aoqi@0: /* shshahma@8611: * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. aoqi@0: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. aoqi@0: * aoqi@0: * This code is free software; you can redistribute it and/or modify it aoqi@0: * under the terms of the GNU General Public License version 2 only, as aoqi@0: * published by the Free Software Foundation. aoqi@0: * aoqi@0: * This code is distributed in the hope that it will be useful, but WITHOUT aoqi@0: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or aoqi@0: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License aoqi@0: * version 2 for more details (a copy is included in the LICENSE file that aoqi@0: * accompanied this code). aoqi@0: * aoqi@0: * You should have received a copy of the GNU General Public License version aoqi@0: * 2 along with this work; if not, write to the Free Software Foundation, aoqi@0: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. aoqi@0: * aoqi@0: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA aoqi@0: * or visit www.oracle.com if you need additional information or have any aoqi@0: * questions. aoqi@0: * aoqi@0: */ aoqi@0: aoqi@0: #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP aoqi@0: #define SHARE_VM_UTILITIES_TASKQUEUE_HPP aoqi@0: aoqi@0: #include "memory/allocation.hpp" aoqi@0: #include "memory/allocation.inline.hpp" aoqi@0: #include "runtime/mutex.hpp" goetz@6911: #include "runtime/orderAccess.inline.hpp" aoqi@0: #include "utilities/stack.hpp" aoqi@0: aoqi@0: // Simple TaskQueue stats that are collected by default in debug builds. aoqi@0: aoqi@0: #if !defined(TASKQUEUE_STATS) && defined(ASSERT) aoqi@0: #define TASKQUEUE_STATS 1 aoqi@0: #elif !defined(TASKQUEUE_STATS) aoqi@0: #define TASKQUEUE_STATS 0 aoqi@0: #endif aoqi@0: aoqi@0: #if TASKQUEUE_STATS aoqi@0: #define TASKQUEUE_STATS_ONLY(code) code aoqi@0: #else aoqi@0: #define TASKQUEUE_STATS_ONLY(code) aoqi@0: #endif // TASKQUEUE_STATS aoqi@0: aoqi@0: #if TASKQUEUE_STATS aoqi@0: class TaskQueueStats { aoqi@0: public: aoqi@0: enum StatId { aoqi@0: push, // number of taskqueue pushes aoqi@0: pop, // number of taskqueue pops aoqi@0: pop_slow, // subset of taskqueue pops that were done slow-path aoqi@0: steal_attempt, // number of taskqueue steal attempts aoqi@0: steal, // number of taskqueue steals aoqi@0: overflow, // number of overflow pushes aoqi@0: overflow_max_len, // max length of overflow stack aoqi@0: last_stat_id aoqi@0: }; aoqi@0: aoqi@0: public: aoqi@0: inline TaskQueueStats() { reset(); } aoqi@0: aoqi@0: inline void record_push() { ++_stats[push]; } aoqi@0: inline void record_pop() { ++_stats[pop]; } aoqi@0: inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; } aoqi@0: inline void record_steal(bool success); aoqi@0: inline void record_overflow(size_t new_length); aoqi@0: aoqi@0: TaskQueueStats & operator +=(const TaskQueueStats & addend); aoqi@0: aoqi@0: inline size_t get(StatId id) const { return _stats[id]; } aoqi@0: inline const size_t* get() const { return _stats; } aoqi@0: aoqi@0: inline void reset(); aoqi@0: aoqi@0: // Print the specified line of the header (does not include a line separator). aoqi@0: static void print_header(unsigned int line, outputStream* const stream = tty, aoqi@0: unsigned int width = 10); aoqi@0: // Print the statistics (does not include a line separator). aoqi@0: void print(outputStream* const stream = tty, unsigned int width = 10) const; aoqi@0: aoqi@0: DEBUG_ONLY(void verify() const;) aoqi@0: aoqi@0: private: aoqi@0: size_t _stats[last_stat_id]; aoqi@0: static const char * const _names[last_stat_id]; aoqi@0: }; aoqi@0: aoqi@0: void TaskQueueStats::record_steal(bool success) { aoqi@0: ++_stats[steal_attempt]; aoqi@0: if (success) ++_stats[steal]; aoqi@0: } aoqi@0: aoqi@0: void TaskQueueStats::record_overflow(size_t new_len) { aoqi@0: ++_stats[overflow]; aoqi@0: if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len; aoqi@0: } aoqi@0: aoqi@0: void TaskQueueStats::reset() { aoqi@0: memset(_stats, 0, sizeof(_stats)); aoqi@0: } aoqi@0: #endif // TASKQUEUE_STATS aoqi@0: aoqi@0: // TaskQueueSuper collects functionality common to all GenericTaskQueue instances. aoqi@0: aoqi@0: