Merge branch 'stable-2.0'
[bpt/guile.git] / libguile / gc.c
1 /* Copyright (C) 1995,1996,1997,1998,1999,2000,2001, 2002, 2003, 2006,
2 * 2008, 2009, 2010, 2011, 2012, 2013 Free Software Foundation, Inc.
3 *
4 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public License
6 * as published by the Free Software Foundation; either version 3 of
7 * the License, or (at your option) any later version.
8 *
9 * This library is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
17 * 02110-1301 USA
18 */
19
20 /* #define DEBUGINFO */
21
22 #ifdef HAVE_CONFIG_H
23 # include <config.h>
24 #endif
25
26 #include "libguile/gen-scmconfig.h"
27
28 #include <stdio.h>
29 #include <errno.h>
30 #include <string.h>
31 #include <stdlib.h>
32 #include <math.h>
33
34 #ifdef __ia64__
35 #include <ucontext.h>
36 extern unsigned long * __libc_ia64_register_backing_store_base;
37 #endif
38
39 #include "libguile/_scm.h"
40 #include "libguile/eval.h"
41 #include "libguile/stime.h"
42 #include "libguile/stackchk.h"
43 #include "libguile/struct.h"
44 #include "libguile/smob.h"
45 #include "libguile/arrays.h"
46 #include "libguile/async.h"
47 #include "libguile/ports.h"
48 #include "libguile/root.h"
49 #include "libguile/simpos.h"
50 #include "libguile/strings.h"
51 #include "libguile/vectors.h"
52 #include "libguile/hashtab.h"
53 #include "libguile/tags.h"
54
55 #include "libguile/validate.h"
56 #include "libguile/deprecation.h"
57 #include "libguile/gc.h"
58 #include "libguile/dynwind.h"
59
60 #include "libguile/bdw-gc.h"
61
62 /* For GC_set_start_callback. */
63 #include <gc/gc_mark.h>
64
65 #ifdef GUILE_DEBUG_MALLOC
66 #include "libguile/debug-malloc.h"
67 #endif
68
69 #ifdef HAVE_UNISTD_H
70 #include <unistd.h>
71 #endif
72
73 /* Size in bytes of the initial heap. This should be about the size of
74 result of 'guile -c "(display (assq-ref (gc-stats)
75 'heap-total-allocated))"'. */
76
77 #define DEFAULT_INITIAL_HEAP_SIZE (128 * 1024 * SIZEOF_SCM_T_BITS)
78
79 /* Set this to != 0 if every cell that is accessed shall be checked:
80 */
81 int scm_debug_cell_accesses_p = 0;
82 int scm_expensive_debug_cell_accesses_p = 0;
83
84 /* Set this to 0 if no additional gc's shall be performed, otherwise set it to
85 * the number of cell accesses after which a gc shall be called.
86 */
87 int scm_debug_cells_gc_interval = 0;
88
89 /* Hash table that keeps a reference to objects the user wants to protect from
90 garbage collection. */
91 static SCM scm_protects;
92
93
94 #if (SCM_DEBUG_CELL_ACCESSES == 1)
95
96
97 /*
98
99 Assert that the given object is a valid reference to a valid cell. This
100 test involves to determine whether the object is a cell pointer, whether
101 this pointer actually points into a heap segment and whether the cell
102 pointed to is not a free cell. Further, additional garbage collections may
103 get executed after a user defined number of cell accesses. This helps to
104 find places in the C code where references are dropped for extremely short
105 periods.
106
107 */
108 void
109 scm_i_expensive_validation_check (SCM cell)
110 {
111 /* If desired, perform additional garbage collections after a user
112 * defined number of cell accesses.
