Add cache line ordering that prevents hardware prefetching, fix cachepc counts read

4 files changed, 30 insertions, 30 deletions

diff --git a/kmod/cache_types.h b/kmod/cache_types.h
index 33da39b..b337d55 100755
--- a/kmod/cache_types.h
+++ b/kmod/cache_types.h
@@ -22,9 +22,8 @@
 #define IS_LAST(flags) GET_BIT(flags, 1)
 #define IS_CACHE_GROUP_INIT(flags) GET_BIT(flags, 2)
 
-// Offset of the next and prev field in the cacheline struct
-#define CL_NEXT_OFFSET 0
-#define CL_PREV_OFFSET 8
+#define CL_NEXT_OFFSET offsetof(struct cacheline, next)
+#define CL_PREV_OFFSET offsetof(struct cacheline, prev)
 
 typedef enum cache_level cache_level;
 typedef enum addressing_type addressing_type;
@@ -53,12 +52,15 @@ struct cacheline {
     cacheline *next;
     cacheline *prev;
 
-    uint16_t cache_set;
-    uint16_t flags;
+    uint32_t cache_set;
+    uint32_t cache_line;
+    uint32_t flags;
 
     // Unused padding to fill cache line
     uint64_t count;
-    char padding[32];
+
+    char padding[24];
 };
 
 static_assert(sizeof(struct cacheline) == CACHELINE_SIZE, "Bad cache line struct size");
+static_assert(CL_NEXT_OFFSET == 0 && CL_PREV_OFFSET == 8);
diff --git a/kmod/cachepc.c b/kmod/cachepc.c
index ed36ac7..9c896f3 100755
--- a/kmod/cachepc.c
+++ b/kmod/cachepc.c
@@ -46,8 +46,6 @@ cache_ctx *
 cachepc_get_ctx(cache_level cache_level)
 {
 	cache_ctx *ctx;
-
-	// printk(KERN_WARNING "CachePC: Getting ctx..\n");
        
 	ctx = kzalloc(sizeof(cache_ctx), GFP_KERNEL);
 	BUG_ON(ctx == NULL);
@@ -72,8 +70,6 @@ cachepc_get_ctx(cache_level cache_level)
 	ctx->set_size = CACHELINE_SIZE * ctx->associativity;
 	ctx->cache_size = ctx->sets * ctx->set_size;
 
-	// printk(KERN_WARNING "CachePC: Getting ctx done\n");
-
 	return ctx;
 }
 
@@ -395,6 +391,7 @@ allocate_cache_ds(cache_ctx *ctx)
 	for (i = 0; i < ctx->nr_of_cachelines; ++i) {
 		cl_ptr_arr[i] = cl_arr + i;
 		cl_ptr_arr[i]->cache_set = get_virt_cache_set(ctx, cl_ptr_arr[i]);
+		cl_ptr_arr[i]->cache_line = i / ctx->sets;
 	}
 
 	return cl_ptr_arr;
diff --git a/kmod/cachepc.h b/kmod/cachepc.h
index 5deb712..8a9521c 100755
--- a/kmod/cachepc.h
+++ b/kmod/cachepc.h
@@ -49,13 +49,12 @@ cachepc_prime(cacheline *head)
 {
     cacheline *curr_cl;
 
-    cachepc_cpuid();
+    //cachepc_cpuid();
     curr_cl = head;
     do {
         curr_cl = curr_cl->next;
-        cachepc_mfence();
     } while(curr_cl != head);
-    cachepc_cpuid();
+    //cachepc_cpuid();
 
     return curr_cl->prev;
 }
@@ -79,13 +78,12 @@ cachepc_prime_rev(cacheline *head)
 {
     cacheline *curr_cl;
 
-    cachepc_cpuid();
+    //cachepc_cpuid();
     curr_cl = head;
     do {
         curr_cl = curr_cl->prev;
-        cachepc_mfence();
     } while(curr_cl != head);
-    cachepc_cpuid();
+    //cachepc_cpuid();
 
     return curr_cl->prev;
 }
@@ -96,16 +94,14 @@ cachepc_probe(cacheline *start_cl)
 	uint64_t pre, post;
 	cacheline *next_cl;
 	cacheline *curr_cl;
-	volatile register uint64_t i asm("r12");
 
 	curr_cl = start_cl;
 
 	do {
 		pre = cachepc_read_pmc(0);
-		pre += cachepc_read_pmc(1);
 
 		cachepc_mfence();
-		cachepc_cpuid();
+		//cachepc_cpuid();
 		
 		asm volatile(
 			"mov 8(%[curr_cl]), %%rax \n\t"              // +8
@@ -123,13 +119,12 @@ cachepc_probe(cacheline *start_cl)
 		);
 
 		cachepc_mfence();
-		cachepc_cpuid();
+		//cachepc_cpuid();
 
 		post = cachepc_read_pmc(0);
-		post += cachepc_read_pmc(1);
 
 		cachepc_mfence();
-		cachepc_cpuid();
+		//cachepc_cpuid();
 
 		/* works across size boundary */
 		curr_cl->count = post - pre;
diff --git a/kmod/util.c b/kmod/util.c
index 166ec3e..1932c9b 100755
--- a/kmod/util.c
+++ b/kmod/util.c
@@ -20,16 +20,22 @@ prng_bytes(uint8_t *dst, size_t size)
 void
 random_perm(uint32_t *arr, uint32_t arr_len)
 {
-	uint32_t i, idx, tmp;
+	uint32_t i, mid; // idx, tmp;
 
-	for (i = arr_len - 1; i > 0; --i) {
-		prng_bytes((void*)&idx, 4);
-		idx = idx % i;
-
-		tmp = arr[idx];
-		arr[idx] = arr[i];
-		arr[i] = tmp;
+	/* defeat stream prefetcher by preventing access direction */
+	mid = arr_len / 2;
+	for (i = 0; i < arr_len; i++) {
+		arr[i] = mid + (i % 2 ? -1 : 1) * ((i + 1) / 2);
 	}
+
+	// for (i = arr_len - 1; i > 0; --i) {
+	// 	prng_bytes((void*)&idx, 4);
+	// 	idx = idx % i;
+
+	// 	tmp = arr[idx];
+	// 	arr[idx] = arr[i];
+	// 	arr[i] = tmp;
+	// }
 }
 
 void