1 /*
   2  * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_RUNTIME_ATOMIC_HPP
  26 #define SHARE_VM_RUNTIME_ATOMIC_HPP
  27 
  28 #include "memory/allocation.hpp"
  29 #include "metaprogramming/conditional.hpp"
  30 #include "metaprogramming/integerTypes.hpp"
  31 #include "metaprogramming/isDerived.hpp"
  32 #include "metaprogramming/isIntegral.hpp"
  33 #include "metaprogramming/isPointer.hpp"
  34 #include "utilities/align.hpp"
  35 #include "utilities/debug.hpp"
  36 #include "utilities/macros.hpp"
  37 
  38 enum cmpxchg_memory_order {
  39   memory_order_relaxed,
  40   // Use value which doesn't interfere with C++2011. We need to be more conservative.
  41   memory_order_conservative = 8
  42 };
  43 
  44 class GeneralizedAtomic : AllStatic {
  45   template<typename T> class Never: public FalseType {};
  46 
  47   template <typename T>
  48   inline static void specialized_store(T store_value, volatile T* dest) {
  49     STATIC_ASSERT(sizeof(T) <= size_t(BytesPerWord)); // Does the machine support atomic wide accesses?
  50     (void)const_cast<T&>(*dest = store_value);
  51   }
  52 
  53   template <typename T>
  54   inline static T specialized_load(const volatile T* dest) {
  55     STATIC_ASSERT(sizeof(T) <= size_t(BytesPerWord)); // Does the machine support atomic wide accesses?
  56     return *dest;
  57   }
  58 
  59   template <typename T>
  60   inline static T specialized_add(T add_value, volatile T* dest) {
  61     STATIC_ASSERT(Never<T>::value);
  62     return add_value;
  63   }
  64 
  65   template <typename T>
  66   inline static void specialized_inc(volatile T* dest) {
  67     add(1, dest);
  68   }
  69 
  70   template <typename T>
  71   inline static void specialized_dec(volatile T* dest) {
  72     add(-1, dest);
  73   }
  74 
  75   template <typename T>
  76   inline static T specialized_xchg(T exchange_value, volatile T* dest) {
  77     STATIC_ASSERT(Never<T>::value);
  78     return exchange_value;
  79   }
  80 
  81   template <typename T>
  82   inline static T specialized_cmpxchg(T exchange_value, volatile T* dest, T compare_value, cmpxchg_memory_order order) {
  83     STATIC_ASSERT(Never<T>::value);
  84     return exchange_value;
  85   }
  86 
  87 public:
  88   template <typename T, typename U>
  89   inline static void store(T store_value, volatile U* dest);
  90 
  91   template <typename T>
  92   inline static T load(volatile T* src);
  93 
  94   template <typename T, typename U>
  95   inline static U add(T add_value, volatile U* dst);
  96 
  97   template <typename T, typename U>
  98   inline static U* add(T add_value, U* volatile* dst);
  99 
 100   template <typename T>
 101   inline static void inc(volatile T* dest);
 102 
 103   template <typename T>
 104   inline static void inc(T* volatile* dest);
 105 
 106   template <typename T>
 107   inline static void dec(volatile T* dest);
 108 
 109   template <typename T>
 110   inline static void dec(T* volatile* dest);
 111 
 112   template <typename T, typename U>
 113   inline static U xchg(T exchange_value, volatile U* dest);
 114 
 115   template <typename T, typename U, typename V>
 116   inline static U cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order);
 117 };
 118 
 119 
 120 // platform specific in-line definitions - must come before shared definitions
 121 
 122 class PlatformAtomic;
 123 
 124 #include OS_CPU_HEADER(atomic)
 125 
 126 typedef Conditional<IsDerived<PlatformAtomic, AllStatic>::value, PlatformAtomic, GeneralizedAtomic>::type AtomicImpl;
 127 
 128 class Atomic : AllStatic {
 129  public:
 130   // Atomic operations on 64-bit types are not available on all 32-bit
 131   // platforms. If atomic ops on 64-bit types are defined here they must only
 132   // be used from code that verifies they are available at runtime and
 133   // can provide an alternative action if not - see supports_cx8() for
 134   // a means to test availability.
 135 
 136   // The memory operations that are mentioned with each of the atomic
 137   // function families come from src/share/vm/runtime/orderAccess.hpp,
 138   // e.g., <fence> is described in that file and is implemented by the
 139   // OrderAccess::fence() function. See that file for the gory details
 140   // on the Memory Access Ordering Model.
 141 
 142   // All of the atomic operations that imply a read-modify-write action
 143   // guarantee a two-way memory barrier across that operation. Historically
 144   // these semantics reflect the strength of atomic operations that are
 145   // provided on SPARC/X86. We assume that strength is necessary unless
 146   // we can prove that a weaker form is sufficiently safe.
