thread_local_storage.cc 10.5 KB
Newer Older
gejun's avatar
gejun committed
1 2 3 4
// Copyright 2014 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

5
#include "butil/threading/thread_local_storage.h"
gejun's avatar
gejun committed
6

7 8
#include "butil/atomicops.h"
#include "butil/logging.h"
gejun's avatar
gejun committed
9

10
using butil::internal::PlatformThreadLocalStorage;
gejun's avatar
gejun committed
11 12 13 14 15 16 17 18 19 20

namespace {
// In order to make TLS destructors work, we need to keep around a function
// pointer to the destructor for each slot. We keep this array of pointers in a
// global (static) array.
// We use the single OS-level TLS slot (giving us one pointer per thread) to
// hold a pointer to a per-thread array (table) of slots that we allocate to
// Chromium consumers.

// g_native_tls_key is the one native TLS that we use.  It stores our table.
21
butil::subtle::AtomicWord g_native_tls_key =
gejun's avatar
gejun committed
22 23 24 25 26 27 28 29 30
    PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES;

// g_last_used_tls_key is the high-water-mark of allocated thread local storage.
// Each allocation is an index into our g_tls_destructors[].  Each such index is
// assigned to the instance variable slot_ in a ThreadLocalStorage::Slot
// instance.  We reserve the value slot_ == 0 to indicate that the corresponding
// instance of ThreadLocalStorage::Slot has been freed (i.e., destructor called,
// etc.).  This reserved use of 0 is then stated as the initial value of
// g_last_used_tls_key, so that the first issued index will be 1.
31
butil::subtle::Atomic32 g_last_used_tls_key = 0;
gejun's avatar
gejun committed
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48

// The maximum number of 'slots' in our thread local storage stack.
const int kThreadLocalStorageSize = 256;

// The maximum number of times to try to clear slots by calling destructors.
// Use pthread naming convention for clarity.
const int kMaxDestructorIterations = kThreadLocalStorageSize;

// An array of destructor function pointers for the slots.  If a slot has a
// destructor, it will be stored in its corresponding entry in this array.
// The elements are volatile to ensure that when the compiler reads the value
// to potentially call the destructor, it does so once, and that value is tested
// for null-ness and then used. Yes, that would be a weird de-optimization,
// but I can imagine some register machines where it was just as easy to
// re-fetch an array element, and I want to be sure a call to free the key
// (i.e., null out the destructor entry) that happens on a separate thread can't
// hurt the racy calls to the destructors on another thread.
49
volatile butil::ThreadLocalStorage::TLSDestructorFunc
gejun's avatar
gejun committed
50 51 52 53 54 55 56 57 58 59
    g_tls_destructors[kThreadLocalStorageSize];

// This function is called to initialize our entire Chromium TLS system.
// It may be called very early, and we need to complete most all of the setup
// (initialization) before calling *any* memory allocator functions, which may
// recursively depend on this initialization.
// As a result, we use Atomics, and avoid anything (like a singleton) that might
// require memory allocations.
void** ConstructTlsVector() {
  PlatformThreadLocalStorage::TLSKey key =
60
      butil::subtle::NoBarrier_Load(&g_native_tls_key);
gejun's avatar
gejun committed
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
  if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) {
    CHECK(PlatformThreadLocalStorage::AllocTLS(&key));

