1 /*
2 * Copyright (c) 1999, 2017, 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/shared/genCollectedHeap.hpp"
27 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
28 #include "logging/log.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "memory/universe.inline.hpp"
31 #include "oops/oop.inline.hpp"
32 #include "runtime/thread.inline.hpp"
33 #include "runtime/threadSMR.hpp"
34 #include "utilities/copy.hpp"
35
36 // Thread-Local Edens support
37
38 // static member initialization
39 size_t ThreadLocalAllocBuffer::_max_size = 0;
40 int ThreadLocalAllocBuffer::_reserve_for_allocation_prefetch = 0;
41 unsigned ThreadLocalAllocBuffer::_target_refills = 0;
42 GlobalTLABStats* ThreadLocalAllocBuffer::_global_stats = NULL;
43
44 void ThreadLocalAllocBuffer::clear_before_allocation() {
45 _slow_refill_waste += (unsigned)remaining();
46 make_parsable(true); // also retire the TLAB
47 }
48
49 void ThreadLocalAllocBuffer::accumulate_statistics_before_gc() {
50 global_stats()->initialize();
51
52 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *thread = jtiwh.next(); ) {
53 thread->tlab().accumulate_statistics();
54 thread->tlab().initialize_statistics();
55 }
56
57 Universe::heap()->accumulate_statistics_all_gclabs();
58
59 // Publish new stats if some allocation occurred.
60 if (global_stats()->allocation() != 0) {
61 global_stats()->publish();
62 global_stats()->print();
63 }
64 }
65
66 void ThreadLocalAllocBuffer::accumulate_statistics() {
67 Thread* thread = myThread();
68 size_t capacity = Universe::heap()->tlab_capacity(thread);
69 size_t used = Universe::heap()->tlab_used(thread);
70
71 _gc_waste += (unsigned)remaining();
72 size_t total_allocated = _gclab ? thread->allocated_bytes_gclab() : thread->allocated_bytes();
73 size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
74 _allocated_before_last_gc = total_allocated;
75
76 print_stats("gc");
77
78 if (_number_of_refills > 0) {
79 // Update allocation history if a reasonable amount of eden was allocated.
80 bool update_allocation_history = used > 0.5 * capacity;
81
82 if (update_allocation_history) {
83 // Average the fraction of eden allocated in a tlab by this
84 // thread for use in the next resize operation.
85 // _gc_waste is not subtracted because it's included in
86 // "used".
87 // The result can be larger than 1.0 due to direct to old allocations.
88 // These allocations should ideally not be counted but since it is not possible
89 // to filter them out here we just cap the fraction to be at most 1.0.
90 double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
91 _allocation_fraction.sample(alloc_frac);
92 }
93 global_stats()->update_allocating_threads();
94 global_stats()->update_number_of_refills(_number_of_refills);
95 global_stats()->update_allocation(_number_of_refills * desired_size());
96 global_stats()->update_gc_waste(_gc_waste);
97 global_stats()->update_slow_refill_waste(_slow_refill_waste);
98 global_stats()->update_fast_refill_waste(_fast_refill_waste);
99
100 } else {
101 assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
102 _slow_refill_waste == 0 && _gc_waste == 0,
103 "tlab stats == 0");
104 }
105 global_stats()->update_slow_allocations(_slow_allocations);
106 }
107
108 // Fills the current tlab with a dummy filler array to create
109 // an illusion of a contiguous Eden and optionally retires the tlab.
110 // Waste accounting should be done in caller as appropriate; see,
111 // for example, clear_before_allocation().
