* [PATCH v6 1/3] KVM: make halt_poll_ns per-vCPU
[not found] <1441178971-3836-1-git-send-email-wanpeng.li@hotmail.com>
@ 2015-09-02 7:29 ` Wanpeng Li
2015-09-02 7:29 ` [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment Wanpeng Li
1 sibling, 0 replies; 12+ messages in thread
From: Wanpeng Li @ 2015-09-02 7:29 UTC (permalink / raw)
To: Paolo Bonzini; +Cc: David Matlack, kvm, linux-kernel, Wanpeng Li
Change halt_poll_ns into per-vCPU variable, seeded from module parameter,
to allow greater flexibility.
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
---
include/linux/kvm_host.h | 1 +
virt/kvm/kvm_main.c | 5 +++--
2 files changed, 4 insertions(+), 2 deletions(-)
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 81089cf..1bef9e2 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -242,6 +242,7 @@ struct kvm_vcpu {
int sigset_active;
sigset_t sigset;
struct kvm_vcpu_stat stat;
+ unsigned int halt_poll_ns;
#ifdef CONFIG_HAS_IOMEM
int mmio_needed;
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index d8db2f8f..c06e57c 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -217,6 +217,7 @@ int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
vcpu->kvm = kvm;
vcpu->vcpu_id = id;
vcpu->pid = NULL;
+ vcpu->halt_poll_ns = 0;
init_waitqueue_head(&vcpu->wq);
kvm_async_pf_vcpu_init(vcpu);
@@ -1930,8 +1931,8 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
bool waited = false;
start = cur = ktime_get();
- if (halt_poll_ns) {
- ktime_t stop = ktime_add_ns(ktime_get(), halt_poll_ns);
+ if (vcpu->halt_poll_ns) {
+ ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
do {
/*
--
1.9.1
^ permalink raw reply related [flat|nested] 12+ messages in thread
* [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
[not found] <1441178971-3836-1-git-send-email-wanpeng.li@hotmail.com>
2015-09-02 7:29 ` [PATCH v6 1/3] KVM: make halt_poll_ns per-vCPU Wanpeng Li
@ 2015-09-02 7:29 ` Wanpeng Li
2015-09-02 18:09 ` David Matlack
1 sibling, 1 reply; 12+ messages in thread
From: Wanpeng Li @ 2015-09-02 7:29 UTC (permalink / raw)
To: Paolo Bonzini; +Cc: David Matlack, kvm, linux-kernel, Wanpeng Li
There is a downside of always-poll since poll is still happened for idle
vCPUs which can waste cpu usage. This patch adds the ability to adjust
halt_poll_ns dynamically, to grow halt_poll_ns when shot halt is detected,
and to shrink halt_poll_ns when long halt is detected.
There are two new kernel parameters for changing the halt_poll_ns:
halt_poll_ns_grow and halt_poll_ns_shrink.
no-poll always-poll dynamic-poll
-----------------------------------------------------------------------
Idle (nohz) vCPU %c0 0.15% 0.3% 0.2%
Idle (250HZ) vCPU %c0 1.1% 4.6%~14% 1.2%
TCP_RR latency 34us 27us 26.7us
"Idle (X) vCPU %c0" is the percent of time the physical cpu spent in
c0 over 60 seconds (each vCPU is pinned to a pCPU). (nohz) means the
guest was tickless. (250HZ) means the guest was ticking at 250HZ.
The big win is with ticking operating systems. Running the linux guest
with nohz=off (and HZ=250), we save 3.4%~12.8% CPUs/second and get close
to no-polling overhead levels by using the dynamic-poll. The savings
should be even higher for higher frequency ticks.
Suggested-by: David Matlack <dmatlack@google.com>
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
---
virt/kvm/kvm_main.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++---
1 file changed, 58 insertions(+), 3 deletions(-)
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index c06e57c..3cff02f 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -66,9 +66,18 @@
MODULE_AUTHOR("Qumranet");
MODULE_LICENSE("GPL");
-static unsigned int halt_poll_ns;
+/* halt polling only reduces halt latency by 5-7 us, 500us is enough */
+static unsigned int halt_poll_ns = 500000;
module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
+/* Default doubles per-vcpu halt_poll_ns. */
+static unsigned int halt_poll_ns_grow = 2;
+module_param(halt_poll_ns_grow, int, S_IRUGO);
+
+/* Default resets per-vcpu halt_poll_ns . */
+static unsigned int halt_poll_ns_shrink;
+module_param(halt_poll_ns_shrink, int, S_IRUGO);
+
/*
* Ordering of locks:
*
@@ -1907,6 +1916,31 @@ void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
}
EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
+static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
+{
+ int val = vcpu->halt_poll_ns;
+
+ /* 10us base */
+ if (val == 0 && halt_poll_ns_grow)
+ val = 10000;
+ else
+ val *= halt_poll_ns_grow;
+
+ vcpu->halt_poll_ns = val;
+}
+
+static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
+{
+ int val = vcpu->halt_poll_ns;
+
+ if (halt_poll_ns_shrink == 0)
+ val = 0;
+ else
+ val /= halt_poll_ns_shrink;
+
+ vcpu->halt_poll_ns = val;
+}
+
static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
{
if (kvm_arch_vcpu_runnable(vcpu)) {
@@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
ktime_t start, cur;
DEFINE_WAIT(wait);
bool waited = false;
+ u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
start = cur = ktime_get();
if (vcpu->halt_poll_ns) {
