archive/jdk
1
0
Fork 0
mirror of https://github.com/openjdk/jdk.git synced 2025-09-22 03:54:33 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

8049737: Contended Locking reorder and cache line bucket

Browse source 
JEP-143/JDK-8046133 - optimization #1 - reorder and cache line bucket.

Co-authored-by: Dave Dice <dave.dice@oracle.com>
Co-authored-by: Karen Kinnear <karen.kinnear@oracle.com>
Reviewed-by: shade, dice, dholmes, dsimms
This commit is contained in:
Daniel D. Daugherty 2014-10-14 10:32:12 -07:00
parent 51866388d1
commit f1ab0fae73
13 changed files with 302 additions and 247 deletions
Download patch file Download diff file Expand all files Collapse all files

294
hotspot/src/share/vm/runtime/objectMonitor.hpp
View file

@ -25,6 +25,7 @@
#ifndef SHARE_VM_RUNTIME_OBJECTMONITOR_HPP
#define SHARE_VM_RUNTIME_OBJECTMONITOR_HPP
#include "memory/padded.hpp"
#include "runtime/os.hpp"
#include "runtime/park.hpp"
#include "runtime/perfData.hpp"
@ -58,21 +59,71 @@ class ObjectWaiter : public StackObj {
// forward declaration to avoid include tracing.hpp
class EventJavaMonitorWait;
// WARNING:
// This is a very sensitive and fragile class. DO NOT make any
// change unless you are fully aware of the underlying semantics.
// This class can not inherit from any other class, because I have
// to let the displaced header be the very first word. Otherwise I
// have to let markOop include this file, which would export the
// monitor data structure to everywhere.
// The ObjectMonitor class implements the heavyweight version of a
// JavaMonitor. The lightweight BasicLock/stack lock version has been
// inflated into an ObjectMonitor. This inflation is typically due to
// contention or use of Object.wait().
//
// The ObjectMonitor class is used to implement JavaMonitors which have
// transformed from the lightweight structure of the thread stack to a
// heavy weight lock due to contention
// It is also used as RawMonitor by the JVMTI
// WARNING: This is a very sensitive and fragile class. DO NOT make any
// changes unless you are fully aware of the underlying semantics.
//
// Class JvmtiRawMonitor currently inherits from ObjectMonitor so
// changes in this class must be careful to not break JvmtiRawMonitor.
// These two subsystems should be separated.
//
// ObjectMonitor Layout Overview/Highlights/Restrictions:
//
// - The _header field must be at offset 0 because the displaced header
// from markOop is stored there. We do not want markOop.hpp to include
// ObjectMonitor.hpp to avoid exposing ObjectMonitor everywhere. This
// means that ObjectMonitor cannot inherit from any other class nor can
// it use any virtual member functions. This restriction is critical to
// the proper functioning of the VM.
// - The _header and _owner fields should be separated by enough space
// to avoid false sharing due to parallel access by different threads.
// This is an advisory recommendation.
// - The general layout of the fields in ObjectMonitor is:
// _header
// <lightly_used_fields>
// <optional padding>
// _owner
// <remaining_fields>
// - The VM assumes write ordering and machine word alignment with
// respect to the _owner field and the <remaining_fields> that can
// be read in parallel by other threads.
// - Generally fields that are accessed closely together in time should
// be placed proximally in space to promote data cache locality. That
// is, temporal locality should condition spatial locality.
// - We have to balance avoiding false sharing with excessive invalidation
// from coherence traffic. As such, we try to cluster fields that tend
// to be _written_ at approximately the same time onto the same data
// cache line.
// - We also have to balance the natural tension between minimizing
// single threaded capacity misses with excessive multi-threaded
// coherency misses. There is no single optimal layout for both
// single-threaded and multi-threaded environments.
//
// - See ObjectMonitor::sanity_checks() for how critical restrictions are
// enforced and advisory recommendations are reported.
// - Adjacent ObjectMonitors should be separated by enough space to avoid
// false sharing. This is handled by the ObjectMonitor allocation code
// in synchronizer.cpp. Also see ObjectSynchronizer::sanity_checks().
//
// Futures notes:
// - Separating _owner from the <remaining_fields> by enough space to
// avoid false sharing might be profitable. Given
// http://blogs.oracle.com/dave/entry/cas_and_cache_trivia_invalidate
// we know that the CAS in monitorenter will invalidate the line
// underlying _owner. We want to avoid an L1 data cache miss on that
// same line for monitorexit. Putting these <remaining_fields>:
// _recursions, _EntryList, _cxq, and _succ, all of which may be
