First patch tries to make the size of 'struct kvm_vcpu_pv_apf_data'
matched with its documentation.
Second patch fixes the wrong description of the MSR_KVM_ASYNC_PF_EN
documentation and some minor improvement.
v1: https://lore.kernel.org/all/[email protected]/T/#m0e12562199923ab58975d4ae9abaeb4a57597893
Xiaoyao Li (2):
x86/kvm/async_pf: Use separate percpu variable to track the enabling
of asyncpf
KVM: x86: Improve documentation of MSR_KVM_ASYNC_PF_EN
Documentation/virt/kvm/x86/msr.rst | 19 +++++++++----------
arch/x86/include/uapi/asm/kvm_para.h | 1 -
arch/x86/kernel/kvm.c | 11 ++++++-----
3 files changed, 15 insertions(+), 16 deletions(-)
base-commit: 2b3f2325e71f09098723727d665e2e8003d455dc
--
2.34.1
Fix some incorrect statement of MSR_KVM_ASYNC_PF_EN documentation and
state clearly the token in 'struct kvm_vcpu_pv_apf_data' of 'page ready'
event is matchted with the token in CR2 in 'page not present' event.
Signed-off-by: Xiaoyao Li <[email protected]>
---
Documentation/virt/kvm/x86/msr.rst | 18 +++++++++---------
1 file changed, 9 insertions(+), 9 deletions(-)
diff --git a/Documentation/virt/kvm/x86/msr.rst b/Documentation/virt/kvm/x86/msr.rst
index f6d70f99a1a7..3aecf2a70e7b 100644
--- a/Documentation/virt/kvm/x86/msr.rst
+++ b/Documentation/virt/kvm/x86/msr.rst
@@ -193,8 +193,8 @@ data:
Asynchronous page fault (APF) control MSR.
Bits 63-6 hold 64-byte aligned physical address of a 64 byte memory area
- which must be in guest RAM and must be zeroed. This memory is expected
- to hold a copy of the following structure::
+ which must be in guest RAM. This memory is expected to hold the
+ following structure::
struct kvm_vcpu_pv_apf_data {
/* Used for 'page not present' events delivered via #PF */
@@ -231,14 +231,14 @@ data:
as regular page fault, guest must reset 'flags' to '0' before it does
something that can generate normal page fault.
- Bytes 5-7 of 64 byte memory location ('token') will be written to by the
+ Bytes 4-7 of 64 byte memory location ('token') will be written to by the
hypervisor at the time of APF 'page ready' event injection. The content
- of these bytes is a token which was previously delivered as 'page not
- present' event. The event indicates the page in now available. Guest is
- supposed to write '0' to 'token' when it is done handling 'page ready'
- event and to write 1' to MSR_KVM_ASYNC_PF_ACK after clearing the location;
- writing to the MSR forces KVM to re-scan its queue and deliver the next
- pending notification.
+ of these bytes is a token which was previously delivered in CR2 as
+ 'page not present' event. The event indicates the page is now available.
+ Guest is supposed to write '0' to 'token' when it is done handling
+ 'page ready' event and to write '1' to MSR_KVM_ASYNC_PF_ACK after
+ clearing the location; writing to the MSR forces KVM to re-scan its
+ queue and deliver the next pending notification.
Note, MSR_KVM_ASYNC_PF_INT MSR specifying the interrupt vector for 'page
ready' APF delivery needs to be written to before enabling APF mechanism
--
2.34.1
Refer to commit fd10cde9294f ("KVM paravirt: Add async PF initialization
to PV guest") and commit 344d9588a9df ("KVM: Add PV MSR to enable
asynchronous page faults delivery"). It turns out that at the time when
asyncpf was introduced, the purpose was defining the shared PV data 'struct
kvm_vcpu_pv_apf_data' with the size of 64 bytes. However, it made a mistake
and defined the size to 68 bytes, which failed to make fit in a cache line
and made the code inconsistent with the documentation.
Below justification quoted from Sean[*]
KVM (the host side) has *never* read kvm_vcpu_pv_apf_data.enabled, and
the documentation clearly states that enabling is based solely on the
bit in the synthetic MSR.
So rather than update the documentation, fix the goof by removing the
enabled filed and use the separate percpu variable instread.
