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"
45 #include "genesis/vector.h"
46 #include "genesis/weak-pointer.h"
47 #include "genesis/fdefn.h"
48 #include "genesis/simple-fun.h"
50 #include "genesis/hash-table.h"
51 #include "genesis/instance.h"
52 #include "genesis/layout.h"
54 #if defined(LUTEX_WIDETAG)
55 #include "pthread-lutex.h"
58 /* forward declarations */
59 page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
67 /* Generations 0-5 are normal collected generations, 6 is only used as
68 * scratch space by the collector, and should never get collected.
71 HIGHEST_NORMAL_GENERATION = 5,
72 PSEUDO_STATIC_GENERATION,
77 /* Should we use page protection to help avoid the scavenging of pages
78 * that don't have pointers to younger generations? */
79 boolean enable_page_protection = 1;
81 /* the minimum size (in bytes) for a large object*/
82 long large_object_size = 4 * PAGE_BYTES;
89 /* the verbosity level. All non-error messages are disabled at level 0;
90 * and only a few rare messages are printed at level 1. */
92 boolean gencgc_verbose = 1;
94 boolean gencgc_verbose = 0;
97 /* FIXME: At some point enable the various error-checking things below
98 * and see what they say. */
100 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
101 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
103 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
105 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
106 boolean pre_verify_gen_0 = 0;
108 /* Should we check for bad pointers after gc_free_heap is called
109 * from Lisp PURIFY? */
110 boolean verify_after_free_heap = 0;
112 /* Should we print a note when code objects are found in the dynamic space
113 * during a heap verify? */
114 boolean verify_dynamic_code_check = 0;
116 /* Should we check code objects for fixup errors after they are transported? */
117 boolean check_code_fixups = 0;
119 /* Should we check that newly allocated regions are zero filled? */
120 boolean gencgc_zero_check = 0;
122 /* Should we check that the free space is zero filled? */
123 boolean gencgc_enable_verify_zero_fill = 0;
125 /* Should we check that free pages are zero filled during gc_free_heap
126 * called after Lisp PURIFY? */
127 boolean gencgc_zero_check_during_free_heap = 0;
129 /* When loading a core, don't do a full scan of the memory for the
130 * memory region boundaries. (Set to true by coreparse.c if the core
131 * contained a pagetable entry).
133 boolean gencgc_partial_pickup = 0;
135 /* If defined, free pages are read-protected to ensure that nothing
139 /* #define READ_PROTECT_FREE_PAGES */
143 * GC structures and variables
146 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
147 unsigned long bytes_allocated = 0;
148 unsigned long auto_gc_trigger = 0;
150 /* the source and destination generations. These are set before a GC starts
152 generation_index_t from_space;
153 generation_index_t new_space;
155 /* Set to 1 when in GC */
156 boolean gc_active_p = 0;
158 /* should the GC be conservative on stack. If false (only right before
159 * saving a core), don't scan the stack / mark pages dont_move. */
160 static boolean conservative_stack = 1;
162 /* An array of page structures is allocated on gc initialization.
163 * This helps quickly map between an address its page structure.
164 * page_table_pages is set from the size of the dynamic space. */
165 page_index_t page_table_pages;
166 struct page *page_table;
168 static inline boolean page_allocated_p(page_index_t page) {
169 return (page_table[page].allocated != FREE_PAGE_FLAG);
172 static inline boolean page_no_region_p(page_index_t page) {
173 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
176 static inline boolean page_allocated_no_region_p(page_index_t page) {
177 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
178 && page_no_region_p(page));
181 static inline boolean page_free_p(page_index_t page) {
182 return (page_table[page].allocated == FREE_PAGE_FLAG);
185 static inline boolean page_boxed_p(page_index_t page) {
186 return (page_table[page].allocated & BOXED_PAGE_FLAG);
189 static inline boolean code_page_p(page_index_t page) {
190 return (page_table[page].allocated & CODE_PAGE_FLAG);
193 static inline boolean page_boxed_no_region_p(page_index_t page) {
194 return page_boxed_p(page) && page_no_region_p(page);
197 static inline boolean page_unboxed_p(page_index_t page) {
198 /* Both flags set == boxed code page */
199 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
200 && !page_boxed_p(page));
203 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
204 return (page_boxed_no_region_p(page)
205 && (page_table[page].bytes_used != 0)
206 && !page_table[page].dont_move
207 && (page_table[page].gen == generation));
210 /* To map addresses to page structures the address of the first page
212 static void *heap_base = NULL;
214 /* Calculate the start address for the given page number. */
216 page_address(page_index_t page_num)
218 return (heap_base + (page_num * PAGE_BYTES));
221 /* Calculate the address where the allocation region associated with
222 * the page starts. */
224 page_region_start(page_index_t page_index)
226 return page_address(page_index)-page_table[page_index].region_start_offset;
229 /* Find the page index within the page_table for the given
230 * address. Return -1 on failure. */
232 find_page_index(void *addr)
234 if (addr >= heap_base) {
235 page_index_t index = ((pointer_sized_uint_t)addr -
236 (pointer_sized_uint_t)heap_base) / PAGE_BYTES;
237 if (index < page_table_pages)
244 npage_bytes(long npages)
246 gc_assert(npages>=0);
247 return ((unsigned long)npages)*PAGE_BYTES;
250 /* Check that X is a higher address than Y and return offset from Y to
253 size_t void_diff(void *x, void *y)
256 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
259 /* a structure to hold the state of a generation */
262 /* the first page that gc_alloc() checks on its next call */
263 page_index_t alloc_start_page;
265 /* the first page that gc_alloc_unboxed() checks on its next call */
266 page_index_t alloc_unboxed_start_page;
268 /* the first page that gc_alloc_large (boxed) considers on its next
269 * call. (Although it always allocates after the boxed_region.) */
270 page_index_t alloc_large_start_page;
272 /* the first page that gc_alloc_large (unboxed) considers on its
273 * next call. (Although it always allocates after the
274 * current_unboxed_region.) */
275 page_index_t alloc_large_unboxed_start_page;
277 /* the bytes allocated to this generation */
278 unsigned long bytes_allocated;
280 /* the number of bytes at which to trigger a GC */
281 unsigned long gc_trigger;
283 /* to calculate a new level for gc_trigger */
284 unsigned long bytes_consed_between_gc;
286 /* the number of GCs since the last raise */
289 /* the average age after which a GC will raise objects to the
293 /* the cumulative sum of the bytes allocated to this generation. It is
294 * cleared after a GC on this generations, and update before new
295 * objects are added from a GC of a younger generation. Dividing by
296 * the bytes_allocated will give the average age of the memory in
297 * this generation since its last GC. */
298 unsigned long cum_sum_bytes_allocated;
300 /* a minimum average memory age before a GC will occur helps
301 * prevent a GC when a large number of new live objects have been
302 * added, in which case a GC could be a waste of time */
303 double min_av_mem_age;
305 /* A linked list of lutex structures in this generation, used for
306 * implementing lutex finalization. */
308 struct lutex *lutexes;
314 /* an array of generation structures. There needs to be one more
315 * generation structure than actual generations as the oldest
316 * generation is temporarily raised then lowered. */
317 struct generation generations[NUM_GENERATIONS];
319 /* the oldest generation that is will currently be GCed by default.
320 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
322 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
324 * Setting this to 0 effectively disables the generational nature of
325 * the GC. In some applications generational GC may not be useful
326 * because there are no long-lived objects.
328 * An intermediate value could be handy after moving long-lived data
329 * into an older generation so an unnecessary GC of this long-lived
330 * data can be avoided. */
331 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
333 /* The maximum free page in the heap is maintained and used to update
334 * ALLOCATION_POINTER which is used by the room function to limit its
335 * search of the heap. XX Gencgc obviously needs to be better
336 * integrated with the Lisp code. */
337 page_index_t last_free_page;
339 #ifdef LISP_FEATURE_SB_THREAD
340 /* This lock is to prevent multiple threads from simultaneously
341 * allocating new regions which overlap each other. Note that the
342 * majority of GC is single-threaded, but alloc() may be called from
343 * >1 thread at a time and must be thread-safe. This lock must be
344 * seized before all accesses to generations[] or to parts of
345 * page_table[] that other threads may want to see */
346 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
347 /* This lock is used to protect non-thread-local allocation. */
348 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
353 * miscellaneous heap functions
356 /* Count the number of pages which are write-protected within the
357 * given generation. */
359 count_write_protect_generation_pages(generation_index_t generation)
362 unsigned long count = 0;
364 for (i = 0; i < last_free_page; i++)
365 if (page_allocated_p(i)
366 && (page_table[i].gen == generation)
367 && (page_table[i].write_protected == 1))
372 /* Count the number of pages within the given generation. */
374 count_generation_pages(generation_index_t generation)
379 for (i = 0; i < last_free_page; i++)
380 if (page_allocated_p(i)
381 && (page_table[i].gen == generation))
388 count_dont_move_pages(void)
392 for (i = 0; i < last_free_page; i++) {
393 if (page_allocated_p(i)
394 && (page_table[i].dont_move != 0)) {
402 /* Work through the pages and add up the number of bytes used for the
403 * given generation. */
405 count_generation_bytes_allocated (generation_index_t gen)
408 unsigned long result = 0;
409 for (i = 0; i < last_free_page; i++) {
410 if (page_allocated_p(i)
411 && (page_table[i].gen == gen))
412 result += page_table[i].bytes_used;
417 /* Return the average age of the memory in a generation. */
419 gen_av_mem_age(generation_index_t gen)
421 if (generations[gen].bytes_allocated == 0)
425 ((double)generations[gen].cum_sum_bytes_allocated)
426 / ((double)generations[gen].bytes_allocated);
429 /* The verbose argument controls how much to print: 0 for normal
430 * level of detail; 1 for debugging. */
432 print_generation_stats(int verbose) /* FIXME: should take FILE argument */
434 generation_index_t i, gens;
436 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
437 #define FPU_STATE_SIZE 27
438 int fpu_state[FPU_STATE_SIZE];
439 #elif defined(LISP_FEATURE_PPC)
440 #define FPU_STATE_SIZE 32
441 long long fpu_state[FPU_STATE_SIZE];
444 /* This code uses the FP instructions which may be set up for Lisp
445 * so they need to be saved and reset for C. */
448 /* highest generation to print */
450 gens = SCRATCH_GENERATION;
452 gens = PSEUDO_STATIC_GENERATION;
454 /* Print the heap stats. */
456 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
458 for (i = 0; i < gens; i++) {
461 long unboxed_cnt = 0;
462 long large_boxed_cnt = 0;
463 long large_unboxed_cnt = 0;
466 for (j = 0; j < last_free_page; j++)
467 if (page_table[j].gen == i) {
469 /* Count the number of boxed pages within the given
471 if (page_boxed_p(j)) {
472 if (page_table[j].large_object)
477 if(page_table[j].dont_move) pinned_cnt++;
478 /* Count the number of unboxed pages within the given
480 if (page_unboxed_p(j)) {
481 if (page_table[j].large_object)
488 gc_assert(generations[i].bytes_allocated
489 == count_generation_bytes_allocated(i));
491 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
493 generations[i].alloc_start_page,
494 generations[i].alloc_unboxed_start_page,
495 generations[i].alloc_large_start_page,
496 generations[i].alloc_large_unboxed_start_page,
502 generations[i].bytes_allocated,
503 (npage_bytes(count_generation_pages(i))
504 - generations[i].bytes_allocated),
505 generations[i].gc_trigger,
506 count_write_protect_generation_pages(i),
507 generations[i].num_gc,
510 fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated);
511 fprintf(stderr," Dynamic-space-size bytes = %lu\n", dynamic_space_size);
513 fpu_restore(fpu_state);
517 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
518 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
521 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
522 * if zeroing it ourselves, i.e. in practice give the memory back to the
523 * OS. Generally done after a large GC.
525 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
527 void *addr = page_address(start), *new_addr;
528 size_t length = npage_bytes(1+end-start);
533 os_invalidate(addr, length);
534 new_addr = os_validate(addr, length);
535 if (new_addr == NULL || new_addr != addr) {
536 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
540 for (i = start; i <= end; i++) {
541 page_table[i].need_to_zero = 0;
545 /* Zero the pages from START to END (inclusive). Generally done just after
546 * a new region has been allocated.
549 zero_pages(page_index_t start, page_index_t end) {
553 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
554 fast_bzero(page_address(start), npage_bytes(1+end-start));
556 bzero(page_address(start), npage_bytes(1+end-start));
561 /* Zero the pages from START to END (inclusive), except for those
562 * pages that are known to already zeroed. Mark all pages in the
563 * ranges as non-zeroed.
566 zero_dirty_pages(page_index_t start, page_index_t end) {
569 for (i = start; i <= end; i++) {
570 if (page_table[i].need_to_zero == 1) {
571 zero_pages(start, end);
576 for (i = start; i <= end; i++) {
577 page_table[i].need_to_zero = 1;
583 * To support quick and inline allocation, regions of memory can be
584 * allocated and then allocated from with just a free pointer and a
585 * check against an end address.
587 * Since objects can be allocated to spaces with different properties
588 * e.g. boxed/unboxed, generation, ages; there may need to be many
589 * allocation regions.
591 * Each allocation region may start within a partly used page. Many
592 * features of memory use are noted on a page wise basis, e.g. the
593 * generation; so if a region starts within an existing allocated page
594 * it must be consistent with this page.
596 * During the scavenging of the newspace, objects will be transported
597 * into an allocation region, and pointers updated to point to this
598 * allocation region. It is possible that these pointers will be
599 * scavenged again before the allocation region is closed, e.g. due to
600 * trans_list which jumps all over the place to cleanup the list. It
601 * is important to be able to determine properties of all objects
602 * pointed to when scavenging, e.g to detect pointers to the oldspace.
603 * Thus it's important that the allocation regions have the correct
604 * properties set when allocated, and not just set when closed. The
605 * region allocation routines return regions with the specified
606 * properties, and grab all the pages, setting their properties
607 * appropriately, except that the amount used is not known.
609 * These regions are used to support quicker allocation using just a
610 * free pointer. The actual space used by the region is not reflected
611 * in the pages tables until it is closed. It can't be scavenged until
614 * When finished with the region it should be closed, which will
615 * update the page tables for the actual space used returning unused
616 * space. Further it may be noted in the new regions which is
617 * necessary when scavenging the newspace.
619 * Large objects may be allocated directly without an allocation
620 * region, the page tables are updated immediately.
622 * Unboxed objects don't contain pointers to other objects and so
623 * don't need scavenging. Further they can't contain pointers to
624 * younger generations so WP is not needed. By allocating pages to
625 * unboxed objects the whole page never needs scavenging or
626 * write-protecting. */
628 /* We are only using two regions at present. Both are for the current
629 * newspace generation. */
630 struct alloc_region boxed_region;
631 struct alloc_region unboxed_region;
633 /* The generation currently being allocated to. */
634 static generation_index_t gc_alloc_generation;
636 static inline page_index_t
637 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
640 if (UNBOXED_PAGE_FLAG == page_type_flag) {
641 return generations[generation].alloc_large_unboxed_start_page;
642 } else if (BOXED_PAGE_FLAG & page_type_flag) {
643 /* Both code and data. */
644 return generations[generation].alloc_large_start_page;
646 lose("bad page type flag: %d", page_type_flag);
649 if (UNBOXED_PAGE_FLAG == page_type_flag) {
650 return generations[generation].alloc_unboxed_start_page;
651 } else if (BOXED_PAGE_FLAG & page_type_flag) {
652 /* Both code and data. */
653 return generations[generation].alloc_start_page;
655 lose("bad page_type_flag: %d", page_type_flag);
661 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
665 if (UNBOXED_PAGE_FLAG == page_type_flag) {
666 generations[generation].alloc_large_unboxed_start_page = page;
667 } else if (BOXED_PAGE_FLAG & page_type_flag) {
668 /* Both code and data. */
669 generations[generation].alloc_large_start_page = page;
671 lose("bad page type flag: %d", page_type_flag);
674 if (UNBOXED_PAGE_FLAG == page_type_flag) {
675 generations[generation].alloc_unboxed_start_page = page;
676 } else if (BOXED_PAGE_FLAG & page_type_flag) {
677 /* Both code and data. */
678 generations[generation].alloc_start_page = page;
680 lose("bad page type flag: %d", page_type_flag);
685 /* Find a new region with room for at least the given number of bytes.
