• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * ptw32_MCS_lock.c
3  *
4  * Description:
5  * This translation unit implements queue-based locks.
6  *
7  * --------------------------------------------------------------------------
8  *
9  *      Pthreads-win32 - POSIX Threads Library for Win32
10  *      Copyright(C) 1998 John E. Bossom
11  *      Copyright(C) 1999,2005 Pthreads-win32 contributors
12  *
13  *      Contact Email: rpj@callisto.canberra.edu.au
14  *
15  *      The current list of contributors is contained
16  *      in the file CONTRIBUTORS included with the source
17  *      code distribution. The list can also be seen at the
18  *      following World Wide Web location:
19  *      http://sources.redhat.com/pthreads-win32/contributors.html
20  *
21  *      This library is free software; you can redistribute it and/or
22  *      modify it under the terms of the GNU Lesser General Public
23  *      License as published by the Free Software Foundation; either
24  *      version 2 of the License, or (at your option) any later version.
25  *
26  *      This library is distributed in the hope that it will be useful,
27  *      but WITHOUT ANY WARRANTY; without even the implied warranty of
28  *      MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
29  *      Lesser General Public License for more details.
30  *
31  *      You should have received a copy of the GNU Lesser General Public
32  *      License along with this library in the file COPYING.LIB;
33  *      if not, write to the Free Software Foundation, Inc.,
34  *      59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
35  */
36 
37 /*
38  * About MCS locks:
39  *
40  * MCS locks are queue-based locks, where the queue nodes are local to the
41  * thread. The 'lock' is nothing more than a global pointer that points to
42  * the last node in the queue, or is NULL if the queue is empty.
43  *
44  * Originally designed for use as spin locks requiring no kernel resources
45  * for synchronisation or blocking, the implementation below has adapted
46  * the MCS spin lock for use as a general mutex that will suspend threads
47  * when there is lock contention.
48  *
49  * Because the queue nodes are thread-local, most of the memory read/write
50  * operations required to add or remove nodes from the queue do not trigger
51  * cache-coherence updates.
52  *
53  * Like 'named' mutexes, MCS locks consume system resources transiently -
54  * they are able to acquire and free resources automatically - but MCS
55  * locks do not require any unique 'name' to identify the lock to all
56  * threads using it.
57  *
58  * Usage of MCS locks:
59  *
60  * - you need a global ptw32_mcs_lock_t instance initialised to 0 or NULL.
61  * - you need a local thread-scope ptw32_mcs_local_node_t instance, which
62  *   may serve several different locks but you need at least one node for
63  *   every lock held concurrently by a thread.
64  *
65  * E.g.:
66  *
67  * ptw32_mcs_lock_t lock1 = 0;
68  * ptw32_mcs_lock_t lock2 = 0;
69  *
70  * void *mythread(void *arg)
71  * {
72  *   ptw32_mcs_local_node_t node;
73  *
74  *   ptw32_mcs_acquire (&lock1, &node);
75  *   ptw32_mcs_lock_release (&node);
76  *
77  *   ptw32_mcs_lock_acquire (&lock2, &node);
78  *   ptw32_mcs_lock_release (&node);
79  *   {
80  *      ptw32_mcs_local_node_t nodex;
81  *
82  *      ptw32_mcs_lock_acquire (&lock1, &node);
83  *      ptw32_mcs_lock_acquire (&lock2, &nodex);
84  *
85  *      ptw32_mcs_lock_release (&nodex);
86  *      ptw32_mcs_lock_release (&node);
87  *   }
88  *   return (void *)0;
89  * }
90  */
91 
92 #include "pthread.h"
93 #include "sched.h"
94 #include "implement.h"
95 
96 /*
97  * ptw32_mcs_flag_set -- notify another thread about an event.
98  *
99  * Set event if an event handle has been stored in the flag, and
100  * set flag to -1 otherwise. Note that -1 cannot be a valid handle value.
