2 * GENerational Conservative Garbage Collector for SBCL
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
17 * For a review of garbage collection techniques (e.g. generational
18 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
19 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
20 * had been accepted for _ACM Computing Surveys_ and was available
21 * as a PostScript preprint through
22 * <http://www.cs.utexas.edu/users/oops/papers.html>
24 * <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
37 #include "interrupt.h"
42 #include "gc-internal.h"
44 #include "pseudo-atomic.h"
46 #include "genesis/vector.h"
47 #include "genesis/weak-pointer.h"
48 #include "genesis/fdefn.h"
49 #include "genesis/simple-fun.h"
51 #include "genesis/hash-table.h"
52 #include "genesis/instance.h"
53 #include "genesis/layout.h"
55 #if defined(LUTEX_WIDETAG)
56 #include "pthread-lutex.h"
59 /* forward declarations */
60 page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
68 /* Generations 0-5 are normal collected generations, 6 is only used as
69 * scratch space by the collector, and should never get collected.
72 HIGHEST_NORMAL_GENERATION = 5,
73 PSEUDO_STATIC_GENERATION,
78 /* Should we use page protection to help avoid the scavenging of pages
79 * that don't have pointers to younger generations? */
80 boolean enable_page_protection = 1;
82 /* the minimum size (in bytes) for a large object*/
83 long large_object_size = 4 * PAGE_BYTES;
90 /* the verbosity level. All non-error messages are disabled at level 0;
91 * and only a few rare messages are printed at level 1. */
93 boolean gencgc_verbose = 1;
95 boolean gencgc_verbose = 0;
98 /* FIXME: At some point enable the various error-checking things below
99 * and see what they say. */
101 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
102 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
104 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
106 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
107 boolean pre_verify_gen_0 = 0;
109 /* Should we check for bad pointers after gc_free_heap is called
110 * from Lisp PURIFY? */
111 boolean verify_after_free_heap = 0;
113 /* Should we print a note when code objects are found in the dynamic space
114 * during a heap verify? */
115 boolean verify_dynamic_code_check = 0;
117 /* Should we check code objects for fixup errors after they are transported? */
118 boolean check_code_fixups = 0;
120 /* Should we check that newly allocated regions are zero filled? */
121 boolean gencgc_zero_check = 0;
123 /* Should we check that the free space is zero filled? */
124 boolean gencgc_enable_verify_zero_fill = 0;
126 /* Should we check that free pages are zero filled during gc_free_heap
127 * called after Lisp PURIFY? */
128 boolean gencgc_zero_check_during_free_heap = 0;
130 /* When loading a core, don't do a full scan of the memory for the
131 * memory region boundaries. (Set to true by coreparse.c if the core
132 * contained a pagetable entry).
134 boolean gencgc_partial_pickup = 0;
136 /* If defined, free pages are read-protected to ensure that nothing
140 /* #define READ_PROTECT_FREE_PAGES */
144 * GC structures and variables
147 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
148 unsigned long bytes_allocated = 0;
149 unsigned long auto_gc_trigger = 0;
151 /* the source and destination generations. These are set before a GC starts
153 generation_index_t from_space;
154 generation_index_t new_space;
156 /* Set to 1 when in GC */
157 boolean gc_active_p = 0;
159 /* should the GC be conservative on stack. If false (only right before
160 * saving a core), don't scan the stack / mark pages dont_move. */
161 static boolean conservative_stack = 1;
163 /* An array of page structures is allocated on gc initialization.
164 * This helps quickly map between an address its page structure.
165 * page_table_pages is set from the size of the dynamic space. */
166 page_index_t page_table_pages;
167 struct page *page_table;
169 static inline boolean page_allocated_p(page_index_t page) {
170 return (page_table[page].allocated != FREE_PAGE_FLAG);
173 static inline boolean page_no_region_p(page_index_t page) {
174 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
177 static inline boolean page_allocated_no_region_p(page_index_t page) {
178 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
179 && page_no_region_p(page));
182 static inline boolean page_free_p(page_index_t page) {
183 return (page_table[page].allocated == FREE_PAGE_FLAG);
186 static inline boolean page_boxed_p(page_index_t page) {
187 return (page_table[page].allocated & BOXED_PAGE_FLAG);
190 static inline boolean code_page_p(page_index_t page) {
191 return (page_table[page].allocated & CODE_PAGE_FLAG);
194 static inline boolean page_boxed_no_region_p(page_index_t page) {
195 return page_boxed_p(page) && page_no_region_p(page);
198 static inline boolean page_unboxed_p(page_index_t page) {
199 /* Both flags set == boxed code page */
200 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
201 && !page_boxed_p(page));
204 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
205 return (page_boxed_no_region_p(page)
206 && (page_table[page].bytes_used != 0)
207 && !page_table[page].dont_move
208 && (page_table[page].gen == generation));
211 /* To map addresses to page structures the address of the first page
213 static void *heap_base = NULL;
215 /* Calculate the start address for the given page number. */
217 page_address(page_index_t page_num)
219 return (heap_base + (page_num * PAGE_BYTES));
222 /* Calculate the address where the allocation region associated with
223 * the page starts. */
225 page_region_start(page_index_t page_index)
227 return page_address(page_index)-page_table[page_index].region_start_offset;
230 /* Find the page index within the page_table for the given
231 * address. Return -1 on failure. */
233 find_page_index(void *addr)
235 if (addr >= heap_base) {
236 page_index_t index = ((pointer_sized_uint_t)addr -
237 (pointer_sized_uint_t)heap_base) / PAGE_BYTES;
238 if (index < page_table_pages)
245 npage_bytes(long npages)
247 gc_assert(npages>=0);
248 return ((unsigned long)npages)*PAGE_BYTES;
251 /* Check that X is a higher address than Y and return offset from Y to
254 size_t void_diff(void *x, void *y)
257 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
260 /* a structure to hold the state of a generation */
263 /* the first page that gc_alloc() checks on its next call */
264 page_index_t alloc_start_page;
266 /* the first page that gc_alloc_unboxed() checks on its next call */
267 page_index_t alloc_unboxed_start_page;
269 /* the first page that gc_alloc_large (boxed) considers on its next
270 * call. (Although it always allocates after the boxed_region.) */
271 page_index_t alloc_large_start_page;
273 /* the first page that gc_alloc_large (unboxed) considers on its
274 * next call. (Although it always allocates after the
275 * current_unboxed_region.) */
276 page_index_t alloc_large_unboxed_start_page;
278 /* the bytes allocated to this generation */
279 unsigned long bytes_allocated;
281 /* the number of bytes at which to trigger a GC */
282 unsigned long gc_trigger;
284 /* to calculate a new level for gc_trigger */
285 unsigned long bytes_consed_between_gc;
287 /* the number of GCs since the last raise */
290 /* the average age after which a GC will raise objects to the
294 /* the cumulative sum of the bytes allocated to this generation. It is
295 * cleared after a GC on this generations, and update before new
296 * objects are added from a GC of a younger generation. Dividing by
297 * the bytes_allocated will give the average age of the memory in
298 * this generation since its last GC. */
299 unsigned long cum_sum_bytes_allocated;
301 /* a minimum average memory age before a GC will occur helps
302 * prevent a GC when a large number of new live objects have been
303 * added, in which case a GC could be a waste of time */
304 double min_av_mem_age;
306 /* A linked list of lutex structures in this generation, used for
307 * implementing lutex finalization. */
309 struct lutex *lutexes;
315 /* an array of generation structures. There needs to be one more
316 * generation structure than actual generations as the oldest
317 * generation is temporarily raised then lowered. */
318 struct generation generations[NUM_GENERATIONS];
320 /* the oldest generation that is will currently be GCed by default.
321 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
323 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
325 * Setting this to 0 effectively disables the generational nature of
326 * the GC. In some applications generational GC may not be useful
327 * because there are no long-lived objects.
