Android IPC: Part 2 - Binder and Service Manager Perspective
Introduction
As mentioned in the previous article, Android uses the Binder for IPC communications. Good to know, the Binder was not created by Google. Its initial appearance was in BeOS, an old OS for mobile devices. After some acquisitions, original developers joined Android and took the Binder with them. The OpenBinder porting to Android was more implementation specific and it is a key component of the current Android OS. The official OpenBinder website is not up anymore, but there are some mirrors like this one that contain precious documentation.
High level overview
Binder is a Kernel module written in C++, mainly responsible to let processes securely, transparently and easily communicate with each other using a client-server architecture. The simplicity on how processes can interact together is awful, a client application just needs to call a method provided by the service (that is the server in the client-side architecture) and everything in between is handled by the Binder. With ‘everything in between’ I mean Location, Delivery and credentials.
When a client needs to talk to a service, he needs to locate the target service (that is, the target process). The Binder is responsible to locate the service, handle the communication, deliver messages and check for caller privileges (Credentials).
The location stage is handled by the servicemanager that acts as the endpoint mapper, it maintains a service directory that maps an interface name to a Binder handle. So, when the Binder receives a request for a specific service, it interrogates the servicemanager. The servicemanager will return a handle to it after some permission checks (for example the AID_ISOLATED mentioned in the first part) scrolling on its service list (aka service directory). If the client has permissions to interact with the requested service, the Binder will proxy the communication and deliver the message to the server, that will elaborate the request and return the result to the Binder, that will turn it to the client as a 'message'. These messages are technically called 'Parcel', containers that are written from both client and server in order to communicate serializing and deserializing necessary data (that can be parameters for clients and return values for services).
This is how IPC communications, at a higher level, are handled by Android.
Binder Introduction
Let's start with the main component of an IPC transaction, the Binder. As we said, the Binder is a small kernel module that lives in the kernel and acts as a messenger for clients and services. Every operation in Android goes through the Binder, and that's why two researchers, Nitay Artenstein and Idam Revivo, took an interesting talk at BH2014, ‘Man in the Binder: He Who Controls IPC, Controls the Droid’ (Youtube Video).
This research demonstrates an advanced post exploitation technique (a rootkit implant) where it is possible to sniff every data that uses IPC, in order to manipulate network traffic and sensitive information by hooking binder calls.
The character device at /dev/binder is read/write by everyone, any process can perform read and write operations on it using ioctl(). The ioctl() responsible to handle the IPC connection from clients (applications) is located in the 'libbinder.so' shared library, that is loaded in each application process. This library is responsible for the client initialization phase, setting up messages (aka Parcel) and talking with the binder module. We will deepen this specific library while talking, in next chapters, about the client and service implementations.
Binder interactions (userland --> kerneland)
First to introduce more concepts, let's first take an introduction on how a basic interaction works from userland to kerneland, from a client (or a service) to the binder kernel module.
As a linux based OS, ioctl system call is used to talk with the kernel module using the special character file '/dev/binder'. The driver accepts different request codes:
BINDER_SET_MAX_THREAD: Set limit thread numbers of a thread pool (deepened in the client/service implementation chapter)
BINDER_SET_CONTEXT_MGR: Set the context manager (the service manager)
BINDER_THREAD_EXIT: A thread exit the thread pool
BINDER_VERSION: Get the Binder version
BINDER_WRITE_READ: The most used code used for client and service requests
We will deepen all commands during these articles, but let's start with the BINDER_WRITE_READ request code.
The binder module source code is at drivers/android/binder.c, here the binder_ioctl() is responsible to dispatch requests received from userland based on above request codes. In the case of a BINDER_WRITE_READ code, the binder_ioctl_write_read() is triggered and parameters are handled from userland to kerneland (and vice versa) using the binder_write_read structure:
struct binder_write_read {
signed long write_size; // size of buffer by the client
signed long write_consumed; // size of buffer by the binder
unsigned long write_buffer;
signed long read_size; // size of buffer by the client
signed long read_consumed; // size of buffer by the binder
unsigned long read_buffer;
};In this structure, we have 2 main divisions: write and read items.
