zgu@7074: /*
zgu@7074:  * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved.
zgu@7074:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
zgu@7074:  *
zgu@7074:  * This code is free software; you can redistribute it and/or modify it
zgu@7074:  * under the terms of the GNU General Public License version 2 only, as
zgu@7074:  * published by the Free Software Foundation.
zgu@7074:  *
zgu@7074:  * This code is distributed in the hope that it will be useful, but WITHOUT
zgu@7074:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
zgu@7074:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
zgu@7074:  * version 2 for more details (a copy is included in the LICENSE file that
zgu@7074:  * accompanied this code).
zgu@7074:  *
zgu@7074:  * You should have received a copy of the GNU General Public License version
zgu@7074:  * 2 along with this work; if not, write to the Free Software Foundation,
zgu@7074:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
zgu@7074:  *
zgu@7074:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
zgu@7074:  * or visit www.oracle.com if you need additional information or have any
zgu@7074:  * questions.
zgu@7074:  *
zgu@7074:  */
zgu@7074: 
zgu@7074: #ifndef SHARE_VM_SERVICES_MALLOC_SITE_TABLE_HPP
zgu@7074: #define SHARE_VM_SERVICES_MALLOC_SITE_TABLE_HPP
zgu@7074: 
zgu@7074: #if INCLUDE_NMT
zgu@7074: 
zgu@7074: #include "memory/allocation.hpp"
zgu@7074: #include "runtime/atomic.hpp"
zgu@7074: #include "services/allocationSite.hpp"
zgu@7074: #include "services/mallocTracker.hpp"
zgu@7074: #include "services/nmtCommon.hpp"
zgu@7074: 
zgu@7074: // MallocSite represents a code path that eventually calls
zgu@7074: // os::malloc() to allocate memory
zgu@7074: class MallocSite : public AllocationSite<MemoryCounter> {
zgu@7074:  public:
zgu@7074:   MallocSite() :
zgu@7074:     AllocationSite<MemoryCounter>(emptyStack) { }
zgu@7074: 
zgu@7074:   MallocSite(const NativeCallStack& stack) :
zgu@7074:     AllocationSite<MemoryCounter>(stack) { }
zgu@7074: 
zgu@7074:   void allocate(size_t size)      { data()->allocate(size);   }
zgu@7074:   void deallocate(size_t size)    { data()->deallocate(size); }
zgu@7074: 
zgu@7074:   // Memory allocated from this code path
zgu@7074:   size_t size()  const { return peek()->size(); }
zgu@7074:   // The number of calls were made
zgu@7074:   size_t count() const { return peek()->count(); }
zgu@7074: };
zgu@7074: 
zgu@7074: // Malloc site hashtable entry
zgu@7074: class MallocSiteHashtableEntry : public CHeapObj<mtNMT> {
zgu@7074:  private:
zgu@7074:   MallocSite                _malloc_site;
zgu@7074:   MallocSiteHashtableEntry* _next;
zgu@7074: 
zgu@7074:  public:
zgu@7074:   MallocSiteHashtableEntry() : _next(NULL) { }
zgu@7074: 
zgu@7074:   MallocSiteHashtableEntry(NativeCallStack stack):
zgu@7074:     _malloc_site(stack), _next(NULL) { }
zgu@7074: 
zgu@7074:   inline const MallocSiteHashtableEntry* next() const {
zgu@7074:     return _next;
zgu@7074:   }
zgu@7074: 
zgu@7074:   // Insert an entry atomically.
zgu@7074:   // Return true if the entry is inserted successfully.
zgu@7074:   // The operation can be failed due to contention from other thread.
zgu@7074:   bool atomic_insert(const MallocSiteHashtableEntry* entry) {
zgu@7074:     return (Atomic::cmpxchg_ptr((void*)entry, (volatile void*)&_next,
zgu@7074:       NULL) == NULL);
zgu@7074:   }
zgu@7074: 
zgu@7074:   void set_callsite(const MallocSite& site) {
zgu@7074:     _malloc_site = site;
zgu@7074:   }
zgu@7074: 
zgu@7074:   inline const MallocSite* peek() const { return &_malloc_site; }
zgu@7074:   inline MallocSite* data()             { return &_malloc_site; }
zgu@7074: 
zgu@7074:   inline long hash() const { return _malloc_site.hash(); }
zgu@7074:   inline bool equals(const NativeCallStack& stack) const {
zgu@7074:     return _malloc_site.equals(stack);
zgu@7074:   }
zgu@7074:   // Allocation/deallocation on this allocation site
zgu@7074:   inline void allocate(size_t size)   { _malloc_site.allocate(size);   }
zgu@7074:   inline void deallocate(size_t size) { _malloc_site.deallocate(size); }
zgu@7074:   // Memory counters
zgu@7074:   inline size_t size() const  { return _malloc_site.size();  }
zgu@7074:   inline size_t count() const { return _malloc_site.count(); }
zgu@7074: };
zgu@7074: 
zgu@7074: // The walker walks every entry on MallocSiteTable
zgu@7074: class MallocSiteWalker : public StackObj {
zgu@7074:  public:
zgu@7074:    virtual bool do_malloc_site(const MallocSite* e) { return false; }
zgu@7074: };
zgu@7074: 
zgu@7074: /*
zgu@7074:  * Native memory tracking call site table.
