tsan: speed up pthread_setname_np

[dvyukov]

Summary

pthread_setname_np does linear search over all thread descriptors
to map pthread_t to the thread descriptor. This has O(N^2) complexity
and becomes much worse in the new tsan runtime that keeps all ever
existed threads in the thread registry.
Replace linear search with direct access if pthread_setname_np
is called for the current thread (a very common case).

[vitalybuka]

compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp

2137

This part looks unrelated

[dvyukov]

removed semi-related part of the diff

[dvyukov]

PTAL

[DavidSpickett]

The new test is failing on our Arch64 bots:
https://lab.llvm.org/buildbot/#/builders/179/builds/1761

That one is a stage 2 failure but same thing happens on other 1 stage bots and in local builds.

Actual output:

starting transient threads...
starting persistent threads...
ThreadSanitizer: CHECK failed: tsan_rtl.cpp:327 "((addr + size)) <= ((TraceMemEnd()))" (0xf06200e03010, 0xf06200000000) (tid=2492946)
    #0 __tsan::CheckUnwind() /home/david.spickett/llvm-project/compiler-rt/lib/tsan/rtl/tsan_rtl.cpp:357 (bench_threads.cpp.tmp+0x4a1b84)
    #1 __sanitizer::CheckFailed(char const*, int, char const*, unsigned long long, unsigned long long) /home/david.spickett/llvm-project/compiler-rt/lib/sanitizer_common/sanitizer_termination.cpp:86 (bench_threads.cpp.tmp+0x430cc4)
    #2 __tsan::MapThreadTrace(unsigned long, unsigned long, char const*) /home/david.spickett/llvm-project/compiler-rt/lib/tsan/rtl/tsan_rtl.cpp:327 (bench_threads.cpp.tmp+0x4a1b00)
    #3 __tsan::CreateThreadContext(unsigned int) /home/david.spickett/llvm-project/compiler-rt/lib/tsan/rtl/tsan_rtl.cpp:82 (bench_threads.cpp.tmp+0x4a10c4)
    #4 __sanitizer::ThreadRegistry::CreateThread(unsigned long, bool, unsigned int, void*) /home/david.spickett/llvm-project/compiler-rt/lib/sanitizer_common/sanitizer_thread_registry.cpp:144 (bench_threads.cpp.tmp+0x42ea9c)
    #5 __tsan::ThreadCreate(__tsan::ThreadState*, unsigned long, unsigned long, bool) /home/david.spickett/llvm-project/compiler-rt/lib/tsan/rtl/tsan_rtl_thread.cpp:122 (bench_threads.cpp.tmp+0x4bb000)
    #6 pthread_create /home/david.spickett/llvm-project/compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp:1042 (bench_threads.cpp.tmp+0x4496bc)
    #7 bench() /home/david.spickett/llvm-project/compiler-rt/test/tsan/bench_threads.cpp:38 (bench_threads.cpp.tmp+0x4c307c)
    #8 main /home/david.spickett/llvm-project/compiler-rt/test/tsan/bench.h:23 (bench_threads.cpp.tmp+0x4c2e20)
    #9 __libc_start_main ??:? (libc.so.6+0x2408c)
    #10 _start ??:? (bench_threads.cpp.tmp+0x41ebc0)

[DavidSpickett]

And I see d0c138ec8aa47ec2acb38dba247f537ca470618f disables it, thanks!

compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp

 extern "C" int pthread_atfork(void (*prepare)(void), void (*parent)(void),
                               void (*child)(void));
 extern "C" int pthread_key_create(unsigned *key, void (*destructor)(void* v));
 extern "C" int pthread_setspecific(unsigned key, const void *v);
 DECLARE_REAL(int, pthread_mutexattr_gettype, void *, void *)
 DECLARE_REAL(int, fflush, __sanitizer_FILE *fp)
 DECLARE_REAL_AND_INTERCEPTOR(void *, malloc, uptr size)
 DECLARE_REAL_AND_INTERCEPTOR(void, free, void *ptr)
+extern "C" int pthread_equal(void *t1, void *t2);
 extern "C" void *pthread_self();
 extern "C" void _exit(int status);
 #if !SANITIZER_NETBSD
 extern "C" int fileno_unlocked(void *stream);
 extern "C" int dirfd(void *dirp);
 #endif
 #if SANITIZER_NETBSD
 extern __sanitizer_FILE __sF[];
TSAN_INTERCEPTOR(int, kill, int pid, int sig) {
   return res;
 }
 
