1 /*
   2  * Copyright (c) 1999, 2015, 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/heapMonitoring.hpp"
  33 #include "runtime/thread.inline.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 (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
  53     thread->tlab().accumulate_statistics();
  54     thread->tlab().initialize_statistics();
  55   }
  56 
  57   // Publish new stats if some allocation occurred.
  58   if (global_stats()->allocation() != 0) {
  59     global_stats()->publish();
  60     global_stats()->print();
  61   }
  62 }
  63 
  64 void ThreadLocalAllocBuffer::accumulate_statistics() {
  65   Thread* thread = myThread();
  66   size_t capacity = Universe::heap()->tlab_capacity(thread);
  67   size_t used     = Universe::heap()->tlab_used(thread);
  68 
  69   _gc_waste += (unsigned)remaining();
  70   size_t total_allocated = thread->allocated_bytes();
  71   size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
  72   _allocated_before_last_gc = total_allocated;
  73 
  74   print_stats("gc");
  75 
  76   if (_number_of_refills > 0) {
  77     // Update allocation history if a reasonable amount of eden was allocated.
  78     bool update_allocation_history = used > 0.5 * capacity;
  79 
  80     if (update_allocation_history) {
  81       // Average the fraction of eden allocated in a tlab by this
  82       // thread for use in the next resize operation.
  83       // _gc_waste is not subtracted because it's included in
  84       // "used".
  85       // The result can be larger than 1.0 due to direct to old allocations.
  86       // These allocations should ideally not be counted but since it is not possible
  87       // to filter them out here we just cap the fraction to be at most 1.0.
  88       double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
  89       _allocation_fraction.sample(alloc_frac);
  90     }
  91     global_stats()->update_allocating_threads();
  92     global_stats()->update_number_of_refills(_number_of_refills);
  93     global_stats()->update_allocation(_number_of_refills * desired_size());
  94     global_stats()->update_gc_waste(_gc_waste);
  95     global_stats()->update_slow_refill_waste(_slow_refill_waste);
  96     global_stats()->update_fast_refill_waste(_fast_refill_waste);
  97 
  98   } else {
  99     assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
 100            _slow_refill_waste == 0 && _gc_waste          == 0,
 101            "tlab stats == 0");
 102   }
 103   global_stats()->update_slow_allocations(_slow_allocations);
 104 }
 105 
 106 // Fills the current tlab with a dummy filler array to create
 107 // an illusion of a contiguous Eden and optionally retires the tlab.
 108 // Waste accounting should be done in caller as appropriate; see,
 109 // for example, clear_before_allocation().
 110 void ThreadLocalAllocBuffer::make_parsable(bool retire, bool zap) {
 111   if (end() != NULL) {
 112     invariants();
 113 
 114     if (retire) {
 115       myThread()->incr_allocated_bytes(used_bytes());
 116     }
 117 
 118     CollectedHeap::fill_with_object(top(), hard_end(), retire && zap);
 119 
 120     if (retire || ZeroTLAB) {  // "Reset" the TLAB
 121       set_start(NULL);
 122       set_top(NULL);
 123       set_pf_top(NULL);
 124       set_end(NULL);
 125       set_actual_end(NULL);
 126       set_slow_path_end(NULL);
 127     }
 128   }
 129   assert(!(retire || ZeroTLAB)  ||
 130          (start() == NULL && end() == NULL && top() == NULL &&
 131           actual_end() == NULL && slow_path_end() == 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();
 139     }
 140   }
 141 }
 142 
 143 void ThreadLocalAllocBuffer::resize() {
 144   // Compute the next tlab size using expected allocation amount
 145   assert(ResizeTLAB, "Should not call this otherwise");
 146   size_t alloc = (size_t)(_allocation_fraction.average() *
 147                           (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize));
 148   size_t new_size = alloc / _target_refills;
 149 
 150   new_size = MIN2(MAX2(new_size, min_size()), max_size());
