src/share/vm/gc_implementation/g1/vm_operations_g1.cpp

Mon, 12 Mar 2012 14:59:00 -0700

author
johnc
date
Mon, 12 Mar 2012 14:59:00 -0700
changeset 3666
64bf7c8270cb
parent 3456
9509c20bba28
child 3710
5c86f8211d1e
permissions
-rw-r--r--

7147724: G1: hang in SurrogateLockerThread::manipulatePLL
Summary: Attempting to initiate a marking cycle when allocating a humongous object can, if a marking cycle is successfully initiated by another thread, result in the allocating thread spinning until the marking cycle is complete. Eliminate a deadlock between the main ConcurrentMarkThread, the SurrogateLocker thread, the VM thread, and a mutator thread waiting on the SecondaryFreeList_lock (while free regions are going to become available) by not manipulating the pending list lock during the prologue and epilogue of the cleanup pause.
Reviewed-by: brutisso, jcoomes, tonyp

     1 /*
     2  * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    25 #include "precompiled.hpp"
    26 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
    27 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
    28 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
    29 #include "gc_implementation/g1/vm_operations_g1.hpp"
    30 #include "gc_implementation/shared/isGCActiveMark.hpp"
    31 #include "gc_implementation/g1/vm_operations_g1.hpp"
    32 #include "runtime/interfaceSupport.hpp"
    34 VM_G1CollectForAllocation::VM_G1CollectForAllocation(
    35                                                   unsigned int gc_count_before,
    36                                                   size_t word_size)
    37   : VM_G1OperationWithAllocRequest(gc_count_before, word_size,
    38                                    GCCause::_allocation_failure) {
    39   guarantee(word_size > 0, "an allocation should always be requested");
    40 }
    42 void VM_G1CollectForAllocation::doit() {
    43   G1CollectedHeap* g1h = G1CollectedHeap::heap();
    44   _result = g1h->satisfy_failed_allocation(_word_size, &_pause_succeeded);
    45   assert(_result == NULL || _pause_succeeded,
    46          "if we get back a result, the pause should have succeeded");
    47 }
    49 void VM_G1CollectFull::doit() {
    50   G1CollectedHeap* g1h = G1CollectedHeap::heap();
    51   GCCauseSetter x(g1h, _gc_cause);
    52   g1h->do_full_collection(false /* clear_all_soft_refs */);
    53 }
    55 VM_G1IncCollectionPause::VM_G1IncCollectionPause(
    56                                       unsigned int   gc_count_before,
    57                                       size_t         word_size,
    58                                       bool           should_initiate_conc_mark,
    59                                       double         target_pause_time_ms,
    60                                       GCCause::Cause gc_cause)
    61   : VM_G1OperationWithAllocRequest(gc_count_before, word_size, gc_cause),
    62     _should_initiate_conc_mark(should_initiate_conc_mark),
    63     _target_pause_time_ms(target_pause_time_ms),
    64     _should_retry_gc(false),
    65     _full_collections_completed_before(0) {
    66   guarantee(target_pause_time_ms > 0.0,
    67             err_msg("target_pause_time_ms = %1.6lf should be positive",
    68                     target_pause_time_ms));
    69   guarantee(word_size == 0 || gc_cause == GCCause::_g1_inc_collection_pause,
    70             "we can only request an allocation if the GC cause is for "
    71             "an incremental GC pause");
    72   _gc_cause = gc_cause;
    73 }
    75 bool VM_G1IncCollectionPause::doit_prologue() {
    76   bool res = VM_GC_Operation::doit_prologue();
    77   if (!res) {
    78     if (_should_initiate_conc_mark) {
    79       // The prologue can fail for a couple of reasons. The first is that another GC
    80       // got scheduled and prevented the scheduling of the initial mark GC. The
    81       // second is that the GC locker may be active and the heap can't be expanded.
    82       // In both cases we want to retry the GC so that the initial mark pause is
    83       // actually scheduled. In the second case, however, we should stall until
    84       // until the GC locker is no longer active and then retry the initial mark GC.
    85       _should_retry_gc = true;
    86     }
    87   }
    88   return res;
    89 }
    91 void VM_G1IncCollectionPause::doit() {
    92   G1CollectedHeap* g1h = G1CollectedHeap::heap();
    93   assert(!_should_initiate_conc_mark ||
    94   ((_gc_cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
    95    (_gc_cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent) ||
    96     _gc_cause == GCCause::_g1_humongous_allocation),
    97          "only a GC locker, a System.gc() or a hum allocation induced GC should start a cycle");
    99   if (_word_size > 0) {
   100     // An allocation has been requested. So, try to do that first.
   101     _result = g1h->attempt_allocation_at_safepoint(_word_size,
   102                                      false /* expect_null_cur_alloc_region */);
   103     if (_result != NULL) {
   104       // If we can successfully allocate before we actually do the
   105       // pause then we will consider this pause successful.
   106       _pause_succeeded = true;
   107       return;
   108     }
   109   }
   111   GCCauseSetter x(g1h, _gc_cause);
   112   if (_should_initiate_conc_mark) {
   113     // It's safer to read full_collections_completed() here, given
   114     // that noone else will be updating it concurrently. Since we'll
   115     // only need it if we're initiating a marking cycle, no point in
   116     // setting it earlier.
   117     _full_collections_completed_before = g1h->full_collections_completed();
   119     // At this point we are supposed to start a concurrent cycle. We
   120     // will do so if one is not already in progress.
   121     bool res = g1h->g1_policy()->force_initial_mark_if_outside_cycle(_gc_cause);
   123     // The above routine returns true if we were able to force the
   124     // next GC pause to be an initial mark; it returns false if a
