aoqi@0: /*
aoqi@0:  * Copyright (c) 1999, 2014, Oracle and/or its affiliates. All rights reserved.
aoqi@0:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
aoqi@0:  *
aoqi@0:  * This code is free software; you can redistribute it and/or modify it
aoqi@0:  * under the terms of the GNU General Public License version 2 only, as
aoqi@0:  * published by the Free Software Foundation.
aoqi@0:  *
aoqi@0:  * This code is distributed in the hope that it will be useful, but WITHOUT
aoqi@0:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
aoqi@0:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
aoqi@0:  * version 2 for more details (a copy is included in the LICENSE file that
aoqi@0:  * accompanied this code).
aoqi@0:  *
aoqi@0:  * You should have received a copy of the GNU General Public License version
aoqi@0:  * 2 along with this work; if not, write to the Free Software Foundation,
aoqi@0:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
aoqi@0:  *
aoqi@0:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
aoqi@0:  * or visit www.oracle.com if you need additional information or have any
aoqi@0:  * questions.
aoqi@0:  *
aoqi@0:  */
aoqi@0: 
aoqi@0: #include "precompiled.hpp"
aoqi@0: #include "memory/genCollectedHeap.hpp"
aoqi@0: #include "memory/resourceArea.hpp"
aoqi@0: #include "memory/threadLocalAllocBuffer.inline.hpp"
aoqi@0: #include "memory/universe.inline.hpp"
aoqi@0: #include "oops/oop.inline.hpp"
aoqi@0: #include "runtime/thread.inline.hpp"
aoqi@0: #include "utilities/copy.hpp"
aoqi@0: 
aoqi@0: PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
aoqi@0: 
aoqi@0: // Thread-Local Edens support
aoqi@0: 
aoqi@0: // static member initialization
aoqi@0: size_t           ThreadLocalAllocBuffer::_max_size       = 0;
aoqi@0: unsigned         ThreadLocalAllocBuffer::_target_refills = 0;
aoqi@0: GlobalTLABStats* ThreadLocalAllocBuffer::_global_stats   = NULL;
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::clear_before_allocation() {
aoqi@0:   _slow_refill_waste += (unsigned)remaining();
aoqi@0:   make_parsable(true);   // also retire the TLAB
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::accumulate_statistics_before_gc() {
aoqi@0:   global_stats()->initialize();
aoqi@0: 
aoqi@0:   for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
aoqi@0:     thread->tlab().accumulate_statistics();
aoqi@0:     thread->tlab().initialize_statistics();
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Publish new stats if some allocation occurred.
aoqi@0:   if (global_stats()->allocation() != 0) {
aoqi@0:     global_stats()->publish();
aoqi@0:     if (PrintTLAB) {
aoqi@0:       global_stats()->print();
aoqi@0:     }
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::accumulate_statistics() {
aoqi@0:   Thread* thread = myThread();
aoqi@0:   size_t capacity = Universe::heap()->tlab_capacity(thread);
aoqi@0:   size_t used     = Universe::heap()->tlab_used(thread);
aoqi@0: 
aoqi@0:   _gc_waste += (unsigned)remaining();
aoqi@0:   size_t total_allocated = thread->allocated_bytes();
aoqi@0:   size_t allocated_since_last_gc = total_allocated - _allocated_before_last_gc;
aoqi@0:   _allocated_before_last_gc = total_allocated;
aoqi@0: 
aoqi@0:   if (PrintTLAB && (_number_of_refills > 0 || Verbose)) {
aoqi@0:     print_stats("gc");
aoqi@0:   }
aoqi@0: 
aoqi@0:   if (_number_of_refills > 0) {
aoqi@0:     // Update allocation history if a reasonable amount of eden was allocated.
aoqi@0:     bool update_allocation_history = used > 0.5 * capacity;
aoqi@0: 
aoqi@0:     if (update_allocation_history) {
aoqi@0:       // Average the fraction of eden allocated in a tlab by this
aoqi@0:       // thread for use in the next resize operation.
aoqi@0:       // _gc_waste is not subtracted because it's included in
aoqi@0:       // "used".
