_sysRefreshAllGpuRecoveryAction mask deadlock fix (massive issue) & more

src/nvidia/src/kernel/compute/fabric.c

@@ -892,12 +892,12 @@ fabricConstruct_IMPL
 
     return NV_OK;
 
-//TODO: Remove the WAR to suppress unused label warning
-goto fail;
-fail:
-    fabricDestruct_IMPL(pFabric);
-    return status;
-}
+// Removed the WAR to suppress unused label warning.
+
+    fail:
+        fabricDestruct_IMPL(pFabric);
+        return status;
+    }
 
 void
 fabricDestruct_IMPL

src/nvidia/src/kernel/core/hal_mgr.c

@@ -77,11 +77,13 @@ halmgrCreateHal_IMPL
 
     NV_ASSERT_OR_RETURN(halImpl < HAL_IMPL_MAXIMUM, NV_ERR_INVALID_ARGUMENT);
 
+    // Guard against duplicate registration, would silently leak the existing object mem leak
+    NV_ASSERT_OR_RETURN(pHalMgr->pHalList[halImpl] == NULL, NV_ERR_INVALID_STATE);
+
     status = objCreate(&pHal, pHalMgr, OBJHAL);
     if (status != NV_OK)
         return status;
 
-    // Store away the object pointer for this particular HAL object
     pHalMgr->pHalList[halImpl] = pHal;
 
     return NV_OK;

src/nvidia/src/kernel/core/system.c

@@ -861,6 +861,8 @@ sysSyncExternalFabricMgmtWAR_IMPL
     return status;
 }
 
+
+// use separate gpuLockMask for acquire and release so gpumgrGetGpuAttachInfo can't clobber it and cause a lock imbalance.
 static void
 _sysRefreshAllGpuRecoveryAction
 (
@@ -872,10 +874,11 @@ _sysRefreshAllGpuRecoveryAction
     NvU32      i;
     NvU32      gpuCount;
     NvU32      gpuIndex;
+    NvU32      gpuLockMask;
     NvU32      gpuMask;
 
     NV_ASSERT_OK_OR_ELSE(status,
-        rmGpuGroupLockAcquire(0, GPU_LOCK_GRP_ALL, GPUS_LOCK_FLAGS_NONE, RM_LOCK_MODULES_NONE, &gpuMask),
+        rmGpuGroupLockAcquire(0, GPU_LOCK_GRP_ALL, GPUS_LOCK_FLAGS_NONE, RM_LOCK_MODULES_NONE, &gpuLockMask),
         return);
 
     gpumgrGetGpuAttachInfo(&gpuCount, &gpuMask);
@@ -888,7 +891,7 @@ _sysRefreshAllGpuRecoveryAction
         }
     }
 
-    rmGpuGroupLockRelease(gpuMask, GPUS_LOCK_FLAGS_NONE);
+    rmGpuGroupLockRelease(gpuLockMask, GPUS_LOCK_FLAGS_NONE);
 }
 
 void

src/nvidia/src/kernel/mem_mgr/console_mem.c

@@ -72,8 +72,13 @@ conmemConstruct_IMPL
     }
 
     NV_ASSERT(pMemDesc->Allocated == 0);
-    memdescAddRef(pMemDesc);
-    pMemDesc->DupCount++;
+
+    // fixed DupCount being incremented manually without a matching decrement on failure.
+
+    if (status != NV_OK) {
+    pMemDesc->DupCount--;
+    memdescDestroy(pMemDesc);
+}
 
     //
     // NV01_MEMORY_FRAMEBUFFER_CONSOLE is just a way to get at the reserved

src/nvidia/src/kernel/power/gpu_boost_mgr.c

@@ -547,6 +547,7 @@ gpuboostmgrGpuItr_IMPL
  * @return NV_ERR_OBJECT_NOT_FOUND if SGBG is not found
  * @return  NV_ERR_INVALID_ARGUMENT Null or incorrect arguments passed in.
  */
+
 NV_STATUS
 gpuboostmgrGetBoostGrpIdFromGpu_IMPL
 (
@@ -556,16 +557,20 @@ gpuboostmgrGetBoostGrpIdFromGpu_IMPL
 )
 {
     NvU32   i;
-    NvU32   index = 0;
+    NvU32   index;
     OBJGPU *pGpuTemp = NULL;
 
     *pBoostGrpId = NV0000_SYNC_GPU_BOOST_INVALID_GROUP_ID;
 
     NV_ASSERT_OR_RETURN(NULL != pGpu, NV_ERR_INVALID_ARGUMENT);
     NV_ASSERT_OR_RETURN(NULL != pBoostGrpId, NV_ERR_INVALID_ARGUMENT);
 
-    for (i = 0; i < pBoostMgr->groupCount; i++)
+    for (i = 0; i < NV0000_SYNC_GPU_BOOST_MAX_GROUPS; i++)
     {
+        if (NV_OK != gpuboostmgrValidateGroupId(pBoostMgr, i))
+            continue;
+
+         index = 0;
         while (NULL != (pGpuTemp = gpuboostmgrGpuItr(pBoostMgr, i, &index)))
         {
             if (pGpuTemp->gpuId == pGpu->gpuId)