This bug is similar to Jann Horn's issue (https://bugs.chromium.org/p/project-zero/issues/detail?id=851) -- credit should go to him.

The hardware service manager allows the registration of HAL services. These services are used by the vendor domain and other core processes, including system_server, surfaceflinger and hwservicemanager.

Similarly to the "regular" service manager ("servicemanager"), the hardware service manager is the context manager node for the "hwbinder" device, allowing it to mediate access to all hardware services registered under it. This is done by allowing its users to list, access or insert services into its registry, identified by a unique full-qualified name and an instance name (see http://androidxref.com/8.0.0_r4/xref/system/libhidl/transport/manager/1.0/IServiceManager.hal).

The "add" binder call allows callers to supply a binder instance to be registered with the hardware service manager. When issued, the call is unpacked by the auto-generated hidl stub, and then passed to "ServiceManager::add" for processing. Here is a snippet from that function (http://androidxref.com/8.0.0_r4/xref/system/hwservicemanager/ServiceManager.cpp#172):

1.  Return<bool> ServiceManager::add(const hidl_string& name, const sp<IBase>& service) {
2.      ...
3.      // TODO(b/34235311): use HIDL way to determine this
4.      // also, this assumes that the PID that is registering is the pid that is the service
5.      pid_t pid = IPCThreadState::self()->getCallingPid();
6. 
7.      auto ret = service->interfaceChain([&](const auto &interfaceChain) {
8.      if (interfaceChain.size() == 0) {
9.          return;
10.    }
11.
12.    // First, verify you're allowed to add() the whole interface hierarchy
13.    for(size_t i = 0; i < interfaceChain.size(); i++) {
14.        std::string fqName = interfaceChain[i];
15.        if (!mAcl.canAdd(fqName, pid)) {
16.            return;
17.        }
18.    }
19.    ...
20.}

As we can see in the snippet above, the function first records the pid of the calling process (populated into the transaction by the binder driver). Then, it issues a (non-oneway) transaction to the given service binder, in order to retrieve the list of interfaces corresponding to the given instance. As the comment correctly notes (lines 3-4), this approach is incorrect, for two reasons:

  1. The given service can be hosted in a different process to the one making the binder call
  2. Recording the pid does not guarantee that the calling process cannot transition from zombie to dead, allowing other processes to take its place

The pid is later used by the AccessControl class in order to perform the access control check, using getpidcon (http://androidxref.com/8.0.0_r4/xref/system/hwservicemanager/AccessControl.cpp#63). Consequently, an attack similar to the one proposed by Jann in the original bug is possible - namely, creating a race condition where the issuing process transitions to dead state, and a new privileged tid to be created in its place, causing the access control checks to be bypassed (by using the privileged process's SELinux context).

Furthermore, this code would have been susceptible to another vulnerability, by James Forshaw (https://bugs.chromium.org/p/project-zero/issues/detail?id=727) - namely, the caller can issue a "oneway" binder transaction in the "add" call, causing the calling pid field recorded by the driver to be zero. In such a case, getpidcon(0) is called, which would have returned the current process's context (the hardware service manager can register several critical services, including the "HIDL manager" and the "Token Manager"). However, this behaviour has since been changed in upstream libselinux (https://patchwork.kernel.org/patch/8395851/), making getpidcon(0) calls invalid, and therefore avoiding this issue.

However, an alternate exploit flow exists, which allows the issue to be exploited deterministically with no race condition required. Since the code above issues a non-oneway binder transaction on the given binder object, this allows the following attack flow to occur:

1. Process A creates a hardware binder service
2. Process A forks to create process B
3. Process B receives binder object from process A
4. Process B registers the binder object with the hardware service manager, by calling the "add" binder call
5. Hardware service manager executes "ServiceManager::add", records process B's pid, calls the (non-oneway) "interfaceChain" binder call on the given binder
6. Process A receives the "interfaceChain" binder call
7. Process A kills process B
8. Process A forks and kills the child processes, until reaching the pid before process B's pid 
9. Process A calls the "loadSoundEffects" binder call on the "audio" service, spawning a new long-lived thread in system_server ("SoundPoolThread")
10. The new thread occupies process B's pid
11. Process A completes the "interfaceChain" transaction
12. Hardware service manager uses system_server's context to perform the ACL check

This attack flow allows a caller to replace any service published by system_server, including "IBase", "ISchedulingPolicyService" and "ISensorManager", or register any other services of behalf of system_server.

Note that in order to pass the binder instance between process A and process B, the "Token Manager" service can be used. This service allows callers to insert binder objects and retrieve 20-byte opaque tokens representing them. Subsequently, callers can supply the same 20-byte token, and retrieve the previously inserted binder object from the service. The service is accessible even to (non-isolated) app contexts (http://androidxref.com/8.0.0_r4/xref/system/sepolicy/private/app.te#188).

I'm attaching a PoC which performs the aforementioned attack flow, resulting in the "IBase" service (default instance) being hijacked. Running the PoC should result in the following output:

        pid=23701
        service manager: 0x7d0b44b000
        token manager: 0x7d0b44b140
        TOKEN: 0502010000000000B78268179E69C3B0EB6AEBFF60D82B42732F0FF853E8773379A005493648BCF1
        05 02 01 00 00 00 00 00 B7 82 68 17 9E 69 C3 B0 EB 6A EB FF 60 D8 2B 42 73 2F 0F F8 53 E8 77 33 79 A0 05 49 36 48 BC F1 
        pid=23702
        service manager: 0x72e544e000
        token manager: 0x72e544e0a0
        token manager returned binder: 0x72e544e140
        Registering service...
        interfaceChain called!
        load: 0
        Killing the child PID: 0
        waitpid: 23702
        Cycling to pid
        unload: 0
        load: 0

After running the PoC, the IBase service will be replaced with our own malicious service. This can be seen be running "lshal":

        All binderized services (registered services through hwservicemanager)
        Interface                                                         Server                     Clients
        ...
        android.hidl.base@1.0::IBase/default                              23701 (<-our pid)          463

Note that this attack can also be launched from an application context (with no required permissions), as apps can access both the "hwbinder" (http://androidxref.com/8.0.0_r4/xref/system/sepolicy/private/app.te#186) and the token service (http://androidxref.com/8.0.0_r4/xref/system/sepolicy/private/app.te#188).

The attached PoC should be built as part of the Android source tree, by extracting the source files into "frameworks/native/cmds/hwservice", and running a build (e.g., "mmm hwservice"). The resulting binary ("hwservice") contains the PoC code.

It should be noted that the hardware service manager uses the PID in all other calls ("get", "getTransport", "list", "listByInterface", "registerForNotifications", "debugDump", "registerPassthroughClient") as well.

These commands are all similarly racy (due to the getpidcon(...) usage), but are harder to exploit, as no binder call takes place prior to the ACL check.


Proof of Concept:
https://gitlab.com/exploit-database/exploitdb-bin-sploits/-/raw/main/bin-sploits/43513.zip