aoqi@0:       // The result can be larger than 1.0 due to direct to old allocations.
aoqi@0:       // These allocations should ideally not be counted but since it is not possible
aoqi@0:       // to filter them out here we just cap the fraction to be at most 1.0.
aoqi@0:       double alloc_frac = MIN2(1.0, (double) allocated_since_last_gc / used);
aoqi@0:       _allocation_fraction.sample(alloc_frac);
aoqi@0:     }
aoqi@0:     global_stats()->update_allocating_threads();
aoqi@0:     global_stats()->update_number_of_refills(_number_of_refills);
aoqi@0:     global_stats()->update_allocation(_number_of_refills * desired_size());
aoqi@0:     global_stats()->update_gc_waste(_gc_waste);
aoqi@0:     global_stats()->update_slow_refill_waste(_slow_refill_waste);
aoqi@0:     global_stats()->update_fast_refill_waste(_fast_refill_waste);
aoqi@0: 
aoqi@0:   } else {
aoqi@0:     assert(_number_of_refills == 0 && _fast_refill_waste == 0 &&
aoqi@0:            _slow_refill_waste == 0 && _gc_waste          == 0,
aoqi@0:            "tlab stats == 0");
aoqi@0:   }
aoqi@0:   global_stats()->update_slow_allocations(_slow_allocations);
aoqi@0: }
aoqi@0: 
aoqi@0: // Fills the current tlab with a dummy filler array to create
aoqi@0: // an illusion of a contiguous Eden and optionally retires the tlab.
aoqi@0: // Waste accounting should be done in caller as appropriate; see,
aoqi@0: // for example, clear_before_allocation().
aoqi@0: void ThreadLocalAllocBuffer::make_parsable(bool retire) {
aoqi@0:   if (end() != NULL) {
aoqi@0:     invariants();
aoqi@0: 
aoqi@0:     if (retire) {
aoqi@0:       myThread()->incr_allocated_bytes(used_bytes());
aoqi@0:     }
aoqi@0: 
aoqi@0:     CollectedHeap::fill_with_object(top(), hard_end(), retire);
aoqi@0: 
aoqi@0:     if (retire || ZeroTLAB) {  // "Reset" the TLAB
aoqi@0:       set_start(NULL);
aoqi@0:       set_top(NULL);
aoqi@0:       set_pf_top(NULL);
aoqi@0:       set_end(NULL);
aoqi@0:     }
aoqi@0:   }
aoqi@0:   assert(!(retire || ZeroTLAB)  ||
aoqi@0:          (start() == NULL && end() == NULL && top() == NULL),
aoqi@0:          "TLAB must be reset");
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::resize_all_tlabs() {
aoqi@0:   if (ResizeTLAB) {
aoqi@0:     for (JavaThread *thread = Threads::first(); thread != NULL; thread = thread->next()) {
aoqi@0:       thread->tlab().resize();
aoqi@0:     }
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::resize() {
aoqi@0:   // Compute the next tlab size using expected allocation amount
aoqi@0:   assert(ResizeTLAB, "Should not call this otherwise");
aoqi@0:   size_t alloc = (size_t)(_allocation_fraction.average() *
aoqi@0:                           (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize));
aoqi@0:   size_t new_size = alloc / _target_refills;
aoqi@0: 
aoqi@0:   new_size = MIN2(MAX2(new_size, min_size()), max_size());
aoqi@0: 
aoqi@0:   size_t aligned_new_size = align_object_size(new_size);
aoqi@0: 
aoqi@0:   if (PrintTLAB && Verbose) {
aoqi@0:     gclog_or_tty->print("TLAB new size: thread: " INTPTR_FORMAT " [id: %2d]"
aoqi@0:                         " refills %d  alloc: %8.6f desired_size: " SIZE_FORMAT " -> " SIZE_FORMAT "\n",
aoqi@0:                         myThread(), myThread()->osthread()->thread_id(),
aoqi@0:                         _target_refills, _allocation_fraction.average(), desired_size(), aligned_new_size);
aoqi@0:   }
aoqi@0:   set_desired_size(aligned_new_size);
aoqi@0:   set_refill_waste_limit(initial_refill_waste_limit());
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::initialize_statistics() {
