Re: [dpdk-stable] [dpdk-dev] [PATCH] eal: generic counter based loop for CPU freq calculation

[Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>]

Hi Jerin,
	Thanks for the comments.

> -----Original Message-----
> From: Jerin Jacob <jerinjacobk@gmail.com>
> Sent: Wednesday, June 24, 2020 7:51 AM
> To: Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>
> Cc: dpdk-dev <dev@dpdk.org>; jerinj@marvell.com;
> hemant.agrawal@nxp.com; Akhil.goyal@nxp.com; ogerlitz@mellanox.com;
> Ajit Khaparde (ajit.khaparde@broadcom.com)
> <ajit.khaparde@broadcom.com>; ruigeng.wang@arm.com; Dharmik Thakkar
> <Dharmik.Thakkar@arm.com>; Phil Yang <Phil.Yang@arm.com>; dpdk stable
> <stable@dpdk.org>
> Subject: Re: [dpdk-dev] [PATCH] eal: generic counter based loop for CPU freq
> calculation
> 
> On Tue, Jun 9, 2020 at 3:04 AM Honnappa Nagarahalli
> <honnappa.nagarahalli@arm.com> wrote:
> >
> > get_tsc_freq uses 'nanosleep' system call to calculate the CPU
> > frequency. However, 'nanosleep' results in the process getting
> > un-scheduled. The kernel saves and restores the PMU state. This
> > ensures that the PMU cycles are not counted towards a sleeping
> > process. When RTE_ARM_EAL_RDTSC_USE_PMU is defined, this results in
> > incorrect CPU frequency calculation. This logic is replaced with
> > generic counter based loop.
> >
> > Bugzilla ID: 450
> > Fixes: af75078fece3 ("first public release")
> 
> The Fix looks good to me.
> 
> The Fixes is not correct. It should be the patch where
> RTE_ARM_EAL_RDTSC_USE_PMU got introduced.
Ok, will dig that out.

> 
> 
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> > Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > Reviewed-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
> > Reviewed-by: Phil Yang <phil.yang@arm.com>
> >
> > ---
> >  lib/librte_eal/arm/include/rte_cycles_64.h | 45 +++++++++++++++++++---
> >  lib/librte_eal/arm/rte_cycles.c            | 24 ++++++++++--
> >  2 files changed, 61 insertions(+), 8 deletions(-)
> >
> > diff --git a/lib/librte_eal/arm/include/rte_cycles_64.h
> > b/lib/librte_eal/arm/include/rte_cycles_64.h
> > index da557b6a1..6fc352036 100644
> > --- a/lib/librte_eal/arm/include/rte_cycles_64.h
> > +++ b/lib/librte_eal/arm/include/rte_cycles_64.h
> > @@ -11,6 +11,36 @@ extern "C" {
> >
> >  #include "generic/rte_cycles.h"
> >
> > +/** Read generic counter frequency */ static inline uint64_t
> 
> I prefer to have __rte_allways_inline
> 
> > +__rte_rd_generic_cntr_freq(void)
> 
> I think, the generic counter is confusing, I think, since the symbol is exposed
> due to placed in header file, it is better to change, __rte_arm64_cntfrq()
Ok, makes sense.

> 
> > +{
> > +       uint64_t freq;
> > +
> > +       asm volatile("mrs %0, cntfrq_el0" : "=r" (freq));
> > +       return freq;
> > +}
> > +
> > +/** Read generic counter */
> > +static inline uint64_t
> 
> Likewise, __rte_arm64_cntvct()
> 
> 
> > +__rte_rd_generic_cntr(void)
> > +{
> > +       uint64_t tsc;
> > +
> > +       asm volatile("mrs %0, cntvct_el0" : "=r" (tsc));
> > +       return tsc;
> > +}
> > +
> > +static inline uint64_t
> > +__rte_rd_generic_cntr_precise(void)
> 
> __rte_arm64_cntfrq_precise()
> 
> > +{
> > +       uint64_t tsc;
> > +
> > +       asm volatile("isb" : : : "memory");
> > +       asm volatile("mrs %0, cntvct_el0" : "=r" (tsc));
> > +       return tsc;
> > +}
> > +
> >  /**
> >   * Read the time base register.
> >   *
> > @@ -25,10 +55,7 @@ extern "C" {
> >  static inline uint64_t
> >  rte_rdtsc(void)
> >  {
> > -       uint64_t tsc;
> > -
> > -       asm volatile("mrs %0, cntvct_el0" : "=r" (tsc));
> > -       return tsc;
> > +       return __rte_rd_generic_cntr();
> >  }
> >  #else
> >  /**
> > @@ -49,14 +76,22 @@ rte_rdtsc(void)
> >   * asm volatile("msr pmcr_el0, %0" : : "r" (val));
> >   *
> >   */
> > +
> > +/** Read PMU cycle counter */
> >  static inline uint64_t
> > -rte_rdtsc(void)
> > +__rte_rd_pmu_cycle_cntr(void)
I will change this to __rte_arm64_pmccntr

> >  {
> >         uint64_t tsc;
> >
> >         asm volatile("mrs %0, pmccntr_el0" : "=r"(tsc));
> >         return tsc;
> >  }
> > +
> > +static inline uint64_t
> > +rte_rdtsc(void)
> > +{
> > +       return __rte_rd_pmu_cycle_cntr(); }
> >  #endif
> >
> >  static inline uint64_t
> > diff --git a/lib/librte_eal/arm/rte_cycles.c
> > b/lib/librte_eal/arm/rte_cycles.c index 3500d523e..92c87a8a4 100644
> > --- a/lib/librte_eal/arm/rte_cycles.c
> > +++ b/lib/librte_eal/arm/rte_cycles.c
> > @@ -3,14 +3,32 @@
> >   */
> >
> >  #include "eal_private.h"
> > +#include "rte_cycles.h"
> >
> >  uint64_t
> >  get_tsc_freq_arch(void)
> >  {
> >  #if defined RTE_ARCH_ARM64 && !defined
> RTE_ARM_EAL_RDTSC_USE_PMU
> > -       uint64_t freq;
> > -       asm volatile("mrs %0, cntfrq_el0" : "=r" (freq));
> > -       return freq;
> > +       return __rte_rd_generic_cntr_freq(); #elif defined
> > +RTE_ARCH_ARM64 && defined RTE_ARM_EAL_RDTSC_USE_PMU
> > +       /* Use the generic counter ticks to calculate the PMU
> > +        * cycle frequency.
> > +        */
> > +       uint64_t gcnt_ticks;
> > +       uint64_t start_ticks, cur_ticks;
> > +       uint64_t start_pmu_cycles, end_pmu_cycles;
> > +
> > +       /* Number of ticks for 1/10 second */
> > +       gcnt_ticks = __rte_rd_generic_cntr_freq() / 10;
> > +
> > +       start_ticks = __rte_rd_generic_cntr_precise();
> > +       start_pmu_cycles = rte_rdtsc_precise();
> > +       do {
> > +               cur_ticks = __rte_rd_generic_cntr();
> > +       } while ((cur_ticks - start_ticks) < gcnt_ticks);
> > +       end_pmu_cycles = rte_rdtsc_precise();
> > +
> > +       return ((end_pmu_cycles - start_pmu_cycles) * 10);
> 
> Good thought. On the plus side, it will reduce the boot time by .9 sec.
> 
> >  #else
> >         return 0;
> 
> With above changes:
> 
> Acked-by: Jerin Jacob <jerinj@marvell.com>
> 
> 
> 
> >  #endif
> > --
> > 2.17.1
> >