329 * An intermediate value could be handy after moving long-lived data
330 * into an older generation so an unnecessary GC of this long-lived
331 * data can be avoided. */
332 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
334 /* The maximum free page in the heap is maintained and used to update
335 * ALLOCATION_POINTER which is used by the room function to limit its
336 * search of the heap. XX Gencgc obviously needs to be better
337 * integrated with the Lisp code. */
338 page_index_t last_free_page;
340 #ifdef LISP_FEATURE_SB_THREAD
341 /* This lock is to prevent multiple threads from simultaneously
342 * allocating new regions which overlap each other. Note that the
343 * majority of GC is single-threaded, but alloc() may be called from
344 * >1 thread at a time and must be thread-safe. This lock must be
345 * seized before all accesses to generations[] or to parts of
346 * page_table[] that other threads may want to see */
347 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
348 /* This lock is used to protect non-thread-local allocation. */
349 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
354 * miscellaneous heap functions
357 /* Count the number of pages which are write-protected within the
358 * given generation. */
360 count_write_protect_generation_pages(generation_index_t generation)
363 unsigned long count = 0;
365 for (i = 0; i < last_free_page; i++)
366 if (page_allocated_p(i)
367 && (page_table[i].gen == generation)
368 && (page_table[i].write_protected == 1))
373 /* Count the number of pages within the given generation. */
375 count_generation_pages(generation_index_t generation)
380 for (i = 0; i < last_free_page; i++)
381 if (page_allocated_p(i)
382 && (page_table[i].gen == generation))
389 count_dont_move_pages(void)
393 for (i = 0; i < last_free_page; i++) {
394 if (page_allocated_p(i)
395 && (page_table[i].dont_move != 0)) {
403 /* Work through the pages and add up the number of bytes used for the
404 * given generation. */
406 count_generation_bytes_allocated (generation_index_t gen)
409 unsigned long result = 0;
410 for (i = 0; i < last_free_page; i++) {
411 if (page_allocated_p(i)
412 && (page_table[i].gen == gen))
413 result += page_table[i].bytes_used;
418 /* Return the average age of the memory in a generation. */
420 gen_av_mem_age(generation_index_t gen)
422 if (generations[gen].bytes_allocated == 0)
426 ((double)generations[gen].cum_sum_bytes_allocated)
427 / ((double)generations[gen].bytes_allocated);
430 /* The verbose argument controls how much to print: 0 for normal
431 * level of detail; 1 for debugging. */
433 print_generation_stats() /* FIXME: should take FILE argument, or construct a string */
435 generation_index_t i;
437 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
438 #define FPU_STATE_SIZE 27
439 int fpu_state[FPU_STATE_SIZE];
440 #elif defined(LISP_FEATURE_PPC)
441 #define FPU_STATE_SIZE 32
442 long long fpu_state[FPU_STATE_SIZE];
445 /* This code uses the FP instructions which may be set up for Lisp
446 * so they need to be saved and reset for C. */
449 /* Print the heap stats. */
451 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
453 for (i = 0; i < SCRATCH_GENERATION; i++) {
456 long unboxed_cnt = 0;
457 long large_boxed_cnt = 0;
458 long large_unboxed_cnt = 0;
461 for (j = 0; j < last_free_page; j++)
462 if (page_table[j].gen == i) {
464 /* Count the number of boxed pages within the given
466 if (page_boxed_p(j)) {
467 if (page_table[j].large_object)
472 if(page_table[j].dont_move) pinned_cnt++;
473 /* Count the number of unboxed pages within the given
475 if (page_unboxed_p(j)) {
476 if (page_table[j].large_object)
483 gc_assert(generations[i].bytes_allocated
484 == count_generation_bytes_allocated(i));
486 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
488 generations[i].alloc_start_page,
489 generations[i].alloc_unboxed_start_page,
490 generations[i].alloc_large_start_page,
491 generations[i].alloc_large_unboxed_start_page,
497 generations[i].bytes_allocated,
498 (npage_bytes(count_generation_pages(i))
499 - generations[i].bytes_allocated),
500 generations[i].gc_trigger,
501 count_write_protect_generation_pages(i),
502 generations[i].num_gc,
505 fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated);
506 fprintf(stderr," Dynamic-space-size bytes = %u\n", dynamic_space_size);
508 fpu_restore(fpu_state);
512 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
513 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
516 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
517 * if zeroing it ourselves, i.e. in practice give the memory back to the
518 * OS. Generally done after a large GC.
520 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
522 void *addr = page_address(start), *new_addr;
523 size_t length = npage_bytes(1+end-start);
528 os_invalidate(addr, length);
529 new_addr = os_validate(addr, length);
530 if (new_addr == NULL || new_addr != addr) {
531 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
535 for (i = start; i <= end; i++) {
536 page_table[i].need_to_zero = 0;
540 /* Zero the pages from START to END (inclusive). Generally done just after
541 * a new region has been allocated.
544 zero_pages(page_index_t start, page_index_t end) {
548 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
549 fast_bzero(page_address(start), npage_bytes(1+end-start));
551 bzero(page_address(start), npage_bytes(1+end-start));
556 /* Zero the pages from START to END (inclusive), except for those
557 * pages that are known to already zeroed. Mark all pages in the
558 * ranges as non-zeroed.
561 zero_dirty_pages(page_index_t start, page_index_t end) {
564 for (i = start; i <= end; i++) {
565 if (page_table[i].need_to_zero == 1) {
566 zero_pages(start, end);
571 for (i = start; i <= end; i++) {
572 page_table[i].need_to_zero = 1;
578 * To support quick and inline allocation, regions of memory can be
579 * allocated and then allocated from with just a free pointer and a
580 * check against an end address.
582 * Since objects can be allocated to spaces with different properties
583 * e.g. boxed/unboxed, generation, ages; there may need to be many
584 * allocation regions.
586 * Each allocation region may start within a partly used page. Many
587 * features of memory use are noted on a page wise basis, e.g. the
588 * generation; so if a region starts within an existing allocated page
589 * it must be consistent with this page.
591 * During the scavenging of the newspace, objects will be transported
592 * into an allocation region, and pointers updated to point to this
593 * allocation region. It is possible that these pointers will be
594 * scavenged again before the allocation region is closed, e.g. due to
595 * trans_list which jumps all over the place to cleanup the list. It
596 * is important to be able to determine properties of all objects
597 * pointed to when scavenging, e.g to detect pointers to the oldspace.
598 * Thus it's important that the allocation regions have the correct
599 * properties set when allocated, and not just set when closed. The
600 * region allocation routines return regions with the specified
601 * properties, and grab all the pages, setting their properties
602 * appropriately, except that the amount used is not known.
604 * These regions are used to support quicker allocation using just a
605 * free pointer. The actual space used by the region is not reflected
606 * in the pages tables until it is closed. It can't be scavenged until
609 * When finished with the region it should be closed, which will
610 * update the page tables for the actual space used returning unused
611 * space. Further it may be noted in the new regions which is
612 * necessary when scavenging the newspace.
614 * Large objects may be allocated directly without an allocation
615 * region, the page tables are updated immediately.
