[ ¼ע ]

PatentTips - Fair scalable reader-writer mutual exclusion

2017-07-14 09:42:12 admin һҳ

ǩdes   style   blog   code   http   ext   

BACKGROUND

The present invention relates generally to multithreaded programming and more specifically to mutual exclusion of readers and writers in a multithreaded programming environment.

Mutual exclusion is a programming technique that ensures that only one program or routine at a time can access some resource such as a memory location an input/output (I/O) port or a file often through the use of semaphores which are flags used in programs to coordinate the activities of more than one program or routine. An object for implementing mutual exclusion (or mutex) may be called a lock.

A reader-writer (RW) lock allows either multiple readers to inspect shared data or a single writer exclusive access for modifying that data. On shared memory multiprocessors the cost of acquiring and releasing these locks can have a large impact on the performance of parallel applications. A major problem with na?ve implementations of these locks where processors spin on a global lock variable waiting for the lock to become available is that the memory containing the lock and the interconnection network to that memory will become contended when the lock is contended.

Various approaches in the prior art implement scalable exclusive locks that is exclusive locks that can become contended without resulting in memory or interconnection contention. These approaches depend either on cache hardware support or on the existence of local memory where accesses to local memory involve lower latency than accesses to remote memory.

In "Scalable Reader-Writer Synchronization for Shared-Memory Multiprocessors" by John M. Mellor-Crummey and Michael L. Scott Proceedings of the 3rd?ACM Symposium on Principles and Practice of Parallel Programming pp. 106-113 Williamsburg Va. April 1991 the authors describe an exclusive lock which uses atomic operations to build a singly linked list of waiting processors. The processor at the head of the list has the lock and new processors add themselves to the list tail. Rather than spinning on a global lock variable each processor spins on a variable in its local memory. A processor releases the lock by zeroing the variable on which the next processor in the queue in spinning.

For the RW variant of this exclusive lock each queue element contains an additional variable to maintain the state of the request. When a new reader request arrives the state of the previous element in the queue is examined to determine if the new request must block. With a RW lock readers must be able to release the lock in any order. Hence the singly linked list of Mellor-Crummey and Scott becomes discontinuous as readers dequeue. To allow for this two global variables were added to their exclusive lock a count of the number of active readers and a pointer to the first writer in the queue. As readers acquire and release the lock they keep the global count of active readers up to date. When releasing the lock if a reader discovers that the reader count is zero it unblocks the writer pointed to by the global variable.

In "A Fair Fast Scalable Reader-Writer Lock" by Orran Krieger Michael Stumm Ron Unrau and Jonathan Hanna Proceedings of the 1993 International Conference on Parallel Processing the authors describe a fair scalable RW locking algorithm derived from Mellor-Crummey and Scotts exclusive locking algorithm. In the Krieger et al. process rather than adding more global state (that can become contended) an extra state needed for a RW lock is distributed across the list associated with the lock. In particular readers are maintained in a doubly linked list. With a doubly linked list instead of synchronizing on a global variable a reader that is releasing the lock can synchronize with its nearest neighbors to remove itself from the queue. This allows readers to dequeue in any order without the list becoming discontinuous. Hence it is not necessary to keep either a global pointer to the first writer or a global count of the number of active readers.

There are at least several disadvantages with the two prior art approaches discussed above. In each of the above approaches queue nodes cannot be allocated on a stack because sometimes a queue node supplied by a caller is read or written by other threads even after the caller has released its lock on the mutex. These approaches require the queue nodes to be allocated on a heap which is slower than stack allocation and may require acquiring other locks on the heap itself. Further these methods require that queue nodes never be freed for the lifetime of the mutex or somehow be atomically reference-counted to determine when it is safe to free them (which is expensive in a multithreaded environment compared to ordinary reads and writes). The approaches also require that a queue node live longer than the time between acquisition and release of the lock. Additionally the Krieger et al. method sometimes allows readers to block other readers when a reader expects to be unblocked by its predecessor when the latter has already seen no successor to unblock which causes the reader to block until all previous readers release the mutex.

