content/browser/zygote_host/zygote_host_impl_linux.cc

   1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "content/browser/zygote_host/zygote_host_impl_linux.h"
   6
   7 #include <string.h>
   8 #include <sys/socket.h>
   9 #include <sys/stat.h>
  10 #include <sys/types.h>
  11 #include <unistd.h>
  12
  13 #include "base/base_switches.h"
  14 #include "base/command_line.h"
  15 #include "base/environment.h"
  16 #include "base/file_util.h"
  17 #include "base/files/file_enumerator.h"
  18 #include "base/files/scoped_file.h"
  19 #include "base/linux_util.h"
  20 #include "base/logging.h"
  21 #include "base/memory/linked_ptr.h"
  22 #include "base/memory/scoped_ptr.h"
  23 #include "base/memory/scoped_vector.h"
  24 #include "base/metrics/histogram.h"
  25 #include "base/path_service.h"
  26 #include "base/posix/eintr_wrapper.h"
  27 #include "base/posix/unix_domain_socket_linux.h"
  28 #include "base/process/launch.h"
  29 #include "base/process/memory.h"
  30 #include "base/process/process_handle.h"
  31 #include "base/strings/string_number_conversions.h"
  32 #include "base/strings/string_util.h"
  33 #include "base/strings/utf_string_conversions.h"
  34 #include "base/time/time.h"
  35 #include "content/browser/renderer_host/render_sandbox_host_linux.h"
  36 #include "content/common/child_process_sandbox_support_impl_linux.h"
  37 #include "content/common/zygote_commands_linux.h"
  38 #include "content/public/browser/content_browser_client.h"
  39 #include "content/public/common/content_switches.h"
  40 #include "content/public/common/result_codes.h"
  41 #include "sandbox/linux/suid/client/setuid_sandbox_client.h"
  42 #include "sandbox/linux/suid/common/sandbox.h"
  43 #include "ui/base/ui_base_switches.h"
  44 #include "ui/gfx/switches.h"
  45
  46 #if defined(USE_TCMALLOC)
  47 #include "third_party/tcmalloc/chromium/src/gperftools/heap-profiler.h"
  48 #endif
  49
  50 namespace content {
  51
  52 // Receive a fixed message on fd and return the sender's PID.
  53 // Returns true if the message received matches the expected message.
  54 static bool ReceiveFixedMessage(int fd,
  55                                 const char* expect_msg,
  56                                 size_t expect_len,
  57                                 base::ProcessId* sender_pid) {
  58   char buf[expect_len + 1];
  59   ScopedVector<base::ScopedFD> fds_vec;
  60
  61   const ssize_t len = UnixDomainSocket::RecvMsgWithPid(
  62       fd, buf, sizeof(buf), &fds_vec, sender_pid);
  63   if (static_cast<size_t>(len) != expect_len)
  64     return false;
  65   if (memcmp(buf, expect_msg, expect_len) != 0)
  66     return false;
  67   if (!fds_vec.empty())
  68     return false;
  69   return true;
  70 }
  71
  72 // static
  73 ZygoteHost* ZygoteHost::GetInstance() {
  74   return ZygoteHostImpl::GetInstance();
  75 }
  76
  77 ZygoteHostImpl::ZygoteHostImpl()
  78     : control_fd_(-1),
  79       control_lock_(),
  80       pid_(-1),
  81       init_(false),
  82       using_suid_sandbox_(false),
  83       sandbox_binary_(),
  84       have_read_sandbox_status_word_(false),
  85       sandbox_status_(0),
  86       child_tracking_lock_(),
  87       list_of_running_zygote_children_(),
  88       should_teardown_after_last_child_exits_(false) {}
  89
  90 ZygoteHostImpl::~ZygoteHostImpl() { TearDown(); }
  91
  92 // static
  93 ZygoteHostImpl* ZygoteHostImpl::GetInstance() {
  94   return Singleton<ZygoteHostImpl>::get();
  95 }
  96
  97 void ZygoteHostImpl::Init(const std::string& sandbox_cmd) {
  98   DCHECK(!init_);
  99   init_ = true;
 100
 101   base::FilePath chrome_path;
 102   CHECK(PathService::Get(base::FILE_EXE, &chrome_path));
 103   base::CommandLine cmd_line(chrome_path);
 104
 105   cmd_line.AppendSwitchASCII(switches::kProcessType, switches::kZygoteProcess);
 106
 107   int fds[2];
 108   CHECK(socketpair(AF_UNIX, SOCK_SEQPACKET, 0, fds) == 0);
 109   CHECK(UnixDomainSocket::EnableReceiveProcessId(fds[0]));
 110   base::FileHandleMappingVector fds_to_map;
 111   fds_to_map.push_back(std::make_pair(fds[1], kZygoteSocketPairFd));
 112
 113   base::LaunchOptions options;
 114   const base::CommandLine& browser_command_line =
 115       *base::CommandLine::ForCurrentProcess();
 116   if (browser_command_line.HasSwitch(switches::kZygoteCmdPrefix)) {
 117     cmd_line.PrependWrapper(
 118         browser_command_line.GetSwitchValueNative(switches::kZygoteCmdPrefix));
 119   }
 120   // Append any switches from the browser process that need to be forwarded on
 121   // to the zygote/renderers.