Write items (write_size, write_consumed, write_buffer) are used to send commands to the binder that it has to execute, meanwhile read items contain transactions from the binder to the clients that they have to execute (them whose ‘ioctl’ the binder).
For example, if a client needs to talk to a service, it will send a binder_write_read command with write items filled. When binder replies back, the client will have read items filled back. The same, a service waiting for client interactions, will receive transactions from the Binder with read items.
While talking about 'clients', I don’t mean only application clients that need to perform a request in an IPC context. A client, in this context, is a process that ioctl() the binder. For example, a service waiting for transactions from an application is a client of the binder, because it calls ioctl() in order to receive actions.
Note that in the case of a client, the ioctl() is performed when it is needed from the application (for example to perform an Inter Process Communication). Meanwhile, the service process has threads waiting in a loop for transactions from the binder.
Inside these read and write attributes we have more commands, that starts with BC_* and BR_*. The difference is in the way that the transaction is going, to or from the Binder. BR are commands received FROM the binder, while BC are commands SENT to the binder. To remind me about this difference, I think to them as they are 'BinderCall' (BC) and 'BinderReceive' (BR), but I think is not an official naming convention, so in case just use it as a reminder.
An example can be the most common TRANSACTION, we can have BC_TRANSACTION and BR_TRANSACTION.
BC_TRANSACTION is used from clients to binder, while BR_TRANSACTION is used from the binder to its clients.
The Servicemanager
As was illustrated in the High Level overview, the servicemanager is responsible for the location stage. When a client needs to interact with a service (using the Binder), the Binder will ask the servicemanager for a handle to that service.
The servicemanager source code is located at frameworks/native/cmds/servicemanager/, where service_manager.c is responsible to initialize itself and handle service related requests. Meanwhile, in binder.c (inside that path, not the kernel module) we find the code responsible to handle the communication with the binder, parsing received requests from it and sending the appropriate replies.
Servicemanager is started at boot time as defined in /init.rc file. This init file is part of the boot image and is responsible to load system partitions and binaries in the boot process:
/.../
# start essential services
start logd
start servicemanager
start hwservicemanager
start vndservicemanager
/.../
# When servicemanager goes down, restart all specified services
service servicemanager /system/bin/servicemanager
class core
user system
group system
critical
onrestart restart healthd
onrestart restart zygote
onrestart restart media
onrestart restart surfaceflinger
onrestart restart drm
onrestart restart perfhub
/.../
When the service manager is started, the main function obtains an handle to the binder ('/dev/binder') and successively call binder_become_context_manager(), that will ioctl the binder with the BINDER_SET_CONTEXT_MGR command, in order to declare itself as the context manager.
The context manager is crucial for the binder as it serves as the service locator. When the binder needs to locate a service, it asks for a handle to its context manager.
Once the registration with the binder it's done, it calls binder_loop (from binder.c) with a callback function parameter. This callback (svcmgr_handler) will be responsible to handle service related requests.
The binder_loop assignment, as the name says, is to start an infinite loop that will receive requests from the binder. Before this loop, it will call the binder with BC_ENTER_LOOPER, to inform the binder that a specific thread is joining the thread pool. The thread pool is a group of threads that are waiting for incoming messages from the Binder, usually services have multiple threads in order to handle multiple requests. By the way, the service manager is a single-threaded service, so this is the first and unique thread.
After this notification, the servicemanager starts its infinite loop that continuously asks the binder (using ioctl) waiting for actions. This is managed using the BINDER_WRITE_READ command (to the binder) with a binder_write_read structure that will be filled by the binder in its read_* items, this is the structure:
struct binder_write_read {
signed long write_size; // size of buffer by the client
signed long write_consumed; // size of buffer by the binder
unsigned long write_buffer;
signed long read_size; // size of buffer by the client
signed long read_consumed; // size of buffer by the binder
unsigned long read_buffer;
};When the binder needs the service manager to perform an action (e.g. getting a handle to a service) it will return to the binder_loop() a binder_write_read structure with read_buffer filled with the requested transaction (and in read_consumed its actual size). These two values are passed over the binder_parse() function that will start to ‘deserialize’ the transaction request:
/ .. /
uintptr_t end = ptr + (uintptr_t) size; // end calculated using the bwr.read_consumed
while (ptr < end) {
uint32_t cmd = *(uint32_t *) ptr; // the command is read from the buffer
ptr += sizeof(uint32_t);
// switch case on the received command
switch(cmd) {
// BR_NOOP is a command
case BR_NOOP:
break;
/../The first 32 bits of the bwr.read_buffer contains the command to be executed (CMD).