    // The TLS_KEY_OUT_OF_INDEXES is used to find out whether the key is set or
    // not in NoBarrier_CompareAndSwap, but Posix doesn't have invalid key, we
    // define an almost impossible value be it.
    // If we really get TLS_KEY_OUT_OF_INDEXES as value of key, just alloc
    // another TLS slot.
    if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) {
      PlatformThreadLocalStorage::TLSKey tmp = key;
      CHECK(PlatformThreadLocalStorage::AllocTLS(&key) &&
            key != PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES);
      PlatformThreadLocalStorage::FreeTLS(tmp);
    }
    // Atomically test-and-set the tls_key.  If the key is
    // TLS_KEY_OUT_OF_INDEXES, go ahead and set it.  Otherwise, do nothing, as
    // another thread already did our dirty work.
    if (PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES !=
79
        butil::subtle::NoBarrier_CompareAndSwap(&g_native_tls_key,
gejun's avatar
gejun committed
80 81 82 83 84
            PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES, key)) {
      // We've been shortcut. Another thread replaced g_native_tls_key first so
      // we need to destroy our index and use the one the other thread got
      // first.
      PlatformThreadLocalStorage::FreeTLS(key);
85
      key = butil::subtle::NoBarrier_Load(&g_native_tls_key);
gejun's avatar
gejun committed
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
    }
  }
  CHECK(!PlatformThreadLocalStorage::GetTLSValue(key));

  // Some allocators, such as TCMalloc, make use of thread local storage.
  // As a result, any attempt to call new (or malloc) will lazily cause such a
  // system to initialize, which will include registering for a TLS key.  If we
  // are not careful here, then that request to create a key will call new back,
  // and we'll have an infinite loop.  We avoid that as follows:
  // Use a stack allocated vector, so that we don't have dependence on our
  // allocator until our service is in place.  (i.e., don't even call new until
  // after we're setup)
  void* stack_allocated_tls_data[kThreadLocalStorageSize];
  memset(stack_allocated_tls_data, 0, sizeof(stack_allocated_tls_data));
  // Ensure that any rentrant calls change the temp version.
  PlatformThreadLocalStorage::SetTLSValue(key, stack_allocated_tls_data);

  // Allocate an array to store our data.
  void** tls_data = new void*[kThreadLocalStorageSize];
  memcpy(tls_data, stack_allocated_tls_data, sizeof(stack_allocated_tls_data));
  PlatformThreadLocalStorage::SetTLSValue(key, tls_data);
  return tls_data;
}

void OnThreadExitInternal(void* value) {
  DCHECK(value);
  void** tls_data = static_cast<void**>(value);
  // Some allocators, such as TCMalloc, use TLS.  As a result, when a thread
  // terminates, one of the destructor calls we make may be to shut down an
  // allocator.  We have to be careful that after we've shutdown all of the
  // known destructors (perchance including an allocator), that we don't call
  // the allocator and cause it to resurrect itself (with no possibly destructor
  // call to follow).  We handle this problem as follows:
  // Switch to using a stack allocated vector, so that we don't have dependence
  // on our allocator after we have called all g_tls_destructors.  (i.e., don't
  // even call delete[] after we're done with destructors.)
  void* stack_allocated_tls_data[kThreadLocalStorageSize];
  memcpy(stack_allocated_tls_data, tls_data, sizeof(stack_allocated_tls_data));
  // Ensure that any re-entrant calls change the temp version.
  PlatformThreadLocalStorage::TLSKey key =
126
      butil::subtle::NoBarrier_Load(&g_native_tls_key);
gejun's avatar
gejun committed
127 128 129 130 131 132 133 134 135 136 137 138 139
  PlatformThreadLocalStorage::SetTLSValue(key, stack_allocated_tls_data);
  delete[] tls_data;  // Our last dependence on an allocator.

  int remaining_attempts = kMaxDestructorIterations;
  bool need_to_scan_destructors = true;
  while (need_to_scan_destructors) {
    need_to_scan_destructors = false;
    // Try to destroy the first-created-slot (which is slot 1) in our last
    // destructor call.  That user was able to function, and define a slot with
    // no other services running, so perhaps it is a basic service (like an
    // allocator) and should also be destroyed last.  If we get the order wrong,
    // then we'll itterate several more times, so it is really not that
    // critical (but it might help).
140 141
    butil::subtle::Atomic32 last_used_tls_key =
        butil::subtle::NoBarrier_Load(&g_last_used_tls_key);
gejun's avatar
gejun committed
142 143 144 145 146
    for (int slot = last_used_tls_key; slot > 0; --slot) {
      void* value = stack_allocated_tls_data[slot];
      if (value == NULL)
        continue;

