pkg/bisect: support clang for crash bisection

Before we hardcoded bisection to use gcc, now the compiler family can
be configured in the bisection config.

8 files changed, 107 insertions, 61 deletions

diff --git a/pkg/bisect/bisect.go b/pkg/bisect/bisect.go
index 0948bc653..1813abaa5 100644
--- a/pkg/bisect/bisect.go
+++ b/pkg/bisect/bisect.go
@@ -19,15 +19,16 @@ import (
 )
 
 type Config struct {
-	Trace     debugtracer.DebugTracer
-	Fix       bool
-	BinDir    string
-	Ccache    string
-	Timeout   time.Duration
-	Kernel    KernelConfig
-	Syzkaller SyzkallerConfig
-	Repro     ReproConfig
-	Manager   *mgrconfig.Config
+	Trace          debugtracer.DebugTracer
+	Fix            bool
+	BisectCompiler string
+	BinDir         string
+	Ccache         string
+	Timeout        time.Duration
+	Kernel         KernelConfig
+	Syzkaller      SyzkallerConfig
+	Repro          ReproConfig
+	Manager        *mgrconfig.Config
 }
 
 type KernelConfig struct {
@@ -375,7 +376,7 @@ func (env *env) commitRangeForFix() (*vcs.Commit, *vcs.Commit, *report.Report, [
 
 func (env *env) commitRangeForBug() (*vcs.Commit, *vcs.Commit, *report.Report, []*testResult, error) {
 	cfg := env.cfg
-	tags, err := env.bisecter.PreviousReleaseTags(cfg.Kernel.Commit)
+	tags, err := env.bisecter.PreviousReleaseTags(cfg.Kernel.Commit, cfg.BisectCompiler)
 	if err != nil {
 		return nil, nil, nil, nil, err
 	}
@@ -420,11 +421,11 @@ func (env *env) build() (*vcs.Commit, string, error) {
 		return nil, "", err
 	}
 
-	bisectEnv, err := env.bisecter.EnvForCommit(env.cfg.BinDir, current.Hash, env.kernelConfig)
+	bisectEnv, err := env.bisecter.EnvForCommit(env.cfg.BisectCompiler, env.cfg.BinDir, current.Hash, env.kernelConfig)
 	if err != nil {
 		return current, "", err
 	}
-	env.log("testing commit %v", current.Hash)
+	env.log("testing commit %v %v", current.Hash, env.cfg.BisectCompiler)
 	buildStart := time.Now()
 	mgr := env.cfg.Manager
 	if err := build.Clean(mgr.TargetOS, mgr.TargetVMArch, mgr.Type, mgr.KernelSrc); err != nil {
diff --git a/pkg/vcs/linux.go b/pkg/vcs/linux.go
index d2371ca5c..8ee801430 100644
--- a/pkg/vcs/linux.go
+++ b/pkg/vcs/linux.go
@@ -42,32 +42,43 @@ func newLinux(dir string, opts []RepoOpt, vmType string) *linux {
 	}
 }
 
