perf tools: Add Intel PT support for PSB periods
authorAdrian Hunter <adrian.hunter@intel.com>
Fri, 17 Jul 2015 16:33:52 +0000 (19:33 +0300)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Mon, 24 Aug 2015 20:45:08 +0000 (17:45 -0300)
The PSB packet is a synchronization packet that provides a starting
point for decoding or recovery from errors.

This patch adds support for a new Intel PT feature that allows the
frequency of PSB packets to be specified.

Support for this feature is indicated by
/sys/bus/event_source/devices/intel_pt/caps/psb_cyc which contains "1"
if the feature is supported and "0" otherwise.

The PSB period can be specified as a PMU config term e.g. perf record -e
intel_pt/psb_period=2/u sleep 1

The default value is 3 or the nearest lower value that is supported.  0
is always supported.

Valid values are given by:

/sys/bus/event_source/devices/intel_pt/caps/psb_periods

which contains a hexadecimal value, the bits of which represent valid
values e.g. bit 2 set means value 2 is valid.

The value is converted to the approximate number of trace bytes between
PSB packets as:

2 ^ (value + 11)

e.g. value 3 means 16KiB bytes between PSBs

If an invalid value is entered, the error message will give a list of
valid values e.g.

$ perf record -e intel_pt/psb_period=15/u uname
Invalid psb_period for intel_pt. Valid values are: 0-5

tools/perf/Documentation/intel-pt.txt is updated in a later patch as
there are a number of new features being added.

For more information about PSB periods refer to the Intel 64 and IA-32
Architectures SDM Chapter 36 Intel Processor Trace from June 2015 or
later.

Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/1437150840-31811-18-git-send-email-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/arch/x86/util/intel-pt.c

index da7d2c15e611b8b492f5d3d350d4abbcdafe6f56..145975b003a74552d62293d4b3bd3330ae225eed 100644 (file)
@@ -99,17 +99,121 @@ static int intel_pt_parse_terms(struct list_head *formats, const char *str,
        return intel_pt_parse_terms_with_default(formats, str, config);
 }
 
-static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu __maybe_unused,
-                                 struct perf_evlist *evlist __maybe_unused)
+static u64 intel_pt_masked_bits(u64 mask, u64 bits)
 {
-       return 256;
+       const u64 top_bit = 1ULL << 63;
+       u64 res = 0;
+       int i;
+
+       for (i = 0; i < 64; i++) {
+               if (mask & top_bit) {
+                       res <<= 1;
+                       if (bits & top_bit)
+                               res |= 1;
+               }
+               mask <<= 1;
+               bits <<= 1;
+       }
+
+       return res;
+}
+
+static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
+                               struct perf_evlist *evlist, u64 *res)
+{
+       struct perf_evsel *evsel;
+       u64 mask;
+
+       *res = 0;
+
+       mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
+       if (!mask)
+               return -EINVAL;
+
+       evlist__for_each(evlist, evsel) {
+               if (evsel->attr.type == intel_pt_pmu->type) {
+                       *res = intel_pt_masked_bits(mask, evsel->attr.config);
+                       return 0;
+               }
+       }
+
+       return -EINVAL;
+}
+
+static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
+                                 struct perf_evlist *evlist)
+{
+       u64 val;
+       int err, topa_multiple_entries;
+       size_t psb_period;
+
+       if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
+                               "%d", &topa_multiple_entries) != 1)
+               topa_multiple_entries = 0;
+
+       /*
+        * Use caps/topa_multiple_entries to indicate early hardware that had
+        * extra frequent PSBs.
+        */
+       if (!topa_multiple_entries) {
+               psb_period = 256;
+               goto out;
+       }
+
+       err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
+       if (err)
+               val = 0;
+
+       psb_period = 1 << (val + 11);
+out:
+       pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
+       return psb_period;
+}
+
+static int intel_pt_pick_bit(int bits, int target)
+{
+       int pos, pick = -1;
+
+       for (pos = 0; bits; bits >>= 1, pos++) {
+               if (bits & 1) {
+                       if (pos <= target || pick < 0)
+                               pick = pos;
+                       if (pos >= target)
+                               break;
+               }
+       }
+
+       return pick;
 }
 
 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
 {
+       char buf[256];
+       int psb_cyc, psb_periods, psb_period;
+       int pos = 0;
        u64 config;
 
