aoqi@0: /*
aoqi@0:  * Copyright (c) 1999, 2012, 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: #ifndef SHARE_VM_MEMORY_THREADLOCALALLOCBUFFER_HPP
aoqi@0: #define SHARE_VM_MEMORY_THREADLOCALALLOCBUFFER_HPP
aoqi@0: 
aoqi@0: #include "gc_implementation/shared/gcUtil.hpp"
aoqi@0: #include "oops/typeArrayOop.hpp"
aoqi@0: #include "runtime/perfData.hpp"
aoqi@0: 
aoqi@0: class GlobalTLABStats;
aoqi@0: 
aoqi@0: // ThreadLocalAllocBuffer: a descriptor for thread-local storage used by
aoqi@0: // the threads for allocation.
aoqi@0: //            It is thread-private at any time, but maybe multiplexed over
aoqi@0: //            time across multiple threads. The park()/unpark() pair is
aoqi@0: //            used to make it available for such multiplexing.
aoqi@0: class ThreadLocalAllocBuffer: public CHeapObj<mtThread> {
aoqi@0:   friend class VMStructs;
aoqi@0: private:
aoqi@0:   HeapWord* _start;                              // address of TLAB
aoqi@0:   HeapWord* _top;                                // address after last allocation
aoqi@0:   HeapWord* _pf_top;                             // allocation prefetch watermark
aoqi@0:   HeapWord* _end;                                // allocation end (excluding alignment_reserve)
aoqi@0:   size_t    _desired_size;                       // desired size   (including alignment_reserve)
aoqi@0:   size_t    _refill_waste_limit;                 // hold onto tlab if free() is larger than this
aoqi@0:   size_t    _allocated_before_last_gc;           // total bytes allocated up until the last gc
aoqi@0: 
aoqi@0:   static size_t   _max_size;                     // maximum size of any TLAB
aoqi@0:   static unsigned _target_refills;               // expected number of refills between GCs
aoqi@0: 
aoqi@0:   unsigned  _number_of_refills;
aoqi@0:   unsigned  _fast_refill_waste;
aoqi@0:   unsigned  _slow_refill_waste;
aoqi@0:   unsigned  _gc_waste;
aoqi@0:   unsigned  _slow_allocations;
aoqi@0: 
aoqi@0:   AdaptiveWeightedAverage _allocation_fraction;  // fraction of eden allocated in tlabs
aoqi@0: 
aoqi@0:   void accumulate_statistics();
aoqi@0:   void initialize_statistics();
aoqi@0: 
aoqi@0:   void set_start(HeapWord* start)                { _start = start; }
aoqi@0:   void set_end(HeapWord* end)                    { _end = end; }
aoqi@0:   void set_top(HeapWord* top)                    { _top = top; }
aoqi@0:   void set_pf_top(HeapWord* pf_top)              { _pf_top = pf_top; }
aoqi@0:   void set_desired_size(size_t desired_size)     { _desired_size = desired_size; }
aoqi@0:   void set_refill_waste_limit(size_t waste)      { _refill_waste_limit = waste;  }
aoqi@0: 
aoqi@0:   size_t initial_refill_waste_limit()            { return desired_size() / TLABRefillWasteFraction; }
aoqi@0: 
aoqi@0:   static int    target_refills()                 { return _target_refills; }
aoqi@0:   size_t initial_desired_size();
aoqi@0: 
aoqi@0:   size_t remaining() const                       { return end() == NULL ? 0 : pointer_delta(hard_end(), top()); }
aoqi@0: 
aoqi@0:   // Make parsable and release it.
aoqi@0:   void reset();
aoqi@0: 
aoqi@0:   // Resize based on amount of allocation, etc.
