1 /*
  2  * Copyright (c) 2021, 2025, 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_UTILITIES_NONBLOCKINGQUEUE_INLINE_HPP
 26 #define SHARE_UTILITIES_NONBLOCKINGQUEUE_INLINE_HPP
 27 
 28 #include "utilities/nonblockingQueue.hpp"
 29 
 30 #include "runtime/atomicAccess.hpp"
 31 
 32 template<typename T, T* volatile* (*next_ptr)(T&)>
 33 T* NonblockingQueue<T, next_ptr>::next(const T& node) {
 34   return AtomicAccess::load(next_ptr(const_cast<T&>(node)));
 35 }
 36 
 37 template<typename T, T* volatile* (*next_ptr)(T&)>
 38 void NonblockingQueue<T, next_ptr>::set_next(T& node, T* new_next) {
 39   AtomicAccess::store(next_ptr(node), new_next);
 40 }
 41 
 42 template<typename T, T* volatile* (*next_ptr)(T&)>
 43 NonblockingQueue<T, next_ptr>::NonblockingQueue() : _head(nullptr), _tail(nullptr) {}
 44 
 45 #ifdef ASSERT
 46 template<typename T, T* volatile* (*next_ptr)(T&)>
 47 NonblockingQueue<T, next_ptr>::~NonblockingQueue() {
 48   assert(_head == nullptr, "precondition");
 49   assert(_tail == nullptr, "precondition");
 50 }
 51 #endif
 52 
 53 // The end_marker must be uniquely associated with the specific queue, in
 54 // case queue elements can make their way through multiple queues.  A
 55 // pointer to the queue itself (after casting) satisfies that requirement.
 56 template<typename T, T* volatile* (*next_ptr)(T&)>
 57 T* NonblockingQueue<T, next_ptr>::end_marker() const {
 58   return const_cast<T*>(reinterpret_cast<const T*>(this));
 59 }
 60 
 61 template<typename T, T* volatile* (*next_ptr)(T&)>
 62 T* NonblockingQueue<T, next_ptr>::first() const {
 63   T* head = AtomicAccess::load(&_head);
 64   return head == nullptr ? end_marker() : head;
 65 }
 66 
 67 template<typename T, T* volatile* (*next_ptr)(T&)>
 68 bool NonblockingQueue<T, next_ptr>::is_end(const T* entry) const {
 69   return entry == end_marker();
 70 }
 71 
 72 template<typename T, T* volatile* (*next_ptr)(T&)>
 73 bool NonblockingQueue<T, next_ptr>::empty() const {
 74   return AtomicAccess::load(&_head) == nullptr;
 75 }
 76 
 77 template<typename T, T* volatile* (*next_ptr)(T&)>
 78 size_t NonblockingQueue<T, next_ptr>::length() const {
 79   size_t result = 0;
 80   for (T* cur = first(); !is_end(cur); cur = next(*cur)) {
 81     ++result;
 82   }
 83   return result;
 84 }
 85 
 86 // An append operation atomically exchanges the new tail with the queue tail.
 87 // It then sets the "next" value of the old tail to the head of the list being
 88 // appended. If the old tail is null then the queue was empty, then the
 89 // head of the list being appended is instead stored in the queue head.
 90 //
 91 // This means there is a period between the exchange and the old tail update
 92 // where the queue sequence is split into two parts, the list from the queue
 93 // head to the old tail, and the list being appended.  If there are concurrent
 94 // push/append operations, each may introduce another such segment.  But they
 95 // all eventually get resolved by their respective updates of their old tail's
 96 // "next" value.  This also means that try_pop operation must handle an object
 97 // differently depending on its "next" value.
 98 //
 99 // A push operation is just a degenerate append, where the object being pushed
100 // is both the head and the tail of the list being appended.
101 template<typename T, T* volatile* (*next_ptr)(T&)>
102 void NonblockingQueue<T, next_ptr>::append(T& first, T& last) {
103   assert(next(last) == nullptr, "precondition");
104   // Make last the new end of the queue.  Any further push/appends will
105   // extend after last.  We will try to extend from the previous end of
106   // queue.
107   set_next(last, end_marker());
108   T* old_tail = AtomicAccess::xchg(&_tail, &last);
109   if (old_tail == nullptr) {
110     // If old_tail is null then the queue was empty, and _head must also be
111     // null. The correctness of this assertion depends on try_pop clearing
112     // first _head then _tail when taking the last entry.
113     assert(AtomicAccess::load(&_head) == nullptr, "invariant");
114     // Fall through to common update of _head.
115   } else if (is_end(AtomicAccess::cmpxchg(next_ptr(*old_tail), end_marker(), &first))) {
116     // Successfully extended the queue list from old_tail to first.  No
117     // other push/append could have competed with us, because we claimed
118     // old_tail for extension.  We won any races with try_pop by changing
119     // away from end-marker.  So we're done.
120     //
121     // Note that ABA is possible here.  A concurrent try_pop could take
122     // old_tail before our update of old_tail's next_ptr, old_tail gets
123     // recycled and re-added to the end of this queue, and then we
124     // successfully cmpxchg, making the list in _tail circular.  Callers
125     // must ensure this can't happen.
126     return;
127   } else {
128     // A concurrent try_pop has claimed old_tail, so it is no longer in the
129     // list. The queue was logically empty.  _head is either null or
130     // old_tail, depending on how far try_pop operations have progressed.
131     DEBUG_ONLY(T* old_head = AtomicAccess::load(&_head);)
132     assert((old_head == nullptr) || (old_head == old_tail), "invariant");
133     // Fall through to common update of _head.