617 * Unboxed objects don't contain pointers to other objects and so
618 * don't need scavenging. Further they can't contain pointers to
619 * younger generations so WP is not needed. By allocating pages to
620 * unboxed objects the whole page never needs scavenging or
621 * write-protecting. */
623 /* We are only using two regions at present. Both are for the current
624 * newspace generation. */
625 struct alloc_region boxed_region;
626 struct alloc_region unboxed_region;
628 /* The generation currently being allocated to. */
629 static generation_index_t gc_alloc_generation;
631 static inline page_index_t
632 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
635 if (UNBOXED_PAGE_FLAG == page_type_flag) {
636 return generations[generation].alloc_large_unboxed_start_page;
637 } else if (BOXED_PAGE_FLAG & page_type_flag) {
638 /* Both code and data. */
639 return generations[generation].alloc_large_start_page;
641 lose("bad page type flag: %d", page_type_flag);
644 if (UNBOXED_PAGE_FLAG == page_type_flag) {
645 return generations[generation].alloc_unboxed_start_page;
646 } else if (BOXED_PAGE_FLAG & page_type_flag) {
647 /* Both code and data. */
648 return generations[generation].alloc_start_page;
650 lose("bad page_type_flag: %d", page_type_flag);
656 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
660 if (UNBOXED_PAGE_FLAG == page_type_flag) {
661 generations[generation].alloc_large_unboxed_start_page = page;
662 } else if (BOXED_PAGE_FLAG & page_type_flag) {
663 /* Both code and data. */
664 generations[generation].alloc_large_start_page = page;
666 lose("bad page type flag: %d", page_type_flag);
669 if (UNBOXED_PAGE_FLAG == page_type_flag) {
670 generations[generation].alloc_unboxed_start_page = page;
671 } else if (BOXED_PAGE_FLAG & page_type_flag) {
672 /* Both code and data. */
673 generations[generation].alloc_start_page = page;
675 lose("bad page type flag: %d", page_type_flag);
680 /* Find a new region with room for at least the given number of bytes.
682 * It starts looking at the current generation's alloc_start_page. So
683 * may pick up from the previous region if there is enough space. This
684 * keeps the allocation contiguous when scavenging the newspace.
686 * The alloc_region should have been closed by a call to
687 * gc_alloc_update_page_tables(), and will thus be in an empty state.
689 * To assist the scavenging functions write-protected pages are not
690 * used. Free pages should not be write-protected.
692 * It is critical to the conservative GC that the start of regions be
693 * known. To help achieve this only small regions are allocated at a
696 * During scavenging, pointers may be found to within the current
697 * region and the page generation must be set so that pointers to the
698 * from space can be recognized. Therefore the generation of pages in
699 * the region are set to gc_alloc_generation. To prevent another
700 * allocation call using the same pages, all the pages in the region
701 * are allocated, although they will initially be empty.
704 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
706 page_index_t first_page;
707 page_index_t last_page;
708 unsigned long bytes_found;
714 "/alloc_new_region for %d bytes from gen %d\n",
715 nbytes, gc_alloc_generation));
718 /* Check that the region is in a reset state. */
719 gc_assert((alloc_region->first_page == 0)
720 && (alloc_region->last_page == -1)
721 && (alloc_region->free_pointer == alloc_region->end_addr));
722 ret = thread_mutex_lock(&free_pages_lock);
724 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
725 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
726 bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
727 + npage_bytes(last_page-first_page);
729 /* Set up the alloc_region. */
730 alloc_region->first_page = first_page;
731 alloc_region->last_page = last_page;
732 alloc_region->start_addr = page_table[first_page].bytes_used
733 + page_address(first_page);
734 alloc_region->free_pointer = alloc_region->start_addr;
735 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
737 /* Set up the pages. */
739 /* The first page may have already been in use. */
740 if (page_table[first_page].bytes_used == 0) {
741 page_table[first_page].allocated = page_type_flag;
742 page_table[first_page].gen = gc_alloc_generation;
743 page_table[first_page].large_object = 0;
744 page_table[first_page].region_start_offset = 0;
747 gc_assert(page_table[first_page].allocated == page_type_flag);
748 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
750 gc_assert(page_table[first_page].gen == gc_alloc_generation);
751 gc_assert(page_table[first_page].large_object == 0);
753 for (i = first_page+1; i <= last_page; i++) {
754 page_table[i].allocated = page_type_flag;
755 page_table[i].gen = gc_alloc_generation;
756 page_table[i].large_object = 0;
757 /* This may not be necessary for unboxed regions (think it was
759 page_table[i].region_start_offset =
760 void_diff(page_address(i),alloc_region->start_addr);
761 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
763 /* Bump up last_free_page. */
764 if (last_page+1 > last_free_page) {
765 last_free_page = last_page+1;
766 /* do we only want to call this on special occasions? like for
768 set_alloc_pointer((lispobj)page_address(last_free_page));
770 ret = thread_mutex_unlock(&free_pages_lock);
773 #ifdef READ_PROTECT_FREE_PAGES
774 os_protect(page_address(first_page),
775 npage_bytes(1+last_page-first_page),
779 /* If the first page was only partial, don't check whether it's
780 * zeroed (it won't be) and don't zero it (since the parts that
781 * we're interested in are guaranteed to be zeroed).
783 if (page_table[first_page].bytes_used) {
787 zero_dirty_pages(first_page, last_page);
789 /* we can do this after releasing free_pages_lock */
790 if (gencgc_zero_check) {
792 for (p = (long *)alloc_region->start_addr;
793 p < (long *)alloc_region->end_addr; p++) {
795 /* KLUDGE: It would be nice to use %lx and explicit casts
796 * (long) in code like this, so that it is less likely to
797 * break randomly when running on a machine with different
798 * word sizes. -- WHN 19991129 */
799 lose("The new region at %x is not zero (start=%p, end=%p).\n",
800 p, alloc_region->start_addr, alloc_region->end_addr);
806 /* If the record_new_objects flag is 2 then all new regions created
809 * If it's 1 then then it is only recorded if the first page of the
810 * current region is <= new_areas_ignore_page. This helps avoid
811 * unnecessary recording when doing full scavenge pass.
813 * The new_object structure holds the page, byte offset, and size of
814 * new regions of objects. Each new area is placed in the array of
815 * these structures pointer to by new_areas. new_areas_index holds the
816 * offset into new_areas.
