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