2 * GENerational Conservative Garbage Collector for SBCL
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
17 * For a review of garbage collection techniques (e.g. generational
18 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
19 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
20 * had been accepted for _ACM Computing Surveys_ and was available
21 * as a PostScript preprint through
22 * <http://www.cs.utexas.edu/users/oops/papers.html>
24 * <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
37 #include "interrupt.h"
42 #include "gc-internal.h"
44 #include "pseudo-atomic.h"
46 #include "genesis/vector.h"
47 #include "genesis/weak-pointer.h"
48 #include "genesis/fdefn.h"
49 #include "genesis/simple-fun.h"
51 #include "genesis/hash-table.h"
52 #include "genesis/instance.h"
53 #include "genesis/layout.h"
55 #if defined(LUTEX_WIDETAG)
56 #include "pthread-lutex.h"
58 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
59 #include "genesis/cons.h"
62 /* forward declarations */
63 page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
71 /* Generations 0-5 are normal collected generations, 6 is only used as
72 * scratch space by the collector, and should never get collected.
75 SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
79 /* Should we use page protection to help avoid the scavenging of pages
80 * that don't have pointers to younger generations? */
81 boolean enable_page_protection = 1;
83 /* the minimum size (in bytes) for a large object*/
84 #if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES)
85 long large_object_size = 4 * GENCGC_ALLOC_GRANULARITY;
86 #elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY)
87 long large_object_size = 4 * GENCGC_CARD_BYTES;
89 long large_object_size = 4 * PAGE_BYTES;
97 /* the verbosity level. All non-error messages are disabled at level 0;
98 * and only a few rare messages are printed at level 1. */
100 boolean gencgc_verbose = 1;
102 boolean gencgc_verbose = 0;
105 /* FIXME: At some point enable the various error-checking things below
106 * and see what they say. */
108 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
109 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
111 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
113 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
114 boolean pre_verify_gen_0 = 0;
116 /* Should we check for bad pointers after gc_free_heap is called
117 * from Lisp PURIFY? */
118 boolean verify_after_free_heap = 0;
120 /* Should we print a note when code objects are found in the dynamic space
121 * during a heap verify? */
122 boolean verify_dynamic_code_check = 0;
124 /* Should we check code objects for fixup errors after they are transported? */
125 boolean check_code_fixups = 0;
127 /* Should we check that newly allocated regions are zero filled? */
128 boolean gencgc_zero_check = 0;
130 /* Should we check that the free space is zero filled? */
131 boolean gencgc_enable_verify_zero_fill = 0;
133 /* Should we check that free pages are zero filled during gc_free_heap
134 * called after Lisp PURIFY? */
135 boolean gencgc_zero_check_during_free_heap = 0;
137 /* When loading a core, don't do a full scan of the memory for the
138 * memory region boundaries. (Set to true by coreparse.c if the core
139 * contained a pagetable entry).
141 boolean gencgc_partial_pickup = 0;
143 /* If defined, free pages are read-protected to ensure that nothing
147 /* #define READ_PROTECT_FREE_PAGES */
151 * GC structures and variables
154 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
155 unsigned long bytes_allocated = 0;
156 unsigned long auto_gc_trigger = 0;
158 /* the source and destination generations. These are set before a GC starts
160 generation_index_t from_space;
161 generation_index_t new_space;
163 /* Set to 1 when in GC */
164 boolean gc_active_p = 0;
166 /* should the GC be conservative on stack. If false (only right before
167 * saving a core), don't scan the stack / mark pages dont_move. */
168 static boolean conservative_stack = 1;
170 /* An array of page structures is allocated on gc initialization.
171 * This helps quickly map between an address its page structure.
172 * page_table_pages is set from the size of the dynamic space. */
173 page_index_t page_table_pages;
174 struct page *page_table;
176 static inline boolean page_allocated_p(page_index_t page) {
177 return (page_table[page].allocated != FREE_PAGE_FLAG);
180 static inline boolean page_no_region_p(page_index_t page) {
181 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
184 static inline boolean page_allocated_no_region_p(page_index_t page) {
185 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
186 && page_no_region_p(page));
189 static inline boolean page_free_p(page_index_t page) {
190 return (page_table[page].allocated == FREE_PAGE_FLAG);
193 static inline boolean page_boxed_p(page_index_t page) {
194 return (page_table[page].allocated & BOXED_PAGE_FLAG);
197 static inline boolean code_page_p(page_index_t page) {
198 return (page_table[page].allocated & CODE_PAGE_FLAG);
201 static inline boolean page_boxed_no_region_p(page_index_t page) {
202 return page_boxed_p(page) && page_no_region_p(page);
205 static inline boolean page_unboxed_p(page_index_t page) {
206 /* Both flags set == boxed code page */
207 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
208 && !page_boxed_p(page));
211 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
212 return (page_boxed_no_region_p(page)
213 && (page_table[page].bytes_used != 0)
214 && !page_table[page].dont_move
215 && (page_table[page].gen == generation));
218 /* To map addresses to page structures the address of the first page
220 static void *heap_base = NULL;
222 /* Calculate the start address for the given page number. */
224 page_address(page_index_t page_num)
226 return (heap_base + (page_num * GENCGC_CARD_BYTES));
229 /* Calculate the address where the allocation region associated with
230 * the page starts. */
232 page_region_start(page_index_t page_index)
234 return page_address(page_index)-page_table[page_index].region_start_offset;
237 /* Find the page index within the page_table for the given
238 * address. Return -1 on failure. */
240 find_page_index(void *addr)
242 if (addr >= heap_base) {
243 page_index_t index = ((pointer_sized_uint_t)addr -
244 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
245 if (index < page_table_pages)
252 npage_bytes(long npages)
254 gc_assert(npages>=0);
255 return ((unsigned long)npages)*GENCGC_CARD_BYTES;
258 /* Check that X is a higher address than Y and return offset from Y to
261 size_t void_diff(void *x, void *y)
264 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
267 /* a structure to hold the state of a generation
269 * CAUTION: If you modify this, make sure to touch up the alien
270 * definition in src/code/gc.lisp accordingly. ...or better yes,
271 * deal with the FIXME there...
275 /* the first page that gc_alloc() checks on its next call */
276 page_index_t alloc_start_page;
278 /* the first page that gc_alloc_unboxed() checks on its next call */
279 page_index_t alloc_unboxed_start_page;
281 /* the first page that gc_alloc_large (boxed) considers on its next
282 * call. (Although it always allocates after the boxed_region.) */
283 page_index_t alloc_large_start_page;
285 /* the first page that gc_alloc_large (unboxed) considers on its
286 * next call. (Although it always allocates after the
287 * current_unboxed_region.) */
288 page_index_t alloc_large_unboxed_start_page;
290 /* the bytes allocated to this generation */
291 unsigned long bytes_allocated;
293 /* the number of bytes at which to trigger a GC */
294 unsigned long gc_trigger;
296 /* to calculate a new level for gc_trigger */
297 unsigned long bytes_consed_between_gc;
299 /* the number of GCs since the last raise */
302 /* the number of GCs to run on the generations before raising objects to the
304 int number_of_gcs_before_promotion;
306 /* the cumulative sum of the bytes allocated to this generation. It is
307 * cleared after a GC on this generations, and update before new
308 * objects are added from a GC of a younger generation. Dividing by
309 * the bytes_allocated will give the average age of the memory in
310 * this generation since its last GC. */
311 unsigned long cum_sum_bytes_allocated;
313 /* a minimum average memory age before a GC will occur helps
314 * prevent a GC when a large number of new live objects have been
315 * added, in which case a GC could be a waste of time */
316 double minimum_age_before_gc;
318 /* A linked list of lutex structures in this generation, used for
319 * implementing lutex finalization. */
321 struct lutex *lutexes;
327 /* an array of generation structures. There needs to be one more
328 * generation structure than actual generations as the oldest
329 * generation is temporarily raised then lowered. */
330 struct generation generations[NUM_GENERATIONS];
332 /* the oldest generation that is will currently be GCed by default.
333 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
335 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
337 * Setting this to 0 effectively disables the generational nature of
338 * the GC. In some applications generational GC may not be useful
339 * because there are no long-lived objects.
341 * An intermediate value could be handy after moving long-lived data
342 * into an older generation so an unnecessary GC of this long-lived
343 * data can be avoided. */
344 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
346 /* The maximum free page in the heap is maintained and used to update
347 * ALLOCATION_POINTER which is used by the room function to limit its
348 * search of the heap. XX Gencgc obviously needs to be better
349 * integrated with the Lisp code. */
350 page_index_t last_free_page;
352 #ifdef LISP_FEATURE_SB_THREAD
353 /* This lock is to prevent multiple threads from simultaneously
354 * allocating new regions which overlap each other. Note that the
355 * majority of GC is single-threaded, but alloc() may be called from
356 * >1 thread at a time and must be thread-safe. This lock must be
357 * seized before all accesses to generations[] or to parts of
358 * page_table[] that other threads may want to see */
359 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
360 /* This lock is used to protect non-thread-local allocation. */
361 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
364 extern unsigned long gencgc_release_granularity;
365 unsigned long gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
367 extern unsigned long gencgc_alloc_granularity;
368 unsigned long gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
372 * miscellaneous heap functions
375 /* Count the number of pages which are write-protected within the
376 * given generation. */
378 count_write_protect_generation_pages(generation_index_t generation)
381 unsigned long count = 0;
383 for (i = 0; i < last_free_page; i++)
384 if (page_allocated_p(i)
385 && (page_table[i].gen == generation)
386 && (page_table[i].write_protected == 1))
391 /* Count the number of pages within the given generation. */
393 count_generation_pages(generation_index_t generation)
398 for (i = 0; i < last_free_page; i++)
399 if (page_allocated_p(i)
400 && (page_table[i].gen == generation))
407 count_dont_move_pages(void)
411 for (i = 0; i < last_free_page; i++) {
412 if (page_allocated_p(i)
413 && (page_table[i].dont_move != 0)) {
421 /* Work through the pages and add up the number of bytes used for the
422 * given generation. */
424 count_generation_bytes_allocated (generation_index_t gen)
427 unsigned long result = 0;
428 for (i = 0; i < last_free_page; i++) {
429 if (page_allocated_p(i)
430 && (page_table[i].gen == gen))
431 result += page_table[i].bytes_used;
436 /* Return the average age of the memory in a generation. */
438 generation_average_age(generation_index_t gen)
440 if (generations[gen].bytes_allocated == 0)
444 ((double)generations[gen].cum_sum_bytes_allocated)
445 / ((double)generations[gen].bytes_allocated);
449 write_generation_stats(FILE *file)
451 generation_index_t i;
453 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
454 #define FPU_STATE_SIZE 27
455 int fpu_state[FPU_STATE_SIZE];
456 #elif defined(LISP_FEATURE_PPC)
457 #define FPU_STATE_SIZE 32
458 long long fpu_state[FPU_STATE_SIZE];
461 /* This code uses the FP instructions which may be set up for Lisp
462 * so they need to be saved and reset for C. */
465 /* Print the heap stats. */
467 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
469 for (i = 0; i < SCRATCH_GENERATION; i++) {
472 long unboxed_cnt = 0;
473 long large_boxed_cnt = 0;
474 long large_unboxed_cnt = 0;
477 for (j = 0; j < last_free_page; j++)
478 if (page_table[j].gen == i) {
480 /* Count the number of boxed pages within the given
482 if (page_boxed_p(j)) {
483 if (page_table[j].large_object)
488 if(page_table[j].dont_move) pinned_cnt++;
489 /* Count the number of unboxed pages within the given
491 if (page_unboxed_p(j)) {
492 if (page_table[j].large_object)
499 gc_assert(generations[i].bytes_allocated
500 == count_generation_bytes_allocated(i));
502 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
504 generations[i].alloc_start_page,
505 generations[i].alloc_unboxed_start_page,
506 generations[i].alloc_large_start_page,
507 generations[i].alloc_large_unboxed_start_page,
513 generations[i].bytes_allocated,
514 (npage_bytes(count_generation_pages(i))
515 - generations[i].bytes_allocated),
516 generations[i].gc_trigger,
517 count_write_protect_generation_pages(i),
518 generations[i].num_gc,
519 generation_average_age(i));
521 fprintf(file," Total bytes allocated = %lu\n", bytes_allocated);
522 fprintf(file," Dynamic-space-size bytes = %lu\n", (unsigned long)dynamic_space_size);
524 fpu_restore(fpu_state);
528 write_heap_exhaustion_report(FILE *file, long available, long requested,
529 struct thread *thread)
532 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
533 gc_active_p ? "garbage collection" : "allocation",
536 write_generation_stats(file);
537 fprintf(file, "GC control variables:\n");
538 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
539 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
540 (SymbolValue(GC_PENDING, thread) == T) ?
541 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
542 "false" : "in progress"));
543 #ifdef LISP_FEATURE_SB_THREAD
544 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
545 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
550 print_generation_stats(void)
552 write_generation_stats(stderr);
555 extern char* gc_logfile;
556 char * gc_logfile = NULL;
559 log_generation_stats(char *logfile, char *header)
562 FILE * log = fopen(logfile, "a");
564 fprintf(log, "%s\n", header);
565 write_generation_stats(log);
568 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
575 report_heap_exhaustion(long available, long requested, struct thread *th)
578 FILE * log = fopen(gc_logfile, "a");
580 write_heap_exhaustion_report(log, available, requested, th);
583 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
587 /* Always to stderr as well. */
588 write_heap_exhaustion_report(stderr, available, requested, th);
592 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
593 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
596 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
597 * if zeroing it ourselves, i.e. in practice give the memory back to the
598 * OS. Generally done after a large GC.
600 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
602 void *addr = page_address(start), *new_addr;
603 size_t length = npage_bytes(1+end-start);
608 gc_assert(length >= gencgc_release_granularity);
609 gc_assert((length % gencgc_release_granularity) == 0);
611 os_invalidate(addr, length);
612 new_addr = os_validate(addr, length);
613 if (new_addr == NULL || new_addr != addr) {
614 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
618 for (i = start; i <= end; i++) {
619 page_table[i].need_to_zero = 0;
623 /* Zero the pages from START to END (inclusive). Generally done just after
624 * a new region has been allocated.
627 zero_pages(page_index_t start, page_index_t end) {
631 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
632 fast_bzero(page_address(start), npage_bytes(1+end-start));
634 bzero(page_address(start), npage_bytes(1+end-start));
640 zero_and_mark_pages(page_index_t start, page_index_t end) {
643 zero_pages(start, end);
644 for (i = start; i <= end; i++)
645 page_table[i].need_to_zero = 0;
648 /* Zero the pages from START to END (inclusive), except for those
649 * pages that are known to already zeroed. Mark all pages in the
650 * ranges as non-zeroed.
653 zero_dirty_pages(page_index_t start, page_index_t end) {
656 for (i = start; i <= end; i++) {
657 if (!page_table[i].need_to_zero) continue;
658 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
663 for (i = start; i <= end; i++) {
664 page_table[i].need_to_zero = 1;
670 * To support quick and inline allocation, regions of memory can be
671 * allocated and then allocated from with just a free pointer and a
672 * check against an end address.
674 * Since objects can be allocated to spaces with different properties
675 * e.g. boxed/unboxed, generation, ages; there may need to be many
676 * allocation regions.
678 * Each allocation region may start within a partly used page. Many
679 * features of memory use are noted on a page wise basis, e.g. the
680 * generation; so if a region starts within an existing allocated page
681 * it must be consistent with this page.
683 * During the scavenging of the newspace, objects will be transported
684 * into an allocation region, and pointers updated to point to this
685 * allocation region. It is possible that these pointers will be
686 * scavenged again before the allocation region is closed, e.g. due to
687 * trans_list which jumps all over the place to cleanup the list. It
688 * is important to be able to determine properties of all objects
689 * pointed to when scavenging, e.g to detect pointers to the oldspace.
690 * Thus it's important that the allocation regions have the correct
691 * properties set when allocated, and not just set when closed. The
692 * region allocation routines return regions with the specified
693 * properties, and grab all the pages, setting their properties
694 * appropriately, except that the amount used is not known.
696 * These regions are used to support quicker allocation using just a
697 * free pointer. The actual space used by the region is not reflected
698 * in the pages tables until it is closed. It can't be scavenged until
701 * When finished with the region it should be closed, which will
702 * update the page tables for the actual space used returning unused
703 * space. Further it may be noted in the new regions which is
704 * necessary when scavenging the newspace.
706 * Large objects may be allocated directly without an allocation
707 * region, the page tables are updated immediately.