134   }
135   // The queue was empty, and first should become the new _head.  The queue
136   // will appear to be empty to any further try_pops until done.
137   AtomicAccess::store(&_head, &first);
138 }
139 
140 template<typename T, T* volatile* (*next_ptr)(T&)>
141 bool NonblockingQueue<T, next_ptr>::try_pop(T** node_ptr) {
142   // We only need memory_order_consume. Upgrade it to "load_acquire"
143   // as the memory_order_consume API is not ready for use yet.
144   T* old_head = AtomicAccess::load_acquire(&_head);
145   if (old_head == nullptr) {
146     *node_ptr = nullptr;
147     return true;                // Queue is empty.
148   }
149 
150   T* next_node = AtomicAccess::load_acquire(next_ptr(*old_head));
151   if (!is_end(next_node)) {
152     // [Clause 1]
153     // There are several cases for next_node.
154     // (1) next_node is the extension of the queue's list.
155     // (2) next_node is null, because a competing try_pop took old_head.
156     // (3) next_node is the extension of some unrelated list, because a
157     // competing try_pop took old_head and put it in some other list.
158     //
159     // Attempt to advance the list, replacing old_head with next_node in
160     // _head.  The success or failure of that attempt, along with the value
161     // of next_node, are used to partially determine which case we're in and
162     // how to proceed.  In particular, advancement will fail for case (3).
163     if (old_head != AtomicAccess::cmpxchg(&_head, old_head, next_node)) {
164       // [Clause 1a]
165       // The cmpxchg to advance the list failed; a concurrent try_pop won
166       // the race and claimed old_head.  This can happen for any of the
167       // next_node cases.
168       return false;
169     } else if (next_node == nullptr) {
170       // [Clause 1b]
171       // The cmpxchg to advance the list succeeded, but a concurrent try_pop
172       // has already claimed old_head (see [Clause 2] - old_head was the last
173       // entry in the list) by nulling old_head's next field.  The advance set
174       // _head to null, "helping" the competing try_pop.  _head will remain
175       // nullptr until a subsequent push/append.  This is a lost race, and we
176       // report it as such for consistency, though we could report the queue
177       // was empty.  We don't attempt to further help [Clause 2] by also
178       // trying to set _tail to nullptr, as that would just ensure that one or
179       // the other cmpxchg is a wasted failure.
180       return false;
181     } else {
182       // [Clause 1c]
183       // Successfully advanced the list and claimed old_head.  next_node was
184       // in the extension of the queue's list.  Return old_head after
185       // unlinking it from next_node.
186       set_next(*old_head, nullptr);
187       *node_ptr = old_head;
188       return true;
189     }
190 
191   } else if (is_end(AtomicAccess::cmpxchg(next_ptr(*old_head), next_node, (T*)nullptr))) {
192     // [Clause 2]
193     // Old_head was the last entry and we've claimed it by setting its next
194     // value to null.  However, this leaves the queue in disarray.  Fix up
195     // the queue, possibly in conjunction with other concurrent operations.
196     // Any further try_pops will consider the queue empty until a
197     // push/append completes by installing a new head.
198 
199     // The order of the two cmpxchgs doesn't matter algorithmically, but
200     // dealing with _head first gives a stronger invariant in append, and is
201     // also consistent with [Clause 1b].
202 
203     // Attempt to change the queue head from old_head to null.  Failure of
204     // the cmpxchg indicates a concurrent operation updated _head first.  That
205     // could be either a push/append or a try_pop in [Clause 1b].
206     AtomicAccess::cmpxchg(&_head, old_head, (T*)nullptr);
207 
208     // Attempt to change the queue tail from old_head to null.  Failure of
209     // the cmpxchg indicates that a concurrent push/append updated _tail first.
210     // That operation will eventually recognize the old tail (our old_head) is
211     // no longer in the list and update _head from the list being appended.
212     AtomicAccess::cmpxchg(&_tail, old_head, (T*)nullptr);
213 
214     // The queue has been restored to order, and we can return old_head.
215     *node_ptr = old_head;
216     return true;
217 
218   } else {
219     // [Clause 3]
220     // Old_head was the last entry in the list, but either a concurrent
221     // try_pop claimed it first or a concurrent push/append extended the
222     // list from it.  Either way, we lost the race to claim it.
223     return false;
224   }
225 }
226 
227 template<typename T, T* volatile* (*next_ptr)(T&)>
228 T* NonblockingQueue<T, next_ptr>::pop() {
229   T* result = nullptr;
230   // Typically try_pop() will succeed without retrying many times, thus we
231   // omit SpinPause in the loop body.  SpinPause or yield may be worthwhile
232   // in rare, highly contended cases, and client code could implement such
233   // with try_pop().
234   while (!try_pop(&result)) {}
235   return result;
236 }
237 
238 template<typename T, T* volatile* (*next_ptr)(T&)>
239 Pair<T*, T*> NonblockingQueue<T, next_ptr>::take_all() {
240   T* tail = AtomicAccess::load(&_tail);
241   if (tail != nullptr) set_next(*tail, nullptr); // Clear end marker.
242   Pair<T*, T*> result(AtomicAccess::load(&_head), tail);
243   AtomicAccess::store(&_head, (T*)nullptr);
244   AtomicAccess::store(&_tail, (T*)nullptr);
245   return result;
246 }
247 
248 #endif // SHARE_UTILITIES_NONBLOCKINGQUEUE_INLINE_HPP