Remove tests and shims for pre-7.2 bdw-gc.
[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/strings.h"
50 #include "libguile/vectors.h"
51 #include "libguile/hashtab.h"
52 #include "libguile/tags.h"
53
54 #include "libguile/private-gc.h"
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 /* Set this to != 0 if every cell that is accessed shall be checked:
74 */
75 int scm_debug_cell_accesses_p = 0;
76 int scm_expensive_debug_cell_accesses_p = 0;
77
78 /* Set this to 0 if no additional gc's shall be performed, otherwise set it to
79 * the number of cell accesses after which a gc shall be called.
80 */
81 int scm_debug_cells_gc_interval = 0;
82
83 /* Hash table that keeps a reference to objects the user wants to protect from
84 garbage collection. */
85 static SCM scm_protects;
86
87
88 #if (SCM_DEBUG_CELL_ACCESSES == 1)
89
90
91 /*
92
93 Assert that the given object is a valid reference to a valid cell. This
94 test involves to determine whether the object is a cell pointer, whether
95 this pointer actually points into a heap segment and whether the cell
96 pointed to is not a free cell. Further, additional garbage collections may
97 get executed after a user defined number of cell accesses. This helps to
98 find places in the C code where references are dropped for extremely short
99 periods.
100
101 */
102 void
103 scm_i_expensive_validation_check (SCM cell)
104 {
105 /* If desired, perform additional garbage collections after a user
106 * defined number of cell accesses.
107 */
108 if (scm_debug_cells_gc_interval)
109 {
110 static unsigned int counter = 0;
111
112 if (counter != 0)
113 {
114 --counter;
115 }
116 else
117 {
118 counter = scm_debug_cells_gc_interval;
119 scm_gc ();
120 }
121 }
122 }
123
124 /* Whether cell validation is already running. */
125 static int scm_i_cell_validation_already_running = 0;
126
127 void
128 scm_assert_cell_valid (SCM cell)
129 {
130 if (!scm_i_cell_validation_already_running && scm_debug_cell_accesses_p)
131 {
132 scm_i_cell_validation_already_running = 1; /* set to avoid recursion */
133
134 /*
135 During GC, no user-code should be run, and the guile core
136 should use non-protected accessors.
137 */
138 if (scm_gc_running_p)
139 return;
140
141 /*
142 Only scm_in_heap_p and rescanning the heap is wildly
143 expensive.
144 */
145 if (scm_expensive_debug_cell_accesses_p)
146 scm_i_expensive_validation_check (cell);
147
148 scm_i_cell_validation_already_running = 0; /* re-enable */
149 }
150 }
151
152
153
154 SCM_DEFINE (scm_set_debug_cell_accesses_x, "set-debug-cell-accesses!", 1, 0, 0,
155 (SCM flag),
156 "If @var{flag} is @code{#f}, cell access checking is disabled.\n"
157 "If @var{flag} is @code{#t}, cheap cell access checking is enabled,\n"
158 "but no additional calls to garbage collection are issued.\n"
159 "If @var{flag} is a number, strict cell access checking is enabled,\n"
160 "with an additional garbage collection after the given\n"
161 "number of cell accesses.\n"
162 "This procedure only exists when the compile-time flag\n"
163 "@code{SCM_DEBUG_CELL_ACCESSES} was set to 1.")
