| 1 |
// Program captrace traces processes and notices when they attempt |
| 2 |
// kernel actions that require Effective capabilities. |
| 3 |
// |
| 4 |
// The reference material for developing this tool was the the book |
| 5 |
// "Linux Observabililty with BPF" by David Calavera and Lorenzo |
| 6 |
// Fontana. |
| 7 |
package main |
| 8 |
|
| 9 |
import ( |
| 10 |
"bufio" |
| 11 |
"flag" |
| 12 |
"fmt" |
| 13 |
"io" |
| 14 |
"log" |
| 15 |
"os" |
| 16 |
"os/exec" |
| 17 |
"strconv" |
| 18 |
"strings" |
| 19 |
"sync" |
| 20 |
"syscall" |
| 21 |
"time" |
| 22 |
|
| 23 |
"kernel.org/pub/linux/libs/security/libcap/cap" |
| 24 |
) |
| 25 |
|
| 26 |
var ( |
| 27 |
bpftrace = flag.String("bpftrace", "bpftrace", "command to launch bpftrace") |
| 28 |
debug = flag.Bool("debug", false, "more output") |
| 29 |
pid = flag.Int("pid", -1, "PID of target process to trace (-1 = trace all)") |
| 30 |
) |
| 31 |
|
| 32 |
type thread struct { |
| 33 |
PPID, Datum int |
| 34 |
Value cap.Value |
| 35 |
Token string |
| 36 |
} |
| 37 |
|
| 38 |
// mu protects these two maps. |
| 39 |
var mu sync.Mutex |
| 40 |
|
| 41 |
// tids tracks which PIDs we are following. |
| 42 |
var tids = make(map[int]int) |
| 43 |
|
| 44 |
// cache tracks in-flight cap_capable invocations. |
| 45 |
var cache = make(map[int]*thread) |
| 46 |
|
| 47 |
// event adds or resolves a capability event. |
| 48 |
func event(add bool, tid int, th *thread) { |
| 49 |
mu.Lock() |
| 50 |
defer mu.Unlock() |
| 51 |
|
| 52 |
if len(tids) != 0 { |
| 53 |
if _, ok := tids[th.PPID]; !ok { |
| 54 |
if *debug { |
| 55 |
log.Printf("dropped %d %d %v event", th.PPID, tid, *th) |
| 56 |
} |
| 57 |
return |
| 58 |
} |
| 59 |
tids[tid] = th.PPID |
| 60 |
tids[th.PPID] = th.PPID |
| 61 |
} |
| 62 |
|
| 63 |
if add { |
| 64 |
cache[tid] = th |
| 65 |
} else { |
| 66 |
if b, ok := cache[tid]; ok { |
| 67 |
detail := "" |
| 68 |
if th.Datum < 0 { |
| 69 |
detail = fmt.Sprintf(" (%v)", syscall.Errno(-th.Datum)) |
| 70 |
} |
| 71 |
task := "" |
| 72 |
if th.PPID != tid { |
| 73 |
task = fmt.Sprintf("+{%d}", tid) |
| 74 |
} |
| 75 |
log.Printf("%-16s %d%s opt=%d %q -> %d%s", b.Token, b.PPID, task, b.Datum, b.Value, th.Datum, detail) |
| 76 |
} |
| 77 |
delete(cache, tid) |
| 78 |
} |
| 79 |
} |
| 80 |
|
| 81 |
// tailTrace tails the bpftrace command output recognizing lines of |
| 82 |
// interest. |
| 83 |
func tailTrace(cmd *exec.Cmd, out io.Reader) { |
| 84 |
launched := false |
| 85 |
sc := bufio.NewScanner(out) |
| 86 |
for sc.Scan() { |
| 87 |
fields := strings.Split(sc.Text(), " ") |
| 88 |
if len(fields) < 4 { |
| 89 |
continue // ignore |
| 90 |
} |
| 91 |
if !launched { |
| 92 |
launched = true |
| 93 |
mu.Unlock() |
| 94 |
} |
| 95 |
switch fields[0] { |
| 96 |
case "CB": |
| 97 |
if len(fields) < 6 { |
| 98 |
continue |
| 99 |
} |
| 100 |
pid, err := strconv.Atoi(fields[1]) |
| 101 |
if err != nil { |
| 102 |
continue |
| 103 |
} |
| 104 |
th := &thread{ |
| 105 |
PPID: pid, |
| 106 |
} |
| 107 |
tid, err := strconv.Atoi(fields[2]) |
| 108 |
if err != nil { |
| 109 |
continue |
| 110 |
} |
| 111 |
c, err := strconv.Atoi(fields[3]) |
| 112 |
if err != nil { |
| 113 |
continue |
| 114 |
} |
| 115 |
th.Value = cap.Value(c) |
| 116 |
aud, err := strconv.Atoi(fields[4]) |
| 117 |
if err != nil { |
| 118 |
continue |
| 119 |
} |
| 120 |
th.Datum = aud |
| 121 |
th.Token = strings.Join(fields[5:], " ") |
| 122 |
event(true, tid, th) |
| 123 |
case "CE": |
| 124 |
if len(fields) < 4 { |
| 125 |
continue |
| 126 |
