hotspot/src/share/vm/runtime/sweeper.cpp
author anoll
Fri, 27 Sep 2013 10:50:55 +0200
changeset 20290 2127dc70bce9
parent 20019 3e9220f93c61
child 21575 6a9645992cee
permissions -rw-r--r--
8020151: PSR:PERF Large performance regressions when code cache is filled Summary: Code cache sweeping based on method hotness; removed speculatively disconnect Reviewed-by: kvn, iveresov

/*
 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 * or visit www.oracle.com if you need additional information or have any
 * questions.
 *
 */

#include "precompiled.hpp"
#include "code/codeCache.hpp"
#include "code/compiledIC.hpp"
#include "code/icBuffer.hpp"
#include "code/nmethod.hpp"
#include "compiler/compileBroker.hpp"
#include "memory/resourceArea.hpp"
#include "oops/method.hpp"
#include "runtime/atomic.hpp"
#include "runtime/compilationPolicy.hpp"
#include "runtime/mutexLocker.hpp"
#include "runtime/os.hpp"
#include "runtime/sweeper.hpp"
#include "runtime/vm_operations.hpp"
#include "trace/tracing.hpp"
#include "utilities/events.hpp"
#include "utilities/xmlstream.hpp"

#ifdef ASSERT

#define SWEEP(nm) record_sweep(nm, __LINE__)
// Sweeper logging code
class SweeperRecord {
 public:
  int traversal;
  int invocation;
  int compile_id;
  long traversal_mark;
  int state;
  const char* kind;
  address vep;
  address uep;
  int line;

  void print() {
      tty->print_cr("traversal = %d invocation = %d compile_id = %d %s uep = " PTR_FORMAT " vep = "
                    PTR_FORMAT " state = %d traversal_mark %d line = %d",
                    traversal,
                    invocation,
                    compile_id,
                    kind == NULL ? "" : kind,
                    uep,
                    vep,
                    state,
                    traversal_mark,
                    line);
  }
};

static int _sweep_index = 0;
static SweeperRecord* _records = NULL;

void NMethodSweeper::report_events(int id, address entry) {
  if (_records != NULL) {
    for (int i = _sweep_index; i < SweeperLogEntries; i++) {
      if (_records[i].uep == entry ||
          _records[i].vep == entry ||
          _records[i].compile_id == id) {
        _records[i].print();
      }
    }
    for (int i = 0; i < _sweep_index; i++) {
      if (_records[i].uep == entry ||
          _records[i].vep == entry ||
          _records[i].compile_id == id) {
        _records[i].print();
      }
    }
  }
}

void NMethodSweeper::report_events() {
  if (_records != NULL) {
    for (int i = _sweep_index; i < SweeperLogEntries; i++) {
      // skip empty records
      if (_records[i].vep == NULL) continue;
      _records[i].print();
    }
    for (int i = 0; i < _sweep_index; i++) {
      // skip empty records
      if (_records[i].vep == NULL) continue;
      _records[i].print();
    }
  }
}

void NMethodSweeper::record_sweep(nmethod* nm, int line) {
  if (_records != NULL) {
    _records[_sweep_index].traversal = _traversals;
    _records[_sweep_index].traversal_mark = nm->_stack_traversal_mark;
    _records[_sweep_index].invocation = _invocations;
    _records[_sweep_index].compile_id = nm->compile_id();
    _records[_sweep_index].kind = nm->compile_kind();
    _records[_sweep_index].state = nm->_state;
    _records[_sweep_index].vep = nm->verified_entry_point();
    _records[_sweep_index].uep = nm->entry_point();
    _records[_sweep_index].line = line;

    _sweep_index = (_sweep_index + 1) % SweeperLogEntries;
  }
}
#else
#define SWEEP(nm)
#endif

nmethod*  NMethodSweeper::_current         = NULL; // Current nmethod
long      NMethodSweeper::_traversals      = 0;    // Nof. stack traversals performed
int       NMethodSweeper::_seen            = 0;    // Nof. nmethods we have currently processed in current pass of CodeCache
int       NMethodSweeper::_flushed_count   = 0;    // Nof. nmethods flushed in current sweep
int       NMethodSweeper::_zombified_count = 0;    // Nof. nmethods made zombie in current sweep
int       NMethodSweeper::_marked_count    = 0;    // Nof. nmethods marked for reclaim in current sweep

volatile int NMethodSweeper::_invocations   = 0; // Nof. invocations left until we are completed with this pass
volatile int NMethodSweeper::_sweep_started = 0; // Whether a sweep is in progress.