 151 
 152   size_t aligned_new_size = align_object_size(new_size);
 153 
 154   log_trace(gc, tlab)("TLAB new size: thread: " INTPTR_FORMAT " [id: %2d]"
 155                       " refills %d  alloc: %8.6f desired_size: " SIZE_FORMAT " -> " SIZE_FORMAT,
 156                       p2i(myThread()), myThread()->osthread()->thread_id(),
 157                       _target_refills, _allocation_fraction.average(), desired_size(), aligned_new_size);
 158 
 159   set_desired_size(aligned_new_size);
 160   set_refill_waste_limit(initial_refill_waste_limit());
 161 }
 162 
 163 void ThreadLocalAllocBuffer::initialize_statistics() {
 164     _number_of_refills = 0;
 165     _fast_refill_waste = 0;
 166     _slow_refill_waste = 0;
 167     _gc_waste          = 0;
 168     _slow_allocations  = 0;
 169 }
 170 
 171 void ThreadLocalAllocBuffer::fill(HeapWord* start,
 172                                   HeapWord* top,
 173                                   size_t    new_size) {
 174   _number_of_refills++;
 175   print_stats("fill");
 176   assert(top <= start + new_size - alignment_reserve(), "size too small");
 177 
 178   // Remember old bytes until sample for the next tlab only if this is our first
 179   // actual refill.
 180   size_t old_bytes_until_sample = 0;
 181   if (_number_of_refills > 1) {
 182     old_bytes_until_sample = bytes_until_sample();
 183   }
 184 
 185   initialize(start, top, start + new_size - alignment_reserve());
 186 
 187   if (old_bytes_until_sample > 0) {
 188     set_bytes_until_sample(old_bytes_until_sample);
 189     set_sample_end();
 190   }
 191 
 192   // Reset amount of internal fragmentation
 193   set_refill_waste_limit(initial_refill_waste_limit());
 194 }
 195 
 196 void ThreadLocalAllocBuffer::initialize(HeapWord* start,
 197                                         HeapWord* top,
 198                                         HeapWord* end) {
 199   set_start(start);
 200   set_top(top);
 201   set_pf_top(top);
 202   set_end(end);
 203   set_actual_end(end);
 204   set_slow_path_end(end);
 205   invariants();
 206   _bytes_until_sample = 0;
 207 }
 208 
 209 void ThreadLocalAllocBuffer::initialize() {
 210   initialize(NULL,                    // start
 211              NULL,                    // top
 212              NULL);                   // end
 213 
 214   set_desired_size(initial_desired_size());
 215 
 216   // Following check is needed because at startup the main (primordial)
 217   // thread is initialized before the heap is.  The initialization for
 218   // this thread is redone in startup_initialization below.
 219   if (Universe::heap() != NULL) {
 220     size_t capacity   = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
 221     double alloc_frac = desired_size() * target_refills() / (double) capacity;
 222     _allocation_fraction.sample(alloc_frac);
 223   }
 224 
 225   set_refill_waste_limit(initial_refill_waste_limit());
 226 
 227   initialize_statistics();
 228 }
 229 
 230 void ThreadLocalAllocBuffer::startup_initialization() {
 231 
 232   // Assuming each thread's active tlab is, on average,
 233   // 1/2 full at a GC
 234   _target_refills = 100 / (2 * TLABWasteTargetPercent);
 235   _target_refills = MAX2(_target_refills, (unsigned)1U);
 236 
 237   _global_stats = new GlobalTLABStats();
 238 
 239 #ifdef COMPILER2
 240   // If the C2 compiler is present, extra space is needed at the end of
 241   // TLABs, otherwise prefetching instructions generated by the C2
 242   // compiler will fault (due to accessing memory outside of heap).
 243   // The amount of space is the max of the number of lines to
 244   // prefetch for array and for instance allocations. (Extra space must be
 245   // reserved to accommodate both types of allocations.)
 246   //
 247   // Only SPARC-specific BIS instructions are known to fault. (Those
 248   // instructions are generated if AllocatePrefetchStyle==3 and
 249   // AllocatePrefetchInstr==1). To be on the safe side, however,
 250   // extra space is reserved for all combinations of
 251   // AllocatePrefetchStyle and AllocatePrefetchInstr.
