zgu@3900: /*
zgu@4890:  * Copyright (c) 2012, 2013, Oracle and/or its affiliates. All rights reserved.
zgu@3900:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
zgu@3900:  *
zgu@3900:  * This code is free software; you can redistribute it and/or modify it
zgu@3900:  * under the terms of the GNU General Public License version 2 only, as
zgu@3900:  * published by the Free Software Foundation.
zgu@3900:  *
zgu@3900:  * This code is distributed in the hope that it will be useful, but WITHOUT
zgu@3900:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
zgu@3900:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
zgu@3900:  * version 2 for more details (a copy is included in the LICENSE file that
zgu@3900:  * accompanied this code).
zgu@3900:  *
zgu@3900:  * You should have received a copy of the GNU General Public License version
zgu@3900:  * 2 along with this work; if not, write to the Free Software Foundation,
zgu@3900:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
zgu@3900:  *
zgu@3900:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
zgu@3900:  * or visit www.oracle.com if you need additional information or have any
zgu@3900:  * questions.
zgu@3900:  *
zgu@3900:  */
zgu@3900: #include "precompiled.hpp"
zgu@3900: 
zgu@4400: #include "oops/instanceKlass.hpp"
zgu@3900: #include "runtime/atomic.hpp"
zgu@3900: #include "runtime/interfaceSupport.hpp"
zgu@3900: #include "runtime/mutexLocker.hpp"
zgu@3900: #include "runtime/safepoint.hpp"
zgu@3900: #include "runtime/threadCritical.hpp"
ctornqvi@4512: #include "runtime/vm_operations.hpp"
zgu@3900: #include "services/memPtr.hpp"
zgu@3900: #include "services/memReporter.hpp"
zgu@3900: #include "services/memTracker.hpp"
zgu@3900: #include "utilities/decoder.hpp"
jprovino@5188: #include "utilities/defaultStream.hpp"
zgu@3900: #include "utilities/globalDefinitions.hpp"
zgu@3900: 
zgu@3900: bool NMT_track_callsite = false;
zgu@3900: 
zgu@3900: // walk all 'known' threads at NMT sync point, and collect their recorders
zgu@3900: void SyncThreadRecorderClosure::do_thread(Thread* thread) {
zgu@3900:   assert(SafepointSynchronize::is_at_safepoint(), "Safepoint required");
zgu@3900:   if (thread->is_Java_thread()) {
zgu@3900:     JavaThread* javaThread = (JavaThread*)thread;
zgu@3900:     MemRecorder* recorder = javaThread->get_recorder();
zgu@3900:     if (recorder != NULL) {
zgu@3900:       MemTracker::enqueue_pending_recorder(recorder);
zgu@3900:       javaThread->set_recorder(NULL);
zgu@3900:     }
zgu@3900:   }
zgu@3900:   _thread_count ++;
zgu@3900: }
zgu@3900: 
zgu@3900: 
zgu@4927: MemRecorder* volatile           MemTracker::_global_recorder = NULL;
zgu@3900: MemSnapshot*                    MemTracker::_snapshot = NULL;
zgu@3900: MemBaseline                     MemTracker::_baseline;
zgu@3936: Mutex*                          MemTracker::_query_lock = NULL;
zgu@4927: MemRecorder* volatile           MemTracker::_merge_pending_queue = NULL;
zgu@4927: MemRecorder* volatile           MemTracker::_pooled_recorders = NULL;
zgu@3900: MemTrackWorker*                 MemTracker::_worker_thread = NULL;
zgu@3900: int                             MemTracker::_sync_point_skip_count = 0;
zgu@3900: MemTracker::NMTLevel            MemTracker::_tracking_level = MemTracker::NMT_off;
zgu@3900: volatile MemTracker::NMTStates  MemTracker::_state = NMT_uninited;
zgu@3900: MemTracker::ShutdownReason      MemTracker::_reason = NMT_shutdown_none;
zgu@3900: int                             MemTracker::_thread_count = 255;
zgu@3900: volatile jint                   MemTracker::_pooled_recorder_count = 0;
ctornqvi@4512: volatile unsigned long          MemTracker::_processing_generation = 0;
ctornqvi@4512: volatile bool                   MemTracker::_worker_thread_idle = false;
zgu@5272: volatile jint                   MemTracker::_pending_op_count = 0;
zgu@4810: volatile bool                   MemTracker::_slowdown_calling_thread = false;
zgu@3900: debug_only(intx                 MemTracker::_main_thread_tid = 0;)
zgu@3994: NOT_PRODUCT(volatile jint       MemTracker::_pending_recorder_count = 0;)
zgu@3900: 
zgu@3900: void MemTracker::init_tracking_options(const char* option_line) {
zgu@3900:   _tracking_level = NMT_off;
zgu@4291:   if (strcmp(option_line, "=summary") == 0) {
zgu@3900:     _tracking_level = NMT_summary;
zgu@4291:   } else if (strcmp(option_line, "=detail") == 0) {
jprovino@5188:     // detail relies on a stack-walking ability that may not
jprovino@5188:     // be available depending on platform and/or compiler flags
jprovino@5402: #if PLATFORM_NATIVE_STACK_WALKING_SUPPORTED
jprovino@5188:       _tracking_level = NMT_detail;
jprovino@5402: #else
jprovino@5188:       jio_fprintf(defaultStream::error_stream(),
jprovino@5402:         "NMT detail is not supported on this platform.  Using NMT summary instead.\n");
jprovino@5188:       _tracking_level = NMT_summary;
jprovino@5402: #endif
zgu@4291:   } else if (strcmp(option_line, "=off") != 0) {
zgu@4291:     vm_exit_during_initialization("Syntax error, expecting -XX:NativeMemoryTracking=[off|summary|detail]", NULL);
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // first phase of bootstrapping, when VM is still in single-threaded mode.
