1 /*
2 * Copyright (c) 1999, 2018, 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/threadLocalAllocBuffer.inline.hpp"
27 #include "logging/log.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "memory/universe.hpp"
30 #include "oops/oop.inline.hpp"
31 #include "runtime/thread.inline.hpp"
32 #include "runtime/threadSMR.hpp"
33 #include "utilities/copy.hpp"
34
35 // Thread-Local Edens support
36
37 // static member initialization
38 size_t ThreadLocalAllocBuffer::_max_size = 0;
39 int ThreadLocalAllocBuffer::_reserve_for_allocation_prefetch = 0;
40 unsigned ThreadLocalAllocBuffer::_target_refills = 0;
41 GlobalTLABStats* ThreadLocalAllocBuffer::_global_stats = NULL;
42
43 void ThreadLocalAllocBuffer::clear_before_allocation() {
44 _slow_refill_waste += (unsigned)remaining();
45 make_parsable(true); // also retire the TLAB
46 }
47
48 void ThreadLocalAllocBuffer::accumulate_statistics_before_gc() {
49 global_stats()->initialize();
50
51 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *thread = jtiwh.next(); ) {
52 thread->tlab().accumulate_statistics();
53 thread->tlab().initialize_statistics();
54 }
55
56 Universe::heap()->accumulate_statistics_all_gclabs();
57
58 // Publish new stats if some allocation occurred.
59 if (global_stats()->allocation() != 0) {
60 global_stats()->publish();
61 global_stats()->print();
62 }
63 }
64
65 void ThreadLocalAllocBuffer::accumulate_statistics() {
66 Thread* thread = myThread();
67 size_t capacity = Universe::heap()->tlab_capacity(thread);
68 size_t used = Universe::heap()->tlab_used(thread);
69
70 _gc_waste += (unsigned)remaining();
71 size_t total_allocated = _gclab ? thread->allocated_bytes_gclab() : thread->allocated_bytes();
72 size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
73 _allocated_before_last_gc = total_allocated;
74
75 print_stats("gc");
76
77 if (_number_of_refills > 0) {
78 // Update allocation history if a reasonable amount of eden was allocated.
79 bool update_allocation_history = used > 0.5 * capacity;
80
81 if (update_allocation_history) {
82 // Average the fraction of eden allocated in a tlab by this
83 // thread for use in the next resize operation.
84 // _gc_waste is not subtracted because it's included in
85 // "used".
86 // The result can be larger than 1.0 due to direct to old allocations.
87 // These allocations should ideally not be counted but since it is not possible
88 // to filter them out here we just cap the fraction to be at most 1.0.
89 double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
90 _allocation_fraction.sample(alloc_frac);
91 }
92 global_stats()->update_allocating_threads();
93 global_stats()->update_number_of_refills(_number_of_refills);
94 global_stats()->update_allocation(_number_of_refills * desired_size());
95 global_stats()->update_gc_waste(_gc_waste);
96 global_stats()->update_slow_refill_waste(_slow_refill_waste);
97 global_stats()->update_fast_refill_waste(_fast_refill_waste);
98
99 } else {
100 assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
101 _slow_refill_waste == 0 && _gc_waste == 0,
102 "tlab stats == 0");
103 }
104 global_stats()->update_slow_allocations(_slow_allocations);
105 }
106
107 // Fills the current tlab with a dummy filler array to create
108 // an illusion of a contiguous Eden and optionally retires the tlab.
109 // Waste accounting should be done in caller as appropriate; see,
110 // for example, clear_before_allocation().