@@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
*/
if (kvm_vcpu_check_block(vcpu) < 0) {
++vcpu->stat.halt_successful_poll;
- goto out;
+ break;
}
cur = ktime_get();
} while (single_task_running() && ktime_before(cur, stop));
+
+ if (ktime_before(cur, stop)) {
+ poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
+ goto out;
+ }
}
for (;;) {
@@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
finish_wait(&vcpu->wq, &wait);
cur = ktime_get();
+ wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
out:
- trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
+ block_ns = poll_ns + wait_ns;
+
+ if (halt_poll_ns) {
+ if (block_ns <= vcpu->halt_poll_ns)
+ ;
+ /* we had a long block, shrink polling */
+ else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
+ shrink_halt_poll_ns(vcpu);
+ /* we had a short halt and our poll time is too small */
+ else if (vcpu->halt_poll_ns < halt_poll_ns &&
+ block_ns < halt_poll_ns)
+ grow_halt_poll_ns(vcpu);
+ }
+
+ trace_kvm_vcpu_wakeup(block_ns, waited);
}
EXPORT_SYMBOL_GPL(kvm_vcpu_block);
--
1.9.1
^ permalink raw reply related [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 7:29 ` [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment Wanpeng Li
@ 2015-09-02 18:09 ` David Matlack
2015-09-02 19:12 ` Paolo Bonzini
` (2 more replies)
0 siblings, 3 replies; 12+ messages in thread
From: David Matlack @ 2015-09-02 18:09 UTC (permalink / raw)
To: Wanpeng Li; +Cc: Paolo Bonzini, kvm list, linux-kernel
On Wed, Sep 2, 2015 at 12:29 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
> There is a downside of always-poll since poll is still happened for idle
> vCPUs which can waste cpu usage. This patch adds the ability to adjust
> halt_poll_ns dynamically, to grow halt_poll_ns when shot halt is detected,
> and to shrink halt_poll_ns when long halt is detected.
>
> There are two new kernel parameters for changing the halt_poll_ns:
> halt_poll_ns_grow and halt_poll_ns_shrink.
>
> no-poll always-poll dynamic-poll
> -----------------------------------------------------------------------
> Idle (nohz) vCPU %c0 0.15% 0.3% 0.2%
> Idle (250HZ) vCPU %c0 1.1% 4.6%~14% 1.2%
> TCP_RR latency 34us 27us 26.7us
>
> "Idle (X) vCPU %c0" is the percent of time the physical cpu spent in
> c0 over 60 seconds (each vCPU is pinned to a pCPU). (nohz) means the
> guest was tickless. (250HZ) means the guest was ticking at 250HZ.
>
> The big win is with ticking operating systems. Running the linux guest
> with nohz=off (and HZ=250), we save 3.4%~12.8% CPUs/second and get close
> to no-polling overhead levels by using the dynamic-poll. The savings
> should be even higher for higher frequency ticks.
>
> Suggested-by: David Matlack <dmatlack@google.com>
> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
> ---
> virt/kvm/kvm_main.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++---
> 1 file changed, 58 insertions(+), 3 deletions(-)
>
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index c06e57c..3cff02f 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -66,9 +66,18 @@
> MODULE_AUTHOR("Qumranet");
> MODULE_LICENSE("GPL");
>
> -static unsigned int halt_poll_ns;
> +/* halt polling only reduces halt latency by 5-7 us, 500us is enough */
> +static unsigned int halt_poll_ns = 500000;
> module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
>
> +/* Default doubles per-vcpu halt_poll_ns. */
> +static unsigned int halt_poll_ns_grow = 2;
> +module_param(halt_poll_ns_grow, int, S_IRUGO);
> +
> +/* Default resets per-vcpu halt_poll_ns . */
> +static unsigned int halt_poll_ns_shrink;
> +module_param(halt_poll_ns_shrink, int, S_IRUGO);
> +
> /*
> * Ordering of locks:
> *
> @@ -1907,6 +1916,31 @@ void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
> }
> EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
>
> +static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
> +{
> + int val = vcpu->halt_poll_ns;
> +
> + /* 10us base */
> + if (val == 0 && halt_poll_ns_grow)
> + val = 10000;
> + else
> + val *= halt_poll_ns_grow;
> +
> + vcpu->halt_poll_ns = val;
> +}
> +
> +static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
> +{
> + int val = vcpu->halt_poll_ns;
> +
> + if (halt_poll_ns_shrink == 0)
> + val = 0;
> + else
> + val /= halt_poll_ns_shrink;
> +
> + vcpu->halt_poll_ns = val;
> +}
> +
> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
> {
> if (kvm_arch_vcpu_runnable(vcpu)) {
> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> ktime_t start, cur;
> DEFINE_WAIT(wait);
> bool waited = false;
> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>
> start = cur = ktime_get();
> if (vcpu->halt_poll_ns) {
> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> */
> if (kvm_vcpu_check_block(vcpu) < 0) {
> ++vcpu->stat.halt_successful_poll;
> - goto out;
> + break;
> }
> cur = ktime_get();
> } while (single_task_running() && ktime_before(cur, stop));
> +
> + if (ktime_before(cur, stop)) {
You can't use 'cur' to tell if the interrupt arrived. single_task_running()
can break us out of the loop before 'stop'.
> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
Put this line before the if(). block_ns should always include the time
spent polling; even if polling does not succeed.
> + goto out;
> + }
> }
>
> for (;;) {
> @@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>
> finish_wait(&vcpu->wq, &wait);
> cur = ktime_get();
> + wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>
> out:
> - trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
> + block_ns = poll_ns + wait_ns;
> +
> + if (halt_poll_ns) {
If you want, you can leave this if() out and save some indentation.
> + if (block_ns <= vcpu->halt_poll_ns)
> + ;
> + /* we had a long block, shrink polling */
> + else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
> + shrink_halt_poll_ns(vcpu);
> + /* we had a short halt and our poll time is too small */
> + else if (vcpu->halt_poll_ns < halt_poll_ns &&
> + block_ns < halt_poll_ns)
> + grow_halt_poll_ns(vcpu);
> + }
> +
> + trace_kvm_vcpu_wakeup(block_ns, waited);
> }
> EXPORT_SYMBOL_GPL(kvm_vcpu_block);
>
> --
> 1.9.1
>
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 18:09 ` David Matlack
@ 2015-09-02 19:12 ` Paolo Bonzini
2015-09-02 19:23 ` David Matlack
2015-09-03 9:23 ` Wanpeng Li
2015-09-04 1:16 ` Wanpeng Li
2 siblings, 1 reply; 12+ messages in thread
From: Paolo Bonzini @ 2015-09-02 19:12 UTC (permalink / raw)
To: David Matlack, Wanpeng Li; +Cc: kvm list, linux-kernel
On 02/09/2015 20:09, David Matlack wrote:
> On Wed, Sep 2, 2015 at 12:29 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
>> There is a downside of always-poll since poll is still happened for idle
>> vCPUs which can waste cpu usage. This patch adds the ability to adjust
>> halt_poll_ns dynamically, to grow halt_poll_ns when shot halt is detected,
>> and to shrink halt_poll_ns when long halt is detected.
>>
>> There are two new kernel parameters for changing the halt_poll_ns:
>> halt_poll_ns_grow and halt_poll_ns_shrink.
>>
>> no-poll always-poll dynamic-poll
>> -----------------------------------------------------------------------
>> Idle (nohz) vCPU %c0 0.15% 0.3% 0.2%
>> Idle (250HZ) vCPU %c0 1.1% 4.6%~14% 1.2%
>> TCP_RR latency 34us 27us 26.7us
>>
>> "Idle (X) vCPU %c0" is the percent of time the physical cpu spent in
>> c0 over 60 seconds (each vCPU is pinned to a pCPU). (nohz) means the
>> guest was tickless. (250HZ) means the guest was ticking at 250HZ.
>>
>> The big win is with ticking operating systems. Running the linux guest
>> with nohz=off (and HZ=250), we save 3.4%~12.8% CPUs/second and get close
>> to no-polling overhead levels by using the dynamic-poll. The savings
>> should be even higher for higher frequency ticks.
>>
>> Suggested-by: David Matlack <dmatlack@google.com>
>> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
>> ---
>> virt/kvm/kvm_main.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++---
>> 1 file changed, 58 insertions(+), 3 deletions(-)
>>
>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> index c06e57c..3cff02f 100644
>> --- a/virt/kvm/kvm_main.c
>> +++ b/virt/kvm/kvm_main.c
>> @@ -66,9 +66,18 @@
>> MODULE_AUTHOR("Qumranet");
>> MODULE_LICENSE("GPL");
>>
>> -static unsigned int halt_poll_ns;
>> +/* halt polling only reduces halt latency by 5-7 us, 500us is enough */
>> +static unsigned int halt_poll_ns = 500000;
>> module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
>>
>> +/* Default doubles per-vcpu halt_poll_ns. */
>> +static unsigned int halt_poll_ns_grow = 2;
>> +module_param(halt_poll_ns_grow, int, S_IRUGO);
>> +
>> +/* Default resets per-vcpu halt_poll_ns . */
>> +static unsigned int halt_poll_ns_shrink;
>> +module_param(halt_poll_ns_shrink, int, S_IRUGO);
>> +
>> /*
>> * Ordering of locks:
>> *
>> @@ -1907,6 +1916,31 @@ void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
>> }
>> EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
>>
>> +static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
>> +{
>> + int val = vcpu->halt_poll_ns;
>> +
>> + /* 10us base */
>> + if (val == 0 && halt_poll_ns_grow)
>> + val = 10000;
>> + else
>> + val *= halt_poll_ns_grow;
>> +
>> + vcpu->halt_poll_ns = val;
>> +}
>> +
>> +static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
>> +{
>> + int val = vcpu->halt_poll_ns;
>> +
>> + if (halt_poll_ns_shrink == 0)
>> + val = 0;
>> + else
>> + val /= halt_poll_ns_shrink;
>> +
>> + vcpu->halt_poll_ns = val;
>> +}
>> +
>> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
>> {
>> if (kvm_arch_vcpu_runnable(vcpu)) {
>> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> ktime_t start, cur;
>> DEFINE_WAIT(wait);
>> bool waited = false;
>> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>>
>> start = cur = ktime_get();
>> if (vcpu->halt_poll_ns) {
>> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> */
>> if (kvm_vcpu_check_block(vcpu) < 0) {
>> ++vcpu->stat.halt_successful_poll;
>> - goto out;
>> + break;
>> }
>> cur = ktime_get();
>> } while (single_task_running() && ktime_before(cur, stop));
>> +
>> + if (ktime_before(cur, stop)) {
>
> You can't use 'cur' to tell if the interrupt arrived. single_task_running()
> can break us out of the loop before 'stop'.