aoqi@0:     _number_of_refills = 0;
aoqi@0:     _fast_refill_waste = 0;
aoqi@0:     _slow_refill_waste = 0;
aoqi@0:     _gc_waste          = 0;
aoqi@0:     _slow_allocations  = 0;
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::fill(HeapWord* start,
aoqi@0:                                   HeapWord* top,
aoqi@0:                                   size_t    new_size) {
aoqi@0:   _number_of_refills++;
aoqi@0:   if (PrintTLAB && Verbose) {
aoqi@0:     print_stats("fill");
aoqi@0:   }
aoqi@0:   assert(top <= start + new_size - alignment_reserve(), "size too small");
aoqi@0:   initialize(start, top, start + new_size - alignment_reserve());
aoqi@0: 
aoqi@0:   // Reset amount of internal fragmentation
aoqi@0:   set_refill_waste_limit(initial_refill_waste_limit());
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::initialize(HeapWord* start,
aoqi@0:                                         HeapWord* top,
aoqi@0:                                         HeapWord* end) {
aoqi@0:   set_start(start);
aoqi@0:   set_top(top);
aoqi@0:   set_pf_top(top);
aoqi@0:   set_end(end);
aoqi@0:   invariants();
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::initialize() {
aoqi@0:   initialize(NULL,                    // start
aoqi@0:              NULL,                    // top
aoqi@0:              NULL);                   // end
aoqi@0: 
aoqi@0:   set_desired_size(initial_desired_size());
aoqi@0: 
aoqi@0:   // Following check is needed because at startup the main (primordial)
aoqi@0:   // thread is initialized before the heap is.  The initialization for
aoqi@0:   // this thread is redone in startup_initialization below.
aoqi@0:   if (Universe::heap() != NULL) {
aoqi@0:     size_t capacity   = Universe::heap()->tlab_capacity(myThread()) / HeapWordSize;
aoqi@0:     double alloc_frac = desired_size() * target_refills() / (double) capacity;
aoqi@0:     _allocation_fraction.sample(alloc_frac);
aoqi@0:   }
aoqi@0: 
aoqi@0:   set_refill_waste_limit(initial_refill_waste_limit());
aoqi@0: 
aoqi@0:   initialize_statistics();
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::startup_initialization() {
aoqi@0: 
aoqi@0:   // Assuming each thread's active tlab is, on average,
aoqi@0:   // 1/2 full at a GC
aoqi@0:   _target_refills = 100 / (2 * TLABWasteTargetPercent);
aoqi@0:   _target_refills = MAX2(_target_refills, (unsigned)1U);
aoqi@0: 
aoqi@0:   _global_stats = new GlobalTLABStats();
aoqi@0: 
aoqi@0:   // During jvm startup, the main (primordial) thread is initialized
aoqi@0:   // before the heap is initialized.  So reinitialize it now.
aoqi@0:   guarantee(Thread::current()->is_Java_thread(), "tlab initialization thread not Java thread");
aoqi@0:   Thread::current()->tlab().initialize();
aoqi@0: 
aoqi@0:   if (PrintTLAB && Verbose) {
aoqi@0:     gclog_or_tty->print("TLAB min: " SIZE_FORMAT " initial: " SIZE_FORMAT " max: " SIZE_FORMAT "\n",
aoqi@0:                         min_size(), Thread::current()->tlab().initial_desired_size(), max_size());
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: size_t ThreadLocalAllocBuffer::initial_desired_size() {
aoqi@0:   size_t init_sz;
aoqi@0: 
aoqi@0:   if (TLABSize > 0) {
aoqi@0:     init_sz = MIN2(TLABSize / HeapWordSize, max_size());
aoqi@0:   } else if (global_stats() == NULL) {
aoqi@0:     // Startup issue - main thread initialized before heap initialized.
aoqi@0:     init_sz = min_size();
aoqi@0:   } else {
aoqi@0:     // Initial size is a function of the average number of allocating threads.
