介紹
Java 6的併發編程包中的SynchronousQueue是一個沒有數據緩衝的BlockingQueue,生產者線程對其的插入操作put必須等待消費者的移除操作take,反過來也一樣。
不像ArrayBlockingQueue或LinkedListBlockingQueue,SynchronousQueue內部並沒有數據緩存空間,你不能調用peek()方法來看隊列中是否有數據元素,因爲數據元素只有當你試着取走的時候纔可能存在,不取走而只想偷窺一下是不行的,當然遍歷這個隊列的操作也是不允許的。隊列頭元素是第一個排隊要插入數據的線程,而不是要交換的數據。數據是在配對的生產者和消費者線程之間直接傳遞的,並不會將數據緩衝數據到隊列中。可以這樣來理解:生產者和消費者互相等待對方,握手,然後一起離開。
SynchronousQueue的一個使用場景是在線程池裏。Executors.newCachedThreadPool()就使用了SynchronousQueue,這個線程池根據需要(新任務到來時)創建新的線程,如果有空閒線程則會重複使用,線程空閒了60秒後會被回收。
實現原理
阻塞隊列的實現方法有許多:
阻塞算法實現
阻塞算法實現通常在內部採用一個鎖來保證多個線程中的put()和take()方法是串行執行的。採用鎖的開銷是比較大的,還會存在一種情況是線程A持有線程B需要的鎖,B必須一直等待A釋放鎖,即使A可能一段時間內因爲B的優先級比較高而得不到時間片運行。所以在高性能的應用中我們常常希望規避鎖的使用。
01 |
public class NativeSynchronousQueue<E> { |
02 |
boolean putting = false; |
03 |
E item = null; |
04 |
05 |
public synchronized E take() throws InterruptedException { |
06 |
while (item == null) |
07 |
wait(); |
08 |
E e = item; |
09 |
item = null; |
10 |
notifyAll(); |
11 |
return e; |
12 |
} |
13 |
14 |
public synchronized void put(E e) throws InterruptedException { |
15 |
if (e==null) return; |
16 |
while (putting) |
17 |
wait(); |
18 |
putting = true; |
19 |
item = e; |
20 |
notifyAll(); |
21 |
while (item!=null) |
22 |
wait(); |
23 |
putting = false; |
24 |
notifyAll(); |
25 |
} |
26 |
} |
信號量實現
經典同步隊列實現採用了三個信號量,代碼很簡單,比較容易理解:
01 |
public class
SemaphoreSynchronousQueue<E> { |
02 |
E item =
null ; |
03 |
Semaphore sync =
new Semaphore( 0 ); |
04 |
Semaphore send =
new Semaphore( 1 ); |
05 |
Semaphore recv =
new Semaphore( 0 ); |
06 |
07 |
public
E take() throws
InterruptedException { |
08 |
recv.acquire(); |
09 |
E x = item; |
10 |
sync.release(); |
11 |
send.release(); |
12 |
return
x; |
13 |
} |
14 |
15 |
public
void put (E x) throws
InterruptedException{ |
16 |
send.acquire(); |
17 |
item = x; |
18 |
recv.release(); |
19 |
sync.acquire(); |
20 |
} |
21 |
} |
在多核機器上,上面方法的同步代價仍然較高,操作系統調度器需要上千個時間片來阻塞或喚醒線程,而上面的實現即使在生產者put()時已經有一個消費者在等待的情況下,阻塞和喚醒的調用仍然需要。
Java 5實現
01 |
public class
Java5SynchronousQueue<E> { |
02 |
ReentrantLock qlock =
new ReentrantLock(); |
03 |
Queue waitingProducers =
new Queue(); |
04 |
Queue waitingConsumers =
new Queue(); |
05 |
06 |
static
class Node extends
AbstractQueuedSynchronizer { |
07 |
E item; |
08 |
Node next; |
09 |
10 |
Node(Object x) { item = x; } |
11 |
void
waitForTake() { /* (uses AQS) */ } |
12 |
E waitForPut() { /* (uses AQS) */
} |
13 |
} |
14 |
15 |
public
E take() { |
16 |
Node node; |
17 |
boolean
mustWait; |
18 |
qlock.lock(); |
19 |
node = waitingProducers.pop(); |
20 |
if (mustWait = (node ==
null )) |
21 |
node = waitingConsumers.push( null ); |
22 |
qlock.unlock(); |
23 |
24 |
if
(mustWait) |
25 |
return
node.waitForPut(); |
26 |
else |
27 |
return
node.item; |
28 |
} |
29 |
30 |
public
void put(E e) { |
31 |
Node node; |
32 |
boolean
mustWait; |
33 |
qlock.lock(); |
34 |
node = waitingConsumers.pop(); |
35 |
if
(mustWait = (node == null )) |
36 |
node = waitingProducers.push(e); |
37 |
qlock.unlock(); |
38 |
39 |
if
(mustWait) |
40 |
node.waitForTake(); |
41 |
else |
42 |
node.item = e; |
43 |
} |
44 |
} |
Java 5的實現相對來說做了一些優化,只使用了一個鎖,使用隊列代替信號量也可以允許發佈者直接發佈數據,而不是要首先從阻塞在信號量處被喚醒。
Java6實現
Java 6的SynchronousQueue的實現採用了一種性能更好的無鎖算法 — 擴展的“Dual stack and Dual queue”算法。性能比Java5的實現有較大提升。競爭機制支持公平和非公平兩種:非公平競爭模式使用的數據結構是後進先出棧(Lifo Stack);公平競爭模式則使用先進先出隊列(Fifo Queue),性能上兩者是相當的,一般情況下,Fifo通常可以支持更大的吞吐量,但Lifo可以更大程度的保持線程的本地化。
代碼實現裏的Dual Queue或Stack內部是用鏈表(LinkedList)來實現的,其節點狀態爲以下三種情況:
- 持有數據 – put()方法的元素
- 持有請求 – take()方法
- 空
這個算法的特點就是任何操作都可以根據節點的狀態判斷執行,而不需要用到鎖。
其核心接口是Transfer,生產者的put或消費者的take都使用這個接口,根據第一個參數來區別是入列(棧)還是出列(棧)。
01 |
/** |
02 |
* Shared internal API for dual stacks and queues. |
03 |
*/ |