112 void ThreadLocalAllocBuffer::make_parsable(bool retire, bool zap) {
113 if (end() != NULL) {
114 invariants();
115
116 if (retire) {
117 if (_gclab) {
118 myThread()->incr_allocated_bytes_gclab(used_bytes());
119 } else {
120 myThread()->incr_allocated_bytes(used_bytes());
121 }
122 }
123
124 HeapWord* obj = Universe::heap()->tlab_post_allocation_setup(top());
125 CollectedHeap::fill_with_object(obj, hard_end(), retire && zap);
126
127 if (retire || ZeroTLAB) { // "Reset" the TLAB
128 set_start(NULL);
129 set_top(NULL);
130 set_pf_top(NULL);
131 set_end(NULL);
132 }
133 }
134 assert(!(retire || ZeroTLAB) ||
135 (start() == NULL && end() == NULL && top() == NULL),
136 "TLAB must be reset");
137 }
138
139 void ThreadLocalAllocBuffer::resize_all_tlabs() {
140 if (ResizeTLAB) {
141 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *thread = jtiwh.next(); ) {
142 thread->tlab().resize();
143 }
144 }
145 }
146
147 void ThreadLocalAllocBuffer::resize() {
148 // Compute the next tlab size using expected allocation amount
149 assert(ResizeTLAB, "Should not call this otherwise");
150 size_t alloc = (size_t)(_allocation_fraction.average() *
151 (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize));
152 size_t new_size = alloc / _target_refills;
153
154 new_size = MIN2(MAX2(new_size, min_size()), max_size());
155
156 size_t aligned_new_size = align_object_size(new_size);
157
158 log_trace(gc, tlab)("TLAB new size: thread: " INTPTR_FORMAT " [id: %2d]"
159 " refills %d alloc: %8.6f desired_size: " SIZE_FORMAT " -> " SIZE_FORMAT,
160 p2i(myThread()), myThread()->osthread()->thread_id(),
161 _target_refills, _allocation_fraction.average(), desired_size(), aligned_new_size);
162
163 set_desired_size(aligned_new_size);
164 set_refill_waste_limit(initial_refill_waste_limit());
165 }
166
167 void ThreadLocalAllocBuffer::initialize_statistics() {
168 _number_of_refills = 0;
169 _fast_refill_waste = 0;
170 _slow_refill_waste = 0;
171 _gc_waste = 0;
172 _slow_allocations = 0;
173 }
174
175 void ThreadLocalAllocBuffer::fill(HeapWord* start,
176 HeapWord* top,
177 size_t new_size) {
178 _number_of_refills++;
179 print_stats("fill");
180 assert(top <= start + new_size - alignment_reserve(), "size too small");
181 initialize(start, top, start + new_size - alignment_reserve());
182
183 // Reset amount of internal fragmentation
184 set_refill_waste_limit(initial_refill_waste_limit());
185 }
186
187 void ThreadLocalAllocBuffer::initialize(HeapWord* start,
188 HeapWord* top,
189 HeapWord* end) {
190 set_start(start);
191 set_top(top);
192 set_pf_top(top);
193 set_end(end);
194 invariants();
195 }
196
197 void ThreadLocalAllocBuffer::initialize(bool gclab) {
198 _initialized = true;
199 _gclab = gclab;
200 initialize(NULL, // start
201 NULL, // top
202 NULL); // end
203
204 set_desired_size(initial_desired_size());
205
206 // Following check is needed because at startup the main
207 // thread is initialized before the heap is. The initialization for
208 // this thread is redone in startup_initialization below.
209 if (Universe::heap() != NULL) {
210 size_t capacity = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
211 double alloc_frac = desired_size() * target_refills() / (double) capacity;
212 _allocation_fraction.sample(alloc_frac);
213 }
214
215 set_refill_waste_limit(initial_refill_waste_limit());
216
217 initialize_statistics();
218 }
219
220 void ThreadLocalAllocBuffer::startup_initialization() {
221
222 // Assuming each thread's active tlab is, on average,
223 // 1/2 full at a GC
224 _target_refills = 100 / (2 * TLABWasteTargetPercent);
225 _target_refills = MAX2(_target_refills, (unsigned)1U);
226
227 _global_stats = new GlobalTLABStats();
228
229 #ifdef COMPILER2
230 // If the C2 compiler is present, extra space is needed at the end of
231 // TLABs, otherwise prefetching instructions generated by the C2
232 // compiler will fault (due to accessing memory outside of heap).
233 // The amount of space is the max of the number of lines to
234 // prefetch for array and for instance allocations. (Extra space must be
235 // reserved to accommodate both types of allocations.)
236 //
237 // Only SPARC-specific BIS instructions are known to fault. (Those
238 // instructions are generated if AllocatePrefetchStyle==3 and
239 // AllocatePrefetchInstr==1). To be on the safe side, however,
240 // extra space is reserved for all combinations of
241 // AllocatePrefetchStyle and AllocatePrefetchInstr.
