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"
59 /* forward declarations */
60 page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
68 /* Generations 0-5 are normal collected generations, 6 is only used as
69 * scratch space by the collector, and should never get collected.
72 SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
76 /* Should we use page protection to help avoid the scavenging of pages
77 * that don't have pointers to younger generations? */
78 boolean enable_page_protection = 1;
80 /* the minimum size (in bytes) for a large object*/
81 long large_object_size = 4 * PAGE_BYTES;
88 /* the verbosity level. All non-error messages are disabled at level 0;
89 * and only a few rare messages are printed at level 1. */
91 boolean gencgc_verbose = 1;
93 boolean gencgc_verbose = 0;
96 /* FIXME: At some point enable the various error-checking things below
97 * and see what they say. */
99 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
100 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
102 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
104 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
105 boolean pre_verify_gen_0 = 0;
107 /* Should we check for bad pointers after gc_free_heap is called
108 * from Lisp PURIFY? */
109 boolean verify_after_free_heap = 0;
111 /* Should we print a note when code objects are found in the dynamic space
112 * during a heap verify? */
113 boolean verify_dynamic_code_check = 0;
115 /* Should we check code objects for fixup errors after they are transported? */
116 boolean check_code_fixups = 0;
118 /* Should we check that newly allocated regions are zero filled? */
119 boolean gencgc_zero_check = 0;
121 /* Should we check that the free space is zero filled? */
122 boolean gencgc_enable_verify_zero_fill = 0;
124 /* Should we check that free pages are zero filled during gc_free_heap
125 * called after Lisp PURIFY? */
126 boolean gencgc_zero_check_during_free_heap = 0;
128 /* When loading a core, don't do a full scan of the memory for the
129 * memory region boundaries. (Set to true by coreparse.c if the core
130 * contained a pagetable entry).
132 boolean gencgc_partial_pickup = 0;
134 /* If defined, free pages are read-protected to ensure that nothing
138 /* #define READ_PROTECT_FREE_PAGES */
142 * GC structures and variables
145 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
146 unsigned long bytes_allocated = 0;
147 unsigned long auto_gc_trigger = 0;
149 /* the source and destination generations. These are set before a GC starts
151 generation_index_t from_space;
152 generation_index_t new_space;
154 /* Set to 1 when in GC */
155 boolean gc_active_p = 0;
157 /* should the GC be conservative on stack. If false (only right before
158 * saving a core), don't scan the stack / mark pages dont_move. */
159 static boolean conservative_stack = 1;
161 /* An array of page structures is allocated on gc initialization.
162 * This helps quickly map between an address its page structure.
163 * page_table_pages is set from the size of the dynamic space. */
164 page_index_t page_table_pages;
165 struct page *page_table;
167 static inline boolean page_allocated_p(page_index_t page) {
168 return (page_table[page].allocated != FREE_PAGE_FLAG);
171 static inline boolean page_no_region_p(page_index_t page) {
172 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
175 static inline boolean page_allocated_no_region_p(page_index_t page) {
176 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
177 && page_no_region_p(page));
180 static inline boolean page_free_p(page_index_t page) {
181 return (page_table[page].allocated == FREE_PAGE_FLAG);
184 static inline boolean page_boxed_p(page_index_t page) {
185 return (page_table[page].allocated & BOXED_PAGE_FLAG);
188 static inline boolean code_page_p(page_index_t page) {
189 return (page_table[page].allocated & CODE_PAGE_FLAG);
192 static inline boolean page_boxed_no_region_p(page_index_t page) {
193 return page_boxed_p(page) && page_no_region_p(page);
196 static inline boolean page_unboxed_p(page_index_t page) {
197 /* Both flags set == boxed code page */
198 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
199 && !page_boxed_p(page));
202 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
203 return (page_boxed_no_region_p(page)
204 && (page_table[page].bytes_used != 0)
205 && !page_table[page].dont_move
206 && (page_table[page].gen == generation));
209 /* To map addresses to page structures the address of the first page
211 static void *heap_base = NULL;
213 /* Calculate the start address for the given page number. */
215 page_address(page_index_t page_num)
217 return (heap_base + (page_num * PAGE_BYTES));
220 /* Calculate the address where the allocation region associated with
221 * the page starts. */
223 page_region_start(page_index_t page_index)
225 return page_address(page_index)-page_table[page_index].region_start_offset;
228 /* Find the page index within the page_table for the given
229 * address. Return -1 on failure. */
231 find_page_index(void *addr)
233 if (addr >= heap_base) {
234 page_index_t index = ((pointer_sized_uint_t)addr -
235 (pointer_sized_uint_t)heap_base) / PAGE_BYTES;
236 if (index < page_table_pages)
243 npage_bytes(long npages)
245 gc_assert(npages>=0);
246 return ((unsigned long)npages)*PAGE_BYTES;
249 /* Check that X is a higher address than Y and return offset from Y to
252 size_t void_diff(void *x, void *y)
255 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
258 /* a structure to hold the state of a generation
260 * CAUTION: If you modify this, make sure to touch up the alien
261 * definition in src/code/gc.lisp accordingly. ...or better yes,
262 * deal with the FIXME there...
266 /* the first page that gc_alloc() checks on its next call */
267 page_index_t alloc_start_page;
269 /* the first page that gc_alloc_unboxed() checks on its next call */
270 page_index_t alloc_unboxed_start_page;
272 /* the first page that gc_alloc_large (boxed) considers on its next
273 * call. (Although it always allocates after the boxed_region.) */
274 page_index_t alloc_large_start_page;
276 /* the first page that gc_alloc_large (unboxed) considers on its
277 * next call. (Although it always allocates after the
278 * current_unboxed_region.) */
279 page_index_t alloc_large_unboxed_start_page;
281 /* the bytes allocated to this generation */
282 unsigned long bytes_allocated;
284 /* the number of bytes at which to trigger a GC */
285 unsigned long gc_trigger;
287 /* to calculate a new level for gc_trigger */
288 unsigned long bytes_consed_between_gc;
290 /* the number of GCs since the last raise */
293 /* the number of GCs to run on the generations before raising objects to the
295 int number_of_gcs_before_promotion;
297 /* the cumulative sum of the bytes allocated to this generation. It is
298 * cleared after a GC on this generations, and update before new
299 * objects are added from a GC of a younger generation. Dividing by
300 * the bytes_allocated will give the average age of the memory in
301 * this generation since its last GC. */
302 unsigned long cum_sum_bytes_allocated;
304 /* a minimum average memory age before a GC will occur helps
305 * prevent a GC when a large number of new live objects have been
306 * added, in which case a GC could be a waste of time */
307 double minimum_age_before_gc;
309 /* A linked list of lutex structures in this generation, used for
310 * implementing lutex finalization. */
312 struct lutex *lutexes;
318 /* an array of generation structures. There needs to be one more
319 * generation structure than actual generations as the oldest
320 * generation is temporarily raised then lowered. */
321 struct generation generations[NUM_GENERATIONS];
323 /* the oldest generation that is will currently be GCed by default.
324 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
326 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
328 * Setting this to 0 effectively disables the generational nature of
329 * the GC. In some applications generational GC may not be useful
330 * because there are no long-lived objects.
332 * An intermediate value could be handy after moving long-lived data
333 * into an older generation so an unnecessary GC of this long-lived
334 * data can be avoided. */
335 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
337 /* The maximum free page in the heap is maintained and used to update
338 * ALLOCATION_POINTER which is used by the room function to limit its
339 * search of the heap. XX Gencgc obviously needs to be better
340 * integrated with the Lisp code. */
341 page_index_t last_free_page;
343 #ifdef LISP_FEATURE_SB_THREAD
344 /* This lock is to prevent multiple threads from simultaneously
345 * allocating new regions which overlap each other. Note that the
346 * majority of GC is single-threaded, but alloc() may be called from
347 * >1 thread at a time and must be thread-safe. This lock must be
348 * seized before all accesses to generations[] or to parts of
349 * page_table[] that other threads may want to see */
350 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
351 /* This lock is used to protect non-thread-local allocation. */
352 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
357 * miscellaneous heap functions
360 /* Count the number of pages which are write-protected within the
361 * given generation. */
363 count_write_protect_generation_pages(generation_index_t generation)
366 unsigned long count = 0;
368 for (i = 0; i < last_free_page; i++)
369 if (page_allocated_p(i)
370 && (page_table[i].gen == generation)
371 && (page_table[i].write_protected == 1))
376 /* Count the number of pages within the given generation. */
378 count_generation_pages(generation_index_t generation)
383 for (i = 0; i < last_free_page; i++)
384 if (page_allocated_p(i)
385 && (page_table[i].gen == generation))
392 count_dont_move_pages(void)
396 for (i = 0; i < last_free_page; i++) {
397 if (page_allocated_p(i)
398 && (page_table[i].dont_move != 0)) {
406 /* Work through the pages and add up the number of bytes used for the
407 * given generation. */
409 count_generation_bytes_allocated (generation_index_t gen)
412 unsigned long result = 0;
413 for (i = 0; i < last_free_page; i++) {
414 if (page_allocated_p(i)
415 && (page_table[i].gen == gen))
416 result += page_table[i].bytes_used;
421 /* Return the average age of the memory in a generation. */
423 generation_average_age(generation_index_t gen)
425 if (generations[gen].bytes_allocated == 0)
429 ((double)generations[gen].cum_sum_bytes_allocated)
430 / ((double)generations[gen].bytes_allocated);
433 /* The verbose argument controls how much to print: 0 for normal
434 * level of detail; 1 for debugging. */
436 print_generation_stats() /* FIXME: should take FILE argument, or construct a string */
438 generation_index_t i;
440 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
441 #define FPU_STATE_SIZE 27
442 int fpu_state[FPU_STATE_SIZE];
443 #elif defined(LISP_FEATURE_PPC)
444 #define FPU_STATE_SIZE 32
445 long long fpu_state[FPU_STATE_SIZE];
448 /* This code uses the FP instructions which may be set up for Lisp
449 * so they need to be saved and reset for C. */
452 /* Print the heap stats. */
454 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
456 for (i = 0; i < SCRATCH_GENERATION; i++) {
459 long unboxed_cnt = 0;
460 long large_boxed_cnt = 0;
461 long large_unboxed_cnt = 0;
464 for (j = 0; j < last_free_page; j++)
465 if (page_table[j].gen == i) {
467 /* Count the number of boxed pages within the given
469 if (page_boxed_p(j)) {
470 if (page_table[j].large_object)
475 if(page_table[j].dont_move) pinned_cnt++;
476 /* Count the number of unboxed pages within the given
478 if (page_unboxed_p(j)) {
479 if (page_table[j].large_object)
486 gc_assert(generations[i].bytes_allocated
487 == count_generation_bytes_allocated(i));
489 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
491 generations[i].alloc_start_page,
492 generations[i].alloc_unboxed_start_page,
493 generations[i].alloc_large_start_page,
494 generations[i].alloc_large_unboxed_start_page,
500 generations[i].bytes_allocated,
501 (npage_bytes(count_generation_pages(i))
502 - generations[i].bytes_allocated),
503 generations[i].gc_trigger,
504 count_write_protect_generation_pages(i),
505 generations[i].num_gc,
506 generation_average_age(i));
508 fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated);
509 fprintf(stderr," Dynamic-space-size bytes = %u\n", dynamic_space_size);
511 fpu_restore(fpu_state);
515 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
516 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
519 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
520 * if zeroing it ourselves, i.e. in practice give the memory back to the
521 * OS. Generally done after a large GC.
523 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
525 void *addr = page_address(start), *new_addr;
526 size_t length = npage_bytes(1+end-start);
531 os_invalidate(addr, length);
532 new_addr = os_validate(addr, length);
533 if (new_addr == NULL || new_addr != addr) {
534 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
538 for (i = start; i <= end; i++) {
539 page_table[i].need_to_zero = 0;
543 /* Zero the pages from START to END (inclusive). Generally done just after
544 * a new region has been allocated.
547 zero_pages(page_index_t start, page_index_t end) {
551 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
552 fast_bzero(page_address(start), npage_bytes(1+end-start));
554 bzero(page_address(start), npage_bytes(1+end-start));
559 /* Zero the pages from START to END (inclusive), except for those
560 * pages that are known to already zeroed. Mark all pages in the
561 * ranges as non-zeroed.
564 zero_dirty_pages(page_index_t start, page_index_t end) {
567 for (i = start; i <= end; i++) {
568 if (page_table[i].need_to_zero == 1) {
569 zero_pages(start, end);
574 for (i = start; i <= end; i++) {
575 page_table[i].need_to_zero = 1;
581 * To support quick and inline allocation, regions of memory can be
582 * allocated and then allocated from with just a free pointer and a
583 * check against an end address.
585 * Since objects can be allocated to spaces with different properties
586 * e.g. boxed/unboxed, generation, ages; there may need to be many
587 * allocation regions.
589 * Each allocation region may start within a partly used page. Many
590 * features of memory use are noted on a page wise basis, e.g. the
591 * generation; so if a region starts within an existing allocated page
592 * it must be consistent with this page.
594 * During the scavenging of the newspace, objects will be transported
595 * into an allocation region, and pointers updated to point to this
596 * allocation region. It is possible that these pointers will be
597 * scavenged again before the allocation region is closed, e.g. due to
598 * trans_list which jumps all over the place to cleanup the list. It
599 * is important to be able to determine properties of all objects
600 * pointed to when scavenging, e.g to detect pointers to the oldspace.
601 * Thus it's important that the allocation regions have the correct
602 * properties set when allocated, and not just set when closed. The
603 * region allocation routines return regions with the specified
604 * properties, and grab all the pages, setting their properties
605 * appropriately, except that the amount used is not known.
607 * These regions are used to support quicker allocation using just a
608 * free pointer. The actual space used by the region is not reflected
609 * in the pages tables until it is closed. It can't be scavenged until
612 * When finished with the region it should be closed, which will
613 * update the page tables for the actual space used returning unused
614 * space. Further it may be noted in the new regions which is
615 * necessary when scavenging the newspace.
617 * Large objects may be allocated directly without an allocation
618 * region, the page tables are updated immediately.
620 * Unboxed objects don't contain pointers to other objects and so
621 * don't need scavenging. Further they can't contain pointers to
622 * younger generations so WP is not needed. By allocating pages to
623 * unboxed objects the whole page never needs scavenging or
624 * write-protecting. */
626 /* We are only using two regions at present. Both are for the current
627 * newspace generation. */
628 struct alloc_region boxed_region;
629 struct alloc_region unboxed_region;
631 /* The generation currently being allocated to. */
632 static generation_index_t gc_alloc_generation;
634 static inline page_index_t
635 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
638 if (UNBOXED_PAGE_FLAG == page_type_flag) {
639 return generations[generation].alloc_large_unboxed_start_page;
640 } else if (BOXED_PAGE_FLAG & page_type_flag) {
641 /* Both code and data. */
642 return generations[generation].alloc_large_start_page;
644 lose("bad page type flag: %d", page_type_flag);
647 if (UNBOXED_PAGE_FLAG == page_type_flag) {
648 return generations[generation].alloc_unboxed_start_page;
649 } else if (BOXED_PAGE_FLAG & page_type_flag) {
650 /* Both code and data. */
651 return generations[generation].alloc_start_page;
653 lose("bad page_type_flag: %d", page_type_flag);
659 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
663 if (UNBOXED_PAGE_FLAG == page_type_flag) {
664 generations[generation].alloc_large_unboxed_start_page = page;
665 } else if (BOXED_PAGE_FLAG & page_type_flag) {
666 /* Both code and data. */
667 generations[generation].alloc_large_start_page = page;
669 lose("bad page type flag: %d", page_type_flag);
672 if (UNBOXED_PAGE_FLAG == page_type_flag) {
673 generations[generation].alloc_unboxed_start_page = page;
674 } else if (BOXED_PAGE_FLAG & page_type_flag) {
675 /* Both code and data. */
676 generations[generation].alloc_start_page = page;
678 lose("bad page type flag: %d", page_type_flag);
683 /* Find a new region with room for at least the given number of bytes.
685 * It starts looking at the current generation's alloc_start_page. So
686 * may pick up from the previous region if there is enough space. This
687 * keeps the allocation contiguous when scavenging the newspace.
689 * The alloc_region should have been closed by a call to
690 * gc_alloc_update_page_tables(), and will thus be in an empty state.
692 * To assist the scavenging functions write-protected pages are not
693 * used. Free pages should not be write-protected.