aoqi@0:     unsigned nof_threads = global_stats()->allocating_threads_avg();
aoqi@0: 
aoqi@0:     init_sz  = (Universe::heap()->tlab_capacity(myThread()) / HeapWordSize) /
aoqi@0:                       (nof_threads * target_refills());
aoqi@0:     init_sz = align_object_size(init_sz);
aoqi@0:     init_sz = MIN2(MAX2(init_sz, min_size()), max_size());
aoqi@0:   }
aoqi@0:   return init_sz;
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::print_stats(const char* tag) {
aoqi@0:   Thread* thrd = myThread();
aoqi@0:   size_t waste = _gc_waste + _slow_refill_waste + _fast_refill_waste;
aoqi@0:   size_t alloc = _number_of_refills * _desired_size;
aoqi@0:   double waste_percent = alloc == 0 ? 0.0 :
aoqi@0:                       100.0 * waste / alloc;
aoqi@0:   size_t tlab_used  = Universe::heap()->tlab_used(thrd);
aoqi@0:   gclog_or_tty->print("TLAB: %s thread: " INTPTR_FORMAT " [id: %2d]"
aoqi@0:                       " desired_size: " SIZE_FORMAT "KB"
aoqi@0:                       " slow allocs: %d  refill waste: " SIZE_FORMAT "B"
aoqi@0:                       " alloc:%8.5f %8.0fKB refills: %d waste %4.1f%% gc: %dB"
aoqi@0:                       " slow: %dB fast: %dB\n",
aoqi@0:                       tag, thrd, thrd->osthread()->thread_id(),
aoqi@0:                       _desired_size / (K / HeapWordSize),
aoqi@0:                       _slow_allocations, _refill_waste_limit * HeapWordSize,
aoqi@0:                       _allocation_fraction.average(),
aoqi@0:                       _allocation_fraction.average() * tlab_used / K,
aoqi@0:                       _number_of_refills, waste_percent,
aoqi@0:                       _gc_waste * HeapWordSize,
aoqi@0:                       _slow_refill_waste * HeapWordSize,
aoqi@0:                       _fast_refill_waste * HeapWordSize);
aoqi@0: }
aoqi@0: 
aoqi@0: void ThreadLocalAllocBuffer::verify() {
aoqi@0:   HeapWord* p = start();
aoqi@0:   HeapWord* t = top();
aoqi@0:   HeapWord* prev_p = NULL;
aoqi@0:   while (p < t) {
aoqi@0:     oop(p)->verify();
aoqi@0:     prev_p = p;
aoqi@0:     p += oop(p)->size();
aoqi@0:   }
aoqi@0:   guarantee(p == top(), "end of last object must match end of space");
aoqi@0: }
aoqi@0: 
aoqi@0: Thread* ThreadLocalAllocBuffer::myThread() {
aoqi@0:   return (Thread*)(((char *)this) +
aoqi@0:                    in_bytes(start_offset()) -
aoqi@0:                    in_bytes(Thread::tlab_start_offset()));
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: GlobalTLABStats::GlobalTLABStats() :
aoqi@0:   _allocating_threads_avg(TLABAllocationWeight) {
aoqi@0: 
aoqi@0:   initialize();
aoqi@0: 
aoqi@0:   _allocating_threads_avg.sample(1); // One allocating thread at startup
aoqi@0: 
aoqi@0:   if (UsePerfData) {
aoqi@0: 
aoqi@0:     EXCEPTION_MARK;
aoqi@0:     ResourceMark rm;
aoqi@0: 
aoqi@0:     char* cname = PerfDataManager::counter_name("tlab", "allocThreads");
aoqi@0:     _perf_allocating_threads =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "fills");
aoqi@0:     _perf_total_refills =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "maxFills");
aoqi@0:     _perf_max_refills =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "alloc");
aoqi@0:     _perf_allocation =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "gcWaste");
aoqi@0:     _perf_gc_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "maxGcWaste");
aoqi@0:     _perf_max_gc_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "slowWaste");
aoqi@0:     _perf_slow_refill_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "maxSlowWaste");
aoqi@0:     _perf_max_slow_refill_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "fastWaste");
