RFR (XL): 8152664 - Support non-continuous CodeBlobs in HotSpot
Volker Simonis
volker.simonis at gmail.com
Fri Apr 8 18:12:19 UTC 2016
Hi Rickard,
I found the problem why your change crashes the VM on ppc (and I'm pretty
sure it will also crash on ARM - @Andrew, maybe you can try it out?). It is
caused by the following code in address NativeCall::get_trampoline() which
is also present on arm64:
address NativeCall::get_trampoline() {
address call_addr = addr_at(0);
CodeBlob *code = CodeCache::find_blob(call_addr);
...
// If the codeBlob is not a nmethod, this is because we get here from the
// CodeBlob constructor, which is called within the nmethod constructor.
return trampoline_stub_Relocation::get_trampoline_for(call_addr,
(nmethod*)code);
}
The comment explains the situation quite well: we're in the CodeBlob
constructor which was called by the CompiledMethod constructor which was
called from the nmethod constructor:
#3 0x00003fffb741b80c in NativeCall::get_trampoline (this=0x3fff607d0fac)
at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/cpu/ppc/vm/nativeInst_ppc.cpp:146
#4 0x00003fffb7596914 in Relocation::pd_call_destination
(this=0x3ffdfe3fcc90, orig_addr=0x3fff603b8a2c "\001") at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/cpu/ppc/vm/relocInfo_ppc.cpp:87
#5 0x00003fffb758f5fc in CallRelocation::fix_relocation_after_move
(this=0x3ffdfe3fcc90, src=0x3ffdfe3fdb40, dest=0x3ffdfe3fcd58) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/relocInfo.cpp:506
#6 0x00003fffb6c48898 in CodeBuffer::relocate_code_to
(this=0x3ffdfe3fdb40, dest=0x3ffdfe3fcd58) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/asm/codeBuffer.cpp:812
#7 0x00003fffb6c48404 in CodeBuffer::copy_code_to (this=0x3ffdfe3fdb40,
dest_blob=0x3fff607d0c10) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/asm/codeBuffer.cpp:748
#8 0x00003fffb6c42670 in CodeBuffer::copy_code_and_locs_to
(this=0x3ffdfe3fdb40, blob=0x3fff607d0c10) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/asm/codeBuffer.hpp:607
#9 0x00003fffb6c3f834 in CodeBlob::CodeBlob (this=0x3fff607d0c10,
name=0x3fffb7a75fd8 "nmethod", layout=..., cb=0x3ffdfe3fdb40,
frame_complete_offset=20, frame_size=14, oop_maps=0x3ffe00049620,
caller_must_gc_arguments=false, subtype=8) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/codeBlob.cpp:117
#10 0x00003fffb6ce52c8 in CompiledMethod::CompiledMethod
(this=0x3fff607d0c10, method=0x3ffe1ddce568, name=0x3fffb7a75fd8 "nmethod",
size=1768, header_size=392, cb=0x3ffdfe3fdb40, frame_complete_offset=20,
frame_size=14, oop_maps=0x3ffe00049620, caller_must_gc_arguments=false) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/compiledMethod.cpp:42
#11 0x00003fffb7421f58 in nmethod::nmethod (this=0x3fff607d0c10,
method=0x3ffe1ddce568, nmethod_size=1768, compile_id=4, entry_bci=-1,
offsets=0x3ffdfe3fdb18, orig_pc_offset=104, debug_info=0x3fffb03d55f0,
dependencies=0x3ffe00049690, code_buffer=0x3ffdfe3fdb40, frame_size=14,
oop_maps=0x3ffe00049620, handler_table=0x3ffdfe3fdad0,
nul_chk_table=0x3ffdfe3fdaf0, compiler=0x3fffb03bc270, comp_level=3) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/nmethod.cpp:706
Now we cast 'code' to 'nmethod' but at this point in time 'code' is still a
CodeBlob from the C++ point of view (i.e. it still has a CodeBlob vtable
(see [1] for an explanation)).
