| 1 | package cgroup |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "os" |
| 8 | "strconv" |
| 9 | "strings" |
| 10 | "sync/atomic" |
| 11 | "time" |
| 12 | ) |
| 13 | |
| 14 | // ShimVerb is the carlos binary's hidden placement verb (Paul's ruling, |
| 15 | // #272, 2026-10-13). carlos-edge.service's RestrictNamespaces=yes makes |
| 16 | // systemd's seccomp filter answer clone3 with ENOSYS, so no child can be |
| 17 | // created inside its cgroup. Instead the child starts as this binary, |
| 18 | // writes its own PID into the target cgroup through a directory fd the |
| 19 | // parent validated, checks where it landed, and only then execs the |
| 20 | // tenant: the tenant's code never runs outside its limits. |
| 21 | const ShimVerb = "__cgexec" |
| 22 | |
| 23 | // ProbeVerb prints the process's own cgroup and exits: the tenant the |
| 24 | // arm-time placement check (VerifyPlacement) execs through the shim. |
| 25 | const ProbeVerb = "__cgprobe" |
| 26 | |
| 27 | var shimHooked atomic.Bool |
| 28 | |
| 29 | // shimForcedForTest makes shimAvailable answer true while |
| 30 | // FakePlacementForTest has swapped the placement for one that runs no shim. |
| 31 | // startViaShim itself still asks shimHooked, so the override can never make |
| 32 | // a test binary without the hook start itself as a shim. |
| 33 | var shimForcedForTest atomic.Bool |
| 34 | |
| 35 | // RunShimIfRequested must be the first statement of main(), and of TestMain |
| 36 | // in a test binary that places processes. It marks this process as able to |
| 37 | // serve the shim (shimAvailable), and when this process IS a shim or a |
| 38 | // probe it runs it and exits; it returns only when it is neither. |
| 39 | func RunShimIfRequested() { |
| 40 | shimHooked.Store(true) |
| 41 | if len(os.Args) < 2 { |
| 42 | return |
| 43 | } |
| 44 | switch os.Args[1] { |
| 45 | case ShimVerb: |
| 46 | os.Exit(runShim(os.Args[2:])) |
| 47 | case ProbeVerb: |
| 48 | os.Exit(runProbe()) |
| 49 | } |
| 50 | } |
| 51 | |
| 52 | // placementError is a start whose process could not be placed in its |
| 53 | // cgroup before exec. Nothing of the tenant ran. |
| 54 | type placementError struct{ err error } |
| 55 | |
| 56 | func (e *placementError) Error() string { return "placement before exec: " + e.err.Error() } |
| 57 | func (e *placementError) Unwrap() error { return e.err } |
| 58 | |
| 59 | func isPlacementError(err error) bool { |
| 60 | var p *placementError |
| 61 | return errors.As(err, &p) |
| 62 | } |
| 63 | |
| 64 | // ErrPlacementInFlight is a cleanup asked for members while a process is |
| 65 | // being placed in its cgroup (Task 2). |
| 66 | var ErrPlacementInFlight = errors.New("cgroup: a process is being placed in this cgroup") |
| 67 | |
| 68 | // shimDeadline bounds each of the parent's two waits on the shim: start to |
| 69 | // placement, and release to exec (Ruling 9). |
| 70 | var shimDeadline = 10 * time.Second |
| 71 | |
| 72 | // shimExe is what the parent starts as the shim: the running build, even |
| 73 | // after a roll has replaced the file on disk. |
| 74 | var shimExe = "/proc/self/exe" |
| 75 | |
| 76 | // shimTestHold is copied into each spec's TestHold. Only test files set it, |
| 77 | // and only a test binary's shim honours it. |
| 78 | var shimTestHold int |
| 79 | |
| 80 | // shimAfterRelease, when set, runs between the release byte and the read of |
| 81 | // the exec status. Only test files set it. |
| 82 | var shimAfterRelease func() |
| 83 | |
| 84 | // shimSpec is what the parent sends the shim over the spec pipe: never in |
| 85 | // the shim's argv or environment, so no secret shows in /proc/<pid>/cmdline |
| 86 | // and no tenant variable shapes the shim before placement. |
| 87 | type shimSpec struct { |
| 88 | Path string |
| 89 | Dir string // the tenant's cmd.Dir; the shim changes to it after placement |
| 90 | Argv []string |
| 91 | Env []string |
| 92 | Expected string // "/..." as /proc/self/cgroup shows it after "0::" |
| 93 | KeepFDs int // fds 0..KeepFDs-1 are the tenant's |
| 94 | TestHold int // test binaries only: 0 none, 1 hold before step 4, 2 hold after step 10 (before R) |
| 95 | // SigIgn is the set of signals the agent ignored when it started the |
| 96 | // shim, as /proc/self/status's SigIgn shows it: bit n-1 is signal n. |
| 97 | // A direct start's child keeps exactly these ignored through its exec |
| 98 | // (the runtime resets only the signals it handles, clearSignalHandlers), |
| 99 | // but the shim's own Go runtime replaces most inherited ignores with its |
| 100 | // handlers, which the tenant's execve then resets to default. So the |
