simplify weak pointer processing
[ghc.git] / rts / sm / MarkWeak.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team 1998-2008
4 *
5 * Weak pointers and weak-like things in the GC
6 *
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
9 *
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
11 *
12 * ---------------------------------------------------------------------------*/
13
14 #include "PosixSource.h"
15 #include "Rts.h"
16
17 #include "MarkWeak.h"
18 #include "GC.h"
19 #include "GCThread.h"
20 #include "Evac.h"
21 #include "Trace.h"
22 #include "Schedule.h"
23 #include "Weak.h"
24 #include "Storage.h"
25
26 /* -----------------------------------------------------------------------------
27 Weak Pointers
28
29 traverse_weak_ptr_list is called possibly many times during garbage
30 collection. It returns a flag indicating whether it did any work
31 (i.e. called evacuate on any live pointers).
32
33 Invariant: traverse_weak_ptr_list is called when the heap is in an
34 idempotent state. That means that there are no pending
35 evacuate/scavenge operations. This invariant helps the weak
36 pointer code decide which weak pointers are dead - if there are no
37 new live weak pointers, then all the currently unreachable ones are
38 dead.
39
40 For generational GC: we just don't try to finalize weak pointers in
41 older generations than the one we're collecting. This could
42 probably be optimised by keeping per-generation lists of weak
43 pointers, but for a few weak pointers this scheme will work.
44
45 There are three distinct stages to processing weak pointers:
46
47 - weak_stage == WeakPtrs
48
49 We process all the weak pointers whos keys are alive (evacuate
50 their values and finalizers), and repeat until we can find no new
51 live keys. If no live keys are found in this pass, then we
52 evacuate the finalizers of all the dead weak pointers in order to
53 run them.
54
55 - weak_stage == WeakThreads
56
57 Now, we discover which *threads* are still alive. Pointers to
58 threads from the all_threads and main thread lists are the
59 weakest of all: a pointers from the finalizer of a dead weak
60 pointer can keep a thread alive. Any threads found to be unreachable
61 are evacuated and placed on the resurrected_threads list so we
62 can send them a signal later.
63
64 - weak_stage == WeakDone
65
66 No more evacuation is done.
67
68 -------------------------------------------------------------------------- */
69
70 /* Which stage of processing various kinds of weak pointer are we at?
71 * (see traverse_weak_ptr_list() below for discussion).
72 */
73 typedef enum { WeakPtrs, WeakThreads, WeakDone } WeakStage;
74 static WeakStage weak_stage;
75
76 /* Weak pointers
77 */
78 StgWeak *old_weak_ptr_list; // also pending finaliser list
79
80 // List of threads found to be unreachable
81 StgTSO *resurrected_threads;
82
83 // List of blocked threads found to have pending throwTos
84 StgTSO *exception_threads;
85
86 static void resurrectUnreachableThreads (generation *gen);
87 static rtsBool tidyThreadList (generation *gen);
88
89 void
90 initWeakForGC(void)
91 {
92 old_weak_ptr_list = weak_ptr_list;
93 weak_ptr_list = NULL;
94 weak_stage = WeakPtrs;
95 resurrected_threads = END_TSO_QUEUE;
96 exception_threads = END_TSO_QUEUE;
97 }
98
99 rtsBool
100 traverseWeakPtrList(void)
101 {
102 StgWeak *w, **last_w, *next_w;
103 StgClosure *new;
104 rtsBool flag = rtsFalse;
105 const StgInfoTable *info;
106
107 switch (weak_stage) {
108
109 case WeakDone:
110 return rtsFalse;
111
112 case WeakPtrs:
113 /* doesn't matter where we evacuate values/finalizers to, since
114 * these pointers are treated as roots (iff the keys are alive).
115 */
116 gct->evac_gen = 0;
117
118 last_w = &old_weak_ptr_list;
119 for (w = old_weak_ptr_list; w != NULL; w = next_w) {
120
121 /* There might be a DEAD_WEAK on the list if finalizeWeak# was
122 * called on a live weak pointer object. Just remove it.
123 */
124 if (w->header.info == &stg_DEAD_WEAK_info) {
125 next_w = ((StgDeadWeak *)w)->link;
126 *last_w = next_w;
127 continue;
128 }
129
130 info = get_itbl(w);
131 switch (info->type) {
132
133 case WEAK:
134 /* Now, check whether the key is reachable.