KVM-as-a-host obviously doesn't enforce anything or consume the size,
and changing the header will only affect guests that are rebuilt against
the new header, so there's no chance of ABI breakage between KVM and its
guests. The only possible breakage is if some other hypervisor is
emulating KVM's async #PF (LOL) and relies on the guest to set
kvm_vcpu_pv_apf_data.enabled. But (a) I highly doubt such a hypervisor
exists, (b) that would arguably be a violation of KVM's "spec", and
(c) the worst case scenario is that the guest would simply lose async
#PF functionality.
[*] https://lore.kernel.org/all/[email protected]/T/#u
Suggested-by: Sean Christopherson <[email protected]>
Signed-off-by: Xiaoyao Li <[email protected]>
---
Documentation/virt/kvm/x86/msr.rst | 1 -
arch/x86/include/uapi/asm/kvm_para.h | 1 -
arch/x86/kernel/kvm.c | 11 ++++++-----
3 files changed, 6 insertions(+), 7 deletions(-)
diff --git a/Documentation/virt/kvm/x86/msr.rst b/Documentation/virt/kvm/x86/msr.rst
index 9315fc385fb0..f6d70f99a1a7 100644
--- a/Documentation/virt/kvm/x86/msr.rst
+++ b/Documentation/virt/kvm/x86/msr.rst
@@ -204,7 +204,6 @@ data:
__u32 token;
__u8 pad[56];
- __u32 enabled;
};
Bits 5-4 of the MSR are reserved and should be zero. Bit 0 is set to 1
diff --git a/arch/x86/include/uapi/asm/kvm_para.h b/arch/x86/include/uapi/asm/kvm_para.h
index 6e64b27b2c1e..605899594ebb 100644
--- a/arch/x86/include/uapi/asm/kvm_para.h
+++ b/arch/x86/include/uapi/asm/kvm_para.h
@@ -142,7 +142,6 @@ struct kvm_vcpu_pv_apf_data {
__u32 token;
__u8 pad[56];
- __u32 enabled;
};
#define KVM_PV_EOI_BIT 0
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index b8ab9ee5896c..388a3fdd3cad 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
early_param("no-steal-acc", parse_no_stealacc);
+static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
static int has_steal_clock = 0;
@@ -244,7 +245,7 @@ noinstr u32 kvm_read_and_reset_apf_flags(void)
{
u32 flags = 0;
- if (__this_cpu_read(apf_reason.enabled)) {
+ if (__this_cpu_read(async_pf_enabled)) {
flags = __this_cpu_read(apf_reason.flags);
__this_cpu_write(apf_reason.flags, 0);
}
@@ -295,7 +296,7 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
inc_irq_stat(irq_hv_callback_count);
- if (__this_cpu_read(apf_reason.enabled)) {
+ if (__this_cpu_read(async_pf_enabled)) {
token = __this_cpu_read(apf_reason.token);
kvm_async_pf_task_wake(token);
__this_cpu_write(apf_reason.token, 0);
@@ -362,7 +363,7 @@ static void kvm_guest_cpu_init(void)
wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
- __this_cpu_write(apf_reason.enabled, 1);
+ __this_cpu_write(async_pf_enabled, 1);
pr_debug("setup async PF for cpu %d\n", smp_processor_id());
}
@@ -383,11 +384,11 @@ static void kvm_guest_cpu_init(void)
static void kvm_pv_disable_apf(void)
{
- if (!__this_cpu_read(apf_reason.enabled))
+ if (!__this_cpu_read(async_pf_enabled))
return;
wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
- __this_cpu_write(apf_reason.enabled, 0);
+ __this_cpu_write(async_pf_enabled, 0);
pr_debug("disable async PF for cpu %d\n", smp_processor_id());
}
--
2.34.1
Xiaoyao Li <[email protected]> writes:
> Refer to commit fd10cde9294f ("KVM paravirt: Add async PF initialization
> to PV guest") and commit 344d9588a9df ("KVM: Add PV MSR to enable
> asynchronous page faults delivery"). It turns out that at the time when
> asyncpf was introduced, the purpose was defining the shared PV data 'struct
> kvm_vcpu_pv_apf_data' with the size of 64 bytes. However, it made a mistake
> and defined the size to 68 bytes, which failed to make fit in a cache line
> and made the code inconsistent with the documentation.
Oh, I actually though it was done on purpose :-) 'enabled' is not
accessed by the host, it's only purpose is to cache MSR_KVM_ASYNC_PF_EN.
