Changes to handle conflict with monitoring: 1. we check if counters are...

Changes to handle conflict with monitoring: 1. we check if counters are already running, and just read them if this is the case. 2. we program them if they are not running. 3. When dcdb stops, they are left running for other agents like EAR.

src/sensors/msr/MSRSensorGroup.cpp

@@ -16,7 +16,6 @@
 #include <stdio.h>
 #include <sys/types.h>
 #include <unistd.h>
-#include <bitset>
 #include <atomic>
 #include <exception>
 #include <utility>
@@ -43,10 +42,7 @@ void MSRSensorGroup::start() {
 		return;
 	}
 
-	if (!program_fixed()) {
-        LOG(error) << "MSR fixed counters already in use! Can not start counters";
-        return;
-    }
+	program_fixed()
 
 	for (auto &kv : cpuToFd) {
 		int cpu = kv.first;
@@ -73,28 +69,10 @@ void MSRSensorGroup::start() {
 }
 
 void MSRSensorGroup::stop() {
-    int temp = _keepRunning;
-
-	_keepRunning = 0;
-	
-	//only close this stuff if we have been properly started before
-    if (temp) {
-        //close file descriptors
-        for (auto &kv: cpuToFd) {
-            close(kv.second);
-            kv.second = -1;
-        }
-
-        //reset counters
-        for (auto& kv : cpuToFd) {
-            //disable counters before writing anything else
-            msr_write(IA32_CR_PERF_GLOBAL_CTRL, 0, kv.first);
-            //reset all bits to 0
-            msr_write(IA32_CR_FIXED_CTR_CTRL, 0, kv.first);
-            //set all global enable bits to 1
-            uint64_t value = (1ULL << 0) + (1ULL << 1) + (1ULL << 2) + (1ULL << 3) + (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
-            msr_write(IA32_CR_PERF_GLOBAL_CTRL, value, kv.first);
-        }
+	//close file descriptors and leave counters running freely
+    for (auto &kv: cpuToFd) {
+    	close(kv.second);
+        kv.second = -1;
     }
 
 	LOG(info) << "Sensorgroup " << _groupName << " stopped.";
@@ -142,8 +120,7 @@ int32_t MSRSensorGroup::msr_write(uint64_t msr_number, uint64_t value, unsigned
  * @return  True if counters programmed successfully, false otherwise, e.g.
  *          because the counters are already in use.
  */
-bool MSRSensorGroup::program_fixed(){
-    unsigned failedCpu = 9999;
+void MSRSensorGroup::program_fixed(){
 
 	for (auto &kv : cpuToFd) {
 		// program core counters
@@ -151,16 +128,16 @@ bool MSRSensorGroup::program_fixed(){
 	    //we do not want to interrupt other services already doing measurements with MSRs
         //therefore check if any fixed counter is currently enabled
 		struct FixedEventControlRegister ctrl_reg;
-        std::bitset<64> cur_reg(ctrl_reg.value);
 
 		msr_read(IA32_CR_FIXED_CTR_CTRL, &ctrl_reg.value, kv.first);
-
-		if (cur_reg[0] || cur_reg[1] || cur_reg[4] || cur_reg[5] || cur_reg[8] || cur_reg[9]) {
-		  //at least one is enabled. Do not overwrite them.
-		  //Instead we revert the already programmed counters and return
-		  failedCpu = kv.first;
-		  break;
+		//are they all enabled?
+		if (ctrl_reg.fields.os0 && ctrl_reg.fields.usr0 && ctrl_reg.fields.os1
+						&& ctrl_reg.fields.usr1 && ctrl_reg.fields.os2
+						&& ctrl_reg.fields.usr2) {
+			//yes! Free running counters were set by someone else => we don't need to program them, just read them.
+			continue;
 		}
+		//not all of them (or none) are enabled => we program them again
 
 		// disable counters while programming
         msr_write(IA32_CR_PERF_GLOBAL_CTRL, 0, kv.first);
@@ -182,32 +159,14 @@ bool MSRSensorGroup::program_fixed(){
 
 		ctrl_reg.fields.reserved1 = 0;
 
+		// program them
 		msr_write(IA32_CR_FIXED_CTR_CTRL, ctrl_reg.value, kv.first);
 
-		// start counting, enable 3 fixed counters
-		//uint64 value = (1ULL << 0) + (1ULL << 1) + (1ULL << 2) + (1ULL << 3) + (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
-		uint64_t value = (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
+		// start counting, enable 3 fixed counters (enable also the programmables counters)
+		uint64 value = (1ULL << 0) + (1ULL << 1) + (1ULL << 2) + (1ULL << 3) + (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
+		//uint64_t value = (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
 		msr_write(IA32_CR_PERF_GLOBAL_CTRL, value, kv.first);
 	}
-
-	//One of the fixed counters was already enabled.
-	//Reset all counters which were already programmed before returning.
-	if (failedCpu != 9999) {
-	  for (auto& kv : cpuToFd) {
-	    if (kv.first != failedCpu) {
-	      //disable counters before writing anything else
-          msr_write(IA32_CR_PERF_GLOBAL_CTRL, 0, kv.first);
-          //reset all bits to 0
-          msr_write(IA32_CR_FIXED_CTR_CTRL, 0, kv.first);
-          //set all global enable bits to 1
-          uint64_t value = (1ULL << 0) + (1ULL << 1) + (1ULL << 2) + (1ULL << 3) + (1ULL << 32) + (1ULL << 33) + (1ULL << 34);
-          msr_write(IA32_CR_PERF_GLOBAL_CTRL, value, kv.first);
-	    } else {
-	      return false;
-	    }
-	  }
-	}
-
 	return true;
 }
 

src/sensors/msr/MSRSensorGroup.h

@@ -35,7 +35,7 @@ private:
 	void read() override;
 	void readAsync() override;
 
-	bool program_fixed();
+	void program_fixed();
 	std::map<unsigned int,int> cpuToFd;
 	int32_t msr_read(uint64_t msr_number, uint64_t * value, unsigned int cpu);
 	int32_t msr_write(uint64_t msr_number, uint64_t value, unsigned int cpu);