zgu@3900: void MemTracker::bootstrap_single_thread() {
zgu@3900:   if (_tracking_level > NMT_off) {
zgu@3900:     assert(_state == NMT_uninited, "wrong state");
zgu@3900: 
zgu@3900:     // NMT is not supported with UseMallocOnly is on. NMT can NOT
zgu@3900:     // handle the amount of malloc data without significantly impacting
zgu@3900:     // runtime performance when this flag is on.
zgu@3900:     if (UseMallocOnly) {
zgu@3900:       shutdown(NMT_use_malloc_only);
zgu@3900:       return;
zgu@3900:     }
zgu@3900: 
zgu@3936:     _query_lock = new (std::nothrow) Mutex(Monitor::max_nonleaf, "NMT_queryLock");
zgu@3936:     if (_query_lock == NULL) {
zgu@3936:       shutdown(NMT_out_of_memory);
zgu@3936:       return;
zgu@3936:     }
zgu@3936: 
zgu@3900:     debug_only(_main_thread_tid = os::current_thread_id();)
zgu@3900:     _state = NMT_bootstrapping_single_thread;
zgu@3900:     NMT_track_callsite = (_tracking_level == NMT_detail && can_walk_stack());
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // second phase of bootstrapping, when VM is about to or already entered multi-theaded mode.
zgu@3900: void MemTracker::bootstrap_multi_thread() {
zgu@3900:   if (_tracking_level > NMT_off && _state == NMT_bootstrapping_single_thread) {
zgu@3900:   // create nmt lock for multi-thread execution
zgu@3900:     assert(_main_thread_tid == os::current_thread_id(), "wrong thread");
zgu@3900:     _state = NMT_bootstrapping_multi_thread;
zgu@3900:     NMT_track_callsite = (_tracking_level == NMT_detail && can_walk_stack());
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // fully start nmt
zgu@3900: void MemTracker::start() {
zgu@3900:   // Native memory tracking is off from command line option
zgu@3900:   if (_tracking_level == NMT_off || shutdown_in_progress()) return;
zgu@3900: 
zgu@3900:   assert(_main_thread_tid == os::current_thread_id(), "wrong thread");
zgu@3900:   assert(_state == NMT_bootstrapping_multi_thread, "wrong state");
zgu@3900: 
zgu@3900:   _snapshot = new (std::nothrow)MemSnapshot();
zgu@4890:   if (_snapshot != NULL) {
zgu@4927:     if (!_snapshot->out_of_memory() && start_worker(_snapshot)) {
zgu@3900:       _state = NMT_started;
zgu@3900:       NMT_track_callsite = (_tracking_level == NMT_detail && can_walk_stack());
zgu@3900:       return;
zgu@3900:     }
zgu@4890: 
zgu@4890:     delete _snapshot;
zgu@4890:     _snapshot = NULL;
zgu@3900:   }
zgu@3900: 
zgu@3900:   // fail to start native memory tracking, shut it down
zgu@3900:   shutdown(NMT_initialization);
zgu@3900: }
zgu@3900: 
zgu@3900: /**
zgu@3900:  * Shutting down native memory tracking.
zgu@3900:  * We can not shutdown native memory tracking immediately, so we just
zgu@3900:  * setup shutdown pending flag, every native memory tracking component
zgu@3900:  * should orderly shut itself down.
zgu@3900:  *
zgu@3900:  * The shutdown sequences:
zgu@3900:  *  1. MemTracker::shutdown() sets MemTracker to shutdown pending state
zgu@3900:  *  2. Worker thread calls MemTracker::final_shutdown(), which transites
zgu@3900:  *     MemTracker to final shutdown state.
zgu@3900:  *  3. At sync point, MemTracker does final cleanup, before sets memory
zgu@3900:  *     tracking level to off to complete shutdown.
zgu@3900:  */
zgu@3900: void MemTracker::shutdown(ShutdownReason reason) {
zgu@3900:   if (_tracking_level == NMT_off) return;
zgu@3900: 
zgu@3900:   if (_state <= NMT_bootstrapping_single_thread) {
zgu@3900:     // we still in single thread mode, there is not contention
zgu@3900:     _state = NMT_shutdown_pending;
zgu@3900:     _reason = reason;
zgu@3900:   } else {
zgu@3900:     // we want to know who initialized shutdown
zgu@3900:     if ((jint)NMT_started == Atomic::cmpxchg((jint)NMT_shutdown_pending,
zgu@3900:                                        (jint*)&_state, (jint)NMT_started)) {
zgu@3900:         _reason = reason;
zgu@3900:     }
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // final phase of shutdown
zgu@3900: void MemTracker::final_shutdown() {
zgu@3900:   // delete all pending recorders and pooled recorders
zgu@3900:   delete_all_pending_recorders();
zgu@3900:   delete_all_pooled_recorders();
zgu@3900: 
zgu@3900:   {
zgu@3900:     // shared baseline and snapshot are the only objects needed to
zgu@3900:     // create query results
zgu@3936:     MutexLockerEx locker(_query_lock, true);
zgu@3900:     // cleanup baseline data and snapshot
zgu@3900:     _baseline.clear();
zgu@3900:     delete _snapshot;
zgu@3900:     _snapshot = NULL;
zgu@3900:   }
zgu@3900: 
zgu@3900:   // shutdown shared decoder instance, since it is only
zgu@3900:   // used by native memory tracking so far.
