arc/mount-passthrough/mount-passthrough.cc

/*
 * Copyright 2016 The Chromium OS Authors. All rights reserved.
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#define FUSE_USE_VERSION 26

#include <base/command_line.h>
#include <base/files/file_path.h>
#include <base/logging.h>
#include <base/notreached.h>
#include <base/strings/string_split.h>
#include <base/strings/stringprintf.h>
#include <brillo/flag_helper.h>
#include <brillo/syslog_logging.h>
#include <dirent.h>
#include <errno.h>
#include <fcntl.h>
#include <fuse/fuse.h>
#include <fuse/fuse_common.h>
#include <fuse/fuse_lowlevel.h>
#include <linux/limits.h>
#include <signal.h>
#include <stdint.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/vfs.h>
#include <unistd.h>

#include <algorithm>
#include <fstream>
#include <sstream>
#include <string>

#define USER_NS_SHIFT 655360
#define CHRONOS_UID 1000
#define CHRONOS_GID 1000
// Android's media_rw UID and GID shifted by USER_NS_SHIFT.
#define AID_MEDIA_RW_UID 656383
#define AID_MEDIA_RW_GID 656383

#define WRAP_FS_CALL(res) ((res) < 0 ? -errno : 0)

namespace {

constexpr uid_t kAndroidAppUidStart = 10000 + USER_NS_SHIFT;
constexpr uid_t kAndroidAppUidEnd = 19999 + USER_NS_SHIFT;

struct FusePrivateData {
  std::string android_app_access_type;
  bool force_group_permission = false;
};

// Given android_app_access_type, figure out the source of /storage mount in
// Android.
std::vector<std::string> get_storage_source(
    const std::string& android_app_access_type) {
  std::string storage_source;
  // Either full (if no Android permission check is needed), read (for Android
  // READ_EXTERNAL_STORAGE permission check), or write (for Android
  // WRITE_EXTERNAL_STORAGE_PERMISSION).
  if (android_app_access_type == "full") {
    return {};
  } else if (android_app_access_type == "read") {
    // We allow apps with both READ_EXTERNAL_STORAGE and WRITE_EXTERNAL_STORAGE
    // to access the read view. This is useful for MyFiles so that we can expose
    // a read-only view (this one) as a second mount point under
    // /mnt/runtime/write.
    return {"/runtime/read", "/runtime/write"};
  } else if (android_app_access_type == "write") {
    return {"/runtime/write"};
  } else {
    NOTREACHED();
    return {"notreached"};
  }
}

// Perform the following checks (only for Android apps):
// 1. if android_app_access_type is read, checks if READ_EXTERNAL_STORAGE
// permission is granted
// 2. if android_app_access_type is write, checks if WRITE_EXTERNAL_STORAGE
// permission is granted
// 3. if android_app_access_type is full, performs no check.
// Caveat: This method is implemented based on Android storage permission that
// uses mount namespace. If Android changes their permission in the future
// release, than this method needs to be adjusted.
int check_allowed() {
  fuse_context* context = fuse_get_context();
  // We only check Android app process for the Android external storage
  // permissions. Other kind of permissions (such as uid/gid) should be checked
  // through the standard Linux permission checks.
  if (context->uid < kAndroidAppUidStart || context->uid > kAndroidAppUidEnd) {
    return 0;
  }

  std::vector<std::string> storage_source =
      get_storage_source(static_cast<FusePrivateData*>(context->private_data)
                             ->android_app_access_type);
  // No check is required because the android_app_access_type is "full".
  if (storage_source.empty()) {
    return 0;
  }

  std::string mountinfo_path =
      base::StringPrintf("/proc/%d/mountinfo", context->pid);
  std::ifstream in(mountinfo_path);
  if (!in.is_open()) {
    PLOG(ERROR) << "Failed to open " << mountinfo_path;
    return -EPERM;
  }
  while (!in.eof()) {
    std::string line;
    std::getline(in, line);
    if (in.bad()) {
      return -EPERM;
    }
    std::vector<std::string> tokens = base::SplitString(
        line, " ", base::KEEP_WHITESPACE, base::SPLIT_WANT_ALL);
    if (tokens.size() < 5) {
      continue;
    }
    std::string source = tokens[3];
    auto source_iterator =
        std::find(storage_source.begin(), storage_source.end(), source);
    std::string target = tokens[4];
    if (source_iterator != storage_source.end() && target == "/storage") {
      return 0;
    }
  }
  return -EPERM;
}

