Fix B6 proc_security_policy: was stubbing copyio instead of real target

Root cause: "most-called BL target" heuristic in _proc_info picked copyio
(4 calls, 0x28C bytes) over the real _proc_security_policy (2 calls,
0x134 bytes). Lowered size filter threshold from 0x300 to 0x200 to
correctly exclude utility functions like copyio. Boot-tested: PASS.
This commit is contained in:
Lakr
2026-03-04 20:42:58 +08:00
parent a65758e34c
commit a3756e29a2
2 changed files with 99 additions and 40 deletions

View File

@@ -1,25 +1,61 @@
# B6 `patch_proc_security_policy`
## How the patch works
- Source: `scripts/patchers/kernel_jb_patch_proc_security.py`.
## Status: FIXED (was PANIC)
## Root cause of failure
The patcher's heuristic picked the **wrong function** to stub:
- Found `_proc_info` correctly via `sub wN,wM,#1; cmp wN,#0x21` switch pattern
- Took the "most-called BL target" within proc_info as `_proc_security_policy`
- The most-called function (4 calls) was actually **copyio** (`sub_FFFFFE0007C4DD48`), a
generic copy-to-userspace utility used everywhere in the kernel (100+ xrefs)
- Stubbing copyio with `mov x0,#0; ret` broke all copyin/copyout operations
- Result: "Process 1 exec of /sbin/launchd failed, errno 2" (can't load launchd binary)
## Fix applied
Changed the heuristic from "most-called BL target" to a filtered approach:
1. Only count BL targets AFTER the switch dispatch (security policy is called within
switch cases, not in the prologue)
2. Filter by function size: skip large functions >0x300 bytes (copyio and other utilities
are large; `_proc_security_policy` is ~0x130 bytes)
3. Skip tiny functions <0x40 bytes (trivial helpers)
## IDA MCP evidence
### The wrong target (copyio)
- VA: `0xFFFFFE0007C4DD48` (file offset `0xC49D48`)
- References "copyio.c" and "copy_ensure_address_space_spec"
- 100+ xrefs from across the entire kernel
- Large function handling address space operations
### The real `_proc_security_policy`
- VA: `0xFFFFFE0008067148` (file offset `0x1063148`)
- Only 6 xrefs, all from proc_info-related functions:
- `sub_FFFFFE0008064A30` (_proc_info main handler) x2
- `sub_FFFFFE0008065540` x1
- `sub_FFFFFE0008065F6C` x1
- `sub_FFFFFE0008066624` x1
- `sub_FFFFFE0008064078` x1
- Function size: ~0x128 bytes
- Behavior: calls current_proc, does MAC policy check via indirect call, returns 0/error
### `_proc_info` function
- VA: `0xFFFFFE0008064A30`, size `0x9F8`
- Switch table at `0xFFFFFE0008064AA0`: `SUB W28, W25, #1; CMP W28, #0x21`
- BL target counts in proc_info:
- copyio (0x7C4DD48): 4 calls (most-called, WRONG target)
- proc_security_policy (0x8067148): 2 calls (correct target)
## How the patch works (fixed version)
- Locator strategy:
1. Try symbol `_proc_security_policy`.
2. If stripped, locate `_proc_info` by a switch-shape signature (`sub wN, wM, #1` + `cmp wN, #0x21`).
3. Inside that function, count BL targets and pick the most-called one as `_proc_security_policy`.
- Patch action: overwrite target function entry with:
- `mov x0, #0`
- `ret`
2. If stripped, locate `_proc_info` by switch-shape signature.
3. Count BL targets after the switch dispatch.
4. Filter candidates by size (0x40-0x300 bytes) to exclude utilities.
5. Pick the best match.
- Patch action: overwrite entry with `mov x0, #0; ret`
## Expected outcome
- Force `proc_security_policy` checks to return success.
## Target
- Targeted logic is the centralized process-security policy gate used by `proc_info` family paths.
## IDA MCP evidence (current state)
- This patch is mainly pattern-driven and symbol-first; this IDB is stripped.
- I validated related immediate patterns and candidate comparison sites in kernel text, but did not obtain a single uniquely provable `_proc_info -> most-called callee` chain within MCP single-call time limits.
- Status: behavior-level analysis confirmed from patch code; exact final function address is pending a longer multi-pass IDA sweep.
- Force `proc_security_policy` checks to return success (allow any process to query proc_info).
## Risk
- Turning this policy function into unconditional success can over-broaden process introspection and privilege surfaces.
- Over-broadens process introspection (any process can read info about any other process).