template aoqi@0: class TaskQueueSuper: public CHeapObj { aoqi@0: protected: aoqi@0: // Internal type for indexing the queue; also used for the tag. aoqi@0: typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t; aoqi@0: fujie@415: #ifdef MIPS64 fujie@415: private: fujie@415: #endif aoqi@0: // The first free element after the last one pushed (mod N). aoqi@0: volatile uint _bottom; aoqi@0: fujie@415: #ifdef MIPS64 fujie@415: protected: fujie@415: inline uint get_bottom() const { fujie@415: return OrderAccess::load_acquire((volatile juint*)&_bottom); fujie@415: } fujie@415: fujie@415: inline void set_bottom(uint new_bottom) { fujie@415: OrderAccess::release_store(&_bottom, new_bottom); fujie@415: } fujie@415: #endif fujie@415: aoqi@0: enum { MOD_N_MASK = N - 1 }; aoqi@0: aoqi@0: class Age { aoqi@0: public: aoqi@0: Age(size_t data = 0) { _data = data; } aoqi@0: Age(const Age& age) { _data = age._data; } aoqi@0: Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; } aoqi@0: fujie@415: #ifndef MIPS64 aoqi@0: Age get() const volatile { return _data; } aoqi@0: void set(Age age) volatile { _data = age._data; } aoqi@8860: aoqi@0: idx_t top() const volatile { return _fields._top; } aoqi@0: idx_t tag() const volatile { return _fields._tag; } fujie@415: #else aoqi@8860: Age get() const volatile { fujie@415: size_t res = OrderAccess::load_ptr_acquire((volatile intptr_t*) &_data); fujie@415: return *(Age*)(&res); fujie@415: } aoqi@8860: void set(Age age) volatile { OrderAccess::release_store_ptr((volatile intptr_t*) &_data, *(size_t*)(&age._data)); } fujie@415: fujie@415: idx_t top() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._top)); } fujie@415: idx_t tag() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._tag)); } fujie@415: #endif aoqi@0: aoqi@0: // Increment top; if it wraps, increment tag also. aoqi@0: void increment() { aoqi@0: _fields._top = increment_index(_fields._top); aoqi@0: if (_fields._top == 0) ++_fields._tag; aoqi@0: } aoqi@0: aoqi@0: Age cmpxchg(const Age new_age, const Age old_age) volatile { aoqi@0: return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data, aoqi@0: (volatile intptr_t *)&_data, aoqi@0: (intptr_t)old_age._data); aoqi@0: } aoqi@0: aoqi@0: bool operator ==(const Age& other) const { return _data == other._data; } aoqi@0: aoqi@0: private: aoqi@0: struct fields { aoqi@0: idx_t _top; aoqi@0: idx_t _tag; aoqi@0: }; aoqi@0: union { aoqi@0: size_t _data; aoqi@0: fields _fields; aoqi@0: }; aoqi@0: }; aoqi@0: aoqi@0: volatile Age _age; aoqi@0: aoqi@0: // These both operate mod N. aoqi@0: static uint increment_index(uint ind) { aoqi@0: return (ind + 1) & MOD_N_MASK; aoqi@0: } aoqi@0: static uint decrement_index(uint ind) { aoqi@0: return (ind - 1) & MOD_N_MASK; aoqi@0: } aoqi@0: aoqi@0: // Returns a number in the range [0..N). If the result is "N-1", it should be aoqi@0: // interpreted as 0. aoqi@0: uint dirty_size(uint bot, uint top) const { aoqi@0: return (bot - top) & MOD_N_MASK; aoqi@0: } aoqi@0: aoqi@0: // Returns the size corresponding to the given "bot" and "top". aoqi@0: uint size(uint bot, uint top) const { aoqi@0: uint sz = dirty_size(bot, top); aoqi@0: // Has the queue "wrapped", so that bottom is less than top? There's a aoqi@0: // complicated special case here. A pair of threads could perform pop_local aoqi@0: // and pop_global operations concurrently, starting from a state in which aoqi@0: // _bottom == _top+1. The pop_local could succeed in decrementing _bottom, aoqi@0: // and the pop_global in incrementing _top (in which case the pop_global aoqi@0: // will be awarded the contested queue element.) The resulting state must aoqi@0: // be interpreted as an empty queue. (We only need to worry about one such aoqi@0: // event: only the queue owner performs pop_local's, and several concurrent aoqi@0: // threads attempting to perform the pop_global will all perform the same aoqi@0: // CAS, and only one can succeed.) Any stealing thread that reads after aoqi@0: // either the increment or decrement will see an empty queue, and will not aoqi@0: // join the competitors. The "sz == -1 || sz == N-1" state will not be aoqi@0: // modified by concurrent queues, so the owner thread can reset the state to aoqi@0: // _bottom == top so subsequent pushes will be performed normally. aoqi@0: return (sz == N - 1) ? 