>
> Below justification quoted from Sean[*]
>
> KVM (the host side) has *never* read kvm_vcpu_pv_apf_data.enabled, and
> the documentation clearly states that enabling is based solely on the
> bit in the synthetic MSR.
>
> So rather than update the documentation, fix the goof by removing the
> enabled filed and use the separate percpu variable instread.
> KVM-as-a-host obviously doesn't enforce anything or consume the size,
> and changing the header will only affect guests that are rebuilt against
> the new header, so there's no chance of ABI breakage between KVM and its
> guests. The only possible breakage is if some other hypervisor is
> emulating KVM's async #PF (LOL) and relies on the guest to set
> kvm_vcpu_pv_apf_data.enabled. But (a) I highly doubt such a hypervisor
> exists, (b) that would arguably be a violation of KVM's "spec", and
> (c) the worst case scenario is that the guest would simply lose async
> #PF functionality.
>
> [*] https://lore.kernel.org/all/[email protected]/T/#u
>
> Suggested-by: Sean Christopherson <[email protected]>
> Signed-off-by: Xiaoyao Li <[email protected]>
> ---
> Documentation/virt/kvm/x86/msr.rst | 1 -
> arch/x86/include/uapi/asm/kvm_para.h | 1 -
> arch/x86/kernel/kvm.c | 11 ++++++-----
> 3 files changed, 6 insertions(+), 7 deletions(-)
>
> diff --git a/Documentation/virt/kvm/x86/msr.rst b/Documentation/virt/kvm/x86/msr.rst
> index 9315fc385fb0..f6d70f99a1a7 100644
> --- a/Documentation/virt/kvm/x86/msr.rst
> +++ b/Documentation/virt/kvm/x86/msr.rst
> @@ -204,7 +204,6 @@ data:
> __u32 token;
>
> __u8 pad[56];
> - __u32 enabled;
> };
>
> Bits 5-4 of the MSR are reserved and should be zero. Bit 0 is set to 1
> diff --git a/arch/x86/include/uapi/asm/kvm_para.h b/arch/x86/include/uapi/asm/kvm_para.h
> index 6e64b27b2c1e..605899594ebb 100644
> --- a/arch/x86/include/uapi/asm/kvm_para.h
> +++ b/arch/x86/include/uapi/asm/kvm_para.h
> @@ -142,7 +142,6 @@ struct kvm_vcpu_pv_apf_data {
> __u32 token;
>
> __u8 pad[56];
> - __u32 enabled;
> };
>
> #define KVM_PV_EOI_BIT 0
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index b8ab9ee5896c..388a3fdd3cad 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
>
> early_param("no-steal-acc", parse_no_stealacc);
>
> +static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
Would it make a difference is we replace this with a cpumask? I realize
that we need to access it on all CPUs from hotpaths but this mask will
rarely change so maybe there's no real perfomance hit?
> static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
> DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
> static int has_steal_clock = 0;
> @@ -244,7 +245,7 @@ noinstr u32 kvm_read_and_reset_apf_flags(void)
> {
> u32 flags = 0;
>
> - if (__this_cpu_read(apf_reason.enabled)) {
> + if (__this_cpu_read(async_pf_enabled)) {
> flags = __this_cpu_read(apf_reason.flags);
> __this_cpu_write(apf_reason.flags, 0);
> }
> @@ -295,7 +296,7 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
>
> inc_irq_stat(irq_hv_callback_count);
>
> - if (__this_cpu_read(apf_reason.enabled)) {
> + if (__this_cpu_read(async_pf_enabled)) {
> token = __this_cpu_read(apf_reason.token);
> kvm_async_pf_task_wake(token);
> __this_cpu_write(apf_reason.token, 0);
> @@ -362,7 +363,7 @@ static void kvm_guest_cpu_init(void)
> wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
>
> wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
> - __this_cpu_write(apf_reason.enabled, 1);
> + __this_cpu_write(async_pf_enabled, 1);
As 'async_pf_enabled' is bool, it would probably be more natural to
write
__this_cpu_write(async_pf_enabled, true);
> pr_debug("setup async PF for cpu %d\n", smp_processor_id());
> }
>
> @@ -383,11 +384,11 @@ static void kvm_guest_cpu_init(void)
>
> static void kvm_pv_disable_apf(void)
> {
> - if (!__this_cpu_read(apf_reason.enabled))
> + if (!__this_cpu_read(async_pf_enabled))
> return;
>
> wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
> - __this_cpu_write(apf_reason.enabled, 0);
> + __this_cpu_write(async_pf_enabled, 0);
... and 'false' here.
