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 #include "precompiled.hpp"
26 #include "gc/g1/g1BarrierSet.hpp"
27 #include "gc/g1/g1ConcurrentRefine.hpp"
28 #include "gc/g1/g1ConcurrentRefineThread.hpp"
29 #include "gc/g1/g1DirtyCardQueue.hpp"
30 #include "logging/log.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/iterator.hpp"
33 #include "runtime/java.hpp"
34 #include "runtime/thread.hpp"
35 #include "utilities/debug.hpp"
36 #include "utilities/globalDefinitions.hpp"
37 #include "utilities/pair.hpp"
38 #include <math.h>
39
40 G1ConcurrentRefineThread* G1ConcurrentRefineThreadControl::create_refinement_thread(uint worker_id, bool initializing) {
41 G1ConcurrentRefineThread* result = NULL;
42 if (initializing || !InjectGCWorkerCreationFailure) {
43 result = new G1ConcurrentRefineThread(_cr, worker_id);
44 }
45 if (result == NULL || result->osthread() == NULL) {
46 log_warning(gc)("Failed to create refinement thread %u, no more %s",
47 worker_id,
48 result == NULL ? "memory" : "OS threads");
49 }
50 return result;
51 }
52
53 G1ConcurrentRefineThreadControl::G1ConcurrentRefineThreadControl() :
54 _cr(NULL),
55 _threads(NULL),
56 _num_max_threads(0)
57 {
58 }
59
60 G1ConcurrentRefineThreadControl::~G1ConcurrentRefineThreadControl() {
61 for (uint i = 0; i < _num_max_threads; i++) {
62 G1ConcurrentRefineThread* t = _threads[i];
63 if (t != NULL) {
64 delete t;
65 }
66 }
67 FREE_C_HEAP_ARRAY(G1ConcurrentRefineThread*, _threads);
68 }
69
70 jint G1ConcurrentRefineThreadControl::initialize(G1ConcurrentRefine* cr, uint num_max_threads) {
71 assert(cr != NULL, "G1ConcurrentRefine must not be NULL");
72 _cr = cr;
73 _num_max_threads = num_max_threads;
74
75 _threads = NEW_C_HEAP_ARRAY_RETURN_NULL(G1ConcurrentRefineThread*, num_max_threads, mtGC);
76 if (_threads == NULL) {
77 vm_shutdown_during_initialization("Could not allocate thread holder array.");
78 return JNI_ENOMEM;
79 }
80
81 for (uint i = 0; i < num_max_threads; i++) {
82 if (UseDynamicNumberOfGCThreads && i != 0 /* Always start first thread. */) {
83 _threads[i] = NULL;
84 } else {
85 _threads[i] = create_refinement_thread(i, true);
86 if (_threads[i] == NULL) {
87 vm_shutdown_during_initialization("Could not allocate refinement threads.");
88 return JNI_ENOMEM;
89 }
90 }
91 }
92 return JNI_OK;
93 }
94
95 void G1ConcurrentRefineThreadControl::maybe_activate_next(uint cur_worker_id) {
96 assert(cur_worker_id < _num_max_threads,
97 "Activating another thread from %u not allowed since there can be at most %u",
98 cur_worker_id, _num_max_threads);
99 if (cur_worker_id == (_num_max_threads - 1)) {
100 // Already the last thread, there is no more thread to activate.
101 return;
102 }
103
104 uint worker_id = cur_worker_id + 1;
105 G1ConcurrentRefineThread* thread_to_activate = _threads[worker_id];
106 if (thread_to_activate == NULL) {
107 // Still need to create the thread...
108 _threads[worker_id] = create_refinement_thread(worker_id, false);
109 thread_to_activate = _threads[worker_id];
110 }
111 if (thread_to_activate != NULL && !thread_to_activate->is_active()) {
112 thread_to_activate->activate();
113 }
114 }
115
116 void G1ConcurrentRefineThreadControl::print_on(outputStream* st) const {
117 for (uint i = 0; i < _num_max_threads; ++i) {
118 if (_threads[i] != NULL) {
119 _threads[i]->print_on(st);
120 st->cr();
121 }
122 }
123 }
124
125 void G1ConcurrentRefineThreadControl::worker_threads_do(ThreadClosure* tc) {
126 for (uint i = 0; i < _num_max_threads; i++) {
127 if (_threads[i] != NULL) {
128 tc->do_thread(_threads[i]);
129 }
130 }
131 }
132
133 void G1ConcurrentRefineThreadControl::stop() {
134 for (uint i = 0; i < _num_max_threads; i++) {
135 if (_threads[i] != NULL) {
136 _threads[i]->stop();
137 }
138 }
139 }
140
141 // Arbitrary but large limits, to simplify some of the zone calculations.
