libsanitizer/sanitizer_common/sanitizer_stoptheworld_linux_libcdep.cc

   1 //===-- sanitizer_stoptheworld_linux_libcdep.cc ---------------------------===//
   2 //
   3 // This file is distributed under the University of Illinois Open Source
   4 // License. See LICENSE.TXT for details.
   5 //
   6 //===----------------------------------------------------------------------===//
   7 //
   8 // See sanitizer_stoptheworld.h for details.
   9 // This implementation was inspired by Markus Gutschke's linuxthreads.cc.
  10 //
  11 //===----------------------------------------------------------------------===//
  12
  13 #include "sanitizer_platform.h"
  14
  15 #if SANITIZER_LINUX && (defined(__x86_64__) || defined(__mips__) || \
  16                         defined(__aarch64__))
  17
  18 #include "sanitizer_stoptheworld.h"
  19
  20 #include "sanitizer_platform_limits_posix.h"
  21 #include "sanitizer_atomic.h"
  22
  23 #include <errno.h>
  24 #include <sched.h> // for CLONE_* definitions
  25 #include <stddef.h>
  26 #include <sys/prctl.h> // for PR_* definitions
  27 #include <sys/ptrace.h> // for PTRACE_* definitions
  28 #include <sys/types.h> // for pid_t
  29 #include <sys/uio.h> // for iovec
  30 #include <elf.h> // for NT_PRSTATUS
  31 #if SANITIZER_ANDROID && defined(__arm__)
  32 # include <linux/user.h>  // for pt_regs
  33 #else
  34 # ifdef __aarch64__
  35 // GLIBC 2.20+ sys/user does not include asm/ptrace.h
  36 #  include <asm/ptrace.h>
  37 # endif
  38 # include <sys/user.h>  // for user_regs_struct
  39 #endif
  40 #include <sys/wait.h> // for signal-related stuff
  41
  42 #ifdef sa_handler
  43 # undef sa_handler
  44 #endif
  45
  46 #ifdef sa_sigaction
  47 # undef sa_sigaction
  48 #endif
  49
  50 #include "sanitizer_common.h"
  51 #include "sanitizer_flags.h"
  52 #include "sanitizer_libc.h"
  53 #include "sanitizer_linux.h"
  54 #include "sanitizer_mutex.h"
  55 #include "sanitizer_placement_new.h"
  56
  57 // This module works by spawning a Linux task which then attaches to every
  58 // thread in the caller process with ptrace. This suspends the threads, and
  59 // PTRACE_GETREGS can then be used to obtain their register state. The callback
  60 // supplied to StopTheWorld() is run in the tracer task while the threads are
  61 // suspended.
  62 // The tracer task must be placed in a different thread group for ptrace to
  63 // work, so it cannot be spawned as a pthread. Instead, we use the low-level
  64 // clone() interface (we want to share the address space with the caller
  65 // process, so we prefer clone() over fork()).
  66 //
  67 // We don't use any libc functions, relying instead on direct syscalls. There
  68 // are two reasons for this:
  69 // 1. calling a library function while threads are suspended could cause a
  70 // deadlock, if one of the treads happens to be holding a libc lock;
  71 // 2. it's generally not safe to call libc functions from the tracer task,
  72 // because clone() does not set up a thread-local storage for it. Any
  73 // thread-local variables used by libc will be shared between the tracer task
  74 // and the thread which spawned it.
  75
  76 COMPILER_CHECK(sizeof(SuspendedThreadID) == sizeof(pid_t));
  77
  78 namespace __sanitizer {
  79
  80 // Structure for passing arguments into the tracer thread.
  81 struct TracerThreadArgument {
  82   StopTheWorldCallback callback;
  83   void *callback_argument;
  84   // The tracer thread waits on this mutex while the parent finishes its
  85   // preparations.
  86   BlockingMutex mutex;
  87   // Tracer thread signals its completion by setting done.
  88   atomic_uintptr_t done;
  89   uptr parent_pid;
  90 };
  91
  92 // This class handles thread suspending/unsuspending in the tracer thread.
