From fe6cc1c675f0ae219fe035dd9d0cc86ae6aebf46 Mon Sep 17 00:00:00 2001
From: Andreas Abel <abel@cs.uni-saarland.de>
Date: Wed, 6 Mar 2019 17:12:10 +0100
Subject: [PATCH] restore OF flag

---
 common/nanoBench.c | 16 +++++++++++++---
 kernel/nb_km.c     | 30 ++++++++++++++++++++++--------
 2 files changed, 35 insertions(+), 11 deletions(-)

diff --git a/common/nanoBench.c b/common/nanoBench.c
index 5b3101f..97264a7 100644
--- a/common/nanoBench.c
+++ b/common/nanoBench.c
@@ -515,7 +515,8 @@ void measurement_template_Intel() {
         ".intel_syntax noprefix                  \n"
         "push rax                                \n"
         "lahf                                    \n"
-        "push rax                                \n"
+        "seto al                                 \n"
+        "push rax                                \n"        
         "push rcx                                \n"
         "push rdx                                \n"
         "push r15                                \n"
@@ -545,7 +546,8 @@ void measurement_template_Intel() {
         "pop rdx; lfence                         \n"
         "pop rcx; lfence                         \n"
         "pop rax; lfence                         \n"
-        "sahf; lfence                            \n"
+        "cmp al, -127; lfence                    \n"
+        "sahf; lfence                            \n"        
         "pop rax;                                \n"
         "lfence                                  \n"
         ".att_syntax noprefix                    ");
@@ -642,6 +644,7 @@ void measurement_template_AMD() {
         ".intel_syntax noprefix                  \n"
         "push rax                                \n"
         "lahf                                    \n"
+        "seto al                                 \n"
         "push rax                                \n"
         "push rcx                                \n"
         "push rdx                                \n"
@@ -682,6 +685,7 @@ void measurement_template_AMD() {
         "pop rdx; lfence                         \n"
         "pop rcx; lfence                         \n"
         "pop rax; lfence                         \n"
+        "cmp al, -127; lfence                    \n"
         "sahf; lfence                            \n"
         "pop rax;                                \n"
         "lfence                                  \n"
@@ -807,6 +811,7 @@ void measurement_FF_template_Intel() {
         ".intel_syntax noprefix                  \n"
         "push rax                                \n"
         "lahf                                    \n"
+        "seto al                                 \n"
         "push rax                                \n"
         "push rcx                                \n"
         "push rdx                                \n"
@@ -836,6 +841,7 @@ void measurement_FF_template_Intel() {
         "pop rdx; lfence                         \n"
         "pop rcx; lfence                         \n"
         "pop rax; lfence                         \n"
+        "cmp al, -127; lfence                    \n"
         "sahf; lfence                            \n"
         "pop rax;                                \n"
         "lfence                                  \n"
@@ -927,8 +933,9 @@ void measurement_FF_template_AMD() {
     asm(".quad "STRINGIFY(MAGIC_BYTES_INIT));
     asm volatile(
         ".intel_syntax noprefix                  \n"
-        "push rax                                \n"
+        "push rax                                \n"        
         "lahf                                    \n"
+        "seto al                                 \n"
         "push rax                                \n"
         "push rdx                                \n"
         "push r15                                \n"
@@ -951,6 +958,7 @@ void measurement_FF_template_AMD() {
         "pop r15; lfence                         \n"
         "pop rdx; lfence                         \n"
         "pop rax; lfence                         \n"
+        "cmp al, -127; lfence                    \n"
         "sahf; lfence                            \n"
         "pop rax;                                \n"
         "lfence                                  \n"
@@ -1031,6 +1039,7 @@ void measurement_RDTSC_template() {
         ".intel_syntax noprefix                  \n"
         "push rax                                \n"
         "lahf                                    \n"
+        "seto al                                 \n"
         "push rax                                \n"
         "push rdx                                \n"
         "push r15                                \n"
@@ -1043,6 +1052,7 @@ void measurement_RDTSC_template() {
         "pop r15; lfence                         \n"
         "pop rdx; lfence                         \n"
         "pop rax; lfence                         \n"
+        "cmp al, -127; lfence                    \n"
         "sahf; lfence                            \n"
         "pop rax;                                \n"
         "lfence                                  \n"
diff --git a/kernel/nb_km.c b/kernel/nb_km.c
index ceb9b10..8f8069d 100644
--- a/kernel/nb_km.c
+++ b/kernel/nb_km.c
@@ -63,11 +63,15 @@ static ssize_t unroll_count_show(struct kobject *kobj, struct kobj_attribute *at
     return sprintf(buf, "%ld\n", unroll_count);
 }
 static ssize_t unroll_count_store(struct kobject *kobj, struct kobj_attribute *attr, const char *buf, size_t count) {
+    long old_unroll_count = unroll_count;
     sscanf(buf, "%ld", &unroll_count);
-    vfree(runtime_code);
-    runtime_code = __vmalloc(PAGE_SIZE + (unroll_count)*PAGE_SIZE*2 + 10000, GFP_KERNEL, PAGE_KERNEL_EXEC);
-    if (!runtime_code) {
-        pr_debug("failed to allocate executable memory\n");
+
+    if (old_unroll_count != unroll_count) {
+        vfree(runtime_code);
+        runtime_code = __vmalloc(PAGE_SIZE + (unroll_count)*PAGE_SIZE*2 + 10000, GFP_KERNEL, PAGE_KERNEL_EXEC);
+        if (!runtime_code) {
+            pr_debug("failed to allocate executable memory\n");
+        }
     }
     return count;
 }
@@ -86,10 +90,18 @@ static ssize_t n_measurements_show(struct kobject *kobj, struct kobj_attribute *
     return sprintf(buf, "%ld\n", n_measurements);
 }
 static ssize_t n_measurements_store(struct kobject *kobj, struct kobj_attribute *attr, const char *buf, size_t count) {
+    long old_n_measurements = n_measurements;
     sscanf(buf, "%ld", &n_measurements);
-    for (int i=0; i<MAX_PROGRAMMABLE_COUNTERS; i++) {
-        kfree(measurement_results[i]);
-        measurement_results[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);
+    
+    if (old_n_measurements != n_measurements) {
+        for (int i=0; i<MAX_PROGRAMMABLE_COUNTERS; i++) {
+            kfree(measurement_results[i]);
+            kfree(measurement_results_base[i]);
+            measurement_results[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);
+            measurement_results_base[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);
+            memset(measurement_results[i], 0, n_measurements*sizeof(int64_t));
+            memset(measurement_results_base[i], 0, n_measurements*sizeof(int64_t));
+        }
     }
     return count;
 }
@@ -322,12 +334,14 @@ static int __init nb_init (void) {
     }
 
     for (int i=0; i<MAX_PROGRAMMABLE_COUNTERS; i++) {
-        measurement_results[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);
+        measurement_results[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);        
         measurement_results_base[i] = kmalloc(n_measurements*sizeof(int64_t), GFP_KERNEL);
         if(!measurement_results[i] || !measurement_results_base[i]){
             printk(KERN_ERR "Could not allocate memory for measurement_results\n");
             return -1;
         }
+        memset(measurement_results[i], 0, n_measurements*sizeof(int64_t));
+        memset(measurement_results_base[i], 0, n_measurements*sizeof(int64_t));
     }
 
     runtime_mem = kmalloc(2*1024*1024, GFP_KERNEL);