report.go

// Copyright 2016 syzkaller project authors. All rights reserved.
// Use of this source code is governed by Apache 2 LICENSE that can be found in the LICENSE file.

// Package report contains functions that process kernel output,
// detect/extract crash messages, symbolize them, etc.
package report

import (
	"bufio"
	"bytes"
	"fmt"
	"regexp"
	"strings"

	"github.com/google/syzkaller/pkg/mgrconfig"
	"github.com/google/syzkaller/pkg/vcs"
	"github.com/google/syzkaller/sys/targets"
)

type reporterImpl interface {
	// ContainsCrash searches kernel console output for oops messages.
	ContainsCrash(output []byte) bool

	// Parse extracts information about oops from console output.
	// Returns nil if no oops found.
	Parse(output []byte) *Report

	// Symbolize symbolizes rep.Report and fills in Maintainers.
	Symbolize(rep *Report) error
}

type Reporter struct {
	impl         reporterImpl
	suppressions []*regexp.Regexp
	typ          string
}

type Report struct {
	// Title contains a representative description of the first oops.
	Title string
	// Alternative titles, used for better deduplication.
	// If two crashes have a non-empty intersection of Title/AltTitles, they are considered the same bug.
	AltTitles []string
	// Bug type (e.g. hang, memory leak, etc).
	Type Type
	// The indicative function name.
	Frame string
	// Report contains whole oops text.
	Report []byte
	// Output contains whole raw console output as passed to Reporter.Parse.
	Output []byte
	// StartPos/EndPos denote region of output with oops message(s).
	StartPos int
	EndPos   int
	// SkipPos is position in output where parsing for the next report should start.
	SkipPos int
	// Suppressed indicates whether the report should not be reported to user.
	Suppressed bool
	// Corrupted indicates whether the report is truncated of corrupted in some other way.
	Corrupted bool
	// CorruptedReason contains reason why the report is marked as corrupted.
	CorruptedReason string
	// Recipients is a list of RecipientInfo with Email, Display Name, and type.
	Recipients vcs.Recipients
	// guiltyFile is the source file that we think is to blame for the crash  (filled in by Symbolize).
	guiltyFile string
	// reportPrefixLen is length of additional prefix lines that we added before actual crash report.
	reportPrefixLen int
}

type Type int

const (
	Unknown Type = iota
	Hang
	MemoryLeak
	DataRace
	UnexpectedReboot
)

func (t Type) String() string {
	switch t {
	case Unknown:
		return "UNKNOWN"
	case Hang:
		return "HANG"
	case MemoryLeak:
		return "LEAK"
	case DataRace:
		return "DATARACE"
	case UnexpectedReboot:
		return "REBOOT"
	default:
		panic("unknown report type")
	}
}

// NewReporter creates reporter for the specified OS/Type.
func NewReporter(cfg *mgrconfig.Config) (*Reporter, error) {
	typ := cfg.TargetOS
	if cfg.Type == "gvisor" {
		typ = cfg.Type
	}
	ctor := ctors[typ]
	if ctor == nil {
		return nil, fmt.Errorf("unknown OS: %v", typ)
	}
	ignores, err := compileRegexps(cfg.Ignores)
	if err != nil {
		return nil, err
	}
	config := &config{
		target:         cfg.SysTarget,
		kernelSrc:      cfg.KernelSrc,
		kernelBuildSrc: cfg.KernelBuildSrc,
		kernelObj:      cfg.KernelObj,
		ignores:        ignores,
	}
	rep, suppressions, err := ctor(config)
	if err != nil {
		return nil, err
	}
	supps, err := compileRegexps(append(suppressions, cfg.Suppressions...))
	if err != nil {
		return nil, err
	}
	return &Reporter{rep, supps, typ}, nil
}

const (
	unexpectedKernelReboot = "unexpected kernel reboot"
	memoryLeakPrefix       = "memory leak in "
	dataRacePrefix         = "KCSAN: data-race"
	corruptedNoFrames      = "extracted no frames"
)

var ctors = map[string]fn{
	targets.Akaros:  ctorAkaros,
	targets.Linux:   ctorLinux,
	"gvisor":        ctorGvisor,
	targets.FreeBSD: ctorFreebsd,
	targets.Darwin:  ctorDarwin,
	targets.NetBSD:  ctorNetbsd,
	targets.OpenBSD: ctorOpenbsd,
	targets.Fuchsia: ctorFuchsia,
	targets.Windows: ctorStub,
}

