36 // Thread-Local Edens support
37
38 // static member initialization
39 size_t ThreadLocalAllocBuffer::_max_size = 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 (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
52 thread->tlab().accumulate_statistics();
53 thread->tlab().initialize_statistics();
54 }
55
56 // Publish new stats if some allocation occurred.
57 if (global_stats()->allocation() != 0) {
58 global_stats()->publish();
59 if (PrintTLAB) {
60 global_stats()->print();
61 }
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 = 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 if (PrintTLAB && (_number_of_refills > 0 || Verbose)) {
76 print_stats("gc");
77 }
78
79 if (_number_of_refills > 0) {
80 // Update allocation history if a reasonable amount of eden was allocated.
81 bool update_allocation_history = used > 0.5 * capacity;
82
83 if (update_allocation_history) {
84 // Average the fraction of eden allocated in a tlab by this
85 // thread for use in the next resize operation.
86 // _gc_waste is not subtracted because it's included in
87 // "used".
88 // The result can be larger than 1.0 due to direct to old allocations.
89 // These allocations should ideally not be counted but since it is not possible
90 // to filter them out here we just cap the fraction to be at most 1.0.
91 double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
98 global_stats()->update_slow_refill_waste(_slow_refill_waste);
99 global_stats()->update_fast_refill_waste(_fast_refill_waste);
100
101 } else {
102 assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
103 _slow_refill_waste == 0 && _gc_waste == 0,
104 "tlab stats == 0");
105 }
106 global_stats()->update_slow_allocations(_slow_allocations);
107 }
108
109 // Fills the current tlab with a dummy filler array to create
110 // an illusion of a contiguous Eden and optionally retires the tlab.
111 // Waste accounting should be done in caller as appropriate; see,
112 // for example, clear_before_allocation().
113 void ThreadLocalAllocBuffer::make_parsable(bool retire) {
114 if (end() != NULL) {
115 invariants();
116
117 if (retire) {
118 myThread()->incr_allocated_bytes(used_bytes());
119 }
120
121 CollectedHeap::fill_with_object(top(), hard_end(), retire);
122
123 if (retire || ZeroTLAB) { // "Reset" the TLAB
124 set_start(NULL);
125 set_top(NULL);
126 set_pf_top(NULL);
127 set_end(NULL);
128 }
129 }
130 assert(!(retire || ZeroTLAB) ||
131 (start() == NULL && end() == NULL && top() == NULL),
132 "TLAB must be reset");
133 }
134
135 void ThreadLocalAllocBuffer::resize_all_tlabs() {
136 if (ResizeTLAB) {
137 for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
138 thread->tlab().resize();
176 if (PrintTLAB && Verbose) {
177 print_stats("fill");
178 }
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() {
197 initialize(NULL, // start
198 NULL, // top
199 NULL); // end
200
201 set_desired_size(initial_desired_size());
202
203 // Following check is needed because at startup the main
204 // thread is initialized before the heap is. The initialization for
205 // this thread is redone in startup_initialization below.
206 if (Universe::heap() != NULL) {
207 size_t capacity = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
208 double alloc_frac = desired_size() * target_refills() / (double) capacity;
209 _allocation_fraction.sample(alloc_frac);
210 }
211
212 set_refill_waste_limit(initial_refill_waste_limit());
213
214 initialize_statistics();
215 }
216
217 void ThreadLocalAllocBuffer::startup_initialization() {
218
219 // Assuming each thread's active tlab is, on average,
220 // 1/2 full at a GC
221 _target_refills = 100 / (2 * TLABWasteTargetPercent);
222 _target_refills = MAX2(_target_refills, (unsigned)1U);
223
224 _global_stats = new GlobalTLABStats();
225
226 // During jvm startup, the main thread is initialized
227 // before the heap is initialized. So reinitialize it now.
228 guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
229 Thread::current()->tlab().initialize();
230
231 if (PrintTLAB && Verbose) {
232 gclog_or_tty->print("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT "\n",
233 min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
234 }
235 }
236
237 size_t ThreadLocalAllocBuffer::initial_desired_size() {
238 size_t init_sz = 0;
239
240 if (TLABSize > 0) {
241 init_sz = TLABSize / HeapWordSize;
242 } else if (global_stats() != NULL) {
243 // Initial size is a function of the average number of allocating threads.
