[
https://issues.apache.org/jira/browse/TS-3104?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Victor updated TS-3104:
-----------------------
Description:
In some cases traffic_cop can't restart traffic_manager properly. We met these
issues at "Ashmanov and partners" (http://en.ashmanov.com/). There are two
places in code which in my opinion need corrections:
1) The logic which decides whether to kill process or group.
2) The main traffic_cop loop: it doesn't reinitialize manager API in case of
failure and this fact leads to constant attempts to connect to manager using
socket id == -1.
I have prepared patches for both issues. Please kindly take a look at them and
let me know your thoughts.
was:
In some cases traffic_cop can't restart traffic_manager properly. We met these
issues at "Ashmanov and partners" (http://en.ashmanov.com/). There are two
places in code which in my opinion need corrections:
1) The logic which decides whether to kill process or group.
2) The main traffic_cop loop: it doesn't reinitialize manager API in case of
failure and this fact leads to constant attempts to connect to manager using
socket id == -1.
I have prepared patches for both issues. Please kindly take a look at them and
let me know your thoughts.
diff --git lib/ts/lockfile.cc lib/ts/lockfile.cc
index f6e9587..dbd7394 100644
--- lib/ts/lockfile.cc
+++ lib/ts/lockfile.cc
@@ -241,6 +241,7 @@ Lockfile::KillGroup(int sig, int initial_sig, const char
*pname)
int err;
pid_t pid;
pid_t holding_pid;
+ pid_t self = getpid();
err = Open(&holding_pid);
if (err == 1) // success getting the lock file
@@ -252,12 +253,20 @@ Lockfile::KillGroup(int sig, int initial_sig, const char
*pname)
pid = getpgid(holding_pid);
} while ((pid < 0) && (errno == EINTR));
- if ((pid < 0) || (pid == getpid()))
+ if ((pid < 0) || (pid == self)) {
+ // Error getting process group,
+ // or we are the group's owner.
+ // Let's kill just holding_pid
pid = holding_pid;
-
- if (pid != 0) {
+ }
+ else if (pid != self) {
+ // We managed to get holding_pid's process group
+ // and this group is not ours.
// This way, we kill the process_group:
pid = -pid;
+ }
+
+ if (pid != 0) {
// In order to get core files from each process, please
// set your core_pattern appropriately.
lockfile_kill_internal(holding_pid, initial_sig, pid, pname, sig);
diff --git cop/TrafficCop.cc cop/TrafficCop.cc
index 307270e..56bc6d2 100644
--- cop/TrafficCop.cc
+++ cop/TrafficCop.cc
@@ -59,6 +59,7 @@ static const char COP_TRACE_FILE[] = "/tmp/traffic_cop.trace";
#define COP_FATAL LOG_ALERT
#define COP_WARNING LOG_ERR
+#define COP_INFO LOG_INFO
#define COP_DEBUG LOG_DEBUG
Diags * g_diags; // link time dependency
@@ -131,6 +132,9 @@ static int child_pid = 0;
static int child_status = 0;
static int sem_id = 11452;
+// manager API is initialized
+static bool mgmt_init = false;
+
AppVersionInfo appVersionInfo;
static char const localhost[] = "127.0.0.1";
@@ -1142,6 +1146,7 @@ test_mgmt_cli_port()
if (TSRecordGetString("proxy.config.manager_binary", &val) != TS_ERR_OKAY) {
cop_log(COP_WARNING, "(cli test) unable to retrieve manager_binary\n");
+ mgmt_init = false;
ret = -1;
} else {
if (strcmp(val, manager_binary) != 0) {
@@ -1544,7 +1549,6 @@ check_no_run()
static void*
check(void *arg)
{
- bool mgmt_init = false;
cop_log_trace("Entering check()\n");
for (;;) {
@@ -1593,6 +1597,7 @@ check(void *arg)
// We do this after the first round of checks, since the first "check"
will spawn traffic_manager
if (!mgmt_init) {
+ cop_log(COP_INFO, "Initializing manager API\n");
TSInit(Layout::get()->runtimedir,
static_cast<TSInitOptionT>(TS_MGMT_OPT_NO_EVENTS | TS_MGMT_OPT_NO_SOCK_TESTS));
mgmt_init = true;
}
> traffic_cop can't restart traffic_manager properly
> --------------------------------------------------
>
> Key: TS-3104
> URL: https://issues.apache.org/jira/browse/TS-3104
> Project: Traffic Server
> Issue Type: Bug
> Components: Cop
> Reporter: Victor
> Attachments: ts-0022-fix-lockfile-killgroup.patch,
> ts-0023-cop-reinit-mgr-api-on-failure.patch
>
>
> In some cases traffic_cop can't restart traffic_manager properly. We met
> these issues at "Ashmanov and partners" (http://en.ashmanov.com/). There are
> two places in code which in my opinion need corrections:
> 1) The logic which decides whether to kill process or group.
> 2) The main traffic_cop loop: it doesn't reinitialize manager API in case of
> failure and this fact leads to constant attempts to connect to manager using
> socket id == -1.
> I have prepared patches for both issues. Please kindly take a look at them
> and let me know your thoughts.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