zgu@3900:   Decoder::shutdown();
zgu@3900: 
zgu@3900:   MemTrackWorker* worker = NULL;
zgu@3900:   {
zgu@3900:     ThreadCritical tc;
zgu@3900:     // can not delete worker inside the thread critical
zgu@3900:     if (_worker_thread != NULL && Thread::current() == _worker_thread) {
zgu@3900:       worker = _worker_thread;
zgu@3900:       _worker_thread = NULL;
zgu@3900:     }
zgu@3900:   }
zgu@3900:   if (worker != NULL) {
zgu@3900:     delete worker;
zgu@3900:   }
zgu@3900:   _state = NMT_final_shutdown;
zgu@3900: }
zgu@3900: 
zgu@3900: // delete all pooled recorders
zgu@3900: void MemTracker::delete_all_pooled_recorders() {
zgu@3900:   // free all pooled recorders
zgu@4927:   MemRecorder* volatile cur_head = _pooled_recorders;
zgu@3900:   if (cur_head != NULL) {
zgu@3900:     MemRecorder* null_ptr = NULL;
zgu@3900:     while (cur_head != NULL && (void*)cur_head != Atomic::cmpxchg_ptr((void*)null_ptr,
zgu@3900:       (void*)&_pooled_recorders, (void*)cur_head)) {
zgu@3900:       cur_head = _pooled_recorders;
zgu@3900:     }
zgu@3900:     if (cur_head != NULL) {
zgu@3900:       delete cur_head;
zgu@3900:       _pooled_recorder_count = 0;
zgu@3900:     }
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // delete all recorders in pending queue
zgu@3900: void MemTracker::delete_all_pending_recorders() {
zgu@3900:   // free all pending recorders
zgu@3900:   MemRecorder* pending_head = get_pending_recorders();
zgu@3900:   if (pending_head != NULL) {
zgu@3900:     delete pending_head;
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * retrieve per-thread recorder of specified thread.
zgu@3900:  * if thread == NULL, it means global recorder
zgu@3900:  */
zgu@3900: MemRecorder* MemTracker::get_thread_recorder(JavaThread* thread) {
zgu@3900:   if (shutdown_in_progress()) return NULL;
zgu@3900: 
zgu@3900:   MemRecorder* rc;
zgu@3900:   if (thread == NULL) {
zgu@3900:     rc = _global_recorder;
zgu@3900:   } else {
zgu@3900:     rc = thread->get_recorder();
zgu@3900:   }
zgu@3900: 
zgu@3900:   if (rc != NULL && rc->is_full()) {
zgu@3900:     enqueue_pending_recorder(rc);
zgu@3900:     rc = NULL;
zgu@3900:   }
zgu@3900: 
zgu@3900:   if (rc == NULL) {
zgu@3900:     rc = get_new_or_pooled_instance();
zgu@3900:     if (thread == NULL) {
zgu@3900:       _global_recorder = rc;
zgu@3900:     } else {
zgu@3900:       thread->set_recorder(rc);
zgu@3900:     }
zgu@3900:   }
zgu@3900:   return rc;
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * get a per-thread recorder from pool, or create a new one if
zgu@3900:  * there is not one available.
zgu@3900:  */
zgu@3900: MemRecorder* MemTracker::get_new_or_pooled_instance() {
zgu@3900:    MemRecorder* cur_head = const_cast<MemRecorder*> (_pooled_recorders);
zgu@3900:    if (cur_head == NULL) {
zgu@3900:      MemRecorder* rec = new (std::nothrow)MemRecorder();
zgu@3900:      if (rec == NULL || rec->out_of_memory()) {
zgu@3900:        shutdown(NMT_out_of_memory);
zgu@3900:        if (rec != NULL) {
zgu@3900:          delete rec;
zgu@3900:          rec = NULL;
zgu@3900:        }
zgu@3900:      }
zgu@3900:      return rec;
zgu@3900:    } else {
zgu@3900:      MemRecorder* next_head = cur_head->next();
zgu@3900:      if ((void*)cur_head != Atomic::cmpxchg_ptr((void*)next_head, (void*)&_pooled_recorders,
zgu@3900:        (void*)cur_head)) {
zgu@3900:        return get_new_or_pooled_instance();
zgu@3900:      }
zgu@3900:      cur_head->set_next(NULL);
zgu@3900:      Atomic::dec(&_pooled_recorder_count);
ctornqvi@4512:      cur_head->set_generation();
zgu@3900:      return cur_head;
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * retrieve all recorders in pending queue, and empty the queue
zgu@3900:  */
zgu@3900: MemRecorder* MemTracker::get_pending_recorders() {
zgu@3900:   MemRecorder* cur_head = const_cast<MemRecorder*>(_merge_pending_queue);
zgu@3900:   MemRecorder* null_ptr = NULL;
zgu@3900:   while ((void*)cur_head != Atomic::cmpxchg_ptr((void*)null_ptr, (void*)&_merge_pending_queue,
zgu@3900:     (void*)cur_head)) {
zgu@3900:     cur_head = const_cast<MemRecorder*>(_merge_pending_queue);
zgu@3900:   }
zgu@3994:   NOT_PRODUCT(Atomic::store(0, &_pending_recorder_count));
zgu@3900:   return cur_head;
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * release a recorder to recorder pool.
zgu@3900:  */
zgu@3900: void MemTracker::release_thread_recorder(MemRecorder* rec) {
zgu@3900:   assert(rec != NULL, "null recorder");
zgu@3900:   // we don't want to pool too many recorders
zgu@3900:   rec->set_next(NULL);
zgu@3900:   if (shutdown_in_progress() || _pooled_recorder_count > _thread_count * 2) {
zgu@3900:     delete rec;
zgu@3900:     return;
zgu@3900:   }
zgu@3900: 
zgu@3900:   rec->clear();
zgu@3900:   MemRecorder* cur_head = const_cast<MemRecorder*>(_pooled_recorders);
zgu@3900:   rec->set_next(cur_head);
zgu@3900:   while ((void*)cur_head != Atomic::cmpxchg_ptr((void*)rec, (void*)&_pooled_recorders,
zgu@3900:     (void*)cur_head)) {
zgu@3900:     cur_head = const_cast<MemRecorder*>(_pooled_recorders);
zgu@3900:     rec->set_next(cur_head);
zgu@3900:   }
zgu@3900:   Atomic::inc(&_pooled_recorder_count);
zgu@3900: }
zgu@3900: 
zgu@3900: // write a record to proper recorder. No lock can be taken from this method
zgu@3900: // down.