111 void ThreadLocalAllocBuffer::make_parsable(bool retire, bool zap) {
112 if (end() != NULL) {
113 invariants();
114
115 if (retire) {
116 if (_gclab) {
117 myThread()->incr_allocated_bytes_gclab(used_bytes());
118 } else {
119 myThread()->incr_allocated_bytes(used_bytes());
120 }
121 }
122
123 HeapWord* obj = Universe::heap()->tlab_post_allocation_setup(top());
124 CollectedHeap::fill_with_object(obj, hard_end(), retire && zap);
125
126 if (retire || ZeroTLAB) { // "Reset" the TLAB
127 set_start(NULL);
128 set_top(NULL);
129 set_pf_top(NULL);
130 set_end(NULL);
131 }
132 }
133 assert(!(retire || ZeroTLAB) ||
134 (start() == NULL && end() == NULL && top() == NULL),
135 "TLAB must be reset");
136 }
137
138 void ThreadLocalAllocBuffer::resize_all_tlabs() {
139 if (ResizeTLAB) {
140 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *thread = jtiwh.next(); ) {
141 thread->tlab().resize();
142 }
143 }
144 }
145
146 void ThreadLocalAllocBuffer::resize() {
147 // Compute the next tlab size using expected allocation amount
148 assert(ResizeTLAB, "Should not call this otherwise");
149 size_t alloc = (size_t)(_allocation_fraction.average() *
150 (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize));
151 size_t new_size = alloc / _target_refills;
152
153 new_size = MIN2(MAX2(new_size, min_size()), max_size());
154
155 size_t aligned_new_size = align_object_size(new_size);
156
157 log_trace(gc, tlab)("TLAB new size: thread: " INTPTR_FORMAT " [id: %2d]"
158 " refills %d alloc: %8.6f desired_size: " SIZE_FORMAT " -> " SIZE_FORMAT,
159 p2i(myThread()), myThread()->osthread()->thread_id(),
160 _target_refills, _allocation_fraction.average(), desired_size(), aligned_new_size);
161
162 set_desired_size(aligned_new_size);
163 set_refill_waste_limit(initial_refill_waste_limit());
164 }
165
166 void ThreadLocalAllocBuffer::initialize_statistics() {
167 _number_of_refills = 0;
168 _fast_refill_waste = 0;
169 _slow_refill_waste = 0;
170 _gc_waste = 0;
171 _slow_allocations = 0;
172 }
173
174 void ThreadLocalAllocBuffer::fill(HeapWord* start,
175 HeapWord* top,
176 size_t new_size) {
177 _number_of_refills++;
178 print_stats("fill");
179 assert(top <= start + new_size - alignment_reserve(), "size too small");
180 initialize(start, top, start + new_size - alignment_reserve());
181
182 // Reset amount of internal fragmentation
183 set_refill_waste_limit(initial_refill_waste_limit());
184 }
185
186 void ThreadLocalAllocBuffer::initialize(HeapWord* start,
187 HeapWord* top,
188 HeapWord* end) {
189 set_start(start);
190 set_top(top);
191 set_pf_top(top);
192 set_end(end);
193 invariants();
194 }
195
196 void ThreadLocalAllocBuffer::initialize(bool gclab) {
197 _initialized = true;
198 _gclab = gclab;
199 initialize(NULL, // start
200 NULL, // top
201 NULL); // end
202
203 set_desired_size(initial_desired_size());
204
205 // Following check is needed because at startup the main
206 // thread is initialized before the heap is. The initialization for
207 // this thread is redone in startup_initialization below.
208 if (Universe::heap() != NULL) {
209 size_t capacity = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
210 double alloc_frac = desired_size() * target_refills() / (double) capacity;
211 _allocation_fraction.sample(alloc_frac);
212 }
213
214 set_refill_waste_limit(initial_refill_waste_limit());
215
216 initialize_statistics();
217 }
218
219 void ThreadLocalAllocBuffer::startup_initialization() {
220
221 // Assuming each thread's active tlab is, on average,
222 // 1/2 full at a GC
223 _target_refills = 100 / (2 * TLABWasteTargetPercent);
224 _target_refills = MAX2(_target_refills, (unsigned)1U);
225
226 _global_stats = new GlobalTLABStats();
227
228 #ifdef COMPILER2
229 // If the C2 compiler is present, extra space is needed at the end of
230 // TLABs, otherwise prefetching instructions generated by the C2
231 // compiler will fault (due to accessing memory outside of heap).
232 // The amount of space is the max of the number of lines to
233 // prefetch for array and for instance allocations. (Extra space must be
234 // reserved to accommodate both types of allocations.)
235 //
236 // Only SPARC-specific BIS instructions are known to fault. (Those
237 // instructions are generated if AllocatePrefetchStyle==3 and
238 // AllocatePrefetchInstr==1). To be on the safe side, however,
239 // extra space is reserved for all combinations of
240 // AllocatePrefetchStyle and AllocatePrefetchInstr.
