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