818 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
819 * later code must detect this and handle it, probably by doing a full
820 * scavenge of a generation. */
821 #define NUM_NEW_AREAS 512
822 static int record_new_objects = 0;
823 static page_index_t new_areas_ignore_page;
829 static struct new_area (*new_areas)[];
830 static long new_areas_index;
833 /* Add a new area to new_areas. */
835 add_new_area(page_index_t first_page, size_t offset, size_t size)
837 unsigned long new_area_start,c;
840 /* Ignore if full. */
841 if (new_areas_index >= NUM_NEW_AREAS)
844 switch (record_new_objects) {
848 if (first_page > new_areas_ignore_page)
857 new_area_start = npage_bytes(first_page) + offset;
859 /* Search backwards for a prior area that this follows from. If
860 found this will save adding a new area. */
861 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
862 unsigned long area_end =
863 npage_bytes((*new_areas)[i].page)
864 + (*new_areas)[i].offset
865 + (*new_areas)[i].size;
867 "/add_new_area S1 %d %d %d %d\n",
868 i, c, new_area_start, area_end));*/
869 if (new_area_start == area_end) {
871 "/adding to [%d] %d %d %d with %d %d %d:\n",
873 (*new_areas)[i].page,
874 (*new_areas)[i].offset,
875 (*new_areas)[i].size,
879 (*new_areas)[i].size += size;
884 (*new_areas)[new_areas_index].page = first_page;
885 (*new_areas)[new_areas_index].offset = offset;
886 (*new_areas)[new_areas_index].size = size;
888 "/new_area %d page %d offset %d size %d\n",
889 new_areas_index, first_page, offset, size));*/
892 /* Note the max new_areas used. */
893 if (new_areas_index > max_new_areas)
894 max_new_areas = new_areas_index;
897 /* Update the tables for the alloc_region. The region may be added to
900 * When done the alloc_region is set up so that the next quick alloc
901 * will fail safely and thus a new region will be allocated. Further
902 * it is safe to try to re-update the page table of this reset
905 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
908 page_index_t first_page;
909 page_index_t next_page;
910 unsigned long bytes_used;
911 unsigned long orig_first_page_bytes_used;
912 unsigned long region_size;
913 unsigned long byte_cnt;
917 first_page = alloc_region->first_page;
919 /* Catch an unused alloc_region. */
920 if ((first_page == 0) && (alloc_region->last_page == -1))
923 next_page = first_page+1;
925 ret = thread_mutex_lock(&free_pages_lock);
927 if (alloc_region->free_pointer != alloc_region->start_addr) {
928 /* some bytes were allocated in the region */
929 orig_first_page_bytes_used = page_table[first_page].bytes_used;
931 gc_assert(alloc_region->start_addr ==
932 (page_address(first_page)
933 + page_table[first_page].bytes_used));
935 /* All the pages used need to be updated */
937 /* Update the first page. */
939 /* If the page was free then set up the gen, and
940 * region_start_offset. */
941 if (page_table[first_page].bytes_used == 0)
942 gc_assert(page_table[first_page].region_start_offset == 0);
943 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
945 gc_assert(page_table[first_page].allocated & page_type_flag);
946 gc_assert(page_table[first_page].gen == gc_alloc_generation);
947 gc_assert(page_table[first_page].large_object == 0);
951 /* Calculate the number of bytes used in this page. This is not
952 * always the number of new bytes, unless it was free. */
954 if ((bytes_used = void_diff(alloc_region->free_pointer,
955 page_address(first_page)))
957 bytes_used = PAGE_BYTES;
960 page_table[first_page].bytes_used = bytes_used;
961 byte_cnt += bytes_used;
964 /* All the rest of the pages should be free. We need to set
965 * their region_start_offset pointer to the start of the
966 * region, and set the bytes_used. */
968 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
969 gc_assert(page_table[next_page].allocated & page_type_flag);
970 gc_assert(page_table[next_page].bytes_used == 0);
971 gc_assert(page_table[next_page].gen == gc_alloc_generation);
972 gc_assert(page_table[next_page].large_object == 0);
974 gc_assert(page_table[next_page].region_start_offset ==
975 void_diff(page_address(next_page),
976 alloc_region->start_addr));
978 /* Calculate the number of bytes used in this page. */
980 if ((bytes_used = void_diff(alloc_region->free_pointer,
981 page_address(next_page)))>PAGE_BYTES) {
982 bytes_used = PAGE_BYTES;
985 page_table[next_page].bytes_used = bytes_used;
986 byte_cnt += bytes_used;
991 region_size = void_diff(alloc_region->free_pointer,
992 alloc_region->start_addr);
993 bytes_allocated += region_size;
994 generations[gc_alloc_generation].bytes_allocated += region_size;
996 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
998 /* Set the generations alloc restart page to the last page of
1000 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1002 /* Add the region to the new_areas if requested. */
1003 if (BOXED_PAGE_FLAG & page_type_flag)
1004 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1008 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1010 gc_alloc_generation));
1013 /* There are no bytes allocated. Unallocate the first_page if
1014 * there are 0 bytes_used. */
1015 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1016 if (page_table[first_page].bytes_used == 0)
1017 page_table[first_page].allocated = FREE_PAGE_FLAG;
1020 /* Unallocate any unused pages. */
1021 while (next_page <= alloc_region->last_page) {
1022 gc_assert(page_table[next_page].bytes_used == 0);
1023 page_table[next_page].allocated = FREE_PAGE_FLAG;
1026 ret = thread_mutex_unlock(&free_pages_lock);
1027 gc_assert(ret == 0);
1029 /* alloc_region is per-thread, we're ok to do this unlocked */
1030 gc_set_region_empty(alloc_region);
1033 static inline void *gc_quick_alloc(long nbytes);
1035 /* Allocate a possibly large object. */
1037 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1039 page_index_t first_page;
1040 page_index_t last_page;
1041 int orig_first_page_bytes_used;
1044 unsigned long bytes_used;
1045 page_index_t next_page;
1048 ret = thread_mutex_lock(&free_pages_lock);
1049 gc_assert(ret == 0);
1051 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1052 if (first_page <= alloc_region->last_page) {
1053 first_page = alloc_region->last_page+1;
1056 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1058 gc_assert(first_page > alloc_region->last_page);
1060 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1062 /* Set up the pages. */
1063 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1065 /* If the first page was free then set up the gen, and
1066 * region_start_offset. */
1067 if (page_table[first_page].bytes_used == 0) {
1068 page_table[first_page].allocated = page_type_flag;
1069 page_table[first_page].gen = gc_alloc_generation;
1070 page_table[first_page].region_start_offset = 0;
1071 page_table[first_page].large_object = 1;
1074 gc_assert(page_table[first_page].allocated == page_type_flag);
1075 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1076 gc_assert(page_table[first_page].large_object == 1);
1080 /* Calc. the number of bytes used in this page. This is not
1081 * always the number of new bytes, unless it was free. */
1083 if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
1084 bytes_used = PAGE_BYTES;
1087 page_table[first_page].bytes_used = bytes_used;
1088 byte_cnt += bytes_used;
1090 next_page = first_page+1;
1092 /* All the rest of the pages should be free. We need to set their
1093 * region_start_offset pointer to the start of the region, and set
1094 * the bytes_used. */
1096 gc_assert(page_free_p(next_page));
1097 gc_assert(page_table[next_page].bytes_used == 0);
1098 page_table[next_page].allocated = page_type_flag;
1099 page_table[next_page].gen = gc_alloc_generation;
1100 page_table[next_page].large_object = 1;
1102 page_table[next_page].region_start_offset =
1103 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1105 /* Calculate the number of bytes used in this page. */
1107 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1108 if (bytes_used > PAGE_BYTES) {
1109 bytes_used = PAGE_BYTES;
1112 page_table[next_page].bytes_used = bytes_used;
1113 page_table[next_page].write_protected=0;
1114 page_table[next_page].dont_move=0;
1115 byte_cnt += bytes_used;
1119 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1121 bytes_allocated += nbytes;
1122 generations[gc_alloc_generation].bytes_allocated += nbytes;
1124 /* Add the region to the new_areas if requested. */
1125 if (BOXED_PAGE_FLAG & page_type_flag)
1126 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1128 /* Bump up last_free_page */
1129 if (last_page+1 > last_free_page) {
1130 last_free_page = last_page+1;
1131 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1133 ret = thread_mutex_unlock(&free_pages_lock);
1134 gc_assert(ret == 0);
1136 #ifdef READ_PROTECT_FREE_PAGES
1137 os_protect(page_address(first_page),
1138 npage_bytes(1+last_page-first_page),
1142 zero_dirty_pages(first_page, last_page);
1144 return page_address(first_page);
1147 static page_index_t gencgc_alloc_start_page = -1;
1150 gc_heap_exhausted_error_or_lose (long available, long requested)
1152 struct thread *thread = arch_os_get_current_thread();
1153 /* Write basic information before doing anything else: if we don't
1154 * call to lisp this is a must, and even if we do there is always
1155 * the danger that we bounce back here before the error has been
1156 * handled, or indeed even printed.
1158 fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
1159 gc_active_p ? "garbage collection" : "allocation",
1160 available, requested);
1161 if (gc_active_p || (available == 0)) {
1162 /* If we are in GC, or totally out of memory there is no way
1163 * to sanely transfer control to the lisp-side of things.
1165 print_generation_stats();
1166 fprintf(stderr, "GC control variables:\n");
1167 fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
1168 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
1169 (SymbolValue(GC_PENDING, thread) == T) ?