Ah, I thought this was on purpose. :)
If !single_task_running(), it is okay to keep vcpu->halt_poll_ns high,
because the physical CPU is not going to be idle anyway. Resetting the
timer as soon as single_task_running() becomes false will not cost much
CPU time.
Does it make sense?
Paolo
>> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>
> Put this line before the if(). block_ns should always include the time
> spent polling; even if polling does not succeed.
>
>> + goto out;
>> + }
>> }
>>
>> for (;;) {
>> @@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>
>> finish_wait(&vcpu->wq, &wait);
>> cur = ktime_get();
>> + wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>>
>> out:
>> - trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
>> + block_ns = poll_ns + wait_ns;
>> +
>> + if (halt_poll_ns) {
>
> If you want, you can leave this if() out and save some indentation.
>
>> + if (block_ns <= vcpu->halt_poll_ns)
>> + ;
>> + /* we had a long block, shrink polling */
>> + else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
>> + shrink_halt_poll_ns(vcpu);
>> + /* we had a short halt and our poll time is too small */
>> + else if (vcpu->halt_poll_ns < halt_poll_ns &&
>> + block_ns < halt_poll_ns)
>> + grow_halt_poll_ns(vcpu);
>> + }
>> +
>> + trace_kvm_vcpu_wakeup(block_ns, waited);
>> }
>> EXPORT_SYMBOL_GPL(kvm_vcpu_block);
>>
>> --
>> 1.9.1
>>
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 19:12 ` Paolo Bonzini
@ 2015-09-02 19:23 ` David Matlack
2015-09-03 7:31 ` Paolo Bonzini
0 siblings, 1 reply; 12+ messages in thread
From: David Matlack @ 2015-09-02 19:23 UTC (permalink / raw)
To: Paolo Bonzini; +Cc: Wanpeng Li, kvm list, linux-kernel
On Wed, Sep 2, 2015 at 12:12 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>
>
> On 02/09/2015 20:09, David Matlack wrote:
>> On Wed, Sep 2, 2015 at 12:29 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
>>> There is a downside of always-poll since poll is still happened for idle
>>> vCPUs which can waste cpu usage. This patch adds the ability to adjust
>>> halt_poll_ns dynamically, to grow halt_poll_ns when shot halt is detected,
>>> and to shrink halt_poll_ns when long halt is detected.
>>>
>>> There are two new kernel parameters for changing the halt_poll_ns:
>>> halt_poll_ns_grow and halt_poll_ns_shrink.
>>>
>>> no-poll always-poll dynamic-poll
>>> -----------------------------------------------------------------------
>>> Idle (nohz) vCPU %c0 0.15% 0.3% 0.2%
>>> Idle (250HZ) vCPU %c0 1.1% 4.6%~14% 1.2%
>>> TCP_RR latency 34us 27us 26.7us
>>>
>>> "Idle (X) vCPU %c0" is the percent of time the physical cpu spent in
>>> c0 over 60 seconds (each vCPU is pinned to a pCPU). (nohz) means the
>>> guest was tickless. (250HZ) means the guest was ticking at 250HZ.
>>>
>>> The big win is with ticking operating systems. Running the linux guest
>>> with nohz=off (and HZ=250), we save 3.4%~12.8% CPUs/second and get close
>>> to no-polling overhead levels by using the dynamic-poll. The savings
>>> should be even higher for higher frequency ticks.
