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   options.fds_to_remap = &fds_to_map;
 162   base::Process process = base::LaunchProcess(cmd_line.argv(), options);
 163   CHECK(process.IsValid()) << "Failed to launch zygote process";
 164   dummy_fd.reset();
 165
 166   if (using_suid_sandbox_) {
 167     // The SUID sandbox will execute the zygote in a new PID namespace, and
 168     // the main zygote process will then fork from there.  Watch now our
 169     // elaborate dance to find and validate the zygote's PID.
 170
 171     // First we receive a message from the zygote boot process.
 172     base::ProcessId boot_pid;
 173     CHECK(ReceiveFixedMessage(
 174         fds[0], kZygoteBootMessage, sizeof(kZygoteBootMessage), &boot_pid));
 175
 176     // Within the PID namespace, the zygote boot process thinks it's PID 1,
 177     // but its real PID can never be 1. This gives us a reliable test that
 178     // the kernel is translating the sender's PID to our namespace.
 179     CHECK_GT(boot_pid, 1)
 180         << "Received invalid process ID for zygote; kernel might be too old? "
 181            "See crbug.com/357670 or try using --"
 182         << switches::kDisableSetuidSandbox << " to workaround.";
 183
 184     // Now receive the message that the zygote's ready to go, along with the
 185     // main zygote process's ID.
 186     CHECK(ReceiveFixedMessage(
 187         fds[0], kZygoteHelloMessage, sizeof(kZygoteHelloMessage), &pid_));
 188     CHECK_GT(pid_, 1);
 189
 190     if (process.pid() != pid_) {
 191       // Reap the sandbox.
 192       base::EnsureProcessGetsReaped(process.pid());
 193     }
 194   } else {
 195     // Not using the SUID sandbox.
 196     // Note that ~base::Process() will reset the internal value, but there's no
 197     // real "handle" on POSIX so that is safe.
 198     pid_ = process.pid();
 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(const std::vector<std::string>& argv,
 292                                   scoped_ptr<FileDescriptorInfo> mapping,
 293                                   const std::string& process_type) {
 294   DCHECK(init_);
 295   Pickle pickle;
 296
 297   int raw_socks[2];
 298   PCHECK(0 == socketpair(AF_UNIX, SOCK_SEQPACKET, 0, raw_socks));
 299   base::ScopedFD my_sock(raw_socks[0]);
 300   base::ScopedFD peer_sock(raw_socks[1]);
 301   CHECK(UnixDomainSocket::EnableReceiveProcessId(my_sock.get()));
 302
 303   pickle.WriteInt(kZygoteCommandFork);
 304   pickle.WriteString(process_type);
 305   pickle.WriteInt(argv.size());
 306   for (std::vector<std::string>::const_iterator
 307        i = argv.begin(); i != argv.end(); ++i)
 308     pickle.WriteString(*i);
 309
 310   // Fork requests contain one file descriptor for the PID oracle, and one
 311   // more for each file descriptor mapping for the child process.
 312   const size_t num_fds_to_send = 1 + mapping->GetMappingSize();
 313   pickle.WriteInt(num_fds_to_send);
 314
 315   std::vector<int> fds;
 316
 317   // First FD to send is peer_sock.
 318   // TODO(morrita): Ideally, this should be part of the mapping so that
 319   // FileDescriptorInfo can manages its lifetime.
 320   fds.push_back(peer_sock.get());
 321
 322   // The rest come from mapping.
 323   for (size_t i = 0; i < mapping->GetMappingSize(); ++i) {
 324     pickle.WriteUInt32(mapping->GetIDAt(i));
 325     fds.push_back(mapping->GetFDAt(i));
 326   }
 327
 328   // Sanity check that we've populated |fds| correctly.
 329   DCHECK_EQ(num_fds_to_send, fds.size());
 330
 331   pid_t pid;
 332   {
 333     base::AutoLock lock(control_lock_);
 334     if (!SendMessage(pickle, &fds))
 335       return base::kNullProcessHandle;
 336     mapping.reset();
 337     peer_sock.reset();
 338
 339     {
 340       char buf[sizeof(kZygoteChildPingMessage) + 1];
 341       ScopedVector<base::ScopedFD> recv_fds;
 342       base::ProcessId real_pid;
 343
 344       ssize_t n = UnixDomainSocket::RecvMsgWithPid(
 345           my_sock.get(), buf, sizeof(buf), &recv_fds, &real_pid);
 346       if (n != sizeof(kZygoteChildPingMessage) ||
 347           0 != memcmp(buf,
 348                       kZygoteChildPingMessage,
 349                       sizeof(kZygoteChildPingMessage))) {
 350         // Zygote children should still be trustworthy when they're supposed to
 351         // ping us, so something's broken if we don't receive a valid ping.