0 : sz; aoqi@0: } aoqi@0: aoqi@0: public: aoqi@0: TaskQueueSuper() : _bottom(0), _age() {} aoqi@0: aoqi@0: // Return true if the TaskQueue contains/does not contain any tasks. aoqi@8860: bool peek() const { fujie@415: #ifdef MIPS64 aoqi@8860: return get_bottom() != _age.top(); fujie@415: #else aoqi@8860: return _bottom != _age.top(); fujie@415: #endif fujie@415: } aoqi@0: bool is_empty() const { return size() == 0; } aoqi@0: aoqi@0: // Return an estimate of the number of elements in the queue. aoqi@0: // The "careful" version admits the possibility of pop_local/pop_global aoqi@0: // races. aoqi@0: uint size() const { fujie@415: #ifdef MIPS64 fujie@415: return size(get_bottom(), _age.top()); fujie@415: #else aoqi@0: return size(_bottom, _age.top()); fujie@415: #endif aoqi@0: } aoqi@0: aoqi@0: uint dirty_size() const { fujie@415: #ifdef MIPS64 fujie@415: return dirty_size(get_bottom(), _age.top()); fujie@415: #else aoqi@0: return dirty_size(_bottom, _age.top()); fujie@415: #endif aoqi@0: } aoqi@0: aoqi@0: void set_empty() { fujie@415: #ifdef MIPS64 fujie@415: set_bottom(0); fujie@415: #else aoqi@0: _bottom = 0; fujie@415: #endif aoqi@0: _age.set(0); aoqi@0: } aoqi@0: aoqi@0: // Maximum number of elements allowed in the queue. This is two less aoqi@0: // than the actual queue size, for somewhat complicated reasons. aoqi@0: uint max_elems() const { return N - 2; } aoqi@0: aoqi@0: // Total size of queue. aoqi@0: static const uint total_size() { return N; } aoqi@0: aoqi@0: TASKQUEUE_STATS_ONLY(TaskQueueStats stats;) aoqi@0: }; aoqi@0: aoqi@0: // aoqi@0: // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double- aoqi@0: // ended-queue (deque), intended for use in work stealing. Queue operations aoqi@0: // are non-blocking. aoqi@0: // aoqi@0: // A queue owner thread performs push() and pop_local() operations on one end aoqi@0: // of the queue, while other threads may steal work using the pop_global() aoqi@0: // method. aoqi@0: // aoqi@0: // The main difference to the original algorithm is that this aoqi@0: // implementation allows wrap-around at the end of its allocated aoqi@0: // storage, which is an array. aoqi@0: // aoqi@0: // The original paper is: aoqi@0: // aoqi@0: // Arora, N. S., Blumofe, R. D., and Plaxton, C. G. aoqi@0: // Thread scheduling for multiprogrammed multiprocessors. aoqi@0: // Theory of Computing Systems 34, 2 (2001), 115-144. aoqi@0: // aoqi@0: // The following paper provides an correctness proof and an aoqi@0: // implementation for weakly ordered memory models including (pseudo-) aoqi@0: // code containing memory barriers for a Chase-Lev deque. Chase-Lev is aoqi@0: // similar to ABP, with the main difference that it allows resizing of the aoqi@0: // underlying storage: aoqi@0: // aoqi@0: // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z. aoqi@0: // Correct and efficient work-stealing for weak memory models aoqi@0: // Proceedings of the 18th ACM SIGPLAN symposium on Principles and aoqi@0: // practice of parallel programming (PPoPP 2013), 69-80 aoqi@0: // aoqi@0: aoqi@0: template aoqi@0: class GenericTaskQueue: public TaskQueueSuper { aoqi@0: ArrayAllocator _array_allocator; aoqi@0: protected: aoqi@0: typedef typename TaskQueueSuper::Age Age; aoqi@0: typedef typename TaskQueueSuper::idx_t idx_t; aoqi@0: fujie@415: #ifndef MIPS64 aoqi@0: using TaskQueueSuper::_bottom; fujie@415: #endif aoqi@0: using TaskQueueSuper::_age; aoqi@0: using TaskQueueSuper::increment_index; aoqi@0: using TaskQueueSuper::decrement_index; aoqi@0: using TaskQueueSuper::dirty_size; aoqi@0: aoqi@0: public: aoqi@0: using TaskQueueSuper::max_elems; aoqi@0: using TaskQueueSuper::size; aoqi@0: aoqi@0: #if TASKQUEUE_STATS aoqi@0: using TaskQueueSuper::stats; aoqi@0: #endif aoqi@0: aoqi@0: private: aoqi@0: // Slow paths for push, pop_local. (pop_global has no fast path.) aoqi@0: bool push_slow(E t, uint dirty_n_elems); aoqi@0: bool pop_local_slow(uint localBot, Age oldAge); aoqi@0: aoqi@0: public: aoqi@0: typedef E element_type; aoqi@0: aoqi@0: // Initializes the queue to empty. aoqi@0: GenericTaskQueue(); aoqi@0: aoqi@0: void initialize(); aoqi@0: aoqi@0: // Push the task "t" on the queue. Returns "false" iff the queue is full. aoqi@0: inline bool push(E t); aoqi@0: aoqi@0: // Attempts to claim a task from the "local" end of the queue (the most aoqi@0: // recently pushed). If successful, returns true and sets t to the task; aoqi@0: // otherwise, returns false (the queue is empty). aoqi@0: inline bool pop_local(volatile E& t); aoqi@0: aoqi@0: // Like pop_local(), but uses the "global" end of the queue (the least aoqi@0: // recently pushed). aoqi@0: bool pop_global(volatile E& t); aoqi@0: aoqi@0: // Delete any resource associated with the queue. aoqi@0: ~GenericTaskQueue(); aoqi@0: aoqi@0: // apply the closure to all elements in the task queue aoqi@0: void oops_do(OopClosure* f); aoqi@0: aoqi@0: private: aoqi@0: // Element array. aoqi@0: volatile E* _elems; aoqi@0: }; aoqi@0: aoqi@0: template aoqi@0: GenericTaskQueue::GenericTaskQueue() { aoqi@0: assert(sizeof(Age) == sizeof(size_t), "Depends on this."); aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: void GenericTaskQueue::initialize() { aoqi@0: _elems = _array_allocator.allocate(N); aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: void GenericTaskQueue::oops_do(OopClosure* f) { aoqi@0: // tty->print_cr("START OopTaskQueue::oops_do"); aoqi@0: uint iters = size(); fujie@415: #ifdef MIPS64 fujie@415: uint index = this->get_bottom(); fujie@415: #else aoqi@0: uint index = _bottom; fujie@415: #endif aoqi@0: for (uint i = 0; i < iters; ++i) { aoqi@0: index = decrement_index(index); aoqi@0: // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T, aoqi@0: // index, &_elems[index], _elems[index]); aoqi@0: E* t = (E*)&_elems[index]; // cast away volatility aoqi@0: oop* p = (oop*)t; aoqi@0: assert((*t)->is_oop_or_null(), "Not an oop or null"); aoqi@0: f->do_oop(p); aoqi@0: } aoqi@0: // tty->print_cr("END OopTaskQueue::oops_do"); aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: bool GenericTaskQueue::push_slow(E t, uint dirty_n_elems) { aoqi@0: if (dirty_n_elems == N - 1) { aoqi@0: // Actually means 0, so do the push. fujie@415: #ifdef MIPS64 fujie@415: uint localBot = this->get_bottom(); fujie@415: #else aoqi@0: uint localBot = _bottom; fujie@415: #endif aoqi@0: // g++ complains if the volatile result of the assignment is aoqi@0: // unused, so we cast the volatile away. We cannot cast directly aoqi@0: // to void, because gcc treats that as not using the result of the aoqi@0: // assignment. However, casting to E& means that we trigger an aoqi@0: // unused-value warning. So, we cast the E& to void. aoqi@0: (void)const_cast(_elems[localBot] = t); fujie@415: #ifdef MIPS64 fujie@415: this->set_bottom(increment_index(localBot)); fujie@415: #else aoqi@0: OrderAccess::release_store(&_bottom, increment_index(localBot)); fujie@415: #endif aoqi@0: TASKQUEUE_STATS_ONLY(stats.record_push()); aoqi@0: return true; aoqi@0: } aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // pop_local_slow() is done by the owning thread and is trying to aoqi@0: // get the last task in the queue. It will compete with pop_global() aoqi@0: // that will be used by other threads. The tag age is incremented aoqi@0: // whenever the queue goes empty which it will do here if this thread aoqi@0: // gets the last task or in pop_global() if the queue wraps (top == 0 aoqi@0: // and pop_global() succeeds, see pop_global()). aoqi@0: template aoqi@0: bool GenericTaskQueue::pop_local_slow(uint localBot, Age oldAge) { aoqi@0: // This queue was observed to contain exactly one element; either this aoqi@0: // thread will claim it, or a competing "pop_global". In either case, aoqi@0: // the queue will be logically empty afterwards. Create a new Age value aoqi@0: // that represents the empty queue for the given value of "_bottom". (We aoqi@0: // must also increment "tag" because of the case where "bottom == 1", aoqi@0: // "top == 0". A pop_global could read the queue element in that case, aoqi@0: // then have the owner thread do a pop followed by another push. Without aoqi@0: // the incrementing of "tag", the