164 #define FUNC_NAME s_scm_set_debug_cell_accesses_x
165 {
166 if (scm_is_false (flag))
167 {
168 scm_debug_cell_accesses_p = 0;
169 }
170 else if (scm_is_eq (flag, SCM_BOOL_T))
171 {
172 scm_debug_cells_gc_interval = 0;
173 scm_debug_cell_accesses_p = 1;
174 scm_expensive_debug_cell_accesses_p = 0;
175 }
176 else
177 {
178 scm_debug_cells_gc_interval = scm_to_signed_integer (flag, 0, INT_MAX);
179 scm_debug_cell_accesses_p = 1;
180 scm_expensive_debug_cell_accesses_p = 1;
181 }
182 return SCM_UNSPECIFIED;
183 }
184 #undef FUNC_NAME
185
186
187 #endif /* SCM_DEBUG_CELL_ACCESSES == 1 */
188
189 \f
190
191 /* Hooks. */
192 scm_t_c_hook scm_before_gc_c_hook;
193 scm_t_c_hook scm_before_mark_c_hook;
194 scm_t_c_hook scm_before_sweep_c_hook;
195 scm_t_c_hook scm_after_sweep_c_hook;
196 scm_t_c_hook scm_after_gc_c_hook;
197
198
199 static void
200 run_before_gc_c_hook (void)
201 {
202 if (!SCM_I_CURRENT_THREAD)
203 /* GC while a thread is spinning up; punt. */
204 return;
205
206 scm_c_hook_run (&scm_before_gc_c_hook, NULL);
207 }
208
209
210 /* GC Statistics Keeping
211 */
212 unsigned long scm_gc_ports_collected = 0;
213 static long gc_time_taken = 0;
214 static long gc_start_time = 0;
215
216 static unsigned long free_space_divisor;
217 static unsigned long minimum_free_space_divisor;
218 static double target_free_space_divisor;
219
220 static unsigned long protected_obj_count = 0;
221
222
223 SCM_SYMBOL (sym_gc_time_taken, "gc-time-taken");
224 SCM_SYMBOL (sym_heap_size, "heap-size");
225 SCM_SYMBOL (sym_heap_free_size, "heap-free-size");
226 SCM_SYMBOL (sym_heap_total_allocated, "heap-total-allocated");
227 SCM_SYMBOL (sym_heap_allocated_since_gc, "heap-allocated-since-gc");
228 SCM_SYMBOL (sym_protected_objects, "protected-objects");
229 SCM_SYMBOL (sym_times, "gc-times");
230
231
232 /* {Scheme Interface to GC}
233 */
234 static SCM
235 tag_table_to_type_alist (void *closure, SCM key, SCM val, SCM acc)
236 {
237 if (scm_is_integer (key))
238 {
239 int c_tag = scm_to_int (key);
240
241 char const * name = scm_i_tag_name (c_tag);
242 if (name != NULL)
243 {
244 key = scm_from_locale_string (name);
245 }
246 else
247 {
248 char s[100];
249 sprintf (s, "tag %d", c_tag);
250 key = scm_from_locale_string (s);
251 }
252 }
253
254 return scm_cons (scm_cons (key, val), acc);
255 }
256
257 SCM_DEFINE (scm_gc_live_object_stats, "gc-live-object-stats", 0, 0, 0,
258 (),
259 "Return an alist of statistics of the current live objects. ")
260 #define FUNC_NAME s_scm_gc_live_object_stats
261 {
262 SCM tab = scm_make_hash_table (scm_from_int (57));
263 SCM alist;
264
265 alist
266 = scm_internal_hash_fold (&tag_table_to_type_alist, NULL, SCM_EOL, tab);
267
268 return alist;
269 }
270 #undef FUNC_NAME
271
272 extern int scm_gc_malloc_yield_percentage;
273 SCM_DEFINE (scm_gc_stats, "gc-stats", 0, 0, 0,
274 (),
275 "Return an association list of statistics about Guile's current\n"
276 "use of storage.\n")
277 #define FUNC_NAME s_scm_gc_stats
278 {
279 SCM answer;
280 GC_word heap_size, free_bytes, unmapped_bytes, bytes_since_gc, total_bytes;
281 size_t gc_times;
282
283 GC_get_heap_usage_safe (&heap_size, &free_bytes, &unmapped_bytes,
284 &bytes_since_gc, &total_bytes);
285 gc_times = GC_get_gc_no ();
286
287 answer =
288 scm_list_n (scm_cons (sym_gc_time_taken, scm_from_long (gc_time_taken)),
289 scm_cons (sym_heap_size, scm_from_size_t (heap_size)),
290 scm_cons (sym_heap_free_size, scm_from_size_t (free_bytes)),
291 scm_cons (sym_heap_total_allocated,
292 scm_from_size_t (total_bytes)),
293 scm_cons (sym_heap_allocated_since_gc,
294 scm_from_size_t (bytes_since_gc)),
295 scm_cons (sym_protected_objects,
296 scm_from_ulong (protected_obj_count)),
297 scm_cons (sym_times, scm_from_size_t (gc_times)),
298 SCM_UNDEFINED);
299
300 return answer;
301 }
302 #undef FUNC_NAME
303
304
305 SCM_DEFINE (scm_gc_dump, "gc-dump", 0, 0, 0,
306 (void),
307 "Dump information about the garbage collector's internal data "
308 "structures and memory usage to the standard output.")
