<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body bgcolor="#ffffff" text="#000000">
    Hi Bengt,<br>
    <br>
    Thanks for volunteering. I was until I saw your message.<br>
    <br>
    Hal: Great catch.<br>
    <br>
    We should probably change the other casts/shifts  in
    heapRegionRemSet.cpp and concurrentMark.cpp at the same time<br>
    <br>
    ./share/vm/gc_implementation/g1/heapRegionRemSet.cpp<br>
        _max_fine_entries = (size_t)(1 << max_entries_log);<br>
    <br>
    ./share/vm/gc_implementation/g1/concurrentMark.cpp<br>
      assert(((size_t)_bm.size() * (size_t)(1 << _shifter)) ==
    _bmWordSize,<br>
    <br>
    even though they shouldn't overflow.<br>
    <br>
    JohnC<br>
    <br>
    On 9/12/2012 12:50 AM, Bengt Rutisson wrote:
    <blockquote cite="mid:50503EAA.30206@oracle.com" type="cite">
      <meta content="text/html; charset=ISO-8859-1"
        http-equiv="Content-Type">
      <div class="moz-cite-prefix"><br>
        Hi Hal Mo,<br>
        <br>
        Nice catch!<br>
        <br>
        I filed "7197906:BlockOffsetArray::power_to_cards_back() needs
        to handle > 32 bit shifts" to track this issue. I can help
        you shepherd this fix in.<br>
        <br>
        Thanks,<br>
        Bengt<br>
        <br>
        <br>
        On 2012-09-12 03:22, Jianhao Mo wrote:<br>
      </div>
      <blockquote
cite="mid:CAKz_je4nJUcCFtV+24DtR3s7Xu4eiJ0=fyaHLhUX1-=OQmWV8Q@mail.gmail.com"
        type="cite">Hi all,<br>
        <br>
        This is Hal Mo<<a moz-do-not-send="true"
          href="mailto:kungu.mjh@taobao.com">kungu.mjh@taobao.com</a>>

        from Alibaba Group(with OCA).<br>
        <br>
        Our hadoop namenode crashed, when we set the heap size to 135G
        using CMS GC.<br>
        Attached please find the crash log(hs_err_pid.log).<br>
        <br>
        I can steadily reproduce the crash on a test machine with 190G
        physical memory, by a simple command:<br>
        $ java -Xmx135g -XX:+UseConcMarkSweepGC<br>
        <br>
        Then I build a debug jvm and use gdb to debug the problem.<br>
        <br>
        call stack<br>
        <br>
        C  [libc.so.6+0x7a9b0]  memset+0x40<br>
        V  [libjvm.so+0x2b6c42]
         BlockOffsetArray::set_remainder_to_point_to_start_incl(unsigned
        long, unsigned long, bool)+0xce<br>
        V  [libjvm.so+0x2b7043]
         BlockOffsetArray::set_remainder_to_point_to_start(HeapWord*,
        HeapWord*, bool)+0x71<br>
        V  [libjvm.so+0x2b728d]
         BlockOffsetArray::BlockOffsetArray(BlockOffsetSharedArray*,
        MemRegion, bool)+0x9f<br>
        V  [libjvm.so+0x3c089f]
         BlockOffsetArrayNonContigSpace::BlockOffsetArrayNonContigSpace(BlockOffsetSharedArray*,

        MemRegion)+0x37<br>
        V  [libjvm.so+0x3be56f]
         CompactibleFreeListSpace::CompactibleFreeListSpace(BlockOffsetSharedArray*,

        MemRegion, bool, FreeBlockDictionary::DictionaryChoice)+0x9b<br>
        V  [libjvm.so+0x3fd2e1]
         ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(ReservedSpace,

        unsigned long, int, CardTableRS*, bool,
        FreeBlockDictionary::DictionaryChoice)+0x1df<br>
        V  [libjvm.so+0x4dc03e]  GenerationSpec::init(ReservedSpace,
        int, GenRemSet*)+0x37c<br>
        V  [libjvm.so+0x4ced40]  GenCollectedHeap::initialize()+0x510<br>
        V  [libjvm.so+0x7c23c3]  Universe::initialize_heap()+0x31d<br>
        V  [libjvm.so+0x7c27ec]  universe_init()+0xa6<br>
        V  [libjvm.so+0x5056e2]  init_globals()+0x34<br>
        V  [libjvm.so+0x7ac926]  Threads::create_vm(JavaVMInitArgs*,
        bool*)+0x23a<br>
        V  [libjvm.so+0x53f3d4]  JNI_CreateJavaVM+0x7a<br>
        <br>
        in function
        BlockOffsetArray::set_remainder_to_point_to_start_inc, inside
        the for loop:<br>
            size_t reach = start_card - 1 + (power_to_cards_back(i+1) -
        1);<br>
        when i = 7, the value of reach was 0. then the loop could not
        break, and <br>
            _array->set_offset_array(start_card_for_region, reach,
        offset, reducing);<br>
        accessed the wrong address, and crashed.<br>
        <br>
        the root cause was <br>
        static size_t power_to_cards_back(uint i) {<br>
            return (size_t)(1 << (LogBase * i));<br>
        }<br>
        the literal 1 is a 32bit int, and 1<<32 overflow. <br>
        <br>
        <br>
        Here was my fix(has been tested), also found in attached file
        cms_large_heap_crash.patch<br>
        <br>
        +++ b/src/share/vm/memory/blockOffsetTable.hpp<br>
        @@ -289,7 +289,7 @@<br>
        };<br>
        <br>
        static size_t power_to_cards_back(uint i) {<br>
        - return (size_t)(1 << (LogBase * i));<br>
        + return (size_t)1 << (LogBase * i);<br>
        }<br>
        static size_t power_to_words_back(uint i) {<br>
        return power_to_cards_back(i) * N_words;<br>
        <br>
        Contributed-by: Hal Mo <<a moz-do-not-send="true"
          href="mailto:kungu.mjh@taobao.com">kungu.mjh@taobao.com</a>>

        <br>
        <br>
        Similar situation also found in G1, but the size is mega(2^20)
        based. 2^(32+20) is too large to overflow.<br>
        <br>
        Krystal remind me, this changeset cover the same code, <a
          moz-do-not-send="true"
href="http://hg.openjdk.java.net/hsx/hotspot-gc/hotspot/rev/c3a720eefe82">http://hg.openjdk.java.net/hsx/hotspot-gc/hotspot/rev/c3a720eefe82</a>
         .<br>
        I do not  build it on visual studio, someone please help to
        review the compatibility with VS.<br>
        <br>
        Regards,<br>
        <br>
        Hal<br>
      </blockquote>
      <br>
    </blockquote>
  </body>
</html>