 252   //
 253   // If the C2 compiler is not present, no space is reserved.
 254 
 255   // +1 for rounding up to next cache line, +1 to be safe
 256   if (is_server_compilation_mode_vm()) {
 257     int lines =  MAX2(AllocatePrefetchLines, AllocateInstancePrefetchLines) + 2;
 258     _reserve_for_allocation_prefetch = (AllocatePrefetchDistance + AllocatePrefetchStepSize * lines) /
 259                                        (int)HeapWordSize;
 260   }
 261 #endif
 262 
 263   // During jvm startup, the main (primordial) thread is initialized
 264   // before the heap is initialized.  So reinitialize it now.
 265   guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
 266   Thread::current()->tlab().initialize();
 267 
 268   log_develop_trace(gc, tlab)("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT,
 269                                min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
 270 }
 271 
 272 size_t ThreadLocalAllocBuffer::initial_desired_size() {
 273   size_t init_sz = 0;
 274 
 275   if (TLABSize > 0) {
 276     init_sz = TLABSize / HeapWordSize;
 277   } else if (global_stats() != NULL) {
 278     // Initial size is a function of the average number of allocating threads.
 279     unsigned nof_threads = global_stats()->allocating_threads_avg();
 280 
 281     init_sz  = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
 282                       (nof_threads * target_refills());
 283     init_sz = align_object_size(init_sz);
 284   }
 285   init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
 286   return init_sz;
 287 }
 288 
 289 void ThreadLocalAllocBuffer::print_stats(const char* tag) {
 290   Log(gc, tlab) log;
 291   if (!log.is_trace()) {
 292     return;
 293   }
 294 
 295   Thread* thrd = myThread();
 296   size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
 297   size_t alloc = _number_of_refills * _desired_size;
 298   double waste_percent = alloc == 0 ? 0.0 :
 299                       100.0 * waste / alloc;
 300   size_t tlab_used  = Universe::heap()->tlab_used(thrd);
 301   log.trace("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
 302             " desired_size: " SIZE_FORMAT "KB"
 303             " slow allocs: %d  refill waste: " SIZE_FORMAT "B"
 304             " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
 305             " slow: %dB fast: %dB",
 306             tag, p2i(thrd), thrd->osthread()->thread_id(),
 307             _desired_size / (K / HeapWordSize),
 308             _slow_allocations, _refill_waste_limit * HeapWordSize,
 309             _allocation_fraction.average(),
 310             _allocation_fraction.average() * tlab_used / K,
 311             _number_of_refills, waste_percent,
 312             _gc_waste * HeapWordSize,
 313             _slow_refill_waste * HeapWordSize,
 314             _fast_refill_waste * HeapWordSize);
 315 }
 316 
 317 void ThreadLocalAllocBuffer::verify() {
 318   HeapWord* p = start();
 319   HeapWord* t = top();
 320   HeapWord* prev_p = NULL;
 321   while (p < t) {
 322     oop(p)->verify();
 323     prev_p = p;
 324     p += oop(p)->size();
 325   }
 326   guarantee(p == top(), "end of last object must match end of space");
 327 }
 328 
 329 void ThreadLocalAllocBuffer::set_sample_end() {
 330   size_t heap_words_remaining = _end - _top;
 331   size_t bytes_left = bytes_until_sample();
 332   size_t words_until_sample = bytes_left / HeapWordSize;
 333 
 334   if (heap_words_remaining > words_until_sample) {
 335     HeapWord* new_end = _top + words_until_sample;
 336     set_end(new_end);
 337     set_slow_path_end(new_end);
 338     set_bytes_until_sample(0);
 339   } else {
 340     bytes_left -= heap_words_remaining * HeapWordSize;
 341     set_bytes_until_sample(bytes_left);
 342   }
 343 }
 344 
 345 void ThreadLocalAllocBuffer::pick_next_sample(size_t diff) {
 346   if (!HeapMonitoring::enabled()) {
 347     return;
 348   }
 349 
 350   if (bytes_until_sample() == 0) {
 351     HeapMonitoring::pick_next_sample(bytes_until_sample_addr());
 352   }
 353 
 354   if (diff > 0) {
 355     // Try to correct sample size by removing extra space from last allocation.
