X-Git-Url: https://git.liburcu.org/?a=blobdiff_plain;f=markers-test%2Ftest-mark-speed-local.c;h=1940cdc36e3a1f8b05ee2de227acbc00aafa1b8a;hb=fade4071bfc7dcaebaad3e7f502c62a152f659a9;hp=76c219b12e59e55f2a9f539d89bc02fcb8f2296d;hpb=9f92c61e0644705f85b810d6586eb4aa8c46b7c1;p=lttv.git

diff --git a/markers-test/test-mark-speed-local.c b/markers-test/test-mark-speed-local.c
index 76c219b1..1940cdc3 100644
--- a/markers-test/test-mark-speed-local.c
+++ b/markers-test/test-mark-speed-local.c
@@ -10,6 +10,29 @@
 #include <asm/ptrace.h>
 #include <asm/system.h>
 
+static void pmc_flush_cache(void)
+  {
+    register int i;
+    /* write back and invalidate cache (a serializing instruction) */
+
+    __asm__ __volatile__ ( "wbinvd" : : : "memory" );
+
+    /* The wbinvd instruction does not wait for the external caches
+     * to be flushed, but only requests that it be done.  The loop
+     * is to be sure that enough time has elapsed, but the compiler
+     * might simplify or even remove it.  The loop bound is for a
+     * 512 KB L2 cache.  On a Pentium Pro/II/III, the loop uses
+     * 2 cycles per iteration.
+     *
+     * Does wbinvd also cause the TLB to be flushed?
+     * A comment in mtrr.c suggests that it does.
+     */
+    for (i = 0; i < 512*1024; i++) {
+      cpu_relax();
+    }
+  }
+
+
 static void noinline test2(const struct marker *mdata,
         void *call_private, ...)
 {
@@ -59,22 +82,24 @@ static void noinline test2(const struct marker *mdata,
 		}							\
 	} while (0)
 
-	//asm volatile ("");
 struct proc_dir_entry *pentry = NULL;
 
-volatile int temp = 10;
-volatile int temp2 = 10;
-
 static inline void test(unsigned long arg, unsigned long arg2)
 {
+	volatile int temp[5];
 #ifdef CACHEFLUSH
-	wbinvd();
+	clflush(&current->pid);
+	//pmc_flush_cache();
 #endif
-	temp = temp * 100 + 60;
-	temp2 = temp << 10;
+	temp[2] = (temp[0] + 60) << 10;
+	temp[3] = (temp[2] + 60) << 10;
+	temp[4] = (temp[3] + 60) << 10;
+	temp[0] = (temp[4] + 60) << 10;
 	//asm volatile ("");
 	//__my_trace_mark(1, kernel_debug_test, NULL, "%d %d %ld %ld", 2, current->pid, arg, arg2);
+	barrier();
 	test2(NULL, NULL, 2, 10, arg, arg2);
+	barrier();	// make sure the compiler does not optimize marker conditions away!
 	//__my_trace_mark(0, kernel_debug_test, NULL, "%d %d %ld %ld", 2, current->pid, arg, arg2);
 }
 
@@ -85,10 +110,14 @@ static int my_open(struct inode *inode, struct file *file)
 	unsigned long flags;
 
 	local_irq_save(flags);
+#ifdef CACHEFLUSH
+	//pmc_flush_cache();	/* initial write back, without cycle count */
+	//msleep(20);	/* wait for L2 flush */
+#endif
 	rdtsc_barrier();
 	cycles1 = get_cycles();
 	rdtsc_barrier();
-	for(i=0; i<200; i++) {
+	for(i=0; i<2000; i++) {
 		test(i, i);
 		test(i, i);
 		test(i, i);