[PATCH v3 0/2] pvclock: more code cleanup

[PATCH v3 0/2] pvclock: more code cleanup

[Paolo Bonzini]

Patch 1 simplifies the pvclock.h API by moving seqcount logic into
two new functions pvclock_read_begin and pvclock_read_retry.
Patch 2 uses the new simplified API in the vDSO.

Please ack the vDSO changes!

Paolo

v2->v3: add unlikely() around read_retry check.

Paolo Bonzini (2):
  pvclock: introduce seqcount-like API
  x86: vdso: use __pvclock_read_cycles

 arch/x86/entry/vdso/vclock_gettime.c | 25 +++++------------------
 arch/x86/include/asm/pvclock.h       | 39 +++++++++++++++++++++---------------
 arch/x86/kernel/pvclock.c            | 17 ++++++----------
 3 files changed, 34 insertions(+), 47 deletions(-)

-- 
1.8.3.1

[PATCH 1/2] pvclock: introduce seqcount-like API

[Paolo Bonzini]

The version field in struct pvclock_vcpu_time_info basically implements
a seqcount.  Wrap it with the usual read_begin and read_retry functions,
and use these APIs instead of peppering the code with smp_rmb()s.
While at it, change it to the more pedantically correct virt_rmb().

With this change, __pvclock_read_cycles can be simplified noticeably.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
	v2->v3: add unlikely() around read_retry check.

 arch/x86/entry/vdso/vclock_gettime.c |  9 ++-------
 arch/x86/include/asm/pvclock.h       | 39 +++++++++++++++++++++---------------
 arch/x86/kernel/pvclock.c            | 17 ++++++----------
 3 files changed, 31 insertions(+), 34 deletions(-)

diff --git a/arch/x86/entry/vdso/vclock_gettime.c b/arch/x86/entry/vdso/vclock_gettime.c
index 2f02d23a05ef..db1e3b4c3693 100644
--- a/arch/x86/entry/vdso/vclock_gettime.c
+++ b/arch/x86/entry/vdso/vclock_gettime.c
@@ -123,9 +123,7 @@ static notrace cycle_t vread_pvclock(int *mode)
 	 */
 
 	do {
-		version = pvti->version;
-
-		smp_rmb();
+		version = pvclock_read_begin(pvti);
 
 		if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
 			*mode = VCLOCK_NONE;
@@ -137,10 +135,7 @@ static notrace cycle_t vread_pvclock(int *mode)
 		pvti_tsc_shift = pvti->tsc_shift;
 		pvti_system_time = pvti->system_time;
 		pvti_tsc = pvti->tsc_timestamp;
-
-		/* Make sure that the version double-check is last. */
-		smp_rmb();
-	} while (unlikely((version & 1) || version != pvti->version));
+	} while (pvclock_read_retry(pvti, version));
 
 	delta = tsc - pvti_tsc;
 	ret = pvti_system_time +
diff --git a/arch/x86/include/asm/pvclock.h b/arch/x86/include/asm/pvclock.h
index 7c1c89598688..d019f0cc80ec 100644
--- a/arch/x86/include/asm/pvclock.h
+++ b/arch/x86/include/asm/pvclock.h
@@ -25,6 +25,24 @@ void pvclock_resume(void);
 
 void pvclock_touch_watchdogs(void);
 
+static __always_inline
+unsigned pvclock_read_begin(const struct pvclock_vcpu_time_info *src)
+{
+	unsigned version = src->version & ~1;
+	/* Make sure that the version is read before the data. */
+	virt_rmb();
+	return version;
+}
+
+static __always_inline
+bool pvclock_read_retry(const struct pvclock_vcpu_time_info *src,
+			unsigned version)
+{
+	/* Make sure that the version is re-read after the data. */
+	virt_rmb();
+	return unlikely(version != src->version);
+}
+
 /*
  * Scale a 64-bit delta by scaling and multiplying by a 32-bit fraction,
  * yielding a 64-bit result.
@@ -69,23 +87,12 @@ static inline u64 pvclock_scale_delta(u64 delta, u32 mul_frac, int shift)
 }
 