   125     // marking cycle is already in progress.
   126     //
   127     // If a marking cycle is already in progress just return and skip the
   128     // pause below - if the reason for requesting this initial mark pause
   129     // was due to a System.gc() then the requesting thread should block in
   130     // doit_epilogue() until the marking cycle is complete.
   131     //
   132     // If this initial mark pause was requested as part of a humongous
   133     // allocation then we know that the marking cycle must just have
   134     // been started by another thread (possibly also allocating a humongous
   135     // object) as there was no active marking cycle when the requesting
   136     // thread checked before calling collect() in
   137     // attempt_allocation_humongous(). Retrying the GC, in this case,
   138     // will cause the requesting thread to spin inside collect() until the
   139     // just started marking cycle is complete - which may be a while. So
   140     // we do NOT retry the GC.
   141     if (!res) {
   142       assert(_word_size == 0, "Concurrent Full GC/Humongous Object IM shouldn't be allocating");
   143       if (_gc_cause != GCCause::_g1_humongous_allocation) {
   144         _should_retry_gc = true;
   145       }
   146       return;
   147     }
   148   }
   150   _pause_succeeded =
   151     g1h->do_collection_pause_at_safepoint(_target_pause_time_ms);
   152   if (_pause_succeeded && _word_size > 0) {
   153     // An allocation had been requested.
   154     _result = g1h->attempt_allocation_at_safepoint(_word_size,
   155                                       true /* expect_null_cur_alloc_region */);
   156   } else {
   157     assert(_result == NULL, "invariant");
   158     if (!_pause_succeeded) {
   159       // Another possible reason reason for the pause to not be successful
   160       // is that, again, the GC locker is active (and has become active
   161       // since the prologue was executed). In this case we should retry
   162       // the pause after waiting for the GC locker to become inactive.
   163       _should_retry_gc = true;
   164     }
   165   }
   166 }
   168 void VM_G1IncCollectionPause::doit_epilogue() {
   169   VM_GC_Operation::doit_epilogue();
   171   // If the pause was initiated by a System.gc() and
   172   // +ExplicitGCInvokesConcurrent, we have to wait here for the cycle
   173   // that just started (or maybe one that was already in progress) to
   174   // finish.
   175   if (_gc_cause == GCCause::_java_lang_system_gc &&
   176       _should_initiate_conc_mark) {
   177     assert(ExplicitGCInvokesConcurrent,
   178            "the only way to be here is if ExplicitGCInvokesConcurrent is set");
   180     G1CollectedHeap* g1h = G1CollectedHeap::heap();
   182     // In the doit() method we saved g1h->full_collections_completed()
   183     // in the _full_collections_completed_before field. We have to
   184     // wait until we observe that g1h->full_collections_completed()
   185     // has increased by at least one. This can happen if a) we started
   186     // a cycle and it completes, b) a cycle already in progress
   187     // completes, or c) a Full GC happens.
   189     // If the condition has already been reached, there's no point in
   190     // actually taking the lock and doing the wait.
   191     if (g1h->full_collections_completed() <=
   192                                           _full_collections_completed_before) {
   193       // The following is largely copied from CMS
   195       Thread* thr = Thread::current();
   196       assert(thr->is_Java_thread(), "invariant");
   197       JavaThread* jt = (JavaThread*)thr;
   198       ThreadToNativeFromVM native(jt);
   200       MutexLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
   201       while (g1h->full_collections_completed() <=
   202                                           _full_collections_completed_before) {
   203         FullGCCount_lock->wait(Mutex::_no_safepoint_check_flag);
   204       }
   205     }
   206   }
   207 }
   209 void VM_CGC_Operation::acquire_pending_list_lock() {
   210   assert(_needs_pll, "don't call this otherwise");
   211   // The caller may block while communicating
   212   // with the SLT thread in order to acquire/release the PLL.
   213   ConcurrentMarkThread::slt()->
   214     manipulatePLL(SurrogateLockerThread::acquirePLL);
   215 }
   217 void VM_CGC_Operation::release_and_notify_pending_list_lock() {
   218   assert(_needs_pll, "don't call this otherwise");
   219   // The caller may block while communicating
   220   // with the SLT thread in order to acquire/release the PLL.
   221   ConcurrentMarkThread::slt()->
   222     manipulatePLL(SurrogateLockerThread::releaseAndNotifyPLL);
   223 }
   225 void VM_CGC_Operation::doit() {
   226   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
   227   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
   228   TraceTime t(_printGCMessage, PrintGC, true, gclog_or_tty);
   229   SharedHeap* sh = SharedHeap::heap();
   230   // This could go away if CollectedHeap gave access to _gc_is_active...
   231   if (sh != NULL) {
   232     IsGCActiveMark x;
   233     _cl->do_void();
   234   } else {
   235     _cl->do_void();
   236   }
   237 }
   239 bool VM_CGC_Operation::doit_prologue() {
   240   // Note the relative order of the locks must match that in
   241   // VM_GC_Operation::doit_prologue() or deadlocks can occur
   242   if (_needs_pll) {
   243     acquire_pending_list_lock();
   244   }
   246   Heap_lock->lock();
   247   SharedHeap::heap()->_thread_holds_heap_lock_for_gc = true;
   248   return true;
   249 }
   251 void VM_CGC_Operation::doit_epilogue() {
   252   // Note the relative order of the unlocks must match that in
   253   // VM_GC_Operation::doit_epilogue()
   254   SharedHeap::heap()->_thread_holds_heap_lock_for_gc = false;
   255   Heap_lock->unlock();
   256   if (_needs_pll) {
   257     release_and_notify_pending_list_lock();
   258   }
   259 }

mercurial