| 1 | //go:build linux |
| 2 | |
| 3 | package host |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "strconv" |
| 10 | "strings" |
| 11 | |
| 12 | "github.com/carlosframework/platform/internal/release" |
| 13 | "github.com/carlosframework/platform/internal/rootstate" |
| 14 | "golang.org/x/sys/unix" |
| 15 | ) |
| 16 | |
| 17 | func generationIdentity(path string) (rootstate.FileIdentity, error) { |
| 18 | var st unix.Stat_t |
| 19 | if err := unix.Stat(path, &st); err != nil { |
| 20 | return rootstate.FileIdentity{}, err |
| 21 | } |
| 22 | return rootstate.FileIdentity{Device: uint64(st.Dev), Inode: st.Ino}, nil |
| 23 | } |
| 24 | |
| 25 | func (k generationKernel) containedTask(ctx context.Context, path string, status []byte) (bool, error) { |
| 26 | raw, err := generationRead(path + "/cgroup") |
| 27 | if err != nil { |
| 28 | return false, err |
| 29 | } |
| 30 | group := strings.TrimSpace(string(raw)) |
| 31 | for _, r := range k.contained { |
| 32 | p := r.Process |
| 33 | if p.Containment != 1 || r.Closed || r.Retirement != "" || group != "0::/"+p.Group { |
| 34 | continue |
| 35 | } |
| 36 | if err := generationContainedStatus(status); err != nil { |
| 37 | return false, err |
| 38 | } |
| 39 | if err := k.containedBinding(ctx, path, r); err != nil { |
| 40 | return false, err |
| 41 | } |
| 42 | return true, nil |
| 43 | } |
| 44 | return false, nil |
| 45 | } |
| 46 | |
| 47 | func generationContainedStatus(raw []byte) error { |
| 48 | fields := map[string]string{} |
| 49 | for _, line := range strings.Split(string(raw), "\n") { |
| 50 | key, value, ok := strings.Cut(line, ":") |
| 51 | if ok { |
| 52 | if _, dup := fields[key]; dup { |
| 53 | return os.ErrPermission |
| 54 | } |
| 55 | fields[key] = strings.TrimSpace(value) |
| 56 | } |
| 57 | } |
| 58 | for _, key := range []string{"CapInh", "CapPrm", "CapEff", "CapBnd", "CapAmb"} { |
| 59 | n, err := strconv.ParseUint(fields[key], 16, 64) |
| 60 | if err != nil || n != 0 { |
| 61 | return os.ErrPermission |
| 62 | } |
| 63 | } |
| 64 | if fields["NoNewPrivs"] != "1" || fields["Seccomp"] != "2" { |
| 65 | return os.ErrPermission |
| 66 | } |
| 67 | return nil |
| 68 | } |
| 69 | |
| 70 | func (k generationKernel) containedBinding(ctx context.Context, path string, r release.GenerationReservation) error { |
| 71 | if err := ctx.Err(); err != nil { |
| 72 | return err |
| 73 | } |
| 74 | b, p := r.Binding, r.Process |
| 75 | boot, err := k.boot() |
| 76 | if err != nil || boot != b.Target.BootID || p.Group != generationBindingGroup(b) { |
| 77 | return os.ErrPermission |
| 78 | } |
| 79 | var group unix.Stat_t |
| 80 | if err := unix.Lstat(filepath.Join(k.cgroups, p.Group), &group); err != nil { |
| 81 | return err |
| 82 | } |
| 83 | if group.Uid != k.owner || group.Mode&0022 != 0 || group.Mode&unix.S_IFMT != unix.S_IFDIR || (rootstate.FileIdentity{Device: uint64(group.Dev), Inode: group.Ino}) != p.GroupIdentity { |
| 84 | return os.ErrPermission |
| 85 | } |
| 86 | main, tenant := filepath.Join(k.proc, strconv.Itoa(p.PID)), filepath.Join(k.proc, strconv.Itoa(p.TenantPID)) |
| 87 | for pid, want := range map[int]string{p.PID: p.Start, p.TenantPID: p.TenantStart} { |
| 88 | start, err := claimLaunchProcessStart(k.proc, pid) |
| 89 | if err != nil || start != want { |
| 90 | return os.ErrPermission |
| 91 | } |
| 92 | raw, err := generationRead(filepath.Join(k.proc, strconv.Itoa(pid), "cgroup")) |
| 93 | if err != nil || strings.TrimSpace(string(raw)) != "0::/"+p.Group { |
| 94 | return os.ErrPermission |
| 95 | } |
| 96 | } |
| 97 | launcher, err := generationIdentity(main + "/exe") |
| 98 | if err != nil || launcher != p.LauncherIdentity { |
| 99 | return os.ErrPermission |
| 100 | } |
| 101 | privatePID := "" |
| 102 | for _, ns := range []string{"mnt", "pid"} { |
| 103 | want, err := os.Readlink(tenant + "/ns/" + ns) |
| 104 | if err != nil || want == "" { |
| 105 | return os.ErrPermission |
| 106 | } |
| 107 | if ns == "pid" { |
| 108 | privatePID = want |
| 109 | } |
| 110 | outer, err := os.Readlink(k.proc + "/self/ns/" + ns) |
| 111 | if err != nil || outer == want { |
| 112 | return os.ErrPermission |
| 113 | } |
| 114 | for _, task := range []string{main, path} { |
| 115 | actual, err := os.Readlink(task + "/ns/" + ns) |
| 116 | if err != nil || actual != want { |
| 117 | return os.ErrPermission |
| 118 | } |
| 119 | } |
| 120 | } |
| 121 | bound := strings.TrimPrefix(k.boundApp, "/") |
| 122 | if bound == "" { |