695 * It is critical to the conservative GC that the start of regions be
696 * known. To help achieve this only small regions are allocated at a
699 * During scavenging, pointers may be found to within the current
700 * region and the page generation must be set so that pointers to the
701 * from space can be recognized. Therefore the generation of pages in
702 * the region are set to gc_alloc_generation. To prevent another
703 * allocation call using the same pages, all the pages in the region
704 * are allocated, although they will initially be empty.
707 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
709 page_index_t first_page;
710 page_index_t last_page;
711 unsigned long bytes_found;
717 "/alloc_new_region for %d bytes from gen %d\n",
718 nbytes, gc_alloc_generation));
721 /* Check that the region is in a reset state. */
722 gc_assert((alloc_region->first_page == 0)
723 && (alloc_region->last_page == -1)
724 && (alloc_region->free_pointer == alloc_region->end_addr));
725 ret = thread_mutex_lock(&free_pages_lock);
727 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
728 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
729 bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
730 + npage_bytes(last_page-first_page);
732 /* Set up the alloc_region. */
733 alloc_region->first_page = first_page;
734 alloc_region->last_page = last_page;
735 alloc_region->start_addr = page_table[first_page].bytes_used
736 + page_address(first_page);
737 alloc_region->free_pointer = alloc_region->start_addr;
738 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
740 /* Set up the pages. */
742 /* The first page may have already been in use. */
743 if (page_table[first_page].bytes_used == 0) {
744 page_table[first_page].allocated = page_type_flag;
745 page_table[first_page].gen = gc_alloc_generation;
746 page_table[first_page].large_object = 0;
747 page_table[first_page].region_start_offset = 0;
750 gc_assert(page_table[first_page].allocated == page_type_flag);
751 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
753 gc_assert(page_table[first_page].gen == gc_alloc_generation);
754 gc_assert(page_table[first_page].large_object == 0);
756 for (i = first_page+1; i <= last_page; i++) {
757 page_table[i].allocated = page_type_flag;
758 page_table[i].gen = gc_alloc_generation;
759 page_table[i].large_object = 0;
760 /* This may not be necessary for unboxed regions (think it was
762 page_table[i].region_start_offset =
763 void_diff(page_address(i),alloc_region->start_addr);
764 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
766 /* Bump up last_free_page. */
767 if (last_page+1 > last_free_page) {
768 last_free_page = last_page+1;
769 /* do we only want to call this on special occasions? like for
771 set_alloc_pointer((lispobj)page_address(last_free_page));
773 ret = thread_mutex_unlock(&free_pages_lock);
776 #ifdef READ_PROTECT_FREE_PAGES
777 os_protect(page_address(first_page),
778 npage_bytes(1+last_page-first_page),
782 /* If the first page was only partial, don't check whether it's
783 * zeroed (it won't be) and don't zero it (since the parts that
784 * we're interested in are guaranteed to be zeroed).
786 if (page_table[first_page].bytes_used) {
790 zero_dirty_pages(first_page, last_page);
792 /* we can do this after releasing free_pages_lock */
793 if (gencgc_zero_check) {
795 for (p = (long *)alloc_region->start_addr;
796 p < (long *)alloc_region->end_addr; p++) {
798 /* KLUDGE: It would be nice to use %lx and explicit casts
799 * (long) in code like this, so that it is less likely to
800 * break randomly when running on a machine with different
801 * word sizes. -- WHN 19991129 */
802 lose("The new region at %x is not zero (start=%p, end=%p).\n",
803 p, alloc_region->start_addr, alloc_region->end_addr);
809 /* If the record_new_objects flag is 2 then all new regions created
812 * If it's 1 then then it is only recorded if the first page of the
813 * current region is <= new_areas_ignore_page. This helps avoid
814 * unnecessary recording when doing full scavenge pass.
816 * The new_object structure holds the page, byte offset, and size of
817 * new regions of objects. Each new area is placed in the array of
818 * these structures pointer to by new_areas. new_areas_index holds the
819 * offset into new_areas.
821 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
822 * later code must detect this and handle it, probably by doing a full
823 * scavenge of a generation. */
824 #define NUM_NEW_AREAS 512
825 static int record_new_objects = 0;
826 static page_index_t new_areas_ignore_page;
832 static struct new_area (*new_areas)[];
833 static long new_areas_index;
836 /* Add a new area to new_areas. */
838 add_new_area(page_index_t first_page, size_t offset, size_t size)
840 unsigned long new_area_start,c;
843 /* Ignore if full. */
844 if (new_areas_index >= NUM_NEW_AREAS)
847 switch (record_new_objects) {
851 if (first_page > new_areas_ignore_page)
860 new_area_start = npage_bytes(first_page) + offset;
862 /* Search backwards for a prior area that this follows from. If
863 found this will save adding a new area. */
864 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
865 unsigned long area_end =
866 npage_bytes((*new_areas)[i].page)
867 + (*new_areas)[i].offset
868 + (*new_areas)[i].size;
870 "/add_new_area S1 %d %d %d %d\n",
871 i, c, new_area_start, area_end));*/
872 if (new_area_start == area_end) {
874 "/adding to [%d] %d %d %d with %d %d %d:\n",
876 (*new_areas)[i].page,
877 (*new_areas)[i].offset,
878 (*new_areas)[i].size,
882 (*new_areas)[i].size += size;
887 (*new_areas)[new_areas_index].page = first_page;
888 (*new_areas)[new_areas_index].offset = offset;
889 (*new_areas)[new_areas_index].size = size;
891 "/new_area %d page %d offset %d size %d\n",
892 new_areas_index, first_page, offset, size));*/
895 /* Note the max new_areas used. */
896 if (new_areas_index > max_new_areas)
897 max_new_areas = new_areas_index;
900 /* Update the tables for the alloc_region. The region may be added to
903 * When done the alloc_region is set up so that the next quick alloc
904 * will fail safely and thus a new region will be allocated. Further
905 * it is safe to try to re-update the page table of this reset
908 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
911 page_index_t first_page;
912 page_index_t next_page;
913 unsigned long bytes_used;
914 unsigned long orig_first_page_bytes_used;
915 unsigned long region_size;
916 unsigned long byte_cnt;
920 first_page = alloc_region->first_page;
922 /* Catch an unused alloc_region. */
923 if ((first_page == 0) && (alloc_region->last_page == -1))
926 next_page = first_page+1;
928 ret = thread_mutex_lock(&free_pages_lock);
930 if (alloc_region->free_pointer != alloc_region->start_addr) {
931 /* some bytes were allocated in the region */
932 orig_first_page_bytes_used = page_table[first_page].bytes_used;
934 gc_assert(alloc_region->start_addr ==
935 (page_address(first_page)
936 + page_table[first_page].bytes_used));
938 /* All the pages used need to be updated */
940 /* Update the first page. */
942 /* If the page was free then set up the gen, and
943 * region_start_offset. */
944 if (page_table[first_page].bytes_used == 0)
945 gc_assert(page_table[first_page].region_start_offset == 0);
946 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
948 gc_assert(page_table[first_page].allocated & page_type_flag);
949 gc_assert(page_table[first_page].gen == gc_alloc_generation);
950 gc_assert(page_table[first_page].large_object == 0);
954 /* Calculate the number of bytes used in this page. This is not
955 * always the number of new bytes, unless it was free. */
957 if ((bytes_used = void_diff(alloc_region->free_pointer,
958 page_address(first_page)))
960 bytes_used = PAGE_BYTES;
963 page_table[first_page].bytes_used = bytes_used;
964 byte_cnt += bytes_used;
967 /* All the rest of the pages should be free. We need to set
968 * their region_start_offset pointer to the start of the
969 * region, and set the bytes_used. */
971 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
972 gc_assert(page_table[next_page].allocated & page_type_flag);
973 gc_assert(page_table[next_page].bytes_used == 0);
974 gc_assert(page_table[next_page].gen == gc_alloc_generation);
975 gc_assert(page_table[next_page].large_object == 0);
977 gc_assert(page_table[next_page].region_start_offset ==
978 void_diff(page_address(next_page),
979 alloc_region->start_addr));
981 /* Calculate the number of bytes used in this page. */
983 if ((bytes_used = void_diff(alloc_region->free_pointer,
984 page_address(next_page)))>PAGE_BYTES) {
985 bytes_used = PAGE_BYTES;
988 page_table[next_page].bytes_used = bytes_used;
989 byte_cnt += bytes_used;
994 region_size = void_diff(alloc_region->free_pointer,
995 alloc_region->start_addr);
996 bytes_allocated += region_size;
997 generations[gc_alloc_generation].bytes_allocated += region_size;
999 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1001 /* Set the generations alloc restart page to the last page of
1003 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1005 /* Add the region to the new_areas if requested. */
1006 if (BOXED_PAGE_FLAG & page_type_flag)
1007 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1011 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1013 gc_alloc_generation));
1016 /* There are no bytes allocated. Unallocate the first_page if
1017 * there are 0 bytes_used. */
1018 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1019 if (page_table[first_page].bytes_used == 0)
1020 page_table[first_page].allocated = FREE_PAGE_FLAG;
1023 /* Unallocate any unused pages. */
1024 while (next_page <= alloc_region->last_page) {
1025 gc_assert(page_table[next_page].bytes_used == 0);
1026 page_table[next_page].allocated = FREE_PAGE_FLAG;
1029 ret = thread_mutex_unlock(&free_pages_lock);
1030 gc_assert(ret == 0);
1032 /* alloc_region is per-thread, we're ok to do this unlocked */
1033 gc_set_region_empty(alloc_region);
1036 static inline void *gc_quick_alloc(long nbytes);
1038 /* Allocate a possibly large object. */
1040 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1042 page_index_t first_page;
1043 page_index_t last_page;
1044 int orig_first_page_bytes_used;
1047 unsigned long bytes_used;
1048 page_index_t next_page;
1051 ret = thread_mutex_lock(&free_pages_lock);
1052 gc_assert(ret == 0);
1054 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1055 if (first_page <= alloc_region->last_page) {
1056 first_page = alloc_region->last_page+1;
1059 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1061 gc_assert(first_page > alloc_region->last_page);
1063 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1065 /* Set up the pages. */
1066 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1068 /* If the first page was free then set up the gen, and
1069 * region_start_offset. */
1070 if (page_table[first_page].bytes_used == 0) {
1071 page_table[first_page].allocated = page_type_flag;
1072 page_table[first_page].gen = gc_alloc_generation;
1073 page_table[first_page].region_start_offset = 0;
1074 page_table[first_page].large_object = 1;
1077 gc_assert(page_table[first_page].allocated == page_type_flag);
1078 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1079 gc_assert(page_table[first_page].large_object == 1);
1083 /* Calc. the number of bytes used in this page. This is not
1084 * always the number of new bytes, unless it was free. */
1086 if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
1087 bytes_used = PAGE_BYTES;
1090 page_table[first_page].bytes_used = bytes_used;
1091 byte_cnt += bytes_used;
1093 next_page = first_page+1;
1095 /* All the rest of the pages should be free. We need to set their
1096 * region_start_offset pointer to the start of the region, and set
1097 * the bytes_used. */
1099 gc_assert(page_free_p(next_page));
1100 gc_assert(page_table[next_page].bytes_used == 0);
1101 page_table[next_page].allocated = page_type_flag;
1102 page_table[next_page].gen = gc_alloc_generation;
1103 page_table[next_page].large_object = 1;
1105 page_table[next_page].region_start_offset =
1106 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1108 /* Calculate the number of bytes used in this page. */
1110 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1111 if (bytes_used > PAGE_BYTES) {
1112 bytes_used = PAGE_BYTES;
1115 page_table[next_page].bytes_used = bytes_used;
1116 page_table[next_page].write_protected=0;
1117 page_table[next_page].dont_move=0;
1118 byte_cnt += bytes_used;
1122 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1124 bytes_allocated += nbytes;
1125 generations[gc_alloc_generation].bytes_allocated += nbytes;
1127 /* Add the region to the new_areas if requested. */
1128 if (BOXED_PAGE_FLAG & page_type_flag)
1129 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1131 /* Bump up last_free_page */
1132 if (last_page+1 > last_free_page) {
1133 last_free_page = last_page+1;
1134 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1136 ret = thread_mutex_unlock(&free_pages_lock);
1137 gc_assert(ret == 0);
1139 #ifdef READ_PROTECT_FREE_PAGES
1140 os_protect(page_address(first_page),
1141 npage_bytes(1+last_page-first_page),
1145 zero_dirty_pages(first_page, last_page);
1147 return page_address(first_page);
1150 static page_index_t gencgc_alloc_start_page = -1;
1153 gc_heap_exhausted_error_or_lose (long available, long requested)
1155 struct thread *thread = arch_os_get_current_thread();
1156 /* Write basic information before doing anything else: if we don't
1157 * call to lisp this is a must, and even if we do there is always
1158 * the danger that we bounce back here before the error has been
1159 * handled, or indeed even printed.
1161 fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
1162 gc_active_p ? "garbage collection" : "allocation",
1163 available, requested);
1164 print_generation_stats();
1165 fprintf(stderr, "GC control variables:\n");
1166 fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
1167 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
1168 (SymbolValue(GC_PENDING, thread) == T) ?
1169 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
1170 "false" : "in progress"));
1171 #ifdef LISP_FEATURE_SB_THREAD
1172 fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n",
1173 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
1175 if (gc_active_p || (available == 0)) {
1176 /* If we are in GC, or totally out of memory there is no way
1177 * to sanely transfer control to the lisp-side of things.
1179 lose("Heap exhausted, game over.");
1182 /* FIXME: assert free_pages_lock held */
1183 (void)thread_mutex_unlock(&free_pages_lock);
1184 gc_assert(get_pseudo_atomic_atomic(thread));
1185 clear_pseudo_atomic_atomic(thread);
1186 if (get_pseudo_atomic_interrupted(thread))
1187 do_pending_interrupt();
1188 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1189 * to running user code at arbitrary places, even in a
1190 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1191 * running out of the heap. So at this point all bets are
1193 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1194 corruption_warning_and_maybe_lose
1195 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1196 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1197 alloc_number(available), alloc_number(requested));
1198 lose("HEAP-EXHAUSTED-ERROR fell through");
1203 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes,
1206 page_index_t first_page, last_page;
1207 page_index_t restart_page = *restart_page_ptr;
1208 long bytes_found = 0;
1209 long most_bytes_found = 0;
1210 /* FIXME: assert(free_pages_lock is held); */
1212 /* Toggled by gc_and_save for heap compaction, normally -1. */
1213 if (gencgc_alloc_start_page != -1) {
1214 restart_page = gencgc_alloc_start_page;
1217 gc_assert(nbytes>=0);
1218 if (((unsigned long)nbytes)>=PAGE_BYTES) {
1219 /* Search for a contiguous free space of at least nbytes,
1220 * aligned on a page boundary. The page-alignment is strictly
1221 * speaking needed only for objects at least large_object_size
1224 first_page = restart_page;
1225 while ((first_page < page_table_pages) &&
1226 page_allocated_p(first_page))
1229 last_page = first_page;
1230 bytes_found = PAGE_BYTES;
1231 while ((bytes_found < nbytes) &&
1232 (last_page < (page_table_pages-1)) &&
1233 page_free_p(last_page+1)) {
1235 bytes_found += PAGE_BYTES;
1236 gc_assert(0 == page_table[last_page].bytes_used);
1237 gc_assert(0 == page_table[last_page].write_protected);
1239 if (bytes_found > most_bytes_found)
1240 most_bytes_found = bytes_found;
1241 restart_page = last_page + 1;
1242 } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
1245 /* Search for a page with at least nbytes of space. We prefer
1246 * not to split small objects on multiple pages, to reduce the
1247 * number of contiguous allocation regions spaning multiple
1248 * pages: this helps avoid excessive conservativism. */
1249 first_page = restart_page;
1250 while (first_page < page_table_pages) {
1251 if (page_free_p(first_page))
1253 gc_assert(0 == page_table[first_page].bytes_used);
1254 bytes_found = PAGE_BYTES;
1257 else if ((page_table[first_page].allocated == page_type_flag) &&
1258 (page_table[first_page].large_object == 0) &&
1259 (page_table[first_page].gen == gc_alloc_generation) &&
1260 (page_table[first_page].write_protected == 0) &&
1261 (page_table[first_page].dont_move == 0))
1263 bytes_found = PAGE_BYTES
1264 - page_table[first_page].bytes_used;
1265 if (bytes_found > most_bytes_found)
1266 most_bytes_found = bytes_found;
1267 if (bytes_found >= nbytes)
1272 last_page = first_page;
1273 restart_page = first_page + 1;
1276 /* Check for a failure */
1277 if (bytes_found < nbytes) {
1278 gc_assert(restart_page >= page_table_pages);
1279 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1282 gc_assert(page_table[first_page].write_protected == 0);
1284 *restart_page_ptr = first_page;
1288 /* Allocate bytes. All the rest of the special-purpose allocation
1289 * functions will eventually call this */
1292 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1295 void *new_free_pointer;
1297 if (nbytes>=large_object_size)
1298 return gc_alloc_large(nbytes, page_type_flag, my_region);
1300 /* Check whether there is room in the current alloc region. */
1301 new_free_pointer = my_region->free_pointer + nbytes;
1303 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1304 my_region->free_pointer, new_free_pointer); */
1306 if (new_free_pointer <= my_region->end_addr) {
1307 /* If so then allocate from the current alloc region. */
1308 void *new_obj = my_region->free_pointer;
1309 my_region->free_pointer = new_free_pointer;
1311 /* Unless a `quick' alloc was requested, check whether the
1312 alloc region is almost empty. */
1314 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1315 /* If so, finished with the current region. */
1316 gc_alloc_update_page_tables(page_type_flag, my_region);
1317 /* Set up a new region. */
1318 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1321 return((void *)new_obj);
1324 /* Else not enough free space in the current region: retry with a
1327 gc_alloc_update_page_tables(page_type_flag, my_region);
1328 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1329 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1332 /* these are only used during GC: all allocation from the mutator calls
1333 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1336 static inline void *
1337 gc_quick_alloc(long nbytes)
1339 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1342 static inline void *
1343 gc_quick_alloc_large(long nbytes)
1345 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1348 static inline void *
1349 gc_alloc_unboxed(long nbytes)
1351 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1354 static inline void *
1355 gc_quick_alloc_unboxed(long nbytes)
1357 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1360 static inline void *
1361 gc_quick_alloc_large_unboxed(long nbytes)
1363 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1367 /* Copy a large boxed object. If the object is in a large object
1368 * region then it is simply promoted, else it is copied. If it's large
1369 * enough then it's copied to a large object region.