  93 class ThreadSuspender {
  94  public:
  95   explicit ThreadSuspender(pid_t pid, TracerThreadArgument *arg)
  96     : arg(arg)
  97     , pid_(pid) {
  98       CHECK_GE(pid, 0);
  99     }
 100   bool SuspendAllThreads();
 101   void ResumeAllThreads();
 102   void KillAllThreads();
 103   SuspendedThreadsList &suspended_threads_list() {
 104     return suspended_threads_list_;
 105   }
 106   TracerThreadArgument *arg;
 107  private:
 108   SuspendedThreadsList suspended_threads_list_;
 109   pid_t pid_;
 110   bool SuspendThread(SuspendedThreadID thread_id);
 111 };
 112
 113 bool ThreadSuspender::SuspendThread(SuspendedThreadID tid) {
 114   // Are we already attached to this thread?
 115   // Currently this check takes linear time, however the number of threads is
 116   // usually small.
 117   if (suspended_threads_list_.Contains(tid))
 118     return false;
 119   int pterrno;
 120   if (internal_iserror(internal_ptrace(PTRACE_ATTACH, tid, nullptr, nullptr),
 121                        &pterrno)) {
 122     // Either the thread is dead, or something prevented us from attaching.
 123     // Log this event and move on.
 124     VReport(1, "Could not attach to thread %d (errno %d).\n", tid, pterrno);
 125     return false;
 126   } else {
 127     VReport(2, "Attached to thread %d.\n", tid);
 128     // The thread is not guaranteed to stop before ptrace returns, so we must
 129     // wait on it. Note: if the thread receives a signal concurrently,
 130     // we can get notification about the signal before notification about stop.
 131     // In such case we need to forward the signal to the thread, otherwise
 132     // the signal will be missed (as we do PTRACE_DETACH with arg=0) and
 133     // any logic relying on signals will break. After forwarding we need to
 134     // continue to wait for stopping, because the thread is not stopped yet.
 135     // We do ignore delivery of SIGSTOP, because we want to make stop-the-world
 136     // as invisible as possible.
 137     for (;;) {
 138       int status;
 139       uptr waitpid_status;
 140       HANDLE_EINTR(waitpid_status, internal_waitpid(tid, &status, __WALL));
 141       int wperrno;
 142       if (internal_iserror(waitpid_status, &wperrno)) {
 143         // Got a ECHILD error. I don't think this situation is possible, but it
 144         // doesn't hurt to report it.
 145         VReport(1, "Waiting on thread %d failed, detaching (errno %d).\n",
 146                 tid, wperrno);
 147         internal_ptrace(PTRACE_DETACH, tid, nullptr, nullptr);
 148         return false;
 149       }
 150       if (WIFSTOPPED(status) && WSTOPSIG(status) != SIGSTOP) {
 151         internal_ptrace(PTRACE_CONT, tid, nullptr,
 152                         (void*)(uptr)WSTOPSIG(status));
 153         continue;
 154       }
 155       break;
 156     }
 157     suspended_threads_list_.Append(tid);
 158     return true;
 159   }
 160 }
 161
 162 void ThreadSuspender::ResumeAllThreads() {
 163   for (uptr i = 0; i < suspended_threads_list_.thread_count(); i++) {
 164     pid_t tid = suspended_threads_list_.GetThreadID(i);
 165     int pterrno;
 166     if (!internal_iserror(internal_ptrace(PTRACE_DETACH, tid, nullptr, nullptr),
 167                           &pterrno)) {
 168       VReport(2, "Detached from thread %d.\n", tid);
 169     } else {
 170       // Either the thread is dead, or we are already detached.
 171       // The latter case is possible, for instance, if this function was called
 172       // from a signal handler.