687 * It starts looking at the current generation's alloc_start_page. So
688 * may pick up from the previous region if there is enough space. This
689 * keeps the allocation contiguous when scavenging the newspace.
691 * The alloc_region should have been closed by a call to
692 * gc_alloc_update_page_tables(), and will thus be in an empty state.
694 * To assist the scavenging functions write-protected pages are not
695 * used. Free pages should not be write-protected.
697 * It is critical to the conservative GC that the start of regions be
698 * known. To help achieve this only small regions are allocated at a
701 * During scavenging, pointers may be found to within the current
702 * region and the page generation must be set so that pointers to the
703 * from space can be recognized. Therefore the generation of pages in
704 * the region are set to gc_alloc_generation. To prevent another
705 * allocation call using the same pages, all the pages in the region
706 * are allocated, although they will initially be empty.
709 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
711 page_index_t first_page;
712 page_index_t last_page;
713 unsigned long bytes_found;
719 "/alloc_new_region for %d bytes from gen %d\n",
720 nbytes, gc_alloc_generation));
723 /* Check that the region is in a reset state. */
724 gc_assert((alloc_region->first_page == 0)
725 && (alloc_region->last_page == -1)
726 && (alloc_region->free_pointer == alloc_region->end_addr));
727 ret = thread_mutex_lock(&free_pages_lock);
729 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
730 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
731 bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
732 + npage_bytes(last_page-first_page);
734 /* Set up the alloc_region. */
735 alloc_region->first_page = first_page;
736 alloc_region->last_page = last_page;
737 alloc_region->start_addr = page_table[first_page].bytes_used
738 + page_address(first_page);
739 alloc_region->free_pointer = alloc_region->start_addr;
740 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
742 /* Set up the pages. */
744 /* The first page may have already been in use. */
745 if (page_table[first_page].bytes_used == 0) {
746 page_table[first_page].allocated = page_type_flag;
747 page_table[first_page].gen = gc_alloc_generation;
748 page_table[first_page].large_object = 0;
749 page_table[first_page].region_start_offset = 0;
752 gc_assert(page_table[first_page].allocated == page_type_flag);
753 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
755 gc_assert(page_table[first_page].gen == gc_alloc_generation);
756 gc_assert(page_table[first_page].large_object == 0);
758 for (i = first_page+1; i <= last_page; i++) {
759 page_table[i].allocated = page_type_flag;
760 page_table[i].gen = gc_alloc_generation;
761 page_table[i].large_object = 0;
762 /* This may not be necessary for unboxed regions (think it was
764 page_table[i].region_start_offset =
765 void_diff(page_address(i),alloc_region->start_addr);
766 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
768 /* Bump up last_free_page. */
769 if (last_page+1 > last_free_page) {
770 last_free_page = last_page+1;
771 /* do we only want to call this on special occasions? like for
773 set_alloc_pointer((lispobj)page_address(last_free_page));
775 ret = thread_mutex_unlock(&free_pages_lock);
778 #ifdef READ_PROTECT_FREE_PAGES
779 os_protect(page_address(first_page),
780 npage_bytes(1+last_page-first_page),
784 /* If the first page was only partial, don't check whether it's
785 * zeroed (it won't be) and don't zero it (since the parts that
786 * we're interested in are guaranteed to be zeroed).
788 if (page_table[first_page].bytes_used) {
792 zero_dirty_pages(first_page, last_page);
794 /* we can do this after releasing free_pages_lock */
795 if (gencgc_zero_check) {
797 for (p = (long *)alloc_region->start_addr;
798 p < (long *)alloc_region->end_addr; p++) {
800 /* KLUDGE: It would be nice to use %lx and explicit casts
801 * (long) in code like this, so that it is less likely to
802 * break randomly when running on a machine with different
803 * word sizes. -- WHN 19991129 */
804 lose("The new region at %x is not zero (start=%p, end=%p).\n",
805 p, alloc_region->start_addr, alloc_region->end_addr);
811 /* If the record_new_objects flag is 2 then all new regions created
814 * If it's 1 then then it is only recorded if the first page of the
815 * current region is <= new_areas_ignore_page. This helps avoid
816 * unnecessary recording when doing full scavenge pass.
818 * The new_object structure holds the page, byte offset, and size of
819 * new regions of objects. Each new area is placed in the array of
820 * these structures pointer to by new_areas. new_areas_index holds the
821 * offset into new_areas.
823 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
824 * later code must detect this and handle it, probably by doing a full
825 * scavenge of a generation. */
826 #define NUM_NEW_AREAS 512
827 static int record_new_objects = 0;
828 static page_index_t new_areas_ignore_page;
834 static struct new_area (*new_areas)[];
835 static long new_areas_index;
838 /* Add a new area to new_areas. */
840 add_new_area(page_index_t first_page, size_t offset, size_t size)
842 unsigned long new_area_start,c;
845 /* Ignore if full. */
846 if (new_areas_index >= NUM_NEW_AREAS)
849 switch (record_new_objects) {
853 if (first_page > new_areas_ignore_page)
862 new_area_start = npage_bytes(first_page) + offset;
864 /* Search backwards for a prior area that this follows from. If
865 found this will save adding a new area. */
866 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
867 unsigned long area_end =
868 npage_bytes((*new_areas)[i].page)
869 + (*new_areas)[i].offset
870 + (*new_areas)[i].size;
872 "/add_new_area S1 %d %d %d %d\n",
873 i, c, new_area_start, area_end));*/
874 if (new_area_start == area_end) {
876 "/adding to [%d] %d %d %d with %d %d %d:\n",
878 (*new_areas)[i].page,
879 (*new_areas)[i].offset,
880 (*new_areas)[i].size,
884 (*new_areas)[i].size += size;
889 (*new_areas)[new_areas_index].page = first_page;
890 (*new_areas)[new_areas_index].offset = offset;
891 (*new_areas)[new_areas_index].size = size;
893 "/new_area %d page %d offset %d size %d\n",
894 new_areas_index, first_page, offset, size));*/
897 /* Note the max new_areas used. */
898 if (new_areas_index > max_new_areas)
899 max_new_areas = new_areas_index;
902 /* Update the tables for the alloc_region. The region may be added to
905 * When done the alloc_region is set up so that the next quick alloc
906 * will fail safely and thus a new region will be allocated. Further
907 * it is safe to try to re-update the page table of this reset
910 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
913 page_index_t first_page;
914 page_index_t next_page;
915 unsigned long bytes_used;
916 unsigned long orig_first_page_bytes_used;
917 unsigned long region_size;
918 unsigned long byte_cnt;
922 first_page = alloc_region->first_page;
924 /* Catch an unused alloc_region. */
925 if ((first_page == 0) && (alloc_region->last_page == -1))
928 next_page = first_page+1;
930 ret = thread_mutex_lock(&free_pages_lock);
932 if (alloc_region->free_pointer != alloc_region->start_addr) {
933 /* some bytes were allocated in the region */
934 orig_first_page_bytes_used = page_table[first_page].bytes_used;
936 gc_assert(alloc_region->start_addr ==
937 (page_address(first_page)
938 + page_table[first_page].bytes_used));
940 /* All the pages used need to be updated */
942 /* Update the first page. */
944 /* If the page was free then set up the gen, and
945 * region_start_offset. */
946 if (page_table[first_page].bytes_used == 0)
947 gc_assert(page_table[first_page].region_start_offset == 0);
948 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
950 gc_assert(page_table[first_page].allocated & page_type_flag);
951 gc_assert(page_table[first_page].gen == gc_alloc_generation);
952 gc_assert(page_table[first_page].large_object == 0);
956 /* Calculate the number of bytes used in this page. This is not
957 * always the number of new bytes, unless it was free. */
959 if ((bytes_used = void_diff(alloc_region->free_pointer,
960 page_address(first_page)))
962 bytes_used = PAGE_BYTES;
965 page_table[first_page].bytes_used = bytes_used;
966 byte_cnt += bytes_used;
969 /* All the rest of the pages should be free. We need to set
970 * their region_start_offset pointer to the start of the
971 * region, and set the bytes_used. */
973 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
974 gc_assert(page_table[next_page].allocated & page_type_flag);
975 gc_assert(page_table[next_page].bytes_used == 0);
976 gc_assert(page_table[next_page].gen == gc_alloc_generation);
977 gc_assert(page_table[next_page].large_object == 0);
979 gc_assert(page_table[next_page].region_start_offset ==
980 void_diff(page_address(next_page),
981 alloc_region->start_addr));
983 /* Calculate the number of bytes used in this page. */
985 if ((bytes_used = void_diff(alloc_region->free_pointer,
986 page_address(next_page)))>PAGE_BYTES) {
987 bytes_used = PAGE_BYTES;
990 page_table[next_page].bytes_used = bytes_used;
991 byte_cnt += bytes_used;
996 region_size = void_diff(alloc_region->free_pointer,
997 alloc_region->start_addr);
998 bytes_allocated += region_size;
999 generations[gc_alloc_generation].bytes_allocated += region_size;
1001 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1003 /* Set the generations alloc restart page to the last page of
1005 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1007 /* Add the region to the new_areas if requested. */
1008 if (BOXED_PAGE_FLAG & page_type_flag)
1009 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1013 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1015 gc_alloc_generation));
1018 /* There are no bytes allocated. Unallocate the first_page if
1019 * there are 0 bytes_used. */
1020 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1021 if (page_table[first_page].bytes_used == 0)
1022 page_table[first_page].allocated = FREE_PAGE_FLAG;
1025 /* Unallocate any unused pages. */
1026 while (next_page <= alloc_region->last_page) {
1027 gc_assert(page_table[next_page].bytes_used == 0);
1028 page_table[next_page].allocated = FREE_PAGE_FLAG;
1031 ret = thread_mutex_unlock(&free_pages_lock);
1032 gc_assert(ret == 0);
1034 /* alloc_region is per-thread, we're ok to do this unlocked */
1035 gc_set_region_empty(alloc_region);
1038 static inline void *gc_quick_alloc(long nbytes);
1040 /* Allocate a possibly large object. */
1042 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1044 page_index_t first_page;
1045 page_index_t last_page;
1046 int orig_first_page_bytes_used;
1050 page_index_t next_page;
1053 ret = thread_mutex_lock(&free_pages_lock);
1054 gc_assert(ret == 0);
1056 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1057 if (first_page <= alloc_region->last_page) {
1058 first_page = alloc_region->last_page+1;
1061 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1063 gc_assert(first_page > alloc_region->last_page);
1065 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1067 /* Set up the pages. */
1068 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1070 /* If the first page was free then set up the gen, and
1071 * region_start_offset. */
1072 if (page_table[first_page].bytes_used == 0) {
1073 page_table[first_page].allocated = page_type_flag;
1074 page_table[first_page].gen = gc_alloc_generation;
1075 page_table[first_page].region_start_offset = 0;
1076 page_table[first_page].large_object = 1;
1079 gc_assert(page_table[first_page].allocated == page_type_flag);
1080 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1081 gc_assert(page_table[first_page].large_object == 1);
1085 /* Calc. the number of bytes used in this page. This is not
1086 * always the number of new bytes, unless it was free. */
1088 if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
1089 bytes_used = PAGE_BYTES;
1092 page_table[first_page].bytes_used = bytes_used;
1093 byte_cnt += bytes_used;
1095 next_page = first_page+1;
1097 /* All the rest of the pages should be free. We need to set their
1098 * region_start_offset pointer to the start of the region, and set
1099 * the bytes_used. */
1101 gc_assert(page_free_p(next_page));
1102 gc_assert(page_table[next_page].bytes_used == 0);
1103 page_table[next_page].allocated = page_type_flag;
1104 page_table[next_page].gen = gc_alloc_generation;
1105 page_table[next_page].large_object = 1;
1107 page_table[next_page].region_start_offset =
1108 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1110 /* Calculate the number of bytes used in this page. */
1112 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1113 if (bytes_used > PAGE_BYTES) {
1114 bytes_used = PAGE_BYTES;
1117 page_table[next_page].bytes_used = bytes_used;
1118 page_table[next_page].write_protected=0;
1119 page_table[next_page].dont_move=0;
1120 byte_cnt += bytes_used;
1124 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1126 bytes_allocated += nbytes;
1127 generations[gc_alloc_generation].bytes_allocated += nbytes;
1129 /* Add the region to the new_areas if requested. */
1130 if (BOXED_PAGE_FLAG & page_type_flag)
1131 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1133 /* Bump up last_free_page */
1134 if (last_page+1 > last_free_page) {
1135 last_free_page = last_page+1;
1136 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1138 ret = thread_mutex_unlock(&free_pages_lock);
1139 gc_assert(ret == 0);
1141 #ifdef READ_PROTECT_FREE_PAGES
1142 os_protect(page_address(first_page),
1143 npage_bytes(1+last_page-first_page),
1147 zero_dirty_pages(first_page, last_page);
1149 return page_address(first_page);
1152 static page_index_t gencgc_alloc_start_page = -1;
1155 gc_heap_exhausted_error_or_lose (long available, long requested)
1157 /* Write basic information before doing anything else: if we don't
1158 * call to lisp this is a must, and even if we do there is always
1159 * the danger that we bounce back here before the error has been
1160 * handled, or indeed even printed.
1162 fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
1163 gc_active_p ? "garbage collection" : "allocation",
1164 available, requested);
1165 if (gc_active_p || (available == 0)) {
1166 /* If we are in GC, or totally out of memory there is no way
1167 * to sanely transfer control to the lisp-side of things.
1169 struct thread *thread = arch_os_get_current_thread();
1170 print_generation_stats(1);
1171 fprintf(stderr, "GC control variables:\n");
1172 fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
1173 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
1174 SymbolValue(GC_PENDING,thread)==NIL ? "false" : "true");
1175 #ifdef LISP_FEATURE_SB_THREAD
1176 fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n",
1177 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
1179 lose("Heap exhausted, game over.");
1182 /* FIXME: assert free_pages_lock held */
1183 (void)thread_mutex_unlock(&free_pages_lock);
1184 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1185 alloc_number(available), alloc_number(requested));
1186 lose("HEAP-EXHAUSTED-ERROR fell through");
1191 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int page_type_flag)
1193 page_index_t first_page, last_page;
1194 page_index_t restart_page = *restart_page_ptr;
1195 long bytes_found = 0;
1196 long most_bytes_found = 0;
1197 /* FIXME: assert(free_pages_lock is held); */
1199 /* Toggled by gc_and_save for heap compaction, normally -1. */
1200 if (gencgc_alloc_start_page != -1) {
1201 restart_page = gencgc_alloc_start_page;
1204 if (nbytes>=PAGE_BYTES) {
1205 /* Search for a contiguous free space of at least nbytes,
1206 * aligned on a page boundary. The page-alignment is strictly
1207 * speaking needed only for objects at least large_object_size
1210 first_page = restart_page;
1211 while ((first_page < page_table_pages) &&
1212 page_allocated_p(first_page))
1215 last_page = first_page;
1216 bytes_found = PAGE_BYTES;
1217 while ((bytes_found < nbytes) &&
1218 (last_page < (page_table_pages-1)) &&
1219 page_free_p(last_page+1)) {
1221 bytes_found += PAGE_BYTES;
1222 gc_assert(0 == page_table[last_page].bytes_used);
1223 gc_assert(0 == page_table[last_page].write_protected);
1225 if (bytes_found > most_bytes_found)
1226 most_bytes_found = bytes_found;
1227 restart_page = last_page + 1;
1228 } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
1231 /* Search for a page with at least nbytes of space. We prefer
1232 * not to split small objects on multiple pages, to reduce the
1233 * number of contiguous allocation regions spaning multiple
1234 * pages: this helps avoid excessive conservativism. */
1235 first_page = restart_page;
1236 while (first_page < page_table_pages) {
1237 if (page_free_p(first_page))
1239 gc_assert(0 == page_table[first_page].bytes_used);
1240 bytes_found = PAGE_BYTES;
1243 else if ((page_table[first_page].allocated == page_type_flag) &&
1244 (page_table[first_page].large_object == 0) &&
1245 (page_table[first_page].gen == gc_alloc_generation) &&
1246 (page_table[first_page].write_protected == 0) &&
1247 (page_table[first_page].dont_move == 0))
1249 bytes_found = PAGE_BYTES
1250 - page_table[first_page].bytes_used;
1251 if (bytes_found > most_bytes_found)
1252 most_bytes_found = bytes_found;
1253 if (bytes_found >= nbytes)
1258 last_page = first_page;
1259 restart_page = first_page + 1;
1262 /* Check for a failure */
1263 if (bytes_found < nbytes) {
1264 gc_assert(restart_page >= page_table_pages);
1265 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1268 gc_assert(page_table[first_page].write_protected == 0);
1270 *restart_page_ptr = first_page;
1274 /* Allocate bytes. All the rest of the special-purpose allocation
1275 * functions will eventually call this */
1278 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1281 void *new_free_pointer;
1283 if (nbytes>=large_object_size)
1284 return gc_alloc_large(nbytes, page_type_flag, my_region);
1286 /* Check whether there is room in the current alloc region. */
1287 new_free_pointer = my_region->free_pointer + nbytes;
1289 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1290 my_region->free_pointer, new_free_pointer); */
1292 if (new_free_pointer <= my_region->end_addr) {
1293 /* If so then allocate from the current alloc region. */
1294 void *new_obj = my_region->free_pointer;
1295 my_region->free_pointer = new_free_pointer;
1297 /* Unless a `quick' alloc was requested, check whether the
1298 alloc region is almost empty. */
1300 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1301 /* If so, finished with the current region. */
1302 gc_alloc_update_page_tables(page_type_flag, my_region);
1303 /* Set up a new region. */
1304 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1307 return((void *)new_obj);
1310 /* Else not enough free space in the current region: retry with a
1313 gc_alloc_update_page_tables(page_type_flag, my_region);
1314 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1315 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1318 /* these are only used during GC: all allocation from the mutator calls
1319 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1322 static inline void *
1323 gc_quick_alloc(long nbytes)
1325 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1328 static inline void *
1329 gc_quick_alloc_large(long nbytes)
1331 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1334 static inline void *
1335 gc_alloc_unboxed(long nbytes)
1337 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1340 static inline void *
1341 gc_quick_alloc_unboxed(long nbytes)
1343 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1346 static inline void *
1347 gc_quick_alloc_large_unboxed(long nbytes)
1349 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1353 /* Copy a large boxed object. If the object is in a large object
1354 * region then it is simply promoted, else it is copied. If it's large
1355 * enough then it's copied to a large object region.