1371 * Vectors may have shrunk. If the object is not copied the space
1372 * needs to be reclaimed, and the page_tables corrected. */
1374 copy_large_object(lispobj object, long nwords)
1378 page_index_t first_page;
1380 gc_assert(is_lisp_pointer(object));
1381 gc_assert(from_space_p(object));
1382 gc_assert((nwords & 0x01) == 0);
1385 /* Check whether it's in a large object region. */
1386 first_page = find_page_index((void *)object);
1387 gc_assert(first_page >= 0);
1389 if (page_table[first_page].large_object) {
1391 /* Promote the object. */
1393 unsigned long remaining_bytes;
1394 page_index_t next_page;
1395 unsigned long bytes_freed;
1396 unsigned long old_bytes_used;
1398 /* Note: Any page write-protection must be removed, else a
1399 * later scavenge_newspace may incorrectly not scavenge these
1400 * pages. This would not be necessary if they are added to the
1401 * new areas, but let's do it for them all (they'll probably
1402 * be written anyway?). */
1404 gc_assert(page_table[first_page].region_start_offset == 0);
1406 next_page = first_page;
1407 remaining_bytes = nwords*N_WORD_BYTES;
1408 while (remaining_bytes > PAGE_BYTES) {
1409 gc_assert(page_table[next_page].gen == from_space);
1410 gc_assert(page_boxed_p(next_page));
1411 gc_assert(page_table[next_page].large_object);
1412 gc_assert(page_table[next_page].region_start_offset ==
1413 npage_bytes(next_page-first_page));
1414 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1415 /* Should have been unprotected by unprotect_oldspace(). */
1416 gc_assert(page_table[next_page].write_protected == 0);
1418 page_table[next_page].gen = new_space;
1420 remaining_bytes -= PAGE_BYTES;
1424 /* Now only one page remains, but the object may have shrunk
1425 * so there may be more unused pages which will be freed. */
1427 /* The object may have shrunk but shouldn't have grown. */
1428 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1430 page_table[next_page].gen = new_space;
1431 gc_assert(page_boxed_p(next_page));
1433 /* Adjust the bytes_used. */
1434 old_bytes_used = page_table[next_page].bytes_used;
1435 page_table[next_page].bytes_used = remaining_bytes;
1437 bytes_freed = old_bytes_used - remaining_bytes;
1439 /* Free any remaining pages; needs care. */
1441 while ((old_bytes_used == PAGE_BYTES) &&
1442 (page_table[next_page].gen == from_space) &&
1443 page_boxed_p(next_page) &&
1444 page_table[next_page].large_object &&
1445 (page_table[next_page].region_start_offset ==
1446 npage_bytes(next_page - first_page))) {
1447 /* Checks out OK, free the page. Don't need to bother zeroing
1448 * pages as this should have been done before shrinking the
1449 * object. These pages shouldn't be write-protected as they
1450 * should be zero filled. */
1451 gc_assert(page_table[next_page].write_protected == 0);
1453 old_bytes_used = page_table[next_page].bytes_used;
1454 page_table[next_page].allocated = FREE_PAGE_FLAG;
1455 page_table[next_page].bytes_used = 0;
1456 bytes_freed += old_bytes_used;
1460 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1462 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1463 bytes_allocated -= bytes_freed;
1465 /* Add the region to the new_areas if requested. */
1466 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1470 /* Get tag of object. */
1471 tag = lowtag_of(object);
1473 /* Allocate space. */
1474 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1476 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1478 /* Return Lisp pointer of new object. */
1479 return ((lispobj) new) | tag;
1483 /* to copy unboxed objects */
1485 copy_unboxed_object(lispobj object, long nwords)
1490 gc_assert(is_lisp_pointer(object));
1491 gc_assert(from_space_p(object));
1492 gc_assert((nwords & 0x01) == 0);
1494 /* Get tag of object. */
1495 tag = lowtag_of(object);
1497 /* Allocate space. */
1498 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1500 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1502 /* Return Lisp pointer of new object. */
1503 return ((lispobj) new) | tag;
1506 /* to copy large unboxed objects
1508 * If the object is in a large object region then it is simply
1509 * promoted, else it is copied. If it's large enough then it's copied
1510 * to a large object region.
1512 * Bignums and vectors may have shrunk. If the object is not copied
1513 * the space needs to be reclaimed, and the page_tables corrected.
1515 * KLUDGE: There's a lot of cut-and-paste duplication between this
1516 * function and copy_large_object(..). -- WHN 20000619 */
1518 copy_large_unboxed_object(lispobj object, long nwords)
1522 page_index_t first_page;
1524 gc_assert(is_lisp_pointer(object));
1525 gc_assert(from_space_p(object));
1526 gc_assert((nwords & 0x01) == 0);
1528 if ((nwords > 1024*1024) && gencgc_verbose) {
1529 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1530 nwords*N_WORD_BYTES));
1533 /* Check whether it's a large object. */
1534 first_page = find_page_index((void *)object);
1535 gc_assert(first_page >= 0);
1537 if (page_table[first_page].large_object) {
1538 /* Promote the object. Note: Unboxed objects may have been
1539 * allocated to a BOXED region so it may be necessary to
1540 * change the region to UNBOXED. */
1541 unsigned long remaining_bytes;
1542 page_index_t next_page;
1543 unsigned long bytes_freed;
1544 unsigned long old_bytes_used;
1546 gc_assert(page_table[first_page].region_start_offset == 0);
1548 next_page = first_page;
1549 remaining_bytes = nwords*N_WORD_BYTES;
1550 while (remaining_bytes > PAGE_BYTES) {
1551 gc_assert(page_table[next_page].gen == from_space);
1552 gc_assert(page_allocated_no_region_p(next_page));
1553 gc_assert(page_table[next_page].large_object);
1554 gc_assert(page_table[next_page].region_start_offset ==
1555 npage_bytes(next_page-first_page));
1556 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1558 page_table[next_page].gen = new_space;
1559 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1560 remaining_bytes -= PAGE_BYTES;
1564 /* Now only one page remains, but the object may have shrunk so
1565 * there may be more unused pages which will be freed. */
1567 /* Object may have shrunk but shouldn't have grown - check. */
1568 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1570 page_table[next_page].gen = new_space;
1571 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1573 /* Adjust the bytes_used. */
1574 old_bytes_used = page_table[next_page].bytes_used;
1575 page_table[next_page].bytes_used = remaining_bytes;
1577 bytes_freed = old_bytes_used - remaining_bytes;
1579 /* Free any remaining pages; needs care. */
1581 while ((old_bytes_used == PAGE_BYTES) &&
1582 (page_table[next_page].gen == from_space) &&
1583 page_allocated_no_region_p(next_page) &&
1584 page_table[next_page].large_object &&
1585 (page_table[next_page].region_start_offset ==
1586 npage_bytes(next_page - first_page))) {
1587 /* Checks out OK, free the page. Don't need to both zeroing
1588 * pages as this should have been done before shrinking the
1589 * object. These pages shouldn't be write-protected, even if
1590 * boxed they should be zero filled. */
1591 gc_assert(page_table[next_page].write_protected == 0);
1593 old_bytes_used = page_table[next_page].bytes_used;
1594 page_table[next_page].allocated = FREE_PAGE_FLAG;
1595 page_table[next_page].bytes_used = 0;
1596 bytes_freed += old_bytes_used;
1600 if ((bytes_freed > 0) && gencgc_verbose) {
1602 "/copy_large_unboxed bytes_freed=%d\n",
1606 generations[from_space].bytes_allocated -=
1607 nwords*N_WORD_BYTES + bytes_freed;
1608 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1609 bytes_allocated -= bytes_freed;
1614 /* Get tag of object. */
1615 tag = lowtag_of(object);
1617 /* Allocate space. */
1618 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1620 /* Copy the object. */
1621 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1623 /* Return Lisp pointer of new object. */
1624 return ((lispobj) new) | tag;
1633 * code and code-related objects
1636 static lispobj trans_fun_header(lispobj object);
1637 static lispobj trans_boxed(lispobj object);
1640 /* Scan a x86 compiled code object, looking for possible fixups that
1641 * have been missed after a move.
1643 * Two types of fixups are needed:
1644 * 1. Absolute fixups to within the code object.
1645 * 2. Relative fixups to outside the code object.
1647 * Currently only absolute fixups to the constant vector, or to the
1648 * code area are checked. */
1650 sniff_code_object(struct code *code, unsigned long displacement)
1652 #ifdef LISP_FEATURE_X86
1653 long nheader_words, ncode_words, nwords;
1655 void *constants_start_addr = NULL, *constants_end_addr;
1656 void *code_start_addr, *code_end_addr;
1657 int fixup_found = 0;
1659 if (!check_code_fixups)
1662 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1664 ncode_words = fixnum_value(code->code_size);
1665 nheader_words = HeaderValue(*(lispobj *)code);
1666 nwords = ncode_words + nheader_words;
1668 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1669 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1670 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1671 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1673 /* Work through the unboxed code. */
1674 for (p = code_start_addr; p < code_end_addr; p++) {
1675 void *data = *(void **)p;
1676 unsigned d1 = *((unsigned char *)p - 1);
1677 unsigned d2 = *((unsigned char *)p - 2);
1678 unsigned d3 = *((unsigned char *)p - 3);
1679 unsigned d4 = *((unsigned char *)p - 4);
1681 unsigned d5 = *((unsigned char *)p - 5);
1682 unsigned d6 = *((unsigned char *)p - 6);
1685 /* Check for code references. */
1686 /* Check for a 32 bit word that looks like an absolute
1687 reference to within the code adea of the code object. */
1688 if ((data >= (code_start_addr-displacement))
1689 && (data < (code_end_addr-displacement))) {
1690 /* function header */
1692 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1694 /* Skip the function header */
1698 /* the case of PUSH imm32 */
1702 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1703 p, d6, d5, d4, d3, d2, d1, data));
1704 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1706 /* the case of MOV [reg-8],imm32 */
1708 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1709 || d2==0x45 || d2==0x46 || d2==0x47)
1713 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1714 p, d6, d5, d4, d3, d2, d1, data));
1715 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1717 /* the case of LEA reg,[disp32] */
1718 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1721 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1722 p, d6, d5, d4, d3, d2, d1, data));
1723 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1727 /* Check for constant references. */
1728 /* Check for a 32 bit word that looks like an absolute
1729 reference to within the constant vector. Constant references
1731 if ((data >= (constants_start_addr-displacement))
1732 && (data < (constants_end_addr-displacement))
1733 && (((unsigned)data & 0x3) == 0)) {
1738 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1739 p, d6, d5, d4, d3, d2, d1, data));
1740 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1743 /* the case of MOV m32,EAX */
1747 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1748 p, d6, d5, d4, d3, d2, d1, data));
1749 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1752 /* the case of CMP m32,imm32 */
1753 if ((d1 == 0x3d) && (d2 == 0x81)) {
1756 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1757 p, d6, d5, d4, d3, d2, d1, data));
1759 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1762 /* Check for a mod=00, r/m=101 byte. */
1763 if ((d1 & 0xc7) == 5) {
1768 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1769 p, d6, d5, d4, d3, d2, d1, data));
1770 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1772 /* the case of CMP reg32,m32 */
1776 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1777 p, d6, d5, d4, d3, d2, d1, data));
1778 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1780 /* the case of MOV m32,reg32 */
1784 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1785 p, d6, d5, d4, d3, d2, d1, data));
1786 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1788 /* the case of MOV reg32,m32 */
1792 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1793 p, d6, d5, d4, d3, d2, d1, data));
1794 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1796 /* the case of LEA reg32,m32 */
1800 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1801 p, d6, d5, d4, d3, d2, d1, data));
1802 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1808 /* If anything was found, print some information on the code
1812 "/compiled code object at %x: header words = %d, code words = %d\n",
1813 code, nheader_words, ncode_words));
1815 "/const start = %x, end = %x\n",
1816 constants_start_addr, constants_end_addr));
1818 "/code start = %x, end = %x\n",
1819 code_start_addr, code_end_addr));
1825 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1827 /* x86-64 uses pc-relative addressing instead of this kludge */
1828 #ifndef LISP_FEATURE_X86_64
1829 long nheader_words, ncode_words, nwords;
1830 void *constants_start_addr, *constants_end_addr;
1831 void *code_start_addr, *code_end_addr;
1832 lispobj fixups = NIL;
1833 unsigned long displacement =
1834 (unsigned long)new_code - (unsigned long)old_code;
1835 struct vector *fixups_vector;
1837 ncode_words = fixnum_value(new_code->code_size);
1838 nheader_words = HeaderValue(*(lispobj *)new_code);
1839 nwords = ncode_words + nheader_words;
1841 "/compiled code object at %x: header words = %d, code words = %d\n",
1842 new_code, nheader_words, ncode_words)); */
1843 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1844 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1845 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1846 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1849 "/const start = %x, end = %x\n",
1850 constants_start_addr,constants_end_addr));
1852 "/code start = %x; end = %x\n",
1853 code_start_addr,code_end_addr));
1856 /* The first constant should be a pointer to the fixups for this
1857 code objects. Check. */
1858 fixups = new_code->constants[0];
1860 /* It will be 0 or the unbound-marker if there are no fixups (as
1861 * will be the case if the code object has been purified, for
1862 * example) and will be an other pointer if it is valid. */
1863 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1864 !is_lisp_pointer(fixups)) {
1865 /* Check for possible errors. */
1866 if (check_code_fixups)
1867 sniff_code_object(new_code, displacement);
1872 fixups_vector = (struct vector *)native_pointer(fixups);
1874 /* Could be pointing to a forwarding pointer. */
1875 /* FIXME is this always in from_space? if so, could replace this code with
1876 * forwarding_pointer_p/forwarding_pointer_value */
1877 if (is_lisp_pointer(fixups) &&
1878 (find_page_index((void*)fixups_vector) != -1) &&
1879 (fixups_vector->header == 0x01)) {
1880 /* If so, then follow it. */
1881 /*SHOW("following pointer to a forwarding pointer");*/
1883 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1886 /*SHOW("got fixups");*/
1888 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1889 /* Got the fixups for the code block. Now work through the vector,
1890 and apply a fixup at each address. */
1891 long length = fixnum_value(fixups_vector->length);
1893 for (i = 0; i < length; i++) {
1894 unsigned long offset = fixups_vector->data[i];
1895 /* Now check the current value of offset. */
1896 unsigned long old_value =
1897 *(unsigned long *)((unsigned long)code_start_addr + offset);
1899 /* If it's within the old_code object then it must be an
1900 * absolute fixup (relative ones are not saved) */
1901 if ((old_value >= (unsigned long)old_code)
1902 && (old_value < ((unsigned long)old_code
1903 + nwords*N_WORD_BYTES)))
1904 /* So add the dispacement. */
1905 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1906 old_value + displacement;
1908 /* It is outside the old code object so it must be a
1909 * relative fixup (absolute fixups are not saved). So
1910 * subtract the displacement. */
1911 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1912 old_value - displacement;
1915 /* This used to just print a note to stderr, but a bogus fixup seems to
1916 * indicate real heap corruption, so a hard hailure is in order. */
1917 lose("fixup vector %p has a bad widetag: %d\n",
1918 fixups_vector, widetag_of(fixups_vector->header));
1921 /* Check for possible errors. */
1922 if (check_code_fixups) {
1923 sniff_code_object(new_code,displacement);
1930 trans_boxed_large(lispobj object)
1933 unsigned long length;
1935 gc_assert(is_lisp_pointer(object));
1937 header = *((lispobj *) native_pointer(object));
1938 length = HeaderValue(header) + 1;
1939 length = CEILING(length, 2);
1941 return copy_large_object(object, length);
1944 /* Doesn't seem to be used, delete it after the grace period. */
1947 trans_unboxed_large(lispobj object)
1950 unsigned long length;
1952 gc_assert(is_lisp_pointer(object));
1954 header = *((lispobj *) native_pointer(object));
1955 length = HeaderValue(header) + 1;
1956 length = CEILING(length, 2);
1958 return copy_large_unboxed_object(object, length);
1964 * Lutexes. Using the normal finalization machinery for finalizing
1965 * lutexes is tricky, since the finalization depends on working lutexes.