>>>
>>> Suggested-by: David Matlack <dmatlack@google.com>
>>> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
>>> ---
>>> virt/kvm/kvm_main.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++---
>>> 1 file changed, 58 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>>> index c06e57c..3cff02f 100644
>>> --- a/virt/kvm/kvm_main.c
>>> +++ b/virt/kvm/kvm_main.c
>>> @@ -66,9 +66,18 @@
>>> MODULE_AUTHOR("Qumranet");
>>> MODULE_LICENSE("GPL");
>>>
>>> -static unsigned int halt_poll_ns;
>>> +/* halt polling only reduces halt latency by 5-7 us, 500us is enough */
>>> +static unsigned int halt_poll_ns = 500000;
>>> module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
>>>
>>> +/* Default doubles per-vcpu halt_poll_ns. */
>>> +static unsigned int halt_poll_ns_grow = 2;
>>> +module_param(halt_poll_ns_grow, int, S_IRUGO);
>>> +
>>> +/* Default resets per-vcpu halt_poll_ns . */
>>> +static unsigned int halt_poll_ns_shrink;
>>> +module_param(halt_poll_ns_shrink, int, S_IRUGO);
>>> +
>>> /*
>>> * Ordering of locks:
>>> *
>>> @@ -1907,6 +1916,31 @@ void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
>>> }
>>> EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
>>>
>>> +static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
>>> +{
>>> + int val = vcpu->halt_poll_ns;
>>> +
>>> + /* 10us base */
>>> + if (val == 0 && halt_poll_ns_grow)
>>> + val = 10000;
>>> + else
>>> + val *= halt_poll_ns_grow;
>>> +
>>> + vcpu->halt_poll_ns = val;
>>> +}
>>> +
>>> +static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
>>> +{
>>> + int val = vcpu->halt_poll_ns;
>>> +
>>> + if (halt_poll_ns_shrink == 0)
>>> + val = 0;
>>> + else
>>> + val /= halt_poll_ns_shrink;
>>> +
>>> + vcpu->halt_poll_ns = val;
>>> +}
>>> +
>>> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
>>> {
>>> if (kvm_arch_vcpu_runnable(vcpu)) {
>>> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>> ktime_t start, cur;
>>> DEFINE_WAIT(wait);
>>> bool waited = false;
>>> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>>>
>>> start = cur = ktime_get();
>>> if (vcpu->halt_poll_ns) {
>>> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>> */
>>> if (kvm_vcpu_check_block(vcpu) < 0) {
>>> ++vcpu->stat.halt_successful_poll;
>>> - goto out;
>>> + break;
>>> }
>>> cur = ktime_get();
>>> } while (single_task_running() && ktime_before(cur, stop));
>>> +
>>> + if (ktime_before(cur, stop)) {
>>
>> You can't use 'cur' to tell if the interrupt arrived. single_task_running()
>> can break us out of the loop before 'stop'.
>
> Ah, I thought this was on purpose. :)
>
> If !single_task_running(), it is okay to keep vcpu->halt_poll_ns high,
> because the physical CPU is not going to be idle anyway. Resetting the
> timer as soon as single_task_running() becomes false will not cost much
> CPU time.
Good point. I agree we can keep halt_poll_ns high in this case.
I actually wasn't thinking about vcpu->halt_poll_ns though. If
single_task_running() breaks us out of the loop we will "goto out" instead
of scheduling. My suspicion is this will cause us to loop calling
kvm_vcpu_block and starve the waiting task (at least until need_resched()),
which would break the "only hog the cpu when idle" aspect of halt-polling.
>
> Does it make sense?
>
> Paolo
>
>>> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>>
>> Put this line before the if(). block_ns should always include the time
>> spent polling; even if polling does not succeed.
>>
>>> + goto out;
>>> + }
>>> }
>>>
>>> for (;;) {
>>> @@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>>
>>> finish_wait(&vcpu->wq, &wait);
>>> cur = ktime_get();
>>> + wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>>>
>>> out:
>>> - trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
>>> + block_ns = poll_ns + wait_ns;
>>> +
>>> + if (halt_poll_ns) {
>>
>> If you want, you can leave this if() out and save some indentation.
>>
>>> + if (block_ns <= vcpu->halt_poll_ns)
>>> + ;
>>> + /* we had a long block, shrink polling */
>>> + else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
>>> + shrink_halt_poll_ns(vcpu);
>>> + /* we had a short halt and our poll time is too small */
>>> + else if (vcpu->halt_poll_ns < halt_poll_ns &&
>>> + block_ns < halt_poll_ns)
>>> + grow_halt_poll_ns(vcpu);
>>> + }
>>> +
>>> + trace_kvm_vcpu_wakeup(block_ns, waited);
>>> }
>>> EXPORT_SYMBOL_GPL(kvm_vcpu_block);
>>>
>>> --
>>> 1.9.1
>>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe kvm" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 19:23 ` David Matlack
@ 2015-09-03 7:31 ` Paolo Bonzini
2015-09-03 8:52 ` Wanpeng Li
0 siblings, 1 reply; 12+ messages in thread
From: Paolo Bonzini @ 2015-09-03 7:31 UTC (permalink / raw)
To: David Matlack; +Cc: Wanpeng Li, kvm list, linux-kernel
On 02/09/2015 21:23, David Matlack wrote:
>
> I actually wasn't thinking about vcpu->halt_poll_ns though. If
> single_task_running() breaks us out of the loop we will "goto out" instead
> of scheduling. My suspicion is this will cause us to loop calling
> kvm_vcpu_block and starve the waiting task (at least until need_resched()),
> which would break the "only hog the cpu when idle" aspect of halt-polling.
That's definitely a bug, yes.
Paolo
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-03 7:31 ` Paolo Bonzini
@ 2015-09-03 8:52 ` Wanpeng Li
0 siblings, 0 replies; 12+ messages in thread
From: Wanpeng Li @ 2015-09-03 8:52 UTC (permalink / raw)
To: Paolo Bonzini, David Matlack; +Cc: kvm list, linux-kernel
On 9/3/15 3:31 PM, Paolo Bonzini wrote:
>
> On 02/09/2015 21:23, David Matlack wrote:
>> I actually wasn't thinking about vcpu->halt_poll_ns though. If
>> single_task_running() breaks us out of the loop we will "goto out" instead
>> of scheduling. My suspicion is this will cause us to loop calling
>> kvm_vcpu_block and starve the waiting task (at least until need_resched()),
>> which would break the "only hog the cpu when idle" aspect of halt-polling.
