Skip to contentSkip to conversation

KVM: lapic: Fix a variety of timer adv issues

PATCH Superseded Sean Christopherson 2 participants 20 messages Apr 16, 2019 – Apr 17, 2019

Subsystems (3)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v3 cover KVM: lapic: Fix a variety of timer adv issues

The recent change to automatically calculate lapic_timer_advance_ns introduced a handful of gnarly bugs, and also exposed a latent bug by virtue of the advancement logic being enabled by default. Inspection and debug revealed several other opportunities for minor improvements.

The primary issue is that the auto-tuning of lapic_timer_advance_ns is completely unbounded, e.g. there's nothing in the logic that prevents the advancement from creeping up to hundreds of milliseconds. Adjusting the timers by large amounts creates major discrepancies in the guest, e.g. a timer that was configured to fire after multiple milliseconds may arrive before the guest executes a single instruction. While technically correct from a time perspective, it breaks a reasonable assumption from the guest that it can execute some number of instructions between timer events.

The other major issue is that the advancement is global, while TSC scaling is done on a per-vCPU basic. Adjusting the advancement at runtime exacerbates this as there is no protection against multiple vCPUs and/or multiple VMs concurrently modifying the advancement value, e.g. it can effectively become corrupted or never stabilize due to getting bounced all over tarnation.

As for the latent bug, when timer advancement was applied to the hv_timer, i.e. the VMX preemption timer, the logic to trigger wait_for_lapic_timer() was not updated. As a result, a timer interrupt emulated via the hv_timer can easily arrive too early from a *time* perspective, as opposed to simply arriving early from a "number of instructions executed" perspective.

v3:
- Split the refactoring of start_hv_timer() and ->set_hv_timer into three separate patches instead of attempting to do a big refactor in a single patch to fix three separate issues.
- Explicitly cancel the hv timer to avoid
- Use a param for "expired" instead of overloading the return value of ->set_hv_timer().
- Check for a pending non-periodic in restart_apic_timer(). [Liran]
- Add more Reviewed-by tags.

v2:
- https://patchwork.kernel.org/cover/10903613/
- Add explicit param to control automatic tuning. [Liran]
- Document the effect of per-vCPU tracking on the module params.
- Use fancy math to convert guest clockcycles to host nanoseconds instead of brute forcing the delay with a for loop. [Liran]
- Refactor start_hv_timer()'s return semantics to move the "expired timer" handling up a level. [Liran and Paolo]
- Add Liran's Reviewed-by tags.

v1: https://patchwork.kernel.org/cover/10899101/

Sean Christopherson (9):
KVM: lapic: Hard cap the auto-calculated timer advancement
KVM: lapic: Convert guest TSC to host time domain when delaying
KVM: lapic: Track lapic timer advance per vCPU
KVM: lapic: Allow user to disable auto-tuning of timer advancement
KVM: lapic: Busy wait for timer to expire when using hv_timer
KVM: lapic: Explicitly cancel the hv timer if it's pre-expired
KVM: lapic: Refactor ->set_hv_timer to use an explicit expired param
KVM: lapic: Check for a pending timer intr prior to start_hv_timer()
KVM: VMX: Skip delta_tsc shift-and-divide if the dividend is zero

 arch/x86/include/asm/kvm_host.h |  3 +-
 arch/x86/kvm/lapic.c            | 81 +++++++++++++++++++++------------
 arch/x86/kvm/lapic.h            |  5 +-
 arch/x86/kvm/vmx/vmx.c          | 15 +++---
 arch/x86/kvm/x86.c              | 11 +++--
 arch/x86/kvm/x86.h              |  2 -
 6 files changed, 73 insertions(+), 44 deletions(-)
Signature · 1 line
2.21.0
#1 lore raw

PATCH v3 1/9 KVM: lapic: Hard cap the auto-calculated timer advancement

To minimize the latency of timer interrupts as observed by the guest, KVM adjusts the values it programs into the host timers to account for the host's overhead of programming and handling the timer event. Now that the timer advancement is automatically tuned during runtime, it's effectively unbounded by default, e.g. if KVM is running as L1 the advancement can measure in hundreds of milliseconds.

Place a somewhat arbitrary hard cap of 5000ns on the auto-calculated advancement, as large advancements can break reasonable assumptions of the guest, e.g. that a timer configured to fire after 1ms won't arrive on the next instruction. Although KVM busy waits to mitigate the timer event arriving too early, complications can arise when shifting the interrupt too far, e.g. vmx.flat/interrupt in kvm-unit-tests will fail when its "host" exits on interrupts (because the INTR is injected before the gets executes STI+HLT). Arguably the unit test is "broken" in the sense that delaying the timer interrupt by 1ms doesn't technically guarantee the interrupt will arrive after STI+HLT, but it's a reasonable assumption that KVM should support.

Furthermore, an unbounded advancement also effectively unbounds the time spent busy waiting, e.g. if the guest programs a timer with a very large delay.

Arguably the advancement logic could simply be disabled when running as L1, but KVM needs to bound the advancement time regardless, e.g. if the TSC is unstable and the calculations get wildly out of whack. And allowing the advancement when running as L1 is a good stress test of sorts for the logic.

Cc Liran Alon <[email protected]>
Cc Wanpeng Li <[email protected]>
Fixes 3b8a5df6c4dc6 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
---
 arch/x86/kvm/lapic.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/kvm/lapic.c +5 -0
index 9bf70cf84564..92446cba9b24 100644
@@ -74,6 +74,7 @@ static bool lapic_timer_advance_adjust_done = false;
 #define LAPIC_TIMER_ADVANCE_ADJUST_DONE 100
 /* step-by-step approximation to mitigate fluctuation */
 #define LAPIC_TIMER_ADVANCE_ADJUST_STEP 8
+#define LAPIC_TIMER_ADVANCE_MAX_NS	5000
 
 static inline int apic_test_vector(int vec, void *bitmap)
 {
@@ -1522,6 +1523,10 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 		}
 		if (abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE)
 			lapic_timer_advance_adjust_done = true;
+		if (unlikely(lapic_timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
+			lapic_timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
+			lapic_timer_advance_adjust_done = true;
+		}
 	}
 }
 
Signature · 1 line
2.21.0
maintainer
#2 lore raw

Re: [PATCH v3 1/9] KVM: lapic: Hard cap the auto-calculated timer advancement

Quoted 50 lines · Sean Christopherson

To minimize the latency of timer interrupts as observed by the guest, KVM adjusts the values it programs into the host timers to account for the host's overhead of programming and handling the timer event. Now that the timer advancement is automatically tuned during runtime, it's effectively unbounded by default, e.g. if KVM is running as L1 the advancement can measure in hundreds of milliseconds.