101  */
102 INLINE void
ptw32_mcs_flag_set(HANDLE * flag)103 ptw32_mcs_flag_set (HANDLE * flag)
104 {
105   HANDLE e = (HANDLE)(PTW32_INTERLOCKED_SIZE)PTW32_INTERLOCKED_COMPARE_EXCHANGE_SIZE(
106 						(PTW32_INTERLOCKED_SIZEPTR)flag,
107 						(PTW32_INTERLOCKED_SIZE)-1,
108 						(PTW32_INTERLOCKED_SIZE)0);
109   if ((HANDLE)0 != e)
110     {
111       /* another thread has already stored an event handle in the flag */
112       SetEvent(e);
113     }
114 }
115 
116 /*
117  * ptw32_mcs_flag_set -- wait for notification from another.
118  *
119  * Store an event handle in the flag and wait on it if the flag has not been
120  * set, and proceed without creating an event otherwise.
121  */
122 INLINE void
ptw32_mcs_flag_wait(HANDLE * flag)123 ptw32_mcs_flag_wait (HANDLE * flag)
124 {
125   if ((PTW32_INTERLOCKED_LONG)0 ==
126         PTW32_INTERLOCKED_EXCHANGE_ADD_SIZE((PTW32_INTERLOCKED_SIZEPTR)flag,
127                                             (PTW32_INTERLOCKED_SIZE)0)) /* MBR fence */
128     {
129       /* the flag is not set. create event. */
130 
131       HANDLE e = CreateEvent(NULL, PTW32_FALSE, PTW32_FALSE, NULL);
132 
133       if ((PTW32_INTERLOCKED_SIZE)0 == PTW32_INTERLOCKED_COMPARE_EXCHANGE_SIZE(
134 			                  (PTW32_INTERLOCKED_SIZEPTR)flag,
135 			                  (PTW32_INTERLOCKED_SIZE)e,
136 			                  (PTW32_INTERLOCKED_SIZE)0))
137 	{
138 	  /* stored handle in the flag. wait on it now. */
139 	  WaitForSingleObject(e, INFINITE);
140 	}
141 
142       CloseHandle(e);
143     }
144 }
145 
146 /*
147  * ptw32_mcs_lock_acquire -- acquire an MCS lock.
148  *
149  * See:
150  * J. M. Mellor-Crummey and M. L. Scott.
151  * Algorithms for Scalable Synchronization on Shared-Memory Multiprocessors.
152  * ACM Transactions on Computer Systems, 9(1):21-65, Feb. 1991.
153  */
154 #if defined(PTW32_BUILD_INLINED)
155 INLINE
156 #endif /* PTW32_BUILD_INLINED */
157 void
ptw32_mcs_lock_acquire(ptw32_mcs_lock_t * lock,ptw32_mcs_local_node_t * node)158 ptw32_mcs_lock_acquire (ptw32_mcs_lock_t * lock, ptw32_mcs_local_node_t * node)
159 {
160   ptw32_mcs_local_node_t  *pred;
161 
162   node->lock = lock;
163   node->nextFlag = 0;
164   node->readyFlag = 0;
165   node->next = 0; /* initially, no successor */
166 
167   /* queue for the lock */
168   pred = (ptw32_mcs_local_node_t *)PTW32_INTERLOCKED_EXCHANGE_PTR((PTW32_INTERLOCKED_PVOID_PTR)lock,
169 								  (PTW32_INTERLOCKED_PVOID)node);
170 
171   if (0 != pred)
172     {
173       /* the lock was not free. link behind predecessor. */
174       pred->next = node;
175       ptw32_mcs_flag_set(&pred->nextFlag);
176       ptw32_mcs_flag_wait(&node->readyFlag);
177     }
178 }
179 
180 /*
181  * ptw32_mcs_lock_release -- release an MCS lock.
182  *
183  * See:
184  * J. M. Mellor-Crummey and M. L. Scott.
185  * Algorithms for Scalable Synchronization on Shared-Memory Multiprocessors.
186  * ACM Transactions on Computer Systems, 9(1):21-65, Feb. 1991.