 TSAN_INTERCEPTOR(int, pthread_kill, void *tid, int sig) {
   SCOPED_TSAN_INTERCEPTOR(pthread_kill, tid, sig);
   ThreadSignalContext *sctx = SigCtx(thr);
   CHECK_NE(sctx, 0);
   int prev = sctx->int_signal_send;
   if (tid == pthread_self()) {
     sctx->int_signal_send = sig;
   }
   int res = REAL(pthread_kill)(tid, sig);
   if (tid == pthread_self()) {
     CHECK_EQ(sctx->int_signal_send, sig);
     sctx->int_signal_send = prev;
   }
   return res;
#define COMMON_INTERCEPTOR_FD_ACCESS(ctx, fd) \
   FdAccess(((TsanInterceptorContext *) ctx)->thr, pc, fd)
 
 #define COMMON_INTERCEPTOR_FD_SOCKET_ACCEPT(ctx, fd, newfd) \
   FdSocketAccept(((TsanInterceptorContext *) ctx)->thr, pc, fd, newfd)
 
 #define COMMON_INTERCEPTOR_SET_THREAD_NAME(ctx, name) \
   ThreadSetName(((TsanInterceptorContext *) ctx)->thr, name)
 
-#define COMMON_INTERCEPTOR_SET_PTHREAD_NAME(ctx, thread, name) \
+#define COMMON_INTERCEPTOR_SET_PTHREAD_NAME(ctx, thread, name)         \
+  if (pthread_equal(pthread_self(), reinterpret_cast<void *>(thread))) \
+    COMMON_INTERCEPTOR_SET_THREAD_NAME(ctx, name);                     \
+  else                                                                 \
     __tsan::ctx->thread_registry.SetThreadNameByUserId(thread, name)
 
 #define COMMON_INTERCEPTOR_BLOCK_REAL(name) BLOCK_REAL(name)
 
 #define COMMON_INTERCEPTOR_ON_EXIT(ctx) \
   OnExit(((TsanInterceptorContext *) ctx)->thr)
 
 #define COMMON_INTERCEPTOR_MUTEX_PRE_LOCK(ctx, m) \
   MutexPreLock(((TsanInterceptorContext *)ctx)->thr, \

compiler-rt/test/tsan/bench_threads.cpp

+// RUN: %clangxx_tsan %s -o %t
+// RUN: %run %t 2>&1 | FileCheck %s
+
+// bench.h needs pthread barriers which are not available on OS X
+// UNSUPPORTED: darwin
+
+#include "bench.h"
+
+void *nop_thread(void *arg) {
+  pthread_setname_np(pthread_self(), "nop_thread");
+  return nullptr;
+}
+
+void thread(int tid) {
+  for (int i = 0; i < bench_niter; i++) {
+    pthread_t th;
+    pthread_create(&th, nullptr, nop_thread, nullptr);
+    pthread_join(th, nullptr);
+  }
+}
+
+void bench() {
+  // Benchmark thread creation/joining in presence of a large number
+  // of threads (both alive and already joined).
+  printf("starting transient threads...\n");
+  for (int i = 0; i < 200; i++) {
+    const int kBatch = 100;
+    pthread_t th[kBatch];
+    for (int j = 0; j < kBatch; j++)
+      pthread_create(&th[j], nullptr, nop_thread, nullptr);
+    for (int j = 0; j < kBatch; j++)
+      pthread_join(th[j], nullptr);
+  }
+  printf("starting persistent threads...\n");
+  const int kLiveThreads = 2000;
+  pthread_t th[kLiveThreads];
+  for (int j = 0; j < kLiveThreads; j++)
+    pthread_create(&th[j], nullptr, nop_thread, nullptr);
+  printf("starting benchmark threads...\n");
+  start_thread_group(bench_nthread, thread);
+  for (int j = 0; j < kLiveThreads; j++)
+    pthread_join(th[j], nullptr);
+}
+
+// CHECK: DONE