1357 * Vectors may have shrunk. If the object is not copied the space
1358 * needs to be reclaimed, and the page_tables corrected. */
1360 copy_large_object(lispobj object, long nwords)
1364 page_index_t first_page;
1366 gc_assert(is_lisp_pointer(object));
1367 gc_assert(from_space_p(object));
1368 gc_assert((nwords & 0x01) == 0);
1371 /* Check whether it's in a large object region. */
1372 first_page = find_page_index((void *)object);
1373 gc_assert(first_page >= 0);
1375 if (page_table[first_page].large_object) {
1377 /* Promote the object. */
1379 unsigned long remaining_bytes;
1380 page_index_t next_page;
1381 unsigned long bytes_freed;
1382 unsigned long old_bytes_used;
1384 /* Note: Any page write-protection must be removed, else a
1385 * later scavenge_newspace may incorrectly not scavenge these
1386 * pages. This would not be necessary if they are added to the
1387 * new areas, but let's do it for them all (they'll probably
1388 * be written anyway?). */
1390 gc_assert(page_table[first_page].region_start_offset == 0);
1392 next_page = first_page;
1393 remaining_bytes = nwords*N_WORD_BYTES;
1394 while (remaining_bytes > PAGE_BYTES) {
1395 gc_assert(page_table[next_page].gen == from_space);
1396 gc_assert(page_boxed_p(next_page));
1397 gc_assert(page_table[next_page].large_object);
1398 gc_assert(page_table[next_page].region_start_offset ==
1399 npage_bytes(next_page-first_page));
1400 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1402 page_table[next_page].gen = new_space;
1404 /* Remove any write-protection. We should be able to rely
1405 * on the write-protect flag to avoid redundant calls. */
1406 if (page_table[next_page].write_protected) {
1407 os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL);
1408 page_table[next_page].write_protected = 0;
1410 remaining_bytes -= PAGE_BYTES;
1414 /* Now only one page remains, but the object may have shrunk
1415 * so there may be more unused pages which will be freed. */
1417 /* The object may have shrunk but shouldn't have grown. */
1418 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1420 page_table[next_page].gen = new_space;
1421 gc_assert(page_boxed_p(next_page));
1423 /* Adjust the bytes_used. */
1424 old_bytes_used = page_table[next_page].bytes_used;
1425 page_table[next_page].bytes_used = remaining_bytes;
1427 bytes_freed = old_bytes_used - remaining_bytes;
1429 /* Free any remaining pages; needs care. */
1431 while ((old_bytes_used == PAGE_BYTES) &&
1432 (page_table[next_page].gen == from_space) &&
1433 page_boxed_p(next_page) &&
1434 page_table[next_page].large_object &&
1435 (page_table[next_page].region_start_offset ==
1436 npage_bytes(next_page - first_page))) {
1437 /* Checks out OK, free the page. Don't need to bother zeroing
1438 * pages as this should have been done before shrinking the
1439 * object. These pages shouldn't be write-protected as they
1440 * should be zero filled. */
1441 gc_assert(page_table[next_page].write_protected == 0);
1443 old_bytes_used = page_table[next_page].bytes_used;
1444 page_table[next_page].allocated = FREE_PAGE_FLAG;
1445 page_table[next_page].bytes_used = 0;
1446 bytes_freed += old_bytes_used;
1450 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1452 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1453 bytes_allocated -= bytes_freed;
1455 /* Add the region to the new_areas if requested. */
1456 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1460 /* Get tag of object. */
1461 tag = lowtag_of(object);
1463 /* Allocate space. */
1464 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1466 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1468 /* Return Lisp pointer of new object. */
1469 return ((lispobj) new) | tag;
1473 /* to copy unboxed objects */
1475 copy_unboxed_object(lispobj object, long nwords)
1480 gc_assert(is_lisp_pointer(object));
1481 gc_assert(from_space_p(object));
1482 gc_assert((nwords & 0x01) == 0);
1484 /* Get tag of object. */
1485 tag = lowtag_of(object);
1487 /* Allocate space. */
1488 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1490 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1492 /* Return Lisp pointer of new object. */
1493 return ((lispobj) new) | tag;
1496 /* to copy large unboxed objects
1498 * If the object is in a large object region then it is simply
1499 * promoted, else it is copied. If it's large enough then it's copied
1500 * to a large object region.
1502 * Bignums and vectors may have shrunk. If the object is not copied
1503 * the space needs to be reclaimed, and the page_tables corrected.
1505 * KLUDGE: There's a lot of cut-and-paste duplication between this
1506 * function and copy_large_object(..). -- WHN 20000619 */
1508 copy_large_unboxed_object(lispobj object, long nwords)
1512 page_index_t first_page;
1514 gc_assert(is_lisp_pointer(object));
1515 gc_assert(from_space_p(object));
1516 gc_assert((nwords & 0x01) == 0);
1518 if ((nwords > 1024*1024) && gencgc_verbose) {
1519 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1520 nwords*N_WORD_BYTES));
1523 /* Check whether it's a large object. */
1524 first_page = find_page_index((void *)object);
1525 gc_assert(first_page >= 0);
1527 if (page_table[first_page].large_object) {
1528 /* Promote the object. Note: Unboxed objects may have been
1529 * allocated to a BOXED region so it may be necessary to
1530 * change the region to UNBOXED. */
1531 unsigned long remaining_bytes;
1532 page_index_t next_page;
1533 unsigned long bytes_freed;
1534 unsigned long old_bytes_used;
1536 gc_assert(page_table[first_page].region_start_offset == 0);
1538 next_page = first_page;
1539 remaining_bytes = nwords*N_WORD_BYTES;
1540 while (remaining_bytes > PAGE_BYTES) {
1541 gc_assert(page_table[next_page].gen == from_space);
1542 gc_assert(page_allocated_no_region_p(next_page));
1543 gc_assert(page_table[next_page].large_object);
1544 gc_assert(page_table[next_page].region_start_offset ==
1545 npage_bytes(next_page-first_page));
1546 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1548 page_table[next_page].gen = new_space;
1549 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1550 remaining_bytes -= PAGE_BYTES;
1554 /* Now only one page remains, but the object may have shrunk so
1555 * there may be more unused pages which will be freed. */
1557 /* Object may have shrunk but shouldn't have grown - check. */
1558 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1560 page_table[next_page].gen = new_space;
1561 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1563 /* Adjust the bytes_used. */
1564 old_bytes_used = page_table[next_page].bytes_used;
1565 page_table[next_page].bytes_used = remaining_bytes;
1567 bytes_freed = old_bytes_used - remaining_bytes;
1569 /* Free any remaining pages; needs care. */
1571 while ((old_bytes_used == PAGE_BYTES) &&
1572 (page_table[next_page].gen == from_space) &&
1573 page_allocated_no_region_p(next_page) &&
1574 page_table[next_page].large_object &&
1575 (page_table[next_page].region_start_offset ==
1576 npage_bytes(next_page - first_page))) {
1577 /* Checks out OK, free the page. Don't need to both zeroing
1578 * pages as this should have been done before shrinking the
1579 * object. These pages shouldn't be write-protected, even if
1580 * boxed they should be zero filled. */
1581 gc_assert(page_table[next_page].write_protected == 0);
1583 old_bytes_used = page_table[next_page].bytes_used;
1584 page_table[next_page].allocated = FREE_PAGE_FLAG;
1585 page_table[next_page].bytes_used = 0;
1586 bytes_freed += old_bytes_used;
1590 if ((bytes_freed > 0) && gencgc_verbose) {
1592 "/copy_large_unboxed bytes_freed=%d\n",
1596 generations[from_space].bytes_allocated -=
1597 nwords*N_WORD_BYTES + bytes_freed;
1598 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1599 bytes_allocated -= bytes_freed;
1604 /* Get tag of object. */
1605 tag = lowtag_of(object);
1607 /* Allocate space. */
1608 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1610 /* Copy the object. */
1611 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1613 /* Return Lisp pointer of new object. */
1614 return ((lispobj) new) | tag;
1623 * code and code-related objects
1626 static lispobj trans_fun_header(lispobj object);
1627 static lispobj trans_boxed(lispobj object);
1630 /* Scan a x86 compiled code object, looking for possible fixups that
1631 * have been missed after a move.
1633 * Two types of fixups are needed:
1634 * 1. Absolute fixups to within the code object.
1635 * 2. Relative fixups to outside the code object.
1637 * Currently only absolute fixups to the constant vector, or to the
1638 * code area are checked. */
1640 sniff_code_object(struct code *code, unsigned long displacement)
1642 #ifdef LISP_FEATURE_X86
1643 long nheader_words, ncode_words, nwords;
1645 void *constants_start_addr = NULL, *constants_end_addr;
1646 void *code_start_addr, *code_end_addr;
1647 int fixup_found = 0;
1649 if (!check_code_fixups)
1652 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1654 ncode_words = fixnum_value(code->code_size);
1655 nheader_words = HeaderValue(*(lispobj *)code);
1656 nwords = ncode_words + nheader_words;
1658 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1659 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1660 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1661 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1663 /* Work through the unboxed code. */
1664 for (p = code_start_addr; p < code_end_addr; p++) {
1665 void *data = *(void **)p;
1666 unsigned d1 = *((unsigned char *)p - 1);
1667 unsigned d2 = *((unsigned char *)p - 2);
1668 unsigned d3 = *((unsigned char *)p - 3);
1669 unsigned d4 = *((unsigned char *)p - 4);
1671 unsigned d5 = *((unsigned char *)p - 5);
1672 unsigned d6 = *((unsigned char *)p - 6);
1675 /* Check for code references. */
1676 /* Check for a 32 bit word that looks like an absolute
1677 reference to within the code adea of the code object. */
1678 if ((data >= (code_start_addr-displacement))
1679 && (data < (code_end_addr-displacement))) {
1680 /* function header */
1682 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1684 /* Skip the function header */
1688 /* the case of PUSH imm32 */
1692 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1693 p, d6, d5, d4, d3, d2, d1, data));
1694 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1696 /* the case of MOV [reg-8],imm32 */
1698 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1699 || d2==0x45 || d2==0x46 || d2==0x47)
1703 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1704 p, d6, d5, d4, d3, d2, d1, data));
1705 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1707 /* the case of LEA reg,[disp32] */
1708 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1711 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1712 p, d6, d5, d4, d3, d2, d1, data));
1713 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1717 /* Check for constant references. */
1718 /* Check for a 32 bit word that looks like an absolute
1719 reference to within the constant vector. Constant references
1721 if ((data >= (constants_start_addr-displacement))
1722 && (data < (constants_end_addr-displacement))
1723 && (((unsigned)data & 0x3) == 0)) {
1728 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1729 p, d6, d5, d4, d3, d2, d1, data));
1730 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1733 /* the case of MOV m32,EAX */
1737 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1738 p, d6, d5, d4, d3, d2, d1, data));
1739 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1742 /* the case of CMP m32,imm32 */
1743 if ((d1 == 0x3d) && (d2 == 0x81)) {
1746 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1747 p, d6, d5, d4, d3, d2, d1, data));
1749 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1752 /* Check for a mod=00, r/m=101 byte. */
1753 if ((d1 & 0xc7) == 5) {
1758 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1759 p, d6, d5, d4, d3, d2, d1, data));
1760 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1762 /* the case of CMP reg32,m32 */
1766 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1767 p, d6, d5, d4, d3, d2, d1, data));
1768 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1770 /* the case of MOV m32,reg32 */
1774 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1775 p, d6, d5, d4, d3, d2, d1, data));
1776 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1778 /* the case of MOV reg32,m32 */
1782 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1783 p, d6, d5, d4, d3, d2, d1, data));
1784 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1786 /* the case of LEA reg32,m32 */
1790 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1791 p, d6, d5, d4, d3, d2, d1, data));
1792 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1798 /* If anything was found, print some information on the code
1802 "/compiled code object at %x: header words = %d, code words = %d\n",
1803 code, nheader_words, ncode_words));
1805 "/const start = %x, end = %x\n",
1806 constants_start_addr, constants_end_addr));
1808 "/code start = %x, end = %x\n",
1809 code_start_addr, code_end_addr));
1815 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1817 /* x86-64 uses pc-relative addressing instead of this kludge */
1818 #ifndef LISP_FEATURE_X86_64
1819 long nheader_words, ncode_words, nwords;
1820 void *constants_start_addr, *constants_end_addr;
1821 void *code_start_addr, *code_end_addr;
1822 lispobj fixups = NIL;
1823 unsigned long displacement =
1824 (unsigned long)new_code - (unsigned long)old_code;
1825 struct vector *fixups_vector;
1827 ncode_words = fixnum_value(new_code->code_size);
1828 nheader_words = HeaderValue(*(lispobj *)new_code);
1829 nwords = ncode_words + nheader_words;
1831 "/compiled code object at %x: header words = %d, code words = %d\n",
1832 new_code, nheader_words, ncode_words)); */
1833 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1834 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1835 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1836 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1839 "/const start = %x, end = %x\n",
1840 constants_start_addr,constants_end_addr));
1842 "/code start = %x; end = %x\n",
1843 code_start_addr,code_end_addr));
1846 /* The first constant should be a pointer to the fixups for this
1847 code objects. Check. */
1848 fixups = new_code->constants[0];
1850 /* It will be 0 or the unbound-marker if there are no fixups (as
1851 * will be the case if the code object has been purified, for
1852 * example) and will be an other pointer if it is valid. */
1853 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1854 !is_lisp_pointer(fixups)) {
1855 /* Check for possible errors. */
1856 if (check_code_fixups)
1857 sniff_code_object(new_code, displacement);
1862 fixups_vector = (struct vector *)native_pointer(fixups);
1864 /* Could be pointing to a forwarding pointer. */
1865 /* FIXME is this always in from_space? if so, could replace this code with
1866 * forwarding_pointer_p/forwarding_pointer_value */
1867 if (is_lisp_pointer(fixups) &&
1868 (find_page_index((void*)fixups_vector) != -1) &&
1869 (fixups_vector->header == 0x01)) {
1870 /* If so, then follow it. */
1871 /*SHOW("following pointer to a forwarding pointer");*/
1873 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1876 /*SHOW("got fixups");*/
1878 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1879 /* Got the fixups for the code block. Now work through the vector,
1880 and apply a fixup at each address. */
1881 long length = fixnum_value(fixups_vector->length);
1883 for (i = 0; i < length; i++) {
1884 unsigned long offset = fixups_vector->data[i];
1885 /* Now check the current value of offset. */
1886 unsigned long old_value =
1887 *(unsigned long *)((unsigned long)code_start_addr + offset);
1889 /* If it's within the old_code object then it must be an
1890 * absolute fixup (relative ones are not saved) */
1891 if ((old_value >= (unsigned long)old_code)
1892 && (old_value < ((unsigned long)old_code
1893 + nwords*N_WORD_BYTES)))
1894 /* So add the dispacement. */
1895 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1896 old_value + displacement;
1898 /* It is outside the old code object so it must be a
1899 * relative fixup (absolute fixups are not saved). So
1900 * subtract the displacement. */
1901 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1902 old_value - displacement;
1905 /* This used to just print a note to stderr, but a bogus fixup seems to
1906 * indicate real heap corruption, so a hard hailure is in order. */
1907 lose("fixup vector %p has a bad widetag: %d\n",
1908 fixups_vector, widetag_of(fixups_vector->header));
1911 /* Check for possible errors. */
1912 if (check_code_fixups) {
1913 sniff_code_object(new_code,displacement);
1920 trans_boxed_large(lispobj object)
1923 unsigned long length;
1925 gc_assert(is_lisp_pointer(object));
1927 header = *((lispobj *) native_pointer(object));
1928 length = HeaderValue(header) + 1;
1929 length = CEILING(length, 2);
1931 return copy_large_object(object, length);
1934 /* Doesn't seem to be used, delete it after the grace period. */
1937 trans_unboxed_large(lispobj object)
1940 unsigned long length;
1942 gc_assert(is_lisp_pointer(object));
1944 header = *((lispobj *) native_pointer(object));
1945 length = HeaderValue(header) + 1;
1946 length = CEILING(length, 2);
1948 return copy_large_unboxed_object(object, length);
1954 * Lutexes. Using the normal finalization machinery for finalizing
1955 * lutexes is tricky, since the finalization depends on working lutexes.