244 unsigned nof_threads = global_stats()->allocating_threads_avg();
245
246 init_sz = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
247 (nof_threads * target_refills());
248 init_sz = align_object_size(init_sz);
249 }
250 init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
251 return init_sz;
252 }
253
254 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
255 Thread* thrd = myThread();
256 size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
257 size_t alloc = _number_of_refills * _desired_size;
258 double waste_percent = alloc == 0 ? 0.0 :
259 100.0 * waste / alloc;
260 size_t tlab_used = Universe::heap()->tlab_used(thrd);
261 gclog_or_tty->print("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
262 " desired_size: " SIZE_FORMAT "KB"
263 " slow allocs: %d refill waste: " SIZE_FORMAT "B"
264 " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
265 " slow: %dB fast: %dB\n",
266 tag, thrd, thrd->osthread()->thread_id(),
267 _desired_size / (K / HeapWordSize),
268 _slow_allocations, _refill_waste_limit * HeapWordSize,
269 _allocation_fraction.average(),
270 _allocation_fraction.average() * tlab_used / K,
271 _number_of_refills, waste_percent,
272 _gc_waste * HeapWordSize,
273 _slow_refill_waste * HeapWordSize,
274 _fast_refill_waste * HeapWordSize);
275 }
276
277 void ThreadLocalAllocBuffer::verify() {
278 HeapWord* p = start();
279 HeapWord* t = top();
280 HeapWord* prev_p = NULL;
281 while (p < t) {
282 oop(p)->verify();
283 prev_p = p;
284 p += oop(p)->size();
285 }
286 guarantee(p == top(), "end of last object must match end of space");
287 }
288
289 Thread* ThreadLocalAllocBuffer::myThread() {
290 return (Thread*)(((char *)this) +
291 in_bytes(start_offset()) -
292 in_bytes(Thread::tlab_start_offset()));
293 }
294
295
296 GlobalTLABStats::GlobalTLABStats() :
297 _allocating_threads_avg(TLABAllocationWeight) {
298
299 initialize();
300
301 _allocating_threads_avg.sample(1); // One allocating thread at startup
302
303 if (UsePerfData) {
304
305 EXCEPTION_MARK;
306 ResourceMark rm;
307
308 char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
309 _perf_allocating_threads =
310 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
311
312 cname = PerfDataManager::counter_name("tlab", "fills");
|
36 // Thread-Local Edens support
37
38 // static member initialization
39 size_t ThreadLocalAllocBuffer::_max_size = 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 (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->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 if (PrintTLAB) {
62 global_stats()->print();
63 }
64 }
65 }
66
67 void ThreadLocalAllocBuffer::accumulate_statistics() {
68 Thread* thread = myThread();
69 size_t capacity = Universe::heap()->tlab_capacity(thread);
70 size_t used = Universe::heap()->tlab_used(thread);
71
72 _gc_waste += (unsigned)remaining();
73 size_t total_allocated = _gclab ? thread->allocated_bytes_gclab() : thread->allocated_bytes();
74 size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
75 _allocated_before_last_gc = total_allocated;
76
77 if (PrintTLAB && (_number_of_refills > 0 || Verbose)) {
78 print_stats("gc");
79 }
80
81 if (_number_of_refills > 0) {
82 // Update allocation history if a reasonable amount of eden was allocated.
83 bool update_allocation_history = used > 0.5 * capacity;
84
85 if (update_allocation_history) {
86 // Average the fraction of eden allocated in a tlab by this
87 // thread for use in the next resize operation.
88 // _gc_waste is not subtracted because it's included in
89 // "used".
90 // The result can be larger than 1.0 due to direct to old allocations.
91 // These allocations should ideally not be counted but since it is not possible
92 // to filter them out here we just cap the fraction to be at most 1.0.
93 double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
100 global_stats()->update_slow_refill_waste(_slow_refill_waste);
101 global_stats()->update_fast_refill_waste(_fast_refill_waste);
102
103 } else {
104 assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
105 _slow_refill_waste == 0 && _gc_waste == 0,
106 "tlab stats == 0");
107 }
108 global_stats()->update_slow_allocations(_slow_allocations);
109 }
110
111 // Fills the current tlab with a dummy filler array to create
112 // an illusion of a contiguous Eden and optionally retires the tlab.
113 // Waste accounting should be done in caller as appropriate; see,
114 // for example, clear_before_allocation().
115 void ThreadLocalAllocBuffer::make_parsable(bool retire) {
116 if (end() != NULL) {
117 invariants();
118
119 if (retire) {
120 if (_gclab) {
121 myThread()->incr_allocated_bytes_gclab(used_bytes());
122 } else {
123 myThread()->incr_allocated_bytes(used_bytes());
124 }
125 }
126
127 CollectedHeap::fill_with_object(top(), hard_end(), retire);
128
129 if (retire || ZeroTLAB) { // "Reset" the TLAB
130 set_start(NULL);
131 set_top(NULL);
132 set_pf_top(NULL);
133 set_end(NULL);
134 }
135 }
136 assert(!(retire || ZeroTLAB) ||
137 (start() == NULL && end() == NULL && top() == NULL),
138 "TLAB must be reset");
139 }
140
141 void ThreadLocalAllocBuffer::resize_all_tlabs() {
142 if (ResizeTLAB) {
143 for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
144 thread->tlab().resize();
182 if (PrintTLAB && Verbose) {
183 print_stats("fill");
184 }
185 assert(top <= start + new_size - alignment_reserve(), "size too small");
186 initialize(start, top, start + new_size - alignment_reserve());
187
188 // Reset amount of internal fragmentation
189 set_refill_waste_limit(initial_refill_waste_limit());
190 }
191
192 void ThreadLocalAllocBuffer::initialize(HeapWord* start,
193 HeapWord* top,
194 HeapWord* end) {
195 set_start(start);
196 set_top(top);
197 set_pf_top(top);
198 set_end(end);
199 invariants();
200 }
201
202 void ThreadLocalAllocBuffer::initialize(bool gclab) {
203 _initialized = true;
204 _gclab = gclab;
205 initialize(NULL, // start
206 NULL, // top
207 NULL); // end
208
209 set_desired_size(initial_desired_size());
210
211 // Following check is needed because at startup the main
212 // thread is initialized before the heap is. The initialization for
213 // this thread is redone in startup_initialization below.