 173       VReport(1, "Could not detach from thread %d (errno %d).\n", tid, pterrno);
 174     }
 175   }
 176 }
 177
 178 void ThreadSuspender::KillAllThreads() {
 179   for (uptr i = 0; i < suspended_threads_list_.thread_count(); i++)
 180     internal_ptrace(PTRACE_KILL, suspended_threads_list_.GetThreadID(i),
 181                     nullptr, nullptr);
 182 }
 183
 184 bool ThreadSuspender::SuspendAllThreads() {
 185   ThreadLister thread_lister(pid_);
 186   bool added_threads;
 187   do {
 188     // Run through the directory entries once.
 189     added_threads = false;
 190     pid_t tid = thread_lister.GetNextTID();
 191     while (tid >= 0) {
 192       if (SuspendThread(tid))
 193         added_threads = true;
 194       tid = thread_lister.GetNextTID();
 195     }
 196     if (thread_lister.error()) {
 197       // Detach threads and fail.
 198       ResumeAllThreads();
 199       return false;
 200     }
 201     thread_lister.Reset();
 202   } while (added_threads);
 203   return true;
 204 }
 205
 206 // Pointer to the ThreadSuspender instance for use in signal handler.
 207 static ThreadSuspender *thread_suspender_instance = nullptr;
 208
 209 // Synchronous signals that should not be blocked.
 210 static const int kSyncSignals[] = { SIGABRT, SIGILL, SIGFPE, SIGSEGV, SIGBUS,
 211                                     SIGXCPU, SIGXFSZ };
 212
 213 static void TracerThreadDieCallback() {
 214   // Generally a call to Die() in the tracer thread should be fatal to the
 215   // parent process as well, because they share the address space.
 216   // This really only works correctly if all the threads are suspended at this
 217   // point. So we correctly handle calls to Die() from within the callback, but
 218   // not those that happen before or after the callback. Hopefully there aren't
 219   // a lot of opportunities for that to happen...
 220   ThreadSuspender *inst = thread_suspender_instance;
 221   if (inst && stoptheworld_tracer_pid == internal_getpid()) {
 222     inst->KillAllThreads();
 223     thread_suspender_instance = nullptr;
 224   }
 225 }
 226
 227 // Signal handler to wake up suspended threads when the tracer thread dies.
 228 static void TracerThreadSignalHandler(int signum, void *siginfo, void *uctx) {
 229   SignalContext ctx = SignalContext::Create(siginfo, uctx);
 230   VPrintf(1, "Tracer caught signal %d: addr=0x%zx pc=0x%zx sp=0x%zx\n",
 231       signum, ctx.addr, ctx.pc, ctx.sp);
 232   ThreadSuspender *inst = thread_suspender_instance;
 233   if (inst) {
 234     if (signum == SIGABRT)
 235       inst->KillAllThreads();
 236     else
 237       inst->ResumeAllThreads();
 238     RAW_CHECK(RemoveDieCallback(TracerThreadDieCallback));
 239     thread_suspender_instance = nullptr;
 240     atomic_store(&inst->arg->done, 1, memory_order_relaxed);
 241   }
 242   internal__exit((signum == SIGABRT) ? 1 : 2);
 243 }
 244
 245 // Size of alternative stack for signal handlers in the tracer thread.
 246 static const int kHandlerStackSize = 4096;
 247
 248 // This function will be run as a cloned task.
 249 static int TracerThread(void* argument) {
 250   TracerThreadArgument *tracer_thread_argument =
 251       (TracerThreadArgument *)argument;
 252
 253   internal_prctl(PR_SET_PDEATHSIG, SIGKILL, 0, 0, 0);
 254   // Check if parent is already dead.
 255   if (internal_getppid() != tracer_thread_argument->parent_pid)
 256     internal__exit(4);
 257
 258   // Wait for the parent thread to finish preparations.
 259   tracer_thread_argument->mutex.Lock();
 260   tracer_thread_argument->mutex.Unlock();
 261
 262   RAW_CHECK(AddDieCallback(TracerThreadDieCallback));
 263
 264   ThreadSuspender thread_suspender(internal_getppid(), tracer_thread_argument);
 265   // Global pointer for the signal handler.
