Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <cover.1593397455.git.zong.li@sifive.com> <8b9f52c19bdb11a4ad741ad1a3695526a1d061b8.1593397455.git.zong.li@sifive.com>
 <CAAhSdy1Jbvs1+X5A_ej39YOevioUA7f3o2fOi=VT_AQWBgLjXw@mail.gmail.com>
In-Reply-To: <CAAhSdy1Jbvs1+X5A_ej39YOevioUA7f3o2fOi=VT_AQWBgLjXw@mail.gmail.com>
From:   Zong Li <zong.li@sifive.com>
Date:   Mon, 29 Jun 2020 12:35:40 +0800
Message-ID: <CANXhq0oeDty5ec9So5cR1W3qoLgHXbMXrb_POoRYoG1E3Eiw9w@mail.gmail.com>
Subject: Re: [RFC PATCH 4/6] riscv: perf: Add raw event support
To:     Anup Patel <anup@brainfault.org>
Cc:     Palmer Dabbelt <palmer@dabbelt.com>,
        Paul Walmsley <paul.walmsley@sifive.com>,
        linux-riscv <linux-riscv@lists.infradead.org>,
        "linux-kernel@vger.kernel.org List" <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Jun 29, 2020 at 12:17 PM Anup Patel <anup@brainfault.org> wrote:
>
> On Mon, Jun 29, 2020 at 8:49 AM Zong Li <zong.li@sifive.com> wrote:
> >
> > Add support for raw events and hardware cache events. Currently, we set
> > the events by writing the mhpmeventN CSRs, it would raise an illegal
> > instruction exception and trap into m-mode to emulate event selector
> > CSRs access. It doesn't make sense because we shouldn't write the
> > m-mode CSRs in s-mode, it would be better that set events through SBI
> > call or the shadow CSRs of s-mode. We would change it later.
> >
> > Signed-off-by: Zong Li <zong.li@sifive.com>
> > ---
> >  arch/riscv/include/asm/perf_event.h |  65 ++++++---
> >  arch/riscv/kernel/perf_event.c      | 204 +++++++++++++++++++++++-----
> >  2 files changed, 215 insertions(+), 54 deletions(-)
> >
> > diff --git a/arch/riscv/include/asm/perf_event.h b/arch/riscv/include/asm/perf_event.h
> > index 062efd3a1d5d..41d515a1f331 100644
> > --- a/arch/riscv/include/asm/perf_event.h
> > +++ b/arch/riscv/include/asm/perf_event.h
> > @@ -14,39 +14,64 @@
> >
> >  #ifdef CONFIG_RISCV_BASE_PMU
> >  #define RISCV_BASE_COUNTERS    2
> > +#define RISCV_EVENT_COUNTERS   29
>
> Same comment as DT documentation related to naming.

Change it as well. Thanks.

>
> Regards,
> Anup
>
>
> > +#define RISCV_TOTAL_COUNTERS   (RISCV_BASE_COUNTERS + RISCV_EVENT_COUNTERS)
> >
> >  /*
> > - * The RISCV_MAX_COUNTERS parameter should be specified.
> > - */
> > -
> > -#define RISCV_MAX_COUNTERS     2
> > -
> > -/*
> > - * These are the indexes of bits in counteren register *minus* 1,
> > - * except for cycle.  It would be coherent if it can directly mapped
> > - * to counteren bit definition, but there is a *time* register at
> > - * counteren[1].  Per-cpu structure is scarce resource here.
> > - *
> >   * According to the spec, an implementation can support counter up to
> >   * mhpmcounter31, but many high-end processors has at most 6 general
> >   * PMCs, we give the definition to MHPMCOUNTER8 here.