| 101 | // shim ignores these again immediately before it execs (whole-branch |
| 102 | // review, finding 2). |
| 103 | SigIgn uint64 |
| 104 | } |
| 105 | |
| 106 | const ( |
| 107 | shimSpecMagic = "carlos-cgexec-2" |
| 108 | // shimSpecMax bounds the encoded spec the shim will read. |
| 109 | shimSpecMax = 16 << 20 |
| 110 | // shimSpecMaxCount bounds each decoded count (argv, env, KeepFDs). |
| 111 | shimSpecMaxCount = 1 << 20 |
| 112 | ) |
| 113 | |
| 114 | // encodeShimSpec writes each field followed by a NUL, in order: the magic, |
| 115 | // Path, Dir, Expected, KeepFDs, TestHold, SigIgn, len(Argv), each argv element, |
| 116 | // len(Env), each env entry. Bytes pass through untouched (invalid UTF-8 |
| 117 | // included); a string holding a NUL is refused, as execve could not carry |
| 118 | // it either. |
| 119 | func encodeShimSpec(s shimSpec) ([]byte, error) { |
| 120 | var b []byte |
| 121 | put := func(field, v string) error { |
| 122 | if strings.IndexByte(v, 0) >= 0 { |
| 123 | return fmt.Errorf("cgroup: shim spec: %s contains a NUL byte", field) |
| 124 | } |
| 125 | b = append(b, v...) |
| 126 | b = append(b, 0) |
| 127 | return nil |
| 128 | } |
| 129 | if err := put("magic", shimSpecMagic); err != nil { |
| 130 | return nil, err |
| 131 | } |
| 132 | for _, f := range []struct{ name, v string }{ |
| 133 | {"path", s.Path}, {"dir", s.Dir}, {"expected", s.Expected}, |
| 134 | {"keep", strconv.Itoa(s.KeepFDs)}, {"hold", strconv.Itoa(s.TestHold)}, |
| 135 | {"sigign", strconv.FormatUint(s.SigIgn, 10)}, |
| 136 | {"argc", strconv.Itoa(len(s.Argv))}, |
| 137 | } { |
| 138 | if err := put(f.name, f.v); err != nil { |
| 139 | return nil, err |
| 140 | } |
| 141 | } |
| 142 | for _, a := range s.Argv { |
| 143 | if err := put("argv", a); err != nil { |
| 144 | return nil, err |
| 145 | } |
| 146 | } |
| 147 | if err := put("envc", strconv.Itoa(len(s.Env))); err != nil { |
| 148 | return nil, err |
| 149 | } |
| 150 | for _, e := range s.Env { |
| 151 | if err := put("env", e); err != nil { |
| 152 | return nil, err |
| 153 | } |
| 154 | } |
| 155 | return b, nil |
| 156 | } |
| 157 | |
| 158 | // decodeShimSpec is encodeShimSpec's inverse. It refuses a wrong magic, a |
| 159 | // buffer that ends early or runs past the last field, and a count above |
| 160 | // shimSpecMaxCount. |
| 161 | func decodeShimSpec(b []byte) (shimSpec, error) { |
| 162 | var s shimSpec |
| 163 | next := func() (string, error) { |
| 164 | i := bytes.IndexByte(b, 0) |
| 165 | if i < 0 { |
| 166 | return "", errors.New("cgroup: shim spec: truncated") |
| 167 | } |
| 168 | v := string(b[:i]) |
| 169 | b = b[i+1:] |
| 170 | return v, nil |
| 171 | } |
| 172 | count := func() (int, error) { |
| 173 | v, err := next() |
| 174 | if err != nil { |
| 175 | return 0, err |
| 176 | } |
| 177 | n, err := strconv.Atoi(v) |
| 178 | if err != nil || n < 0 || n > shimSpecMaxCount { |
| 179 | return 0, fmt.Errorf("cgroup: shim spec: bad count %q", v) |
| 180 | } |
| 181 | return n, nil |
| 182 | } |
| 183 | magic, err := next() |
| 184 | if err != nil { |
| 185 | return s, err |
| 186 | } |
| 187 | if magic != shimSpecMagic { |
| 188 | return s, fmt.Errorf("cgroup: shim spec: magic %q, want %q", magic, shimSpecMagic) |
| 189 | } |
| 190 | for _, p := range []*string{&s.Path, &s.Dir, &s.Expected} { |
| 191 | if *p, err = next(); err != nil { |
| 192 | return s, err |
| 193 | } |
| 194 | } |
| 195 | if s.KeepFDs, err = count(); err != nil { |
| 196 | return s, err |
| 197 | } |
| 198 | if s.TestHold, err = count(); err != nil { |
| 199 | return s, err |
| 200 | } |
| 201 | sigign, err := next() |
| 202 | if err != nil { |
| 203 | return s, err |
| 204 | } |
| 205 | if s.SigIgn, err = strconv.ParseUint(sigign, 10, 64); err != nil { |
| 206 | return s, fmt.Errorf("cgroup: shim spec: bad ignored-signal set %q", sigign) |
| 207 | } |
| 208 | argc, err := count() |
| 209 | if err != nil { |
| 210 | return s, err |
| 211 | } |
| 212 | s.Argv = []string{} |
| 213 | for range argc { |
| 214 | a, err := next() |
| 215 | if err != nil { |
| 216 | return s, err |
| 217 | } |
| 218 | s.Argv = append(s.Argv, a) |
| 219 | } |
| 220 | envc, err := count() |
| 221 | if err != nil { |
| 222 | return s, err |
| 223 | } |
| 224 | s.Env = []string{} |
| 225 | for range envc { |
| 226 | e, err := next() |
| 227 | if err != nil { |
| 228 | return s, err |
| 229 | } |
| 230 | s.Env = append(s.Env, e) |
| 231 | } |
| 232 | if len(b) != 0 { |
| 233 | return s, fmt.Errorf("cgroup: shim spec: %d trailing bytes", len(b)) |
| 234 | } |
| 235 | return s, nil |
| 236 | } |
| 237 | |
amadan