1966 * So we track the lutexes in the GC and finalize them manually.
1969 #if defined(LUTEX_WIDETAG)
1972 * Start tracking LUTEX in the GC, by adding it to the linked list of
1973 * lutexes in the nursery generation. The caller is responsible for
1974 * locking, and GCs must be inhibited until the registration is
1978 gencgc_register_lutex (struct lutex *lutex) {
1979 int index = find_page_index(lutex);
1980 generation_index_t gen;
1983 /* This lutex is in static space, so we don't need to worry about
1989 gen = page_table[index].gen;
1991 gc_assert(gen >= 0);
1992 gc_assert(gen < NUM_GENERATIONS);
1994 head = generations[gen].lutexes;
2001 generations[gen].lutexes = lutex;
2005 * Stop tracking LUTEX in the GC by removing it from the appropriate
2006 * linked lists. This will only be called during GC, so no locking is
2010 gencgc_unregister_lutex (struct lutex *lutex) {
2012 lutex->prev->next = lutex->next;
2014 generations[lutex->gen].lutexes = lutex->next;
2018 lutex->next->prev = lutex->prev;
2027 * Mark all lutexes in generation GEN as not live.
2030 unmark_lutexes (generation_index_t gen) {
2031 struct lutex *lutex = generations[gen].lutexes;
2035 lutex = lutex->next;
2040 * Finalize all lutexes in generation GEN that have not been marked live.
2043 reap_lutexes (generation_index_t gen) {
2044 struct lutex *lutex = generations[gen].lutexes;
2047 struct lutex *next = lutex->next;
2049 lutex_destroy((tagged_lutex_t) lutex);
2050 gencgc_unregister_lutex(lutex);
2057 * Mark LUTEX as live.
2060 mark_lutex (lispobj tagged_lutex) {
2061 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
2067 * Move all lutexes in generation FROM to generation TO.
2070 move_lutexes (generation_index_t from, generation_index_t to) {
2071 struct lutex *tail = generations[from].lutexes;
2073 /* Nothing to move */
2077 /* Change the generation of the lutexes in FROM. */
2078 while (tail->next) {
2084 /* Link the last lutex in the FROM list to the start of the TO list */
2085 tail->next = generations[to].lutexes;
2087 /* And vice versa */
2088 if (generations[to].lutexes) {
2089 generations[to].lutexes->prev = tail;
2092 /* And update the generations structures to match this */
2093 generations[to].lutexes = generations[from].lutexes;
2094 generations[from].lutexes = NULL;
2098 scav_lutex(lispobj *where, lispobj object)
2100 mark_lutex((lispobj) where);
2102 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2106 trans_lutex(lispobj object)
2108 struct lutex *lutex = (struct lutex *) native_pointer(object);
2110 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2111 gc_assert(is_lisp_pointer(object));
2112 copied = copy_object(object, words);
2114 /* Update the links, since the lutex moved in memory. */
2116 lutex->next->prev = (struct lutex *) native_pointer(copied);
2120 lutex->prev->next = (struct lutex *) native_pointer(copied);
2122 generations[lutex->gen].lutexes =
2123 (struct lutex *) native_pointer(copied);
2130 size_lutex(lispobj *where)
2132 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2134 #endif /* LUTEX_WIDETAG */
2141 /* XX This is a hack adapted from cgc.c. These don't work too
2142 * efficiently with the gencgc as a list of the weak pointers is
2143 * maintained within the objects which causes writes to the pages. A
2144 * limited attempt is made to avoid unnecessary writes, but this needs
2146 #define WEAK_POINTER_NWORDS \
2147 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2150 scav_weak_pointer(lispobj *where, lispobj object)
2152 /* Since we overwrite the 'next' field, we have to make
2153 * sure not to do so for pointers already in the list.
2154 * Instead of searching the list of weak_pointers each
2155 * time, we ensure that next is always NULL when the weak
2156 * pointer isn't in the list, and not NULL otherwise.
2157 * Since we can't use NULL to denote end of list, we
2158 * use a pointer back to the same weak_pointer.
2160 struct weak_pointer * wp = (struct weak_pointer*)where;
2162 if (NULL == wp->next) {
2163 wp->next = weak_pointers;
2165 if (NULL == wp->next)
2169 /* Do not let GC scavenge the value slot of the weak pointer.
2170 * (That is why it is a weak pointer.) */
2172 return WEAK_POINTER_NWORDS;
2177 search_read_only_space(void *pointer)
2179 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2180 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2181 if ((pointer < (void *)start) || (pointer >= (void *)end))
2183 return (gc_search_space(start,
2184 (((lispobj *)pointer)+2)-start,
2185 (lispobj *) pointer));
2189 search_static_space(void *pointer)
2191 lispobj *start = (lispobj *)STATIC_SPACE_START;
2192 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2193 if ((pointer < (void *)start) || (pointer >= (void *)end))
2195 return (gc_search_space(start,
2196 (((lispobj *)pointer)+2)-start,
2197 (lispobj *) pointer));
2200 /* a faster version for searching the dynamic space. This will work even
2201 * if the object is in a current allocation region. */
2203 search_dynamic_space(void *pointer)
2205 page_index_t page_index = find_page_index(pointer);
2208 /* The address may be invalid, so do some checks. */
2209 if ((page_index == -1) || page_free_p(page_index))
2211 start = (lispobj *)page_region_start(page_index);
2212 return (gc_search_space(start,
2213 (((lispobj *)pointer)+2)-start,
2214 (lispobj *)pointer));
2217 /* Helper for valid_lisp_pointer_p and
2218 * possibly_valid_dynamic_space_pointer.
2220 * pointer is the pointer to validate, and start_addr is the address
2221 * of the enclosing object.
2224 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2226 if (!is_lisp_pointer((lispobj)pointer)) {
2230 /* Check that the object pointed to is consistent with the pointer
2232 switch (lowtag_of((lispobj)pointer)) {
2233 case FUN_POINTER_LOWTAG:
2234 /* Start_addr should be the enclosing code object, or a closure
2236 switch (widetag_of(*start_addr)) {
2237 case CODE_HEADER_WIDETAG:
2238 /* This case is probably caught above. */
2240 case CLOSURE_HEADER_WIDETAG:
2241 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2242 if ((unsigned long)pointer !=
2243 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2244 if (gencgc_verbose) {
2247 pointer, start_addr, *start_addr));
2253 if (gencgc_verbose) {
2256 pointer, start_addr, *start_addr));
2261 case LIST_POINTER_LOWTAG:
2262 if ((unsigned long)pointer !=
2263 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2264 if (gencgc_verbose) {
2267 pointer, start_addr, *start_addr));
2271 /* Is it plausible cons? */
2272 if ((is_lisp_pointer(start_addr[0]) ||
2273 is_lisp_immediate(start_addr[0])) &&
2274 (is_lisp_pointer(start_addr[1]) ||
2275 is_lisp_immediate(start_addr[1])))
2278 if (gencgc_verbose) {
2281 pointer, start_addr, *start_addr));
2285 case INSTANCE_POINTER_LOWTAG:
2286 if ((unsigned long)pointer !=
2287 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2288 if (gencgc_verbose) {
2291 pointer, start_addr, *start_addr));
2295 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2296 if (gencgc_verbose) {
2299 pointer, start_addr, *start_addr));
2304 case OTHER_POINTER_LOWTAG:
2306 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
2307 /* The all-architecture test below is good as far as it goes,
2308 * but an LRA object is similar to a FUN-POINTER: It is
2309 * embedded within a CODE-OBJECT pointed to by start_addr, and
2310 * cannot be found by simply walking the heap, therefore we
2311 * need to check for it. -- AB, 2010-Jun-04 */
2312 if ((widetag_of(start_addr[0]) == CODE_HEADER_WIDETAG)) {
2313 lispobj *potential_lra =
2314 (lispobj *)(((unsigned long)pointer) - OTHER_POINTER_LOWTAG);
2315 if ((widetag_of(potential_lra[0]) == RETURN_PC_HEADER_WIDETAG) &&
2316 ((potential_lra - HeaderValue(potential_lra[0])) == start_addr)) {
2317 return 1; /* It's as good as we can verify. */
2322 if ((unsigned long)pointer !=
2323 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2324 if (gencgc_verbose) {
2327 pointer, start_addr, *start_addr));
2331 /* Is it plausible? Not a cons. XXX should check the headers. */
2332 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2333 if (gencgc_verbose) {
2336 pointer, start_addr, *start_addr));
2340 switch (widetag_of(start_addr[0])) {
2341 case UNBOUND_MARKER_WIDETAG:
2342 case NO_TLS_VALUE_MARKER_WIDETAG:
2343 case CHARACTER_WIDETAG:
2344 #if N_WORD_BITS == 64
2345 case SINGLE_FLOAT_WIDETAG:
2347 if (gencgc_verbose) {
2350 pointer, start_addr, *start_addr));
2354 /* only pointed to by function pointers? */
2355 case CLOSURE_HEADER_WIDETAG:
2356 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2357 if (gencgc_verbose) {
2360 pointer, start_addr, *start_addr));
2364 case INSTANCE_HEADER_WIDETAG:
2365 if (gencgc_verbose) {
2368 pointer, start_addr, *start_addr));
2372 /* the valid other immediate pointer objects */
2373 case SIMPLE_VECTOR_WIDETAG:
2375 case COMPLEX_WIDETAG:
2376 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2377 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2379 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2380 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2382 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2383 case COMPLEX_LONG_FLOAT_WIDETAG:
2385 case SIMPLE_ARRAY_WIDETAG:
2386 case COMPLEX_BASE_STRING_WIDETAG:
2387 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2388 case COMPLEX_CHARACTER_STRING_WIDETAG:
2390 case COMPLEX_VECTOR_NIL_WIDETAG:
2391 case COMPLEX_BIT_VECTOR_WIDETAG:
2392 case COMPLEX_VECTOR_WIDETAG:
2393 case COMPLEX_ARRAY_WIDETAG:
2394 case VALUE_CELL_HEADER_WIDETAG:
2395 case SYMBOL_HEADER_WIDETAG:
2397 case CODE_HEADER_WIDETAG:
2398 case BIGNUM_WIDETAG:
2399 #if N_WORD_BITS != 64
2400 case SINGLE_FLOAT_WIDETAG:
2402 case DOUBLE_FLOAT_WIDETAG:
2403 #ifdef LONG_FLOAT_WIDETAG
2404 case LONG_FLOAT_WIDETAG:
2406 case SIMPLE_BASE_STRING_WIDETAG:
2407 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2408 case SIMPLE_CHARACTER_STRING_WIDETAG:
2410 case SIMPLE_BIT_VECTOR_WIDETAG:
2411 case SIMPLE_ARRAY_NIL_WIDETAG:
2412 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2413 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2414 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2415 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2416 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2417 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2418 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2419 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2421 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2422 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2423 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2424 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2426 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2427 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2429 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2430 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2432 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2433 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2435 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2436 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2438 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2439 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2441 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2442 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2444 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2445 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2447 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2448 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2450 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2451 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2452 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2453 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2455 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2456 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2458 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2459 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2461 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2462 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2465 case WEAK_POINTER_WIDETAG:
2466 #ifdef LUTEX_WIDETAG
2472 if (gencgc_verbose) {
2475 pointer, start_addr, *start_addr));
2481 if (gencgc_verbose) {
2484 pointer, start_addr, *start_addr));
2493 /* Used by the debugger to validate possibly bogus pointers before
2494 * calling MAKE-LISP-OBJ on them.
2496 * FIXME: We would like to make this perfect, because if the debugger
2497 * constructs a reference to a bugs lisp object, and it ends up in a
2498 * location scavenged by the GC all hell breaks loose.
2500 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2501 * and return true for all valid pointers, this could actually be eager
2502 * and lie about a few pointers without bad results... but that should
2503 * be reflected in the name.
2506 valid_lisp_pointer_p(lispobj *pointer)
2509 if (((start=search_dynamic_space(pointer))!=NULL) ||
2510 ((start=search_static_space(pointer))!=NULL) ||
2511 ((start=search_read_only_space(pointer))!=NULL))
2512 return looks_like_valid_lisp_pointer_p(pointer, start);
2517 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2519 /* Is there any possibility that pointer is a valid Lisp object
2520 * reference, and/or something else (e.g. subroutine call return
2521 * address) which should prevent us from moving the referred-to thing?
2522 * This is called from preserve_pointers() */
2524 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2526 lispobj *start_addr;
2528 /* Find the object start address. */
2529 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2533 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2536 /* Adjust large bignum and vector objects. This will adjust the
2537 * allocated region if the size has shrunk, and move unboxed objects
2538 * into unboxed pages. The pages are not promoted here, and the
2539 * promoted region is not added to the new_regions; this is really
2540 * only designed to be called from preserve_pointer(). Shouldn't fail
2541 * if this is missed, just may delay the moving of objects to unboxed
2542 * pages, and the freeing of pages. */
2544 maybe_adjust_large_object(lispobj *where)
2546 page_index_t first_page;
2547 page_index_t next_page;
2550 unsigned long remaining_bytes;
2551 unsigned long bytes_freed;
2552 unsigned long old_bytes_used;
2556 /* Check whether it's a vector or bignum object. */
2557 switch (widetag_of(where[0])) {
2558 case SIMPLE_VECTOR_WIDETAG:
2559 boxed = BOXED_PAGE_FLAG;
2561 case BIGNUM_WIDETAG:
2562 case SIMPLE_BASE_STRING_WIDETAG:
2563 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2564 case SIMPLE_CHARACTER_STRING_WIDETAG:
2566 case SIMPLE_BIT_VECTOR_WIDETAG:
2567 case SIMPLE_ARRAY_NIL_WIDETAG:
2568 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2569 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2570 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2571 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2572 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2573 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2574 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2575 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2577 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2578 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2579 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2580 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2582 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2583 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2585 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2586 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2588 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2589 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2591 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2592 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2594 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2595 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2597 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2598 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2600 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2601 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2603 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2604 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2606 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2607 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2608 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2609 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2611 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2612 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2614 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2615 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2617 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2618 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2620 boxed = UNBOXED_PAGE_FLAG;
2626 /* Find its current size. */
2627 nwords = (sizetab[widetag_of(where[0])])(where);
2629 first_page = find_page_index((void *)where);
2630 gc_assert(first_page >= 0);
2632 /* Note: Any page write-protection must be removed, else a later
2633 * scavenge_newspace may incorrectly not scavenge these pages.