// fetched in the inflated unlock path, on a different cache line
// would make them immune to CAS-based invalidation from the _owner
// field.
//
// - The _recursions field should be of type int, or int32_t but not
// intptr_t. There's no reason to use a 64-bit type for this field
// in a 64-bit JVM.
class ObjectMonitor {
public:
@ -84,7 +135,84 @@ class ObjectMonitor {
OM_TIMED_OUT // Object.wait() timed out
};
private:
friend class ObjectSynchronizer;
friend class ObjectWaiter;
friend class VMStructs;
volatile markOop _header; // displaced object header word - mark
void* volatile _object; // backward object pointer - strong root
public:
ObjectMonitor * FreeNext; // Free list linkage
private:
DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE,
sizeof(volatile markOop) + sizeof(void * volatile) +
sizeof(ObjectMonitor *));
protected: // protected for JvmtiRawMonitor
void * volatile _owner; // pointer to owning thread OR BasicLock
volatile jlong _previous_owner_tid; // thread id of the previous owner of the monitor
volatile intptr_t _recursions; // recursion count, 0 for first entry
ObjectWaiter * volatile _EntryList; // Threads blocked on entry or reentry.
// The list is actually composed of WaitNodes,
// acting as proxies for Threads.
private:
ObjectWaiter * volatile _cxq; // LL of recently-arrived threads blocked on entry.
Thread * volatile _succ; // Heir presumptive thread - used for futile wakeup throttling
Thread * volatile _Responsible;
volatile int _Spinner; // for exit->spinner handoff optimization
volatile int _SpinFreq; // Spin 1-out-of-N attempts: success rate
volatile int _SpinClock;
volatile intptr_t _SpinState; // MCS/CLH list of spinners
volatile int _SpinDuration;
volatile jint _count; // reference count to prevent reclamation/deflation
// at stop-the-world time. See deflate_idle_monitors().
// _count is approximately |_WaitSet| + |_EntryList|
protected:
ObjectWaiter * volatile _WaitSet; // LL of threads wait()ing on the monitor
volatile jint _waiters; // number of waiting threads
private:
volatile int _WaitSetLock; // protects Wait Queue - simple spinlock
public:
static void Initialize();
static PerfCounter * _sync_ContendedLockAttempts;
static PerfCounter * _sync_FutileWakeups;
static PerfCounter * _sync_Parks;
static PerfCounter * _sync_EmptyNotifications;
static PerfCounter * _sync_Notifications;
static PerfCounter * _sync_SlowEnter;
static PerfCounter * _sync_SlowExit;
static PerfCounter * _sync_SlowNotify;
static PerfCounter * _sync_SlowNotifyAll;
static PerfCounter * _sync_FailedSpins;
static PerfCounter * _sync_SuccessfulSpins;
static PerfCounter * _sync_PrivateA;
static PerfCounter * _sync_PrivateB;
static PerfCounter * _sync_MonInCirculation;
static PerfCounter * _sync_MonScavenged;
static PerfCounter * _sync_Inflations;
static PerfCounter * _sync_Deflations;
static PerfLongVariable * _sync_MonExtant;
static int Knob_Verbose;
static int Knob_VerifyInUse;
static int Knob_SpinLimit;
void* operator new (size_t size) throw() {
return AllocateHeap(size, mtInternal);
}
void* operator new[] (size_t size) throw() {
return operator new (size);
}
void operator delete(void* p) {
FreeHeap(p, mtInternal);
}
void operator delete[] (void *p) {
operator delete(p);
}
// TODO-FIXME: the "offset" routines should return a type of off_t instead of int ...
// ByteSize would also be an appropriate type.
static int header_offset_in_bytes() { return offset_of(ObjectMonitor, _header); }
@ -100,14 +228,11 @@ class ObjectMonitor {
static int Responsible_offset_in_bytes() { return offset_of(ObjectMonitor, _Responsible); }
static int Spinner_offset_in_bytes() { return offset_of(ObjectMonitor, _Spinner); }
public:
// Eventually we'll make provisions for multiple callbacks, but
// now one will suffice.
static int (*SpinCallbackFunction)(intptr_t, int);
static intptr_t SpinCallbackArgument;
public:
markOop header() const;
void set_header(markOop hdr);
@ -123,39 +248,22 @@ class ObjectMonitor {
void* owner() const;
void set_owner(void* owner);
intptr_t waiters() const;
jint waiters() const;
intptr_t count() const;
void set_count(intptr_t count);
intptr_t contentions() const;
jint count() const;
void set_count(jint count);
jint contentions() const;
intptr_t recursions() const { return _recursions; }
// JVM/DI GetMonitorInfo() needs this
// JVM/TI GetObjectMonitorUsage() needs this:
ObjectWaiter* first_waiter() { return _WaitSet; }
ObjectWaiter* next_waiter(ObjectWaiter* o) { return o->_next; }
Thread* thread_of_waiter(ObjectWaiter* o) { return o->_thread; }
// initialize the monitor, exception the semaphore, all other fields
// are simple integers or pointers
ObjectMonitor() {
_header = NULL;
_count = 0;
_waiters = 0;
_recursions = 0;
_object = NULL;
