How to use the ARM PMU in GEM5? - arm

Question

I had a problem initializing the PMU in gem5 for an arm full system with the starter_fs.py in --cpu hpi.
i followed the instructions of this post Using perf_event with the ARM PMU inside gem5 and i managed to solve my problem. I added the patch and configure the system. I am not using perf. I try to access directly the registers and read them. As i see GEM5 has only some register events implemented. Can we add the others as well as :
for example EXC_TAKEN is not implemented. Is the following the way to add them?
self.addEvent(ProbeEvent(self,0x09, cpu, "EXC_TAKEN"))
#0x09: EXC_TAKEN ???
Also, reading the pmu event registers i manage to read them and extract the events but the pmccntr cycle register always returns zero? How gem5 increments this register? What are the steps to read the cycle reggister?
a code that i use to read using perf is the following:
#include <stdio.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <unistd.h>
#include <sys/syscall.h>
#include <linux/perf_event.h>
#define NUM_NODES 100
#define NONE 9999
struct _NODE
{
int iDist;
int iPrev;
};
typedef struct _NODE NODE;
struct _QITEM
{
int iNode;
int iDist;
int iPrev;
struct _QITEM *qNext;
};
typedef struct _QITEM QITEM;
QITEM *qHead = NULL;
int AdjMatrix[NUM_NODES][NUM_NODES];
int g_qCount = 0;
NODE rgnNodes[NUM_NODES];
int ch;
int iPrev, iNode;
int i, iCost, iDist;
void print_path (NODE *rgnNodes, int chNode)
{
if (rgnNodes[chNode].iPrev != NONE)
{
//print_path(rgnNodes, rgnNodes[chNode].iPrev);
}
//printf (" %d", chNode);
fflush(stdout);
}
void enqueue (int iNode, int iDist, int iPrev)
{
QITEM *qNew = (QITEM *) malloc(sizeof(QITEM));
QITEM *qLast = qHead;
if (!qNew)
{
//fprintf(stderr, "Out of memory.\n");
exit(1);
}
qNew->iNode = iNode;
qNew->iDist = iDist;
qNew->iPrev = iPrev;
qNew->qNext = NULL;
if (!qLast)
{
qHead = qNew;
}
else
{
while (qLast->qNext) qLast = qLast->qNext;
qLast->qNext = qNew;
}
g_qCount++;
// ASSERT(g_qCount);
}
void dequeue (int *piNode, int *piDist, int *piPrev)
{
QITEM *qKill = qHead;
if (qHead)
{
// ASSERT(g_qCount);
*piNode = qHead->iNode;
*piDist = qHead->iDist;
*piPrev = qHead->iPrev;
qHead = qHead->qNext;
free(qKill);
g_qCount--;
}
}
int qcount (void)
{
return(g_qCount);
}
int dijkstra(int chStart, int chEnd)
{
for (ch = 0; ch < NUM_NODES; ch++)
{
rgnNodes[ch].iDist = NONE;
rgnNodes[ch].iPrev = NONE;
}
if (chStart == chEnd)
{
//printf("Shortest path is 0 in cost. Just stay where you are.\n");
}
else
{
rgnNodes[chStart].iDist = 0;
rgnNodes[chStart].iPrev = NONE;
enqueue (chStart, 0, NONE);
while (qcount() > 0)
{
dequeue (&iNode, &iDist, &iPrev);
for (i = 0; i < NUM_NODES; i++)
{
if ((iCost = AdjMatrix[iNode][i]) != NONE)
{
if ((NONE == rgnNodes[i].iDist) ||
(rgnNodes[i].iDist > (iCost + iDist)))
{
rgnNodes[i].iDist = iDist + iCost;
rgnNodes[i].iPrev = iNode;
enqueue (i, iDist + iCost, iNode);
}
}
}
}
//printf("Shortest path is %d in cost. ", rgnNodes[chEnd].iDist);
//printf("Path is: ");
//print_path(rgnNodes, chEnd);
//printf("\n");
}
}