 122   // Should this list be obtained from browser_render_process_host.cc?
 123   static const char* kForwardSwitches[] = {
 124     switches::kAllowSandboxDebugging,
 125     switches::kLoggingLevel,
 126     switches::kEnableLogging,  // Support, e.g., --enable-logging=stderr.
 127     switches::kV,
 128     switches::kVModule,
 129     switches::kRegisterPepperPlugins,
 130     switches::kDisableSeccompFilterSandbox,
 131
 132     // Zygote process needs to know what resources to have loaded when it
 133     // becomes a renderer process.
 134     switches::kForceDeviceScaleFactor,
 135
 136     switches::kNoSandbox,
 137   };
 138   cmd_line.CopySwitchesFrom(browser_command_line, kForwardSwitches,
 139                             arraysize(kForwardSwitches));
 140
 141   GetContentClient()->browser()->AppendExtraCommandLineSwitches(&cmd_line, -1);
 142
 143   sandbox_binary_ = sandbox_cmd.c_str();
 144
 145   // A non empty sandbox_cmd means we want a SUID sandbox.
 146   using_suid_sandbox_ = !sandbox_cmd.empty();
 147
 148   // Start up the sandbox host process and get the file descriptor for the
 149   // renderers to talk to it.
 150   const int sfd = RenderSandboxHostLinux::GetInstance()->GetRendererSocket();
 151   fds_to_map.push_back(std::make_pair(sfd, GetSandboxFD()));
 152
 153   base::ScopedFD dummy_fd;
 154   if (using_suid_sandbox_) {
 155     scoped_ptr<sandbox::SetuidSandboxClient>
 156         sandbox_client(sandbox::SetuidSandboxClient::Create());
 157     sandbox_client->PrependWrapper(&cmd_line);
 158     sandbox_client->SetupLaunchOptions(&options, &fds_to_map, &dummy_fd);
 159     sandbox_client->SetupLaunchEnvironment();
 160   }
 161
 162   base::ProcessHandle process = -1;
 163   options.fds_to_remap = &fds_to_map;
 164   base::LaunchProcess(cmd_line.argv(), options, &process);
 165   CHECK(process != -1) << "Failed to launch zygote process";
 166   dummy_fd.reset();
 167
 168   if (using_suid_sandbox_) {
 169     // The SUID sandbox will execute the zygote in a new PID namespace, and
 170     // the main zygote process will then fork from there.  Watch now our
 171     // elaborate dance to find and validate the zygote's PID.
 172
 173     // First we receive a message from the zygote boot process.
 174     base::ProcessId boot_pid;
 175     CHECK(ReceiveFixedMessage(
 176         fds[0], kZygoteBootMessage, sizeof(kZygoteBootMessage), &boot_pid));
 177
 178     // Within the PID namespace, the zygote boot process thinks it's PID 1,
 179     // but its real PID can never be 1. This gives us a reliable test that
 180     // the kernel is translating the sender's PID to our namespace.
 181     CHECK_GT(boot_pid, 1)
 182         << "Received invalid process ID for zygote; kernel might be too old? "
 183            "See crbug.com/357670 or try using --"
 184         << switches::kDisableSetuidSandbox << " to workaround.";
 185
 186     // Now receive the message that the zygote's ready to go, along with the
 187     // main zygote process's ID.