309 #define FUNC_NAME s_scm_gc_dump
310 {
311 GC_dump ();
312
313 return SCM_UNSPECIFIED;
314 }
315 #undef FUNC_NAME
316
317
318 SCM_DEFINE (scm_object_address, "object-address", 1, 0, 0,
319 (SCM obj),
320 "Return an integer that for the lifetime of @var{obj} is uniquely\n"
321 "returned by this function for @var{obj}")
322 #define FUNC_NAME s_scm_object_address
323 {
324 return scm_from_ulong (SCM_UNPACK (obj));
325 }
326 #undef FUNC_NAME
327
328
329 SCM_DEFINE (scm_gc_disable, "gc-disable", 0, 0, 0,
330 (),
331 "Disables the garbage collector. Nested calls are permitted. "
332 "GC is re-enabled once @code{gc-enable} has been called the "
333 "same number of times @code{gc-disable} was called.")
334 #define FUNC_NAME s_scm_gc_disable
335 {
336 GC_disable ();
337 return SCM_UNSPECIFIED;
338 }
339 #undef FUNC_NAME
340
341 SCM_DEFINE (scm_gc_enable, "gc-enable", 0, 0, 0,
342 (),
343 "Enables the garbage collector.")
344 #define FUNC_NAME s_scm_gc_enable
345 {
346 GC_enable ();
347 return SCM_UNSPECIFIED;
348 }
349 #undef FUNC_NAME
350
351
352 SCM_DEFINE (scm_gc, "gc", 0, 0, 0,
353 (),
354 "Scans all of SCM objects and reclaims for further use those that are\n"
355 "no longer accessible.")
356 #define FUNC_NAME s_scm_gc
357 {
358 scm_i_gc ("call");
359 /* If you're calling scm_gc(), you probably want synchronous
360 finalization. */
361 GC_invoke_finalizers ();
362 return SCM_UNSPECIFIED;
363 }
364 #undef FUNC_NAME
365
366 void
367 scm_i_gc (const char *what)
368 {
369 GC_gcollect ();
370 }
371
372
373 \f
374 /* {GC Protection Helper Functions}
375 */
376
377
378 /*
379 * If within a function you need to protect one or more scheme objects from
380 * garbage collection, pass them as parameters to one of the
381 * scm_remember_upto_here* functions below. These functions don't do
382 * anything, but since the compiler does not know that they are actually
383 * no-ops, it will generate code that calls these functions with the given
384 * parameters. Therefore, you can be sure that the compiler will keep those
385 * scheme values alive (on the stack or in a register) up to the point where
386 * scm_remember_upto_here* is called. In other words, place the call to
387 * scm_remember_upto_here* _behind_ the last code in your function, that
388 * depends on the scheme object to exist.
389 *
390 * Example: We want to make sure that the string object str does not get
391 * garbage collected during the execution of 'some_function' in the code
392 * below, because otherwise the characters belonging to str would be freed and
393 * 'some_function' might access freed memory. To make sure that the compiler
394 * keeps str alive on the stack or in a register such that it is visible to
395 * the conservative gc we add the call to scm_remember_upto_here_1 _after_ the
396 * call to 'some_function'. Note that this would not be necessary if str was
397 * used anyway after the call to 'some_function'.
398 * char *chars = scm_i_string_chars (str);
399 * some_function (chars);
400 * scm_remember_upto_here_1 (str); // str will be alive up to this point.
401 */
402
403 /* Remove any macro versions of these while defining the functions.
404 Functions are always included in the library, for upward binary
405 compatibility and in case combinations of GCC and non-GCC are used. */
406 #undef scm_remember_upto_here_1
407 #undef scm_remember_upto_here_2
408
409 void
410 scm_remember_upto_here_1 (SCM obj SCM_UNUSED)
411 {
412 /* Empty. Protects a single object from garbage collection. */
413 }
414
415 void
416 scm_remember_upto_here_2 (SCM obj1 SCM_UNUSED, SCM obj2 SCM_UNUSED)
417 {
418 /* Empty. Protects two objects from garbage collection. */
419 }
420
421 void
422 scm_remember_upto_here (SCM obj SCM_UNUSED, ...)