aoqi@0:     _perf_fast_refill_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "maxFastWaste");
aoqi@0:     _perf_max_fast_refill_waste =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_Bytes, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "slowAlloc");
aoqi@0:     _perf_slow_allocations =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
aoqi@0: 
aoqi@0:     cname = PerfDataManager::counter_name("tlab", "maxSlowAlloc");
aoqi@0:     _perf_max_slow_allocations =
aoqi@0:       PerfDataManager::create_variable(SUN_GC, cname, PerfData::U_None, CHECK);
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: void GlobalTLABStats::initialize() {
aoqi@0:   // Clear counters summarizing info from all threads
aoqi@0:   _allocating_threads      = 0;
aoqi@0:   _total_refills           = 0;
aoqi@0:   _max_refills             = 0;
aoqi@0:   _total_allocation        = 0;
aoqi@0:   _total_gc_waste          = 0;
aoqi@0:   _max_gc_waste            = 0;
aoqi@0:   _total_slow_refill_waste = 0;
aoqi@0:   _max_slow_refill_waste   = 0;
aoqi@0:   _total_fast_refill_waste = 0;
aoqi@0:   _max_fast_refill_waste   = 0;
aoqi@0:   _total_slow_allocations  = 0;
aoqi@0:   _max_slow_allocations    = 0;
aoqi@0: }
aoqi@0: 
aoqi@0: void GlobalTLABStats::publish() {
aoqi@0:   _allocating_threads_avg.sample(_allocating_threads);
aoqi@0:   if (UsePerfData) {
aoqi@0:     _perf_allocating_threads   ->set_value(_allocating_threads);
aoqi@0:     _perf_total_refills        ->set_value(_total_refills);
aoqi@0:     _perf_max_refills          ->set_value(_max_refills);
aoqi@0:     _perf_allocation           ->set_value(_total_allocation);
aoqi@0:     _perf_gc_waste             ->set_value(_total_gc_waste);
aoqi@0:     _perf_max_gc_waste         ->set_value(_max_gc_waste);
aoqi@0:     _perf_slow_refill_waste    ->set_value(_total_slow_refill_waste);
aoqi@0:     _perf_max_slow_refill_waste->set_value(_max_slow_refill_waste);
aoqi@0:     _perf_fast_refill_waste    ->set_value(_total_fast_refill_waste);
aoqi@0:     _perf_max_fast_refill_waste->set_value(_max_fast_refill_waste);
aoqi@0:     _perf_slow_allocations     ->set_value(_total_slow_allocations);
aoqi@0:     _perf_max_slow_allocations ->set_value(_max_slow_allocations);
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: void GlobalTLABStats::print() {
aoqi@0:   size_t waste = _total_gc_waste + _total_slow_refill_waste + _total_fast_refill_waste;
aoqi@0:   double waste_percent = _total_allocation == 0 ? 0.0 :
aoqi@0:                          100.0 * waste / _total_allocation;
aoqi@0:   gclog_or_tty->print("TLAB totals: thrds: %d  refills: %d max: %d"
aoqi@0:                       " slow allocs: %d max %d waste: %4.1f%%"
aoqi@0:                       " gc: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
aoqi@0:                       " slow: " SIZE_FORMAT "B max: " SIZE_FORMAT "B"
aoqi@0:                       " fast: " SIZE_FORMAT "B max: " SIZE_FORMAT "B\n",
aoqi@0:                       _allocating_threads,
aoqi@0:                       _total_refills, _max_refills,
aoqi@0:                       _total_slow_allocations, _max_slow_allocations,
aoqi@0:                       waste_percent,
aoqi@0:                       _total_gc_waste * HeapWordSize,
aoqi@0:                       _max_gc_waste * HeapWordSize,
aoqi@0:                       _total_slow_refill_waste * HeapWordSize,
aoqi@0:                       _max_slow_refill_waste * HeapWordSize,
aoqi@0:                       _total_fast_refill_waste * HeapWordSize,
aoqi@0:                       _max_fast_refill_waste * HeapWordSize);
aoqi@0: }