Since KernelSU Manager can now be built for 32-bit, theres this problematic
setup where userspace is 32-bit (armeabi-v7a) and kernel is 64bit (aarch64).
On 64-bit kernels with CONFIG_COMPAT=y, 32-bit userspace passes 32-bit pointers.
These values are interpreted as 64-bit pointers without proper casting and that
results in invalid or near-null memory access.
This patch adds proper compat-mode handling with the ff changes:
- introduce a dedicated struct (`sepol_compat_data`) using u32 fields
- use `compat_ptr()` to safely convert 32-bit user pointers to kernel pointers
- adding a runtime `ksu_is_compat` flag to dynamically select between struct layouts
This prevents a near-null pointer dereference when handling SELinux
policy updates from 32-bit ksud in a 64-bit kernel.
Truth table:
kernel 32 + ksud 32, struct is u32, no compat_ptr
kernel 64 + ksud 32, struct is u32, yes compat_ptr
kernel 64 + ksud 64, struct is u64, no compat_ptr
Preprocessor check
64BIT=y COMPAT=y: define both structs, select dynamically
64BIT=y COMPAT=n: struct u64
64BIT=n: struct u32
Tested-by: ...
Tested-by: ...
Tested-by: ...
Signed-off-by: backslashxx <118538522+backslashxx@users.noreply.github.com>
1. Replace `do_execveat_common` with `sys_execve` and `sys_execveat`
2. Replace `input_handle_event` with `input_event` and
`input_inject_event`
Tested on android12-5.10-2024-04, android13-5.15-2024-04.
android14-6.1-2024-04
input-event-codes.h:
Input: add input-event-codes header file
(f902dd8934)
This was in 4.4-rc, so 4.4.0 or above has it else no.
aio.h:
fs: move struct kiocb to fs.h
(e2e40f2c1e)
Below this version, we need to explicitly include aio.h for struct kiocb
This was in 4.1-rc, so 4.0 or below should do the include
uaccess.h, sched.h was present for long times, but 4.10 splited out to
include/sched/ but the current ifdef is not including uaccess.h for
lower versions than 4.4. Fix it.
Basic support for the case that init_task.mnt_ns != zygote.mnt_ns(WSA),
just copy nsproxy and fs pointers for solve #276.
Note the copy in `apk_sign.c` is not required but suggested for
secure(ensure the checked mnt_ns is what ns android running, not created
by user, although many distributions does not have user ns.).
Tested with latest release on Win10 19045.3086(with WSAPatch).
Further review required for:
- [x] Security of this operation (without locking).
- [x] The impact of these modifications on other Android distributions.
Hi @tiann.
Thanks for the great project, I had great fun playing around with it.
This PR mainly tries to further minimize the possible delays caused by
KernelSU hooking.
There are 3 major changes:
- Processes with 0 < UID < 2000 are blocked straight-up before going
through the allow_list.
I don't see any need for such processes to be interested in root, and
this allows returning early before going through a more expensive
lookup.
If there's an expected breakage due to this change, I'll remove it. Let
me know.
- A page-sized (4K) bitmap is added.
This allows O(1) lookup for UID <= 32767.
This speeds up `ksu_is_allow_uid()` by about 4.8x by sacrificing a 4K
memory. IMHO, a good trade-off.
Most notably, this reduces the 99.999% result previously from worrying
milliseconds scale to microseconds scale.
For UID > 32767, another page-sized (4K) sequential array is used to
cache allow_list.
Compared to the previous PR #557, this new approach gives another nice
25% performance boost in average, 63-96% boost in worst cases.
Benchmark results are available at
https://docs.google.com/spreadsheets/d/1w_tO1zRLPNMFRer49pL1TQfL6ndEhilRrDU1XFIcWXY/edit?usp=sharing
Thanks!
---------
Signed-off-by: Juhyung Park <qkrwngud825@gmail.com>
