1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/services/mallocSiteTable.cpp Wed Aug 27 08:19:12 2014 -0400
1.3 @@ -0,0 +1,261 @@
1.4 +/*
1.5 + * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +#include "precompiled.hpp"
1.28 +
1.29 +
1.30 +#include "memory/allocation.inline.hpp"
1.31 +#include "runtime/atomic.hpp"
1.32 +#include "services/mallocSiteTable.hpp"
1.33 +
1.34 +/*
1.35 + * Early os::malloc() calls come from initializations of static variables, long before entering any
1.36 + * VM code. Upon the arrival of the first os::malloc() call, malloc site hashtable has to be
1.37 + * initialized, along with the allocation site for the hashtable entries.
1.38 + * To ensure that malloc site hashtable can be initialized without triggering any additional os::malloc()
1.39 + * call, the hashtable bucket array and hashtable entry allocation site have to be static.
1.40 + * It is not a problem for hashtable bucket, since it is an array of pointer type, C runtime just
1.41 + * allocates a block memory and zero the memory for it.
1.42 + * But for hashtable entry allocation site object, things get tricky. C runtime not only allocates
1.43 + * memory for it, but also calls its constructor at some later time. If we initialize the allocation site
1.44 + * at the first os::malloc() call, the object will be reinitialized when its constructor is called
1.45 + * by C runtime.
1.46 + * To workaround above issue, we declare a static size_t array with the size of the CallsiteHashtableEntry,
1.47 + * the memory is used to instantiate CallsiteHashtableEntry for the hashtable entry allocation site.
1.48 + * Given it is a primitive type array, C runtime will do nothing other than assign the memory block for the variable,
1.49 + * which is exactly what we want.
1.50 + * The same trick is also applied to create NativeCallStack object for CallsiteHashtableEntry memory allocation.
1.51 + *
1.52 + * Note: C++ object usually aligns to particular alignment, depends on compiler implementation, we declare
1.53 + * the memory as size_t arrays, to ensure the memory is aligned to native machine word alignment.
1.54 + */
1.55 +
1.56 +// Reserve enough memory for NativeCallStack and MallocSiteHashtableEntry objects
1.57 +size_t MallocSiteTable::_hash_entry_allocation_stack[CALC_OBJ_SIZE_IN_TYPE(NativeCallStack, size_t)];
1.58 +size_t MallocSiteTable::_hash_entry_allocation_site[CALC_OBJ_SIZE_IN_TYPE(MallocSiteHashtableEntry, size_t)];
1.59 +
1.60 +// Malloc site hashtable buckets
1.61 +MallocSiteHashtableEntry* MallocSiteTable::_table[MallocSiteTable::table_size];
1.62 +
1.63 +// concurrent access counter
1.64 +volatile int MallocSiteTable::_access_count = 0;
1.65 +
1.66 +// Tracking hashtable contention
1.67 +NOT_PRODUCT(int MallocSiteTable::_peak_count = 0;)
1.68 +
1.69 +
1.70 +/*
1.71 + * Initialize malloc site table.
1.72 + * Hashtable entry is malloc'd, so it can cause infinite recursion.
1.73 + * To avoid above problem, we pre-initialize a hash entry for
1.74 + * this allocation site.
1.75 + * The method is called during C runtime static variable initialization
1.76 + * time, it is in single-threaded mode from JVM perspective.
