multi-tenant: fix loader dead lock

A dead lock could occur at start up, where a loader thread would
get stuck on it's condition variable, while the main thread was
polling the loaders task results.

The vector to the dead lock is as follows:

main	                        loader
DetectEngineMultiTenantSetup
-DetectLoaderSetupLoadTenant
--DetectLoaderQueueTask
---lock loader
---add task
---unlock loader
	                        lock loader
	                        check/exec tasks
	                        unlock loader
---wake up threads
	                        lock ctrl mutx
	                        cond wait ctrl
	                        unlock ctrl
-DetectLoadersSync
--lock loader
--check tasks
--unlock loader

Between the main thread unlocking the loader and waking up the
threads, it is possible that the loader has already moved ahead
but not yet entered its conditional wait. The main thread sends
its condition signal, but since the loader isn't yet waiting on
it the signal is ignored. Then when the loader does enter its
conditional wait, the signal is not sent again.

This patch updates the logic to send signals much more often.
It also makes sure that the signal is sent under lock, as the
API requires.

Bug: #6768.

Co-authored-by: Shivani Bhardwaj <shivani@oisf.net>

diff --git a/src/detect-engine-loader.c b/src/detect-engine-loader.c
index e41f27779f5dce5b5bf1bea00959923735289766..0cdb453388ab9616ce913d2a0c5ba133b60b98c2 100644 (file)
--- a/src/detect-engine-loader.c
+++ b/src/detect-engine-loader.c
@@ -456,6 +456,12 @@ int DetectLoadersSync(void)
                 done = true;
             }
             SCMutexUnlock(&loader->m);
+            if (!done) {
+                /* nudge thread in case it's sleeping */
+                SCCtrlMutexLock(loader->tv->ctrl_mutex);
+                pthread_cond_broadcast(loader->tv->ctrl_cond);
+                SCCtrlMutexUnlock(loader->tv->ctrl_mutex);
+            }
         }
         SCMutexLock(&loader->m);
         if (loader->result != 0) {
@@ -511,7 +517,9 @@ static void TmThreadWakeupDetectLoaderThreads(void)
         while (tv != NULL) {
             if (strncmp(tv->name,"DL#",3) == 0) {
                 BUG_ON(tv->ctrl_cond == NULL);
+                SCCtrlMutexLock(tv->ctrl_mutex);
                 pthread_cond_broadcast(tv->ctrl_cond);
+                SCCtrlMutexUnlock(tv->ctrl_mutex);
             }
             tv = tv->next;
         }
@@ -555,6 +563,9 @@ static TmEcode DetectLoaderThreadInit(ThreadVars *t, const void *initdata, void
     /* pass thread data back to caller */
     *data = ftd;
 
+    DetectLoaderControl *loader = &loaders[ftd->instance];
+    loader->tv = t;
+
     return TM_ECODE_OK;
 }
 

diff --git a/src/detect-engine-loader.h b/src/detect-engine-loader.h
index 7ffb8c8648a08c8cb4ebf0edaf7db7fc32de4a74..f43ff9a5491c11112a2dd5847b08517d1639dd14 100644 (file)
--- a/src/detect-engine-loader.h
+++ b/src/detect-engine-loader.h
@@ -43,9 +43,14 @@ typedef struct DetectLoaderTask_ {
 
 typedef struct DetectLoaderControl_ {
     int id;
-    int result;     /* 0 for ok, error otherwise */
-    SCMutex m;
-    TAILQ_HEAD(, DetectLoaderTask_) task_list;
+    ThreadVars *tv; /**< loader threads threadvars - for waking them up */
+
+    /** struct to group members and mutex */
+    struct {
+        SCMutex m;  /**< mutex protects result and task_list */
+        int result; /**< 0 for ok, error otherwise */
+        TAILQ_HEAD(, DetectLoaderTask_) task_list;
+    };
 } DetectLoaderControl;
 
 int DetectLoaderQueueTask(int loader_id, LoaderFunc Func, void *func_ctx, LoaderFreeFunc FreeFunc);