jint      NMethodSweeper::_locked_seen               = 0;
jint      NMethodSweeper::_not_entrant_seen_on_stack = 0;
bool      NMethodSweeper::_request_mark_phase        = false;

int       NMethodSweeper::_total_nof_methods_reclaimed = 0;
jlong     NMethodSweeper::_total_time_sweeping         = 0;
jlong     NMethodSweeper::_total_time_this_sweep       = 0;
jlong     NMethodSweeper::_peak_sweep_time             = 0;
jlong     NMethodSweeper::_peak_sweep_fraction_time    = 0;
int       NMethodSweeper::_hotness_counter_reset_val   = 0;


class MarkActivationClosure: public CodeBlobClosure {
public:
  virtual void do_code_blob(CodeBlob* cb) {
    if (cb->is_nmethod()) {
      nmethod* nm = (nmethod*)cb;
      nm->set_hotness_counter(NMethodSweeper::hotness_counter_reset_val());
      // If we see an activation belonging to a non_entrant nmethod, we mark it.
      if (nm->is_not_entrant()) {
        nm->mark_as_seen_on_stack();
      }
    }
  }
};
static MarkActivationClosure mark_activation_closure;

class SetHotnessClosure: public CodeBlobClosure {
public:
  virtual void do_code_blob(CodeBlob* cb) {
    if (cb->is_nmethod()) {
      nmethod* nm = (nmethod*)cb;
      nm->set_hotness_counter(NMethodSweeper::hotness_counter_reset_val());
    }
  }
};
static SetHotnessClosure set_hotness_closure;


int NMethodSweeper::hotness_counter_reset_val() {
  if (_hotness_counter_reset_val == 0) {
    _hotness_counter_reset_val = (ReservedCodeCacheSize < M) ? 1 : (ReservedCodeCacheSize / M) * 2;
  }
  return _hotness_counter_reset_val;
}
bool NMethodSweeper::sweep_in_progress() {
  return (_current != NULL);
}

// Scans the stacks of all Java threads and marks activations of not-entrant methods.
// No need to synchronize access, since 'mark_active_nmethods' is always executed at a
// safepoint.
void NMethodSweeper::mark_active_nmethods() {
  assert(SafepointSynchronize::is_at_safepoint(), "must be executed at a safepoint");
  // If we do not want to reclaim not-entrant or zombie methods there is no need
  // to scan stacks
  if (!MethodFlushing) {
    return;
  }

  // Check for restart
  assert(CodeCache::find_blob_unsafe(_current) == _current, "Sweeper nmethod cached state invalid");
  if (!sweep_in_progress() && need_marking_phase()) {
    _seen        = 0;
    _invocations = NmethodSweepFraction;
    _current     = CodeCache::first_nmethod();
    _traversals  += 1;
    _total_time_this_sweep = 0;

    if (PrintMethodFlushing) {
      tty->print_cr("### Sweep: stack traversal %d", _traversals);
    }
    Threads::nmethods_do(&mark_activation_closure);

    // reset the flags since we started a scan from the beginning.
    reset_nmethod_marking();
    _locked_seen = 0;
    _not_entrant_seen_on_stack = 0;
  } else {
    // Only set hotness counter
    Threads::nmethods_do(&set_hotness_closure);
  }

  OrderAccess::storestore();
}

void NMethodSweeper::possibly_sweep() {
  assert(JavaThread::current()->thread_state() == _thread_in_vm, "must run in vm mode");
  if (!MethodFlushing || !sweep_in_progress()) {
    return;
  }

  if (_invocations > 0) {
    // Only one thread at a time will sweep
    jint old = Atomic::cmpxchg( 1, &_sweep_started, 0 );
    if (old != 0) {
      return;
    }
#ifdef ASSERT
    if (LogSweeper && _records == NULL) {
      // Create the ring buffer for the logging code
      _records = NEW_C_HEAP_ARRAY(SweeperRecord, SweeperLogEntries, mtGC);
      memset(_records, 0, sizeof(SweeperRecord) * SweeperLogEntries);
    }
#endif
    if (_invocations > 0) {
      sweep_code_cache();
      _invocations--;
    }
    _sweep_started = 0;
  }
}

void NMethodSweeper::sweep_code_cache() {

  jlong sweep_start_counter = os::elapsed_counter();

