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