Mercurial > hg > graal-jvmci-8
view src/share/vm/memory/allocation.cpp @ 1145:e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
Summary: Autonomic per-worker free block cache sizing, tunable coalition policies, fixes to per-size block statistics, retuned gain and bandwidth of some feedback loop filters to allow quicker reactivity to abrupt changes in ambient demand, and other heuristics to reduce fragmentation of the CMS old gen. Also tightened some assertions, including those related to locking.
Reviewed-by: jmasa
author | ysr |
---|---|
date | Wed, 23 Dec 2009 09:23:54 -0800 |
parents | a61af66fc99e |
children | c18cbe5936b8 |
line wrap: on
line source
/* * Copyright 1997-2005 Sun Microsystems, Inc. All Rights Reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, * CA 95054 USA or visit www.sun.com if you need additional information or * have any questions. * */ # include "incls/_precompiled.incl" # include "incls/_allocation.cpp.incl" void* CHeapObj::operator new(size_t size){ return (void *) AllocateHeap(size, "CHeapObj-new"); } void CHeapObj::operator delete(void* p){ FreeHeap(p); } void* StackObj::operator new(size_t size) { ShouldNotCallThis(); return 0; }; void StackObj::operator delete(void* p) { ShouldNotCallThis(); }; void* _ValueObj::operator new(size_t size) { ShouldNotCallThis(); return 0; }; void _ValueObj::operator delete(void* p) { ShouldNotCallThis(); }; void* ResourceObj::operator new(size_t size, allocation_type type) { address res; switch (type) { case C_HEAP: res = (address)AllocateHeap(size, "C_Heap: ResourceOBJ"); break; case RESOURCE_AREA: res = (address)operator new(size); break; default: ShouldNotReachHere(); } // Set allocation type in the resource object for assertion checks. DEBUG_ONLY(((ResourceObj *)res)->_allocation = type;) return res; } void ResourceObj::operator delete(void* p) { assert(((ResourceObj *)p)->allocated_on_C_heap(), "delete only allowed for C_HEAP objects"); FreeHeap(p); } void trace_heap_malloc(size_t size, const char* name, void* p) { // A lock is not needed here - tty uses a lock internally tty->print_cr("Heap malloc " INTPTR_FORMAT " %7d %s", p, size, name == NULL ? "" : name); } void trace_heap_free(void* p) { // A lock is not needed here - tty uses a lock internally tty->print_cr("Heap free " INTPTR_FORMAT, p); } bool warn_new_operator = false; // see vm_main //-------------------------------------------------------------------------------------- // ChunkPool implementation // MT-safe pool of chunks to reduce malloc/free thrashing // NB: not using Mutex because pools are used before Threads are initialized class ChunkPool { Chunk* _first; // first cached Chunk; its first word points to next chunk size_t _num_chunks; // number of unused chunks in pool size_t _num_used; // number of chunks currently checked out const size_t _size; // size of each chunk (must be uniform) // Our three static pools static ChunkPool* _large_pool; static ChunkPool* _medium_pool; static ChunkPool* _small_pool; // return first element or null void* get_first() { Chunk* c = _first; if (_first) { _first = _first->next(); _num_chunks--; } return c; } public: // All chunks in a ChunkPool has the same size ChunkPool(size_t size) : _size(size) { _first = NULL; _num_chunks = _num_used = 0; } // Allocate a new chunk from the pool (might expand the pool) void* allocate(size_t bytes) { assert(bytes == _size, "bad size"); void* p = NULL; { ThreadCritical tc; _num_used++; p = get_first(); if (p == NULL) p = os::malloc(bytes); } if (p == NULL) vm_exit_out_of_memory(bytes, "ChunkPool::allocate"); return p; } // Return a chunk to the pool void free(Chunk* chunk) { assert(chunk->length() + Chunk::aligned_overhead_size() == _size, "bad size"); ThreadCritical tc; _num_used--; // Add chunk to list chunk->set_next(_first); _first = chunk; _num_chunks++; } // Prune the pool void free_all_but(size_t n) { // if we have more than n chunks, free all of them