242 //
243 // If the C2 compiler is not present, no space is reserved.
244
245 // +1 for rounding up to next cache line, +1 to be safe
246 if (is_server_compilation_mode_vm()) {
247 int lines = MAX2(AllocatePrefetchLines, AllocateInstancePrefetchLines) + 2;
248 _reserve_for_allocation_prefetch = (AllocatePrefetchDistance + AllocatePrefetchStepSize * lines) /
249 (int)HeapWordSize;
250 }
251 #endif
252
253 // During jvm startup, the main thread is initialized
254 // before the heap is initialized. So reinitialize it now.
255 guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
256 Thread::current()->tlab().initialize(false);
257 Thread::current()->gclab().initialize(true);
258
259 log_develop_trace(gc, tlab)("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT,
260 min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
261 }
262
263 size_t ThreadLocalAllocBuffer::initial_desired_size() {
264 size_t init_sz = 0;
265
266 if (TLABSize > 0) {
267 init_sz = TLABSize / HeapWordSize;
268 } else if (global_stats() != NULL) {
269 // Initial size is a function of the average number of allocating threads.
270 unsigned nof_threads = global_stats()->allocating_threads_avg();
271
272 init_sz = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
273 (nof_threads * target_refills());
274 init_sz = align_object_size(init_sz);
275 }
276 init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
277 return init_sz;
278 }
279
280 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
281 Log(gc, tlab) log;
282 if (!log.is_trace()) {
283 return;
284 }
285
286 Thread* thrd = myThread();
287 size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
288 size_t alloc = _number_of_refills * _desired_size;
289 double waste_percent = percent_of(waste, alloc);
290 size_t tlab_used = Universe::heap()->tlab_used(thrd);
291 log.trace("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
292 " desired_size: " SIZE_FORMAT "KB"
293 " slow allocs: %d refill waste: " SIZE_FORMAT "B"
294 " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
295 " slow: %dB fast: %dB",
296 tag, p2i(thrd), thrd->osthread()->thread_id(),
297 _desired_size / (K / HeapWordSize),
298 _slow_allocations, _refill_waste_limit * HeapWordSize,
299 _allocation_fraction.average(),
300 _allocation_fraction.average() * tlab_used / K,
301 _number_of_refills, waste_percent,
302 _gc_waste * HeapWordSize,
303 _slow_refill_waste * HeapWordSize,
304 _fast_refill_waste * HeapWordSize);
305 }
306
307 void ThreadLocalAllocBuffer::verify() {
308 HeapWord* p = start();
309 HeapWord* t = top();
310 HeapWord* prev_p = NULL;
311 while (p < t) {
312 oop(p)->verify();
313 prev_p = p;
314 p += oop(p)->size();
315 }
316 guarantee(p == top(), "end of last object must match end of space");
317 }
318
319 Thread* ThreadLocalAllocBuffer::myThread() {
320 ByteSize gclab_offset = Thread::gclab_start_offset();
321 ByteSize tlab_offset = Thread::tlab_start_offset();
322 ByteSize offs = _gclab ? gclab_offset : tlab_offset;
323 Thread* thread = (Thread*)(((char *)this) +
324 in_bytes(start_offset()) - in_bytes(offs));
325 #ifdef ASSERT
326 assert(this == (_gclab ? &thread->gclab() : &thread->tlab()), "must be");
327 #endif
328 return thread;
329 }
330
331 size_t ThreadLocalAllocBuffer::end_reserve() {
332 int reserve_size = typeArrayOopDesc::header_size(T_INT);
333 return MAX2(reserve_size, _reserve_for_allocation_prefetch);
334 }
335
336 void ThreadLocalAllocBuffer::rollback(size_t size) {
337 HeapWord* old_top = top();
338 if (old_top != NULL) { // Pathological case: we accept that we can't rollback.