135 */
136 new = isAlive(w->key);
137 if (new != NULL) {
138 w->key = new;
139 // evacuate the value and finalizer
140 evacuate(&w->value);
141 evacuate(&w->finalizer);
142 // remove this weak ptr from the old_weak_ptr list
143 *last_w = w->link;
144 // and put it on the new weak ptr list
145 next_w = w->link;
146 w->link = weak_ptr_list;
147 weak_ptr_list = w;
148 flag = rtsTrue;
149
150 debugTrace(DEBUG_weak,
151 "weak pointer still alive at %p -> %p",
152 w, w->key);
153 continue;
154 }
155 else {
156 last_w = &(w->link);
157 next_w = w->link;
158 continue;
159 }
160
161 default:
162 barf("traverseWeakPtrList: not WEAK");
163 }
164 }
165
166 /* If we didn't make any changes, then we can go round and kill all
167 * the dead weak pointers. The old_weak_ptr list is used as a list
168 * of pending finalizers later on.
169 */
170 if (flag == rtsFalse) {
171 for (w = old_weak_ptr_list; w; w = w->link) {
172 evacuate(&w->finalizer);
173 }
174
175 // Next, move to the WeakThreads stage after fully
176 // scavenging the finalizers we've just evacuated.
177 weak_stage = WeakThreads;
178 }
179
180 return rtsTrue;
181
182 case WeakThreads:
183 /* Now deal with the step->threads lists, which behave somewhat like
184 * the weak ptr list. If we discover any threads that are about to
185 * become garbage, we wake them up and administer an exception.
186 */
187 {
188 nat g;
189
190 // Traverse thread lists for generations we collected...
191 // ToDo when we have one gen per capability:
192 // for (n = 0; n < n_capabilities; n++) {
193 // if (tidyThreadList(&nurseries[n])) {
194 // flag = rtsTrue;
195 // }
196 // }
197 for (g = 0; g <= N; g++) {
198 if (tidyThreadList(&generations[g])) {
199 flag = rtsTrue;
200 }
201 }
202
203 /* If we evacuated any threads, we need to go back to the scavenger.
204 */
205 if (flag) return rtsTrue;
206
207 /* And resurrect any threads which were about to become garbage.
208 */
209 {
210 nat g;
211 for (g = 0; g <= N; g++) {
212 resurrectUnreachableThreads(&generations[g]);
213 }
214 }
215
216 /* Finally, we can update the blackhole_queue. This queue
217 * simply strings together TSOs blocked on black holes, it is
218 * not intended to keep anything alive. Hence, we do not follow
219 * pointers on the blackhole_queue until now, when we have
220 * determined which TSOs are otherwise reachable. We know at
221 * this point that all TSOs have been evacuated, however.
222 */
223 {
224 StgTSO **pt;
225 for (pt = &blackhole_queue; *pt != END_TSO_QUEUE; pt = &((*pt)->_link)) {
226 *pt = (StgTSO *)isAlive((StgClosure *)*pt);
227 ASSERT(*pt != NULL);
228 }
229 }
230
231 weak_stage = WeakDone; // *now* we're done,
232 return rtsTrue; // but one more round of scavenging, please
233 }
234
235 default:
236 barf("traverse_weak_ptr_list");
237 return rtsTrue;
238 }
239 }
240
241 static void resurrectUnreachableThreads (generation *gen)
242 {
243 StgTSO *t, *tmp, *next;
244
245 for (t = gen->old_threads; t != END_TSO_QUEUE; t = next) {
246 next = t->global_link;
247
248 // ThreadFinished and ThreadComplete: we have to keep
249 // these on the all_threads list until they
250 // become garbage, because they might get
251 // pending exceptions.
252 switch (t->what_next) {
253 case ThreadKilled:
254 case ThreadComplete:
255 continue;
256 default:
257 tmp = t;
258 evacuate((StgClosure **)&tmp);
259 tmp->global_link = resurrected_threads;
260 resurrected_threads = tmp;
261 }
262 }
263 }
264
265 static rtsBool tidyThreadList (generation *gen)
266 {
267 StgTSO *t, *tmp, *next, **prev;
268 rtsBool flag = rtsFalse;
269
270 prev = &gen->old_threads;
271
272 for (t = gen->old_threads; t != END_TSO_QUEUE; t = next) {
273
274 tmp = (StgTSO *)isAlive((StgClosure *)t);
275
276 if (tmp != NULL) {
277 t = tmp;
278 }
279
280 ASSERT(get_itbl(t)->type == TSO);
281 if (t->what_next == ThreadRelocated) {
282 next = t->_link;
283 *prev = next;
284 continue;
285 }
286
287 next = t->global_link;
288
289 // This is a good place to check for blocked
290 // exceptions. It might be the case that a thread is
291 // blocked on delivering an exception to a thread that
292 // is also blocked - we try to ensure that this
293 // doesn't happen in throwTo(), but it's too hard (or
294 // impossible) to close all the race holes, so we
295 // accept that some might get through and deal with
296 // them here. A GC will always happen at some point,
297 // even if the system is otherwise deadlocked.