_owner = NULL;
_WaitSet = NULL;
_WaitSetLock = 0;
_Responsible = NULL;
_succ = NULL;
_cxq = NULL;
FreeNext = NULL;
_EntryList = NULL;
_SpinFreq = 0;
_SpinClock = 0;
OwnerIsThread = 0;
_previous_owner_tid = 0;
}
protected:
// We don't typically expect or want the ctors or dtors to run.
// normal ObjectMonitors are type-stable and immortal.
ObjectMonitor() { ::memset((void *)this, 0, sizeof(*this)); }
~ObjectMonitor() {
// TODO: Add asserts ...
@ -169,7 +277,7 @@ class ObjectMonitor {
// _cxq == 0 _succ == NULL _owner == NULL _waiters == 0
// _count == 0 EntryList == NULL
// _recursions == 0 _WaitSet == NULL
// TODO: assert (is_busy()|_recursions) == 0
assert(((is_busy()|_recursions) == 0), "freeing inuse monitor");
_succ = NULL;
_EntryList = NULL;
_cxq = NULL;
@ -177,7 +285,6 @@ class ObjectMonitor {
_recursions = 0;
_SpinFreq = 0;
_SpinClock = 0;
OwnerIsThread = 0;
}
public:
@ -221,7 +328,6 @@ class ObjectMonitor {
int TrySpin_Fixed(Thread * Self);
int TrySpin_VaryFrequency(Thread * Self);
int TrySpin_VaryDuration(Thread * Self);
void ctAsserts();
void ExitEpilog(Thread * Self, ObjectWaiter * Wakee);
bool ExitSuspendEquivalent(JavaThread * Self);
void post_monitor_wait_event(EventJavaMonitorWait * event,
@ -229,102 +335,6 @@ class ObjectMonitor {
jlong timeout,
bool timedout);
private:
friend class ObjectSynchronizer;
friend class ObjectWaiter;
friend class VMStructs;
// WARNING: this must be the very first word of ObjectMonitor
// This means this class can't use any virtual member functions.
volatile markOop _header; // displaced object header word - mark
void* volatile _object; // backward object pointer - strong root
double SharingPad[1]; // temp to reduce false sharing
// All the following fields must be machine word aligned
// The VM assumes write ordering wrt these fields, which can be
// read from other threads.
protected: // protected for jvmtiRawMonitor
void * volatile _owner; // pointer to owning thread OR BasicLock
volatile jlong _previous_owner_tid; // thread id of the previous owner of the monitor
volatile intptr_t _recursions; // recursion count, 0 for first entry
private:
int OwnerIsThread; // _owner is (Thread *) vs SP/BasicLock
ObjectWaiter * volatile _cxq; // LL of recently-arrived threads blocked on entry.
// The list is actually composed of WaitNodes, acting
// as proxies for Threads.
protected:
ObjectWaiter * volatile _EntryList; // Threads blocked on entry or reentry.
private:
Thread * volatile _succ; // Heir presumptive thread - used for futile wakeup throttling
Thread * volatile _Responsible;
int _PromptDrain; // rqst to drain cxq into EntryList ASAP
volatile int _Spinner; // for exit->spinner handoff optimization
volatile int _SpinFreq; // Spin 1-out-of-N attempts: success rate
volatile int _SpinClock;
volatile int _SpinDuration;
volatile intptr_t _SpinState; // MCS/CLH list of spinners
// TODO-FIXME: _count, _waiters and _recursions should be of
// type int, or int32_t but not intptr_t. There's no reason
// to use 64-bit fields for these variables on a 64-bit JVM.
volatile intptr_t _count; // reference count to prevent reclamation/deflation
// at stop-the-world time. See deflate_idle_monitors().
// _count is approximately |_WaitSet| + |_EntryList|
protected:
volatile intptr_t _waiters; // number of waiting threads
private:
protected:
ObjectWaiter * volatile _WaitSet; // LL of threads wait()ing on the monitor
private:
volatile int _WaitSetLock; // protects Wait Queue - simple spinlock
public:
int _QMix; // Mixed prepend queue discipline
ObjectMonitor * FreeNext; // Free list linkage
intptr_t StatA, StatsB;
public:
static void Initialize();
static PerfCounter * _sync_ContendedLockAttempts;
static PerfCounter * _sync_FutileWakeups;
static PerfCounter * _sync_Parks;
static PerfCounter * _sync_EmptyNotifications;
static PerfCounter * _sync_Notifications;
static PerfCounter * _sync_SlowEnter;
static PerfCounter * _sync_SlowExit;
static PerfCounter * _sync_SlowNotify;
static PerfCounter * _sync_SlowNotifyAll;
static PerfCounter * _sync_FailedSpins;
static PerfCounter * _sync_SuccessfulSpins;
static PerfCounter * _sync_PrivateA;
static PerfCounter * _sync_PrivateB;
static PerfCounter * _sync_MonInCirculation;
static PerfCounter * _sync_MonScavenged;
static PerfCounter * _sync_Inflations;
static PerfCounter * _sync_Deflations;
static PerfLongVariable * _sync_MonExtant;
public:
static int Knob_Verbose;
static int Knob_VerifyInUse;
static int Knob_SpinLimit;
void* operator new (size_t size) throw() {
return AllocateHeap(size, mtInternal);
}
void* operator new[] (size_t size) throw() {
return operator new (size);
}
void operator delete(void* p) {
FreeHeap(p, mtInternal);
}
void operator delete[] (void *p) {
operator delete(p);
}
};
#undef TEVENT