>
> pr_debug("disable async PF for cpu %d\n", smp_processor_id());
> }
--
Vitaly
On Wed, Oct 25, 2023, Vitaly Kuznetsov wrote:
> Xiaoyao Li <[email protected]> writes:
> > diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> > index b8ab9ee5896c..388a3fdd3cad 100644
> > --- a/arch/x86/kernel/kvm.c
> > +++ b/arch/x86/kernel/kvm.c
> > @@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
> >
> > early_param("no-steal-acc", parse_no_stealacc);
> >
> > +static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
>
> Would it make a difference is we replace this with a cpumask? I realize
> that we need to access it on all CPUs from hotpaths but this mask will
> rarely change so maybe there's no real perfomance hit?
FWIW, I personally prefer per-CPU booleans from a readability perspective. I
doubt there is a meaningful performance difference for a bitmap vs. individual
booleans, the check is already gated by a static key, i.e. kernels that are NOT
running as KVM guests don't care.
Actually, if there's performance gains to be had, optimizing kvm_read_and_reset_apf_flags()
to read the "enabled" flag if and only if it's necessary is a more likely candidate.
Assuming the host isn't being malicious/stupid, then apf_reason.flags will be '0'
if PV async #PFs are disabled. The only question is whether or not apf_reason.flags
is predictable enough for the CPU.
Aha! In practice, the CPU already needs to resolve a branch based on apf_reason.flags,
it's just "hidden" up in __kvm_handle_async_pf().
If we really want to micro-optimize, provide an __always_inline inner helper so
that __kvm_handle_async_pf() doesn't need to make a CALL just to read the flags.
Then in the common case where a #PF isn't due to the host swapping out a page,
the paravirt happy path doesn't need a taken branch and never reads the enabled
variable. E.g. the below generates:
0xffffffff81939ed0 <+0>: 41 54 push %r12
0xffffffff81939ed2 <+2>: 31 c0 xor %eax,%eax
0xffffffff81939ed4 <+4>: 55 push %rbp
0xffffffff81939ed5 <+5>: 53 push %rbx
0xffffffff81939ed6 <+6>: 48 83 ec 08 sub $0x8,%rsp
0xffffffff81939eda <+10>: 65 8b 2d df 81 6f 7e mov %gs:0x7e6f81df(%rip),%ebp # 0x320c0 <apf_reason>
0xffffffff81939ee1 <+17>: 85 ed test %ebp,%ebp
0xffffffff81939ee3 <+19>: 75 09 jne 0xffffffff81939eee <__kvm_handle_async_pf+30>
0xffffffff81939ee5 <+21>: 48 83 c4 08 add $0x8,%rsp
0xffffffff81939ee9 <+25>: 5b pop %rbx
0xffffffff81939eea <+26>: 5d pop %rbp
0xffffffff81939eeb <+27>: 41 5c pop %r12
0xffffffff81939eed <+29>: c3 ret
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index b8ab9ee5896c..b24133dc0731 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -240,22 +240,29 @@ void kvm_async_pf_task_wake(u32 token)
}
EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
-noinstr u32 kvm_read_and_reset_apf_flags(void)
+static __always_inline u32 __kvm_read_and_reset_apf_flags(void)
{
- u32 flags = 0;
+ u32 flags = __this_cpu_read(apf_reason.flags);
- if (__this_cpu_read(apf_reason.enabled)) {
- flags = __this_cpu_read(apf_reason.flags);
- __this_cpu_write(apf_reason.flags, 0);
+ if (unlikely(flags)) {
+ if (likely(__this_cpu_read(apf_reason.enabled)))
+ __this_cpu_write(apf_reason.flags, 0);
+ else
+ flags = 0;
}
return flags;
}
+
+u32 kvm_read_and_reset_apf_flags(void)
+{
+ return __kvm_read_and_reset_apf_flags();
+}
EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
{
- u32 flags = kvm_read_and_reset_apf_flags();
+ u32 flags = __kvm_read_and_reset_apf_flags();
irqentry_state_t state;
if (!flags)
> > static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
> > DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
> > static int has_steal_clock = 0;
> > @@ -244,7 +245,7 @@ noinstr u32 kvm_read_and_reset_apf_flags(void)
> > {
> > u32 flags = 0;
> >
> > - if (__this_cpu_read(apf_reason.enabled)) {
> > + if (__this_cpu_read(async_pf_enabled)) {
> > flags = __this_cpu_read(apf_reason.flags);
> > __this_cpu_write(apf_reason.flags, 0);
> > }
> > @@ -295,7 +296,7 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
> >
> > inc_irq_stat(irq_hv_callback_count);
> >
> > - if (__this_cpu_read(apf_reason.enabled)) {
> > + if (__this_cpu_read(async_pf_enabled)) {
