perf tools: Skip PERF_RECORD_MMAP event synthesis for kernel threads
authorNamhyung Kim <namhyung@kernel.org>
Tue, 2 Feb 2021 09:01:17 +0000 (18:01 +0900)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Wed, 3 Feb 2021 16:10:44 +0000 (13:10 -0300)
To synthesize information to resolve sample IPs, it needs to scan task
and mmap info from the /proc filesystem.  For each process, it opens
(and reads) status and maps file respectively.  But as kernel threads
don't have memory maps so we can skip the maps file.

To find kernel threads, check "VmPeak:" line in /proc/<PID>/status file.
It's about the peak virtual memory usage so only user-level tasks have
that.  Note that it's possible to miss the line due to partial reads.
So we should double-check if it's a really kernel thread when there's no
VmPeak line.

Thus check "Threads:" line (which follows the VmPeak line whether or not
it exists) to be sure it's read enough data - just in case of deeply
nested pid namespaces or large number of supplementary groups are
involved.

This is for user process:

  $ head -40 /proc/1/status
  Name: systemd
  Umask: 0000
  State: S (sleeping)
  Tgid: 1
  Ngid: 0
  Pid: 1
  PPid: 0
  TracerPid: 0
  Uid: 0 0 0 0
  Gid: 0 0 0 0
  FDSize: 256
  Groups:
  NStgid: 1
  NSpid: 1
  NSpgid: 1
  NSsid: 1
  VmPeak:   234192 kB           <-- here
  VmSize:   169964 kB
  VmLck:        0 kB
  VmPin:        0 kB
  VmHWM:    29528 kB
  VmRSS:     6104 kB
  RssAnon:     2756 kB
  RssFile:     3348 kB
  RssShmem:        0 kB
  VmData:    19776 kB
  VmStk:     1036 kB
  VmExe:      784 kB
  VmLib:     9532 kB
  VmPTE:      116 kB
  VmSwap:     2400 kB
  HugetlbPages:        0 kB
  CoreDumping: 0
  THP_enabled: 1
  Threads: 1                     <-- and here
  SigQ: 1/62808
  SigPnd: 0000000000000000
  ShdPnd: 0000000000000000
  SigBlk: 7be3c0fe28014a03
  SigIgn: 0000000000001000

And this is for kernel thread:

  $ head -20 /proc/2/status
  Name: kthreadd
  Umask: 0000
  State: S (sleeping)
  Tgid: 2
  Ngid: 0
  Pid: 2
  PPid: 0
  TracerPid: 0
  Uid: 0 0 0 0
  Gid: 0 0 0 0
  FDSize: 64
  Groups:
  NStgid: 2
  NSpid: 2
  NSpgid: 0
  NSsid: 0
  Threads: 1                     <-- here
  SigQ: 1/62808
  SigPnd: 0000000000000000
  ShdPnd: 0000000000000000

Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: https://lore.kernel.org/r/20210202090118.2008551-3-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/util/synthetic-events.c

index 0cc9986..abea688 100644 (file)
@@ -70,13 +70,13 @@ int perf_tool__process_synth_event(struct perf_tool *tool,
  * the comm, tgid and ppid.
  */
 static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len,
-                                   pid_t *tgid, pid_t *ppid)
+                                   pid_t *tgid, pid_t *ppid, bool *kernel)
 {
        char bf[4096];
        int fd;
        size_t size = 0;
        ssize_t n;
-       char *name, *tgids, *ppids;
+       char *name, *tgids, *ppids, *vmpeak, *threads;
 
        *tgid = -1;
        *ppid = -1;
@@ -102,8 +102,14 @@ static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len
        bf[n] = '\0';
 
        name = strstr(bf, "Name:");
-       tgids = strstr(bf, "Tgid:");
-       ppids = strstr(bf, "PPid:");
+       tgids = strstr(name ?: bf, "Tgid:");
+       ppids = strstr(tgids ?: bf, "PPid:");
+       vmpeak = strstr(ppids ?: bf, "VmPeak:");
+
+       if (vmpeak)
+               threads = NULL;
+       else
+               threads = strstr(ppids ?: bf, "Threads:");
 
        if (name) {
                char *nl;
@@ -136,12 +142,17 @@ static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len
                pr_debug("PPid: string not found for pid %d\n", tid);
        }
 
+       if (!vmpeak && threads)
+               *kernel = true;
+       else
+               *kernel = false;
+
        return 0;
 }
 
 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, pid_t tid,
                                    struct machine *machine,
-                                   pid_t *tgid, pid_t *ppid)
+                                   pid_t *tgid, pid_t *ppid, bool *kernel)
 {
        size_t size;
 
@@ -152,7 +163,7 @@ static int perf_event__prepare_comm(union perf_event *event, pid_t pid, pid_t ti
        if (machine__is_host(machine)) {
                if (perf_event__get_comm_ids(pid, tid, event->comm.comm,
                                             sizeof(event->comm.comm),
-                                            tgid, ppid) != 0) {
+                                            tgid, ppid, kernel) != 0) {
                        return -1;
                }
        } else {
@@ -182,8 +193,10 @@ pid_t perf_event__synthesize_comm(struct perf_tool *tool,
                                         struct machine *machine)
 {
        pid_t tgid, ppid;
+       bool kernel_thread;
 
-       if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid) != 0)
+       if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid,
+                                    &kernel_thread) != 0)
                return -1;
 
        if (perf_tool__process_synth_event(tool, event, machine, process) != 0)
@@ -743,6 +756,7 @@ static int __event__synthesize_thread(union perf_event *comm_event,
        while ((dirent = readdir(tasks)) != NULL) {
                char *end;
                pid_t _pid;
+               bool kernel_thread;
 
                _pid = strtol(dirent->d_name, &end, 10);
                if (*end)
@@ -750,7 +764,7 @@ static int __event__synthesize_thread(union perf_event *comm_event,
 
                rc = -1;
                if (perf_event__prepare_comm(comm_event, pid, _pid, machine,
-                                            &tgid, &ppid) != 0)
+                                            &tgid, &ppid, &kernel_thread) != 0)
                        break;
 
                if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
@@ -768,7 +782,7 @@ static int __event__synthesize_thread(union perf_event *comm_event,
                        break;
 
                rc = 0;
-               if (_pid == pid) {
+               if (_pid == pid && !kernel_thread) {
                        /* process the parent's maps too */
                        rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
                                                process, machine, mmap_data);