1.77 + */
1.78 +bool MallocSiteTable::initialize() {
1.79 + assert(sizeof(_hash_entry_allocation_stack) >= sizeof(NativeCallStack), "Sanity Check");
1.80 + assert(sizeof(_hash_entry_allocation_site) >= sizeof(MallocSiteHashtableEntry),
1.81 + "Sanity Check");
1.82 + assert((size_t)table_size <= MAX_MALLOCSITE_TABLE_SIZE, "Hashtable overflow");
1.83 +
1.84 + // Fake the call stack for hashtable entry allocation
1.85 + assert(NMT_TrackingStackDepth > 1, "At least one tracking stack");
1.86 +
1.87 + // Create pseudo call stack for hashtable entry allocation
1.88 + address pc[3];
1.89 + if (NMT_TrackingStackDepth >= 3) {
1.90 + pc[2] = (address)MallocSiteTable::allocation_at;
1.91 + }
1.92 + if (NMT_TrackingStackDepth >= 2) {
1.93 + pc[1] = (address)MallocSiteTable::lookup_or_add;
1.94 + }
1.95 + pc[0] = (address)MallocSiteTable::new_entry;
1.96 +
1.97 + // Instantiate NativeCallStack object, have to use placement new operator. (see comments above)
1.98 + NativeCallStack* stack = ::new ((void*)_hash_entry_allocation_stack)
1.99 + NativeCallStack(pc, MIN2(((int)(sizeof(pc) / sizeof(address))), ((int)NMT_TrackingStackDepth)));
1.100 +
1.101 + // Instantiate hash entry for hashtable entry allocation callsite
1.102 + MallocSiteHashtableEntry* entry = ::new ((void*)_hash_entry_allocation_site)
1.103 + MallocSiteHashtableEntry(*stack);
1.104 +
1.105 + // Add the allocation site to hashtable.
1.106 + int index = hash_to_index(stack->hash());
1.107 + _table[index] = entry;
1.108 +
1.109 + return true;
1.110 +}
1.111 +
1.112 +// Walks entries in the hashtable.
1.113 +// It stops walk if the walker returns false.
1.114 +bool MallocSiteTable::walk(MallocSiteWalker* walker) {
1.115 + MallocSiteHashtableEntry* head;
1.116 + for (int index = 0; index < table_size; index ++) {
1.117 + head = _table[index];
1.118 + while (head != NULL) {
1.119 + if (!walker->do_malloc_site(head->peek())) {
1.120 + return false;
1.121 + }
1.122 + head = (MallocSiteHashtableEntry*)head->next();
1.123 + }
1.124 + }
1.125 + return true;
1.126 +}
1.127 +
1.128 +/*
1.129 + * The hashtable does not have deletion policy on individual entry,
1.130 + * and each linked list node is inserted via compare-and-swap,
1.131 + * so each linked list is stable, the contention only happens
1.132 + * at the end of linked list.
1.133 + * This method should not return NULL under normal circumstance.
1.134 + * If NULL is returned, it indicates:
1.135 + * 1. Out of memory, it cannot allocate new hash entry.
1.136 + * 2. Overflow hash bucket.
1.137 + * Under any of above circumstances, caller should handle the situation.
1.138 + */
1.139 +MallocSite* MallocSiteTable::lookup_or_add(const NativeCallStack& key, size_t* bucket_idx,
1.140 + size_t* pos_idx) {
1.141 + int index = hash_to_index(key.hash());
1.142 + assert(index >= 0, "Negative index");
1.143 + *bucket_idx = (size_t)index;
1.144 + *pos_idx = 0;
1.145 +
1.146 + // First entry for this hash bucket
1.147 + if (_table[index] == NULL) {
1.148 + MallocSiteHashtableEntry* entry = new_entry(key);
1.149 + // OOM check
1.150 + if (entry == NULL) return NULL;
1.151 +
1.152 + // swap in the head
1.153 + if (Atomic::cmpxchg_ptr((void*)entry, (volatile void *)&_table[index], NULL) == NULL) {
1.154 + return entry->data();
1.155 + }
1.156 +
1.157 + delete entry;
1.158 + }
1.159 +
1.160 + MallocSiteHashtableEntry* head = _table[index];
1.161 + while (head != NULL && (*pos_idx) <= MAX_BUCKET_LENGTH) {
1.162 + MallocSite* site = head->data();
1.163 + if (site->equals(key)) {
1.164 + // found matched entry
1.165 + return head->data();
1.166 + }
1.167 +
1.168 + if (head->next() == NULL && (*pos_idx) < MAX_BUCKET_LENGTH) {
1.169 + MallocSiteHashtableEntry* entry = new_entry(key);
1.170 + // OOM check
1.171 + if (entry == NULL) return NULL;
1.172 + if (head->atomic_insert(entry)) {
1.173 + (*pos_idx) ++;
1.174 + return entry->data();
1.175 + }
1.176 + // contended, other thread won
1.177 + delete entry;
1.178 + }
1.179 + head = (MallocSiteHashtableEntry*)head->next();
1.180 + (*pos_idx) ++;
1.181 + }
1.182 + return NULL;
1.183 +}
1.184 +
1.185 +// Access malloc site
1.186 +MallocSite* MallocSiteTable::malloc_site(size_t bucket_idx, size_t pos_idx) {
1.187 + assert(bucket_idx < table_size, "Invalid bucket index");
1.188 + MallocSiteHashtableEntry* head = _table[bucket_idx];
1.189 + for (size_t index = 0; index < pos_idx && head != NULL;
1.190 + index ++, head = (MallocSiteHashtableEntry*)head->next());
1.191 + assert(head != NULL, "Invalid position index");
1.192 + return head->data();
1.193 +}
1.194 +
1.195 +// Allocates MallocSiteHashtableEntry object. Special call stack
1.196 +// (pre-installed allocation site) has to be used to avoid infinite
1.197 +// recursion.