> > token = __this_cpu_read(apf_reason.token);
> > kvm_async_pf_task_wake(token);
> > __this_cpu_write(apf_reason.token, 0);
> > @@ -362,7 +363,7 @@ static void kvm_guest_cpu_init(void)
> > wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
> >
> > wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
> > - __this_cpu_write(apf_reason.enabled, 1);
> > + __this_cpu_write(async_pf_enabled, 1);
>
> As 'async_pf_enabled' is bool, it would probably be more natural to
> write
>
> __this_cpu_write(async_pf_enabled, true);
+1000
On 10/25/2023 5:10 PM, Vitaly Kuznetsov wrote:
> Xiaoyao Li <[email protected]> writes:
>
>> Refer to commit fd10cde9294f ("KVM paravirt: Add async PF initialization
>> to PV guest") and commit 344d9588a9df ("KVM: Add PV MSR to enable
>> asynchronous page faults delivery"). It turns out that at the time when
>> asyncpf was introduced, the purpose was defining the shared PV data 'struct
>> kvm_vcpu_pv_apf_data' with the size of 64 bytes. However, it made a mistake
>> and defined the size to 68 bytes, which failed to make fit in a cache line
>> and made the code inconsistent with the documentation.
>
> Oh, I actually though it was done on purpose :-) 'enabled' is not
> accessed by the host, it's only purpose is to cache MSR_KVM_ASYNC_PF_EN.
I didn't find any clue to show it was on purpose, so thought it was a
mistake. Anyway, if the fact is it was done on purpose and people now
still accept it. I can drop this patch, and write another one to
document it's intentional instead.
>>
>> Below justification quoted from Sean[*]
>>
>> KVM (the host side) has *never* read kvm_vcpu_pv_apf_data.enabled, and
>> the documentation clearly states that enabling is based solely on the
>> bit in the synthetic MSR.
>>
>> So rather than update the documentation, fix the goof by removing the
>> enabled filed and use the separate percpu variable instread.
>> KVM-as-a-host obviously doesn't enforce anything or consume the size,
>> and changing the header will only affect guests that are rebuilt against
>> the new header, so there's no chance of ABI breakage between KVM and its
>> guests. The only possible breakage is if some other hypervisor is
>> emulating KVM's async #PF (LOL) and relies on the guest to set
>> kvm_vcpu_pv_apf_data.enabled. But (a) I highly doubt such a hypervisor
>> exists, (b) that would arguably be a violation of KVM's "spec", and
>> (c) the worst case scenario is that the guest would simply lose async
>> #PF functionality.
>>
>> [*] https://lore.kernel.org/all/[email protected]/T/#u
>>
>> Suggested-by: Sean Christopherson <[email protected]>
>> Signed-off-by: Xiaoyao Li <[email protected]>
>> ---
>> Documentation/virt/kvm/x86/msr.rst | 1 -
>> arch/x86/include/uapi/asm/kvm_para.h | 1 -
>> arch/x86/kernel/kvm.c | 11 ++++++-----
>> 3 files changed, 6 insertions(+), 7 deletions(-)
>>
>> diff --git a/Documentation/virt/kvm/x86/msr.rst b/Documentation/virt/kvm/x86/msr.rst
>> index 9315fc385fb0..f6d70f99a1a7 100644
>> --- a/Documentation/virt/kvm/x86/msr.rst
>> +++ b/Documentation/virt/kvm/x86/msr.rst
>> @@ -204,7 +204,6 @@ data:
>> __u32 token;
>>
>> __u8 pad[56];
>> - __u32 enabled;
>> };
>>
>> Bits 5-4 of the MSR are reserved and should be zero. Bit 0 is set to 1
>> diff --git a/arch/x86/include/uapi/asm/kvm_para.h b/arch/x86/include/uapi/asm/kvm_para.h
>> index 6e64b27b2c1e..605899594ebb 100644
>> --- a/arch/x86/include/uapi/asm/kvm_para.h
>> +++ b/arch/x86/include/uapi/asm/kvm_para.h
>> @@ -142,7 +142,6 @@ struct kvm_vcpu_pv_apf_data {
>> __u32 token;
>>
>> __u8 pad[56];
>> - __u32 enabled;
>> };
>>
>> #define KVM_PV_EOI_BIT 0
>> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
>> index b8ab9ee5896c..388a3fdd3cad 100644
>> --- a/arch/x86/kernel/kvm.c
>> +++ b/arch/x86/kernel/kvm.c
>> @@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
>>
>> early_param("no-steal-acc", parse_no_stealacc);
>>
>> +static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
>
> Would it make a difference is we replace this with a cpumask? I realize
> that we need to access it on all CPUs from hotpaths but this mask will
> rarely change so maybe there's no real perfomance hit?
