Mon, 12 Mar 2012 14:59:00 -0700
7147724: G1: hang in SurrogateLockerThread::manipulatePLL
Summary: Attempting to initiate a marking cycle when allocating a humongous object can, if a marking cycle is successfully initiated by another thread, result in the allocating thread spinning until the marking cycle is complete. Eliminate a deadlock between the main ConcurrentMarkThread, the SurrogateLocker thread, the VM thread, and a mutator thread waiting on the SecondaryFreeList_lock (while free regions are going to become available) by not manipulating the pending list lock during the prologue and epilogue of the cleanup pause.
Reviewed-by: brutisso, jcoomes, tonyp
1 /*
2 * Copyright (c) 2001, 2011, 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 */
25 #include "precompiled.hpp"
26 #include "gc_implementation/g1/ptrQueue.hpp"
27 #include "memory/allocation.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "runtime/mutex.hpp"
30 #include "runtime/mutexLocker.hpp"
31 #ifdef TARGET_OS_FAMILY_linux
32 # include "thread_linux.inline.hpp"
33 #endif
34 #ifdef TARGET_OS_FAMILY_solaris
35 # include "thread_solaris.inline.hpp"
36 #endif
37 #ifdef TARGET_OS_FAMILY_windows
38 # include "thread_windows.inline.hpp"
39 #endif
40 #ifdef TARGET_OS_FAMILY_bsd
41 # include "thread_bsd.inline.hpp"
42 #endif
44 PtrQueue::PtrQueue(PtrQueueSet* qset, bool perm, bool active) :
45 _qset(qset), _buf(NULL), _index(0), _active(active),
46 _perm(perm), _lock(NULL)
47 {}
49 void PtrQueue::flush() {
50 if (!_perm && _buf != NULL) {
51 if (_index == _sz) {
52 // No work to do.
53 qset()->deallocate_buffer(_buf);
54 } else {
55 // We must NULL out the unused entries, then enqueue.
56 for (size_t i = 0; i < _index; i += oopSize) {
57 _buf[byte_index_to_index((int)i)] = NULL;
58 }
59 qset()->enqueue_complete_buffer(_buf);
60 }
61 _buf = NULL;
62 _index = 0;
63 }
64 }
67 static int byte_index_to_index(int ind) {
68 assert((ind % oopSize) == 0, "Invariant.");
69 return ind / oopSize;
70 }
72 static int index_to_byte_index(int byte_ind) {
73 return byte_ind * oopSize;
74 }
76 void PtrQueue::enqueue_known_active(void* ptr) {
77 assert(0 <= _index && _index <= _sz, "Invariant.");
78 assert(_index == 0 || _buf != NULL, "invariant");
80 while (_index == 0) {
81 handle_zero_index();
82 }
84 assert(_index > 0, "postcondition");
85 _index -= oopSize;
86 _buf[byte_index_to_index((int)_index)] = ptr;
87 assert(0 <= _index && _index <= _sz, "Invariant.");
88 }
90 void PtrQueue::locking_enqueue_completed_buffer(void** buf) {
91 assert(_lock->owned_by_self(), "Required.");
93 // We have to unlock _lock (which may be Shared_DirtyCardQ_lock) before
94 // we acquire DirtyCardQ_CBL_mon inside enqeue_complete_buffer as they
95 // have the same rank and we may get the "possible deadlock" message
96 _lock->unlock();
98 qset()->enqueue_complete_buffer(buf);
99 // We must relock only because the caller will unlock, for the normal
100 // case.
101 _lock->lock_without_safepoint_check();
102 }
105 PtrQueueSet::PtrQueueSet(bool notify_when_complete) :
106 _max_completed_queue(0),
107 _cbl_mon(NULL), _fl_lock(NULL),
108 _notify_when_complete(notify_when_complete),
109 _sz(0),
110 _completed_buffers_head(NULL),
111 _completed_buffers_tail(NULL),
112 _n_completed_buffers(0),
113 _process_completed_threshold(0), _process_completed(false),
114 _buf_free_list(NULL), _buf_free_list_sz(0)
115 {
116 _fl_owner = this;
117 }
119 void** PtrQueueSet::allocate_buffer() {
120 assert(_sz > 0, "Didn't set a buffer size.");
121 MutexLockerEx x(_fl_owner->_fl_lock, Mutex::_no_safepoint_check_flag);
122 if (_fl_owner->_buf_free_list != NULL) {
123 void** res = BufferNode::make_buffer_from_node(_fl_owner->_buf_free_list);
124 _fl_owner->_buf_free_list = _fl_owner->_buf_free_list->next();
125 _fl_owner->_buf_free_list_sz--;
126 return res;
127 } else {
128 // Allocate space for the BufferNode in front of the buffer.