 266   thread_suspender_instance = &thread_suspender;
 267
 268   // Alternate stack for signal handling.
 269   InternalScopedBuffer<char> handler_stack_memory(kHandlerStackSize);
 270   struct sigaltstack handler_stack;
 271   internal_memset(&handler_stack, 0, sizeof(handler_stack));
 272   handler_stack.ss_sp = handler_stack_memory.data();
 273   handler_stack.ss_size = kHandlerStackSize;
 274   internal_sigaltstack(&handler_stack, nullptr);
 275
 276   // Install our handler for synchronous signals. Other signals should be
 277   // blocked by the mask we inherited from the parent thread.
 278   for (uptr i = 0; i < ARRAY_SIZE(kSyncSignals); i++) {
 279     __sanitizer_sigaction act;
 280     internal_memset(&act, 0, sizeof(act));
 281     act.sigaction = TracerThreadSignalHandler;
 282     act.sa_flags = SA_ONSTACK | SA_SIGINFO;
 283     internal_sigaction_norestorer(kSyncSignals[i], &act, 0);
 284   }
 285
 286   int exit_code = 0;
 287   if (!thread_suspender.SuspendAllThreads()) {
 288     VReport(1, "Failed suspending threads.\n");
 289     exit_code = 3;
 290   } else {
 291     tracer_thread_argument->callback(thread_suspender.suspended_threads_list(),
 292                                      tracer_thread_argument->callback_argument);
 293     thread_suspender.ResumeAllThreads();
 294     exit_code = 0;
 295   }
 296   RAW_CHECK(RemoveDieCallback(TracerThreadDieCallback));
 297   thread_suspender_instance = nullptr;
 298   atomic_store(&tracer_thread_argument->done, 1, memory_order_relaxed);
 299   return exit_code;
 300 }
 301
 302 class ScopedStackSpaceWithGuard {
 303  public:
 304   explicit ScopedStackSpaceWithGuard(uptr stack_size) {
 305     stack_size_ = stack_size;
 306     guard_size_ = GetPageSizeCached();
 307     // FIXME: Omitting MAP_STACK here works in current kernels but might break
 308     // in the future.
 309     guard_start_ = (uptr)MmapOrDie(stack_size_ + guard_size_,
 310                                    "ScopedStackWithGuard");
 311     CHECK(MprotectNoAccess((uptr)guard_start_, guard_size_));
 312   }
 313   ~ScopedStackSpaceWithGuard() {
 314     UnmapOrDie((void *)guard_start_, stack_size_ + guard_size_);
 315   }
 316   void *Bottom() const {
 317     return (void *)(guard_start_ + stack_size_ + guard_size_);
 318   }
 319
 320  private:
 321   uptr stack_size_;
 322   uptr guard_size_;
 323   uptr guard_start_;
 324 };
 325
 326 // We have a limitation on the stack frame size, so some stuff had to be moved
 327 // into globals.
 328 static __sanitizer_sigset_t blocked_sigset;
 329 static __sanitizer_sigset_t old_sigset;
 330
 331 class StopTheWorldScope {
 332  public:
 333   StopTheWorldScope() {
 334     // Make this process dumpable. Processes that are not dumpable cannot be
 335     // attached to.
 336     process_was_dumpable_ = internal_prctl(PR_GET_DUMPABLE, 0, 0, 0, 0);
 337     if (!process_was_dumpable_)
 338       internal_prctl(PR_SET_DUMPABLE, 1, 0, 0, 0);
 339   }
 340
 341   ~StopTheWorldScope() {
 342     // Restore the dumpable flag.
 343     if (!process_was_dumpable_)
 344       internal_prctl(PR_SET_DUMPABLE, 0, 0, 0, 0);
 345   }
 346
 347  private:
 348   int process_was_dumpable_;
 349 };
 350
 351 // When sanitizer output is being redirected to file (i.e. by using log_path),
 352 // the tracer should write to the parent's log instead of trying to open a new
 353 // file. Alert the logging code to the fact that we have a tracer.