423 {
424 /* Empty. Protects any number of objects from garbage collection. */
425 }
426
427 /*
428 These crazy functions prevent garbage collection
429 of arguments after the first argument by
430 ensuring they remain live throughout the
431 function because they are used in the last
432 line of the code block.
433 It'd be better to have a nice compiler hint to
434 aid the conservative stack-scanning GC. --03/09/00 gjb */
435 SCM
436 scm_return_first (SCM elt, ...)
437 {
438 return elt;
439 }
440
441 int
442 scm_return_first_int (int i, ...)
443 {
444 return i;
445 }
446
447
448 SCM
449 scm_permanent_object (SCM obj)
450 {
451 return (scm_gc_protect_object (obj));
452 }
453
454
455 /* Protect OBJ from the garbage collector. OBJ will not be freed, even if all
456 other references are dropped, until the object is unprotected by calling
457 scm_gc_unprotect_object (OBJ). Calls to scm_gc_protect/unprotect_object nest,
458 i. e. it is possible to protect the same object several times, but it is
459 necessary to unprotect the object the same number of times to actually get
460 the object unprotected. It is an error to unprotect an object more often
461 than it has been protected before. The function scm_protect_object returns
462 OBJ.
463 */
464
465 /* Implementation note: For every object X, there is a counter which
466 scm_gc_protect_object (X) increments and scm_gc_unprotect_object (X) decrements.
467 */
468
469
470
471 SCM
472 scm_gc_protect_object (SCM obj)
473 {
474 SCM handle;
475
476 /* This critical section barrier will be replaced by a mutex. */
477 /* njrev: Indeed; if my comment above is correct, there is the same
478 critsec/mutex inconsistency here. */
479 SCM_CRITICAL_SECTION_START;
480
481 handle = scm_hashq_create_handle_x (scm_protects, obj, scm_from_int (0));
482 SCM_SETCDR (handle, scm_sum (SCM_CDR (handle), scm_from_int (1)));
483
484 protected_obj_count ++;
485
486 SCM_CRITICAL_SECTION_END;
487
488 return obj;
489 }
490
491
492 /* Remove any protection for OBJ established by a prior call to
493 scm_protect_object. This function returns OBJ.
494
495 See scm_protect_object for more information. */
496 SCM
497 scm_gc_unprotect_object (SCM obj)
498 {
499 SCM handle;
500
501 /* This critical section barrier will be replaced by a mutex. */
502 /* njrev: and again. */
503 SCM_CRITICAL_SECTION_START;
504
505 if (scm_gc_running_p)
506 {
507 fprintf (stderr, "scm_unprotect_object called during GC.\n");
508 abort ();
509 }
510
511 handle = scm_hashq_get_handle (scm_protects, obj);
512
513 if (scm_is_false (handle))
514 {
515 fprintf (stderr, "scm_unprotect_object called on unprotected object\n");
516 abort ();
517 }
518 else
519 {
520 SCM count = scm_difference (SCM_CDR (handle), scm_from_int (1));
521 if (scm_is_eq (count, scm_from_int (0)))
522 scm_hashq_remove_x (scm_protects, obj);
523 else
524 SCM_SETCDR (handle, count);
525 }
526 protected_obj_count --;
527
528 SCM_CRITICAL_SECTION_END;
529
530 return obj;
531 }
532
533 void
534 scm_gc_register_root (SCM *p)
535 {
536 /* Nothing. */
537 }
538
539 void
540 scm_gc_unregister_root (SCM *p)
541 {
542 /* Nothing. */
543 }
544
545 void
546 scm_gc_register_roots (SCM *b, unsigned long n)
547 {
548 SCM *p = b;
549 for (; p < b + n; ++p)
550 scm_gc_register_root (p);
551 }
552
553 void
554 scm_gc_unregister_roots (SCM *b, unsigned long n)
555 {
556 SCM *p = b;
557 for (; p < b + n; ++p)
558 scm_gc_unregister_root (p);