ThreadCritical tc; if (_num_chunks > n) { // free chunks at end of queue, for better locality Chunk* cur = _first; for (size_t i = 0; i < (n - 1) && cur != NULL; i++) cur = cur->next(); if (cur != NULL) { Chunk* next = cur->next(); cur->set_next(NULL); cur = next; // Free all remaining chunks while(cur != NULL) { next = cur->next(); os::free(cur); _num_chunks--; cur = next; } } } } // Accessors to preallocated pool's static ChunkPool* large_pool() { assert(_large_pool != NULL, "must be initialized"); return _large_pool; } static ChunkPool* medium_pool() { assert(_medium_pool != NULL, "must be initialized"); return _medium_pool; } static ChunkPool* small_pool() { assert(_small_pool != NULL, "must be initialized"); return _small_pool; } static void initialize() { _large_pool = new ChunkPool(Chunk::size + Chunk::aligned_overhead_size()); _medium_pool = new ChunkPool(Chunk::medium_size + Chunk::aligned_overhead_size()); _small_pool = new ChunkPool(Chunk::init_size + Chunk::aligned_overhead_size()); } }; ChunkPool* ChunkPool::_large_pool = NULL; ChunkPool* ChunkPool::_medium_pool = NULL; ChunkPool* ChunkPool::_small_pool = NULL; void chunkpool_init() { ChunkPool::initialize(); } //-------------------------------------------------------------------------------------- // ChunkPoolCleaner implementation class ChunkPoolCleaner : public PeriodicTask { enum { CleaningInterval = 5000, // cleaning interval in ms BlocksToKeep = 5 // # of extra blocks to keep }; public: ChunkPoolCleaner() : PeriodicTask(CleaningInterval) {} void task() { ChunkPool::small_pool()->free_all_but(BlocksToKeep); ChunkPool::medium_pool()->free_all_but(BlocksToKeep); ChunkPool::large_pool()->free_all_but(BlocksToKeep); } }; //-------------------------------------------------------------------------------------- // Chunk implementation void* Chunk::operator new(size_t requested_size, size_t length) { // requested_size is equal to sizeof(Chunk) but in order for the arena // allocations to come out aligned as expected the size must be aligned // to expected arean alignment. // expect requested_size but if sizeof(Chunk) doesn't match isn't proper size we must align it. assert(ARENA_ALIGN(requested_size) == aligned_overhead_size(), "Bad alignment"); size_t bytes = ARENA_ALIGN(requested_size) + length; switch (length) { case Chunk::size: return ChunkPool::large_pool()->allocate(bytes); case Chunk::medium_size: return ChunkPool::medium_pool()->allocate(bytes); case Chunk::init_size: return ChunkPool::small_pool()->allocate(bytes); default: { void *p = os::malloc(bytes); if (p == NULL) vm_exit_out_of_memory(bytes, "Chunk::new"); return p; } } } void Chunk::operator delete(void* p) { Chunk* c = (Chunk*)p; switch (c->length()) { case Chunk::size: ChunkPool::large_pool()->free(c); break; case Chunk::medium_size: ChunkPool::medium_pool()->free(c); break; case Chunk::init_size: ChunkPool::small_pool()->free(c); break; default: os::free(c); } } Chunk::Chunk(size_t length) : _len(length) { _next = NULL; // Chain on the linked list } void Chunk::chop() { Chunk *k = this; while( k ) { Chunk *tmp = k->next(); // clear out this chunk (to detect allocation bugs) if (ZapResourceArea) memset(k->bottom(), badResourceValue, k->length()); delete k; // Free chunk (was malloc'd) k = tmp; } } void Chunk::next_chop() { _next->chop(); _next = NULL; } void Chunk::start_chunk_pool_cleaner_task() { #ifdef ASSERT static bool task_created = false; assert(!task_created, "should not start chuck pool cleaner twice"); task_created = true; #endif ChunkPoolCleaner* cleaner = new ChunkPoolCleaner(); cleaner->enroll(); } //------------------------------Arena------------------------------------------ Arena::Arena(size_t init_size) { size_t round_size = (sizeof (char *)) - 1; init_size = (init_size+round_size) & ~round_size; _first = _chunk = new (init_size) Chunk(init_size); _hwm = _chunk->bottom(); // Save the cached hwm, max _max = _chunk->top(); set_size_in_bytes(init_size); } Arena::Arena() { _first = _chunk = new (Chunk::init_size) Chunk(Chunk::init_size); _hwm = _chunk->bottom(); // Save the cached hwm, max _max = _chunk->top(); set_size_in_bytes(Chunk::init_size); } Arena::Arena(Arena *a) : _chunk(a->_chunk), _hwm(a->_hwm), _max(a->_max), _first(a->_first) { set_size_in_bytes(a->size_in_bytes()); } Arena *Arena::move_contents(Arena *copy) { copy->destruct_contents(); copy->_chunk = _chunk; copy->_hwm = _hwm; copy->_max = _max; copy->_first = _first; copy->set_size_in_bytes(size_in_bytes()); // Destroy original arena reset(); return copy; // Return Arena with contents } Arena::~Arena() { destruct_contents(); } // Destroy this arenas contents and reset to empty void Arena::destruct_contents() { if (UseMallocOnly && _first != NULL) { char* end = _first->next() ? _first->top() : _hwm; free_malloced_objects(_first, _first->bottom(), end, _hwm); } _first->chop(); reset(); } // Total of all Chunks in arena size_t Arena::used() const { size_t sum = _chunk->length() - (_max-_hwm); // Size leftover in this Chunk register Chunk *k = _first; while( k != _chunk) { // Whilst have Chunks in a row sum += k->length(); // Total size of this Chunk k = k->next(); // Bump along to next Chunk } return sum; // Return total consumed space. } // Grow a new Chunk void* Arena::grow( size_t x ) { // Get minimal required size. Either real big, or even bigger for giant objs size_t len = MAX2(x, (size_t) Chunk::size); Chunk *k = _chunk; // Get filled-up chunk address _chunk = new (len) Chunk(len); if (_chunk == NULL) vm_exit_out_of_memory(len * Chunk::aligned_overhead_size(), "Arena::grow"); if (k) k->set_next(_chunk); // Append new chunk to end of linked list else _first = _chunk; _hwm = _chunk->bottom(); // Save the cached hwm, max _max = _chunk->top(); set_size_in_bytes(size_in_bytes() + len); void* result = _hwm; _hwm += x; return result; } // Reallocate storage in Arena. void *Arena::Arealloc(void* old_ptr, size_t old_size, size_t new_size) { assert(new_size >= 0, "bad size"); if (new_size == 0) return NULL; #ifdef ASSERT if (UseMallocOnly) { // always allocate a new object (otherwise we'll free this one twice) char* copy = (char*)Amalloc(new_size); size_t n = MIN2(old_size, new_size); if (n > 0) memcpy(copy, old_ptr, n); Afree(old_ptr,old_size); // Mostly done to keep stats accurate return copy; } #endif char *c_old = (char*)old_ptr; // Handy name // Stupid fast special case if( new_size <= old_size ) { // Shrink in-place if( c_old+old_size == _hwm) // Attempt to free the excess bytes _hwm = c_old+new_size; // Adjust hwm return c_old; } // make sure that new_size is legal size_t corrected_new_size = ARENA_ALIGN(new_size); // See if we can resize in-place if( (c_old+old_size == _hwm) && // Adjusting recent thing (c_old+corrected_new_size <= _max) ) { // Still fits where it sits _hwm = c_old+corrected_new_size; // Adjust hwm return c_old; // Return old pointer } // Oops, got to relocate guts void *new_ptr = Amalloc(new_size); memcpy( new_ptr, c_old, old_size ); Afree(c_old,old_size); // Mostly done to keep stats accurate return new_ptr; } // Determine if pointer belongs to this Arena or not. bool Arena::contains( const void *ptr ) const { #ifdef ASSERT if (UseMallocOnly) { // really slow, but not easy to make fast if (_chunk == NULL) return false; char** bottom = (char**)_chunk->bottom(); for (char** p = (char**)_hwm - 1; p >= bottom; p--) { if (*p == ptr) return true; } for (Chunk *c = _first; c != NULL; c = c->next()) { if (c == _chunk) continue; // current chunk has been processed char** bottom = (char**)c->bottom(); for (char** p = (char**)c->top() - 1; p >= bottom; p--) { if (*p == ptr) return true; } } return false; } #endif if( (void*)_chunk->bottom() <= ptr && ptr < (void*)_hwm ) return true; // Check for in this chunk for (Chunk *c = _first; c; c = c->next()) { if (c == _chunk) continue; // current chunk has been processed if ((void*)c->bottom() <= ptr && ptr < (void*)c->top()) { return true; // Check for every chunk in Arena } } return false; // Not in any Chunk, so not in Arena } #ifdef ASSERT void* Arena::malloc(size_t size) { assert(UseMallocOnly, "shouldn't call"); // use malloc, but save pointer in res. area for later freeing char** save = (char**)internal_malloc_4(sizeof(char*)); return (*save = (char*)os::malloc(size)); } // for debugging with UseMallocOnly void* Arena::internal_malloc_4(size_t x) { assert( (x&(sizeof(char*)-1)) == 0, "misaligned size" ); if (_hwm + x > _max) { return grow(x); } else { char *old = _hwm; _hwm += x; return old; } } #endif //-------------------------------------------------------------------------------------- // Non-product code #ifndef PRODUCT // The global operator new should never be called since it will usually indicate // a memory leak. Use CHeapObj as the base class of such objects to make it explicit // that they're allocated on the C heap. // Commented out in product version to avoid conflicts with third-party C++ native code. // %% note this is causing a problem on solaris debug build. the global // new is being called from jdk source and causing data corruption. // src/share/native/sun/awt/font/fontmanager/textcache/hsMemory.cpp::hsSoftNew // define CATCH_OPERATOR_NEW_USAGE if you want to use this. #ifdef CATCH_OPERATOR_NEW_USAGE void* operator new(size_t size){ static bool warned = false; if (!warned && warn_new_operator) warning("should not call global (default) operator new"); warned = true; return (void *) AllocateHeap(size, "global operator new"); } #endif void AllocatedObj::print() const { print_on(tty); } void AllocatedObj::print_value() const { print_value_on(tty); } void AllocatedObj::print_on(outputStream* st) const { st->print_cr("AllocatedObj(" INTPTR_FORMAT ")", this); } void AllocatedObj::print_value_on(outputStream* st) const { st->print("AllocatedObj(" INTPTR_FORMAT ")", this); } size_t Arena::_bytes_allocated = 0; AllocStats::AllocStats() { start_mallocs = os::num_mallocs; start_frees = os::num_frees; start_malloc_bytes = os::alloc_bytes; start_res_bytes = Arena::_bytes_allocated; } int AllocStats::num_mallocs() { return os::num_mallocs - start_mallocs; } size_t AllocStats::alloc_bytes() { return os::alloc_bytes - start_malloc_bytes; } size_t AllocStats::resource_bytes() { return Arena::_bytes_allocated - start_res_bytes; } int AllocStats::num_frees() { return os::num_frees - start_frees; } void AllocStats::print() { tty->print("%d mallocs (%ldK), %d frees, %ldK resrc", num_mallocs(), alloc_bytes()/K, num_frees(), resource_bytes()/K); } // debugging code inline void Arena::free_all(char** start, char** end) { for (char** p = start; p < end; p++) if (*p) os::free(*p); } void Arena::free_malloced_objects(Chunk* chunk, char* hwm, char* max, char* hwm2) { assert(UseMallocOnly, "should not call"); // free all objects malloced since resource mark was created; resource area // contains their addresses if (chunk->next()) { // this chunk is full, and some others too for (Chunk* c = chunk->next(); c != NULL; c = c->next()) { char* top = c->top(); if (c->next() == NULL) { top = hwm2; // last junk is only used up to hwm2 assert(c->contains(hwm2), "bad hwm2"); } free_all((char**)c->bottom(), (char**)top); } assert(chunk->contains(hwm), "bad hwm"); assert(chunk->contains(max), "bad max"); free_all((char**)hwm, (char**)max); } else { // this chunk was partially used assert(chunk->contains(hwm), "bad hwm"); assert(chunk->contains(hwm2), "bad hwm2"); free_all((char**)hwm, (char**)hwm2); } } ReallocMark::ReallocMark() { #ifdef ASSERT Thread *thread = ThreadLocalStorage::get_thread_slow(); _nesting = thread->resource_area()->nesting(); #endif } void ReallocMark::check() { #ifdef ASSERT if (_nesting != Thread::current()->resource_area()->nesting()) { fatal("allocation bug: array could grow within nested ResourceMark"); } #endif } #endif // Non-product