There is a huge list of handled commands: BR_NOOP, BR_TRANSACTION_COMPLETE, BR_INCREFS - BR_ACQUIRE - BR_RELEASE, BR_DECREFS, BR_DEAD_BINDER, BR_FAILED_REPLY - BR_DEAD_REPLY, BR_TRANSACTION, BR_REPLY.
You can find a lot more BR_* commands, but these are the only ones handled by the servicemanager. For example, a normal service can receive a SPWAN_LOOPER command from the binder, that requests the service to spawn a new thread in order to handle more requests. We said that the servicemanager is single thread, so there is no sense to receive this type of requests, so they are not handled. We will better deepen on these commands that are used by other services in IPCThreadState.cpp in next articles.
After having extracted the command from the binder_read_write structure, this one is inserted in a switch case where above commands are managed. The most interesting one is the BR_TRANSACTION, because it means that the binder needs to retrieve a service handle or register a new service.
BR_TRANSACTION
Following the source code, we can encounter and deppen some essentials structures, such as the binder_transaction data that is casted from bwr.read_buffer (now referenced in the local function as ptr) + sizeof(uint32_t), that’s because the first 32 bits are dedicated to the command constant.
struct binder_transaction_data *txn = (struct binder_transaction_data *)
ptr;
This is the binder_transaction_data structure:
//https://android.googlesource.com/kernel/msm/+/android-6.0.1_r0.74/drivers/staging/android/uapi/binder.h
struct binder_transaction_data {
/* The first two are only used for bcTRANSACTION and brTRANSACTION,
* identifying the target and contents of the transaction.
*/
union {
__u32 handle; /* target descriptor of command transaction */
binder_uintptr_t ptr; /* target descriptor of return transaction */
// in BR_TRANSACTION this must be BINDER_SERVICE_MANAGER or the service_manager return -1
} target;
binder_uintptr_t cookie; /* target object cookie */
__u32 code; /* transaction command. */ // e.g. SVC_MGR_GET_SERVICE
/* General information about the transaction. */
__u32 flags;
pid_t sender_pid;
uid_t sender_euid;
binder_size_t data_size; /* number of bytes of data */
binder_size_t offsets_size; /* number of bytes of offsets */
/* If this transaction is inline, the data immediately
* follows here; otherwise, it ends with a pointer to
* the data buffer.
*/
union {
struct {
/* transaction data */
binder_uintptr_t buffer;
/* offsets from buffer to flat_binder_object structs */
binder_uintptr_t offsets;
} ptr;
__u8 buf[8];
} data;
};
This structure contains necessary information about the incoming request, such as the sender PID and UID to check permissions for a service, the target descriptor and the transaction command for the service manager (for example PING_TRANSACTION or SVC_MGR_CHECK_SERVICE).
This binder_transaction_data structure initializes a new binder_io (binder I/O) structure using bio_init_from_txn(), that will copy data and offsets from binder_transaction_data to this new one.
struct binder_io
{
char *data; /* pointer to read/write from */
binder_size_t *offs; /* array of offsets */
size_t data_avail; /* bytes available in data buffer */
size_t offs_avail; /* entries available in offsets array */
char *data0; /* start of data buffer */
binder_size_t *offs0; /* start of offsets buffer */
uint32_t flags;
uint32_t unused;
};
bio_* functions refers to operations on the binder_io structure, here is an example on how the that structure is filled from binder_transaction_data:
void bio_init_from_txn(struct binder_io *bio, struct binder_transaction_data *txn)
{
bio->data = bio->data0 = (char *)(intptr_t)txn->data.ptr.buffer;
bio->offs = bio->offs0 = (binder_size_t *)(intptr_t)txn->data.ptr.offsets;
bio->data_avail = txn->data_size;
bio->offs_avail = txn->offsets_size / sizeof(size_t);
bio->flags = BIO_F_SHARED;
}As we can see, buffer and offsets (including their size) of binder_transaction_data are filled in their relative binder_io structure, and both structures are passed over the service manager callback function (the svcmgr_handler() function defined in service_manager.c while calling binder_loop()):
res = func(bs, txn, &msg, reply);
// func -> binder_handle defined in service_manager.c - binder_loop(bs, svcmgr_handler);
// bs -> binder_state
// txn -> binder_transaction_data
// msg -> binder_io initialized from binder_transaction_data
// reply -> an empty binder_io that will contain the reply from the service managerNow, the BR_TRANSACTION is inside the svcmgr_handler().