pop_global's CAS could succeed, aoqi@0: // allowing it to believe it has claimed the stale element.) aoqi@0: Age newAge((idx_t)localBot, oldAge.tag() + 1); aoqi@0: // Perhaps a competing pop_global has already incremented "top", in which aoqi@0: // case it wins the element. aoqi@0: if (localBot == oldAge.top()) { aoqi@0: // No competing pop_global has yet incremented "top"; we'll try to aoqi@0: // install new_age, thus claiming the element. aoqi@0: Age tempAge = _age.cmpxchg(newAge, oldAge); aoqi@0: if (tempAge == oldAge) { aoqi@0: // We win. aoqi@0: assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); aoqi@0: TASKQUEUE_STATS_ONLY(stats.record_pop_slow()); aoqi@0: return true; aoqi@0: } aoqi@0: } aoqi@0: // We lose; a completing pop_global gets the element. But the queue is empty aoqi@0: // and top is greater than bottom. Fix this representation of the empty queue aoqi@0: // to become the canonical one. aoqi@0: _age.set(newAge); aoqi@0: assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: bool GenericTaskQueue::pop_global(volatile E& t) { aoqi@0: Age oldAge = _age.get(); aoqi@0: // Architectures with weak memory model require a barrier here aoqi@0: // to guarantee that bottom is not older than age, aoqi@0: // which is crucial for the correctness of the algorithm. aoqi@0: #if !(defined SPARC || defined IA32 || defined AMD64) aoqi@0: OrderAccess::fence(); aoqi@0: #endif fujie@415: #ifdef MIPS64 fujie@415: uint localBot = this->get_bottom(); fujie@415: #else aoqi@0: uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom); fujie@415: #endif aoqi@0: uint n_elems = size(localBot, oldAge.top()); aoqi@0: if (n_elems == 0) { aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // g++ complains if the volatile result of the assignment is aoqi@0: // unused, so we cast the volatile away. We cannot cast directly aoqi@0: // to void, because gcc treats that as not using the result of the aoqi@0: // assignment. However, casting to E& means that we trigger an aoqi@0: // unused-value warning. So, we cast the E& to void. aoqi@0: (void) const_cast(t = _elems[oldAge.top()]); aoqi@0: Age newAge(oldAge); aoqi@0: newAge.increment(); aoqi@0: Age resAge = _age.cmpxchg(newAge, oldAge); aoqi@0: aoqi@0: // Note that using "_bottom" here might fail, since a pop_local might aoqi@0: // have decremented it. aoqi@0: assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity"); aoqi@0: return resAge == oldAge; aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: GenericTaskQueue::~GenericTaskQueue() { aoqi@0: FREE_C_HEAP_ARRAY(E, _elems, F); aoqi@0: } aoqi@0: aoqi@0: // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for aoqi@0: // elements that do not fit in the TaskQueue. aoqi@0: // aoqi@0: // This class hides two methods from super classes: aoqi@0: // aoqi@0: // push() - push onto the task queue or, if that fails, onto the overflow stack aoqi@0: // is_empty() - return true if both the TaskQueue and overflow stack are empty aoqi@0: // aoqi@0: // Note that size() is not hidden--it returns the number of elements in the aoqi@0: // TaskQueue, and does not include the size of the overflow stack. This aoqi@0: // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. aoqi@0: template aoqi@0: class OverflowTaskQueue: public GenericTaskQueue aoqi@0: { aoqi@0: public: aoqi@0: typedef Stack overflow_t; aoqi@0: typedef GenericTaskQueue taskqueue_t; aoqi@0: aoqi@0: TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) aoqi@0: aoqi@0: // Push task t onto the queue or onto the overflow stack. Return true. aoqi@0: inline bool push(E t); aoqi@0: shshahma@8611: // Try to push task t onto the queue only. Returns true if successful, false otherwise. shshahma@8611: inline bool try_push_to_taskqueue(E t); shshahma@8611: aoqi@0: // Attempt to pop from the overflow stack; return true if anything was popped. aoqi@0: inline bool pop_overflow(E& t); aoqi@0: aoqi@0: inline overflow_t* overflow_stack() { return &_overflow_stack; } aoqi@0: aoqi@0: inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } aoqi@0: inline bool overflow_empty() const { return _overflow_stack.is_empty(); } aoqi@0: inline bool is_empty() const { aoqi@0: