content/common/sandbox_linux/sandbox_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 <dirent.h>
   6 #include <fcntl.h>
   7 #include <sys/resource.h>
   8 #include <sys/stat.h>
   9 #include <sys/time.h>
  10 #include <sys/types.h>
  11 #include <unistd.h>
  12
  13 #include <limits>
  14
  15 #include "base/bind.h"
  16 #include "base/callback_helpers.h"
  17 #include "base/command_line.h"
  18 #include "base/debug/stack_trace.h"
  19 #include "base/files/scoped_file.h"
  20 #include "base/logging.h"
  21 #include "base/memory/scoped_ptr.h"
  22 #include "base/memory/singleton.h"
  23 #include "base/posix/eintr_wrapper.h"
  24 #include "base/strings/string_number_conversions.h"
  25 #include "base/sys_info.h"
  26 #include "base/time/time.h"
  27 #include "build/build_config.h"
  28 #include "content/common/sandbox_linux/sandbox_linux.h"
  29 #include "content/common/sandbox_linux/sandbox_seccomp_bpf_linux.h"
  30 #include "content/public/common/content_switches.h"
  31 #include "content/public/common/sandbox_linux.h"
  32 #include "sandbox/linux/services/credentials.h"
  33 #include "sandbox/linux/services/thread_helpers.h"
  34 #include "sandbox/linux/services/yama.h"
  35 #include "sandbox/linux/suid/client/setuid_sandbox_client.h"
  36
  37 using sandbox::Yama;
  38
  39 namespace {
  40
  41 struct FDCloser {
  42   inline void operator()(int* fd) const {
  43     DCHECK(fd);
  44     PCHECK(0 == IGNORE_EINTR(close(*fd)));
  45     *fd = -1;
  46   }
  47 };
  48
  49 void LogSandboxStarted(const std::string& sandbox_name) {
  50   const CommandLine& command_line = *CommandLine::ForCurrentProcess();
  51   const std::string process_type =
  52       command_line.GetSwitchValueASCII(switches::kProcessType);
  53   const std::string activated_sandbox =
  54       "Activated " + sandbox_name + " sandbox for process type: " +
  55       process_type + ".";
  56 #if defined(OS_CHROMEOS)
  57   LOG(WARNING) << activated_sandbox;
  58 #else
  59   VLOG(1) << activated_sandbox;
  60 #endif
  61 }
  62
  63 bool AddResourceLimit(int resource, rlim_t limit) {
  64   struct rlimit old_rlimit;
  65   if (getrlimit(resource, &old_rlimit))
  66     return false;
  67   // Make sure we don't raise the existing limit.
  68   const struct rlimit new_rlimit = {
  69       std::min(old_rlimit.rlim_cur, limit),
  70       std::min(old_rlimit.rlim_max, limit)
  71       };
  72   int rc = setrlimit(resource, &new_rlimit);
  73   return rc == 0;
  74 }
  75
  76 bool IsRunningTSAN() {
  77 #if defined(THREAD_SANITIZER)
  78   return true;
  79 #else
  80   return false;
  81 #endif
  82 }
  83
  84 // Try to open /proc/self/task/ with the help of |proc_fd|. |proc_fd| can be
  85 // -1. Will return -1 on error and set errno like open(2).
  86 int OpenProcTaskFd(int proc_fd) {
  87   int proc_self_task = -1;
  88   if (proc_fd >= 0) {
  89     // If a handle to /proc is available, use it. This allows to bypass file
  90     // system restrictions.
  91     proc_self_task = openat(proc_fd, "self/task/", O_RDONLY | O_DIRECTORY);
  92   } else {
  93     // Otherwise, make an attempt to access the file system directly.
  94     proc_self_task = open("/proc/self/task/", O_RDONLY | O_DIRECTORY);
  95   }
  96   return proc_self_task;
  97 }
  98
  99 }  // namespace
 100
 101 namespace content {
 102
 103 LinuxSandbox::LinuxSandbox()
 104     : proc_fd_(-1),
 105       seccomp_bpf_started_(false),
 106       sandbox_status_flags_(kSandboxLinuxInvalid),
 107       pre_initialized_(false),
 108       seccomp_bpf_supported_(false),
 109       yama_is_enforcing_(false),
 110       setuid_sandbox_client_(sandbox::SetuidSandboxClient::Create()) {
 111   if (setuid_sandbox_client_ == NULL) {
 112     LOG(FATAL) << "Failed to instantiate the setuid sandbox client.";
 113   }
 114 }
 115
 116 LinuxSandbox::~LinuxSandbox() {
 117 }
 118
 119 LinuxSandbox* LinuxSandbox::GetInstance() {
 120   LinuxSandbox* instance = Singleton<LinuxSandbox>::get();
 121   CHECK(instance);
 122   return instance;
 123 }
 124
 125 #if defined(ADDRESS_SANITIZER) && defined(OS_LINUX)
 126 // ASan API call to notify the tool the sandbox is going to be turned on.