 352         LOG(ERROR) << "Did not receive ping from zygote child";
 353         NOTREACHED();
 354         real_pid = -1;
 355       }
 356       my_sock.reset();
 357
 358       // Always send PID back to zygote.
 359       Pickle pid_pickle;
 360       pid_pickle.WriteInt(kZygoteCommandForkRealPID);
 361       pid_pickle.WriteInt(real_pid);
 362       if (!SendMessage(pid_pickle, NULL))
 363         return base::kNullProcessHandle;
 364     }
 365
 366     // Read the reply, which pickles the PID and an optional UMA enumeration.
 367     static const unsigned kMaxReplyLength = 2048;
 368     char buf[kMaxReplyLength];
 369     const ssize_t len = ReadReply(buf, sizeof(buf));
 370
 371     Pickle reply_pickle(buf, len);
 372     PickleIterator iter(reply_pickle);
 373     if (len <= 0 || !iter.ReadInt(&pid))
 374       return base::kNullProcessHandle;
 375
 376     // If there is a nonempty UMA name string, then there is a UMA
 377     // enumeration to record.
 378     std::string uma_name;
 379     int uma_sample;
 380     int uma_boundary_value;
 381     if (iter.ReadString(&uma_name) &&
 382         !uma_name.empty() &&
 383         iter.ReadInt(&uma_sample) &&
 384         iter.ReadInt(&uma_boundary_value)) {
 385       // We cannot use the UMA_HISTOGRAM_ENUMERATION macro here,
 386       // because that's only for when the name is the same every time.
 387       // Here we're using whatever name we got from the other side.
 388       // But since it's likely that the same one will be used repeatedly
 389       // (even though it's not guaranteed), we cache it here.
 390       static base::HistogramBase* uma_histogram;
 391       if (!uma_histogram || uma_histogram->histogram_name() != uma_name) {
 392         uma_histogram = base::LinearHistogram::FactoryGet(
 393             uma_name, 1,
 394             uma_boundary_value,
 395             uma_boundary_value + 1,
 396             base::HistogramBase::kUmaTargetedHistogramFlag);
 397       }
 398       uma_histogram->Add(uma_sample);
 399     }
 400
 401     if (pid <= 0)
 402       return base::kNullProcessHandle;
 403   }
 404
 405 #if !defined(OS_OPENBSD)
 406   // This is just a starting score for a renderer or extension (the
 407   // only types of processes that will be started this way).  It will
 408   // get adjusted as time goes on.  (This is the same value as
 409   // chrome::kLowestRendererOomScore in chrome/chrome_constants.h, but
 410   // that's not something we can include here.)
 411   const int kLowestRendererOomScore = 300;
 412   AdjustRendererOOMScore(pid, kLowestRendererOomScore);
 413 #endif
 414
 415   ZygoteChildBorn(pid);
 416   return pid;
 417 }
 418
 419 #if !defined(OS_OPENBSD)
 420 void ZygoteHostImpl::AdjustRendererOOMScore(base::ProcessHandle pid,
 421                                             int score) {
 422   // 1) You can't change the oom_score_adj of a non-dumpable process
 423   //    (EPERM) unless you're root. Because of this, we can't set the
 424   //    oom_adj from the browser process.
 425   //
 426   // 2) We can't set the oom_score_adj before entering the sandbox
 427   //    because the zygote is in the sandbox and the zygote is as
 428   //    critical as the browser process. Its oom_adj value shouldn't
 429   //    be changed.
 430   //
 431   // 3) A non-dumpable process can't even change its own oom_score_adj
 432   //    because it's root owned 0644. The sandboxed processes don't
 433   //    even have /proc, but one could imagine passing in a descriptor
 434   //    from outside.
 435   //
 436   // So, in the normal case, we use the SUID binary to change it for us.
 437   // However, Fedora (and other SELinux systems) don't like us touching other
 438   // process's oom_score_adj (or oom_adj) values
 439   // (https://bugzilla.redhat.com/show_bug.cgi?id=581256).
