Previously, magic mount creates its own mirror devices and mount
mirror mount points. With these mirror mount points, magic mount
can get the original files and directory trees. However, some
devices use overlayfs to modify some mount points, and thus after
magic mount, the overlayed files are missing because the mirror
mount points do not contain the overlayed files. To address this
issue and make magic mount more compatible, this patch refactors
how magic mount works.
The new workflows are as follows:
1. make MAGISKTMP a private mount point so that we can create the
private mount points there
2. for mirror mount points, we instead of creating our own mirror
devices and mount the mirror mount points, we "copy" the
original mount points by recursively mounting /
3. to prevent magic mount affecting the mirror mount points, we
recursively set the mirror mount points private
4. to trace the mount points we created for reverting mounts, we
again make the mirror mount points shared, and by this way we
create a new peer group for each mirror mount points
5. as for tracing the newly created tmpfs mount point by magic
mount, we create a dedicated tmpfs mount point for them, namely
worker mount point, and obviously, it is shared as in a newly
created peer group for tracing
6. when reverting mount points by magic mount, we can then trace
the peer group id and unmount the mount points whose peer group
ids are created by us
The advantages are as follows:
1. it is more compatible, (e.g., with overlayfs, fix#2359)
2. it can mount more partitions for which previous implementation
cannot create mirror mount points (fix#3338)
It turns out that decompressing and recompressing the kernel is enough to break booting on many devices that use MT6763.
Fix#5124, fix#6204, fix#6566
Co-authored-by: LoveSy <shana@zju.edu.cn>
Co-authored-by: 南宫雪珊 <vvb2060@gmail.com>
We resolve available partitions for sepolicy.rules when patching
boot and bind mount the partition by magiskinit.
For older devices, the previous logic won't work because the part name
is never readable.
Co-authored-by: topjohnwu <topjohnwu@gmail.com>
Some Android 11+ devices have the /sbin partition but not accessible by the global shell (`PATH` doesn't contain `/sbin`). Not only custom ROMs but also some stock ROMs have the same behavior so I believe it is something we need to deal with.
Fix#6427, fix#4309, fix#5728, fix#3593
If remote process died, `xreadlink` fails and leaves `buf` uninitialized. Then the daemon calls `str_ends`, creates a temp `std::string_view` with the uninitialized buffer and undefined behavior occurs.
`daemon_entry` calls `getprop` which initializes sysprop impl and checks whether we need to load persistent property file. On FDE devices, magiskd starts before /data is actually decrypted, and the check always fails. Thus `persist_getprop("persist.sys.safemode")` will always fail.
With Android 13 GKI kernels, the boot partition has no ramdisk, so
Magisk constructs one from scratch. In this scenario, there's no backup
init binary at /.backup/init. For normal boot, magiskinit will symlink
/init -> /system/bin/init if needed. This commit implements the same
for booting into recovery. Before, magiskinit would just exec itself
over and over again because it couldn't restore the backup init.
Signed-off-by: Andrew Gunnerson <chillermillerlong@hotmail.com>
Meizu devices using 2SI won't switch root to /system and still on rootfs, and /init is the 1st stage's, which cannot handle the 2nd stage. So we have to manually execute /system/bin/init for the 2nd stage.
Many Amlogic devices (e.g. FireTV 2nd gen Cube, Vero 4k+, MI Smart Speaker, etc.) are A-only with androidboot.slot_suffix=normal argument. I think "normal" actually means A-only in this case so just ignore it.
Fix topjohnwu#5806
The hijacked load node does not need to be a FIFO. A FIFO is only
required for blocking init's control flow, which is already achieved
by hijacking the enforce node.
Previously `read_string()` calls `std::string.resize()` with a int read from remote process. When I/O error occurs, -1 will be used for resizing the string, `std::bad_alloc` is thrown and since magisk is compiled with `-fno-exceptions`, it will crash the whole daemon process.
May fix topjohnwu#5681
Since Android 13, sepolicy are also loaded from APEX modules. Part
of the change is to run restorecon before SELinux is set to enforce.
In order to support this situation, we also hijack plat_file_contexts
if necessary to properly order our operations.
Original idea credits to @yujincheng08, close#5603
On older Android versions, pre-mounting selinuxfs will lead to errors,
so we have to use a different method to block init's control flow.
Since all devices that falls in this catagory must both:
1. Be Android 8.0 - 9.0
2. Have early mount fstab in its device tree
We can actually use the same FIFO trick, but this time not on selinuxfs,
but on the read-only device tree nodes in sysfs or procfs. By mocking
the fstab/compatible node in the device tree, we can block init when
it attempts to do early mount; at that point, we can then mock selinuxfs
as we normally would, successfully hijack and inject patched sepolicy.
In the current implementation, Magisk will either have to recreate
all early mount implementation (for legacy SAR and rootfs devices) or
delegate early mount to first stage init (for 2SI devices) to access
required partitions for loading sepolicy. It then has to recreate the
split sepolicy loading implementation in-house, apply patches, then
dump the compiled + patched policies into monolithic format somewhere.
Finally, it patches the original init to force it to load the sepolicy
file we just created.
With the increasing complexity involved in early mount and split
sepolicy (there is even APEX module involved in the future!),
it is about time to rethink Magisk's sepolicy strategy as rebuilding
init's functionality is not scalable and easy to maintain.