> >   */
> > -#define RISCV_PMU_CYCLE                0
> > -#define RISCV_PMU_INSTRET      1
> > -#define RISCV_PMU_MHPMCOUNTER3 2
> > -#define RISCV_PMU_MHPMCOUNTER4 3
> > -#define RISCV_PMU_MHPMCOUNTER5 4
> > -#define RISCV_PMU_MHPMCOUNTER6 5
> > -#define RISCV_PMU_MHPMCOUNTER7 6
> > -#define RISCV_PMU_MHPMCOUNTER8 7
> > +#define RISCV_PMU_CYCLE                        0
> > +#define RISCV_PMU_INSTRET              2
> > +#define RISCV_PMU_HPMCOUNTER3          3
> > +#define RISCV_PMU_HPMCOUNTER4          4
> > +#define RISCV_PMU_HPMCOUNTER5          5
> > +#define RISCV_PMU_HPMCOUNTER6          6
> > +#define RISCV_PMU_HPMCOUNTER7          7
> > +#define RISCV_PMU_HPMCOUNTER8          8
> > +
> > +#define RISCV_PMU_HPMCOUNTER_FIRST     3
> > +#define RISCV_PMU_HPMCOUNTER_LAST                                      \
> > +       (RISCV_PMU_HPMCOUNTER_FIRST + riscv_pmu->num_counters - 1)
> >
> >  #define RISCV_OP_UNSUPP                (-EOPNOTSUPP)
> >
> > +/* Hardware cache event encoding */
> > +#define PERF_HW_CACHE_TYPE             0
> > +#define PERF_HW_CACHE_OP               8
> > +#define PERF_HW_CACHE_RESULT           16
> > +#define PERF_HW_CACHE_MASK             0xff
> > +
> > +/* config_base encoding */
> > +#define RISCV_PMU_TYPE_MASK            0x3
> > +#define RISCV_PMU_TYPE_BASE            0x1
> > +#define RISCV_PMU_TYPE_EVENT           0x2
> > +#define RISCV_PMU_EXCLUDE_MASK         0xc
> > +#define RISCV_PMU_EXCLUDE_USER         0x3
> > +#define RISCV_PMU_EXCLUDE_KERNEL       0x4
> > +
> > +/*
> > + * Currently, machine-mode supports emulation of mhpmeventN. Setting mhpmeventN
> > + * to raise an illegal instruction exception to set event types in machine-mode.
> > + * Eventually, we should set event types through standard SBI call or the shadow
> > + * CSRs of supervisor-mode, because it is weird for writing CSR of machine-mode
> > + * explicitly in supervisor-mode. These macro should be removed in the future.
> > + */
> > +#define CSR_MHPMEVENT3 0x323
> > +#define CSR_MHPMEVENT4 0x324
> > +#define CSR_MHPMEVENT5 0x325
> > +#define CSR_MHPMEVENT6 0x326
> > +#define CSR_MHPMEVENT7 0x327
> > +#define CSR_MHPMEVENT8 0x328
> > +
> >  struct cpu_hw_events {
> >         /* # currently enabled events*/
> >         int                     n_events;
> >         /* currently enabled events */
> > -       struct perf_event       *events[RISCV_MAX_COUNTERS];
> > +       struct perf_event       *events[RISCV_EVENT_COUNTERS];
> > +       /* bitmap of used event counters */
> > +       unsigned long           used_cntr_mask;
> >         /* vendor-defined PMU data */
> >         void                    *platform;
> >  };
> > diff --git a/arch/riscv/kernel/perf_event.c b/arch/riscv/kernel/perf_event.c
> > index c835f0362d94..0cfcd6f1e57b 100644
> > --- a/arch/riscv/kernel/perf_event.c
> > +++ b/arch/riscv/kernel/perf_event.c
> > @@ -139,6 +139,53 @@ static const int riscv_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
> >         },
> >  };
> >
> > +/*
> > + * Methods for checking and getting PMU information
> > + */
> > +
> > +static inline int is_base_counter(int idx)
> > +{
> > +       return (idx == RISCV_PMU_CYCLE || idx == RISCV_PMU_INSTRET);
> > +}
> > +
> > +static inline int is_event_counter(int idx)
> > +{
> > +       return (idx >= RISCV_PMU_HPMCOUNTER_FIRST &&
> > +               idx <= RISCV_PMU_HPMCOUNTER_LAST);
> > +}
> > +
> > +static inline int get_available_counter(struct perf_event *event)
> > +{
> > +       struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
> > +       struct hw_perf_event *hwc = &event->hw;