 356     if (bytes_until_sample() > diff * HeapWordSize) {
 357       set_bytes_until_sample(bytes_until_sample() - diff * HeapWordSize);
 358     }
 359   }
 360 
 361   set_sample_end();
 362 
 363   log_trace(gc, tlab)("TLAB picked next sample: thread: " INTPTR_FORMAT " [id: %2d]"
 364                       " start: %p  top: %p end: %p actual_end: %p slow_path_end: %p",
 365                       p2i(myThread()), myThread()->osthread()->thread_id(),
 366                       start(), top(), end(),
 367                       actual_end(), slow_path_end());
 368 }
 369 
 370 Thread* ThreadLocalAllocBuffer::myThread() {
 371   return (Thread*)(((char *)this) +
 372                    in_bytes(start_offset()) -
 373                    in_bytes(Thread::tlab_start_offset()));
 374 }
 375 
 376 void ThreadLocalAllocBuffer::set_back_actual_end() {
 377   // Did a fast TLAB refill occur?
 378   if (_slow_path_end != _end) {
 379     // Fix up the actual end to be now the end of this TLAB.
 380     _slow_path_end = _end;
 381     _actual_end = _end;
 382   } else {
 383     _end = _actual_end;
 384   }
 385 }
 386 
 387 void ThreadLocalAllocBuffer::handle_sample(Thread* thread, HeapWord* result,
 388                                            size_t size) {
 389   if (!HeapMonitoring::enabled()) {
 390     return;
 391   }
 392 
 393   size_t size_in_bytes = size * HeapWordSize;
 394   if (bytes_until_sample() > size_in_bytes) {
 395     set_bytes_until_sample(bytes_until_sample() - size_in_bytes);
 396   } else {
 397     // Technically this is not exactly right, we probably should remember how many bytes are
 398     // negative probably to then reduce our next sample size.
 399     set_bytes_until_sample(0);
 400   }
 401 
 402   // Should we sample now?
 403   if (should_sample()) {
 404     HeapMonitoring::object_alloc_do_sample(thread,
 405                                            reinterpret_cast<oopDesc*>(result),
 406                                            size_in_bytes);
 407     set_back_actual_end();
 408     pick_next_sample();
 409   }
 410 }
 411 
 412 HeapWord* ThreadLocalAllocBuffer::hard_end() {
 413   // Did a fast TLAB refill occur?
 414   if (_slow_path_end != _end) {
 415     // Fix up the actual end to be now the end of this TLAB.