 147 
 148   // Atomically store to a location
 149   // See comment above about using 64-bit atomics on 32-bit platforms
 150   template <typename T, typename U>
 151   inline static void store(T store_value, volatile U* dest);
 152 
 153   // The store_ptr() member functions are deprecated. Use store() instead.
 154   static void store_ptr(intptr_t store_value, volatile intptr_t* dest) {
 155     store(store_value, dest);
 156   }
 157 
 158   static void store_ptr(void*    store_value, volatile void*     dest) {
 159     store((intptr_t)store_value, (volatile intptr_t*)dest);
 160   }
 161 
 162   // Atomically load from a location
 163   // See comment above about using 64-bit atomics on 32-bit platforms
 164   template <typename T>
 165   inline static T load(volatile T* src);
 166 
 167   // Atomically add to a location. Returns updated value. add*() provide:
 168   // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
 169   // add(I1 v, I* d)
 170   // add(I1 v, P* d)
 171   // where I, I1 are integral types, P is a pointer type.
 172   // Functional behavior is modelled on *dest += add_value.
 173   template <typename T, typename U>
 174   inline static U add(T add_value, volatile U* dst);
 175 
 176   // The add_ptr() member functions are deprecated. Use add() instead.
 177   static intptr_t add_ptr(intptr_t add_value, volatile intptr_t* dest) {
 178     return add(add_value, dest);
 179   }
 180 
 181   static void*    add_ptr(intptr_t add_value, volatile void*     dest) {
 182     return (void*)add(add_value, (volatile intptr_t*)dest);
 183   }
 184 
 185   // Atomically increment location. inc*() provide:
 186   // <fence> increment-dest <membar StoreLoad|StoreStore>
 187   // Functional behavior is modelled on *dest++
 188   template <typename T>
 189   inline static void inc(volatile T* dest);
 190 
 191   // The inc_ptr member functions are deprecated. Use inc() instead.
 192   static void inc_ptr(volatile intptr_t* dest) {
 193     inc(dest);
 194   }
 195 
 196   static void inc_ptr(volatile void*     dest) {
 197     inc((volatile intptr_t*)dest);
 198   }
 199 
 200   // Atomically decrement a location. dec*() provide:
 201   // <fence> decrement-dest <membar StoreLoad|StoreStore>
 202   // Functional behavior is modelled on *dest--
 203   template <typename T>
 204   inline static void dec(volatile T* dest);
 205 
 206   // The dec_ptr member functions are deprecated. Use dec() instead.
 207   static void dec_ptr(volatile intptr_t* dest) {
 208     dec(dest);
 209   }
 210 
 211   static void dec_ptr(volatile void*     dest) {
 212     dec((volatile intptr_t*)dest);
 213   }
 214 
 215   // Performs atomic exchange of *dest with exchange_value. Returns old
 216   // prior value of *dest. xchg*() provide:
 217   // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
 218   template <typename T, typename U>
 219   inline static U xchg(T exchange_value, volatile U* dest);
 220 
 221   // The xchg_ptr() member functions are deprecated. Use xchg() instead.
 222   static intptr_t xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) {
 223     return xchg(exchange_value, dest);
 224   }
 225 
 226   static void*    xchg_ptr(void*    exchange_value, volatile void*     dest) {
 227     return (void*)xchg((intptr_t)exchange_value, (volatile intptr_t*)dest);
 228   }
 229 
 230   // Performs atomic compare of *dest and compare_value, and exchanges
 231   // *dest with exchange_value if the comparison succeeded. Returns prior
 232   // value of *dest. cmpxchg*() provide:
 233   // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
 234   // See comment above about using 64-bit atomics on 32-bit platforms
 235   template <typename T, typename U, typename V>
 236   inline static U cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order = memory_order_conservative);
 237 
 238   // The cmpxchg_ptr member functions are deprecated. Use cmpxchg() instead.