129 char *b = NEW_C_HEAP_ARRAY(char, _sz + BufferNode::aligned_size());
130 return BufferNode::make_buffer_from_block(b);
131 }
132 }
134 void PtrQueueSet::deallocate_buffer(void** buf) {
135 assert(_sz > 0, "Didn't set a buffer size.");
136 MutexLockerEx x(_fl_owner->_fl_lock, Mutex::_no_safepoint_check_flag);
137 BufferNode *node = BufferNode::make_node_from_buffer(buf);
138 node->set_next(_fl_owner->_buf_free_list);
139 _fl_owner->_buf_free_list = node;
140 _fl_owner->_buf_free_list_sz++;
141 }
143 void PtrQueueSet::reduce_free_list() {
144 assert(_fl_owner == this, "Free list reduction is allowed only for the owner");
145 // For now we'll adopt the strategy of deleting half.
146 MutexLockerEx x(_fl_lock, Mutex::_no_safepoint_check_flag);
147 size_t n = _buf_free_list_sz / 2;
148 while (n > 0) {
149 assert(_buf_free_list != NULL, "_buf_free_list_sz must be wrong.");
150 void* b = BufferNode::make_block_from_node(_buf_free_list);
151 _buf_free_list = _buf_free_list->next();
152 FREE_C_HEAP_ARRAY(char, b);
153 _buf_free_list_sz --;
154 n--;
155 }
156 }
158 void PtrQueue::handle_zero_index() {
159 assert(_index == 0, "Precondition.");
161 // This thread records the full buffer and allocates a new one (while
162 // holding the lock if there is one).
163 if (_buf != NULL) {
164 if (!should_enqueue_buffer()) {
165 assert(_index > 0, "the buffer can only be re-used if it's not full");
166 return;
167 }
169 if (_lock) {
170 assert(_lock->owned_by_self(), "Required.");
172 // The current PtrQ may be the shared dirty card queue and
173 // may be being manipulated by more than one worker thread
174 // during a pause. Since the enqueuing of the completed
175 // buffer unlocks the Shared_DirtyCardQ_lock more than one
176 // worker thread can 'race' on reading the shared queue attributes
177 // (_buf and _index) and multiple threads can call into this
178 // routine for the same buffer. This will cause the completed
179 // buffer to be added to the CBL multiple times.
181 // We "claim" the current buffer by caching value of _buf in
182 // a local and clearing the field while holding _lock. When
183 // _lock is released (while enqueueing the completed buffer)
184 // the thread that acquires _lock will skip this code,
185 // preventing the subsequent the multiple enqueue, and
186 // install a newly allocated buffer below.
188 void** buf = _buf; // local pointer to completed buffer
189 _buf = NULL; // clear shared _buf field
191 locking_enqueue_completed_buffer(buf); // enqueue completed buffer
193 // While the current thread was enqueuing the buffer another thread
194 // may have a allocated a new buffer and inserted it into this pointer
195 // queue. If that happens then we just return so that the current
196 // thread doesn't overwrite the buffer allocated by the other thread
197 // and potentially losing some dirtied cards.
199 if (_buf != NULL) return;
200 } else {
201 if (qset()->process_or_enqueue_complete_buffer(_buf)) {
202 // Recycle the buffer. No allocation.
203 _sz = qset()->buffer_size();
204 _index = _sz;
205 return;
206 }
207 }
208 }
209 // Reallocate the buffer
210 _buf = qset()->allocate_buffer();
211 _sz = qset()->buffer_size();
212 _index = _sz;
213 assert(0 <= _index && _index <= _sz, "Invariant.");
214 }
216 bool PtrQueueSet::process_or_enqueue_complete_buffer(void** buf) {
217 if (Thread::current()->is_Java_thread()) {
218 // We don't lock. It is fine to be epsilon-precise here.
