1 /*
2 * Copyright (c) 2001, 2020, 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
93 if (num_max_threads > 0) {
94 G1BarrierSet::dirty_card_queue_set().set_primary_refinement_thread(_threads[0]);
95 }
96
97 return JNI_OK;
98 }
99
100 void G1ConcurrentRefineThreadControl::maybe_activate_next(uint cur_worker_id) {
101 assert(cur_worker_id < _num_max_threads,
102 "Activating another thread from %u not allowed since there can be at most %u",
103 cur_worker_id, _num_max_threads);
104 if (cur_worker_id == (_num_max_threads - 1)) {
105 // Already the last thread, there is no more thread to activate.
106 return;
107 }
108
109 uint worker_id = cur_worker_id + 1;
110 G1ConcurrentRefineThread* thread_to_activate = _threads[worker_id];
111 if (thread_to_activate == NULL) {
112 // Still need to create the thread...
113 _threads[worker_id] = create_refinement_thread(worker_id, false);
114 thread_to_activate = _threads[worker_id];
115 }
116 if (thread_to_activate != NULL) {
117 thread_to_activate->activate();
118 }
119 }
120
121 void G1ConcurrentRefineThreadControl::print_on(outputStream* st) const {
122 for (uint i = 0; i < _num_max_threads; ++i) {
123 if (_threads[i] != NULL) {
124 _threads[i]->print_on(st);
125 st->cr();
126 }
127 }
128 }
129
130 void G1ConcurrentRefineThreadControl::worker_threads_do(ThreadClosure* tc) {
131 for (uint i = 0; i < _num_max_threads; i++) {
132 if (_threads[i] != NULL) {
133 tc->do_thread(_threads[i]);
134 }
135 }
136 }
137
138 void G1ConcurrentRefineThreadControl::stop() {
139 for (uint i = 0; i < _num_max_threads; i++) {
140 if (_threads[i] != NULL) {
141 _threads[i]->stop();
142 }
143 }
144 }
145
146 // Arbitrary but large limits, to simplify some of the zone calculations.
147 // The general idea is to allow expressions like
148 // MIN2(x OP y, max_XXX_zone)
149 // without needing to check for overflow in "x OP y", because the
150 // ranges for x and y have been restricted.
151 STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2));
152 const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort);
153 const size_t max_green_zone = max_yellow_zone / 2;
154 const size_t max_red_zone = INT_MAX; // For dcqs.set_max_cards.
155 STATIC_ASSERT(max_yellow_zone <= max_red_zone);
156
157 // Range check assertions for green zone values.
158 #define assert_zone_constraints_g(green) \
159 do { \
160 size_t azc_g_green = (green); \
161 assert(azc_g_green <= max_green_zone, \
162 "green exceeds max: " SIZE_FORMAT, azc_g_green); \
163 } while (0)
164
165 // Range check assertions for green and yellow zone values.
166 #define assert_zone_constraints_gy(green, yellow) \
167 do { \
168 size_t azc_gy_green = (green); \
169 size_t azc_gy_yellow = (yellow); \
170 assert_zone_constraints_g(azc_gy_green); \
171 assert(azc_gy_yellow <= max_yellow_zone, \
172 "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow); \
173 assert(azc_gy_green <= azc_gy_yellow, \
174 "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")", \
175 azc_gy_green, azc_gy_yellow); \
176 } while (0)
177
178 // Range check assertions for green, yellow, and red zone values.
179 #define assert_zone_constraints_gyr(green, yellow, red) \
180 do { \
181 size_t azc_gyr_green = (green); \
182 size_t azc_gyr_yellow = (yellow); \
183 size_t azc_gyr_red = (red); \
184 assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow); \
185 assert(azc_gyr_red <= max_red_zone, \
186 "red exceeds max: " SIZE_FORMAT, azc_gyr_red); \
187 assert(azc_gyr_yellow <= azc_gyr_red, \
188 "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")", \
189 azc_gyr_yellow, azc_gyr_red); \
190 } while (0)
191
192 // Logging tag sequence for refinement control updates.
193 #define CTRL_TAGS gc, ergo, refine
194
195 // For logging zone values, ensuring consistency of level and tags.
196 #define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__)
197
198 static size_t buffers_to_cards(size_t value) {
199 return value * G1UpdateBufferSize;
200 }
201
202 // Package for pair of refinement thread activation and deactivation
203 // thresholds. The activation and deactivation levels are resp. the first
204 // and second values of the pair.
205 typedef Pair<size_t, size_t> Thresholds;
206 inline size_t activation_level(const Thresholds& t) { return t.first; }
207 inline size_t deactivation_level(const Thresholds& t) { return t.second; }
208
209 static Thresholds calc_thresholds(size_t green_zone,
210 size_t yellow_zone,
211 uint worker_id) {
212 double yellow_size = yellow_zone - green_zone;
213 double step = yellow_size / G1ConcurrentRefine::max_num_threads();
214 if (worker_id == 0) {
215 // Potentially activate worker 0 more aggressively, to keep
216 // available buffers near green_zone value. When yellow_size is
217 // large we don't want to allow a full step to accumulate before
218 // doing any processing, as that might lead to significantly more
219 // than green_zone buffers to be processed during pause. So limit
220 // to an extra half buffer per pause-time processing thread.