 354 struct ScopedSetTracerPID {
 355   explicit ScopedSetTracerPID(uptr tracer_pid) {
 356     stoptheworld_tracer_pid = tracer_pid;
 357     stoptheworld_tracer_ppid = internal_getpid();
 358   }
 359   ~ScopedSetTracerPID() {
 360     stoptheworld_tracer_pid = 0;
 361     stoptheworld_tracer_ppid = 0;
 362   }
 363 };
 364
 365 void StopTheWorld(StopTheWorldCallback callback, void *argument) {
 366   StopTheWorldScope in_stoptheworld;
 367   // Prepare the arguments for TracerThread.
 368   struct TracerThreadArgument tracer_thread_argument;
 369   tracer_thread_argument.callback = callback;
 370   tracer_thread_argument.callback_argument = argument;
 371   tracer_thread_argument.parent_pid = internal_getpid();
 372   atomic_store(&tracer_thread_argument.done, 0, memory_order_relaxed);
 373   const uptr kTracerStackSize = 2 * 1024 * 1024;
 374   ScopedStackSpaceWithGuard tracer_stack(kTracerStackSize);
 375   // Block the execution of TracerThread until after we have set ptrace
 376   // permissions.
 377   tracer_thread_argument.mutex.Lock();
 378   // Signal handling story.
 379   // We don't want async signals to be delivered to the tracer thread,
 380   // so we block all async signals before creating the thread. An async signal
 381   // handler can temporary modify errno, which is shared with this thread.
 382   // We ought to use pthread_sigmask here, because sigprocmask has undefined
 383   // behavior in multithreaded programs. However, on linux sigprocmask is
 384   // equivalent to pthread_sigmask with the exception that pthread_sigmask
 385   // does not allow to block some signals used internally in pthread
 386   // implementation. We are fine with blocking them here, we are really not
 387   // going to pthread_cancel the thread.
 388   // The tracer thread should not raise any synchronous signals. But in case it
 389   // does, we setup a special handler for sync signals that properly kills the
 390   // parent as well. Note: we don't pass CLONE_SIGHAND to clone, so handlers
 391   // in the tracer thread won't interfere with user program. Double note: if a
 392   // user does something along the lines of 'kill -11 pid', that can kill the
 393   // process even if user setup own handler for SEGV.
 394   // Thing to watch out for: this code should not change behavior of user code
 395   // in any observable way. In particular it should not override user signal
 396   // handlers.
 397   internal_sigfillset(&blocked_sigset);
 398   for (uptr i = 0; i < ARRAY_SIZE(kSyncSignals); i++)
 399     internal_sigdelset(&blocked_sigset, kSyncSignals[i]);
 400   int rv = internal_sigprocmask(SIG_BLOCK, &blocked_sigset, &old_sigset);
 401   CHECK_EQ(rv, 0);
 402   uptr tracer_pid = internal_clone(
 403       TracerThread, tracer_stack.Bottom(),
 404       CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_UNTRACED,
 405       &tracer_thread_argument, nullptr /* parent_tidptr */,
 406       nullptr /* newtls */, nullptr /* child_tidptr */);
 407   internal_sigprocmask(SIG_SETMASK, &old_sigset, 0);
 408   int local_errno = 0;
 409   if (internal_iserror(tracer_pid, &local_errno)) {
 410     VReport(1, "Failed spawning a tracer thread (errno %d).\n", local_errno);
 411     tracer_thread_argument.mutex.Unlock();
 412   } else {
 413     ScopedSetTracerPID scoped_set_tracer_pid(tracer_pid);
 414     // On some systems we have to explicitly declare that we want to be traced
 415     // by the tracer thread.
 416 #ifdef PR_SET_PTRACER
 417     internal_prctl(PR_SET_PTRACER, tracer_pid, 0, 0, 0);
 418 #endif
 419     // Allow the tracer thread to start.
 420     tracer_thread_argument.mutex.Unlock();
 421     // NOTE: errno is shared between this thread and the tracer thread.
