// Copyright 2015/2016 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 prog import ( "bytes" "fmt" "math" "math/rand" "path/filepath" "sort" "strings" "github.com/google/syzkaller/pkg/ifuzz" _ "github.com/google/syzkaller/pkg/ifuzz/generated" // pull in generated instruction descriptions ) const ( // "Recommended" number of calls in programs that we try to aim at during fuzzing. RecommendedCalls = 20 // "Recommended" max number of calls in programs. // If we receive longer programs from hub/corpus we discard them. MaxCalls = 40 ) type randGen struct { *rand.Rand target *Target inCreateResource bool recDepth map[string]int } func newRand(target *Target, rs rand.Source) *randGen { return &randGen{ Rand: rand.New(rs), target: target, recDepth: make(map[string]int), } } func (r *randGen) rand(n int) uint64 { return uint64(r.Intn(n)) } func (r *randGen) randRange(begin, end uint64) uint64 { return begin + uint64(r.Intn(int(end-begin+1))) } func (r *randGen) bin() bool { return r.Intn(2) == 0 } func (r *randGen) oneOf(n int) bool { return r.Intn(n) == 0 } func (r *randGen) rand64() uint64 { v := uint64(r.Int63()) if r.bin() { v |= 1 << 63 } return v } var ( // Some potentially interesting integers. specialInts = []uint64{ 0, 1, 31, 32, 63, 64, 127, 128, 129, 255, 256, 257, 511, 512, 1023, 1024, 1025, 2047, 2048, 4095, 4096, (1 << 15) - 1, (1 << 15), (1 << 15) + 1, (1 << 16) - 1, (1 << 16), (1 << 16) + 1, (1 << 31) - 1, (1 << 31), (1 << 31) + 1, (1 << 32) - 1, (1 << 32), (1 << 32) + 1, } // The indexes (exclusive) for the maximum specialInts values that fit in 1, 2, ... 8 bytes. specialIntIndex [9]int ) func init() { sort.Slice(specialInts, func(i, j int) bool { return specialInts[i] < specialInts[j] }) for i := range specialIntIndex { bitSize := uint64(8 * i) specialIntIndex[i] = sort.Search(len(specialInts), func(i int) bool { return specialInts[i]>>bitSize != 0 }) } } func (r *randGen) randInt64() uint64 { return r.randInt(64) } func (r *randGen) randInt(bits uint64) uint64 { v := r.rand64() switch { case r.nOutOf(100, 182): v %= 10 case bits >= 8 && r.nOutOf(50, 82): v = specialInts[r.Intn(specialIntIndex[bits/8])] case r.nOutOf(10, 32): v %= 256 case r.nOutOf(10, 22): v %= 4 << 10 case r.nOutOf(10, 12): v %= 64 << 10 default: v %= 1 << 31 } switch { case r.nOutOf(100, 107): case r.nOutOf(5, 7): v = uint64(-int64(v)) default: v <<= uint(r.Intn(int(bits))) } return truncateToBitSize(v, bits) } func truncateToBitSize(v, bitSize uint64) uint64 { if bitSize == 0 || bitSize > 64 { panic(fmt.Sprintf("invalid bitSize value: %d", bitSize)) } return v & uint64(1<>= 1 } else { flag <<= 1 } } v ^= flag } return v } func (r *randGen) filename(s *state, typ *BufferType) string { fn := r.filenameImpl(s) if fn != "" && fn[len(fn)-1] == 0 { panic(fmt.Sprintf("zero-terminated filename: %q", fn)) } if escapingFilename(fn) { panic(fmt.Sprintf("sandbox escaping file name %q, s.files are %v", fn, s.files)) } if !typ.Varlen() { size := typ.Size() if uint64(len(fn)) < size { fn += string(make([]byte, size-uint64(len(fn)))) } fn = fn[:size] } else if !typ.NoZ { fn += "\x00" } return fn } func escapingFilename(file string) bool { file = filepath.Clean(file) return len(file) >= 1 && file[0] == '/' || len(file) >= 2 && file[0] == '.' && file[1] == '.' } var specialFiles = []string{"", "."} func (r *randGen) filenameImpl(s *state) string { if r.oneOf(100) { return specialFiles[r.Intn(len(specialFiles))] } if len(s.files) == 0 || r.oneOf(10) { // Generate a new name. dir := "." if r.oneOf(2) && len(s.files) != 0 { dir = r.randFromMap(s.files) if dir != "" && dir[len(dir)-1] == 0 { dir = dir[:len(dir)-1] } if r.oneOf(10) && filepath.Clean(dir)[0] != '.' { dir += "/.." } } for i := 0; ; i++ { f := fmt.Sprintf("%v/file%v", dir, i) if !s.files[f] { return f } } } return r.randFromMap(s.files) } func (r *randGen) randFromMap(m map[string]bool) string { files := make([]string, 0, len(m)) for f := range m { files = append(files, f) } sort.Strings(files) return files[r.Intn(len(files))] } func (r *randGen) randString(s *state, t *BufferType) []byte { if len(t.Values) != 0 { return []byte(t.Values[r.Intn(len(t.Values))]) } if len(s.strings) != 0 && r.bin() { // Return an existing string. // TODO(dvyukov): make s.strings indexed by string SubKind. return []byte(r.randFromMap(s.strings)) } punct := []byte{'!', '@', '#', '$', '%', '^', '&', '*', '(', ')', '-', '+', '\\', '/', ':', '.', ',', '-', '\'', '[', ']', '{', '}'} buf := new(bytes.Buffer) for r.nOutOf(3, 4) { if r.nOutOf(10, 11) { buf.Write([]byte{punct[r.Intn(len(punct))]}) } else { buf.Write([]byte{byte(r.Intn(256))}) } } if r.oneOf(100) == t.NoZ { buf.Write([]byte{0}) } return buf.Bytes() } func (r *randGen) allocAddr(s *state, typ Type, dir Dir, size uint64, data Arg) *PointerArg { return MakePointerArg(typ, dir, s.ma.alloc(r, size), data) } func (r *randGen) allocVMA(s *state, typ Type, dir Dir, numPages uint64) *PointerArg { page := s.va.alloc(r, numPages) return MakeVmaPointerArg(typ, dir, page*r.target.PageSize, numPages*r.target.PageSize) } func (r *randGen) createResource(s *state, res *ResourceType, dir Dir) (arg Arg, calls []*Call) { if r.inCreateResource { return nil, nil } r.inCreateResource = true defer func() { r.inCreateResource = false }() kind := res.Desc.Name // We may have no resources, but still be in createResource due to ANYRES. if len(r.target.resourceMap) != 0 && r.oneOf(1000) { // Spoof resource subkind. var all []string for kind1 := range r.target.resourceMap { if r.target.isCompatibleResource(res.Desc.Kind[0], kind1) { all = append(all, kind1) } } if len(all) == 0 { panic(fmt.Sprintf("got no spoof resources for %v in %v/%v", kind, r.target.OS, r.target.Arch)) } sort.Strings(all) kind = all[r.Intn(len(all))] } // Find calls that produce the necessary resources. metas0 := r.target.resourceCtors[kind] // TODO: reduce priority of less specialized ctors. var metas []*Syscall for _, meta := range metas0 { if s.ct.Enabled(meta.ID) { metas = append(metas, meta) } } if len(metas) == 0 { return res.DefaultArg(dir), nil } // Now we have a set of candidate calls that can create the necessary resource. for i := 0; i < 1e3; i++ { // Generate one of them. meta := metas[r.Intn(len(metas))] calls := r.generateParticularCall(s, meta) s1 := newState(r.target, s.ct, nil) s1.analyze(calls[len(calls)-1]) // Now see if we have what we want. var allres []*ResultArg for kind1, res1 := range s1.resources { if r.target.isCompatibleResource(kind, kind1) { allres = append(allres, res1...) } } sort.SliceStable(allres, func(i, j int) bool { return allres[i].Type().Name() < allres[j].Type().Name() }) if len(allres) != 0 { // Bingo! arg := MakeResultArg(res, dir, allres[r.Intn(len(allres))], 0) return arg, calls } // Discard unsuccessful calls. // Note: s.ma/va have already noted allocations of the new objects // in discarded syscalls, ideally we should recreate state // by analyzing the program again. for _, c := range calls { ForeachArg(c, func(arg Arg, _ *ArgCtx) { if a, ok := arg.(*ResultArg); ok && a.Res != nil { delete(a.Res.uses, a) } }) } } // Generally we can loop several times, e.g. when we choose a call that returns // the resource in an array, but then generateArg generated that array of zero length. // But we must succeed eventually. var ctors []string for _, meta := range metas { ctors = append(ctors, meta.Name) } panic(fmt.Sprintf("failed to create a resource %v with %v", res.Desc.Kind[0], strings.Join(ctors, ", "))) } func (r *randGen) generateText(kind TextKind) []byte { switch kind { case TextTarget: if cfg := createTargetIfuzzConfig(r.target); cfg != nil { return ifuzz.Generate(cfg, r.Rand) } fallthrough case TextArm64: // Just a stub, need something better. text := make([]byte, 50) for i := range text { text[i] = byte(r.Intn(256)) } return text default: cfg := createIfuzzConfig(kind) return ifuzz.Generate(cfg, r.Rand) } } func (r *randGen) mutateText(kind TextKind, text []byte) []byte { switch kind { case TextTarget: if cfg := createTargetIfuzzConfig(r.target); cfg != nil { return ifuzz.Mutate(cfg, r.Rand, text) } fallthrough case TextArm64: return mutateData(r, text, 40, 60) default: cfg := createIfuzzConfig(kind) return ifuzz.Mutate(cfg, r.Rand, text) } } func createTargetIfuzzConfig(target *Target) *ifuzz.Config { cfg := &ifuzz.Config{ Len: 10, Priv: false, Exec: true, MemRegions: []ifuzz.MemRegion{ {Start: target.DataOffset, Size: target.NumPages * target.PageSize}, }, } for _, p := range target.SpecialPointers { cfg.MemRegions = append(cfg.MemRegions, ifuzz.MemRegion{ Start: p & ^target.PageSize, Size: p & ^target.PageSize + target.PageSize, }) } switch target.Arch { case "amd64": cfg.Mode = ifuzz.ModeLong64 case "386": cfg.Mode = ifuzz.ModeProt32 default: return nil } return cfg } func createIfuzzConfig(kind TextKind) *ifuzz.Config { cfg := &ifuzz.Config{ Len: 10, Priv: true, Exec: true, MemRegions: []ifuzz.MemRegion{ {Start: 0 << 12, Size: 1 << 12}, {Start: 1 << 12, Size: 1 << 12}, {Start: 2 << 12, Size: 1 << 12}, {Start: 3 << 12, Size: 1 << 12}, {Start: 4 << 12, Size: 1 << 12}, {Start: 5 << 12, Size: 1 << 12}, {Start: 6 << 12, Size: 1 << 12}, {Start: 7 << 12, Size: 1 << 12}, {Start: 8 << 12, Size: 1 << 12}, {Start: 9 << 12, Size: 1 << 12}, {Start: 0xfec00000, Size: 0x100}, // ioapic }, } switch kind { case TextX86Real: cfg.Mode = ifuzz.ModeReal16 case TextX86bit16: cfg.Mode = ifuzz.ModeProt16 case TextX86bit32: cfg.Mode = ifuzz.ModeProt32 case TextX86bit64: cfg.Mode = ifuzz.ModeLong64 default: panic("unknown text kind") } return cfg } // nOutOf returns true n out of outOf times. func (r *randGen) nOutOf(n, outOf int) bool { if n <= 0 || n >= outOf { panic("bad probability") } v := r.Intn(outOf) return v < n } func (r *randGen) generateCall(s *state, p *Prog, insertionPoint int) []*Call { biasCall := -1 if insertionPoint > 0 { // Choosing the base call is based on the insertion point of the new calls sequence. biasCall = p.Calls[r.Intn(insertionPoint)].Meta.ID } idx := s.ct.choose(r.Rand, biasCall) meta := r.target.Syscalls[idx] return r.generateParticularCall(s, meta) } func (r *randGen) generateParticularCall(s *state, meta *Syscall) (calls []*Call) { if meta.Attrs.Disabled { panic(fmt.Sprintf("generating disabled call %v", meta.Name)) } c := &Call{ Meta: meta, Ret: MakeReturnArg(meta.Ret), } c.Args, calls = r.generateArgs(s, meta.Args, DirIn) r.target.assignSizesCall(c) return append(calls, c) } // GenerateAllSyzProg generates a program that contains all pseudo syz_ calls for testing. func (target *Target) GenerateAllSyzProg(rs rand.Source) *Prog { p := &Prog{ Target: target, } r := newRand(target, rs) s := newState(target, target.DefaultChoiceTable(), nil) handled := make(map[string]bool) for _, meta := range target.Syscalls { if !strings.HasPrefix(meta.CallName, "syz_") || handled[meta.CallName] || meta.Attrs.Disabled { continue } handled[meta.CallName] = true calls := r.generateParticularCall(s, meta) for _, c := range calls { s.analyze(c) p.Calls = append(p.Calls, c) } } if err := p.validate(); err != nil { panic(err) } return p } // DataMmapProg creates program that maps data segment. // Also used for testing as the simplest program. func (target *Target) DataMmapProg() *Prog { return &Prog{ Target: target, Calls: target.MakeDataMmap(), } } func (r *randGen) generateArgs(s *state, fields []Field, dir Dir) ([]Arg, []*Call) { var calls []*Call args := make([]Arg, len(fields)) // Generate all args. Size args have the default value 0 for now. for i, field := range fields { arg, calls1 := r.generateArg(s, field.Type, dir) if arg == nil { panic(fmt.Sprintf("generated arg is nil for field '%v', fields: %+v", field.Type.Name(), fields)) } args[i] = arg calls = append(calls, calls1...) } return args, calls } func (r *randGen) generateArg(s *state, typ Type, dir Dir) (arg Arg, calls []*Call) { return r.generateArgImpl(s, typ, dir, false) } func (r *randGen) generateArgImpl(s *state, typ Type, dir Dir, ignoreSpecial bool) (arg Arg, calls []*Call) { if dir == DirOut { // No need to generate something interesting for output scalar arguments. // But we still need to generate the argument itself so that it can be referenced // in subsequent calls. For the same reason we do generate pointer/array/struct // output arguments (their elements can be referenced in subsequent calls). switch typ.