1 /*
   2  * Copyright (c) 2001, 2019, 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_GC_SHARED_PTRQUEUE_HPP
  26 #define SHARE_GC_SHARED_PTRQUEUE_HPP
  27 
  28 #include "memory/padded.hpp"
  29 #include "utilities/align.hpp"
  30 #include "utilities/debug.hpp"
  31 #include "utilities/lockFreeStack.hpp"
  32 #include "utilities/sizes.hpp"
  33 
  34 class Mutex;
  35 
  36 // There are various techniques that require threads to be able to log
  37 // addresses.  For example, a generational write barrier might log
  38 // the addresses of modified old-generation objects.  This type supports
  39 // this operation.
  40 
  41 class BufferNode;
  42 class PtrQueueSet;
  43 class PtrQueue {
  44   friend class VMStructs;
  45 
  46   // Noncopyable - not defined.
  47   PtrQueue(const PtrQueue&);
  48   PtrQueue& operator=(const PtrQueue&);
  49 
  50   // The ptr queue set to which this queue belongs.
  51   PtrQueueSet* const _qset;
  52 
  53   // Whether updates should be logged.
  54   bool _active;
  55 
  56   // If true, the queue is permanent, and doesn't need to deallocate
  57   // its buffer in the destructor (since that obtains a lock which may not
  58   // be legally locked by then.
  59   const bool _permanent;
  60 
  61   // The (byte) index at which an object was last enqueued.  Starts at
  62   // capacity_in_bytes (indicating an empty buffer) and goes towards zero.
  63   // Value is always pointer-size aligned.
  64   size_t _index;
  65 
  66   // Size of the current buffer, in bytes.
  67   // Value is always pointer-size aligned.
  68   size_t _capacity_in_bytes;
  69 
  70   static const size_t _element_size = sizeof(void*);
  71 
  72   // Get the capacity, in bytes.  The capacity must have been set.
  73   size_t capacity_in_bytes() const {
  74     assert(_capacity_in_bytes > 0, "capacity not set");
  75     return _capacity_in_bytes;
  76   }
  77 
  78   void set_capacity(size_t entries) {
  79     size_t byte_capacity = index_to_byte_index(entries);
  80     assert(_capacity_in_bytes == 0 || _capacity_in_bytes == byte_capacity,
  81            "changing capacity " SIZE_FORMAT " -> " SIZE_FORMAT,
  82            _capacity_in_bytes, byte_capacity);
  83     _capacity_in_bytes = byte_capacity;
  84   }
  85 
  86   static size_t byte_index_to_index(size_t ind) {
  87     assert(is_aligned(ind, _element_size), "precondition");
  88     return ind / _element_size;
  89   }
  90 
  91   static size_t index_to_byte_index(size_t ind) {
  92     return ind * _element_size;
  93   }
  94 
  95 protected:
  96   // The buffer.
  97   void** _buf;
  98 
  99   size_t index() const {
 100     return byte_index_to_index(_index);
 101   }
 102 
 103   void set_index(size_t new_index) {
 104     size_t byte_index = index_to_byte_index(new_index);
 105     assert(byte_index <= capacity_in_bytes(), "precondition");
 106     _index = byte_index;
 107   }
 108 
 109   size_t capacity() const {
 110     return byte_index_to_index(capacity_in_bytes());
 111   }
 112 
 113   // If there is a lock associated with this buffer, this is that lock.
 114   Mutex* _lock;
 115 
 116   PtrQueueSet* qset() { return _qset; }
 117   bool is_permanent() const { return _permanent; }
 118 
 119   // Process queue entries and release resources.
 120   void flush_impl();
 121 
 122   // Initialize this queue to contain a null buffer, and be part of the
 123   // given PtrQueueSet.
 124   PtrQueue(PtrQueueSet* qset, bool permanent = false, bool active = false);
 125 
 126   // Requires queue flushed or permanent.
