Re: [PATCH v1 2/7] perf cs_etm: Use pid tracing explicitly instead of contextid

From: Mathieu Poirier
Date: Fri Jan 15 2021 - 17:45:17 EST

Next message: Douglas Anderson: "[PATCH v2 0/5] drm/panel-simple: Patches for N116BCA-EA1"
Previous message: Palmer Dabbelt: "Re: [PATCH 3/4] RISC-V: Fix L1_CACHE_BYTES for RV32"
In reply to: Leo Yan: "[PATCH v1 2/7] perf cs_etm: Use pid tracing explicitly instead of contextid"
Next in thread: Leo Yan: "[PATCH v1 3/7] perf cs-etm: Calculate per CPU metadata array size"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Sat, Jan 09, 2021 at 03:44:30PM +0800, Leo Yan wrote:
> From: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
>
> If the kernel is running at EL2, the pid of the task is exposed
> via VMID instead of the CONTEXTID. Add support for this in the
> perf tool.
>
> By default the perf tool requests contextid and timestamp for
> task bound events. Instead of hard coding contextid, switch
> to "pid" config exposed by the kernel. While at it, define new
> independent macros (rather than using the "config" bits) for
> requesting the "pid" and "timestamp" for cs_etm_set_option(),
> since the PID config is now dynamic depending on the kernel
> exception level.
>
> Cc: Mike Leach <mike.leach@xxxxxxxxxx>
> Cc: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> Cc: Al Grant <al.grant@xxxxxxx>
> Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
> ---
> tools/include/linux/coresight-pmu.h | 11 +++--
> tools/perf/arch/arm/util/cs-etm.c | 68 ++++++++++++++++++++++-------
> 2 files changed, 59 insertions(+), 20 deletions(-)
>
> diff --git a/tools/include/linux/coresight-pmu.h b/tools/include/linux/coresight-pmu.h
> index b0e35eec6499..927c6285ce5d 100644
> --- a/tools/include/linux/coresight-pmu.h
> +++ b/tools/include/linux/coresight-pmu.h
> @@ -11,16 +11,19 @@
> #define CORESIGHT_ETM_PMU_SEED 0x10
>
> /* ETMv3.5/PTM's ETMCR config bit */
> -#define ETM_OPT_CYCACC 12
> -#define ETM_OPT_CTXTID 14
> -#define ETM_OPT_TS 28
> -#define ETM_OPT_RETSTK 29
> +#define ETM_OPT_CYCACC 12
> +#define ETM_OPT_CTXTID 14
> +#define ETM_OPT_CTXTID_IN_VMID 15
> +#define ETM_OPT_TS 28
> +#define ETM_OPT_RETSTK 29
>
> /* ETMv4 CONFIGR programming bits for the ETM OPTs */
> #define ETM4_CFG_BIT_CYCACC 4
> #define ETM4_CFG_BIT_CTXTID 6
> +#define ETM4_CFG_BIT_VMID 7
> #define ETM4_CFG_BIT_TS 11
> #define ETM4_CFG_BIT_RETSTK 12
> +#define ETM4_CFG_BIT_VMID_OPT 15

Wasn't this done in the previous patch?

>
> static inline int coresight_get_trace_id(int cpu)
> {
> diff --git a/tools/perf/arch/arm/util/cs-etm.c b/tools/perf/arch/arm/util/cs-etm.c
> index cad7bf783413..fad7b6e13ccc 100644
> --- a/tools/perf/arch/arm/util/cs-etm.c
> +++ b/tools/perf/arch/arm/util/cs-etm.c
> @@ -59,14 +59,15 @@ static const char *metadata_etmv4_ro[CS_ETMV4_PRIV_MAX] = {
>
> static bool cs_etm_is_etmv4(struct auxtrace_record *itr, int cpu);
>
> -static int cs_etm_set_context_id(struct auxtrace_record *itr,
> - struct evsel *evsel, int cpu)
> +static int cs_etm_set_pid(struct auxtrace_record *itr,
> + struct evsel *evsel, int cpu)
> {
> struct cs_etm_recording *ptr;
> struct perf_pmu *cs_etm_pmu;
> char path[PATH_MAX];
> int err = -EINVAL;
> u32 val;
> + u64 pid_fmt;
>
> ptr = container_of(itr, struct cs_etm_recording, itr);
> cs_etm_pmu = ptr->cs_etm_pmu;
> @@ -86,21 +87,50 @@ static int cs_etm_set_context_id(struct auxtrace_record *itr,
> goto out;
> }
>
> + pid_fmt = perf_pmu__format_bits(&cs_etm_pmu->format, "pid");

If we go with my suggestion from patch 1 this would become "contextid".

I think what follows below is all good.

