cmd/pprof: go1.3 profiles and tip profiles show different information

[rsc]

pprof has a list of internal runtime functions to squelch.
that list needs to be updated for tip.
in particular, old profiles did not show malloc - instead a malloc is accounted to the
funtion in which it appears. but at tip profiles are showing malloc again.

[rsc]

Comment 1:

Related: the malloc that shows up is all stack-less because it is on the g0 stack.
Probably want to charge g0 ticks to the curg if there is one. Something like
diff -r 284a98e61e32 src/runtime/proc.c
--- a/src/runtime/proc.c    Wed Oct 15 19:33:15 2014 -0400
+++ b/src/runtime/proc.c    Wed Oct 15 23:13:50 2014 -0400
@@ -2436,7 +2436,7 @@
 void
 runtime·sigprof(uint8 *pc, uint8 *sp, uint8 *lr, G *gp, M *mp)
 {
-   int32 n;
+   int32 n, off;
    bool traceback;
    // Do not use global m in this function, use mp instead.
    // On windows one m is sending reports about all the g's, so m means a wrong thing.
@@ -2530,9 +2530,20 @@
       ((uint8*)runtime·gogo <= pc && pc < (uint8*)runtime·gogo + RuntimeGogoBytes))
        traceback = false;
 
+   off = 0;
+   if(gp == mp->g0 && mp->curg != nil) {
+       stk[0] = (uintptr)pc;
+       off = 1;
+       gp = mp->curg;
+       pc = (uint8*)gp->sched.pc;
+       sp = (uint8*)gp->sched.sp;
+       lr = 0;
+       traceback = true;
+   }
+
    n = 0;
    if(traceback)
-       n = runtime·gentraceback((uintptr)pc, (uintptr)sp, (uintptr)lr, gp, 0, stk,
nelem(stk), nil, nil, false);
+       n = runtime·gentraceback((uintptr)pc, (uintptr)sp, (uintptr)lr, gp, 0, stk+off,
nelem(stk)-off, nil, nil, false);
    if(!traceback || n <= 0) {
        // Normal traceback is impossible or has failed.
        // See if it falls into several common cases.

[dvyukov]

Comment 2:

There is also related issue
https://golang.org/issue/8867

[gopherbot]

Comment 3:

CL https://golang.org/cl/159990043 mentions this issue.

[rsc]

Comment 4:

This issue was updated by revision 7ed8723.

This does not fully address issue #8942 but it does make
the profiles much more useful, until that issue can be
fixed completely.
LGTM=dvyukov
R=r, dvyukov
CC=golang-codereviews
https://golang.org/cl/159990043

[rsc]

Comment 5:

This issue was updated by revision 1ba977c.

This does not fully address issue #8942 but it does make
the profiles much more useful, until that issue can be
fixed completely.
LGTM=dvyukov
R=r, dvyukov
CC=golang-codereviews
https://golang.org/cl/159990043
»»»
TBR=dvyukov
CC=golang-codereviews
https://golang.org/cl/160030043

[rsc]

Comment 6:

Although it is still true that they show different information, they're still useful, so
they're good enough for now. Revisit for 1.5.

Labels changed: added release-go1.5, removed release-go1.4.

[rsc]

I was just looking at pprof profiles and it seems okay now.
And the stack traces are useful on the system stack too.