 188     CHECK(ReceiveFixedMessage(
 189         fds[0], kZygoteHelloMessage, sizeof(kZygoteHelloMessage), &pid_));
 190     CHECK_GT(pid_, 1);
 191
 192     if (process != pid_) {
 193       // Reap the sandbox.
 194       base::EnsureProcessGetsReaped(process);
 195     }
 196   } else {
 197     // Not using the SUID sandbox.
 198     pid_ = process;
 199   }
 200
 201   close(fds[1]);
 202   control_fd_ = fds[0];
 203
 204   Pickle pickle;
 205   pickle.WriteInt(kZygoteCommandGetSandboxStatus);
 206   if (!SendMessage(pickle, NULL))
 207     LOG(FATAL) << "Cannot communicate with zygote";
 208   // We don't wait for the reply. We'll read it in ReadReply.
 209 }
 210
 211 void ZygoteHostImpl::TearDownAfterLastChild() {
 212   bool do_teardown = false;
 213   {
 214     base::AutoLock lock(child_tracking_lock_);
 215     should_teardown_after_last_child_exits_ = true;
 216     do_teardown = list_of_running_zygote_children_.empty();
 217   }
 218   if (do_teardown) {
 219     TearDown();
 220   }
 221 }
 222
 223 // Note: this is also called from the destructor.
 224 void ZygoteHostImpl::TearDown() {
 225   base::AutoLock lock(control_lock_);
 226   if (control_fd_ > -1) {
 227     // Closing the IPC channel will act as a notification to exit
 228     // to the Zygote.
 229     if (IGNORE_EINTR(close(control_fd_))) {
 230       PLOG(ERROR) << "Could not close Zygote control channel.";
 231       NOTREACHED();
 232     }
 233     control_fd_ = -1;
 234   }
 235 }
 236
 237 void ZygoteHostImpl::ZygoteChildBorn(pid_t process) {
 238   base::AutoLock lock(child_tracking_lock_);
 239   bool new_element_inserted =
 240       list_of_running_zygote_children_.insert(process).second;
 241   DCHECK(new_element_inserted);
 242 }
 243
 244 void ZygoteHostImpl::ZygoteChildDied(pid_t process) {
 245   bool do_teardown = false;
 246   {
 247     base::AutoLock lock(child_tracking_lock_);
 248     size_t num_erased = list_of_running_zygote_children_.erase(process);
 249     DCHECK_EQ(1U, num_erased);
 250     do_teardown = should_teardown_after_last_child_exits_ &&
 251                   list_of_running_zygote_children_.empty();
 252   }
 253   if (do_teardown) {
 254     TearDown();
 255   }
 256 }
 257
 258 bool ZygoteHostImpl::SendMessage(const Pickle& data,
 259                                  const std::vector<int>* fds) {
 260   DCHECK_NE(-1, control_fd_);
 261   CHECK(data.size() <= kZygoteMaxMessageLength)
 262       << "Trying to send too-large message to zygote (sending " << data.size()
 263       << " bytes, max is " << kZygoteMaxMessageLength << ")";
 264   CHECK(!fds || fds->size() <= UnixDomainSocket::kMaxFileDescriptors)
 265       << "Trying to send message with too many file descriptors to zygote "
 266       << "(sending " << fds->size() << ", max is "
 267       << UnixDomainSocket::kMaxFileDescriptors << ")";
 268
 269   return UnixDomainSocket::SendMsg(control_fd_,
 270                                    data.data(), data.size(),
 271                                    fds ? *fds : std::vector<int>());
 272 }
 273
 274 ssize_t ZygoteHostImpl::ReadReply(void* buf, size_t buf_len) {
 275   DCHECK_NE(-1, control_fd_);
 276   // At startup we send a kZygoteCommandGetSandboxStatus request to the zygote,
 277   // but don't wait for the reply. Thus, the first time that we read from the
 278   // zygote, we get the reply to that request.