 static __always_inline
-unsigned __pvclock_read_cycles(const struct pvclock_vcpu_time_info *src,
-			       cycle_t *cycles, u8 *flags)
+cycle_t __pvclock_read_cycles(const struct pvclock_vcpu_time_info *src)
 {
-	unsigned version;
-	cycle_t offset;
-	u64 delta;
-
-	version = src->version;
-	/* Make the latest version visible */
-	smp_rmb();
-
-	delta = rdtsc_ordered() - src->tsc_timestamp;
-	offset = pvclock_scale_delta(delta, src->tsc_to_system_mul,
-				   src->tsc_shift);
-	*cycles = src->system_time + offset;
-	*flags = src->flags;
-	return version;
+	u64 delta = rdtsc_ordered() - src->tsc_timestamp;
+	cycle_t offset = pvclock_scale_delta(delta, src->tsc_to_system_mul,
+					     src->tsc_shift);
+	return src->system_time + offset;
 }
 
 struct pvclock_vsyscall_time_info {
diff --git a/arch/x86/kernel/pvclock.c b/arch/x86/kernel/pvclock.c
index 06c58ce46762..3599404e3089 100644
--- a/arch/x86/kernel/pvclock.c
+++ b/arch/x86/kernel/pvclock.c
@@ -64,14 +64,9 @@ u8 pvclock_read_flags(struct pvclock_vcpu_time_info *src)
 	u8 flags;
 
 	do {
-		version = src->version;
-		/* Make the latest version visible */
-		smp_rmb();
-
+		version = pvclock_read_begin(src);
 		flags = src->flags;
-		/* Make sure that the version double-check is last. */
-		smp_rmb();
-	} while ((src->version & 1) || version != src->version);
+	} while (pvclock_read_retry(src, version));
 
 	return flags & valid_flags;
 }
@@ -84,10 +79,10 @@ cycle_t pvclock_clocksource_read(struct pvclock_vcpu_time_info *src)
 	u8 flags;
 
 	do {
-		version = __pvclock_read_cycles(src, &ret, &flags);
-		/* Make sure that the version double-check is last. */
-		smp_rmb();
-	} while ((src->version & 1) || version != src->version);
+		version = pvclock_read_begin(src);
+		ret = __pvclock_read_cycles(src);
+		flags = src->flags;
+	} while (pvclock_read_retry(src, version));
 
 	if (unlikely((flags & PVCLOCK_GUEST_STOPPED) != 0)) {
 		src->flags &= ~PVCLOCK_GUEST_STOPPED;
-- 
1.8.3.1

[PATCH 2/2] x86: vdso: use __pvclock_read_cycles

[Paolo Bonzini]

The new simplified __pvclock_read_cycles does the same computation
as vread_pvclock, except that (because it takes the pvclock_vcpu_time_info
pointer) it has to be moved inside the loop.  Since the loop is expected to
never roll, this makes no difference.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 arch/x86/entry/vdso/vclock_gettime.c | 16 +++-------------
 1 file changed, 3 insertions(+), 13 deletions(-)

diff --git a/arch/x86/entry/vdso/vclock_gettime.c b/arch/x86/entry/vdso/vclock_gettime.c
index db1e3b4c3693..94d54d0defa7 100644
--- a/arch/x86/entry/vdso/vclock_gettime.c
+++ b/arch/x86/entry/vdso/vclock_gettime.c
@@ -96,9 +96,8 @@ static notrace cycle_t vread_pvclock(int *mode)
 {
 	const struct pvclock_vcpu_time_info *pvti = &get_pvti0()->pvti;
 	cycle_t ret;
-	u64 tsc, pvti_tsc;
-	u64 last, delta, pvti_system_time;
-	u32 version, pvti_tsc_to_system_mul, pvti_tsc_shift;
+	u64 last;
+	u32 version;
 
 	/*
 	 * Note: The kernel and hypervisor must guarantee that cpu ID
@@ -130,18 +129,9 @@ static notrace cycle_t vread_pvclock(int *mode)
 			return 0;
 		}
 
-		tsc = rdtsc_ordered();
-		pvti_tsc_to_system_mul = pvti->tsc_to_system_mul;
-		pvti_tsc_shift = pvti->tsc_shift;
-		pvti_system_time = pvti->system_time;
-		pvti_tsc = pvti->tsc_timestamp;
+		ret = __pvclock_read_cycles(pvti);
 	} while (pvclock_read_retry(pvti, version));
 
