add some performance counters and remove bug in the physical to virtual address translation

arch/x86/include/asm/svm.h

@@ -85,6 +85,8 @@ static inline void svm_flush(void)
 #endif
 }
 
+int svm_statistics(void);
+
 #endif
 
 #ifdef __cplusplus

arch/x86/mm/svm.c

@@ -48,6 +48,9 @@ static volatile uint8_t*	page_owner = NULL;
 static size_t phys2virt[SHARED_PAGES] = {[0 ... SHARED_PAGES-1] = 0};
 static size_t shmbegin = 0;
 static int my_ue = 0;
+static uint32_t emit[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
+static uint32_t request[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
+static uint32_t forward[RCCE_MAXNP] = {[0 ... RCCE_MAXNP-1] = 0};
 
 int svm_init(void)
 {
@@ -119,6 +122,7 @@ int svm_access_request(size_t addr)
 	/* send ping request */
 	iRCCE_mail_send(2*sizeof(size_t), ICC_TAG_SVMREQUEST, 0, payload, remote_rank);
 
+	request[remote_rank]++;
         NOP8;
         icc_send_irq(remote_rank);
 
@@ -134,8 +138,7 @@ int svm_access_request(size_t addr)
 
 void* svmmalloc(size_t size)
 {
-	size_t phyaddr;
-	size_t viraddr;
+	size_t phyaddr, viraddr, i;
 	uint32_t flags;
 	uint32_t map_flags = MAP_KERNEL_SPACE|MAP_MPE|MAP_SVM;
 
@@ -158,7 +161,8 @@ void* svmmalloc(size_t size)
 	}
 
 	viraddr = map_region(0, phyaddr, size >> PAGE_SHIFT, map_flags);
-	phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT] = viraddr;
+	for(i=0; i<size; i+=PAGE_SIZE)
+		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = viraddr + i;
 
 	kprintf("svmmalloc: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, viraddr, size);
 
@@ -167,7 +171,7 @@ void* svmmalloc(size_t size)
 
 void svmfree(void* addr, size_t size)
 {
-	size_t phyaddr;
+	size_t phyaddr, i;
 	uint32_t flags;
 
 	if (BUILTIN_EXPECT(!addr || !size, 0))
@@ -181,7 +185,8 @@ void svmfree(void* addr, size_t size)
 	kprintf("svmfree: phyaddr 0x%x, viraddr 0x%x, size 0x%x\n", phyaddr, addr, size);
 
 	unmap_region((size_t) addr, size >> PAGE_SHIFT);
-	phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT] = 0;
+	for(i=0; i<size; i+=PAGE_SIZE)
+		phys2virt[(phyaddr + i - shmbegin) >> PAGE_SHIFT] = 0;
 
 	// iRCCE is not thread save => disable interrupts
 	flags = irq_nested_disable();
@@ -219,8 +224,10 @@ int svm_emit_page(size_t phyaddr, int ue)
 		/* send ping request */
 		iRCCE_mail_send(2*sizeof(size_t), ICC_TAG_SVMREQUEST, 0, payload, remote_rank);
 
-		 NOP8;
+		NOP8;
 		icc_send_irq(remote_rank);
+
+		forward[remote_rank]++;
 	} else {
 		size_t viraddr = phys2virt[(phyaddr - shmbegin) >> PAGE_SHIFT];
 
@@ -230,9 +237,29 @@ int svm_emit_page(size_t phyaddr, int ue)
 		page_owner[pageid] = ue;
 		// need to write to another line to make sure the write combine buffer gets flushed
 		*(int *)RCCE_fool_write_combine_buffer = 1;
+
+		emit[ue]++;
 	}
 
 	return 0;
 }
 
+int svm_statistics(void)
+{
+	uint32_t i;
+
+	kprintf("emit\t:");
+	for(i=0; i<RCCE_MAXNP; i++)
+		kprintf("\t%u", emit[i]);
+	kprintf("\nrequest\t:");
+	for(i=0; i<RCCE_MAXNP; i++) 
+		kprintf("\t%u", request[i]);
+	kprintf("\nforward\t:");
+	for(i=0; i<RCCE_MAXNP; i++)
+		kprintf("\t%u", forward[i]);
+	kputs("\n");
+
+	return 0;
+}
+
 #endif

kernel/tests.c

@@ -131,7 +131,7 @@ static int svm_test(void *arg)
 	my_ue = RCCE_ue();
 	num_ues = RCCE_num_ues();
 
-#if 1
+#if 0
 	if (!my_ue) {
 		// allocate and initialize SVM region
 		A[0] = (int*) kmalloc(3*N*N*sizeof(int));
@@ -192,15 +192,15 @@ static int svm_test(void *arg)
 
 	svm_flush();
 
-	// Now, we need only read access on A and B
-	change_page_permissions((size_t) A[0], (size_t) (A[0]+2*N*N), VMA_CACHEABLE|VMA_READ);
-	RCCE_barrier(&RCCE_COMM_WORLD);
-
 	kputs("Start parallel calculation...\n");
 
 	start = rdtsc();
 	start = rdtsc();
 
+	// Now, we need only read access on A and B
+	change_page_permissions((size_t) A[0], (size_t) (A[0]+2*N*N), VMA_CACHEABLE|VMA_READ);
+	RCCE_barrier(&RCCE_COMM_WORLD);
+
 	// start calculation
 	for(i=my_ue*(N/num_ues); i<(my_ue+1)*(N/num_ues); i++)
 		for(j=0; j<N; j++)
@@ -233,6 +233,8 @@ static int svm_test(void *arg)
 
 	svmfree((void*) A[0], 3*N*sizeof(int));
 
+	svm_statistics();
+
 	return 0;
 }
 #endif