Place a somewhat arbitrary hard cap of 5000ns on the auto-calculated advancement, as large advancements can break reasonable assumptions of the guest, e.g. that a timer configured to fire after 1ms won't arrive on the next instruction. Although KVM busy waits to mitigate the timer event arriving too early, complications can arise when shifting the interrupt too far, e.g. vmx.flat/interrupt in kvm-unit-tests will fail when its "host" exits on interrupts (because the INTR is injected before the gets executes STI+HLT). Arguably the unit test is "broken" in the sense that delaying the timer interrupt by 1ms doesn't technically guarantee the interrupt will arrive after STI+HLT, but it's a reasonable assumption that KVM should support.

Furthermore, an unbounded advancement also effectively unbounds the time spent busy waiting, e.g. if the guest programs a timer with a very large delay.

Arguably the advancement logic could simply be disabled when running as L1, but KVM needs to bound the advancement time regardless, e.g. if the TSC is unstable and the calculations get wildly out of whack. And allowing the advancement when running as L1 is a good stress test of sorts for the logic.

Cc Liran Alon <[email protected]>
Cc Wanpeng Li <[email protected]>
Fixes 3b8a5df6c4dc6 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
---
 arch/x86/kvm/lapic.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/kvm/lapic.c +1 -0
index 9bf70cf84564..92446cba9b24 100644
@@ -74,6 +74,7 @@ static bool lapic_timer_advance_adjust_done = false;
 #define LAPIC_TIMER_ADVANCE_ADJUST_DONE 100
 /* step-by-step approximation to mitigate fluctuation */
 #define LAPIC_TIMER_ADVANCE_ADJUST_STEP 8
+#define LAPIC_TIMER_ADVANCE_MAX_NS	5000
 
 static inline int apic_test_vector(int vec, void *bitmap)
 {

arch/x86/kvm/lapic.c void wait_lapic_expire(struct kvm_vcpu *vcpu) · line 1529

Quoted 7 lines
arch/x86/kvm/lapic.c +4 -0
@@ -1522,6 +1523,10 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 		}
 		if (abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE)
 			lapic_timer_advance_adjust_done = true;
+		if (unlikely(lapic_timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
+			lapic_timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
+			lapic_timer_advance_adjust_done = true;
+		}

I would treat this case as "advancing the timer has failed miserably" and reset lapic_timer_advance_ns to 0.

Paolo

#3 lore raw

Re: [PATCH v3 1/9] KVM: lapic: Hard cap the auto-calculated timer advancement

Quoted 48 lines · Paolo Bonzini
Quoted 47 lines · Sean Christopherson

To minimize the latency of timer interrupts as observed by the guest, KVM adjusts the values it programs into the host timers to account for the host's overhead of programming and handling the timer event. Now that the timer advancement is automatically tuned during runtime, it's effectively unbounded by default, e.g. if KVM is running as L1 the advancement can measure in hundreds of milliseconds.

Place a somewhat arbitrary hard cap of 5000ns on the auto-calculated advancement, as large advancements can break reasonable assumptions of the guest, e.g. that a timer configured to fire after 1ms won't arrive on the next instruction. Although KVM busy waits to mitigate the timer event arriving too early, complications can arise when shifting the interrupt too far, e.g. vmx.flat/interrupt in kvm-unit-tests will fail when its "host" exits on interrupts (because the INTR is injected before the gets executes STI+HLT). Arguably the unit test is "broken" in the sense that delaying the timer interrupt by 1ms doesn't technically guarantee the interrupt will arrive after STI+HLT, but it's a reasonable assumption that KVM should support.

Furthermore, an unbounded advancement also effectively unbounds the time spent busy waiting, e.g. if the guest programs a timer with a very large delay.

Arguably the advancement logic could simply be disabled when running as L1, but KVM needs to bound the advancement time regardless, e.g. if the TSC is unstable and the calculations get wildly out of whack. And allowing the advancement when running as L1 is a good stress test of sorts for the logic.

Cc Liran Alon <[email protected]>
Cc Wanpeng Li <[email protected]>
Fixes 3b8a5df6c4dc6 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
---
 arch/x86/kvm/lapic.c | 5 +++++
 1 file changed, 5 insertions(+)
arch/x86/kvm/lapic.c +1 -0
index 9bf70cf84564..92446cba9b24 100644
@@ -74,6 +74,7 @@ static bool lapic_timer_advance_adjust_done = false;
 #define LAPIC_TIMER_ADVANCE_ADJUST_DONE 100
 /* step-by-step approximation to mitigate fluctuation */
 #define LAPIC_TIMER_ADVANCE_ADJUST_STEP 8
+#define LAPIC_TIMER_ADVANCE_MAX_NS	5000
 
Quoted 13 lines

arch/x86/kvm/lapic.c void wait_lapic_expire(struct kvm_vcpu *vcpu) · line 1529

Quoted 10 lines
arch/x86/kvm/lapic.c +4 -0
 static inline int apic_test_vector(int vec, void *bitmap)
 {
@@ -1522,6 +1523,10 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 		}
 		if (abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE)
 			lapic_timer_advance_adjust_done = true;
+		if (unlikely(lapic_timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
+			lapic_timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
+			lapic_timer_advance_adjust_done = true;
+		}

I would treat this case as "advancing the timer has failed miserably" and reset lapic_timer_advance_ns to 0.

As in, disable the feature, correct? That works for me.

maintainer
#4 lore raw

Re: [PATCH v3 1/9] KVM: lapic: Hard cap the auto-calculated timer advancement

Quoted 11 lines · Sean Christopherson
Quoted 10 lines

arch/x86/kvm/lapic.c void wait_lapic_expire(struct kvm_vcpu *vcpu) · line 1529

Quoted 8 lines
arch/x86/kvm/lapic.c +4 -0
@@ -1522,6 +1523,10 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 		}
 		if (abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE)
 			lapic_timer_advance_adjust_done = true;
+		if (unlikely(lapic_timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
+			lapic_timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
+			lapic_timer_advance_adjust_done = true;
+		}

I would treat this case as "advancing the timer has failed miserably" and reset lapic_timer_advance_ns to 0.