Thus there is a need for further advances in multithreaded programming techniques to overcome these and other disadvantages.

DETAILED DESCRIPTION

An embodiment of the present invention is a reader-writer mutex process that enforces mutual exclusion of readers and writers in a way that overcomes deficiencies in previous reader-writer mutexes. Processing threads acquire either "reader locks" or "writer locks" on a mutex before entering a critical section and release their lock afterwards. The mutexs job is to ensure that if a writer is in the critical section it is the only thread there. Typically a reader-writer mutex serves to protect the consistency of a data structure. Threads wishing to modify the structure do so only after acquiring a writer lock. Threads wishing to read the structure acquire a reader lock. The advantage of a reader-writer mutex over a plain mutex is that multiple readers can proceed concurrently and hence it is important that readers never block readers.

Reference in the specification to "one embodiment" or "an embodiment" of the present invention means that a particular feature structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus the appearances of the phrase "in one embodiment" appearing in various places throughout the specification are not necessarily all referring to the same embodiment.

A thread may have to wait for other threads to leave before it enters a critical section. The invention is fair because it services threads in "first come first serve" order which ensures that no thread gets stuck waiting forever. A waiting thread should not consume memory bandwidth; otherwise it will slow down other threads. Waiting is done by "local spinning" which means reading a cache line that no other processor will touch frequently until another thread updates the cache line to say "proceed".

The RW mutex may be implemented by using a queue. The queue includes zero or more queue nodes that are created and destroyed as needed. The mutex comprises a single word that points to the tail of the queue. In embodiments of the present invention and as used herein a "word" defines the amount of memory needed to store a pointer value in the processing system. Processing threads may put a request onto the queue when they require either read or write access to a critical section protected by the mutex. A request for access is represented by a queue node. Each thread provides the queue node on which it will spin. When a queue node is the head of the queue the thread that created the queue node is granted access to the critical section. In embodiments of the present invention the queue node is created on the stack of a thread. The RW mutex points to the tail of the queue to indicate where a next thread desiring access to the critical section may be linked.

FIG. 1 is a diagram of an example queue for a reader writer mutex according to an embodiment of the present invention. FIG. 1 shows four threads at various stages. Thread?1?100?and thread?2?102?are readers who are inside the critical section. Thread?1?100?includes a stack?104. Thread?1?has previously created queue node?1?106?on its stack?104. This queue node?1106?is currently the head of the queue for reader writer mutex?108. Thread?1?also includes instruction pointer?110?to specify the current instruction being executed by the thread. When thread?1?is in the critical section the instruction pointer references memory within the critical section. Similarly Thread?2?102?includes a stack?112. Thread?2?has previously created queue node?2?114?on its stack?112?and linked queue node?2?into the queue for the RW mutex. Thread?2?also includes instruction pointer?116?to specify the current instruction being executed by the thread. When thread?2?is in the critical section the instruction pointer references memory within the critical section. Thread?3?118?is a writer who is waiting to enter the critical section after the two prior readers (i.e. threads?1?and?2) leave it. Thread?3?includes a stack?120?containing queue node?3?122 as well as instruction pointer?124. Thread?4126?is a reader who is constructing queue node?4?128?on its stack?130?in preparation to adding it to the queue. After thread?4?adds queue node?4?128?to the queue thread?4?will wait until the prior writer (thread?3) signals that it has left the critical section. Once queue node?4?128?is added to the queue reader writer mutex?108?will be changed to point to queue node?4?128?as the new tail of the queue. Thread?4?also includes instruction pointer?130.

bubuko.com

FIG. 2 is a flow diagram of using a queue with a reader writer mutex according to an embodiment of the present invention. At block?200 a thread desiring access to a critical section creates a queue node on the stack of the thread. This is the step that thread?4?126?is at in the example of FIG. 1. At block?202 the thread adds the newly created queue node to the tail of the queue. At block?204 the thread waits to enter the critical section by spinning on its queue node. This is the step that thread?3?118?is at in the example of FIG. 1. At block?206 when a thread acquires the lock the thread is allowed to enter the critical section. This is the step that thread?1?100?and thread?2?102?are at in the example of FIG. 1. At block?208 the thread exits the critical section. At block?210 the thread removes the queue node from the queue. Finally at block?212 the thread deletes the queue node from the threads stack.