>
>> static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
>> DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
>> static int has_steal_clock = 0;
>> @@ -244,7 +245,7 @@ noinstr u32 kvm_read_and_reset_apf_flags(void)
>> {
>> u32 flags = 0;
>>
>> - if (__this_cpu_read(apf_reason.enabled)) {
>> + if (__this_cpu_read(async_pf_enabled)) {
>> flags = __this_cpu_read(apf_reason.flags);
>> __this_cpu_write(apf_reason.flags, 0);
>> }
>> @@ -295,7 +296,7 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
>>
>> inc_irq_stat(irq_hv_callback_count);
>>
>> - if (__this_cpu_read(apf_reason.enabled)) {
>> + if (__this_cpu_read(async_pf_enabled)) {
>> token = __this_cpu_read(apf_reason.token);
>> kvm_async_pf_task_wake(token);
>> __this_cpu_write(apf_reason.token, 0);
>> @@ -362,7 +363,7 @@ static void kvm_guest_cpu_init(void)
>> wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
>>
>> wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
>> - __this_cpu_write(apf_reason.enabled, 1);
>> + __this_cpu_write(async_pf_enabled, 1);
>
> As 'async_pf_enabled' is bool, it would probably be more natural to
> write
>
> __this_cpu_write(async_pf_enabled, true);
>
>> pr_debug("setup async PF for cpu %d\n", smp_processor_id());
>> }
>>
>> @@ -383,11 +384,11 @@ static void kvm_guest_cpu_init(void)
>>
>> static void kvm_pv_disable_apf(void)
>> {
>> - if (!__this_cpu_read(apf_reason.enabled))
>> + if (!__this_cpu_read(async_pf_enabled))
>> return;
>>
>> wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
>> - __this_cpu_write(apf_reason.enabled, 0);
>> + __this_cpu_write(async_pf_enabled, 0);
>
> ... and 'false' here.
sure, I can do it in a v3, if v3 is needed.
>>
>> pr_debug("disable async PF for cpu %d\n", smp_processor_id());
>> }
>
On 10/25/2023 10:22 PM, Sean Christopherson wrote:
> On Wed, Oct 25, 2023, Vitaly Kuznetsov wrote:
>> Xiaoyao Li <[email protected]> writes:
>>> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
>>> index b8ab9ee5896c..388a3fdd3cad 100644
>>> --- a/arch/x86/kernel/kvm.c
>>> +++ b/arch/x86/kernel/kvm.c
>>> @@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
>>>
>>> early_param("no-steal-acc", parse_no_stealacc);
>>>
>>> +static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
>>
>> Would it make a difference is we replace this with a cpumask? I realize
>> that we need to access it on all CPUs from hotpaths but this mask will
>> rarely change so maybe there's no real perfomance hit?
>
> FWIW, I personally prefer per-CPU booleans from a readability perspective. I
> doubt there is a meaningful performance difference for a bitmap vs. individual
> booleans, the check is already gated by a static key, i.e. kernels that are NOT
> running as KVM guests don't care.
I agree with it.
> Actually, if there's performance gains to be had, optimizing kvm_read_and_reset_apf_flags()
> to read the "enabled" flag if and only if it's necessary is a more likely candidate.
> Assuming the host isn't being malicious/stupid, then apf_reason.flags will be '0'
> if PV async #PFs are disabled. The only question is whether or not apf_reason.flags
> is predictable enough for the CPU.