187  */
188 #if defined(PTW32_BUILD_INLINED)
189 INLINE
190 #endif /* PTW32_BUILD_INLINED */
191 void
ptw32_mcs_lock_release(ptw32_mcs_local_node_t * node)192 ptw32_mcs_lock_release (ptw32_mcs_local_node_t * node)
193 {
194   ptw32_mcs_lock_t *lock = node->lock;
195   ptw32_mcs_local_node_t *next =
196     (ptw32_mcs_local_node_t *)
197       PTW32_INTERLOCKED_EXCHANGE_ADD_SIZE((PTW32_INTERLOCKED_SIZEPTR)&node->next, (PTW32_INTERLOCKED_SIZE)0); /* MBR fence */
198 
199   if (0 == next)
200     {
201       /* no known successor */
202 
203       if (node == (ptw32_mcs_local_node_t *)
204 	  PTW32_INTERLOCKED_COMPARE_EXCHANGE_PTR((PTW32_INTERLOCKED_PVOID_PTR)lock,
205 						 (PTW32_INTERLOCKED_PVOID)0,
206 						 (PTW32_INTERLOCKED_PVOID)node))
207 	{
208 	  /* no successor, lock is free now */
209 	  return;
210 	}
211 
212       /* A successor has started enqueueing behind us so wait for them to link to us */
213       ptw32_mcs_flag_wait(&node->nextFlag);
214       next = (ptw32_mcs_local_node_t *)
215 	PTW32_INTERLOCKED_EXCHANGE_ADD_SIZE((PTW32_INTERLOCKED_SIZEPTR)&node->next, (PTW32_INTERLOCKED_SIZE)0); /* MBR fence */
216     }
217 
218   /* pass the lock */
219   ptw32_mcs_flag_set(&next->readyFlag);
220 }
221 
222 /*
223   * ptw32_mcs_lock_try_acquire
224  */
225 #if defined(PTW32_BUILD_INLINED)
226 INLINE
227 #endif /* PTW32_BUILD_INLINED */
228 int
ptw32_mcs_lock_try_acquire(ptw32_mcs_lock_t * lock,ptw32_mcs_local_node_t * node)229 ptw32_mcs_lock_try_acquire (ptw32_mcs_lock_t * lock, ptw32_mcs_local_node_t * node)
230 {
231   node->lock = lock;
232   node->nextFlag = 0;
233   node->readyFlag = 0;
234   node->next = 0; /* initially, no successor */
235 
236   return ((PTW32_INTERLOCKED_PVOID)PTW32_INTERLOCKED_COMPARE_EXCHANGE_PTR((PTW32_INTERLOCKED_PVOID_PTR)lock,
237                                                         (PTW32_INTERLOCKED_PVOID)node,
238                                                         (PTW32_INTERLOCKED_PVOID)0)
239                                  == (PTW32_INTERLOCKED_PVOID)0) ? 0 : EBUSY;
240 }
241 
242 /*
243  * ptw32_mcs_node_transfer -- move an MCS lock local node, usually from thread
244  * space to, for example, global space so that another thread can release
245  * the lock on behalf of the current lock owner.
246  *
247  * Example: used in pthread_barrier_wait where we want the last thread out of
248  * the barrier to release the lock owned by the last thread to enter the barrier
249  * (the one that releases all threads but not necessarily the last to leave).
250  *
251  * Should only be called by the thread that has the lock.
252  */
253 #if defined(PTW32_BUILD_INLINED)
254 INLINE
255 #endif /* PTW32_BUILD_INLINED */
256 void
ptw32_mcs_node_transfer(ptw32_mcs_local_node_t * new_node,ptw32_mcs_local_node_t * old_node)257 ptw32_mcs_node_transfer (ptw32_mcs_local_node_t * new_node, ptw32_mcs_local_node_t * old_node)
258 {
259   new_node->lock = old_node->lock;
260   new_node->nextFlag = 0; /* Not needed - used only in initial Acquire */
261   new_node->readyFlag = 0; /* Not needed - we were waiting on this */
262   new_node->next = 0;
263 
264   if ((ptw32_mcs_local_node_t *)PTW32_INTERLOCKED_COMPARE_EXCHANGE_PTR((PTW32_INTERLOCKED_PVOID_PTR)new_node->lock,
265                                                                        (PTW32_INTERLOCKED_PVOID)new_node,
266                                                                        (PTW32_INTERLOCKED_PVOID)old_node)
267        != old_node)
268     {
269       /*
270        * A successor has queued after us, so wait for them to link to us
271        */
272       while (old_node->next == 0)
273         {
274           sched_yield();
275         }
276       new_node->next = old_node->next;
277     }
278 }
279