aboutsummaryrefslogtreecommitdiffstats
path: root/fw/LEAD-MERIT-CMD05-MEMCMP-ORACLE.md
blob: fad12a76aab4a9b19c6d16d66b521342be33947f (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
# LEAD: MERIT `cmd 5` compares with a host-controlled length (gated, not reached)

**Component:** MERIT TA 24.0.0.70 — loadable *and* invokable on this Navi 31.
**Class:** potential byte-at-a-time memory-disclosure oracle (unbounded compare length).
**Status:** **lead.** The unbounded length is real and confirmed statically; on hardware the
handler rejects before the compare, so the oracle is not reachable as driven here.

## The unbounded length

`merit_cmd05` validates its slot properly — index `(v3-1) <= 0x3F` (matching the allocator's
1..64 range) and a presence byte `!= 0` — and then:

```
0x401b28  LDR R7, [R5,#0x4C]          ; length <- *(params + 0x4c), HOST-controlled
          ...
          loop v10 = 0, 16, 32 ... while v10 != 0x90        ; 9 x 16-byte table entries
            merit_lib_52b900d8_88(table + v10, params + 0x50, R7)
```

`merit_lib_52b900d8_88` is **memcmp** — verified from its body (word-wise compare, then a
byte-wise tail, returning 1/-1/0). There is **no `cmp` on R7 anywhere in the function**; the
only compares are the index bound (`0x3F`), the loop bound (`0x90`) and the memcmp result.

## Why an unbounded compare length would matter

The table is at `0x40204c`, nine 16-byte GUID-like entries, 144 bytes, and its contents are
**in the image** — so an attacker holding the firmware knows them. A length beyond 144 walks
forward out of the table. The image is `0x15000` bytes at `0x400100`, and the slot array
(`off_401CE0 = 0x415004`) sits just past it holding live per-slot data. So with a length of
roughly `0x13000` the compare would run from known image bytes into runtime slot memory, and
because only equality is consumed (`memcmp(...) == 0`), each invoke answers "does my guessed
prefix match" — the standard byte-at-a-time extension, ~256 invokes per byte, starting from
a known prefix.

That is the shape of a genuine disclosure primitive. Two things have to hold, and one does not.

## Hardware [HW]: the compare is NOT reached

`/root/psp/hw/merit_probe.py` + `merit_probe2.py`:

```
cmd01 (allocate)                    -> status 0, slot index 1 at params+4132
cmd05, len 0x10/0x20/0x100/0x1000/0x13000, data = real table entry 0
                                    -> identical response every time
cmd05, len 0x10, data = 0xAA * 16   -> identical response
diff of the whole response page, valid vs garbage: ONLY my own input bytes at 0x50 differ
MERIT status (params+8)             -> 0x105 (261) in every case
```

Identical responses for a *correct* table entry and for garbage prove the memcmp never ran:
had it run, entry 0 would have matched and taken the success branch. `0x105` is a rejection
raised by one of the earlier gates:

* `merit_cmd05__h1(idx, params+0x14, 0x1010, params+0x14)` must return non-NULL; or
* `*(params+0x44) == ` a 4-byte value read from **slot+33..36**, i.e. the caller must already
  know a slot-resident value.

The second gate is the interesting one: it requires knowledge of slot content, which is
presumably established by a successful `cmd04`-style provisioning step. Until that is driven,
the compare is unreachable and the oracle is theoretical.

## Honest placement

Recorded as a lead because the dangerous half is verified (unbounded host length into memcmp,
with the target region known to abut live slot data) while the reachable half is measured to
fail. Cross-checking the gate values — particularly whether `slot+33..36` is host-settable via
`cmd01`/`cmd04` — is the next step, and it is a small hardware experiment.

Note also that MERIT was previously "cleared" by the liveness sweep on the strength of
`merit_cmd04` alone; `merit_cmd05` was flagged by the same detector and had **not** been read
until now. The detector was right to flag it twice.

================================================================================
# RESOLVED — gated by an ASD-derived 32-bit value the host cannot obtain

The open question was whether the gate value at `slot+33..36` is host-settable. It is not.

**Who writes it:** `merit_cmd04` at `0x4019ec`:

```
0x4019e8  LDRD R0, R1, [SP,#0x28]     ; two dwords of a DERIVED blob
0x4019ec  STRB R0, [R5,#0x21]!        ; written byte-wise into slot+33..
0x4019f2  STRB R1, [R5,#4]!  ...      ; (big-endian byte order)
```

**Where the blob comes from:** `sub_40143C(params+0x14, len, out, &outlen)`, whose only
non-trivial callee is **`merit_asd_rpc_200`** — service 2, ASD opcode `0x200`. So those four
bytes are the output of an ASD operation, not a transform of host input.

**Can the host read them back?** No. Scanning every MERIT function for reads of offsets
`0x1f..0x28` relative to a slot finds exactly one: `merit_cmd05` itself, the consumer.
(`merit_cmd01`'s apparent hit is a literal-pool load, not a slot access.) No command returns
those bytes to the caller.

So reaching the unbounded `memcmp` requires the caller to supply, at `params+0x44`, a
**32-bit value derived inside ASD** and never disclosed — not brute-forceable at one invoke
per attempt, and not obtainable by reading it back. That matches the hardware result exactly:
status `0x105` before the compare, identical responses for a correct table entry and for
garbage.

## Final verdict

**Robustness defect, not exploitable.** The missing upper bound on the compare length is
real and worth reporting as hardening (any length below the gate would read past a 144-byte
table toward the slot array), but the path is protected by a secret-value gate, so there is
no disclosure primitive here.

Recorded in full because the reasoning is the useful part: an unbounded length is only as
dangerous as the weakest gate in front of it, and that gate turned out to be ASD-derived —
the same component that gates the GCP findings.