2634 * This would not be necessary if they are added to the new areas,
2635 * but lets do it for them all (they'll probably be written
2638 gc_assert(page_table[first_page].region_start_offset == 0);
2640 next_page = first_page;
2641 remaining_bytes = nwords*N_WORD_BYTES;
2642 while (remaining_bytes > PAGE_BYTES) {
2643 gc_assert(page_table[next_page].gen == from_space);
2644 gc_assert(page_allocated_no_region_p(next_page));
2645 gc_assert(page_table[next_page].large_object);
2646 gc_assert(page_table[next_page].region_start_offset ==
2647 npage_bytes(next_page-first_page));
2648 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
2650 page_table[next_page].allocated = boxed;
2652 /* Shouldn't be write-protected at this stage. Essential that the
2654 gc_assert(!page_table[next_page].write_protected);
2655 remaining_bytes -= PAGE_BYTES;
2659 /* Now only one page remains, but the object may have shrunk so
2660 * there may be more unused pages which will be freed. */
2662 /* Object may have shrunk but shouldn't have grown - check. */
2663 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2665 page_table[next_page].allocated = boxed;
2666 gc_assert(page_table[next_page].allocated ==
2667 page_table[first_page].allocated);
2669 /* Adjust the bytes_used. */
2670 old_bytes_used = page_table[next_page].bytes_used;
2671 page_table[next_page].bytes_used = remaining_bytes;
2673 bytes_freed = old_bytes_used - remaining_bytes;
2675 /* Free any remaining pages; needs care. */
2677 while ((old_bytes_used == PAGE_BYTES) &&
2678 (page_table[next_page].gen == from_space) &&
2679 page_allocated_no_region_p(next_page) &&
2680 page_table[next_page].large_object &&
2681 (page_table[next_page].region_start_offset ==
2682 npage_bytes(next_page - first_page))) {
2683 /* It checks out OK, free the page. We don't need to both zeroing
2684 * pages as this should have been done before shrinking the
2685 * object. These pages shouldn't be write protected as they
2686 * should be zero filled. */
2687 gc_assert(page_table[next_page].write_protected == 0);
2689 old_bytes_used = page_table[next_page].bytes_used;
2690 page_table[next_page].allocated = FREE_PAGE_FLAG;
2691 page_table[next_page].bytes_used = 0;
2692 bytes_freed += old_bytes_used;
2696 if ((bytes_freed > 0) && gencgc_verbose) {
2698 "/maybe_adjust_large_object() freed %d\n",
2702 generations[from_space].bytes_allocated -= bytes_freed;
2703 bytes_allocated -= bytes_freed;
2708 /* Take a possible pointer to a Lisp object and mark its page in the
2709 * page_table so that it will not be relocated during a GC.
2711 * This involves locating the page it points to, then backing up to
2712 * the start of its region, then marking all pages dont_move from there
2713 * up to the first page that's not full or has a different generation
2715 * It is assumed that all the page static flags have been cleared at
2716 * the start of a GC.
2718 * It is also assumed that the current gc_alloc() region has been
2719 * flushed and the tables updated. */
2722 preserve_pointer(void *addr)
2724 page_index_t addr_page_index = find_page_index(addr);
2725 page_index_t first_page;
2727 unsigned int region_allocation;
2729 /* quick check 1: Address is quite likely to have been invalid. */
2730 if ((addr_page_index == -1)
2731 || page_free_p(addr_page_index)
2732 || (page_table[addr_page_index].bytes_used == 0)
2733 || (page_table[addr_page_index].gen != from_space)
2734 /* Skip if already marked dont_move. */
2735 || (page_table[addr_page_index].dont_move != 0))
2737 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2738 /* (Now that we know that addr_page_index is in range, it's
2739 * safe to index into page_table[] with it.) */
2740 region_allocation = page_table[addr_page_index].allocated;
2742 /* quick check 2: Check the offset within the page.
2745 if (((unsigned long)addr & (PAGE_BYTES - 1)) >
2746 page_table[addr_page_index].bytes_used)
2749 /* Filter out anything which can't be a pointer to a Lisp object
2750 * (or, as a special case which also requires dont_move, a return
2751 * address referring to something in a CodeObject). This is
2752 * expensive but important, since it vastly reduces the
2753 * probability that random garbage will be bogusly interpreted as
2754 * a pointer which prevents a page from moving. */
2755 if (!(code_page_p(addr_page_index)
2756 || (is_lisp_pointer((lispobj)addr) &&
2757 possibly_valid_dynamic_space_pointer(addr))))
2760 /* Find the beginning of the region. Note that there may be
2761 * objects in the region preceding the one that we were passed a
2762 * pointer to: if this is the case, we will write-protect all the
2763 * previous objects' pages too. */
2766 /* I think this'd work just as well, but without the assertions.
2767 * -dan 2004.01.01 */
2768 first_page = find_page_index(page_region_start(addr_page_index))
2770 first_page = addr_page_index;
2771 while (page_table[first_page].region_start_offset != 0) {
2773 /* Do some checks. */
2774 gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
2775 gc_assert(page_table[first_page].gen == from_space);
2776 gc_assert(page_table[first_page].allocated == region_allocation);
2780 /* Adjust any large objects before promotion as they won't be
2781 * copied after promotion. */
2782 if (page_table[first_page].large_object) {
2783 maybe_adjust_large_object(page_address(first_page));
2784 /* If a large object has shrunk then addr may now point to a
2785 * free area in which case it's ignored here. Note it gets
2786 * through the valid pointer test above because the tail looks
2788 if (page_free_p(addr_page_index)
2789 || (page_table[addr_page_index].bytes_used == 0)
2790 /* Check the offset within the page. */
2791 || (((unsigned long)addr & (PAGE_BYTES - 1))
2792 > page_table[addr_page_index].bytes_used)) {
2794 "weird? ignore ptr 0x%x to freed area of large object\n",
2798 /* It may have moved to unboxed pages. */
2799 region_allocation = page_table[first_page].allocated;
2802 /* Now work forward until the end of this contiguous area is found,
2803 * marking all pages as dont_move. */
2804 for (i = first_page; ;i++) {
2805 gc_assert(page_table[i].allocated == region_allocation);
2807 /* Mark the page static. */
2808 page_table[i].dont_move = 1;
2810 /* Move the page to the new_space. XX I'd rather not do this
2811 * but the GC logic is not quite able to copy with the static
2812 * pages remaining in the from space. This also requires the
2813 * generation bytes_allocated counters be updated. */
2814 page_table[i].gen = new_space;
2815 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2816 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2818 /* It is essential that the pages are not write protected as
2819 * they may have pointers into the old-space which need
2820 * scavenging. They shouldn't be write protected at this
2822 gc_assert(!page_table[i].write_protected);
2824 /* Check whether this is the last page in this contiguous block.. */
2825 if ((page_table[i].bytes_used < PAGE_BYTES)
2826 /* ..or it is PAGE_BYTES and is the last in the block */
2828 || (page_table[i+1].bytes_used == 0) /* next page free */
2829 || (page_table[i+1].gen != from_space) /* diff. gen */
2830 || (page_table[i+1].region_start_offset == 0))
2834 /* Check that the page is now static. */
2835 gc_assert(page_table[addr_page_index].dont_move != 0);
2838 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2841 /* If the given page is not write-protected, then scan it for pointers
2842 * to younger generations or the top temp. generation, if no
2843 * suspicious pointers are found then the page is write-protected.
2845 * Care is taken to check for pointers to the current gc_alloc()
2846 * region if it is a younger generation or the temp. generation. This
2847 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2848 * the gc_alloc_generation does not need to be checked as this is only
2849 * called from scavenge_generation() when the gc_alloc generation is
2850 * younger, so it just checks if there is a pointer to the current
2853 * We return 1 if the page was write-protected, else 0. */
2855 update_page_write_prot(page_index_t page)
2857 generation_index_t gen = page_table[page].gen;
2860 void **page_addr = (void **)page_address(page);
2861 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2863 /* Shouldn't be a free page. */
2864 gc_assert(page_allocated_p(page));
2865 gc_assert(page_table[page].bytes_used != 0);
2867 /* Skip if it's already write-protected, pinned, or unboxed */
2868 if (page_table[page].write_protected
2869 /* FIXME: What's the reason for not write-protecting pinned pages? */
2870 || page_table[page].dont_move
2871 || page_unboxed_p(page))
2874 /* Scan the page for pointers to younger generations or the
2875 * top temp. generation. */
2877 for (j = 0; j < num_words; j++) {
2878 void *ptr = *(page_addr+j);
2879 page_index_t index = find_page_index(ptr);
2881 /* Check that it's in the dynamic space */
2883 if (/* Does it point to a younger or the temp. generation? */
2884 (page_allocated_p(index)
2885 && (page_table[index].bytes_used != 0)
2886 && ((page_table[index].gen < gen)
2887 || (page_table[index].gen == SCRATCH_GENERATION)))
2889 /* Or does it point within a current gc_alloc() region? */
2890 || ((boxed_region.start_addr <= ptr)
2891 && (ptr <= boxed_region.free_pointer))
2892 || ((unboxed_region.start_addr <= ptr)
2893 && (ptr <= unboxed_region.free_pointer))) {
2900 /* Write-protect the page. */
2901 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2903 os_protect((void *)page_addr,
2905 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2907 /* Note the page as protected in the page tables. */
2908 page_table[page].write_protected = 1;
2914 /* Scavenge all generations from FROM to TO, inclusive, except for
2915 * new_space which needs special handling, as new objects may be
2916 * added which are not checked here - use scavenge_newspace generation.
2918 * Write-protected pages should not have any pointers to the
2919 * from_space so do need scavenging; thus write-protected pages are
2920 * not always scavenged. There is some code to check that these pages
2921 * are not written; but to check fully the write-protected pages need
2922 * to be scavenged by disabling the code to skip them.
2924 * Under the current scheme when a generation is GCed the younger
2925 * generations will be empty. So, when a generation is being GCed it
2926 * is only necessary to scavenge the older generations for pointers
2927 * not the younger. So a page that does not have pointers to younger
2928 * generations does not need to be scavenged.
2930 * The write-protection can be used to note pages that don't have
2931 * pointers to younger pages. But pages can be written without having
2932 * pointers to younger generations. After the pages are scavenged here
2933 * they can be scanned for pointers to younger generations and if
2934 * there are none the page can be write-protected.
2936 * One complication is when the newspace is the top temp. generation.
2938 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2939 * that none were written, which they shouldn't be as they should have
2940 * no pointers to younger generations. This breaks down for weak
2941 * pointers as the objects contain a link to the next and are written
2942 * if a weak pointer is scavenged. Still it's a useful check. */
2944 scavenge_generations(generation_index_t from, generation_index_t to)
2951 /* Clear the write_protected_cleared flags on all pages. */
2952 for (i = 0; i < page_table_pages; i++)
2953 page_table[i].write_protected_cleared = 0;
2956 for (i = 0; i < last_free_page; i++) {
2957 generation_index_t generation = page_table[i].gen;
2959 && (page_table[i].bytes_used != 0)
2960 && (generation != new_space)
2961 && (generation >= from)
2962 && (generation <= to)) {
2963 page_index_t last_page,j;
2964 int write_protected=1;
2966 /* This should be the start of a region */
2967 gc_assert(page_table[i].region_start_offset == 0);
2969 /* Now work forward until the end of the region */
2970 for (last_page = i; ; last_page++) {
2972 write_protected && page_table[last_page].write_protected;
2973 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2974 /* Or it is PAGE_BYTES and is the last in the block */
2975 || (!page_boxed_p(last_page+1))
2976 || (page_table[last_page+1].bytes_used == 0)
2977 || (page_table[last_page+1].gen != generation)
2978 || (page_table[last_page+1].region_start_offset == 0))
2981 if (!write_protected) {
2982 scavenge(page_address(i),
2983 ((unsigned long)(page_table[last_page].bytes_used
2984 + npage_bytes(last_page-i)))
2987 /* Now scan the pages and write protect those that
2988 * don't have pointers to younger generations. */
2989 if (enable_page_protection) {
2990 for (j = i; j <= last_page; j++) {
2991 num_wp += update_page_write_prot(j);
2994 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2996 "/write protected %d pages within generation %d\n",
2997 num_wp, generation));
3005 /* Check that none of the write_protected pages in this generation
3006 * have been written to. */
3007 for (i = 0; i < page_table_pages; i++) {
3008 if (page_allocated_p(i)
3009 && (page_table[i].bytes_used != 0)
3010 && (page_table[i].gen == generation)
3011 && (page_table[i].write_protected_cleared != 0)) {
3012 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
3014 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
3015 page_table[i].bytes_used,
3016 page_table[i].region_start_offset,
3017 page_table[i].dont_move));
3018 lose("write to protected page %d in scavenge_generation()\n", i);
3025 /* Scavenge a newspace generation. As it is scavenged new objects may
3026 * be allocated to it; these will also need to be scavenged. This
3027 * repeats until there are no more objects unscavenged in the
3028 * newspace generation.
3030 * To help improve the efficiency, areas written are recorded by
3031 * gc_alloc() and only these scavenged. Sometimes a little more will be
3032 * scavenged, but this causes no harm. An easy check is done that the
3033 * scavenged bytes equals the number allocated in the previous
3036 * Write-protected pages are not scanned except if they are marked
3037 * dont_move in which case they may have been promoted and still have
3038 * pointers to the from space.
3040 * Write-protected pages could potentially be written by alloc however
3041 * to avoid having to handle re-scavenging of write-protected pages
3042 * gc_alloc() does not write to write-protected pages.
3044 * New areas of objects allocated are recorded alternatively in the two
3045 * new_areas arrays below. */
3046 static struct new_area new_areas_1[NUM_NEW_AREAS];
3047 static struct new_area new_areas_2[NUM_NEW_AREAS];
3049 /* Do one full scan of the new space generation. This is not enough to
3050 * complete the job as new objects may be added to the generation in
3051 * the process which are not scavenged. */
3053 scavenge_newspace_generation_one_scan(generation_index_t generation)
3058 "/starting one full scan of newspace generation %d\n",
3060 for (i = 0; i < last_free_page; i++) {
3061 /* Note that this skips over open regions when it encounters them. */
3063 && (page_table[i].bytes_used != 0)
3064 && (page_table[i].gen == generation)
3065 && ((page_table[i].write_protected == 0)
3066 /* (This may be redundant as write_protected is now
3067 * cleared before promotion.) */
3068 || (page_table[i].dont_move == 1))) {
3069 page_index_t last_page;
3072 /* The scavenge will start at the region_start_offset of
3075 * We need to find the full extent of this contiguous
3076 * block in case objects span pages.
3078 * Now work forward until the end of this contiguous area
3079 * is found. A small area is preferred as there is a
3080 * better chance of its pages being write-protected. */
3081 for (last_page = i; ;last_page++) {
3082 /* If all pages are write-protected and movable,
3083 * then no need to scavenge */
3084 all_wp=all_wp && page_table[last_page].write_protected &&
3085 !page_table[last_page].dont_move;
3087 /* Check whether this is the last page in this
3088 * contiguous block */
3089 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3090 /* Or it is PAGE_BYTES and is the last in the block */
3091 || (!page_boxed_p(last_page+1))
3092 || (page_table[last_page+1].bytes_used == 0)
3093 || (page_table[last_page+1].gen != generation)
3094 || (page_table[last_page+1].region_start_offset == 0))
3098 /* Do a limited check for write-protected pages. */
3100 long nwords = (((unsigned long)
3101 (page_table[last_page].bytes_used
3102 + npage_bytes(last_page-i)
3103 + page_table[i].region_start_offset))
3105 new_areas_ignore_page = last_page;
3107 scavenge(page_region_start(i), nwords);
3114 "/done with one full scan of newspace generation %d\n",
3118 /* Do a complete scavenge of the newspace generation. */
3120 scavenge_newspace_generation(generation_index_t generation)
3124 /* the new_areas array currently being written to by gc_alloc() */
3125 struct new_area (*current_new_areas)[] = &new_areas_1;
3126 long current_new_areas_index;
3128 /* the new_areas created by the previous scavenge cycle */
3129 struct new_area (*previous_new_areas)[] = NULL;
3130 long previous_new_areas_index;
3132 /* Flush the current regions updating the tables. */
3133 gc_alloc_update_all_page_tables();
3135 /* Turn on the recording of new areas by gc_alloc(). */
3136 new_areas = current_new_areas;
3137 new_areas_index = 0;
3139 /* Don't need to record new areas that get scavenged anyway during
3140 * scavenge_newspace_generation_one_scan. */
3141 record_new_objects = 1;
3143 /* Start with a full scavenge. */
3144 scavenge_newspace_generation_one_scan(generation);
3146 /* Record all new areas now. */
3147 record_new_objects = 2;
3149 /* Give a chance to weak hash tables to make other objects live.