  _flushed_count   = 0;
  _zombified_count = 0;
  _marked_count    = 0;

  if (PrintMethodFlushing && Verbose) {
    tty->print_cr("### Sweep at %d out of %d. Invocations left: %d", _seen, CodeCache::nof_nmethods(), _invocations);
  }

  if (!CompileBroker::should_compile_new_jobs()) {
    // If we have turned off compilations we might as well do full sweeps
    // in order to reach the clean state faster. Otherwise the sleeping compiler
    // threads will slow down sweeping.
    _invocations = 1;
  }

  // We want to visit all nmethods after NmethodSweepFraction
  // invocations so divide the remaining number of nmethods by the
  // remaining number of invocations.  This is only an estimate since
  // the number of nmethods changes during the sweep so the final
  // stage must iterate until it there are no more nmethods.
  int todo = (CodeCache::nof_nmethods() - _seen) / _invocations;
  int swept_count = 0;


  assert(!SafepointSynchronize::is_at_safepoint(), "should not be in safepoint when we get here");
  assert(!CodeCache_lock->owned_by_self(), "just checking");

  int freed_memory = 0;
  {
    MutexLockerEx mu(CodeCache_lock, Mutex::_no_safepoint_check_flag);

    // The last invocation iterates until there are no more nmethods
    for (int i = 0; (i < todo || _invocations == 1) && _current != NULL; i++) {
      swept_count++;
      if (SafepointSynchronize::is_synchronizing()) { // Safepoint request
        if (PrintMethodFlushing && Verbose) {
          tty->print_cr("### Sweep at %d out of %d, invocation: %d, yielding to safepoint", _seen, CodeCache::nof_nmethods(), _invocations);
        }
        MutexUnlockerEx mu(CodeCache_lock, Mutex::_no_safepoint_check_flag);

        assert(Thread::current()->is_Java_thread(), "should be java thread");
        JavaThread* thread = (JavaThread*)Thread::current();
        ThreadBlockInVM tbivm(thread);
        thread->java_suspend_self();
      }
      // Since we will give up the CodeCache_lock, always skip ahead
      // to the next nmethod.  Other blobs can be deleted by other
      // threads but nmethods are only reclaimed by the sweeper.
      nmethod* next = CodeCache::next_nmethod(_current);

      // Now ready to process nmethod and give up CodeCache_lock
      {
        MutexUnlockerEx mu(CodeCache_lock, Mutex::_no_safepoint_check_flag);
        freed_memory += process_nmethod(_current);
      }
      _seen++;
      _current = next;
    }
  }

  assert(_invocations > 1 || _current == NULL, "must have scanned the whole cache");

  if (!sweep_in_progress() && !need_marking_phase() && (_locked_seen || _not_entrant_seen_on_stack)) {
    // we've completed a scan without making progress but there were
    // nmethods we were unable to process either because they were
    // locked or were still on stack. We don't have to aggressively
    // clean them up so just stop scanning. We could scan once more
    // but that complicates the control logic and it's unlikely to
    // matter much.
    if (PrintMethodFlushing) {
      tty->print_cr("### Couldn't make progress on some nmethods so stopping sweep");
    }
  }

  jlong sweep_end_counter = os::elapsed_counter();
  jlong sweep_time = sweep_end_counter - sweep_start_counter;
  _total_time_sweeping  += sweep_time;
  _total_time_this_sweep += sweep_time;
  _peak_sweep_fraction_time = MAX2(sweep_time, _peak_sweep_fraction_time);
  _total_nof_methods_reclaimed += _flushed_count;

  EventSweepCodeCache event(UNTIMED);
  if (event.should_commit()) {
    event.set_starttime(sweep_start_counter);
    event.set_endtime(sweep_end_counter);
    event.set_sweepIndex(_traversals);
    event.set_sweepFractionIndex(NmethodSweepFraction - _invocations + 1);
    event.set_sweptCount(swept_count);
    event.set_flushedCount(_flushed_count);
    event.set_markedCount(_marked_count);
    event.set_zombifiedCount(_zombified_count);
    event.commit();
  }