113 */
114 if (scm_debug_cells_gc_interval)
115 {
116 static unsigned int counter = 0;
117
118 if (counter != 0)
119 {
120 --counter;
121 }
122 else
123 {
124 counter = scm_debug_cells_gc_interval;
125 scm_gc ();
126 }
127 }
128 }
129
130 /* Whether cell validation is already running. */
131 static int scm_i_cell_validation_already_running = 0;
132
133 void
134 scm_assert_cell_valid (SCM cell)
135 {
136 if (!scm_i_cell_validation_already_running && scm_debug_cell_accesses_p)
137 {
138 scm_i_cell_validation_already_running = 1; /* set to avoid recursion */
139
140 /*
141 During GC, no user-code should be run, and the guile core
142 should use non-protected accessors.
143 */
144 if (scm_gc_running_p)
145 return;
146
147 /*
148 Only scm_in_heap_p and rescanning the heap is wildly
149 expensive.
150 */
151 if (scm_expensive_debug_cell_accesses_p)
152 scm_i_expensive_validation_check (cell);
153
154 scm_i_cell_validation_already_running = 0; /* re-enable */
155 }
156 }
157
158
159
160 SCM_DEFINE (scm_set_debug_cell_accesses_x, "set-debug-cell-accesses!", 1, 0, 0,
161 (SCM flag),
162 "If @var{flag} is @code{#f}, cell access checking is disabled.\n"
163 "If @var{flag} is @code{#t}, cheap cell access checking is enabled,\n"
164 "but no additional calls to garbage collection are issued.\n"
165 "If @var{flag} is a number, strict cell access checking is enabled,\n"
166 "with an additional garbage collection after the given\n"
167 "number of cell accesses.\n"
168 "This procedure only exists when the compile-time flag\n"
169 "@code{SCM_DEBUG_CELL_ACCESSES} was set to 1.")
170 #define FUNC_NAME s_scm_set_debug_cell_accesses_x
171 {
172 if (scm_is_false (flag))
173 {
174 scm_debug_cell_accesses_p = 0;
175 }
176 else if (scm_is_eq (flag, SCM_BOOL_T))
177 {
178 scm_debug_cells_gc_interval = 0;
179 scm_debug_cell_accesses_p = 1;
180 scm_expensive_debug_cell_accesses_p = 0;
181 }
182 else
183 {
184 scm_debug_cells_gc_interval = scm_to_signed_integer (flag, 0, INT_MAX);
185 scm_debug_cell_accesses_p = 1;
186 scm_expensive_debug_cell_accesses_p = 1;
187 }
188 return SCM_UNSPECIFIED;
189 }
190 #undef FUNC_NAME
191
192
193 #endif /* SCM_DEBUG_CELL_ACCESSES == 1 */
194
195 \f
196
197 /* Hooks. */
198 scm_t_c_hook scm_before_gc_c_hook;
199 scm_t_c_hook scm_before_mark_c_hook;
200 scm_t_c_hook scm_before_sweep_c_hook;
201 scm_t_c_hook scm_after_sweep_c_hook;
202 scm_t_c_hook scm_after_gc_c_hook;
203
204
205 static void
206 run_before_gc_c_hook (void)
207 {
208 if (!SCM_I_CURRENT_THREAD)
209 /* GC while a thread is spinning up; punt. */
210 return;
211
212 scm_c_hook_run (&scm_before_gc_c_hook, NULL);
213 }
214
215
216 /* GC Statistics Keeping
217 */
218 unsigned long scm_gc_ports_collected = 0;
219 static long gc_time_taken = 0;
220 static long gc_start_time = 0;
221
222 static unsigned long free_space_divisor;
223 static unsigned long minimum_free_space_divisor;
224 static double target_free_space_divisor;
225