Later on, in RelocIterator::initialize() we call virtual methods on the
nmethod which still has the vtable of a "CodeBlob" and this fails badly:
#0 SingletonBlob::print_on (this=0x3fff607d0c10, st=0x0) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/codeBlob.cpp:584
#1 0x00003fffb758d51c in RelocIterator::initialize (this=0x3ffdfe3fc928,
nm=0x3fff607d0c10, begin=0x3fff607d0fac "\001", limit=0x0) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/relocInfo.cpp:144
#2 0x00003fffb6ace56c in RelocIterator::RelocIterator
(this=0x3ffdfe3fc928, nm=0x3fff607d0c10, begin=0x3fff607d0fac "\001",
limit=0x0) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/relocInfo.hpp:1378
#3 0x00003fffb75919dc in trampoline_stub_Relocation::get_trampoline_for
(call=0x3fff607d0fac "\001", code=0x3fff607d0c10) at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/share/vm/code/relocInfo.cpp:849
#4 0x00003fffb741b80c in NativeCall::get_trampoline (this=0x3fff607d0fac)
at
/usr/work/d046063/OpenJDK/jdk9-hs-comp/hotspot/src/cpu/ppc/vm/nativeInst_ppc.cpp:146
As you can see, we actually want to call nmethod::stub_begin() at
relocInfo.cpp:144
142 _section_start[CodeBuffer::SECT_CONSTS] = nm->consts_begin();
143 _section_start[CodeBuffer::SECT_INSTS ] = nm->insts_begin() ;
144 _section_start[CodeBuffer::SECT_STUBS ] = nm->stub_begin() ;
but we actually end up in SingletonBlob::print_on() which is a completely
different method. Notice that the call to nm->consts_begin() before also
fails, but it doesn't crash the VM because it happens to call
SingletonBlob::verify() which has no bad side effect. The call to
nm->insts_begin() in line 143 is non-virtual and thus works fine. Here are
the corresponding vtable slots in the CodeBlob vtable for consts_begin()
and stub_begin()
(gdb) p &nmethod::consts_begin
$76 = &virtual table offset 42
(gdb) p &nmethod::stub_begin
$77 = &virtual table offset 44
(gdb) p ((*(void ***)nm) + 1)[42]
$86 = (void *) 0x3fffb6c41df8 <SingletonBlob::verify()>
(gdb) p ((*(void ***)nm) + 1)[44]
$87 = (void *) 0x3fffb6c41e64 <SingletonBlob::print_on(outputStream*) const>
As you can see, 'nm' is indeed a "CodeBlob" at this point:
(gdb) p *(void ***)nm
$91 = (void **) 0x3fffb7befa00 <vtable for CodeBlob+16>
(gdb) p nm->print()
[CodeBlob (0x00003fff607d1090)]
Framesize: 14
The offending calls succeeded before your change, because they where not
virtual. Any idea how we can fix this with the new class hierarchy?
Regards,
Volker
[1]
http://stackoverflow.com/questions/6591859/when-does-the-vptr-pointing-to-vtable-get-initialized-for-a-polymorphic-class
On Thu, Apr 7, 2016 at 5:50 PM, Volker Simonis <volker.simonis at gmail.com>
wrote:
> Hi Rickard,
>
> I'd also like to know what's the rational behind this quite large
> change. Do you expect some performance or memory consumption
> improvements or is this a prerequisite for another change which is
> still to come?
>
> The change itself currently doesn't work on ppc64 (neither on Linux
> nor on AIX). I get the following crash during the build when the newly
> built Hotspot is JIT-compiling java.lang.String::charAt on C1 :
>
> #
> # A fatal error has been detected by the Java Runtime Environment:
> #
> # SIGSEGV (0xb) at pc=0x00001000012a44d0, pid=35331, tid=35404
> #
> # JRE version: OpenJDK Runtime Environment (9.0) (slowdebug build
> 9-internal+0-2016-04-07-162501.d046063.jdk9-hs-comp)
> # Java VM: OpenJDK 64-Bit Server VM (slowdebug
> 9-internal+0-2016-04-07-162501.d046063.jdk9-hs-comp, mixed mode,
> tiered, compressed oo
> ps, serial gc, linux-ppc64le)
> # Problematic frame:
> # V [libjvm.so+0xf744d0] outputStream::do_vsnprintf_and_write(char
> const*, char*, bool)+0x40
> #
> # No core dump will be written. Core dumps have been disabled. To
> enable core dumping, try "ulimit -c unlimited" before starting Java
> again
> #
> # If you would like to submit a bug report, please visit:
> # http://bugreport.java.com/bugreport/crash.jsp
> #
>
> --------------- S U M M A R Y ------------
>
> Command Line:
> -Dapplication.home=/sapmnt/ld9510/a/d046063/output-jdk9-hs-comp-dbg/jdk
> -Xms8m -XX:+UseSerialGC -Xms32M -Xmx512M -Djdk.
