On Tue, Jun 07, 2011 at 05:56:04PM -0600, David Ahern wrote: > Use reftime event for initial correlation of perf_clock to > time-of-day. Add timekeeping trace events to event list to > capture jumps in time-of-day. > > Signed-off-by: David Ahern <[email protected]> > --- > tools/perf/Documentation/perf-record.txt | 3 + > tools/perf/builtin-record.c | 102 > +++++++++++++++++++++++++++++- > 2 files changed, 104 insertions(+), 1 deletions(-) > > diff --git a/tools/perf/Documentation/perf-record.txt > b/tools/perf/Documentation/perf-record.txt > index 5a520f8..e4c87ba 100644 > --- a/tools/perf/Documentation/perf-record.txt > +++ b/tools/perf/Documentation/perf-record.txt > @@ -148,6 +148,9 @@ an empty cgroup (monitor all the time) using, e.g., -G > foo,,bar. Cgroups must ha > corresponding events, i.e., they always refer to events defined earlier on > the command > line. > > +--tod:: > +Collect data for displaying time-of-day strings when printing events. > + > SEE ALSO > -------- > linkperf:perf-stat[1], linkperf:perf-list[1] > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c > index 8e2c857..4f8d5f2 100644 > --- a/tools/perf/builtin-record.c > +++ b/tools/perf/builtin-record.c > @@ -32,6 +32,9 @@ > > #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) > > +#define TRACE_TOD_SUBSYSTEM "timekeeping" > +#define TRACE_TOD_SUBSYSTEM_LEN 11 > + > enum write_mode_t { > WRITE_FORCE, > WRITE_APPEND > @@ -65,6 +68,8 @@ static bool sample_address > = false; > static bool sample_time = false; > static bool no_buildid = false; > static bool no_buildid_cache = false; > +static bool want_tod = false; > +static u64 tod_sample_type; > static struct perf_evlist *evsel_list; > > static long samples = 0; > @@ -215,6 +220,9 @@ static void config_attr(struct perf_evsel *evsel, struct > perf_evlist *evlist) > attr->sample_type |= PERF_SAMPLE_CPU; > } > > + if (want_tod) > + attr->sample_type |= tod_sample_type; > + > if (nodelay) { > attr->watermark = 0; > attr->wakeup_events = 1; > @@ -248,6 +256,86 @@ static bool perf_evlist__equal(struct perf_evlist > *evlist, > return true; > } > > +static int perf_event__synthesize_reftime(perf_event__handler_t process) > +{ > + union perf_event ev; > + struct timespec tp; > + > + memset(&ev, 0, sizeof(ev)); > + > + if (gettimeofday(&ev.reftime.tv, NULL) != 0) { > + error("gettimeofday failed.\n"); > + return -1; > + } > + if (clock_gettime(CLOCK_MONOTONIC, &tp) != 0) { > + error("clock_gettime(CLOCK_MONOTONIC) failed.\n"); > + return -1; > + } > + ev.reftime.nsec = (u64) tp.tv_sec * NSEC_PER_SEC + (u64) tp.tv_nsec; > + > + ev.header.type = PERF_RECORD_REFTIME; > + ev.header.size = sizeof(ev.reftime); > + > + return process(&ev, NULL, session); > +} > + > +static int add_timeofday_events(void) > +{ > + int rc, i, len; > + struct perf_event_attr attr; > + struct perf_evsel *evsel; > + > + /* events that modify xtime */ > + const char *tod_events[] = {"settimeofday", > + "timekeeping_inject_offset", > + "timekeeping_inject_sleeptime", > + NULL}; > + > + i = 0; > + rc = -1; > + while (tod_events[i]) { > + memset(&attr, 0, sizeof(attr)); > + len = strlen(tod_events[i]); > + > + if (parse_single_tracepoint_event(TRACE_TOD_SUBSYSTEM, > + tod_events[i], len, &attr, NULL) == EVT_FAILED) { > + error("Failed to parse event %s\n", tod_events[i]); > + goto out; > + } > + > + evsel = perf_evsel__new(&attr, evsel_list->nr_entries); > + if (evsel == NULL) > + return -1; > + > + perf_evlist__add(evsel_list, evsel); > + > + /* +2 for ':' delimiter and string terminator */ > + evsel->name = calloc(TRACE_TOD_SUBSYSTEM_LEN + len + 2, 1); > + if (!evsel->name) > + return -1; > + > + sprintf(evsel->name, "timekeeping:%s", tod_events[i]); > + > + tod_sample_type |= attr.sample_type; > + > + ++i; > + } > + > + /* > + * right now sample_type for all samples needs to be the same. > + * tracepoints are collected at sample period 1 and hence do not > + * request the period with the sample. However, default for record > + * is cycles at a frequency. So, until this sample_type mess is > + * fixed.... > + */ > + if (freq) > + tod_sample_type |= PERF_SAMPLE_PERIOD; > + > + rc = 0; > +out: > + return rc; > +}
So I feel uncomfortable with this tod_sample_type hack. I think we can't really continue with this fixed sample_type per session given the kind of hacks that involves. One thing we could do is to split session->sample_type into an array with one sample type per event type (hardware, breakpoint, software, tracepoint). And then each builtin tool can provide their constraints on top of these values: - builtin-report wants sample_type[HARDWARE] == sample_type[SOFTWARE] == sample_type[TRACEPOINT] == sample_type[BREAKPOINT] although that may be tunable by the time but we can start with that. - builtin-script has no specific constraints, except that sample_type[i] meets what the user passed as a parameter - etc.. Constraints can probably default to sample_type[i] == sample_type[i+1] to mimic the current behaviour. Then tools can override that. What do you think? -- To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