> That's definitely a bug, yes.
Ok, I will send out v7 to fix this in this sunday since there is
vacation in my country currently.
Regards,
Wanpeng Li
>
> Paolo
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 18:09 ` David Matlack
2015-09-02 19:12 ` Paolo Bonzini
@ 2015-09-03 9:23 ` Wanpeng Li
2015-09-03 16:07 ` David Matlack
2015-09-04 1:16 ` Wanpeng Li
2 siblings, 1 reply; 12+ messages in thread
From: Wanpeng Li @ 2015-09-03 9:23 UTC (permalink / raw)
To: David Matlack; +Cc: Paolo Bonzini, kvm list, linux-kernel
On 9/3/15 2:09 AM, David Matlack wrote:
[...]
>> +
>> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
>> {
>> if (kvm_arch_vcpu_runnable(vcpu)) {
>> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> ktime_t start, cur;
>> DEFINE_WAIT(wait);
>> bool waited = false;
>> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>>
>> start = cur = ktime_get();
>> if (vcpu->halt_poll_ns) {
>> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> */
>> if (kvm_vcpu_check_block(vcpu) < 0) {
>> ++vcpu->stat.halt_successful_poll;
>> - goto out;
>> + break;
>> }
>> cur = ktime_get();
>> } while (single_task_running() && ktime_before(cur, stop));
>> +
>> + if (ktime_before(cur, stop)) {
> You can't use 'cur' to tell if the interrupt arrived. single_task_running()
> can break us out of the loop before 'stop'.
>
>> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
> Put this line before the if(). block_ns should always include the time
> spent polling; even if polling does not succeed.
How about something like:
@@ -1941,10 +1976,14 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
*/
if (kvm_vcpu_check_block(vcpu) < 0) {
++vcpu->stat.halt_successful_poll;
- goto out;
+ break;
}
cur = ktime_get();
} while (single_task_running() && ktime_before(cur, stop));
+
+ poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
+ if (ktime_before(cur, stop) && single_task_running())
+ goto out;
}
>
>> + goto out;
>> + }
>> }
>>
>> for (;;) {
>> @@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>
>> finish_wait(&vcpu->wq, &wait);
>> cur = ktime_get();
>> + wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>>
>> out:
>> - trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
>> + block_ns = poll_ns + wait_ns;
>> +
>> + if (halt_poll_ns) {
> If you want, you can leave this if() out and save some indentation.
Then we will miss the tracepoint.
Regards,
Wanpeng Li
>
>> + if (block_ns <= vcpu->halt_poll_ns)
>> + ;
>> + /* we had a long block, shrink polling */
>> + else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
>> + shrink_halt_poll_ns(vcpu);
>> + /* we had a short halt and our poll time is too small */
>> + else if (vcpu->halt_poll_ns < halt_poll_ns &&
>> + block_ns < halt_poll_ns)
>> + grow_halt_poll_ns(vcpu);
>> + }
>> +
>> + trace_kvm_vcpu_wakeup(block_ns, waited);
>> }
>> EXPORT_SYMBOL_GPL(kvm_vcpu_block);
>>
>> --
>> 1.9.1
>>
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-03 9:23 ` Wanpeng Li
@ 2015-09-03 16:07 ` David Matlack
2015-09-04 0:15 ` Wanpeng Li
0 siblings, 1 reply; 12+ messages in thread
From: David Matlack @ 2015-09-03 16:07 UTC (permalink / raw)
To: Wanpeng Li; +Cc: Paolo Bonzini, kvm list, linux-kernel
On Thu, Sep 3, 2015 at 2:23 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
>
> How about something like:
>
> @@ -1941,10 +1976,14 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> */
> if (kvm_vcpu_check_block(vcpu) < 0) {
> ++vcpu->stat.halt_successful_poll;
> - goto out;
> + break;
> }
> cur = ktime_get();
> } while (single_task_running() && ktime_before(cur, stop));
> +
> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
> + if (ktime_before(cur, stop) && single_task_running())
> + goto out;
I would prefer an explicit signal (e.g. set a bool to true before breaking out
of the loop, and check it here) to avoid duplicating the loop exit condition.
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-03 16:07 ` David Matlack
@ 2015-09-04 0:15 ` Wanpeng Li
0 siblings, 0 replies; 12+ messages in thread
From: Wanpeng Li @ 2015-09-04 0:15 UTC (permalink / raw)
To: David Matlack; +Cc: Paolo Bonzini, kvm list, linux-kernel
On 9/4/15 12:07 AM, David Matlack wrote:
> On Thu, Sep 3, 2015 at 2:23 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
>> How about something like:
>>
>> @@ -1941,10 +1976,14 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> */
>> if (kvm_vcpu_check_block(vcpu) < 0) {
>> ++vcpu->stat.halt_successful_poll;
>> - goto out;
>> + break;
>> }
>> cur = ktime_get();
>> } while (single_task_running() && ktime_before(cur, stop));
>> +
>> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>> + if (ktime_before(cur, stop) && single_task_running())
>> + goto out;
> I would prefer an explicit signal (e.g. set a bool to true before breaking out
> of the loop, and check it here) to avoid duplicating the loop exit condition.