 239   inline static intptr_t cmpxchg_ptr(intptr_t exchange_value, volatile intptr_t*  dest,
 240                                      intptr_t compare_value, cmpxchg_memory_order order = memory_order_conservative) {
 241     return cmpxchg(exchange_value, dest, compare_value, order);
 242   }
 243 
 244   inline static void*    cmpxchg_ptr(void*    exchange_value, volatile void*      dest,
 245                                      void*    compare_value, cmpxchg_memory_order order = memory_order_conservative) {
 246     return (void*)cmpxchg((intptr_t)exchange_value, (volatile intptr_t*)dest, (intptr_t)compare_value, order);
 247   }
 248 };
 249 
 250 // internal implementation
 251 
 252 template <typename T, typename U>
 253 inline void GeneralizedAtomic::store(T store_value, volatile U* dest) {
 254   typedef typename IntegerTypes::Signed<U>::type Raw;
 255   U store_value_cast = store_value;
 256   specialized_store(IntegerTypes::cast_to_signed(store_value_cast), reinterpret_cast<volatile Raw*>(dest));
 257 }
 258 
 259 template <typename T>
 260 inline T GeneralizedAtomic::load(volatile T* src) {
 261   typedef typename IntegerTypes::Signed<T>::type Raw;
 262   return IntegerTypes::cast<T>(specialized_load(reinterpret_cast<const volatile Raw*>(src)));
 263 }
 264 
 265 template <typename T, typename U>
 266 inline U GeneralizedAtomic::add(T add_value, volatile U* dst) {
 267   STATIC_ASSERT(IsIntegral<T>::value);
 268   STATIC_ASSERT(IsIntegral<U>::value);
 269   typedef typename IntegerTypes::Signed<U>::type Raw;
 270   // Allow -Wconversion or the like to complain about unsafe conversions.
 271   U value = add_value;
 272   Raw raw_value = IntegerTypes::cast_to_signed(value);
 273   Raw result = specialized_add(raw_value, reinterpret_cast<volatile Raw*>(dst));
 274   return IntegerTypes::cast<U>(result);
 275 }
 276 
 277 template <typename T, typename U>
 278 inline U* GeneralizedAtomic::add(T add_value, U* volatile* dst) {
 279   STATIC_ASSERT(IsIntegral<T>::value);
 280   typedef typename IntegerTypes::Signed<U*>::type Raw;
 281   ptrdiff_t value = add_value;
 282   Raw raw_value = IntegerTypes::cast_to_signed(value * sizeof(U));
 283   Raw result = specialized_add(raw_value, reinterpret_cast<volatile Raw*>(dst));
 284   return IntegerTypes::cast<U*>(result);
 285 }
 286 
 287 template <typename T>
 288 inline void GeneralizedAtomic::inc(volatile T* src) {
 289   STATIC_ASSERT(IsIntegral<T>::value);
 290   typedef typename IntegerTypes::Signed<T>::type Raw;
 291   specialized_inc(reinterpret_cast<volatile Raw*>(src));
 292 }
 293 
 294 template <typename T>
 295 inline void GeneralizedAtomic::inc(T* volatile* src) {
 296   if (sizeof(T) != 1) {
 297     add(1, src);
 298   } else {
 299     typedef typename IntegerTypes::Signed<T*>::type Raw;
 300     specialized_inc(reinterpret_cast<volatile Raw*>(src));
 301   }
 302 }
 303 
 304 template <typename T>
 305 inline void GeneralizedAtomic::dec(volatile T* src) {
 306   STATIC_ASSERT(IsIntegral<T>::value);
 307   typedef typename IntegerTypes::Signed<T>::type Raw;
 308   specialized_dec(reinterpret_cast<volatile Raw*>(src));
 309 }
 310 
 311 template <typename T>
 312 inline void GeneralizedAtomic::dec(T* volatile* src) {
 313   if (sizeof(T) != 1) {
 314     add(-1, src);
 315   } else {
 316     typedef typename IntegerTypes::Signed<T*>::type Raw;
 317     specialized_dec(reinterpret_cast<volatile Raw*>(src));
 318   }
 319 }
 320 
 321 template <typename T, typename U>
 322 inline U GeneralizedAtomic::xchg(T exchange_value, volatile U* dest) {
 323   typedef typename IntegerTypes::Signed<U>::type Raw;
 324   U exchange_value_cast = exchange_value;
 325   Raw result = specialized_xchg(IntegerTypes::cast_to_signed(exchange_value_cast),
 326                                 reinterpret_cast<volatile Raw*>(dest));
 327   return IntegerTypes::cast<U>(result);
 328 }
 329 
 330 template <typename T, typename U, typename V>
 331 inline U GeneralizedAtomic::cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order) {
 332   typedef typename IntegerTypes::Signed<U>::type Raw;
 333   U exchange_value_cast = exchange_value;
 334   U compare_value_cast = compare_value;
 335   Raw result = specialized_cmpxchg(IntegerTypes::cast_to_signed(exchange_value_cast),
 336                                    reinterpret_cast<volatile Raw*>(dest),
 337                                    IntegerTypes::cast_to_signed(compare_value_cast), order);
 338   return IntegerTypes::cast<U>(result);
 339 }
 340 
 341 template <typename T, typename U>
 342 inline void Atomic::store(T store_value, volatile U* dest) {
 343   AtomicImpl::store(store_value, dest);
 344 }
 345 
 346 template <typename T>
 347 inline T Atomic::load(volatile T* src) {
 348   return AtomicImpl::load(src);
 349 }
 350 
 351 template <typename T, typename U>
 352 inline U Atomic::add(T add_value, volatile U* dst) {
 353   return AtomicImpl::add(add_value, dst);
 354 }
 355 
 356 template <typename T>
 357 inline void Atomic::inc(volatile T* src) {
 358   AtomicImpl::inc(src);
 359 }
 360 
 361 template <typename T>
 362 inline void Atomic::dec(volatile T* src) {
 363   AtomicImpl::dec(src);
 364 }
 365 
 366 template <typename T, typename U>
 367 inline U Atomic::xchg(T exchange_value, volatile U* dest) {
 368   return AtomicImpl::xchg(exchange_value, dest);
 369 }
 370 
 371 template <typename T, typename U, typename V>
 372 inline U Atomic::cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order) {
 373   return AtomicImpl::cmpxchg(exchange_value, dest, compare_value, order);
 374 }
 375 
 376 // shared in-line definitions
 377 
 378 #ifndef VM_HAS_SPECIALIZED_CMPXCHG_BYTE
 379 /*
 380  * This is the default implementation of byte-sized cmpxchg. It emulates 8-bit-sized cmpxchg
 381  * in terms of 32-bit-sized cmpxchg. Platforms may override this by defining their own inline definition
 382  * as well as defining VM_HAS_SPECIALIZED_CMPXCHG_BYTE. This will cause the platform specific
 383  * implementation to be used instead.