 127   ~PtrQueue();
 128 
 129 public:
 130 
 131   // Associate a lock with a ptr queue.
 132   void set_lock(Mutex* lock) { _lock = lock; }
 133 
 134   // Forcibly set empty.
 135   void reset() {
 136     if (_buf != NULL) {
 137       _index = capacity_in_bytes();
 138     }
 139   }
 140 
 141   void enqueue(volatile void* ptr) {
 142     enqueue((void*)(ptr));
 143   }
 144 
 145   // Enqueues the given "obj".
 146   void enqueue(void* ptr) {
 147     if (!_active) return;
 148     else enqueue_known_active(ptr);
 149   }
 150 
 151   // This method is called when we're doing the zero index handling
 152   // and gives a chance to the queues to do any pre-enqueueing
 153   // processing they might want to do on the buffer. It should return
 154   // true if the buffer should be enqueued, or false if enough
 155   // entries were cleared from it so that it can be re-used. It should
 156   // not return false if the buffer is still full (otherwise we can
 157   // get into an infinite loop).
 158   virtual bool should_enqueue_buffer() { return true; }
 159   void handle_zero_index();
 160 
 161   void enqueue_known_active(void* ptr);
 162 
 163   // Return the size of the in-use region.
 164   size_t size() const {
 165     size_t result = 0;
 166     if (_buf != NULL) {
 167       assert(_index <= capacity_in_bytes(), "Invariant");
 168       result = byte_index_to_index(capacity_in_bytes() - _index);
 169     }
 170     return result;
 171   }
 172 
 173   bool is_empty() const {
 174     return _buf == NULL || capacity_in_bytes() == _index;
 175   }
 176 
 177   // Set the "active" property of the queue to "b".  An enqueue to an
 178   // inactive thread is a no-op.  Setting a queue to inactive resets its
 179   // log to the empty state.
 180   void set_active(bool b) {
 181     _active = b;
 182     if (!b && _buf != NULL) {
 183       reset();
 184     } else if (b && _buf != NULL) {
 185       assert(index() == capacity(),
 186              "invariant: queues are empty when activated.");
 187     }
 188   }
 189 
 190   bool is_active() const { return _active; }
 191 
 192   // To support compiler.
 193 
 194 protected:
 195   template<typename Derived>
 196   static ByteSize byte_offset_of_index() {
 197     return byte_offset_of(Derived, _index);
 198   }
 199 
 200   static ByteSize byte_width_of_index() { return in_ByteSize(sizeof(size_t)); }
 201 
 202   template<typename Derived>
 203   static ByteSize byte_offset_of_buf() {
 204     return byte_offset_of(Derived, _buf);
 205   }
 206 
 207   static ByteSize byte_width_of_buf() { return in_ByteSize(_element_size); }
 208 
 209   template<typename Derived>
 210   static ByteSize byte_offset_of_active() {
 211     return byte_offset_of(Derived, _active);
 212   }
 213 
 214   static ByteSize byte_width_of_active() { return in_ByteSize(sizeof(bool)); }
 215 
 216 };
 217 
 218 class BufferNode {
 219   size_t _index;
 220   BufferNode* volatile _next;
 221   void* _buffer[1];             // Pseudo flexible array member.
 222 
 223   BufferNode() : _index(0), _next(NULL) { }
 224   ~BufferNode() { }
 225 
 226   static size_t buffer_offset() {
 227     return offset_of(BufferNode, _buffer);
 228   }
 229 
 230   static BufferNode* volatile* next_ptr(BufferNode& bn) { return &bn._next; }
 231 
 232 AIX_ONLY(public:)               // xlC 12 on AIX doesn't implement C++ DR45.
 233   // Allocate a new BufferNode with the "buffer" having size elements.
 234   static BufferNode* allocate(size_t size);
 235 
 236   // Free a BufferNode.
