| 1 | package host |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "context" |
| 6 | "io" |
| 7 | "net" |
| 8 | "net/netip" |
| 9 | "os" |
| 10 | "path/filepath" |
| 11 | "slices" |
| 12 | "strings" |
| 13 | |
| 14 | "github.com/carlosframework/platform/internal/instancectl" |
| 15 | "github.com/carlosframework/platform/internal/provision" |
| 16 | "github.com/carlosframework/platform/internal/rootstate" |
| 17 | ) |
| 18 | |
| 19 | func claimReservedDropin(s claimUnitShape) []byte { |
| 20 | source := filepath.Join("/run/carlos-control/exe/.claim", provision.InstanceID(s.Host)) |
| 21 | instance := provision.InstanceID(s.Host) |
| 22 | if s.Unit == instancectl.UnitName(instancectl.RoleRun, s.Host) { |
| 23 | instance = "%I" |
| 24 | } |
| 25 | return []byte("[Service]\nBindReadOnlyPaths=" + source + "/launch.json:" + filepath.Join(provision.InstanceDir("/data", s.Host), instancectl.LaunchFileName(instancectl.RoleRun)) + "\nBindReadOnlyPaths=" + source + "/exec.sock:" + instancectl.ClaimExecSocket + "\nBindReadOnlyPaths=" + source + "/permit.json:" + instancectl.ClaimExecSnapshot + "\nBindReadOnlyPaths=/proc/sys/kernel/random/boot_id:" + instancectl.ClaimExecBootID + "\nExecStart=\nExecStart=" + instancectl.SystemBinary + " instance-exec -claim-reservation=" + instancectl.ClaimExecSnapshot + " run " + instance + "\n") |
| 26 | } |
| 27 | |
| 28 | type claimEffectiveRefusal struct { |
| 29 | field string |
| 30 | cause error |
| 31 | } |
| 32 | |
| 33 | func (e claimEffectiveRefusal) Error() string { return e.cause.Error() } |
| 34 | func (e claimEffectiveRefusal) Unwrap() error { return e.cause } |
| 35 | |
| 36 | func claimEffectiveUnit(ctx context.Context, s claimUnitShape, ctl func(context.Context, ...string) (string, error)) error { |
| 37 | if ctl == nil { |
| 38 | return claimEffectiveRefusal{"unknown", os.ErrPermission} |
| 39 | } |
| 40 | ordinary := s.Unit == instancectl.UnitName(instancectl.RoleRun, s.Host) |
| 41 | fragment := "/etc/systemd/system/" + s.Unit |
| 42 | var definition []byte |
| 43 | if ordinary { |
| 44 | fragment = "/etc/systemd/system/carlos-instance@.service" |
| 45 | var err error |
| 46 | definition, err = unitsFS.ReadFile("units/carlos-instance@.service") |
| 47 | if err != nil { |
| 48 | return err |
| 49 | } |
| 50 | f, err := rootstate.OpenTrustedFile(fragment) |
| 51 | if err != nil { |
| 52 | return err |
| 53 | } |
| 54 | body, err := io.ReadAll(io.LimitReader(f, rootstate.MaxRecord+1)) |
| 55 | f.Close() |
| 56 | if err != nil { |
| 57 | return err |
| 58 | } |
| 59 | if !bytes.Equal(body, definition) { |
| 60 | return os.ErrPermission |
| 61 | } |
| 62 | } else { |
| 63 | var err error |
| 64 | definition, err = claimBoundUnitDefinition(s) |
| 65 | if err != nil { |
| 66 | return err |
| 67 | } |
| 68 | } |
| 69 | scalar := map[string]string{"User": "carlos", "Group": "carlos", "NoNewPrivileges": "yes", "ProtectSystem": "strict", "ProtectHome": "yes", "PrivateTmp": "yes", "PrivateDevices": "yes", "PrivateIPC": "yes", "ProtectControlGroups": "yes", "ProtectProc": "invisible", "ProcSubset": "pid", "PrivatePIDs": "yes", "RestrictNamespaces": "yes", "CapabilityBoundingSet": "", "MemoryDenyWriteExecute": "yes", "ProtectKernelTunables": "yes", "ProtectKernelModules": "yes", "ProtectKernelLogs": "yes", "ProtectClock": "yes", "ProtectHostname": "yes", "RemoveIPC": "yes", "LockPersonality": "yes", "RestrictRealtime": "yes", "Restart": "no", "Slice": instancectl.SliceName(s.Host)} |
| 70 | if s.Addr != "" { |
| 71 | host, port, e := net.SplitHostPort(s.Addr) |
| 72 | if e != nil { |
| 73 | return e |
| 74 | } |
| 75 | family := "ipv4" |
| 76 | if net.ParseIP(host).To4() == nil { |
| 77 | family = "ipv6" |
| 78 | } |
| 79 | scalar["SocketBindAllow"] = family + ":tcp:" + port |
| 80 | scalar["SocketBindDeny"] = "any" |
| 81 | } |
| 82 | keys := []string{"FragmentPath", "DropInPaths", "NeedDaemonReload", "ExecStart", "IPAddressAllow", "IPAddressDeny", "BindsTo", "PartOf"} |
| 83 | for key := range scalar { |
| 84 | keys = append(keys, key) |
| 85 | } |
| 86 | slices.Sort(keys) |
| 87 | raw, err := ctl(ctx, "show", "-p", strings.Join(keys, ","), s.Unit, "--all") |