} |
| 127 |
pid, err := strconv.Atoi(fields[1]) |
| 128 |
if err != nil { |
| 129 |
continue |
| 130 |
} |
| 131 |
th := &thread{ |
| 132 |
PPID: pid, |
| 133 |
} |
| 134 |
tid, err := strconv.Atoi(fields[2]) |
| 135 |
if err != nil { |
| 136 |
continue |
| 137 |
} |
| 138 |
aud, err := strconv.Atoi(fields[3]) |
| 139 |
if err != nil { |
| 140 |
continue |
| 141 |
} |
| 142 |
th.Datum = aud |
| 143 |
event(false, tid, th) |
| 144 |
default: |
| 145 |
if *debug { |
| 146 |
fmt.Println("unparsable:", fields) |
| 147 |
} |
| 148 |
} |
| 149 |
} |
| 150 |
if err := sc.Err(); err != nil { |
| 151 |
log.Fatalf("scanning failed: %v", err) |
| 152 |
} |
| 153 |
} |
| 154 |
|
| 155 |
// tracer invokes bpftool it returns an error if the invocation fails. |
| 156 |
func tracer() (*exec.Cmd, error) { |
| 157 |
cmd := exec.Command(*bpftrace, "-e", `kprobe:cap_capable { |
| 158 |
printf("CB %d %d %d %d %s\n", pid, tid, arg2, arg3, comm); |
| 159 |
} |
| 160 |
kretprobe:cap_capable { |
| 161 |
printf("CE %d %d %d\n", pid, tid, retval); |
| 162 |
}`) |
| 163 |
out, err := cmd.StdoutPipe() |
| 164 |
cmd.Stderr = os.Stderr |
| 165 |
if err != nil { |
| 166 |
return nil, fmt.Errorf("unable to create stdout for %q: %v", *bpftrace, err) |
| 167 |
} |
| 168 |
mu.Lock() // Unlocked on first ouput from tracer. |
| 169 |
if err := cmd.Start(); err != nil { |
| 170 |
return nil, fmt.Errorf("failed to start %q: %v", *bpftrace, err) |
| 171 |
} |
| 172 |
go tailTrace(cmd, out) |
| 173 |
return cmd, nil |
| 174 |
} |
| 175 |
|
| 176 |
func main() { |
| 177 |
flag.Usage = func() { |
| 178 |
fmt.Fprintf(flag.CommandLine.Output(), `Usage: %s [options] [command ...] |
| 179 |
|
| 180 |
This tool monitors cap_capable() kernel execution to summarize when |
| 181 |
Effective Flag capabilities are checked in a running process{thread}. |
| 182 |
The monitoring is performed indirectly using the bpftrace tool. |
| 183 |
|
| 184 |
Each line logged has a timestamp at which the tracing program is able to |
| 185 |
summarize the return value of the check. A return value of " -> 0" implies |
| 186 |
the check succeeded and confirms the process{thread} does have the |
| 187 |
specified Effective capability. |
| 188 |
|
| 189 |
The listed "opt=" value indicates some auditing context for why the |
| 190 |
kernel needed to check the capability was Effective. |
| 191 |
|
| 192 |
Options: |
| 193 |
`, os.Args[0]) |
| 194 |
flag.PrintDefaults() |
| 195 |
} |
| 196 |
flag.Parse() |
| 197 |
|
| 198 |
tr, err := tracer() |
| 199 |
if err != nil { |
| 200 |
log.Fatalf("failed to start tracer: %v", err) |
| 201 |
} |
| 202 |
|
| 203 |
mu.Lock() |
| 204 |
|
| 205 |
if *pid != -1 { |
| 206 |
tids[*pid] = *pid |
| 207 |
} else if len(flag.Args()) != 0 { |
| 208 |
args := flag.Args() |
| 209 |
cmd := exec.Command(args[0], args[1:]...) |
| 210 |
cmd.Stdin = os.Stdin |
| 211 |
cmd.Stdout = os.Stdout |
| 212 |
cmd.Stderr = os.Stderr |
| 213 |
if err := cmd.Start(); err != nil { |
| 214 |
log.Fatalf("failed to start %v: %v", flag.Args(), err) |
| 215 |
} |
| 216 |
tids[cmd.Process.Pid] = cmd.Process.Pid |
| 217 |
|
| 218 |
// waiting for the trace to complete is racy, so we sleep |
| 219 |
// to obtain the last events then kill the tracer and wait |
| 220 |
// for it to exit. Defers are in reverse order. |
| 221 |
defer tr.Wait() |
| 222 |
defer tr.Process.Kill() |
| 223 |
defer time.Sleep(1 * time.Second) |
| 224 |
|
| 225 |
tr = cmd |
| 226 |
} |
| 227 |
|
| 228 |
mu.Unlock() |
| 229 |
tr.Wait() |
| 230 |
} |