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/enableIf.hpp"
  30 #include "metaprogramming/integerTypes.hpp"
  31 #include "metaprogramming/isIntegral.hpp"
  32 #include "metaprogramming/isSame.hpp"
  33 #include "metaprogramming/removeCV.hpp"
  34 #include "utilities/align.hpp"
  35 #include "utilities/macros.hpp"
  36 
  37 enum cmpxchg_memory_order {
  38   memory_order_relaxed,
  39   // Use value which doesn't interfere with C++2011. We need to be more conservative.
  40   memory_order_conservative = 8
  41 };
  42 
  43 class Atomic : AllStatic {
  44  public:
  45   // Atomic operations on jlong types are not available on all 32-bit
  46   // platforms. If atomic ops on jlongs are defined here they must only
  47   // be used from code that verifies they are available at runtime and
  48   // can provide an alternative action if not - see supports_cx8() for
  49   // a means to test availability.
  50 
  51   // The memory operations that are mentioned with each of the atomic
  52   // function families come from src/share/vm/runtime/orderAccess.hpp,
  53   // e.g., <fence> is described in that file and is implemented by the
  54   // OrderAccess::fence() function. See that file for the gory details
  55   // on the Memory Access Ordering Model.
  56 
  57   // All of the atomic operations that imply a read-modify-write action
  58   // guarantee a two-way memory barrier across that operation. Historically
  59   // these semantics reflect the strength of atomic operations that are
  60   // provided on SPARC/X86. We assume that strength is necessary unless
  61   // we can prove that a weaker form is sufficiently safe.
  62 
  63   // Atomically store to a location
  64   inline static void store    (jbyte    store_value, jbyte*    dest);
  65   inline static void store    (jshort   store_value, jshort*   dest);
  66   inline static void store    (jint     store_value, jint*     dest);
  67   // See comment above about using jlong atomics on 32-bit platforms
  68   inline static void store    (jlong    store_value, jlong*    dest);
  69   inline static void store_ptr(intptr_t store_value, intptr_t* dest);
  70   inline static void store_ptr(void*    store_value, void*     dest);
  71 
  72   inline static void store    (jbyte    store_value, volatile jbyte*    dest);
  73   inline static void store    (jshort   store_value, volatile jshort*   dest);
  74   inline static void store    (jint     store_value, volatile jint*     dest);
  75   // See comment above about using jlong atomics on 32-bit platforms
  76   inline static void store    (jlong    store_value, volatile jlong*    dest);
  77   inline static void store_ptr(intptr_t store_value, volatile intptr_t* dest);
  78   inline static void store_ptr(void*    store_value, volatile void*     dest);
  79 
  80   // See comment above about using jlong atomics on 32-bit platforms
  81   inline static jlong load(const volatile jlong* src);
  82 
  83   // Atomically add to a location. Returns updated value. add*() provide:
  84   // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
  85   inline static jshort   add    (jshort   add_value, volatile jshort*   dest);
  86   inline static jint     add    (jint     add_value, volatile jint*     dest);
  87   inline static size_t   add    (size_t   add_value, volatile size_t*   dest);
  88   inline static intptr_t add_ptr(intptr_t add_value, volatile intptr_t* dest);
  89   inline static void*    add_ptr(intptr_t add_value, volatile void*     dest);
  90 
  91   // Atomically increment location. inc*() provide:
  92   // <fence> increment-dest <membar StoreLoad|StoreStore>
  93   inline static void inc    (volatile jint*     dest);
  94   inline static void inc    (volatile jshort*   dest);
  95   inline static void inc    (volatile size_t*   dest);
  96   inline static void inc_ptr(volatile intptr_t* dest);
  97   inline static void inc_ptr(volatile void*     dest);
  98 
  99   // Atomically decrement a location. dec*() provide:
 100   // <fence> decrement-dest <membar StoreLoad|StoreStore>
 101   inline static void dec    (volatile jint*     dest);
 102   inline static void dec    (volatile jshort*   dest);
 103   inline static void dec    (volatile size_t*   dest);
 104   inline static void dec_ptr(volatile intptr_t* dest);
 105   inline static void dec_ptr(volatile void*     dest);
 106 
 107   // Performs atomic exchange of *dest with exchange_value. Returns old
 108   // prior value of *dest. xchg*() provide:
 109   // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
 110   inline static jint         xchg    (jint         exchange_value, volatile jint*         dest);
 111   inline static unsigned int xchg    (unsigned int exchange_value, volatile unsigned int* dest);
 112   inline static intptr_t     xchg_ptr(intptr_t     exchange_value, volatile intptr_t*     dest);
 113   inline static void*        xchg_ptr(void*        exchange_value, volatile void*         dest);
 114 
 115   // Performs atomic compare of *dest and compare_value, and exchanges
 116   // *dest with exchange_value if the comparison succeeded. Returns prior
 117   // value of *dest. cmpxchg*() provide:
 118   // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
 119 
 120   template<typename T, typename D, typename U>
 121   inline static D cmpxchg(T exchange_value,
 122                           D volatile* dest,
 123                           U compare_value,
 124                           cmpxchg_memory_order order = memory_order_conservative);
 125 
 126   // Performs atomic compare of *dest and NULL, and replaces *dest
 127   // with exchange_value if the comparison succeeded.  Returns true if
 128   // the comparison succeeded and the exchange occurred.  This is
 129   // often used as part of lazy initialization, as a lock-free
 130   // alternative to the Double-Checked Locking Pattern.