return taskqueue_empty() && overflow_empty(); aoqi@0: } aoqi@0: aoqi@0: private: aoqi@0: overflow_t _overflow_stack; aoqi@0: }; aoqi@0: aoqi@0: template aoqi@0: bool OverflowTaskQueue::push(E t) aoqi@0: { aoqi@0: if (!taskqueue_t::push(t)) { aoqi@0: overflow_stack()->push(t); aoqi@0: TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size())); aoqi@0: } aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: bool OverflowTaskQueue::pop_overflow(E& t) aoqi@0: { aoqi@0: if (overflow_empty()) return false; aoqi@0: t = overflow_stack()->pop(); aoqi@0: return true; aoqi@0: } aoqi@0: shshahma@8611: template shshahma@8611: bool OverflowTaskQueue::try_push_to_taskqueue(E t) { shshahma@8611: return taskqueue_t::push(t); shshahma@8611: } aoqi@0: class TaskQueueSetSuper { aoqi@0: protected: aoqi@0: static int randomParkAndMiller(int* seed0); aoqi@0: public: aoqi@0: // Returns "true" if some TaskQueue in the set contains a task. aoqi@0: virtual bool peek() = 0; aoqi@0: }; aoqi@0: aoqi@0: template class TaskQueueSetSuperImpl: public CHeapObj, public TaskQueueSetSuper { aoqi@0: }; aoqi@0: aoqi@0: template aoqi@0: class GenericTaskQueueSet: public TaskQueueSetSuperImpl { aoqi@0: private: aoqi@0: uint _n; aoqi@0: T** _queues; aoqi@0: aoqi@0: public: aoqi@0: typedef typename T::element_type E; aoqi@0: aoqi@0: GenericTaskQueueSet(int n) : _n(n) { aoqi@0: typedef T* GenericTaskQueuePtr; aoqi@0: _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F); aoqi@0: for (int i = 0; i < n; i++) { aoqi@0: _queues[i] = NULL; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: bool steal_best_of_2(uint queue_num, int* seed, E& t); aoqi@0: aoqi@0: void register_queue(uint i, T* q); aoqi@0: aoqi@0: T* queue(uint n); aoqi@0: aoqi@0: // The thread with queue number "queue_num" (and whose random number seed is aoqi@0: // at "seed") is trying to steal a task from some other queue. (It may try aoqi@0: // several queues, according to some configuration parameter.) If some steal aoqi@0: // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns aoqi@0: // false. aoqi@0: bool steal(uint queue_num, int* seed, E& t); aoqi@0: aoqi@0: bool peek(); aoqi@0: }; aoqi@0: aoqi@0: template void aoqi@0: GenericTaskQueueSet::register_queue(uint i, T* q) { aoqi@0: assert(i < _n, "index out of range."); aoqi@0: _queues[i] = q; aoqi@0: } aoqi@0: aoqi@0: template T* aoqi@0: GenericTaskQueueSet::queue(uint i) { aoqi@0: return _queues[i]; aoqi@0: } aoqi@0: aoqi@0: template bool aoqi@0: GenericTaskQueueSet::steal(uint queue_num, int* seed, E& t) { aoqi@0: for (uint i = 0; i < 2 * _n; i++) { aoqi@0: if (steal_best_of_2(queue_num, seed, t)) { aoqi@0: TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true)); aoqi@0: return true; aoqi@0: } aoqi@0: } aoqi@0: TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false)); aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: template bool aoqi@0: GenericTaskQueueSet::steal_best_of_2(uint queue_num, int* seed, E& t) { aoqi@0: if (_n > 2) { aoqi@25: uint k1 = queue_num; aoqi@25: while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; aoqi@25: uint k2 = queue_num; aoqi@25: while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; aoqi@25: // Sample both and try the larger. aoqi@25: uint sz1 = _queues[k1]->size(); aoqi@25: uint sz2 = _queues[k2]->size(); aoqi@25: if (sz2 > sz1) return _queues[k2]->pop_global(t); aoqi@25: else return _queues[k1]->pop_global(t); aoqi@0: } else if (_n == 2) { aoqi@0: // Just try the other one. aoqi@0: uint k = (queue_num + 1) % 2; aoqi@0: return _queues[k]->pop_global(t); aoqi@0: } else { aoqi@0: assert(_n == 1, "can't be zero."); aoqi@0: return false; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: bool GenericTaskQueueSet::peek() { aoqi@0: // Try all the queues. aoqi@0: for (uint j = 0; j < _n; j++) { aoqi@0: if (_queues[j]->peek()) aoqi@0: return true; aoqi@0: } aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // When to terminate from the termination protocol. aoqi@0: class TerminatorTerminator: public CHeapObj { aoqi@0: public: aoqi@0: virtual bool should_exit_termination() = 0; aoqi@0: }; aoqi@0: aoqi@0: // A class to aid in the termination