 237   static void deallocate(BufferNode* node);
 238 
 239 public:
 240   typedef LockFreeStack<BufferNode, &next_ptr> Stack;
 241 
 242   BufferNode* next() const     { return _next;  }
 243   void set_next(BufferNode* n) { _next = n;     }
 244   size_t index() const         { return _index; }
 245   void set_index(size_t i)     { _index = i; }
 246 
 247   // Return the BufferNode containing the buffer, after setting its index.
 248   static BufferNode* make_node_from_buffer(void** buffer, size_t index) {
 249     BufferNode* node =
 250       reinterpret_cast<BufferNode*>(
 251         reinterpret_cast<char*>(buffer) - buffer_offset());
 252     node->set_index(index);
 253     return node;
 254   }
 255 
 256   // Return the buffer for node.
 257   static void** make_buffer_from_node(BufferNode *node) {
 258     // &_buffer[0] might lead to index out of bounds warnings.
 259     return reinterpret_cast<void**>(
 260       reinterpret_cast<char*>(node) + buffer_offset());
 261   }
 262 
 263   class Allocator;              // Free-list based allocator.
 264   class TestSupport;            // Unit test support.
 265 };
 266 
 267 // Allocation is based on a lock-free free list of nodes, linked through
 268 // BufferNode::_next (see BufferNode::Stack).  To solve the ABA problem,
 269 // popping a node from the free list is performed within a GlobalCounter
 270 // critical section, and pushing nodes onto the free list is done after
 271 // a GlobalCounter synchronization associated with the nodes to be pushed.
 272 // This is documented behavior so that other parts of the node life-cycle
 273 // can depend on and make use of it too.
 274 class BufferNode::Allocator {
 275   friend class TestSupport;
 276 
 277   // Since we don't expect many instances, and measured >15% speedup
 278   // on stress gtest, padding seems like a good tradeoff here.
 279 #define DECLARE_PADDED_MEMBER(Id, Type, Name) \
 280   Type Name; DEFINE_PAD_MINUS_SIZE(Id, DEFAULT_CACHE_LINE_SIZE, sizeof(Type))
 281 
 282   const size_t _buffer_size;
 283   char _name[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)]; // Use name as padding.
 284   DECLARE_PADDED_MEMBER(1, Stack, _pending_list);
 285   DECLARE_PADDED_MEMBER(2, Stack, _free_list);
 286   DECLARE_PADDED_MEMBER(3, volatile size_t, _pending_count);
 287   DECLARE_PADDED_MEMBER(4, volatile size_t, _free_count);
 288   DECLARE_PADDED_MEMBER(5, volatile bool, _transfer_lock);
 289 
 290 #undef DECLARE_PADDED_MEMBER
 291 
 292   void delete_list(BufferNode* list);
 293   bool try_transfer_pending();
 294 
 295 public:
 296   Allocator(const char* name, size_t buffer_size);
 297   ~Allocator();
 298 
 299   const char* name() const { return _name; }
 300   size_t buffer_size() const { return _buffer_size; }
 301   size_t free_count() const;
 302   BufferNode* allocate();
 303   void release(BufferNode* node);
 304 
 305   // Deallocate some of the available buffers.  remove_goal is the target
 306   // number to remove.  Returns the number actually deallocated, which may
 307   // be less than the goal if there were fewer available.
 308   size_t reduce_free_list(size_t remove_goal);
 309 };
 310 
 311 // A PtrQueueSet represents resources common to a set of pointer queues.
 312 // In particular, the individual queues allocate buffers from this shared
 313 // set, and return completed buffers to the set.
 314 class PtrQueueSet {
 315   BufferNode::Allocator* _allocator;
 316 
 317   Monitor* _cbl_mon;  // Protects the fields below.
 318   BufferNode* _completed_buffers_head;
 319   BufferNode* _completed_buffers_tail;
 320   size_t _n_completed_buffers;
 321 
 322   size_t _process_completed_buffers_threshold;
 323   volatile bool _process_completed_buffers;
 324 
 325   // If true, notify_all on _cbl_mon when the threshold is reached.