aoqi@0:   void resize();
aoqi@0: 
aoqi@0:   void invariants() const { assert(top() >= start() && top() <= end(), "invalid tlab"); }
aoqi@0: 
aoqi@0:   void initialize(HeapWord* start, HeapWord* top, HeapWord* end);
aoqi@0: 
aoqi@0:   void print_stats(const char* tag);
aoqi@0: 
aoqi@0:   Thread* myThread();
aoqi@0: 
aoqi@0:   // statistics
aoqi@0: 
aoqi@0:   int number_of_refills() const { return _number_of_refills; }
aoqi@0:   int fast_refill_waste() const { return _fast_refill_waste; }
aoqi@0:   int slow_refill_waste() const { return _slow_refill_waste; }
aoqi@0:   int gc_waste() const          { return _gc_waste; }
aoqi@0:   int slow_allocations() const  { return _slow_allocations; }
aoqi@0: 
aoqi@0:   static GlobalTLABStats* _global_stats;
aoqi@0:   static GlobalTLABStats* global_stats() { return _global_stats; }
aoqi@0: 
aoqi@0: public:
aoqi@0:   ThreadLocalAllocBuffer() : _allocation_fraction(TLABAllocationWeight), _allocated_before_last_gc(0) {
aoqi@0:     // do nothing.  tlabs must be inited by initialize() calls
aoqi@0:   }
aoqi@0: 
aoqi@0:   static const size_t min_size()                 { return align_object_size(MinTLABSize / HeapWordSize); }
aoqi@0:   static const size_t max_size()                 { assert(_max_size != 0, "max_size not set up"); return _max_size; }
aoqi@0:   static void set_max_size(size_t max_size)      { _max_size = max_size; }
aoqi@0: 
aoqi@0:   HeapWord* start() const                        { return _start; }
aoqi@0:   HeapWord* end() const                          { return _end; }
aoqi@0:   HeapWord* hard_end() const                     { return _end + alignment_reserve(); }
aoqi@0:   HeapWord* top() const                          { return _top; }
aoqi@0:   HeapWord* pf_top() const                       { return _pf_top; }
aoqi@0:   size_t desired_size() const                    { return _desired_size; }
aoqi@0:   size_t used() const                            { return pointer_delta(top(), start()); }
aoqi@0:   size_t used_bytes() const                      { return pointer_delta(top(), start(), 1); }
aoqi@0:   size_t free() const                            { return pointer_delta(end(), top()); }
aoqi@0:   // Don't discard tlab if remaining space is larger than this.
aoqi@0:   size_t refill_waste_limit() const              { return _refill_waste_limit; }
aoqi@0: 
aoqi@0:   // Allocate size HeapWords. The memory is NOT initialized to zero.
aoqi@0:   inline HeapWord* allocate(size_t size);
aoqi@0: 
aoqi@0:   // Reserve space at the end of TLAB
aoqi@0:   static size_t end_reserve() {
aoqi@0:     int reserve_size = typeArrayOopDesc::header_size(T_INT);
aoqi@0:     return MAX2(reserve_size, VM_Version::reserve_for_allocation_prefetch());
aoqi@0:   }
aoqi@0:   static size_t alignment_reserve()              { return align_object_size(end_reserve()); }
aoqi@0:   static size_t alignment_reserve_in_bytes()     { return alignment_reserve() * HeapWordSize; }
aoqi@0: 
aoqi@0:   // Return tlab size or remaining space in eden such that the
aoqi@0:   // space is large enough to hold obj_size and necessary fill space.
aoqi@0:   // Otherwise return 0;
aoqi@0:   inline size_t compute_size(size_t obj_size);
aoqi@0: 
aoqi@0:   // Record slow allocation
aoqi@0:   inline void record_slow_allocation(size_t obj_size);
aoqi@0: 
aoqi@0:   // Initialization at startup
aoqi@0:   static void startup_initialization();
aoqi@0: 
aoqi@0:   // Make an in-use tlab parsable, optionally also retiring it.
aoqi@0:   void make_parsable(bool retire);
aoqi@0: 
aoqi@0:   // Retire in-use tlab before allocation of a new tlab
aoqi@0:   void clear_before_allocation();
aoqi@0: 
aoqi@0:   // Accumulate statistics across all tlabs before gc
aoqi@0:   static void accumulate_statistics_before_gc();
aoqi@0: 
aoqi@0:   // Resize tlabs for all threads
aoqi@0:   static void resize_all_tlabs();
aoqi@0: 
aoqi@0:   void fill(HeapWord* start, HeapWord* top, size_t new_size);
aoqi@0:   void initialize();
aoqi@0: 
aoqi@0:   static size_t refill_waste_limit_increment()   { return TLABWasteIncrement; }
aoqi@0: 
aoqi@0:   // Code generation support
aoqi@0:   static ByteSize start_offset()                 { return byte_offset_of(ThreadLocalAllocBuffer, _start); }
