1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 2009
5 * Work-stealing Deque data structure
7 * The implementation uses Double-Ended Queues with lock-free access
8 * (thereby often called "deque") as described in
10 * D.Chase and Y.Lev, Dynamic Circular Work-Stealing Deque.
11 * SPAA'05, July 2005, Las Vegas, USA.
12 * ACM 1-58113-986-1/05/0007
14 * Author: Jost Berthold MSRC 07-09/2008
16 * The DeQue is held as a circular array with known length. Positions
17 * of top (read-end) and bottom (write-end) always increase, and the
18 * array is accessed with indices modulo array-size. While this bears
19 * the risk of overflow, we assume that (with 64 bit indices), a
20 * program must run very long to reach that point.
22 * The write end of the queue (position bottom) can only be used with
23 * mutual exclusion, i.e. by exactly one caller at a time. At this
24 * end, new items can be enqueued using pushBottom()/newSpark(), and
25 * removed using popBottom()/reclaimSpark() (the latter implying a cas
26 * synchronisation with potential concurrent readers for the case of
29 * Multiple readers can steal from the read end (position top), and
30 * are synchronised without a lock, based on a cas of the top
31 * position. One reader wins, the others return NULL for a failure.
33 * Both popBottom and steal also return NULL when the queue is empty.
35 * ---------------------------------------------------------------------------*/
40 #include "SMP.h" // for cas
42 #if defined(THREADED_RTS)
44 #define CASTOP(addr,old,new) ((old) == cas(((StgPtr)addr),(old),(new)))
46 /* -----------------------------------------------------------------------------
48 * -------------------------------------------------------------------------- */
50 /* internal helpers ... */
57 /* StgWord is unsigned anyway, only catch 0 */
59 barf("DeQue,roundUp2: invalid size 0 requested");
61 /* at least 1 bit set, shift up to its place */
63 rounded = rounded << 1;
64 } while (0 != (val = val>>1));
74 realsize = roundUp2(size); /* to compute modulo as a bitwise & */
76 q = (WSDeque*) stgMallocBytes(sizeof(WSDeque), /* admin fields */
78 q->elements = stgMallocBytes(realsize * sizeof(StgClosurePtr), /* dataspace */
79 "newWSDeque:data space");
82 q->topBound=0; /* read by writer, updated each time top is read */
84 q->size = realsize; /* power of 2 */
85 q->moduloSize = realsize - 1; /* n % size == n & moduloSize */
87 ASSERT_WSDEQUE_INVARIANTS(q);
91 /* -----------------------------------------------------------------------------
93 * -------------------------------------------------------------------------- */
96 freeWSDeque (WSDeque *q)
102 /* -----------------------------------------------------------------------------
104 * popWSDeque: remove an element from the write end of the queue.
105 * Returns the removed spark, and NULL if a race is lost or the pool
108 * If only one spark is left in the pool, we synchronise with
109 * concurrently stealing threads by using cas to modify the top field.
110 * This routine should NEVER be called by a task which does not own
113 * -------------------------------------------------------------------------- */
116 popWSDeque (WSDeque *q)
118 /* also a bit tricky, has to avoid concurrent steal() calls by
119 accessing top with cas, when there is only one element left */
125 ASSERT_WSDEQUE_INVARIANTS(q);
128 /* "decrement b as a test, see what happens" */
130 pos = (q->elements) + (b & (q->moduloSize));
131 t = q->top; /* using topBound would give an *upper* bound, we
132 need a lower bound. We use the real top here, but
133 can update the topBound value */
136 if (currSize < 0) { /* was empty before decrementing b, set b
137 consistently and abort */
142 if (currSize > 0) { /* no danger, still elements in buffer after b-- */
145 /* otherwise, has someone meanwhile stolen the same (last) element?
146 Check and increment top value to know */
147 if ( !(CASTOP(&(q->top),t,t+1)) ) {
148 removed = NULL; /* no success, but continue adjusting bottom */
150 q->bottom = t+1; /* anyway, empty now. Adjust bottom consistently. */
151 q->topBound = t+1; /* ...and cached top value as well */
153 ASSERT_WSDEQUE_INVARIANTS(q);
154 ASSERT(q->bottom >= q->top);
159 /* -----------------------------------------------------------------------------
161 * -------------------------------------------------------------------------- */
164 stealWSDeque_ (WSDeque *q)
172 // Can't do this on someone else's spark pool:
173 // ASSERT_WSDEQUE_INVARIANTS(q);
178 // NB. b and t are unsigned; we need a signed value for the test
180 if ((long)b - (long)t <= 0 ) {
181 return NULL; /* already looks empty, abort */
184 /* now access array, see pushBottom() */
185 arraybase = q->elements;
187 pos = arraybase + (t & sz);
190 /* now decide whether we have won */
191 if ( !(CASTOP(&(q->top),t,t+1)) ) {
192 /* lost the race, someon else has changed top in the meantime */
194 } /* else: OK, top has been incremented by the cas call */
196 // Can't do this on someone else's spark pool:
197 // ASSERT_WSDEQUE_INVARIANTS(q);
203 stealWSDeque (WSDeque *q)
208 stolen = stealWSDeque_(q);
209 } while (stolen == NULL && !looksEmptyWSDeque(q));
217 /* enqueue an element. Should always succeed by resizing the array
218 (not implemented yet, silently fails in that case). */
220 pushWSDeque (WSDeque* q, void * elem)
224 StgWord sz = q->moduloSize;
225 StgWord b = q->bottom;
227 ASSERT_WSDEQUE_INVARIANTS(q);
229 /* we try to avoid reading q->top (accessed by all) and use
230 q->topBound (accessed only by writer) instead.
231 This is why we do not just call empty(q) here.
234 if ( (StgInt)b - (StgInt)t >= (StgInt)sz ) {
235 /* NB. 1. sz == q->size - 1, thus ">="
236 2. signed comparison, it is possible that t > b
238 /* could be full, check the real top value in this case */
241 if (b - t >= sz) { /* really no space left :-( */
242 /* reallocate the array, copying the values. Concurrent steal()s
243 will in the meantime use the old one and modify only top.
244 This means: we cannot safely free the old space! Can keep it
245 on a free list internally here...
247 Potential bug in combination with steal(): if array is
248 replaced, it is unclear which one concurrent steal operations
249 use. Must read the array base address in advance in steal().
251 #if defined(DISCARD_NEW)
252 ASSERT_WSDEQUE_INVARIANTS(q);
253 return rtsFalse; // we didn't push anything
255 /* could make room by incrementing the top position here. In
256 * this case, should use CASTOP. If this fails, someone else has
257 * removed something, and new room will be available.
259 ASSERT_WSDEQUE_INVARIANTS(q);
263 pos = (q->elements) + (b & sz);
267 ASSERT_WSDEQUE_INVARIANTS(q);