709 * Unboxed objects don't contain pointers to other objects and so
710 * don't need scavenging. Further they can't contain pointers to
711 * younger generations so WP is not needed. By allocating pages to
712 * unboxed objects the whole page never needs scavenging or
713 * write-protecting. */
715 /* We are only using two regions at present. Both are for the current
716 * newspace generation. */
717 struct alloc_region boxed_region;
718 struct alloc_region unboxed_region;
720 /* The generation currently being allocated to. */
721 static generation_index_t gc_alloc_generation;
723 static inline page_index_t
724 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
727 if (UNBOXED_PAGE_FLAG == page_type_flag) {
728 return generations[generation].alloc_large_unboxed_start_page;
729 } else if (BOXED_PAGE_FLAG & page_type_flag) {
730 /* Both code and data. */
731 return generations[generation].alloc_large_start_page;
733 lose("bad page type flag: %d", page_type_flag);
736 if (UNBOXED_PAGE_FLAG == page_type_flag) {
737 return generations[generation].alloc_unboxed_start_page;
738 } else if (BOXED_PAGE_FLAG & page_type_flag) {
739 /* Both code and data. */
740 return generations[generation].alloc_start_page;
742 lose("bad page_type_flag: %d", page_type_flag);
748 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
752 if (UNBOXED_PAGE_FLAG == page_type_flag) {
753 generations[generation].alloc_large_unboxed_start_page = page;
754 } else if (BOXED_PAGE_FLAG & page_type_flag) {
755 /* Both code and data. */
756 generations[generation].alloc_large_start_page = page;
758 lose("bad page type flag: %d", page_type_flag);
761 if (UNBOXED_PAGE_FLAG == page_type_flag) {
762 generations[generation].alloc_unboxed_start_page = page;
763 } else if (BOXED_PAGE_FLAG & page_type_flag) {
764 /* Both code and data. */
765 generations[generation].alloc_start_page = page;
767 lose("bad page type flag: %d", page_type_flag);
772 /* Find a new region with room for at least the given number of bytes.
774 * It starts looking at the current generation's alloc_start_page. So
775 * may pick up from the previous region if there is enough space. This
776 * keeps the allocation contiguous when scavenging the newspace.
778 * The alloc_region should have been closed by a call to
779 * gc_alloc_update_page_tables(), and will thus be in an empty state.
781 * To assist the scavenging functions write-protected pages are not
782 * used. Free pages should not be write-protected.
784 * It is critical to the conservative GC that the start of regions be
785 * known. To help achieve this only small regions are allocated at a
788 * During scavenging, pointers may be found to within the current
789 * region and the page generation must be set so that pointers to the
790 * from space can be recognized. Therefore the generation of pages in
791 * the region are set to gc_alloc_generation. To prevent another
792 * allocation call using the same pages, all the pages in the region
793 * are allocated, although they will initially be empty.
796 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
798 page_index_t first_page;
799 page_index_t last_page;
800 unsigned long bytes_found;
806 "/alloc_new_region for %d bytes from gen %d\n",
807 nbytes, gc_alloc_generation));
810 /* Check that the region is in a reset state. */
811 gc_assert((alloc_region->first_page == 0)
812 && (alloc_region->last_page == -1)
813 && (alloc_region->free_pointer == alloc_region->end_addr));
814 ret = thread_mutex_lock(&free_pages_lock);
816 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
817 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
818 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
819 + npage_bytes(last_page-first_page);
821 /* Set up the alloc_region. */
822 alloc_region->first_page = first_page;
823 alloc_region->last_page = last_page;
824 alloc_region->start_addr = page_table[first_page].bytes_used
825 + page_address(first_page);
826 alloc_region->free_pointer = alloc_region->start_addr;
827 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
829 /* Set up the pages. */
831 /* The first page may have already been in use. */
832 if (page_table[first_page].bytes_used == 0) {
833 page_table[first_page].allocated = page_type_flag;
834 page_table[first_page].gen = gc_alloc_generation;
835 page_table[first_page].large_object = 0;
836 page_table[first_page].region_start_offset = 0;
839 gc_assert(page_table[first_page].allocated == page_type_flag);
840 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
842 gc_assert(page_table[first_page].gen == gc_alloc_generation);
843 gc_assert(page_table[first_page].large_object == 0);
845 for (i = first_page+1; i <= last_page; i++) {
846 page_table[i].allocated = page_type_flag;
847 page_table[i].gen = gc_alloc_generation;
848 page_table[i].large_object = 0;
849 /* This may not be necessary for unboxed regions (think it was
851 page_table[i].region_start_offset =
852 void_diff(page_address(i),alloc_region->start_addr);
853 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
855 /* Bump up last_free_page. */
856 if (last_page+1 > last_free_page) {
857 last_free_page = last_page+1;
858 /* do we only want to call this on special occasions? like for
860 set_alloc_pointer((lispobj)page_address(last_free_page));
862 ret = thread_mutex_unlock(&free_pages_lock);
865 #ifdef READ_PROTECT_FREE_PAGES
866 os_protect(page_address(first_page),
867 npage_bytes(1+last_page-first_page),
871 /* If the first page was only partial, don't check whether it's
872 * zeroed (it won't be) and don't zero it (since the parts that
873 * we're interested in are guaranteed to be zeroed).
875 if (page_table[first_page].bytes_used) {
879 zero_dirty_pages(first_page, last_page);
881 /* we can do this after releasing free_pages_lock */
882 if (gencgc_zero_check) {
884 for (p = (long *)alloc_region->start_addr;
885 p < (long *)alloc_region->end_addr; p++) {
887 /* KLUDGE: It would be nice to use %lx and explicit casts
888 * (long) in code like this, so that it is less likely to
889 * break randomly when running on a machine with different
890 * word sizes. -- WHN 19991129 */
891 lose("The new region at %x is not zero (start=%p, end=%p).\n",
892 p, alloc_region->start_addr, alloc_region->end_addr);
898 /* If the record_new_objects flag is 2 then all new regions created
901 * If it's 1 then then it is only recorded if the first page of the
902 * current region is <= new_areas_ignore_page. This helps avoid
903 * unnecessary recording when doing full scavenge pass.
905 * The new_object structure holds the page, byte offset, and size of
906 * new regions of objects. Each new area is placed in the array of
907 * these structures pointer to by new_areas. new_areas_index holds the
908 * offset into new_areas.
910 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
911 * later code must detect this and handle it, probably by doing a full
912 * scavenge of a generation. */
913 #define NUM_NEW_AREAS 512
914 static int record_new_objects = 0;
915 static page_index_t new_areas_ignore_page;
921 static struct new_area (*new_areas)[];
922 static long new_areas_index;
925 /* Add a new area to new_areas. */
927 add_new_area(page_index_t first_page, size_t offset, size_t size)
929 unsigned long new_area_start,c;
932 /* Ignore if full. */
933 if (new_areas_index >= NUM_NEW_AREAS)
936 switch (record_new_objects) {
940 if (first_page > new_areas_ignore_page)
949 new_area_start = npage_bytes(first_page) + offset;
951 /* Search backwards for a prior area that this follows from. If
952 found this will save adding a new area. */
953 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
954 unsigned long area_end =
955 npage_bytes((*new_areas)[i].page)
956 + (*new_areas)[i].offset
957 + (*new_areas)[i].size;
959 "/add_new_area S1 %d %d %d %d\n",
960 i, c, new_area_start, area_end));*/
961 if (new_area_start == area_end) {
963 "/adding to [%d] %d %d %d with %d %d %d:\n",
965 (*new_areas)[i].page,
966 (*new_areas)[i].offset,
967 (*new_areas)[i].size,
971 (*new_areas)[i].size += size;
976 (*new_areas)[new_areas_index].page = first_page;
977 (*new_areas)[new_areas_index].offset = offset;
978 (*new_areas)[new_areas_index].size = size;
980 "/new_area %d page %d offset %d size %d\n",
981 new_areas_index, first_page, offset, size));*/
984 /* Note the max new_areas used. */
985 if (new_areas_index > max_new_areas)
986 max_new_areas = new_areas_index;
989 /* Update the tables for the alloc_region. The region may be added to
992 * When done the alloc_region is set up so that the next quick alloc
993 * will fail safely and thus a new region will be allocated. Further
994 * it is safe to try to re-update the page table of this reset
997 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
1000 page_index_t first_page;
1001 page_index_t next_page;
1002 unsigned long bytes_used;
1003 unsigned long orig_first_page_bytes_used;
1004 unsigned long region_size;
1005 unsigned long byte_cnt;
1009 first_page = alloc_region->first_page;
1011 /* Catch an unused alloc_region. */
1012 if ((first_page == 0) && (alloc_region->last_page == -1))
1015 next_page = first_page+1;
1017 ret = thread_mutex_lock(&free_pages_lock);
1018 gc_assert(ret == 0);
1019 if (alloc_region->free_pointer != alloc_region->start_addr) {
1020 /* some bytes were allocated in the region */
1021 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1023 gc_assert(alloc_region->start_addr ==
1024 (page_address(first_page)
1025 + page_table[first_page].bytes_used));
1027 /* All the pages used need to be updated */
1029 /* Update the first page. */
1031 /* If the page was free then set up the gen, and
1032 * region_start_offset. */
1033 if (page_table[first_page].bytes_used == 0)
1034 gc_assert(page_table[first_page].region_start_offset == 0);
1035 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1037 gc_assert(page_table[first_page].allocated & page_type_flag);
1038 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1039 gc_assert(page_table[first_page].large_object == 0);
1043 /* Calculate the number of bytes used in this page. This is not
1044 * always the number of new bytes, unless it was free. */
1046 if ((bytes_used = void_diff(alloc_region->free_pointer,
1047 page_address(first_page)))
1048 >GENCGC_CARD_BYTES) {
1049 bytes_used = GENCGC_CARD_BYTES;
1052 page_table[first_page].bytes_used = bytes_used;
1053 byte_cnt += bytes_used;
1056 /* All the rest of the pages should be free. We need to set
1057 * their region_start_offset pointer to the start of the
1058 * region, and set the bytes_used. */
1060 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1061 gc_assert(page_table[next_page].allocated & page_type_flag);
1062 gc_assert(page_table[next_page].bytes_used == 0);
1063 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1064 gc_assert(page_table[next_page].large_object == 0);
1066 gc_assert(page_table[next_page].region_start_offset ==
1067 void_diff(page_address(next_page),
1068 alloc_region->start_addr));
1070 /* Calculate the number of bytes used in this page. */
1072 if ((bytes_used = void_diff(alloc_region->free_pointer,
1073 page_address(next_page)))>GENCGC_CARD_BYTES) {
1074 bytes_used = GENCGC_CARD_BYTES;
1077 page_table[next_page].bytes_used = bytes_used;
1078 byte_cnt += bytes_used;
1083 region_size = void_diff(alloc_region->free_pointer,
1084 alloc_region->start_addr);
1085 bytes_allocated += region_size;
1086 generations[gc_alloc_generation].bytes_allocated += region_size;
1088 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1090 /* Set the generations alloc restart page to the last page of
1092 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1094 /* Add the region to the new_areas if requested. */
1095 if (BOXED_PAGE_FLAG & page_type_flag)
1096 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1100 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1102 gc_alloc_generation));
1105 /* There are no bytes allocated. Unallocate the first_page if
1106 * there are 0 bytes_used. */
1107 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1108 if (page_table[first_page].bytes_used == 0)
1109 page_table[first_page].allocated = FREE_PAGE_FLAG;
1112 /* Unallocate any unused pages. */
1113 while (next_page <= alloc_region->last_page) {
1114 gc_assert(page_table[next_page].bytes_used == 0);
1115 page_table[next_page].allocated = FREE_PAGE_FLAG;
1118 ret = thread_mutex_unlock(&free_pages_lock);
1119 gc_assert(ret == 0);
1121 /* alloc_region is per-thread, we're ok to do this unlocked */
1122 gc_set_region_empty(alloc_region);
1125 static inline void *gc_quick_alloc(long nbytes);
1127 /* Allocate a possibly large object. */
1129 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1131 page_index_t first_page;
1132 page_index_t last_page;
1133 int orig_first_page_bytes_used;
1136 unsigned long bytes_used;
1137 page_index_t next_page;
1140 ret = thread_mutex_lock(&free_pages_lock);
1141 gc_assert(ret == 0);
1143 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1144 if (first_page <= alloc_region->last_page) {
1145 first_page = alloc_region->last_page+1;
1148 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1150 gc_assert(first_page > alloc_region->last_page);
1152 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1154 /* Set up the pages. */
1155 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1157 /* If the first page was free then set up the gen, and
1158 * region_start_offset. */
1159 if (page_table[first_page].bytes_used == 0) {
1160 page_table[first_page].allocated = page_type_flag;
1161 page_table[first_page].gen = gc_alloc_generation;
1162 page_table[first_page].region_start_offset = 0;
1163 page_table[first_page].large_object = 1;
1166 gc_assert(page_table[first_page].allocated == page_type_flag);
1167 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1168 gc_assert(page_table[first_page].large_object == 1);
1172 /* Calc. the number of bytes used in this page. This is not
1173 * always the number of new bytes, unless it was free. */
1175 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1176 bytes_used = GENCGC_CARD_BYTES;
1179 page_table[first_page].bytes_used = bytes_used;
1180 byte_cnt += bytes_used;
1182 next_page = first_page+1;
1184 /* All the rest of the pages should be free. We need to set their
1185 * region_start_offset pointer to the start of the region, and set
1186 * the bytes_used. */
1188 gc_assert(page_free_p(next_page));
1189 gc_assert(page_table[next_page].bytes_used == 0);
1190 page_table[next_page].allocated = page_type_flag;
1191 page_table[next_page].gen = gc_alloc_generation;
1192 page_table[next_page].large_object = 1;
1194 page_table[next_page].region_start_offset =
1195 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1197 /* Calculate the number of bytes used in this page. */
1199 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1200 if (bytes_used > GENCGC_CARD_BYTES) {
1201 bytes_used = GENCGC_CARD_BYTES;
1204 page_table[next_page].bytes_used = bytes_used;
1205 page_table[next_page].write_protected=0;
1206 page_table[next_page].dont_move=0;
1207 byte_cnt += bytes_used;
1211 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1213 bytes_allocated += nbytes;
1214 generations[gc_alloc_generation].bytes_allocated += nbytes;
1216 /* Add the region to the new_areas if requested. */
1217 if (BOXED_PAGE_FLAG & page_type_flag)
1218 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1220 /* Bump up last_free_page */
1221 if (last_page+1 > last_free_page) {
1222 last_free_page = last_page+1;
1223 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1225 ret = thread_mutex_unlock(&free_pages_lock);
1226 gc_assert(ret == 0);
1228 #ifdef READ_PROTECT_FREE_PAGES
1229 os_protect(page_address(first_page),
1230 npage_bytes(1+last_page-first_page),
1234 zero_dirty_pages(first_page, last_page);
1236 return page_address(first_page);
1239 static page_index_t gencgc_alloc_start_page = -1;
1242 gc_heap_exhausted_error_or_lose (long available, long requested)
1244 struct thread *thread = arch_os_get_current_thread();
1245 /* Write basic information before doing anything else: if we don't
1246 * call to lisp this is a must, and even if we do there is always
1247 * the danger that we bounce back here before the error has been
1248 * handled, or indeed even printed.
1250 report_heap_exhaustion(available, requested, thread);
1251 if (gc_active_p || (available == 0)) {
1252 /* If we are in GC, or totally out of memory there is no way
1253 * to sanely transfer control to the lisp-side of things.
1255 lose("Heap exhausted, game over.");
1258 /* FIXME: assert free_pages_lock held */
1259 (void)thread_mutex_unlock(&free_pages_lock);
1260 gc_assert(get_pseudo_atomic_atomic(thread));
1261 clear_pseudo_atomic_atomic(thread);
1262 if (get_pseudo_atomic_interrupted(thread))
1263 do_pending_interrupt();
1264 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1265 * to running user code at arbitrary places, even in a
1266 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1267 * running out of the heap. So at this point all bets are
1269 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1270 corruption_warning_and_maybe_lose
1271 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1272 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1273 alloc_number(available), alloc_number(requested));
1274 lose("HEAP-EXHAUSTED-ERROR fell through");
1279 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes,
1282 page_index_t first_page, last_page;
1283 page_index_t restart_page = *restart_page_ptr;
1284 long nbytes_goal = nbytes;
1285 long bytes_found = 0;
1286 long most_bytes_found = 0;
1287 page_index_t most_bytes_found_from, most_bytes_found_to;
1288 int small_object = nbytes < GENCGC_CARD_BYTES;
1289 /* FIXME: assert(free_pages_lock is held); */
1291 if (nbytes_goal < gencgc_alloc_granularity)
1292 nbytes_goal = gencgc_alloc_granularity;
1294 /* Toggled by gc_and_save for heap compaction, normally -1. */
1295 if (gencgc_alloc_start_page != -1) {
1296 restart_page = gencgc_alloc_start_page;
1299 gc_assert(nbytes>=0);
1300 /* Search for a page with at least nbytes of space. We prefer
1301 * not to split small objects on multiple pages, to reduce the
1302 * number of contiguous allocation regions spaning multiple
1303 * pages: this helps avoid excessive conservativism.