The binder_transaction_data.ptr must contain BINDER_SERVICE_MANAGER in order to continue (otherwise return -1) and binder_transaction_data.code contains the service manager command. These service commands (dispatched in a switch condition) can be:
PING_TRANSACTION: It is a ping to the servicemanager, so return 0.
SVC_MGR_GET_SERVICE - SVC_MGR_CHECK_SERVICE: Get a handle to a service. They follow the same switch path
SVC_MGR_ADD_SERVICE - Add a new service
SVC_MGR_LIST_SERVICES - List all available services
Let's start to dig into SVC_MGR_GET_SERVICE.
SVC_MGR_GET_SERVICE
This service command occurs when the binder needs a service handle based on a service name (requested from a client).
The service name is taken from the binder_io structure (referred as `msg` in the source) using bio_get_string16().
We have different functions as bio_get_* (bio_get_uint32, bio_get_string16, _bio_get_obj, bio_get_ref). They are all primitives of bio_get() that retrieves the requested data type from (binder_io)bio->data. The same for bio_put_* functions in order to insert data in a binder_io structure while replying to a command.
do_find_service() is called with the service name and the caller UID and PID from the binder_transaction_data structure and immediately call find_svc(), that will iterate its service singularly linked list and return an `svcinfo` structure if match the requested service name:
{
struct svcinfo *next; // pointer to the next service
uint32_t handle;
struct binder_death death;
int allow_isolated;
uint32_t dumpsys_priority;
size_t len;
uint16_t name[0];
} svcinfo;
The svcinfo structure mainly contains information about the target service.
If the service matches the svcinfo.name item, the structure is returned to the do_find_service() function, that is responsible to perform extra checks.
The first check is about process isolation. As we were talking in the first part of this series, some services are not allowed to be called from isolated apps (such as. web browsers):
if (!si->allow_isolated) {
uid_t appid = uid % AID_USER;
if (appid >= AID_ISOLATED_START && appid <= AID_ISOLATED_END) {
return 0;
}
}In this piece of code, the UID retrieved from binder_transaction_data struct (coming from the binder) is verified against AID_ISOLATED_START and AID_ISOLATED_END. These UIDs (a range from 99000 to 99999) are associated with isolated processes and they can interact only with services with svcinfo.allow_isolated set to True.
If this check is passed, a SELINUX permission checks if the sender is allowed to retrieve the service, and the handler is returned to the main switch case in the service handler. The returned handle will be put inside the binder_io reply using `bio_put_ref()` and return 0, meaning everything is fine. Later on we will see how the message is sent back to the binder.
SVC_MGR_LIST_SERVICES
We can also list available services with the SVC_MGR_LIST_SERVICES command, that will iterate through the service list (svclist) and put the result in the binder_io reply message using bio_put_string16(). There is also an interesting condition on dumpsys_priority. The priority, that can be defined while registering a new service, can be of three levels: CRITICAL, HIGH and NORMAL. While listing all services, we can choose to dump only services with a specific priority (specified in the svclist structure).
For example, using the dumpsys utility in Android, we can specify the desired level:
adb > `dumpsys -l --priority CRITICAL`
Currently running services:
SurfaceFlinger
activity
cpuinfo
input
notification
window
`dumpsys -l --priority HIGH`
Currently running services:
connectivity
meminfo
adb > `dumpsys -l --priority NORMAL`
Currently running services:
activity
connectivity
notificationSVC_MGR_ADD_SERVICE
If the requested command from the binder is SVC_MGR_ADD_SERVICE, the binder is proxying a client request to register a new service. Details about the new service are taken from the binder_io message (binder_io->data). Service attributes are the service name, the priority level (dumpsys_priority), the handle and if it is permitted to interact with the service from isolated apps (allow_isolation). The function do_add_service() is called with this information and the caller UID and PID from the binder_transaction_data message.