type config struct {
	target         *targets.Target
	kernelSrc      string
	kernelBuildSrc string
	kernelObj      string
	ignores        []*regexp.Regexp
}

type fn func(cfg *config) (reporterImpl, []string, error)

func compileRegexps(list []string) ([]*regexp.Regexp, error) {
	compiled := make([]*regexp.Regexp, len(list))
	for i, str := range list {
		re, err := regexp.Compile(str)
		if err != nil {
			return nil, fmt.Errorf("failed to compile %q: %v", str, err)
		}
		compiled[i] = re
	}
	return compiled, nil
}

func (reporter *Reporter) Parse(output []byte) *Report {
	return reporter.ParseFrom(output, 0)
}

func (reporter *Reporter) ParseFrom(output []byte, minReportPos int) *Report {
	rep := reporter.impl.Parse(output[minReportPos:])
	if rep == nil {
		return nil
	}
	rep.Output = output
	rep.StartPos += minReportPos
	rep.EndPos += minReportPos
	rep.Title = sanitizeTitle(replaceTable(dynamicTitleReplacement, rep.Title))
	for i, title := range rep.AltTitles {
		rep.AltTitles[i] = sanitizeTitle(replaceTable(dynamicTitleReplacement, title))
	}
	rep.Suppressed = matchesAny(rep.Output, reporter.suppressions)
	if bytes.Contains(rep.Output, gceConsoleHangup) {
		rep.Corrupted = true
	}
	rep.Type = extractReportType(rep)
	if match := reportFrameRe.FindStringSubmatch(rep.Title); match != nil {
		rep.Frame = match[1]
	}
	rep.SkipPos = len(output)
	if pos := bytes.IndexByte(rep.Output[rep.StartPos:], '\n'); pos != -1 {
		rep.SkipPos = rep.StartPos + pos
	}
	if rep.EndPos < rep.SkipPos {
		// This generally should not happen.
		// But openbsd does some hacks with /r/n which may lead to off-by-one EndPos.
		rep.EndPos = rep.SkipPos
	}
	return rep
}

func (reporter *Reporter) ContainsCrash(output []byte) bool {
	return reporter.impl.ContainsCrash(output)
}

func (reporter *Reporter) Symbolize(rep *Report) error {
	return reporter.impl.Symbolize(rep)
}

func extractReportType(rep *Report) Type {
	// Type/frame extraction logic should be integrated with oops types.
	// But for now we do this more ad-hoc analysis here to at least isolate
	// the rest of the code base from report parsing.
	if rep.Title == unexpectedKernelReboot {
		return UnexpectedReboot
	}
	if strings.HasPrefix(rep.Title, memoryLeakPrefix) {
		return MemoryLeak
	}
	if strings.HasPrefix(rep.Title, dataRacePrefix) {
		return DataRace
	}
	if strings.HasPrefix(rep.Title, "INFO: rcu detected stall") ||
		strings.HasPrefix(rep.Title, "INFO: task hung") ||
		strings.HasPrefix(rep.Title, "BUG: soft lockup") ||
		strings.HasPrefix(rep.Title, "INFO: task can't die") {
		return Hang
	}
	return Unknown
}

func IsSuppressed(reporter *Reporter, output []byte) bool {
	return matchesAny(output, reporter.suppressions) ||
		bytes.Contains(output, gceConsoleHangup)
}

// ParseAll returns all successive reports in output.
func ParseAll(reporter *Reporter, output []byte) (reports []*Report) {
	skipPos := 0
	for {
		rep := reporter.ParseFrom(output, skipPos)
		if rep == nil {
			return
		}
		reports = append(reports, rep)
		skipPos = rep.SkipPos
	}
}

// GCE console connection sometimes fails with this message.
// The message frequently happens right after a kernel panic.
// So if we see it in output where we recognized a crash, we mark the report as corrupted
// because the crash message is usually truncated (maybe we don't even have the title line).
// If we see it in no output/lost connection reports then we mark them as suppressed instead
// because the crash itself may have been caused by the console connection error.
var gceConsoleHangup = []byte("serialport: VM disconnected.")