aoqi@0:   static ByteSize end_offset()                   { return byte_offset_of(ThreadLocalAllocBuffer, _end  ); }
aoqi@0:   static ByteSize top_offset()                   { return byte_offset_of(ThreadLocalAllocBuffer, _top  ); }
aoqi@0:   static ByteSize pf_top_offset()                { return byte_offset_of(ThreadLocalAllocBuffer, _pf_top  ); }
aoqi@0:   static ByteSize size_offset()                  { return byte_offset_of(ThreadLocalAllocBuffer, _desired_size ); }
aoqi@0:   static ByteSize refill_waste_limit_offset()    { return byte_offset_of(ThreadLocalAllocBuffer, _refill_waste_limit ); }
aoqi@0: 
aoqi@0:   static ByteSize number_of_refills_offset()     { return byte_offset_of(ThreadLocalAllocBuffer, _number_of_refills ); }
aoqi@0:   static ByteSize fast_refill_waste_offset()     { return byte_offset_of(ThreadLocalAllocBuffer, _fast_refill_waste ); }
aoqi@0:   static ByteSize slow_allocations_offset()      { return byte_offset_of(ThreadLocalAllocBuffer, _slow_allocations ); }
aoqi@0: 
aoqi@0:   void verify();
aoqi@0: };
aoqi@0: 
aoqi@0: class GlobalTLABStats: public CHeapObj<mtThread> {
aoqi@0: private:
aoqi@0: 
aoqi@0:   // Accumulate perfdata in private variables because
aoqi@0:   // PerfData should be write-only for security reasons
aoqi@0:   // (see perfData.hpp)
aoqi@0:   unsigned _allocating_threads;
aoqi@0:   unsigned _total_refills;
aoqi@0:   unsigned _max_refills;
aoqi@0:   size_t   _total_allocation;
aoqi@0:   size_t   _total_gc_waste;
aoqi@0:   size_t   _max_gc_waste;
aoqi@0:   size_t   _total_slow_refill_waste;
aoqi@0:   size_t   _max_slow_refill_waste;
aoqi@0:   size_t   _total_fast_refill_waste;
aoqi@0:   size_t   _max_fast_refill_waste;
aoqi@0:   unsigned _total_slow_allocations;
aoqi@0:   unsigned _max_slow_allocations;
aoqi@0: 
aoqi@0:   PerfVariable* _perf_allocating_threads;
aoqi@0:   PerfVariable* _perf_total_refills;
aoqi@0:   PerfVariable* _perf_max_refills;
aoqi@0:   PerfVariable* _perf_allocation;
aoqi@0:   PerfVariable* _perf_gc_waste;
aoqi@0:   PerfVariable* _perf_max_gc_waste;
aoqi@0:   PerfVariable* _perf_slow_refill_waste;
aoqi@0:   PerfVariable* _perf_max_slow_refill_waste;
aoqi@0:   PerfVariable* _perf_fast_refill_waste;
aoqi@0:   PerfVariable* _perf_max_fast_refill_waste;
aoqi@0:   PerfVariable* _perf_slow_allocations;
aoqi@0:   PerfVariable* _perf_max_slow_allocations;
aoqi@0: 
aoqi@0:   AdaptiveWeightedAverage _allocating_threads_avg;
aoqi@0: 
aoqi@0: public:
aoqi@0:   GlobalTLABStats();
aoqi@0: 
aoqi@0:   // Initialize all counters
aoqi@0:   void initialize();
aoqi@0: 
aoqi@0:   // Write all perf counters to the perf_counters
aoqi@0:   void publish();
aoqi@0: 
aoqi@0:   void print();
aoqi@0: 
aoqi@0:   // Accessors
aoqi@0:   unsigned allocating_threads_avg() {
aoqi@0:     return MAX2((unsigned)(_allocating_threads_avg.average() + 0.5), 1U);
aoqi@0:   }
aoqi@0: 
aoqi@0:   size_t allocation() {
aoqi@0:     return _total_allocation;
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Update methods
aoqi@0: 
aoqi@0:   void update_allocating_threads() {
aoqi@0:     _allocating_threads++;
aoqi@0:   }
aoqi@0:   void update_number_of_refills(unsigned value) {
aoqi@0:     _total_refills += value;
aoqi@0:     _max_refills    = MAX2(_max_refills, value);
aoqi@0:   }
aoqi@0:   void update_allocation(size_t value) {
aoqi@0:     _total_allocation += value;
aoqi@0:   }
aoqi@0:   void update_gc_waste(size_t value) {
aoqi@0:     _total_gc_waste += value;
aoqi@0:     _max_gc_waste    = MAX2(_max_gc_waste, value);
aoqi@0:   }
aoqi@0:   void update_fast_refill_waste(size_t value) {
aoqi@0:     _total_fast_refill_waste += value;
aoqi@0:     _max_fast_refill_waste    = MAX2(_max_fast_refill_waste, value);
aoqi@0:   }
aoqi@0:   void update_slow_refill_waste(size_t value) {
aoqi@0:     _total_slow_refill_waste += value;
aoqi@0:     _max_slow_refill_waste    = MAX2(_max_slow_refill_waste, value);
aoqi@0:   }
aoqi@0:   void update_slow_allocations(unsigned value) {
aoqi@0:     _total_slow_allocations += value;
aoqi@0:     _max_slow_allocations    = MAX2(_max_slow_allocations, value);
aoqi@0:   }
aoqi@0: };
aoqi@0: 
aoqi@0: #endif // SHARE_VM_MEMORY_THREADLOCALALLOCBUFFER_HPP