 384  */
 385 template <>
 386 inline int8_t GeneralizedAtomic::specialized_cmpxchg<int8_t>(int8_t exchange_value, volatile int8_t* dest,
 387                                                              int8_t compare_value, cmpxchg_memory_order order) {
 388   volatile int32_t* dest_int =
 389       reinterpret_cast<volatile int32_t*>(align_down(dest, sizeof(int32_t)));
 390   size_t offset = pointer_delta(dest, dest_int, 1);
 391   int32_t cur = *dest_int;
 392   int8_t* cur_as_bytes = reinterpret_cast<int8_t*>(&cur);
 393 
 394   // current value may not be what we are looking for, so force it
 395   // to that value so the initial cmpxchg will fail if it is different
 396   cur_as_bytes[offset] = compare_value;
 397 
 398   // always execute a real cmpxchg so that we get the required memory
 399   // barriers even on initial failure
 400   do {
 401     // value to swap in matches current value ...
 402     int32_t new_value = cur;
 403     // ... except for the one byte we want to update
 404     reinterpret_cast<int8_t*>(&new_value)[offset] = exchange_value;
 405 
 406     int32_t res = cmpxchg(new_value, dest_int, cur, order);
 407     if (res == cur) break; // success
 408 
 409     // at least one byte in the int changed value, so update
 410     // our view of the current int
 411     cur = res;
 412     // if our byte is still as cur we loop and try again
 413   } while (cur_as_bytes[offset] == compare_value);
 414 
 415   return cur_as_bytes[offset];
 416 }
 417 
 418 #endif // VM_HAS_SPECIALIZED_CMPXCHG_BYTE
 419 
 420 template <>
 421 inline int16_t GeneralizedAtomic::specialized_add<int16_t>(int16_t add_value, volatile int16_t* dest) {
 422   // Most platforms do not support atomic add on a 2-byte value. However,
 423   // if the value occupies the most significant 16 bits of an aligned 32-bit
 424   // word, then we can do this with an atomic add of (add_value << 16)
 425   // to the 32-bit word.
 426   //
 427   // The least significant parts of this 32-bit word will never be affected, even
 428   // in case of overflow/underflow.
 429   //
 430   // Use the ATOMIC_SHORT_PAIR macro (see macros.hpp) to get the desired alignment.
 431 #ifdef VM_LITTLE_ENDIAN
 432   assert((intx(dest) & 0x03) == 0x02, "wrong alignment");
 433   int32_t new_value = Atomic::add(int32_t(add_value) << 16, (volatile int32_t*)(dest-1));
 434 #else
 435   assert((intx(dest) & 0x03) == 0x00, "wrong alignment");
 436   int32_t new_value = Atomic::add(int32_t(add_value) << 16, (volatile int32_t*)(dest));
 437 #endif
 438   return (int16_t)(new_value >> 16); // preserves sign
 439 }
 440 
 441 template <>
 442 inline void GeneralizedAtomic::specialized_inc<int16_t>(volatile int16_t* dest) {
 443   (void)add(int16_t(1), dest);
 444 }
 445 
 446 template <>
 447 inline void GeneralizedAtomic::specialized_dec<int16_t>(volatile int16_t* dest) {
 448   (void)add(int16_t(-1), dest);
 449 }
 450 
 451 #endif // SHARE_VM_RUNTIME_ATOMIC_HPP