#ifdef ASSERT
  if(PrintMethodFlushing) {
    tty->print_cr("### sweeper:      sweep time(%d): " INT64_FORMAT, _invocations, (jlong)sweep_time);
  }
#endif

  if (_invocations == 1) {
    _peak_sweep_time = MAX2(_peak_sweep_time, _total_time_this_sweep);
    log_sweep("finished");
  }

  // Sweeper is the only case where memory is released, check here if it
  // is time to restart the compiler. Only checking if there is a certain
  // amount of free memory in the code cache might lead to re-enabling
  // compilation although no memory has been released. For example, there are
  // cases when compilation was disabled although there is 4MB (or more) free
  // memory in the code cache. The reason is code cache fragmentation. Therefore,
  // it only makes sense to re-enable compilation if we have actually freed memory.
  // Note that typically several kB are released for sweeping 16MB of the code
  // cache. As a result, 'freed_memory' > 0 to restart the compiler.
  if (UseCodeCacheFlushing && (!CompileBroker::should_compile_new_jobs() && (freed_memory > 0))) {
    CompileBroker::set_should_compile_new_jobs(CompileBroker::run_compilation);
    log_sweep("restart_compiler");
  }
}

class NMethodMarker: public StackObj {
 private:
  CompilerThread* _thread;
 public:
  NMethodMarker(nmethod* nm) {
    _thread = CompilerThread::current();
    if (!nm->is_zombie() && !nm->is_unloaded()) {
      // Only expose live nmethods for scanning
      _thread->set_scanned_nmethod(nm);
    }
  }
  ~NMethodMarker() {
    _thread->set_scanned_nmethod(NULL);
  }
};

void NMethodSweeper::release_nmethod(nmethod *nm) {
  // Clean up any CompiledICHolders
  {
    ResourceMark rm;
    MutexLocker ml_patch(CompiledIC_lock);
    RelocIterator iter(nm);
    while (iter.next()) {
      if (iter.type() == relocInfo::virtual_call_type) {
        CompiledIC::cleanup_call_site(iter.virtual_call_reloc());
      }
    }
  }

  MutexLockerEx mu(CodeCache_lock, Mutex::_no_safepoint_check_flag);
  nm->flush();
}

int NMethodSweeper::process_nmethod(nmethod *nm) {
  assert(!CodeCache_lock->owned_by_self(), "just checking");

  int freed_memory = 0;
  // Make sure this nmethod doesn't get unloaded during the scan,
  // since safepoints may happen during acquired below locks.
  NMethodMarker nmm(nm);
  SWEEP(nm);

  // Skip methods that are currently referenced by the VM
  if (nm->is_locked_by_vm()) {
    // But still remember to clean-up inline caches for alive nmethods
    if (nm->is_alive()) {
      // Clean inline caches that point to zombie/non-entrant methods
      MutexLocker cl(CompiledIC_lock);
      nm->cleanup_inline_caches();
      SWEEP(nm);
    } else {
      _locked_seen++;
      SWEEP(nm);
    }
    return freed_memory;
  }