zgu@5272: void MemTracker::write_tracking_record(address addr, MEMFLAGS flags,
zgu@5272:     size_t size, jint seq, address pc, JavaThread* thread) {
zgu@3900: 
zgu@3935:     MemRecorder* rc = get_thread_recorder(thread);
zgu@3900:     if (rc != NULL) {
zgu@5272:       rc->record(addr, flags, size, seq, pc);
zgu@3900:     }
zgu@3900: }
zgu@3900: 
zgu@3900: /**
zgu@3900:  * enqueue a recorder to pending queue
zgu@3900:  */
zgu@3900: void MemTracker::enqueue_pending_recorder(MemRecorder* rec) {
zgu@3900:   assert(rec != NULL, "null recorder");
zgu@3900: 
zgu@3900:   // we are shutting down, so just delete it
zgu@3900:   if (shutdown_in_progress()) {
zgu@3900:     rec->set_next(NULL);
zgu@3900:     delete rec;
zgu@3900:     return;
zgu@3900:   }
zgu@3900: 
zgu@3900:   MemRecorder* cur_head = const_cast<MemRecorder*>(_merge_pending_queue);
zgu@3900:   rec->set_next(cur_head);
zgu@3900:   while ((void*)cur_head != Atomic::cmpxchg_ptr((void*)rec, (void*)&_merge_pending_queue,
zgu@3900:     (void*)cur_head)) {
zgu@3900:     cur_head = const_cast<MemRecorder*>(_merge_pending_queue);
zgu@3900:     rec->set_next(cur_head);
zgu@3900:   }
zgu@3994:   NOT_PRODUCT(Atomic::inc(&_pending_recorder_count);)
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * The method is called at global safepoint
zgu@3900:  * during it synchronization process.
zgu@3900:  *   1. enqueue all JavaThreads' per-thread recorders
zgu@3900:  *   2. enqueue global recorder
zgu@3900:  *   3. retrieve all pending recorders
zgu@3900:  *   4. reset global sequence number generator
zgu@3900:  *   5. call worker's sync
zgu@3900:  */
zgu@3900: #define MAX_SAFEPOINTS_TO_SKIP     128
zgu@3900: #define SAFE_SEQUENCE_THRESHOLD    30
zgu@3900: #define HIGH_GENERATION_THRESHOLD  60
zgu@4810: #define MAX_RECORDER_THREAD_RATIO  30
zgu@5410: #define MAX_RECORDER_PER_THREAD    100
zgu@3900: 
zgu@3900: void MemTracker::sync() {
zgu@3900:   assert(_tracking_level > NMT_off, "NMT is not enabled");
zgu@3900:   assert(SafepointSynchronize::is_at_safepoint(), "Safepoint required");
zgu@3900: 
zgu@3900:   // Some GC tests hit large number of safepoints in short period of time
zgu@3900:   // without meaningful activities. We should prevent going to
zgu@3900:   // sync point in these cases, which can potentially exhaust generation buffer.
zgu@3900:   // Here is the factots to determine if we should go into sync point:
zgu@3900:   // 1. not to overflow sequence number
zgu@3900:   // 2. if we are in danger to overflow generation buffer
zgu@3900:   // 3. how many safepoints we already skipped sync point
zgu@3900:   if (_state == NMT_started) {
zgu@3900:     // worker thread is not ready, no one can manage generation
zgu@3900:     // buffer, so skip this safepoint
zgu@3900:     if (_worker_thread == NULL) return;
zgu@3900: 
zgu@3900:     if (_sync_point_skip_count < MAX_SAFEPOINTS_TO_SKIP) {
zgu@3900:       int per_seq_in_use = SequenceGenerator::peek() * 100 / max_jint;
zgu@3900:       int per_gen_in_use = _worker_thread->generations_in_use() * 100 / MAX_GENERATIONS;
zgu@3900:       if (per_seq_in_use < SAFE_SEQUENCE_THRESHOLD && per_gen_in_use >= HIGH_GENERATION_THRESHOLD) {
zgu@3900:         _sync_point_skip_count ++;
zgu@3900:         return;
zgu@3900:       }
zgu@3900:     }
zgu@3900:     {
zgu@3900:       // This method is running at safepoint, with ThreadCritical lock,
zgu@3900:       // it should guarantee that NMT is fully sync-ed.
zgu@3900:       ThreadCritical tc;
zgu@3935: 
zgu@5272:       // We can NOT execute NMT sync-point if there are pending tracking ops.
zgu@5272:       if (_pending_op_count == 0) {
zgu@5272:         SequenceGenerator::reset();
zgu@5272:         _sync_point_skip_count = 0;
zgu@4193: 
zgu@5272:         // walk all JavaThreads to collect recorders
zgu@5272:         SyncThreadRecorderClosure stc;
zgu@5272:         Threads::threads_do(&stc);
zgu@3935: 
zgu@5272:         _thread_count = stc.get_thread_count();
zgu@5272:         MemRecorder* pending_recorders = get_pending_recorders();
zgu@3935: 
zgu@5272:         if (_global_recorder != NULL) {
zgu@5272:           _global_recorder->set_next(pending_recorders);
zgu@5272:           pending_recorders = _global_recorder;
zgu@5272:           _global_recorder = NULL;
zgu@5272:         }
zgu@5272: 
zgu@5272:         // see if NMT has too many outstanding recorder instances, it usually
zgu@5272:         // means that worker thread is lagging behind in processing them.