 127 extern "C" void __sanitizer_sandbox_on_notify(void *reserved);
 128 #endif
 129
 130 void LinuxSandbox::PreinitializeSandbox() {
 131   CHECK(!pre_initialized_);
 132   seccomp_bpf_supported_ = false;
 133 #if defined(ADDRESS_SANITIZER) && defined(OS_LINUX)
 134   // ASan needs to open some resources before the sandbox is enabled.
 135   // This should not fork, not launch threads, not open a directory.
 136   __sanitizer_sandbox_on_notify(/*reserved*/ NULL);
 137 #endif
 138
 139 #if !defined(NDEBUG)
 140   // The in-process stack dumping needs to open /proc/self/maps and cache
 141   // its contents before the sandbox is enabled.  It also pre-opens the
 142   // object files that are already loaded in the process address space.
 143   base::debug::EnableInProcessStackDumpingForSandbox();
 144
 145   // Open proc_fd_ only in Debug mode so that forgetting to close it doesn't
 146   // produce a sandbox escape in Release mode.
 147   proc_fd_ = open("/proc", O_DIRECTORY | O_RDONLY | O_CLOEXEC);
 148   CHECK_GE(proc_fd_, 0);
 149 #endif  // !defined(NDEBUG)
 150   // We "pre-warm" the code that detects supports for seccomp BPF.
 151   if (SandboxSeccompBPF::IsSeccompBPFDesired()) {
 152     if (!SandboxSeccompBPF::SupportsSandbox()) {
 153       VLOG(1) << "Lacking support for seccomp-bpf sandbox.";
 154     } else {
 155       seccomp_bpf_supported_ = true;
 156     }
 157   }
 158
 159   // Yama is a "global", system-level status. We assume it will not regress
 160   // after startup.
 161   const int yama_status = Yama::GetStatus();
 162   yama_is_enforcing_ = (yama_status & Yama::STATUS_PRESENT) &&
 163                        (yama_status & Yama::STATUS_ENFORCING);
 164   pre_initialized_ = true;
 165 }
 166
 167 bool LinuxSandbox::InitializeSandbox() {
 168   LinuxSandbox* linux_sandbox = LinuxSandbox::GetInstance();
 169   return linux_sandbox->InitializeSandboxImpl();
 170 }
 171
 172 void LinuxSandbox::StopThread(base::Thread* thread) {
 173   LinuxSandbox* linux_sandbox = LinuxSandbox::GetInstance();
 174   linux_sandbox->StopThreadImpl(thread);
 175 }
 176
 177 int LinuxSandbox::GetStatus() {
 178   CHECK(pre_initialized_);
 179   if (kSandboxLinuxInvalid == sandbox_status_flags_) {
 180     // Initialize sandbox_status_flags_.
 181     sandbox_status_flags_ = 0;
 182     if (setuid_sandbox_client_->IsSandboxed()) {
 183       sandbox_status_flags_ |= kSandboxLinuxSUID;
 184       if (setuid_sandbox_client_->IsInNewPIDNamespace())
 185         sandbox_status_flags_ |= kSandboxLinuxPIDNS;
 186       if (setuid_sandbox_client_->IsInNewNETNamespace())
 187         sandbox_status_flags_ |= kSandboxLinuxNetNS;
 188     }
 189
 190     // We report whether the sandbox will be activated when renderers, workers
 191     // and PPAPI plugins go through sandbox initialization.
 192     if (seccomp_bpf_supported() &&
 193         SandboxSeccompBPF::ShouldEnableSeccompBPF(switches::kRendererProcess)) {
 194       sandbox_status_flags_ |= kSandboxLinuxSeccompBPF;
 195     }
 196
 197     if (yama_is_enforcing_) {
 198       sandbox_status_flags_ |= kSandboxLinuxYama;
 199     }
 200   }
 201
 202   return sandbox_status_flags_;
 203 }
 204
 205 // Threads are counted via /proc/self/task. This is a little hairy because of
 206 // PID namespaces and existing sandboxes, so "self" must really be used instead
 207 // of using the pid.