226 static unsigned long protected_obj_count = 0;
227
228
229 SCM_SYMBOL (sym_gc_time_taken, "gc-time-taken");
230 SCM_SYMBOL (sym_heap_size, "heap-size");
231 SCM_SYMBOL (sym_heap_free_size, "heap-free-size");
232 SCM_SYMBOL (sym_heap_total_allocated, "heap-total-allocated");
233 SCM_SYMBOL (sym_heap_allocated_since_gc, "heap-allocated-since-gc");
234 SCM_SYMBOL (sym_protected_objects, "protected-objects");
235 SCM_SYMBOL (sym_times, "gc-times");
236
237
238 /* {Scheme Interface to GC}
239 */
240 extern int scm_gc_malloc_yield_percentage;
241 SCM_DEFINE (scm_gc_stats, "gc-stats", 0, 0, 0,
242 (),
243 "Return an association list of statistics about Guile's current\n"
244 "use of storage.\n")
245 #define FUNC_NAME s_scm_gc_stats
246 {
247 SCM answer;
248 GC_word heap_size, free_bytes, unmapped_bytes, bytes_since_gc, total_bytes;
249 size_t gc_times;
250
251 GC_get_heap_usage_safe (&heap_size, &free_bytes, &unmapped_bytes,
252 &bytes_since_gc, &total_bytes);
253 gc_times = GC_get_gc_no ();
254
255 answer =
256 scm_list_n (scm_cons (sym_gc_time_taken, scm_from_long (gc_time_taken)),
257 scm_cons (sym_heap_size, scm_from_size_t (heap_size)),
258 scm_cons (sym_heap_free_size, scm_from_size_t (free_bytes)),
259 scm_cons (sym_heap_total_allocated,
260 scm_from_size_t (total_bytes)),
261 scm_cons (sym_heap_allocated_since_gc,
262 scm_from_size_t (bytes_since_gc)),
263 scm_cons (sym_protected_objects,
264 scm_from_ulong (protected_obj_count)),
265 scm_cons (sym_times, scm_from_size_t (gc_times)),
266 SCM_UNDEFINED);
267
268 return answer;
269 }
270 #undef FUNC_NAME
271
272
273 SCM_DEFINE (scm_gc_dump, "gc-dump", 0, 0, 0,
274 (void),
275 "Dump information about the garbage collector's internal data "
276 "structures and memory usage to the standard output.")
277 #define FUNC_NAME s_scm_gc_dump
278 {
279 GC_dump ();
280
281 return SCM_UNSPECIFIED;
282 }
283 #undef FUNC_NAME
284
285
286 SCM_DEFINE (scm_object_address, "object-address", 1, 0, 0,
287 (SCM obj),
288 "Return an integer that for the lifetime of @var{obj} is uniquely\n"
289 "returned by this function for @var{obj}")
290 #define FUNC_NAME s_scm_object_address
291 {
292 return scm_from_ulong (SCM_UNPACK (obj));
293 }
294 #undef FUNC_NAME
295
296
297 SCM_DEFINE (scm_gc_disable, "gc-disable", 0, 0, 0,
298 (),
299 "Disables the garbage collector. Nested calls are permitted. "
300 "GC is re-enabled once @code{gc-enable} has been called the "
301 "same number of times @code{gc-disable} was called.")
302 #define FUNC_NAME s_scm_gc_disable
303 {
304 GC_disable ();
305 return SCM_UNSPECIFIED;
306 }
307 #undef FUNC_NAME
308
309 SCM_DEFINE (scm_gc_enable, "gc-enable", 0, 0, 0,
310 (),
311 "Enables the garbage collector.")
312 #define FUNC_NAME s_scm_gc_enable
313 {
314 GC_enable ();
315 return SCM_UNSPECIFIED;
316 }
317 #undef FUNC_NAME
318
319
320 SCM_DEFINE (scm_gc, "gc", 0, 0, 0,
321 (),
322 "Scans all of SCM objects and reclaims for further use those that are\n"
323 "no longer accessible.")
