]> Cypherpunks repositories - gostls13.git/commitdiff
runtime: shift timeHistogram buckets and allow negative durations
authorMichael Anthony Knyszek <mknyszek@google.com>
Tue, 22 Dec 2020 17:47:43 +0000 (17:47 +0000)
committerMichael Knyszek <mknyszek@google.com>
Wed, 23 Dec 2020 17:31:18 +0000 (17:31 +0000)
Today, timeHistogram, when copied, has the wrong set of counts for the
bucket that should represent (-inf, 0), when in fact it contains [0, 1).
In essence, the buckets are all shifted over by one from where they're
supposed to be.

But this also means that the existence of the overflow bucket is wrong:
the top bucket is supposed to extend to infinity, and what we're really
missing is an underflow bucket to represent the range (-inf, 0).

We could just always zero this bucket and continue ignoring negative
durations, but that likely isn't prudent.

timeHistogram is intended to be used with differences in nanotime, but
depending on how a platform is implemented (or due to a bug in that
platform) it's possible to get a negative duration without having done
anything wrong. We should just be resilient to that and be able to
detect it.

So this change removes the overflow bucket and replaces it with an
underflow bucket, and timeHistogram no longer panics when faced with a
negative duration.

Fixes #43328.
Fixes #43329.

Change-Id: If336425d7d080fd37bf071e18746800e22d38108
Reviewed-on: https://go-review.googlesource.com/c/go/+/279468
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
src/runtime/export_test.go
src/runtime/histogram.go
src/runtime/histogram_test.go
src/runtime/metrics.go

index 44551dcaf141ef859383313c6c3c64081b26fdb8..22fef3134f54a91bd2c151e198ec7ebed72976f1 100644 (file)
@@ -1201,12 +1201,12 @@ type TimeHistogram timeHistogram
 
 // Counts returns the counts for the given bucket, subBucket indices.
 // Returns true if the bucket was valid, otherwise returns the counts
-// for the overflow bucket and false.
+// for the underflow bucket and false.
 func (th *TimeHistogram) Count(bucket, subBucket uint) (uint64, bool) {
        t := (*timeHistogram)(th)
        i := bucket*TimeHistNumSubBuckets + subBucket
        if i >= uint(len(t.counts)) {
-               return t.overflow, false
+               return t.underflow, false
        }
        return t.counts[i], true
 }
index 4020969eb993a590a121d475b6cb66b285f06fb9..d48e856cd0d8531f63b39b4607de553e4551ece9 100644 (file)
@@ -69,17 +69,15 @@ const (
 // for concurrent use. It is also safe to read all the values
 // atomically.
 type timeHistogram struct {
-       counts   [timeHistNumSuperBuckets * timeHistNumSubBuckets]uint64
-       overflow uint64
+       counts    [timeHistNumSuperBuckets * timeHistNumSubBuckets]uint64
+       underflow uint64
 }
 
 // record adds the given duration to the distribution.
-//
-// Although the duration is an int64 to facilitate ease-of-use
-// with e.g. nanotime, the duration must be non-negative.
 func (h *timeHistogram) record(duration int64) {
        if duration < 0 {
-               throw("timeHistogram encountered negative duration")
+               atomic.Xadd64(&h.underflow, 1)
+               return
        }
        // The index of the exponential bucket is just the index
        // of the highest set bit adjusted for how many bits we
@@ -92,15 +90,17 @@ func (h *timeHistogram) record(duration int64) {
                superBucket = uint(sys.Len64(uint64(duration))) - timeHistSubBucketBits
                if superBucket*timeHistNumSubBuckets >= uint(len(h.counts)) {
                        // The bucket index we got is larger than what we support, so
-                       // add into the special overflow bucket.
-                       atomic.Xadd64(&h.overflow, 1)
-                       return
+                       // include this count in the highest bucket, which extends to
+                       // infinity.
+                       superBucket = timeHistNumSuperBuckets - 1
+                       subBucket = timeHistNumSubBuckets - 1
+               } else {
+                       // The linear subbucket index is just the timeHistSubBucketsBits
+                       // bits after the top bit. To extract that value, shift down
+                       // the duration such that we leave the top bit and the next bits
+                       // intact, then extract the index.
+                       subBucket = uint((duration >> (superBucket - 1)) % timeHistNumSubBuckets)
                }
-               // The linear subbucket index is just the timeHistSubBucketsBits
-               // bits after the top bit. To extract that value, shift down
-               // the duration such that we leave the top bit and the next bits
-               // intact, then extract the index.
-               subBucket = uint((duration >> (superBucket - 1)) % timeHistNumSubBuckets)
        } else {
                subBucket = uint(duration)
        }
@@ -128,7 +128,7 @@ func timeHistogramMetricsBuckets() []float64 {
                // index to combine it with the bucketMin.
                subBucketShift := uint(0)
                if i > 1 {
-                       // The first two buckets are exact with respect to integers,
+                       // The first two super buckets are exact with respect to integers,
                        // so we'll never have to shift the sub-bucket index. Thereafter,
                        // we shift up by 1 with each subsequent bucket.
                        subBucketShift = uint(i - 2)
index 5f5b28f7848d3f1b1b353f444345f94b1ee9029e..dbc64fa5592407bc728a5f0c1d70d30048e2b0b6 100644 (file)
@@ -5,6 +5,7 @@
 package runtime_test
 