1305 * For other objects, we guarantee that they start on their own
1308 first_page = restart_page;
1309 while (first_page < page_table_pages) {
1311 if (page_free_p(first_page)) {
1312 gc_assert(0 == page_table[first_page].bytes_used);
1313 bytes_found = GENCGC_CARD_BYTES;
1314 } else if (small_object &&
1315 (page_table[first_page].allocated == page_type_flag) &&
1316 (page_table[first_page].large_object == 0) &&
1317 (page_table[first_page].gen == gc_alloc_generation) &&
1318 (page_table[first_page].write_protected == 0) &&
1319 (page_table[first_page].dont_move == 0)) {
1320 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1321 if (bytes_found < nbytes) {
1322 if (bytes_found > most_bytes_found)
1323 most_bytes_found = bytes_found;
1332 gc_assert(page_table[first_page].write_protected == 0);
1333 for (last_page = first_page+1;
1334 ((last_page < page_table_pages) &&
1335 page_free_p(last_page) &&
1336 (bytes_found < nbytes_goal));
1338 bytes_found += GENCGC_CARD_BYTES;
1339 gc_assert(0 == page_table[last_page].bytes_used);
1340 gc_assert(0 == page_table[last_page].write_protected);
1343 if (bytes_found > most_bytes_found) {
1344 most_bytes_found = bytes_found;
1345 most_bytes_found_from = first_page;
1346 most_bytes_found_to = last_page;
1348 if (bytes_found >= nbytes_goal)
1351 first_page = last_page;
1354 bytes_found = most_bytes_found;
1355 restart_page = first_page + 1;
1357 /* Check for a failure */
1358 if (bytes_found < nbytes) {
1359 gc_assert(restart_page >= page_table_pages);
1360 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1363 *restart_page_ptr = most_bytes_found_from;
1364 return most_bytes_found_to-1;
1367 /* Allocate bytes. All the rest of the special-purpose allocation
1368 * functions will eventually call this */
1371 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1374 void *new_free_pointer;
1376 if (nbytes>=large_object_size)
1377 return gc_alloc_large(nbytes, page_type_flag, my_region);
1379 /* Check whether there is room in the current alloc region. */
1380 new_free_pointer = my_region->free_pointer + nbytes;
1382 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1383 my_region->free_pointer, new_free_pointer); */
1385 if (new_free_pointer <= my_region->end_addr) {
1386 /* If so then allocate from the current alloc region. */
1387 void *new_obj = my_region->free_pointer;
1388 my_region->free_pointer = new_free_pointer;
1390 /* Unless a `quick' alloc was requested, check whether the
1391 alloc region is almost empty. */
1393 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1394 /* If so, finished with the current region. */
1395 gc_alloc_update_page_tables(page_type_flag, my_region);
1396 /* Set up a new region. */
1397 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1400 return((void *)new_obj);
1403 /* Else not enough free space in the current region: retry with a
1406 gc_alloc_update_page_tables(page_type_flag, my_region);
1407 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1408 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1411 /* these are only used during GC: all allocation from the mutator calls
1412 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1415 static inline void *
1416 gc_quick_alloc(long nbytes)
1418 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1421 static inline void *
1422 gc_quick_alloc_large(long nbytes)
1424 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1427 static inline void *
1428 gc_alloc_unboxed(long nbytes)
1430 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1433 static inline void *
1434 gc_quick_alloc_unboxed(long nbytes)
1436 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1439 static inline void *
1440 gc_quick_alloc_large_unboxed(long nbytes)
1442 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1446 /* Copy a large boxed object. If the object is in a large object
1447 * region then it is simply promoted, else it is copied. If it's large
1448 * enough then it's copied to a large object region.
1450 * Vectors may have shrunk. If the object is not copied the space
1451 * needs to be reclaimed, and the page_tables corrected. */
1453 copy_large_object(lispobj object, long nwords)
1457 page_index_t first_page;
1459 gc_assert(is_lisp_pointer(object));
1460 gc_assert(from_space_p(object));
1461 gc_assert((nwords & 0x01) == 0);
1464 /* Check whether it's in a large object region. */
1465 first_page = find_page_index((void *)object);
1466 gc_assert(first_page >= 0);
1468 if (page_table[first_page].large_object) {
1470 /* Promote the object. */
1472 unsigned long remaining_bytes;
1473 page_index_t next_page;
1474 unsigned long bytes_freed;
1475 unsigned long old_bytes_used;
1477 /* Note: Any page write-protection must be removed, else a
1478 * later scavenge_newspace may incorrectly not scavenge these
1479 * pages. This would not be necessary if they are added to the
1480 * new areas, but let's do it for them all (they'll probably
1481 * be written anyway?). */
1483 gc_assert(page_table[first_page].region_start_offset == 0);
1485 next_page = first_page;
1486 remaining_bytes = nwords*N_WORD_BYTES;
1487 while (remaining_bytes > GENCGC_CARD_BYTES) {
1488 gc_assert(page_table[next_page].gen == from_space);
1489 gc_assert(page_boxed_p(next_page));
1490 gc_assert(page_table[next_page].large_object);
1491 gc_assert(page_table[next_page].region_start_offset ==
1492 npage_bytes(next_page-first_page));
1493 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1494 /* Should have been unprotected by unprotect_oldspace(). */
1495 gc_assert(page_table[next_page].write_protected == 0);
1497 page_table[next_page].gen = new_space;
1499 remaining_bytes -= GENCGC_CARD_BYTES;
1503 /* Now only one page remains, but the object may have shrunk
1504 * so there may be more unused pages which will be freed. */
1506 /* The object may have shrunk but shouldn't have grown. */
1507 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1509 page_table[next_page].gen = new_space;
1510 gc_assert(page_boxed_p(next_page));
1512 /* Adjust the bytes_used. */
1513 old_bytes_used = page_table[next_page].bytes_used;
1514 page_table[next_page].bytes_used = remaining_bytes;
1516 bytes_freed = old_bytes_used - remaining_bytes;
1518 /* Free any remaining pages; needs care. */
1520 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1521 (page_table[next_page].gen == from_space) &&
1522 page_boxed_p(next_page) &&
1523 page_table[next_page].large_object &&
1524 (page_table[next_page].region_start_offset ==
1525 npage_bytes(next_page - first_page))) {
1526 /* Checks out OK, free the page. Don't need to bother zeroing
1527 * pages as this should have been done before shrinking the
1528 * object. These pages shouldn't be write-protected as they
1529 * should be zero filled. */
1530 gc_assert(page_table[next_page].write_protected == 0);
1532 old_bytes_used = page_table[next_page].bytes_used;
1533 page_table[next_page].allocated = FREE_PAGE_FLAG;
1534 page_table[next_page].bytes_used = 0;
1535 bytes_freed += old_bytes_used;
1539 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1541 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1542 bytes_allocated -= bytes_freed;
1544 /* Add the region to the new_areas if requested. */
1545 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1549 /* Get tag of object. */
1550 tag = lowtag_of(object);
1552 /* Allocate space. */
1553 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1555 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1557 /* Return Lisp pointer of new object. */
1558 return ((lispobj) new) | tag;
1562 /* to copy unboxed objects */
1564 copy_unboxed_object(lispobj object, long nwords)
1569 gc_assert(is_lisp_pointer(object));
1570 gc_assert(from_space_p(object));
1571 gc_assert((nwords & 0x01) == 0);
1573 /* Get tag of object. */
1574 tag = lowtag_of(object);
1576 /* Allocate space. */
1577 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1579 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1581 /* Return Lisp pointer of new object. */
1582 return ((lispobj) new) | tag;
1585 /* to copy large unboxed objects
1587 * If the object is in a large object region then it is simply
1588 * promoted, else it is copied. If it's large enough then it's copied
1589 * to a large object region.
1591 * Bignums and vectors may have shrunk. If the object is not copied
1592 * the space needs to be reclaimed, and the page_tables corrected.
1594 * KLUDGE: There's a lot of cut-and-paste duplication between this
1595 * function and copy_large_object(..). -- WHN 20000619 */
1597 copy_large_unboxed_object(lispobj object, long nwords)
1601 page_index_t first_page;
1603 gc_assert(is_lisp_pointer(object));
1604 gc_assert(from_space_p(object));
1605 gc_assert((nwords & 0x01) == 0);
1607 if ((nwords > 1024*1024) && gencgc_verbose) {
1608 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1609 nwords*N_WORD_BYTES));
1612 /* Check whether it's a large object. */
1613 first_page = find_page_index((void *)object);
1614 gc_assert(first_page >= 0);
1616 if (page_table[first_page].large_object) {
1617 /* Promote the object. Note: Unboxed objects may have been
1618 * allocated to a BOXED region so it may be necessary to
1619 * change the region to UNBOXED. */
1620 unsigned long remaining_bytes;
1621 page_index_t next_page;
1622 unsigned long bytes_freed;
1623 unsigned long old_bytes_used;
1625 gc_assert(page_table[first_page].region_start_offset == 0);
1627 next_page = first_page;
1628 remaining_bytes = nwords*N_WORD_BYTES;
1629 while (remaining_bytes > GENCGC_CARD_BYTES) {
1630 gc_assert(page_table[next_page].gen == from_space);
1631 gc_assert(page_allocated_no_region_p(next_page));
1632 gc_assert(page_table[next_page].large_object);
1633 gc_assert(page_table[next_page].region_start_offset ==
1634 npage_bytes(next_page-first_page));
1635 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1637 page_table[next_page].gen = new_space;
1638 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1639 remaining_bytes -= GENCGC_CARD_BYTES;
1643 /* Now only one page remains, but the object may have shrunk so
1644 * there may be more unused pages which will be freed. */
1646 /* Object may have shrunk but shouldn't have grown - check. */
1647 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1649 page_table[next_page].gen = new_space;
1650 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1652 /* Adjust the bytes_used. */
1653 old_bytes_used = page_table[next_page].bytes_used;
1654 page_table[next_page].bytes_used = remaining_bytes;
1656 bytes_freed = old_bytes_used - remaining_bytes;
1658 /* Free any remaining pages; needs care. */
1660 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1661 (page_table[next_page].gen == from_space) &&
1662 page_allocated_no_region_p(next_page) &&
1663 page_table[next_page].large_object &&
1664 (page_table[next_page].region_start_offset ==
1665 npage_bytes(next_page - first_page))) {
1666 /* Checks out OK, free the page. Don't need to both zeroing
1667 * pages as this should have been done before shrinking the
1668 * object. These pages shouldn't be write-protected, even if
1669 * boxed they should be zero filled. */
1670 gc_assert(page_table[next_page].write_protected == 0);
1672 old_bytes_used = page_table[next_page].bytes_used;
1673 page_table[next_page].allocated = FREE_PAGE_FLAG;
1674 page_table[next_page].bytes_used = 0;
1675 bytes_freed += old_bytes_used;
1679 if ((bytes_freed > 0) && gencgc_verbose) {
1681 "/copy_large_unboxed bytes_freed=%d\n",
1685 generations[from_space].bytes_allocated -=
1686 nwords*N_WORD_BYTES + bytes_freed;
1687 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1688 bytes_allocated -= bytes_freed;
1693 /* Get tag of object. */
1694 tag = lowtag_of(object);
1696 /* Allocate space. */
1697 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1699 /* Copy the object. */
1700 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1702 /* Return Lisp pointer of new object. */
1703 return ((lispobj) new) | tag;
1712 * code and code-related objects
1715 static lispobj trans_fun_header(lispobj object);
1716 static lispobj trans_boxed(lispobj object);
1719 /* Scan a x86 compiled code object, looking for possible fixups that
1720 * have been missed after a move.
1722 * Two types of fixups are needed:
1723 * 1. Absolute fixups to within the code object.
1724 * 2. Relative fixups to outside the code object.
1726 * Currently only absolute fixups to the constant vector, or to the
1727 * code area are checked. */
1729 sniff_code_object(struct code *code, unsigned long displacement)
1731 #ifdef LISP_FEATURE_X86
1732 long nheader_words, ncode_words, nwords;
1734 void *constants_start_addr = NULL, *constants_end_addr;
1735 void *code_start_addr, *code_end_addr;
1736 int fixup_found = 0;
1738 if (!check_code_fixups)
1741 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1743 ncode_words = fixnum_value(code->code_size);
1744 nheader_words = HeaderValue(*(lispobj *)code);
1745 nwords = ncode_words + nheader_words;
1747 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1748 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1749 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1750 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1752 /* Work through the unboxed code. */
1753 for (p = code_start_addr; p < code_end_addr; p++) {
1754 void *data = *(void **)p;
1755 unsigned d1 = *((unsigned char *)p - 1);
1756 unsigned d2 = *((unsigned char *)p - 2);
1757 unsigned d3 = *((unsigned char *)p - 3);
1758 unsigned d4 = *((unsigned char *)p - 4);
1760 unsigned d5 = *((unsigned char *)p - 5);
1761 unsigned d6 = *((unsigned char *)p - 6);
1764 /* Check for code references. */
1765 /* Check for a 32 bit word that looks like an absolute
1766 reference to within the code adea of the code object. */
1767 if ((data >= (code_start_addr-displacement))
1768 && (data < (code_end_addr-displacement))) {
1769 /* function header */
1771 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1773 /* Skip the function header */
1777 /* the case of PUSH imm32 */
1781 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1782 p, d6, d5, d4, d3, d2, d1, data));
1783 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1785 /* the case of MOV [reg-8],imm32 */
1787 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1788 || d2==0x45 || d2==0x46 || d2==0x47)
1792 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1793 p, d6, d5, d4, d3, d2, d1, data));
1794 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1796 /* the case of LEA reg,[disp32] */
1797 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1800 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1801 p, d6, d5, d4, d3, d2, d1, data));
1802 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1806 /* Check for constant references. */
1807 /* Check for a 32 bit word that looks like an absolute
1808 reference to within the constant vector. Constant references
1810 if ((data >= (constants_start_addr-displacement))
1811 && (data < (constants_end_addr-displacement))
1812 && (((unsigned)data & 0x3) == 0)) {
1817 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1818 p, d6, d5, d4, d3, d2, d1, data));
1819 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1822 /* the case of MOV m32,EAX */
1826 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1827 p, d6, d5, d4, d3, d2, d1, data));
1828 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1831 /* the case of CMP m32,imm32 */
1832 if ((d1 == 0x3d) && (d2 == 0x81)) {
1835 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1836 p, d6, d5, d4, d3, d2, d1, data));
1838 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1841 /* Check for a mod=00, r/m=101 byte. */
1842 if ((d1 & 0xc7) == 5) {
1847 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1848 p, d6, d5, d4, d3, d2, d1, data));
1849 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1851 /* the case of CMP reg32,m32 */
1855 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1856 p, d6, d5, d4, d3, d2, d1, data));
1857 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1859 /* the case of MOV m32,reg32 */
1863 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1864 p, d6, d5, d4, d3, d2, d1, data));
1865 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1867 /* the case of MOV reg32,m32 */
1871 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1872 p, d6, d5, d4, d3, d2, d1, data));
1873 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1875 /* the case of LEA reg32,m32 */
1879 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1880 p, d6, d5, d4, d3, d2, d1, data));
1881 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1887 /* If anything was found, print some information on the code
1891 "/compiled code object at %x: header words = %d, code words = %d\n",
1892 code, nheader_words, ncode_words));
1894 "/const start = %x, end = %x\n",
1895 constants_start_addr, constants_end_addr));
1897 "/code start = %x, end = %x\n",
1898 code_start_addr, code_end_addr));
1904 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1906 /* x86-64 uses pc-relative addressing instead of this kludge */
1907 #ifndef LISP_FEATURE_X86_64
1908 long nheader_words, ncode_words, nwords;
1909 void *constants_start_addr, *constants_end_addr;
1910 void *code_start_addr, *code_end_addr;
1911 lispobj fixups = NIL;
1912 unsigned long displacement =
1913 (unsigned long)new_code - (unsigned long)old_code;
1914 struct vector *fixups_vector;
1916 ncode_words = fixnum_value(new_code->code_size);
1917 nheader_words = HeaderValue(*(lispobj *)new_code);
1918 nwords = ncode_words + nheader_words;
1920 "/compiled code object at %x: header words = %d, code words = %d\n",
1921 new_code, nheader_words, ncode_words)); */
1922 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1923 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1924 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1925 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1928 "/const start = %x, end = %x\n",
1929 constants_start_addr,constants_end_addr));
1931 "/code start = %x; end = %x\n",
1932 code_start_addr,code_end_addr));
1935 /* The first constant should be a pointer to the fixups for this
1936 code objects. Check. */
1937 fixups = new_code->constants[0];
1939 /* It will be 0 or the unbound-marker if there are no fixups (as
1940 * will be the case if the code object has been purified, for
1941 * example) and will be an other pointer if it is valid. */
1942 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1943 !is_lisp_pointer(fixups)) {
1944 /* Check for possible errors. */
1945 if (check_code_fixups)
1946 sniff_code_object(new_code, displacement);
1951 fixups_vector = (struct vector *)native_pointer(fixups);
1953 /* Could be pointing to a forwarding pointer. */
1954 /* FIXME is this always in from_space? if so, could replace this code with
1955 * forwarding_pointer_p/forwarding_pointer_value */
1956 if (is_lisp_pointer(fixups) &&
1957 (find_page_index((void*)fixups_vector) != -1) &&
1958 (fixups_vector->header == 0x01)) {
1959 /* If so, then follow it. */
1960 /*SHOW("following pointer to a forwarding pointer");*/
1962 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1965 /*SHOW("got fixups");*/
1967 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1968 /* Got the fixups for the code block. Now work through the vector,
1969 and apply a fixup at each address. */
1970 long length = fixnum_value(fixups_vector->length);
1972 for (i = 0; i < length; i++) {
1973 unsigned long offset = fixups_vector->data[i];
1974 /* Now check the current value of offset. */
1975 unsigned long old_value =
1976 *(unsigned long *)((unsigned long)code_start_addr + offset);
1978 /* If it's within the old_code object then it must be an
1979 * absolute fixup (relative ones are not saved) */
1980 if ((old_value >= (unsigned long)old_code)
1981 && (old_value < ((unsigned long)old_code
1982 + nwords*N_WORD_BYTES)))
1983 /* So add the dispacement. */
1984 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1985 old_value + displacement;
1987 /* It is outside the old code object so it must be a
1988 * relative fixup (absolute fixups are not saved). So
1989 * subtract the displacement. */
1990 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1991 old_value - displacement;
1994 /* This used to just print a note to stderr, but a bogus fixup seems to
1995 * indicate real heap corruption, so a hard hailure is in order. */
1996 lose("fixup vector %p has a bad widetag: %d\n",
1997 fixups_vector, widetag_of(fixups_vector->header));
2000 /* Check for possible errors. */
2001 if (check_code_fixups) {
2002 sniff_code_object(new_code,displacement);
2009 trans_boxed_large(lispobj object)
2012 unsigned long length;
2014 gc_assert(is_lisp_pointer(object));
2016 header = *((lispobj *) native_pointer(object));
2017 length = HeaderValue(header) + 1;
2018 length = CEILING(length, 2);
2020 return copy_large_object(object, length);
2023 /* Doesn't seem to be used, delete it after the grace period. */
2026 trans_unboxed_large(lispobj object)
2029 unsigned long length;
2031 gc_assert(is_lisp_pointer(object));
2033 header = *((lispobj *) native_pointer(object));
2034 length = HeaderValue(header) + 1;
2035 length = CEILING(length, 2);
2037 return copy_large_unboxed_object(object, length);
2043 * Lutexes. Using the normal finalization machinery for finalizing
2044 * lutexes is tricky, since the finalization depends on working lutexes.