As in, disable the feature, correct? That works for me.

Yes.

Paolo

#5 lore raw

PATCH v3 2/9 KVM: lapic: Convert guest TSC to host time domain when delaying

In reply to Sean Christopherson

To minimize the latency of timer interrupts as observed by the guest, KVM adjusts the values it programs into the host timers to account for the host's overhead of programming and handling the timer event. In the event that the adjustments are too aggressive, i.e. the timer fires earlier than the guest expects, KVM busy waits immediately prior to entering the guest.

Currently, KVM manually converts the delay from nanoseconds to clock cycles. But, the conversion is done in the guest's time domain, while the delay occurs in the host's time domain, i.e. the delay may not be accurate and could wait too little or too long.

Convert the delay from guest clock cycles to host nanoseconds and use ndelay() instead of __delay() to provide more accurate timing. This also avoids the need to convert lapic_timer_advance_ns, which is used to cap the delay, to guest clock cycles.

---
 arch/x86/kvm/lapic.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
arch/x86/kvm/lapic.c +6 -4
index 92446cba9b24..5891c0badfa6 100644
@@ -1502,10 +1502,12 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 	guest_tsc = kvm_read_l1_tsc(vcpu, rdtsc());
 	trace_kvm_wait_lapic_expire(vcpu->vcpu_id, guest_tsc - tsc_deadline);
 
-	/* __delay is delay_tsc whenever the hardware has TSC, thus always.  */
-	if (guest_tsc < tsc_deadline)
-		__delay(min(tsc_deadline - guest_tsc,
-			nsec_to_cycles(vcpu, lapic_timer_advance_ns)));
+	/* ndelay uses delay_tsc whenever the hardware has TSC, thus always. */
+	if (guest_tsc < tsc_deadline) {
+		ns = (tsc_deadline - guest_tsc) * 1000000ULL;
+		do_div(ns, vcpu->arch.virtual_tsc_khz);
+		ndelay(min_t(u32, ns, lapic_timer_advance_ns));
+	}
 
 	if (!lapic_timer_advance_adjust_done) {
 		/* too early */
Signature · 1 line
2.21.0
maintainer
#6 lore raw

Re: [PATCH v3 2/9] KVM: lapic: Convert guest TSC to host time domain when delaying

arch/x86/kvm/lapic.c void wait_lapic_expire(struct kvm_vcpu *vcpu) · line 1510

Quoted 7 lines · Sean Christopherson
arch/x86/kvm/lapic.c +6 -0
+	/* ndelay uses delay_tsc whenever the hardware has TSC, thus always. */
+	if (guest_tsc < tsc_deadline) {
+		ns = (tsc_deadline - guest_tsc) * 1000000ULL;
+		do_div(ns, vcpu->arch.virtual_tsc_khz);
+		ndelay(min_t(u32, ns, lapic_timer_advance_ns));
+	}

This min_t would allow the timer to expire in advance of the indicated deadline. If this happens, wouldn't it be a bug elsewhere?

Paolo

#7 lore raw

Re: [PATCH v3 2/9] KVM: lapic: Convert guest TSC to host time domain when delaying

Quoted 11 lines · Paolo Bonzini

arch/x86/kvm/lapic.c void wait_lapic_expire(struct kvm_vcpu *vcpu) · line 1510

Quoted 7 lines · Sean Christopherson
arch/x86/kvm/lapic.c +6 -0
+	/* ndelay uses delay_tsc whenever the hardware has TSC, thus always. */
+	if (guest_tsc < tsc_deadline) {
+		ns = (tsc_deadline - guest_tsc) * 1000000ULL;
+		do_div(ns, vcpu->arch.virtual_tsc_khz);
+		ndelay(min_t(u32, ns, lapic_timer_advance_ns));
+	}

This min_t would allow the timer to expire in advance of the indicated deadline. If this happens, wouldn't it be a bug elsewhere?

Maybe?

In this exact variation of the code, no, since userspace can modify lapic_timer_advance_ns while the vCPU is running. Obviously that case goes away when lapic_timer_advance_ns is no longer directly consumed by vCPUS.

Originally, the check was added because lockup occurred if the guest's TSC was sent backwards[1]. The discussion surrounding the bugfix patch[2] is probably more interesting/amusing/relevant. We're basically rehashing that discussion. :-)

Anyways, capping at a KVM controlled value is prudent since not doing so can hang the host if something does go wrong. As for the __delay() vs ndelay() discussion, I'll try to dig more into how/why I saw the timer fire with an incorrect TSC, i.e. why ndelay() "works" but __delay() does not.

https://bugzilla.redhat.com/show_bug.cgi?id=1337667
https://patchwork.kernel.org/patch/9130687/

maintainer
#8 lore raw

Re: [PATCH v3 2/9] KVM: lapic: Convert guest TSC to host time domain when delaying

Quoted 4 lines · Sean Christopherson

In this exact variation of the code, no, since userspace can modify lapic_timer_advance_ns while the vCPU is running. Obviously that case goes away when lapic_timer_advance_ns is no longer directly consumed by vCPUS.

Indeed, and if we weren't about to add per-vCPU lapic_timer_advance_ns, the old suggestion of making lapic_timer_advance_ns read-only would stand.

Quoted 2 lines

Anyways, capping at a KVM controlled value is prudent since not doing so can hang the host if something does go wrong.

Yes---for the per-vCPU lapic_timer_advance_ns it is completely reasonable.

Paolo

#9 lore raw

PATCH v3 3/9 KVM: lapic: Track lapic timer advance per vCPU

In reply to Sean Christopherson

Automatically adjusting the globally-shared timer advancement could corrupt the timer, e.g. if multiple vCPUs are concurrently adjusting the advancement value. That could be partially fixed by using a local variable for the arithmetic, but it would still be susceptible to a race when setting timer_advance_adjust_done.

And because virtual_tsc_khz and tsc_scaling_ratio are per-vCPU, the correct calibration for a given vCPU may not apply to all vCPUs.