zgu@5272:         if (!AutoShutdownNMT) {
zgu@5272:           _slowdown_calling_thread = (MemRecorder::_instance_count > MAX_RECORDER_THREAD_RATIO * _thread_count);
zgu@5410:         } else {
zgu@5410:           // If auto shutdown is on, enforce MAX_RECORDER_PER_THREAD threshold to prevent OOM
zgu@5410:           if (MemRecorder::_instance_count >= _thread_count * MAX_RECORDER_PER_THREAD) {
zgu@5410:             shutdown(NMT_out_of_memory);
zgu@5410:           }
zgu@5272:         }
zgu@5272: 
zgu@5272:         // check _worker_thread with lock to avoid racing condition
zgu@5272:         if (_worker_thread != NULL) {
zgu@5272:           _worker_thread->at_sync_point(pending_recorders, InstanceKlass::number_of_instance_classes());
zgu@5272:         }
zgu@5272:         assert(SequenceGenerator::peek() == 1, "Should not have memory activities during sync-point");
zgu@5272:       } else {
zgu@5272:         _sync_point_skip_count ++;
zgu@3900:       }
zgu@3900:     }
zgu@3900:   }
zgu@3900: 
zgu@3900:   // now, it is the time to shut whole things off
zgu@3900:   if (_state == NMT_final_shutdown) {
zgu@3900:     // walk all JavaThreads to delete all recorders
zgu@3900:     SyncThreadRecorderClosure stc;
zgu@3900:     Threads::threads_do(&stc);
zgu@3900:     // delete global recorder
zgu@3900:     {
zgu@3900:       ThreadCritical tc;
zgu@3900:       if (_global_recorder != NULL) {
zgu@3900:         delete _global_recorder;
zgu@3900:         _global_recorder = NULL;
zgu@3900:       }
zgu@3900:     }
zgu@3935:     MemRecorder* pending_recorders = get_pending_recorders();
zgu@3935:     if (pending_recorders != NULL) {
zgu@3935:       delete pending_recorders;
zgu@3935:     }
zgu@3935:     // try at a later sync point to ensure MemRecorder instance drops to zero to
zgu@3935:     // completely shutdown NMT
zgu@3935:     if (MemRecorder::_instance_count == 0) {
zgu@3935:       _state = NMT_shutdown;
zgu@3935:       _tracking_level = NMT_off;
zgu@3935:     }
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * Start worker thread.
zgu@3900:  */
zgu@4927: bool MemTracker::start_worker(MemSnapshot* snapshot) {
zgu@4927:   assert(_worker_thread == NULL && _snapshot != NULL, "Just Check");
zgu@4927:   _worker_thread = new (std::nothrow) MemTrackWorker(snapshot);
zgu@4927:   if (_worker_thread == NULL) {
zgu@4927:     return false;
zgu@4927:   } else if (_worker_thread->has_error()) {
zgu@4927:     delete _worker_thread;
zgu@4927:     _worker_thread = NULL;
zgu@3900:     return false;
zgu@3900:   }
zgu@3900:   _worker_thread->start();
zgu@3900:   return true;
zgu@3900: }
zgu@3900: 
zgu@3900: /*
zgu@3900:  * We need to collect a JavaThread's per-thread recorder
zgu@3900:  * before it exits.
zgu@3900:  */
zgu@3900: void MemTracker::thread_exiting(JavaThread* thread) {
zgu@3900:   if (is_on()) {
zgu@3900:     MemRecorder* rec = thread->get_recorder();
zgu@3900:     if (rec != NULL) {
zgu@3900:       enqueue_pending_recorder(rec);
zgu@3900:       thread->set_recorder(NULL);
zgu@3900:     }
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: // baseline current memory snapshot
zgu@3900: bool MemTracker::baseline() {
zgu@4980:   MutexLocker lock(_query_lock);
zgu@3900:   MemSnapshot* snapshot = get_snapshot();
zgu@3900:   if (snapshot != NULL) {
zgu@3900:     return _baseline.baseline(*snapshot, false);
zgu@3900:   }
zgu@3900:   return false;
zgu@3900: }
zgu@3900: 
zgu@3900: // print memory usage from current snapshot
zgu@3900: bool MemTracker::print_memory_usage(BaselineOutputer& out, size_t unit, bool summary_only) {
zgu@3900:   MemBaseline  baseline;
zgu@4980:   MutexLocker  lock(_query_lock);
zgu@3900:   MemSnapshot* snapshot = get_snapshot();
zgu@3900:   if (snapshot != NULL && baseline.baseline(*snapshot, summary_only)) {
zgu@3900:     BaselineReporter reporter(out, unit);
zgu@3900:     reporter.report_baseline(baseline, summary_only);
zgu@3900:     return true;
zgu@3900:   }
zgu@3900:   return false;
zgu@3900: }
zgu@3900: 
ctornqvi@4512: // Whitebox API for blocking until the current generation of NMT data has been merged
ctornqvi@4512: bool MemTracker::wbtest_wait_for_data_merge() {
ctornqvi@4512:   // NMT can't be shutdown while we're holding _query_lock
zgu@4980:   MutexLocker lock(_query_lock);
ctornqvi@4512:   assert(_worker_thread != NULL, "Invalid query");
ctornqvi@4512:   // the generation at query time, so NMT will spin till this generation is processed
ctornqvi@4512:   unsigned long generation_at_query_time = SequenceGenerator::current_generation();
ctornqvi@4512:   unsigned long current_processing_generation = _processing_generation;
ctornqvi@4512:   // if generation counter overflown
ctornqvi@4512:   bool generation_overflown = (generation_at_query_time < current_processing_generation);
ctornqvi@4512:   long generations_to_wrap = MAX_UNSIGNED_LONG - current_processing_generation;
ctornqvi@4512:   // spin
ctornqvi@4512:   while (!shutdown_in_progress()) {
ctornqvi@4512:     if (!generation_overflown) {
ctornqvi@4512:       if (current_processing_generation > generation_at_query_time) {
ctornqvi@4512:         return true;
ctornqvi@4512:       }
ctornqvi@4512:     } else {
ctornqvi@4512:       assert(generations_to_wrap >= 0, "Sanity check");
ctornqvi@4512:       long current_generations_to_wrap = MAX_UNSIGNED_LONG - current_processing_generation;
ctornqvi@4512:       assert(current_generations_to_wrap >= 0, "Sanity check");
ctornqvi@4512:       // to overflow an unsigned long should take long time, so to_wrap check should be sufficient
ctornqvi@4512:       if (current_generations_to_wrap > generations_to_wrap &&
ctornqvi@4512:           current_processing_generation > generation_at_query_time) {
ctornqvi@4512:         return true;
ctornqvi@4512:       }
ctornqvi@4512:     }
ctornqvi@4512: 
ctornqvi@4512:     // if worker thread is idle, but generation is not advancing, that means
ctornqvi@4512:     // there is not safepoint to let NMT advance generation, force one.