324 #define FUNC_NAME s_scm_gc
325 {
326 scm_i_gc ("call");
327 /* If you're calling scm_gc(), you probably want synchronous
328 finalization. */
329 GC_invoke_finalizers ();
330 return SCM_UNSPECIFIED;
331 }
332 #undef FUNC_NAME
333
334 void
335 scm_i_gc (const char *what)
336 {
337 GC_gcollect ();
338 }
339
340
341 \f
342 /* {GC Protection Helper Functions}
343 */
344
345
346 /*
347 * If within a function you need to protect one or more scheme objects from
348 * garbage collection, pass them as parameters to one of the
349 * scm_remember_upto_here* functions below. These functions don't do
350 * anything, but since the compiler does not know that they are actually
351 * no-ops, it will generate code that calls these functions with the given
352 * parameters. Therefore, you can be sure that the compiler will keep those
353 * scheme values alive (on the stack or in a register) up to the point where
354 * scm_remember_upto_here* is called. In other words, place the call to
355 * scm_remember_upto_here* _behind_ the last code in your function, that
356 * depends on the scheme object to exist.
357 *
358 * Example: We want to make sure that the string object str does not get
359 * garbage collected during the execution of 'some_function' in the code
360 * below, because otherwise the characters belonging to str would be freed and
361 * 'some_function' might access freed memory. To make sure that the compiler
362 * keeps str alive on the stack or in a register such that it is visible to
363 * the conservative gc we add the call to scm_remember_upto_here_1 _after_ the
364 * call to 'some_function'. Note that this would not be necessary if str was
365 * used anyway after the call to 'some_function'.
366 * char *chars = scm_i_string_chars (str);
367 * some_function (chars);
368 * scm_remember_upto_here_1 (str); // str will be alive up to this point.
369 */
370
371 /* Remove any macro versions of these while defining the functions.
372 Functions are always included in the library, for upward binary
373 compatibility and in case combinations of GCC and non-GCC are used. */
374 #undef scm_remember_upto_here_1
375 #undef scm_remember_upto_here_2
376
377 void
378 scm_remember_upto_here_1 (SCM obj SCM_UNUSED)
379 {
380 /* Empty. Protects a single object from garbage collection. */
381 }
382
383 void
384 scm_remember_upto_here_2 (SCM obj1 SCM_UNUSED, SCM obj2 SCM_UNUSED)
385 {
386 /* Empty. Protects two objects from garbage collection. */
387 }
388
389 void
390 scm_remember_upto_here (SCM obj SCM_UNUSED, ...)
391 {
392 /* Empty. Protects any number of objects from garbage collection. */
393 }
394
395 /*
396 These crazy functions prevent garbage collection
397 of arguments after the first argument by
398 ensuring they remain live throughout the
399 function because they are used in the last
400 line of the code block.
401 It'd be better to have a nice compiler hint to
402 aid the conservative stack-scanning GC. --03/09/00 gjb */
403 SCM
404 scm_return_first (SCM elt, ...)
405 {
406 return elt;
407 }
408
409 int
410 scm_return_first_int (int i, ...)
411 {
412 return i;
413 }
414
415
416 SCM
417 scm_permanent_object (SCM obj)
418 {
419 return (scm_gc_protect_object (obj));
420 }
421
422
423 /* Protect OBJ from the garbage collector. OBJ will not be freed, even if all
424 other references are dropped, until the object is unprotected by calling
425 scm_gc_unprotect_object (OBJ). Calls to scm_gc_protect/unprotect_object nest,
426 i. e. it is possible to protect the same object several times, but it is
427 necessary to unprotect the object the same number of times to actually get
428 the object unprotected. It is an error to unprotect an object more often
429 than it has been protected before. The function scm_protect_object returns
430 OBJ.
431 */
432
433 /* Implementation note: For every object X, there is a counter which
434 scm_gc_protect_object (X) increments and scm_gc_unprotect_object (X) decrements.
435 */
436
437
438
439 SCM
440 scm_gc_protect_object (SCM obj)
441 {
442 SCM handle;
443
444 /* This critical section barrier will be replaced by a mutex. */
445 /* njrev: Indeed; if my comment above is correct, there is the same
446 critsec/mutex inconsistency here. */
447 SCM_CRITICAL_SECTION_START;
448
449 handle = scm_hashq_create_handle_x (scm_protects, obj, scm_from_int (0));
450 SCM_SETCDR (handle, scm_sum (SCM_CDR (handle), scm_from_int (1)));
451
452 protected_obj_count ++;
453
454 SCM_CRITICAL_SECTION_END;
455
456 return obj;
457 }
458
459
460 /* Remove any protection for OBJ established by a prior call to
461 scm_protect_object. This function returns OBJ.