Furthermore, lapic_timer_advance_ns is marked __read_mostly, which is effectively violated when finding a stable advancement takes an extended amount of timer.

Opportunistically change the definition of lapic_timer_advance_ns to a u32 so that it matches the style of struct kvm_timer. Explicitly pass the param to kvm_create_lapic() so that it doesn't have to be exposed to lapic.c, thus reducing the probability of unintentionally using the global value instead of the per-vCPU value.

Fixes 3b8a5df6c4dc6 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Cc Liran Alon <[email protected]>
Cc Wanpeng Li <[email protected]>
---
 arch/x86/kvm/lapic.c   | 34 ++++++++++++++++++----------------
 arch/x86/kvm/lapic.h   |  4 +++-
 arch/x86/kvm/vmx/vmx.c |  4 +++-
 arch/x86/kvm/x86.c     |  7 +++----
 arch/x86/kvm/x86.h     |  2 --
 5 files changed, 27 insertions(+), 24 deletions(-)
arch/x86/kvm/lapic.c +18 -16
index 5891c0badfa6..3cd62edf8084 100644
@@ -70,7 +70,6 @@
 #define APIC_BROADCAST			0xFF
 #define X2APIC_BROADCAST		0xFFFFFFFFul
 
-static bool lapic_timer_advance_adjust_done = false;
 #define LAPIC_TIMER_ADVANCE_ADJUST_DONE 100
 /* step-by-step approximation to mitigate fluctuation */
 #define LAPIC_TIMER_ADVANCE_ADJUST_STEP 8
@@ -1486,6 +1485,7 @@ static bool lapic_timer_int_injected(struct kvm_vcpu *vcpu)
 void wait_lapic_expire(struct kvm_vcpu *vcpu)
 {
 	struct kvm_lapic *apic = vcpu->arch.apic;
+	u32 timer_advance_ns = apic->lapic_timer.timer_advance_ns;
 	u64 guest_tsc, tsc_deadline, ns;
 
 	if (!lapic_in_kernel(vcpu))
@@ -1506,35 +1506,37 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
 	if (guest_tsc < tsc_deadline) {
 		ns = (tsc_deadline - guest_tsc) * 1000000ULL;
 		do_div(ns, vcpu->arch.virtual_tsc_khz);
-		ndelay(min_t(u32, ns, lapic_timer_advance_ns));
+		ndelay(min_t(u32, ns, timer_advance_ns));
 	}
 
-	if (!lapic_timer_advance_adjust_done) {
+	if (!apic->lapic_timer.timer_advance_adjust_done) {
 		/* too early */
 		if (guest_tsc < tsc_deadline) {
 			ns = (tsc_deadline - guest_tsc) * 1000000ULL;
 			do_div(ns, vcpu->arch.virtual_tsc_khz);
-			lapic_timer_advance_ns -= min((unsigned int)ns,
-				lapic_timer_advance_ns / LAPIC_TIMER_ADVANCE_ADJUST_STEP);
+			timer_advance_ns -= min((u32)ns,
+				timer_advance_ns / LAPIC_TIMER_ADVANCE_ADJUST_STEP);
 		} else {
 		/* too late */
 			ns = (guest_tsc - tsc_deadline) * 1000000ULL;
 			do_div(ns, vcpu->arch.virtual_tsc_khz);
-			lapic_timer_advance_ns += min((unsigned int)ns,
-				lapic_timer_advance_ns / LAPIC_TIMER_ADVANCE_ADJUST_STEP);
+			timer_advance_ns += min((u32)ns,
+				timer_advance_ns / LAPIC_TIMER_ADVANCE_ADJUST_STEP);
 		}
 		if (abs(guest_tsc - tsc_deadline) < LAPIC_TIMER_ADVANCE_ADJUST_DONE)
-			lapic_timer_advance_adjust_done = true;
-		if (unlikely(lapic_timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
-			lapic_timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
-			lapic_timer_advance_adjust_done = true;
+			apic->lapic_timer.timer_advance_adjust_done = true;
+		if (unlikely(timer_advance_ns > LAPIC_TIMER_ADVANCE_MAX_NS)) {
+			timer_advance_ns = LAPIC_TIMER_ADVANCE_MAX_NS;
+			apic->lapic_timer.timer_advance_adjust_done = true;
 		}
+		apic->lapic_timer.timer_advance_ns = timer_advance_ns;
 	}
 }
 
 static void start_sw_tscdeadline(struct kvm_lapic *apic)
 {
-	u64 guest_tsc, tscdeadline = apic->lapic_timer.tscdeadline;
+	struct kvm_timer *ktimer = &apic->lapic_timer;
+	u64 guest_tsc, tscdeadline = ktimer->tscdeadline;
 	u64 ns = 0;
 	ktime_t expire;
 	struct kvm_vcpu *vcpu = apic->vcpu;
@@ -1553,9 +1555,8 @@ static void start_sw_tscdeadline(struct kvm_lapic *apic)
 		ns = (tscdeadline - guest_tsc) * 1000000ULL;
 		do_div(ns, this_tsc_khz);
 		expire = ktime_add_ns(now, ns);
-		expire = ktime_sub_ns(expire, lapic_timer_advance_ns);
-		hrtimer_start(&apic->lapic_timer.timer,
-				expire, HRTIMER_MODE_ABS_PINNED);
+		expire = ktime_sub_ns(expire, ktimer->timer_advance_ns);
+		hrtimer_start(&ktimer->timer, expire, HRTIMER_MODE_ABS_PINNED);
 	} else
 		apic_timer_expired(apic);
 
@@ -2262,7 +2263,7 @@ static enum hrtimer_restart apic_timer_fn(struct hrtimer *data)
 		return HRTIMER_NORESTART;
 }
 
-int kvm_create_lapic(struct kvm_vcpu *vcpu)
+int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns)
 {
 	struct kvm_lapic *apic;
 
@@ -2286,6 +2287,7 @@ int kvm_create_lapic(struct kvm_vcpu *vcpu)
 	hrtimer_init(&apic->lapic_timer.timer, CLOCK_MONOTONIC,
 		     HRTIMER_MODE_ABS_PINNED);
 	apic->lapic_timer.timer.function = apic_timer_fn;
+	apic->lapic_timer.timer_advance_ns = timer_advance_ns;
 