 208 bool LinuxSandbox::IsSingleThreaded() const {
 209   bool is_single_threaded = false;
 210   base::ScopedFD proc_self_task(OpenProcTaskFd(proc_fd_));
 211
 212 // In Debug mode, it's mandatory to be able to count threads to catch bugs.
 213 #if !defined(NDEBUG)
 214   // Using CHECK here since we want to check all the cases where
 215   // !defined(NDEBUG)
 216   // gets built.
 217   CHECK(proc_self_task.is_valid())
 218       << "Could not count threads, the sandbox was not "
 219       << "pre-initialized properly.";
 220 #endif  // !defined(NDEBUG)
 221
 222   if (!proc_self_task.is_valid()) {
 223     // Pretend to be monothreaded if it can't be determined (for instance the
 224     // setuid sandbox is already engaged but no proc_fd_ is available).
 225     is_single_threaded = true;
 226   } else {
 227     is_single_threaded =
 228         sandbox::ThreadHelpers::IsSingleThreaded(proc_self_task.get());
 229   }
 230
 231   return is_single_threaded;
 232 }
 233
 234 bool LinuxSandbox::seccomp_bpf_started() const {
 235   return seccomp_bpf_started_;
 236 }
 237
 238 sandbox::SetuidSandboxClient*
 239     LinuxSandbox::setuid_sandbox_client() const {
 240   return setuid_sandbox_client_.get();
 241 }
 242
 243 // For seccomp-bpf, we use the SandboxSeccompBPF class.
 244 bool LinuxSandbox::StartSeccompBPF(const std::string& process_type) {
 245   CHECK(!seccomp_bpf_started_);
 246   CHECK(pre_initialized_);
 247   if (seccomp_bpf_supported())
 248     seccomp_bpf_started_ = SandboxSeccompBPF::StartSandbox(process_type);
 249
 250   if (seccomp_bpf_started_)
 251     LogSandboxStarted("seccomp-bpf");
 252
 253   return seccomp_bpf_started_;
 254 }
 255
 256 bool LinuxSandbox::InitializeSandboxImpl() {
 257   CommandLine* command_line = CommandLine::ForCurrentProcess();
 258   const std::string process_type =
 259       command_line->GetSwitchValueASCII(switches::kProcessType);
 260
 261   // We need to make absolutely sure that our sandbox is "sealed" before
 262   // returning.
 263   // Unretained() since the current object is a Singleton.
 264   base::ScopedClosureRunner sandbox_sealer(
 265       base::Bind(&LinuxSandbox::SealSandbox, base::Unretained(this)));
 266   // Make sure that this function enables sandboxes as promised by GetStatus().
 267   // Unretained() since the current object is a Singleton.
 268   base::ScopedClosureRunner sandbox_promise_keeper(
 269       base::Bind(&LinuxSandbox::CheckForBrokenPromises,
 270                  base::Unretained(this),
 271                  process_type));
 272
 273   // No matter what, it's always an error to call InitializeSandbox() after
 274   // threads have been created.
 275   if (!IsSingleThreaded()) {
 276     std::string error_message = "InitializeSandbox() called with multiple "
 277                                 "threads in process " + process_type;
 278     // TSAN starts a helper thread, so we don't start the sandbox and don't
 279     // even report an error about it.
 280     if (IsRunningTSAN())
 281       return false;
 282
 283     // The GPU process is allowed to call InitializeSandbox() with threads.
 284     bool sandbox_failure_fatal = process_type != switches::kGpuProcess;
 285     // This can be disabled with the '--gpu-sandbox-failures-fatal' flag.
 286     // Setting the flag with no value or any value different than 'yes' or 'no'
 287     // is equal to setting '--gpu-sandbox-failures-fatal=yes'.
 288     if (process_type == switches::kGpuProcess &&
 289         command_line->HasSwitch(switches::kGpuSandboxFailuresFatal)) {
 290       const std::string switch_value =
 291           command_line->GetSwitchValueASCII(switches::kGpuSandboxFailuresFatal);
 292       sandbox_failure_fatal = switch_value != "no";
 293     }
 294
 295     if (sandbox_failure_fatal)
 296       LOG(FATAL) << error_message;
 297
 298     LOG(ERROR) << error_message;
 299     return false;
 300   }
 301
 302   // Only one thread is running, pre-initialize if not already done.
 303   if (!pre_initialized_)
 304     PreinitializeSandbox();
 305
 306   DCHECK(!HasOpenDirectories()) <<
 307       "InitializeSandbox() called after unexpected directories have been " <<
 308       "opened. This breaks the security of the setuid sandbox.";
 309
 310   // Attempt to limit the future size of the address space of the process.