zgu@7074:  * The table is only needed when detail tracking is enabled.
zgu@7074:  */
zgu@7074: class MallocSiteTable : AllStatic {
zgu@7074:  private:
zgu@7074:   // The number of hash bucket in this hashtable. The number should
zgu@7074:   // be tuned if malloc activities changed significantly.
zgu@7074:   // The statistics data can be obtained via Jcmd
zgu@7074:   // jcmd <pid> VM.native_memory statistics.
zgu@7074: 
zgu@7074:   // Currently, (number of buckets / number of entires) ratio is
zgu@7074:   // about 1 / 6
zgu@7074:   enum {
zgu@7074:     table_base_size = 128,   // The base size is calculated from statistics to give
zgu@7074:                              // table ratio around 1:6
zgu@7074:     table_size = (table_base_size * NMT_TrackingStackDepth - 1)
zgu@7074:   };
zgu@7074: 
zgu@7074: 
zgu@7074:   // This is a very special lock, that allows multiple shared accesses (sharedLock), but
zgu@7074:   // once exclusive access (exclusiveLock) is requested, all shared accesses are
zgu@7074:   // rejected forever.
zgu@7074:   class AccessLock : public StackObj {
zgu@7074:     enum LockState {
zgu@7074:       NoLock,
zgu@7074:       SharedLock,
zgu@7074:       ExclusiveLock
zgu@7074:     };
zgu@7074: 
zgu@7074:    private:
zgu@7074:     // A very large negative number. The only possibility to "overflow"
zgu@7074:     // this number is when there are more than -min_jint threads in
zgu@7074:     // this process, which is not going to happen in foreseeable future.
zgu@7074:     const static int _MAGIC_ = min_jint;
zgu@7074: 
zgu@7074:     LockState      _lock_state;
zgu@7074:     volatile int*  _lock;
zgu@7074:    public:
zgu@7074:     AccessLock(volatile int* lock) :
zgu@7074:       _lock(lock), _lock_state(NoLock) {
zgu@7074:     }
zgu@7074: 
zgu@7074:     ~AccessLock() {
zgu@7074:       if (_lock_state == SharedLock) {
zgu@7074:         Atomic::dec((volatile jint*)_lock);
zgu@7074:       }
zgu@7074:     }
zgu@7074:     // Acquire shared lock.
zgu@7074:     // Return true if shared access is granted.
zgu@7074:     inline bool sharedLock() {
zgu@7074:       jint res = Atomic::add(1, _lock);
zgu@7074:       if (res < 0) {
zgu@7074:         Atomic::add(-1, _lock);
zgu@7074:         return false;
zgu@7074:       }
zgu@7074:       _lock_state = SharedLock;
zgu@7074:       return true;
zgu@7074:     }
zgu@7074:     // Acquire exclusive lock
zgu@7074:     void exclusiveLock();
zgu@7074:  };
zgu@7074: 
zgu@7074:  public:
zgu@7074:   static bool initialize();
zgu@7074:   static void shutdown();
zgu@7074: 
zgu@7074:   NOT_PRODUCT(static int access_peak_count() { return _peak_count; })
zgu@7074: 
zgu@7074:   // Number of hash buckets
zgu@7074:   static inline int hash_buckets()      { return (int)table_size; }
zgu@7074: 
zgu@7074:   // Access and copy a call stack from this table. Shared lock should be
zgu@7074:   // acquired before access the entry.
zgu@7074:   static inline bool access_stack(NativeCallStack& stack, size_t bucket_idx,
zgu@7074:     size_t pos_idx) {
zgu@7074:     AccessLock locker(&_access_count);
zgu@7074:     if (locker.sharedLock()) {
zgu@7074:       NOT_PRODUCT(_peak_count = MAX2(_peak_count, _access_count);)
zgu@7074:       MallocSite* site = malloc_site(bucket_idx, pos_idx);
zgu@7074:       if (site != NULL) {
zgu@7074:         stack = *site->call_stack();
zgu@7074:         return true;
zgu@7074:       }
zgu@7074:     }
zgu@7074:     return false;
zgu@7074:   }
zgu@7074: 
zgu@7074:   // Record a new allocation from specified call path.