1956 * So we track the lutexes in the GC and finalize them manually.
1959 #if defined(LUTEX_WIDETAG)
1962 * Start tracking LUTEX in the GC, by adding it to the linked list of
1963 * lutexes in the nursery generation. The caller is responsible for
1964 * locking, and GCs must be inhibited until the registration is
1968 gencgc_register_lutex (struct lutex *lutex) {
1969 int index = find_page_index(lutex);
1970 generation_index_t gen;
1973 /* This lutex is in static space, so we don't need to worry about
1979 gen = page_table[index].gen;
1981 gc_assert(gen >= 0);
1982 gc_assert(gen < NUM_GENERATIONS);
1984 head = generations[gen].lutexes;
1991 generations[gen].lutexes = lutex;
1995 * Stop tracking LUTEX in the GC by removing it from the appropriate
1996 * linked lists. This will only be called during GC, so no locking is
2000 gencgc_unregister_lutex (struct lutex *lutex) {
2002 lutex->prev->next = lutex->next;
2004 generations[lutex->gen].lutexes = lutex->next;
2008 lutex->next->prev = lutex->prev;
2017 * Mark all lutexes in generation GEN as not live.
2020 unmark_lutexes (generation_index_t gen) {
2021 struct lutex *lutex = generations[gen].lutexes;
2025 lutex = lutex->next;
2030 * Finalize all lutexes in generation GEN that have not been marked live.
2033 reap_lutexes (generation_index_t gen) {
2034 struct lutex *lutex = generations[gen].lutexes;
2037 struct lutex *next = lutex->next;
2039 lutex_destroy((tagged_lutex_t) lutex);
2040 gencgc_unregister_lutex(lutex);
2047 * Mark LUTEX as live.
2050 mark_lutex (lispobj tagged_lutex) {
2051 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
2057 * Move all lutexes in generation FROM to generation TO.
2060 move_lutexes (generation_index_t from, generation_index_t to) {
2061 struct lutex *tail = generations[from].lutexes;
2063 /* Nothing to move */
2067 /* Change the generation of the lutexes in FROM. */
2068 while (tail->next) {
2074 /* Link the last lutex in the FROM list to the start of the TO list */
2075 tail->next = generations[to].lutexes;
2077 /* And vice versa */
2078 if (generations[to].lutexes) {
2079 generations[to].lutexes->prev = tail;
2082 /* And update the generations structures to match this */
2083 generations[to].lutexes = generations[from].lutexes;
2084 generations[from].lutexes = NULL;
2088 scav_lutex(lispobj *where, lispobj object)
2090 mark_lutex((lispobj) where);
2092 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2096 trans_lutex(lispobj object)
2098 struct lutex *lutex = (struct lutex *) native_pointer(object);
2100 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2101 gc_assert(is_lisp_pointer(object));
2102 copied = copy_object(object, words);
2104 /* Update the links, since the lutex moved in memory. */
2106 lutex->next->prev = (struct lutex *) native_pointer(copied);
2110 lutex->prev->next = (struct lutex *) native_pointer(copied);
2112 generations[lutex->gen].lutexes =
2113 (struct lutex *) native_pointer(copied);
2120 size_lutex(lispobj *where)
2122 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2124 #endif /* LUTEX_WIDETAG */
2131 /* XX This is a hack adapted from cgc.c. These don't work too
2132 * efficiently with the gencgc as a list of the weak pointers is
2133 * maintained within the objects which causes writes to the pages. A
2134 * limited attempt is made to avoid unnecessary writes, but this needs
2136 #define WEAK_POINTER_NWORDS \
2137 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2140 scav_weak_pointer(lispobj *where, lispobj object)
2142 /* Since we overwrite the 'next' field, we have to make
2143 * sure not to do so for pointers already in the list.
2144 * Instead of searching the list of weak_pointers each
2145 * time, we ensure that next is always NULL when the weak
2146 * pointer isn't in the list, and not NULL otherwise.
2147 * Since we can't use NULL to denote end of list, we
2148 * use a pointer back to the same weak_pointer.
2150 struct weak_pointer * wp = (struct weak_pointer*)where;
2152 if (NULL == wp->next) {
2153 wp->next = weak_pointers;
2155 if (NULL == wp->next)
2159 /* Do not let GC scavenge the value slot of the weak pointer.
2160 * (That is why it is a weak pointer.) */
2162 return WEAK_POINTER_NWORDS;
2167 search_read_only_space(void *pointer)
2169 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2170 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2171 if ((pointer < (void *)start) || (pointer >= (void *)end))
2173 return (gc_search_space(start,
2174 (((lispobj *)pointer)+2)-start,
2175 (lispobj *) pointer));
2179 search_static_space(void *pointer)
2181 lispobj *start = (lispobj *)STATIC_SPACE_START;
2182 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2183 if ((pointer < (void *)start) || (pointer >= (void *)end))
2185 return (gc_search_space(start,
2186 (((lispobj *)pointer)+2)-start,
2187 (lispobj *) pointer));
2190 /* a faster version for searching the dynamic space. This will work even
2191 * if the object is in a current allocation region. */
2193 search_dynamic_space(void *pointer)
2195 page_index_t page_index = find_page_index(pointer);
2198 /* The address may be invalid, so do some checks. */
2199 if ((page_index == -1) || page_free_p(page_index))
2201 start = (lispobj *)page_region_start(page_index);
2202 return (gc_search_space(start,
2203 (((lispobj *)pointer)+2)-start,
2204 (lispobj *)pointer));
2207 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2209 /* Helper for valid_lisp_pointer_p and
2210 * possibly_valid_dynamic_space_pointer.
2212 * pointer is the pointer to validate, and start_addr is the address
2213 * of the enclosing object.
2216 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2218 if (!is_lisp_pointer((lispobj)pointer)) {
2222 /* Check that the object pointed to is consistent with the pointer
2224 switch (lowtag_of((lispobj)pointer)) {
2225 case FUN_POINTER_LOWTAG:
2226 /* Start_addr should be the enclosing code object, or a closure
2228 switch (widetag_of(*start_addr)) {
2229 case CODE_HEADER_WIDETAG:
2230 /* This case is probably caught above. */
2232 case CLOSURE_HEADER_WIDETAG:
2233 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2234 if ((unsigned long)pointer !=
2235 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2236 if (gencgc_verbose) {
2239 pointer, start_addr, *start_addr));
2245 if (gencgc_verbose) {
2248 pointer, start_addr, *start_addr));
2253 case LIST_POINTER_LOWTAG:
2254 if ((unsigned long)pointer !=
2255 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2256 if (gencgc_verbose) {
2259 pointer, start_addr, *start_addr));
2263 /* Is it plausible cons? */
2264 if ((is_lisp_pointer(start_addr[0]) ||
2265 is_lisp_immediate(start_addr[0])) &&
2266 (is_lisp_pointer(start_addr[1]) ||
2267 is_lisp_immediate(start_addr[1])))
2270 if (gencgc_verbose) {
2273 pointer, start_addr, *start_addr));
2277 case INSTANCE_POINTER_LOWTAG:
2278 if ((unsigned long)pointer !=
2279 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2280 if (gencgc_verbose) {
2283 pointer, start_addr, *start_addr));
2287 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2288 if (gencgc_verbose) {
2291 pointer, start_addr, *start_addr));
2296 case OTHER_POINTER_LOWTAG:
2297 if ((unsigned long)pointer !=
2298 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2299 if (gencgc_verbose) {
2302 pointer, start_addr, *start_addr));
2306 /* Is it plausible? Not a cons. XXX should check the headers. */
2307 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2308 if (gencgc_verbose) {
2311 pointer, start_addr, *start_addr));
2315 switch (widetag_of(start_addr[0])) {
2316 case UNBOUND_MARKER_WIDETAG:
2317 case NO_TLS_VALUE_MARKER_WIDETAG:
2318 case CHARACTER_WIDETAG:
2319 #if N_WORD_BITS == 64
2320 case SINGLE_FLOAT_WIDETAG:
2322 if (gencgc_verbose) {
2325 pointer, start_addr, *start_addr));
2329 /* only pointed to by function pointers? */
2330 case CLOSURE_HEADER_WIDETAG:
2331 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2332 if (gencgc_verbose) {
2335 pointer, start_addr, *start_addr));
2339 case INSTANCE_HEADER_WIDETAG:
2340 if (gencgc_verbose) {
2343 pointer, start_addr, *start_addr));
2347 /* the valid other immediate pointer objects */
2348 case SIMPLE_VECTOR_WIDETAG:
2350 case COMPLEX_WIDETAG:
2351 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2352 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2354 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2355 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2357 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2358 case COMPLEX_LONG_FLOAT_WIDETAG:
2360 case SIMPLE_ARRAY_WIDETAG:
2361 case COMPLEX_BASE_STRING_WIDETAG:
2362 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2363 case COMPLEX_CHARACTER_STRING_WIDETAG:
2365 case COMPLEX_VECTOR_NIL_WIDETAG:
2366 case COMPLEX_BIT_VECTOR_WIDETAG:
2367 case COMPLEX_VECTOR_WIDETAG:
2368 case COMPLEX_ARRAY_WIDETAG:
2369 case VALUE_CELL_HEADER_WIDETAG:
2370 case SYMBOL_HEADER_WIDETAG:
2372 case CODE_HEADER_WIDETAG:
2373 case BIGNUM_WIDETAG:
2374 #if N_WORD_BITS != 64
2375 case SINGLE_FLOAT_WIDETAG:
2377 case DOUBLE_FLOAT_WIDETAG:
2378 #ifdef LONG_FLOAT_WIDETAG
2379 case LONG_FLOAT_WIDETAG:
2381 case SIMPLE_BASE_STRING_WIDETAG:
2382 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2383 case SIMPLE_CHARACTER_STRING_WIDETAG:
2385 case SIMPLE_BIT_VECTOR_WIDETAG:
2386 case SIMPLE_ARRAY_NIL_WIDETAG:
2387 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2388 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2389 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2390 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2391 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2392 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2393 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2394 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2396 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2397 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2398 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2399 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2401 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2402 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2404 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2405 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2407 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2408 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2410 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2411 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2413 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2414 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2416 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2417 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2419 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2420 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2422 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2423 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2425 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2426 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2427 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2428 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2430 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2431 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2433 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2434 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2436 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2437 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2440 case WEAK_POINTER_WIDETAG:
2441 #ifdef LUTEX_WIDETAG
2447 if (gencgc_verbose) {
2450 pointer, start_addr, *start_addr));
2456 if (gencgc_verbose) {
2459 pointer, start_addr, *start_addr));
2468 /* Used by the debugger to validate possibly bogus pointers before
2469 * calling MAKE-LISP-OBJ on them.
2471 * FIXME: We would like to make this perfect, because if the debugger
2472 * constructs a reference to a bugs lisp object, and it ends up in a
2473 * location scavenged by the GC all hell breaks loose.
2475 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2476 * and return true for all valid pointers, this could actually be eager
2477 * and lie about a few pointers without bad results... but that should
2478 * be reflected in the name.
2481 valid_lisp_pointer_p(lispobj *pointer)
2484 if (((start=search_dynamic_space(pointer))!=NULL) ||
2485 ((start=search_static_space(pointer))!=NULL) ||
2486 ((start=search_read_only_space(pointer))!=NULL))
2487 return looks_like_valid_lisp_pointer_p(pointer, start);
2492 /* Is there any possibility that pointer is a valid Lisp object
2493 * reference, and/or something else (e.g. subroutine call return
2494 * address) which should prevent us from moving the referred-to thing?
2495 * This is called from preserve_pointers() */
2497 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2499 lispobj *start_addr;
2501 /* Find the object start address. */
2502 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2506 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2509 /* Adjust large bignum and vector objects. This will adjust the
2510 * allocated region if the size has shrunk, and move unboxed objects
2511 * into unboxed pages. The pages are not promoted here, and the
2512 * promoted region is not added to the new_regions; this is really
2513 * only designed to be called from preserve_pointer(). Shouldn't fail
2514 * if this is missed, just may delay the moving of objects to unboxed
2515 * pages, and the freeing of pages. */
2517 maybe_adjust_large_object(lispobj *where)
2519 page_index_t first_page;
2520 page_index_t next_page;
2523 unsigned long remaining_bytes;
2524 unsigned long bytes_freed;
2525 unsigned long old_bytes_used;
2529 /* Check whether it's a vector or bignum object. */
2530 switch (widetag_of(where[0])) {
2531 case SIMPLE_VECTOR_WIDETAG:
2532 boxed = BOXED_PAGE_FLAG;
2534 case BIGNUM_WIDETAG:
2535 case SIMPLE_BASE_STRING_WIDETAG:
2536 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2537 case SIMPLE_CHARACTER_STRING_WIDETAG:
2539 case SIMPLE_BIT_VECTOR_WIDETAG:
2540 case SIMPLE_ARRAY_NIL_WIDETAG:
2541 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2542 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2543 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2544 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2545 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2546 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2547 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2548 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2550 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2551 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2552 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2553 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2555 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2556 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2558 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2559 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2561 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2562 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2564 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2565 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2567 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2568 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2570 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2571 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2573 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2574 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2576 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2577 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2579 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2580 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2581 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2582 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2584 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2585 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2587 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2588 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2590 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2591 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2593 boxed = UNBOXED_PAGE_FLAG;
2599 /* Find its current size. */
2600 nwords = (sizetab[widetag_of(where[0])])(where);
2602 first_page = find_page_index((void *)where);
2603 gc_assert(first_page >= 0);
2605 /* Note: Any page write-protection must be removed, else a later
2606 * scavenge_newspace may incorrectly not scavenge these pages.