142 // The general idea is to allow expressions like
143 // MIN2(x OP y, max_XXX_zone)
144 // without needing to check for overflow in "x OP y", because the
145 // ranges for x and y have been restricted.
146 STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2));
147 const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort);
148 const size_t max_green_zone = max_yellow_zone / 2;
149 const size_t max_red_zone = INT_MAX; // For dcqs.set_max_cards.
150 STATIC_ASSERT(max_yellow_zone <= max_red_zone);
151
152 // Range check assertions for green zone values.
153 #define assert_zone_constraints_g(green) \
154 do { \
155 size_t azc_g_green = (green); \
156 assert(azc_g_green <= max_green_zone, \
157 "green exceeds max: " SIZE_FORMAT, azc_g_green); \
158 } while (0)
159
160 // Range check assertions for green and yellow zone values.
161 #define assert_zone_constraints_gy(green, yellow) \
162 do { \
163 size_t azc_gy_green = (green); \
164 size_t azc_gy_yellow = (yellow); \
165 assert_zone_constraints_g(azc_gy_green); \
166 assert(azc_gy_yellow <= max_yellow_zone, \
167 "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow); \
168 assert(azc_gy_green <= azc_gy_yellow, \
169 "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")", \
170 azc_gy_green, azc_gy_yellow); \
171 } while (0)
172
173 // Range check assertions for green, yellow, and red zone values.
174 #define assert_zone_constraints_gyr(green, yellow, red) \
175 do { \
176 size_t azc_gyr_green = (green); \
177 size_t azc_gyr_yellow = (yellow); \
178 size_t azc_gyr_red = (red); \
179 assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow); \
180 assert(azc_gyr_red <= max_red_zone, \
181 "red exceeds max: " SIZE_FORMAT, azc_gyr_red); \
182 assert(azc_gyr_yellow <= azc_gyr_red, \
183 "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")", \
184 azc_gyr_yellow, azc_gyr_red); \
185 } while (0)
186
187 // Logging tag sequence for refinement control updates.
188 #define CTRL_TAGS gc, ergo, refine
189
190 // For logging zone values, ensuring consistency of level and tags.
191 #define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__)
192
193 // Package for pair of refinement thread activation and deactivation
194 // thresholds. The activation and deactivation levels are resp. the first
195 // and second values of the pair.
196 typedef Pair<size_t, size_t> Thresholds;
197 inline size_t activation_level(const Thresholds& t) { return t.first; }
198 inline size_t deactivation_level(const Thresholds& t) { return t.second; }
199
200 static Thresholds calc_thresholds(size_t green_zone,
201 size_t yellow_zone,
202 uint worker_id) {
203 double yellow_size = yellow_zone - green_zone;
204 double step = yellow_size / G1ConcurrentRefine::max_num_threads();
205 if (worker_id == 0) {
206 // Potentially activate worker 0 more aggressively, to keep
207 // available buffers near green_zone value. When yellow_size is
208 // large we don't want to allow a full step to accumulate before
209 // doing any processing, as that might lead to significantly more
210 // than green_zone buffers to be processed during scanning.