241 //
242 // If the C2 compiler is not present, no space is reserved.
243
244 // +1 for rounding up to next cache line, +1 to be safe
245 if (is_server_compilation_mode_vm()) {
246 int lines = MAX2(AllocatePrefetchLines, AllocateInstancePrefetchLines) + 2;
247 _reserve_for_allocation_prefetch = (AllocatePrefetchDistance + AllocatePrefetchStepSize * lines) /
248 (int)HeapWordSize;
249 }
250 #endif
251
252 // During jvm startup, the main thread is initialized
253 // before the heap is initialized. So reinitialize it now.
254 guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
255 Thread::current()->tlab().initialize(false);
256 Thread::current()->gclab().initialize(true);
257
258 log_develop_trace(gc, tlab)("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT,
259 min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
260 }
261
262 size_t ThreadLocalAllocBuffer::initial_desired_size() {
263 size_t init_sz = 0;
264
265 if (TLABSize > 0) {
266 init_sz = TLABSize / HeapWordSize;
267 } else if (global_stats() != NULL) {
268 // Initial size is a function of the average number of allocating threads.
269 unsigned nof_threads = global_stats()->allocating_threads_avg();
270
271 init_sz = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
272 (nof_threads * target_refills());
273 init_sz = align_object_size(init_sz);
274 }
275 init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
276 return init_sz;
277 }
278
279 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
280 Log(gc, tlab) log;
281 if (!log.is_trace()) {
282 return;
283 }
284
285 Thread* thrd = myThread();
286 size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
287 size_t alloc = _number_of_refills * _desired_size;
288 double waste_percent = percent_of(waste, alloc);
289 size_t tlab_used = Universe::heap()->tlab_used(thrd);
290 log.trace("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
291 " desired_size: " SIZE_FORMAT "KB"
292 " slow allocs: %d refill waste: " SIZE_FORMAT "B"
293 " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
294 " slow: %dB fast: %dB",
295 tag, p2i(thrd), thrd->osthread()->thread_id(),
296 _desired_size / (K / HeapWordSize),
297 _slow_allocations, _refill_waste_limit * HeapWordSize,
298 _allocation_fraction.average(),
299 _allocation_fraction.average() * tlab_used / K,
300 _number_of_refills, waste_percent,
301 _gc_waste * HeapWordSize,
302 _slow_refill_waste * HeapWordSize,
303 _fast_refill_waste * HeapWordSize);
304 }
305
306 void ThreadLocalAllocBuffer::verify() {
307 HeapWord* p = start();
308 HeapWord* t = top();
309 HeapWord* prev_p = NULL;
310 while (p < t) {
311 oop(p)->verify();
312 prev_p = p;
313 p += oop(p)->size();
314 }
315 guarantee(p == top(), "end of last object must match end of space");
316 }
317
318 Thread* ThreadLocalAllocBuffer::myThread() {
319 ByteSize gclab_offset = Thread::gclab_start_offset();
320 ByteSize tlab_offset = Thread::tlab_start_offset();
321 ByteSize offs = _gclab ? gclab_offset : tlab_offset;
322 Thread* thread = (Thread*)(((char *)this) +
323 in_bytes(start_offset()) - in_bytes(offs));
324 #ifdef ASSERT
325 assert(this == (_gclab ? &thread->gclab() : &thread->tlab()), "must be");
326 #endif
327 return thread;
328 }
329
330 size_t ThreadLocalAllocBuffer::end_reserve() {
331 int reserve_size = typeArrayOopDesc::header_size(T_INT);
332 return MAX2(reserve_size, _reserve_for_allocation_prefetch);
333 }
334
335 void ThreadLocalAllocBuffer::rollback(size_t size) {
336 HeapWord* old_top = top();
337 if (old_top != NULL) { // Pathological case: we accept that we can't rollback.