2607 * This would not be necessary if they are added to the new areas,
2608 * but lets do it for them all (they'll probably be written
2611 gc_assert(page_table[first_page].region_start_offset == 0);
2613 next_page = first_page;
2614 remaining_bytes = nwords*N_WORD_BYTES;
2615 while (remaining_bytes > PAGE_BYTES) {
2616 gc_assert(page_table[next_page].gen == from_space);
2617 gc_assert(page_allocated_no_region_p(next_page));
2618 gc_assert(page_table[next_page].large_object);
2619 gc_assert(page_table[next_page].region_start_offset ==
2620 npage_bytes(next_page-first_page));
2621 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
2623 page_table[next_page].allocated = boxed;
2625 /* Shouldn't be write-protected at this stage. Essential that the
2627 gc_assert(!page_table[next_page].write_protected);
2628 remaining_bytes -= PAGE_BYTES;
2632 /* Now only one page remains, but the object may have shrunk so
2633 * there may be more unused pages which will be freed. */
2635 /* Object may have shrunk but shouldn't have grown - check. */
2636 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2638 page_table[next_page].allocated = boxed;
2639 gc_assert(page_table[next_page].allocated ==
2640 page_table[first_page].allocated);
2642 /* Adjust the bytes_used. */
2643 old_bytes_used = page_table[next_page].bytes_used;
2644 page_table[next_page].bytes_used = remaining_bytes;
2646 bytes_freed = old_bytes_used - remaining_bytes;
2648 /* Free any remaining pages; needs care. */
2650 while ((old_bytes_used == PAGE_BYTES) &&
2651 (page_table[next_page].gen == from_space) &&
2652 page_allocated_no_region_p(next_page) &&
2653 page_table[next_page].large_object &&
2654 (page_table[next_page].region_start_offset ==
2655 npage_bytes(next_page - first_page))) {
2656 /* It checks out OK, free the page. We don't need to both zeroing
2657 * pages as this should have been done before shrinking the
2658 * object. These pages shouldn't be write protected as they
2659 * should be zero filled. */
2660 gc_assert(page_table[next_page].write_protected == 0);
2662 old_bytes_used = page_table[next_page].bytes_used;
2663 page_table[next_page].allocated = FREE_PAGE_FLAG;
2664 page_table[next_page].bytes_used = 0;
2665 bytes_freed += old_bytes_used;
2669 if ((bytes_freed > 0) && gencgc_verbose) {
2671 "/maybe_adjust_large_object() freed %d\n",
2675 generations[from_space].bytes_allocated -= bytes_freed;
2676 bytes_allocated -= bytes_freed;
2681 /* Take a possible pointer to a Lisp object and mark its page in the
2682 * page_table so that it will not be relocated during a GC.
2684 * This involves locating the page it points to, then backing up to
2685 * the start of its region, then marking all pages dont_move from there
2686 * up to the first page that's not full or has a different generation
2688 * It is assumed that all the page static flags have been cleared at
2689 * the start of a GC.
2691 * It is also assumed that the current gc_alloc() region has been
2692 * flushed and the tables updated. */
2695 preserve_pointer(void *addr)
2697 page_index_t addr_page_index = find_page_index(addr);
2698 page_index_t first_page;
2700 unsigned int region_allocation;
2702 /* quick check 1: Address is quite likely to have been invalid. */
2703 if ((addr_page_index == -1)
2704 || page_free_p(addr_page_index)
2705 || (page_table[addr_page_index].bytes_used == 0)
2706 || (page_table[addr_page_index].gen != from_space)
2707 /* Skip if already marked dont_move. */
2708 || (page_table[addr_page_index].dont_move != 0))
2710 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2711 /* (Now that we know that addr_page_index is in range, it's
2712 * safe to index into page_table[] with it.) */
2713 region_allocation = page_table[addr_page_index].allocated;
2715 /* quick check 2: Check the offset within the page.
2718 if (((unsigned long)addr & (PAGE_BYTES - 1)) >
2719 page_table[addr_page_index].bytes_used)
2722 /* Filter out anything which can't be a pointer to a Lisp object
2723 * (or, as a special case which also requires dont_move, a return
2724 * address referring to something in a CodeObject). This is
2725 * expensive but important, since it vastly reduces the
2726 * probability that random garbage will be bogusly interpreted as
2727 * a pointer which prevents a page from moving. */
2728 if (!(code_page_p(addr_page_index)
2729 || (is_lisp_pointer(addr) &&
2730 possibly_valid_dynamic_space_pointer(addr))))
2733 /* Find the beginning of the region. Note that there may be
2734 * objects in the region preceding the one that we were passed a
2735 * pointer to: if this is the case, we will write-protect all the
2736 * previous objects' pages too. */
2739 /* I think this'd work just as well, but without the assertions.
2740 * -dan 2004.01.01 */
2741 first_page = find_page_index(page_region_start(addr_page_index))
2743 first_page = addr_page_index;
2744 while (page_table[first_page].region_start_offset != 0) {
2746 /* Do some checks. */
2747 gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
2748 gc_assert(page_table[first_page].gen == from_space);
2749 gc_assert(page_table[first_page].allocated == region_allocation);
2753 /* Adjust any large objects before promotion as they won't be
2754 * copied after promotion. */
2755 if (page_table[first_page].large_object) {
2756 maybe_adjust_large_object(page_address(first_page));
2757 /* If a large object has shrunk then addr may now point to a
2758 * free area in which case it's ignored here. Note it gets
2759 * through the valid pointer test above because the tail looks
2761 if (page_free_p(addr_page_index)
2762 || (page_table[addr_page_index].bytes_used == 0)
2763 /* Check the offset within the page. */
2764 || (((unsigned long)addr & (PAGE_BYTES - 1))
2765 > page_table[addr_page_index].bytes_used)) {
2767 "weird? ignore ptr 0x%x to freed area of large object\n",
2771 /* It may have moved to unboxed pages. */
2772 region_allocation = page_table[first_page].allocated;
2775 /* Now work forward until the end of this contiguous area is found,
2776 * marking all pages as dont_move. */
2777 for (i = first_page; ;i++) {
2778 gc_assert(page_table[i].allocated == region_allocation);
2780 /* Mark the page static. */
2781 page_table[i].dont_move = 1;
2783 /* Move the page to the new_space. XX I'd rather not do this
2784 * but the GC logic is not quite able to copy with the static
2785 * pages remaining in the from space. This also requires the
2786 * generation bytes_allocated counters be updated. */
2787 page_table[i].gen = new_space;
2788 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2789 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2791 /* It is essential that the pages are not write protected as
2792 * they may have pointers into the old-space which need
2793 * scavenging. They shouldn't be write protected at this
2795 gc_assert(!page_table[i].write_protected);
2797 /* Check whether this is the last page in this contiguous block.. */
2798 if ((page_table[i].bytes_used < PAGE_BYTES)
2799 /* ..or it is PAGE_BYTES and is the last in the block */
2801 || (page_table[i+1].bytes_used == 0) /* next page free */
2802 || (page_table[i+1].gen != from_space) /* diff. gen */
2803 || (page_table[i+1].region_start_offset == 0))
2807 /* Check that the page is now static. */
2808 gc_assert(page_table[addr_page_index].dont_move != 0);
2811 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2814 /* If the given page is not write-protected, then scan it for pointers
2815 * to younger generations or the top temp. generation, if no
2816 * suspicious pointers are found then the page is write-protected.
2818 * Care is taken to check for pointers to the current gc_alloc()
2819 * region if it is a younger generation or the temp. generation. This
2820 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2821 * the gc_alloc_generation does not need to be checked as this is only
2822 * called from scavenge_generation() when the gc_alloc generation is
2823 * younger, so it just checks if there is a pointer to the current
2826 * We return 1 if the page was write-protected, else 0. */
2828 update_page_write_prot(page_index_t page)
2830 generation_index_t gen = page_table[page].gen;
2833 void **page_addr = (void **)page_address(page);
2834 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2836 /* Shouldn't be a free page. */
2837 gc_assert(page_allocated_p(page));
2838 gc_assert(page_table[page].bytes_used != 0);
2840 /* Skip if it's already write-protected, pinned, or unboxed */
2841 if (page_table[page].write_protected
2842 /* FIXME: What's the reason for not write-protecting pinned pages? */
2843 || page_table[page].dont_move
2844 || page_unboxed_p(page))
2847 /* Scan the page for pointers to younger generations or the
2848 * top temp. generation. */
2850 for (j = 0; j < num_words; j++) {
2851 void *ptr = *(page_addr+j);
2852 page_index_t index = find_page_index(ptr);
2854 /* Check that it's in the dynamic space */
2856 if (/* Does it point to a younger or the temp. generation? */
2857 (page_allocated_p(index)
2858 && (page_table[index].bytes_used != 0)
2859 && ((page_table[index].gen < gen)
2860 || (page_table[index].gen == SCRATCH_GENERATION)))
2862 /* Or does it point within a current gc_alloc() region? */
2863 || ((boxed_region.start_addr <= ptr)
2864 && (ptr <= boxed_region.free_pointer))
2865 || ((unboxed_region.start_addr <= ptr)
2866 && (ptr <= unboxed_region.free_pointer))) {
2873 /* Write-protect the page. */
2874 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2876 os_protect((void *)page_addr,
2878 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2880 /* Note the page as protected in the page tables. */
2881 page_table[page].write_protected = 1;
2887 /* Scavenge all generations from FROM to TO, inclusive, except for
2888 * new_space which needs special handling, as new objects may be
2889 * added which are not checked here - use scavenge_newspace generation.
2891 * Write-protected pages should not have any pointers to the
2892 * from_space so do need scavenging; thus write-protected pages are
2893 * not always scavenged. There is some code to check that these pages
2894 * are not written; but to check fully the write-protected pages need
2895 * to be scavenged by disabling the code to skip them.
2897 * Under the current scheme when a generation is GCed the younger
2898 * generations will be empty. So, when a generation is being GCed it
2899 * is only necessary to scavenge the older generations for pointers
2900 * not the younger. So a page that does not have pointers to younger
2901 * generations does not need to be scavenged.
2903 * The write-protection can be used to note pages that don't have
2904 * pointers to younger pages. But pages can be written without having
2905 * pointers to younger generations. After the pages are scavenged here
2906 * they can be scanned for pointers to younger generations and if
2907 * there are none the page can be write-protected.
2909 * One complication is when the newspace is the top temp. generation.
2911 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2912 * that none were written, which they shouldn't be as they should have
2913 * no pointers to younger generations. This breaks down for weak
2914 * pointers as the objects contain a link to the next and are written
2915 * if a weak pointer is scavenged. Still it's a useful check. */
2917 scavenge_generations(generation_index_t from, generation_index_t to)
2924 /* Clear the write_protected_cleared flags on all pages. */
2925 for (i = 0; i < page_table_pages; i++)
2926 page_table[i].write_protected_cleared = 0;
2929 for (i = 0; i < last_free_page; i++) {
2930 generation_index_t generation = page_table[i].gen;
2932 && (page_table[i].bytes_used != 0)
2933 && (generation != new_space)
2934 && (generation >= from)
2935 && (generation <= to)) {
2936 page_index_t last_page,j;
2937 int write_protected=1;
2939 /* This should be the start of a region */
2940 gc_assert(page_table[i].region_start_offset == 0);
2942 /* Now work forward until the end of the region */
2943 for (last_page = i; ; last_page++) {
2945 write_protected && page_table[last_page].write_protected;
2946 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2947 /* Or it is PAGE_BYTES and is the last in the block */
2948 || (!page_boxed_p(last_page+1))
2949 || (page_table[last_page+1].bytes_used == 0)
2950 || (page_table[last_page+1].gen != generation)
2951 || (page_table[last_page+1].region_start_offset == 0))
2954 if (!write_protected) {
2955 scavenge(page_address(i),
2956 ((unsigned long)(page_table[last_page].bytes_used
2957 + npage_bytes(last_page-i)))
2960 /* Now scan the pages and write protect those that
2961 * don't have pointers to younger generations. */
2962 if (enable_page_protection) {
2963 for (j = i; j <= last_page; j++) {
2964 num_wp += update_page_write_prot(j);
2967 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2969 "/write protected %d pages within generation %d\n",
2970 num_wp, generation));
2978 /* Check that none of the write_protected pages in this generation
2979 * have been written to. */
2980 for (i = 0; i < page_table_pages; i++) {
2981 if (page_allocated_p(i)
2982 && (page_table[i].bytes_used != 0)
2983 && (page_table[i].gen == generation)
2984 && (page_table[i].write_protected_cleared != 0)) {
2985 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2987 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
2988 page_table[i].bytes_used,
2989 page_table[i].region_start_offset,
2990 page_table[i].dont_move));
2991 lose("write to protected page %d in scavenge_generation()\n", i);
2998 /* Scavenge a newspace generation. As it is scavenged new objects may
2999 * be allocated to it; these will also need to be scavenged. This
3000 * repeats until there are no more objects unscavenged in the
3001 * newspace generation.
3003 * To help improve the efficiency, areas written are recorded by
3004 * gc_alloc() and only these scavenged. Sometimes a little more will be
3005 * scavenged, but this causes no harm. An easy check is done that the
3006 * scavenged bytes equals the number allocated in the previous
3009 * Write-protected pages are not scanned except if they are marked
3010 * dont_move in which case they may have been promoted and still have
3011 * pointers to the from space.
3013 * Write-protected pages could potentially be written by alloc however
3014 * to avoid having to handle re-scavenging of write-protected pages
3015 * gc_alloc() does not write to write-protected pages.
3017 * New areas of objects allocated are recorded alternatively in the two
3018 * new_areas arrays below. */
3019 static struct new_area new_areas_1[NUM_NEW_AREAS];
3020 static struct new_area new_areas_2[NUM_NEW_AREAS];
3022 /* Do one full scan of the new space generation. This is not enough to
3023 * complete the job as new objects may be added to the generation in
3024 * the process which are not scavenged. */
3026 scavenge_newspace_generation_one_scan(generation_index_t generation)
3031 "/starting one full scan of newspace generation %d\n",
3033 for (i = 0; i < last_free_page; i++) {
3034 /* Note that this skips over open regions when it encounters them. */
3036 && (page_table[i].bytes_used != 0)
3037 && (page_table[i].gen == generation)
3038 && ((page_table[i].write_protected == 0)
3039 /* (This may be redundant as write_protected is now
3040 * cleared before promotion.) */
3041 || (page_table[i].dont_move == 1))) {
3042 page_index_t last_page;
3045 /* The scavenge will start at the region_start_offset of
3048 * We need to find the full extent of this contiguous
3049 * block in case objects span pages.
3051 * Now work forward until the end of this contiguous area
3052 * is found. A small area is preferred as there is a
3053 * better chance of its pages being write-protected. */
3054 for (last_page = i; ;last_page++) {
3055 /* If all pages are write-protected and movable,
3056 * then no need to scavenge */
3057 all_wp=all_wp && page_table[last_page].write_protected &&
3058 !page_table[last_page].dont_move;
3060 /* Check whether this is the last page in this
3061 * contiguous block */
3062 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3063 /* Or it is PAGE_BYTES and is the last in the block */
3064 || (!page_boxed_p(last_page+1))
3065 || (page_table[last_page+1].bytes_used == 0)
3066 || (page_table[last_page+1].gen != generation)
3067 || (page_table[last_page+1].region_start_offset == 0))
3071 /* Do a limited check for write-protected pages. */
3073 long nwords = (((unsigned long)
3074 (page_table[last_page].bytes_used
3075 + npage_bytes(last_page-i)
3076 + page_table[i].region_start_offset))
3078 new_areas_ignore_page = last_page;
3080 scavenge(page_region_start(i), nwords);
3087 "/done with one full scan of newspace generation %d\n",
3091 /* Do a complete scavenge of the newspace generation. */
3093 scavenge_newspace_generation(generation_index_t generation)
3097 /* the new_areas array currently being written to by gc_alloc() */
3098 struct new_area (*current_new_areas)[] = &new_areas_1;
3099 long current_new_areas_index;
3101 /* the new_areas created by the previous scavenge cycle */
3102 struct new_area (*previous_new_areas)[] = NULL;
3103 long previous_new_areas_index;
3105 /* Flush the current regions updating the tables. */
3106 gc_alloc_update_all_page_tables();
3108 /* Turn on the recording of new areas by gc_alloc(). */
3109 new_areas = current_new_areas;
3110 new_areas_index = 0;
3112 /* Don't need to record new areas that get scavenged anyway during
3113 * scavenge_newspace_generation_one_scan. */
3114 record_new_objects = 1;
3116 /* Start with a full scavenge. */
3117 scavenge_newspace_generation_one_scan(generation);
3119 /* Record all new areas now. */
3120 record_new_objects = 2;
3122 /* Give a chance to weak hash tables to make other objects live.