 	/*
 	 * APIC is created enabled. This will prevent kvm_lapic_set_base from
arch/x86/kvm/lapic.h +3 -1
index ff6ef9c3d760..3e97f8a68967 100644
@@ -31,8 +31,10 @@ struct kvm_timer {
 	u32 timer_mode_mask;
 	u64 tscdeadline;
 	u64 expired_tscdeadline;
+	u32 timer_advance_ns;
 	atomic_t pending;			/* accumulated triggered timers */
 	bool hv_timer_in_use;
+	bool timer_advance_adjust_done;
 };
 
 struct kvm_lapic {
@@ -62,7 +64,7 @@ struct kvm_lapic {
 
 struct dest_map;
 
-int kvm_create_lapic(struct kvm_vcpu *vcpu);
+int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns);
 void kvm_free_lapic(struct kvm_vcpu *vcpu);
 
 int kvm_apic_has_interrupt(struct kvm_vcpu *vcpu);
arch/x86/kvm/vmx/vmx.c +3 -1
index 5f90616da275..a3211e3ee679 100644
@@ -7031,6 +7031,7 @@ static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
 {
 	struct vcpu_vmx *vmx;
 	u64 tscl, guest_tscl, delta_tsc, lapic_timer_advance_cycles;
+	struct kvm_timer *ktimer = &vcpu->arch.apic->lapic_timer;
 
 	if (kvm_mwait_in_guest(vcpu->kvm))
 		return -EOPNOTSUPP;
@@ -7039,7 +7040,8 @@ static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
 	tscl = rdtsc();
 	guest_tscl = kvm_read_l1_tsc(vcpu, tscl);
 	delta_tsc = max(guest_deadline_tsc, guest_tscl) - guest_tscl;
-	lapic_timer_advance_cycles = nsec_to_cycles(vcpu, lapic_timer_advance_ns);
+	lapic_timer_advance_cycles = nsec_to_cycles(vcpu,
+						    ktimer->timer_advance_ns);
 
 	if (delta_tsc > lapic_timer_advance_cycles)
 		delta_tsc -= lapic_timer_advance_cycles;
arch/x86/kvm/x86.c +3 -4
index 38440316a806..b303a21a2bc2 100644
@@ -137,9 +137,8 @@ static u32 __read_mostly tsc_tolerance_ppm = 250;
 module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR);
 
 /* lapic timer advance (tscdeadline mode only) in nanoseconds */
-unsigned int __read_mostly lapic_timer_advance_ns = 1000;
+static u32 __read_mostly lapic_timer_advance_ns = 1000;
 module_param(lapic_timer_advance_ns, uint, S_IRUGO | S_IWUSR);
-EXPORT_SYMBOL_GPL(lapic_timer_advance_ns);
 
 static bool __read_mostly vector_hashing = true;
 module_param(vector_hashing, bool, S_IRUGO);
@@ -7891,7 +7890,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	}
 
 	trace_kvm_entry(vcpu->vcpu_id);
-	if (lapic_timer_advance_ns)
+	if (vcpu->arch.apic->lapic_timer.timer_advance_ns)
 		wait_lapic_expire(vcpu);
 	guest_enter_irqoff();
 
@@ -9079,7 +9078,7 @@ int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
 
 	if (irqchip_in_kernel(vcpu->kvm)) {
 		vcpu->arch.apicv_active = kvm_x86_ops->get_enable_apicv(vcpu);
-		r = kvm_create_lapic(vcpu);
+		r = kvm_create_lapic(vcpu, lapic_timer_advance_ns);
 		if (r < 0)
 			goto fail_mmu_destroy;
 	} else
arch/x86/kvm/x86.h +0 -2
index eda954441213..a470ff0868c5 100644
@@ -294,8 +294,6 @@ extern u64 kvm_supported_xcr0(void);
 
 extern unsigned int min_timer_period_us;
 
-extern unsigned int lapic_timer_advance_ns;
-
 extern bool enable_vmware_backdoor;
 
 extern struct static_key kvm_no_apic_vcpu;
Signature · 1 line
2.21.0
#10 lore raw

PATCH v3 4/9 KVM: lapic: Allow user to disable auto-tuning of timer advancement

In reply to Sean Christopherson

Add a new KVM param, lapic_timer_advance_autotune, to allow userspace to disable auto-tuning the timer advancement, e.g. to manually tune the delay.

Note, this does not completely restore the original behavior of lapic_timer_advance_ns. Prior to tracking the advancement per vCPU, which is necessary to support autotuning, userspace could adjust lapic_timer_advance_ns for *running* vCPU. With per-vCPU tracking, the module params are snapshotted at vCPU creation, i.e. applying a new advancement effectively requires restarting a VM.

Dynamically updating a running vCPU is possible, e.g. a helper could be added to retrieve the desired delay, choosing between the global module param and the per-VCPU value depending on whether or not auto-tuning is (globally) enabled, but introduces a great deal of complexity. The wrapper itself is not complex, but understanding and documenting the effects of dynamically toggling auto-tuning and/or adjusting the timer advancement is nigh impossible since the behavior would be dependent on KVM's implementation as well as compiler optimizations. In other words, providing stable behavior would require extremely careful consideration now and in the future.

Given that the expected use of a manually-tuned timer advancement is to "tune once, run many", use the vastly simpler approach of recognizing changes to the module params only when creating a new vCPU.