147
      butil::ThreadLocalStorage::TLSDestructorFunc destructor =
gejun's avatar
gejun committed
148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169
          g_tls_destructors[slot];
      if (destructor == NULL)
        continue;
      stack_allocated_tls_data[slot] = NULL;  // pre-clear the slot.
      destructor(value);
      // Any destructor might have called a different service, which then set
      // a different slot to a non-NULL value.  Hence we need to check
      // the whole vector again.  This is a pthread standard.
      need_to_scan_destructors = true;
    }
    if (--remaining_attempts <= 0) {
      NOTREACHED();  // Destructors might not have been called.
      break;
    }
  }

  // Remove our stack allocated vector.
  PlatformThreadLocalStorage::SetTLSValue(key, NULL);
}

}  // namespace

170
namespace butil {
gejun's avatar
gejun committed
171 172 173 174 175 176

namespace internal {

#if defined(OS_WIN)
void PlatformThreadLocalStorage::OnThreadExit() {
  PlatformThreadLocalStorage::TLSKey key =
177
      butil::subtle::NoBarrier_Load(&g_native_tls_key);
gejun's avatar
gejun committed
178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
  if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES)
    return;
  void *tls_data = GetTLSValue(key);
  // Maybe we have never initialized TLS for this thread.
  if (!tls_data)
    return;
  OnThreadExitInternal(tls_data);
}
#elif defined(OS_POSIX)
void PlatformThreadLocalStorage::OnThreadExit(void* value) {
  OnThreadExitInternal(value);
}
#endif  // defined(OS_WIN)

}  // namespace internal

ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) {
  initialized_ = false;
  slot_ = 0;
  Initialize(destructor);
}

bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) {
  PlatformThreadLocalStorage::TLSKey key =
202
      butil::subtle::NoBarrier_Load(&g_native_tls_key);
gejun's avatar
gejun committed
203 204 205 206 207
  if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES ||
      !PlatformThreadLocalStorage::GetTLSValue(key))
    ConstructTlsVector();

  // Grab a new slot.
208
  slot_ = butil::subtle::NoBarrier_AtomicIncrement(&g_last_used_tls_key, 1);
gejun's avatar
gejun committed
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230
  DCHECK_GT(slot_, 0);
  CHECK_LT(slot_, kThreadLocalStorageSize);

  // Setup our destructor.
  g_tls_destructors[slot_] = destructor;
  initialized_ = true;
  return true;
}

void ThreadLocalStorage::StaticSlot::Free() {
  // At this time, we don't reclaim old indices for TLS slots.
  // So all we need to do is wipe the destructor.
  DCHECK_GT(slot_, 0);
  DCHECK_LT(slot_, kThreadLocalStorageSize);
  g_tls_destructors[slot_] = NULL;
  slot_ = 0;
  initialized_ = false;
}

void* ThreadLocalStorage::StaticSlot::Get() const {
  void** tls_data = static_cast<void**>(
      PlatformThreadLocalStorage::GetTLSValue(
231
          butil::subtle::NoBarrier_Load(&g_native_tls_key)));
gejun's avatar
gejun committed
232 233 234 235 236 237 238 239 240 241
  if (!tls_data)
    tls_data = ConstructTlsVector();
  DCHECK_GT(slot_, 0);
  DCHECK_LT(slot_, kThreadLocalStorageSize);
  return tls_data[slot_];
}

void ThreadLocalStorage::StaticSlot::Set(void* value) {
  void** tls_data = static_cast<void**>(
      PlatformThreadLocalStorage::GetTLSValue(
242
          butil::subtle::NoBarrier_Load(&g_native_tls_key)));
gejun's avatar
gejun committed
243 244 245 246 247 248 249
  if (!tls_data)
    tls_data = ConstructTlsVector();
  DCHECK_GT(slot_, 0);
  DCHECK_LT(slot_, kThreadLocalStorageSize);
  tls_data[slot_] = value;
}

250
}  // namespace butil