 279   if (!have_read_sandbox_status_word_) {
 280     if (HANDLE_EINTR(read(control_fd_, &sandbox_status_,
 281                           sizeof(sandbox_status_))) !=
 282         sizeof(sandbox_status_)) {
 283       return -1;
 284     }
 285     have_read_sandbox_status_word_ = true;
 286   }
 287
 288   return HANDLE_EINTR(read(control_fd_, buf, buf_len));
 289 }
 290
 291 pid_t ZygoteHostImpl::ForkRequest(
 292     const std::vector<std::string>& argv,
 293     const std::vector<FileDescriptorInfo>& mapping,
 294     const std::string& process_type) {
 295   DCHECK(init_);
 296   Pickle pickle;
 297
 298   int raw_socks[2];
 299   PCHECK(0 == socketpair(AF_UNIX, SOCK_SEQPACKET, 0, raw_socks));
 300   base::ScopedFD my_sock(raw_socks[0]);
 301   base::ScopedFD peer_sock(raw_socks[1]);
 302   CHECK(UnixDomainSocket::EnableReceiveProcessId(my_sock.get()));
 303
 304   pickle.WriteInt(kZygoteCommandFork);
 305   pickle.WriteString(process_type);
 306   pickle.WriteInt(argv.size());
 307   for (std::vector<std::string>::const_iterator
 308        i = argv.begin(); i != argv.end(); ++i)
 309     pickle.WriteString(*i);
 310
 311   // Fork requests contain one file descriptor for the PID oracle, and one
 312   // more for each file descriptor mapping for the child process.
 313   const size_t num_fds_to_send = 1 + mapping.size();
 314   pickle.WriteInt(num_fds_to_send);
 315
 316   std::vector<int> fds;
 317   ScopedVector<base::ScopedFD> autoclose_fds;
 318
 319   // First FD to send is peer_sock.
 320   fds.push_back(peer_sock.get());
 321   autoclose_fds.push_back(new base::ScopedFD(peer_sock.Pass()));
 322
 323   // The rest come from mapping.
 324   for (std::vector<FileDescriptorInfo>::const_iterator
 325        i = mapping.begin(); i != mapping.end(); ++i) {
 326     pickle.WriteUInt32(i->id);
 327     fds.push_back(i->fd.fd);
 328     if (i->fd.auto_close) {
 329       // Auto-close means we need to close the FDs after they have been passed
 330       // to the other process.
 331       autoclose_fds.push_back(new base::ScopedFD(i->fd.fd));
 332     }
 333   }
 334
 335   // Sanity check that we've populated |fds| correctly.
 336   DCHECK_EQ(num_fds_to_send, fds.size());
 337
 338   pid_t pid;
 339   {
 340     base::AutoLock lock(control_lock_);
 341     if (!SendMessage(pickle, &fds))
 342       return base::kNullProcessHandle;
 343     autoclose_fds.clear();
 344
 345     {
 346       char buf[sizeof(kZygoteChildPingMessage) + 1];
 347       ScopedVector<base::ScopedFD> recv_fds;
 348       base::ProcessId real_pid;
 349
 350       ssize_t n = UnixDomainSocket::RecvMsgWithPid(
 351           my_sock.get(), buf, sizeof(buf), &recv_fds, &real_pid);
 352       if (n != sizeof(kZygoteChildPingMessage) ||
 353           0 != memcmp(buf,
 354                       kZygoteChildPingMessage,
 355                       sizeof(kZygoteChildPingMessage))) {
 356         // Zygote children should still be trustworthy when they're supposed to
 357         // ping us, so something's broken if we don't receive a valid ping.
 358         LOG(ERROR) << "Did not receive ping from zygote child";
 359         NOTREACHED();
 360         real_pid = -1;
 361       }
 362       my_sock.reset();
 363
 364       // Always send PID back to zygote.
 365       Pickle pid_pickle;
 366       pid_pickle.WriteInt(kZygoteCommandForkRealPID);
 367       pid_pickle.WriteInt(real_pid);
 368       if (!SendMessage(pid_pickle, NULL))
 369         return base::kNullProcessHandle;
 370     }
 371
 372     // Read the reply, which pickles the PID and an optional UMA enumeration.
