Some code to help instrument kernel for profiling..

af216111 · Silas Boyd-Wickizer · 56eccf91 · af216111 · af216111 · af216111
--- a/Makefile
+++ b/Makefile
@@ -46,6 +46,7 @@ OBJS = \
 	picirq.o \
 	pipe.o \
 	proc.o \
+	prof.o \
 	rcu.o \
 	sched.o \
 	spinlock.o \

--- a/console.c
+++ b/console.c
@@ -308,6 +308,18 @@ consoleintr(int (*getc)(void))
    case C('W'):  // Work queue stats
      wq_dump();
      break;
+    case C('L'):  // Prof stats
+      profdump();
+      break;
+    case C('K'):  // Prof enable
+      profreset();
+      cprintf("prof enabled\n");
+      profenable = 1;
+      break;
+    case C('I'):  // Prof disable
+      profenable = 0;
+      cprintf("prof disabled\n");
+      break;
    default:
      if(c != 0 && input.e-input.r < INPUT_BUF){
        c = (c == '\r') ? '\n' : c;

--- a/exec.c
+++ b/exec.c
@@ -13,12 +13,15 @@
 #include "elf.h"
 #include "cpu.h"
 #include "vm.h"
+#include "prof.h"
 #define USTACKPAGES 2
 #define BRK (USERTOP >> 1)
 static const int odp = 1;
+DEFINE_PROFCTR(dosegment_prof);
 struct eargs {
  struct inode *ip;
  struct vmap *vmap;
@@ -32,6 +35,7 @@ dosegment(uptr a0, u64 a1)
  struct vmnode *vmn = NULL;
  struct proghdr ph;
+  prof_start(dosegment_prof);
  if(readi(args->ip, (char*)&ph, off, sizeof(ph)) != sizeof(ph))
    goto bad;
  if(ph.type != ELF_PROG_LOAD) {
@@ -63,6 +67,7 @@ dosegment(uptr a0, u64 a1)
  if(vmap_insert(args->vmap, vmn, ph.vaddr) < 0)
    goto bad;
+  prof_end(dosegment_prof);
  return;
 bad:

--- a/kernel.h
+++ b/kernel.h
@@ -204,6 +204,11 @@ int             wait(void);
 void            yield(void);
 void            migrate(struct proc *);
+// prof.c
+extern int profenable;
+void            profreset(void);
+void            profdump(void);
 // rcu.c
 void            rcuinit(void);
 void            rcu_begin_write(struct spinlock *);

--- a/kernel.ld
+++ b/kernel.ld
@@ -13,6 +13,12 @@ SECTIONS
 		*(.rodata .rodata.* .gnu.linkonce.r.*)
 	}
 	. = ALIGN(0x1000);
+	PROVIDE(sprof = .);
+	.prof : {
+		*(.prof)
+	}
+	PROVIDE(eprof = .);
+	. = ALIGN(0x1000);
 	.data : {
 		*(.data)
 	}

--- a/prof.c
+++ b/prof.c
+#include "param.h"
+#include "types.h"
+#include "spinlock.h"
+#include "condvar.h"
+#include "fs.h"
+#include "file.h"
+#include "prof.h"
+#include "kernel.h"
+extern profctr_t sprof[];
+extern profctr_t eprof[];
+int profenable;
+void
+profreset(void)
+{
+  profctr_t *p = sprof;
+  for (; p != eprof; p++) {
+    memset(p->rec, 0, sizeof(p->rec));
+  }
+}
+static void
+profsum(struct profctr *ctr, u64 *tot, u64 *cnt)
+{
+  for (int i = 0; i < NCPU; i++) {
+    *tot += ctr->rec[i].tot;
+    *cnt += ctr->rec[i].cnt;
+  }
+}
+void
+profdump(void)
+{
+  profctr_t *p = sprof;
+  for (; p != eprof; p++) {
+    u64 tot = 0, cnt = 0;
+    profsum(p, &tot, &cnt);
+    if (cnt)
+      cprintf("%s %lu\n", p->name, tot/cnt);
+  }
+}
--- a/prof.h
+++ b/prof.h
+struct profrec {
+  u64 tot;
+  u64 cnt;
+  __padout__;
+} __mpalign__;
+typedef struct profctr {
+  const char *name;
+  struct profrec rec[NCPU] __mpalign__;
+  __padout__;
+} profctr_t;
+#define DEFINE_PROFCTR(xname) \
+  profctr_t xname __attribute__((section(".prof"))) = { .name = #xname };
+#define prof_start(name) \
+  u64 __prof##name = profenable ? rdtsc() : 0;
+#define prof_end(name) do { \
+  if (profenable) { \
+    u64 __eprof##name = rdtsc(); \
+    u64 __profid = mycpu()->id; \
+    name.rec[__profid].tot += __eprof##name - __prof##name; \
+    name.rec[__profid].cnt++; \
+  } \
+} while (0)