 131   template<typename T, typename D>
 132   inline static bool replace_if_null(T* value, D* volatile* dest,
 133                                      cmpxchg_memory_order order = memory_order_conservative);
 134 
 135   inline static intptr_t cmpxchg_ptr(intptr_t exchange_value,
 136                                      volatile intptr_t* dest,
 137                                      intptr_t compare_value,
 138                                      cmpxchg_memory_order order = memory_order_conservative) {
 139     return cmpxchg(exchange_value, dest, compare_value, order);
 140   }
 141 
 142   inline static void* cmpxchg_ptr(void* exchange_value,
 143                                   volatile void* dest,
 144                                   void* compare_value,
 145                                   cmpxchg_memory_order order = memory_order_conservative) {
 146     return cmpxchg(exchange_value,
 147                    reinterpret_cast<void* volatile*>(dest),
 148                    compare_value,
 149                    order);
 150   }
 151 
 152 private:
 153   // Test whether From is implicitly convertible to To.
 154   // From and To must be pointer types.
 155   // Note: Provides the limited subset of C++11 std::is_convertible
 156   // that is needed here.
 157   template<typename From, typename To> struct IsPointerConvertible;
 158 
 159   // Dispatch handler for cmpxchg.  Provides type-based validity
 160   // checking and limited conversions around calls to the
 161   // platform-specific implementation layer provided by
 162   // PlatformCmpxchg.
 163   template<typename T, typename D, typename U, typename Enable = void>
 164   struct CmpxchgImpl;
 165 
 166   // Platform-specific implementation of cmpxchg.  Support for sizes
 167   // of 1, 4, and 8 are required.  The class is a function object that
 168   // must be default constructable, with these requirements:
 169   //
 170   // - dest is of type T*.
 171   // - exchange_value and compare_value are of type T.
 172   // - order is of type cmpxchg_memory_order.
 173   // - platform_cmpxchg is an object of type PlatformCmpxchg<sizeof(T)>.
 174   //
 175   // Then
 176   //   platform_cmpxchg(exchange_value, dest, compare_value, order)
 177   // must be a valid expression, returning a result convertible to T.
 178   //
 179   // A default definition is provided, which declares a function template
 180   //   T operator()(T, T volatile*, T, cmpxchg_memory_order) const
 181   //
 182   // For each required size, a platform must either provide an
 183   // appropriate definition of that function, or must entirely
 184   // specialize the class template for that size.
 185   template<size_t byte_size> struct PlatformCmpxchg;
 186 
 187   // Support for platforms that implement some variants of cmpxchg
 188   // using a (typically out of line) non-template helper function.
 189   // The generic arguments passed to PlatformCmpxchg need to be
 190   // translated to the appropriate type for the helper function, the
 191   // helper invoked on the translated arguments, and the result
 192   // translated back.  Type is the parameter / return type of the
 193   // helper function.
 194   template<typename Type, typename Fn, typename T>
 195   static T cmpxchg_using_helper(Fn fn,
 196                                 T exchange_value,
 197                                 T volatile* dest,
 198                                 T compare_value);
 199 
 200   // Support platforms that do not provide Read-Modify-Write
 201   // byte-level atomic access. To use, derive PlatformCmpxchg<1> from
 202   // this class.
 203 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 204   struct CmpxchgByteUsingInt;
 205 private:
 206 };
 207 
 208 template<typename From, typename To>
 209 struct Atomic::IsPointerConvertible<From*, To*> : AllStatic {
 210   // Determine whether From* is implicitly convertible to To*, using
 211   // the "sizeof trick".
 212   typedef char yes;
 213   typedef char (&no)[2];
 214 
 215   static yes test(To*);
 216   static no test(...);
 217   static From* test_value;
 218 
 219   static const bool value = (sizeof(yes) == sizeof(test(test_value)));
 220 };
 221 
 222 // Define the class before including platform file, which may specialize
 223 // the operator definition.  No generic definition of specializations
 224 // of the operator template are provided, nor are there any generic
 225 // specializations of the class.  The platform file is responsible for
 226 // providing those.