3123 * FIXME: The algorithm implemented here for weak hash table gcing
3124 * is O(W^2+N) as Bruno Haible warns in
3125 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3126 * see "Implementation 2". */
3127 scav_weak_hash_tables();
3129 /* Flush the current regions updating the tables. */
3130 gc_alloc_update_all_page_tables();
3132 /* Grab new_areas_index. */
3133 current_new_areas_index = new_areas_index;
3136 "The first scan is finished; current_new_areas_index=%d.\n",
3137 current_new_areas_index));*/
3139 while (current_new_areas_index > 0) {
3140 /* Move the current to the previous new areas */
3141 previous_new_areas = current_new_areas;
3142 previous_new_areas_index = current_new_areas_index;
3144 /* Scavenge all the areas in previous new areas. Any new areas
3145 * allocated are saved in current_new_areas. */
3147 /* Allocate an array for current_new_areas; alternating between
3148 * new_areas_1 and 2 */
3149 if (previous_new_areas == &new_areas_1)
3150 current_new_areas = &new_areas_2;
3152 current_new_areas = &new_areas_1;
3154 /* Set up for gc_alloc(). */
3155 new_areas = current_new_areas;
3156 new_areas_index = 0;
3158 /* Check whether previous_new_areas had overflowed. */
3159 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3161 /* New areas of objects allocated have been lost so need to do a
3162 * full scan to be sure! If this becomes a problem try
3163 * increasing NUM_NEW_AREAS. */
3164 if (gencgc_verbose) {
3165 SHOW("new_areas overflow, doing full scavenge");
3168 /* Don't need to record new areas that get scavenged
3169 * anyway during scavenge_newspace_generation_one_scan. */
3170 record_new_objects = 1;
3172 scavenge_newspace_generation_one_scan(generation);
3174 /* Record all new areas now. */
3175 record_new_objects = 2;
3177 scav_weak_hash_tables();
3179 /* Flush the current regions updating the tables. */
3180 gc_alloc_update_all_page_tables();
3184 /* Work through previous_new_areas. */
3185 for (i = 0; i < previous_new_areas_index; i++) {
3186 page_index_t page = (*previous_new_areas)[i].page;
3187 size_t offset = (*previous_new_areas)[i].offset;
3188 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3189 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3190 scavenge(page_address(page)+offset, size);
3193 scav_weak_hash_tables();
3195 /* Flush the current regions updating the tables. */
3196 gc_alloc_update_all_page_tables();
3199 current_new_areas_index = new_areas_index;
3202 "The re-scan has finished; current_new_areas_index=%d.\n",
3203 current_new_areas_index));*/
3206 /* Turn off recording of areas allocated by gc_alloc(). */
3207 record_new_objects = 0;
3210 /* Check that none of the write_protected pages in this generation
3211 * have been written to. */
3212 for (i = 0; i < page_table_pages; i++) {
3213 if (page_allocated_p(i)
3214 && (page_table[i].bytes_used != 0)
3215 && (page_table[i].gen == generation)
3216 && (page_table[i].write_protected_cleared != 0)
3217 && (page_table[i].dont_move == 0)) {
3218 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3219 i, generation, page_table[i].dont_move);
3225 /* Un-write-protect all the pages in from_space. This is done at the
3226 * start of a GC else there may be many page faults while scavenging
3227 * the newspace (I've seen drive the system time to 99%). These pages
3228 * would need to be unprotected anyway before unmapping in
3229 * free_oldspace; not sure what effect this has on paging.. */
3231 unprotect_oldspace(void)
3235 for (i = 0; i < last_free_page; i++) {
3236 if (page_allocated_p(i)
3237 && (page_table[i].bytes_used != 0)
3238 && (page_table[i].gen == from_space)) {
3241 page_start = (void *)page_address(i);
3243 /* Remove any write-protection. We should be able to rely
3244 * on the write-protect flag to avoid redundant calls. */
3245 if (page_table[i].write_protected) {
3246 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3247 page_table[i].write_protected = 0;
3253 /* Work through all the pages and free any in from_space. This
3254 * assumes that all objects have been copied or promoted to an older
3255 * generation. Bytes_allocated and the generation bytes_allocated
3256 * counter are updated. The number of bytes freed is returned. */
3257 static unsigned long
3260 unsigned long bytes_freed = 0;
3261 page_index_t first_page, last_page;
3266 /* Find a first page for the next region of pages. */
3267 while ((first_page < last_free_page)
3268 && (page_free_p(first_page)
3269 || (page_table[first_page].bytes_used == 0)
3270 || (page_table[first_page].gen != from_space)))
3273 if (first_page >= last_free_page)
3276 /* Find the last page of this region. */
3277 last_page = first_page;
3280 /* Free the page. */
3281 bytes_freed += page_table[last_page].bytes_used;
3282 generations[page_table[last_page].gen].bytes_allocated -=
3283 page_table[last_page].bytes_used;
3284 page_table[last_page].allocated = FREE_PAGE_FLAG;
3285 page_table[last_page].bytes_used = 0;
3287 /* Remove any write-protection. We should be able to rely
3288 * on the write-protect flag to avoid redundant calls. */
3290 void *page_start = (void *)page_address(last_page);
3292 if (page_table[last_page].write_protected) {
3293 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3294 page_table[last_page].write_protected = 0;
3299 while ((last_page < last_free_page)
3300 && page_allocated_p(last_page)
3301 && (page_table[last_page].bytes_used != 0)
3302 && (page_table[last_page].gen == from_space));
3304 #ifdef READ_PROTECT_FREE_PAGES
3305 os_protect(page_address(first_page),
3306 npage_bytes(last_page-first_page),
3309 first_page = last_page;
3310 } while (first_page < last_free_page);
3312 bytes_allocated -= bytes_freed;
3317 /* Print some information about a pointer at the given address. */
3319 print_ptr(lispobj *addr)
3321 /* If addr is in the dynamic space then out the page information. */
3322 page_index_t pi1 = find_page_index((void*)addr);
3325 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
3326 (unsigned long) addr,
3328 page_table[pi1].allocated,
3329 page_table[pi1].gen,
3330 page_table[pi1].bytes_used,
3331 page_table[pi1].region_start_offset,
3332 page_table[pi1].dont_move);
3333 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3347 verify_space(lispobj *start, size_t words)
3349 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3350 int is_in_readonly_space =
3351 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3352 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3356 lispobj thing = *(lispobj*)start;
3358 if (is_lisp_pointer(thing)) {
3359 page_index_t page_index = find_page_index((void*)thing);
3360 long to_readonly_space =
3361 (READ_ONLY_SPACE_START <= thing &&
3362 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3363 long to_static_space =
3364 (STATIC_SPACE_START <= thing &&
3365 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3367 /* Does it point to the dynamic space? */
3368 if (page_index != -1) {
3369 /* If it's within the dynamic space it should point to a used
3370 * page. XX Could check the offset too. */
3371 if (page_allocated_p(page_index)
3372 && (page_table[page_index].bytes_used == 0))
3373 lose ("Ptr %x @ %x sees free page.\n", thing, start);
3374 /* Check that it doesn't point to a forwarding pointer! */
3375 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3376 lose("Ptr %x @ %x sees forwarding ptr.\n", thing, start);
3378 /* Check that its not in the RO space as it would then be a
3379 * pointer from the RO to the dynamic space. */
3380 if (is_in_readonly_space) {
3381 lose("ptr to dynamic space %x from RO space %x\n",
3384 /* Does it point to a plausible object? This check slows
3385 * it down a lot (so it's commented out).
3387 * "a lot" is serious: it ate 50 minutes cpu time on
3388 * my duron 950 before I came back from lunch and
3391 * FIXME: Add a variable to enable this
3394 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3395 lose("ptr %x to invalid object %x\n", thing, start);
3399 /* Verify that it points to another valid space. */
3400 if (!to_readonly_space && !to_static_space) {
3401 lose("Ptr %x @ %x sees junk.\n", thing, start);
3405 if (!(fixnump(thing))) {
3407 switch(widetag_of(*start)) {
3410 case SIMPLE_VECTOR_WIDETAG:
3412 case COMPLEX_WIDETAG:
3413 case SIMPLE_ARRAY_WIDETAG:
3414 case COMPLEX_BASE_STRING_WIDETAG:
3415 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3416 case COMPLEX_CHARACTER_STRING_WIDETAG:
3418 case COMPLEX_VECTOR_NIL_WIDETAG:
3419 case COMPLEX_BIT_VECTOR_WIDETAG:
3420 case COMPLEX_VECTOR_WIDETAG:
3421 case COMPLEX_ARRAY_WIDETAG:
3422 case CLOSURE_HEADER_WIDETAG:
3423 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3424 case VALUE_CELL_HEADER_WIDETAG:
3425 case SYMBOL_HEADER_WIDETAG:
3426 case CHARACTER_WIDETAG:
3427 #if N_WORD_BITS == 64
3428 case SINGLE_FLOAT_WIDETAG:
3430 case UNBOUND_MARKER_WIDETAG:
3435 case INSTANCE_HEADER_WIDETAG:
3438 long ntotal = HeaderValue(thing);
3439 lispobj layout = ((struct instance *)start)->slots[0];
3444 nuntagged = ((struct layout *)
3445 native_pointer(layout))->n_untagged_slots;
3446 verify_space(start + 1,
3447 ntotal - fixnum_value(nuntagged));
3451 case CODE_HEADER_WIDETAG:
3453 lispobj object = *start;
3455 long nheader_words, ncode_words, nwords;
3457 struct simple_fun *fheaderp;
3459 code = (struct code *) start;
3461 /* Check that it's not in the dynamic space.
3462 * FIXME: Isn't is supposed to be OK for code
3463 * objects to be in the dynamic space these days? */
3464 if (is_in_dynamic_space
3465 /* It's ok if it's byte compiled code. The trace
3466 * table offset will be a fixnum if it's x86
3467 * compiled code - check.
3469 * FIXME: #^#@@! lack of abstraction here..
3470 * This line can probably go away now that
3471 * there's no byte compiler, but I've got
3472 * too much to worry about right now to try
3473 * to make sure. -- WHN 2001-10-06 */
3474 && fixnump(code->trace_table_offset)
3475 /* Only when enabled */
3476 && verify_dynamic_code_check) {
3478 "/code object at %x in the dynamic space\n",
3482 ncode_words = fixnum_value(code->code_size);
3483 nheader_words = HeaderValue(object);
3484 nwords = ncode_words + nheader_words;
3485 nwords = CEILING(nwords, 2);
3486 /* Scavenge the boxed section of the code data block */
3487 verify_space(start + 1, nheader_words - 1);
3489 /* Scavenge the boxed section of each function
3490 * object in the code data block. */
3491 fheaderl = code->entry_points;
3492 while (fheaderl != NIL) {
3494 (struct simple_fun *) native_pointer(fheaderl);
3495 gc_assert(widetag_of(fheaderp->header) ==
3496 SIMPLE_FUN_HEADER_WIDETAG);
3497 verify_space(&fheaderp->name, 1);
3498 verify_space(&fheaderp->arglist, 1);
3499 verify_space(&fheaderp->type, 1);
3500 fheaderl = fheaderp->next;
3506 /* unboxed objects */
3507 case BIGNUM_WIDETAG:
3508 #if N_WORD_BITS != 64
3509 case SINGLE_FLOAT_WIDETAG:
3511 case DOUBLE_FLOAT_WIDETAG:
3512 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3513 case LONG_FLOAT_WIDETAG:
3515 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3516 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3518 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3519 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3521 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3522 case COMPLEX_LONG_FLOAT_WIDETAG:
3524 case SIMPLE_BASE_STRING_WIDETAG:
3525 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3526 case SIMPLE_CHARACTER_STRING_WIDETAG:
3528 case SIMPLE_BIT_VECTOR_WIDETAG:
3529 case SIMPLE_ARRAY_NIL_WIDETAG:
3530 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3531 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3532 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3533 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3534 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3535 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3536 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
3537 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
3539 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3540 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3541 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
3542 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
3544 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3545 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3547 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3548 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3550 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3551 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3553 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3554 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3556 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
3557 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
3559 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3560 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3562 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
3563 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
3565 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3566 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3568 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3569 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3570 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3571 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3573 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3574 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3576 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3577 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3579 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3580 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3583 case WEAK_POINTER_WIDETAG:
3584 #ifdef LUTEX_WIDETAG
3587 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3588 case NO_TLS_VALUE_MARKER_WIDETAG:
3590 count = (sizetab[widetag_of(*start)])(start);
3594 lose("Unhandled widetag 0x%x at 0x%x\n",
3595 widetag_of(*start), start);
3607 /* FIXME: It would be nice to make names consistent so that
3608 * foo_size meant size *in* *bytes* instead of size in some
3609 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3610 * Some counts of lispobjs are called foo_count; it might be good
3611 * to grep for all foo_size and rename the appropriate ones to
3613 long read_only_space_size =
3614 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3615 - (lispobj*)READ_ONLY_SPACE_START;
3616 long static_space_size =
3617 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3618 - (lispobj*)STATIC_SPACE_START;
3620 for_each_thread(th) {
3621 long binding_stack_size =
3622 (lispobj*)get_binding_stack_pointer(th)
3623 - (lispobj*)th->binding_stack_start;
3624 verify_space(th->binding_stack_start, binding_stack_size);
3626 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3627 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3631 verify_generation(generation_index_t generation)
3635 for (i = 0; i < last_free_page; i++) {
3636 if (page_allocated_p(i)
3637 && (page_table[i].bytes_used != 0)
3638 && (page_table[i].gen == generation)) {
3639 page_index_t last_page;
3640 int region_allocation = page_table[i].allocated;
3642 /* This should be the start of a contiguous block */
3643 gc_assert(page_table[i].region_start_offset == 0);
3645 /* Need to find the full extent of this contiguous block in case
3646 objects span pages. */
3648 /* Now work forward until the end of this contiguous area is
3650 for (last_page = i; ;last_page++)
3651 /* Check whether this is the last page in this contiguous
3653 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3654 /* Or it is PAGE_BYTES and is the last in the block */
3655 || (page_table[last_page+1].allocated != region_allocation)
3656 || (page_table[last_page+1].bytes_used == 0)
3657 || (page_table[last_page+1].gen != generation)
3658 || (page_table[last_page+1].region_start_offset == 0))
3661 verify_space(page_address(i),
3663 (page_table[last_page].bytes_used
3664 + npage_bytes(last_page-i)))
3671 /* Check that all the free space is zero filled. */
3673 verify_zero_fill(void)
3677 for (page = 0; page < last_free_page; page++) {
3678 if (page_free_p(page)) {
3679 /* The whole page should be zero filled. */
3680 long *start_addr = (long *)page_address(page);
3683 for (i = 0; i < size; i++) {
3684 if (start_addr[i] != 0) {
3685 lose("free page not zero at %x\n", start_addr + i);
3689 long free_bytes = PAGE_BYTES - page_table[page].bytes_used;
3690 if (free_bytes > 0) {
3691 long *start_addr = (long *)((unsigned long)page_address(page)
3692 + page_table[page].bytes_used);
3693 long size = free_bytes / N_WORD_BYTES;
3695 for (i = 0; i < size; i++) {
3696 if (start_addr[i] != 0) {
3697 lose("free region not zero at %x\n", start_addr + i);
3705 /* External entry point for verify_zero_fill */
3707 gencgc_verify_zero_fill(void)
3709 /* Flush the alloc regions updating the tables. */
3710 gc_alloc_update_all_page_tables();
3711 SHOW("verifying zero fill");
3716 verify_dynamic_space(void)
3718 generation_index_t i;
3720 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3721 verify_generation(i);
3723 if (gencgc_enable_verify_zero_fill)
3727 /* Write-protect all the dynamic boxed pages in the given generation. */
3729 write_protect_generation_pages(generation_index_t generation)
3733 gc_assert(generation < SCRATCH_GENERATION);
3735 for (start = 0; start < last_free_page; start++) {
3736 if (protect_page_p(start, generation)) {
3740 /* Note the page as protected in the page tables. */
3741 page_table[start].write_protected = 1;
3743 for (last = start + 1; last < last_free_page; last++) {
3744 if (!protect_page_p(last, generation))
3746 page_table[last].write_protected = 1;
3749 page_start = (void *)page_address(start);
3751 os_protect(page_start,
3752 npage_bytes(last - start),
3753 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3759 if (gencgc_verbose > 1) {
3761 "/write protected %d of %d pages in generation %d\n",
3762 count_write_protect_generation_pages(generation),
3763 count_generation_pages(generation),
3768 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3771 scavenge_control_stack()
3773 unsigned long control_stack_size;
3775 /* This is going to be a big problem when we try to port threads
3777 struct thread *th = arch_os_get_current_thread();
3778 lispobj *control_stack =
3779 (lispobj *)(th->control_stack_start);
3781 control_stack_size = current_control_stack_pointer - control_stack;
3782 scavenge(control_stack, control_stack_size);
3785 /* Scavenging Interrupt Contexts */
3787 static int boxed_registers[] = BOXED_REGISTERS;
3790 scavenge_interrupt_context(os_context_t * context)
3796 unsigned long lip_offset;
3797 int lip_register_pair;
3799 unsigned long pc_code_offset;
3801 #ifdef ARCH_HAS_LINK_REGISTER
3802 unsigned long lr_code_offset;
3804 #ifdef ARCH_HAS_NPC_REGISTER
3805 unsigned long npc_code_offset;
3809 /* Find the LIP's register pair and calculate it's offset */
3810 /* before we scavenge the context. */
3813 * I (RLT) think this is trying to find the boxed register that is
3814 * closest to the LIP address, without going past it. Usually, it's
3815 * reg_CODE or reg_LRA. But sometimes, nothing can be found.
3817 lip = *os_context_register_addr(context, reg_LIP);
3818 lip_offset = 0x7FFFFFFF;
3819 lip_register_pair = -1;
3820 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3825 index = boxed_registers[i];
3826 reg = *os_context_register_addr(context, index);
3827 if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
3829 if (offset < lip_offset) {
3830 lip_offset = offset;
3831 lip_register_pair = index;
3835 #endif /* reg_LIP */
3837 /* Compute the PC's offset from the start of the CODE */
3839 pc_code_offset = *os_context_pc_addr(context)
3840 - *os_context_register_addr(context, reg_CODE);
3841 #ifdef ARCH_HAS_NPC_REGISTER
3842 npc_code_offset = *os_context_npc_addr(context)
3843 - *os_context_register_addr(context, reg_CODE);
3844 #endif /* ARCH_HAS_NPC_REGISTER */
3846 #ifdef ARCH_HAS_LINK_REGISTER
3848 *os_context_lr_addr(context) -
3849 *os_context_register_addr(context, reg_CODE);
3852 /* Scanvenge all boxed registers in the context. */
3853 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3857 index = boxed_registers[i];
3858 foo = *os_context_register_addr(context, index);
3860 *os_context_register_addr(context, index) = foo;
3862 scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
3869 * But what happens if lip_register_pair is -1?
3870 * *os_context_register_addr on Solaris (see
3871 * solaris_register_address in solaris-os.c) will return
3872 * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
3873 * that what we really want? My guess is that that is not what we
3874 * want, so if lip_register_pair is -1, we don't touch reg_LIP at
3875 * all. But maybe it doesn't really matter if LIP is trashed?
3877 if (lip_register_pair >= 0) {
3878 *os_context_register_addr(context, reg_LIP) =
3879 *os_context_register_addr(context, lip_register_pair)
3882 #endif /* reg_LIP */
3884 /* Fix the PC if it was in from space */
3885 if (from_space_p(*os_context_pc_addr(context)))
3886 *os_context_pc_addr(context) =
3887 *os_context_register_addr(context, reg_CODE) + pc_code_offset;
3889 #ifdef ARCH_HAS_LINK_REGISTER
3890 /* Fix the LR ditto; important if we're being called from
3891 * an assembly routine that expects to return using blr, otherwise
3893 if (from_space_p(*os_context_lr_addr(context)))
3894 *os_context_lr_addr(context) =
3895 *os_context_register_addr(context, reg_CODE) + lr_code_offset;
3898 #ifdef ARCH_HAS_NPC_REGISTER
3899 if (from_space_p(*os_context_npc_addr(context)))
3900 *os_context_npc_addr(context) =
3901 *os_context_register_addr(context, reg_CODE) + npc_code_offset;
3902 #endif /* ARCH_HAS_NPC_REGISTER */
3906 scavenge_interrupt_contexts(void)
3909 os_context_t *context;
3911 struct thread *th=arch_os_get_current_thread();
3913 index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
3915 #if defined(DEBUG_PRINT_CONTEXT_INDEX)
3916 printf("Number of active contexts: %d\n", index);
3919 for (i = 0; i < index; i++) {
3920 context = th->interrupt_contexts[i];
3921 scavenge_interrupt_context(context);
3927 #if defined(LISP_FEATURE_SB_THREAD)
3929 preserve_context_registers (os_context_t *c)
3932 /* On Darwin the signal context isn't a contiguous block of memory,
3933 * so just preserve_pointering its contents won't be sufficient.
3935 #if defined(LISP_FEATURE_DARWIN)
3936 #if defined LISP_FEATURE_X86
3937 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3938 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3939 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3940 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3941 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3942 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3943 preserve_pointer((void*)*os_context_pc_addr(c));
3944 #elif defined LISP_FEATURE_X86_64
3945 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3946 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3947 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3948 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3949 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3950 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3951 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3952 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3953 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3954 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3955 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3956 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3957 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3958 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3959 preserve_pointer((void*)*os_context_pc_addr(c));
3961 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3964 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3965 preserve_pointer(*ptr);
3970 /* Garbage collect a generation. If raise is 0 then the remains of the
3971 * generation are not raised to the next generation. */
3973 garbage_collect_generation(generation_index_t generation, int raise)
3975 unsigned long bytes_freed;
3977 unsigned long static_space_size;
3978 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3981 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3983 /* The oldest generation can't be raised. */
3984 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3986 /* Check if weak hash tables were processed in the previous GC. */
3987 gc_assert(weak_hash_tables == NULL);
3989 /* Initialize the weak pointer list. */
3990 weak_pointers = NULL;
3992 #ifdef LUTEX_WIDETAG
3993 unmark_lutexes(generation);
3996 /* When a generation is not being raised it is transported to a
3997 * temporary generation (NUM_GENERATIONS), and lowered when
3998 * done. Set up this new generation. There should be no pages
3999 * allocated to it yet. */
4001 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
4004 /* Set the global src and dest. generations */
4005 from_space = generation;
4007 new_space = generation+1;
4009 new_space = SCRATCH_GENERATION;
4011 /* Change to a new space for allocation, resetting the alloc_start_page */
4012 gc_alloc_generation = new_space;
4013 generations[new_space].alloc_start_page = 0;
4014 generations[new_space].alloc_unboxed_start_page = 0;
4015 generations[new_space].alloc_large_start_page = 0;
4016 generations[new_space].alloc_large_unboxed_start_page = 0;
4018 /* Before any pointers are preserved, the dont_move flags on the
4019 * pages need to be cleared. */
4020 for (i = 0; i < last_free_page; i++)
4021 if(page_table[i].gen==from_space)
4022 page_table[i].dont_move = 0;
4024 /* Un-write-protect the old-space pages. This is essential for the
4025 * promoted pages as they may contain pointers into the old-space
4026 * which need to be scavenged. It also helps avoid unnecessary page
4027 * faults as forwarding pointers are written into them. They need to
4028 * be un-protected anyway before unmapping later. */
4029 unprotect_oldspace();
4031 /* Scavenge the stacks' conservative roots. */
4033 /* there are potentially two stacks for each thread: the main
4034 * stack, which may contain Lisp pointers, and the alternate stack.
4035 * We don't ever run Lisp code on the altstack, but it may
4036 * host a sigcontext with lisp objects in it */
4038 /* what we need to do: (1) find the stack pointer for the main
4039 * stack; scavenge it (2) find the interrupt context on the
4040 * alternate stack that might contain lisp values, and scavenge
4043 /* we assume that none of the preceding applies to the thread that
4044 * initiates GC. If you ever call GC from inside an altstack
4045 * handler, you will lose. */
4047 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4048 /* And if we're saving a core, there's no point in being conservative. */
4049 if (conservative_stack) {
4050 for_each_thread(th) {
4052 void **esp=(void **)-1;
4053 #ifdef LISP_FEATURE_SB_THREAD
4055 if(th==arch_os_get_current_thread()) {
4056 /* Somebody is going to burn in hell for this, but casting
4057 * it in two steps shuts gcc up about strict aliasing. */
4058 esp = (void **)((void *)&raise);
4061 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
4062 for(i=free-1;i>=0;i--) {
4063 os_context_t *c=th->interrupt_contexts[i];
4064 esp1 = (void **) *os_context_register_addr(c,reg_SP);
4065 if (esp1>=(void **)th->control_stack_start &&
4066 esp1<(void **)th->control_stack_end) {
4067 if(esp1<esp) esp=esp1;
4068 preserve_context_registers(c);
4073 esp = (void **)((void *)&raise);
4075 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
4076 preserve_pointer(*ptr);
4083 if (gencgc_verbose > 1) {
4084 long num_dont_move_pages = count_dont_move_pages();
4086 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
4087 num_dont_move_pages,
4088 npage_bytes(num_dont_move_pages);
4092 /* Scavenge all the rest of the roots. */
4094 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4096 * If not x86, we need to scavenge the interrupt context(s) and the
4099 scavenge_interrupt_contexts();
4100 scavenge_control_stack();
4103 /* Scavenge the Lisp functions of the interrupt handlers, taking
4104 * care to avoid SIG_DFL and SIG_IGN. */
4105 for (i = 0; i < NSIG; i++) {
4106 union interrupt_handler handler = interrupt_handlers[i];
4107 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4108 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4109 scavenge((lispobj *)(interrupt_handlers + i), 1);
4112 /* Scavenge the binding stacks. */
4115 for_each_thread(th) {
4116 long len= (lispobj *)get_binding_stack_pointer(th) -
4117 th->binding_stack_start;
4118 scavenge((lispobj *) th->binding_stack_start,len);
4119 #ifdef LISP_FEATURE_SB_THREAD
4120 /* do the tls as well */
4121 len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
4122 (sizeof (struct thread))/(sizeof (lispobj));
4123 scavenge((lispobj *) (th+1),len);
4128 /* The original CMU CL code had scavenge-read-only-space code
4129 * controlled by the Lisp-level variable
4130 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4131 * wasn't documented under what circumstances it was useful or
4132 * safe to turn it on, so it's been turned off in SBCL. If you
4133 * want/need this functionality, and can test and document it,
4134 * please submit a patch. */
4136 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4137 unsigned long read_only_space_size =
4138 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4139 (lispobj*)READ_ONLY_SPACE_START;
4141 "/scavenge read only space: %d bytes\n",
4142 read_only_space_size * sizeof(lispobj)));
4143 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4147 /* Scavenge static space. */
4149 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4150 (lispobj *)STATIC_SPACE_START;
4151 if (gencgc_verbose > 1) {
4153 "/scavenge static space: %d bytes\n",
4154 static_space_size * sizeof(lispobj)));
4156 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4158 /* All generations but the generation being GCed need to be
4159 * scavenged. The new_space generation needs special handling as
4160 * objects may be moved in - it is handled separately below. */
4161 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4163 /* Finally scavenge the new_space generation. Keep going until no
4164 * more objects are moved into the new generation */
4165 scavenge_newspace_generation(new_space);
4167 /* FIXME: I tried reenabling this check when debugging unrelated
4168 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4169 * Since the current GC code seems to work well, I'm guessing that
4170 * this debugging code is just stale, but I haven't tried to
4171 * figure it out. It should be figured out and then either made to
4172 * work or just deleted. */
4173 #define RESCAN_CHECK 0
4175 /* As a check re-scavenge the newspace once; no new objects should
4178 long old_bytes_allocated = bytes_allocated;
4179 long bytes_allocated;
4181 /* Start with a full scavenge. */
4182 scavenge_newspace_generation_one_scan(new_space);
4184 /* Flush the current regions, updating the tables. */
4185 gc_alloc_update_all_page_tables();
4187 bytes_allocated = bytes_allocated - old_bytes_allocated;
4189 if (bytes_allocated != 0) {
4190 lose("Rescan of new_space allocated %d more bytes.\n",
4196 scan_weak_hash_tables();
4197 scan_weak_pointers();
4199 /* Flush the current regions, updating the tables. */
4200 gc_alloc_update_all_page_tables();
4202 /* Free the pages in oldspace, but not those marked dont_move. */
4203 bytes_freed = free_oldspace();
4205 /* If the GC is not raising the age then lower the generation back
4206 * to its normal generation number */
4208 for (i = 0; i < last_free_page; i++)
4209 if ((page_table[i].bytes_used != 0)
4210 && (page_table[i].gen == SCRATCH_GENERATION))
4211 page_table[i].gen = generation;
4212 gc_assert(generations[generation].bytes_allocated == 0);
4213 generations[generation].bytes_allocated =
4214 generations[SCRATCH_GENERATION].bytes_allocated;
4215 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4218 /* Reset the alloc_start_page for generation. */
4219 generations[generation].alloc_start_page = 0;
4220 generations[generation].alloc_unboxed_start_page = 0;
4221 generations[generation].alloc_large_start_page = 0;
4222 generations[generation].alloc_large_unboxed_start_page = 0;
4224 if (generation >= verify_gens) {
4225 if (gencgc_verbose) {
4229 verify_dynamic_space();
4232 /* Set the new gc trigger for the GCed generation. */
4233 generations[generation].gc_trigger =
4234 generations[generation].bytes_allocated
4235 + generations[generation].bytes_consed_between_gc;
4238 generations[generation].num_gc = 0;
4240 ++generations[generation].num_gc;
4242 #ifdef LUTEX_WIDETAG
4243 reap_lutexes(generation);
4245 move_lutexes(generation, generation+1);
4249 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4251 update_dynamic_space_free_pointer(void)
4253 page_index_t last_page = -1, i;
4255 for (i = 0; i < last_free_page; i++)
4256 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
4259 last_free_page = last_page+1;
4261 set_alloc_pointer((lispobj)(page_address(last_free_page)));
4262 return 0; /* dummy value: return something ... */
4266 remap_free_pages (page_index_t from, page_index_t to)
4268 page_index_t first_page, last_page;
4270 for (first_page = from; first_page <= to; first_page++) {
4271 if (page_allocated_p(first_page) ||
4272 (page_table[first_page].need_to_zero == 0)) {
4276 last_page = first_page + 1;
4277 while (page_free_p(last_page) &&
4279 (page_table[last_page].need_to_zero == 1)) {
4283 /* There's a mysterious Solaris/x86 problem with using mmap
4284 * tricks for memory zeroing. See sbcl-devel thread
4285 * "Re: patch: standalone executable redux".
4287 #if defined(LISP_FEATURE_SUNOS)
4288 zero_pages(first_page, last_page-1);
4290 zero_pages_with_mmap(first_page, last_page-1);
4293 first_page = last_page;
4297 generation_index_t small_generation_limit = 1;
4299 /* GC all generations newer than last_gen, raising the objects in each
4300 * to the next older generation - we finish when all generations below
4301 * last_gen are empty. Then if last_gen is due for a GC, or if
4302 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4303 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4305 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4306 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4308 collect_garbage(generation_index_t last_gen)
4310 generation_index_t gen = 0, i;
4313 /* The largest value of last_free_page seen since the time
4314 * remap_free_pages was called. */
4315 static page_index_t high_water_mark = 0;
4317 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4321 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4323 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4328 /* Flush the alloc regions updating the tables. */
4329 gc_alloc_update_all_page_tables();
4331 /* Verify the new objects created by Lisp code. */
4332 if (pre_verify_gen_0) {
4333 FSHOW((stderr, "pre-checking generation 0\n"));
4334 verify_generation(0);
4337 if (gencgc_verbose > 1)
4338 print_generation_stats(0);
4341 /* Collect the generation. */
4343 if (gen >= gencgc_oldest_gen_to_gc) {
4344 /* Never raise the oldest generation. */
4349 || (generations[gen].num_gc >= generations[gen].trigger_age);
4352 if (gencgc_verbose > 1) {
4354 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4357 generations[gen].bytes_allocated,
4358 generations[gen].gc_trigger,
4359 generations[gen].num_gc));
4362 /* If an older generation is being filled, then update its
4365 generations[gen+1].cum_sum_bytes_allocated +=
4366 generations[gen+1].bytes_allocated;
4369 garbage_collect_generation(gen, raise);
4371 /* Reset the memory age cum_sum. */
4372 generations[gen].cum_sum_bytes_allocated = 0;
4374 if (gencgc_verbose > 1) {
4375 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4376 print_generation_stats(0);
4380 } while ((gen <= gencgc_oldest_gen_to_gc)
4381 && ((gen < last_gen)
4382 || ((gen <= gencgc_oldest_gen_to_gc)
4384 && (generations[gen].bytes_allocated
4385 > generations[gen].gc_trigger)
4386 && (gen_av_mem_age(gen)
4387 > generations[gen].min_av_mem_age))));
4389 /* Now if gen-1 was raised all generations before gen are empty.