462
463 See scm_protect_object for more information. */
464 SCM
465 scm_gc_unprotect_object (SCM obj)
466 {
467 SCM handle;
468
469 /* This critical section barrier will be replaced by a mutex. */
470 /* njrev: and again. */
471 SCM_CRITICAL_SECTION_START;
472
473 if (scm_gc_running_p)
474 {
475 fprintf (stderr, "scm_unprotect_object called during GC.\n");
476 abort ();
477 }
478
479 handle = scm_hashq_get_handle (scm_protects, obj);
480
481 if (scm_is_false (handle))
482 {
483 fprintf (stderr, "scm_unprotect_object called on unprotected object\n");
484 abort ();
485 }
486 else
487 {
488 SCM count = scm_difference (SCM_CDR (handle), scm_from_int (1));
489 if (scm_is_eq (count, scm_from_int (0)))
490 scm_hashq_remove_x (scm_protects, obj);
491 else
492 SCM_SETCDR (handle, count);
493 }
494 protected_obj_count --;
495
496 SCM_CRITICAL_SECTION_END;
497
498 return obj;
499 }
500
501 void
502 scm_gc_register_root (SCM *p)
503 {
504 /* Nothing. */
505 }
506
507 void
508 scm_gc_unregister_root (SCM *p)
509 {
510 /* Nothing. */
511 }
512
513 void
514 scm_gc_register_roots (SCM *b, unsigned long n)
515 {
516 SCM *p = b;
517 for (; p < b + n; ++p)
518 scm_gc_register_root (p);
519 }
520
521 void
522 scm_gc_unregister_roots (SCM *b, unsigned long n)
523 {
524 SCM *p = b;
525 for (; p < b + n; ++p)
526 scm_gc_unregister_root (p);
527 }
528
529 \f
530
531
532 void
533 scm_storage_prehistory ()
534 {
535 GC_set_all_interior_pointers (0);
536
537 free_space_divisor = scm_getenv_int ("GC_FREE_SPACE_DIVISOR", 3);
538 minimum_free_space_divisor = free_space_divisor;
539 target_free_space_divisor = free_space_divisor;
540 GC_set_free_space_divisor (free_space_divisor);
541 GC_set_finalize_on_demand (1);
542
543 GC_INIT ();
544
545 GC_expand_hp (DEFAULT_INITIAL_HEAP_SIZE);
546
547 /* We only need to register a displacement for those types for which the
548 higher bits of the type tag are used to store a pointer (that is, a
549 pointer to an 8-octet aligned region). For `scm_tc3_struct', this is
550 handled in `scm_alloc_struct ()'. */
551 GC_REGISTER_DISPLACEMENT (scm_tc3_cons);
552 /* GC_REGISTER_DISPLACEMENT (scm_tc3_unused); */
553
554 /* Sanity check. */
555 if (!GC_is_visible (&scm_protects))
556 abort ();
557
558 scm_c_hook_init (&scm_before_gc_c_hook, 0, SCM_C_HOOK_NORMAL);
559 scm_c_hook_init (&scm_before_mark_c_hook, 0, SCM_C_HOOK_NORMAL);
560 scm_c_hook_init (&scm_before_sweep_c_hook, 0, SCM_C_HOOK_NORMAL);
561 scm_c_hook_init (&scm_after_sweep_c_hook, 0, SCM_C_HOOK_NORMAL);
562 scm_c_hook_init (&scm_after_gc_c_hook, 0, SCM_C_HOOK_NORMAL);
563 }
564
565 scm_i_pthread_mutex_t scm_i_gc_admin_mutex = SCM_I_PTHREAD_MUTEX_INITIALIZER;
566
567 void
568 scm_init_gc_protect_object ()
569 {
570 scm_protects = scm_c_make_hash_table (31);
571
572 #if 0
573 /* We can't have a cleanup handler since we have no thread to run it
574 in. */
575
576 #ifdef HAVE_ATEXIT
577 atexit (cleanup);
578 #else
579 #ifdef HAVE_ON_EXIT
580 on_exit (cleanup, 0);
581 #endif
582 #endif
583
584 #endif
585 }
586
587 \f
588
589 SCM scm_after_gc_hook;
590
591 static SCM after_gc_async_cell;
592
593 /* The function after_gc_async_thunk causes the execution of the
594 * after-gc-hook. It is run after the gc, as soon as the asynchronous
595 * events are handled by the evaluator.
