runtime: fix malloc sampling bug

The malloc sample trigger was not being set in a
new m, so the first allocation in each new m - the
goroutine structure - was being sampled with
probability 1 instead of probability sizeof(G)/rate,
an oversampling of about 5000x for the default
rate of 1 MB.  This bug made pprof graphs show
far more G allocations than there actually were.

R=golang-dev, r
CC=golang-dev
https://golang.org/cl/5224041

diff --git a/src/pkg/runtime/malloc.goc b/src/pkg/runtime/malloc.goc
index 84e0ac4795eb595a1fc6b21b44d73f0570736c3f..6d2f65b3c94570956bfe102a558cfcafe38c3eed 100644 (file)
--- a/src/pkg/runtime/malloc.goc
+++ b/src/pkg/runtime/malloc.goc
@@ -80,6 +80,7 @@ runtime·mallocgc(uintptr size, uint32 flag, int32 dogc, int32 zeroed)
                        m->mcache->next_sample -= size;
                else {
                        // pick next profile time
+                       // If you change this, also change allocmcache.
                        if(rate > 0x3fffffff)   // make 2*rate not overflow
                                rate = 0x3fffffff;
                        m->mcache->next_sample = runtime·fastrand1() % (2*rate);
@@ -205,6 +206,7 @@ runtime·mlookup(void *v, byte **base, uintptr *size, MSpan **sp)
 MCache*
 runtime·allocmcache(void)
 {
+       int32 rate;
        MCache *c;
 
        runtime·lock(&runtime·mheap);
@@ -212,6 +214,13 @@ runtime·allocmcache(void)
        mstats.mcache_inuse = runtime·mheap.cachealloc.inuse;
        mstats.mcache_sys = runtime·mheap.cachealloc.sys;
        runtime·unlock(&runtime·mheap);
+
+       // Set first allocation sample size.
+       rate = runtime·MemProfileRate;
+       if(rate > 0x3fffffff)   // make 2*rate not overflow
+               rate = 0x3fffffff;
+       c->next_sample = runtime·fastrand1() % (2*rate);
+
        return c;
 }