298 //
299 // If an unreachable thread has blocked
300 // exceptions, we really want to perform the
301 // blocked exceptions rather than throwing
302 // BlockedIndefinitely exceptions. This is the
303 // only place we can discover such threads.
304 // The target thread might even be
305 // ThreadFinished or ThreadKilled. Bugs here
306 // will only be seen when running on a
307 // multiprocessor.
308 if (t->blocked_exceptions != END_TSO_QUEUE) {
309 if (tmp == NULL) {
310 evacuate((StgClosure **)&t);
311 flag = rtsTrue;
312 }
313 t->global_link = exception_threads;
314 exception_threads = t;
315 *prev = next;
316 continue;
317 }
318
319 if (tmp == NULL) {
320 // not alive (yet): leave this thread on the
321 // old_all_threads list.
322 prev = &(t->global_link);
323 }
324 else {
325 // alive
326 *prev = next;
327
328 // move this thread onto the correct threads list.
329 generation *new_gen;
330 new_gen = Bdescr((P_)t)->gen;
331 t->global_link = new_gen->threads;
332 new_gen->threads = t;
333 }
334 }
335
336 return flag;
337 }
338
339 /* -----------------------------------------------------------------------------
340 The blackhole queue
341
342 Threads on this list behave like weak pointers during the normal
343 phase of garbage collection: if the blackhole is reachable, then
344 the thread is reachable too.
345 -------------------------------------------------------------------------- */
346 rtsBool
347 traverseBlackholeQueue (void)
348 {
349 StgTSO *prev, *t, *tmp;
350 rtsBool flag;
351 nat type;
352
353 flag = rtsFalse;
354 prev = NULL;
355
356 for (t = blackhole_queue; t != END_TSO_QUEUE; prev=t, t = t->_link) {
357 // if the thread is not yet alive...
358 if (! (tmp = (StgTSO *)isAlive((StgClosure*)t))) {
359 // if the closure it is blocked on is either (a) a
360 // reachable BLAKCHOLE or (b) not a BLACKHOLE, then we
361 // make the thread alive.
362 if (!isAlive(t->block_info.closure)) {
363 type = get_itbl(t->block_info.closure)->type;
364 if (type == BLACKHOLE || type == CAF_BLACKHOLE) {
365 continue;
366 }
367 }
368 evacuate((StgClosure **)&t);
369 if (prev) {
370 prev->_link = t;
371 } else {
372 blackhole_queue = t;
373 }
374 // no write barrier when on the blackhole queue,
375 // because we traverse the whole queue on every GC.
376 flag = rtsTrue;
377 }
378 }
379 return flag;
380 }
381
382 /* -----------------------------------------------------------------------------
383 Evacuate every weak pointer object on the weak_ptr_list, and update
384 the link fields.
385
386 ToDo: with a lot of weak pointers, this will be expensive. We
387 should have a per-GC weak pointer list, just like threads.
388 -------------------------------------------------------------------------- */
389
390 void
391 markWeakPtrList ( void )
392 {
393 StgWeak *w, **last_w;
394
395 last_w = &weak_ptr_list;
396 for (w = weak_ptr_list; w; w = w->link) {
397 // w might be WEAK, EVACUATED, or DEAD_WEAK (actually CON_STATIC) here
398 ASSERT(IS_FORWARDING_PTR(w->header.info)
399 || w->header.info == &stg_DEAD_WEAK_info
400 || get_itbl(w)->type == WEAK);
401 evacuate((StgClosure **)last_w);
402 w = *last_w;
403 if (w->header.info == &stg_DEAD_WEAK_info) {
404 last_w = &(((StgDeadWeak*)w)->link);
405 } else {
406 last_w = &(w->link);
407 }
408 }
409 }
410