-func (ctx *linux) PreviousReleaseTags(commit string) ([]string, error) {
+func (ctx *linux) PreviousReleaseTags(commit, bisectCompiler string) ([]string, error) {
 	tags, err := ctx.git.previousReleaseTags(commit, false, false, false)
 	if err != nil {
 		return nil, err
 	}
+
+	cutoff := ""
+	if bisectCompiler == "gcc" {
+		// Initially we tried to stop at 3.8 because:
+		// v3.8 does not work with modern perl, and as we go further in history
+		// make stops to work, then binutils, glibc, etc. So we stop at v3.8.
+		// Up to that point we only need an ancient gcc.
+		//
+		// But kernels don't boot starting from 4.0 and back.
+		// That was fixed by 99124e4db5b7b70daeaaf1d88a6a8078a0004c6e,
+		// and it can be cherry-picked into 3.14..4.0 but it conflicts for 3.13 and older.
+		//
+		// But starting from 4.0 our user-space binaries start crashing with
+		// assorted errors which suggests process memory corruption by kernel.
+		//
+		// We used to use 4.1 as the oldest tested release (it works in general).
+		// However, there is correlation between how far back we go and probability
+		// of getting correct result (see #1532). So we now stop at 4.6.
+		// 4.6 is somewhat arbitrary, we've seen lots of wrong results in 4.5..4.6 range,
+		// but there is definitive reason for 4.6. Most likely later we want to bump it
+		// even more (as new releases are produced). Next good candidate may be 4.11
+		// because then we won't need gcc 5.5.
+		cutoff = "v4.5"
+	} else if bisectCompiler == "clang" {
+		// v5.3 was the first release with solid clang support, however I was able to
+		// compile v5.1..v5.3 using a newer defconfig + make oldconfig. Everything older
+		// would require further cherry-picks.
+		cutoff = "v5.2"
+	}
+
 	for i, tag := range tags {
-		if tag == "v4.5" {
-			// Initially we tried to stop at 3.8 because:
-			// v3.8 does not work with modern perl, and as we go further in history
-			// make stops to work, then binutils, glibc, etc. So we stop at v3.8.
-			// Up to that point we only need an ancient gcc.
-			//
-			// But kernels don't boot starting from 4.0 and back.
-			// That was fixed by 99124e4db5b7b70daeaaf1d88a6a8078a0004c6e,
-			// and it can be cherry-picked into 3.14..4.0 but it conflicts for 3.13 and older.
-			//
-			// But starting from 4.0 our user-space binaries start crashing with
-			// assorted errors which suggests process memory corruption by kernel.
-			//
-			// We used to use 4.1 as the oldest tested release (it works in general).
-			// However, there is correlation between how far back we go and probability
-			// of getting correct result (see #1532). So we now stop at 4.6.
-			// 4.6 is somewhat arbitrary, we've seen lots of wrong results in 4.5..4.6 range,
-			// but there is definitive reason for 4.6. Most likely later we want to bump it
-			// even more (as new releases are produced). Next good candidate may be 4.11
-			// because then we won't need gcc 5.5.
+		if tag == cutoff {
 			tags = tags[:i]
 			break
 		}
@@ -121,7 +132,7 @@ func gitReleaseTagToInt(tag string, includeRC bool) uint64 {
 	return v1*1e9 + v2*1e6 + rc*1e3 + v3
 }
 
-func (ctx *linux) EnvForCommit(binDir, commit string, kernelConfig []byte) (*BisectEnv, error) {
+func (ctx *linux) EnvForCommit(bisectCompiler, binDir, commit string, kernelConfig []byte) (*BisectEnv, error) {
 	tagList, err := ctx.previousReleaseTags(commit, true, false, false)
 	if err != nil {
 		return nil, err
@@ -135,8 +146,18 @@ func (ctx *linux) EnvForCommit(binDir, commit string, kernelConfig []byte) (*Bis
 		return nil, err
 	}
 	linuxAlterConfigs(cf, tags)
+
+	compiler := ""
+	if bisectCompiler == "gcc" {
+		compiler = filepath.Join(binDir, "gcc-"+linuxGCCVersion(tags), "bin", "gcc")
+	} else if bisectCompiler == "clang" {
+		compiler = filepath.Join(binDir, "llvm-"+linuxClangVersion(tags), "bin", "clang")
+	} else {
+		return nil, fmt.Errorf("unsupported bisect compiler: %v", bisectCompiler)
+	}
+
 	env := &BisectEnv{
-		Compiler:     filepath.Join(binDir, "gcc-"+linuxCompilerVersion(tags), "bin", "gcc"),
+		Compiler:     compiler,
 		KernelConfig: cf.Serialize(),
 	}
 
@@ -159,7 +180,18 @@ func (ctx *linux) EnvForCommit(binDir, commit string, kernelConfig []byte) (*Bis
 	return env, nil
 }
 
-func linuxCompilerVersion(tags map[string]bool) string {
+func linuxClangVersion(tags map[string]bool) string {
+	switch {
+	case tags["v5.9"]:
+		return "14.0.6"
+	default:
+		// everything before v5.3 might not work great
+		// everything before v5.1 does not work
+		return "9.0.1"
+	}
+}
+
+func linuxGCCVersion(tags map[string]bool) string {
 	switch {
 	case tags["v5.9"]:
 		return "10.1.0"
diff --git a/pkg/vcs/testos.go b/pkg/vcs/testos.go
index baadb1a2a..0053a528e 100644
--- a/pkg/vcs/testos.go
+++ b/pkg/vcs/testos.go
@@ -22,11 +22,11 @@ func newTestos(dir string, opts []RepoOpt) *testos {
 	}
 }
 
