Some hacking on the sampling profiler

d164ddab · Silas Boyd-Wickizer · fbecbf0c · d164ddab · d164ddab · d164ddab
--- a/Makefile
+++ b/Makefile
@@ -102,7 +102,8 @@ UPROGS= \
 	_dirbench \
 	_usertests \
 	_lockstat \
-	_preadtest
+	_preadtest \
+	_perf
 UPROGS := $(addprefix $(O)/, $(UPROGS))

 all: $(O)/kernel
@@ -155,6 +156,9 @@ $(O)/_%: $(O)/_%.unstripped
 $(O)/mkfs: mkfs.c fs.h
 	gcc -m32 -Werror -Wall -o $@ mkfs.c

+$(O)/print-perf: print-perf.c sampler.h
+	gcc -m32 -Werror -Wall -o $@ print-perf.c
+
 $(O)/fs.img: $(O)/mkfs README $(UPROGS)
 	@echo "  MKFS   $@"
 	$(Q)$(O)/mkfs $@ README $(UPROGS)

--- a/perf.cc
+++ b/perf.cc
+#include "types.h"
+#include "stat.h"
+#include "user.h"
+#include "fcntl.h"
+#include "sampler.h"
+
+static u64 selector = 
+  0UL << 32 |
+  1 << 24 | 
+  1 << 22 | 
+  1 << 20 |
+  1 << 17 | 
+  1 << 16 | 
+  0x00 << 8 | 
+  0x76;
+static u64 period = 100000;
+
+static void
+conf(int fd, sampop_t op)
+{
+  struct sampconf c = { 
+    op: op,
+    selector: selector,
+    period: period,
+  };
+
+  if (write(fd, &c, sizeof(c)) != sizeof(c))
+    die("perf: write failed");
+  close(fd);
+}
+
+static void
+save(void)
+{
+  static char buf[4096];
+  int fd, sfd;
+  int r;
+
+  fd = open("/dev/sampler", O_RDONLY);
+  if (fd < 0)
+    die("perf: open failed");
+
+  unlink("perf.data");
+  sfd = open("perf.data", O_RDWR|O_CREATE);
+  if (sfd < 0)
+    die("perf: open failed");
+
+  while (1) {
+    r = read(fd, buf, sizeof(buf));
+    if (r < 0)
+      die("perf: read failed");
+    if (r == 0)
+      break;
+    if (write(sfd, buf, r) != r)
+      die("perf: truncated write");
+  }
+
+  close(sfd);
+  close(fd);
+}
+
+int
+main(int ac, const char *av[])
+{
+  int fd = open("/dev/sampler", O_RDWR);
+  if (fd < 0)
+    die("perf: open failed");
+
+  int pid = fork(0);
+  if (pid < 0)
+    die("perf: fork failed");
+
+  if (pid == 0) {
+    conf(fd, SAMP_ENABLE);
+    exec(av[1], av+1);
+    die("perf: exec failed");
+  }
+  
+  wait();
+  conf(fd, SAMP_DISABLE);
+  save();
+  exit();
+}
--- a/print-perf.c
+++ b/print-perf.c
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "types.h"
+#include "sampler.h"
+
+int
+main(int ac, char **av)
+{
+  static const char *pathname = "perf.data";
+  struct logheader *header;
+  struct stat buf;
+  void *x;
+  int fd;
+  int i;
+
+  if (ac > 1)
+    pathname = av[1];
+
+  fd = open(pathname, O_RDONLY);
+  if (fd < 0) {
+    perror("open");
+    exit(EXIT_FAILURE);
+  }
+  
+  if (fstat(fd, &buf) < 0) {
+    perror("fstat");
+    exit(EXIT_FAILURE);
+  }
+
+  x = mmap(0, buf.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
+  if (x == MAP_FAILED) {
+    perror("mmap");
+    exit(EXIT_FAILURE);
+  }
+  header = x;
+
+  printf("ncpus %u\n", header->ncpus);
+  for (i = 0; i < header->ncpus; i++) {
+    printf(" offset %lu size %lu\n",
+           header->cpu[i].offset, header->cpu[i].size);
+  }
+
+  return 0;
+}
--- a/sampler.cc
+++ b/sampler.cc
@@ -15,17 +15,8 @@ extern "C" {
 #include "sampler.h"
 }

-static const u64 debug_sel = 
-  0UL << 32 |
-  1 << 24 | 
-  1 << 22 | 
-  1 << 20 |
-  1 << 17 | 
-  1 << 16 | 
-  0x00 << 8 | 
-  0x76;
-
-static const u64 debug_cnt = 100000;
+static volatile u64 selector;
+static volatile u64 period;

 struct pmu {
  void (*config)(u64 ctr, u64 sel, u64 val);  
@@ -84,20 +75,21 @@ void
 sampconf(void)
 {
  pushcli();
-  pmu.config(0, debug_sel, -debug_cnt);
+  pmu.config(0, selector, -period);
  popcli();
 }

 void
 sampstart(void)
 {
+  pushcli();
  for(struct cpu *c = cpus; c < cpus+ncpu; c++) {
    if(c == cpus+cpunum())
      continue;
    lapic_sampconf(c->id);
  }
-
  sampconf();
+  popcli();
 }

 static int
@@ -129,7 +121,7 @@ sampintr(struct trapframe *tf)
    return 0;

  if (samplog(tf))
-    pmu.config(0, debug_sel, -debug_cnt);
+    pmu.config(0, selector, -period);

  return 1;
 }
@@ -205,6 +197,35 @@ sampread(struct inode *ip, char *dst, u32 off, u32 n)
  return ret;
 }

+static int
+sampwrite(struct inode *ip, char *buf, u32 off, u32 n)
+{
+  struct sampconf *conf;
+
+  if (n != sizeof(*conf))
+    return -1;
+  conf = (struct sampconf*)buf;
+
+  switch(conf->op) {
+  case SAMP_ENABLE:
+    selector = conf->selector;
+    period = conf->period;
+    cprintf("sampwrite: selector %lu period %lu\n",
+            selector, period);
+    sampstart();
+    break;
+  case SAMP_DISABLE:
+    selector = 0;
+    period = 0;
+    for (int i = 0; i < NCPU; i++)
+      cprintf("sampwrite: count %lu\n", pmulog[i].count);
+    sampstart();
+    break;
+  }
+
+  return n;
+}
+
 void
 initsamp(void)
 {
@@ -234,6 +255,6 @@ initsamp(void)
  pmulog[cpunum()].event = (pmuevent*) p;
  pmulog[cpunum()].capacity = PERFSIZE / sizeof(struct pmuevent);

-  devsw[SAMPLER].write = 0;
+  devsw[SAMPLER].write = sampwrite;
  devsw[SAMPLER].read = sampread;
 }
--- a/sampler.h
+++ b/sampler.h
+typedef enum { SAMP_ENABLE, SAMP_DISABLE } sampop_t;
+
+struct sampconf {
+  sampop_t op;
+  u64 selector;
+  u64 period;
+};
+
 struct pmuevent {
  u64 rip;
 };

--- a/spinlock.cc
+++ b/spinlock.cc
@@ -114,7 +114,7 @@ static struct spinlock lockstat_lock = {
 void
 lockstat_init(struct spinlock *lk)
 {
-    lk->stat = (struct klockstat*) kmalloc(sizeof(*lk->stat));
+  lk->stat = (struct klockstat*) kmalloc(sizeof(*lk->stat));
  if (lk->stat == NULL)
    return;