559 }
560
561 \f
562
563
564 /*
565 MOVE THIS FUNCTION. IT DOES NOT HAVE ANYTHING TODO WITH GC.
566 */
567
568 /* Get an integer from an environment variable. */
569 int
570 scm_getenv_int (const char *var, int def)
571 {
572 char *end = 0;
573 char *val = getenv (var);
574 long res = def;
575 if (!val)
576 return def;
577 res = strtol (val, &end, 10);
578 if (end == val)
579 return def;
580 return res;
581 }
582
583 void
584 scm_storage_prehistory ()
585 {
586 GC_set_all_interior_pointers (0);
587
588 free_space_divisor = scm_getenv_int ("GC_FREE_SPACE_DIVISOR", 3);
589 minimum_free_space_divisor = free_space_divisor;
590 target_free_space_divisor = free_space_divisor;
591 GC_set_free_space_divisor (free_space_divisor);
592 GC_set_finalize_on_demand (1);
593
594 GC_INIT ();
595
596 #if (! ((defined GC_VERSION_MAJOR) && (GC_VERSION_MAJOR >= 7))) \
597 && (defined SCM_I_GSC_USE_PTHREAD_THREADS)
598 /* When using GC 6.8, this call is required to initialize thread-local
599 freelists (shouldn't be necessary with GC 7.0). */
600 GC_init ();
601 #endif
602
603 GC_expand_hp (SCM_DEFAULT_INIT_HEAP_SIZE_2);
604
605 /* We only need to register a displacement for those types for which the
606 higher bits of the type tag are used to store a pointer (that is, a
607 pointer to an 8-octet aligned region). For `scm_tc3_struct', this is
608 handled in `scm_alloc_struct ()'. */
609 GC_REGISTER_DISPLACEMENT (scm_tc3_cons);
610 /* GC_REGISTER_DISPLACEMENT (scm_tc3_unused); */
611
612 /* Sanity check. */
613 if (!GC_is_visible (&scm_protects))
614 abort ();
615
616 scm_c_hook_init (&scm_before_gc_c_hook, 0, SCM_C_HOOK_NORMAL);
617 scm_c_hook_init (&scm_before_mark_c_hook, 0, SCM_C_HOOK_NORMAL);
618 scm_c_hook_init (&scm_before_sweep_c_hook, 0, SCM_C_HOOK_NORMAL);
619 scm_c_hook_init (&scm_after_sweep_c_hook, 0, SCM_C_HOOK_NORMAL);
620 scm_c_hook_init (&scm_after_gc_c_hook, 0, SCM_C_HOOK_NORMAL);
621 }
622
623 scm_i_pthread_mutex_t scm_i_gc_admin_mutex = SCM_I_PTHREAD_MUTEX_INITIALIZER;
624
625 void
626 scm_init_gc_protect_object ()
627 {
628 scm_protects = scm_c_make_hash_table (31);
629
630 #if 0
631 /* We can't have a cleanup handler since we have no thread to run it
632 in. */
633
634 #ifdef HAVE_ATEXIT
635 atexit (cleanup);
636 #else
637 #ifdef HAVE_ON_EXIT
638 on_exit (cleanup, 0);
639 #endif
640 #endif
641
642 #endif
643 }
644
645 \f
646
647 SCM scm_after_gc_hook;
648
649 static SCM after_gc_async_cell;
650
651 /* The function after_gc_async_thunk causes the execution of the
652 * after-gc-hook. It is run after the gc, as soon as the asynchronous
653 * events are handled by the evaluator.
654 */
655 static SCM
656 after_gc_async_thunk (void)
657 {
658 /* Fun, no? Hook-run *and* run-hook? */
659 scm_c_hook_run (&scm_after_gc_c_hook, NULL);
660 scm_c_run_hook (scm_after_gc_hook, SCM_EOL);
661 return SCM_UNSPECIFIED;
662 }
663
664
665 /* The function queue_after_gc_hook is run by the scm_before_gc_c_hook
666 * at the end of the garbage collection. The only purpose of this
667 * function is to mark the after_gc_async (which will eventually lead to
668 * the execution of the after_gc_async_thunk).