type replacement struct {
	match       *regexp.Regexp
	replacement string
}

func replaceTable(replacements []replacement, str string) string {
	for _, repl := range replacements {
		for stop := false; !stop; {
			newStr := repl.match.ReplaceAllString(str, repl.replacement)
			stop = newStr == str
			str = newStr
		}
	}
	return str
}

var dynamicTitleReplacement = []replacement{
	{
		// Executor PIDs are not interesting.
		regexp.MustCompile(`syz-executor\.?[0-9]+((/|:)[0-9]+)?`),
		"syz-executor",
	},
	{
		// Executor process IDs are dynamic and are not interesting.
		regexp.MustCompile(`syzkaller[0-9]+((/|:)[0-9]+)?`),
		"syzkaller",
	},
	{
		// Replace that everything looks like an address with "ADDR",
		// addresses in descriptions can't be good regardless of the oops regexps.
		regexp.MustCompile(`([^a-zA-Z0])(?:0x)?[0-9a-f]{6,}`),
		"${1}ADDR",
	},
	{
		// Replace IP addresses.
		regexp.MustCompile(`([0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3})`),
		"IP",
	},
	{
		// Replace that everything looks like a file line number with "LINE".
		regexp.MustCompile(`(\.\w+)(:[0-9]+)+`),
		"${1}:LINE",
	},
	{
		// Replace all raw references to runctions (e.g. "ip6_fragment+0x1052/0x2d80")
		// with just function name ("ip6_fragment"). Offsets and sizes are not stable.
		regexp.MustCompile(`([a-zA-Z][a-zA-Z0-9_.]+)\+0x[0-9a-z]+/0x[0-9a-z]+`),
		"${1}",
	},
	{
		// CPU numbers are not interesting.
		regexp.MustCompile(`CPU#[0-9]+`),
		"CPU",
	},
	{
		// Replace with "NUM" everything that looks like a decimal number and has not
		// been replaced yet. It might require multiple replacement executions as the
		// matching substrings may overlap (e.g. "0,1,2").
		regexp.MustCompile(`(\W)(\d+)(\W|$)`),
		"${1}NUM${3}",
	},
}

func sanitizeTitle(title string) string {
	const maxTitleLen = 120 // Corrupted/intermixed lines can be very long.
	res := make([]byte, 0, len(title))
	prev := byte(' ')
	for i := 0; i < len(title) && i < maxTitleLen; i++ {
		ch := title[i]
		switch {
		case ch == '\t':
			ch = ' '
		case ch < 0x20 || ch >= 0x7f:
			continue
		}
		if ch == ' ' && prev == ' ' {
			continue
		}
		res = append(res, ch)
		prev = ch
	}
	return strings.TrimSpace(string(res))
}

type oops struct {
	header       []byte
	formats      []oopsFormat
	suppressions []*regexp.Regexp
}

type oopsFormat struct {
	title *regexp.Regexp
	// If title is matched but report is not, the report is considered corrupted.
	report *regexp.Regexp
	// Format string to create report title.
	// Strings captured by title (or by report if present) are passed as input.
	// If stack is not nil, extracted function name is passed as an additional last argument.
	fmt string
	// Alternative titles used for better crash deduplication.
	// Format is the same as for fmt.
	alt []string
	// If not nil, a function name is extracted from the report and passed to fmt.
	// If not nil but frame extraction fails, the report is considered corrupted.
	stack        *stackFmt
	noStackTrace bool
	corrupted    bool
}

type stackFmt struct {
	// parts describe how guilty stack frame must be extracted from the report.
	// parts are matched consecutively potentially capturing frames.
	// parts can be of 3 types:
	//  - non-capturing regexp, matched against report and advances current position
	//  - capturing regexp, same as above, but also yields a frame
	//  - special value parseStackTrace means that a stack trace must be parsed
	//    starting from current position
	parts []*regexp.Regexp
	// If parts2 is present it is tried when parts matching fails.
	parts2 []*regexp.Regexp
	// Skip these functions in stack traces (matched as substring).
	skip []string
	// Custom frame extractor (optional).
	// Accepts set of all frames, returns guilty frame and corruption reason.
	extractor frameExtractor
}

type frameExtractor func(frames []string) (frame, corrupted string)

var parseStackTrace *regexp.Regexp

func compile(re string) *regexp.Regexp {
	re = strings.Replace(re, "{{ADDR}}", "0x[0-9a-f]+", -1)
	re = strings.Replace(re, "{{PC}}", "\\[\\<?(?:0x)?[0-9a-f]+\\>?\\]", -1)
	re = strings.Replace(re, "{{FUNC}}", "([a-zA-Z0-9_]+)(?:\\.|\\+)", -1)
	re = strings.Replace(re, "{{SRC}}", "([a-zA-Z0-9-_/.]+\\.[a-z]+:[0-9]+)", -1)
	return regexp.MustCompile(re)
}

