[dpdk-dev] [PATCH v6] eal/arm64: fix rdtsc precise version

[dpdk-dev] [PATCH v6] eal/arm64: fix rdtsc precise version

[Linhaifeng]

In order to get more accurate the cntvct_el0 reading,
SW must invoke isb and arch_counter_enforce_ordering.

Reference of linux kernel:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/
linux.git/tree/arch/arm64/include/asm/arch_timer.h?h=v5.5#n220

Fixes: ccad39ea0712 ("eal/arm: add cpu cycle operations for ARMv8")
Cc: stable@dpdk.org

Reviewed-by: Jerin Jacob <jerinjacobk@gmail.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Signed-off-by: Haifeng Lin <haifeng.lin@huawei.com>
---
 .../common/include/arch/arm/rte_cycles_64.h   | 26 +++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)

diff --git a/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h b/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
index 68e7c7338..6b0df5b0a 100644
--- a/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
+++ b/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
@@ -59,11 +59,33 @@ rte_rdtsc(void)
 }
 #endif
 
+static inline void
+isb(void)
+{
+	asm volatile("isb" : : : "memory");
+}
+
+static inline void
+__rte_arm64_cntvct_el0_enforce_ordering(uint64_t val)
+{
+	uint64_t tmp;
+
+	asm volatile(
+	"	eor	%0, %1, %1\n"
+	"	add	%0, sp, %0\n"
+	"	ldr	xzr, [%0]"
+	: "=r" (tmp) : "r" (val));
+}
+
 static inline uint64_t
 rte_rdtsc_precise(void)
 {
-	rte_mb();
-	return rte_rdtsc();
+	uint64_t tsc;
+
+	isb();
+	tsc = rte_rdtsc();
+	__rte_arm64_cntvct_el0_enforce_ordering(tsc);
+	return tsc;
 }
 
 static inline uint64_t
-- 
2.24.1.windows.2

Re: [dpdk-dev] [PATCH v6] eal/arm64: fix rdtsc precise version

[David Marchand]

On Wed, Mar 11, 2020 at 8:24 AM Linhaifeng <haifeng.lin@huawei.com> wrote:
>
> In order to get more accurate the cntvct_el0 reading,
> SW must invoke isb and arch_counter_enforce_ordering.
>
> Reference of linux kernel:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/
> linux.git/tree/arch/arm64/include/asm/arch_timer.h?h=v5.5#n220
>
> Fixes: ccad39ea0712 ("eal/arm: add cpu cycle operations for ARMv8")
> Cc: stable@dpdk.org
>
> Reviewed-by: Jerin Jacob <jerinjacobk@gmail.com>
> Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> Signed-off-by: Haifeng Lin <haifeng.lin@huawei.com>
> ---
>  .../common/include/arch/arm/rte_cycles_64.h   | 26 +++++++++++++++++--
>  1 file changed, 24 insertions(+), 2 deletions(-)
>
> diff --git a/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h b/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
> index 68e7c7338..6b0df5b0a 100644
> --- a/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
> +++ b/lib/librte_eal/common/include/arch/arm/rte_cycles_64.h
> @@ -59,11 +59,33 @@ rte_rdtsc(void)
>  }
>  #endif
>
> +static inline void
> +isb(void)
> +{
> +       asm volatile("isb" : : : "memory");
> +}

NAK.

Don't export badly named stuff like this.

> +
> +static inline void
> +__rte_arm64_cntvct_el0_enforce_ordering(uint64_t val)
> +{
> +       uint64_t tmp;
> +
> +       asm volatile(
> +       "       eor     %0, %1, %1\n"
> +       "       add     %0, sp, %0\n"
> +       "       ldr     xzr, [%0]"
> +       : "=r" (tmp) : "r" (val));
> +}
> +

I can see no point in exporting this.
Please move this inline of rte_rdtsc_precise().
If one day, ARM needs this elsewhere, we can reevaluate and introduce
a helper, but I don't see this atm.


>  static inline uint64_t
>  rte_rdtsc_precise(void)
>  {
> -       rte_mb();
> -       return rte_rdtsc();
> +       uint64_t tsc;
> +
> +       isb();
> +       tsc = rte_rdtsc();
> +       __rte_arm64_cntvct_el0_enforce_ordering(tsc);
> +       return tsc;
>  }
>
>  static inline uint64_t
> --
> 2.24.1.windows.2
>

-- 
David Marchand

Re: [dpdk-dev] [PATCH v6] eal/arm64: fix rdtsc precise version

[Linhaifeng]

> > +static inline void
> > +isb(void)
> > +{
> > +       asm volatile("isb" : : : "memory"); }
> 
> NAK.
> 
> Don't export badly named stuff like this.
> 

Just use asm volatile("isb" : : : "memory") in rte_rdtsc_precise or which file I should use to define this maco

> > +
> > +static inline void
> > +__rte_arm64_cntvct_el0_enforce_ordering(uint64_t val) {
> > +       uint64_t tmp;
> > +
> > +       asm volatile(
> > +       "       eor     %0, %1, %1\n"
> > +       "       add     %0, sp, %0\n"
> > +       "       ldr     xzr, [%0]"
> > +       : "=r" (tmp) : "r" (val));
> > +}
> > +
> 
> I can see no point in exporting this.
> Please move this inline of rte_rdtsc_precise().
> If one day, ARM needs this elsewhere, we can reevaluate and introduce a
> helper, but I don't see this atm.
> 
Ok. I will remove it. 
> 
> --
> David Marchand