perf trace: Add support for pagefault tracing
authorStanislav Fomichev <stfomichev@yandex-team.ru>
Thu, 26 Jun 2014 16:14:25 +0000 (20:14 +0400)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Thu, 26 Jun 2014 19:07:43 +0000 (16:07 -0300)
This patch adds optional pagefault tracing support to 'perf trace'.

Using -F/--pf option user can specify whether he wants minor, major or
all pagefault events to be traced. This patch adds only live mode,
record and replace will come in a separate patch.

Example output:

  1756272.905 ( 0.000 ms): curl/5937 majfault [0x7fa7261978b6] => /usr/lib/x86_64-linux-gnu/libkrb5.so.26.0.0@0x85288 (d.)
  1862866.036 ( 0.000 ms): wget/8460 majfault [__clear_user+0x3f] => 0x659cb4 (?k)

Signed-off-by: Stanislav Fomichev <stfomichev@yandex-team.ru>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1403799268-1367-3-git-send-email-stfomichev@yandex-team.ru
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/Documentation/perf-trace.txt
tools/perf/builtin-trace.c

index fae38d9a44a4a0750ed069c9264e3e2cd96b59ea..72397d9aa2ec5e1b79abfcdaf1a75551921053cd 100644 (file)
@@ -107,6 +107,45 @@ the thread executes on the designated CPUs. Default is to monitor all CPUs.
        Show tool stats such as number of times fd->pathname was discovered thru
        hooking the open syscall return + vfs_getname or via reading /proc/pid/fd, etc.
 
+-F=[all|min|maj]::
+--pf=[all|min|maj]::
+       Trace pagefaults. Optionally, you can specify whether you want minor,
+       major or all pagefaults. Default value is maj.
+
+PAGEFAULTS
+----------
+
+When tracing pagefaults, the format of the trace is as follows:
+
+<min|maj>fault [<ip.symbol>+<ip.offset>] => <addr.dso@addr.offset> (<map type><addr level>).
+
+- min/maj indicates whether fault event is minor or major;
+- ip.symbol shows symbol for instruction pointer (the code that generated the
+  fault); if no debug symbols available, perf trace will print raw IP;
+- addr.dso shows DSO for the faulted address;
+- map type is either 'd' for non-executable maps or 'x' for executable maps;
+- addr level is either 'k' for kernel dso or '.' for user dso.
+
+For symbols resolution you may need to install debugging symbols.
+
+Please be aware that duration is currently always 0 and doesn't reflect actual
+time it took for fault to be handled!
+
+When --verbose specified, perf trace tries to print all available information
+for both IP and fault address in the form of dso@symbol+offset.
+
+EXAMPLES
+--------
+
+Trace syscalls, major and minor pagefaults:
+
+ $ perf trace -F all
+
+  1416.547 ( 0.000 ms): python/20235 majfault [CRYPTO_push_info_+0x0] => /lib/x86_64-linux-gnu/libcrypto.so.1.0.0@0x61be0 (x.)
+
+  As you can see, there was major pagefault in python process, from
+  CRYPTO_push_info_ routine which faulted somewhere in libcrypto.so.
+
 SEE ALSO
 --------
 linkperf:perf-record[1], linkperf:perf-script[1]
index 4a9e26b731febdd334e56155d75b66830ff5ae34..1985c3b8cc0640daf9ede158ee86a29732920eb4 100644 (file)
@@ -1178,6 +1178,9 @@ fail:
        return NULL;
 }
 
+#define TRACE_PFMAJ            (1 << 0)
+#define TRACE_PFMIN            (1 << 1)
+
 struct trace {
        struct perf_tool        tool;
        struct {
@@ -1212,6 +1215,7 @@ struct trace {
        bool                    summary_only;
        bool                    show_comm;
        bool                    show_tool_stats;
+       int                     trace_pgfaults;
 };
 
 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
@@ -1773,6 +1777,68 @@ out_dump:
        return 0;
 }
 