> > +       unsigned long config_base = hwc->config_base & RISCV_PMU_TYPE_MASK;
> > +       unsigned long mask;
> > +       int ret;
> > +
> > +       switch (config_base) {
> > +       case RISCV_PMU_TYPE_BASE:
> > +               ret = hwc->config;
> > +               if (WARN_ON_ONCE(!is_base_counter(ret)))
> > +                       return -ENOSPC;
> > +               break;
> > +       case RISCV_PMU_TYPE_EVENT:
> > +               mask = ~cpuc->used_cntr_mask;
> > +               ret = find_next_bit(&mask, RISCV_PMU_HPMCOUNTER_LAST, 3);
> > +               if (WARN_ON_ONCE(!is_event_counter(ret)))
> > +                       return -ENOSPC;
> > +               break;
> > +       default:
> > +               return -ENOENT;
> > +       }
> > +
> > +       __set_bit(ret, &cpuc->used_cntr_mask);
> > +
> > +       return ret;
> > +}
> > +
> > +/*
> > + * Map generic hardware event
> > + */
> >  static int riscv_map_hw_event(u64 config)
> >  {
> >         if (config >= riscv_pmu->max_events)
> > @@ -147,32 +194,28 @@ static int riscv_map_hw_event(u64 config)
> >         return riscv_pmu->hw_events[config];
> >  }
> >
> > -static int riscv_map_cache_decode(u64 config, unsigned int *type,
> > -                          unsigned int *op, unsigned int *result)
> > -{
> > -       return -ENOENT;
> > -}
> > -
> > +/*
> > + * Map generic hardware cache event
> > + */
> >  static int riscv_map_cache_event(u64 config)
> >  {
> >         unsigned int type, op, result;
> > -       int err = -ENOENT;
> > -               int code;
> > +       int ret;
> >
> > -       err = riscv_map_cache_decode(config, &type, &op, &result);
> > -       if (!riscv_pmu->cache_events || err)
> > -               return err;
> > +       type    = (config >> PERF_HW_CACHE_TYPE) & PERF_HW_CACHE_MASK;
> > +       op      = (config >> PERF_HW_CACHE_OP) & PERF_HW_CACHE_MASK;
> > +       result  = (config >> PERF_HW_CACHE_RESULT) & PERF_HW_CACHE_MASK;
> >
> >         if (type >= PERF_COUNT_HW_CACHE_MAX ||
> >             op >= PERF_COUNT_HW_CACHE_OP_MAX ||
> >             result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
> >                 return -EINVAL;
> >
> > -       code = (*riscv_pmu->cache_events)[type][op][result];
> > -       if (code == RISCV_OP_UNSUPP)
> > +       ret = riscv_cache_event_map[type][op][result];
> > +       if (ret == RISCV_OP_UNSUPP)
> >                 return -EINVAL;
> >
> > -       return code;
> > +       return ret == RISCV_OP_UNSUPP ? -ENOENT : ret;
> >  }
> >
> >  /*
> > @@ -190,8 +233,27 @@ static inline u64 read_counter(int idx)
> >         case RISCV_PMU_INSTRET:
> >                 val = csr_read(CSR_INSTRET);
> >                 break;
> > +       case RISCV_PMU_HPMCOUNTER3:
> > +               val = csr_read(CSR_HPMCOUNTER3);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER4:
> > +               val = csr_read(CSR_HPMCOUNTER4);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER5:
> > +               val = csr_read(CSR_HPMCOUNTER5);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER6:
> > +               val = csr_read(CSR_HPMCOUNTER6);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER7:
> > +               val = csr_read(CSR_HPMCOUNTER7);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER8:
> > +               val = csr_read(CSR_HPMCOUNTER8);
>
> This is broken for RV32 because for RV32 we have to read two
> CSRs to get a counter value.

Oh yes, thanks for your reminder. Add them in the next version.

>
> Also, for correctly reading a 64bit counter on RV32 we have
> to read just like get_cycles64() does for RV32.