Fix it in v7, thanks for your review, David! ;-)
Regards,
Wanpeng Li
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-02 18:09 ` David Matlack
2015-09-02 19:12 ` Paolo Bonzini
2015-09-03 9:23 ` Wanpeng Li
@ 2015-09-04 1:16 ` Wanpeng Li
2015-09-04 1:42 ` Wanpeng Li
2 siblings, 1 reply; 12+ messages in thread
From: Wanpeng Li @ 2015-09-04 1:16 UTC (permalink / raw)
To: David Matlack; +Cc: Paolo Bonzini, kvm list, linux-kernel
On 9/3/15 2:09 AM, David Matlack wrote:
> On Wed, Sep 2, 2015 at 12:29 AM, Wanpeng Li <wanpeng.li@hotmail.com> wrote:
>> There is a downside of always-poll since poll is still happened for idle
>> vCPUs which can waste cpu usage. This patch adds the ability to adjust
>> halt_poll_ns dynamically, to grow halt_poll_ns when shot halt is detected,
>> and to shrink halt_poll_ns when long halt is detected.
>>
>> There are two new kernel parameters for changing the halt_poll_ns:
>> halt_poll_ns_grow and halt_poll_ns_shrink.
>>
>> no-poll always-poll dynamic-poll
>> -----------------------------------------------------------------------
>> Idle (nohz) vCPU %c0 0.15% 0.3% 0.2%
>> Idle (250HZ) vCPU %c0 1.1% 4.6%~14% 1.2%
>> TCP_RR latency 34us 27us 26.7us
>>
>> "Idle (X) vCPU %c0" is the percent of time the physical cpu spent in
>> c0 over 60 seconds (each vCPU is pinned to a pCPU). (nohz) means the
>> guest was tickless. (250HZ) means the guest was ticking at 250HZ.
>>
>> The big win is with ticking operating systems. Running the linux guest
>> with nohz=off (and HZ=250), we save 3.4%~12.8% CPUs/second and get close
>> to no-polling overhead levels by using the dynamic-poll. The savings
>> should be even higher for higher frequency ticks.
>>
>> Suggested-by: David Matlack <dmatlack@google.com>
>> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
>> ---
>> virt/kvm/kvm_main.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++---
>> 1 file changed, 58 insertions(+), 3 deletions(-)
>>
>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> index c06e57c..3cff02f 100644
>> --- a/virt/kvm/kvm_main.c
>> +++ b/virt/kvm/kvm_main.c
>> @@ -66,9 +66,18 @@
>> MODULE_AUTHOR("Qumranet");
>> MODULE_LICENSE("GPL");
>>
>> -static unsigned int halt_poll_ns;
>> +/* halt polling only reduces halt latency by 5-7 us, 500us is enough */
>> +static unsigned int halt_poll_ns = 500000;
>> module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
>>
>> +/* Default doubles per-vcpu halt_poll_ns. */
>> +static unsigned int halt_poll_ns_grow = 2;
>> +module_param(halt_poll_ns_grow, int, S_IRUGO);
>> +
>> +/* Default resets per-vcpu halt_poll_ns . */
>> +static unsigned int halt_poll_ns_shrink;
>> +module_param(halt_poll_ns_shrink, int, S_IRUGO);
>> +
>> /*
>> * Ordering of locks:
>> *
>> @@ -1907,6 +1916,31 @@ void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
>> }
>> EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
>>
>> +static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
>> +{
>> + int val = vcpu->halt_poll_ns;
>> +
>> + /* 10us base */
>> + if (val == 0 && halt_poll_ns_grow)
>> + val = 10000;
>> + else
>> + val *= halt_poll_ns_grow;
>> +
>> + vcpu->halt_poll_ns = val;
>> +}
>> +
>> +static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
>> +{
>> + int val = vcpu->halt_poll_ns;
>> +
>> + if (halt_poll_ns_shrink == 0)
>> + val = 0;
>> + else
>> + val /= halt_poll_ns_shrink;
>> +
>> + vcpu->halt_poll_ns = val;
>> +}
>> +
>> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
>> {
>> if (kvm_arch_vcpu_runnable(vcpu)) {
>> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> ktime_t start, cur;
>> DEFINE_WAIT(wait);
>> bool waited = false;
>> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>>
>> start = cur = ktime_get();
>> if (vcpu->halt_poll_ns) {
>> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> */
>> if (kvm_vcpu_check_block(vcpu) < 0) {
>> ++vcpu->stat.halt_successful_poll;
>> - goto out;
>> + break;
>> }
>> cur = ktime_get();
>> } while (single_task_running() && ktime_before(cur, stop));
>> +
>> + if (ktime_before(cur, stop)) {
> You can't use 'cur' to tell if the interrupt arrived. single_task_running()
> can break us out of the loop before 'stop'.
>
>> + poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
> Put this line before the if(). block_ns should always include the time
> spent polling; even if polling does not succeed.