>
> Aha! In practice, the CPU already needs to resolve a branch based on apf_reason.flags,
> it's just "hidden" up in __kvm_handle_async_pf().
>
> If we really want to micro-optimize, provide an __always_inline inner helper so
> that __kvm_handle_async_pf() doesn't need to make a CALL just to read the flags.
> Then in the common case where a #PF isn't due to the host swapping out a page,
> the paravirt happy path doesn't need a taken branch and never reads the enabled
> variable. E.g. the below generates:
If this is wanted. It can be a separate patch, irrelevant with this
series, I think.
> 0xffffffff81939ed0 <+0>: 41 54 push %r12
> 0xffffffff81939ed2 <+2>: 31 c0 xor %eax,%eax
> 0xffffffff81939ed4 <+4>: 55 push %rbp
> 0xffffffff81939ed5 <+5>: 53 push %rbx
> 0xffffffff81939ed6 <+6>: 48 83 ec 08 sub $0x8,%rsp
> 0xffffffff81939eda <+10>: 65 8b 2d df 81 6f 7e mov %gs:0x7e6f81df(%rip),%ebp # 0x320c0 <apf_reason>
> 0xffffffff81939ee1 <+17>: 85 ed test %ebp,%ebp
> 0xffffffff81939ee3 <+19>: 75 09 jne 0xffffffff81939eee <__kvm_handle_async_pf+30>
> 0xffffffff81939ee5 <+21>: 48 83 c4 08 add $0x8,%rsp
> 0xffffffff81939ee9 <+25>: 5b pop %rbx
> 0xffffffff81939eea <+26>: 5d pop %rbp
> 0xffffffff81939eeb <+27>: 41 5c pop %r12
> 0xffffffff81939eed <+29>: c3 ret
>
>
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index b8ab9ee5896c..b24133dc0731 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -240,22 +240,29 @@ void kvm_async_pf_task_wake(u32 token)
> }
> EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
>
> -noinstr u32 kvm_read_and_reset_apf_flags(void)
> +static __always_inline u32 __kvm_read_and_reset_apf_flags(void)
> {
> - u32 flags = 0;
> + u32 flags = __this_cpu_read(apf_reason.flags);
>
> - if (__this_cpu_read(apf_reason.enabled)) {
> - flags = __this_cpu_read(apf_reason.flags);
> - __this_cpu_write(apf_reason.flags, 0);
> + if (unlikely(flags)) {
> + if (likely(__this_cpu_read(apf_reason.enabled)))
> + __this_cpu_write(apf_reason.flags, 0);
> + else
> + flags = 0;
> }
>
> return flags;
> }
> +
> +u32 kvm_read_and_reset_apf_flags(void)
> +{
> + return __kvm_read_and_reset_apf_flags();
> +}
> EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
>
> noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
> {
> - u32 flags = kvm_read_and_reset_apf_flags();
> + u32 flags = __kvm_read_and_reset_apf_flags();
> irqentry_state_t state;
>
> if (!flags)
>
>>> static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
>>> DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
>>> static int has_steal_clock = 0;
>>> @@ -244,7 +245,7 @@ noinstr u32 kvm_read_and_reset_apf_flags(void)
>>> {
>>> u32 flags = 0;
>>>
>>> - if (__this_cpu_read(apf_reason.enabled)) {
>>> + if (__this_cpu_read(async_pf_enabled)) {
>>> flags = __this_cpu_read(apf_reason.flags);
>>> __this_cpu_write(apf_reason.flags, 0);
>>> }
>>> @@ -295,7 +296,7 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
>>>
>>> inc_irq_stat(irq_hv_callback_count);
>>>
>>> - if (__this_cpu_read(apf_reason.enabled)) {
>>> + if (__this_cpu_read(async_pf_enabled)) {
>>> token = __this_cpu_read(apf_reason.token);
>>> kvm_async_pf_task_wake(token);
>>> __this_cpu_write(apf_reason.token, 0);
>>> @@ -362,7 +363,7 @@ static void kvm_guest_cpu_init(void)
>>> wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
>>>
>>> wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
>>> - __this_cpu_write(apf_reason.enabled, 1);
>>> + __this_cpu_write(async_pf_enabled, 1);
>>
>> As 'async_pf_enabled' is bool, it would probably be more natural to
>> write
>>
>> __this_cpu_write(async_pf_enabled, true);
>
> +1000
On Mon, Oct 30, 2023, Xiaoyao Li wrote:
> On 10/25/2023 10:22 PM, Sean Christopherson wrote:
> > On Wed, Oct 25, 2023, Vitaly Kuznetsov wrote:
> > > Xiaoyao Li <[email protected]> writes:
> > > > diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> > > > index b8ab9ee5896c..388a3fdd3cad 100644
> > > > --- a/arch/x86/kernel/kvm.c
> > > > +++ b/arch/x86/kernel/kvm.c
> > > > @@ -65,6 +65,7 @@ static int __init parse_no_stealacc(char *arg)
> > > > early_param("no-steal-acc", parse_no_stealacc);
> > > > +static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
> > >
> > > Would it make a difference is we replace this with a cpumask? I realize
> > > that we need to access it on all CPUs from hotpaths but this mask will
> > > rarely change so maybe there's no real perfomance hit?