338 set_top(old_top - size);
339 }
340 }
341
342
343 GlobalTLABStats::GlobalTLABStats() :
344 _allocating_threads_avg(TLABAllocationWeight) {
345
346 initialize();
347
348 _allocating_threads_avg.sample(1); // One allocating thread at startup
349
350 if (UsePerfData) {
351
352 EXCEPTION_MARK;
353 ResourceMark rm;
354
355 char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
356 _perf_allocating_threads =
357 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
358
359 cname = PerfDataManager::counter_name("tlab", "fills");
360 _perf_total_refills =
361 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
362
363 cname = PerfDataManager::counter_name("tlab", "maxFills");
364 _perf_max_refills =
365 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
366
367 cname = PerfDataManager::counter_name("tlab", "alloc");
368 _perf_allocation =
369 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
370
371 cname = PerfDataManager::counter_name("tlab", "gcWaste");
372 _perf_gc_waste =
373 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
374
375 cname = PerfDataManager::counter_name("tlab", "maxGcWaste");
376 _perf_max_gc_waste =
377 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
378
379 cname = PerfDataManager::counter_name("tlab", "slowWaste");
380 _perf_slow_refill_waste =
381 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
382
383 cname = PerfDataManager::counter_name("tlab", "maxSlowWaste");
384 _perf_max_slow_refill_waste =
385 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
386
387 cname = PerfDataManager::counter_name("tlab", "fastWaste");
388 _perf_fast_refill_waste =
389 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
390
391 cname = PerfDataManager::counter_name("tlab", "maxFastWaste");
392 _perf_max_fast_refill_waste =
393 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
394
395 cname = PerfDataManager::counter_name("tlab", "slowAlloc");
396 _perf_slow_allocations =
397 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
398
399 cname = PerfDataManager::counter_name("tlab", "maxSlowAlloc");
400 _perf_max_slow_allocations =
401 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
402 }
403 }
404
405 void GlobalTLABStats::initialize() {
406 // Clear counters summarizing info from all threads
407 _allocating_threads = 0;
408 _total_refills = 0;
409 _max_refills = 0;
410 _total_allocation = 0;
411 _total_gc_waste = 0;
412 _max_gc_waste = 0;
413 _total_slow_refill_waste = 0;
414 _max_slow_refill_waste = 0;
415 _total_fast_refill_waste = 0;
416 _max_fast_refill_waste = 0;
417 _total_slow_allocations = 0;
418 _max_slow_allocations = 0;
419 }
420
421 void GlobalTLABStats::publish() {
422 _allocating_threads_avg.sample(_allocating_threads);
423 if (UsePerfData) {
424 _perf_allocating_threads ->set_value(_allocating_threads);
425 _perf_total_refills ->set_value(_total_refills);
426 _perf_max_refills ->set_value(_max_refills);
427 _perf_allocation ->set_value(_total_allocation);
428 _perf_gc_waste ->set_value(_total_gc_waste);
429 _perf_max_gc_waste ->set_value(_max_gc_waste);
430 _perf_slow_refill_waste ->set_value(_total_slow_refill_waste);
431 _perf_max_slow_refill_waste->set_value(_max_slow_refill_waste);
432 _perf_fast_refill_waste ->set_value(_total_fast_refill_waste);
433 _perf_max_fast_refill_waste->set_value(_max_fast_refill_waste);
434 _perf_slow_allocations ->set_value(_total_slow_allocations);
435 _perf_max_slow_allocations ->set_value(_max_slow_allocations);
436 }
437 }
438
439 void GlobalTLABStats::print() {
440 Log(gc, tlab) log;
441 if (!log.is_debug()) {
442 return;
443 }
444
445 size_t waste = _total_gc_waste + _total_slow_refill_waste + _total_fast_refill_waste;
446 double waste_percent = percent_of(waste, _total_allocation);
447 log.debug("TLAB totals: thrds: %d refills: %d max: %d"
448 " slow allocs: %d max %d waste: %4.1f%%"
449 " gc: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
450 " slow: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
451 " fast: " SIZE_FORMAT "B max: " SIZE_FORMAT "B",
452 _allocating_threads,
453 _total_refills, _max_refills,
454 _total_slow_allocations, _max_slow_allocations,
455 waste_percent,
456 _total_gc_waste * HeapWordSize,
457 _max_gc_waste * HeapWordSize,
458 _total_slow_refill_waste * HeapWordSize,
459 _max_slow_refill_waste * HeapWordSize,
460 _total_fast_refill_waste * HeapWordSize,
461 _max_fast_refill_waste * HeapWordSize);
462 }