aoqi@0: /*
aoqi@0:  * Copyright (c) 2006, 2013, 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_GC_IMPLEMENTATION_SHARED_MUTABLENUMASPACE_HPP
aoqi@0: #define SHARE_VM_GC_IMPLEMENTATION_SHARED_MUTABLENUMASPACE_HPP
aoqi@0: 
aoqi@0: #include "utilities/macros.hpp"
aoqi@0: #if INCLUDE_ALL_GCS
aoqi@0: #include "gc_implementation/shared/gcUtil.hpp"
aoqi@0: #include "gc_implementation/shared/mutableSpace.hpp"
aoqi@0: #endif // INCLUDE_ALL_GCS
aoqi@0: 
aoqi@0: /*
aoqi@0:  *    The NUMA-aware allocator (MutableNUMASpace) is basically a modification
aoqi@0:  * of MutableSpace which preserves interfaces but implements different
aoqi@0:  * functionality. The space is split into chunks for each locality group
aoqi@0:  * (resizing for adaptive size policy is also supported). For each thread
aoqi@0:  * allocations are performed in the chunk corresponding to the home locality
aoqi@0:  * group of the thread. Whenever any chunk fills-in the young generation
aoqi@0:  * collection occurs.
aoqi@0:  *   The chunks can be also be adaptively resized. The idea behind the adaptive
aoqi@0:  * sizing is to reduce the loss of the space in the eden due to fragmentation.
aoqi@0:  * The main cause of fragmentation is uneven allocation rates of threads.
aoqi@0:  * The allocation rate difference between locality groups may be caused either by
aoqi@0:  * application specifics or by uneven LWP distribution by the OS. Besides,
aoqi@0:  * application can have less threads then the number of locality groups.
aoqi@0:  * In order to resize the chunk we measure the allocation rate of the
aoqi@0:  * application between collections. After that we reshape the chunks to reflect
aoqi@0:  * the allocation rate pattern. The AdaptiveWeightedAverage exponentially
aoqi@0:  * decaying average is used to smooth the measurements. The NUMASpaceResizeRate
aoqi@0:  * parameter is used to control the adaptation speed by restricting the number of
aoqi@0:  * bytes that can be moved during the adaptation phase.
aoqi@0:  *   Chunks may contain pages from a wrong locality group. The page-scanner has
aoqi@0:  * been introduced to address the problem. Remote pages typically appear due to
aoqi@0:  * the memory shortage in the target locality group. Besides Solaris would
aoqi@0:  * allocate a large page from the remote locality group even if there are small
aoqi@0:  * local pages available. The page-scanner scans the pages right after the
aoqi@0:  * collection and frees remote pages in hope that subsequent reallocation would
aoqi@0:  * be more successful. This approach proved to be useful on systems with high
aoqi@0:  * load where multiple processes are competing for the memory.
aoqi@0:  */
aoqi@0: 
aoqi@0: class MutableNUMASpace : public MutableSpace {
aoqi@0:   friend class VMStructs;
aoqi@0: 
aoqi@0:   class LGRPSpace : public CHeapObj<mtGC> {
aoqi@0:     int _lgrp_id;
aoqi@0:     MutableSpace* _space;
aoqi@0:     MemRegion _invalid_region;
aoqi@0:     AdaptiveWeightedAverage *_alloc_rate;
aoqi@0:     bool _allocation_failed;
aoqi@0: 
aoqi@0:     struct SpaceStats {
aoqi@0:       size_t _local_space, _remote_space, _unbiased_space, _uncommited_space;
aoqi@0:       size_t _large_pages, _small_pages;
aoqi@0: 
aoqi@0:       SpaceStats() {
aoqi@0:         _local_space = 0;
aoqi@0:         _remote_space = 0;
aoqi@0:         _unbiased_space = 0;
aoqi@0:         _uncommited_space = 0;
aoqi@0:         _large_pages = 0;
aoqi@0:         _small_pages = 0;
aoqi@0:       }
aoqi@0:     };
aoqi@0: 
aoqi@0:     SpaceStats _space_stats;
aoqi@0: 
aoqi@0:     char* _last_page_scanned;
aoqi@0:     char* last_page_scanned()            { return _last_page_scanned; }
aoqi@0:     void set_last_page_scanned(char* p)  { _last_page_scanned = p;    }
aoqi@0:    public:
aoqi@0:     LGRPSpace(int l, size_t alignment) : _lgrp_id(l), _last_page_scanned(NULL), _allocation_failed(false) {
aoqi@0:       _space = new MutableSpace(alignment);
aoqi@0:       _alloc_rate = new AdaptiveWeightedAverage(NUMAChunkResizeWeight);
aoqi@0:     }
aoqi@0:     ~LGRPSpace() {
aoqi@0:       delete _space;
aoqi@0:       delete _alloc_rate;
aoqi@0:     }
aoqi@0: 
aoqi@0:     void add_invalid_region(MemRegion r) {
aoqi@0:       if (!_invalid_region.is_empty()) {
aoqi@0:       _invalid_region.set_start(MIN2(_invalid_region.start(), r.start()));
aoqi@0:       _invalid_region.set_end(MAX2(_invalid_region.end(), r.end()));
aoqi@0:       } else {
aoqi@0:       _invalid_region = r;
aoqi@0:       }
aoqi@0:     }
aoqi@0: 
aoqi@0:     static bool equals(void* lgrp_id_value, LGRPSpace* p) {
aoqi@0:       return *(int*)lgrp_id_value == p->lgrp_id();
aoqi@0:     }
aoqi@0: 
aoqi@0:     // Report a failed allocation.
