Add paravirt support for __vdso_gettimeofday().

Fixed two bugs in tests/time that cancelled each other out:
- loop should take 1 cycle, not 2
- though it was on split byteQ lines, and did take 2,
depending on surrounding code.

Removed the manual envp walk to get the VDSO base.

TESTED=integration

Change-Id: Ia2d66f3be222b1f11f5339cfc093c02a1110fd73

scripts/run_tests.py

@@ -54,7 +54,7 @@ def setUp(self):
 
     def tearDown(self):
         if self.clean_run_dir:
-            shutil.rmtree(self.xio.GetRunDir())
+            shutil.rmtree(self.run_dir)
         if self.background_bmk:
             self.background_bmk.kill()
 
@@ -504,6 +504,7 @@ def setDriverParams(self):
         self.xio.AddBmks(bmk_cfg)
         self.xio.AddPinOptions()
         self.xio.AddPintoolOptions(num_cores=1)
+        self.xio.AddROIOptions()
         self.xio.AddZestoOptions(os.path.join(self.xio.GetTreeDir(),
                                               "xiosim/config", "N.cfg"))
 
@@ -512,9 +513,36 @@ def setUp(self):
         self.expected_vals.append((xs.PerfStatRE("all_insn"), 3000000.0))
 
         # Set up expected output from the simulated program.
-        # 2M cycles / 3.2GHz = 625us.
+        # 1M cycles / 3.2GHz = 317us.
+        elapsed_re = "Elapsed: (%s) sec" % xs.DECIMAL_RE
+        self.bmk_expected_vals.append((elapsed_re, 0.000317))
+
+    def runTest(self):
+        self.runAndValidate()
+
+class TimeVDSOTest(XIOSimTest):
+    ''' End-to-end test for __vdso_gettimeofday(). '''
+    def setDriverParams(self):
+        bmk_cfg = self.writeTestBmkConfig("time")
+        self.xio.AddBmks(bmk_cfg)
+        self.xio.AddPinOptions()
+        self.xio.AddPintoolOptions(num_cores=1)
+        self.xio.AddROIOptions()
+        self.xio.AddTraceFile("trace.out")
+        self.xio.AddZestoOptions(os.path.join(self.xio.GetTreeDir(),
+                                              "xiosim/config", "N.cfg"))
+
+    def setUp(self):
+        super(TimeVDSOTest, self).setUp()
+        if self.xio.TARGET_ARCH == "piii":
+            self.tearDown()
+            self.skipTest("VDSO not used for gettimeofday() on i686")
+        self.expected_vals.append((xs.PerfStatRE("all_insn"), 3000000.0))
+
+        # Set up expected output from the simulated program.
+        # 1M cycles / 3.2GHz = 317us.
         elapsed_re = "Elapsed: (%s) sec" % xs.DECIMAL_RE
-        self.bmk_expected_vals.append((elapsed_re, 0.000625))
+        self.bmk_expected_vals.append((elapsed_re, 0.000317))
 
     def runTest(self):
         self.runAndValidate()

scripts/xiosim_driver.py

@@ -78,6 +78,9 @@ def AddReplaceOptions(self, func):
     def AddTraceFile(self, file):
         self.cmd += "-trace %s " % file
 
+    def DisableSpeculation(self):
+        self.cmd += "-speculation false "
+
     def AddZestoOptions(self, cfg):
         if self.bridge_dirs:
             self.cmd += "-buffer_bridge_dirs %s " % self.bridge_dirs

tests/Makefile

@@ -24,7 +24,7 @@ piii/prefetch: prefetch.cpp
 piii/ignore: ignore.cpp
 	g++ -m32 -std=c++1y -static -O0 -o $@ $<
 piii/time: time.cpp
-	g++ -m32 -O1 -static -o $@ $<
+	g++ -m32 -O1 -falign-functions=64 -static -o $@ $<
 piii/loop: loop.cpp
 	g++ -m32 -O0 -static -o $@ $<
 piii/rdtsc: rdtsc.cpp
@@ -49,7 +49,9 @@ k8/prefetch: prefetch.cpp
 k8/ignore: ignore.cpp
 	g++ -std=c++1y -static -O0 -o $@ $<
 k8/time: time.cpp
-	g++ -O1 -static -o $@ $<
+	g++ -O1 -falign-functions=64 -static -o $@ $<
+k8/time_vdso: time.cpp
+	g++ -O1 -falign-functions=64 -o $@ $<
 k8/loop: loop.cpp
 	g++ -O0 -static -o $@ $<
 k8/rdtsc: rdtsc.cpp

tests/time.cpp

@@ -2,14 +2,19 @@
  *
  * Simulator should return simulated time instead of host time.
  *
- * Compile: g++ -O1 -m32 -static -o time time.c
  */
 