Fixes 3b8a5df6c4dc6 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Cc Liran Alon <[email protected]>
Cc Wanpeng Li <[email protected]>
---
 arch/x86/kvm/lapic.c | 4 +++-
 arch/x86/kvm/lapic.h | 3 ++-
 arch/x86/kvm/x86.c   | 6 +++++-
 3 files changed, 10 insertions(+), 3 deletions(-)
arch/x86/kvm/lapic.c +3 -1
index 3cd62edf8084..37d3489f68c9 100644
@@ -2263,7 +2263,8 @@ static enum hrtimer_restart apic_timer_fn(struct hrtimer *data)
 		return HRTIMER_NORESTART;
 }
 
-int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns)
+int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns,
+		     bool timer_advance_autotune)
 {
 	struct kvm_lapic *apic;
 
@@ -2288,6 +2289,7 @@ int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns)
 		     HRTIMER_MODE_ABS_PINNED);
 	apic->lapic_timer.timer.function = apic_timer_fn;
 	apic->lapic_timer.timer_advance_ns = timer_advance_ns;
+	apic->lapic_timer.timer_advance_adjust_done = !timer_advance_autotune;
 
 	/*
 	 * APIC is created enabled. This will prevent kvm_lapic_set_base from
arch/x86/kvm/lapic.h +2 -1
index 3e97f8a68967..7fa1aed02c14 100644
@@ -64,7 +64,8 @@ struct kvm_lapic {
 
 struct dest_map;
 
-int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns);
+int kvm_create_lapic(struct kvm_vcpu *vcpu, u32 timer_advance_ns,
+		     bool timer_advance_autotune);
 void kvm_free_lapic(struct kvm_vcpu *vcpu);
 
 int kvm_apic_has_interrupt(struct kvm_vcpu *vcpu);
arch/x86/kvm/x86.c +5 -1
index b303a21a2bc2..60d80e354f62 100644
@@ -140,6 +140,9 @@ module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR);
 static u32 __read_mostly lapic_timer_advance_ns = 1000;
 module_param(lapic_timer_advance_ns, uint, S_IRUGO | S_IWUSR);
 
+static bool __read_mostly lapic_timer_advance_autotune = true;
+module_param(lapic_timer_advance_autotune, bool, S_IRUGO | S_IWUSR);
+
 static bool __read_mostly vector_hashing = true;
 module_param(vector_hashing, bool, S_IRUGO);
 
@@ -9078,7 +9081,8 @@ int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
 
 	if (irqchip_in_kernel(vcpu->kvm)) {
 		vcpu->arch.apicv_active = kvm_x86_ops->get_enable_apicv(vcpu);
-		r = kvm_create_lapic(vcpu, lapic_timer_advance_ns);
+		r = kvm_create_lapic(vcpu, lapic_timer_advance_ns,
+				     lapic_timer_advance_autotune);
 		if (r < 0)
 			goto fail_mmu_destroy;
 	} else
Signature · 1 line
2.21.0
maintainer
#11 lore raw

Re: [PATCH v3 4/9] KVM: lapic: Allow user to disable auto-tuning of timer advancement

arch/x86/kvm/x86.c module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR); · line 145

Quoted 3 lines · Sean Christopherson
arch/x86/kvm/x86.c +3 -0
+static bool __read_mostly lapic_timer_advance_autotune = true;
+module_param(lapic_timer_advance_autotune, bool, S_IRUGO | S_IWUSR);
+

Could we just have lapic_timer_advance_ns = 0 mean "adaptive tuning, starting from a default of 1000" and anything else mean "no tuning, just use this value"?

Paolo

#12 lore raw

Re: [PATCH v3 4/9] KVM: lapic: Allow user to disable auto-tuning of timer advancement

Quoted 8 lines · Paolo Bonzini

arch/x86/kvm/x86.c module_param(tsc_tolerance_ppm, uint, S_IRUGO | S_IWUSR); · line 145

Quoted 3 lines · Sean Christopherson
arch/x86/kvm/x86.c +3 -0
+static bool __read_mostly lapic_timer_advance_autotune = true;
+module_param(lapic_timer_advance_autotune, bool, S_IRUGO | S_IWUSR);
+

Could we just have lapic_timer_advance_ns = 0 mean "adaptive tuning, starting from a default of 1000" and anything else mean "no tuning, just use this value"?

Originally I implemented a variation of that, except I disabled autotuning if lapic_timer_advance_ns was set to anything other than the default. Liran requested a separate param since the behavior was admittedly a bit sneaky, and I didn't have a strong opinion either way, so here we are.

I think 'lapic_timer_advance_ns = 0' should be used to disable the advancement. What if we make the param a signed int and have '-1' select the "adaptive tuning ..." option? That could also be the default. It saves a param and is hopefully more intuitive than what I originally proposed.

#13 lore raw

PATCH v3 5/9 KVM: lapic: Busy wait for timer to expire when using hv_timer

In reply to Sean Christopherson

...now that VMX's preemption timer, i.e. the hv_timer, also adjusts its programmed time based on lapic_timer_advance_ns. Without the delay, a guest can see a timer interrupt arrive before the requested time when KVM is using the hv_timer to emulate the guest's interrupt.

Fixes c5ce8235cffa0 ("KVM: VMX: Optimize tscdeadline timer latency")
Cc Wanpeng Li <[email protected]>
---
 arch/x86/kvm/lapic.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/lapic.c +1 -1
index 37d3489f68c9..fa2dbca5ad6a 100644
@@ -1455,7 +1455,7 @@ static void apic_timer_expired(struct kvm_lapic *apic)
 	if (swait_active(q))
 		swake_up_one(q);
 
-	if (apic_lvtt_tscdeadline(apic))
+	if (apic_lvtt_tscdeadline(apic) || ktimer->hv_timer_in_use)
 		ktimer->expired_tscdeadline = ktimer->tscdeadline;
 }
 
Signature · 1 line
2.21.0
#14 lore raw

PATCH v3 6/9 KVM: lapic: Explicitly cancel the hv timer if it's pre-expired

In reply to Sean Christopherson

Explicitly call cancel_hv_timer() instead of returning %false to coerce restart_apic_timer() into canceling it by way of start_sw_timer().

Functionally, the existing code is correct in the sense that it doesn't doing anything visibily wrong, e.g. generate spurious interrupts or miss an interrupt. But it's extremely confusing and inefficient, e.g. there are multiple extraneous calls to apic_timer_expired() that effectively get dropped due to @timer_pending being %true.