4390 * If it wasn't raised then all generations before gen-1 are empty.
4392 * Now objects within this gen's pages cannot point to younger
4393 * generations unless they are written to. This can be exploited
4394 * by write-protecting the pages of gen; then when younger
4395 * generations are GCed only the pages which have been written
4400 gen_to_wp = gen - 1;
4402 /* There's not much point in WPing pages in generation 0 as it is
4403 * never scavenged (except promoted pages). */
4404 if ((gen_to_wp > 0) && enable_page_protection) {
4405 /* Check that they are all empty. */
4406 for (i = 0; i < gen_to_wp; i++) {
4407 if (generations[i].bytes_allocated)
4408 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4411 write_protect_generation_pages(gen_to_wp);
4414 /* Set gc_alloc() back to generation 0. The current regions should
4415 * be flushed after the above GCs. */
4416 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4417 gc_alloc_generation = 0;
4419 /* Save the high-water mark before updating last_free_page */
4420 if (last_free_page > high_water_mark)
4421 high_water_mark = last_free_page;
4423 update_dynamic_space_free_pointer();
4425 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4427 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4430 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4433 if (gen > small_generation_limit) {
4434 if (last_free_page > high_water_mark)
4435 high_water_mark = last_free_page;
4436 remap_free_pages(0, high_water_mark);
4437 high_water_mark = 0;
4442 SHOW("returning from collect_garbage");
4445 /* This is called by Lisp PURIFY when it is finished. All live objects
4446 * will have been moved to the RO and Static heaps. The dynamic space
4447 * will need a full re-initialization. We don't bother having Lisp
4448 * PURIFY flush the current gc_alloc() region, as the page_tables are
4449 * re-initialized, and every page is zeroed to be sure. */
4455 if (gencgc_verbose > 1) {
4456 SHOW("entering gc_free_heap");
4459 for (page = 0; page < page_table_pages; page++) {
4460 /* Skip free pages which should already be zero filled. */
4461 if (page_allocated_p(page)) {
4462 void *page_start, *addr;
4464 /* Mark the page free. The other slots are assumed invalid
4465 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4466 * should not be write-protected -- except that the
4467 * generation is used for the current region but it sets
4469 page_table[page].allocated = FREE_PAGE_FLAG;
4470 page_table[page].bytes_used = 0;
4472 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4473 * about this change. */
4474 /* Zero the page. */
4475 page_start = (void *)page_address(page);
4477 /* First, remove any write-protection. */
4478 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
4479 page_table[page].write_protected = 0;
4481 os_invalidate(page_start,PAGE_BYTES);
4482 addr = os_validate(page_start,PAGE_BYTES);
4483 if (addr == NULL || addr != page_start) {
4484 lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n",
4489 page_table[page].write_protected = 0;
4491 } else if (gencgc_zero_check_during_free_heap) {
4492 /* Double-check that the page is zero filled. */
4495 gc_assert(page_free_p(page));
4496 gc_assert(page_table[page].bytes_used == 0);
4497 page_start = (long *)page_address(page);
4498 for (i=0; i<1024; i++) {
4499 if (page_start[i] != 0) {
4500 lose("free region not zero at %x\n", page_start + i);
4506 bytes_allocated = 0;
4508 /* Initialize the generations. */
4509 for (page = 0; page < NUM_GENERATIONS; page++) {
4510 generations[page].alloc_start_page = 0;
4511 generations[page].alloc_unboxed_start_page = 0;
4512 generations[page].alloc_large_start_page = 0;
4513 generations[page].alloc_large_unboxed_start_page = 0;
4514 generations[page].bytes_allocated = 0;
4515 generations[page].gc_trigger = 2000000;
4516 generations[page].num_gc = 0;
4517 generations[page].cum_sum_bytes_allocated = 0;
4518 generations[page].lutexes = NULL;
4521 if (gencgc_verbose > 1)
4522 print_generation_stats(0);
4524 /* Initialize gc_alloc(). */
4525 gc_alloc_generation = 0;
4527 gc_set_region_empty(&boxed_region);
4528 gc_set_region_empty(&unboxed_region);
4531 set_alloc_pointer((lispobj)((char *)heap_base));
4533 if (verify_after_free_heap) {
4534 /* Check whether purify has left any bad pointers. */
4535 FSHOW((stderr, "checking after free_heap\n"));
4545 /* Compute the number of pages needed for the dynamic space.
4546 * Dynamic space size should be aligned on page size. */
4547 page_table_pages = dynamic_space_size/PAGE_BYTES;
4548 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4550 page_table = calloc(page_table_pages, sizeof(struct page));
4551 gc_assert(page_table);
4554 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4555 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4557 #ifdef LUTEX_WIDETAG
4558 scavtab[LUTEX_WIDETAG] = scav_lutex;
4559 transother[LUTEX_WIDETAG] = trans_lutex;
4560 sizetab[LUTEX_WIDETAG] = size_lutex;
4563 heap_base = (void*)DYNAMIC_SPACE_START;
4565 /* Initialize each page structure. */
4566 for (i = 0; i < page_table_pages; i++) {
4567 /* Initialize all pages as free. */
4568 page_table[i].allocated = FREE_PAGE_FLAG;
4569 page_table[i].bytes_used = 0;
4571 /* Pages are not write-protected at startup. */
4572 page_table[i].write_protected = 0;
4575 bytes_allocated = 0;
4577 /* Initialize the generations.
4579 * FIXME: very similar to code in gc_free_heap(), should be shared */
4580 for (i = 0; i < NUM_GENERATIONS; i++) {
4581 generations[i].alloc_start_page = 0;
4582 generations[i].alloc_unboxed_start_page = 0;
4583 generations[i].alloc_large_start_page = 0;
4584 generations[i].alloc_large_unboxed_start_page = 0;
4585 generations[i].bytes_allocated = 0;
4586 generations[i].gc_trigger = 2000000;
4587 generations[i].num_gc = 0;
4588 generations[i].cum_sum_bytes_allocated = 0;
4589 /* the tune-able parameters */
4590 generations[i].bytes_consed_between_gc = 2000000;
4591 generations[i].trigger_age = 1;
4592 generations[i].min_av_mem_age = 0.75;
4593 generations[i].lutexes = NULL;
4596 /* Initialize gc_alloc. */
4597 gc_alloc_generation = 0;
4598 gc_set_region_empty(&boxed_region);
4599 gc_set_region_empty(&unboxed_region);
4604 /* Pick up the dynamic space from after a core load.
4606 * The ALLOCATION_POINTER points to the end of the dynamic space.
4610 gencgc_pickup_dynamic(void)
4612 page_index_t page = 0;
4613 void *alloc_ptr = (void *)get_alloc_pointer();
4614 lispobj *prev=(lispobj *)page_address(page);
4615 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4617 lispobj *first,*ptr= (lispobj *)page_address(page);
4618 page_table[page].allocated = BOXED_PAGE_FLAG;
4619 page_table[page].gen = gen;
4620 page_table[page].bytes_used = PAGE_BYTES;
4621 page_table[page].large_object = 0;
4622 page_table[page].write_protected = 0;
4623 page_table[page].write_protected_cleared = 0;
4624 page_table[page].dont_move = 0;
4625 page_table[page].need_to_zero = 1;
4627 if (!gencgc_partial_pickup) {
4628 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4629 if(ptr == first) prev=ptr;
4630 page_table[page].region_start_offset =
4631 page_address(page) - (void *)prev;
4634 } while (page_address(page) < alloc_ptr);
4636 #ifdef LUTEX_WIDETAG
4637 /* Lutexes have been registered in generation 0 by coreparse, and
4638 * need to be moved to the right one manually.
4640 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4643 last_free_page = page;
4645 generations[gen].bytes_allocated = npage_bytes(page);
4646 bytes_allocated = npage_bytes(page);
4648 gc_alloc_update_all_page_tables();
4649 write_protect_generation_pages(gen);
4653 gc_initialize_pointers(void)
4655 gencgc_pickup_dynamic();
4659 /* alloc(..) is the external interface for memory allocation. It
4660 * allocates to generation 0. It is not called from within the garbage
4661 * collector as it is only external uses that need the check for heap
4662 * size (GC trigger) and to disable the interrupts (interrupts are
4663 * always disabled during a GC).
4665 * The vops that call alloc(..) assume that the returned space is zero-filled.
4666 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4668 * The check for a GC trigger is only performed when the current
4669 * region is full, so in most cases it's not needed. */
4671 static inline lispobj *
4672 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4673 struct thread *thread)
4675 #ifndef LISP_FEATURE_WIN32
4676 lispobj alloc_signal;
4679 void *new_free_pointer;
4681 gc_assert(nbytes>0);
4683 /* Check for alignment allocation problems. */
4684 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4685 && ((nbytes & LOWTAG_MASK) == 0));
4687 /* Must be inside a PA section. */
4688 gc_assert(get_pseudo_atomic_atomic(thread));
4690 /* maybe we can do this quickly ... */
4691 new_free_pointer = region->free_pointer + nbytes;
4692 if (new_free_pointer <= region->end_addr) {
4693 new_obj = (void*)(region->free_pointer);
4694 region->free_pointer = new_free_pointer;
4695 return(new_obj); /* yup */
4698 /* we have to go the long way around, it seems. Check whether we
4699 * should GC in the near future
4701 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4702 /* Don't flood the system with interrupts if the need to gc is
4703 * already noted. This can happen for example when SUB-GC
4704 * allocates or after a gc triggered in a WITHOUT-GCING. */
4705 if (SymbolValue(GC_PENDING,thread) == NIL) {
4706 /* set things up so that GC happens when we finish the PA
4708 SetSymbolValue(GC_PENDING,T,thread);
4709 if (SymbolValue(GC_INHIBIT,thread) == NIL)
4710 set_pseudo_atomic_interrupted(thread);
4713 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4715 #ifndef LISP_FEATURE_WIN32
4716 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4717 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4718 if ((signed long) alloc_signal <= 0) {
4719 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4720 #ifdef LISP_FEATURE_SB_THREAD
4721 kill_thread_safely(thread->os_thread, SIGPROF);
4726 SetSymbolValue(ALLOC_SIGNAL,
4727 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4737 general_alloc(long nbytes, int page_type_flag)
4739 struct thread *thread = arch_os_get_current_thread();
4740 /* Select correct region, and call general_alloc_internal with it.
4741 * For other then boxed allocation we must lock first, since the
4742 * region is shared. */
4743 if (BOXED_PAGE_FLAG & page_type_flag) {
4744 #ifdef LISP_FEATURE_SB_THREAD
4745 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4747 struct alloc_region *region = &boxed_region;
4749 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4750 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4752 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4753 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4754 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4757 lose("bad page type flag: %d", page_type_flag);
4764 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4768 * shared support for the OS-dependent signal handlers which
4769 * catch GENCGC-related write-protect violations
4771 void unhandled_sigmemoryfault(void* addr);
4773 /* Depending on which OS we're running under, different signals might
4774 * be raised for a violation of write protection in the heap. This
4775 * function factors out the common generational GC magic which needs
4776 * to invoked in this case, and should be called from whatever signal
4777 * handler is appropriate for the OS we're running under.
4779 * Return true if this signal is a normal generational GC thing that
4780 * we were able to handle, or false if it was abnormal and control
4781 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4784 gencgc_handle_wp_violation(void* fault_addr)
4786 page_index_t page_index = find_page_index(fault_addr);
4788 #ifdef QSHOW_SIGNALS
4789 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4790 fault_addr, page_index));
4793 /* Check whether the fault is within the dynamic space. */
4794 if (page_index == (-1)) {
4796 /* It can be helpful to be able to put a breakpoint on this
4797 * case to help diagnose low-level problems. */
4798 unhandled_sigmemoryfault(fault_addr);
4800 /* not within the dynamic space -- not our responsibility */
4804 if (page_table[page_index].write_protected) {
4805 /* Unprotect the page. */
4806 os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
4807 page_table[page_index].write_protected_cleared = 1;
4808 page_table[page_index].write_protected = 0;
4810 /* The only acceptable reason for this signal on a heap
4811 * access is that GENCGC write-protected the page.
4812 * However, if two CPUs hit a wp page near-simultaneously,
4813 * we had better not have the second one lose here if it
4814 * does this test after the first one has already set wp=0
4816 if(page_table[page_index].write_protected_cleared != 1)
4817 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4818 page_index, boxed_region.first_page,
4819 boxed_region.last_page);
4821 /* Don't worry, we can handle it. */
4825 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4826 * it's not just a case of the program hitting the write barrier, and
4827 * are about to let Lisp deal with it. It's basically just a
4828 * convenient place to set a gdb breakpoint. */
4830 unhandled_sigmemoryfault(void *addr)
4833 void gc_alloc_update_all_page_tables(void)
4835 /* Flush the alloc regions updating the tables. */
4838 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4839 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4840 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4844 gc_set_region_empty(struct alloc_region *region)
4846 region->first_page = 0;
4847 region->last_page = -1;
4848 region->start_addr = page_address(0);
4849 region->free_pointer = page_address(0);
4850 region->end_addr = page_address(0);
4854 zero_all_free_pages()
4858 for (i = 0; i < last_free_page; i++) {
4859 if (page_free_p(i)) {
4860 #ifdef READ_PROTECT_FREE_PAGES
4861 os_protect(page_address(i),
4870 /* Things to do before doing a final GC before saving a core (without
4873 * + Pages in large_object pages aren't moved by the GC, so we need to
4874 * unset that flag from all pages.
4875 * + The pseudo-static generation isn't normally collected, but it seems
4876 * reasonable to collect it at least when saving a core. So move the
4877 * pages to a normal generation.
4880 prepare_for_final_gc ()
4883 for (i = 0; i < last_free_page; i++) {
4884 page_table[i].large_object = 0;
4885 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4886 int used = page_table[i].bytes_used;
4887 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4888 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4889 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4895 /* Do a non-conservative GC, and then save a core with the initial
4896 * function being set to the value of the static symbol
4897 * SB!VM:RESTART-LISP-FUNCTION */
4899 gc_and_save(char *filename, boolean prepend_runtime,
4900 boolean save_runtime_options)
4903 void *runtime_bytes = NULL;
4904 size_t runtime_size;
4906 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4911 conservative_stack = 0;
4913 /* The filename might come from Lisp, and be moved by the now
4914 * non-conservative GC. */
4915 filename = strdup(filename);
4917 /* Collect twice: once into relatively high memory, and then back
4918 * into low memory. This compacts the retained data into the lower
4919 * pages, minimizing the size of the core file.
4921 prepare_for_final_gc();
4922 gencgc_alloc_start_page = last_free_page;
4923 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4925 prepare_for_final_gc();
4926 gencgc_alloc_start_page = -1;
4927 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4929 if (prepend_runtime)
4930 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4932 /* The dumper doesn't know that pages need to be zeroed before use. */
4933 zero_all_free_pages();
4934 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4935 prepend_runtime, save_runtime_options);
4936 /* Oops. Save still managed to fail. Since we've mangled the stack
4937 * beyond hope, there's not much we can do.
4938 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4939 * going to be rather unsatisfactory too... */
4940 lose("Attempt to save core after non-conservative GC failed.\n");