aoqi@0:     void set_allocation_failed() { _allocation_failed = true;  }
aoqi@0: 
aoqi@0:     void sample() {
aoqi@0:       // If there was a failed allocation make allocation rate equal
aoqi@0:       // to the size of the whole chunk. This ensures the progress of
aoqi@0:       // the adaptation process.
aoqi@0:       size_t alloc_rate_sample;
aoqi@0:       if (_allocation_failed) {
aoqi@0:         alloc_rate_sample = space()->capacity_in_bytes();
aoqi@0:         _allocation_failed = false;
aoqi@0:       } else {
aoqi@0:         alloc_rate_sample = space()->used_in_bytes();
aoqi@0:       }
aoqi@0:       alloc_rate()->sample(alloc_rate_sample);
aoqi@0:     }
aoqi@0: 
aoqi@0:     MemRegion invalid_region() const                { return _invalid_region;      }
aoqi@0:     void set_invalid_region(MemRegion r)            { _invalid_region = r;         }
aoqi@0:     int lgrp_id() const                             { return _lgrp_id;             }
aoqi@0:     MutableSpace* space() const                     { return _space;               }
aoqi@0:     AdaptiveWeightedAverage* alloc_rate() const     { return _alloc_rate;          }
aoqi@0:     void clear_alloc_rate()                         { _alloc_rate->clear();        }
aoqi@0:     SpaceStats* space_stats()                       { return &_space_stats;        }
aoqi@0:     void clear_space_stats()                        { _space_stats = SpaceStats(); }
aoqi@0: 
aoqi@0:     void accumulate_statistics(size_t page_size);
aoqi@0:     void scan_pages(size_t page_size, size_t page_count);
aoqi@0:   };
aoqi@0: 
aoqi@0:   GrowableArray<LGRPSpace*>* _lgrp_spaces;
aoqi@0:   size_t _page_size;
aoqi@0:   unsigned _adaptation_cycles, _samples_count;
aoqi@0: 
aoqi@0:   void set_page_size(size_t psz)                     { _page_size = psz;          }
aoqi@0:   size_t page_size() const                           { return _page_size;         }
aoqi@0: 
aoqi@0:   unsigned adaptation_cycles()                       { return _adaptation_cycles; }
aoqi@0:   void set_adaptation_cycles(int v)                  { _adaptation_cycles = v;    }
aoqi@0: 
aoqi@0:   unsigned samples_count()                           { return _samples_count;     }
aoqi@0:   void increment_samples_count()                     { ++_samples_count;          }
aoqi@0: 
aoqi@0:   size_t _base_space_size;
aoqi@0:   void set_base_space_size(size_t v)                 { _base_space_size = v;      }
aoqi@0:   size_t base_space_size() const                     { return _base_space_size;   }
aoqi@0: 
aoqi@0:   // Check if the NUMA topology has changed. Add and remove spaces if needed.
aoqi@0:   // The update can be forced by setting the force parameter equal to true.
aoqi@0:   bool update_layout(bool force);
aoqi@0:   // Bias region towards the lgrp.
aoqi@0:   void bias_region(MemRegion mr, int lgrp_id);
aoqi@0:   // Free pages in a given region.