1170 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
1171 "false" : "in progress"));
1172 #ifdef LISP_FEATURE_SB_THREAD
1173 fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n",
1174 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
1176 lose("Heap exhausted, game over.");
1179 /* FIXME: assert free_pages_lock held */
1180 (void)thread_mutex_unlock(&free_pages_lock);
1181 gc_assert(get_pseudo_atomic_atomic(thread));
1182 clear_pseudo_atomic_atomic(thread);
1183 if (get_pseudo_atomic_interrupted(thread))
1184 do_pending_interrupt();
1185 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1186 * to running user code at arbitrary places, even in a
1187 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1188 * running out of the heap. So at this point all bets are
1190 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1191 corruption_warning_and_maybe_lose
1192 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1193 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1194 alloc_number(available), alloc_number(requested));
1195 lose("HEAP-EXHAUSTED-ERROR fell through");
1200 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes,
1203 page_index_t first_page, last_page;
1204 page_index_t restart_page = *restart_page_ptr;
1205 long bytes_found = 0;
1206 long most_bytes_found = 0;
1207 /* FIXME: assert(free_pages_lock is held); */
1209 /* Toggled by gc_and_save for heap compaction, normally -1. */
1210 if (gencgc_alloc_start_page != -1) {
1211 restart_page = gencgc_alloc_start_page;
1214 gc_assert(nbytes>=0);
1215 if (((unsigned long)nbytes)>=PAGE_BYTES) {
1216 /* Search for a contiguous free space of at least nbytes,
1217 * aligned on a page boundary. The page-alignment is strictly
1218 * speaking needed only for objects at least large_object_size
1221 first_page = restart_page;
1222 while ((first_page < page_table_pages) &&
1223 page_allocated_p(first_page))
1226 last_page = first_page;
1227 bytes_found = PAGE_BYTES;
1228 while ((bytes_found < nbytes) &&
1229 (last_page < (page_table_pages-1)) &&
1230 page_free_p(last_page+1)) {
1232 bytes_found += PAGE_BYTES;
1233 gc_assert(0 == page_table[last_page].bytes_used);
1234 gc_assert(0 == page_table[last_page].write_protected);
1236 if (bytes_found > most_bytes_found)
1237 most_bytes_found = bytes_found;
1238 restart_page = last_page + 1;
1239 } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
1242 /* Search for a page with at least nbytes of space. We prefer
1243 * not to split small objects on multiple pages, to reduce the
1244 * number of contiguous allocation regions spaning multiple
1245 * pages: this helps avoid excessive conservativism. */
1246 first_page = restart_page;
1247 while (first_page < page_table_pages) {
1248 if (page_free_p(first_page))
1250 gc_assert(0 == page_table[first_page].bytes_used);
1251 bytes_found = PAGE_BYTES;
1254 else if ((page_table[first_page].allocated == page_type_flag) &&
1255 (page_table[first_page].large_object == 0) &&
1256 (page_table[first_page].gen == gc_alloc_generation) &&
1257 (page_table[first_page].write_protected == 0) &&
1258 (page_table[first_page].dont_move == 0))
1260 bytes_found = PAGE_BYTES
1261 - page_table[first_page].bytes_used;
1262 if (bytes_found > most_bytes_found)
1263 most_bytes_found = bytes_found;
1264 if (bytes_found >= nbytes)
1269 last_page = first_page;
1270 restart_page = first_page + 1;
1273 /* Check for a failure */
1274 if (bytes_found < nbytes) {
1275 gc_assert(restart_page >= page_table_pages);
1276 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1279 gc_assert(page_table[first_page].write_protected == 0);
1281 *restart_page_ptr = first_page;
1285 /* Allocate bytes. All the rest of the special-purpose allocation
1286 * functions will eventually call this */
1289 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1292 void *new_free_pointer;
1294 if (nbytes>=large_object_size)
1295 return gc_alloc_large(nbytes, page_type_flag, my_region);
1297 /* Check whether there is room in the current alloc region. */
1298 new_free_pointer = my_region->free_pointer + nbytes;
1300 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1301 my_region->free_pointer, new_free_pointer); */
1303 if (new_free_pointer <= my_region->end_addr) {
1304 /* If so then allocate from the current alloc region. */
1305 void *new_obj = my_region->free_pointer;
1306 my_region->free_pointer = new_free_pointer;
1308 /* Unless a `quick' alloc was requested, check whether the
1309 alloc region is almost empty. */
1311 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1312 /* If so, finished with the current region. */
1313 gc_alloc_update_page_tables(page_type_flag, my_region);
1314 /* Set up a new region. */
1315 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1318 return((void *)new_obj);
1321 /* Else not enough free space in the current region: retry with a
1324 gc_alloc_update_page_tables(page_type_flag, my_region);
1325 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1326 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1329 /* these are only used during GC: all allocation from the mutator calls
1330 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1333 static inline void *
1334 gc_quick_alloc(long nbytes)
1336 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1339 static inline void *
1340 gc_quick_alloc_large(long nbytes)
1342 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1345 static inline void *
1346 gc_alloc_unboxed(long nbytes)
1348 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1351 static inline void *
1352 gc_quick_alloc_unboxed(long nbytes)
1354 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1357 static inline void *
1358 gc_quick_alloc_large_unboxed(long nbytes)
1360 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1364 /* Copy a large boxed object. If the object is in a large object
1365 * region then it is simply promoted, else it is copied. If it's large
1366 * enough then it's copied to a large object region.
1368 * Vectors may have shrunk. If the object is not copied the space
1369 * needs to be reclaimed, and the page_tables corrected. */
1371 copy_large_object(lispobj object, long nwords)
1375 page_index_t first_page;
1377 gc_assert(is_lisp_pointer(object));
1378 gc_assert(from_space_p(object));
1379 gc_assert((nwords & 0x01) == 0);
1382 /* Check whether it's in a large object region. */
1383 first_page = find_page_index((void *)object);
1384 gc_assert(first_page >= 0);
1386 if (page_table[first_page].large_object) {
1388 /* Promote the object. */
1390 unsigned long remaining_bytes;
1391 page_index_t next_page;
1392 unsigned long bytes_freed;
1393 unsigned long old_bytes_used;
1395 /* Note: Any page write-protection must be removed, else a
1396 * later scavenge_newspace may incorrectly not scavenge these
1397 * pages. This would not be necessary if they are added to the
1398 * new areas, but let's do it for them all (they'll probably
1399 * be written anyway?). */
1401 gc_assert(page_table[first_page].region_start_offset == 0);
1403 next_page = first_page;
1404 remaining_bytes = nwords*N_WORD_BYTES;
1405 while (remaining_bytes > PAGE_BYTES) {
1406 gc_assert(page_table[next_page].gen == from_space);
1407 gc_assert(page_boxed_p(next_page));
1408 gc_assert(page_table[next_page].large_object);
1409 gc_assert(page_table[next_page].region_start_offset ==
1410 npage_bytes(next_page-first_page));
1411 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1413 page_table[next_page].gen = new_space;
1415 /* Remove any write-protection. We should be able to rely
1416 * on the write-protect flag to avoid redundant calls. */
1417 if (page_table[next_page].write_protected) {
1418 os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL);
1419 page_table[next_page].write_protected = 0;
1421 remaining_bytes -= PAGE_BYTES;
1425 /* Now only one page remains, but the object may have shrunk
1426 * so there may be more unused pages which will be freed. */
1428 /* The object may have shrunk but shouldn't have grown. */
1429 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1431 page_table[next_page].gen = new_space;
1432 gc_assert(page_boxed_p(next_page));
1434 /* Adjust the bytes_used. */
1435 old_bytes_used = page_table[next_page].bytes_used;
1436 page_table[next_page].bytes_used = remaining_bytes;
1438 bytes_freed = old_bytes_used - remaining_bytes;
1440 /* Free any remaining pages; needs care. */
1442 while ((old_bytes_used == PAGE_BYTES) &&
1443 (page_table[next_page].gen == from_space) &&
1444 page_boxed_p(next_page) &&
1445 page_table[next_page].large_object &&
1446 (page_table[next_page].region_start_offset ==
1447 npage_bytes(next_page - first_page))) {
1448 /* Checks out OK, free the page. Don't need to bother zeroing
1449 * pages as this should have been done before shrinking the
1450 * object. These pages shouldn't be write-protected as they
1451 * should be zero filled. */
1452 gc_assert(page_table[next_page].write_protected == 0);
1454 old_bytes_used = page_table[next_page].bytes_used;
1455 page_table[next_page].allocated = FREE_PAGE_FLAG;
1456 page_table[next_page].bytes_used = 0;
1457 bytes_freed += old_bytes_used;
1461 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1463 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1464 bytes_allocated -= bytes_freed;
1466 /* Add the region to the new_areas if requested. */
1467 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1471 /* Get tag of object. */
1472 tag = lowtag_of(object);
1474 /* Allocate space. */
1475 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1477 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1479 /* Return Lisp pointer of new object. */
1480 return ((lispobj) new) | tag;
1484 /* to copy unboxed objects */
1486 copy_unboxed_object(lispobj object, long nwords)
1491 gc_assert(is_lisp_pointer(object));
1492 gc_assert(from_space_p(object));
1493 gc_assert((nwords & 0x01) == 0);
1495 /* Get tag of object. */
1496 tag = lowtag_of(object);
1498 /* Allocate space. */
1499 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1501 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1503 /* Return Lisp pointer of new object. */
1504 return ((lispobj) new) | tag;
1507 /* to copy large unboxed objects
1509 * If the object is in a large object region then it is simply
1510 * promoted, else it is copied. If it's large enough then it's copied
1511 * to a large object region.