  if (nm->is_zombie()) {
    // If it is the first time we see nmethod then we mark it. Otherwise,
    // we reclaim it. When we have seen a zombie method twice, we know that
    // there are no inline caches that refer to it.
    if (nm->is_marked_for_reclamation()) {
      assert(!nm->is_locked_by_vm(), "must not flush locked nmethods");
      if (PrintMethodFlushing && Verbose) {
        tty->print_cr("### Nmethod %3d/" PTR_FORMAT " (marked for reclamation) being flushed", nm->compile_id(), nm);
      }
      freed_memory = nm->total_size();
      release_nmethod(nm);
      _flushed_count++;
    } else {
      if (PrintMethodFlushing && Verbose) {
        tty->print_cr("### Nmethod %3d/" PTR_FORMAT " (zombie) being marked for reclamation", nm->compile_id(), nm);
      }
      nm->mark_for_reclamation();
      request_nmethod_marking();
      _marked_count++;
      SWEEP(nm);
    }
  } else if (nm->is_not_entrant()) {
    // If there are no current activations of this method on the
    // stack we can safely convert it to a zombie method
    if (nm->can_not_entrant_be_converted()) {
      if (PrintMethodFlushing && Verbose) {
        tty->print_cr("### Nmethod %3d/" PTR_FORMAT " (not entrant) being made zombie", nm->compile_id(), nm);
      }
      nm->make_zombie();
      request_nmethod_marking();
      _zombified_count++;
      SWEEP(nm);
    } else {
      // Still alive, clean up its inline caches
      MutexLocker cl(CompiledIC_lock);
      nm->cleanup_inline_caches();
      // we coudn't transition this nmethod so don't immediately
      // request a rescan.  If this method stays on the stack for a
      // long time we don't want to keep rescanning the code cache.
      _not_entrant_seen_on_stack++;
      SWEEP(nm);
    }
  } else if (nm->is_unloaded()) {
    // Unloaded code, just make it a zombie
    if (PrintMethodFlushing && Verbose) {
      tty->print_cr("### Nmethod %3d/" PTR_FORMAT " (unloaded) being made zombie", nm->compile_id(), nm);
    }
    if (nm->is_osr_method()) {
      SWEEP(nm);
      // No inline caches will ever point to osr methods, so we can just remove it
      freed_memory = nm->total_size();
      release_nmethod(nm);
      _flushed_count++;
    } else {
      nm->make_zombie();
      request_nmethod_marking();
      _zombified_count++;
      SWEEP(nm);
    }
  } else {
    if (UseCodeCacheFlushing) {
      if (!nm->is_locked_by_vm() && !nm->is_osr_method() && !nm->is_native_method()) {
        // Do not make native methods and OSR-methods not-entrant
        nm->dec_hotness_counter();
        // Get the initial value of the hotness counter. This value depends on the
        // ReservedCodeCacheSize
        int reset_val = hotness_counter_reset_val();
        int time_since_reset = reset_val - nm->hotness_counter();
        double threshold = -reset_val + (CodeCache::reverse_free_ratio() * NmethodSweepActivity);
        // The less free space in the code cache we have - the bigger reverse_free_ratio() is.
        // I.e., 'threshold' increases with lower available space in the code cache and a higher
        // NmethodSweepActivity. If the current hotness counter - which decreases from its initial
        // value until it is reset by stack walking - is smaller than the computed threshold, the
        // corresponding nmethod is considered for removal.
        if ((NmethodSweepActivity > 0) && (nm->hotness_counter() < threshold) && (time_since_reset > 10)) {
          // A method is marked as not-entrant if the method is
          // 1) 'old enough': nm->hotness_counter() < threshold
          // 2) The method was in_use for a minimum amount of time: (time_since_reset > 10)
          //    The second condition is necessary if we are dealing with very small code cache
          //    sizes (e.g., <10m) and the code cache size is too small to hold all hot methods.
          //    The second condition ensures that methods are not immediately made not-entrant
          //    after compilation.
          nm->make_not_entrant();
          request_nmethod_marking();
        }
      }
    }
    // Clean-up all inline caches that point to zombie/non-reentrant methods
    MutexLocker cl(CompiledIC_lock);
    nm->cleanup_inline_caches();
    SWEEP(nm);
  }
  return freed_memory;
}

// Print out some state information about the current sweep and the
// state of the code cache if it's requested.
void NMethodSweeper::log_sweep(const char* msg, const char* format, ...) {
  if (PrintMethodFlushing) {
    stringStream s;
    // Dump code cache state into a buffer before locking the tty,
    // because log_state() will use locks causing lock conflicts.
    CodeCache::log_state(&s);

    ttyLocker ttyl;
    tty->print("### sweeper: %s ", msg);
    if (format != NULL) {
      va_list ap;
      va_start(ap, format);
      tty->vprint(format, ap);
      va_end(ap);
    }
    tty->print_cr(s.as_string());
  }

  if (LogCompilation && (xtty != NULL)) {
    stringStream s;
    // Dump code cache state into a buffer before locking the tty,
    // because log_state() will use locks causing lock conflicts.
    CodeCache::log_state(&s);

    ttyLocker ttyl;
    xtty->begin_elem("sweeper state='%s' traversals='" INTX_FORMAT "' ", msg, (intx)traversal_count());
    if (format != NULL) {
      va_list ap;
      va_start(ap, format);
      xtty->vprint(format, ap);
      va_end(ap);
    }
    xtty->print(s.as_string());
    xtty->stamp();
    xtty->end_elem();
  }
}