 440   //
 441   // The offical way to get the SELinux mode is selinux_getenforcemode, but I
 442   // don't want to add another library to the build as it's sure to cause
 443   // problems with other, non-SELinux distros.
 444   //
 445   // So we just check for files in /selinux. This isn't foolproof, but it's not
 446   // bad and it's easy.
 447
 448   static bool selinux;
 449   static bool selinux_valid = false;
 450
 451   if (!selinux_valid) {
 452     const base::FilePath kSelinuxPath("/selinux");
 453     base::FileEnumerator en(kSelinuxPath, false, base::FileEnumerator::FILES);
 454     bool has_selinux_files = !en.Next().empty();
 455
 456     selinux = access(kSelinuxPath.value().c_str(), X_OK) == 0 &&
 457               has_selinux_files;
 458     selinux_valid = true;
 459   }
 460
 461   if (using_suid_sandbox_ && !selinux) {
 462 #if defined(USE_TCMALLOC)
 463     // If heap profiling is running, these processes are not exiting, at least
 464     // on ChromeOS. The easiest thing to do is not launch them when profiling.
 465     // TODO(stevenjb): Investigate further and fix.
 466     if (IsHeapProfilerRunning())
 467       return;
 468 #endif
 469     std::vector<std::string> adj_oom_score_cmdline;
 470     adj_oom_score_cmdline.push_back(sandbox_binary_);
 471     adj_oom_score_cmdline.push_back(sandbox::kAdjustOOMScoreSwitch);
 472     adj_oom_score_cmdline.push_back(base::Int64ToString(pid));
 473     adj_oom_score_cmdline.push_back(base::IntToString(score));
 474
 475     base::Process sandbox_helper_process;
 476     base::LaunchOptions options;
 477
 478     // sandbox_helper_process is a setuid binary.
 479     options.allow_new_privs = true;
 480
 481     sandbox_helper_process =
 482         base::LaunchProcess(adj_oom_score_cmdline, options);
 483     if (sandbox_helper_process.IsValid())
 484       base::EnsureProcessGetsReaped(sandbox_helper_process.pid());
 485   } else if (!using_suid_sandbox_) {
 486     if (!base::AdjustOOMScore(pid, score))
 487       PLOG(ERROR) << "Failed to adjust OOM score of renderer with pid " << pid;
 488   }
 489 }
 490 #endif
 491
 492 void ZygoteHostImpl::EnsureProcessTerminated(pid_t process) {
 493   DCHECK(init_);
 494   Pickle pickle;
 495
 496   pickle.WriteInt(kZygoteCommandReap);
 497   pickle.WriteInt(process);
 498   if (!SendMessage(pickle, NULL))
 499     LOG(ERROR) << "Failed to send Reap message to zygote";
 500   ZygoteChildDied(process);
 501 }
 502
 503 base::TerminationStatus ZygoteHostImpl::GetTerminationStatus(
 504     base::ProcessHandle handle,
 505     bool known_dead,
 506     int* exit_code) {
 507   DCHECK(init_);
 508   Pickle pickle;
 509   pickle.WriteInt(kZygoteCommandGetTerminationStatus);
 510   pickle.WriteBool(known_dead);
 511   pickle.WriteInt(handle);
 512
 513   static const unsigned kMaxMessageLength = 128;
 514   char buf[kMaxMessageLength];
 515   ssize_t len;
 516   {
 517     base::AutoLock lock(control_lock_);
 518     if (!SendMessage(pickle, NULL))
 519       LOG(ERROR) << "Failed to send GetTerminationStatus message to zygote";
 520     len = ReadReply(buf, sizeof(buf));
 521   }
 522
 523   // Set this now to handle the error cases.
 524   if (exit_code)
 525     *exit_code = RESULT_CODE_NORMAL_EXIT;
 526   int status = base::TERMINATION_STATUS_NORMAL_TERMINATION;
 527
 528   if (len == -1) {
 529     LOG(WARNING) << "Error reading message from zygote: " << errno;
 530   } else if (len == 0) {
 531     LOG(WARNING) << "Socket closed prematurely.";
 532   } else {
 533     Pickle read_pickle(buf, len);
 534     int tmp_status, tmp_exit_code;
 535     PickleIterator iter(read_pickle);
 536     if (!iter.ReadInt(&tmp_status) || !iter.ReadInt(&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