> >
> > FWIW, I personally prefer per-CPU booleans from a readability perspective. I
> > doubt there is a meaningful performance difference for a bitmap vs. individual
> > booleans, the check is already gated by a static key, i.e. kernels that are NOT
> > running as KVM guests don't care.
>
> I agree with it.
>
> > Actually, if there's performance gains to be had, optimizing kvm_read_and_reset_apf_flags()
> > to read the "enabled" flag if and only if it's necessary is a more likely candidate.
> > Assuming the host isn't being malicious/stupid, then apf_reason.flags will be '0'
> > if PV async #PFs are disabled. The only question is whether or not apf_reason.flags
> > is predictable enough for the CPU.
> >
> > Aha! In practice, the CPU already needs to resolve a branch based on apf_reason.flags,
> > it's just "hidden" up in __kvm_handle_async_pf().
> >
> > If we really want to micro-optimize, provide an __always_inline inner helper so
> > that __kvm_handle_async_pf() doesn't need to make a CALL just to read the flags.
> > Then in the common case where a #PF isn't due to the host swapping out a page,
> > the paravirt happy path doesn't need a taken branch and never reads the enabled
> > variable. E.g. the below generates:
>
> If this is wanted. It can be a separate patch, irrelevant with this series,
> I think.
Yes, it's definitely beyond the scope of this series.
On Wed, 25 Oct 2023 01:59:12 -0400, Xiaoyao Li wrote:
> First patch tries to make the size of 'struct kvm_vcpu_pv_apf_data'
> matched with its documentation.
>
> Second patch fixes the wrong description of the MSR_KVM_ASYNC_PF_EN
> documentation and some minor improvement.
>
> v1: https://lore.kernel.org/all/[email protected]/T/#m0e12562199923ab58975d4ae9abaeb4a57597893
>
> [...]
Applied to kvm-x86 asyncpf_abi. I'll send a pull request (for 6.9) to Paolo
"soon" to ensure we get his eyeballs on the ABI change.
[1/2] x86/kvm/async_pf: Use separate percpu variable to track the enabling of asyncpf
https://github.com/kvm-x86/linux/commit/ccb2280ec2f9
[2/2] KVM: x86: Improve documentation of MSR_KVM_ASYNC_PF_EN
https://github.com/kvm-x86/linux/commit/df01f0a1165c
--
https://github.com/kvm-x86/linux/tree/next
On 2/7/2024 5:36 AM, Sean Christopherson wrote:
> On Wed, 25 Oct 2023 01:59:12 -0400, Xiaoyao Li wrote:
>> First patch tries to make the size of 'struct kvm_vcpu_pv_apf_data'
>> matched with its documentation.
>>
>> Second patch fixes the wrong description of the MSR_KVM_ASYNC_PF_EN
>> documentation and some minor improvement.
>>
>> v1: https://lore.kernel.org/all/[email protected]/T/#m0e12562199923ab58975d4ae9abaeb4a57597893
>>
>> [...]
>
> Applied to kvm-x86 asyncpf_abi. I'll send a pull request (for 6.9) to Paolo
> "soon" to ensure we get his eyeballs on the ABI change.
Thanks!
> [1/2] x86/kvm/async_pf: Use separate percpu variable to track the enabling of asyncpf
> https://github.com/kvm-x86/linux/commit/ccb2280ec2f9
> [2/2] KVM: x86: Improve documentation of MSR_KVM_ASYNC_PF_EN
> https://github.com/kvm-x86/linux/commit/df01f0a1165c
>
> --
> https://github.com/kvm-x86/linux/tree/next