bubuko.com

FIG. 3 is a diagram of the detailed structure of a queue node?300?according to an embodiment of the present invention. A queue node comprises a data structure representing a request to acquire the mutex and access a critical section and includes at least five fields. The NodeState field?302?specifies information about or to the owning thread. According to embodiments of the present invention the value may be Writer Reader UnblockNext or ActiveReader. The Writer state indicates that the thread plans to write data in the critical section. The Reader state indicates that the thread plans to read data in the critical section. According to embodiments of the present invention the UnblockNext state lets a successor thread signal that it has only partially linked in its own queue node and therefore the current thread should wait until the successor queue node is fully linked. The ActiveReader state indicates that the thread is currently in the critical section and actively reading data. The Next field?304?is a pointer to the next queue node in the queue (towards the tail of the queue). The Next field points to a queue node that is stored in another threads stack. The WaitState field?306?specifies the waiting status of the thread (whether a thread is to wait or proceed). According to embodiments of the present invention the types may be Spin Going or Busy. The Spin state indicates that the thread is waiting. The Going state indicates that the thread can access the critical section. According to embodiments of the present invention the Busy state indicates that the queue node cannot be deleted yet. This avoids premature destruction of a queue node before all threads are done accessing it. The Previous (Prev) field?308?is a pointer to the previous queue node in the queue (towards the head of the queue). According to embodiments of the present invention Flag field?309?may comprise an otherwise unused bit of the Previous field (a low order bit for example) as a flag bit to coordinate removal of a queue node from the queue. The EL field?310?comprises a simple spin mutex used to assist atomicity of deletion of readers from the middle of the queue.

bubuko.com

Table I shows the internal structure of a ReaderWriterMutex and a QueueNode. To aid in comparison with the prior art underscored text indicates code added by embodiments of the present invention and strikeout text indicates code removed by the present invention relative to the code disclosed in the Krieger et al. reference. Stylistic and obvious modifications are not marked.

TABLE I

1 enum NodeState_t {

2????WRITER READER UNBLOCKNEXT ACTIVEREADER

3 };

4 enum WaitState_t {

5????SPIN GOING BUSY

6 };

7 struct QueueNode {

8????QueueNode* Prev;

9????QueueNode* Next;

10????NodeState_t NodeState;

11????WaitState_t WaitState;

12????SpinLock EL;

13 };

14 struct ReaderWriterMutex {

15????QueueNode* Qtail;

16 };

FIG. 4 is a diagram of the example queue of FIG. 1 detailing sample queue node values according to an embodiment of the present invention. The reader writer mutex structure?108?comprises a pointer to the queue tail (Qtail). Queue node?1?106?has a NodeState field value of ActiveReader and the WaitState field of queue node?1?is set to Going because the thread that created queue node?1?has access to the critical region and is a reader. The Next field points to the next queue node in the queue (i.e. queue node?2?114). The Prev field points to null because this queue node is the head of the queue. The EL field of queue node?1?is Unlocked. Similarly Queue node?2?114?has a NodeState field value of ActiveReader and the WaitState field of queue node?2?is set to Going because the thread that created queue node?2?has access to the critical region and is a reader. The Next field points to the next queue node in the queue (i.e. queue node?3?122). The Prev field points to queue node?1?because that queue node is ahead of queue node?2?in the queue. The EL field of queue node?2?is Unlocked.

bubuko.com

For queue node?3?122 the NodeState field value is Writer. The Next field points to null because queue node?3?is the current tail of the queue. The WaitState field is set to Spin because queue node?3?is waiting for its turn to access the critical section. The Prev field points to the previous queue node in the queue (i.e. queue node?2?114). The EL field of queue node?3?is Unlocked.