| 123 | bound = "opt/carlos/app" |
| 124 | } |
| 125 | id, err := generationIdentity(filepath.Join(path, "root", bound)) |
| 126 | if err != nil || id != b.Executable { |
| 127 | return os.ErrPermission |
| 128 | } |
| 129 | tree := filepath.Dir(b.Tree) |
| 130 | if b.Binding.Account != "" && b.Binding.Account != release.OpsAccount { |
| 131 | tree = filepath.Dir(tree) |
| 132 | } |
| 133 | if filepath.Base(tree) != "versions" || b.Path != release.VersionPath(filepath.Dir(tree), b.Binding.Account, b.Version, b.Binding.App) { |
| 134 | return os.ErrPermission |
| 135 | } |
| 136 | // The canonical version store must be hidden, not merely read-only. |
| 137 | if _, err := os.Stat(filepath.Join(path, "root", tree)); !os.IsNotExist(err) { |
| 138 | return os.ErrPermission |
| 139 | } |
| 140 | raw, err := generationRead(path + "/mountinfo") |
| 141 | if err != nil { |
| 142 | return err |
| 143 | } |
| 144 | boundMount := false |
| 145 | for _, line := range strings.Split(strings.TrimSpace(string(raw)), "\n") { |
| 146 | f := strings.Fields(line) |
| 147 | if len(f) < 10 { |
| 148 | return os.ErrPermission |
| 149 | } |
| 150 | source, e1 := generationUnescape(f[3]) |
| 151 | point, e2 := generationUnescape(f[4]) |
| 152 | if e1 != nil || e2 != nil || !filepath.IsAbs(source) || !filepath.IsAbs(point) { |
| 153 | return os.ErrPermission |
| 154 | } |
| 155 | separator := -1 |
| 156 | for i := 6; i < len(f); i++ { |
| 157 | if f[i] == "-" { |
| 158 | separator = i |
| 159 | break |
| 160 | } |
| 161 | } |
| 162 | if separator < 6 || len(f) != separator+4 { |
| 163 | return os.ErrPermission |
| 164 | } |
| 165 | if f[separator+1] == "proc" { |
| 166 | if source == "/" { |
| 167 | visible, err := os.Readlink(filepath.Join(path, "root", point, "1/ns/pid")) |
| 168 | if err != nil || visible != privatePID { |
| 169 | return os.ErrPermission |
| 170 | } |
| 171 | } else if source != "/sys" && !strings.HasPrefix(source, "/sys/") { |
| 172 | return os.ErrPermission |
| 173 | } |
| 174 | } |
| 175 | major, minor, ok := strings.Cut(f[2], ":") |
| 176 | maj, e3 := strconv.ParseUint(major, 10, 32) |
| 177 | min, e4 := strconv.ParseUint(minor, 10, 32) |
| 178 | if !ok || e3 != nil || e4 != nil { |
| 179 | return os.ErrPermission |
| 180 | } |
| 181 | if unix.Mkdev(uint32(maj), uint32(min)) != b.Executable.Device { |
| 182 | continue |
| 183 | } |
| 184 | // The one reserved file is the only staged bind this namespace may hold. |
| 185 | if source == b.Path && point == "/"+bound { |
| 186 | boundMount = true |
| 187 | continue |
| 188 | } |
| 189 | below, e1 := filepath.Rel(tree, source) |
| 190 | above, e2 := filepath.Rel(source, tree) |
| 191 | if (e1 == nil && filepath.IsLocal(below)) || (e2 == nil && filepath.IsLocal(above) && point != "/") { |
| 192 | return os.ErrPermission |
| 193 | } |
| 194 | } |
| 195 | if !boundMount { |
| 196 | return os.ErrPermission |
| 197 | } |
| 198 | return nil |
| 199 | } |
| 200 | |
| 201 | func (k generationKernel) replacedGroup(ctx context.Context, old release.GenerationReservation) error { |
| 202 | if !old.Closed || !isolatedGeneration(old) || old.Process == (release.GenerationProcess{}) { |
| 203 | return os.ErrPermission |
| 204 | } |
| 205 | boot, err := k.boot() |
| 206 | if err != nil { |
| 207 | return err |
| 208 | } |
| 209 | for _, next := range k.contained { |
| 210 | if next.Closed || next.Retirement != "" || !isolatedGeneration(next) || next.Process.Containment != 1 || |
| 211 | next.Binding.Binding != old.Binding.Binding || next.Binding.Attempt == old.Binding.Attempt || |
| 212 | next.Process.Invocation == "" || next.Process.Invocation == old.Process.Invocation || |
| 213 | next.Process.Group != old.Process.Group || next.Binding.Target.BootID != boot { |
| 214 | continue |
| 215 | } |
| 216 | if old.Binding.Target.BootID == boot && next.Process.GroupIdentity == old.Process.GroupIdentity { |
| 217 | continue |
| 218 | } |
| 219 | path := filepath.Join(k.proc, strconv.Itoa(next.Process.PID)) |
| 220 | status, err := generationRead(path + "/status") |
| 221 | if err != nil { |
| 222 | return err |
| 223 | } |
| 224 | k.contained = []release.GenerationReservation{next} |
| 225 | ok, err := k.containedTask(ctx, path, status) |
| 226 | if err != nil || !ok { |
| 227 | return os.ErrPermission |
| 228 | } |
| 229 | return k.emptyGroups(ctx, filepath.Join(k.cgroups, next.Process.Group)) |
| 230 | } |
| 231 | return os.ErrPermission |
| 232 | } |
| 233 | |
amadan