zgu@7074:   // Return true if the allocation is recorded successfully, bucket_idx
zgu@7074:   // and pos_idx are also updated to indicate the entry where the allocation
zgu@7074:   // information was recorded.
zgu@7074:   // Return false only occurs under rare scenarios:
zgu@7074:   //  1. out of memory
zgu@7074:   //  2. overflow hash bucket
zgu@7074:   static inline bool allocation_at(const NativeCallStack& stack, size_t size,
zgu@7074:     size_t* bucket_idx, size_t* pos_idx) {
zgu@7074:     AccessLock locker(&_access_count);
zgu@7074:     if (locker.sharedLock()) {
zgu@7074:       NOT_PRODUCT(_peak_count = MAX2(_peak_count, _access_count);)
zgu@7074:       MallocSite* site = lookup_or_add(stack, bucket_idx, pos_idx);
zgu@7074:       if (site != NULL) site->allocate(size);
zgu@7074:       return site != NULL;
zgu@7074:     }
zgu@7074:     return false;
zgu@7074:   }
zgu@7074: 
zgu@7074:   // Record memory deallocation. bucket_idx and pos_idx indicate where the allocation
zgu@7074:   // information was recorded.
zgu@7074:   static inline bool deallocation_at(size_t size, size_t bucket_idx, size_t pos_idx) {
zgu@7074:     AccessLock locker(&_access_count);
zgu@7074:     if (locker.sharedLock()) {
zgu@7074:       NOT_PRODUCT(_peak_count = MAX2(_peak_count, _access_count);)
zgu@7074:       MallocSite* site = malloc_site(bucket_idx, pos_idx);
zgu@7074:       if (site != NULL) {
zgu@7074:         site->deallocate(size);
zgu@7074:         return true;
zgu@7074:       }
zgu@7074:     }
zgu@7074:     return false;
zgu@7074:   }
zgu@7074: 
zgu@7074:   // Walk this table.
zgu@7074:   static bool walk_malloc_site(MallocSiteWalker* walker);
zgu@7074: 
zgu@7074:  private:
zgu@7074:   static MallocSiteHashtableEntry* new_entry(const NativeCallStack& key);
zgu@7074:   static void reset();
zgu@7074: 
zgu@7074:   // Delete a bucket linked list
zgu@7074:   static void delete_linked_list(MallocSiteHashtableEntry* head);
zgu@7074: 
zgu@7074:   static MallocSite* lookup_or_add(const NativeCallStack& key, size_t* bucket_idx, size_t* pos_idx);
zgu@7074:   static MallocSite* malloc_site(size_t bucket_idx, size_t pos_idx);
zgu@7074:   static bool walk(MallocSiteWalker* walker);
zgu@7074: 
zgu@7074:   static inline int hash_to_index(int  hash) {
zgu@7074:     hash = (hash > 0) ? hash : (-hash);
zgu@7074:     return (hash % table_size);
zgu@7074:   }
zgu@7074: 
zgu@7074:   static inline const NativeCallStack* hash_entry_allocation_stack() {
zgu@7074:     return (NativeCallStack*)_hash_entry_allocation_stack;
zgu@7074:   }
zgu@7074: 
zgu@7074:  private:
zgu@7074:   // Counter for counting concurrent access
zgu@7074:   static volatile int                _access_count;
zgu@7074: 
zgu@7074:   // The callsite hashtable. It has to be a static table,
zgu@7074:   // since malloc call can come from C runtime linker.
zgu@7074:   static MallocSiteHashtableEntry*   _table[table_size];
zgu@7074: 
zgu@7074: 
zgu@7074:   // Reserve enough memory for placing the objects
zgu@7074: 
zgu@7074:   // The memory for hashtable entry allocation stack object
zgu@7074:   static size_t _hash_entry_allocation_stack[CALC_OBJ_SIZE_IN_TYPE(NativeCallStack, size_t)];
zgu@7074:   // The memory for hashtable entry allocation callsite object
zgu@7074:   static size_t _hash_entry_allocation_site[CALC_OBJ_SIZE_IN_TYPE(MallocSiteHashtableEntry, size_t)];
zgu@7074:   NOT_PRODUCT(static int     _peak_count;)
zgu@7074: };
zgu@7074: 
zgu@7074: #endif // INCLUDE_NMT
zgu@7074: #endif // SHARE_VM_SERVICES_MALLOC_SITE_TABLE_HPP