apps.plugin, properly handle the kernel issues that does not include grandchildren and not-waited-for children into the parent resources
Costa Tsaousis committed
Jul 2, 2016 at 02:46 UTC
87d093eac8e346e807692abf577ad1e0c176cc8c
1 file changed
+52
-4
src/apps_plugin.c
+52
-4
@@ -1323,6 +1323,7 @@ int collect_data_for_all_processes_from_proc(void)
1323
// check: update_apps_groups_statistics()
1324
1325
void link_all_processes_to_their_parents(void) {
1326
+ struct pid_stat *init = all_pids[1];
1327
struct pid_stat *p = NULL;
1328
1329
// link all children to their parents
@@ -1372,10 +1373,57 @@ void link_all_processes_to_their_parents(void) {
1373
pp->fix_cutime += p->last_utime + p->last_cutime + p->last_fix_cutime;
1374
pp->fix_cstime += p->last_stime + p->last_cstime + p->last_fix_cstime;
1375
1375
- if(unlikely(pp->cminflt < pp->fix_cminflt)) pp->fix_cminflt = pp->cminflt;
1376
- if(unlikely(pp->cmajflt < pp->fix_cmajflt)) pp->fix_cmajflt = pp->cmajflt;
1377
- if(unlikely(pp->cutime < pp->fix_cutime)) pp->fix_cutime = pp->cutime;
1378
- if(unlikely(pp->cstime < pp->fix_cstime)) pp->fix_cstime = pp->cstime;
1376
+ // The known exited children (the ones we track) may have
1377
+ // contributed more than the value accumulated into the process
1378
+ // by the kernel.
1379
+ // This can happen if the parent process has not waited-for
1380
+ // its children (check: man 2 times).
1381
+ // In this case, the kernel adds these resources to init (pid 1).
1382
+ //
1383
+ // The following code, attempts to fix this.
1384
+ // Without this code, the charts will have random spikes
1385
+ // for example, when an SSH session ends (sshd forks a child
1386
+ // to serve the session, but when this session ends, sshd
1387
+ // does not wait-for its child, thus all the resources of the
1388
+ // ssh session get added to init, resulting in a huge spike on
1389
+ // the charts).
1390
+
1391
+ if(unlikely(pp->cminflt < pp->fix_cminflt)) {
1392
+ if(likely(init && pp != init)) {
1393
+ unsigned long long have = pp->fix_cminflt - pp->cminflt;
1394
+ unsigned long long max = init->cminflt - init->fix_cminflt;
1395
+ if(have > max) have = max;
1396
+ init->fix_cminflt += have;
1397
+ }
1398
+ pp->fix_cminflt = pp->cminflt;
1399
+ }
1400
+ if(unlikely(pp->cmajflt < pp->fix_cmajflt)) {
1401
+ if(likely(init && pp != init)) {
1402
+ unsigned long long have = pp->fix_cmajflt - pp->cmajflt;
1403
+ unsigned long long max = init->cmajflt - init->fix_cmajflt;
1404
+ if(have > max) have = max;
1405
+ init->fix_cmajflt += have;
1406
+ }
1407
+ pp->fix_cmajflt = pp->cmajflt;
1408
+ }
1409
+ if(unlikely(pp->cutime < pp->fix_cutime)) {
1410
+ if(likely(init && pp != init)) {
1411
+ unsigned long long have = pp->fix_cutime - pp->cutime;
1412
+ unsigned long long max = init->cutime - init->fix_cutime;
1413
+ if(have > max) have = max;
1414
+ init->fix_cutime += have;
1415
+ }
1416
+ pp->fix_cutime = pp->cutime;
1417
+ }
1418
+ if(unlikely(pp->cstime < pp->fix_cstime)) {
1419
+ if(likely(init && pp != init)) {
1420
+ unsigned long long have = pp->fix_cstime - pp->cstime;
1421
+ unsigned long long max = init->cstime - init->fix_cstime;
1422
+ if(have > max) have = max;
1423
+ init->fix_cstime += have;
1424
+ }
1425
+ pp->fix_cstime = pp->cstime;
1426
+ }
1427
1428
if(unlikely(debug))
1429
fprintf(stderr, "apps.plugin: \tupdating child metrics of %d (%s, %s) to its parent %d (%s, %s). Parent has now: utime=%llu, stime=%llu, minflt=%llu, majflt=%llu, cutime=%llu, cstime=%llu, cminflt=%llu, cmajflt=%llu, fix_cutime=%llu, fix_cstime=%llu, fix_cminflt=%llu, fix_cmajflt=%llu\n", p->pid, p->comm, p->updated?"running":"exited", pp->pid, pp->comm, pp->updated?"running":"exited", pp->utime, pp->stime, pp->minflt, pp->majflt, pp->cutime, pp->cstime, pp->cminflt, pp->cmajflt, pp->fix_cutime, pp->fix_cstime, pp->fix_cminflt, pp->fix_cmajflt);