1513 * Bignums and vectors may have shrunk. If the object is not copied
1514 * the space needs to be reclaimed, and the page_tables corrected.
1516 * KLUDGE: There's a lot of cut-and-paste duplication between this
1517 * function and copy_large_object(..). -- WHN 20000619 */
1519 copy_large_unboxed_object(lispobj object, long nwords)
1523 page_index_t first_page;
1525 gc_assert(is_lisp_pointer(object));
1526 gc_assert(from_space_p(object));
1527 gc_assert((nwords & 0x01) == 0);
1529 if ((nwords > 1024*1024) && gencgc_verbose) {
1530 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1531 nwords*N_WORD_BYTES));
1534 /* Check whether it's a large object. */
1535 first_page = find_page_index((void *)object);
1536 gc_assert(first_page >= 0);
1538 if (page_table[first_page].large_object) {
1539 /* Promote the object. Note: Unboxed objects may have been
1540 * allocated to a BOXED region so it may be necessary to
1541 * change the region to UNBOXED. */
1542 unsigned long remaining_bytes;
1543 page_index_t next_page;
1544 unsigned long bytes_freed;
1545 unsigned long old_bytes_used;
1547 gc_assert(page_table[first_page].region_start_offset == 0);
1549 next_page = first_page;
1550 remaining_bytes = nwords*N_WORD_BYTES;
1551 while (remaining_bytes > PAGE_BYTES) {
1552 gc_assert(page_table[next_page].gen == from_space);
1553 gc_assert(page_allocated_no_region_p(next_page));
1554 gc_assert(page_table[next_page].large_object);
1555 gc_assert(page_table[next_page].region_start_offset ==
1556 npage_bytes(next_page-first_page));
1557 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1559 page_table[next_page].gen = new_space;
1560 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1561 remaining_bytes -= PAGE_BYTES;
1565 /* Now only one page remains, but the object may have shrunk so
1566 * there may be more unused pages which will be freed. */
1568 /* Object may have shrunk but shouldn't have grown - check. */
1569 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1571 page_table[next_page].gen = new_space;
1572 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1574 /* Adjust the bytes_used. */
1575 old_bytes_used = page_table[next_page].bytes_used;
1576 page_table[next_page].bytes_used = remaining_bytes;
1578 bytes_freed = old_bytes_used - remaining_bytes;
1580 /* Free any remaining pages; needs care. */
1582 while ((old_bytes_used == PAGE_BYTES) &&
1583 (page_table[next_page].gen == from_space) &&
1584 page_allocated_no_region_p(next_page) &&
1585 page_table[next_page].large_object &&
1586 (page_table[next_page].region_start_offset ==
1587 npage_bytes(next_page - first_page))) {
1588 /* Checks out OK, free the page. Don't need to both zeroing
1589 * pages as this should have been done before shrinking the
1590 * object. These pages shouldn't be write-protected, even if
1591 * boxed they should be zero filled. */
1592 gc_assert(page_table[next_page].write_protected == 0);
1594 old_bytes_used = page_table[next_page].bytes_used;
1595 page_table[next_page].allocated = FREE_PAGE_FLAG;
1596 page_table[next_page].bytes_used = 0;
1597 bytes_freed += old_bytes_used;
1601 if ((bytes_freed > 0) && gencgc_verbose) {
1603 "/copy_large_unboxed bytes_freed=%d\n",
1607 generations[from_space].bytes_allocated -=
1608 nwords*N_WORD_BYTES + bytes_freed;
1609 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1610 bytes_allocated -= bytes_freed;
1615 /* Get tag of object. */
1616 tag = lowtag_of(object);
1618 /* Allocate space. */
1619 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1621 /* Copy the object. */
1622 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1624 /* Return Lisp pointer of new object. */
1625 return ((lispobj) new) | tag;
1634 * code and code-related objects
1637 static lispobj trans_fun_header(lispobj object);
1638 static lispobj trans_boxed(lispobj object);
1641 /* Scan a x86 compiled code object, looking for possible fixups that
1642 * have been missed after a move.
1644 * Two types of fixups are needed:
1645 * 1. Absolute fixups to within the code object.
1646 * 2. Relative fixups to outside the code object.