2045 * So we track the lutexes in the GC and finalize them manually.
2048 #if defined(LUTEX_WIDETAG)
2051 * Start tracking LUTEX in the GC, by adding it to the linked list of
2052 * lutexes in the nursery generation. The caller is responsible for
2053 * locking, and GCs must be inhibited until the registration is
2057 gencgc_register_lutex (struct lutex *lutex) {
2058 int index = find_page_index(lutex);
2059 generation_index_t gen;
2062 /* This lutex is in static space, so we don't need to worry about
2068 gen = page_table[index].gen;
2070 gc_assert(gen >= 0);
2071 gc_assert(gen < NUM_GENERATIONS);
2073 head = generations[gen].lutexes;
2080 generations[gen].lutexes = lutex;
2084 * Stop tracking LUTEX in the GC by removing it from the appropriate
2085 * linked lists. This will only be called during GC, so no locking is
2089 gencgc_unregister_lutex (struct lutex *lutex) {
2091 lutex->prev->next = lutex->next;
2093 generations[lutex->gen].lutexes = lutex->next;
2097 lutex->next->prev = lutex->prev;
2106 * Mark all lutexes in generation GEN as not live.
2109 unmark_lutexes (generation_index_t gen) {
2110 struct lutex *lutex = generations[gen].lutexes;
2114 lutex = lutex->next;
2119 * Finalize all lutexes in generation GEN that have not been marked live.
2122 reap_lutexes (generation_index_t gen) {
2123 struct lutex *lutex = generations[gen].lutexes;
2126 struct lutex *next = lutex->next;
2128 lutex_destroy((tagged_lutex_t) lutex);
2129 gencgc_unregister_lutex(lutex);
2136 * Mark LUTEX as live.
2139 mark_lutex (lispobj tagged_lutex) {
2140 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
2146 * Move all lutexes in generation FROM to generation TO.
2149 move_lutexes (generation_index_t from, generation_index_t to) {
2150 struct lutex *tail = generations[from].lutexes;
2152 /* Nothing to move */
2156 /* Change the generation of the lutexes in FROM. */
2157 while (tail->next) {
2163 /* Link the last lutex in the FROM list to the start of the TO list */
2164 tail->next = generations[to].lutexes;
2166 /* And vice versa */
2167 if (generations[to].lutexes) {
2168 generations[to].lutexes->prev = tail;
2171 /* And update the generations structures to match this */
2172 generations[to].lutexes = generations[from].lutexes;
2173 generations[from].lutexes = NULL;
2177 scav_lutex(lispobj *where, lispobj object)
2179 mark_lutex((lispobj) where);
2181 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2185 trans_lutex(lispobj object)
2187 struct lutex *lutex = (struct lutex *) native_pointer(object);
2189 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2190 gc_assert(is_lisp_pointer(object));
2191 copied = copy_object(object, words);
2193 /* Update the links, since the lutex moved in memory. */
2195 lutex->next->prev = (struct lutex *) native_pointer(copied);
2199 lutex->prev->next = (struct lutex *) native_pointer(copied);
2201 generations[lutex->gen].lutexes =
2202 (struct lutex *) native_pointer(copied);
2209 size_lutex(lispobj *where)
2211 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2213 #endif /* LUTEX_WIDETAG */
2220 /* XX This is a hack adapted from cgc.c. These don't work too
2221 * efficiently with the gencgc as a list of the weak pointers is
2222 * maintained within the objects which causes writes to the pages. A
2223 * limited attempt is made to avoid unnecessary writes, but this needs
2225 #define WEAK_POINTER_NWORDS \
2226 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2229 scav_weak_pointer(lispobj *where, lispobj object)
2231 /* Since we overwrite the 'next' field, we have to make
2232 * sure not to do so for pointers already in the list.
2233 * Instead of searching the list of weak_pointers each
2234 * time, we ensure that next is always NULL when the weak
2235 * pointer isn't in the list, and not NULL otherwise.
2236 * Since we can't use NULL to denote end of list, we
2237 * use a pointer back to the same weak_pointer.
2239 struct weak_pointer * wp = (struct weak_pointer*)where;
2241 if (NULL == wp->next) {
2242 wp->next = weak_pointers;
2244 if (NULL == wp->next)
2248 /* Do not let GC scavenge the value slot of the weak pointer.
2249 * (That is why it is a weak pointer.) */
2251 return WEAK_POINTER_NWORDS;
2256 search_read_only_space(void *pointer)
2258 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2259 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2260 if ((pointer < (void *)start) || (pointer >= (void *)end))
2262 return (gc_search_space(start,
2263 (((lispobj *)pointer)+2)-start,
2264 (lispobj *) pointer));
2268 search_static_space(void *pointer)
2270 lispobj *start = (lispobj *)STATIC_SPACE_START;
2271 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2272 if ((pointer < (void *)start) || (pointer >= (void *)end))
2274 return (gc_search_space(start,
2275 (((lispobj *)pointer)+2)-start,
2276 (lispobj *) pointer));
2279 /* a faster version for searching the dynamic space. This will work even
2280 * if the object is in a current allocation region. */
2282 search_dynamic_space(void *pointer)
2284 page_index_t page_index = find_page_index(pointer);
2287 /* The address may be invalid, so do some checks. */
2288 if ((page_index == -1) || page_free_p(page_index))
2290 start = (lispobj *)page_region_start(page_index);
2291 return (gc_search_space(start,
2292 (((lispobj *)pointer)+2)-start,
2293 (lispobj *)pointer));
2296 /* Helper for valid_lisp_pointer_p and
2297 * possibly_valid_dynamic_space_pointer.
2299 * pointer is the pointer to validate, and start_addr is the address
2300 * of the enclosing object.
2303 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2305 if (!is_lisp_pointer((lispobj)pointer)) {
2309 /* Check that the object pointed to is consistent with the pointer
2311 switch (lowtag_of((lispobj)pointer)) {
2312 case FUN_POINTER_LOWTAG:
2313 /* Start_addr should be the enclosing code object, or a closure
2315 switch (widetag_of(*start_addr)) {
2316 case CODE_HEADER_WIDETAG:
2317 /* Make sure we actually point to a function in the code object,
2318 * as opposed to a random point there. */
2319 if (SIMPLE_FUN_HEADER_WIDETAG==widetag_of(*(pointer-FUN_POINTER_LOWTAG)))
2323 case CLOSURE_HEADER_WIDETAG:
2324 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2325 if ((unsigned long)pointer !=
2326 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2327 if (gencgc_verbose) {
2330 pointer, start_addr, *start_addr));
2336 if (gencgc_verbose) {
2339 pointer, start_addr, *start_addr));
2344 case LIST_POINTER_LOWTAG:
2345 if ((unsigned long)pointer !=
2346 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2347 if (gencgc_verbose) {
2350 pointer, start_addr, *start_addr));
2354 /* Is it plausible cons? */
2355 if ((is_lisp_pointer(start_addr[0]) ||
2356 is_lisp_immediate(start_addr[0])) &&
2357 (is_lisp_pointer(start_addr[1]) ||
2358 is_lisp_immediate(start_addr[1])))
2361 if (gencgc_verbose) {
2364 pointer, start_addr, *start_addr));
2368 case INSTANCE_POINTER_LOWTAG:
2369 if ((unsigned long)pointer !=
2370 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2371 if (gencgc_verbose) {
2374 pointer, start_addr, *start_addr));
2378 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2379 if (gencgc_verbose) {
2382 pointer, start_addr, *start_addr));
2387 case OTHER_POINTER_LOWTAG:
2389 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
2390 /* The all-architecture test below is good as far as it goes,
2391 * but an LRA object is similar to a FUN-POINTER: It is
2392 * embedded within a CODE-OBJECT pointed to by start_addr, and
2393 * cannot be found by simply walking the heap, therefore we
2394 * need to check for it. -- AB, 2010-Jun-04 */
2395 if ((widetag_of(start_addr[0]) == CODE_HEADER_WIDETAG)) {
2396 lispobj *potential_lra =
2397 (lispobj *)(((unsigned long)pointer) - OTHER_POINTER_LOWTAG);
2398 if ((widetag_of(potential_lra[0]) == RETURN_PC_HEADER_WIDETAG) &&
2399 ((potential_lra - HeaderValue(potential_lra[0])) == start_addr)) {
2400 return 1; /* It's as good as we can verify. */
2405 if ((unsigned long)pointer !=
2406 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2407 if (gencgc_verbose) {
2410 pointer, start_addr, *start_addr));
2414 /* Is it plausible? Not a cons. XXX should check the headers. */
2415 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2416 if (gencgc_verbose) {
2419 pointer, start_addr, *start_addr));
2423 switch (widetag_of(start_addr[0])) {
2424 case UNBOUND_MARKER_WIDETAG:
2425 case NO_TLS_VALUE_MARKER_WIDETAG:
2426 case CHARACTER_WIDETAG:
2427 #if N_WORD_BITS == 64
2428 case SINGLE_FLOAT_WIDETAG:
2430 if (gencgc_verbose) {
2433 pointer, start_addr, *start_addr));
2437 /* only pointed to by function pointers? */
2438 case CLOSURE_HEADER_WIDETAG:
2439 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2440 if (gencgc_verbose) {
2443 pointer, start_addr, *start_addr));
2447 case INSTANCE_HEADER_WIDETAG:
2448 if (gencgc_verbose) {
2451 pointer, start_addr, *start_addr));
2455 /* the valid other immediate pointer objects */
2456 case SIMPLE_VECTOR_WIDETAG:
2458 case COMPLEX_WIDETAG:
2459 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2460 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2462 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2463 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2465 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2466 case COMPLEX_LONG_FLOAT_WIDETAG:
2468 case SIMPLE_ARRAY_WIDETAG:
2469 case COMPLEX_BASE_STRING_WIDETAG:
2470 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2471 case COMPLEX_CHARACTER_STRING_WIDETAG:
2473 case COMPLEX_VECTOR_NIL_WIDETAG:
2474 case COMPLEX_BIT_VECTOR_WIDETAG:
2475 case COMPLEX_VECTOR_WIDETAG:
2476 case COMPLEX_ARRAY_WIDETAG:
2477 case VALUE_CELL_HEADER_WIDETAG:
2478 case SYMBOL_HEADER_WIDETAG:
2480 case CODE_HEADER_WIDETAG:
2481 case BIGNUM_WIDETAG:
2482 #if N_WORD_BITS != 64
2483 case SINGLE_FLOAT_WIDETAG:
2485 case DOUBLE_FLOAT_WIDETAG:
2486 #ifdef LONG_FLOAT_WIDETAG
2487 case LONG_FLOAT_WIDETAG:
2489 case SIMPLE_BASE_STRING_WIDETAG:
2490 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2491 case SIMPLE_CHARACTER_STRING_WIDETAG:
2493 case SIMPLE_BIT_VECTOR_WIDETAG:
2494 case SIMPLE_ARRAY_NIL_WIDETAG:
2495 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2496 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2497 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2498 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2499 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2500 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2502 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2504 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2505 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2506 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2507 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2509 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2510 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2512 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2513 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2515 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2516 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2519 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2521 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2522 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2524 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2525 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2527 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2528 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2529 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2530 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2532 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2533 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2535 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2536 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2538 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2539 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2542 case WEAK_POINTER_WIDETAG:
2543 #ifdef LUTEX_WIDETAG
2549 if (gencgc_verbose) {
2552 pointer, start_addr, *start_addr));
2558 if (gencgc_verbose) {
2561 pointer, start_addr, *start_addr));
2570 /* Used by the debugger to validate possibly bogus pointers before
2571 * calling MAKE-LISP-OBJ on them.
2573 * FIXME: We would like to make this perfect, because if the debugger
2574 * constructs a reference to a bugs lisp object, and it ends up in a
2575 * location scavenged by the GC all hell breaks loose.
2577 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2578 * and return true for all valid pointers, this could actually be eager
2579 * and lie about a few pointers without bad results... but that should
2580 * be reflected in the name.
2583 valid_lisp_pointer_p(lispobj *pointer)
2586 if (((start=search_dynamic_space(pointer))!=NULL) ||
2587 ((start=search_static_space(pointer))!=NULL) ||
2588 ((start=search_read_only_space(pointer))!=NULL))
2589 return looks_like_valid_lisp_pointer_p(pointer, start);
2594 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2596 /* Is there any possibility that pointer is a valid Lisp object
2597 * reference, and/or something else (e.g. subroutine call return
2598 * address) which should prevent us from moving the referred-to thing?