221 step = MIN2(step, buffers_to_cards(ParallelGCThreads) / 2.0);
222 }
223 size_t activate_offset = static_cast<size_t>(ceil(step * (worker_id + 1)));
224 size_t deactivate_offset = static_cast<size_t>(floor(step * worker_id));
225 return Thresholds(green_zone + activate_offset,
226 green_zone + deactivate_offset);
227 }
228
229 G1ConcurrentRefine::G1ConcurrentRefine(size_t green_zone,
230 size_t yellow_zone,
231 size_t red_zone,
232 size_t min_yellow_zone_size) :
233 _thread_control(),
234 _green_zone(green_zone),
235 _yellow_zone(yellow_zone),
236 _red_zone(red_zone),
237 _min_yellow_zone_size(min_yellow_zone_size)
238 {
239 assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone);
240 }
241
242 jint G1ConcurrentRefine::initialize() {
243 return _thread_control.initialize(this, max_num_threads());
244 }
245
246 static size_t calc_min_yellow_zone_size() {
247 size_t step = buffers_to_cards(G1ConcRefinementThresholdStep);
248 uint n_workers = G1ConcurrentRefine::max_num_threads();
249 if ((max_yellow_zone / step) < n_workers) {
250 return max_yellow_zone;
251 } else {
252 return step * n_workers;
253 }
254 }
255
256 static size_t calc_init_green_zone() {
257 size_t green = G1ConcRefinementGreenZone;
258 if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) {
259 green = ParallelGCThreads;
260 }
261 green = buffers_to_cards(green);
262 return MIN2(green, max_green_zone);
263 }
264
265 static size_t calc_init_yellow_zone(size_t green, size_t min_size) {
266 size_t config = buffers_to_cards(G1ConcRefinementYellowZone);
267 size_t size = 0;
268 if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) {
269 size = green * 2;
270 } else if (green < config) {
271 size = config - green;
272 }
273 size = MAX2(size, min_size);
274 size = MIN2(size, max_yellow_zone);
275 return MIN2(green + size, max_yellow_zone);
276 }
277
278 static size_t calc_init_red_zone(size_t green, size_t yellow) {
279 size_t size = yellow - green;
280 if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) {
281 size_t config = buffers_to_cards(G1ConcRefinementRedZone);
282 if (yellow < config) {
283 size = MAX2(size, config - yellow);
284 }
285 }
286 return MIN2(yellow + size, max_red_zone);
287 }
288
289 G1ConcurrentRefine* G1ConcurrentRefine::create(jint* ecode) {
290 size_t min_yellow_zone_size = calc_min_yellow_zone_size();
291 size_t green_zone = calc_init_green_zone();
292 size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size);
293 size_t red_zone = calc_init_red_zone(green_zone, yellow_zone);
294
295 LOG_ZONES("Initial Refinement Zones: "
296 "green: " SIZE_FORMAT ", "
297 "yellow: " SIZE_FORMAT ", "
298 "red: " SIZE_FORMAT ", "
299 "min yellow size: " SIZE_FORMAT,
300 green_zone, yellow_zone, red_zone, min_yellow_zone_size);
301
302 G1ConcurrentRefine* cr = new G1ConcurrentRefine(green_zone,
303 yellow_zone,
304 red_zone,
305 min_yellow_zone_size);
306
307 if (cr == NULL) {
308 *ecode = JNI_ENOMEM;
309 vm_shutdown_during_initialization("Could not create G1ConcurrentRefine");
310 return NULL;
311 }
312
313 *ecode = cr->initialize();
314 return cr;
315 }
316
317 void G1ConcurrentRefine::stop() {
318 _thread_control.stop();
319 }
320
321 G1ConcurrentRefine::~G1ConcurrentRefine() {
322 }
323
324 void G1ConcurrentRefine::threads_do(ThreadClosure *tc) {
325 _thread_control.worker_threads_do(tc);
326 }
327
328 uint G1ConcurrentRefine::max_num_threads() {
329 return G1ConcRefinementThreads;
330 }
331
332 void G1ConcurrentRefine::print_threads_on(outputStream* st) const {
333 _thread_control.print_on(st);
334 }
335
336 static size_t calc_new_green_zone(size_t green,
337 double logged_cards_scan_time,
338 size_t processed_logged_cards,
339 double goal_ms) {
340 // Adjust green zone based on whether we're meeting the time goal.
341 // Limit to max_green_zone.