 326   bool _notify_when_complete;
 327 
 328   // Maximum number of elements allowed on completed queue: after that,
 329   // enqueuer does the work itself.
 330   size_t _max_completed_buffers;
 331   size_t _completed_buffers_padding;
 332 
 333   void assert_completed_buffers_list_len_correct_locked() NOT_DEBUG_RETURN;
 334 
 335 protected:
 336   bool _all_active;
 337 
 338   // A mutator thread does the the work of processing a buffer.
 339   // Returns "true" iff the work is complete (and the buffer may be
 340   // deallocated).
 341   virtual bool mut_process_buffer(BufferNode* node) {
 342     ShouldNotReachHere();
 343     return false;
 344   }
 345 
 346   // Create an empty ptr queue set.
 347   PtrQueueSet(bool notify_when_complete = false);
 348   ~PtrQueueSet();
 349 
 350   // Because of init-order concerns, we can't pass these as constructor
 351   // arguments.
 352   void initialize(Monitor* cbl_mon, BufferNode::Allocator* allocator);
 353 
 354   // For (unlocked!) iteration over the completed buffers.
 355   BufferNode* completed_buffers_head() const { return _completed_buffers_head; }
 356 
 357   // Deallocate all of the completed buffers.
 358   void abandon_completed_buffers();
 359 
 360 public:
 361 
 362   // Return the buffer for a BufferNode of size buffer_size().
 363   void** allocate_buffer();
 364 
 365   // Return an empty buffer to the free list.  The node is required
 366   // to have been allocated with a size of buffer_size().
 367   void deallocate_buffer(BufferNode* node);
 368 
 369   // A completed buffer is a buffer the mutator is finished with, and
 370   // is ready to be processed by the collector.  It need not be full.
 371 
 372   // Adds node to the completed buffer list.
 373   void enqueue_completed_buffer(BufferNode* node);
 374 
 375   // If the number of completed buffers is > stop_at, then remove and
 376   // return a completed buffer from the list.  Otherwise, return NULL.
 377   BufferNode* get_completed_buffer(size_t stop_at = 0);
 378 
 379   // To be invoked by the mutator.
 380   bool process_or_enqueue_completed_buffer(BufferNode* node);
 381 
 382   bool process_completed_buffers() { return _process_completed_buffers; }
 383   void set_process_completed_buffers(bool x) { _process_completed_buffers = x; }
 384 
 385   bool is_active() { return _all_active; }
 386 
 387   size_t buffer_size() const {
 388     return _allocator->buffer_size();
 389   }
 390 
 391   // Get/Set the number of completed buffers that triggers log processing.
 392   // Log processing should be done when the number of buffers exceeds the
 393   // threshold.
 394   void set_process_completed_buffers_threshold(size_t sz) {
 395     _process_completed_buffers_threshold = sz;
 396   }
 397   size_t process_completed_buffers_threshold() const {
 398     return _process_completed_buffers_threshold;
 399   }
 400   static const size_t ProcessCompletedBuffersThresholdNever = ~size_t(0);
 401 
 402   size_t completed_buffers_num() const { return _n_completed_buffers; }
 403 
 404   void merge_bufferlists(PtrQueueSet* src);
 405 
 406   void set_max_completed_buffers(size_t m) {
 407     _max_completed_buffers = m;
 408   }
 409   size_t max_completed_buffers() const {
 410     return _max_completed_buffers;
 411   }
 412   static const size_t MaxCompletedBuffersUnlimited = ~size_t(0);
 413 
 414   void set_completed_buffers_padding(size_t padding) {
 415     _completed_buffers_padding = padding;
 416   }
 417   size_t completed_buffers_padding() const {
 418     return _completed_buffers_padding;
 419   }
 420 
 421   // Notify the consumer if the number of buffers crossed the threshold
 422   void notify_if_necessary();
 423 };
 424 
 425 #endif // SHARE_GC_SHARED_PTRQUEUE_HPP