 227 template<size_t byte_size>
 228 struct Atomic::PlatformCmpxchg VALUE_OBJ_CLASS_SPEC {
 229   template<typename T>
 230   T operator()(T exchange_value,
 231                T volatile* dest,
 232                T compare_value,
 233                cmpxchg_memory_order order) const;
 234 };
 235 
 236 // Define the class before including platform file, which may use this
 237 // as a base class, requiring it be complete.  The definition is later
 238 // in this file, near the other definitions related to cmpxchg.
 239 struct Atomic::CmpxchgByteUsingInt VALUE_OBJ_CLASS_SPEC {
 240   template<typename T>
 241   T operator()(T exchange_value,
 242                T volatile* dest,
 243                T compare_value,
 244                cmpxchg_memory_order order) const;
 245 };
 246 
 247 // platform specific in-line definitions - must come before shared definitions
 248 
 249 #include OS_CPU_HEADER(atomic)
 250 
 251 // shared in-line definitions
 252 
 253 // size_t casts...
 254 #if (SIZE_MAX != UINTPTR_MAX)
 255 #error size_t is not WORD_SIZE, interesting platform, but missing implementation here
 256 #endif
 257 
 258 inline size_t Atomic::add(size_t add_value, volatile size_t* dest) {
 259   return (size_t) add_ptr((intptr_t) add_value, (volatile intptr_t*) dest);
 260 }
 261 
 262 inline void Atomic::inc(volatile size_t* dest) {
 263   inc_ptr((volatile intptr_t*) dest);
 264 }
 265 
 266 inline void Atomic::dec(volatile size_t* dest) {
 267   dec_ptr((volatile intptr_t*) dest);
 268 }
 269 
 270 template<typename T, typename D, typename U>
 271 inline D Atomic::cmpxchg(T exchange_value,
 272                          D volatile* dest,
 273                          U compare_value,
 274                          cmpxchg_memory_order order) {
 275   return CmpxchgImpl<T, D, U>()(exchange_value, dest, compare_value, order);
 276 }
 277 
 278 template<typename T, typename D>
 279 inline bool Atomic::replace_if_null(T* value, D* volatile* dest,
 280                                     cmpxchg_memory_order order) {
 281   // Presently using a trivial implementation in terms of cmpxchg.
 282   // Consider adding platform support, to permit the use of compiler
 283   // intrinsics like gcc's __sync_bool_compare_and_swap.
 284   D* expected_null = NULL;
 285   return expected_null == cmpxchg(value, dest, expected_null, order);
 286 }
 287 
 288 // Handle cmpxchg for integral and enum types.
 289 //
 290 // All the involved types must be identical.
 291 template<typename T>
 292 struct Atomic::CmpxchgImpl<
 293   T, T, T,
 294   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 295   VALUE_OBJ_CLASS_SPEC
 296 {
 297   T operator()(T exchange_value, T volatile* dest, T compare_value,
 298                cmpxchg_memory_order order) const {
 299     // Forward to the platform handler for the size of T.
 300     return PlatformCmpxchg<sizeof(T)>()(exchange_value,
 301                                         dest,
 302                                         compare_value,
 303                                         order);
 304   }
 305 };
 306 
 307 // Handle cmpxchg for pointer types.
 308 //
 309 // The destination's type and the compare_value type must be the same,
 310 // ignoring cv-qualifiers; we don't care about the cv-qualifiers of
 311 // the compare_value.
 312 //
 313 // The exchange_value must be implicitly convertible to the
 314 // destination's type; it must be type-correct to store the
 315 // exchange_value in the destination.
 316 template<typename T, typename D, typename U>
 317 struct Atomic::CmpxchgImpl<
 318   T*, D*, U*,
 319   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value &&
 320                     IsSame<typename RemoveCV<D>::type,
 321                            typename RemoveCV<U>::type>::value>::type>
 322   VALUE_OBJ_CLASS_SPEC
 323 {
 324   D* operator()(T* exchange_value, D* volatile* dest, U* compare_value,
 325                cmpxchg_memory_order order) const {
 326     // Allow derived to base conversion, and adding cv-qualifiers.
 327     D* new_value = exchange_value;
 328     // Don't care what the CV qualifiers for compare_value are,
 329     // but we need to match D* when calling platform support.
 330     D* old_value = const_cast<D*>(compare_value);
 331     return PlatformCmpxchg<sizeof(D*)>()(new_value, dest, old_value, order);
 332   }
 333 };
 334 
 335 // Handle cmpxchg for types that have a translator.
 336 //
 337 // All the involved types must be identical.
 338 //
 339 // This translates the original call into a call on the decayed
 340 // arguments, and returns the recovered result of that translated
 341 // call.