In this commit, instead of building sepolicy ourselves, we mock
selinuxfs with FIFO files connected to a pre-init daemon, waiting
for the actual init process to directly write the sepolicy file into
MagiskInit. We then patch the file and load it into the kernel. Some
FIFO tricks has to be used to hijack the original init process's
control flow and prevent race conditions, details are directly in the
comments in code.
At the moment, only system-as-root (read-only root) support is added.
Support for legacy rootfs devices will come with a follow up commit.
Design credit to @yujincheng08
Close#5146. Fix#5491, fix#3752
Previously, Magisk changes the mount point from /system to /system_root
by patching fstab to prevent the original init from changing root.
The reason why we want to prevent the original init from switching the
root directory is because it will then be read-only, making patching
and injecting magiskinit into the boot chain difficult.
This commit (ab)uses the fact that the /data folder will never be part
of early mount (because it is handled very late in the boot by vold),
so that we can use it as the mount point of tmpfs to store files.
Some advantages of this method:
- No need to switch root manually
- No need to modify fstab, which significantly improves compatibility
e.g. avoid hacks for weird devices like those using oplus.fstab,
and avoid hacking init to bypass fstab in device trees
- Supports skip_mount.cfg
- Support DSU
In the constructor of mmap_data, there are two possible values when fails: nullptr if fstat() fails, and MAP_FAILED if mmap() fails, but mmap_data treated MAP_FAILED as valid address and crashes.
Samsung FDE devices with the "persist.sys.zygote.early=true" property will cause Zygote to start before post-fs-data. According to Magisk's document, the post-fs-data phase should always happen before Zygote is started. Features assuming this behavior (like Zygisk and modules that need to control zygote) will not work. To avoid breaking existing modules, we simply invalidate this property to prevent this non-standard behavior from happening
Fix#5299, fix#5328, fix#5308
Co-authored-by: LoveSy <shana@zju.edu.cn>
* Further fix `oplus.fstab` support
In some oneplus devices, `oplus.fstab` does exists but `init` never
loaded it and those entries in `oplus.fstab` are written directly to
`fstab.qcom`. Previous implementation will introduce duplicate entries
to `fstab.qcom` and brick the device. This commit filters those entries
from `oplus.fstab` that are already in `fstab.qcom` and further filters
duplicated entries in `oplus.fstab` (keep only the last entry).
Fix#5016
* Fix UB
Since we moved entry, we need to explicitly copy its member.
For c++23 we can use `auto{}`.
- Use ftruncate64 instead of ftruncate to workaround seccomp
- Cast uint32_t to off64_t before making it negative
Note: Using ftruncate with a modern NDK libc should actually be
fine as the syscall wrapper in bionic will use ftruncate64 internally.
However, since we are using the libc.a from r10e built for Gingerbread,
seccomp wasn't a thing back then, and also the ftruncate64 symbol is
missing; we have to create our own wrapper and call it instead on
32-bit ABIs.
Props to @jnotuo for discovering the overflow bug and seccomp issue
Fix#3703, close#4915
`operator==` of string_view will create a tmp `string_view`.
It's an UB if the `const char *` is a nullptr.
`fdt_get_name` however will return a nullptr.
Samsung Galaxy A21S and Galaxy M12, probably others, are hdr_v2 boot.img with 2SI judging by the ramdisk contents, but the dtb contains an extra cmdline with skip_initramfs present, even though this shouldn't exist on 2SI and the kernel apparently doesn't even contain a skip_initramfs function
I can't find examples of other devices where skip_initramfs is present in the dtb other than these so patch it out like we do the kernel
Co-authored-by: topjohnwu <topjohnwu@gmail.com>
Custom ROM bring-ups of legacy Sony devices contain the following:
/init (symlink to /bin/init_sony)
/init.real (the "real" Android init)
/bin/init_sony (this was /sbin/init_sony on Android <11)
Kernel loads the ramdisk and starts /init -> /bin/init_sony
/bin/init_sony does low-level device setup (see: https://github.com/LineageOS/android_device_sony_common/blob/lineage-18.1/init/init_main.cpp)
/bin/init_sony unlinks /init and renames /init.real to /init
/bin/init_sony starts /init
Since init_sony needs to run first magiskinit needs to replace init.real instead, so add workarounds based on detection of init.real to boot patcher and uninstaller
Thanks @115ek and @bleckdeth
Fixes#3636
Co-authored-by: topjohnwu <topjohnwu@gmail.com>
Fix topjohnwu#4810
> [ 2.927463] [1: init: 1] magiskinit: Replace [/system/etc/selinux/plat_sepolicy.cil] -> [xxx]
[ 2.936801] [1: init: 1] magiskinit: write failed with 14: Bad address
Since topjohnwu#4596, magisk fails to patch `/init`, xwrite() fails with EFAULT, break the original `/init` file and make the device unbootable. Reverting this commit for legacy rootfs devices fixes the problem. I think this is a Samsung kernel magic since currently I can't reproduce this on other devices or find something special in the log currently we have.
- The lambda here infers its return type as `std::string`,
and since `info` is `const`, the labmda copies `info.name`
and returns a `std::string&&`. After captured by the
`std::string_view`, the `std::string&&` return value
deconstructs and makes `std::string_view` refers to a
dangling pointer.