 311   LimitAddressSpace(process_type);
 312
 313   // Try to enable seccomp-bpf.
 314   bool seccomp_bpf_started = StartSeccompBPF(process_type);
 315
 316   return seccomp_bpf_started;
 317 }
 318
 319 void LinuxSandbox::StopThreadImpl(base::Thread* thread) {
 320   DCHECK(thread);
 321   StopThreadAndEnsureNotCounted(thread);
 322 }
 323
 324 bool LinuxSandbox::seccomp_bpf_supported() const {
 325   CHECK(pre_initialized_);
 326   return seccomp_bpf_supported_;
 327 }
 328
 329 bool LinuxSandbox::LimitAddressSpace(const std::string& process_type) {
 330   (void) process_type;
 331 #if !defined(ADDRESS_SANITIZER)
 332   CommandLine* command_line = CommandLine::ForCurrentProcess();
 333   if (command_line->HasSwitch(switches::kNoSandbox)) {
 334     return false;
 335   }
 336
 337   // Limit the address space to 4GB.
 338   // This is in the hope of making some kernel exploits more complex and less
 339   // reliable. It also limits sprays a little on 64-bit.
 340   rlim_t address_space_limit = std::numeric_limits<uint32_t>::max();
 341 #if defined(__LP64__)
 342   // On 64 bits, V8 and possibly others will reserve massive memory ranges and
 343   // rely on on-demand paging for allocation.  Unfortunately, even
 344   // MADV_DONTNEED ranges  count towards RLIMIT_AS so this is not an option.
 345   // See crbug.com/169327 for a discussion.
 346   // On the GPU process, irrespective of V8, we can exhaust a 4GB address space
 347   // under normal usage, see crbug.com/271119
 348   // For now, increase limit to 16GB for renderer and worker and gpu processes
 349   // to accomodate.
 350   if (process_type == switches::kRendererProcess ||
 351       process_type == switches::kWorkerProcess ||
 352       process_type == switches::kGpuProcess) {
 353     address_space_limit = 1L << 34;
 354   }
 355 #endif  // defined(__LP64__)
 356
 357   // On all platforms, add a limit to the brk() heap that would prevent
 358   // allocations that can't be index by an int.
 359   const rlim_t kNewDataSegmentMaxSize = std::numeric_limits<int>::max();
 360
 361   bool limited_as = AddResourceLimit(RLIMIT_AS, address_space_limit);
 362   bool limited_data = AddResourceLimit(RLIMIT_DATA, kNewDataSegmentMaxSize);
 363
 364   // Cache the resource limit before turning on the sandbox.
 365   base::SysInfo::AmountOfVirtualMemory();
 366
 367   return limited_as && limited_data;
 368 #else
 369   base::SysInfo::AmountOfVirtualMemory();
 370   return false;
 371 #endif  // !defined(ADDRESS_SANITIZER)
 372 }
 373
 374 bool LinuxSandbox::HasOpenDirectories() const {
 375   return sandbox::Credentials().HasOpenDirectory(proc_fd_);
 376 }
 377
 378 void LinuxSandbox::SealSandbox() {
 379   if (proc_fd_ >= 0) {
 380     int ret = IGNORE_EINTR(close(proc_fd_));
 381     CHECK_EQ(0, ret);
 382     proc_fd_ = -1;
 383   }
 384 }
 385
 386 void LinuxSandbox::CheckForBrokenPromises(const std::string& process_type) {
 387   // Make sure that any promise made with GetStatus() wasn't broken.
 388   bool promised_seccomp_bpf_would_start = false;
 389   if (process_type == switches::kRendererProcess ||
 390       process_type == switches::kWorkerProcess ||
 391       process_type == switches::kPpapiPluginProcess) {
 392     promised_seccomp_bpf_would_start =
 393         (sandbox_status_flags_ != kSandboxLinuxInvalid) &&
 394         (GetStatus() & kSandboxLinuxSeccompBPF);
 395   }
 396   if (promised_seccomp_bpf_would_start) {
 397     CHECK(seccomp_bpf_started_);
 398   }
 399 }
 400
 401 void LinuxSandbox::StopThreadAndEnsureNotCounted(base::Thread* thread) const {
 402   DCHECK(thread);
 403   base::ScopedFD proc_self_task(OpenProcTaskFd(proc_fd_));
 404   PCHECK(proc_self_task.is_valid());
 405   CHECK(sandbox::ThreadHelpers::StopThreadAndWatchProcFS(proc_self_task.get(),
 406                                                          thread));
 407 }
 408
 409 }  // namespace content