int passthrough_create(const char* path,
                       mode_t mode,
                       struct fuse_file_info* fi) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  // Ignore specified |mode| and always use a fixed mode since we do not allow
  // chmod anyway. Note that we explicitly set the umask to 0022 in main().
  int fd = open(path, fi->flags, 0644);
  if (fd < 0) {
    return -errno;
  }
  fi->fh = fd;
  return 0;
}

int passthrough_fgetattr(const char*,
                         struct stat* buf,
                         struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  // File owner is overridden by uid/gid options passed to fuse.
  return WRAP_FS_CALL(fstat(fd, buf));
}

int passthrough_fsync(const char*, int datasync, struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  return datasync ? WRAP_FS_CALL(fdatasync(fd)) : WRAP_FS_CALL(fsync(fd));
}

int passthrough_fsyncdir(const char*, int datasync, struct fuse_file_info* fi) {
  DIR* dirp = reinterpret_cast<DIR*>(fi->fh);
  int fd = dirfd(dirp);
  return datasync ? WRAP_FS_CALL(fdatasync(fd)) : WRAP_FS_CALL(fsync(fd));
}

int passthrough_ftruncate(const char*, off_t size, struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  return WRAP_FS_CALL(ftruncate(fd, size));
}

int passthrough_getattr(const char* path, struct stat* buf) {
  // File owner is overridden by uid/gid options passed to fuse.
  // Unfortunately, we dont have check_allowed() here because getattr is called
  // by kernel VFS during fstat (which receives fd). We couldn't prohibit such
  // fd calls to happen, so we need to relax this.
  return WRAP_FS_CALL(lstat(path, buf));
}

int passthrough_mkdir(const char* path, mode_t mode) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }

  // When |force_group_permission| is true, forcefully grant full group access
  // permission so that Android's MediaProvider can access the new directory.
  if (static_cast<FusePrivateData*>(fuse_get_context()->private_data)
          ->force_group_permission) {
    mode |= S_IRWXG;
  }

  return WRAP_FS_CALL(mkdir(path, mode));
}

int passthrough_open(const char* path, struct fuse_file_info* fi) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  int fd = open(path, fi->flags);
  if (fd < 0) {
    return -errno;
  }
  fi->fh = fd;
  return 0;
}

int passthrough_opendir(const char* path, struct fuse_file_info* fi) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  DIR* dirp = opendir(path);
  if (!dirp) {
    return -errno;
  }
  fi->fh = reinterpret_cast<uint64_t>(dirp);
  return 0;
}

int passthrough_read(
    const char*, char* buf, size_t size, off_t off, struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  int res = pread(fd, buf, size, off);
  if (res < 0) {
    return -errno;
  }
  return res;
}

int passthrough_read_buf(const char*,
                         struct fuse_bufvec** srcp,
                         size_t size,
                         off_t off,
                         struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  struct fuse_bufvec* src =
      static_cast<struct fuse_bufvec*>(malloc(sizeof(struct fuse_bufvec)));
  *src = FUSE_BUFVEC_INIT(size);
  src->buf[0].flags =
      static_cast<fuse_buf_flags>(FUSE_BUF_IS_FD | FUSE_BUF_FD_SEEK);
  src->buf[0].fd = fd;
  src->buf[0].pos = off;
  *srcp = src;
  return 0;
}

int passthrough_readdir(const char*,
                        void* buf,
                        fuse_fill_dir_t filler,
                        off_t off,
                        struct fuse_file_info* fi) {
  // TODO(nya): This implementation returns all files at once and thus
  // inefficient. Make use of offset and be better to memory.
  DIR* dirp = reinterpret_cast<DIR*>(fi->fh);
  errno = 0;
  for (;;) {
    struct dirent* entry = readdir(dirp);
    if (entry == nullptr) {
      break;
    }
    // Only IF part of st_mode matters. See fill_dir() in fuse.c.
    struct stat stbuf = {};
    stbuf.st_mode = DTTOIF(entry->d_type);
    filler(buf, entry->d_name, &stbuf, 0);
  }
  return -errno;
}

int passthrough_release(const char*, struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  return WRAP_FS_CALL(close(fd));
}

int passthrough_releasedir(const char*, struct fuse_file_info* fi) {
  DIR* dirp = reinterpret_cast<DIR*>(fi->fh);
  return WRAP_FS_CALL(closedir(dirp));
}

int passthrough_rename(const char* oldpath, const char* newpath) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(rename(oldpath, newpath));
}

int passthrough_rmdir(const char* path) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(rmdir(path));
}

int passthrough_statfs(const char* path, struct statvfs* buf) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(statvfs(path, buf));
}