-       intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &config);
+       pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
+
+       if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
+                               &psb_cyc) != 1)
+               psb_cyc = 1;
+
+       if (psb_cyc) {
+               if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
+                                       &psb_periods) != 1)
+                       psb_periods = 0;
+               if (psb_periods) {
+                       psb_period = intel_pt_pick_bit(psb_periods, 3);
+                       pos += scnprintf(buf + pos, sizeof(buf) - pos,
+                                        ",psb_period=%d", psb_period);
+               }
+       }
+
+       pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
+
+       intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
+
        return config;
 }
 
@@ -239,6 +343,103 @@ static int intel_pt_track_switches(struct perf_evlist *evlist)
        return 0;
 }
 
+static void intel_pt_valid_str(char *str, size_t len, u64 valid)
+{
+       unsigned int val, last = 0, state = 1;
+       int p = 0;
+
+       str[0] = '\0';
+
+       for (val = 0; val <= 64; val++, valid >>= 1) {
+               if (valid & 1) {
+                       last = val;
+                       switch (state) {
+                       case 0:
+                               p += scnprintf(str + p, len - p, ",");
+                               /* Fall through */
+                       case 1:
+                               p += scnprintf(str + p, len - p, "%u", val);
+                               state = 2;
+                               break;
+                       case 2:
+                               state = 3;
+                               break;
+                       case 3:
+                               state = 4;
+                               break;
+                       default:
+                               break;
+                       }
+               } else {
+                       switch (state) {
+                       case 3:
+                               p += scnprintf(str + p, len - p, ",%u", last);
+                               state = 0;
+                               break;
+                       case 4:
+                               p += scnprintf(str + p, len - p, "-%u", last);
+                               state = 0;
+                               break;
+                       default:
+                               break;
+                       }
+                       if (state != 1)
+                               state = 0;
+               }
+       }
+}
+
+static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
+                                   const char *caps, const char *name,
+                                   const char *supported, u64 config)
+{
+       char valid_str[256];
+       unsigned int shift;
+       unsigned long long valid;
+       u64 bits;
+       int ok;
+
+       if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
+               valid = 0;
+
+       if (supported &&
+           perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
+               valid = 0;
+
+       valid |= 1;
+
+       bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
+
+       config &= bits;
+
+       for (shift = 0; bits && !(bits & 1); shift++)
+               bits >>= 1;
+
+       config >>= shift;
+
+       if (config > 63)
+               goto out_err;
+
+       if (valid & (1 << config))
+               return 0;
+out_err:
+       intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
+       pr_err("Invalid %s for %s. Valid values are: %s\n",
+              name, INTEL_PT_PMU_NAME, valid_str);
+       return -EINVAL;
+}
+
+static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
+                                   struct perf_evsel *evsel)
+{
+       if (!evsel)
+               return 0;
+
+       return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
+                                       "psb_period", "caps/psb_cyc",
+                                       evsel->attr.config);
+}
+
 static int intel_pt_recording_options(struct auxtrace_record *itr,
                                      struct perf_evlist *evlist,
                                      struct record_opts *opts)
@@ -251,6 +452,7 @@ static int intel_pt_recording_options(struct auxtrace_record *itr,
        const struct cpu_map *cpus = evlist->cpus;
        bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
        u64 tsc_bit;
+       int err;
 
        ptr->evlist = evlist;
        ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
@@ -281,6 +483,10 @@ static int intel_pt_recording_options(struct auxtrace_record *itr,
        if (!opts->full_auxtrace)
                return 0;
 
+       err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
+       if (err)
+               return err;
+
        /* Set default sizes for snapshot mode */
        if (opts->auxtrace_snapshot_mode) {
                size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
@@ -366,8 +572,6 @@ static int intel_pt_recording_options(struct auxtrace_record *itr,
         * threads.
         */
        if (have_timing_info && !cpu_map__empty(cpus)) {
-               int err;
-
                err = intel_pt_track_switches(evlist);
                if (err == -EPERM)
                        pr_debug2("Unable to select sched:sched_switch\n");
@@ -394,7 +598,6 @@ static int intel_pt_recording_options(struct auxtrace_record *itr,
        /* Add dummy event to keep tracking */
        if (opts->full_auxtrace) {
                struct perf_evsel *tracking_evsel;
-               int err;
 
                err = parse_events(evlist, "dummy:u", NULL);
                if (err)