211 step = MIN2(step, ParallelGCThreads / 2.0);
212 }
213 size_t activate_offset = static_cast<size_t>(ceil(step * (worker_id + 1)));
214 size_t deactivate_offset = static_cast<size_t>(floor(step * worker_id));
215 return Thresholds(green_zone + activate_offset,
216 green_zone + deactivate_offset);
217 }
218
219 G1ConcurrentRefine::G1ConcurrentRefine(size_t green_zone,
220 size_t yellow_zone,
221 size_t red_zone,
222 size_t min_yellow_zone_size) :
223 _thread_control(),
224 _green_zone(green_zone),
225 _yellow_zone(yellow_zone),
226 _red_zone(red_zone),
227 _min_yellow_zone_size(min_yellow_zone_size)
228 {
229 assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone);
230 }
231
232 jint G1ConcurrentRefine::initialize() {
233 return _thread_control.initialize(this, max_num_threads());
234 }
235
236 static size_t buffers_to_cards(size_t value) {
237 return value * G1UpdateBufferSize;
238 }
239
240 static size_t calc_min_yellow_zone_size() {
241 size_t step = buffers_to_cards(G1ConcRefinementThresholdStep);
242 uint n_workers = G1ConcurrentRefine::max_num_threads();
243 if ((max_yellow_zone / step) < n_workers) {
244 return max_yellow_zone;
245 } else {
246 return step * n_workers;
247 }
248 }
249
250 static size_t calc_init_green_zone() {
251 size_t green = G1ConcRefinementGreenZone;
252 if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) {
253 green = ParallelGCThreads;
254 }
255 green = buffers_to_cards(green);
256 return MIN2(green, max_green_zone);
257 }
258
259 static size_t calc_init_yellow_zone(size_t green, size_t min_size) {
260 size_t config = buffers_to_cards(G1ConcRefinementYellowZone);
261 size_t size = 0;
262 if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) {
263 size = green * 2;
264 } else if (green < config) {
265 size = config - green;
266 }
267 size = MAX2(size, min_size);
268 size = MIN2(size, max_yellow_zone);
269 return MIN2(green + size, max_yellow_zone);
270 }
271
272 static size_t calc_init_red_zone(size_t green, size_t yellow) {
273 size_t size = yellow - green;
274 if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) {
275 size_t config = buffers_to_cards(G1ConcRefinementRedZone);
276 if (yellow < config) {
277 size = MAX2(size, config - yellow);
278 }
279 }
280 return MIN2(yellow + size, max_red_zone);
281 }
282
283 G1ConcurrentRefine* G1ConcurrentRefine::create(jint* ecode) {
284 size_t min_yellow_zone_size = calc_min_yellow_zone_size();
285 size_t green_zone = calc_init_green_zone();
286 size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size);
287 size_t red_zone = calc_init_red_zone(green_zone, yellow_zone);
288
289 LOG_ZONES("Initial Refinement Zones: "
290 "green: " SIZE_FORMAT ", "
291 "yellow: " SIZE_FORMAT ", "
292 "red: " SIZE_FORMAT ", "
293 "min yellow size: " SIZE_FORMAT,
294 green_zone, yellow_zone, red_zone, min_yellow_zone_size);
295
296 G1ConcurrentRefine* cr = new G1ConcurrentRefine(green_zone,
297 yellow_zone,
298 red_zone,
299 min_yellow_zone_size);
300
301 if (cr == NULL) {
302 *ecode = JNI_ENOMEM;
303 vm_shutdown_during_initialization("Could not create G1ConcurrentRefine");
304 return NULL;
305 }
306
307 *ecode = cr->initialize();
308 return cr;
309 }
310
311 void G1ConcurrentRefine::stop() {
312 _thread_control.stop();
313 }
314
315 G1ConcurrentRefine::~G1ConcurrentRefine() {
316 }
317
318 void G1ConcurrentRefine::threads_do(ThreadClosure *tc) {
319 _thread_control.worker_threads_do(tc);
320 }
321
322 uint G1ConcurrentRefine::max_num_threads() {
323 return G1ConcRefinementThreads;
324 }
325
326 void G1ConcurrentRefine::print_threads_on(outputStream* st) const {
327 _thread_control.print_on(st);
328 }
329
330 static size_t calc_new_green_zone(size_t green,
331 double logged_cards_scan_time,
332 size_t processed_logged_cards,
333 double goal_ms) {
334 // Adjust green zone based on whether we're meeting the time goal.
335 // Limit to max_green_zone.