int passthrough_truncate(const char* path, off_t size) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(truncate(path, size));
}

int passthrough_unlink(const char* path) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(unlink(path));
}

int passthrough_utimens(const char* path, const struct timespec tv[2]) {
  int check_allowed_result = check_allowed();
  if (check_allowed_result < 0) {
    return check_allowed_result;
  }
  return WRAP_FS_CALL(utimensat(AT_FDCWD, path, tv, 0));
}

int passthrough_write(const char*,
                      const char* buf,
                      size_t size,
                      off_t off,
                      struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  int res = pwrite(fd, buf, size, off);
  if (res < 0) {
    return -errno;
  }
  return res;
}

int passthrough_write_buf(const char*,
                          struct fuse_bufvec* src,
                          off_t off,
                          struct fuse_file_info* fi) {
  int fd = static_cast<int>(fi->fh);
  struct fuse_bufvec dst = FUSE_BUFVEC_INIT(fuse_buf_size(src));
  dst.buf[0].flags =
      static_cast<fuse_buf_flags>(FUSE_BUF_IS_FD | FUSE_BUF_FD_SEEK);
  dst.buf[0].fd = fd;
  dst.buf[0].pos = off;
  return fuse_buf_copy(&dst, src, static_cast<fuse_buf_copy_flags>(0));
}

void setup_passthrough_ops(struct fuse_operations* passthrough_ops) {
  memset(passthrough_ops, 0, sizeof(*passthrough_ops));
#define FILL_OP(name) passthrough_ops->name = passthrough_##name
  FILL_OP(create);
  FILL_OP(fgetattr);
  FILL_OP(fsync);
  FILL_OP(fsyncdir);
  FILL_OP(ftruncate);
  FILL_OP(getattr);
  FILL_OP(mkdir);
  FILL_OP(open);
  FILL_OP(opendir);
  FILL_OP(read);
  FILL_OP(read_buf);
  FILL_OP(readdir);
  FILL_OP(release);
  FILL_OP(releasedir);
  FILL_OP(rename);
  FILL_OP(rmdir);
  FILL_OP(statfs);
  FILL_OP(truncate);
  FILL_OP(unlink);
  FILL_OP(utimens);
  FILL_OP(write);
  FILL_OP(write_buf);
#undef FILL_OP
  passthrough_ops->flag_nullpath_ok = 1;
  passthrough_ops->flag_nopath = 1;
}

}  // namespace

int main(int argc, char** argv) {
  DEFINE_string(source, "", "Source path of FUSE mount (required)");
  DEFINE_string(dest, "", "Target path of FUSE mount (required)");
  DEFINE_string(fuse_umask, "",
                "Umask to set filesystem permissions in FUSE (required)");
  DEFINE_int32(fuse_uid, -1, "UID set as file owner in FUSE (required)");
  DEFINE_int32(fuse_gid, -1, "GID set as file group in FUSE (required)");
  DEFINE_string(
      android_app_access_type, "",
      "What type of permission checks should be done for Android apps."
      " Must be either full, read, or write (required)");

  // TODO(b/123669632): Remove the argument |force_group_permission| and related
  // logic once we start to run the daemon as MediaProvider UID and GID from
  // mount-passthrough-jailed-play.
  DEFINE_bool(force_group_permission, false,
              "Forcefully grant full group access permission for newly created"
              " directories (optional)");

  // Use "arc-" prefix so that the log is recorded in /var/log/arc.log.
  brillo::OpenLog("arc-mount-passthrough", true /*log_pid*/);
  brillo::InitLog(brillo::kLogToSyslog | brillo::kLogHeader |
                  brillo::kLogToStderrIfTty);

  brillo::FlagHelper::Init(argc, argv, "mount-passthrough");

  if (FLAGS_source.empty()) {
    LOG(ERROR) << "--source must be specified.";
    return 1;
  }
  if (FLAGS_dest.empty()) {
    LOG(ERROR) << "--dest must be specified.";
    return 1;
  }
  if (FLAGS_fuse_umask.empty()) {
    LOG(ERROR) << "--fuse_umask must be specified.";
    return 1;
  }
  if (FLAGS_fuse_uid < 0) {
    LOG(ERROR) << "--fuse_uid must be specified as a non-negative integer.";
    return 1;
  }
  if (FLAGS_fuse_gid < 0) {
    LOG(ERROR) << "--fuse_gid must be specified as a non-negative integer.";
    return 1;
  }
  if (FLAGS_android_app_access_type.empty()) {
    LOG(ERROR) << "--android_app_access_type must be specified.";
    return 1;
  }
  if (FLAGS_android_app_access_type != "full" &&
      FLAGS_android_app_access_type != "read" &&
      FLAGS_android_app_access_type != "write") {
    LOG(ERROR) << "Invalid android_app_access_type: "
               << FLAGS_android_app_access_type
               << ". It must be either full, read, or write.";
    return 1;
  }