 342 template<typename T>
 343 struct Atomic::CmpxchgImpl<
 344   T, T, T,
 345   typename EnableIf<IntegerTypes::Translate<T>::value>::type>
 346   VALUE_OBJ_CLASS_SPEC
 347 {
 348   T operator()(T exchange_value, T volatile* dest, T compare_value,
 349                cmpxchg_memory_order order) const {
 350     typedef IntegerTypes::Translate<T> Translator;
 351     typedef typename Translator::Decayed Decayed;
 352     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 353     return Translator::recover(
 354       cmpxchg(Translator::decay(exchange_value),
 355               reinterpret_cast<Decayed volatile*>(dest),
 356               Translator::decay(compare_value),
 357               order));
 358   }
 359 };
 360 
 361 template<typename Type, typename Fn, typename T>
 362 inline T Atomic::cmpxchg_using_helper(Fn fn,
 363                                       T exchange_value,
 364                                       T volatile* dest,
 365                                       T compare_value) {
 366   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 367   return IntegerTypes::cast<T>(
 368     fn(IntegerTypes::cast<Type>(exchange_value),
 369        reinterpret_cast<Type volatile*>(dest),
 370        IntegerTypes::cast<Type>(compare_value)));
 371 }
 372 
 373 template<typename T>
 374 inline T Atomic::CmpxchgByteUsingInt::operator()(T exchange_value,
 375                                                  T volatile* dest,
 376                                                  T compare_value,
 377                                                  cmpxchg_memory_order order) const {
 378   STATIC_ASSERT(sizeof(T) == sizeof(uint8_t));
 379   uint8_t canon_exchange_value = exchange_value;
 380   uint8_t canon_compare_value = compare_value;
 381   volatile uint32_t* aligned_dest
 382     = reinterpret_cast<volatile uint32_t*>(align_down(dest, sizeof(uint32_t)));
 383   size_t offset = pointer_delta(dest, aligned_dest, 1);
 384   uint32_t cur = *aligned_dest;
 385   uint8_t* cur_as_bytes = reinterpret_cast<uint8_t*>(&cur);
 386 
 387   // current value may not be what we are looking for, so force it
 388   // to that value so the initial cmpxchg will fail if it is different
 389   cur_as_bytes[offset] = canon_compare_value;
 390 
 391   // always execute a real cmpxchg so that we get the required memory
 392   // barriers even on initial failure
 393   do {
 394     // value to swap in matches current value ...
 395     uint32_t new_value = cur;
 396     // ... except for the one jbyte we want to update
 397     reinterpret_cast<uint8_t*>(&new_value)[offset] = canon_exchange_value;
 398 
 399     uint32_t res = cmpxchg(new_value, aligned_dest, cur, order);
 400     if (res == cur) break;      // success
 401 
 402     // at least one byte in the int changed value, so update
 403     // our view of the current int
 404     cur = res;
 405     // if our byte is still as cur we loop and try again
 406   } while (cur_as_bytes[offset] == canon_compare_value);
 407 
 408   return IntegerTypes::cast<T>(cur_as_bytes[offset]);
 409 }
 410 
 411 inline unsigned Atomic::xchg(unsigned int exchange_value, volatile unsigned int* dest) {
 412   assert(sizeof(unsigned int) == sizeof(jint), "more work to do");
 413   return (unsigned int)Atomic::xchg((jint)exchange_value, (volatile jint*)dest);
 414 }
 415 
 416 inline jshort Atomic::add(jshort add_value, volatile jshort* dest) {
 417   // Most platforms do not support atomic add on a 2-byte value. However,
 418   // if the value occupies the most significant 16 bits of an aligned 32-bit
 419   // word, then we can do this with an atomic add of (add_value << 16)
 420   // to the 32-bit word.
 421   //
 422   // The least significant parts of this 32-bit word will never be affected, even
 423   // in case of overflow/underflow.
 424   //
 425   // Use the ATOMIC_SHORT_PAIR macro (see macros.hpp) to get the desired alignment.
 426 #ifdef VM_LITTLE_ENDIAN
 427   assert((intx(dest) & 0x03) == 0x02, "wrong alignment");
 428   jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest-1));
 429 #else
 430   assert((intx(dest) & 0x03) == 0x00, "wrong alignment");
 431   jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest));
 432 #endif
 433   return (jshort)(new_value >> 16); // preserves sign
 434 }
 435 
 436 inline void Atomic::inc(volatile jshort* dest) {
 437   (void)add(1, dest);
 438 }
 439 
 440 inline void Atomic::dec(volatile jshort* dest) {
 441   (void)add(-1, dest);
 442 }
 443 
 444 #endif // SHARE_VM_RUNTIME_ATOMIC_HPP