func containsCrash(output []byte, oopses []*oops, ignores []*regexp.Regexp) bool {
	for pos := 0; pos < len(output); {
		next := bytes.IndexByte(output[pos:], '\n')
		if next != -1 {
			next += pos
		} else {
			next = len(output)
		}
		for _, oops := range oopses {
			if matchOops(output[pos:next], oops, ignores) {
				return true
			}
		}
		pos = next + 1
	}
	return false
}

func matchOops(line []byte, oops *oops, ignores []*regexp.Regexp) bool {
	match := bytes.Index(line, oops.header)
	if match == -1 {
		return false
	}
	if matchesAny(line, oops.suppressions) {
		return false
	}
	if matchesAny(line, ignores) {
		return false
	}
	return true
}

func extractDescription(output []byte, oops *oops, params *stackParams) (
	desc, corrupted string, altTitles []string, format oopsFormat) {
	startPos := len(output)
	matchedTitle := false
	for _, f := range oops.formats {
		match := f.title.FindSubmatchIndex(output)
		if match == nil || match[0] > startPos {
			continue
		}
		if match[0] == startPos && desc != "" {
			continue
		}
		if match[0] < startPos {
			desc = ""
			format = oopsFormat{}
			startPos = match[0]
		}
		matchedTitle = true
		if f.report != nil {
			match = f.report.FindSubmatchIndex(output)
			if match == nil {
				continue
			}
		}
		var args []interface{}
		for i := 2; i < len(match); i += 2 {
			args = append(args, string(output[match[i]:match[i+1]]))
		}
		corrupted = ""
		if f.stack != nil {
			frame := ""
			frame, corrupted = extractStackFrame(params, f.stack, output[match[0]:])
			if frame == "" {
				frame = "corrupted"
				if corrupted == "" {
					corrupted = "extracted no stack frame"
				}
			}
			args = append(args, frame)
		}
		desc = fmt.Sprintf(f.fmt, args...)
		for _, alt := range f.alt {
			altTitles = append(altTitles, fmt.Sprintf(alt, args...))
		}
		format = f
	}
	if desc == "" {
		// If we are here and matchedTitle is set, it means that we've matched
		// a title of an oops but not full report regexp or stack trace,
		// which means the report was corrupted.
		if matchedTitle {
			corrupted = "matched title but not report regexp"
		}
		pos := bytes.Index(output, oops.header)
		if pos == -1 {
			return
		}
		end := bytes.IndexByte(output[pos:], '\n')
		if end == -1 {
			end = len(output)
		} else {
			end += pos
		}
		desc = string(output[pos:end])
	}
	if corrupted == "" && format.corrupted {
		corrupted = "report format is marked as corrupted"
	}
	return
}

type stackParams struct {
	// stackStartRes matches start of stack traces.
	stackStartRes []*regexp.Regexp
	// frameRes match different formats of lines containing kernel frames (capture function name).
	frameRes []*regexp.Regexp
	// skipPatterns match functions that must be unconditionally skipped.
	skipPatterns []string
	// If we looked at any lines that match corruptedLines during report analysis,
	// then the report is marked as corrupted.
	corruptedLines []*regexp.Regexp
	// Prefixes that need to be removed from frames.
	// E.g. syscall prefixes as different arches have different prefixes.
	stripFramePrefixes []string
}

func extractStackFrame(params *stackParams, stack *stackFmt, output []byte) (string, string) {
	skip := append([]string{}, params.skipPatterns...)
	skip = append(skip, stack.skip...)
	var skipRe *regexp.Regexp
	if len(skip) != 0 {
		skipRe = regexp.MustCompile(strings.Join(skip, "|"))
	}
	extractor := stack.extractor
	if extractor == nil {
		extractor = func(frames []string) (string, string) {
			return frames[0], ""
		}
	}
	frame, corrupted := extractStackFrameImpl(params, output, skipRe, stack.parts, extractor)
	if frame != "" || len(stack.parts2) == 0 {
		return frame, corrupted
	}
	return extractStackFrameImpl(params, output, skipRe, stack.parts2, extractor)
}