| 88 | if err != nil { |
| 89 | return err |
| 90 | } |
| 91 | props := map[string]string{} |
| 92 | for _, line := range strings.Split(raw, "\n") { |
| 93 | if key, value, ok := strings.Cut(line, "="); ok { |
| 94 | if _, seen := props[key]; seen { |
| 95 | return claimEffectiveRefusal{"unknown", os.ErrPermission} |
| 96 | } |
| 97 | props[key] = value |
| 98 | } |
| 99 | } |
| 100 | for _, key := range keys { |
| 101 | if _, ok := props[key]; !ok { |
| 102 | return claimEffectiveRefusal{"missing:" + key, os.ErrPermission} |
| 103 | } |
| 104 | } |
| 105 | if props["FragmentPath"] != fragment || props["NeedDaemonReload"] != "no" { |
| 106 | return claimEffectiveRefusal{"unknown", os.ErrPermission} |
| 107 | } |
| 108 | for key, want := range scalar { |
| 109 | if props[key] != want && !(key == "SocketBindAllow" && props[key] == strings.Replace(want, ":tcp:", ":tcp", 1)) { |
| 110 | return claimEffectiveRefusal{"value:" + key + "=" + props[key], os.ErrPermission} |
| 111 | } |
| 112 | } |
| 113 | if !ordinary && (strings.Contains(props["BindsTo"], "carlos-edge.service") || strings.Contains(props["PartOf"], "carlos-edge.service")) { |
| 114 | return claimEffectiveRefusal{"edge-coupling", os.ErrPermission} |
| 115 | } |
| 116 | executable := "/usr/bin/false" |
| 117 | if ordinary { |
| 118 | executable = instancectl.SystemBinary |
| 119 | } |
| 120 | if props["DropInPaths"] != "" { |
| 121 | path := filepath.Join("/run/systemd/system", s.Unit+".d", "60-claim-launch.conf") |
| 122 | if props["DropInPaths"] != path { |
| 123 | return claimEffectiveRefusal{"dropins", os.ErrPermission} |
| 124 | } |
| 125 | f, err := openClaimUnitDropin(s.Unit) |
| 126 | if err != nil { |
| 127 | return claimEffectiveRefusal{"dropin-file", err} |
| 128 | } |
| 129 | body, err := io.ReadAll(io.LimitReader(f, rootstate.MaxRecord+1)) |
| 130 | f.Close() |
| 131 | if err != nil { |
| 132 | return err |
| 133 | } |
| 134 | if !bytes.Equal(body, claimReservedDropin(s)) { |
| 135 | return claimEffectiveRefusal{"dropin-bytes", os.ErrPermission} |
| 136 | } |
| 137 | executable = instancectl.SystemBinary |
| 138 | } |
| 139 | if strings.Count(props["ExecStart"], "path=") != 1 || !strings.Contains(props["ExecStart"], "path="+executable+" ;") { |
| 140 | return claimEffectiveRefusal{"unknown", os.ErrPermission} |
| 141 | } |
| 142 | for _, key := range []string{"IPAddressAllow", "IPAddressDeny"} { |
| 143 | want := "" |
| 144 | for _, line := range strings.Split(string(definition), "\n") { |
| 145 | if strings.HasPrefix(line, key+"=") { |
| 146 | want += " " + strings.TrimPrefix(line, key+"=") |
| 147 | } |
| 148 | } |
| 149 | a, e := claimIPSet(want) |
| 150 | if e != nil { |
| 151 | return e |
| 152 | } |
| 153 | b, e := claimIPSet(props[key]) |
| 154 | if e != nil { |
| 155 | return e |
| 156 | } |
| 157 | if !slices.Equal(a, b) { |
| 158 | return claimEffectiveRefusal{"network:" + key, os.ErrPermission} |
| 159 | } |
| 160 | } |
| 161 | return nil |
| 162 | } |
| 163 | |
| 164 | func claimIPSet(raw string) ([]string, error) { |
| 165 | var out []string |
| 166 | raw = strings.ReplaceAll(raw, "localhost", "127.0.0.0/8 ::1/128") |
| 167 | raw = strings.ReplaceAll(raw, "multicast", "224.0.0.0/4 ff00::/8") |
| 168 | for _, value := range strings.Fields(raw) { |
| 169 | p, e := netip.ParsePrefix(value) |
| 170 | if e != nil { |
| 171 | return nil, os.ErrPermission |
| 172 | } |
| 173 | out = append(out, p.Masked().String()) |
| 174 | } |
| 175 | slices.Sort(out) |
| 176 | out = slices.Compact(out) |
| 177 | // systemd removes redundant subnets when attaching the running unit's |
| 178 | // filter. Compare address sets, including before and after that reduction. |
| 179 | kept := make([]string, 0, len(out)) |
| 180 | for _, value := range out { |
| 181 | p := netip.MustParsePrefix(value) |
| 182 | covered := false |
| 183 | for _, other := range out { |
| 184 | q := netip.MustParsePrefix(other) |
| 185 | if q.Bits() < p.Bits() && q.Contains(p.Addr()) { |
| 186 | covered = true |
| 187 | break |
| 188 | } |
| 189 | } |
| 190 | if !covered { |
| 191 | kept = append(kept, value) |
| 192 | } |
| 193 | } |
| 194 | return kept, nil |
| 195 | } |
| 196 | |
amadan