669 */
670 static void *
671 queue_after_gc_hook (void * hook_data SCM_UNUSED,
672 void *fn_data SCM_UNUSED,
673 void *data SCM_UNUSED)
674 {
675 /* If cell access debugging is enabled, the user may choose to perform
676 * additional garbage collections after an arbitrary number of cell
677 * accesses. We don't want the scheme level after-gc-hook to be performed
678 * for each of these garbage collections for the following reason: The
679 * execution of the after-gc-hook causes cell accesses itself. Thus, if the
680 * after-gc-hook was performed with every gc, and if the gc was performed
681 * after a very small number of cell accesses, then the number of cell
682 * accesses during the execution of the after-gc-hook will suffice to cause
683 * the execution of the next gc. Then, guile would keep executing the
684 * after-gc-hook over and over again, and would never come to do other
685 * things.
686 *
687 * To overcome this problem, if cell access debugging with additional
688 * garbage collections is enabled, the after-gc-hook is never run by the
689 * garbage collecter. When running guile with cell access debugging and the
690 * execution of the after-gc-hook is desired, then it is necessary to run
691 * the hook explicitly from the user code. This has the effect, that from
692 * the scheme level point of view it seems that garbage collection is
693 * performed with a much lower frequency than it actually is. Obviously,
694 * this will not work for code that depends on a fixed one to one
695 * relationship between the execution counts of the C level garbage
696 * collection hooks and the execution count of the scheme level
697 * after-gc-hook.
698 */
699
700 #if (SCM_DEBUG_CELL_ACCESSES == 1)
701 if (scm_debug_cells_gc_interval == 0)
702 #endif
703 {
704 scm_i_thread *t = SCM_I_CURRENT_THREAD;
705
706 if (scm_is_false (SCM_CDR (after_gc_async_cell)))
707 {
708 SCM_SETCDR (after_gc_async_cell, t->active_asyncs);
709 t->active_asyncs = after_gc_async_cell;
710 t->pending_asyncs = 1;
711 }
712 }
713
714 return NULL;
715 }
716
717 \f
718
719 static void *
720 start_gc_timer (void * hook_data SCM_UNUSED,
721 void *fn_data SCM_UNUSED,
722 void *data SCM_UNUSED)
723 {
724 if (!gc_start_time)
725 gc_start_time = scm_c_get_internal_run_time ();
726
727 return NULL;
728 }
729
730 static void *
731 accumulate_gc_timer (void * hook_data SCM_UNUSED,
732 void *fn_data SCM_UNUSED,
733 void *data SCM_UNUSED)
734 {
735 if (gc_start_time)
736 {
737 long now = scm_c_get_internal_run_time ();
738 gc_time_taken += now - gc_start_time;
739 gc_start_time = 0;
740 }
741
742 return NULL;
743 }
744
745 /* Return some idea of the memory footprint of a process, in bytes.
746 Currently only works on Linux systems. */
747 static size_t
748 get_image_size (void)
749 {
750 unsigned long size, resident, share;
751 size_t ret = 0;
752
753 FILE *fp = fopen ("/proc/self/statm", "r");
754
755 if (fp && fscanf (fp, "%lu %lu %lu", &size, &resident, &share) == 3)
756 ret = resident * 4096;
757
758 if (fp)
759 fclose (fp);
760
761 return ret;
762 }
763
764 /* These are discussed later. */
765 static size_t bytes_until_gc;
766 static scm_i_pthread_mutex_t bytes_until_gc_lock = SCM_I_PTHREAD_MUTEX_INITIALIZER;
767
768 /* Make GC run more frequently when the process image size is growing,
769 measured against the number of bytes allocated through the GC.