(type) { case *IntType, *FlagsType, *ConstType, *ProcType, *VmaType, *ResourceType: return typ.DefaultArg(dir), nil } } if typ.Optional() && r.oneOf(5) { if res, ok := typ.(*ResourceType); ok { v := res.Desc.Values[r.Intn(len(res.Desc.Values))] return MakeResultArg(typ, dir, nil, v), nil } return typ.DefaultArg(dir), nil } // Allow infinite recursion for optional pointers. if pt, ok := typ.(*PtrType); ok && typ.Optional() { switch pt.Elem.(type) { case *StructType, *ArrayType, *UnionType: name := pt.Elem.Name() r.recDepth[name]++ defer func() { r.recDepth[name]-- if r.recDepth[name] == 0 { delete(r.recDepth, name) } }() if r.recDepth[name] >= 3 { return MakeSpecialPointerArg(typ, dir, 0), nil } } } if !ignoreSpecial && dir != DirOut { switch typ.(type) { case *StructType, *UnionType: if gen := r.target.SpecialTypes[typ.Name()]; gen != nil { return gen(&Gen{r, s}, typ, dir, nil) } } } return typ.generate(r, s, dir) } func (a *ResourceType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { if r.oneOf(3) { arg = r.existingResource(s, a, dir) if arg != nil { return } } if r.nOutOf(2, 3) { arg, calls = r.resourceCentric(s, a, dir) if arg != nil { return } } if r.nOutOf(4, 5) { arg, calls = r.createResource(s, a, dir) if arg != nil { return } } special := a.SpecialValues() arg = MakeResultArg(a, dir, nil, special[r.Intn(len(special))]) return } func (a *BufferType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { switch a.Kind { case BufferBlobRand, BufferBlobRange: sz := r.randBufLen() if a.Kind == BufferBlobRange { sz = r.randRange(a.RangeBegin, a.RangeEnd) } if dir == DirOut { return MakeOutDataArg(a, dir, sz), nil } data := make([]byte, sz) for i := range data { data[i] = byte(r.Intn(256)) } return MakeDataArg(a, dir, data), nil case BufferString: data := r.randString(s, a) if dir == DirOut { return MakeOutDataArg(a, dir, uint64(len(data))), nil } return MakeDataArg(a, dir, data), nil case BufferFilename: if dir == DirOut { var sz uint64 switch { case !a.Varlen(): sz = a.Size() case r.nOutOf(1, 3): sz = r.rand(100) case r.nOutOf(1, 2): sz = 108 // UNIX_PATH_MAX default: sz = 4096 // PATH_MAX } return MakeOutDataArg(a, dir, sz), nil } return MakeDataArg(a, dir, []byte(r.filename(s, a))), nil case BufferText: if dir == DirOut { return MakeOutDataArg(a, dir, uint64(r.Intn(100))), nil } return MakeDataArg(a, dir, r.generateText(a.Text)), nil default: panic("unknown buffer kind") } } func (a *VmaType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { npages := r.randPageCount() if a.RangeBegin != 0 || a.RangeEnd != 0 { npages = a.RangeBegin + uint64(r.Intn(int(a.RangeEnd-a.RangeBegin+1))) } return r.allocVMA(s, a, dir, npages), nil } func (a *FlagsType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { return MakeConstArg(a, dir, r.flags(a.Vals, a.BitMask, 0)), nil } func (a *ConstType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { return MakeConstArg(a, dir, a.Val), nil } func (a *IntType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { bits := a.TypeBitSize() v := r.randInt(bits) switch a.Kind { case IntRange: v = r.randRangeInt(a.RangeBegin, a.RangeEnd, bits, a.Align) } return MakeConstArg(a, dir, v), nil } func (a *ProcType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { return MakeConstArg(a, dir, r.rand(int(a.ValuesPerProc))), nil } func (a *ArrayType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { var count uint64 switch a.Kind { case ArrayRandLen: count = r.randArrayLen() case