214 if (Universe::heap() != NULL) {
215 size_t capacity = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
216 double alloc_frac = desired_size() * target_refills() / (double) capacity;
217 _allocation_fraction.sample(alloc_frac);
218 }
219
220 set_refill_waste_limit(initial_refill_waste_limit());
221
222 initialize_statistics();
223 }
224
225 void ThreadLocalAllocBuffer::startup_initialization() {
226
227 // Assuming each thread's active tlab is, on average,
228 // 1/2 full at a GC
229 _target_refills = 100 / (2 * TLABWasteTargetPercent);
230 _target_refills = MAX2(_target_refills, (unsigned)1U);
231
232 _global_stats = new GlobalTLABStats();
233
234 // During jvm startup, the main thread is initialized
235 // before the heap is initialized. So reinitialize it now.
236 guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
237 Thread::current()->tlab().initialize(false);
238 if (UseShenandoahGC) {
239 Thread::current()->gclab().initialize(true);
240 }
241
242 if (PrintTLAB && Verbose) {
243 gclog_or_tty->print("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT "\n",
244 min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
245 }
246 }
247
248 size_t ThreadLocalAllocBuffer::initial_desired_size() {
249 size_t init_sz = 0;
250
251 if (TLABSize > 0) {
252 init_sz = TLABSize / HeapWordSize;
253 } else if (global_stats() != NULL) {
254 // Initial size is a function of the average number of allocating threads.
255 unsigned nof_threads = global_stats()->allocating_threads_avg();
256
257 init_sz = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
258 (nof_threads * target_refills());
259 init_sz = align_object_size(init_sz);
260 }
261 init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
262 return init_sz;
263 }
264
265 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
266 Thread* thrd = myThread();
267 size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
268 size_t alloc = _number_of_refills * _desired_size;
269 double waste_percent = alloc == 0 ? 0.0 :
270 100.0 * waste / alloc;
271 size_t tlab_used = Universe::heap()->tlab_used(thrd);
272 gclog_or_tty->print("TLAB: %s %s thread: " INTPTR_FORMAT " [id: %2d]"
273 " desired_size: " SIZE_FORMAT "KB"
274 " slow allocs: %d refill waste: " SIZE_FORMAT "B"
275 " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
276 " slow: %dB fast: %dB\n",
277 tag, _gclab ? "gclab" : "tlab ", p2i(thrd), thrd->osthread()->thread_id(),
278 _desired_size / (K / HeapWordSize),
279 _slow_allocations, _refill_waste_limit * HeapWordSize,
280 _allocation_fraction.average(),
281 _allocation_fraction.average() * tlab_used / K,
282 _number_of_refills, waste_percent,
283 _gc_waste * HeapWordSize,
284 _slow_refill_waste * HeapWordSize,
285 _fast_refill_waste * HeapWordSize);
286 }
287
288 void ThreadLocalAllocBuffer::verify() {
289 HeapWord* p = start();
290 HeapWord* t = top();
291 HeapWord* prev_p = NULL;
292 while (p < t) {
293 oop(p)->verify();
294 prev_p = p;
295 p += oop(p)->size();
296 }
297 guarantee(p == top(), "end of last object must match end of space");
298 }
299
300 Thread* ThreadLocalAllocBuffer::myThread() {
301 ByteSize gclab_offset = Thread::gclab_start_offset();
302 ByteSize tlab_offset = Thread::tlab_start_offset();
303 ByteSize offs = _gclab ? gclab_offset : tlab_offset;
304 Thread* thread = (Thread*)(((char *)this) +
305 in_bytes(start_offset()) - in_bytes(offs));
306 #ifdef ASSERT
307 assert(this == (_gclab ? &thread->gclab() : &thread->tlab()), "must be");
308 #endif
309 return thread;
310 }
311
312 void ThreadLocalAllocBuffer::rollback(size_t size) {
313 HeapWord* old_top = top();
314 if (old_top != NULL) { // Pathological case: we accept that we can't rollback.
315 set_top(old_top - size);
316 }
317 }
318
319
320 GlobalTLABStats::GlobalTLABStats() :
321 _allocating_threads_avg(TLABAllocationWeight) {
322
323 initialize();
324
325 _allocating_threads_avg.sample(1); // One allocating thread at startup
326
327 if (UsePerfData) {
328
329 EXCEPTION_MARK;
330 ResourceMark rm;
331
332 char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
333 _perf_allocating_threads =
334 PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
335
336 cname = PerfDataManager::counter_name("tlab", "fills");
|