---
 arch/x86/kvm/lapic.c | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)
arch/x86/kvm/lapic.c +15 -7
index fa2dbca5ad6a..bdcf04689a80 100644
@@ -1701,17 +1701,25 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 	hrtimer_cancel(&ktimer->timer);
 
 	/*
-	 * Also recheck ktimer->pending, in case the sw timer triggered in
-	 * the window.  For periodic timer, leave the hv timer running for
-	 * simplicity, and the deadline will be recomputed on the next vmexit.
+	 * To simplify handling the periodic timer, leave the hv timer running
+	 * even if the deadline timer has expired, i.e. rely on the resulting
+	 * VM-Exit to recompute the periodic timer's target expiration.
 	 */
-	if (!apic_lvtt_period(apic) && (r || atomic_read(&ktimer->pending))) {
-		if (r)
+	if (!apic_lvtt_period(apic)) {
+		/*
+		 * Cancel the hv timer if the sw timer fired while the hv timer
+		 * was being programmed, or if the hv timer itself expired.
+		 */
+		if (atomic_read(&ktimer->pending)) {
+			cancel_hv_timer(apic);
+		} else if (r) {
 			apic_timer_expired(apic);
-		return false;
+			cancel_hv_timer(apic);
+		}
 	}
 
-	trace_kvm_hv_timer_state(apic->vcpu->vcpu_id, true);
+	trace_kvm_hv_timer_state(apic->vcpu->vcpu_id, ktimer->hv_timer_in_use);
+
 	return true;
 }
 
Signature · 1 line
2.21.0
#15 lore raw

PATCH v3 7/9 KVM: lapic: Refactor ->set_hv_timer to use an explicit expired param

In reply to Sean Christopherson

Refactor kvm_x86_ops->set_hv_timer to use an explicit parameter for stating that the timer has expired. Overloading the return value is unnecessarily clever, e.g. can lead to confusion over the proper return value from start_hv_timer() when r==1.

---
 arch/x86/include/asm/kvm_host.h |  3 ++-
 arch/x86/kvm/lapic.c            | 10 +++++-----
 arch/x86/kvm/vmx/vmx.c          |  6 ++++--
 3 files changed, 11 insertions(+), 8 deletions(-)
arch/x86/include/asm/kvm_host.h +2 -1
index a9d03af34030..501a8209285b 100644
@@ -1167,7 +1167,8 @@ struct kvm_x86_ops {
 			      uint32_t guest_irq, bool set);
 	void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);
 
-	int (*set_hv_timer)(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc);
+	int (*set_hv_timer)(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc,
+			    bool *expired);
 	void (*cancel_hv_timer)(struct kvm_vcpu *vcpu);
 
 	void (*setup_mce)(struct kvm_vcpu *vcpu);
arch/x86/kvm/lapic.c +5 -5
index bdcf04689a80..9a09a41000c3 100644
@@ -1681,7 +1681,8 @@ static void cancel_hv_timer(struct kvm_lapic *apic)
 static bool start_hv_timer(struct kvm_lapic *apic)
 {
 	struct kvm_timer *ktimer = &apic->lapic_timer;
-	int r;
+	struct kvm_vcpu *vcpu = apic->vcpu;
+	bool expired;
 
 	WARN_ON(preemptible());
 	if (!kvm_x86_ops->set_hv_timer)
@@ -1693,8 +1694,7 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 	if (!ktimer->tscdeadline)
 		return false;
 
-	r = kvm_x86_ops->set_hv_timer(apic->vcpu, ktimer->tscdeadline);
-	if (r < 0)
+	if (kvm_x86_ops->set_hv_timer(vcpu, ktimer->tscdeadline, &expired))
 		return false;
 
 	ktimer->hv_timer_in_use = true;
@@ -1712,13 +1712,13 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 		 */
 		if (atomic_read(&ktimer->pending)) {
 			cancel_hv_timer(apic);
-		} else if (r) {
+		} else if (expired) {
 			apic_timer_expired(apic);
 			cancel_hv_timer(apic);
 		}
 	}
 
-	trace_kvm_hv_timer_state(apic->vcpu->vcpu_id, ktimer->hv_timer_in_use);
+	trace_kvm_hv_timer_state(vcpu->vcpu_id, ktimer->hv_timer_in_use);
 
 	return true;
 }
arch/x86/kvm/vmx/vmx.c +4 -2
index a3211e3ee679..f3c326e3c324 100644
@@ -7027,7 +7027,8 @@ static inline int u64_shl_div_u64(u64 a, unsigned int shift,
 	return 0;
 }
 
-static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
+static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc,
+			    bool *expired)
 {
 	struct vcpu_vmx *vmx;
 	u64 tscl, guest_tscl, delta_tsc, lapic_timer_advance_cycles;
@@ -7066,7 +7067,8 @@ static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
 		return -ERANGE;
 
 	vmx->hv_deadline_tsc = tscl + delta_tsc;
-	return delta_tsc == 0;
+	*expired = !delta_tsc;
+	return 0;
 }
 
 static void vmx_cancel_hv_timer(struct kvm_vcpu *vcpu)
Signature · 1 line
2.21.0
maintainer
#16 lore raw

Re: [PATCH v3 7/9] KVM: lapic: Refactor ->set_hv_timer to use an explicit expired param

Quoted 4 lines · Sean Christopherson

Refactor kvm_x86_ops->set_hv_timer to use an explicit parameter for stating that the timer has expired. Overloading the return value is unnecessarily clever, e.g. can lead to confusion over the proper return value from start_hv_timer() when r==1.

I guess it's also in the eye of the beholder; I don't like out parameters very much in general but this time I think I don't disagree enough to argue. ;)

Paolo

Quoted 88 lines
---
 arch/x86/include/asm/kvm_host.h |  3 ++-
 arch/x86/kvm/lapic.c            | 10 +++++-----
 arch/x86/kvm/vmx/vmx.c          |  6 ++++--
 3 files changed, 11 insertions(+), 8 deletions(-)
arch/x86/include/asm/kvm_host.h +2 -1
index a9d03af34030..501a8209285b 100644
@@ -1167,7 +1167,8 @@ struct kvm_x86_ops {
 			      uint32_t guest_irq, bool set);
 	void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);
 
-	int (*set_hv_timer)(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc);
+	int (*set_hv_timer)(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc,
+			    bool *expired);
 	void (*cancel_hv_timer)(struct kvm_vcpu *vcpu);
 
 	void (*setup_mce)(struct kvm_vcpu *vcpu);
arch/x86/kvm/lapic.c +5 -5
index bdcf04689a80..9a09a41000c3 100644
@@ -1681,7 +1681,8 @@ static void cancel_hv_timer(struct kvm_lapic *apic)
 static bool start_hv_timer(struct kvm_lapic *apic)
 {
 	struct kvm_timer *ktimer = &apic->lapic_timer;
-	int r;
+	struct kvm_vcpu *vcpu = apic->vcpu;
+	bool expired;
 