  const uid_t daemon_uid = getuid();
  if (daemon_uid != CHRONOS_UID && daemon_uid != AID_MEDIA_RW_UID) {
    LOG(ERROR) << "This daemon must run as chronos or Android's media_rw user.";
    return 1;
  }

  const gid_t daemon_gid = getgid();
  if (daemon_gid != CHRONOS_GID && daemon_gid != AID_MEDIA_RW_GID) {
    LOG(ERROR) << "This daemon must run as chronos or Android's media_rw"
               << " group.";
    return 1;
  }

  struct fuse_operations passthrough_ops;
  setup_passthrough_ops(&passthrough_ops);

  const std::string fuse_subdir_opt("subdir=" + FLAGS_source);
  const std::string fuse_uid_opt(
      "uid=" + std::to_string(FLAGS_fuse_uid + USER_NS_SHIFT));
  const std::string fuse_gid_opt(
      "gid=" + std::to_string(FLAGS_fuse_gid + USER_NS_SHIFT));
  const std::string fuse_umask_opt("umask=" + FLAGS_fuse_umask);
  LOG(INFO) << "subdir_opt(" << fuse_subdir_opt << ") "
            << "uid_opt(" << fuse_uid_opt << ") "
            << "gid_opt(" << fuse_gid_opt << ") "
            << "umask_opt(" << fuse_umask_opt << ")";

  const char* fuse_argv[] = {
      argv[0],
      FLAGS_dest.c_str(),
      "-f",
      "-o",
      "allow_other",
      "-o",
      "default_permissions",
      // Never cache attr/dentry since our backend storage is not exclusive to
      // this process.
      "-o",
      "attr_timeout=0",
      "-o",
      "entry_timeout=0",
      "-o",
      "negative_timeout=0",
      "-o",
      "ac_attr_timeout=0",
      "-o",
      "fsname=passthrough",
      "-o",
      fuse_uid_opt.c_str(),
      "-o",
      fuse_gid_opt.c_str(),
      "-o",
      "modules=subdir",
      "-o",
      fuse_subdir_opt.c_str(),
      "-o",
      "direct_io",
      "-o",
      fuse_umask_opt.c_str(),
      "-o",
      "noexec",
  };
  int fuse_argc = sizeof(fuse_argv) / sizeof(fuse_argv[0]);

  const mode_t daemon_umask = FLAGS_force_group_permission ? 0002 : 0022;
  umask(daemon_umask);

  FusePrivateData private_data;
  private_data.android_app_access_type = FLAGS_android_app_access_type;
  private_data.force_group_permission = FLAGS_force_group_permission;

  // The code below does the same thing as fuse_main() except that it ignores
  // signals during shutdown to perform clean shutdown. b/183343552
  // TODO(hashimoto): Stop using deprecated libfuse functions b/185322557.
  char* mountpoint = nullptr;
  int multithreaded = 0;
  struct fuse* fuse = fuse_setup(fuse_argc, const_cast<char**>(fuse_argv),
                                 &passthrough_ops, sizeof(passthrough_ops),
                                 &mountpoint, &multithreaded, &private_data);
  if (fuse == nullptr)
    return 1;

  int res = 0;
  if (multithreaded)
    res = fuse_loop_mt(fuse);
  else
    res = fuse_loop(fuse);

  // The code below does the same thing fuse_teardown() except that it ignores
  // signals instead of calling fuse_remove_signal_handlers().

  // Ignore signals after this point. We're already shutting down.
  struct sigaction sa = {};
  sigemptyset(&sa.sa_mask);
  sa.sa_flags = 0;
  sa.sa_handler = SIG_IGN;
  sigaction(SIGHUP, &sa, nullptr);
  sigaction(SIGINT, &sa, nullptr);
  sigaction(SIGTERM, &sa, nullptr);
  sigaction(SIGPIPE, &sa, nullptr);

  struct fuse_session* se = fuse_get_session(fuse);
  struct fuse_chan* ch = fuse_session_next_chan(se, nullptr);
  fuse_unmount(mountpoint, ch);
  fuse_destroy(fuse);
  free(mountpoint);

  return res == -1 ? 1 : 0;
}