ctornqvi@4512:     if (_worker_thread_idle) {
ctornqvi@4512:       VM_ForceSafepoint vfs;
ctornqvi@4512:       VMThread::execute(&vfs);
ctornqvi@4512:     }
ctornqvi@4512:     MemSnapshot* snapshot = get_snapshot();
ctornqvi@4512:     if (snapshot == NULL) {
ctornqvi@4512:       return false;
ctornqvi@4512:     }
ctornqvi@4512:     snapshot->wait(1000);
ctornqvi@4512:     current_processing_generation = _processing_generation;
ctornqvi@4512:   }
ctornqvi@4512:   // We end up here if NMT is shutting down before our data has been merged
ctornqvi@4512:   return false;
ctornqvi@4512: }
ctornqvi@4512: 
zgu@3900: // compare memory usage between current snapshot and baseline
zgu@3900: bool MemTracker::compare_memory_usage(BaselineOutputer& out, size_t unit, bool summary_only) {
zgu@4980:   MutexLocker lock(_query_lock);
zgu@3900:   if (_baseline.baselined()) {
zgu@3900:     MemBaseline baseline;
zgu@3900:     MemSnapshot* snapshot = get_snapshot();
zgu@3900:     if (snapshot != NULL && baseline.baseline(*snapshot, summary_only)) {
zgu@3900:       BaselineReporter reporter(out, unit);
zgu@3900:       reporter.diff_baselines(baseline, _baseline, summary_only);
zgu@3900:       return true;
zgu@3900:     }
zgu@3900:   }
zgu@3900:   return false;
zgu@3900: }
zgu@3900: 
zgu@3900: #ifndef PRODUCT
zgu@3900: void MemTracker::walk_stack(int toSkip, char* buf, int len) {
zgu@3900:   int cur_len = 0;
zgu@3900:   char tmp[1024];
zgu@3900:   address pc;
zgu@3900: 
zgu@3900:   while (cur_len < len) {
zgu@3900:     pc = os::get_caller_pc(toSkip + 1);
zgu@3900:     if (pc != NULL && os::dll_address_to_function_name(pc, tmp, sizeof(tmp), NULL)) {
zgu@3900:       jio_snprintf(&buf[cur_len], (len - cur_len), "%s\n", tmp);
zgu@3900:       cur_len = (int)strlen(buf);
zgu@3900:     } else {
zgu@3900:       buf[cur_len] = '\0';
zgu@3900:       break;
zgu@3900:     }
zgu@3900:     toSkip ++;
zgu@3900:   }
zgu@3900: }
zgu@3900: 
zgu@3900: void MemTracker::print_tracker_stats(outputStream* st) {
zgu@3900:   st->print_cr("\nMemory Tracker Stats:");
zgu@3900:   st->print_cr("\tMax sequence number = %d", SequenceGenerator::max_seq_num());
zgu@3900:   st->print_cr("\tthead count = %d", _thread_count);
zgu@3900:   st->print_cr("\tArena instance = %d", Arena::_instance_count);
zgu@3900:   st->print_cr("\tpooled recorder count = %d", _pooled_recorder_count);
zgu@3900:   st->print_cr("\tqueued recorder count = %d", _pending_recorder_count);
zgu@3900:   st->print_cr("\tmemory recorder instance count = %d", MemRecorder::_instance_count);
zgu@3900:   if (_worker_thread != NULL) {
zgu@3900:     st->print_cr("\tWorker thread:");
zgu@3900:     st->print_cr("\t\tSync point count = %d", _worker_thread->_sync_point_count);
zgu@3900:     st->print_cr("\t\tpending recorder count = %d", _worker_thread->count_pending_recorders());
zgu@3900:     st->print_cr("\t\tmerge count = %d", _worker_thread->_merge_count);
zgu@3900:   } else {
zgu@3900:     st->print_cr("\tWorker thread is not started");
zgu@3900:   }
zgu@3900:   st->print_cr(" ");
zgu@3900: 
zgu@3900:   if (_snapshot != NULL) {
zgu@3900:     _snapshot->print_snapshot_stats(st);
zgu@3900:   } else {
zgu@3900:     st->print_cr("No snapshot");
zgu@3900:   }
zgu@3900: }
zgu@3900: #endif
zgu@3900: 
zgu@5272: 
zgu@5272: // Tracker Implementation
zgu@5272: 
zgu@5272: /*
zgu@5272:  * Create a tracker.
zgu@5272:  * This is a fairly complicated constructor, as it has to make two important decisions:
zgu@5272:  *   1) Does it need to take ThreadCritical lock to write tracking record
zgu@5272:  *   2) Does it need to pre-reserve a sequence number for the tracking record
zgu@5272:  *
zgu@5272:  * The rules to determine if ThreadCritical is needed:
zgu@5272:  *   1. When nmt is in single-threaded bootstrapping mode, no lock is needed as VM
zgu@5272:  *      still in single thread mode.
zgu@5272:  *   2. For all threads other than JavaThread, ThreadCritical is needed
zgu@5272:  *      to write to recorders to global recorder.
zgu@5272:  *   3. For JavaThreads that are no longer visible by safepoint, also
zgu@5272:  *      need to take ThreadCritical and records are written to global
zgu@5272:  *      recorders, since these threads are NOT walked by Threads.do_thread().
zgu@5272:  *   4. JavaThreads that are running in safepoint-safe states do not stop
zgu@5272:  *      for safepoints, ThreadCritical lock should be taken to write
zgu@5272:  *      memory records.
zgu@5272:  *   5. JavaThreads that are running in VM state do not need any lock and
zgu@5272:  *      records are written to per-thread recorders.
zgu@5272:  *   6. For a thread has yet to attach VM 'Thread', they need to take
zgu@5272:  *      ThreadCritical to write to global recorder.