 	WARN_ON(preemptible());
 	if (!kvm_x86_ops->set_hv_timer)
@@ -1693,8 +1694,7 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 	if (!ktimer->tscdeadline)
 		return false;
 
-	r = kvm_x86_ops->set_hv_timer(apic->vcpu, ktimer->tscdeadline);
-	if (r < 0)
+	if (kvm_x86_ops->set_hv_timer(vcpu, ktimer->tscdeadline, &expired))
 		return false;
 
 	ktimer->hv_timer_in_use = true;
@@ -1712,13 +1712,13 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 		 */
 		if (atomic_read(&ktimer->pending)) {
 			cancel_hv_timer(apic);
-		} else if (r) {
+		} else if (expired) {
 			apic_timer_expired(apic);
 			cancel_hv_timer(apic);
 		}
 	}
 
-	trace_kvm_hv_timer_state(apic->vcpu->vcpu_id, ktimer->hv_timer_in_use);
+	trace_kvm_hv_timer_state(vcpu->vcpu_id, ktimer->hv_timer_in_use);
 
 	return true;
 }
arch/x86/kvm/vmx/vmx.c +4 -2
index a3211e3ee679..f3c326e3c324 100644
@@ -7027,7 +7027,8 @@ static inline int u64_shl_div_u64(u64 a, unsigned int shift,
 	return 0;
 }
 
-static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
+static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc,
+			    bool *expired)
 {
 	struct vcpu_vmx *vmx;
 	u64 tscl, guest_tscl, delta_tsc, lapic_timer_advance_cycles;
@@ -7066,7 +7067,8 @@ static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc)
 		return -ERANGE;
 
 	vmx->hv_deadline_tsc = tscl + delta_tsc;
-	return delta_tsc == 0;
+	*expired = !delta_tsc;
+	return 0;
 }
 
 static void vmx_cancel_hv_timer(struct kvm_vcpu *vcpu)
#17 lore raw

PATCH v3 8/9 KVM: lapic: Check for a pending timer intr prior to start_hv_timer()

In reply to Sean Christopherson

Checking for a pending non-periodic interrupt in start_hv_timer() leads to restart_apic_timer() making an unnecessary call to start_sw_timer() due to start_hv_timer() returning false.

Alternatively, start_hv_timer() could return %true when there is a pending non-periodic interrupt, but that approach is less intuitive, i.e. would require a beefy comment to explain an otherwise simple check.

---
 arch/x86/kvm/lapic.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
arch/x86/kvm/lapic.c +5 -3
index 9a09a41000c3..32b05ca2f70a 100644
@@ -1688,9 +1688,6 @@ static bool start_hv_timer(struct kvm_lapic *apic)
 	if (!kvm_x86_ops->set_hv_timer)
 		return false;
 
-	if (!apic_lvtt_period(apic) && atomic_read(&ktimer->pending))
-		return false;
-
 	if (!ktimer->tscdeadline)
 		return false;
 
@@ -1743,8 +1740,13 @@ static void start_sw_timer(struct kvm_lapic *apic)
 static void restart_apic_timer(struct kvm_lapic *apic)
 {
 	preempt_disable();
+
+	if (!apic_lvtt_period(apic) && atomic_read(&apic->lapic_timer.pending))
+		goto out;
+
 	if (!start_hv_timer(apic))
 		start_sw_timer(apic);
+out:
 	preempt_enable();
 }
 
Signature · 1 line
2.21.0
#18 lore raw

PATCH v3 9/9 KVM: VMX: Skip delta_tsc shift-and-divide if the dividend is zero

In reply to Sean Christopherson

Ten percent of nothin' is... let me do the math here. Nothin' into nothin', carry the nothin'...

---
 arch/x86/kvm/vmx/vmx.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
arch/x86/kvm/vmx/vmx.c +2 -3
index f3c326e3c324..58961a35fe6f 100644
@@ -7051,10 +7051,9 @@ static int vmx_set_hv_timer(struct kvm_vcpu *vcpu, u64 guest_deadline_tsc,
 
 	/* Convert to host delta tsc if tsc scaling is enabled */
 	if (vcpu->arch.tsc_scaling_ratio != kvm_default_tsc_scaling_ratio &&
-			u64_shl_div_u64(delta_tsc,
+	    delta_tsc && u64_shl_div_u64(delta_tsc,
 				kvm_tsc_scaling_ratio_frac_bits,
-				vcpu->arch.tsc_scaling_ratio,
-				&delta_tsc))
+				vcpu->arch.tsc_scaling_ratio, &delta_tsc))
 		return -ERANGE;
 
 	/*
Signature · 1 line
2.21.0
maintainer
#19 lore raw
In reply to Sean Christopherson
Quoted 21 lines · Sean Christopherson

The recent change to automatically calculate lapic_timer_advance_ns introduced a handful of gnarly bugs, and also exposed a latent bug by virtue of the advancement logic being enabled by default. Inspection and debug revealed several other opportunities for minor improvements.

The primary issue is that the auto-tuning of lapic_timer_advance_ns is completely unbounded, e.g. there's nothing in the logic that prevents the advancement from creeping up to hundreds of milliseconds. Adjusting the timers by large amounts creates major discrepancies in the guest, e.g. a timer that was configured to fire after multiple milliseconds may arrive before the guest executes a single instruction. While technically correct from a time perspective, it breaks a reasonable assumption from the guest that it can execute some number of instructions between timer events.

The other major issue is that the advancement is global, while TSC scaling is done on a per-vCPU basic. Adjusting the advancement at runtime exacerbates this as there is no protection against multiple vCPUs and/or multiple VMs concurrently modifying the advancement value, e.g. it can effectively become corrupted or never stabilize due to getting bounced all over tarnation.

Quoted 5 lines

As for the latent bug, when timer advancement was applied to the hv_timer, i.e. the VMX preemption timer, the logic to trigger wait_for_lapic_timer() was not updated. As a result, a timer interrupt emulated via the hv_timer can easily arrive too early from a *time* perspective, as opposed to simply arriving early from a "number of instructions executed" perspective.

In a surprising twist of events, I've queued the last 5 patches. For the first four, I still plan to get them in 5.1 but I had some comments.
They will have to wait for next week since I'll have a (short) Easter
vacation, but there should be plenty of time.

Paolo