found a useful comments in arch/x86/kernel/irq.c:

/* These two declarations are only used in check_irq_vectors_for_cpu_disable()
* below, which is protected by stop_machine(). Putting them on the stack
* results in a stack frame overflow. Dynamically allocating could result in a
* failure so declare these two cpumasks as global.
*/
static struct cpumask affinity_new, online_new;

That code no longer exists.. Also not entirely sure how it would be helpful.

What you probably want to do is have a per-cpu cpumask, since flush_tlb_others() is called with preemption disabled. But you probably don't want an unconditionally allocated one, since most kernels will not in fact be PV.

So you'll want something like:

        static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
Quoted 8 lines

And then you need something like:

        for_each_possible_cpu(cpu) {
                zalloc_cpumask_var_node(per_cpu_ptr(&__pb_tlb_mask, cpu),
                                        GFP_KERNEL, cpu_to_node(cpu));
        }

before you set the pv-op or so.

Thanks Peterz, :) I will have a try.

Regards,
Wanpeng Li

#23 lore raw
In reply to Peter Zijlstra

2017-11-09 18:12+0100, Peter Zijlstra:

Quoted 14 lines · Radim Krcmar
Quoted 13 lines

2017-11-09 17:17+0100, Peter Zijlstra:

Quoted 8 lines · Radim Krcmar
Quoted 4 lines

2017-11-09 10:53-0500, Pankaj Gupta:

Quoted 1 line

2] PV TLB should also behave as per option PV_DEDICATED for better performance.

Right,

Shouldn't KVM do flush_tlb_other() in any case? Not sure how PV_DEDICATED can help with that.

It will, the suggestion was based on recent extension of the flush_tlb_others implementaion, https://lkml.org/lkml/2017/11/8/1146.

Quoted 8 lines
Quoted 5 lines

PV_TLB_FLUSH allows a guest to set a flush bit instead of sending flush IPI if the target VCPU is not running. This would be a waste of time with PV_DEDICATED as all VCPUs are expected to always running.

With PV_DEDICATED, the guest should keep using native_flush_tlb_others.

Is saving that for_each_cpu() really worth the effort compared to the cost of actually doing the IPIs and CR3 write?

It is one line for a few percent (hopefully better for AMD with AVIC).

Still, keeping the decision completely on userspace would be cleaner.

Quoted 1 line

Also, you should not put cpumask_t on stack, that's 'broken'.

Good catch, thanks.

#24 lore raw
In reply to Wanpeng Li

2017-11-09 20:43+0800, Wanpeng Li:

Quoted 37 lines

2017-11-07 4:26 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 34 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>
Quoted 1 line

---

Quoted 1 line

You should also add a cpuid flag in kvm part.

It is better without that. The flag has no dependency on KVM (kernel hypervisor) code.

#25 lore raw

2017-11-10 0:00 GMT+08:00 Radim Krcmar <[email protected]>:

Quoted 38 lines

2017-11-09 20:43+0800, Wanpeng Li:

Quoted 37 lines

2017-11-07 4:26 GMT+08:00 Eduardo Valentin <[email protected]>:

Quoted 35 lines

Currently, the existing qspinlock implementation will fallback to test-and-set if the hypervisor has not set the PV_UNHALT flag.

This patch gives the opportunity to guest kernels to select between test-and-set and the regular queueu fair lock implementation based on the PV_DEDICATED KVM feature flag. When the PV_DEDICATED flag is not set, the code will still fall back to test-and-set, but when the PV_DEDICATED flag is set, the code will use the regular queue spinlock implementation.

With this patch, when in autoselect mode, the guest will use the default spinlock implementation based on host feature flags as follows:

PV_DEDICATED = 1, PV_UNHALT = anything: default is qspinlock PV_DEDICATED = 0, PV_UNHALT = 1: default is pvqspinlock PV_DEDICATED = 0, PV_UNHALT = 0: default is tas

Cc Paolo Bonzini <[email protected]>
Cc "Radim Krčmář" <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Waiman Long <[email protected]>
Cc Jan H. Schoenherr <[email protected]>
Cc Anthony Liguori <[email protected]>

---

Quoted 4 lines
Quoted 1 line

You should also add a cpuid flag in kvm part.

It is better without that. The flag has no dependency on KVM (kernel hypervisor) code.

Do you mean -cpu host, +xxxx,I think it will result in "warning: host doesn't support requested feature: CPUID.40000001H:eax.XXXXXX"

Regards,
Wanpeng Li

maintainer
#26 lore raw
Quoted 5 lines · Wanpeng Li
Quoted 3 lines
Quoted 1 line

You should also add a cpuid flag in kvm part.

It is better without that. The flag has no dependency on KVM (kernel hypervisor) code.

Do you mean -cpu host, +xxxx,I think it will result in "warning: host doesn't support requested feature: CPUID.40000001H:eax.XXXXXX"

There are some exceptions where QEMU overrides the values of KVM_GET_SUPPORTED_CPUID.

I think it is better to add the flag to KVM *and* to QEMU's override in kvm_arch_get_supported_cpuid (target/i386/kvm.c).

Paolo