596 */
597 static SCM
598 after_gc_async_thunk (void)
599 {
600 /* Fun, no? Hook-run *and* run-hook? */
601 scm_c_hook_run (&scm_after_gc_c_hook, NULL);
602 scm_c_run_hook (scm_after_gc_hook, SCM_EOL);
603 return SCM_UNSPECIFIED;
604 }
605
606
607 /* The function queue_after_gc_hook is run by the scm_before_gc_c_hook
608 * at the end of the garbage collection. The only purpose of this
609 * function is to mark the after_gc_async (which will eventually lead to
610 * the execution of the after_gc_async_thunk).
611 */
612 static void *
613 queue_after_gc_hook (void * hook_data SCM_UNUSED,
614 void *fn_data SCM_UNUSED,
615 void *data SCM_UNUSED)
616 {
617 /* If cell access debugging is enabled, the user may choose to perform
618 * additional garbage collections after an arbitrary number of cell
619 * accesses. We don't want the scheme level after-gc-hook to be performed
620 * for each of these garbage collections for the following reason: The
621 * execution of the after-gc-hook causes cell accesses itself. Thus, if the
622 * after-gc-hook was performed with every gc, and if the gc was performed
623 * after a very small number of cell accesses, then the number of cell
624 * accesses during the execution of the after-gc-hook will suffice to cause
625 * the execution of the next gc. Then, guile would keep executing the
626 * after-gc-hook over and over again, and would never come to do other
627 * things.
628 *
629 * To overcome this problem, if cell access debugging with additional
630 * garbage collections is enabled, the after-gc-hook is never run by the
631 * garbage collecter. When running guile with cell access debugging and the
632 * execution of the after-gc-hook is desired, then it is necessary to run
633 * the hook explicitly from the user code. This has the effect, that from
634 * the scheme level point of view it seems that garbage collection is
635 * performed with a much lower frequency than it actually is. Obviously,
636 * this will not work for code that depends on a fixed one to one
637 * relationship between the execution counts of the C level garbage
638 * collection hooks and the execution count of the scheme level
639 * after-gc-hook.
640 */
641
642 #if (SCM_DEBUG_CELL_ACCESSES == 1)
643 if (scm_debug_cells_gc_interval == 0)
644 #endif
645 {
646 scm_i_thread *t = SCM_I_CURRENT_THREAD;
647
648 if (scm_is_false (SCM_CDR (after_gc_async_cell)))
649 {
650 SCM_SETCDR (after_gc_async_cell, t->active_asyncs);
651 t->active_asyncs = after_gc_async_cell;
652 t->pending_asyncs = 1;
653 }
654 }
655
656 return NULL;
657 }
658
659 \f
660
661 static void *
662 start_gc_timer (void * hook_data SCM_UNUSED,
663 void *fn_data SCM_UNUSED,
664 void *data SCM_UNUSED)
665 {
666 if (!gc_start_time)
667 gc_start_time = scm_c_get_internal_run_time ();
668
669 return NULL;
670 }
671
672 static void *
673 accumulate_gc_timer (void * hook_data SCM_UNUSED,
674 void *fn_data SCM_UNUSED,
675 void *data SCM_UNUSED)
676 {
677 if (gc_start_time)
678 {
679 long now = scm_c_get_internal_run_time ();
680 gc_time_taken += now - gc_start_time;
681 gc_start_time = 0;
682 }
683
684 return NULL;
685 }
686
687 /* Return some idea of the memory footprint of a process, in bytes.