-	delta = tsc - pvti_tsc;
-	ret = pvti_system_time +
-		pvclock_scale_delta(delta, pvti_tsc_to_system_mul,
-				    pvti_tsc_shift);
-
 	/* refer to vread_tsc() comment for rationale */
 	last = gtod->cycle_last;
 
-- 
1.8.3.1

Re: [PATCH 1/2] pvclock: introduce seqcount-like API

[Andy Lutomirski]

I think you forget to pass -v3 to git format-patch :)

On Mon, Jun 27, 2016 at 7:19 AM, Paolo Bonzini <pbonzini@redhat.com> wrote:
> The version field in struct pvclock_vcpu_time_info basically implements
> a seqcount.  Wrap it with the usual read_begin and read_retry functions,
> and use these APIs instead of peppering the code with smp_rmb()s.
> While at it, change it to the more pedantically correct virt_rmb().
>
> With this change, __pvclock_read_cycles can be simplified noticeably.
>
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>         v2->v3: add unlikely() around read_retry check.
>
>  arch/x86/entry/vdso/vclock_gettime.c |  9 ++-------
>  arch/x86/include/asm/pvclock.h       | 39 +++++++++++++++++++++---------------
>  arch/x86/kernel/pvclock.c            | 17 ++++++----------
>  3 files changed, 31 insertions(+), 34 deletions(-)
>
> diff --git a/arch/x86/entry/vdso/vclock_gettime.c b/arch/x86/entry/vdso/vclock_gettime.c
> index 2f02d23a05ef..db1e3b4c3693 100644
> --- a/arch/x86/entry/vdso/vclock_gettime.c
> +++ b/arch/x86/entry/vdso/vclock_gettime.c
> @@ -123,9 +123,7 @@ static notrace cycle_t vread_pvclock(int *mode)
>          */
>
>         do {
> -               version = pvti->version;
> -
> -               smp_rmb();
> +               version = pvclock_read_begin(pvti);
>
>                 if (unlikely(!(pvti->flags & PVCLOCK_TSC_STABLE_BIT))) {
>                         *mode = VCLOCK_NONE;
> @@ -137,10 +135,7 @@ static notrace cycle_t vread_pvclock(int *mode)
>                 pvti_tsc_shift = pvti->tsc_shift;
>                 pvti_system_time = pvti->system_time;
>                 pvti_tsc = pvti->tsc_timestamp;
> -
> -               /* Make sure that the version double-check is last. */
> -               smp_rmb();
> -       } while (unlikely((version & 1) || version != pvti->version));
> +       } while (pvclock_read_retry(pvti, version));
>
>         delta = tsc - pvti_tsc;
>         ret = pvti_system_time +
> diff --git a/arch/x86/include/asm/pvclock.h b/arch/x86/include/asm/pvclock.h
> index 7c1c89598688..d019f0cc80ec 100644
> --- a/arch/x86/include/asm/pvclock.h
> +++ b/arch/x86/include/asm/pvclock.h
> @@ -25,6 +25,24 @@ void pvclock_resume(void);
>
>  void pvclock_touch_watchdogs(void);
>
> +static __always_inline
> +unsigned pvclock_read_begin(const struct pvclock_vcpu_time_info *src)
> +{
> +       unsigned version = src->version & ~1;

I see your cute optimization here....

> +       /* Make sure that the version is read before the data. */
> +       virt_rmb();
> +       return version;
> +}
> +
> +static __always_inline
> +bool pvclock_read_retry(const struct pvclock_vcpu_time_info *src,
> +                       unsigned version)
> +{
> +       /* Make sure that the version is re-read after the data. */
> +       virt_rmb();
> +       return unlikely(version != src->version);

...and here.

LGTM.

Re: [PATCH 2/2] x86: vdso: use __pvclock_read_cycles

[Andy Lutomirski]

On Mon, Jun 27, 2016 at 7:19 AM, Paolo Bonzini <pbonzini@redhat.com> wrote:
> The new simplified __pvclock_read_cycles does the same computation
> as vread_pvclock, except that (because it takes the pvclock_vcpu_time_info
> pointer) it has to be moved inside the loop.  Since the loop is expected to
> never roll, this makes no difference.

Acked-by: Andy Lutomirski <luto@kernel.org>