342 const double inc_k = 1.1, dec_k = 0.9;
343 if (logged_cards_scan_time > goal_ms) {
344 if (green > 0) {
345 green = static_cast<size_t>(green * dec_k);
346 }
347 } else if (logged_cards_scan_time < goal_ms &&
348 processed_logged_cards > green) {
349 green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0));
350 green = MIN2(green, max_green_zone);
351 }
352 return green;
353 }
354
355 static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) {
356 size_t size = green * 2;
357 size = MAX2(size, min_yellow_size);
358 return MIN2(green + size, max_yellow_zone);
359 }
360
361 static size_t calc_new_red_zone(size_t green, size_t yellow) {
362 return MIN2(yellow + (yellow - green), max_red_zone);
363 }
364
365 void G1ConcurrentRefine::update_zones(double logged_cards_scan_time,
366 size_t processed_logged_cards,
367 double goal_ms) {
368 log_trace( CTRL_TAGS )("Updating Refinement Zones: "
369 "logged cards scan time: %.3fms, "
370 "processed cards: " SIZE_FORMAT ", "
371 "goal time: %.3fms",
372 logged_cards_scan_time,
373 processed_logged_cards,
374 goal_ms);
375
376 _green_zone = calc_new_green_zone(_green_zone,
377 logged_cards_scan_time,
378 processed_logged_cards,
379 goal_ms);
380 _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size);
381 _red_zone = calc_new_red_zone(_green_zone, _yellow_zone);
382
383 assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone);
384 LOG_ZONES("Updated Refinement Zones: "
385 "green: " SIZE_FORMAT ", "
386 "yellow: " SIZE_FORMAT ", "
387 "red: " SIZE_FORMAT,
388 _green_zone, _yellow_zone, _red_zone);
389 }
390
391 void G1ConcurrentRefine::adjust(double logged_cards_scan_time,
392 size_t processed_logged_cards,
393 double goal_ms) {
394 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
395
396 if (G1UseAdaptiveConcRefinement) {
397 update_zones(logged_cards_scan_time, processed_logged_cards, goal_ms);
398
399 // Change the barrier params
400 if (max_num_threads() == 0) {
401 // Disable dcqs notification when there are no threads to notify.
402 dcqs.set_process_cards_threshold(G1DirtyCardQueueSet::ProcessCardsThresholdNever);
403 } else {
404 // Worker 0 is the primary; wakeup is via dcqs notification.
405 STATIC_ASSERT(max_yellow_zone <= INT_MAX);
406 size_t activate = activation_threshold(0);
407 dcqs.set_process_cards_threshold(activate);
408 }
409 dcqs.set_max_cards(red_zone());
410 }
411
412 size_t curr_queue_size = dcqs.num_cards();
413 if ((dcqs.max_cards() > 0) &&
414 (curr_queue_size >= yellow_zone())) {
415 dcqs.set_max_cards_padding(curr_queue_size);
416 } else {
417 dcqs.set_max_cards_padding(0);
418 }
419 dcqs.notify_if_necessary();
420 }
421
422 G1ConcurrentRefine::RefinementStats G1ConcurrentRefine::total_refinement_stats() const {
423 struct CollectData : public ThreadClosure {
424 Tickspan _total_time;
425 size_t _total_cards;
426 CollectData() : _total_time(), _total_cards(0) {}
427 virtual void do_thread(Thread* t) {
428 G1ConcurrentRefineThread* crt = static_cast<G1ConcurrentRefineThread*>(t);
429 _total_time += crt->total_refinement_time();
430 _total_cards += crt->total_refined_cards();
431 }
432 } collector;
433 // Cast away const so we can call non-modifying closure on threads.
434 const_cast<G1ConcurrentRefine*>(this)->threads_do(&collector);
435 return RefinementStats(collector._total_time, collector._total_cards);
436 }
437
438 size_t G1ConcurrentRefine::activation_threshold(uint worker_id) const {
439 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
440 return activation_level(thresholds);
441 }
442
443 size_t G1ConcurrentRefine::deactivation_threshold(uint worker_id) const {
444 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
445 return deactivation_level(thresholds);
446 }
447
448 uint G1ConcurrentRefine::worker_id_offset() {
449 return G1DirtyCardQueueSet::num_par_ids();
450 }
451
452 void G1ConcurrentRefine::maybe_activate_more_threads(uint worker_id, size_t num_cur_cards) {
453 if (num_cur_cards > activation_threshold(worker_id + 1)) {
454 _thread_control.maybe_activate_next(worker_id);
455 }
456 }
457
458 bool G1ConcurrentRefine::do_refinement_step(uint worker_id,
459 size_t* total_refined_cards) {
460 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
461
462 size_t curr_cards = dcqs.num_cards();
463 // If the number of the cards falls down into the yellow zone,
464 // that means that the transition period after the evacuation pause has ended.
465 // Since the value written to the DCQS is the same for all threads, there is no
466 // need to synchronize.
467 if (dcqs.max_cards_padding() > 0 && curr_cards <= yellow_zone()) {
468 dcqs.set_max_cards_padding(0);
469 }
470
471 maybe_activate_more_threads(worker_id, curr_cards);
472
473 // Process the next buffer, if there are enough left.
474 return dcqs.refine_completed_buffer_concurrently(worker_id + worker_id_offset(),
475 deactivation_threshold(worker_id),
476 total_refined_cards);
477 }