> module.main=jdk.jlink jdk.jlink/jdk.tools.jmod.Main create
> --module-version 9-internal --os-name Linux --os-arch ppc64le
> --os-version
> 2.6 --modulepath /priv/d046063/output-jdk9-hs-comp-dbg/images/jmods
> --hash-dependencies .* --exclude **_the.* --libs
>
> /priv/d046063/output-jdk9-hs-comp-dbg/support/modules_libs-stripped/java.base
> --cmds
> /priv/d046063/output-jdk9-hs-comp-dbg/support/modules_cmds-stripped/java.base
> --config
> /priv/d046063/output-jdk9-hs-comp-dbg/support/modules_conf/java.base
> --class-path /priv/d046063/output-jdk9-hs-comp-dbg/jdk/modules/java.base
> /priv/d046063/output-jdk9-hs-comp-dbg/support/jmods/java.base.jmod
>
> Host: ld9510, POWER8E (raw), altivec supported, 48 cores, 61G, #
> Please check /etc/os-release for details about this release.
> Time: Thu Apr 7 16:28:55 2016 CEST elapsed time: 0 seconds (0d 0h 0m 0s)
>
> --------------- T H R E A D ---------------
>
> Current thread (0x000010000429c800): JavaThread "C1 CompilerThread10"
> daemon [_thread_in_vm, id=35404,
> stack(0x000010006a800000,0x000010006ac00000)]
>
>
> Current CompileTask:
> C1: 761 3 3 java.lang.String::charAt (25 bytes)
>
> Stack: [0x000010006a800000,0x000010006ac00000],
> sp=0x000010006abfc6c0, free space=4081k
> Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native
> code)
> V [libjvm.so+0xf744d0] outputStream::do_vsnprintf_and_write(char
> const*, char*, bool)+0x40
> V [libjvm.so+0xf74668] outputStream::print_cr(char const*, ...)+0x68
> V [libjvm.so+0x72189c] CodeBlob::print_on(outputStream*) const+0x50
> V [libjvm.so+0x723bdc] RuntimeBlob::print_on(outputStream*) const+0x40
> V [libjvm.so+0x721eb0] SingletonBlob::print_on(outputStream*) const+0x4c
> V [libjvm.so+0x106d51c] RelocIterator::initialize(CompiledMethod*,
> unsigned char*, unsigned char*)+0x170
> V [libjvm.so+0x5ae56c] RelocIterator::RelocIterator(CompiledMethod*,
> unsigned char*, unsigned char*)+0x78
> V [libjvm.so+0x10719dc]
> trampoline_stub_Relocation::get_trampoline_for(unsigned char*,
> nmethod*)+0x78
> V [libjvm.so+0xefb80c] NativeCall::get_trampoline()+0x110
> V [libjvm.so+0x1076914] Relocation::pd_call_destination(unsigned
> char*)+0x150
> V [libjvm.so+0x106f5fc]
> CallRelocation::fix_relocation_after_move(CodeBuffer const*,
> CodeBuffer*)+0x74
> V [libjvm.so+0x728898] CodeBuffer::relocate_code_to(CodeBuffer*)
> const+0x390
> V [libjvm.so+0x728404] CodeBuffer::copy_code_to(CodeBlob*)+0x134
> V [libjvm.so+0x722670] CodeBuffer::copy_code_and_locs_to(CodeBlob*)+0x84
> V [libjvm.so+0x71f834] CodeBlob::CodeBlob(char const*,
> CodeBlobLayout const&, CodeBuffer*, int, int, OopMapSet*, bool,
> int)+0x320
> V [libjvm.so+0x7c52c8] CompiledMethod::CompiledMethod(Method*, char
> const*, int, int, CodeBuffer*, int, int, OopMapSet*, bool)+0xd8
> V [libjvm.so+0xf01f58] nmethod::nmethod(Method*, int, int, int,
> CodeOffsets*, int, DebugInformationRecorder*, Dependencies*,
> CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*,
> ImplicitExceptionTable*, AbstractCompiler*, int)+0xe0
> V [libjvm.so+0xf01610] nmethod::new_nmethod(methodHandle const&,
> int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*,
> CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*,
> ImplicitExceptionTable*, AbstractCompiler*, int)+0x2c4
> V [libjvm.so+0x632970] ciEnv::register_method(ciMethod*, int,
> CodeOffsets*, int, CodeBuffer*, int, OopMapSet*,
> ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*,
> bool, bool, RTMState)+0x560
> V [libjvm.so+0x48ee00] Compilation::install_code(int)+0x264
> V [libjvm.so+0x48eff8] Compilation::compile_method()+0x184
> V [libjvm.so+0x48f7a8] Compilation::Compilation(AbstractCompiler*,
> ciEnv*, ciMethod*, int, BufferBlob*, DirectiveSet*)+0x288
> V [libjvm.so+0x4980d0] Compiler::compile_method(ciEnv*, ciMethod*,
> int, DirectiveSet*)+0xc8
> V [libjvm.so+0x7b188c]
> CompileBroker::invoke_compiler_on_method(CompileTask*)+0x590
> V [libjvm.so+0x7b07bc] CompileBroker::compiler_thread_loop()+0x310
> V [libjvm.so+0x11a614c] compiler_thread_entry(JavaThread*, Thread*)+0xa0
> V [libjvm.so+0x119f3a8] JavaThread::thread_main_inner()+0x1b4
> V [libjvm.so+0x119f1a4] JavaThread::run()+0x1b8
> V [libjvm.so+0xf53d90] java_start(Thread*)+0x204
> C [libpthread.so.0+0x8a64] start_thread+0xf4
> C [libc.so.6+0x1032a0] clone+0x98
>
> I haven't identified the exact cause (will analyze it tomorrow) but
> the stack trace indicates that it is indeed related to your changes.
>
> Besides that I have some comments:
>
> codeBuffer.hpp:
>
> 472 CodeSection* insts() { return &_insts; }
> 475 const CodeSection* insts() const { return &_insts; }
>
> - do we really need both versions?
>
> codeBlob.hpp:
>
> 135 nmethod* as_nmethod_or_null() const { return
> is_nmethod() ? (nmethod*) this : NULL; }
> 136 nmethod* as_nmethod() const {
> assert(is_nmethod(), "must be nmethod"); return (nmethod*) this; }
> 137 CompiledMethod* as_compiled_method_or_null() const { return
> is_compiled() ? (CompiledMethod*) this : NULL; }
> 138 CompiledMethod* as_compiled_method() const {
> assert(is_compiled(), "must be compiled"); return (CompiledMethod*)
> this; }
> 139 CodeBlob* as_codeblob_or_null() const { return
> (CodeBlob*) this; }
>
> - I don't like this code. You make the getters 'const' which
> implicitely makes 'this' a "pointer to const" but then the returned
> pointer is a normal pointer to a non-const object and therefore you
> have to statically cast away the "pointer to const" (that's why you
> need the cast even in the case where you return a CodeBlob*). So
> either remove the const qualifier from the method declarations or make
> them return "pointers to const". And by the way, as_codeblob_or_null()
> doesn't seemed to be used anywhere in the code, why do we need it at
> all?
>
> - Why do we need the non-virtual methods is_nmethod() and
> is_compiled() to manually simulate virtual behavior. Why can't we
> simply make them virtual and implement them accordingly in nmathod and
> CompiledMethod?
>
> Regards,
> Volker
>
> On Thu, Apr 7, 2016 at 2:12 PM, Rickard Bäckman
> <rickard.backman at oracle.com> wrote:
> > Hi,
> >
> > can I please have review for this patch please?
> >
> > So far CodeBlobs have required all the data (metadata, oops, code, etc)
> > to be in one continuous blob With this patch we are looking to change
> > that. It's been done by changing offsets in CodeBlob to addresses,
> > making some methods virtual to allow different behavior and also
> > creating a couple of new classes. CompiledMethod now sits inbetween
> > CodeBlob and nmethod.
> >
> > CR: https://bugs.openjdk.java.net/browse/JDK-8152664
> > Webrev: http://cr.openjdk.java.net/~rbackman/8152664/
> >
> > Thanks
> > /R
>
More information about the hotspot-dev
mailing list