View File

@@ -1,6 +1,6 @@
"""Mixin: KernelJBPatchProcSecurityMixin."""
from .kernel_jb_base import ARM64_OP_IMM, MOV_X0_0, RET, Counter
from .kernel_jb_base import ARM64_OP_IMM, MOV_X0_0, RET, Counter, _rd32, struct
class KernelJBPatchProcSecurityMixin:
@@ -8,8 +8,10 @@ class KernelJBPatchProcSecurityMixin:
"""Stub _proc_security_policy: mov x0,#0; ret.
Anchor: find _proc_info via its distinctive switch-table pattern
(sub wN,wM,#1; cmp wN,#0x21), then identify the most-called BL
target within that function — that's _proc_security_policy.
(sub wN,wM,#1; cmp wN,#0x21), then identify _proc_security_policy
among BL targets — it's called 2+ times, is a small function
(<0x200 bytes), and is NOT called from the proc_info prologue
(it's called within switch cases, not before the switch dispatch).
"""
self._log("\n[JB] _proc_security_policy: mov x0,#0; ret")
@@ -23,6 +25,7 @@ class KernelJBPatchProcSecurityMixin:
# Find _proc_info by its distinctive switch table
# Pattern: sub wN, wM, #1; cmp wN, #0x21 (33 = max proc_info callnum)
proc_info_func = -1
switch_off = -1
ks, ke = self.kern_text
for off in range(ks, ke - 8, 4):
d = self._disas_at(off, 2)
@@ -31,21 +34,18 @@ class KernelJBPatchProcSecurityMixin:
i0, i1 = d[0], d[1]
if i0.mnemonic != "sub" or i1.mnemonic != "cmp":
continue
# sub wN, wM, #1
if len(i0.operands) < 3:
continue
if i0.operands[2].type != ARM64_OP_IMM or i0.operands[2].imm != 1:
continue
# cmp wN, #0x21
if len(i1.operands) < 2:
continue
if i1.operands[1].type != ARM64_OP_IMM or i1.operands[1].imm != 0x21:
continue
# Verify same register
if i0.operands[0].reg != i1.operands[0].reg:
continue
# Found it — find function start
proc_info_func = self.find_function_start(off)
switch_off = off
break
if proc_info_func < 0:
@@ -54,31 +54,54 @@ class KernelJBPatchProcSecurityMixin:
proc_info_end = self._find_func_end(proc_info_func, 0x4000)
self._log(
f" [+] _proc_info at 0x{proc_info_func:X} (size 0x{proc_info_end - proc_info_func:X})"
f" [+] _proc_info at 0x{proc_info_func:X} "
f"(size 0x{proc_info_end - proc_info_func:X})"
)
# Count BL targets within _proc_info — the most frequent one
# is _proc_security_policy (called once per switch case)
# Count BL targets within _proc_info (only AFTER the switch dispatch,
# since security policy is called from switch cases not the prologue)
bl_targets = Counter()
for off in range(proc_info_func, proc_info_end, 4):
for off in range(switch_off, proc_info_end, 4):
target = self._is_bl(off)
if target >= 0 and ks <= target < ke:
bl_targets[target] += 1
if not bl_targets:
self._log(" [-] no BL targets found in _proc_info")
self._log(" [-] no BL targets found in _proc_info switch cases")
return False
# The security policy check is called the most (once per case)
most_called = bl_targets.most_common(1)[0]
foff = most_called[0]
count = most_called[1]
self._log(f" [+] most-called BL target: 0x{foff:X} ({count} calls)")
# Find _proc_security_policy among candidates.
# It's called 2+ times, is a small function (<0x300 bytes),
# and is NOT a utility like copyio (which is much larger).
for foff, count in bl_targets.most_common():
if count < 2:
break
if count < 3:
self._log(" [-] most-called target has too few calls")
return False
func_end = self._find_func_end(foff, 0x400)
func_size = func_end - foff
self.emit(foff, MOV_X0_0, "mov x0,#0 [_proc_security_policy]")
self.emit(foff + 4, RET, "ret [_proc_security_policy]")
return True
self._log(
f" [*] candidate 0x{foff:X}: {count} calls, "
f"size 0x{func_size:X}"
)
# Skip large functions (utilities like copyio are ~0x28C bytes)
if func_size > 0x200:
self._log(f" [-] skipped (too large, likely utility)")
continue
# Skip tiny functions (< 0x40 bytes, likely trivial helpers)
if func_size < 0x40:
self._log(f" [-] skipped (too small)")
continue
self._log(
f" [+] identified _proc_security_policy at 0x{foff:X} "
f"({count} calls, size 0x{func_size:X})"
)
self.emit(foff, MOV_X0_0, "mov x0,#0 [_proc_security_policy]")
self.emit(foff + 4, RET, "ret [_proc_security_policy]")
return True
self._log(" [-] _proc_security_policy not identified among BL targets")
return False