ArrayRangeLen: count = r.randRange(a.RangeBegin, a.RangeEnd) } var inner []Arg for i := uint64(0); i < count; i++ { arg1, calls1 := r.generateArg(s, a.Elem, dir) inner = append(inner, arg1) calls = append(calls, calls1...) } return MakeGroupArg(a, dir, inner), calls } func (a *StructType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { args, calls := r.generateArgs(s, a.Fields, dir) group := MakeGroupArg(a, dir, args) return group, calls } func (a *UnionType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { index := r.Intn(len(a.Fields)) optType := a.Fields[index].Type opt, calls := r.generateArg(s, optType, dir) return MakeUnionArg(a, dir, opt, index), calls } func (a *PtrType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { if r.oneOf(1000) { index := r.rand(len(r.target.SpecialPointers)) return MakeSpecialPointerArg(a, dir, index), nil } inner, calls := r.generateArg(s, a.Elem, a.ElemDir) arg = r.allocAddr(s, a, dir, inner.Size(), inner) return arg, calls } func (a *LenType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { // Updated later in assignSizesCall. return MakeConstArg(a, dir, 0), nil } func (a *CsumType) generate(r *randGen, s *state, dir Dir) (arg Arg, calls []*Call) { // Filled at runtime by executor. return MakeConstArg(a, dir, 0), nil } func (r *randGen) existingResource(s *state, res *ResourceType, dir Dir) Arg { alltypes := make([][]*ResultArg, 0, len(s.resources)) for _, res1 := range s.resources { alltypes = append(alltypes, res1) } sort.Slice(alltypes, func(i, j int) bool { return alltypes[i][0].Type().Name() < alltypes[j][0].Type().Name() }) var allres []*ResultArg for _, res1 := range alltypes { name1 := res1[0].Type().Name() if r.target.isCompatibleResource(res.Desc.Name, name1) || r.oneOf(50) && r.target.isCompatibleResource(res.Desc.Kind[0], name1) { allres = append(allres, res1...) } } if len(allres) == 0 { return nil } return MakeResultArg(res, dir, allres[r.Intn(len(allres))], 0) } // Finds a compatible resource with the type `t` and the calls that initialize that resource. func (r *randGen) resourceCentric(s *state, t *ResourceType, dir Dir) (arg Arg, calls []*Call) { var p *Prog var resource *ResultArg for idx := range r.Perm(len(s.corpus)) { p = s.corpus[idx].Clone() resources := getCompatibleResources(p, t.TypeName, r) if len(resources) > 0 { resource = resources[r.Intn(len(resources))] break } } // No compatible resource was found. if resource == nil { return nil, nil } // Set that stores the resources that appear in the same calls with the selected resource. relatedRes := map[*ResultArg]bool{resource: true} // Remove unrelated calls from the program. for idx := len(p.Calls) - 1; idx >= 0; idx-- { includeCall := false var newResources []*ResultArg ForeachArg(p.Calls[idx], func(arg Arg, _ *ArgCtx) { if a, ok := arg.(*ResultArg); ok { if a.Res != nil && !relatedRes[a.Res] { newResources = append(newResources, a.Res) } if relatedRes[a] || relatedRes[a.Res] { includeCall = true } } }) if !includeCall { p.removeCall(idx) } else { for _, res := range newResources { relatedRes[res] = true } } } // Selects a biased random length of the returned calls (more calls could offer more // interesting programs). The values returned (n = len(calls): n, n-1, ..., 2. biasedLen := 2 + r.biasedRand(len(calls)-1, 10) // Removes the references that are not used anymore. for i := biasedLen; i < len(calls); i++ { p.removeCall(i) } return MakeResultArg(t, dir, resource, 0), p.Calls } func getCompatibleResources(p *Prog, resourceType string, r *randGen) (resources []*ResultArg) { for _, c := range p.Calls { ForeachArg(c, func(arg Arg, _ *ArgCtx) { // Collect only initialized resources (the ones that are already used in other calls). a, ok := arg.(*ResultArg) if !ok || len(a.uses) == 0 || a.Dir() != DirOut { return } if !r.target.isCompatibleResource(resourceType, a.Type().Name()) { return } resources = append(resources, a) }) } return resources }