3150 * FIXME: The algorithm implemented here for weak hash table gcing
3151 * is O(W^2+N) as Bruno Haible warns in
3152 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3153 * see "Implementation 2". */
3154 scav_weak_hash_tables();
3156 /* Flush the current regions updating the tables. */
3157 gc_alloc_update_all_page_tables();
3159 /* Grab new_areas_index. */
3160 current_new_areas_index = new_areas_index;
3163 "The first scan is finished; current_new_areas_index=%d.\n",
3164 current_new_areas_index));*/
3166 while (current_new_areas_index > 0) {
3167 /* Move the current to the previous new areas */
3168 previous_new_areas = current_new_areas;
3169 previous_new_areas_index = current_new_areas_index;
3171 /* Scavenge all the areas in previous new areas. Any new areas
3172 * allocated are saved in current_new_areas. */
3174 /* Allocate an array for current_new_areas; alternating between
3175 * new_areas_1 and 2 */
3176 if (previous_new_areas == &new_areas_1)
3177 current_new_areas = &new_areas_2;
3179 current_new_areas = &new_areas_1;
3181 /* Set up for gc_alloc(). */
3182 new_areas = current_new_areas;
3183 new_areas_index = 0;
3185 /* Check whether previous_new_areas had overflowed. */
3186 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3188 /* New areas of objects allocated have been lost so need to do a
3189 * full scan to be sure! If this becomes a problem try
3190 * increasing NUM_NEW_AREAS. */
3191 if (gencgc_verbose) {
3192 SHOW("new_areas overflow, doing full scavenge");
3195 /* Don't need to record new areas that get scavenged
3196 * anyway during scavenge_newspace_generation_one_scan. */
3197 record_new_objects = 1;
3199 scavenge_newspace_generation_one_scan(generation);
3201 /* Record all new areas now. */
3202 record_new_objects = 2;
3204 scav_weak_hash_tables();
3206 /* Flush the current regions updating the tables. */
3207 gc_alloc_update_all_page_tables();
3211 /* Work through previous_new_areas. */
3212 for (i = 0; i < previous_new_areas_index; i++) {
3213 page_index_t page = (*previous_new_areas)[i].page;
3214 size_t offset = (*previous_new_areas)[i].offset;
3215 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3216 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3217 scavenge(page_address(page)+offset, size);
3220 scav_weak_hash_tables();
3222 /* Flush the current regions updating the tables. */
3223 gc_alloc_update_all_page_tables();
3226 current_new_areas_index = new_areas_index;
3229 "The re-scan has finished; current_new_areas_index=%d.\n",
3230 current_new_areas_index));*/
3233 /* Turn off recording of areas allocated by gc_alloc(). */
3234 record_new_objects = 0;
3237 /* Check that none of the write_protected pages in this generation
3238 * have been written to. */
3239 for (i = 0; i < page_table_pages; i++) {
3240 if (page_allocated_p(i)
3241 && (page_table[i].bytes_used != 0)
3242 && (page_table[i].gen == generation)
3243 && (page_table[i].write_protected_cleared != 0)
3244 && (page_table[i].dont_move == 0)) {
3245 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3246 i, generation, page_table[i].dont_move);
3252 /* Un-write-protect all the pages in from_space. This is done at the
3253 * start of a GC else there may be many page faults while scavenging
3254 * the newspace (I've seen drive the system time to 99%). These pages
3255 * would need to be unprotected anyway before unmapping in
3256 * free_oldspace; not sure what effect this has on paging.. */
3258 unprotect_oldspace(void)
3261 void *region_addr = 0;
3262 void *page_addr = 0;
3263 unsigned long region_bytes = 0;
3265 for (i = 0; i < last_free_page; i++) {
3266 if (page_allocated_p(i)
3267 && (page_table[i].bytes_used != 0)
3268 && (page_table[i].gen == from_space)) {
3270 /* Remove any write-protection. We should be able to rely
3271 * on the write-protect flag to avoid redundant calls. */
3272 if (page_table[i].write_protected) {
3273 page_table[i].write_protected = 0;
3274 page_addr = page_address(i);
3277 region_addr = page_addr;
3278 region_bytes = PAGE_BYTES;
3279 } else if (region_addr + region_bytes == page_addr) {
3280 /* Region continue. */
3281 region_bytes += PAGE_BYTES;
3283 /* Unprotect previous region. */
3284 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
3285 /* First page in new region. */
3286 region_addr = page_addr;
3287 region_bytes = PAGE_BYTES;
3293 /* Unprotect last region. */
3294 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
3298 /* Work through all the pages and free any in from_space. This
3299 * assumes that all objects have been copied or promoted to an older
3300 * generation. Bytes_allocated and the generation bytes_allocated
3301 * counter are updated. The number of bytes freed is returned. */
3302 static unsigned long
3305 unsigned long bytes_freed = 0;
3306 page_index_t first_page, last_page;
3311 /* Find a first page for the next region of pages. */
3312 while ((first_page < last_free_page)
3313 && (page_free_p(first_page)
3314 || (page_table[first_page].bytes_used == 0)
3315 || (page_table[first_page].gen != from_space)))
3318 if (first_page >= last_free_page)
3321 /* Find the last page of this region. */
3322 last_page = first_page;
3325 /* Free the page. */
3326 bytes_freed += page_table[last_page].bytes_used;
3327 generations[page_table[last_page].gen].bytes_allocated -=
3328 page_table[last_page].bytes_used;
3329 page_table[last_page].allocated = FREE_PAGE_FLAG;
3330 page_table[last_page].bytes_used = 0;
3331 /* Should already be unprotected by unprotect_oldspace(). */
3332 gc_assert(!page_table[last_page].write_protected);
3335 while ((last_page < last_free_page)
3336 && page_allocated_p(last_page)
3337 && (page_table[last_page].bytes_used != 0)
3338 && (page_table[last_page].gen == from_space));
3340 #ifdef READ_PROTECT_FREE_PAGES
3341 os_protect(page_address(first_page),
3342 npage_bytes(last_page-first_page),
3345 first_page = last_page;
3346 } while (first_page < last_free_page);
3348 bytes_allocated -= bytes_freed;
3353 /* Print some information about a pointer at the given address. */
3355 print_ptr(lispobj *addr)
3357 /* If addr is in the dynamic space then out the page information. */
3358 page_index_t pi1 = find_page_index((void*)addr);
3361 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
3362 (unsigned long) addr,
3364 page_table[pi1].allocated,
3365 page_table[pi1].gen,
3366 page_table[pi1].bytes_used,
3367 page_table[pi1].region_start_offset,
3368 page_table[pi1].dont_move);
3369 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3383 is_in_stack_space(lispobj ptr)
3385 /* For space verification: Pointers can be valid if they point
3386 * to a thread stack space. This would be faster if the thread
3387 * structures had page-table entries as if they were part of
3388 * the heap space. */
3390 for_each_thread(th) {
3391 if ((th->control_stack_start <= (lispobj *)ptr) &&
3392 (th->control_stack_end >= (lispobj *)ptr)) {
3400 verify_space(lispobj *start, size_t words)
3402 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3403 int is_in_readonly_space =
3404 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3405 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3409 lispobj thing = *(lispobj*)start;
3411 if (is_lisp_pointer(thing)) {
3412 page_index_t page_index = find_page_index((void*)thing);
3413 long to_readonly_space =
3414 (READ_ONLY_SPACE_START <= thing &&
3415 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3416 long to_static_space =
3417 (STATIC_SPACE_START <= thing &&
3418 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3420 /* Does it point to the dynamic space? */
3421 if (page_index != -1) {
3422 /* If it's within the dynamic space it should point to a used
3423 * page. XX Could check the offset too. */
3424 if (page_allocated_p(page_index)
3425 && (page_table[page_index].bytes_used == 0))
3426 lose ("Ptr %p @ %p sees free page.\n", thing, start);
3427 /* Check that it doesn't point to a forwarding pointer! */
3428 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3429 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
3431 /* Check that its not in the RO space as it would then be a
3432 * pointer from the RO to the dynamic space. */
3433 if (is_in_readonly_space) {
3434 lose("ptr to dynamic space %p from RO space %x\n",
3437 /* Does it point to a plausible object? This check slows
3438 * it down a lot (so it's commented out).
3440 * "a lot" is serious: it ate 50 minutes cpu time on
3441 * my duron 950 before I came back from lunch and
3444 * FIXME: Add a variable to enable this
3447 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3448 lose("ptr %p to invalid object %p\n", thing, start);
3452 extern void funcallable_instance_tramp;
3453 /* Verify that it points to another valid space. */
3454 if (!to_readonly_space && !to_static_space
3455 && (thing != (lispobj)&funcallable_instance_tramp)
3456 && !is_in_stack_space(thing)) {
3457 lose("Ptr %p @ %p sees junk.\n", thing, start);
3461 if (!(fixnump(thing))) {
3463 switch(widetag_of(*start)) {
3466 case SIMPLE_VECTOR_WIDETAG:
3468 case COMPLEX_WIDETAG:
3469 case SIMPLE_ARRAY_WIDETAG:
3470 case COMPLEX_BASE_STRING_WIDETAG:
3471 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3472 case COMPLEX_CHARACTER_STRING_WIDETAG:
3474 case COMPLEX_VECTOR_NIL_WIDETAG:
3475 case COMPLEX_BIT_VECTOR_WIDETAG:
3476 case COMPLEX_VECTOR_WIDETAG:
3477 case COMPLEX_ARRAY_WIDETAG:
3478 case CLOSURE_HEADER_WIDETAG:
3479 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3480 case VALUE_CELL_HEADER_WIDETAG:
3481 case SYMBOL_HEADER_WIDETAG:
3482 case CHARACTER_WIDETAG:
3483 #if N_WORD_BITS == 64
3484 case SINGLE_FLOAT_WIDETAG:
3486 case UNBOUND_MARKER_WIDETAG:
3491 case INSTANCE_HEADER_WIDETAG:
3494 long ntotal = HeaderValue(thing);
3495 lispobj layout = ((struct instance *)start)->slots[0];
3500 nuntagged = ((struct layout *)
3501 native_pointer(layout))->n_untagged_slots;
3502 verify_space(start + 1,
3503 ntotal - fixnum_value(nuntagged));
3507 case CODE_HEADER_WIDETAG:
3509 lispobj object = *start;
3511 long nheader_words, ncode_words, nwords;
3513 struct simple_fun *fheaderp;
3515 code = (struct code *) start;
3517 /* Check that it's not in the dynamic space.
3518 * FIXME: Isn't is supposed to be OK for code
3519 * objects to be in the dynamic space these days? */
3520 if (is_in_dynamic_space
3521 /* It's ok if it's byte compiled code. The trace
3522 * table offset will be a fixnum if it's x86
3523 * compiled code - check.
3525 * FIXME: #^#@@! lack of abstraction here..
3526 * This line can probably go away now that
3527 * there's no byte compiler, but I've got
3528 * too much to worry about right now to try
3529 * to make sure. -- WHN 2001-10-06 */
3530 && fixnump(code->trace_table_offset)
3531 /* Only when enabled */
3532 && verify_dynamic_code_check) {
3534 "/code object at %p in the dynamic space\n",
3538 ncode_words = fixnum_value(code->code_size);
3539 nheader_words = HeaderValue(object);
3540 nwords = ncode_words + nheader_words;
3541 nwords = CEILING(nwords, 2);
3542 /* Scavenge the boxed section of the code data block */
3543 verify_space(start + 1, nheader_words - 1);
3545 /* Scavenge the boxed section of each function
3546 * object in the code data block. */
3547 fheaderl = code->entry_points;
3548 while (fheaderl != NIL) {
3550 (struct simple_fun *) native_pointer(fheaderl);
3551 gc_assert(widetag_of(fheaderp->header) ==
3552 SIMPLE_FUN_HEADER_WIDETAG);
3553 verify_space(&fheaderp->name, 1);
3554 verify_space(&fheaderp->arglist, 1);
3555 verify_space(&fheaderp->type, 1);
3556 fheaderl = fheaderp->next;
3562 /* unboxed objects */
3563 case BIGNUM_WIDETAG:
3564 #if N_WORD_BITS != 64
3565 case SINGLE_FLOAT_WIDETAG:
3567 case DOUBLE_FLOAT_WIDETAG:
3568 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3569 case LONG_FLOAT_WIDETAG:
3571 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3572 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3574 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3575 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3577 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3578 case COMPLEX_LONG_FLOAT_WIDETAG:
3580 case SIMPLE_BASE_STRING_WIDETAG:
3581 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3582 case SIMPLE_CHARACTER_STRING_WIDETAG:
3584 case SIMPLE_BIT_VECTOR_WIDETAG:
3585 case SIMPLE_ARRAY_NIL_WIDETAG:
3586 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3587 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3588 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3589 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3590 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3591 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3592 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
3593 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
3595 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3596 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3597 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
3598 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
3600 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3601 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3603 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3604 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3606 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3607 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3609 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3610 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3612 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
3613 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
3615 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3616 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3618 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
3619 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
3621 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3622 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3624 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3625 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3626 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3627 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3629 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3630 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3632 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3633 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3635 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3636 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3639 case WEAK_POINTER_WIDETAG:
3640 #ifdef LUTEX_WIDETAG
3643 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3644 case NO_TLS_VALUE_MARKER_WIDETAG:
3646 count = (sizetab[widetag_of(*start)])(start);
3650 lose("Unhandled widetag %p at %p\n",
3651 widetag_of(*start), start);
3663 /* FIXME: It would be nice to make names consistent so that
3664 * foo_size meant size *in* *bytes* instead of size in some
3665 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3666 * Some counts of lispobjs are called foo_count; it might be good
3667 * to grep for all foo_size and rename the appropriate ones to
3669 long read_only_space_size =
3670 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3671 - (lispobj*)READ_ONLY_SPACE_START;
3672 long static_space_size =
3673 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3674 - (lispobj*)STATIC_SPACE_START;
3676 for_each_thread(th) {
3677 long binding_stack_size =
3678 (lispobj*)get_binding_stack_pointer(th)
3679 - (lispobj*)th->binding_stack_start;
3680 verify_space(th->binding_stack_start, binding_stack_size);
3682 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3683 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3687 verify_generation(generation_index_t generation)
3691 for (i = 0; i < last_free_page; i++) {
3692 if (page_allocated_p(i)
3693 && (page_table[i].bytes_used != 0)
3694 && (page_table[i].gen == generation)) {
3695 page_index_t last_page;
3696 int region_allocation = page_table[i].allocated;
3698 /* This should be the start of a contiguous block */
3699 gc_assert(page_table[i].region_start_offset == 0);
3701 /* Need to find the full extent of this contiguous block in case
3702 objects span pages. */
3704 /* Now work forward until the end of this contiguous area is
3706 for (last_page = i; ;last_page++)
3707 /* Check whether this is the last page in this contiguous
3709 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3710 /* Or it is PAGE_BYTES and is the last in the block */
3711 || (page_table[last_page+1].allocated != region_allocation)
3712 || (page_table[last_page+1].bytes_used == 0)
3713 || (page_table[last_page+1].gen != generation)
3714 || (page_table[last_page+1].region_start_offset == 0))
3717 verify_space(page_address(i),
3719 (page_table[last_page].bytes_used
3720 + npage_bytes(last_page-i)))
3727 /* Check that all the free space is zero filled. */
3729 verify_zero_fill(void)
3733 for (page = 0; page < last_free_page; page++) {
3734 if (page_free_p(page)) {
3735 /* The whole page should be zero filled. */
3736 long *start_addr = (long *)page_address(page);
3739 for (i = 0; i < size; i++) {
3740 if (start_addr[i] != 0) {
3741 lose("free page not zero at %x\n", start_addr + i);
3745 long free_bytes = PAGE_BYTES - page_table[page].bytes_used;
3746 if (free_bytes > 0) {
3747 long *start_addr = (long *)((unsigned long)page_address(page)
3748 + page_table[page].bytes_used);
3749 long size = free_bytes / N_WORD_BYTES;
3751 for (i = 0; i < size; i++) {
3752 if (start_addr[i] != 0) {
3753 lose("free region not zero at %x\n", start_addr + i);
3761 /* External entry point for verify_zero_fill */
3763 gencgc_verify_zero_fill(void)
3765 /* Flush the alloc regions updating the tables. */
3766 gc_alloc_update_all_page_tables();
3767 SHOW("verifying zero fill");
3772 verify_dynamic_space(void)
3774 generation_index_t i;
3776 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3777 verify_generation(i);
3779 if (gencgc_enable_verify_zero_fill)
3783 /* Write-protect all the dynamic boxed pages in the given generation. */
3785 write_protect_generation_pages(generation_index_t generation)
3789 gc_assert(generation < SCRATCH_GENERATION);
3791 for (start = 0; start < last_free_page; start++) {
3792 if (protect_page_p(start, generation)) {
3796 /* Note the page as protected in the page tables. */
3797 page_table[start].write_protected = 1;
3799 for (last = start + 1; last < last_free_page; last++) {
3800 if (!protect_page_p(last, generation))
3802 page_table[last].write_protected = 1;
3805 page_start = (void *)page_address(start);
3807 os_protect(page_start,
3808 npage_bytes(last - start),
3809 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3815 if (gencgc_verbose > 1) {
3817 "/write protected %d of %d pages in generation %d\n",
3818 count_write_protect_generation_pages(generation),
3819 count_generation_pages(generation),
3824 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3827 scavenge_control_stack()
3829 unsigned long control_stack_size;
3831 /* This is going to be a big problem when we try to port threads
3833 struct thread *th = arch_os_get_current_thread();
3834 lispobj *control_stack =
3835 (lispobj *)(th->control_stack_start);
3837 control_stack_size = current_control_stack_pointer - control_stack;
3838 scavenge(control_stack, control_stack_size);
3840 /* Scrub the unscavenged control stack space, so that we can't run
3841 * into any stale pointers in a later GC. */
3842 scrub_control_stack();
3845 /* Scavenging Interrupt Contexts */
3847 static int boxed_registers[] = BOXED_REGISTERS;
3850 scavenge_interrupt_context(os_context_t * context)
3856 unsigned long lip_offset;
3857 int lip_register_pair;
3859 unsigned long pc_code_offset;
3861 #ifdef ARCH_HAS_LINK_REGISTER
3862 unsigned long lr_code_offset;
3864 #ifdef ARCH_HAS_NPC_REGISTER
3865 unsigned long npc_code_offset;
3869 /* Find the LIP's register pair and calculate it's offset */
3870 /* before we scavenge the context. */
3873 * I (RLT) think this is trying to find the boxed register that is
3874 * closest to the LIP address, without going past it. Usually, it's
3875 * reg_CODE or reg_LRA. But sometimes, nothing can be found.