Ah, I remember why I keep poll_ns calculation in if(). Actually poll_ns
has already been accumulated into wait_ns(see how wait_ns is calculated)
if polling does not succeed. If put poll_ns calculation before if() I
need to reset variable 'start' to ktime_get().
Regards,
Wanpeng Li
>
>> + goto out;
>> + }
>> }
>>
>> for (;;) {
>> @@ -1959,9 +1999,24 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>
>> finish_wait(&vcpu->wq, &wait);
>> cur = ktime_get();
>> + wait_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>>
>> out:
>> - trace_kvm_vcpu_wakeup(ktime_to_ns(cur) - ktime_to_ns(start), waited);
>> + block_ns = poll_ns + wait_ns;
>> +
>> + if (halt_poll_ns) {
> If you want, you can leave this if() out and save some indentation.
>
>> + if (block_ns <= vcpu->halt_poll_ns)
>> + ;
>> + /* we had a long block, shrink polling */
>> + else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
>> + shrink_halt_poll_ns(vcpu);
>> + /* we had a short halt and our poll time is too small */
>> + else if (vcpu->halt_poll_ns < halt_poll_ns &&
>> + block_ns < halt_poll_ns)
>> + grow_halt_poll_ns(vcpu);
>> + }
>> +
>> + trace_kvm_vcpu_wakeup(block_ns, waited);
>> }
>> EXPORT_SYMBOL_GPL(kvm_vcpu_block);
>>
>> --
>> 1.9.1
>>
^ permalink raw reply [flat|nested] 12+ messages in thread
* Re: [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment
2015-09-04 1:16 ` Wanpeng Li
@ 2015-09-04 1:42 ` Wanpeng Li
0 siblings, 0 replies; 12+ messages in thread
From: Wanpeng Li @ 2015-09-04 1:42 UTC (permalink / raw)
To: David Matlack; +Cc: Paolo Bonzini, kvm list, linux-kernel
On 9/4/15 9:16 AM, Wanpeng Li wrote:
> [...]
>>> +
>>> static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
>>> {
>>> if (kvm_arch_vcpu_runnable(vcpu)) {
>>> @@ -1929,6 +1963,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>> ktime_t start, cur;
>>> DEFINE_WAIT(wait);
>>> bool waited = false;
>>> + u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
>>>
>>> start = cur = ktime_get();
>>> if (vcpu->halt_poll_ns) {
>>> @@ -1941,10 +1976,15 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>>> */
>>> if (kvm_vcpu_check_block(vcpu) < 0) {
>>> ++vcpu->stat.halt_successful_poll;
>>> - goto out;
>>> + break;
>>> }
>>> cur = ktime_get();
>>> } while (single_task_running() && ktime_before(cur,
>>> stop));
>>> +
>>> + if (ktime_before(cur, stop)) {
>> You can't use 'cur' to tell if the interrupt arrived.
>> single_task_running()
>> can break us out of the loop before 'stop'.
>>
>>> + poll_ns = ktime_to_ns(cur) -
>>> ktime_to_ns(start);
>> Put this line before the if(). block_ns should always include the time
>> spent polling; even if polling does not succeed.
>
> Ah, I remember why I keep poll_ns calculation in if(). Actually
> poll_ns has already been accumulated into wait_ns(see how wait_ns is
> calculated) if polling does not succeed. If put poll_ns calculation
> before if() I need to reset variable 'start' to ktime_get().
So how about something like:
@@ -1928,7 +1962,8 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
{
ktime_t start, cur;
DEFINE_WAIT(wait);
- bool waited = false;
+ bool polled = false, waited = false;
+ u64 poll_ns = 0, wait_ns = 0, block_ns = 0;
start = cur = ktime_get();
if (vcpu->halt_poll_ns) {
@@ -1940,11 +1975,17 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
* arrives.
*/
if (kvm_vcpu_check_block(vcpu) < 0) {
+ polled = true;
++vcpu->stat.halt_successful_poll;
- goto out;
+ break;
}
cur = ktime_get();
} while (single_task_running() && ktime_before(cur, stop));
+
+ if (polled) {
+ poll_ns = ktime_to_ns(cur) - ktime_to_ns(start);
+ goto out;
+ }
}
Regards,
Wanpeng Li
^ permalink raw reply [flat|nested] 12+ messages in thread
end of thread, other threads:[~2015-09-04 1:42 UTC | newest]
Thread overview: 12+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
[not found] <1441178971-3836-1-git-send-email-wanpeng.li@hotmail.com>
2015-09-02 7:29 ` [PATCH v6 1/3] KVM: make halt_poll_ns per-vCPU Wanpeng Li
2015-09-02 7:29 ` [PATCH v6 2/3] KVM: dynamic halt_poll_ns adjustment Wanpeng Li
2015-09-02 18:09 ` David Matlack
2015-09-02 19:12 ` Paolo Bonzini
2015-09-02 19:23 ` David Matlack
2015-09-03 7:31 ` Paolo Bonzini
2015-09-03 8:52 ` Wanpeng Li
2015-09-03 9:23 ` Wanpeng Li
2015-09-03 16:07 ` David Matlack
2015-09-04 0:15 ` Wanpeng Li
2015-09-04 1:16 ` Wanpeng Li
2015-09-04 1:42 ` Wanpeng Li
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.