int main(int argc, char *argv[]) {
int diff = 0;
uint64_t num_cycles_nominal=0;
uint64_t num_cycles_attack=0;
uint64_t counter_cpu_cycles = 0;
//system("./load-module");
int i,j,k;
FILE *fp;
static int perf_fd_cpu_cycles;
static struct perf_event_attr attr_cpu_cycles;
attr_cpu_cycles.size = sizeof(attr_cpu_cycles);
attr_cpu_cycles.exclude_kernel = 1;
attr_cpu_cycles.exclude_hv = 1;
attr_cpu_cycles.exclude_callchain_kernel = 1;
attr_cpu_cycles.type = PERF_TYPE_RAW;
attr_cpu_cycles.config = 0x11;
/* Open the file descriptor corresponding to this counter. The counter
should start at this moment. */
if ((perf_fd_cpu_cycles = syscall(__NR_perf_event_open, &attr_cpu_cycles, 0, -1, -1, 0)) == -1)
fprintf(stderr, "perf_event_open fail %d %d: %s\n", perf_fd_cpu_cycles, errno, strerror(errno));
if (argc<2) {
//fprintf(stderr, "Usage: dijkstra <filename>\n");
//fprintf(stderr, "Only supports matrix size is #define'd.\n");
}
/* open the adjacency matrix file */
fp = fopen (argv[1],"r");
/* make a fully connected matrix */
for (i=0;i<NUM_NODES;i++) {
for (j=0;j<NUM_NODES;j++) {
/* make it more sparce */
fscanf(fp,"%d",&k);
AdjMatrix[i][j]= k;
}
}
/* Get and close the performance counters. */
read(perf_fd_cpu_cycles, &counter_cpu_cycles, sizeof(counter_cpu_cycles));
//close(perf_fd_cpu_cycles);
printf("Number of cpu_cycles before: %d\n", counter_cpu_cycles);
num_cycles_nominal = counter_cpu_cycles;
/* Get and close the performance counters. */
read(perf_fd_cpu_cycles, &counter_cpu_cycles, sizeof(counter_cpu_cycles));
//close(perf_fd_cpu_cycles);
printf("Number of cpu_cycles after attack: %d\n", counter_cpu_cycles);
num_cycles_attack = counter_cpu_cycles - num_cycles_nominal;
/* finds 10 shortest paths between nodes */
for (i=0,j=NUM_NODES/2;i<100;i++,j++) {
j=j%NUM_NODES;
dijkstra(i,j);
}
read(perf_fd_cpu_cycles, &counter_cpu_cycles, sizeof(counter_cpu_cycles));
close(perf_fd_cpu_cycles);
printf("Number of cpu_cycles end: %d\n", counter_cpu_cycles);
num_cycles_nominal = counter_cpu_cycles - num_cycles_attack;
printf("Number of cpu_cycles nominal: %d\n", num_cycles_nominal);
printf("Number of cpu_cycles attack: %d\n", num_cycles_attack);
exit(0);
}
the problem is that i can read the branch misses with perf having 0x10 instead 0f 0x11 (cycle counters RAW EVENT in GEM5) but using 0x11 for reading the cycles i get zero. When i try to reverse engineer the increment of cycle counter i do the following comments:
when simple/atomic or simple/timing i see that updateCycleCounter is called from the base.hh, also for the 03 cpu model. When HPI and considering that hpi is a MinorCPU model i see that updateCycleCounter is called only in POWER_STATE_ON, but i didnt find in the code a POWER_STATE_ON reference updateCycleCounter(CPU_STATE_ON) which will update the cycle counter. Please help me verify this assumption.
*****The problem was that in the MinorCPU the updateCycleCounter wasnt called for the CPU_STATE_ON which updates the ActiveCycles. It was fixed by the following patch https://gem5-review.googlesource.com/c/public/gem5/+/38095 .