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 popWSDeque and stealWSDeque also return NULL when the queue is empty.
35 * Testing: see testsuite/tests/ghc-regress/rts/testwsdeque.c. If
36 * there's anything wrong with the deque implementation, this test
37 * will probably catch it.
39 * ---------------------------------------------------------------------------*/
41 #include "PosixSource.h"
47 #define CASTOP(addr,old,new) ((old) == cas(((StgPtr)addr),(old),(new)))
49 /* -----------------------------------------------------------------------------
51 * -------------------------------------------------------------------------- */
53 /* internal helpers ... */
60 /* StgWord is unsigned anyway, only catch 0 */
62 barf("DeQue,roundUp2: invalid size 0 requested");
64 /* at least 1 bit set, shift up to its place */
66 rounded = rounded << 1;
67 } while (0 != (val = val>>1));
77 realsize = roundUp2(size); /* to compute modulo as a bitwise & */
79 q = (WSDeque*) stgMallocBytes(sizeof(WSDeque), /* admin fields */
81 q->elements = stgMallocBytes(realsize * sizeof(StgClosurePtr), /* dataspace */
82 "newWSDeque:data space");
85 q->topBound=0; /* read by writer, updated each time top is read */
87 q->size = realsize; /* power of 2 */
88 q->moduloSize = realsize - 1; /* n % size == n & moduloSize */
90 ASSERT_WSDEQUE_INVARIANTS(q);
94 /* -----------------------------------------------------------------------------
96 * -------------------------------------------------------------------------- */
99 freeWSDeque (WSDeque *q)
101 stgFree(q->elements);
105 /* -----------------------------------------------------------------------------
107 * popWSDeque: remove an element from the write end of the queue.
108 * Returns the removed spark, and NULL if a race is lost or the pool
111 * If only one spark is left in the pool, we synchronise with
112 * concurrently stealing threads by using cas to modify the top field.
113 * This routine should NEVER be called by a task which does not own
116 * -------------------------------------------------------------------------- */
119 popWSDeque (WSDeque *q)
121 /* also a bit tricky, has to avoid concurrent steal() calls by
122 accessing top with cas, when there is only one element left */
127 ASSERT_WSDEQUE_INVARIANTS(q);
131 // "decrement b as a test, see what happens"
135 // very important that the following read of q->top does not occur
136 // before the earlier write to q->bottom.
137 store_load_barrier();
139 t = q->top; /* using topBound would give an *upper* bound, we
140 need a lower bound. We use the real top here, but
141 can update the topBound value */
143 currSize = (long)b - (long)t;
144 if (currSize < 0) { /* was empty before decrementing b, set b
145 consistently and abort */
150 // read the element at b
151 removed = q->elements[b & q->moduloSize];
153 if (currSize > 0) { /* no danger, still elements in buffer after b-- */
154 // debugBelch("popWSDeque: t=%ld b=%ld = %ld\n", t, b, removed);
157 /* otherwise, has someone meanwhile stolen the same (last) element?
158 Check and increment top value to know */
159 if ( !(CASTOP(&(q->top),t,t+1)) ) {
160 removed = NULL; /* no success, but continue adjusting bottom */
162 q->bottom = t+1; /* anyway, empty now. Adjust bottom consistently. */
163 q->topBound = t+1; /* ...and cached top value as well */
165 ASSERT_WSDEQUE_INVARIANTS(q);
166 ASSERT(q->bottom >= q->top);
168 // debugBelch("popWSDeque: t=%ld b=%ld = %ld\n", t, b, removed);
173 /* -----------------------------------------------------------------------------
175 * -------------------------------------------------------------------------- */
178 stealWSDeque_ (WSDeque *q)
183 // Can't do this on someone else's spark pool:
184 // ASSERT_WSDEQUE_INVARIANTS(q);
186 // NB. these loads must be ordered, otherwise there is a race
187 // between steal and pop.
192 // NB. b and t are unsigned; we need a signed value for the test
193 // below, because it is possible that t > b during a
194 // concurrent popWSQueue() operation.
195 if ((long)b - (long)t <= 0 ) {
196 return NULL; /* already looks empty, abort */
199 /* now access array, see pushBottom() */
200 stolen = q->elements[t & q->moduloSize];
202 /* now decide whether we have won */
203 if ( !(CASTOP(&(q->top),t,t+1)) ) {
204 /* lost the race, someon else has changed top in the meantime */
206 } /* else: OK, top has been incremented by the cas call */
208 // debugBelch("stealWSDeque_: t=%d b=%d\n", t, b);
210 // Can't do this on someone else's spark pool:
211 // ASSERT_WSDEQUE_INVARIANTS(q);
217 stealWSDeque (WSDeque *q)
222 stolen = stealWSDeque_(q);
223 } while (stolen == NULL && !looksEmptyWSDeque(q));
228 /* -----------------------------------------------------------------------------
230 * -------------------------------------------------------------------------- */
234 /* enqueue an element. Should always succeed by resizing the array
235 (not implemented yet, silently fails in that case). */
237 pushWSDeque (WSDeque* q, void * elem)
241 StgWord sz = q->moduloSize;
243 ASSERT_WSDEQUE_INVARIANTS(q);
245 /* we try to avoid reading q->top (accessed by all) and use
246 q->topBound (accessed only by writer) instead.
247 This is why we do not just call empty(q) here.
251 if ( (StgInt)b - (StgInt)t >= (StgInt)sz ) {
252 /* NB. 1. sz == q->size - 1, thus ">="
253 2. signed comparison, it is possible that t > b
255 /* could be full, check the real top value in this case */
258 if (b - t >= sz) { /* really no space left :-( */
259 /* reallocate the array, copying the values. Concurrent steal()s
260 will in the meantime use the old one and modify only top.
261 This means: we cannot safely free the old space! Can keep it
262 on a free list internally here...
264 Potential bug in combination with steal(): if array is
265 replaced, it is unclear which one concurrent steal operations
266 use. Must read the array base address in advance in steal().
268 #if defined(DISCARD_NEW)
269 ASSERT_WSDEQUE_INVARIANTS(q);
270 return rtsFalse; // we didn't push anything
272 /* could make room by incrementing the top position here. In
273 * this case, should use CASTOP. If this fails, someone else has
274 * removed something, and new room will be available.
276 ASSERT_WSDEQUE_INVARIANTS(q);
281 q->elements[b & sz] = elem;
284 ASSERT_WSDEQUE_INVARIANTS(q);