339 set_top(old_top - size);
340 }
341 }
342
343
344 GlobalTLABStats::GlobalTLABStats() :
345 _allocating_threads_avg(TLABAllocationWeight) {
346
347 initialize();
348
349 _allocating_threads_avg.sample(1); // One allocating thread at startup
350
351 if (UsePerfData) {
352
353 EXCEPTION_MARK;
354 ResourceMark rm;
355
356 char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
357 _perf_allocating_threads =
358 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
359
360 cname = PerfDataManager::counter_name("tlab", "fills");
361 _perf_total_refills =
362 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
363
364 cname = PerfDataManager::counter_name("tlab", "maxFills");
365 _perf_max_refills =
366 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
367
368 cname = PerfDataManager::counter_name("tlab", "alloc");
369 _perf_allocation =
370 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
371
372 cname = PerfDataManager::counter_name("tlab", "gcWaste");
373 _perf_gc_waste =
374 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
375
376 cname = PerfDataManager::counter_name("tlab", "maxGcWaste");
377 _perf_max_gc_waste =
378 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
379
380 cname = PerfDataManager::counter_name("tlab", "slowWaste");
381 _perf_slow_refill_waste =
382 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
383
384 cname = PerfDataManager::counter_name("tlab", "maxSlowWaste");
385 _perf_max_slow_refill_waste =
386 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
387
388 cname = PerfDataManager::counter_name("tlab", "fastWaste");
389 _perf_fast_refill_waste =
390 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
391
392 cname = PerfDataManager::counter_name("tlab", "maxFastWaste");
393 _perf_max_fast_refill_waste =
394 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
395
396 cname = PerfDataManager::counter_name("tlab", "slowAlloc");
397 _perf_slow_allocations =
398 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
399
400 cname = PerfDataManager::counter_name("tlab", "maxSlowAlloc");
401 _perf_max_slow_allocations =
402 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
403 }
404 }
405
406 void GlobalTLABStats::initialize() {
407 // Clear counters summarizing info from all threads
408 _allocating_threads = 0;
409 _total_refills = 0;
410 _max_refills = 0;
411 _total_allocation = 0;
412 _total_gc_waste = 0;
413 _max_gc_waste = 0;
414 _total_slow_refill_waste = 0;
415 _max_slow_refill_waste = 0;
416 _total_fast_refill_waste = 0;
417 _max_fast_refill_waste = 0;
418 _total_slow_allocations = 0;
419 _max_slow_allocations = 0;
420 }
421
422 void GlobalTLABStats::publish() {
423 _allocating_threads_avg.sample(_allocating_threads);
424 if (UsePerfData) {
425 _perf_allocating_threads ->set_value(_allocating_threads);
426 _perf_total_refills ->set_value(_total_refills);
427 _perf_max_refills ->set_value(_max_refills);
428 _perf_allocation ->set_value(_total_allocation);
429 _perf_gc_waste ->set_value(_total_gc_waste);
430 _perf_max_gc_waste ->set_value(_max_gc_waste);
431 _perf_slow_refill_waste ->set_value(_total_slow_refill_waste);
432 _perf_max_slow_refill_waste->set_value(_max_slow_refill_waste);
433 _perf_fast_refill_waste ->set_value(_total_fast_refill_waste);
434 _perf_max_fast_refill_waste->set_value(_max_fast_refill_waste);
435 _perf_slow_allocations ->set_value(_total_slow_allocations);
436 _perf_max_slow_allocations ->set_value(_max_slow_allocations);
437 }
438 }
439
440 void GlobalTLABStats::print() {
441 Log(gc, tlab) log;
442 if (!log.is_debug()) {
443 return;
444 }
445
446 size_t waste = _total_gc_waste + _total_slow_refill_waste + _total_fast_refill_waste;
447 double waste_percent = percent_of(waste, _total_allocation);
448 log.debug("TLAB totals: thrds: %d refills: %d max: %d"
449 " slow allocs: %d max %d waste: %4.1f%%"
450 " gc: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
451 " slow: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
452 " fast: " SIZE_FORMAT "B max: " SIZE_FORMAT "B",
453 _allocating_threads,
454 _total_refills, _max_refills,
455 _total_slow_allocations, _max_slow_allocations,
456 waste_percent,
457 _total_gc_waste * HeapWordSize,
458 _max_gc_waste * HeapWordSize,
459 _total_slow_refill_waste * HeapWordSize,
460 _max_slow_refill_waste * HeapWordSize,
461 _total_fast_refill_waste * HeapWordSize,
462 _max_fast_refill_waste * HeapWordSize);
463 }