 #include <sys/time.h>
 #include <sys/times.h>
 #include <stdio.h>
 #include <unistd.h>
 
+extern "C" void xiosim_roi_begin() __attribute__ ((noinline));
+extern "C" void xiosim_roi_end() __attribute__ ((noinline));
+
+void xiosim_roi_begin() { __asm__ __volatile__ ("":::"memory"); }
+void xiosim_roi_end() { __asm__ __volatile__ ("":::"memory"); }
+
 const int ITER = 1000000;
 
 double times_test() {
@@ -30,34 +35,39 @@ double times_test() {
   return (end.tms_utime - start.tms_utime);
 }
 
+unsigned loop() {
+  unsigned counter = 0;
+  __asm__ __volatile ("1: addl $1, %0;"
+                      "cmpl %2, %1;"
+                      "jb 1b;"
+                      : "=a"(counter)
+                      : "0"(counter), "i"(ITER)
+                      : "memory");
+  return counter;
+}
+
 double gettimeofday_test() {
   struct timeval start, end;
   gettimeofday(&start, NULL);
-  printf("Start time of day: %d s %d us\n", start.tv_sec, start.tv_usec);
 
-  unsigned int counter = 0;
-  __asm__ __volatile ("loop: addl $1, %0;"
-                      "cmpl %1, %0;"
-                      "jb loop;"
-                      : "=a"(counter)
-                      : "r"(ITER)
-                      : "memory");
+  unsigned counter = loop();
 
   gettimeofday(&end, NULL);
 
   /* 1M iterations of a 3 instruction loop -> 3M instructions.
-   * With pipelining we get 2 cycles per iteration for a total of 2M cycles.
+   * With pipelining we get 1 cycles per iteration for a total of 1M cycles.
    * Default NHM config is 3.2GHz, so we should expect a time difference
-   * of around 2M / 3.2GHz = 625us.
+   * of around 1M / 3.2GHz = 317us.
    */
   printf("Counter value = %d\n", counter);
-  printf("End time of day: %d s %d us\n", end.tv_sec, end.tv_usec);
   return (end.tv_sec - start.tv_sec) + 1.0*(end.tv_usec - start.tv_usec)/(1e6);
 }
 
 int main() {
   // For now, just test gettimeofday() because we haven't completed times().
+  xiosim_roi_begin();
   double elapsed = gettimeofday_test();
+  xiosim_roi_end();
   printf("Elapsed: %3.8f sec\n", elapsed);
   return 0;
 }

xiosim/pintool/BUILD

@@ -39,6 +39,8 @@ cc_binary(
         "ignore_ins.h",
         "ildjit.cpp",
         "ildjit.h",
+        "paravirt.cpp",
+        "paravirt.h",
         "replace_function.cpp",
         "replace_function.h",
         "roi.cpp",
@@ -51,6 +53,8 @@ cc_binary(
         "syscall_handling.h",
         "utils.cpp",
         "utils.h",
+        "vdso.cpp",
+        "vdso.h",
     ],
     linkopts = [
         "-Wl,-Bsymbolic",

xiosim/pintool/feeder.h

@@ -24,7 +24,6 @@ class handshake_container_t;
 
 extern KNOB<BOOL> KnobILDJIT;
 extern KNOB<int> KnobNumCores;
-extern KNOB<BOOL> KnobTimingVirtualization;
 
 /* A list of the threads in this feeder. */
 extern list<THREADID> thread_list;
@@ -45,6 +44,9 @@ extern int asid;
 /* Xed machine mode state for when we need to encode/decode things. */
 extern xed_state_t dstate;
 
+/* Host TSC values for timing virtualization. */
+extern tick_t* initial_timestamps;
+
 #define ATOMIC_ITERATE(_list, _it, _lock)                                                          \
     for (lk_lock(&(_lock), 1), (_it) = (_list).begin(), lk_unlock(&(_lock)); [&] {                 \
         lk_lock(&(_lock), 1);                                                                      \
@@ -214,6 +216,11 @@ VOID ScheduleThread(THREADID tid);
  * Check out base_allocator.h for the API. */
 int AllocateCores(std::vector<double> scaling, double serial_runtime);
 
+/* Helper to check if producer thread @tid will grab instruction at @pc.
+ * If return value is false, we can skip instrumentaion. Can hog execution
+ * if producers are disabled by producer_sleep. */
+BOOL CheckIgnoreConditions(THREADID tid, ADDRINT pc);
+
 /* Insert instrumentation that we didn't add so we can skip ILDJIT compilation even faster. */
 VOID doLateILDJITInstrumentation();