of a set of parallel tasks using aoqi@0: // TaskQueueSet's for work stealing. aoqi@0: aoqi@0: #undef TRACESPINNING aoqi@0: aoqi@0: class ParallelTaskTerminator: public StackObj { aoqi@0: private: aoqi@0: int _n_threads; aoqi@0: TaskQueueSetSuper* _queue_set; aoqi@0: int _offered_termination; aoqi@0: aoqi@0: #ifdef TRACESPINNING aoqi@0: static uint _total_yields; aoqi@0: static uint _total_spins; aoqi@0: static uint _total_peeks; aoqi@0: #endif aoqi@0: aoqi@0: bool peek_in_queue_set(); aoqi@0: protected: aoqi@0: virtual void yield(); aoqi@0: void sleep(uint millis); aoqi@0: aoqi@0: public: aoqi@0: aoqi@0: // "n_threads" is the number of threads to be terminated. "queue_set" is a aoqi@0: // queue sets of work queues of other threads. aoqi@0: ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set); aoqi@0: aoqi@0: // The current thread has no work, and is ready to terminate if everyone aoqi@0: // else is. If returns "true", all threads are terminated. If returns aoqi@0: // "false", available work has been observed in one of the task queues, aoqi@0: // so the global task is not complete. aoqi@0: bool offer_termination() { aoqi@0: return offer_termination(NULL); aoqi@0: } aoqi@0: aoqi@0: // As above, but it also terminates if the should_exit_termination() aoqi@0: // method of the terminator parameter returns true. If terminator is aoqi@0: // NULL, then it is ignored. aoqi@0: bool offer_termination(TerminatorTerminator* terminator); aoqi@0: aoqi@0: // Reset the terminator, so that it may be reused again. aoqi@0: // The caller is responsible for ensuring that this is done aoqi@0: // in an MT-safe manner, once the previous round of use of aoqi@0: // the terminator is finished. aoqi@0: void reset_for_reuse(); aoqi@0: // Same as above but the number of parallel threads is set to the aoqi@0: // given number. aoqi@0: void reset_for_reuse(int n_threads); aoqi@0: aoqi@0: #ifdef TRACESPINNING aoqi@0: static uint total_yields() { return _total_yields; } aoqi@0: static uint total_spins() { return _total_spins; } aoqi@0: static uint total_peeks() { return _total_peeks; } aoqi@0: static void print_termination_counts(); aoqi@0: #endif aoqi@0: }; aoqi@0: aoqi@0: template inline bool aoqi@0: GenericTaskQueue::push(E t) { fujie@415: #ifdef MIPS64 fujie@415: uint localBot = this->get_bottom(); fujie@415: #else aoqi@0: uint localBot = _bottom; fujie@415: #endif aoqi@0: assert(localBot < N, "_bottom out of range."); aoqi@0: idx_t top = _age.top(); aoqi@0: uint dirty_n_elems = dirty_size(localBot, top); aoqi@0: assert(dirty_n_elems < N, "n_elems out of range."); aoqi@0: if (dirty_n_elems < max_elems()) { aoqi@0: // g++ complains if the volatile result of the assignment is aoqi@0: // unused, so we cast the volatile away. We cannot cast directly aoqi@0: // to void, because gcc treats that as not using the result of the aoqi@0: // assignment. However, casting to E& means that we trigger an aoqi@0: // unused-value warning. So, we cast the E& to void. aoqi@0: (void) const_cast(_elems[localBot] = t); fujie@415: #ifdef MIPS64 fujie@415: this->set_bottom(increment_index(localBot)); fujie@415: #else aoqi@0: OrderAccess::release_store(&_bottom, increment_index(localBot)); fujie@415: #endif aoqi@0: TASKQUEUE_STATS_ONLY(stats.record_push()); aoqi@0: return true; aoqi@0: } else { aoqi@0: return push_slow(t, dirty_n_elems); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: template inline bool aoqi@0: GenericTaskQueue::pop_local(volatile E& t) { fujie@415: #ifdef MIPS64 fujie@415: uint localBot = this->get_bottom(); fujie@415: #else aoqi@0: uint localBot = _bottom; fujie@415: #endif aoqi@0: // This value cannot be N-1. That can only occur as a result of aoqi@0: // the assignment to bottom in this method. If it does, this method aoqi@0: // resets the size to 0 before the next call (which is sequential, aoqi@0: // since this is pop_local.) aoqi@0: uint dirty_n_elems = dirty_size(localBot, _age.top()); aoqi@0: assert(dirty_n_elems != N - 1, "Shouldn't be possible..."); aoqi@0: if (dirty_n_elems == 0) return false; aoqi@0: localBot = decrement_index(localBot); fujie@415: #ifdef MIPS64 fujie@415: this->set_bottom(localBot); fujie@415: #else aoqi@0: _bottom = localBot; fujie@415: #endif aoqi@0: // This is necessary to prevent any read below from being reordered aoqi@0: // before the store just above. aoqi@0: OrderAccess::fence(); aoqi@0: // g++ complains if the volatile result of the assignment is aoqi@0: // unused, so we cast the volatile away. We cannot cast directly aoqi@0: // to void, because gcc treats that as not using the result of the aoqi@0: // assignment. However, casting to E& means that we trigger an aoqi@0: // unused-value warning. So, we cast the E& to void. aoqi@0: (void) const_cast(t = _elems[localBot]); aoqi@0: // This is a second read of "age"; the "size()" above is the first. aoqi@0: // If there's still at least one element in the queue, based on the aoqi@0: // "_bottom" and "age" we've read, then there can be no interference with aoqi@0: // a "pop_global" operation, and we're done. aoqi@0: idx_t tp = _age.top(); // XXX aoqi@0: if (size(localBot, tp) > 0) { aoqi@0: assert(dirty_size(localBot, tp) != N - 1, "sanity"); aoqi@0: TASKQUEUE_STATS_ONLY(stats.record_pop()); aoqi@0: return true; aoqi@0: } else { aoqi@0: // Otherwise, the queue contained exactly one element; we take the slow aoqi@0: // path. aoqi@0: return pop_local_slow(localBot, _age.get()); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: typedef GenericTaskQueue OopTaskQueue; aoqi@0: typedef GenericTaskQueueSet OopTaskQueueSet; aoqi@0: aoqi@0: #ifdef _MSC_VER aoqi@0: #pragma warning(push) aoqi@0: // warning C4522: multiple assignment operators specified aoqi@0: #pragma warning(disable:4522) aoqi@0: #endif aoqi@0: aoqi@0: // This is a container class for either an oop* or a narrowOop*. aoqi@0: // Both are pushed onto a task queue and the consumer will test is_narrow() aoqi@0: // to determine which should be processed. aoqi@0: class StarTask { aoqi@0: void* _holder; // either union oop* or narrowOop* aoqi@0: aoqi@0: enum { COMPRESSED_OOP_MASK = 1 }; aoqi@0: aoqi@0: public: aoqi@0: StarTask(narrowOop* p) { aoqi@0: assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); aoqi@0: _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); aoqi@0: } aoqi@0: StarTask(oop* p) { aoqi@0: assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); aoqi@0: _holder = (void*)p; aoqi@0: } aoqi@0: StarTask() { _holder = NULL; } aoqi@0: operator oop*() { return (oop*)_holder; } aoqi@0: operator narrowOop*() { aoqi@0: return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); aoqi@0: } aoqi@0: aoqi@0: StarTask& operator=(const StarTask& t) { aoqi@0: _holder = t._holder; aoqi@0: return *this; aoqi@0: } aoqi@0: volatile StarTask& operator=(const volatile StarTask& t) volatile { aoqi@0: _holder = t._holder; aoqi@0: return *this; aoqi@0: } aoqi@0: aoqi@0: bool is_narrow() const { aoqi@0: return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); aoqi@0: } aoqi@0: }; aoqi@0: aoqi@0: class ObjArrayTask aoqi@0: { aoqi@0: public: aoqi@0: ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } aoqi@0: ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { aoqi@0: assert(idx <= size_t(max_jint), "too big"); aoqi@0: } aoqi@0: ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } aoqi@0: aoqi@0: ObjArrayTask& operator =(const ObjArrayTask& t) { aoqi@0: _obj = t._obj; aoqi@0: _index = t._index; aoqi@0: return *this; aoqi@0: } aoqi@0: volatile ObjArrayTask& aoqi@0: operator =(const volatile ObjArrayTask& t) volatile { aoqi@0: (void)const_cast(_obj = t._obj); aoqi@0: _index = t._index; aoqi@0: return *this; aoqi@0: } aoqi@0: aoqi@0: inline oop obj() const { return _obj; } aoqi@0: inline int index() const { return _index; } aoqi@0: aoqi@0: DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. aoqi@0: aoqi@0: private: aoqi@0: oop _obj; aoqi@0: int _index; aoqi@0: }; aoqi@0: aoqi@0: #ifdef _MSC_VER aoqi@0: #pragma warning(pop) aoqi@0: #endif aoqi@0: aoqi@0: typedef OverflowTaskQueue OopStarTaskQueue; aoqi@0: typedef GenericTaskQueueSet OopStarTaskQueueSet; aoqi@0: aoqi@0: typedef OverflowTaskQueue RegionTaskQueue; aoqi@0: typedef GenericTaskQueueSet RegionTaskQueueSet; aoqi@0: aoqi@0: aoqi@0: #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP