Zygote: clean up comment for isolateAppData()
This comment has several inaccuracies, such as saying that tmpfs is
mounted on /mnt/expand (it's actually /mnt/expand/$volume/user{,_de}),
and referring to DE storage when CE storage is meant. Fix these
inaccuracies, and rework it to explain more clearly what's going on.
Change-Id: I97106f1c2551906c12a96f17934db9253a7fe605
This commit is contained in:
@@ -1177,38 +1177,40 @@ static void relabelAllDirs(const char* path, const char* context, fail_fn_t fail
|
||||
}
|
||||
|
||||
/**
|
||||
* Make other apps data directory not visible in CE, DE storage.
|
||||
* Hide the CE and DE data directories of non-related apps.
|
||||
*
|
||||
* Apps without app data isolation can detect if another app is installed on system,
|
||||
* by "touching" other apps data directory like /data/data/com.whatsapp, if it returns
|
||||
* "Permission denied" it means apps installed, otherwise it returns "File not found".
|
||||
* Traditional file permissions or SELinux can only block accessing those directories but
|
||||
* can't fix fingerprinting like this.
|
||||
* We fix it by "overlaying" data directory, and only relevant app data packages exists
|
||||
* in data directories.
|
||||
* Without this, apps can detect if any app is installed by trying to "touch" the app's CE
|
||||
* or DE data directory, e.g. /data/data/com.whatsapp. This fails with EACCES if the app
|
||||
* is installed, or ENOENT if it's not. Traditional file permissions or SELinux can only
|
||||
* block accessing those directories but can't fix fingerprinting like this.
|
||||
*
|
||||
* Instead, we hide non-related apps' data directories from the filesystem entirely by
|
||||
* mounting tmpfs instances over their parent directories and bind-mounting in just the
|
||||
* needed app data directories. This is done in a private mount namespace.
|
||||
*
|
||||
* Steps:
|
||||
* 1). Collect a list of all related apps (apps with same uid and allowlisted apps) data info
|
||||
* (package name, data stored volume uuid, and inode number of its CE data directory)
|
||||
* 2). Mount tmpfs on /data/data, /data/user(_de) and /mnt/expand, so apps no longer
|
||||
* able to access apps data directly.
|
||||
* 3). For each related app, create its app data directory and bind mount the actual content
|
||||
* from apps data mirror directory. This works on both CE and DE storage, as DE storage
|
||||
* is always available even storage is FBE locked, while we use inode number to find
|
||||
* the encrypted DE directory in mirror so we can still bind mount it successfully.
|
||||
*
|
||||
* Example:
|
||||
* 0). Assuming com.android.foo CE data is stored in /data/data and no shared uid
|
||||
* 1). Mount a tmpfs on /data/data, /data/user, /data/user_de, /mnt/expand
|
||||
* List = ["com.android.foo", "null" (volume uuid "null"=default),
|
||||
* 123456 (inode number)]
|
||||
* 2). On DE storage, we create a directory /data/user_de/0/com.com.android.foo, and bind
|
||||
* mount (in the app's mount namespace) it from /data_mirror/data_de/0/com.android.foo.
|
||||
* 3). We do similar for CE storage. But in direct boot mode, as /data_mirror/data_ce/0/ is
|
||||
* encrypted, we can't find a directory with name com.android.foo on it, so we will
|
||||
* use the inode number to find the right directory instead, which that directory content will
|
||||
* be decrypted after storage is decrypted.
|
||||
* (1) Collect a list of all related apps (apps with same uid and allowlisted apps) data info
|
||||
* (package name, data stored volume uuid, and inode number of its CE data directory)
|
||||
* (2) Mount tmpfs on /data/data and /data/user{,_de}, and on /mnt/expand/$volume/user{,_de}
|
||||
* for all adoptable storage volumes. This hides all app data directories.
|
||||
* (3) For each related app, create stubs for its data directories in the relevant tmpfs
|
||||
* instances, then bind mount in the actual directories from /data_mirror. This works
|
||||
* for both the CE and DE directories. DE storage is always unlocked, whereas the
|
||||
* app's CE directory can be found via inode number if CE storage is locked.
|
||||
*
|
||||
* Example assuming user 0, app "com.android.foo", no shared uid, and no adoptable storage:
|
||||
* (1) Info = ["com.android.foo", "null" (volume uuid "null"=default), "123456" (inode number)]
|
||||
* (2) Mount tmpfs on /data/data, /data/user, and /data/user_de.
|
||||
* (3) For DE storage, create a directory /data/user_de/0/com.android.foo and bind mount
|
||||
* /data_mirror/data_de/0/com.android.foo onto it.
|
||||
* (4) Do similar for CE storage. But if the device is in direct boot mode, then CE
|
||||
* storage will be locked, so the app's CE data directory won't exist at the usual
|
||||
* path /data_mirror/data_ce/0/com.android.foo. It will still exist in
|
||||
* /data_mirror/data_ce/0, but its filename will be an unpredictable no-key name. In
|
||||
* this case, we use the inode number to find the right directory instead. Note that
|
||||
* the bind-mounted app CE data directory will remain locked. It will be unlocked
|
||||
* automatically if/when the user's CE storage is unlocked, since adding an encryption
|
||||
* key takes effect on a whole filesystem instance including all its mounts.
|
||||
*/
|
||||
static void isolateAppData(JNIEnv* env, const std::vector<std::string>& merged_data_info_list,
|
||||
uid_t uid, const char* process_name,
|
||||
|
||||
Reference in New Issue
Block a user