219 if (_max_completed_queue == 0 || _max_completed_queue > 0 &&
220 _n_completed_buffers >= _max_completed_queue + _completed_queue_padding) {
221 bool b = mut_process_buffer(buf);
222 if (b) {
223 // True here means that the buffer hasn't been deallocated and the caller may reuse it.
224 return true;
225 }
226 }
227 }
228 // The buffer will be enqueued. The caller will have to get a new one.
229 enqueue_complete_buffer(buf);
230 return false;
231 }
233 void PtrQueueSet::enqueue_complete_buffer(void** buf, size_t index) {
234 MutexLockerEx x(_cbl_mon, Mutex::_no_safepoint_check_flag);
235 BufferNode* cbn = BufferNode::new_from_buffer(buf);
236 cbn->set_index(index);
237 if (_completed_buffers_tail == NULL) {
238 assert(_completed_buffers_head == NULL, "Well-formedness");
239 _completed_buffers_head = cbn;
240 _completed_buffers_tail = cbn;
241 } else {
242 _completed_buffers_tail->set_next(cbn);
243 _completed_buffers_tail = cbn;
244 }
245 _n_completed_buffers++;
247 if (!_process_completed && _process_completed_threshold >= 0 &&
248 _n_completed_buffers >= _process_completed_threshold) {
249 _process_completed = true;
250 if (_notify_when_complete)
251 _cbl_mon->notify();
252 }
253 debug_only(assert_completed_buffer_list_len_correct_locked());
254 }
256 int PtrQueueSet::completed_buffers_list_length() {
257 int n = 0;
258 BufferNode* cbn = _completed_buffers_head;
259 while (cbn != NULL) {
260 n++;
261 cbn = cbn->next();
262 }
263 return n;
264 }
266 void PtrQueueSet::assert_completed_buffer_list_len_correct() {
267 MutexLockerEx x(_cbl_mon, Mutex::_no_safepoint_check_flag);
268 assert_completed_buffer_list_len_correct_locked();
269 }
271 void PtrQueueSet::assert_completed_buffer_list_len_correct_locked() {
272 guarantee(completed_buffers_list_length() == _n_completed_buffers,
273 "Completed buffer length is wrong.");
274 }
276 void PtrQueueSet::set_buffer_size(size_t sz) {
277 assert(_sz == 0 && sz > 0, "Should be called only once.");
278 _sz = sz * oopSize;
279 }
281 // Merge lists of buffers. Notify the processing threads.
282 // The source queue is emptied as a result. The queues
283 // must share the monitor.
284 void PtrQueueSet::merge_bufferlists(PtrQueueSet *src) {
285 assert(_cbl_mon == src->_cbl_mon, "Should share the same lock");
286 MutexLockerEx x(_cbl_mon, Mutex::_no_safepoint_check_flag);
287 if (_completed_buffers_tail == NULL) {
288 assert(_completed_buffers_head == NULL, "Well-formedness");
289 _completed_buffers_head = src->_completed_buffers_head;
290 _completed_buffers_tail = src->_completed_buffers_tail;
291 } else {
292 assert(_completed_buffers_head != NULL, "Well formedness");
293 if (src->_completed_buffers_head != NULL) {
294 _completed_buffers_tail->set_next(src->_completed_buffers_head);
295 _completed_buffers_tail = src->_completed_buffers_tail;
296 }
297 }
298 _n_completed_buffers += src->_n_completed_buffers;
300 src->_n_completed_buffers = 0;
301 src->_completed_buffers_head = NULL;
302 src->_completed_buffers_tail = NULL;
304 assert(_completed_buffers_head == NULL && _completed_buffers_tail == NULL ||
305 _completed_buffers_head != NULL && _completed_buffers_tail != NULL,
306 "Sanity");
307 }
309 void PtrQueueSet::notify_if_necessary() {
310 MutexLockerEx x(_cbl_mon, Mutex::_no_safepoint_check_flag);
311 if (_n_completed_buffers >= _process_completed_threshold || _max_completed_queue == 0) {
312 _process_completed = true;
313 if (_notify_when_complete)
314 _cbl_mon->notify();
315 }
316 }