> /*
> - * TRCIDR2.CIDSIZE, bit [9-5], indicates whether contextID tracing
> - * is supported:
> - * 0b00000 Context ID tracing is not supported.
> - * 0b00100 Maximum of 32-bit Context ID size.
> - * All other values are reserved.
> + * If the kernel doesn't support the "pid" format (older kernel),
> + * fall back to using the CTXTID.
> */
> - val = BMVAL(val, 5, 9);
> - if (!val || val != 0x4) {
> + if (!pid_fmt)
> + pid_fmt = 1ULL << ETM_OPT_CTXTID;
> +
> + switch (pid_fmt) {
> + case (1ULL << ETM_OPT_CTXTID):
> + /*
> + * TRCIDR2.CIDSIZE, bit [9-5], indicates whether contextID
> + * tracing is supported:
> + * 0b00000 Context ID tracing is not supported.
> + * 0b00100 Maximum of 32-bit Context ID size.
> + * All other values are reserved.
> + */
> + val = BMVAL(val, 5, 9);
> + if (!val || val != 0x4) {
> + err = -EINVAL;
> + goto out;
> + }
> + break;
> + case (1ULL << ETM_OPT_CTXTID_IN_VMID):
> + /*
> + * TRCIDR2.VMIDOPT[30:29] != 0 and
> + * TRCIDR2.VMIDSIZE[14:10] == 0b00100 (32bit virtual contextid)
> + * We can't support CONTEXTIDR in VMID if the size of the
> + * virtual context id is < 32bit.
> + * Any value of VMIDSIZE >= 4 (i.e, > 32bit) is fine for us.
> + */
> + if (!BMVAL(val, 29, 30) || BMVAL(val, 10, 14) < 4) {
> + err = -EINVAL;
> + goto out;
> + }
> + break;
> + default:
> err = -EINVAL;
> goto out;
> }
>
> +
> /* All good, let the kernel know */
> - evsel->core.attr.config |= (1 << ETM_OPT_CTXTID);
> + evsel->core.attr.config |= pid_fmt;
> err = 0;
>
> out:
> @@ -156,6 +186,10 @@ static int cs_etm_set_timestamp(struct auxtrace_record *itr,
> return err;
> }
>
> +#define ETM_SET_OPT_PID (1 << 0)
> +#define ETM_SET_OPT_TS (1 << 1)
> +#define ETM_SET_OPT_MASK (ETM_SET_OPT_PID | ETM_SET_OPT_TS)
> +

I don't think we need this as part of the current patchset. Back in the days I
thought it would be a good idea to use the ETMv3.5/PTM configuration bits but in
hindsight it wasn't. That's what happens when your crystal ball isn't working.

That being said we can start treating the perf_event_attr::config bits
arbitrarily. Someone pointed out that it might get confusing but I don't think
it is anything that can't be overcomed with good in-line documentation.

> static int cs_etm_set_option(struct auxtrace_record *itr,
> struct evsel *evsel, u32 option)
> {
> @@ -169,17 +203,17 @@ static int cs_etm_set_option(struct auxtrace_record *itr,
> !cpu_map__has(online_cpus, i))
> continue;
>
> - if (option & ETM_OPT_CTXTID) {
> - err = cs_etm_set_context_id(itr, evsel, i);
> + if (option & ETM_SET_OPT_PID) {
> + err = cs_etm_set_pid(itr, evsel, i);
> if (err)
> goto out;
> }
> - if (option & ETM_OPT_TS) {
> + if (option & ETM_SET_OPT_TS) {
> err = cs_etm_set_timestamp(itr, evsel, i);
> if (err)
> goto out;
> }
> - if (option & ~(ETM_OPT_CTXTID | ETM_OPT_TS))
> + if (option & ~(ETM_SET_OPT_MASK))
> /* Nothing else is currently supported */
> goto out;
> }
> @@ -406,7 +440,7 @@ static int cs_etm_recording_options(struct auxtrace_record *itr,
> evsel__set_sample_bit(cs_etm_evsel, CPU);
>
> err = cs_etm_set_option(itr, cs_etm_evsel,
> - ETM_OPT_CTXTID | ETM_OPT_TS);
> + ETM_SET_OPT_PID | ETM_SET_OPT_TS);
> if (err)
> goto out;
> }
> @@ -485,7 +519,9 @@ static u64 cs_etmv4_get_config(struct auxtrace_record *itr)
> config |= BIT(ETM4_CFG_BIT_TS);
> if (config_opts & BIT(ETM_OPT_RETSTK))
> config |= BIT(ETM4_CFG_BIT_RETSTK);
> -
> + if (config_opts & BIT(ETM_OPT_CTXTID_IN_VMID))
> + config |= BIT(ETM4_CFG_BIT_VMID) |
> + BIT(ETM4_CFG_BIT_VMID_OPT);
> return config;
> }
>
> --
> 2.25.1
>

Next message: Douglas Anderson: "[PATCH v2 0/5] drm/panel-simple: Patches for N116BCA-EA1"
Previous message: Palmer Dabbelt: "Re: [PATCH 3/4] RISC-V: Fix L1_CACHE_BYTES for RV32"
In reply to: Leo Yan: "[PATCH v1 2/7] perf cs_etm: Use pid tracing explicitly instead of contextid"
Next in thread: Leo Yan: "[PATCH v1 3/7] perf cs-etm: Calculate per CPU metadata array size"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]