For queue node?4?128 the NodeState field value is Reader. The Next and Prev fields point to null because the queue node is newly created and not yet linked into the queue. The WaitState field is set to Spin because queue node?4?is waiting for its turn to access the critical section. The Flag is set to zero. The EL field of queue node?4?is Unlocked.

Table II shows a typical calling sequence for a reader in a processing thread. Similar procedures calls may be used for a writer. In embodiments of the present invention the calling thread can allocate its queue node on its stack. The techniques disclosed in the Krieger et al. and Mellor-Crummey references do not permit this for readers.

????TABLE II

1 ReaderWriterMutex m;

2 {

3????QueueNode n; // create queue node on stack

4????AcquireForRead (&m &n);

5????..... // no thread has writer lock on mutex here.... //

6????ReleaseForRead (&m &n);

7 }

Tables III through VII describe additional operations of embodiments of the present invention. Table III defines the atomic operations required by embodiments of the present invention. These atomic operations are FetchAndStore CompareAndSwap and FetchAndAdd. These operations exist on most modern computing hardware notably processors commercially available from Intel Corporation such as the 32-bit processor architecture known as IA-32 and the 64-bit architecture known as Itanium Processor Family (IPF). In Table III "T" indicates the type of the memory location being atomically updated and "*" indicates a type pointer.

TABLE III

1 T FetchAndStore (T* location T value){

2????T result;

3????Atomic {

4????result = *location;

5????*location = value;

6????}

7????Return result;

8 }

9

10 T CompareAndSwap (T* location T value T comparand){

11????T result;

12????Atomic {

13????result = *location;

14????If (result == comparand)

15????*location = value;

16????}

17????Return result;

18 }

19

20 T FetchAndAdd (T* location T value){

21????T result;

22????Atomic {

23????result = *location;

24????*location = result + value;

25????}

26????Return result;

27 }

Tables IV and V show how a "writer lock" is acquired and released. These procedures are similar to methods disclosed in the Krieger et al. reference and are provided herein to aid in understanding how they interact with the techniques of embodiments of the present invention. The notation &x indicates construction of a pointer that points to location x. When a thread attempts to acquire a write lock on the mutex (Table IV) the thread enqueues its queue node on the queue and if it is not the only node in the queue waits for its predecessor to set its WaitState to Going. When the thread releases the lock (Table V) and there is a successor queue node the thread signals the next thread to go by changing the latters WaitState to Going.

TABLE IV

1 Procedure AcquireForWrite (ReaderWriterMutex* m QueueNode* node){

2????node?>Prev = NULL;

3????node?>Next = NULL;

4????node?>WaitState = SPIN;

5????node?>NodeState = WRITER;

6

7????QueueNode* pred = FetchAndStore (&m.Qtail node);

8????If pred != NULL then

9????pred?>Next = Node;

10????Wait until node?>WaitState == GOING

11 }

?

TABLE V

1 Procedure ReleaseForWrite (ReaderWriterMutex* m QueueNode* node) {

2 ????If node?>Next == NULL then

3 ????If node == CompareAndSwap (m?>Qtail NULL node) then

4 ????Goto Done;

5 ????Wait until node?>Next != NULL;

6 ????node?>Next?>Prev = NULL;

7 ????node?>Next?>WaitState = GOING;

8 ????Done:;

9 ????}

Table VI shows how a reader lock is acquired. The thread enqueues its queue node in the queue using a FetchAndStore procedure call. If the thread has a predecessor (which was the old queue node at the tail of the queue) it atomically checks if the predecessor has a NodeState of READER and if so advances the NodeState field to a state of UnblockNext. If the predecessor is not an ActiveReader the thread spins on its own queue nodes WaitState field until it is set to Going by the preceding thread: Then the thread tries to atomically transition its NodeState from Reader to ActiveReader. If this fails the thread indicates that it has a successor that needs to be unblocked. To unblock its successor the thread waits until the successor queue node is linked into the queue (by waiting for "nodeNext!=NULL") sets the successor queue nodes WaitState to Going and then the current thread becomes an ActiveReader.

?