770
771 If Guile is allocating at a GC-managed heap size H, libgc will tend
772 to limit the process image size to H*N. But if at the same time the
773 user program is mallocating at a rate M bytes per GC-allocated byte,
774 then the process stabilizes at H*N*M -- assuming that collecting data
775 will result in malloc'd data being freed. It doesn't take a very
776 large M for this to be a bad situation. To limit the image size,
777 Guile should GC more often -- the bigger the M, the more often.
778
779 Numeric functions that produce bigger and bigger integers are
780 pessimal, because M is an increasing function of time. Here is an
781 example of such a function:
782
783 (define (factorial n)
784 (define (fac n acc)
785 (if (<= n 1)
786 acc
787 (fac (1- n) (* n acc))))
788 (fac n 1))
789
790 It is possible for a process to grow for reasons that will not be
791 solved by faster GC. In that case M will be estimated as
792 artificially high for a while, and so GC will happen more often on
793 the Guile side. But when it stabilizes, Guile can ease back the GC
794 frequency.
795
796 The key is to measure process image growth, not mallocation rate.
797 For maximum effectiveness, Guile reacts quickly to process growth,
798 and exponentially backs down when the process stops growing.
799
800 See http://thread.gmane.org/gmane.lisp.guile.devel/12552/focus=12936
801 for further discussion.
802 */
803 static void *
804 adjust_gc_frequency (void * hook_data SCM_UNUSED,
805 void *fn_data SCM_UNUSED,
806 void *data SCM_UNUSED)
807 {
808 static size_t prev_image_size = 0;
809 static size_t prev_bytes_alloced = 0;
810 size_t image_size;
811 size_t bytes_alloced;
812
813 scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
814 bytes_until_gc = GC_get_heap_size ();
815 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
816
817 image_size = get_image_size ();
818 bytes_alloced = GC_get_total_bytes ();
819
820 #define HEURISTICS_DEBUG 0
821
822 #if HEURISTICS_DEBUG
823 fprintf (stderr, "prev image / alloced: %lu / %lu\n", prev_image_size, prev_bytes_alloced);
824 fprintf (stderr, " image / alloced: %lu / %lu\n", image_size, bytes_alloced);
825 fprintf (stderr, "divisor %lu / %f\n", free_space_divisor, target_free_space_divisor);
826 #endif
827
828 if (prev_image_size && bytes_alloced != prev_bytes_alloced)
829 {
830 double growth_rate, new_target_free_space_divisor;
831 double decay_factor = 0.5;
832 double hysteresis = 0.1;
833
834 growth_rate = ((double) image_size - prev_image_size)
835 / ((double)bytes_alloced - prev_bytes_alloced);
836
837 #if HEURISTICS_DEBUG
838 fprintf (stderr, "growth rate %f\n", growth_rate);
839 #endif
840
841 new_target_free_space_divisor = minimum_free_space_divisor;
842
843 if (growth_rate > 0)
844 new_target_free_space_divisor *= 1.0 + growth_rate;
845
846 #if HEURISTICS_DEBUG
847 fprintf (stderr, "new divisor %f\n", new_target_free_space_divisor);
848 #endif
849
850 if (new_target_free_space_divisor < target_free_space_divisor)
851 /* Decay down. */
852 target_free_space_divisor =
853 (decay_factor * target_free_space_divisor
854 + (1.0 - decay_factor) * new_target_free_space_divisor);
855 else
856 /* Jump up. */
857 target_free_space_divisor = new_target_free_space_divisor;
858
859 #if HEURISTICS_DEBUG
860 fprintf (stderr, "new target divisor %f\n", target_free_space_divisor);
861 #endif
862
863 if (free_space_divisor + 0.5 + hysteresis < target_free_space_divisor
864 || free_space_divisor - 0.5 - hysteresis > target_free_space_divisor)
865 {
866 free_space_divisor = lround (target_free_space_divisor);
867 #if HEURISTICS_DEBUG
868 fprintf (stderr, "new divisor %lu\n", free_space_divisor);
869 #endif
870 GC_set_free_space_divisor (free_space_divisor);
871 }
872 }
873
874 prev_image_size = image_size;
875 prev_bytes_alloced = bytes_alloced;
876
877 return NULL;
878 }
879
880 /* The adjust_gc_frequency routine handles transients in the process
881 image size. It can't handle instense non-GC-managed steady-state
882 allocation though, as it decays the FSD at steady-state down to its
883 minimum value.
884
885 The only real way to handle continuous, high non-GC allocation is to
886 let the GC know about it. This routine can handle non-GC allocation
887 rates that are similar in size to the GC-managed heap size.