336 const double inc_k = 1.1, dec_k = 0.9;
337 if (logged_cards_scan_time > goal_ms) {
338 if (green > 0) {
339 green = static_cast<size_t>(green * dec_k);
340 }
341 } else if (logged_cards_scan_time < goal_ms &&
342 processed_logged_cards > green) {
343 green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0));
344 green = MIN2(green, max_green_zone);
345 }
346 return green;
347 }
348
349 static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) {
350 size_t size = green * 2;
351 size = MAX2(size, min_yellow_size);
352 return MIN2(green + size, max_yellow_zone);
353 }
354
355 static size_t calc_new_red_zone(size_t green, size_t yellow) {
356 return MIN2(yellow + (yellow - green), max_red_zone);
357 }
358
359 void G1ConcurrentRefine::update_zones(double logged_cards_scan_time,
360 size_t processed_logged_cards,
361 double goal_ms) {
362 log_trace( CTRL_TAGS )("Updating Refinement Zones: "
363 "logged cards scan time: %.3fms, "
364 "processed cards: " SIZE_FORMAT ", "
365 "goal time: %.3fms",
366 logged_cards_scan_time,
367 processed_logged_cards,
368 goal_ms);
369
370 _green_zone = calc_new_green_zone(_green_zone,
371 logged_cards_scan_time,
372 processed_logged_cards,
373 goal_ms);
374 _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size);
375 _red_zone = calc_new_red_zone(_green_zone, _yellow_zone);
376
377 assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone);
378 LOG_ZONES("Updated Refinement Zones: "
379 "green: " SIZE_FORMAT ", "
380 "yellow: " SIZE_FORMAT ", "
381 "red: " SIZE_FORMAT,
382 _green_zone, _yellow_zone, _red_zone);
383 }
384
385 void G1ConcurrentRefine::adjust(double logged_cards_scan_time,
386 size_t processed_logged_cards,
387 double goal_ms) {
388 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
389
390 if (G1UseAdaptiveConcRefinement) {
391 update_zones(logged_cards_scan_time, processed_logged_cards, goal_ms);
392
393 // Change the barrier params
394 if (max_num_threads() == 0) {
395 // Disable dcqs notification when there are no threads to notify.
396 dcqs.set_process_cards_threshold(G1DirtyCardQueueSet::ProcessCardsThresholdNever);
397 } else {
398 // Worker 0 is the primary; wakeup is via dcqs notification.
399 STATIC_ASSERT(max_yellow_zone <= INT_MAX);
400 size_t activate = activation_threshold(0);
401 dcqs.set_process_cards_threshold(activate);
402 }
403 dcqs.set_max_cards(red_zone());
404 }
405
406 size_t curr_queue_size = dcqs.num_cards();
407 if ((dcqs.max_cards() > 0) &&
408 (curr_queue_size >= yellow_zone())) {
409 dcqs.set_max_cards_padding(curr_queue_size);
410 } else {
411 dcqs.set_max_cards_padding(0);
412 }
413 dcqs.notify_if_necessary();
414 }
415
416 G1ConcurrentRefine::RefinementStats G1ConcurrentRefine::total_refinement_stats() const {
417 struct CollectData : public ThreadClosure {
418 Tickspan _total_time;
419 size_t _total_cards;
420 CollectData() : _total_time(), _total_cards(0) {}
421 virtual void do_thread(Thread* t) {
422 G1ConcurrentRefineThread* crt = static_cast<G1ConcurrentRefineThread*>(t);
423 _total_time += crt->total_refinement_time();
424 _total_cards += crt->total_refined_cards();
425 }
426 } collector;
427 // Cast away const so we can call non-modifying closure on threads.
428 const_cast<G1ConcurrentRefine*>(this)->threads_do(&collector);
429 return RefinementStats(collector._total_time, collector._total_cards);
430 }
431
432 size_t G1ConcurrentRefine::activation_threshold(uint worker_id) const {
433 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
434 return activation_level(thresholds);
435 }
436
437 size_t G1ConcurrentRefine::deactivation_threshold(uint worker_id) const {
438 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
439 return deactivation_level(thresholds);
440 }
441
442 uint G1ConcurrentRefine::worker_id_offset() {
443 return G1DirtyCardQueueSet::num_par_ids();
444 }
445
446 void G1ConcurrentRefine::maybe_activate_more_threads(uint worker_id, size_t num_cur_buffers) {
447 if (num_cur_buffers > activation_threshold(worker_id + 1)) {
448 _thread_control.maybe_activate_next(worker_id);
449 }
450 }
451
452 bool G1ConcurrentRefine::do_refinement_step(uint worker_id,
453 size_t* total_refined_cards) {
454 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
455
456 size_t curr_cards = dcqs.num_cards();
457 // If the number of the cards falls down into the yellow zone,
458 // that means that the transition period after the evacuation pause has ended.
459 // Since the value written to the DCQS is the same for all threads, there is no
460 // need to synchronize.
461 if (dcqs.max_cards_padding() > 0 && curr_cards <= yellow_zone()) {
462 dcqs.set_max_cards_padding(0);
463 }
464
465 maybe_activate_more_threads(worker_id, curr_cards);
466
467 // Process the next buffer, if there are enough left.
468 return dcqs.refine_completed_buffer_concurrently(worker_id + worker_id_offset(),
469 deactivation_threshold(worker_id),
470 total_refined_cards);
471 }