TABLE VI

? 2005 Intel Corporation

1 Procedure AcquireForRead (ReaderWriterMutex* m QueueNode* node) {

2????node?>Prev = NULL;

3????node?>Next = NULL;

4????node?>WaitState = SPIN;

5????node?>NodeState = READER;

6????// pred is a pointer to a queue node in the queue

7????QueueNode* pred = FetchAndStore (&m.Qtail node);

8????If pred != NULL then

9????node?>Prev = pred;

10

11????// Problem: Once Pred?>Next becomes non-null its owner

12????// may proceed to destroy *pred.

13????// Improvement: Read Pred?>NodeState before writing to

14????// Pred?>Next because once Pred?>Next becomes non-null

15????// the owner of the pred node may destroy it.

16????NodeState_t predstate = CompareAndSwap(

17????&pred?>NodeState UNBLOCKNEXT

18????READER);

19????// Problem: Our predecessor might be a reader that executes

20????// lines 33-34 below before we have a chance to set Pred?>Next

21????// in which case we regrettably spin until all predecessors release

22????// the lock (the last such predecessor executes line 53 of Table VII).

23????// Improvement: The CompareAndSwap above atomically updates

24????// Pred?>NodeState to UNBLOCKNEXT if it was a READER which

25????// tells the predecessor to wait for us to set "next"

26????Pred?>Next = node;

27????If (predstate != ACTIVEREADER)

28????Wait until node?>WaitState == GOING;

29????// Improvement: The CompareAndSwap fails if our successor has

30????// changed us to UNBLOCKNEXT which signals us that we should wait

31????// until the successor links itself in and then signal the successor to get

32????// going.

33

35

36????NodeState_t oldstate = CompareAndSwap (&node?>NodeState

36????ACTIVEREADER READER);

37????If oldstate != READER then

38????Wait until node?>Next != NULL;

39????node?>NodeState = ACTIVEREADER;

40????node?>Next?>WaitState = GOING;

41????// this solves the readers block readers problem

41 }

Table VII shows how a reader lock is released. The method depends on having an unused bit FLAG in the "Prev" pointer field (e.g. if the queue nodes are 16-bit aligned then the least significant bit of the pointer can be used for FLAG.) The FLAG bit is used to coordinate situations where two threads are trying to remove adjacent nodes in the queue. The thread uses a procedure call to FetchAndAdd to atomically fetch "Prev" and simultaneously mark it as "in use" by setting the FLAG bit. (Logically it is a "FetchAndOr" but Intel processors support only FetchAndAdd and in this case the two operations are identical because the FLAG bit is always zero before the FetchAndAdd).

TABLE VII

? 2005 Intel Corporation

1 Procedure ReleaseForRead (ReaderWriterMutex* m QueueNode* node) {

2 Retry:

3????// Problem: Node pred might be destroyed after we read "node?>Prev"

4????// causing the subsequent lock/unlock operations on the predecessor

5????// to be accessing/modifying freed memory.

6????// Improvement: Instead of reading "node?>Prev" we automatically fetch

7????// it and mark it as in use by us.

8????QueueNode* pred = FetchAndAdd (&node?>Prev FLAG);

9????If pred != NULL Then

10????// Problem: lock/unlock retry loop is inefficient because of the

11????// iterated lock/unlock behavior.

12

13

14

15????// Improvement: the FLAG bit stops the predecessor from

16????// prematurely finishing indeed the predecessor waits for us

17????// to release its lock on pred?>EL.

18????Try to acquire lock on pred?>EL

19????If could not acquire the lock then

20????// Wait for predecessor to unlink itself

21????Wait until node?>Prev != pred + FLAG;

22????// Now predecessor is waiting for us to release it

23????Release lock on pred?>EL;

24????Goto Retry;

25????node?>Prev = pred;

26????Acquire lock on node?>EL;

27????pred?>Next = NULL;

28????If node?>Next == NULL and

29????node != CompareAndSwap (m?>Qtail pred node) then

30????wait until node?>Next != NULL;

31????tmp = NULL;

32????if node?>Next != NULL then

33????tmp = FetchAndStore (&(node?>Next?>Prev) pred)

34????pred?>Next = node?>Next;