func extractStackFrameImpl(params *stackParams, output []byte, skipRe *regexp.Regexp,
	parts []*regexp.Regexp, extractor frameExtractor) (string, string) {
	s := bufio.NewScanner(bytes.NewReader(output))
	var frames []string
nextPart:
	for _, part := range parts {
		if part == parseStackTrace {
			for s.Scan() {
				ln := s.Bytes()
				if matchesAny(ln, params.corruptedLines) {
					break nextPart
				}
				if matchesAny(ln, params.stackStartRes) {
					continue nextPart
				}
				var match [][]byte
				for _, re := range params.frameRes {
					match = re.FindSubmatch(ln)
					if match != nil {
						break
					}
				}
				frames = appendStackFrame(frames, match, params, skipRe)
			}
		} else {
			for s.Scan() {
				ln := s.Bytes()
				if matchesAny(ln, params.corruptedLines) {
					break nextPart
				}
				match := part.FindSubmatch(ln)
				if match == nil {
					continue
				}
				frames = appendStackFrame(frames, match, params, skipRe)
				break
			}
		}
	}
	if len(frames) == 0 {
		return "", corruptedNoFrames
	}
	return extractor(frames)
}

func appendStackFrame(frames []string, match [][]byte, params *stackParams, skipRe *regexp.Regexp) []string {
	if len(match) < 2 {
		return frames
	}
	for _, frame := range match[1:] {
		if frame != nil && (skipRe == nil || !skipRe.Match(frame)) {
			frameName := string(frame)
			for _, prefix := range params.stripFramePrefixes {
				frameName = strings.TrimPrefix(frameName, prefix)
			}
			frames = append(frames, frameName)
			break
		}
	}
	return frames
}

func simpleLineParser(output []byte, oopses []*oops, params *stackParams, ignores []*regexp.Regexp) *Report {
	rep := &Report{
		Output: output,
	}
	var oops *oops
	for pos := 0; pos < len(output); {
		next := bytes.IndexByte(output[pos:], '\n')
		if next != -1 {
			next += pos
		} else {
			next = len(output)
		}
		line := output[pos:next]
		for _, oops1 := range oopses {
			if matchOops(line, oops1, ignores) {
				oops = oops1
				rep.StartPos = pos
				rep.EndPos = next
				break
			}
		}
		if oops != nil {
			break
		}
		pos = next + 1
	}
	if oops == nil {
		return nil
	}
	title, corrupted, altTitles, _ := extractDescription(output[rep.StartPos:], oops, params)
	rep.Title = title
	rep.AltTitles = altTitles
	rep.Report = output[rep.StartPos:]
	rep.Corrupted = corrupted != ""
	rep.CorruptedReason = corrupted
	return rep
}

func matchesAny(line []byte, res []*regexp.Regexp) bool {
	for _, re := range res {
		if re.Match(line) {
			return true
		}
	}
	return false
}

// replace replaces [start:end] in where with what, inplace.
func replace(where []byte, start, end int, what []byte) []byte {
	if len(what) >= end-start {
		where = append(where, what[end-start:]...)
		copy(where[start+len(what):], where[end:])
		copy(where[start:], what)
	} else {
		copy(where[start+len(what):], where[end:])
		where = where[:len(where)-(end-start-len(what))]
		copy(where[start:], what)
	}
	return where
}

var (
	filenameRe    = regexp.MustCompile(`([a-zA-Z0-9_\-\./]*[a-zA-Z0-9_\-]+\.(c|h)):[0-9]+`)
	reportFrameRe = regexp.MustCompile(`.* in ([a-zA-Z0-9_]+)`)
)

// These are produced by syzkaller itself.
// But also catches crashes in Go programs in gvisor/fuchsia.
var commonOopses = []*oops{
	{
		// Errors produced by executor's fail function.
		[]byte("SYZFAIL:"),
		[]oopsFormat{
			{
				title:        compile("SYZFAIL:(.*)"),
				fmt:          "SYZFAIL:%[1]v",
				noStackTrace: true,
			},
		},
		[]*regexp.Regexp{},
	},
	{
		[]byte("panic:"),
		[]oopsFormat{
			{
				// This is gvisor-specific, but we need to handle it here since we handle "panic:" here.
				title:        compile("panic: Sentry detected .* stuck task"),
				fmt:          "panic: Sentry detected stuck tasks",
				noStackTrace: true,
			},
			{
				title:        compile("panic:(.*)"),
				fmt:          "panic:%[1]v",
				noStackTrace: true,
			},
		},
		[]*regexp.Regexp{
			// This can match some kernel functions (skb_panic, skb_over_panic).
			compile("_panic:"),
			// Android prints this sometimes during boot.
			compile("xlog_status:"),
			compile(`ddb\.onpanic:`),
			compile(`evtlog_status:`),
		},
	},
}