1648 * Currently only absolute fixups to the constant vector, or to the
1649 * code area are checked. */
1651 sniff_code_object(struct code *code, unsigned long displacement)
1653 #ifdef LISP_FEATURE_X86
1654 long nheader_words, ncode_words, nwords;
1656 void *constants_start_addr = NULL, *constants_end_addr;
1657 void *code_start_addr, *code_end_addr;
1658 int fixup_found = 0;
1660 if (!check_code_fixups)
1663 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1665 ncode_words = fixnum_value(code->code_size);
1666 nheader_words = HeaderValue(*(lispobj *)code);
1667 nwords = ncode_words + nheader_words;
1669 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1670 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1671 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1672 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1674 /* Work through the unboxed code. */
1675 for (p = code_start_addr; p < code_end_addr; p++) {
1676 void *data = *(void **)p;
1677 unsigned d1 = *((unsigned char *)p - 1);
1678 unsigned d2 = *((unsigned char *)p - 2);
1679 unsigned d3 = *((unsigned char *)p - 3);
1680 unsigned d4 = *((unsigned char *)p - 4);
1682 unsigned d5 = *((unsigned char *)p - 5);
1683 unsigned d6 = *((unsigned char *)p - 6);
1686 /* Check for code references. */
1687 /* Check for a 32 bit word that looks like an absolute
1688 reference to within the code adea of the code object. */
1689 if ((data >= (code_start_addr-displacement))
1690 && (data < (code_end_addr-displacement))) {
1691 /* function header */
1693 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1695 /* Skip the function header */
1699 /* the case of PUSH imm32 */
1703 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1704 p, d6, d5, d4, d3, d2, d1, data));
1705 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1707 /* the case of MOV [reg-8],imm32 */
1709 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1710 || d2==0x45 || d2==0x46 || d2==0x47)
1714 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1715 p, d6, d5, d4, d3, d2, d1, data));
1716 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1718 /* the case of LEA reg,[disp32] */
1719 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1722 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1723 p, d6, d5, d4, d3, d2, d1, data));
1724 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1728 /* Check for constant references. */
1729 /* Check for a 32 bit word that looks like an absolute
1730 reference to within the constant vector. Constant references
1732 if ((data >= (constants_start_addr-displacement))
1733 && (data < (constants_end_addr-displacement))
1734 && (((unsigned)data & 0x3) == 0)) {
1739 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1740 p, d6, d5, d4, d3, d2, d1, data));
1741 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1744 /* the case of MOV m32,EAX */
1748 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1749 p, d6, d5, d4, d3, d2, d1, data));
1750 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1753 /* the case of CMP m32,imm32 */
1754 if ((d1 == 0x3d) && (d2 == 0x81)) {
1757 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1758 p, d6, d5, d4, d3, d2, d1, data));
1760 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1763 /* Check for a mod=00, r/m=101 byte. */
1764 if ((d1 & 0xc7) == 5) {
1769 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1770 p, d6, d5, d4, d3, d2, d1, data));
1771 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1773 /* the case of CMP reg32,m32 */
1777 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1778 p, d6, d5, d4, d3, d2, d1, data));
1779 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1781 /* the case of MOV m32,reg32 */
1785 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1786 p, d6, d5, d4, d3, d2, d1, data));
1787 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1789 /* the case of MOV reg32,m32 */
1793 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1794 p, d6, d5, d4, d3, d2, d1, data));
1795 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1797 /* the case of LEA reg32,m32 */
1801 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1802 p, d6, d5, d4, d3, d2, d1, data));
1803 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1809 /* If anything was found, print some information on the code
1813 "/compiled code object at %x: header words = %d, code words = %d\n",
1814 code, nheader_words, ncode_words));
1816 "/const start = %x, end = %x\n",
1817 constants_start_addr, constants_end_addr));
1819 "/code start = %x, end = %x\n",
1820 code_start_addr, code_end_addr));
1826 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1828 /* x86-64 uses pc-relative addressing instead of this kludge */
1829 #ifndef LISP_FEATURE_X86_64
1830 long nheader_words, ncode_words, nwords;
1831 void *constants_start_addr, *constants_end_addr;
1832 void *code_start_addr, *code_end_addr;
1833 lispobj fixups = NIL;
1834 unsigned long displacement =
1835 (unsigned long)new_code - (unsigned long)old_code;
1836 struct vector *fixups_vector;
1838 ncode_words = fixnum_value(new_code->code_size);
1839 nheader_words = HeaderValue(*(lispobj *)new_code);
1840 nwords = ncode_words + nheader_words;
1842 "/compiled code object at %x: header words = %d, code words = %d\n",
1843 new_code, nheader_words, ncode_words)); */
1844 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1845 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1846 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1847 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1850 "/const start = %x, end = %x\n",
1851 constants_start_addr,constants_end_addr));
1853 "/code start = %x; end = %x\n",
1854 code_start_addr,code_end_addr));
1857 /* The first constant should be a pointer to the fixups for this
1858 code objects. Check. */
1859 fixups = new_code->constants[0];
1861 /* It will be 0 or the unbound-marker if there are no fixups (as
1862 * will be the case if the code object has been purified, for
1863 * example) and will be an other pointer if it is valid. */
1864 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1865 !is_lisp_pointer(fixups)) {
1866 /* Check for possible errors. */
1867 if (check_code_fixups)
1868 sniff_code_object(new_code, displacement);
1873 fixups_vector = (struct vector *)native_pointer(fixups);
1875 /* Could be pointing to a forwarding pointer. */
1876 /* FIXME is this always in from_space? if so, could replace this code with
1877 * forwarding_pointer_p/forwarding_pointer_value */
1878 if (is_lisp_pointer(fixups) &&
1879 (find_page_index((void*)fixups_vector) != -1) &&
1880 (fixups_vector->header == 0x01)) {
1881 /* If so, then follow it. */
1882 /*SHOW("following pointer to a forwarding pointer");*/
1884 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1887 /*SHOW("got fixups");*/
1889 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1890 /* Got the fixups for the code block. Now work through the vector,
1891 and apply a fixup at each address. */
1892 long length = fixnum_value(fixups_vector->length);
1894 for (i = 0; i < length; i++) {
1895 unsigned long offset = fixups_vector->data[i];
1896 /* Now check the current value of offset. */
1897 unsigned long old_value =
1898 *(unsigned long *)((unsigned long)code_start_addr + offset);
1900 /* If it's within the old_code object then it must be an
1901 * absolute fixup (relative ones are not saved) */
1902 if ((old_value >= (unsigned long)old_code)
1903 && (old_value < ((unsigned long)old_code
1904 + nwords*N_WORD_BYTES)))
1905 /* So add the dispacement. */
1906 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1907 old_value + displacement;
1909 /* It is outside the old code object so it must be a
1910 * relative fixup (absolute fixups are not saved). So
1911 * subtract the displacement. */
1912 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1913 old_value - displacement;
1916 /* This used to just print a note to stderr, but a bogus fixup seems to
1917 * indicate real heap corruption, so a hard hailure is in order. */
1918 lose("fixup vector %p has a bad widetag: %d\n",
1919 fixups_vector, widetag_of(fixups_vector->header));
1922 /* Check for possible errors. */
1923 if (check_code_fixups) {
1924 sniff_code_object(new_code,displacement);
1931 trans_boxed_large(lispobj object)
1934 unsigned long length;
1936 gc_assert(is_lisp_pointer(object));
1938 header = *((lispobj *) native_pointer(object));
1939 length = HeaderValue(header) + 1;
1940 length = CEILING(length, 2);
1942 return copy_large_object(object, length);
1945 /* Doesn't seem to be used, delete it after the grace period. */
1948 trans_unboxed_large(lispobj object)
1951 unsigned long length;
1953 gc_assert(is_lisp_pointer(object));
1955 header = *((lispobj *) native_pointer(object));
1956 length = HeaderValue(header) + 1;
1957 length = CEILING(length, 2);
1959 return copy_large_unboxed_object(object, length);
1965 * Lutexes. Using the normal finalization machinery for finalizing
1966 * lutexes is tricky, since the finalization depends on working lutexes.
1967 * So we track the lutexes in the GC and finalize them manually.
1970 #if defined(LUTEX_WIDETAG)
1973 * Start tracking LUTEX in the GC, by adding it to the linked list of
1974 * lutexes in the nursery generation. The caller is responsible for
1975 * locking, and GCs must be inhibited until the registration is
1979 gencgc_register_lutex (struct lutex *lutex) {
1980 int index = find_page_index(lutex);
1981 generation_index_t gen;
1984 /* This lutex is in static space, so we don't need to worry about
1990 gen = page_table[index].gen;
1992 gc_assert(gen >= 0);
1993 gc_assert(gen < NUM_GENERATIONS);
1995 head = generations[gen].lutexes;
2002 generations[gen].lutexes = lutex;
2006 * Stop tracking LUTEX in the GC by removing it from the appropriate
2007 * linked lists. This will only be called during GC, so no locking is
2011 gencgc_unregister_lutex (struct lutex *lutex) {
2013 lutex->prev->next = lutex->next;
2015 generations[lutex->gen].lutexes = lutex->next;
2019 lutex->next->prev = lutex->prev;
2028 * Mark all lutexes in generation GEN as not live.
2031 unmark_lutexes (generation_index_t gen) {
2032 struct lutex *lutex = generations[gen].lutexes;
2036 lutex = lutex->next;
2041 * Finalize all lutexes in generation GEN that have not been marked live.
2044 reap_lutexes (generation_index_t gen) {
2045 struct lutex *lutex = generations[gen].lutexes;
2048 struct lutex *next = lutex->next;
2050 lutex_destroy((tagged_lutex_t) lutex);
2051 gencgc_unregister_lutex(lutex);
2058 * Mark LUTEX as live.
