Release coprocessor registers when a task is deleted.

components/freertos/include/freertos/portable.h

@@ -210,6 +210,7 @@ BaseType_t xPortInIsrContext();
 #if( portUSING_MPU_WRAPPERS == 1 )
 	struct xMEMORY_REGION;
 	void vPortStoreTaskMPUSettings( xMPU_SETTINGS *xMPUSettings, const struct xMEMORY_REGION * const xRegions, StackType_t *pxBottomOfStack, uint32_t usStackDepth ) PRIVILEGED_FUNCTION;
+	void vPortReleaseTaskMPUSettings( xMPU_SETTINGS *xMPUSettings );
 #endif
 
 /* Multi-core: get current core ID */

components/freertos/include/freertos/portmacro.h

@@ -172,6 +172,7 @@ typedef struct {
 #define portASSERT_IF_IN_ISR()        vPortAssertIfInISR()
 void vPortAssertIfInISR();
 
+
 #define portCRITICAL_NESTING_IN_TCB 1 
 
 /*
@@ -313,6 +314,10 @@ typedef struct {
 	#define PRIVILEGED_DATA
 #endif
 
+
+void _xt_coproc_release(void * coproc_sa_base);
+
+
 // porttrace
 #if configUSE_TRACE_FACILITY_2
 #include "porttrace.h"

components/freertos/port.c

@@ -253,6 +253,13 @@ void vPortStoreTaskMPUSettings( xMPU_SETTINGS *xMPUSettings, const struct xMEMOR
 	 */
 	#endif
 }
+
+void vPortReleaseTaskMPUSettings( xMPU_SETTINGS *xMPUSettings )
+{
+	/* If task has live floating point registers somewhere, release them */
+	_xt_coproc_release( xMPUSettings->coproc_area );
+}
+
 #endif
 
 /*

components/freertos/tasks.c

@@ -3550,9 +3550,16 @@ static void prvCheckTasksWaitingTermination( void )
 
 				{
 					pxTCB = ( TCB_t * ) listGET_OWNER_OF_HEAD_ENTRY( ( &xTasksWaitingTermination ) );
-					( void ) uxListRemove( &( pxTCB->xGenericListItem ) );
-					--uxCurrentNumberOfTasks;
-					--uxTasksDeleted;
+					/* We only want to kill tasks that ran on this core because e.g. _xt_coproc_release needs to
+					   be called on the core the process is pinned on, if any */
+					if( pxTCB->xCoreID == tskNO_AFFINITY || pxTCB->xCoreID == xPortGetCoreID()) {
+						( void ) uxListRemove( &( pxTCB->xGenericListItem ) );
+						--uxCurrentNumberOfTasks;
+						--uxTasksDeleted;
+					} else {
+						/* Need to wait until the idle task on the other processor kills that task first. */
+						break;
+					}
 				}
 				
 				#if ( configNUM_THREAD_LOCAL_STORAGE_POINTERS > 0 ) && ( configTHREAD_LOCAL_STORAGE_DELETE_CALLBACKS )
@@ -3770,6 +3777,10 @@ BaseType_t xTaskGetAffinity( TaskHandle_t xTask )
 		}
 		#endif /* configUSE_NEWLIB_REENTRANT */
 
+		#if ( portUSING_MPU_WRAPPERS == 1 )
+			vPortReleaseTaskMPUSettings( &( pxTCB->xMPUSettings) );
+		#endif
+
 		#if( ( configSUPPORT_DYNAMIC_ALLOCATION == 1 ) && ( configSUPPORT_STATIC_ALLOCATION == 0 ) && ( portUSING_MPU_WRAPPERS == 0 ) )
 		{
 			/* The task can only have been allocated dynamically - free both

components/freertos/xtensa_context.S

@@ -378,6 +378,9 @@ May be called when a thread terminates or completes but does not delete
 the co-proc save area, to avoid the exception handler having to save the 
 thread's co-proc state before another thread can use it (optimization).
 
+Needs to be called on the processor the thread was running on. Unpinned threads
+won't have an entry here because they get pinned as soon they use a coprocessor.
+
 Entry Conditions:
     A2  = Pointer to base of co-processor state save area.
 

components/freertos/xtensa_vectors.S

@@ -931,7 +931,6 @@ _xt_coproc_exc:
     addx4   a0, a5, a0                      /* a0 = &_xt_coproc_mask[n] */
     l32i    a0, a0, 0                       /* a0 = (n << 16) | (1 << n) */
 
-    /* TODO: Remove this as soon as coprocessor state moving works across cores - JD */
     /* FPU operations are incompatible with non-pinned tasks. If we have a FPU operation
        here, to keep the entire thing from crashing, it's better to pin the task to whatever
        core we're running on now. */
@@ -944,14 +943,21 @@ _xt_coproc_exc:
 
     /* Grab correct xt_coproc_owner_sa for this core */
     movi    a2, XCHAL_CP_MAX << 2
-    mull    a2, a2, a3
+    mull    a2, a2, a3                      /* multiply by current processor id */
     movi    a3, _xt_coproc_owner_sa         /* a3 = base of owner array */
-    add     a3, a3, a2
+    add     a3, a3, a2                      /* a3 = owner area needed for this processor */
 
     extui   a2, a0, 0, 16                   /* coprocessor bitmask portion */
     or      a4, a4, a2                      /* a4 = CPENABLE | (1 << n) */
     wsr     a4, CPENABLE
 
+/* 
+Keep loading _xt_coproc_owner_sa[n] atomic (=load once, then use that value
+everywhere): _xt_coproc_release assumes it works like this in order not to need
+locking.
+*/
+
+
     /* Get old coprocessor owner thread (save area ptr) and assign new one.  */
     addx4   a3,  a5, a3                      /* a3 = &_xt_coproc_owner_sa[n] */
     l32i    a2,  a3, 0                       /* a2 = old owner's save area */