35????release lock on pred?>EL;

36????else

37????acquire lock on node?>EL;

38????if node?>Next == NULL then

39???? if node == CompareAndSwap (m?>Qtail NULL node) then

40????goto Unlock_Self;

41????else

42????wait until node?>Next != NULL;

43????// Problem: setting WaitState to GOING may cause the owner of

44????// the next node to destroy it causing our write to

45????// node?>Next?Prev to write it into freed memory

46

47

48????// Improvement: temporarily mark WaitState as in use and set it to

49????// GOING after reading Next?>Prev. Use fetch-and-store here so

50????// that we know if the previous value had its FLAG bit set.

51????node?>Next?>WaitState = BUSY;

52????tmp =FetchAndStore (&node?>Next?>Prev NULL);

53????node?>Next?>WaitState = GOING;

54

55 Unlock_Self:

56????// Improvement: Detect if our successor is unlinking itself and if so wait

57????// for it to release our spin lock on Node?>EL.

58????If tmp does not have its FLAG bit set then

59????Release lock on node?>EL;

60????Else

61????Wait until lock on node?>EL is released by our successor;

62????// Problem: exiting too soon might cause us to destroy our node while

63????// other threads are accessing it.

64????// Improvement: wait until we know that our data are not in use by our

65????// predecessor.

66????Wait until node?>WaitState != BUSY;

67 }

If there is a predecessor then the thread tries to acquire a spin lock on the predecessor. If it cannot that means that the owner of the predecessor is trying to remove the predecessor from the queue and the current thread should wait until that removal is done and retry with the new predecessor. Interestingly it detects that the removal is done by spinning until "nodePrev!=pred+FLAG" and then releases the lock on PredEL that the predecessor thread acquired. The present protocol is unusual because a thread releases a lock acquired by another thread. Otherwise having successfully acquired a spin lock on the predecessor the thread acquires a spin lock on its own queue node and removes it from the queue taking care to update the queue tail pointer ("mQtail") if necessary. It then releases the spin lock on the predecessor and goes to the "Unlock_Self" label.

If the result of the FetchAndAdd procedure call indicates that there is not a predecessor the thread acquires a spin lock on its own queue node. If the thread has no successor then the thread attempts to change the queue to empty using "CompareAndSwap (mQtail NULL node"). If successful processing continues at the "Unlock_Self" label. If not that means that another thread is busy appending a successor queue node to the queue and so the thread waits until "nodeNext!=NULL". The thread then signals the next thread that it can proceed. The signaling has to be done carefully because either setting a Prev field to NULL or setting WaitState to Going may cause the owning thread to proceed possibly to the point where the thread destroys the corresponding queue node before the current thread is done with it. To circumvent this problem the current thread temporarily sets the successors WaitState to Busy before setting the successors Prev pointer to NULL and setting WaitState to Going. Note too that this process uses FetchAndStore to set the successors Prev field which is essential to the last "Unlock_Self" step.

The "Unlock_Self" step in Table VII completes the protocol. Here variable tmp is the value of "nodeNextPrev" before the current node was unlinked. If the FLAG bit of tmp is not set the current thread releases the spin lock the current thread acquired on its queue node. But if the FLAG bit is set this indicates that the owner of the successor queue node is in the retry loop and the current thread must wait for the successor thread to release the lock on the queue node that the current thread acquired. Finally the thread waits until nodeWaitState!=Busy after which the thread knows that it is safe to return (and destroy the node).