+static void print_location(FILE *f, struct perf_sample *sample,
+                          struct addr_location *al,
+                          bool print_dso, bool print_sym)
+{
+
+       if ((verbose || print_dso) && al->map)
+               fprintf(f, "%s@", al->map->dso->long_name);
+
+       if ((verbose || print_sym) && al->sym)
+               fprintf(f, "%s+0x%lx", al->sym->name,
+                       al->addr - al->sym->start);
+       else if (al->map)
+               fprintf(f, "0x%lx", al->addr);
+       else
+               fprintf(f, "0x%lx", sample->addr);
+}
+
+static int trace__pgfault(struct trace *trace,
+                         struct perf_evsel *evsel,
+                         union perf_event *event,
+                         struct perf_sample *sample)
+{
+       struct thread *thread;
+       u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
+       struct addr_location al;
+       char map_type = 'd';
+
+       thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
+
+       thread__find_addr_location(thread, trace->host, cpumode, MAP__FUNCTION,
+                             sample->ip, &al);
+
+       trace__fprintf_entry_head(trace, thread, 0, sample->time, trace->output);
+
+       fprintf(trace->output, "%sfault [",
+               evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
+               "maj" : "min");
+
+       print_location(trace->output, sample, &al, false, true);
+
+       fprintf(trace->output, "] => ");
+
+       thread__find_addr_location(thread, trace->host, cpumode, MAP__VARIABLE,
+                                  sample->addr, &al);
+
+       if (!al.map) {
+               thread__find_addr_location(thread, trace->host, cpumode,
+                                          MAP__FUNCTION, sample->addr, &al);
+
+               if (al.map)
+                       map_type = 'x';
+               else
+                       map_type = '?';
+       }
+
+       print_location(trace->output, sample, &al, true, false);
+
+       fprintf(trace->output, " (%c%c)\n", map_type, al.level);
+
+       return 0;
+}
+
 static bool skip_sample(struct trace *trace, struct perf_sample *sample)
 {
        if ((trace->pid_list && intlist__find(trace->pid_list, sample->pid)) ||
@@ -1887,6 +1953,30 @@ static void perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
        perf_evlist__add(evlist, evsel);
 }
 
+static int perf_evlist__add_pgfault(struct perf_evlist *evlist,
+                                   u64 config)
+{
+       struct perf_evsel *evsel;
+       struct perf_event_attr attr = {
+               .type = PERF_TYPE_SOFTWARE,
+               .mmap_data = 1,
+               .sample_period = 1,
+       };
+
+       attr.config = config;
+
+       event_attr_init(&attr);
+
+       evsel = perf_evsel__new(&attr);
+       if (!evsel)
+               return -ENOMEM;
+
+       evsel->handler = trace__pgfault;
+       perf_evlist__add(evlist, evsel);
+
+       return 0;
+}
+
 static int trace__run(struct trace *trace, int argc, const char **argv)
 {
        struct perf_evlist *evlist = perf_evlist__new();
@@ -1907,6 +1997,14 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
 
        perf_evlist__add_vfs_getname(evlist);
 
+       if ((trace->trace_pgfaults & TRACE_PFMAJ) &&
+           perf_evlist__add_pgfault(evlist, PERF_COUNT_SW_PAGE_FAULTS_MAJ))
+               goto out_error_tp;
+
+       if ((trace->trace_pgfaults & TRACE_PFMIN) &&
+           perf_evlist__add_pgfault(evlist, PERF_COUNT_SW_PAGE_FAULTS_MIN))
+               goto out_error_tp;
+
        if (trace->sched &&
                perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
                                trace__sched_stat_runtime))
@@ -1987,7 +2085,8 @@ again:
                                goto next_event;
                        }
 
-                       if (sample.raw_data == NULL) {
+                       if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
+                           sample.raw_data == NULL) {
                                fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
                                       perf_evsel__name(evsel), sample.tid,
                                       sample.cpu, sample.raw_size);
@@ -2269,6 +2368,23 @@ static int trace__open_output(struct trace *trace, const char *filename)
        return trace->output == NULL ? -errno : 0;
 }
 
+static int parse_pagefaults(const struct option *opt, const char *str,
+                           int unset __maybe_unused)
+{
+       int *trace_pgfaults = opt->value;
+
+       if (strcmp(str, "all") == 0)
+               *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
+       else if (strcmp(str, "maj") == 0)
+               *trace_pgfaults |= TRACE_PFMAJ;
+       else if (strcmp(str, "min") == 0)
+               *trace_pgfaults |= TRACE_PFMIN;
+       else
+               return -1;
+
+       return 0;
+}
+
 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
 {
        const char * const trace_usage[] = {
@@ -2335,6 +2451,8 @@ int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
                    "Show only syscall summary with statistics"),
        OPT_BOOLEAN('S', "with-summary", &trace.summary,
                    "Show all syscalls and summary with statistics"),
+       OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
+                    "Trace pagefaults", parse_pagefaults, "maj"),
        OPT_END()
        };
        int err;
@@ -2349,6 +2467,11 @@ int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
        if (trace.summary_only)
                trace.summary = trace.summary_only;
 
+       if (trace.trace_pgfaults) {
+               trace.opts.sample_address = true;
+               trace.opts.sample_time = true;
+       }
+
        if (output_name != NULL) {
                err = trace__open_output(&trace, output_name);
                if (err < 0) {