3877 lip = *os_context_register_addr(context, reg_LIP);
3878 lip_offset = 0x7FFFFFFF;
3879 lip_register_pair = -1;
3880 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3885 index = boxed_registers[i];
3886 reg = *os_context_register_addr(context, index);
3887 if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
3889 if (offset < lip_offset) {
3890 lip_offset = offset;
3891 lip_register_pair = index;
3895 #endif /* reg_LIP */
3897 /* Compute the PC's offset from the start of the CODE */
3899 pc_code_offset = *os_context_pc_addr(context)
3900 - *os_context_register_addr(context, reg_CODE);
3901 #ifdef ARCH_HAS_NPC_REGISTER
3902 npc_code_offset = *os_context_npc_addr(context)
3903 - *os_context_register_addr(context, reg_CODE);
3904 #endif /* ARCH_HAS_NPC_REGISTER */
3906 #ifdef ARCH_HAS_LINK_REGISTER
3908 *os_context_lr_addr(context) -
3909 *os_context_register_addr(context, reg_CODE);
3912 /* Scanvenge all boxed registers in the context. */
3913 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3917 index = boxed_registers[i];
3918 foo = *os_context_register_addr(context, index);
3920 *os_context_register_addr(context, index) = foo;
3922 scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
3929 * But what happens if lip_register_pair is -1?
3930 * *os_context_register_addr on Solaris (see
3931 * solaris_register_address in solaris-os.c) will return
3932 * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
3933 * that what we really want? My guess is that that is not what we
3934 * want, so if lip_register_pair is -1, we don't touch reg_LIP at
3935 * all. But maybe it doesn't really matter if LIP is trashed?
3937 if (lip_register_pair >= 0) {
3938 *os_context_register_addr(context, reg_LIP) =
3939 *os_context_register_addr(context, lip_register_pair)
3942 #endif /* reg_LIP */
3944 /* Fix the PC if it was in from space */
3945 if (from_space_p(*os_context_pc_addr(context)))
3946 *os_context_pc_addr(context) =
3947 *os_context_register_addr(context, reg_CODE) + pc_code_offset;
3949 #ifdef ARCH_HAS_LINK_REGISTER
3950 /* Fix the LR ditto; important if we're being called from
3951 * an assembly routine that expects to return using blr, otherwise
3953 if (from_space_p(*os_context_lr_addr(context)))
3954 *os_context_lr_addr(context) =
3955 *os_context_register_addr(context, reg_CODE) + lr_code_offset;
3958 #ifdef ARCH_HAS_NPC_REGISTER
3959 if (from_space_p(*os_context_npc_addr(context)))
3960 *os_context_npc_addr(context) =
3961 *os_context_register_addr(context, reg_CODE) + npc_code_offset;
3962 #endif /* ARCH_HAS_NPC_REGISTER */
3966 scavenge_interrupt_contexts(void)
3969 os_context_t *context;
3971 struct thread *th=arch_os_get_current_thread();
3973 index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
3975 #if defined(DEBUG_PRINT_CONTEXT_INDEX)
3976 printf("Number of active contexts: %d\n", index);
3979 for (i = 0; i < index; i++) {
3980 context = th->interrupt_contexts[i];
3981 scavenge_interrupt_context(context);
3987 #if defined(LISP_FEATURE_SB_THREAD)
3989 preserve_context_registers (os_context_t *c)
3992 /* On Darwin the signal context isn't a contiguous block of memory,
3993 * so just preserve_pointering its contents won't be sufficient.
3995 #if defined(LISP_FEATURE_DARWIN)
3996 #if defined LISP_FEATURE_X86
3997 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3998 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3999 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
4000 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
4001 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
4002 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
4003 preserve_pointer((void*)*os_context_pc_addr(c));
4004 #elif defined LISP_FEATURE_X86_64
4005 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
4006 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
4007 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
4008 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
4009 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
4010 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
4011 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
4012 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
4013 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
4014 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
4015 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
4016 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
4017 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
4018 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
4019 preserve_pointer((void*)*os_context_pc_addr(c));
4021 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
4024 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
4025 preserve_pointer(*ptr);
4030 /* Garbage collect a generation. If raise is 0 then the remains of the
4031 * generation are not raised to the next generation. */
4033 garbage_collect_generation(generation_index_t generation, int raise)
4035 unsigned long bytes_freed;
4037 unsigned long static_space_size;
4038 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4041 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
4043 /* The oldest generation can't be raised. */
4044 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
4046 /* Check if weak hash tables were processed in the previous GC. */
4047 gc_assert(weak_hash_tables == NULL);
4049 /* Initialize the weak pointer list. */
4050 weak_pointers = NULL;
4052 #ifdef LUTEX_WIDETAG
4053 unmark_lutexes(generation);
4056 /* When a generation is not being raised it is transported to a
4057 * temporary generation (NUM_GENERATIONS), and lowered when
4058 * done. Set up this new generation. There should be no pages
4059 * allocated to it yet. */
4061 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
4064 /* Set the global src and dest. generations */
4065 from_space = generation;
4067 new_space = generation+1;
4069 new_space = SCRATCH_GENERATION;
4071 /* Change to a new space for allocation, resetting the alloc_start_page */
4072 gc_alloc_generation = new_space;
4073 generations[new_space].alloc_start_page = 0;
4074 generations[new_space].alloc_unboxed_start_page = 0;
4075 generations[new_space].alloc_large_start_page = 0;
4076 generations[new_space].alloc_large_unboxed_start_page = 0;
4078 /* Before any pointers are preserved, the dont_move flags on the
4079 * pages need to be cleared. */
4080 for (i = 0; i < last_free_page; i++)
4081 if(page_table[i].gen==from_space)
4082 page_table[i].dont_move = 0;
4084 /* Un-write-protect the old-space pages. This is essential for the
4085 * promoted pages as they may contain pointers into the old-space
4086 * which need to be scavenged. It also helps avoid unnecessary page
4087 * faults as forwarding pointers are written into them. They need to
4088 * be un-protected anyway before unmapping later. */
4089 unprotect_oldspace();
4091 /* Scavenge the stacks' conservative roots. */
4093 /* there are potentially two stacks for each thread: the main
4094 * stack, which may contain Lisp pointers, and the alternate stack.
4095 * We don't ever run Lisp code on the altstack, but it may
4096 * host a sigcontext with lisp objects in it */
4098 /* what we need to do: (1) find the stack pointer for the main
4099 * stack; scavenge it (2) find the interrupt context on the
4100 * alternate stack that might contain lisp values, and scavenge
4103 /* we assume that none of the preceding applies to the thread that
4104 * initiates GC. If you ever call GC from inside an altstack
4105 * handler, you will lose. */
4107 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4108 /* And if we're saving a core, there's no point in being conservative. */
4109 if (conservative_stack) {
4110 for_each_thread(th) {
4112 void **esp=(void **)-1;
4113 #ifdef LISP_FEATURE_SB_THREAD
4115 if(th==arch_os_get_current_thread()) {
4116 /* Somebody is going to burn in hell for this, but casting
4117 * it in two steps shuts gcc up about strict aliasing. */
4118 esp = (void **)((void *)&raise);
4121 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
4122 for(i=free-1;i>=0;i--) {
4123 os_context_t *c=th->interrupt_contexts[i];
4124 esp1 = (void **) *os_context_register_addr(c,reg_SP);
4125 if (esp1>=(void **)th->control_stack_start &&
4126 esp1<(void **)th->control_stack_end) {
4127 if(esp1<esp) esp=esp1;
4128 preserve_context_registers(c);
4133 esp = (void **)((void *)&raise);
4135 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
4136 preserve_pointer(*ptr);
4143 if (gencgc_verbose > 1) {
4144 long num_dont_move_pages = count_dont_move_pages();
4146 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
4147 num_dont_move_pages,
4148 npage_bytes(num_dont_move_pages));
4152 /* Scavenge all the rest of the roots. */
4154 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4156 * If not x86, we need to scavenge the interrupt context(s) and the
4159 scavenge_interrupt_contexts();
4160 scavenge_control_stack();
4163 /* Scavenge the Lisp functions of the interrupt handlers, taking
4164 * care to avoid SIG_DFL and SIG_IGN. */
4165 for (i = 0; i < NSIG; i++) {
4166 union interrupt_handler handler = interrupt_handlers[i];
4167 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4168 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4169 scavenge((lispobj *)(interrupt_handlers + i), 1);
4172 /* Scavenge the binding stacks. */
4175 for_each_thread(th) {
4176 long len= (lispobj *)get_binding_stack_pointer(th) -
4177 th->binding_stack_start;
4178 scavenge((lispobj *) th->binding_stack_start,len);
4179 #ifdef LISP_FEATURE_SB_THREAD
4180 /* do the tls as well */
4181 len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
4182 (sizeof (struct thread))/(sizeof (lispobj));
4183 scavenge((lispobj *) (th+1),len);
4188 /* The original CMU CL code had scavenge-read-only-space code
4189 * controlled by the Lisp-level variable
4190 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4191 * wasn't documented under what circumstances it was useful or
4192 * safe to turn it on, so it's been turned off in SBCL. If you
4193 * want/need this functionality, and can test and document it,
4194 * please submit a patch. */
4196 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4197 unsigned long read_only_space_size =
4198 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4199 (lispobj*)READ_ONLY_SPACE_START;
4201 "/scavenge read only space: %d bytes\n",
4202 read_only_space_size * sizeof(lispobj)));
4203 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4207 /* Scavenge static space. */
4209 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4210 (lispobj *)STATIC_SPACE_START;
4211 if (gencgc_verbose > 1) {
4213 "/scavenge static space: %d bytes\n",
4214 static_space_size * sizeof(lispobj)));
4216 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4218 /* All generations but the generation being GCed need to be
4219 * scavenged. The new_space generation needs special handling as
4220 * objects may be moved in - it is handled separately below. */
4221 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4223 /* Finally scavenge the new_space generation. Keep going until no
4224 * more objects are moved into the new generation */
4225 scavenge_newspace_generation(new_space);
4227 /* FIXME: I tried reenabling this check when debugging unrelated
4228 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4229 * Since the current GC code seems to work well, I'm guessing that
4230 * this debugging code is just stale, but I haven't tried to
4231 * figure it out. It should be figured out and then either made to
4232 * work or just deleted. */
4233 #define RESCAN_CHECK 0
4235 /* As a check re-scavenge the newspace once; no new objects should
4238 long old_bytes_allocated = bytes_allocated;
4239 long bytes_allocated;
4241 /* Start with a full scavenge. */
4242 scavenge_newspace_generation_one_scan(new_space);
4244 /* Flush the current regions, updating the tables. */
4245 gc_alloc_update_all_page_tables();
4247 bytes_allocated = bytes_allocated - old_bytes_allocated;
4249 if (bytes_allocated != 0) {
4250 lose("Rescan of new_space allocated %d more bytes.\n",
4256 scan_weak_hash_tables();
4257 scan_weak_pointers();
4259 /* Flush the current regions, updating the tables. */
4260 gc_alloc_update_all_page_tables();
4262 /* Free the pages in oldspace, but not those marked dont_move. */
4263 bytes_freed = free_oldspace();
4265 /* If the GC is not raising the age then lower the generation back
4266 * to its normal generation number */
4268 for (i = 0; i < last_free_page; i++)
4269 if ((page_table[i].bytes_used != 0)
4270 && (page_table[i].gen == SCRATCH_GENERATION))
4271 page_table[i].gen = generation;
4272 gc_assert(generations[generation].bytes_allocated == 0);
4273 generations[generation].bytes_allocated =
4274 generations[SCRATCH_GENERATION].bytes_allocated;
4275 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4278 /* Reset the alloc_start_page for generation. */
4279 generations[generation].alloc_start_page = 0;
4280 generations[generation].alloc_unboxed_start_page = 0;
4281 generations[generation].alloc_large_start_page = 0;
4282 generations[generation].alloc_large_unboxed_start_page = 0;
4284 if (generation >= verify_gens) {
4285 if (gencgc_verbose) {
4289 verify_dynamic_space();
4292 /* Set the new gc trigger for the GCed generation. */
4293 generations[generation].gc_trigger =
4294 generations[generation].bytes_allocated
4295 + generations[generation].bytes_consed_between_gc;
4298 generations[generation].num_gc = 0;
4300 ++generations[generation].num_gc;
4302 #ifdef LUTEX_WIDETAG
4303 reap_lutexes(generation);
4305 move_lutexes(generation, generation+1);
4309 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4311 update_dynamic_space_free_pointer(void)
4313 page_index_t last_page = -1, i;
4315 for (i = 0; i < last_free_page; i++)
4316 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
4319 last_free_page = last_page+1;
4321 set_alloc_pointer((lispobj)(page_address(last_free_page)));
4322 return 0; /* dummy value: return something ... */
4326 remap_free_pages (page_index_t from, page_index_t to)
4328 page_index_t first_page, last_page;
4330 for (first_page = from; first_page <= to; first_page++) {
4331 if (page_allocated_p(first_page) ||
4332 (page_table[first_page].need_to_zero == 0)) {
4336 last_page = first_page + 1;
4337 while (page_free_p(last_page) &&
4339 (page_table[last_page].need_to_zero == 1)) {
4343 /* There's a mysterious Solaris/x86 problem with using mmap
4344 * tricks for memory zeroing. See sbcl-devel thread
4345 * "Re: patch: standalone executable redux".
4347 #if defined(LISP_FEATURE_SUNOS)
4348 zero_pages(first_page, last_page-1);
4350 zero_pages_with_mmap(first_page, last_page-1);
4353 first_page = last_page;
4357 generation_index_t small_generation_limit = 1;
4359 /* GC all generations newer than last_gen, raising the objects in each
4360 * to the next older generation - we finish when all generations below
4361 * last_gen are empty. Then if last_gen is due for a GC, or if
4362 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4363 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4365 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4366 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4368 collect_garbage(generation_index_t last_gen)
4370 generation_index_t gen = 0, i;
4373 /* The largest value of last_free_page seen since the time
4374 * remap_free_pages was called. */
4375 static page_index_t high_water_mark = 0;
4377 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4381 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4383 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4388 /* Flush the alloc regions updating the tables. */
4389 gc_alloc_update_all_page_tables();
4391 /* Verify the new objects created by Lisp code. */
4392 if (pre_verify_gen_0) {
4393 FSHOW((stderr, "pre-checking generation 0\n"));
4394 verify_generation(0);
4397 if (gencgc_verbose > 1)
4398 print_generation_stats();
4401 /* Collect the generation. */
4403 if (gen >= gencgc_oldest_gen_to_gc) {
4404 /* Never raise the oldest generation. */
4409 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
4412 if (gencgc_verbose > 1) {
4414 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4417 generations[gen].bytes_allocated,
4418 generations[gen].gc_trigger,
4419 generations[gen].num_gc));
4422 /* If an older generation is being filled, then update its
4425 generations[gen+1].cum_sum_bytes_allocated +=
4426 generations[gen+1].bytes_allocated;
4429 garbage_collect_generation(gen, raise);
4431 /* Reset the memory age cum_sum. */
4432 generations[gen].cum_sum_bytes_allocated = 0;
4434 if (gencgc_verbose > 1) {
4435 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4436 print_generation_stats();
4440 } while ((gen <= gencgc_oldest_gen_to_gc)
4441 && ((gen < last_gen)
4442 || ((gen <= gencgc_oldest_gen_to_gc)
4444 && (generations[gen].bytes_allocated
4445 > generations[gen].gc_trigger)
4446 && (generation_average_age(gen)
4447 > generations[gen].minimum_age_before_gc))));
4449 /* Now if gen-1 was raised all generations before gen are empty.