>
> static inline u64 get_cycles64(void)
> {
>     u32 hi, lo;
>
>     do {
>         hi = get_cycles_hi();
>         lo = get_cycles();
>     } while (hi != get_cycles_hi());
>
>     return ((u64)hi << 32) | lo;
> }
>
> Regards,
> Anup
>
>
> > +               break;
> >         default:
> > -               WARN_ON_ONCE(idx < 0 || idx > RISCV_MAX_COUNTERS);
> > +               WARN_ON_ONCE(idx < RISCV_PMU_CYCLE ||
> > +                            idx > RISCV_TOTAL_COUNTERS);
> >                 return -EINVAL;
> >         }
> >
> > @@ -204,6 +266,68 @@ static inline void write_counter(int idx, u64 value)
> >         WARN_ON_ONCE(1);
> >  }
> >
> > +static inline void write_event(int idx, u64 value)
> > +{
> > +       /* TODO: We shouldn't write CSR of m-mode explicitly here. Ideally,
> > +        * it need to set the event selector by SBI call or the s-mode
> > +        * shadow CSRs of them. Exploit illegal instruction exception to
> > +        * emulate mhpmcounterN access in m-mode.
> > +        */
> > +       switch (idx) {
> > +       case RISCV_PMU_HPMCOUNTER3:
> > +               csr_write(CSR_MHPMEVENT3, value);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER4:
> > +               csr_write(CSR_MHPMEVENT4, value);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER5:
> > +               csr_write(CSR_MHPMEVENT5, value);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER6:
> > +               csr_write(CSR_MHPMEVENT6, value);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER7:
> > +               csr_write(CSR_MHPMEVENT7, value);
> > +               break;
> > +       case RISCV_PMU_HPMCOUNTER8:
> > +               csr_write(CSR_MHPMEVENT8, value);
> > +               break;
> > +       default:
> > +               WARN_ON_ONCE(idx < RISCV_PMU_HPMCOUNTER3 ||
> > +                            idx > RISCV_TOTAL_COUNTERS);
> > +               return;
> > +       }
> > +}

I was also wondering if you have any suggestions about the PMU SBI
extension as I mentioned in the cover letter. Currently, we set the
event selectors by emulation of OpenSBI, so just write the m-mode CSRs
as above.

> > +
> > +/*
> > + * Enable and disable event counters
> > + */
> > +
> > +static inline void riscv_pmu_enable_event(struct perf_event *event)
> > +{
> > +       struct hw_perf_event *hwc = &event->hw;
> > +       int idx = hwc->idx;
> > +
> > +       if (is_event_counter(idx))
> > +               write_event(idx, hwc->config);
> > +
> > +       /*
> > +        * Since we cannot write to counters, this serves as an initialization
> > +        * to the delta-mechanism in pmu->read(); otherwise, the delta would be
> > +        * wrong when pmu->read is called for the first time.
> > +        */
> > +       local64_set(&hwc->prev_count, read_counter(hwc->idx));
> > +}
> > +
> > +static inline void riscv_pmu_disable_event(struct perf_event *event)
> > +{
> > +       struct hw_perf_event *hwc = &event->hw;
> > +       int idx = hwc->idx;
> > +
> > +       if (is_event_counter(idx))
> > +               write_event(idx, 0);
> > +}
> > +
> >  /*
> >   * pmu->read: read and update the counter
> >   *
> > @@ -232,6 +356,7 @@ static void riscv_pmu_read(struct perf_event *event)
> >          */
> >         delta = (new_raw_count - prev_raw_count) &
> >                 ((1ULL << riscv_pmu->counter_width) - 1);
> > +
> >         local64_add(delta, &event->count);
> >         /*
> >          * Something like local64_sub(delta, &hwc->period_left) here is
> > @@ -252,6 +377,11 @@ static void riscv_pmu_stop(struct perf_event *event, int flags)
> >  {
> >         struct hw_perf_event *hwc = &event->hw;
> >
> > +       if (WARN_ON_ONCE(hwc->idx == -1))
> > +               return;
> > +
> > +       riscv_pmu_disable_event(event);
> > +
> >         WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
> >         hwc->state |= PERF_HES_STOPPED;
> >
> > @@ -271,6 +401,9 @@ static void riscv_pmu_start(struct perf_event *event, int flags)
> >         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
> >                 return;
> >
> > +       if (WARN_ON_ONCE(hwc->idx == -1))
> > +               return;
> > +
> >         if (flags & PERF_EF_RELOAD) {
> >                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
> >
> > @@ -281,14 +414,10 @@ static void riscv_pmu_start(struct perf_event *event, int flags)
> >         }
> >
> >         hwc->state = 0;
> > -       perf_event_update_userpage(event);
> >
> > -       /*
> > -        * Since we cannot write to counters, this serves as an initialization
> > -        * to the delta-mechanism in pmu->read(); otherwise, the delta would be
> > -        * wrong when pmu->read is called for the first time.