2599 * This is called from preserve_pointers() */
2601 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2603 lispobj *start_addr;
2605 /* Find the object start address. */
2606 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2610 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2613 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2615 /* Adjust large bignum and vector objects. This will adjust the
2616 * allocated region if the size has shrunk, and move unboxed objects
2617 * into unboxed pages. The pages are not promoted here, and the
2618 * promoted region is not added to the new_regions; this is really
2619 * only designed to be called from preserve_pointer(). Shouldn't fail
2620 * if this is missed, just may delay the moving of objects to unboxed
2621 * pages, and the freeing of pages. */
2623 maybe_adjust_large_object(lispobj *where)
2625 page_index_t first_page;
2626 page_index_t next_page;
2629 unsigned long remaining_bytes;
2630 unsigned long bytes_freed;
2631 unsigned long old_bytes_used;
2635 /* Check whether it's a vector or bignum object. */
2636 switch (widetag_of(where[0])) {
2637 case SIMPLE_VECTOR_WIDETAG:
2638 boxed = BOXED_PAGE_FLAG;
2640 case BIGNUM_WIDETAG:
2641 case SIMPLE_BASE_STRING_WIDETAG:
2642 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2643 case SIMPLE_CHARACTER_STRING_WIDETAG:
2645 case SIMPLE_BIT_VECTOR_WIDETAG:
2646 case SIMPLE_ARRAY_NIL_WIDETAG:
2647 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2648 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2649 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2650 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2651 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2652 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2654 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2656 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2657 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2658 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2659 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2661 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2662 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2664 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2665 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2667 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2668 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2671 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2673 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2674 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2676 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2677 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2679 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2680 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2681 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2682 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2684 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2685 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2687 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2688 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2690 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2691 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2693 boxed = UNBOXED_PAGE_FLAG;
2699 /* Find its current size. */
2700 nwords = (sizetab[widetag_of(where[0])])(where);
2702 first_page = find_page_index((void *)where);
2703 gc_assert(first_page >= 0);
2705 /* Note: Any page write-protection must be removed, else a later
2706 * scavenge_newspace may incorrectly not scavenge these pages.
2707 * This would not be necessary if they are added to the new areas,
2708 * but lets do it for them all (they'll probably be written
2711 gc_assert(page_table[first_page].region_start_offset == 0);
2713 next_page = first_page;
2714 remaining_bytes = nwords*N_WORD_BYTES;
2715 while (remaining_bytes > GENCGC_CARD_BYTES) {
2716 gc_assert(page_table[next_page].gen == from_space);
2717 gc_assert(page_allocated_no_region_p(next_page));
2718 gc_assert(page_table[next_page].large_object);
2719 gc_assert(page_table[next_page].region_start_offset ==
2720 npage_bytes(next_page-first_page));
2721 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2723 page_table[next_page].allocated = boxed;
2725 /* Shouldn't be write-protected at this stage. Essential that the
2727 gc_assert(!page_table[next_page].write_protected);
2728 remaining_bytes -= GENCGC_CARD_BYTES;
2732 /* Now only one page remains, but the object may have shrunk so
2733 * there may be more unused pages which will be freed. */
2735 /* Object may have shrunk but shouldn't have grown - check. */
2736 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2738 page_table[next_page].allocated = boxed;
2739 gc_assert(page_table[next_page].allocated ==
2740 page_table[first_page].allocated);
2742 /* Adjust the bytes_used. */
2743 old_bytes_used = page_table[next_page].bytes_used;
2744 page_table[next_page].bytes_used = remaining_bytes;
2746 bytes_freed = old_bytes_used - remaining_bytes;
2748 /* Free any remaining pages; needs care. */
2750 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2751 (page_table[next_page].gen == from_space) &&
2752 page_allocated_no_region_p(next_page) &&
2753 page_table[next_page].large_object &&
2754 (page_table[next_page].region_start_offset ==
2755 npage_bytes(next_page - first_page))) {
2756 /* It checks out OK, free the page. We don't need to both zeroing
2757 * pages as this should have been done before shrinking the
2758 * object. These pages shouldn't be write protected as they
2759 * should be zero filled. */
2760 gc_assert(page_table[next_page].write_protected == 0);
2762 old_bytes_used = page_table[next_page].bytes_used;
2763 page_table[next_page].allocated = FREE_PAGE_FLAG;
2764 page_table[next_page].bytes_used = 0;
2765 bytes_freed += old_bytes_used;
2769 if ((bytes_freed > 0) && gencgc_verbose) {
2771 "/maybe_adjust_large_object() freed %d\n",
2775 generations[from_space].bytes_allocated -= bytes_freed;
2776 bytes_allocated -= bytes_freed;
2781 /* Take a possible pointer to a Lisp object and mark its page in the
2782 * page_table so that it will not be relocated during a GC.
2784 * This involves locating the page it points to, then backing up to
2785 * the start of its region, then marking all pages dont_move from there
2786 * up to the first page that's not full or has a different generation
2788 * It is assumed that all the page static flags have been cleared at
2789 * the start of a GC.
2791 * It is also assumed that the current gc_alloc() region has been
2792 * flushed and the tables updated. */
2795 preserve_pointer(void *addr)
2797 page_index_t addr_page_index = find_page_index(addr);
2798 page_index_t first_page;
2800 unsigned int region_allocation;
2802 /* quick check 1: Address is quite likely to have been invalid. */
2803 if ((addr_page_index == -1)
2804 || page_free_p(addr_page_index)
2805 || (page_table[addr_page_index].bytes_used == 0)
2806 || (page_table[addr_page_index].gen != from_space)
2807 /* Skip if already marked dont_move. */
2808 || (page_table[addr_page_index].dont_move != 0))
2810 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2811 /* (Now that we know that addr_page_index is in range, it's
2812 * safe to index into page_table[] with it.) */
2813 region_allocation = page_table[addr_page_index].allocated;
2815 /* quick check 2: Check the offset within the page.
2818 if (((unsigned long)addr & (GENCGC_CARD_BYTES - 1)) >
2819 page_table[addr_page_index].bytes_used)
2822 /* Filter out anything which can't be a pointer to a Lisp object
2823 * (or, as a special case which also requires dont_move, a return
2824 * address referring to something in a CodeObject). This is
2825 * expensive but important, since it vastly reduces the
2826 * probability that random garbage will be bogusly interpreted as
2827 * a pointer which prevents a page from moving.
2829 * This only needs to happen on x86oids, where this is used for
2830 * conservative roots. Non-x86oid systems only ever call this
2831 * function on known-valid lisp objects. */
2832 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2833 if (!(code_page_p(addr_page_index)
2834 || (is_lisp_pointer((lispobj)addr) &&
2835 possibly_valid_dynamic_space_pointer(addr))))
2839 /* Find the beginning of the region. Note that there may be
2840 * objects in the region preceding the one that we were passed a
2841 * pointer to: if this is the case, we will write-protect all the
2842 * previous objects' pages too. */
2845 /* I think this'd work just as well, but without the assertions.
2846 * -dan 2004.01.01 */
2847 first_page = find_page_index(page_region_start(addr_page_index))
2849 first_page = addr_page_index;
2850 while (page_table[first_page].region_start_offset != 0) {
2852 /* Do some checks. */
2853 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2854 gc_assert(page_table[first_page].gen == from_space);
2855 gc_assert(page_table[first_page].allocated == region_allocation);
2859 /* Adjust any large objects before promotion as they won't be
2860 * copied after promotion. */
2861 if (page_table[first_page].large_object) {
2862 maybe_adjust_large_object(page_address(first_page));
2863 /* If a large object has shrunk then addr may now point to a
2864 * free area in which case it's ignored here. Note it gets
2865 * through the valid pointer test above because the tail looks
2867 if (page_free_p(addr_page_index)
2868 || (page_table[addr_page_index].bytes_used == 0)
2869 /* Check the offset within the page. */
2870 || (((unsigned long)addr & (GENCGC_CARD_BYTES - 1))
2871 > page_table[addr_page_index].bytes_used)) {
2873 "weird? ignore ptr 0x%x to freed area of large object\n",
2877 /* It may have moved to unboxed pages. */
2878 region_allocation = page_table[first_page].allocated;
2881 /* Now work forward until the end of this contiguous area is found,
2882 * marking all pages as dont_move. */
2883 for (i = first_page; ;i++) {
2884 gc_assert(page_table[i].allocated == region_allocation);
2886 /* Mark the page static. */
2887 page_table[i].dont_move = 1;
2889 /* Move the page to the new_space. XX I'd rather not do this
2890 * but the GC logic is not quite able to copy with the static
2891 * pages remaining in the from space. This also requires the
2892 * generation bytes_allocated counters be updated. */
2893 page_table[i].gen = new_space;
2894 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2895 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2897 /* It is essential that the pages are not write protected as
2898 * they may have pointers into the old-space which need
2899 * scavenging. They shouldn't be write protected at this
2901 gc_assert(!page_table[i].write_protected);
2903 /* Check whether this is the last page in this contiguous block.. */
2904 if ((page_table[i].bytes_used < GENCGC_CARD_BYTES)
2905 /* ..or it is CARD_BYTES and is the last in the block */
2907 || (page_table[i+1].bytes_used == 0) /* next page free */
2908 || (page_table[i+1].gen != from_space) /* diff. gen */
2909 || (page_table[i+1].region_start_offset == 0))
2913 /* Check that the page is now static. */
2914 gc_assert(page_table[addr_page_index].dont_move != 0);
2917 /* If the given page is not write-protected, then scan it for pointers
2918 * to younger generations or the top temp. generation, if no
2919 * suspicious pointers are found then the page is write-protected.
2921 * Care is taken to check for pointers to the current gc_alloc()
2922 * region if it is a younger generation or the temp. generation. This
2923 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2924 * the gc_alloc_generation does not need to be checked as this is only
2925 * called from scavenge_generation() when the gc_alloc generation is
2926 * younger, so it just checks if there is a pointer to the current
2929 * We return 1 if the page was write-protected, else 0. */
2931 update_page_write_prot(page_index_t page)
2933 generation_index_t gen = page_table[page].gen;
2936 void **page_addr = (void **)page_address(page);
2937 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2939 /* Shouldn't be a free page. */
2940 gc_assert(page_allocated_p(page));
2941 gc_assert(page_table[page].bytes_used != 0);
2943 /* Skip if it's already write-protected, pinned, or unboxed */
2944 if (page_table[page].write_protected
2945 /* FIXME: What's the reason for not write-protecting pinned pages? */
2946 || page_table[page].dont_move
2947 || page_unboxed_p(page))
2950 /* Scan the page for pointers to younger generations or the
2951 * top temp. generation. */
2953 for (j = 0; j < num_words; j++) {
2954 void *ptr = *(page_addr+j);
2955 page_index_t index = find_page_index(ptr);
2957 /* Check that it's in the dynamic space */
2959 if (/* Does it point to a younger or the temp. generation? */
2960 (page_allocated_p(index)
2961 && (page_table[index].bytes_used != 0)
2962 && ((page_table[index].gen < gen)
2963 || (page_table[index].gen == SCRATCH_GENERATION)))
2965 /* Or does it point within a current gc_alloc() region? */
2966 || ((boxed_region.start_addr <= ptr)
2967 && (ptr <= boxed_region.free_pointer))
2968 || ((unboxed_region.start_addr <= ptr)
2969 && (ptr <= unboxed_region.free_pointer))) {
2976 /* Write-protect the page. */
2977 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2979 os_protect((void *)page_addr,
2981 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2983 /* Note the page as protected in the page tables. */
2984 page_table[page].write_protected = 1;
2990 /* Scavenge all generations from FROM to TO, inclusive, except for
2991 * new_space which needs special handling, as new objects may be
2992 * added which are not checked here - use scavenge_newspace generation.
2994 * Write-protected pages should not have any pointers to the
2995 * from_space so do need scavenging; thus write-protected pages are
2996 * not always scavenged. There is some code to check that these pages
2997 * are not written; but to check fully the write-protected pages need
2998 * to be scavenged by disabling the code to skip them.
3000 * Under the current scheme when a generation is GCed the younger
3001 * generations will be empty. So, when a generation is being GCed it
3002 * is only necessary to scavenge the older generations for pointers
3003 * not the younger. So a page that does not have pointers to younger
3004 * generations does not need to be scavenged.
3006 * The write-protection can be used to note pages that don't have
3007 * pointers to younger pages. But pages can be written without having
3008 * pointers to younger generations. After the pages are scavenged here
3009 * they can be scanned for pointers to younger generations and if
3010 * there are none the page can be write-protected.
3012 * One complication is when the newspace is the top temp. generation.
3014 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
3015 * that none were written, which they shouldn't be as they should have
3016 * no pointers to younger generations. This breaks down for weak
3017 * pointers as the objects contain a link to the next and are written
3018 * if a weak pointer is scavenged. Still it's a useful check. */
3020 scavenge_generations(generation_index_t from, generation_index_t to)
3027 /* Clear the write_protected_cleared flags on all pages. */
3028 for (i = 0; i < page_table_pages; i++)
3029 page_table[i].write_protected_cleared = 0;
3032 for (i = 0; i < last_free_page; i++) {
3033 generation_index_t generation = page_table[i].gen;
3035 && (page_table[i].bytes_used != 0)
3036 && (generation != new_space)
3037 && (generation >= from)
3038 && (generation <= to)) {
3039 page_index_t last_page,j;
3040 int write_protected=1;
3042 /* This should be the start of a region */
3043 gc_assert(page_table[i].region_start_offset == 0);
3045 /* Now work forward until the end of the region */
3046 for (last_page = i; ; last_page++) {
3048 write_protected && page_table[last_page].write_protected;
3049 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3050 /* Or it is CARD_BYTES and is the last in the block */
3051 || (!page_boxed_p(last_page+1))
3052 || (page_table[last_page+1].bytes_used == 0)
3053 || (page_table[last_page+1].gen != generation)
3054 || (page_table[last_page+1].region_start_offset == 0))
3057 if (!write_protected) {
3058 scavenge(page_address(i),
3059 ((unsigned long)(page_table[last_page].bytes_used
3060 + npage_bytes(last_page-i)))
3063 /* Now scan the pages and write protect those that
3064 * don't have pointers to younger generations. */
3065 if (enable_page_protection) {
3066 for (j = i; j <= last_page; j++) {
3067 num_wp += update_page_write_prot(j);
3070 if ((gencgc_verbose > 1) && (num_wp != 0)) {
3072 "/write protected %d pages within generation %d\n",
3073 num_wp, generation));
3081 /* Check that none of the write_protected pages in this generation
3082 * have been written to. */
3083 for (i = 0; i < page_table_pages; i++) {
3084 if (page_allocated_p(i)
3085 && (page_table[i].bytes_used != 0)
3086 && (page_table[i].gen == generation)
3087 && (page_table[i].write_protected_cleared != 0)) {
3088 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
3090 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
3091 page_table[i].bytes_used,
3092 page_table[i].region_start_offset,
3093 page_table[i].dont_move));
3094 lose("write to protected page %d in scavenge_generation()\n", i);
3101 /* Scavenge a newspace generation. As it is scavenged new objects may
3102 * be allocated to it; these will also need to be scavenged. This
3103 * repeats until there are no more objects unscavenged in the
3104 * newspace generation.
3106 * To help improve the efficiency, areas written are recorded by
3107 * gc_alloc() and only these scavenged. Sometimes a little more will be
3108 * scavenged, but this causes no harm. An easy check is done that the
3109 * scavenged bytes equals the number allocated in the previous
3112 * Write-protected pages are not scanned except if they are marked
3113 * dont_move in which case they may have been promoted and still have
3114 * pointers to the from space.
3116 * Write-protected pages could potentially be written by alloc however
3117 * to avoid having to handle re-scavenging of write-protected pages
3118 * gc_alloc() does not write to write-protected pages.
3120 * New areas of objects allocated are recorded alternatively in the two
3121 * new_areas arrays below. */
3122 static struct new_area new_areas_1[NUM_NEW_AREAS];
3123 static struct new_area new_areas_2[NUM_NEW_AREAS];
3125 /* Do one full scan of the new space generation. This is not enough to
3126 * complete the job as new objects may be added to the generation in
3127 * the process which are not scavenged. */
3129 scavenge_newspace_generation_one_scan(generation_index_t generation)
3134 "/starting one full scan of newspace generation %d\n",
3136 for (i = 0; i < last_free_page; i++) {
3137 /* Note that this skips over open regions when it encounters them. */
3139 && (page_table[i].bytes_used != 0)
3140 && (page_table[i].gen == generation)
3141 && ((page_table[i].write_protected == 0)
3142 /* (This may be redundant as write_protected is now
3143 * cleared before promotion.) */
3144 || (page_table[i].dont_move == 1))) {
3145 page_index_t last_page;
3148 /* The scavenge will start at the region_start_offset of
3151 * We need to find the full extent of this contiguous
3152 * block in case objects span pages.