4450 * If it wasn't raised then all generations before gen-1 are empty.
4452 * Now objects within this gen's pages cannot point to younger
4453 * generations unless they are written to. This can be exploited
4454 * by write-protecting the pages of gen; then when younger
4455 * generations are GCed only the pages which have been written
4460 gen_to_wp = gen - 1;
4462 /* There's not much point in WPing pages in generation 0 as it is
4463 * never scavenged (except promoted pages). */
4464 if ((gen_to_wp > 0) && enable_page_protection) {
4465 /* Check that they are all empty. */
4466 for (i = 0; i < gen_to_wp; i++) {
4467 if (generations[i].bytes_allocated)
4468 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4471 write_protect_generation_pages(gen_to_wp);
4474 /* Set gc_alloc() back to generation 0. The current regions should
4475 * be flushed after the above GCs. */
4476 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4477 gc_alloc_generation = 0;
4479 /* Save the high-water mark before updating last_free_page */
4480 if (last_free_page > high_water_mark)
4481 high_water_mark = last_free_page;
4483 update_dynamic_space_free_pointer();
4485 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4487 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4490 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4493 if (gen > small_generation_limit) {
4494 if (last_free_page > high_water_mark)
4495 high_water_mark = last_free_page;
4496 remap_free_pages(0, high_water_mark);
4497 high_water_mark = 0;
4502 SHOW("returning from collect_garbage");
4505 /* This is called by Lisp PURIFY when it is finished. All live objects
4506 * will have been moved to the RO and Static heaps. The dynamic space
4507 * will need a full re-initialization. We don't bother having Lisp
4508 * PURIFY flush the current gc_alloc() region, as the page_tables are
4509 * re-initialized, and every page is zeroed to be sure. */
4515 if (gencgc_verbose > 1) {
4516 SHOW("entering gc_free_heap");
4519 for (page = 0; page < page_table_pages; page++) {
4520 /* Skip free pages which should already be zero filled. */
4521 if (page_allocated_p(page)) {
4522 void *page_start, *addr;
4524 /* Mark the page free. The other slots are assumed invalid
4525 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4526 * should not be write-protected -- except that the
4527 * generation is used for the current region but it sets
4529 page_table[page].allocated = FREE_PAGE_FLAG;
4530 page_table[page].bytes_used = 0;
4532 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4533 * about this change. */
4534 /* Zero the page. */
4535 page_start = (void *)page_address(page);
4537 /* First, remove any write-protection. */
4538 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
4539 page_table[page].write_protected = 0;
4541 os_invalidate(page_start,PAGE_BYTES);
4542 addr = os_validate(page_start,PAGE_BYTES);
4543 if (addr == NULL || addr != page_start) {
4544 lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n",
4549 page_table[page].write_protected = 0;
4551 } else if (gencgc_zero_check_during_free_heap) {
4552 /* Double-check that the page is zero filled. */
4555 gc_assert(page_free_p(page));
4556 gc_assert(page_table[page].bytes_used == 0);
4557 page_start = (long *)page_address(page);
4558 for (i=0; i<1024; i++) {
4559 if (page_start[i] != 0) {
4560 lose("free region not zero at %x\n", page_start + i);
4566 bytes_allocated = 0;
4568 /* Initialize the generations. */
4569 for (page = 0; page < NUM_GENERATIONS; page++) {
4570 generations[page].alloc_start_page = 0;
4571 generations[page].alloc_unboxed_start_page = 0;
4572 generations[page].alloc_large_start_page = 0;
4573 generations[page].alloc_large_unboxed_start_page = 0;
4574 generations[page].bytes_allocated = 0;
4575 generations[page].gc_trigger = 2000000;
4576 generations[page].num_gc = 0;
4577 generations[page].cum_sum_bytes_allocated = 0;
4578 generations[page].lutexes = NULL;
4581 if (gencgc_verbose > 1)
4582 print_generation_stats();
4584 /* Initialize gc_alloc(). */
4585 gc_alloc_generation = 0;
4587 gc_set_region_empty(&boxed_region);
4588 gc_set_region_empty(&unboxed_region);
4591 set_alloc_pointer((lispobj)((char *)heap_base));
4593 if (verify_after_free_heap) {
4594 /* Check whether purify has left any bad pointers. */
4595 FSHOW((stderr, "checking after free_heap\n"));
4605 /* Compute the number of pages needed for the dynamic space.
4606 * Dynamic space size should be aligned on page size. */
4607 page_table_pages = dynamic_space_size/PAGE_BYTES;
4608 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4610 page_table = calloc(page_table_pages, sizeof(struct page));
4611 gc_assert(page_table);
4614 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4615 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4617 #ifdef LUTEX_WIDETAG
4618 scavtab[LUTEX_WIDETAG] = scav_lutex;
4619 transother[LUTEX_WIDETAG] = trans_lutex;
4620 sizetab[LUTEX_WIDETAG] = size_lutex;
4623 heap_base = (void*)DYNAMIC_SPACE_START;
4625 /* Initialize each page structure. */
4626 for (i = 0; i < page_table_pages; i++) {
4627 /* Initialize all pages as free. */
4628 page_table[i].allocated = FREE_PAGE_FLAG;
4629 page_table[i].bytes_used = 0;
4631 /* Pages are not write-protected at startup. */
4632 page_table[i].write_protected = 0;
4635 bytes_allocated = 0;
4637 /* Initialize the generations.
4639 * FIXME: very similar to code in gc_free_heap(), should be shared */
4640 for (i = 0; i < NUM_GENERATIONS; i++) {
4641 generations[i].alloc_start_page = 0;
4642 generations[i].alloc_unboxed_start_page = 0;
4643 generations[i].alloc_large_start_page = 0;
4644 generations[i].alloc_large_unboxed_start_page = 0;
4645 generations[i].bytes_allocated = 0;
4646 generations[i].gc_trigger = 2000000;
4647 generations[i].num_gc = 0;
4648 generations[i].cum_sum_bytes_allocated = 0;
4649 /* the tune-able parameters */
4650 generations[i].bytes_consed_between_gc = 2000000;
4651 generations[i].number_of_gcs_before_promotion = 1;
4652 generations[i].minimum_age_before_gc = 0.75;
4653 generations[i].lutexes = NULL;
4656 /* Initialize gc_alloc. */
4657 gc_alloc_generation = 0;
4658 gc_set_region_empty(&boxed_region);
4659 gc_set_region_empty(&unboxed_region);
4664 /* Pick up the dynamic space from after a core load.
4666 * The ALLOCATION_POINTER points to the end of the dynamic space.
4670 gencgc_pickup_dynamic(void)
4672 page_index_t page = 0;
4673 void *alloc_ptr = (void *)get_alloc_pointer();
4674 lispobj *prev=(lispobj *)page_address(page);
4675 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4677 lispobj *first,*ptr= (lispobj *)page_address(page);
4679 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4680 /* It is possible, though rare, for the saved page table
4681 * to contain free pages below alloc_ptr. */
4682 page_table[page].gen = gen;
4683 page_table[page].bytes_used = PAGE_BYTES;
4684 page_table[page].large_object = 0;
4685 page_table[page].write_protected = 0;
4686 page_table[page].write_protected_cleared = 0;
4687 page_table[page].dont_move = 0;
4688 page_table[page].need_to_zero = 1;
4691 if (!gencgc_partial_pickup) {
4692 page_table[page].allocated = BOXED_PAGE_FLAG;
4693 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4696 page_table[page].region_start_offset =
4697 page_address(page) - (void *)prev;
4700 } while (page_address(page) < alloc_ptr);
4702 #ifdef LUTEX_WIDETAG
4703 /* Lutexes have been registered in generation 0 by coreparse, and
4704 * need to be moved to the right one manually.
4706 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4709 last_free_page = page;
4711 generations[gen].bytes_allocated = npage_bytes(page);
4712 bytes_allocated = npage_bytes(page);
4714 gc_alloc_update_all_page_tables();
4715 write_protect_generation_pages(gen);
4719 gc_initialize_pointers(void)
4721 gencgc_pickup_dynamic();
4725 /* alloc(..) is the external interface for memory allocation. It
4726 * allocates to generation 0. It is not called from within the garbage
4727 * collector as it is only external uses that need the check for heap
4728 * size (GC trigger) and to disable the interrupts (interrupts are
4729 * always disabled during a GC).
4731 * The vops that call alloc(..) assume that the returned space is zero-filled.
4732 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4734 * The check for a GC trigger is only performed when the current
4735 * region is full, so in most cases it's not needed. */
4737 static inline lispobj *
4738 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4739 struct thread *thread)
4741 #ifndef LISP_FEATURE_WIN32
4742 lispobj alloc_signal;
4745 void *new_free_pointer;
4747 gc_assert(nbytes>0);
4749 /* Check for alignment allocation problems. */
4750 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4751 && ((nbytes & LOWTAG_MASK) == 0));
4753 /* Must be inside a PA section. */
4754 gc_assert(get_pseudo_atomic_atomic(thread));
4756 /* maybe we can do this quickly ... */
4757 new_free_pointer = region->free_pointer + nbytes;
4758 if (new_free_pointer <= region->end_addr) {
4759 new_obj = (void*)(region->free_pointer);
4760 region->free_pointer = new_free_pointer;
4761 return(new_obj); /* yup */
4764 /* we have to go the long way around, it seems. Check whether we
4765 * should GC in the near future
4767 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4768 /* Don't flood the system with interrupts if the need to gc is
4769 * already noted. This can happen for example when SUB-GC
4770 * allocates or after a gc triggered in a WITHOUT-GCING. */
4771 if (SymbolValue(GC_PENDING,thread) == NIL) {
4772 /* set things up so that GC happens when we finish the PA
4774 SetSymbolValue(GC_PENDING,T,thread);
4775 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4776 set_pseudo_atomic_interrupted(thread);
4777 #ifdef LISP_FEATURE_PPC
4778 /* PPC calls alloc() from a trap or from pa_alloc(),
4779 * look up the most context if it's from a trap. */
4781 os_context_t *context =
4782 thread->interrupt_data->allocation_trap_context;
4783 maybe_save_gc_mask_and_block_deferrables
4784 (context ? os_context_sigmask_addr(context) : NULL);
4787 maybe_save_gc_mask_and_block_deferrables(NULL);
4792 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4794 #ifndef LISP_FEATURE_WIN32
4795 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4796 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4797 if ((signed long) alloc_signal <= 0) {
4798 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4801 SetSymbolValue(ALLOC_SIGNAL,
4802 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4812 general_alloc(long nbytes, int page_type_flag)
4814 struct thread *thread = arch_os_get_current_thread();
4815 /* Select correct region, and call general_alloc_internal with it.
4816 * For other then boxed allocation we must lock first, since the
4817 * region is shared. */
4818 if (BOXED_PAGE_FLAG & page_type_flag) {
4819 #ifdef LISP_FEATURE_SB_THREAD
4820 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4822 struct alloc_region *region = &boxed_region;
4824 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4825 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4827 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4828 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4829 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4832 lose("bad page type flag: %d", page_type_flag);
4839 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4840 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4844 * shared support for the OS-dependent signal handlers which
4845 * catch GENCGC-related write-protect violations
4847 void unhandled_sigmemoryfault(void* addr);
4849 /* Depending on which OS we're running under, different signals might
4850 * be raised for a violation of write protection in the heap. This
4851 * function factors out the common generational GC magic which needs
4852 * to invoked in this case, and should be called from whatever signal
4853 * handler is appropriate for the OS we're running under.
4855 * Return true if this signal is a normal generational GC thing that
4856 * we were able to handle, or false if it was abnormal and control
4857 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4860 gencgc_handle_wp_violation(void* fault_addr)
4862 page_index_t page_index = find_page_index(fault_addr);
4865 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4866 fault_addr, page_index));
4869 /* Check whether the fault is within the dynamic space. */
4870 if (page_index == (-1)) {
4872 /* It can be helpful to be able to put a breakpoint on this
4873 * case to help diagnose low-level problems. */
4874 unhandled_sigmemoryfault(fault_addr);
4876 /* not within the dynamic space -- not our responsibility */
4881 ret = thread_mutex_lock(&free_pages_lock);
4882 gc_assert(ret == 0);
4883 if (page_table[page_index].write_protected) {
4884 /* Unprotect the page. */
4885 os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
4886 page_table[page_index].write_protected_cleared = 1;
4887 page_table[page_index].write_protected = 0;
4889 /* The only acceptable reason for this signal on a heap
4890 * access is that GENCGC write-protected the page.
4891 * However, if two CPUs hit a wp page near-simultaneously,
4892 * we had better not have the second one lose here if it
4893 * does this test after the first one has already set wp=0
4895 if(page_table[page_index].write_protected_cleared != 1)
4896 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4897 page_index, boxed_region.first_page,
4898 boxed_region.last_page);
4900 ret = thread_mutex_unlock(&free_pages_lock);
4901 gc_assert(ret == 0);
4902 /* Don't worry, we can handle it. */
4906 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4907 * it's not just a case of the program hitting the write barrier, and
4908 * are about to let Lisp deal with it. It's basically just a
4909 * convenient place to set a gdb breakpoint. */
4911 unhandled_sigmemoryfault(void *addr)
4914 void gc_alloc_update_all_page_tables(void)
4916 /* Flush the alloc regions updating the tables. */
4919 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4920 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4921 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4925 gc_set_region_empty(struct alloc_region *region)
4927 region->first_page = 0;
4928 region->last_page = -1;
4929 region->start_addr = page_address(0);
4930 region->free_pointer = page_address(0);
4931 region->end_addr = page_address(0);
4935 zero_all_free_pages()
4939 for (i = 0; i < last_free_page; i++) {
4940 if (page_free_p(i)) {
4941 #ifdef READ_PROTECT_FREE_PAGES
4942 os_protect(page_address(i),
4951 /* Things to do before doing a final GC before saving a core (without
4954 * + Pages in large_object pages aren't moved by the GC, so we need to
4955 * unset that flag from all pages.
4956 * + The pseudo-static generation isn't normally collected, but it seems
4957 * reasonable to collect it at least when saving a core. So move the
4958 * pages to a normal generation.
4961 prepare_for_final_gc ()
4964 for (i = 0; i < last_free_page; i++) {
4965 page_table[i].large_object = 0;
4966 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4967 int used = page_table[i].bytes_used;
4968 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4969 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4970 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4976 /* Do a non-conservative GC, and then save a core with the initial
4977 * function being set to the value of the static symbol
4978 * SB!VM:RESTART-LISP-FUNCTION */
4980 gc_and_save(char *filename, boolean prepend_runtime,
4981 boolean save_runtime_options)
4984 void *runtime_bytes = NULL;
4985 size_t runtime_size;
4987 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4992 conservative_stack = 0;
4994 /* The filename might come from Lisp, and be moved by the now
4995 * non-conservative GC. */
4996 filename = strdup(filename);
4998 /* Collect twice: once into relatively high memory, and then back
4999 * into low memory. This compacts the retained data into the lower
5000 * pages, minimizing the size of the core file.
5002 prepare_for_final_gc();
5003 gencgc_alloc_start_page = last_free_page;
5004 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
5006 prepare_for_final_gc();
5007 gencgc_alloc_start_page = -1;
5008 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
5010 if (prepend_runtime)
5011 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
5013 /* The dumper doesn't know that pages need to be zeroed before use. */
5014 zero_all_free_pages();
5015 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
5016 prepend_runtime, save_runtime_options);
5017 /* Oops. Save still managed to fail. Since we've mangled the stack
5018 * beyond hope, there's not much we can do.
5019 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
5020 * going to be rather unsatisfactory too... */
5021 lose("Attempt to save core after non-conservative GC failed.\n");