zgu@5272:  *
zgu@5272:  *  The memory operations that need pre-reserve sequence numbers:
zgu@5272:  *    The memory operations that "release" memory blocks and the
zgu@5272:  *    operations can fail, need to pre-reserve sequence number. They
zgu@5272:  *    are realloc, uncommit and release.
zgu@5272:  *
zgu@5272:  *  The reason for pre-reserve sequence number, is to prevent race condition:
zgu@5272:  *    Thread 1                      Thread 2
zgu@5272:  *    <release>
zgu@5272:  *                                  <allocate>
zgu@5272:  *                                  <write allocate record>
zgu@5272:  *   <write release record>
zgu@5272:  *   if Thread 2 happens to obtain the memory address Thread 1 just released,
zgu@5272:  *   then NMT can mistakenly report the memory is free.
zgu@5272:  *
zgu@5272:  *  Noticeably, free() does not need pre-reserve sequence number, because the call
zgu@5272:  *  does not fail, so we can alway write "release" record before the memory is actaully
zgu@5272:  *  freed.
zgu@5272:  *
zgu@5272:  *  For realloc, uncommit and release, following coding pattern should be used:
zgu@5272:  *
zgu@5272:  *     MemTracker::Tracker tkr = MemTracker::get_realloc_tracker();
zgu@5272:  *     ptr = ::realloc(...);
zgu@5272:  *     if (ptr == NULL) {
zgu@5272:  *       tkr.record(...)
zgu@5272:  *     } else {
zgu@5272:  *       tkr.discard();
zgu@5272:  *     }
zgu@5272:  *
zgu@5272:  *     MemTracker::Tracker tkr = MemTracker::get_virtual_memory_uncommit_tracker();
zgu@5272:  *     if (uncommit(...)) {
zgu@5272:  *       tkr.record(...);
zgu@5272:  *     } else {
zgu@5272:  *       tkr.discard();
zgu@5272:  *     }
zgu@5272:  *
zgu@5272:  *     MemTracker::Tracker tkr = MemTracker::get_virtual_memory_release_tracker();
zgu@5272:  *     if (release(...)) {
zgu@5272:  *       tkr.record(...);
zgu@5272:  *     } else {
zgu@5272:  *       tkr.discard();
zgu@5272:  *     }
zgu@5272:  *
zgu@5272:  * Since pre-reserved sequence number is only good for the generation that it is acquired,
zgu@5272:  * when there is pending Tracker that reserved sequence number, NMT sync-point has
zgu@5272:  * to be skipped to prevent from advancing generation. This is done by inc and dec
zgu@5272:  * MemTracker::_pending_op_count, when MemTracker::_pending_op_count > 0, NMT sync-point is skipped.
zgu@5272:  * Not all pre-reservation of sequence number will increment pending op count. For JavaThreads
zgu@5272:  * that honor safepoints, safepoint can not occur during the memory operations, so the
zgu@5272:  * pre-reserved sequence number won't cross the generation boundry.
zgu@5272:  */
zgu@5272: MemTracker::Tracker::Tracker(MemoryOperation op, Thread* thr) {
zgu@5272:   _op = NoOp;
zgu@5272:   _seq = 0;
zgu@5272:   if (MemTracker::is_on()) {
zgu@5272:     _java_thread = NULL;
zgu@5272:     _op = op;
zgu@5272: 
zgu@5272:     // figure out if ThreadCritical lock is needed to write this operation
zgu@5272:     // to MemTracker
zgu@5272:     if (MemTracker::is_single_threaded_bootstrap()) {
zgu@5272:       thr = NULL;
zgu@5272:     } else if (thr == NULL) {
zgu@5272:       // don't use Thread::current(), since it is possible that
zgu@5272:       // the calling thread has yet to attach to VM 'Thread',
zgu@5272:       // which will result assertion failure
zgu@5272:       thr = ThreadLocalStorage::thread();
zgu@5272:     }
zgu@5272: 
zgu@5272:     if (thr != NULL) {
zgu@5272:       // Check NMT load
zgu@5272:       MemTracker::check_NMT_load(thr);
zgu@5272: 
zgu@5272:       if (thr->is_Java_thread() && ((JavaThread*)thr)->is_safepoint_visible()) {
zgu@5272:         _java_thread = (JavaThread*)thr;
zgu@5272:         JavaThreadState  state = _java_thread->thread_state();
zgu@5272:         // JavaThreads that are safepoint safe, can run through safepoint,
zgu@5272:         // so ThreadCritical is needed to ensure no threads at safepoint create
zgu@5272:         // new records while the records are being gathered and the sequence number is changing
zgu@5272:         _need_thread_critical_lock =
zgu@5272:           SafepointSynchronize::safepoint_safe(_java_thread, state);
zgu@5272:       } else {
zgu@5272:         _need_thread_critical_lock = true;
zgu@5272:       }
zgu@5272:     } else {
zgu@5272:        _need_thread_critical_lock
zgu@5272:          = !MemTracker::is_single_threaded_bootstrap();
zgu@5272:     }
zgu@5272: 
zgu@5272:     // see if we need to pre-reserve sequence number for this operation
zgu@5272:     if (_op == Realloc || _op == Uncommit || _op == Release) {
zgu@5272:       if (_need_thread_critical_lock) {
zgu@5272:         ThreadCritical tc;
zgu@5272:         MemTracker::inc_pending_op_count();
zgu@5272:         _seq = SequenceGenerator::next();
zgu@5272:       } else {
zgu@5272:         // for the threads that honor safepoints, no safepoint can occur
zgu@5272:         // during the lifespan of tracker, so we don't need to increase
zgu@5272:         // pending op count.