3154 * Now work forward until the end of this contiguous area
3155 * is found. A small area is preferred as there is a
3156 * better chance of its pages being write-protected. */
3157 for (last_page = i; ;last_page++) {
3158 /* If all pages are write-protected and movable,
3159 * then no need to scavenge */
3160 all_wp=all_wp && page_table[last_page].write_protected &&
3161 !page_table[last_page].dont_move;
3163 /* Check whether this is the last page in this
3164 * contiguous block */
3165 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3166 /* Or it is CARD_BYTES and is the last in the block */
3167 || (!page_boxed_p(last_page+1))
3168 || (page_table[last_page+1].bytes_used == 0)
3169 || (page_table[last_page+1].gen != generation)
3170 || (page_table[last_page+1].region_start_offset == 0))
3174 /* Do a limited check for write-protected pages. */
3176 long nwords = (((unsigned long)
3177 (page_table[last_page].bytes_used
3178 + npage_bytes(last_page-i)
3179 + page_table[i].region_start_offset))
3181 new_areas_ignore_page = last_page;
3183 scavenge(page_region_start(i), nwords);
3190 "/done with one full scan of newspace generation %d\n",
3194 /* Do a complete scavenge of the newspace generation. */
3196 scavenge_newspace_generation(generation_index_t generation)
3200 /* the new_areas array currently being written to by gc_alloc() */
3201 struct new_area (*current_new_areas)[] = &new_areas_1;
3202 long current_new_areas_index;
3204 /* the new_areas created by the previous scavenge cycle */
3205 struct new_area (*previous_new_areas)[] = NULL;
3206 long previous_new_areas_index;
3208 /* Flush the current regions updating the tables. */
3209 gc_alloc_update_all_page_tables();
3211 /* Turn on the recording of new areas by gc_alloc(). */
3212 new_areas = current_new_areas;
3213 new_areas_index = 0;
3215 /* Don't need to record new areas that get scavenged anyway during
3216 * scavenge_newspace_generation_one_scan. */
3217 record_new_objects = 1;
3219 /* Start with a full scavenge. */
3220 scavenge_newspace_generation_one_scan(generation);
3222 /* Record all new areas now. */
3223 record_new_objects = 2;
3225 /* Give a chance to weak hash tables to make other objects live.
3226 * FIXME: The algorithm implemented here for weak hash table gcing
3227 * is O(W^2+N) as Bruno Haible warns in
3228 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3229 * see "Implementation 2". */
3230 scav_weak_hash_tables();
3232 /* Flush the current regions updating the tables. */
3233 gc_alloc_update_all_page_tables();
3235 /* Grab new_areas_index. */
3236 current_new_areas_index = new_areas_index;
3239 "The first scan is finished; current_new_areas_index=%d.\n",
3240 current_new_areas_index));*/
3242 while (current_new_areas_index > 0) {
3243 /* Move the current to the previous new areas */
3244 previous_new_areas = current_new_areas;
3245 previous_new_areas_index = current_new_areas_index;
3247 /* Scavenge all the areas in previous new areas. Any new areas
3248 * allocated are saved in current_new_areas. */
3250 /* Allocate an array for current_new_areas; alternating between
3251 * new_areas_1 and 2 */
3252 if (previous_new_areas == &new_areas_1)
3253 current_new_areas = &new_areas_2;
3255 current_new_areas = &new_areas_1;
3257 /* Set up for gc_alloc(). */
3258 new_areas = current_new_areas;
3259 new_areas_index = 0;
3261 /* Check whether previous_new_areas had overflowed. */
3262 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3264 /* New areas of objects allocated have been lost so need to do a
3265 * full scan to be sure! If this becomes a problem try
3266 * increasing NUM_NEW_AREAS. */
3267 if (gencgc_verbose) {
3268 SHOW("new_areas overflow, doing full scavenge");
3271 /* Don't need to record new areas that get scavenged
3272 * anyway during scavenge_newspace_generation_one_scan. */
3273 record_new_objects = 1;
3275 scavenge_newspace_generation_one_scan(generation);
3277 /* Record all new areas now. */
3278 record_new_objects = 2;
3280 scav_weak_hash_tables();
3282 /* Flush the current regions updating the tables. */
3283 gc_alloc_update_all_page_tables();
3287 /* Work through previous_new_areas. */
3288 for (i = 0; i < previous_new_areas_index; i++) {
3289 page_index_t page = (*previous_new_areas)[i].page;
3290 size_t offset = (*previous_new_areas)[i].offset;
3291 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3292 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3293 scavenge(page_address(page)+offset, size);
3296 scav_weak_hash_tables();
3298 /* Flush the current regions updating the tables. */
3299 gc_alloc_update_all_page_tables();
3302 current_new_areas_index = new_areas_index;
3305 "The re-scan has finished; current_new_areas_index=%d.\n",
3306 current_new_areas_index));*/
3309 /* Turn off recording of areas allocated by gc_alloc(). */
3310 record_new_objects = 0;
3313 /* Check that none of the write_protected pages in this generation
3314 * have been written to. */
3315 for (i = 0; i < page_table_pages; i++) {
3316 if (page_allocated_p(i)
3317 && (page_table[i].bytes_used != 0)
3318 && (page_table[i].gen == generation)
3319 && (page_table[i].write_protected_cleared != 0)
3320 && (page_table[i].dont_move == 0)) {
3321 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3322 i, generation, page_table[i].dont_move);
3328 /* Un-write-protect all the pages in from_space. This is done at the
3329 * start of a GC else there may be many page faults while scavenging
3330 * the newspace (I've seen drive the system time to 99%). These pages
3331 * would need to be unprotected anyway before unmapping in
3332 * free_oldspace; not sure what effect this has on paging.. */
3334 unprotect_oldspace(void)
3337 void *region_addr = 0;
3338 void *page_addr = 0;
3339 unsigned long region_bytes = 0;
3341 for (i = 0; i < last_free_page; i++) {
3342 if (page_allocated_p(i)
3343 && (page_table[i].bytes_used != 0)
3344 && (page_table[i].gen == from_space)) {
3346 /* Remove any write-protection. We should be able to rely
3347 * on the write-protect flag to avoid redundant calls. */
3348 if (page_table[i].write_protected) {
3349 page_table[i].write_protected = 0;
3350 page_addr = page_address(i);
3353 region_addr = page_addr;
3354 region_bytes = GENCGC_CARD_BYTES;
3355 } else if (region_addr + region_bytes == page_addr) {
3356 /* Region continue. */
3357 region_bytes += GENCGC_CARD_BYTES;
3359 /* Unprotect previous region. */
3360 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
3361 /* First page in new region. */
3362 region_addr = page_addr;
3363 region_bytes = GENCGC_CARD_BYTES;
3369 /* Unprotect last region. */
3370 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
3374 /* Work through all the pages and free any in from_space. This
3375 * assumes that all objects have been copied or promoted to an older
3376 * generation. Bytes_allocated and the generation bytes_allocated
3377 * counter are updated. The number of bytes freed is returned. */
3378 static unsigned long
3381 unsigned long bytes_freed = 0;
3382 page_index_t first_page, last_page;
3387 /* Find a first page for the next region of pages. */
3388 while ((first_page < last_free_page)
3389 && (page_free_p(first_page)
3390 || (page_table[first_page].bytes_used == 0)
3391 || (page_table[first_page].gen != from_space)))
3394 if (first_page >= last_free_page)
3397 /* Find the last page of this region. */
3398 last_page = first_page;
3401 /* Free the page. */
3402 bytes_freed += page_table[last_page].bytes_used;
3403 generations[page_table[last_page].gen].bytes_allocated -=
3404 page_table[last_page].bytes_used;
3405 page_table[last_page].allocated = FREE_PAGE_FLAG;
3406 page_table[last_page].bytes_used = 0;
3407 /* Should already be unprotected by unprotect_oldspace(). */
3408 gc_assert(!page_table[last_page].write_protected);
3411 while ((last_page < last_free_page)
3412 && page_allocated_p(last_page)
3413 && (page_table[last_page].bytes_used != 0)
3414 && (page_table[last_page].gen == from_space));
3416 #ifdef READ_PROTECT_FREE_PAGES
3417 os_protect(page_address(first_page),
3418 npage_bytes(last_page-first_page),
3421 first_page = last_page;
3422 } while (first_page < last_free_page);
3424 bytes_allocated -= bytes_freed;
3429 /* Print some information about a pointer at the given address. */
3431 print_ptr(lispobj *addr)
3433 /* If addr is in the dynamic space then out the page information. */
3434 page_index_t pi1 = find_page_index((void*)addr);
3437 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
3438 (unsigned long) addr,
3440 page_table[pi1].allocated,
3441 page_table[pi1].gen,
3442 page_table[pi1].bytes_used,
3443 page_table[pi1].region_start_offset,
3444 page_table[pi1].dont_move);
3445 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3459 is_in_stack_space(lispobj ptr)
3461 /* For space verification: Pointers can be valid if they point
3462 * to a thread stack space. This would be faster if the thread
3463 * structures had page-table entries as if they were part of
3464 * the heap space. */
3466 for_each_thread(th) {
3467 if ((th->control_stack_start <= (lispobj *)ptr) &&
3468 (th->control_stack_end >= (lispobj *)ptr)) {
3476 verify_space(lispobj *start, size_t words)
3478 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3479 int is_in_readonly_space =
3480 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3481 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3485 lispobj thing = *(lispobj*)start;
3487 if (is_lisp_pointer(thing)) {
3488 page_index_t page_index = find_page_index((void*)thing);
3489 long to_readonly_space =
3490 (READ_ONLY_SPACE_START <= thing &&
3491 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3492 long to_static_space =
3493 (STATIC_SPACE_START <= thing &&
3494 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3496 /* Does it point to the dynamic space? */
3497 if (page_index != -1) {
3498 /* If it's within the dynamic space it should point to a used
3499 * page. XX Could check the offset too. */
3500 if (page_allocated_p(page_index)
3501 && (page_table[page_index].bytes_used == 0))
3502 lose ("Ptr %p @ %p sees free page.\n", thing, start);
3503 /* Check that it doesn't point to a forwarding pointer! */
3504 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3505 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
3507 /* Check that its not in the RO space as it would then be a
3508 * pointer from the RO to the dynamic space. */
3509 if (is_in_readonly_space) {
3510 lose("ptr to dynamic space %p from RO space %x\n",
3513 /* Does it point to a plausible object? This check slows
3514 * it down a lot (so it's commented out).
3516 * "a lot" is serious: it ate 50 minutes cpu time on
3517 * my duron 950 before I came back from lunch and
3520 * FIXME: Add a variable to enable this
3523 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3524 lose("ptr %p to invalid object %p\n", thing, start);
3528 extern void funcallable_instance_tramp;
3529 /* Verify that it points to another valid space. */
3530 if (!to_readonly_space && !to_static_space
3531 && (thing != (lispobj)&funcallable_instance_tramp)
3532 && !is_in_stack_space(thing)) {
3533 lose("Ptr %p @ %p sees junk.\n", thing, start);
3537 if (!(fixnump(thing))) {
3539 switch(widetag_of(*start)) {
3542 case SIMPLE_VECTOR_WIDETAG:
3544 case COMPLEX_WIDETAG:
3545 case SIMPLE_ARRAY_WIDETAG:
3546 case COMPLEX_BASE_STRING_WIDETAG:
3547 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3548 case COMPLEX_CHARACTER_STRING_WIDETAG:
3550 case COMPLEX_VECTOR_NIL_WIDETAG:
3551 case COMPLEX_BIT_VECTOR_WIDETAG:
3552 case COMPLEX_VECTOR_WIDETAG:
3553 case COMPLEX_ARRAY_WIDETAG:
3554 case CLOSURE_HEADER_WIDETAG:
3555 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3556 case VALUE_CELL_HEADER_WIDETAG:
3557 case SYMBOL_HEADER_WIDETAG:
3558 case CHARACTER_WIDETAG:
3559 #if N_WORD_BITS == 64
3560 case SINGLE_FLOAT_WIDETAG:
3562 case UNBOUND_MARKER_WIDETAG:
3567 case INSTANCE_HEADER_WIDETAG:
3570 long ntotal = HeaderValue(thing);
3571 lispobj layout = ((struct instance *)start)->slots[0];
3576 nuntagged = ((struct layout *)
3577 native_pointer(layout))->n_untagged_slots;
3578 verify_space(start + 1,
3579 ntotal - fixnum_value(nuntagged));
3583 case CODE_HEADER_WIDETAG:
3585 lispobj object = *start;
3587 long nheader_words, ncode_words, nwords;
3589 struct simple_fun *fheaderp;
3591 code = (struct code *) start;
3593 /* Check that it's not in the dynamic space.
3594 * FIXME: Isn't is supposed to be OK for code
3595 * objects to be in the dynamic space these days? */
3596 if (is_in_dynamic_space
3597 /* It's ok if it's byte compiled code. The trace
3598 * table offset will be a fixnum if it's x86
3599 * compiled code - check.
3601 * FIXME: #^#@@! lack of abstraction here..
3602 * This line can probably go away now that
3603 * there's no byte compiler, but I've got
3604 * too much to worry about right now to try
3605 * to make sure. -- WHN 2001-10-06 */
3606 && fixnump(code->trace_table_offset)
3607 /* Only when enabled */
3608 && verify_dynamic_code_check) {
3610 "/code object at %p in the dynamic space\n",
3614 ncode_words = fixnum_value(code->code_size);
3615 nheader_words = HeaderValue(object);
3616 nwords = ncode_words + nheader_words;
3617 nwords = CEILING(nwords, 2);
3618 /* Scavenge the boxed section of the code data block */
3619 verify_space(start + 1, nheader_words - 1);
3621 /* Scavenge the boxed section of each function
3622 * object in the code data block. */
3623 fheaderl = code->entry_points;
3624 while (fheaderl != NIL) {
3626 (struct simple_fun *) native_pointer(fheaderl);
3627 gc_assert(widetag_of(fheaderp->header) ==
3628 SIMPLE_FUN_HEADER_WIDETAG);
3629 verify_space(&fheaderp->name, 1);
3630 verify_space(&fheaderp->arglist, 1);
3631 verify_space(&fheaderp->type, 1);
3632 fheaderl = fheaderp->next;
3638 /* unboxed objects */
3639 case BIGNUM_WIDETAG:
3640 #if N_WORD_BITS != 64
3641 case SINGLE_FLOAT_WIDETAG:
3643 case DOUBLE_FLOAT_WIDETAG:
3644 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3645 case LONG_FLOAT_WIDETAG:
3647 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3648 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3650 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3651 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3653 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3654 case COMPLEX_LONG_FLOAT_WIDETAG:
3656 case SIMPLE_BASE_STRING_WIDETAG:
3657 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3658 case SIMPLE_CHARACTER_STRING_WIDETAG:
3660 case SIMPLE_BIT_VECTOR_WIDETAG:
3661 case SIMPLE_ARRAY_NIL_WIDETAG:
3662 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3663 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3664 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3665 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3666 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3667 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3669 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3671 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3672 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3673 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3674 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3676 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3677 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3679 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3680 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3682 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3683 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3686 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3688 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3689 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3691 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3692 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3694 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3695 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3696 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3697 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3699 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3700 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3702 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3703 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3705 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3706 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3709 case WEAK_POINTER_WIDETAG:
3710 #ifdef LUTEX_WIDETAG
3713 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3714 case NO_TLS_VALUE_MARKER_WIDETAG:
3716 count = (sizetab[widetag_of(*start)])(start);
3720 lose("Unhandled widetag %p at %p\n",
3721 widetag_of(*start), start);
3733 /* FIXME: It would be nice to make names consistent so that
3734 * foo_size meant size *in* *bytes* instead of size in some
3735 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3736 * Some counts of lispobjs are called foo_count; it might be good
3737 * to grep for all foo_size and rename the appropriate ones to
3739 long read_only_space_size =
3740 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3741 - (lispobj*)READ_ONLY_SPACE_START;
3742 long static_space_size =
3743 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3744 - (lispobj*)STATIC_SPACE_START;
3746 for_each_thread(th) {
3747 long binding_stack_size =
3748 (lispobj*)get_binding_stack_pointer(th)
3749 - (lispobj*)th->binding_stack_start;
3750 verify_space(th->binding_stack_start, binding_stack_size);
3752 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3753 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3757 verify_generation(generation_index_t generation)
3761 for (i = 0; i < last_free_page; i++) {
3762 if (page_allocated_p(i)
3763 && (page_table[i].bytes_used != 0)
3764 && (page_table[i].gen == generation)) {
3765 page_index_t last_page;
3766 int region_allocation = page_table[i].allocated;
3768 /* This should be the start of a contiguous block */
3769 gc_assert(page_table[i].region_start_offset == 0);
3771 /* Need to find the full extent of this contiguous block in case
3772 objects span pages. */
3774 /* Now work forward until the end of this contiguous area is
3776 for (last_page = i; ;last_page++)
3777 /* Check whether this is the last page in this contiguous
3779 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3780 /* Or it is CARD_BYTES and is the last in the block */
3781 || (page_table[last_page+1].allocated != region_allocation)
3782 || (page_table[last_page+1].bytes_used == 0)
3783 || (page_table[last_page+1].gen != generation)
3784 || (page_table[last_page+1].region_start_offset == 0))
3787 verify_space(page_address(i),
3789 (page_table[last_page].bytes_used
3790 + npage_bytes(last_page-i)))
3797 /* Check that all the free space is zero filled. */
3799 verify_zero_fill(void)
3803 for (page = 0; page < last_free_page; page++) {
3804 if (page_free_p(page)) {
3805 /* The whole page should be zero filled. */
3806 long *start_addr = (long *)page_address(page);
3809 for (i = 0; i < size; i++) {
3810 if (start_addr[i] != 0) {
3811 lose("free page not zero at %x\n", start_addr + i);
3815 long free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3816 if (free_bytes > 0) {
3817 long *start_addr = (long *)((unsigned long)page_address(page)
3818 + page_table[page].bytes_used);
3819 long size = free_bytes / N_WORD_BYTES;
3821 for (i = 0; i < size; i++) {
3822 if (start_addr[i] != 0) {
3823 lose("free region not zero at %x\n", start_addr + i);
3831 /* External entry point for verify_zero_fill */
3833 gencgc_verify_zero_fill(void)
3835 /* Flush the alloc regions updating the tables. */
3836 gc_alloc_update_all_page_tables();
3837 SHOW("verifying zero fill");
3842 verify_dynamic_space(void)
3844 generation_index_t i;
3846 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3847 verify_generation(i);
3849 if (gencgc_enable_verify_zero_fill)
3853 /* Write-protect all the dynamic boxed pages in the given generation. */
3855 write_protect_generation_pages(generation_index_t generation)
3859 gc_assert(generation < SCRATCH_GENERATION);
3861 for (start = 0; start < last_free_page; start++) {
3862 if (protect_page_p(start, generation)) {
3866 /* Note the page as protected in the page tables. */
3867 page_table[start].write_protected = 1;
3869 for (last = start + 1; last < last_free_page; last++) {
3870 if (!protect_page_p(last, generation))
3872 page_table[last].write_protected = 1;
3875 page_start = (void *)page_address(start);
3877 os_protect(page_start,
3878 npage_bytes(last - start),
3879 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3885 if (gencgc_verbose > 1) {
3887 "/write protected %d of %d pages in generation %d\n",
3888 count_write_protect_generation_pages(generation),
3889 count_generation_pages(generation),
3894 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3896 scavenge_control_stack(struct thread *th)
3898 lispobj *control_stack =
3899 (lispobj *)(th->control_stack_start);
3900 unsigned long control_stack_size =
3901 access_control_stack_pointer(th) - control_stack;
3903 scavenge(control_stack, control_stack_size);
3907 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3909 preserve_context_registers (os_context_t *c)
3912 /* On Darwin the signal context isn't a contiguous block of memory,
3913 * so just preserve_pointering its contents won't be sufficient.