 373     static const unsigned kMaxReplyLength = 2048;
 374     char buf[kMaxReplyLength];
 375     const ssize_t len = ReadReply(buf, sizeof(buf));
 376
 377     Pickle reply_pickle(buf, len);
 378     PickleIterator iter(reply_pickle);
 379     if (len <= 0 || !reply_pickle.ReadInt(&iter, &pid))
 380       return base::kNullProcessHandle;
 381
 382     // If there is a nonempty UMA name string, then there is a UMA
 383     // enumeration to record.
 384     std::string uma_name;
 385     int uma_sample;
 386     int uma_boundary_value;
 387     if (reply_pickle.ReadString(&iter, &uma_name) &&
 388         !uma_name.empty() &&
 389         reply_pickle.ReadInt(&iter, &uma_sample) &&
 390         reply_pickle.ReadInt(&iter, &uma_boundary_value)) {
 391       // We cannot use the UMA_HISTOGRAM_ENUMERATION macro here,
 392       // because that's only for when the name is the same every time.
 393       // Here we're using whatever name we got from the other side.
 394       // But since it's likely that the same one will be used repeatedly
 395       // (even though it's not guaranteed), we cache it here.
 396       static base::HistogramBase* uma_histogram;
 397       if (!uma_histogram || uma_histogram->histogram_name() != uma_name) {
 398         uma_histogram = base::LinearHistogram::FactoryGet(
 399             uma_name, 1,
 400             uma_boundary_value,
 401             uma_boundary_value + 1,
 402             base::HistogramBase::kUmaTargetedHistogramFlag);
 403       }
 404       uma_histogram->Add(uma_sample);
 405     }
 406
 407     if (pid <= 0)
 408       return base::kNullProcessHandle;
 409   }
 410
 411 #if !defined(OS_OPENBSD)
 412   // This is just a starting score for a renderer or extension (the
 413   // only types of processes that will be started this way).  It will
 414   // get adjusted as time goes on.  (This is the same value as
 415   // chrome::kLowestRendererOomScore in chrome/chrome_constants.h, but
 416   // that's not something we can include here.)
 417   const int kLowestRendererOomScore = 300;
 418   AdjustRendererOOMScore(pid, kLowestRendererOomScore);
 419 #endif
 420
 421   ZygoteChildBorn(pid);
 422   return pid;
 423 }
 424
 425 #if !defined(OS_OPENBSD)
 426 void ZygoteHostImpl::AdjustRendererOOMScore(base::ProcessHandle pid,
 427                                             int score) {
 428   // 1) You can't change the oom_score_adj of a non-dumpable process
 429   //    (EPERM) unless you're root. Because of this, we can't set the
 430   //    oom_adj from the browser process.
 431   //
 432   // 2) We can't set the oom_score_adj before entering the sandbox
 433   //    because the zygote is in the sandbox and the zygote is as
 434   //    critical as the browser process. Its oom_adj value shouldn't
 435   //    be changed.
 436   //
 437   // 3) A non-dumpable process can't even change its own oom_score_adj
 438   //    because it's root owned 0644. The sandboxed processes don't
 439   //    even have /proc, but one could imagine passing in a descriptor
 440   //    from outside.
 441   //
 442   // So, in the normal case, we use the SUID binary to change it for us.
 443   // However, Fedora (and other SELinux systems) don't like us touching other
 444   // process's oom_score_adj (or oom_adj) values
 445   // (https://bugzilla.redhat.com/show_bug.cgi?id=581256).
 446   //
 447   // The offical way to get the SELinux mode is selinux_getenforcemode, but I
 448   // don't want to add another library to the build as it's sure to cause
 449   // problems with other, non-SELinux distros.
 450   //
 451   // So we just check for files in /selinux. This isn't foolproof, but it's not
 452   // bad and it's easy.
 453
 454   static bool selinux;
 455   static bool selinux_valid = false;
 456
 457   if (!selinux_valid) {
 458     const base::FilePath kSelinuxPath("/selinux");
 459     base::FileEnumerator en(kSelinuxPath, false, base::FileEnumerator::FILES);
 460     bool has_selinux_files = !en.Next().empty();
 461
 462     selinux = access(kSelinuxPath.value().c_str(), X_OK) == 0 &&
 463               has_selinux_files;
 464     selinux_valid = true;
 465   }
 466
 467   if (using_suid_sandbox_ && !selinux) {
 468 #if defined(USE_TCMALLOC)
 469     // If heap profiling is running, these processes are not exiting, at least
 470     // on ChromeOS. The easiest thing to do is not launch them when profiling.