688 Currently only works on Linux systems. */
689 static size_t
690 get_image_size (void)
691 {
692 unsigned long size, resident, share;
693 size_t ret = 0;
694
695 FILE *fp = fopen ("/proc/self/statm", "r");
696
697 if (fp && fscanf (fp, "%lu %lu %lu", &size, &resident, &share) == 3)
698 ret = resident * 4096;
699
700 if (fp)
701 fclose (fp);
702
703 return ret;
704 }
705
706 /* These are discussed later. */
707 static size_t bytes_until_gc = DEFAULT_INITIAL_HEAP_SIZE;
708 static scm_i_pthread_mutex_t bytes_until_gc_lock = SCM_I_PTHREAD_MUTEX_INITIALIZER;
709
710 /* Make GC run more frequently when the process image size is growing,
711 measured against the number of bytes allocated through the GC.
712
713 If Guile is allocating at a GC-managed heap size H, libgc will tend
714 to limit the process image size to H*N. But if at the same time the
715 user program is mallocating at a rate M bytes per GC-allocated byte,
716 then the process stabilizes at H*N*M -- assuming that collecting data
717 will result in malloc'd data being freed. It doesn't take a very
718 large M for this to be a bad situation. To limit the image size,
719 Guile should GC more often -- the bigger the M, the more often.
720
721 Numeric functions that produce bigger and bigger integers are
722 pessimal, because M is an increasing function of time. Here is an
723 example of such a function:
724
725 (define (factorial n)
726 (define (fac n acc)
727 (if (<= n 1)
728 acc
729 (fac (1- n) (* n acc))))
730 (fac n 1))
731
732 It is possible for a process to grow for reasons that will not be
733 solved by faster GC. In that case M will be estimated as
734 artificially high for a while, and so GC will happen more often on
735 the Guile side. But when it stabilizes, Guile can ease back the GC
736 frequency.
737
738 The key is to measure process image growth, not mallocation rate.
739 For maximum effectiveness, Guile reacts quickly to process growth,
740 and exponentially backs down when the process stops growing.
741
742 See http://thread.gmane.org/gmane.lisp.guile.devel/12552/focus=12936
743 for further discussion.
744 */
745 static void *
746 adjust_gc_frequency (void * hook_data SCM_UNUSED,
747 void *fn_data SCM_UNUSED,
748 void *data SCM_UNUSED)
749 {
750 static size_t prev_image_size = 0;
751 static size_t prev_bytes_alloced = 0;
752 size_t image_size;
753 size_t bytes_alloced;
754
755 scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
756 bytes_until_gc = GC_get_heap_size ();
757 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
758
759 image_size = get_image_size ();
760 bytes_alloced = GC_get_total_bytes ();
761
762 #define HEURISTICS_DEBUG 0
763
764 #if HEURISTICS_DEBUG
765 fprintf (stderr, "prev image / alloced: %lu / %lu\n", prev_image_size, prev_bytes_alloced);
766 fprintf (stderr, " image / alloced: %lu / %lu\n", image_size, bytes_alloced);
767 fprintf (stderr, "divisor %lu / %f\n", free_space_divisor, target_free_space_divisor);
768 #endif
769
770 if (prev_image_size && bytes_alloced != prev_bytes_alloced)
771 {
772 double growth_rate, new_target_free_space_divisor;
773 double decay_factor = 0.5;
774 double hysteresis = 0.1;
775
776 growth_rate = ((double) image_size - prev_image_size)
777 / ((double)bytes_alloced - prev_bytes_alloced);
778
779 #if HEURISTICS_DEBUG
