pkg/host: improve KMEMLEAK support

Rewind kmemleak fd before reading it second time,
otherwise we will read truncated reports.

Auto-learn what leak reports we've already seen
and ignore them in future. This is required because
there are some false positives and some fire too frequently.
So now we will hit each leak only once per manager run,
but we still will try to reproduce them.

2 files changed, 22 insertions, 38 deletions

diff --git a/pkg/host/host.go b/pkg/host/host.go
index b23dd9ab2..36c42de65 100644
--- a/pkg/host/host.go
+++ b/pkg/host/host.go
@@ -65,7 +65,7 @@ type Features [numFeatures]Feature
 
 var checkFeature [numFeatures]func() string
 var setupFeature [numFeatures]func() error
-var callbFeature [numFeatures]func()
+var callbFeature [numFeatures]func(leakFrames [][]byte)
 
 func unconditionallyEnabled() string { return "" }
 
@@ -104,11 +104,11 @@ func Check(target *prog.Target) (*Features, error) {
 
 // Setup enables and does any one-time setup for the requested features on the host.
 // Note: this can be called multiple times and must be idempotent.
-func Setup(target *prog.Target, features *Features) (func(), error) {
+func Setup(target *prog.Target, features *Features) (func(leakFrames [][]byte), error) {
 	if target.OS == "akaros" || target.OS == "test" {
 		return nil, nil
 	}
-	var callback func()
+	var callback func([][]byte)
 	for n, setup := range setupFeature {
 		if setup == nil || !features[n].Enabled {
 			continue
@@ -117,15 +117,15 @@ func Setup(target *prog.Target, features *Features) (func(), error) {
 			return nil, err
 		}
 		cb := callbFeature[n]
-		if cb != nil {
-			prev := callback
-			callback = func() {
-				cb()
-				if prev != nil {
-					prev()
-				}
+		if cb == nil {
+			continue
+		}
+		prev := callback
+		callback = func(leakFrames [][]byte) {
+			cb(leakFrames)
+			if prev != nil {
+				prev(leakFrames)
 			}
-
 		}
 	}
 	return callback, nil
diff --git a/pkg/host/host_linux.go b/pkg/host/host_linux.go
index 8214b35af..5c5db793b 100644
--- a/pkg/host/host_linux.go
+++ b/pkg/host/host_linux.go
@@ -469,7 +469,7 @@ func setupLeakChecking() error {
 	return nil
 }
 
-func callbackLeakChecking() {
+func callbackLeakChecking(leakFrames [][]byte) {
 	start := time.Now()
 	fd, err := syscall.Open("/sys/kernel/debug/kmemleak", syscall.O_RDWR, 0)
 	if err != nil {
@@ -506,11 +506,15 @@ func callbackLeakChecking() {
 		if _, err := syscall.Write(fd, []byte("scan")); err != nil {
 			panic(err)
 		}
+		if _, err := syscall.Seek(fd, 0, 0); err != nil {
+			panic(err)
+		}
 		n, err := syscall.Read(fd, buf)
 		if err != nil {
 			panic(err)
 		}
 		nleaks := 0
+	nextLeak:
 		for buf = buf[:n]; len(buf) != 0; {
 			end := bytes.Index(buf[1:], []byte("unreferenced object"))
 			if end != -1 {
@@ -520,14 +524,18 @@ func callbackLeakChecking() {
 			}
 			report := buf[:end]
 			buf = buf[end:]
-			if kmemleakIgnore(report) {
-				continue
+			for _, frame := range leakFrames {
+				if bytes.Contains(report, frame) {
+					continue nextLeak
+				}
 			}
 			// BUG in output should be recognized by manager.
 			fmt.Printf("BUG: memory leak\n%s\n", report)
 			nleaks++
 		}
 		if nleaks != 0 {
+			// If we exit right away, dying executors will dump lots of garbage to console.
+			time.Sleep(time.Hour)
 			os.Exit(1)
 		}
 	}
@@ -536,30 +544,6 @@ func callbackLeakChecking() {
 	}
 }
 
-func kmemleakIgnore(report []byte) bool {
-	// kmemleak has a bunch of false positives (at least what looks like
-	// false positives at first glance). So we are conservative with what we report.
-	// First, we filter out any allocations that don't come from executor processes.
-	// Second, we ignore a bunch of functions entirely.
-	// Ideally, someone should debug/fix all these cases and remove ignores.
-	if !bytes.Contains(report, []byte(`comm "syz-executor`)) {
-		return true
-	}
-	for _, ignore := range []string{
-		" copy_process",
-		" do_execveat_common",
-		" __ext4_",
-		" get_empty_filp",
-		" do_filp_open",
-		" new_inode",
-	} {
-		if bytes.Contains(report, []byte(ignore)) {
-			return true
-		}
-	}
-	return false
-}
-
 func checkSandboxNamespace() string {
 	if err := osutil.IsAccessible("/proc/self/ns/user"); err != nil {
 		return err.Error()