-func (ctx *testos) PreviousReleaseTags(commit string) ([]string, error) {
+func (ctx *testos) PreviousReleaseTags(commit, bisectCompiler string) ([]string, error) {
 	return ctx.git.previousReleaseTags(commit, false, false, false)
 }
 
-func (ctx *testos) EnvForCommit(binDir, commit string, kernelConfig []byte) (*BisectEnv, error) {
+func (ctx *testos) EnvForCommit(bisectCompiler, binDir, commit string, kernelConfig []byte) (*BisectEnv, error) {
 	return &BisectEnv{KernelConfig: kernelConfig}, nil
 }
 
diff --git a/pkg/vcs/vcs.go b/pkg/vcs/vcs.go
index ae7bc8508..4e42b04a0 100644
--- a/pkg/vcs/vcs.go
+++ b/pkg/vcs/vcs.go
@@ -78,11 +78,11 @@ type Bisecter interface {
 
 	// PreviousReleaseTags returns list of preceding release tags that are reachable from the given commit.
 	// If the commit itself has a release tag, this tag is not included.
-	PreviousReleaseTags(commit string) ([]string, error)
+	PreviousReleaseTags(commit, bisectCompiler string) ([]string, error)
 
 	IsRelease(commit string) (bool, error)
 
-	EnvForCommit(binDir, commit string, kernelConfig []byte) (*BisectEnv, error)
+	EnvForCommit(bisectCompiler, binDir, commit string, kernelConfig []byte) (*BisectEnv, error)
 }
 
 type ConfigMinimizer interface {
diff --git a/syz-ci/jobs.go b/syz-ci/jobs.go
index c77cb3f26..5a4addb02 100644
--- a/syz-ci/jobs.go
+++ b/syz-ci/jobs.go
@@ -412,10 +412,11 @@ func (jp *JobProcessor) bisect(job *Job, mgrcfg *mgrconfig.Config) error {
 		// compete with patch testing jobs (it's bad delaying patch testing).
 		// When/if bisection jobs don't compete with patch testing,
 		// it makes sense to increase this to 12-24h.
-		Timeout: 8 * time.Hour,
-		Fix:     req.Type == dashapi.JobBisectFix,
-		BinDir:  jp.cfg.BisectBinDir,
-		Ccache:  jp.cfg.Ccache,
+		Timeout:        8 * time.Hour,
+		Fix:            req.Type == dashapi.JobBisectFix,
+		BisectCompiler: mgr.mgrcfg.BisectCompiler,
+		BinDir:         jp.cfg.BisectBinDir,
+		Ccache:         jp.cfg.Ccache,
 		Kernel: bisect.KernelConfig{
 			Repo:           mgr.mgrcfg.Repo,
 			Branch:         mgr.mgrcfg.Branch,
diff --git a/syz-ci/syz-ci.go b/syz-ci/syz-ci.go
index 54d9b877c..944c0c745 100644
--- a/syz-ci/syz-ci.go
+++ b/syz-ci/syz-ci.go
@@ -102,10 +102,13 @@ type Config struct {
 	CoverUploadPath string `json:"cover_upload_path"`
 	// Path to upload corpus.db from managers (optional).
 	// Supported protocols: GCS (gs://) and HTTP PUT (http:// or https://).
-	CorpusUploadPath string           `json:"corpus_upload_path"`
-	BisectBinDir     string           `json:"bisect_bin_dir"`
-	Ccache           string           `json:"ccache"`
-	Managers         []*ManagerConfig `json:"managers"`
+	CorpusUploadPath string `json:"corpus_upload_path"`
+	// BinDir must point to a dir that contains compilers required to build
+	// older versions of the kernel. For linux, it needs to include several
+	// compiler versions.
+	BisectBinDir string           `json:"bisect_bin_dir"`
+	Ccache       string           `json:"ccache"`
+	Managers     []*ManagerConfig `json:"managers"`
 	// Poll period for jobs in seconds (optional, defaults to 10 seconds)
 	JobPollPeriod int `json:"job_poll_period"`
 	// Poll period for commits in seconds (optional, defaults to 3600 seconds)
@@ -149,12 +152,15 @@ type ManagerConfig struct {
 	DashboardKey    string `json:"dashboard_key"`
 	Repo            string `json:"repo"`
 	// Short name of the repo (e.g. "linux-next"), used only for reporting.
-	RepoAlias    string `json:"repo_alias"`
-	Branch       string `json:"branch"` // Defaults to "master".
-	Compiler     string `json:"compiler"`
-	Ccache       string `json:"ccache"`
-	Userspace    string `json:"userspace"`
-	KernelConfig string `json:"kernel_config"`
+	RepoAlias string `json:"repo_alias"`
+	// Currently either 'gcc' or 'clang'. Note that pkg/bisect requires
+	// explicit plumbing for every os/compiler combination.
+	BisectCompiler string `json:"bisect_compiler"` // Defaults to "gcc"
+	Branch         string `json:"branch"`          // Defaults to "master".
+	Compiler       string `json:"compiler"`
+	Ccache         string `json:"ccache"`
+	Userspace      string `json:"userspace"`
+	KernelConfig   string `json:"kernel_config"`
 	// Baseline config for bisection, see pkg/bisect.KernelConfig.BaselineConfig.
 	KernelBaselineConfig string `json:"kernel_baseline_config"`
 	// File with kernel cmdline values (optional).
@@ -380,6 +386,9 @@ func loadManagerConfig(cfg *Config, mgr *ManagerConfig) error {
 	if !managerNameRe.MatchString(mgr.Name) {
 		return fmt.Errorf("param 'managers.name' has bad value: %q", mgr.Name)
 	}
+	if mgr.BisectCompiler == "" {
+		mgr.BisectCompiler = "gcc"
+	}
 	if mgr.Branch == "" {
 		mgr.Branch = "master"
 	}
diff --git a/tools/syz-bisect/bisect.go b/tools/syz-bisect/bisect.go
index 1c8692700..9712ddcd3 100644
--- a/tools/syz-bisect/bisect.go
+++ b/tools/syz-bisect/bisect.go
@@ -43,10 +43,12 @@ var (
 )
 