888 */
889
890 void
891 scm_gc_register_allocation (size_t size)
892 {
893 scm_i_pthread_mutex_lock (&bytes_until_gc_lock);
894 if (bytes_until_gc - size > bytes_until_gc)
895 {
896 bytes_until_gc = GC_get_heap_size ();
897 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
898 GC_gcollect ();
899 }
900 else
901 {
902 bytes_until_gc -= size;
903 scm_i_pthread_mutex_unlock (&bytes_until_gc_lock);
904 }
905 }
906
907
908 \f
909
910 char const *
911 scm_i_tag_name (scm_t_bits tag)
912 {
913 switch (tag & 0x7f) /* 7 bits */
914 {
915 case scm_tcs_struct:
916 return "struct";
917 case scm_tcs_cons_imcar:
918 return "cons (immediate car)";
919 case scm_tcs_cons_nimcar:
920 return "cons (non-immediate car)";
921 case scm_tc7_pointer:
922 return "foreign";
923 case scm_tc7_hashtable:
924 return "hashtable";
925 case scm_tc7_weak_set:
926 return "weak-set";
927 case scm_tc7_weak_table:
928 return "weak-table";
929 case scm_tc7_fluid:
930 return "fluid";
931 case scm_tc7_dynamic_state:
932 return "dynamic state";
933 case scm_tc7_frame:
934 return "frame";
935 case scm_tc7_vm_cont:
936 return "vm continuation";
937 case scm_tc7_wvect:
938 return "weak vector";
939 case scm_tc7_vector:
940 return "vector";
941 case scm_tc7_number:
942 switch (tag)
943 {
944 case scm_tc16_real:
945 return "real";
946 break;
947 case scm_tc16_big:
948 return "bignum";
949 break;
950 case scm_tc16_complex:
951 return "complex number";
952 break;
953 case scm_tc16_fraction:
954 return "fraction";
955 break;
956 }
957 break;
958 case scm_tc7_string:
959 return "string";
960 break;
961 case scm_tc7_stringbuf:
962 return "string buffer";
963 break;
964 case scm_tc7_symbol:
965 return "symbol";
966 break;
967 case scm_tc7_variable:
968 return "variable";
969 break;
970 case scm_tc7_port:
971 return "port";
972 break;
973 case scm_tc7_smob:
974 {
975 int k = 0xff & (tag >> 8);
976 return (scm_smobs[k].name);
977 }
978 break;
979 }
980
981 return NULL;
982 }
983
984
985
986 \f
987 void
988 scm_init_gc ()
989 {
990 /* `GC_INIT ()' was invoked in `scm_storage_prehistory ()'. */
991
992 scm_after_gc_hook = scm_make_hook (SCM_INUM0);
993 scm_c_define ("after-gc-hook", scm_after_gc_hook);
994
995 /* When the async is to run, the cdr of the gc_async pair gets set to
996 the asyncs queue of the current thread. */
997 after_gc_async_cell = scm_cons (scm_c_make_gsubr ("%after-gc-thunk", 0, 0, 0,
998 after_gc_async_thunk),
999 SCM_BOOL_F);
1000
1001 scm_c_hook_add (&scm_before_gc_c_hook, queue_after_gc_hook, NULL, 0);
1002 scm_c_hook_add (&scm_before_gc_c_hook, start_gc_timer, NULL, 0);
1003 scm_c_hook_add (&scm_after_gc_c_hook, accumulate_gc_timer, NULL, 0);
1004
1005 /* GC_get_heap_usage does not take a lock, and so can run in the GC
1006 start hook. */
1007 scm_c_hook_add (&scm_before_gc_c_hook, adjust_gc_frequency, NULL, 0);
1008
1009 GC_set_start_callback (run_before_gc_c_hook);
1010
1011 #include "libguile/gc.x"
1012 }
1013
1014
1015 void
1016 scm_gc_sweep (void)
1017 #define FUNC_NAME "scm_gc_sweep"
1018 {
1019 /* FIXME */
1020 fprintf (stderr, "%s: doing nothing\n", FUNC_NAME);
1021 }
1022 #undef FUNC_NAME
1023
1024 /*
1025 Local Variables:
1026 c-file-style: "gnu"
1027 End:
1028 */