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.
* Support deodexed ROM: This should not be done and dexpreopt is mandatory since P
Xposed: Xposed handles them just fine, at least in the latest version 89.3
suMiscL6: For whatever audio mods, a leftover of phh time
Liveboot and suBackL6: Was for CF.lumen and LiveBoot, not needed now
* Also cleanup binder sepolicies since we allow all binder transactions.
Introduce new domain `magisk_client` and new file type `magisk_exec`.
Connection to magiskd's always-on socket is restricted to magisk_client
only. Whitelisted process domains can transit to magisk_client through
executing files labelled magisk_exec. The main magisk binary shall be
the only file labelled as magisk_exec throughout the whole system.
All processes thus are no longer allowed to connect to magiskd directly
without going through the proper magisk binary.
Connection failures are silenced from audit logs with dontaudit rules,
so crazy processes which traverse through all unix domain sockets to try
connection can no longer check logcat to know the actual reason behind
EACCES, leaking the denied process policy (which is u:r:magisk:s0).
This also allows us to remove many rules that open up holes in
untrusted_app domains that were used to make remote shell work properly.
Since all processes establishing the remote shell are now restricted to
the magisk_client domain, all these rules are moved to magisk_client.
This makes Magisk require fewer compromises in Android's security model.
Note: as of this commit, requesting new root access via Magisk Manager
will stop working as Magisk Manager can no longer communicate with
magiskd directly. This will be addressed in a future commit that
involves changes in both native and application side.
For match-all-type rules (e.g. "allow magisk * * *" used in Magisk),
we used to iterate and apply rules on all existing types. However, this
is actually unnecessary as all selinux types should have at least 1
attributes assigned to it (process types "domain", file context types
"file_type" etc.). This means in order to create rules that applies to
all types, we actually only need to create rules for all attributes.
This optimization SIGNIFICANTLY reduces the patched sepolicy that is
loaded into the kernel when running Magisk. For example on Pixel 4 XL
running Android R DP4, the sepolicy sizes are
patched (before) : 3455948
patched (after) : 843176
stock : 630229
The active sepolicy size actually impacts the performance of every single
operation in the operating system, because the larger the policies gets,
the longer it takes for the kernel to lookup and match rules.
topjohnwu