zgu@5272:         _seq = SequenceGenerator::next();
zgu@5272:       }
zgu@5272:     }
zgu@5272:   }
zgu@5272: }
zgu@5272: 
zgu@5272: void MemTracker::Tracker::discard() {
zgu@5272:   if (MemTracker::is_on() && _seq != 0) {
zgu@5272:     if (_need_thread_critical_lock) {
zgu@5272:       ThreadCritical tc;
zgu@5272:       MemTracker::dec_pending_op_count();
zgu@5272:     }
zgu@5272:     _seq = 0;
zgu@5272:   }
zgu@5272: }
zgu@5272: 
zgu@5272: 
zgu@5272: void MemTracker::Tracker::record(address old_addr, address new_addr, size_t size,
zgu@5272:   MEMFLAGS flags, address pc) {
zgu@5272:   assert(old_addr != NULL && new_addr != NULL, "Sanity check");
zgu@5272:   assert(_op == Realloc || _op == NoOp, "Wrong call");
zgu@5272:   if (MemTracker::is_on() && NMT_CAN_TRACK(flags) && _op != NoOp) {
zgu@5272:     assert(_seq > 0, "Need pre-reserve sequence number");
zgu@5272:     if (_need_thread_critical_lock) {
zgu@5272:       ThreadCritical tc;
zgu@5272:       // free old address, use pre-reserved sequence number
zgu@5272:       MemTracker::write_tracking_record(old_addr, MemPointerRecord::free_tag(),
zgu@5272:         0, _seq, pc, _java_thread);
zgu@5272:       MemTracker::write_tracking_record(new_addr, flags | MemPointerRecord::malloc_tag(),
zgu@5272:         size, SequenceGenerator::next(), pc, _java_thread);
zgu@5272:       // decrement MemTracker pending_op_count
zgu@5272:       MemTracker::dec_pending_op_count();
zgu@5272:     } else {
zgu@5272:       // free old address, use pre-reserved sequence number
zgu@5272:       MemTracker::write_tracking_record(old_addr, MemPointerRecord::free_tag(),
zgu@5272:         0, _seq, pc, _java_thread);
zgu@5272:       MemTracker::write_tracking_record(new_addr, flags | MemPointerRecord::malloc_tag(),
zgu@5272:         size, SequenceGenerator::next(), pc, _java_thread);
zgu@5272:     }
zgu@5272:     _seq = 0;
zgu@5272:   }
zgu@5272: }
zgu@5272: 
zgu@5272: void MemTracker::Tracker::record(address addr, size_t size, MEMFLAGS flags, address pc) {
zgu@5272:   // OOM already?
zgu@5272:   if (addr == NULL) return;
zgu@5272: 
zgu@5272:   if (MemTracker::is_on() && NMT_CAN_TRACK(flags) && _op != NoOp) {
zgu@5272:     bool pre_reserved_seq = (_seq != 0);
zgu@5272:     address  pc = CALLER_CALLER_PC;
zgu@5272:     MEMFLAGS orig_flags = flags;
zgu@5272: 
zgu@5272:     // or the tagging flags
zgu@5272:     switch(_op) {
zgu@5272:       case Malloc:
zgu@5272:         flags |= MemPointerRecord::malloc_tag();
zgu@5272:         break;
zgu@5272:       case Free:
zgu@5272:         flags = MemPointerRecord::free_tag();
zgu@5272:         break;
zgu@5272:       case Realloc:
zgu@5272:         fatal("Use the other Tracker::record()");
zgu@5272:         break;
zgu@5272:       case Reserve:
zgu@5272:       case ReserveAndCommit:
zgu@5272:         flags |= MemPointerRecord::virtual_memory_reserve_tag();
zgu@5272:         break;
zgu@5272:       case Commit:
zgu@5272:         flags = MemPointerRecord::virtual_memory_commit_tag();
zgu@5272:         break;
zgu@5272:       case Type:
zgu@5272:         flags |= MemPointerRecord::virtual_memory_type_tag();
zgu@5272:         break;
zgu@5272:       case Uncommit:
zgu@5272:         assert(pre_reserved_seq, "Need pre-reserve sequence number");
zgu@5272:         flags = MemPointerRecord::virtual_memory_uncommit_tag();
zgu@5272:         break;
zgu@5272:       case Release:
zgu@5272:         assert(pre_reserved_seq, "Need pre-reserve sequence number");
zgu@5272:         flags = MemPointerRecord::virtual_memory_release_tag();
zgu@5272:         break;
zgu@5272:       case ArenaSize:
zgu@5272:         // a bit of hack here, add a small postive offset to arena
zgu@5272:         // address for its size record, so the size record is sorted
zgu@5272:         // right after arena record.
zgu@5272:         flags = MemPointerRecord::arena_size_tag();
zgu@5272:         addr += sizeof(void*);
zgu@5272:         break;
zgu@5272:       case StackRelease:
zgu@5272:         flags = MemPointerRecord::virtual_memory_release_tag();
zgu@5272:         break;
zgu@5272:       default:
zgu@5272:         ShouldNotReachHere();
zgu@5272:     }
zgu@5272: 
zgu@5272:     // write memory tracking record
zgu@5272:     if (_need_thread_critical_lock) {
zgu@5272:       ThreadCritical tc;
zgu@5272:       if (_seq == 0) _seq = SequenceGenerator::next();
zgu@5272:       MemTracker::write_tracking_record(addr, flags, size, _seq, pc, _java_thread);
zgu@5272:       if (_op == ReserveAndCommit) {
zgu@5272:         MemTracker::write_tracking_record(addr, orig_flags | MemPointerRecord::virtual_memory_commit_tag(),
zgu@5272:           size, SequenceGenerator::next(), pc, _java_thread);
zgu@5272:       }
zgu@5272:       if (pre_reserved_seq) MemTracker::dec_pending_op_count();
zgu@5272:     } else {
zgu@5272:       if (_seq == 0) _seq = SequenceGenerator::next();
zgu@5272:       MemTracker::write_tracking_record(addr, flags, size, _seq, pc, _java_thread);
zgu@5272:       if (_op == ReserveAndCommit) {
zgu@5272:         MemTracker::write_tracking_record(addr, orig_flags | MemPointerRecord::virtual_memory_commit_tag(),
zgu@5272:           size, SequenceGenerator::next(), pc, _java_thread);
zgu@5272:       }
zgu@5272:     }
zgu@5272:     _seq = 0;
zgu@5272:   }
zgu@5272: }
zgu@5272: