[Cbe-oss-dev] [PATCH 05/28]MARS/base: cache workload module

[Yuji Mano]

This patch caches the currently loaded workload module info inside the kernel so
the workload module only needs to be reloaded from shared storage to MPU storage
if different from the one currently loaded in MPU storage.

This patch adds kernel's custom version of memcmp/memcpy routines to keep code
size to a minimum.

Signed-off-by: Yuji Mano <yuji.mano at am.sony.com>
---
 base/src/mpu/kernel/kernel.c |   76 ++++++++++++++++++++++++++++++++++++-------
 1 file changed, 64 insertions(+), 12 deletions(-)

--- a/base/src/mpu/kernel/kernel.c
+++ b/base/src/mpu/kernel/kernel.c
@@ -37,7 +37,6 @@
 
 #include <stddef.h>
 #include <stdint.h>
-#include <string.h>
 #include <spu_mfcio.h>
 
 #include "config.h"
@@ -70,10 +69,42 @@ static uint64_t workload_ea;
 static struct mars_workload_context schedule_workload;
 static uint16_t schedule_workload_id;
 
-/* module entry */
+/* workload module cached */
+static struct mars_workload_module cached_workload_module = {0, 0, 0, 0};
+
+/* workload module entry */
 typedef void (*module_entry)(
 	const struct mars_kernel_syscalls *kernel_syscalls);
 
+static int kernel_memcmp(const void *s1, const void *s2, int size)
+{
+	unsigned char *ptr_1 = (unsigned char *)s1;
+	unsigned char *ptr_2 = (unsigned char *)s2;
+	unsigned char *ptr_e = (unsigned char *)ptr_1 + size;
+
+	while (ptr_1 < ptr_e) {
+		if (*ptr_1 != *ptr_2)
+			return 1;
+		ptr_1++;
+		ptr_2++;
+	}
+
+	return 0;
+}
+
+static void kernel_memcpy(void *dst, const void *src, int size)
+{
+	unsigned char *ptr_1 = (unsigned char *)dst;
+	unsigned char *ptr_2 = (unsigned char *)src;
+	unsigned char *ptr_e = (unsigned char *)ptr_1 + size;
+
+	while (ptr_1 < ptr_e) {
+		*ptr_1 = *ptr_2;
+		ptr_1++;
+		ptr_2++;
+	}
+}
+
 static void dma_put(void *ls, uint64_t ea, uint32_t size)
 {
 	mfc_put((volatile void *)ls, ea, size, MARS_KERNEL_DMA_TAG, 0, 0);
@@ -629,6 +660,35 @@ static void __attribute__((noinline)) ru
 	((module_entry)workload.module.entry)(&kernel_syscalls);
 }
 
+static void load_workload_module(void)
+{
+	__vector unsigned char *bss_ptr, *bss_end;
+
+	/* 0 the bss section */
+	bss_ptr = (__vector unsigned char *)(MARS_WORKLOAD_MODULE_BASE_ADDR +
+					     workload.module.exec_size);
+	bss_end = (__vector unsigned char *)((void *)bss_ptr +
+					     workload.module.bss_size);
+
+	while (bss_ptr < bss_end)
+		*bss_ptr++ = spu_splats((unsigned char)0);
+
+	/* only reload the workload module if different from cached */
+	if (kernel_memcmp(&cached_workload_module, &workload.module,
+			  sizeof(struct mars_workload_module))) {
+		/* load the exec code into mpu storage from host storage */
+		dma_get((void *)MARS_WORKLOAD_MODULE_BASE_ADDR,
+			workload.module.exec_ea, workload.module.exec_size);
+
+		/* store the current cached workload module ea */
+		kernel_memcpy(&cached_workload_module, &workload.module,
+			      sizeof(struct mars_workload_module));
+	}
+
+	/* sync before executing loaded code */
+	spu_sync();
+}
+
 static int scheduler(void)
 {
 	int status;
@@ -650,16 +710,8 @@ static int scheduler(void)
 	if (status == MARS_WORKLOAD_RESERVED_NONE)
 		return MARS_KERNEL_STATUS_IDLE;
 
-	/* load the exec code into mpu storage from host storage */
-	dma_get((void *)MARS_WORKLOAD_MODULE_BASE_ADDR,
-		workload.module.exec_ea, workload.module.exec_size);
-
-	/* 0 the bss section */
-	memset((void *)MARS_WORKLOAD_MODULE_BASE_ADDR +
-	       workload.module.exec_size, 0, workload.module.bss_size);
-
-	/* sync before executing loaded code */
-	spu_sync();
+	/* load the workload module */
+	load_workload_module();
 
 	/* run workload */
 	run_workload();