Fix: sessiond: kernel error accounting fd still open when unloading modules

[lttng-tools.git] / src / bin / lttng-sessiond / main.c

diff --git a/src/bin/lttng-sessiond/main.c b/src/bin/lttng-sessiond/main.c
index bb84be1ad1475ee491ad9d2e807b481f56f522fb..e0b5595f3b25adbd6fb98332420d9cc458435060 100644 (file)
--- a/src/bin/lttng-sessiond/main.c
+++ b/src/bin/lttng-sessiond/main.c
@@ -74,6 +74,7 @@
 #include "register.h"
 #include "manage-apps.h"
 #include "manage-kernel.h"
+#include "modprobe.h"
 
 static const char *help_msg =
 #ifdef LTTNG_EMBED_HELP
@@ -1887,7 +1888,7 @@ stop_threads:
        sessiond_cleanup();
 
        /*
-        * Wait for all pending call_rcu work to complete tearing shutting down
+        * Wait for all pending call_rcu work to complete before shutting down
         * the notification thread. This call_rcu work includes shutting down
         * UST apps and event notifier pipes.
         */
@@ -1899,12 +1900,26 @@ stop_threads:
        }
 
        /*
-        * Teardown of error accounting needs be done after the teardown of the
-        * notification thread as all error buckets must have been released by
-        * their users (conditions).
+        * Error accounting teardown has to be done after the teardown of all
+        * event notifier pipes to ensure that no tracer may try to use the
+        * error accounting facilities.
         */
        event_notifier_error_accounting_fini();
 
+       /*
+        * Unloading the kernel modules needs to be done after all kernel
+        * ressources have been released. In our case, this includes the
+        * notification fd, the event notifier group fd, error accounting fd,
+        * all event and event notifier fds, etc.
+        *
+        * In short, at this point, we need to have called close() on all fds
+        * received from the kernel tracer.
+        */
+       if (is_root && !config.no_kernel) {
+               DBG("Unloading kernel modules");
+               modprobe_remove_lttng_all();
+       }
+
        /*
         * Ensure all prior call_rcu are done. call_rcu callbacks may push
         * hash tables to the ht_cleanup thread. Therefore, we ensure that