FIG. 5 is a diagram illustrating the use of a NodeState value UnblockNext according to an embodiment of the present invention. In this example there are three threads that participate each with a queue node on its stack. The thread owning queue node?4?502?has already entered the critical section as a reader. Two other threads have recently added their queue nodes (queue node?5?504?and queue node?6?506) to the queue pointed to by reader writer mutex?500. The owner of queue node?5?attempted to change its NodeState from Reader to ActiveReader. However in the meantime before that attempt the owner of queue node?6506?changed that NodeState (of queue node?5) from Reader to UnblockNext so as to notify the owner of queue node?5?that it has a successor and should wait until its Next field becomes non-null (i.e. is set as shown by the dashed arrow). The notifications are all done without any internal locking.

bubuko.com

Embodiments of the present invention have at least three advantages over prior reader-writer mutexes that spin on local memory. First it allows the queue nodes to be allocated on the callers stack. Previous reader-writer mutexes (such as Kreiger et al. and Mellor-Crummey et al.) do not allow the queue nodes to be stack-allocated because sometimes a queue node supplied by a caller is read or written by other threads even after the caller has released its lock on the mutex. The known prior art requires that the queue nodes be heap-allocated which is generally slower than stack allocation and indeed may require acquiring oilier locks on the heap itself. Furthermore the known prior art requires that the queue nodes never be freed for the lifetime of the mutex or somehow be atomically reference-counted to determine when it is safe to free them. Atomic reference counting itself is relatively expensive in a multithreaded environment compared to ordinary reads and writes. Embodiments of the present invention guarantee that a thread will not exit the mutex-release procedure until other threads no longer access its queue node; thus a thread can allocate its queue nodes on its own stack.

Second embodiments of the present invention never allow readers to block other readers as Krieger et al. sometimes does when a reader expects to be unblocked by its predecessor when the latter has already seen no successor to unblock which causes the reader to block until all previous readers release the mutex.

Third embodiments of the present invention use only a single word for the mutex whereas Mellor-Crummey et al. requires three words which makes this prior art design much more costly in applications that have fine-grained locking. The reason is that typically in fine-grained situations there needs to be a mutex per data element but only a lock (or a few locks) held by a thread. So it pays to keep the mutex object small at the cost of perhaps making the lock object bigger.

Another advantage is that embodiments of the present invention fit well with the C++ "resource acquisition is initialization" idiom that is widely used for writing exception-safe code. In this idiom as applied to mutexes an object is constructed that represents acquisition of a lock and destruction of the object releases the lock. If an exception is thrown in the critical section the object is implicitly destroyed thus automatically releasing the lock. This applies to embodiments of the present invention because the object can be (or contain as a member) the queue node itself. Neither the Kreiger et al. nor the Mellor-Crummey et al. prior art references permit this economy because they require that the queue node live longer than the time between acquisition and release of the lock.

Embodiments of the present invention enable stack-allocation of queue nodes by having a protocol that never references a queue node that might already be destroyed by another thread. This arises from the use of several techniques. First use atomic operations and a mark bit in the Prev pointer so that the owner of a node can handshake with the owner of the previous node. Using atomic operations and a mark bit has been used before for lockless algorithms for removing an item from a linked list (such as in Timothy Harris "A Pragmatic Implementation of Non-Blocking Linked-Lists" Proc. Of 2001 IEEE Symposium on Distributed Computing") but in embodiments of the present invention atomic operations and the mark bit are used not only to handshake removal but also to control which thread releases a lock. Furthermore in the prior lockless algorithms links in the item may be used by other threads even after it is removed from the list and thus complicated algorithms (e.g. Harris "timer snapshots") are required to determine when the item can be destroyed. Second a thread releases a lock held by another thread. Third an extra WaitState Busy is added that signals to the owner of a queue node that it should not yet destroy the node. Fourth judicious ordering of reads and writes is used.

Furthermore embodiments of the present invention manage to use only a single word for the mutex portion and yet never cause readers to block readers. Krieger et al. has the single-word feature but has the reader blocks readers problem. Mellor-Crummey does not have the reader blocks readers problem but requires three words for the mutex.

Although the operations detailed herein may be described as a sequential process some of the operations may in fact be performed in parallel or concurrently. In addition in some embodiments the order of the operations may be rearranged.

SRC=http://www.freepatentsonline.com/8707324.html

PatentTips - Fair scalable reader-writer mutual exclusionbubuko.com

PatentTips - Fair scalable reader-writer mutual exclusion

ǩdes   style   blog   code   http   ext   


Դhttp://www.bozhiyue.com/yuyan/2017/0714/1484245.html
һҳ    ط һƪ   һƪ