04 |
static
abstract class
Transferer { |
05 |
/** |
06 |
* Performs a put or take. |
07 |
* |
08 |
* @param e if non-null, the item to be handed to a consumer; |
09 |
* if null, requests that transfer return an item |
10 |
* offered by producer. |
11 |
* @param timed if this operation should timeout |
12 |
* @param nanos the timeout, in nanoseconds |
13 |
* @return if non-null, the item provided or received; if null, |
14 |
* the operation failed due to timeout or interrupt -- |
15 |
* the caller can distinguish which of these occurred |
16 |
* by checking Thread.interrupted. |
17 |
*/ |
18 |
abstract
Object transfer(Object e, boolean
timed, long
nanos); |
19 |
} |
TransferQueue實現如下(摘自Java 6源代碼),入列和出列都基於Spin和CAS方法:
01 |
/** |
02 |
* Puts or takes an item. |
03 |
*/ |
04 |
Object transfer(Object e,
boolean timed, long
nanos) { |
05 |
/* Basic algorithm is to loop trying to take either of |
06 |
* two actions: |
07 |
* |
08 |
* 1. If queue apparently empty or holding same-mode nodes, |
09 |
* try to add node to queue of waiters, wait to be |
10 |
* fulfilled (or cancelled) and return matching item. |
11 |
* |
12 |
* 2. If queue apparently contains waiting items, and this |
13 |
* call is of complementary mode, try to fulfill by CAS'ing |
14 |
* item field of waiting node and dequeuing it, and then |
15 |
* returning matching item. |
16 |
* |
17 |
* In each case, along the way, check for and try to help |
18 |
* advance head and tail on behalf of other stalled/slow |
19 |
* threads. |
20 |
* |
21 |
* The loop starts off with a null check guarding against |
22 |
* seeing uninitialized head or tail values. This never |
23 |
* happens in current SynchronousQueue, but could if |
24 |
* callers held non-volatile/final ref to the |
25 |
* transferer. The check is here anyway because it places |
26 |
* null checks at top of loop, which is usually faster |
27 |
* than having them implicitly interspersed. |
28 |
*/ |
29 |
30 |
QNode s =
null ; // constructed/reused as needed |
31 |
boolean
isData = (e != null ); |
32 |
33 |
for
(;;) { |
34 |
QNode t = tail; |
35 |
QNode h = head; |
36 |
if
(t == null
|| h == null )
// saw uninitialized value |
37 |
continue ;
// spin |
38 |
39 |
if
(h == t || t.isData == isData) { // empty or same-mode |
40 |
QNode tn = t.next; |
41 |
if
(t != tail) // inconsistent read |
42 |
continue ; |
43 |
if
(tn != null ) {
// lagging tail |
44 |
advanceTail(t, tn); |
45 |
continue ; |
46 |
} |
47 |
if
(timed && nanos <= 0 )
// can't wait |
48 |
return
null ; |
49 |
if
(s == null ) |
50 |
s =
new QNode(e, isData); |
51 |
if
(!t.casNext( null , s))
// failed to link in |
52 |
continue ; |
53 |
54 |
advanceTail(t, s);
// swing tail and wait |
55 |
Object x = awaitFulfill(s, e, timed, nanos); |
56 |
if
(x == s) { // wait was cancelled |
57 |
clean(t, s); |
58 |
return
null ; |
59 |
} |
60 |
61 |
if
(!s.isOffList()) { // not already unlinked |
62 |
advanceHead(t, s);
// unlink if head |
63 |
if
(x != null )
// and forget fields |
64 |
s.item = s; |
65 |
s.waiter =
null ; |
66 |
} |
67 |
return
(x != null )? x : e; |
68 |
69 |
}
else {
// complementary-mode |
70 |
QNode m = h.next;
// node to fulfill |
71 |
if
(t != tail || m == null
|| h != head) |
72 |
continue ;
// inconsistent read |
73 |
74 |
Object x = m.item; |
75 |
if
(isData == (x != null ) ||
// m already fulfilled |
76 |
x == m ||
// m cancelled |
77 |
!m.casItem(x, e)) {
// lost CAS |
78 |
advanceHead(h, m);
// dequeue and retry |
79 |
continue ; |
80 |
} |
81 |
82 |
advanceHead(h, m);
// successfully fulfilled |
83 |
LockSupport.unpark(m.waiter); |
84 |
return
(x != null )? x : e; |
85 |
} |
86 |
} |
87 |
} |