> > -        */
> > -       local64_set(&hwc->prev_count, read_counter(hwc->idx));
> > +       riscv_pmu_enable_event(event);
> > +
> > +       perf_event_update_userpage(event);
> >  }
> >
> >  /*
> > @@ -298,21 +427,18 @@ static int riscv_pmu_add(struct perf_event *event, int flags)
> >  {
> >         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
> >         struct hw_perf_event *hwc = &event->hw;
> > +       int count_idx;
> >
> >         if (cpuc->n_events == riscv_pmu->num_counters)
> >                 return -ENOSPC;
> >
> > -       /*
> > -        * We don't have general conunters, so no binding-event-to-counter
> > -        * process here.
> > -        *
> > -        * Indexing using hwc->config generally not works, since config may
> > -        * contain extra information, but here the only info we have in
> > -        * hwc->config is the event index.
> > -        */
> > -       hwc->idx = hwc->config;
> > -       cpuc->events[hwc->idx] = event;
> > +       count_idx = get_available_counter(event);
> > +       if (count_idx < 0)
> > +               return -ENOSPC;
> > +
> >         cpuc->n_events++;
> > +       hwc->idx = count_idx;
> > +       cpuc->events[hwc->idx] = event;
> >
> >         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
> >
> > @@ -330,8 +456,10 @@ static void riscv_pmu_del(struct perf_event *event, int flags)
> >         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
> >         struct hw_perf_event *hwc = &event->hw;
> >
> > -       cpuc->events[hwc->idx] = NULL;
> >         cpuc->n_events--;
> > +       __clear_bit(hwc->idx, &cpuc->used_cntr_mask);
> > +
> > +       cpuc->events[hwc->idx] = NULL;
> >         riscv_pmu->pmu->stop(event, PERF_EF_UPDATE);
> >         perf_event_update_userpage(event);
> >  }
> > @@ -385,6 +513,7 @@ static int riscv_event_init(struct perf_event *event)
> >  {
> >         struct perf_event_attr *attr = &event->attr;
> >         struct hw_perf_event *hwc = &event->hw;
> > +       unsigned long config_base = 0;
> >         int err;
> >         int code;
> >
> > @@ -406,11 +535,17 @@ static int riscv_event_init(struct perf_event *event)
> >                 code = riscv_pmu->map_cache_event(attr->config);
> >                 break;
> >         case PERF_TYPE_RAW:
> > -               return -EOPNOTSUPP;
> > +               code = attr->config;
> > +               break;
> >         default:
> >                 return -ENOENT;
> >         }
> >
> > +       if (is_base_counter(code))
> > +               config_base |= RISCV_PMU_TYPE_BASE;
> > +       else
> > +               config_base |= RISCV_PMU_TYPE_EVENT;
> > +
> >         event->destroy = riscv_event_destroy;
> >         if (code < 0) {
> >                 event->destroy(event);
> > @@ -424,6 +559,7 @@ static int riscv_event_init(struct perf_event *event)
> >          * But since we don't have such support, later in pmu->add(), we just
> >          * use hwc->config as the index instead.
> >          */
> > +       hwc->config_base = config_base;
> >         hwc->config = code;
> >         hwc->idx = -1;
> >
> > --
> > 2.27.0
> >