This function is responsible to check for caller permissions (the process that requests the registration), starting by checking its UID to avoid the creation of a new service from standard applications. This is accomplished by checking if the AID_APP is over 10000.
In Android, installed applications start from UID 10000, so the condition is aimed to prevent an user application from installing a new service (or override an existing one). That also means that the privileged system user (with UID 1000) can register a new service.
If this condition is satisfied, a SELINUX check controls that the caller process has 'add' permissions. If the caller process has rights to register a new service, find_svc() checks if the service name has been already registered. If it already exists, the service handle is overridden with the new one, and svcinfo_death() called.
Before going in depth with this function behaviour, let's introduce the scenario where the service does not exist:
struct svcinfo *si;
/../
si->handle = handle;
si->len = len;
memcpy(si->name, s, (len + 1) * sizeof(uint16_t));
si->name[len] = '\0';
si->death.func = (void*) svcinfo_death;
si->death.ptr = si;
si->allow_isolated = allow_isolated;
si->dumpsys_priority = dumpsys_priority;
si->next = svclist;
svclist = si;
/../
The code is pretty self-explanatory, it is populating the new structure with input values and updates its service list with `si->next = svclist` and `svclist = si ` (linked list behavior). And here, we are back with the death that we were talking some lines above.
The binder_death structure, part of the svclist, contains two items, func and ptr. The ptr is a pointer to its service structure (itself), and the func is a function pointer pointing to svcinfo_death().
This death function sets the service handle to 0 and informs the binder that the service is dead using a BC_RELEASE with the service handle as parameter, so the binder can release this reference. The binder can use this information to also inform associated clients that the service is down using BR_BINDER_DOWN, if clients requested for it (by sending a BC_REQUEST_DEATH_NOTIFICATION for the service to the binder).
On the other side, when a service is registered or overridden, a BC_ACQUIRE with the service handle as parameter is sent to the binder, also with the BC_REQUEST_DEATH_NOTIFICATION in case the service goes down (for example if its crashes).
Comeback to the service handler
When one of these described commands are executed, the Binder usually expects a reply back.
While handling commands, SVC_MGR_ADD_SERVICE puts 0 in reply message if success (bio_put_uint32(reply, 0);) or simply return -1 if something fail, and the binder will receive an empty reply (that was previously initialized using bio_init()).
SVC_MGR_GET_SERVICE and SVC_MGR_LIST_SERVICES act in the same way if something goes wrong (-1 and empty reply packet), or they will return 0 to the function after have filled the reply packet with necessary values (handle in case of get service, and the list of services for the service list command).
When the service handler returns, the execution flow comes back inside the `binder_parse()` function (in the BR_TRANSACTION switch case) with the reply packet and the result value of the servicemanager handler. Based on the binder_transaction_data.flags, if TF_ONE_WAY is set, means that is an asynchronous call, the binder does not expect a reply, so the servicemanager will inform the binder to free the binder_transaction_data.ptr.buffer with a BC_FREE_BUFFER command (internally using the binder_free_buffer() function). If it's not an asynchronous call, it will send the reply back to the binder using binder_send_reply() that will send a BC_REPLY command.
Also, as you could notice, all these functions (binder_send_reply, binder_free_buffer, ..) are meant to be easily called inside the source code, and will perform all setup operations to interact with the binder with the final ioctl(). Let's take a simple example of the binder_free_buffer mentioned before.