 471     // TODO(stevenjb): Investigate further and fix.
 472     if (IsHeapProfilerRunning())
 473       return;
 474 #endif
 475     std::vector<std::string> adj_oom_score_cmdline;
 476     adj_oom_score_cmdline.push_back(sandbox_binary_);
 477     adj_oom_score_cmdline.push_back(sandbox::kAdjustOOMScoreSwitch);
 478     adj_oom_score_cmdline.push_back(base::Int64ToString(pid));
 479     adj_oom_score_cmdline.push_back(base::IntToString(score));
 480
 481     base::ProcessHandle sandbox_helper_process;
 482     base::LaunchOptions options;
 483
 484     // sandbox_helper_process is a setuid binary.
 485     options.allow_new_privs = true;
 486
 487     if (base::LaunchProcess(adj_oom_score_cmdline, options,
 488                             &sandbox_helper_process)) {
 489       base::EnsureProcessGetsReaped(sandbox_helper_process);
 490     }
 491   } else if (!using_suid_sandbox_) {
 492     if (!base::AdjustOOMScore(pid, score))
 493       PLOG(ERROR) << "Failed to adjust OOM score of renderer with pid " << pid;
 494   }
 495 }
 496 #endif
 497
 498 void ZygoteHostImpl::EnsureProcessTerminated(pid_t process) {
 499   DCHECK(init_);
 500   Pickle pickle;
 501
 502   pickle.WriteInt(kZygoteCommandReap);
 503   pickle.WriteInt(process);
 504   if (!SendMessage(pickle, NULL))
 505     LOG(ERROR) << "Failed to send Reap message to zygote";
 506   ZygoteChildDied(process);
 507 }
 508
 509 base::TerminationStatus ZygoteHostImpl::GetTerminationStatus(
 510     base::ProcessHandle handle,
 511     bool known_dead,
 512     int* exit_code) {
 513   DCHECK(init_);
 514   Pickle pickle;
 515   pickle.WriteInt(kZygoteCommandGetTerminationStatus);
 516   pickle.WriteBool(known_dead);
 517   pickle.WriteInt(handle);
 518
 519   static const unsigned kMaxMessageLength = 128;
 520   char buf[kMaxMessageLength];
 521   ssize_t len;
 522   {
 523     base::AutoLock lock(control_lock_);
 524     if (!SendMessage(pickle, NULL))
 525       LOG(ERROR) << "Failed to send GetTerminationStatus message to zygote";
 526     len = ReadReply(buf, sizeof(buf));
 527   }
 528
 529   // Set this now to handle the error cases.
 530   if (exit_code)
 531     *exit_code = RESULT_CODE_NORMAL_EXIT;
 532   int status = base::TERMINATION_STATUS_NORMAL_TERMINATION;
 533
 534   if (len == -1) {
 535     LOG(WARNING) << "Error reading message from zygote: " << errno;
 536   } else if (len == 0) {
 537     LOG(WARNING) << "Socket closed prematurely.";
 538   } else {
 539     Pickle read_pickle(buf, len);
 540     int tmp_status, tmp_exit_code;
 541     PickleIterator iter(read_pickle);
 542     if (!read_pickle.ReadInt(&iter, &tmp_status) ||
 543         !read_pickle.ReadInt(&iter, &tmp_exit_code)) {
 544       LOG(WARNING)
 545           << "Error parsing GetTerminationStatus response from zygote.";
 546     } else {
 547       if (exit_code)
 548         *exit_code = tmp_exit_code;
 549       status = tmp_status;
 550     }
 551   }
 552
 553   if (status != base::TERMINATION_STATUS_STILL_RUNNING) {
 554     ZygoteChildDied(handle);
 555   }
 556   return static_cast<base::TerminationStatus>(status);
 557 }
 558
 559 pid_t ZygoteHostImpl::GetPid() const {
 560   return pid_;
 561 }
 562
 563 int ZygoteHostImpl::GetSandboxStatus() const {
 564   if (have_read_sandbox_status_word_)
 565     return sandbox_status_;
 566   return 0;
 567 }
 568
 569 }  // namespace content