 type Config struct {
+	// Currently either 'gcc' or 'clang'. Note that pkg/bisect requires
+	// explicit plumbing for every os/compiler combination.
+	BisectCompiler string `json:"bisect_compiler"`
 	// BinDir must point to a dir that contains compilers required to build
 	// older versions of the kernel. For linux, it needs to include several
-	// gcc versions. A working archive can be downloaded from:
-	// https://storage.googleapis.com/syzkaller/bisect_bin.tar.gz
+	// compiler versions.
 	BinDir        string `json:"bin_dir"`
 	Ccache        string `json:"ccache"`
 	KernelRepo    string `json:"kernel_repo"`
@@ -94,9 +96,10 @@ func main() {
 			TraceWriter: os.Stdout,
 			OutDir:      *flagCrash,
 		},
-		Fix:    *flagFix,
-		BinDir: mycfg.BinDir,
-		Ccache: mycfg.Ccache,
+		Fix:            *flagFix,
+		BisectCompiler: mycfg.BisectCompiler,
+		BinDir:         mycfg.BinDir,
+		Ccache:         mycfg.Ccache,
 		Kernel: bisect.KernelConfig{
 			Repo:      mycfg.KernelRepo,
 			Branch:    mycfg.KernelBranch,
diff --git a/tools/syz-testbuild/testbuild.go b/tools/syz-testbuild/testbuild.go
index 9fe1dd3f1..be2718b3a 100644
--- a/tools/syz-testbuild/testbuild.go
+++ b/tools/syz-testbuild/testbuild.go
@@ -101,7 +101,7 @@ func main() {
 		tool.Fail(err)
 	}
 	log.Printf("HEAD is on %v %v", head.Hash, head.Title)
-	tags, err := bisecter.PreviousReleaseTags(head.Hash)
+	tags, err := bisecter.PreviousReleaseTags(head.Hash, "gcc")
 	if err != nil {
 		tool.Fail(err)
 	}
@@ -125,7 +125,7 @@ func main() {
 }
 
 func test(repo vcs.Repo, bisecter vcs.Bisecter, kernelConfig []byte, env instance.Env, com *vcs.Commit) {
-	bisectEnv, err := bisecter.EnvForCommit(*flagBisectBin, com.Hash, kernelConfig)
+	bisectEnv, err := bisecter.EnvForCommit("gcc", *flagBisectBin, com.Hash, kernelConfig)
 	if err != nil {
 		tool.Fail(err)
 	}