void binder_free_buffer(struct binder_state *bs,
binder_uintptr_t buffer_to_free)
{
struct {
uint32_t cmd_free;
binder_uintptr_t buffer;
} __attribute__((packed)) data;
data.cmd_free = BC_FREE_BUFFER;
data.buffer = buffer_to_free;
binder_write(bs, &data, sizeof(data));
}This function, previously used by the service manager handler to inform the binder to free a buffer, will setup, using a data structure, a cmd_free with BC_FREE_BUFFER on it and the buffer to free, then call binder_write(). binder_write is the final function that will put received input inside a binder_write_read.write_buffer structure before ioctl the binder with the BINDER_WRITE_READ command:
int binder_write(struct binder_state *bs, void *data, size_t len)
{
struct binder_write_read bwr;
int res;
bwr.write_size = len;
bwr.write_consumed = 0;
bwr.write_buffer = (uintptr_t) data;
bwr.read_size = 0;
bwr.read_consumed = 0;
bwr.read_buffer = 0;
res = ioctl(bs->fd, BINDER_WRITE_READ, &bwr);
if (res < 0) {
fprintf(stderr,"binder_write: ioctl failed (%s)\n",
strerror(errno));
}
return res;
}We can note the differences on the usage of the binder_write_read structure now and before.
When we were expecting an action from the binder (in the binder_loop) the received action was inside the read_buffer (that contains the BR_* command). Now, the binder needs to perform actions based on our input, so we are using the write_buffer (with a a BC_* command).
Said that, we can come back inside the binder_send_reply(), that is responsible to send the reply of performed actions to the Binder. This is the source code:
void binder_send_reply(struct binder_state *bs,
struct binder_io *reply,
binder_uintptr_t buffer_to_free,
int status)
{
struct {
uint32_t cmd_free;
binder_uintptr_t buffer;
uint32_t cmd_reply;
struct binder_transaction_data txn;
} __attribute__((packed)) data;
data.cmd_free = BC_FREE_BUFFER;
data.buffer = buffer_to_free;
data.cmd_reply = BC_REPLY;
data.txn.target.ptr = 0;
data.txn.cookie = 0;
data.txn.code = 0;
if (status) {
// the svcmgr_handler return -1
data.txn.flags = ;
data.txn.data_size = sizeof(int);
data.txn.offsets_size = 0;
data.txn.data.ptr.buffer = (uintptr_t)&status;
data.txn.data.ptr.offsets = 0;
} else {
// the svcmgr_handler return 0
data.txn.flags = 0;
data.txn.data_size = reply->data - reply->data0;
data.txn.offsets_size = ((char*) reply->offs) - ((char*) reply->offs0);
data.txn.data.ptr.buffer = (uintptr_t)reply->data0;
data.txn.data.ptr.offsets = (uintptr_t)reply->offs0;
}
binder_write(bs, &data, sizeof(data));
}Note the declared data structure, it contains an integer cmd_free (that will be BC_FREE_BUFFER), the buffer, the cmd_reply (that will be BC_REPLY) and a binder_transaction_data structure.
The buffer to free is binder_transaction_data.data.ptr.buffer (previously casted in binder_io. it contains 'parameters', for example a service name for the servicemanager) and then the structure is filled based on the status value.
The status value is the return value from the servicemanager handler (svcmgr_handler) that can be 0 if everything was fine (and the reply was filled) or -1 if something went wrong.
If the result is -1, this result is copied inside the data.txn.data.ptr.buffer (so inside the binder_transaction_data of the data structure).
If the result of the service manager handler was fine (0), the binder_transaction_data is filled with reply's data/offset buffers and passed over the binder_write() function, that, as explained before, will take the data structure and put it in binder_write_read.write_buffer before calling ioctl() with the BINDER_WRITE_READ command.
Little resume
The servicemanager is started by the init process (as defined in /init.rc) and first of all it becomes the context manager for the binder. Then he notices the binder that will enter an infinite loop (BC_ENTER_LOOPER) and starts to read and parse operations delivered from the binder. When such events are related to service lookup or service registration (SVC_MGR_GET_SERVICE and SVC_MGR_ADD_SERVICE) the binder request the servicemanager for a BR_TRANSACTION with one of these commands inside its binder_transaction_data structure. The servicemanager checks for necessary rights on the caller process (information sent from the binder) and, in the case of a service lookup, returns an handle to the binder. When it's done, the reply is sent to the binder using ioctl with BINDER_WRITE_READ with the reply inside the write_buffer and the BC_REPLY command.
Conclusions
In this post, we concentrated on transactions between the Binder and the servicemanager. This is a crucial component for IPC. In the following blog post, we will deepen the client and service perspective for IPC transactions.
(Alessandro Groppo)
_0_o.png)