780 fprintf (stderr, "growth rate %f\n", growth_rate);
781 #endif
782
783 new_target_free_space_divisor = minimum_free_space_divisor;
784
785 if (growth_rate > 0)
786 new_target_free_space_divisor *= 1.0 + growth_rate;
787
788 #if HEURISTICS_DEBUG
789 fprintf (stderr, "new divisor %f\n", new_target_free_space_divisor);
790 #endif
791
792 if (new_target_free_space_divisor < target_free_space_divisor)
793 /* Decay down. */
794 target_free_space_divisor =
795 (decay_factor * target_free_space_divisor
796 + (1.0 - decay_factor) * new_target_free_space_divisor);
797 else
798 /* Jump up. */
799 target_free_space_divisor = new_target_free_space_divisor;
800
801 #if HEURISTICS_DEBUG
802 fprintf (stderr, "new target divisor %f\n", target_free_space_divisor);
803 #endif
804
805 if (free_space_divisor + 0.5 + hysteresis < target_free_space_divisor
806 || free_space_divisor - 0.5 - hysteresis > target_free_space_divisor)
807 {
808 free_space_divisor = lround (target_free_space_divisor);
809 #if HEURISTICS_DEBUG
810 fprintf (stderr, "new divisor %lu\n", free_space_divisor);
811 #endif
812 GC_set_free_space_divisor (free_space_divisor);
813 }
814 }
815
816 prev_image_size = image_size;
817 prev_bytes_alloced = bytes_alloced;
818
819 return NULL;
820 }
821
822 /* The adjust_gc_frequency routine handles transients in the process
823 image size. It can't handle instense non-GC-managed steady-state
824 allocation though, as it decays the FSD at steady-state down to its
825 minimum value.
826
827 The only real way to handle continuous, high non-GC allocation is to
828 let the GC know about it. This routine can handle non-GC allocation
829 rates that are similar in size to the GC-managed heap size.
830 */
831
832 void
833 scm_gc_register_allocation (size_t size)
834 {
835 scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
836 if (bytes_until_gc - size > bytes_until_gc)
837 {
838 bytes_until_gc = GC_get_heap_size ();
839 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
840 GC_gcollect ();
841 }
842 else
843 {
844 bytes_until_gc -= size;
845 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
846 }
847 }
848
849
850 \f
851 void
852 scm_init_gc ()
853 {
854 /* `GC_INIT ()' was invoked in `scm_storage_prehistory ()'. */
855
856 scm_after_gc_hook = scm_make_hook (SCM_INUM0);
857 scm_c_define ("after-gc-hook", scm_after_gc_hook);
858
859 /* When the async is to run, the cdr of the gc_async pair gets set to
860 the asyncs queue of the current thread. */
861 after_gc_async_cell = scm_cons (scm_c_make_gsubr ("%after-gc-thunk", 0, 0, 0,
862 after_gc_async_thunk),
863 SCM_BOOL_F);
864
865 scm_c_hook_add (&scm_before_gc_c_hook, queue_after_gc_hook, NULL, 0);
866 scm_c_hook_add (&scm_before_gc_c_hook, start_gc_timer, NULL, 0);
867 scm_c_hook_add (&scm_after_gc_c_hook, accumulate_gc_timer, NULL, 0);
868
869 /* GC_get_heap_usage does not take a lock, and so can run in the GC
870 start hook. */
871 scm_c_hook_add (&scm_before_gc_c_hook, adjust_gc_frequency, NULL, 0);
872
873 GC_set_start_callback (run_before_gc_c_hook);
874
875 #include "libguile/gc.x"
876 }
877
878
879 void
880 scm_gc_sweep (void)
881 #define FUNC_NAME "scm_gc_sweep"
882 {
883 /* FIXME */
884 fprintf (stderr, "%s: doing nothing\n", FUNC_NAME);
885 }
886 #undef FUNC_NAME
887
888 /*
889 Local Variables:
890 c-file-style: "gnu"
891 End:
892 */