Merge branch 'svm' of git.lfbs.rwth-aachen.de:metalsvm into svm

apps/tests.c

@@ -114,6 +114,8 @@ static int mail_noise(void*arg) {
 	return 0;
 }
 
+/* N has to be multiple of UEs */
+
 //#define N	1024
 //#define N 	512
 #define N	128
@@ -204,9 +206,9 @@ static int svm_test(void *arg)
 	svm_flags |= SVM_L2;
 #endif
 
-	A[0] = (int*) svmmalloc(3*N*N*sizeof(int), svm_flags);
+	A[0] = (int*) svm_malloc(3*N*N*sizeof(int), svm_flags);
 
-	if (!my_ue)
+	if (!my_ue) 
 		memset((void*) A[0], 0x00, 3*N*N*sizeof(int));
 
 	// initialize matrices
@@ -215,6 +217,7 @@ static int svm_test(void *arg)
 		B[i] = A[0] + (i*N + N*N);
 		C[i] = A[0] + (i*N + 2*N*N);
 	}
+	
 	if (!my_ue) {
 		for(i=0; i<N; i++) {
 			A[i][i] = 1;
@@ -223,9 +226,8 @@ static int svm_test(void *arg)
 		}
 	}
 
-	svm_flush(0);
+	svm_barrier(svm_flags);
 
-	RCCE_barrier(&RCCE_COMM_WORLD);
 	kputs("Start parallel calculation...\n");
 
 	start = rdtsc();
@@ -245,9 +247,8 @@ static int svm_test(void *arg)
 		}
 	}
 
-	svm_flush(0);
-
-	RCCE_barrier(&RCCE_COMM_WORLD);
+	svm_barrier(svm_flags);
+	
 	end = rdtsc();
 
 	kputs("Check results...\n");
@@ -270,7 +271,7 @@ static int svm_test(void *arg)
 
 	kprintf("Calculation time (par): %llu ms (%llu ticks)\n", (end-start)/(1000ULL*get_cpu_frequency()), end-start);
 
-	svmfree((void*) A[0], 3*N*sizeof(int));
+	svm_free((void*) A[0], 3*N*sizeof(int));
 
 	svm_statistics();
 
@@ -339,7 +340,7 @@ int test_init(void)
 	//create_kernel_task(NULL, mail_noise, NULL, NORMAL_PRIO);
 	create_kernel_task(NULL, svm_test, NULL, NORMAL_PRIO);
 	//create_kernel_task(NULL, pi, NULL, NORMAL_PRIO);
-	create_kernel_task(NULL, laplace, NULL, NORMAL_PRIO);
+	//create_kernel_task(NULL, laplace, NULL, NORMAL_PRIO);
 	//create_user_task(NULL, "/bin/hello", argv);
 	//create_user_task(NULL, "/bin/tests", argv);
 	//create_user_task(NULL, "/bin/jacobi", argv);

arch/x86/include/asm/svm.h

@@ -78,16 +78,20 @@ int svm_access_request(size_t addr);
  */
 int svm_emit_page(size_t addr, int ue);
 
+#ifdef CONFIG_ROCKCREEK
 /* @brief invalidate the cache entries for all SVM regions
  */
+#ifndef SVM_WB
 static inline void svm_invalidate(void)
 {
 	asm volatile ( ".byte 0x0f; .byte 0x0a;\n" ); // CL1FLUSHMB
 }
+#else
+void svm_invalidate(void);
+#endif
 
 /* *brief flushs the cache for all SVM regions
  */
-#ifdef CONFIG_ROCKCREEK
 #ifndef SVM_WB
 static inline void svm_flush(void)
 {

arch/x86/mm/page.c

@@ -784,7 +784,7 @@ int arch_paging_init(void)
 	kprintf("Map configuration registers at 0x%x\n", viraddr);
 
 	// map SCC's message passing buffers
-	viraddr = map_region(MPB_X0_Y0, MPB_X0_Y0, (MPB_OWN-MPB_X0_Y0+64*1024*1024) >> PAGE_SHIFT, MAP_KERNEL_SPACE|MAP_MPE);
+	viraddr = map_region(MPB_X0_Y0, MPB_X0_Y0, (MPB_OWN-MPB_X0_Y0+16*1024*1024) >> PAGE_SHIFT, MAP_KERNEL_SPACE|MAP_MPE);
 	kprintf("Map message passing buffers at 0x%x\n", viraddr);
 
 	// map the FPGA registers

arch/x86/mm/svm.c

@@ -209,9 +209,7 @@ int svm_access_request(size_t addr)
 	return ret;
 }
 
-#if 0
 static atomic_int32_t size_counter = ATOMIC_INIT(0);
-#endif
 
 void* svm_malloc(size_t size, uint32_t consistency)
 {
@@ -432,6 +430,20 @@ void svm_flush(void)
  */
 
 #ifdef SVM_WB
+
+void svm_invalidate(void)
+{
+	task_t* task = per_core(current_task);
+
+	if(task->flags & TASK_L2) {
+ 		asm volatile ( ".byte 0x0f; .byte 0x0a;\n" ); // CL1FLUSHMB
+	} else {
+		/* no action needed svm_flush already invalidates cache */
+		return;
+	}
+
+}
+
 void svm_flush( size_t phyaddr )
 {
 	task_t* task = per_core(current_task);
@@ -518,18 +530,19 @@ wrong_addr:
 }
 #endif
 
-int svm_barrier(uint32_t type)
+int svm_barrier(uint32_t flags)
 {
 	int i;
 	RCCE_COMM *comm = &RCCE_COMM_WORLD;
 	static int index = 0;
 
-	if (type == SVM_LAZYRELEASE) {
+	if( flags & SVM_LAZYRELEASE ) {
+		kputs("svm barrier has to flush and invalidate!\n");
 		svm_flush(0);
 		svm_invalidate();
 	}
 
-#if 1
+#if 0
 	// Lubachevsky barrier with flags
 	index = !index;
 	if (incregs[AIREG1].counter > (comm->size - 2)) {