 422     // internal_waitpid() may call syscall() which can access/spoil errno,
 423     // so we can't call it now. Instead we for the tracer thread to finish using
 424     // the spin loop below. Man page for sched_yield() says "In the Linux
 425     // implementation, sched_yield() always succeeds", so let's hope it does not
 426     // spoil errno. Note that this spin loop runs only for brief periods before
 427     // the tracer thread has suspended us and when it starts unblocking threads.
 428     while (atomic_load(&tracer_thread_argument.done, memory_order_relaxed) == 0)
 429       sched_yield();
 430     // Now the tracer thread is about to exit and does not touch errno,
 431     // wait for it.
 432     for (;;) {
 433       uptr waitpid_status = internal_waitpid(tracer_pid, nullptr, __WALL);
 434       if (!internal_iserror(waitpid_status, &local_errno))
 435         break;
 436       if (local_errno == EINTR)
 437         continue;
 438       VReport(1, "Waiting on the tracer thread failed (errno %d).\n",
 439               local_errno);
 440       break;
 441     }
 442   }
 443 }
 444
 445 // Platform-specific methods from SuspendedThreadsList.
 446 #if SANITIZER_ANDROID && defined(__arm__)
 447 typedef pt_regs regs_struct;
 448 #define REG_SP ARM_sp
 449
 450 #elif SANITIZER_LINUX && defined(__arm__)
 451 typedef user_regs regs_struct;
 452 #define REG_SP uregs[13]
 453
 454 #elif defined(__i386__) || defined(__x86_64__)
 455 typedef user_regs_struct regs_struct;
 456 #if defined(__i386__)
 457 #define REG_SP esp
 458 #else
 459 #define REG_SP rsp
 460 #endif
 461
 462 #elif defined(__powerpc__) || defined(__powerpc64__)
 463 typedef pt_regs regs_struct;
 464 #define REG_SP gpr[PT_R1]
 465
 466 #elif defined(__mips__)
 467 typedef struct user regs_struct;
 468 #define REG_SP regs[EF_REG29]
 469
 470 #elif defined(__aarch64__)
 471 typedef struct user_pt_regs regs_struct;
 472 #define REG_SP sp
 473 #define ARCH_IOVEC_FOR_GETREGSET
 474
 475 #else
 476 #error "Unsupported architecture"
 477 #endif // SANITIZER_ANDROID && defined(__arm__)
 478
 479 int SuspendedThreadsList::GetRegistersAndSP(uptr index,
 480                                             uptr *buffer,
 481                                             uptr *sp) const {
 482   pid_t tid = GetThreadID(index);
 483   regs_struct regs;
 484   int pterrno;
 485 #ifdef ARCH_IOVEC_FOR_GETREGSET
 486   struct iovec regset_io;
 487   regset_io.iov_base = &regs;
 488   regset_io.iov_len = sizeof(regs_struct);
 489   bool isErr = internal_iserror(internal_ptrace(PTRACE_GETREGSET, tid,
 490                                 (void*)NT_PRSTATUS, (void*)&regset_io),
 491                                 &pterrno);
 492 #else
 493   bool isErr = internal_iserror(internal_ptrace(PTRACE_GETREGS, tid, nullptr,
 494                                 &regs), &pterrno);
 495 #endif
 496   if (isErr) {
 497     VReport(1, "Could not get registers from thread %d (errno %d).\n", tid,
 498             pterrno);
 499     return -1;
 500   }
 501
 502   *sp = regs.REG_SP;
 503   internal_memcpy(buffer, &regs, sizeof(regs));
 504   return 0;
 505 }
 506
 507 uptr SuspendedThreadsList::RegisterCount() {
 508   return sizeof(regs_struct) / sizeof(uptr);
 509 }
 510 } // namespace __sanitizer
 511
 512 #endif // SANITIZER_LINUX && (defined(__x86_64__) || defined(__mips__)
 513        // || defined(__aarch64__)