3915 #if defined(LISP_FEATURE_DARWIN)
3916 #if defined LISP_FEATURE_X86
3917 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3918 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3919 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3920 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3921 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3922 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3923 preserve_pointer((void*)*os_context_pc_addr(c));
3924 #elif defined LISP_FEATURE_X86_64
3925 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3926 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3927 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3928 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3929 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3930 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3931 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3932 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3933 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3934 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3935 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3936 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3937 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3938 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3939 preserve_pointer((void*)*os_context_pc_addr(c));
3941 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3944 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3945 preserve_pointer(*ptr);
3950 /* Garbage collect a generation. If raise is 0 then the remains of the
3951 * generation are not raised to the next generation. */
3953 garbage_collect_generation(generation_index_t generation, int raise)
3955 unsigned long bytes_freed;
3957 unsigned long static_space_size;
3960 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3962 /* The oldest generation can't be raised. */
3963 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3965 /* Check if weak hash tables were processed in the previous GC. */
3966 gc_assert(weak_hash_tables == NULL);
3968 /* Initialize the weak pointer list. */
3969 weak_pointers = NULL;
3971 #ifdef LUTEX_WIDETAG
3972 unmark_lutexes(generation);
3975 /* When a generation is not being raised it is transported to a
3976 * temporary generation (NUM_GENERATIONS), and lowered when
3977 * done. Set up this new generation. There should be no pages
3978 * allocated to it yet. */
3980 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3983 /* Set the global src and dest. generations */
3984 from_space = generation;
3986 new_space = generation+1;
3988 new_space = SCRATCH_GENERATION;
3990 /* Change to a new space for allocation, resetting the alloc_start_page */
3991 gc_alloc_generation = new_space;
3992 generations[new_space].alloc_start_page = 0;
3993 generations[new_space].alloc_unboxed_start_page = 0;
3994 generations[new_space].alloc_large_start_page = 0;
3995 generations[new_space].alloc_large_unboxed_start_page = 0;
3997 /* Before any pointers are preserved, the dont_move flags on the
3998 * pages need to be cleared. */
3999 for (i = 0; i < last_free_page; i++)
4000 if(page_table[i].gen==from_space)
4001 page_table[i].dont_move = 0;
4003 /* Un-write-protect the old-space pages. This is essential for the
4004 * promoted pages as they may contain pointers into the old-space
4005 * which need to be scavenged. It also helps avoid unnecessary page
4006 * faults as forwarding pointers are written into them. They need to
4007 * be un-protected anyway before unmapping later. */
4008 unprotect_oldspace();
4010 /* Scavenge the stacks' conservative roots. */
4012 /* there are potentially two stacks for each thread: the main
4013 * stack, which may contain Lisp pointers, and the alternate stack.
4014 * We don't ever run Lisp code on the altstack, but it may
4015 * host a sigcontext with lisp objects in it */
4017 /* what we need to do: (1) find the stack pointer for the main
4018 * stack; scavenge it (2) find the interrupt context on the
4019 * alternate stack that might contain lisp values, and scavenge
4022 /* we assume that none of the preceding applies to the thread that
4023 * initiates GC. If you ever call GC from inside an altstack
4024 * handler, you will lose. */
4026 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4027 /* And if we're saving a core, there's no point in being conservative. */
4028 if (conservative_stack) {
4029 for_each_thread(th) {
4031 void **esp=(void **)-1;
4032 #ifdef LISP_FEATURE_SB_THREAD
4034 if(th==arch_os_get_current_thread()) {
4035 /* Somebody is going to burn in hell for this, but casting
4036 * it in two steps shuts gcc up about strict aliasing. */
4037 esp = (void **)((void *)&raise);
4040 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
4041 for(i=free-1;i>=0;i--) {
4042 os_context_t *c=th->interrupt_contexts[i];
4043 esp1 = (void **) *os_context_register_addr(c,reg_SP);
4044 if (esp1>=(void **)th->control_stack_start &&
4045 esp1<(void **)th->control_stack_end) {
4046 if(esp1<esp) esp=esp1;
4047 preserve_context_registers(c);
4052 esp = (void **)((void *)&raise);
4054 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
4055 preserve_pointer(*ptr);
4060 /* Non-x86oid systems don't have "conservative roots" as such, but
4061 * the same mechanism is used for objects pinned for use by alien
4063 for_each_thread(th) {
4064 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
4065 while (pin_list != NIL) {
4066 struct cons *list_entry =
4067 (struct cons *)native_pointer(pin_list);
4068 preserve_pointer(list_entry->car);
4069 pin_list = list_entry->cdr;
4075 if (gencgc_verbose > 1) {
4076 long num_dont_move_pages = count_dont_move_pages();
4078 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
4079 num_dont_move_pages,
4080 npage_bytes(num_dont_move_pages));
4084 /* Scavenge all the rest of the roots. */
4086 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4088 * If not x86, we need to scavenge the interrupt context(s) and the
4093 for_each_thread(th) {
4094 scavenge_interrupt_contexts(th);
4095 scavenge_control_stack(th);
4098 /* Scrub the unscavenged control stack space, so that we can't run
4099 * into any stale pointers in a later GC (this is done by the
4100 * stop-for-gc handler in the other threads). */
4101 scrub_control_stack();
4105 /* Scavenge the Lisp functions of the interrupt handlers, taking
4106 * care to avoid SIG_DFL and SIG_IGN. */
4107 for (i = 0; i < NSIG; i++) {
4108 union interrupt_handler handler = interrupt_handlers[i];
4109 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4110 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4111 scavenge((lispobj *)(interrupt_handlers + i), 1);
4114 /* Scavenge the binding stacks. */
4117 for_each_thread(th) {
4118 long len= (lispobj *)get_binding_stack_pointer(th) -
4119 th->binding_stack_start;
4120 scavenge((lispobj *) th->binding_stack_start,len);
4121 #ifdef LISP_FEATURE_SB_THREAD
4122 /* do the tls as well */
4123 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
4124 (sizeof (struct thread))/(sizeof (lispobj));
4125 scavenge((lispobj *) (th+1),len);
4130 /* The original CMU CL code had scavenge-read-only-space code
4131 * controlled by the Lisp-level variable
4132 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4133 * wasn't documented under what circumstances it was useful or
4134 * safe to turn it on, so it's been turned off in SBCL. If you
4135 * want/need this functionality, and can test and document it,
4136 * please submit a patch. */
4138 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4139 unsigned long read_only_space_size =
4140 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4141 (lispobj*)READ_ONLY_SPACE_START;
4143 "/scavenge read only space: %d bytes\n",
4144 read_only_space_size * sizeof(lispobj)));
4145 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4149 /* Scavenge static space. */
4151 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4152 (lispobj *)STATIC_SPACE_START;
4153 if (gencgc_verbose > 1) {
4155 "/scavenge static space: %d bytes\n",
4156 static_space_size * sizeof(lispobj)));
4158 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4160 /* All generations but the generation being GCed need to be
4161 * scavenged. The new_space generation needs special handling as
4162 * objects may be moved in - it is handled separately below. */
4163 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4165 /* Finally scavenge the new_space generation. Keep going until no
4166 * more objects are moved into the new generation */
4167 scavenge_newspace_generation(new_space);
4169 /* FIXME: I tried reenabling this check when debugging unrelated
4170 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4171 * Since the current GC code seems to work well, I'm guessing that
4172 * this debugging code is just stale, but I haven't tried to
4173 * figure it out. It should be figured out and then either made to
4174 * work or just deleted. */
4175 #define RESCAN_CHECK 0
4177 /* As a check re-scavenge the newspace once; no new objects should
4180 long old_bytes_allocated = bytes_allocated;
4181 long bytes_allocated;
4183 /* Start with a full scavenge. */
4184 scavenge_newspace_generation_one_scan(new_space);
4186 /* Flush the current regions, updating the tables. */
4187 gc_alloc_update_all_page_tables();
4189 bytes_allocated = bytes_allocated - old_bytes_allocated;
4191 if (bytes_allocated != 0) {
4192 lose("Rescan of new_space allocated %d more bytes.\n",
4198 scan_weak_hash_tables();
4199 scan_weak_pointers();
4201 /* Flush the current regions, updating the tables. */
4202 gc_alloc_update_all_page_tables();
4204 /* Free the pages in oldspace, but not those marked dont_move. */
4205 bytes_freed = free_oldspace();
4207 /* If the GC is not raising the age then lower the generation back
4208 * to its normal generation number */
4210 for (i = 0; i < last_free_page; i++)
4211 if ((page_table[i].bytes_used != 0)
4212 && (page_table[i].gen == SCRATCH_GENERATION))
4213 page_table[i].gen = generation;
4214 gc_assert(generations[generation].bytes_allocated == 0);
4215 generations[generation].bytes_allocated =
4216 generations[SCRATCH_GENERATION].bytes_allocated;
4217 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4220 /* Reset the alloc_start_page for generation. */
4221 generations[generation].alloc_start_page = 0;
4222 generations[generation].alloc_unboxed_start_page = 0;
4223 generations[generation].alloc_large_start_page = 0;
4224 generations[generation].alloc_large_unboxed_start_page = 0;
4226 if (generation >= verify_gens) {
4227 if (gencgc_verbose) {
4231 verify_dynamic_space();
4234 /* Set the new gc trigger for the GCed generation. */
4235 generations[generation].gc_trigger =
4236 generations[generation].bytes_allocated
4237 + generations[generation].bytes_consed_between_gc;
4240 generations[generation].num_gc = 0;
4242 ++generations[generation].num_gc;
4244 #ifdef LUTEX_WIDETAG
4245 reap_lutexes(generation);
4247 move_lutexes(generation, generation+1);
4251 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4253 update_dynamic_space_free_pointer(void)
4255 page_index_t last_page = -1, i;
4257 for (i = 0; i < last_free_page; i++)
4258 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
4261 last_free_page = last_page+1;
4263 set_alloc_pointer((lispobj)(page_address(last_free_page)));
4264 return 0; /* dummy value: return something ... */
4268 remap_page_range (page_index_t from, page_index_t to)
4270 /* There's a mysterious Solaris/x86 problem with using mmap
4271 * tricks for memory zeroing. See sbcl-devel thread
4272 * "Re: patch: standalone executable redux".
4274 #if defined(LISP_FEATURE_SUNOS)
4275 zero_and_mark_pages(from, to);
4278 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
4279 release_mask = release_granularity-1,
4281 aligned_from = (from+release_mask)&~release_mask,
4282 aligned_end = (end&~release_mask);
4284 if (aligned_from < aligned_end) {
4285 zero_pages_with_mmap(aligned_from, aligned_end-1);
4286 if (aligned_from != from)
4287 zero_and_mark_pages(from, aligned_from-1);
4288 if (aligned_end != end)
4289 zero_and_mark_pages(aligned_end, end-1);
4291 zero_and_mark_pages(from, to);
4297 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
4299 page_index_t first_page, last_page;
4302 return remap_page_range(from, to);
4304 for (first_page = from; first_page <= to; first_page++) {
4305 if (page_allocated_p(first_page) ||
4306 (page_table[first_page].need_to_zero == 0))
4309 last_page = first_page + 1;
4310 while (page_free_p(last_page) &&
4311 (last_page <= to) &&
4312 (page_table[last_page].need_to_zero == 1))
4315 remap_page_range(first_page, last_page-1);
4317 first_page = last_page;
4321 generation_index_t small_generation_limit = 1;
4323 /* GC all generations newer than last_gen, raising the objects in each
4324 * to the next older generation - we finish when all generations below
4325 * last_gen are empty. Then if last_gen is due for a GC, or if
4326 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4327 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4329 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4330 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4332 collect_garbage(generation_index_t last_gen)
4334 generation_index_t gen = 0, i;
4337 /* The largest value of last_free_page seen since the time
4338 * remap_free_pages was called. */
4339 static page_index_t high_water_mark = 0;
4341 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4342 log_generation_stats(gc_logfile, "=== GC Start ===");
4346 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4348 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4353 /* Flush the alloc regions updating the tables. */
4354 gc_alloc_update_all_page_tables();
4356 /* Verify the new objects created by Lisp code. */
4357 if (pre_verify_gen_0) {
4358 FSHOW((stderr, "pre-checking generation 0\n"));
4359 verify_generation(0);
4362 if (gencgc_verbose > 1)
4363 print_generation_stats();
4366 /* Collect the generation. */
4368 if (gen >= gencgc_oldest_gen_to_gc) {
4369 /* Never raise the oldest generation. */
4374 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
4377 if (gencgc_verbose > 1) {
4379 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4382 generations[gen].bytes_allocated,
4383 generations[gen].gc_trigger,
4384 generations[gen].num_gc));
4387 /* If an older generation is being filled, then update its
4390 generations[gen+1].cum_sum_bytes_allocated +=
4391 generations[gen+1].bytes_allocated;
4394 garbage_collect_generation(gen, raise);
4396 /* Reset the memory age cum_sum. */
4397 generations[gen].cum_sum_bytes_allocated = 0;
4399 if (gencgc_verbose > 1) {
4400 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4401 print_generation_stats();
4405 } while ((gen <= gencgc_oldest_gen_to_gc)
4406 && ((gen < last_gen)
4407 || ((gen <= gencgc_oldest_gen_to_gc)
4409 && (generations[gen].bytes_allocated
4410 > generations[gen].gc_trigger)
4411 && (generation_average_age(gen)
4412 > generations[gen].minimum_age_before_gc))));
4414 /* Now if gen-1 was raised all generations before gen are empty.
4415 * If it wasn't raised then all generations before gen-1 are empty.
