Merge pull request #110 from mariux64/verify-job-pids-after-restart

Verify job pids after restart

mx_proc.c

@@ -484,3 +484,27 @@ pid_t mx_proc_get_parent(pid_t pid) {
         return -1;
     return parent;
 }
+
+/*
+ * get "comm" value of a process.
+ * buf MUST point to char[16] array
+ * return NULL on any error, otherwise pointer to buf
+ */
+char *mx_proc_get_comm(pid_t pid, char *buf) {
+    char commfilename[32];  // "/proc/2147483647/comm" = 22
+    FILE *commfile;
+    char *p;
+
+    snprintf(commfilename, sizeof(commfilename), "/proc/%d/comm", pid);
+    commfile = fopen(commfilename, "r");
+    if (commfile == NULL)
+        return NULL;
+    p = fgets(buf, 16, commfile);
+    fclose(commfile);
+    if (p == NULL)
+        return NULL;
+    p = strrchr(buf, '\n');
+    if (p != NULL)
+        *p = '\0';
+    return (char *)buf;
+}

mx_proc.h

@@ -86,4 +86,5 @@ int mx_proc_tree_free(struct mx_proc_tree **tree);
 
 struct mx_proc_info *mx_proc_tree_proc_info(struct mx_proc_tree *tree, pid_t pid);
 pid_t mx_proc_get_parent(pid_t pid);
+char *mx_proc_get_comm(pid_t pid, char *buf);
 #endif

mxqd.c

@@ -1187,6 +1187,16 @@ int user_process(struct mxq_group_list *glist, struct mxq_job *job)
 
 static const char REAPER_PNAME[] = "mxqd reaper";
 
+static int is_reaper(pid_t pid) {
+    char comm[16];
+    if (mx_proc_get_comm(pid, comm) == NULL)
+        return 0;
+    if (strcmp(comm, REAPER_PNAME) == 0)
+        return 1;
+    else
+        return 0;
+}
+
 int reaper_process(struct mxq_server *server,struct mxq_group_list *glist, struct mxq_job *job) {
     pid_t pid;
     struct rusage rusage;
@@ -1380,6 +1390,8 @@ unsigned long start_job(struct mxq_group_list *glist)
         job->host_pid = getpid();
 
         mx_log_debug("starting reaper process.");
+        mx_funlock(server->flock);
+        server->flock = NULL;
         mx_mysql_finish(&server->mysql);
 
         res = reaper_process(server, glist, job);
@@ -2285,8 +2297,6 @@ static int lost_scan_one(struct mxq_server *server)
 
     struct mxq_job *job;
 
-    int res;
-
     for (ulist = server->users; ulist; ulist = ulist->next) {
         for (glist = ulist->groups; glist; glist = glist->next) {
             for (jlist = glist->jobs; jlist; jlist = jlist->next) {
@@ -2304,13 +2314,9 @@ static int lost_scan_one(struct mxq_server *server)
                     continue;
                 }
 
-                res = kill(job->host_pid, 0);
-                if (res >= 0)
+                if (is_reaper(job->host_pid))
                     continue;
 
-                if (errno != ESRCH)
-                    return -errno;
-
                 if (!fspool_file_exists(server, job->job_id)) {
                     mx_log_warning("pid %u: process is gone. setting job status of job %lu to unknown.",
                                 jlist->job.host_pid,