 import (
+       "math"
        . "runtime"
        "testing"
 )
@@ -32,8 +33,8 @@ func TestTimeHistogram(t *testing.T) {
                        h.Record(base + v)
                }
        }
-       // Hit the overflow bucket.
-       h.Record(int64(^uint64(0) >> 1))
+       // Hit the underflow bucket.
+       h.Record(int64(-1))
 
        // Check to make sure there's exactly one count in each
        // bucket.
@@ -41,7 +42,7 @@ func TestTimeHistogram(t *testing.T) {
                for j := uint(0); j < TimeHistNumSubBuckets; j++ {
                        c, ok := h.Count(i, j)
                        if !ok {
-                               t.Errorf("hit overflow bucket unexpectedly: (%d, %d)", i, j)
+                               t.Errorf("hit underflow bucket unexpectedly: (%d, %d)", i, j)
                        } else if c != 1 {
                                t.Errorf("bucket (%d, %d) has count that is not 1: %d", i, j, c)
                        }
@@ -49,10 +50,21 @@ func TestTimeHistogram(t *testing.T) {
        }
        c, ok := h.Count(TimeHistNumSuperBuckets, 0)
        if ok {
-               t.Errorf("expected to hit overflow bucket: (%d, %d)", TimeHistNumSuperBuckets, 0)
+               t.Errorf("expected to hit underflow bucket: (%d, %d)", TimeHistNumSuperBuckets, 0)
        }
        if c != 1 {
-               t.Errorf("overflow bucket has count that is not 1: %d", c)
+               t.Errorf("underflow bucket has count that is not 1: %d", c)
        }
+
+       // Check overflow behavior.
+       // By hitting a high value, we should just be adding into the highest bucket.
+       h.Record(math.MaxInt64)
+       c, ok = h.Count(TimeHistNumSuperBuckets-1, TimeHistNumSubBuckets-1)
+       if !ok {
+               t.Error("hit underflow bucket in highest bucket unexpectedly")
+       } else if c != 2 {
+               t.Errorf("highest has count that is not 2: %d", c)
+       }
+
        dummyTimeHistogram = TimeHistogram{}
 }
index af86a0f03aa00a99adc78f6faa662e54f7f67da4..1d191e62985d5e2a3d10c9bd1302a7d17465ffe5 100644 (file)
@@ -116,9 +116,9 @@ func initMetrics() {
                "/gc/pauses:seconds": {
                        compute: func(_ *statAggregate, out *metricValue) {
                                hist := out.float64HistOrInit(timeHistBuckets)
-                               hist.counts[len(hist.counts)-1] = atomic.Load64(&memstats.gcPauseDist.overflow)
+                               hist.counts[0] = atomic.Load64(&memstats.gcPauseDist.underflow)
                                for i := range hist.buckets {
-                                       hist.counts[i] = atomic.Load64(&memstats.gcPauseDist.counts[i])
+                                       hist.counts[i+1] = atomic.Load64(&memstats.gcPauseDist.counts[i])
                                }
                        },
                },