aoqi@0:   void free_region(MemRegion mr);
aoqi@0:   // Get current chunk size.
aoqi@0:   size_t current_chunk_size(int i);
aoqi@0:   // Get default chunk size (equally divide the space).
aoqi@0:   size_t default_chunk_size();
aoqi@0:   // Adapt the chunk size to follow the allocation rate.
aoqi@0:   size_t adaptive_chunk_size(int i, size_t limit);
aoqi@0:   // Scan and free invalid pages.
aoqi@0:   void scan_pages(size_t page_count);
aoqi@0:   // Return the bottom_region and the top_region. Align them to page_size() boundary.
aoqi@0:   // |------------------new_region---------------------------------|
aoqi@0:   // |----bottom_region--|---intersection---|------top_region------|
aoqi@0:   void select_tails(MemRegion new_region, MemRegion intersection,
aoqi@0:                     MemRegion* bottom_region, MemRegion *top_region);
aoqi@0:   // Try to merge the invalid region with the bottom or top region by decreasing
aoqi@0:   // the intersection area. Return the invalid_region aligned to the page_size()
aoqi@0:   // boundary if it's inside the intersection. Return non-empty invalid_region
aoqi@0:   // if it lies inside the intersection (also page-aligned).
aoqi@0:   // |------------------new_region---------------------------------|
aoqi@0:   // |----------------|-------invalid---|--------------------------|
aoqi@0:   // |----bottom_region--|---intersection---|------top_region------|
aoqi@0:   void merge_regions(MemRegion new_region, MemRegion* intersection,
aoqi@0:                      MemRegion *invalid_region);
aoqi@0: 
aoqi@0:  public:
aoqi@0:   GrowableArray<LGRPSpace*>* lgrp_spaces() const     { return _lgrp_spaces;       }
aoqi@0:   MutableNUMASpace(size_t alignment);
aoqi@0:   virtual ~MutableNUMASpace();
aoqi@0:   // Space initialization.
aoqi@0:   virtual void initialize(MemRegion mr, bool clear_space, bool mangle_space, bool setup_pages = SetupPages);
aoqi@0:   // Update space layout if necessary. Do all adaptive resizing job.
aoqi@0:   virtual void update();
aoqi@0:   // Update allocation rate averages.
aoqi@0:   virtual void accumulate_statistics();
aoqi@0: 
aoqi@0:   virtual void clear(bool mangle_space);
aoqi@0:   virtual void mangle_unused_area() PRODUCT_RETURN;
aoqi@0:   virtual void mangle_unused_area_complete() PRODUCT_RETURN;
aoqi@0:   virtual void mangle_region(MemRegion mr) PRODUCT_RETURN;
aoqi@0:   virtual void check_mangled_unused_area(HeapWord* limit) PRODUCT_RETURN;
aoqi@0:   virtual void check_mangled_unused_area_complete() PRODUCT_RETURN;
aoqi@0:   virtual void set_top_for_allocations(HeapWord* v) PRODUCT_RETURN;
aoqi@0:   virtual void set_top_for_allocations() PRODUCT_RETURN;
aoqi@0: 
aoqi@0:   virtual void ensure_parsability();
aoqi@0:   virtual size_t used_in_words() const;
aoqi@0:   virtual size_t free_in_words() const;
aoqi@0: 
aoqi@0:   using MutableSpace::capacity_in_words;
aoqi@0:   virtual size_t capacity_in_words(Thread* thr) const;
aoqi@0:   virtual size_t tlab_capacity(Thread* thr) const;
aoqi@0:   virtual size_t tlab_used(Thread* thr) const;
aoqi@0:   virtual size_t unsafe_max_tlab_alloc(Thread* thr) const;
aoqi@0: 
aoqi@0:   // Allocation (return NULL if full)
aoqi@0:   virtual HeapWord* allocate(size_t word_size);
aoqi@0:   virtual HeapWord* cas_allocate(size_t word_size);
aoqi@0: 
aoqi@0:   // Debugging
aoqi@0:   virtual void print_on(outputStream* st) const;
aoqi@0:   virtual void print_short_on(outputStream* st) const;
aoqi@0:   virtual void verify();
aoqi@0: 
aoqi@0:   virtual void set_top(HeapWord* value);
aoqi@0: };
aoqi@0: 
aoqi@0: #endif // SHARE_VM_GC_IMPLEMENTATION_SHARED_MUTABLENUMASPACE_HPP