1.198 +MallocSiteHashtableEntry* MallocSiteTable::new_entry(const NativeCallStack& key) {
1.199 + void* p = AllocateHeap(sizeof(MallocSiteHashtableEntry), mtNMT,
1.200 + *hash_entry_allocation_stack(), AllocFailStrategy::RETURN_NULL);
1.201 + return ::new (p) MallocSiteHashtableEntry(key);
1.202 +}
1.203 +
1.204 +void MallocSiteTable::reset() {
1.205 + for (int index = 0; index < table_size; index ++) {
1.206 + MallocSiteHashtableEntry* head = _table[index];
1.207 + _table[index] = NULL;
1.208 + delete_linked_list(head);
1.209 + }
1.210 +}
1.211 +
1.212 +void MallocSiteTable::delete_linked_list(MallocSiteHashtableEntry* head) {
1.213 + MallocSiteHashtableEntry* p;
1.214 + while (head != NULL) {
1.215 + p = head;
1.216 + head = (MallocSiteHashtableEntry*)head->next();
1.217 + if (p != (MallocSiteHashtableEntry*)_hash_entry_allocation_site) {
1.218 + delete p;
1.219 + }
1.220 + }
1.221 +}
1.222 +
1.223 +void MallocSiteTable::shutdown() {
1.224 + AccessLock locker(&_access_count);
1.225 + locker.exclusiveLock();
1.226 + reset();
1.227 +}
1.228 +
1.229 +bool MallocSiteTable::walk_malloc_site(MallocSiteWalker* walker) {
1.230 + assert(walker != NULL, "NuLL walker");
1.231 + AccessLock locker(&_access_count);
1.232 + if (locker.sharedLock()) {
1.233 + NOT_PRODUCT(_peak_count = MAX2(_peak_count, _access_count);)
1.234 + return walk(walker);
1.235 + }
1.236 + return false;
1.237 +}
1.238 +
1.239 +
1.240 +void MallocSiteTable::AccessLock::exclusiveLock() {
1.241 + jint target;
1.242 + jint val;
1.243 +
1.244 + assert(_lock_state != ExclusiveLock, "Can only call once");
1.245 + assert(*_lock >= 0, "Can not content exclusive lock");
1.246 +
1.247 + // make counter negative to block out shared locks
1.248 + do {
1.249 + val = *_lock;
1.250 + target = _MAGIC_ + *_lock;
1.251 + } while (Atomic::cmpxchg(target, _lock, val) != val);
1.252 +
1.253 + // wait for all readers to exit
1.254 + while (*_lock != _MAGIC_) {
1.255 +#ifdef _WINDOWS
1.256 + os::naked_short_sleep(1);
1.257 +#else
1.258 + os::NakedYield();
1.259 +#endif
1.260 + }
1.261 + _lock_state = ExclusiveLock;
1.262 +}
1.263 +
1.264 +