 416     _slow_path_end = _end;
 417     _actual_end = _end;
 418   }
 419 
 420   return _actual_end + alignment_reserve();
 421 }
 422 GlobalTLABStats::GlobalTLABStats() :
 423   _allocating_threads_avg(TLABAllocationWeight) {
 424 
 425   initialize();
 426 
 427   _allocating_threads_avg.sample(1); // One allocating thread at startup
 428 
 429   if (UsePerfData) {
 430 
 431     EXCEPTION_MARK;
 432     ResourceMark rm;
 433 
 434     char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
 435     _perf_allocating_threads =
 436       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
 437 
 438     cname = PerfDataManager::counter_name("tlab", "fills");
 439     _perf_total_refills =
 440       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
 441 
 442     cname = PerfDataManager::counter_name("tlab", "maxFills");
 443     _perf_max_refills =
 444       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
 445 
 446     cname = PerfDataManager::counter_name("tlab", "alloc");
 447     _perf_allocation =
 448       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 449 
 450     cname = PerfDataManager::counter_name("tlab", "gcWaste");
 451     _perf_gc_waste =
 452       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 453 
 454     cname = PerfDataManager::counter_name("tlab", "maxGcWaste");
 455     _perf_max_gc_waste =
 456       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 457 
 458     cname = PerfDataManager::counter_name("tlab", "slowWaste");
 459     _perf_slow_refill_waste =
 460       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 461 
 462     cname = PerfDataManager::counter_name("tlab", "maxSlowWaste");
 463     _perf_max_slow_refill_waste =
 464       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 465 
 466     cname = PerfDataManager::counter_name("tlab", "fastWaste");
 467     _perf_fast_refill_waste =
 468       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 469 
 470     cname = PerfDataManager::counter_name("tlab", "maxFastWaste");
 471     _perf_max_fast_refill_waste =
 472       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
 473 
 474     cname = PerfDataManager::counter_name("tlab", "slowAlloc");
 475     _perf_slow_allocations =
 476       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
 477 
 478     cname = PerfDataManager::counter_name("tlab", "maxSlowAlloc");
 479     _perf_max_slow_allocations =
 480       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
 481   }
 482 }
 483 
 484 void GlobalTLABStats::initialize() {
 485   // Clear counters summarizing info from all threads
 486   _allocating_threads      = 0;
 487   _total_refills           = 0;
 488   _max_refills             = 0;
 489   _total_allocation        = 0;
 490   _total_gc_waste          = 0;
 491   _max_gc_waste            = 0;
 492   _total_slow_refill_waste = 0;
 493   _max_slow_refill_waste   = 0;
 494   _total_fast_refill_waste = 0;
 495   _max_fast_refill_waste   = 0;
 496   _total_slow_allocations  = 0;
 497   _max_slow_allocations    = 0;
 498 }
 499 
 500 void GlobalTLABStats::publish() {
 501   _allocating_threads_avg.sample(_allocating_threads);
 502   if (UsePerfData) {
 503     _perf_allocating_threads   ->set_value(_allocating_threads);
 504     _perf_total_refills        ->set_value(_total_refills);
 505     _perf_max_refills          ->set_value(_max_refills);
 506     _perf_allocation           ->set_value(_total_allocation);
 507     _perf_gc_waste             ->set_value(_total_gc_waste);
 508     _perf_max_gc_waste         ->set_value(_max_gc_waste);
 509     _perf_slow_refill_waste    ->set_value(_total_slow_refill_waste);
 510     _perf_max_slow_refill_waste->set_value(_max_slow_refill_waste);
 511     _perf_fast_refill_waste    ->set_value(_total_fast_refill_waste);
 512     _perf_max_fast_refill_waste->set_value(_max_fast_refill_waste);
 513     _perf_slow_allocations     ->set_value(_total_slow_allocations);
 514     _perf_max_slow_allocations ->set_value(_max_slow_allocations);
 515   }
 516 }
 517 
 518 void GlobalTLABStats::print() {
 519   Log(gc, tlab) log;
 520   if (!log.is_debug()) {
 521     return;
 522   }
 523 
 524   size_t waste = _total_gc_waste + _total_slow_refill_waste + _total_fast_refill_waste;
 525   double waste_percent = _total_allocation == 0 ? 0.0 :
 526                          100.0 * waste / _total_allocation;
 527   log.debug("TLAB totals: thrds: %d  refills: %d max: %d"
 528             " slow allocs: %d max %d waste: %4.1f%%"
 529             " gc: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
 530             " slow: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
 531             " fast: " SIZE_FORMAT "B max: " SIZE_FORMAT "B",
 532             _allocating_threads,
 533             _total_refills, _max_refills,
 534             _total_slow_allocations, _max_slow_allocations,
 535             waste_percent,
 536             _total_gc_waste * HeapWordSize,
 537             _max_gc_waste * HeapWordSize,
 538             _total_slow_refill_waste * HeapWordSize,
 539             _max_slow_refill_waste * HeapWordSize,
 540             _total_fast_refill_waste * HeapWordSize,
 541             _max_fast_refill_waste * HeapWordSize);
 542 }