4417 * Now objects within this gen's pages cannot point to younger
4418 * generations unless they are written to. This can be exploited
4419 * by write-protecting the pages of gen; then when younger
4420 * generations are GCed only the pages which have been written
4425 gen_to_wp = gen - 1;
4427 /* There's not much point in WPing pages in generation 0 as it is
4428 * never scavenged (except promoted pages). */
4429 if ((gen_to_wp > 0) && enable_page_protection) {
4430 /* Check that they are all empty. */
4431 for (i = 0; i < gen_to_wp; i++) {
4432 if (generations[i].bytes_allocated)
4433 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4436 write_protect_generation_pages(gen_to_wp);
4439 /* Set gc_alloc() back to generation 0. The current regions should
4440 * be flushed after the above GCs. */
4441 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4442 gc_alloc_generation = 0;
4444 /* Save the high-water mark before updating last_free_page */
4445 if (last_free_page > high_water_mark)
4446 high_water_mark = last_free_page;
4448 update_dynamic_space_free_pointer();
4450 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4452 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4455 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4458 if (gen > small_generation_limit) {
4459 if (last_free_page > high_water_mark)
4460 high_water_mark = last_free_page;
4461 remap_free_pages(0, high_water_mark, 0);
4462 high_water_mark = 0;
4467 log_generation_stats(gc_logfile, "=== GC End ===");
4468 SHOW("returning from collect_garbage");
4471 /* This is called by Lisp PURIFY when it is finished. All live objects
4472 * will have been moved to the RO and Static heaps. The dynamic space
4473 * will need a full re-initialization. We don't bother having Lisp
4474 * PURIFY flush the current gc_alloc() region, as the page_tables are
4475 * re-initialized, and every page is zeroed to be sure. */
4479 page_index_t page, last_page;
4481 if (gencgc_verbose > 1) {
4482 SHOW("entering gc_free_heap");
4485 for (page = 0; page < page_table_pages; page++) {
4486 /* Skip free pages which should already be zero filled. */
4487 if (page_allocated_p(page)) {
4488 void *page_start, *addr;
4489 for (last_page = page;
4490 (last_page < page_table_pages) && page_allocated_p(last_page);
4492 /* Mark the page free. The other slots are assumed invalid
4493 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4494 * should not be write-protected -- except that the
4495 * generation is used for the current region but it sets
4497 page_table[page].allocated = FREE_PAGE_FLAG;
4498 page_table[page].bytes_used = 0;
4499 page_table[page].write_protected = 0;
4502 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4503 * about this change. */
4504 page_start = (void *)page_address(page);
4505 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
4506 remap_free_pages(page, last_page-1, 1);
4509 } else if (gencgc_zero_check_during_free_heap) {
4510 /* Double-check that the page is zero filled. */
4513 gc_assert(page_free_p(page));
4514 gc_assert(page_table[page].bytes_used == 0);
4515 page_start = (long *)page_address(page);
4516 for (i=0; i<GENCGC_CARD_BYTES/sizeof(long); i++) {
4517 if (page_start[i] != 0) {
4518 lose("free region not zero at %x\n", page_start + i);
4524 bytes_allocated = 0;
4526 /* Initialize the generations. */
4527 for (page = 0; page < NUM_GENERATIONS; page++) {
4528 generations[page].alloc_start_page = 0;
4529 generations[page].alloc_unboxed_start_page = 0;
4530 generations[page].alloc_large_start_page = 0;
4531 generations[page].alloc_large_unboxed_start_page = 0;
4532 generations[page].bytes_allocated = 0;
4533 generations[page].gc_trigger = 2000000;
4534 generations[page].num_gc = 0;
4535 generations[page].cum_sum_bytes_allocated = 0;
4536 generations[page].lutexes = NULL;
4539 if (gencgc_verbose > 1)
4540 print_generation_stats();
4542 /* Initialize gc_alloc(). */
4543 gc_alloc_generation = 0;
4545 gc_set_region_empty(&boxed_region);
4546 gc_set_region_empty(&unboxed_region);
4549 set_alloc_pointer((lispobj)((char *)heap_base));
4551 if (verify_after_free_heap) {
4552 /* Check whether purify has left any bad pointers. */
4553 FSHOW((stderr, "checking after free_heap\n"));
4563 /* Compute the number of pages needed for the dynamic space.
4564 * Dynamic space size should be aligned on page size. */
4565 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4566 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4568 /* The page_table must be allocated using "calloc" to initialize
4569 * the page structures correctly. There used to be a separate
4570 * initialization loop (now commented out; see below) but that was
4571 * unnecessary and did hurt startup time. */
4572 page_table = calloc(page_table_pages, sizeof(struct page));
4573 gc_assert(page_table);
4576 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4577 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4579 #ifdef LUTEX_WIDETAG
4580 scavtab[LUTEX_WIDETAG] = scav_lutex;
4581 transother[LUTEX_WIDETAG] = trans_lutex;
4582 sizetab[LUTEX_WIDETAG] = size_lutex;
4585 heap_base = (void*)DYNAMIC_SPACE_START;
4587 /* The page structures are initialized implicitly when page_table
4588 * is allocated with "calloc" above. Formerly we had the following
4589 * explicit initialization here (comments converted to C99 style
4590 * for readability as C's block comments don't nest):
4592 * // Initialize each page structure.
4593 * for (i = 0; i < page_table_pages; i++) {
4594 * // Initialize all pages as free.
4595 * page_table[i].allocated = FREE_PAGE_FLAG;
4596 * page_table[i].bytes_used = 0;
4598 * // Pages are not write-protected at startup.
4599 * page_table[i].write_protected = 0;
4602 * Without this loop the image starts up much faster when dynamic
4603 * space is large -- which it is on 64-bit platforms already by
4604 * default -- and when "calloc" for large arrays is implemented
4605 * using copy-on-write of a page of zeroes -- which it is at least
4606 * on Linux. In this case the pages that page_table_pages is stored
4607 * in are mapped and cleared not before the corresponding part of
4608 * dynamic space is used. For example, this saves clearing 16 MB of
4609 * memory at startup if the page size is 4 KB and the size of
4610 * dynamic space is 4 GB.
4611 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4612 * asserted below: */
4614 /* Compile time assertion: If triggered, declares an array
4615 * of dimension -1 forcing a syntax error. The intent of the
4616 * assignment is to avoid an "unused variable" warning. */
4617 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4618 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4621 bytes_allocated = 0;
4623 /* Initialize the generations.
4625 * FIXME: very similar to code in gc_free_heap(), should be shared */
4626 for (i = 0; i < NUM_GENERATIONS; i++) {
4627 generations[i].alloc_start_page = 0;
4628 generations[i].alloc_unboxed_start_page = 0;
4629 generations[i].alloc_large_start_page = 0;
4630 generations[i].alloc_large_unboxed_start_page = 0;
4631 generations[i].bytes_allocated = 0;
4632 generations[i].gc_trigger = 2000000;
4633 generations[i].num_gc = 0;
4634 generations[i].cum_sum_bytes_allocated = 0;
4635 /* the tune-able parameters */
4636 generations[i].bytes_consed_between_gc = 2000000;
4637 generations[i].number_of_gcs_before_promotion = 1;
4638 generations[i].minimum_age_before_gc = 0.75;
4639 generations[i].lutexes = NULL;
4642 /* Initialize gc_alloc. */
4643 gc_alloc_generation = 0;
4644 gc_set_region_empty(&boxed_region);
4645 gc_set_region_empty(&unboxed_region);
4650 /* Pick up the dynamic space from after a core load.
4652 * The ALLOCATION_POINTER points to the end of the dynamic space.
4656 gencgc_pickup_dynamic(void)
4658 page_index_t page = 0;
4659 void *alloc_ptr = (void *)get_alloc_pointer();
4660 lispobj *prev=(lispobj *)page_address(page);
4661 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4663 lispobj *first,*ptr= (lispobj *)page_address(page);
4665 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4666 /* It is possible, though rare, for the saved page table
4667 * to contain free pages below alloc_ptr. */
4668 page_table[page].gen = gen;
4669 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4670 page_table[page].large_object = 0;
4671 page_table[page].write_protected = 0;
4672 page_table[page].write_protected_cleared = 0;
4673 page_table[page].dont_move = 0;
4674 page_table[page].need_to_zero = 1;
4677 if (!gencgc_partial_pickup) {
4678 page_table[page].allocated = BOXED_PAGE_FLAG;
4679 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4682 page_table[page].region_start_offset =
4683 page_address(page) - (void *)prev;
4686 } while (page_address(page) < alloc_ptr);
4688 #ifdef LUTEX_WIDETAG
4689 /* Lutexes have been registered in generation 0 by coreparse, and
4690 * need to be moved to the right one manually.
4692 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4695 last_free_page = page;
4697 generations[gen].bytes_allocated = npage_bytes(page);
4698 bytes_allocated = npage_bytes(page);
4700 gc_alloc_update_all_page_tables();
4701 write_protect_generation_pages(gen);
4705 gc_initialize_pointers(void)
4707 gencgc_pickup_dynamic();
4711 /* alloc(..) is the external interface for memory allocation. It
4712 * allocates to generation 0. It is not called from within the garbage
4713 * collector as it is only external uses that need the check for heap
4714 * size (GC trigger) and to disable the interrupts (interrupts are
4715 * always disabled during a GC).
4717 * The vops that call alloc(..) assume that the returned space is zero-filled.
4718 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4720 * The check for a GC trigger is only performed when the current
4721 * region is full, so in most cases it's not needed. */
4723 static inline lispobj *
4724 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4725 struct thread *thread)
4727 #ifndef LISP_FEATURE_WIN32
4728 lispobj alloc_signal;
4731 void *new_free_pointer;
4733 gc_assert(nbytes>0);
4735 /* Check for alignment allocation problems. */
4736 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4737 && ((nbytes & LOWTAG_MASK) == 0));
4739 /* Must be inside a PA section. */
4740 gc_assert(get_pseudo_atomic_atomic(thread));
4742 /* maybe we can do this quickly ... */
4743 new_free_pointer = region->free_pointer + nbytes;
4744 if (new_free_pointer <= region->end_addr) {
4745 new_obj = (void*)(region->free_pointer);
4746 region->free_pointer = new_free_pointer;
4747 return(new_obj); /* yup */
4750 /* we have to go the long way around, it seems. Check whether we
4751 * should GC in the near future
4753 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4754 /* Don't flood the system with interrupts if the need to gc is
4755 * already noted. This can happen for example when SUB-GC
4756 * allocates or after a gc triggered in a WITHOUT-GCING. */
4757 if (SymbolValue(GC_PENDING,thread) == NIL) {
4758 /* set things up so that GC happens when we finish the PA
4760 SetSymbolValue(GC_PENDING,T,thread);
4761 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4762 set_pseudo_atomic_interrupted(thread);
4763 #ifdef LISP_FEATURE_PPC
4764 /* PPC calls alloc() from a trap or from pa_alloc(),
4765 * look up the most context if it's from a trap. */
4767 os_context_t *context =
4768 thread->interrupt_data->allocation_trap_context;
4769 maybe_save_gc_mask_and_block_deferrables
4770 (context ? os_context_sigmask_addr(context) : NULL);
4773 maybe_save_gc_mask_and_block_deferrables(NULL);
4778 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4780 #ifndef LISP_FEATURE_WIN32
4781 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4782 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4783 if ((signed long) alloc_signal <= 0) {
4784 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4787 SetSymbolValue(ALLOC_SIGNAL,
4788 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4798 general_alloc(long nbytes, int page_type_flag)
4800 struct thread *thread = arch_os_get_current_thread();
4801 /* Select correct region, and call general_alloc_internal with it.
4802 * For other then boxed allocation we must lock first, since the
4803 * region is shared. */
4804 if (BOXED_PAGE_FLAG & page_type_flag) {
4805 #ifdef LISP_FEATURE_SB_THREAD
4806 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4808 struct alloc_region *region = &boxed_region;
4810 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4811 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4813 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4814 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4815 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4818 lose("bad page type flag: %d", page_type_flag);
4825 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4826 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4830 * shared support for the OS-dependent signal handlers which
4831 * catch GENCGC-related write-protect violations
4833 void unhandled_sigmemoryfault(void* addr);
4835 /* Depending on which OS we're running under, different signals might
4836 * be raised for a violation of write protection in the heap. This
4837 * function factors out the common generational GC magic which needs
4838 * to invoked in this case, and should be called from whatever signal
4839 * handler is appropriate for the OS we're running under.
4841 * Return true if this signal is a normal generational GC thing that
4842 * we were able to handle, or false if it was abnormal and control
4843 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4846 gencgc_handle_wp_violation(void* fault_addr)
4848 page_index_t page_index = find_page_index(fault_addr);
4851 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4852 fault_addr, page_index));
4855 /* Check whether the fault is within the dynamic space. */
4856 if (page_index == (-1)) {
4858 /* It can be helpful to be able to put a breakpoint on this
4859 * case to help diagnose low-level problems. */
4860 unhandled_sigmemoryfault(fault_addr);
4862 /* not within the dynamic space -- not our responsibility */
4867 ret = thread_mutex_lock(&free_pages_lock);
4868 gc_assert(ret == 0);
4869 if (page_table[page_index].write_protected) {
4870 /* Unprotect the page. */
4871 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4872 page_table[page_index].write_protected_cleared = 1;
4873 page_table[page_index].write_protected = 0;
4875 /* The only acceptable reason for this signal on a heap
4876 * access is that GENCGC write-protected the page.
4877 * However, if two CPUs hit a wp page near-simultaneously,
4878 * we had better not have the second one lose here if it
4879 * does this test after the first one has already set wp=0
4881 if(page_table[page_index].write_protected_cleared != 1)
4882 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4883 page_index, boxed_region.first_page,
4884 boxed_region.last_page);
4886 ret = thread_mutex_unlock(&free_pages_lock);
4887 gc_assert(ret == 0);
4888 /* Don't worry, we can handle it. */
4892 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4893 * it's not just a case of the program hitting the write barrier, and
4894 * are about to let Lisp deal with it. It's basically just a
4895 * convenient place to set a gdb breakpoint. */
4897 unhandled_sigmemoryfault(void *addr)
4900 void gc_alloc_update_all_page_tables(void)
4902 /* Flush the alloc regions updating the tables. */
4905 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4906 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4907 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4911 gc_set_region_empty(struct alloc_region *region)
4913 region->first_page = 0;
4914 region->last_page = -1;
4915 region->start_addr = page_address(0);
4916 region->free_pointer = page_address(0);
4917 region->end_addr = page_address(0);
4921 zero_all_free_pages()
4925 for (i = 0; i < last_free_page; i++) {
4926 if (page_free_p(i)) {
4927 #ifdef READ_PROTECT_FREE_PAGES
4928 os_protect(page_address(i),
4937 /* Things to do before doing a final GC before saving a core (without
4940 * + Pages in large_object pages aren't moved by the GC, so we need to
4941 * unset that flag from all pages.
4942 * + The pseudo-static generation isn't normally collected, but it seems
4943 * reasonable to collect it at least when saving a core. So move the
4944 * pages to a normal generation.
4947 prepare_for_final_gc ()
4950 for (i = 0; i < last_free_page; i++) {
4951 page_table[i].large_object = 0;
4952 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4953 int used = page_table[i].bytes_used;
4954 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4955 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4956 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4962 /* Do a non-conservative GC, and then save a core with the initial
4963 * function being set to the value of the static symbol
4964 * SB!VM:RESTART-LISP-FUNCTION */
4966 gc_and_save(char *filename, boolean prepend_runtime,
4967 boolean save_runtime_options,
4968 boolean compressed, int compression_level)
4971 void *runtime_bytes = NULL;
4972 size_t runtime_size;
4974 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4979 conservative_stack = 0;
4981 /* The filename might come from Lisp, and be moved by the now
4982 * non-conservative GC. */
4983 filename = strdup(filename);
4985 /* Collect twice: once into relatively high memory, and then back
4986 * into low memory. This compacts the retained data into the lower
4987 * pages, minimizing the size of the core file.
4989 prepare_for_final_gc();
4990 gencgc_alloc_start_page = last_free_page;
4991 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4993 prepare_for_final_gc();
4994 gencgc_alloc_start_page = -1;
4995 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4997 if (prepend_runtime)
4998 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
5000 /* The dumper doesn't know that pages need to be zeroed before use. */
5001 zero_all_free_pages();
5002 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
5003 prepend_runtime, save_runtime_options,
5004 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
5005 /* Oops. Save still managed to fail. Since we've mangled the stack
5006 * beyond hope, there's not much we can do.
5007 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
5008 * going to be rather unsatisfactory too... */
5009 lose("Attempt to save core after non-conservative GC failed.\n");