2061 mark_lutex (lispobj tagged_lutex) {
2062 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
2068 * Move all lutexes in generation FROM to generation TO.
2071 move_lutexes (generation_index_t from, generation_index_t to) {
2072 struct lutex *tail = generations[from].lutexes;
2074 /* Nothing to move */
2078 /* Change the generation of the lutexes in FROM. */
2079 while (tail->next) {
2085 /* Link the last lutex in the FROM list to the start of the TO list */
2086 tail->next = generations[to].lutexes;
2088 /* And vice versa */
2089 if (generations[to].lutexes) {
2090 generations[to].lutexes->prev = tail;
2093 /* And update the generations structures to match this */
2094 generations[to].lutexes = generations[from].lutexes;
2095 generations[from].lutexes = NULL;
2099 scav_lutex(lispobj *where, lispobj object)
2101 mark_lutex((lispobj) where);
2103 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2107 trans_lutex(lispobj object)
2109 struct lutex *lutex = (struct lutex *) native_pointer(object);
2111 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2112 gc_assert(is_lisp_pointer(object));
2113 copied = copy_object(object, words);
2115 /* Update the links, since the lutex moved in memory. */
2117 lutex->next->prev = (struct lutex *) native_pointer(copied);
2121 lutex->prev->next = (struct lutex *) native_pointer(copied);
2123 generations[lutex->gen].lutexes =
2124 (struct lutex *) native_pointer(copied);
2131 size_lutex(lispobj *where)
2133 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2135 #endif /* LUTEX_WIDETAG */
2142 /* XX This is a hack adapted from cgc.c. These don't work too
2143 * efficiently with the gencgc as a list of the weak pointers is
2144 * maintained within the objects which causes writes to the pages. A
2145 * limited attempt is made to avoid unnecessary writes, but this needs
2147 #define WEAK_POINTER_NWORDS \
2148 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2151 scav_weak_pointer(lispobj *where, lispobj object)
2153 /* Since we overwrite the 'next' field, we have to make
2154 * sure not to do so for pointers already in the list.
2155 * Instead of searching the list of weak_pointers each
2156 * time, we ensure that next is always NULL when the weak
2157 * pointer isn't in the list, and not NULL otherwise.
2158 * Since we can't use NULL to denote end of list, we
2159 * use a pointer back to the same weak_pointer.
2161 struct weak_pointer * wp = (struct weak_pointer*)where;
2163 if (NULL == wp->next) {
2164 wp->next = weak_pointers;
2166 if (NULL == wp->next)
2170 /* Do not let GC scavenge the value slot of the weak pointer.
2171 * (That is why it is a weak pointer.) */
2173 return WEAK_POINTER_NWORDS;
2178 search_read_only_space(void *pointer)
2180 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2181 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2182 if ((pointer < (void *)start) || (pointer >= (void *)end))
2184 return (gc_search_space(start,
2185 (((lispobj *)pointer)+2)-start,
2186 (lispobj *) pointer));
2190 search_static_space(void *pointer)
2192 lispobj *start = (lispobj *)STATIC_SPACE_START;
2193 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2194 if ((pointer < (void *)start) || (pointer >= (void *)end))
2196 return (gc_search_space(start,
2197 (((lispobj *)pointer)+2)-start,
2198 (lispobj *) pointer));
2201 /* a faster version for searching the dynamic space. This will work even
2202 * if the object is in a current allocation region. */
2204 search_dynamic_space(void *pointer)
2206 page_index_t page_index = find_page_index(pointer);
2209 /* The address may be invalid, so do some checks. */
2210 if ((page_index == -1) || page_free_p(page_index))
2212 start = (lispobj *)page_region_start(page_index);
2213 return (gc_search_space(start,
2214 (((lispobj *)pointer)+2)-start,
2215 (lispobj *)pointer));
2218 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2220 /* Helper for valid_lisp_pointer_p and
2221 * possibly_valid_dynamic_space_pointer.
2223 * pointer is the pointer to validate, and start_addr is the address
2224 * of the enclosing object.
2227 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2229 if (!is_lisp_pointer((lispobj)pointer)) {
2233 /* Check that the object pointed to is consistent with the pointer
2235 switch (lowtag_of((lispobj)pointer)) {
2236 case FUN_POINTER_LOWTAG:
2237 /* Start_addr should be the enclosing code object, or a closure
2239 switch (widetag_of(*start_addr)) {
2240 case CODE_HEADER_WIDETAG:
2241 /* This case is probably caught above. */
2243 case CLOSURE_HEADER_WIDETAG:
2244 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2245 if ((unsigned long)pointer !=
2246 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2247 if (gencgc_verbose) {
2250 pointer, start_addr, *start_addr));
2256 if (gencgc_verbose) {
2259 pointer, start_addr, *start_addr));
2264 case LIST_POINTER_LOWTAG:
2265 if ((unsigned long)pointer !=
2266 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2267 if (gencgc_verbose) {
2270 pointer, start_addr, *start_addr));
2274 /* Is it plausible cons? */
2275 if ((is_lisp_pointer(start_addr[0]) ||
2276 is_lisp_immediate(start_addr[0])) &&
2277 (is_lisp_pointer(start_addr[1]) ||
2278 is_lisp_immediate(start_addr[1])))
2281 if (gencgc_verbose) {
2284 pointer, start_addr, *start_addr));
2288 case INSTANCE_POINTER_LOWTAG:
2289 if ((unsigned long)pointer !=
2290 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2291 if (gencgc_verbose) {
2294 pointer, start_addr, *start_addr));
2298 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2299 if (gencgc_verbose) {
2302 pointer, start_addr, *start_addr));
2307 case OTHER_POINTER_LOWTAG:
2308 if ((unsigned long)pointer !=
2309 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2310 if (gencgc_verbose) {
2313 pointer, start_addr, *start_addr));
2317 /* Is it plausible? Not a cons. XXX should check the headers. */
2318 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2319 if (gencgc_verbose) {
2322 pointer, start_addr, *start_addr));
2326 switch (widetag_of(start_addr[0])) {
2327 case UNBOUND_MARKER_WIDETAG:
2328 case NO_TLS_VALUE_MARKER_WIDETAG:
2329 case CHARACTER_WIDETAG:
2330 #if N_WORD_BITS == 64
2331 case SINGLE_FLOAT_WIDETAG:
2333 if (gencgc_verbose) {
2336 pointer, start_addr, *start_addr));
2340 /* only pointed to by function pointers? */
2341 case CLOSURE_HEADER_WIDETAG:
2342 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2343 if (gencgc_verbose) {
2346 pointer, start_addr, *start_addr));
2350 case INSTANCE_HEADER_WIDETAG:
2351 if (gencgc_verbose) {
2354 pointer, start_addr, *start_addr));
2358 /* the valid other immediate pointer objects */
2359 case SIMPLE_VECTOR_WIDETAG:
2361 case COMPLEX_WIDETAG:
2362 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2363 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2365 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2366 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2368 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2369 case COMPLEX_LONG_FLOAT_WIDETAG:
2371 case SIMPLE_ARRAY_WIDETAG:
2372 case COMPLEX_BASE_STRING_WIDETAG:
2373 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2374 case COMPLEX_CHARACTER_STRING_WIDETAG:
2376 case COMPLEX_VECTOR_NIL_WIDETAG:
2377 case COMPLEX_BIT_VECTOR_WIDETAG:
2378 case COMPLEX_VECTOR_WIDETAG:
2379 case COMPLEX_ARRAY_WIDETAG:
2380 case VALUE_CELL_HEADER_WIDETAG:
2381 case SYMBOL_HEADER_WIDETAG:
2383 case CODE_HEADER_WIDETAG:
2384 case BIGNUM_WIDETAG:
2385 #if N_WORD_BITS != 64
2386 case SINGLE_FLOAT_WIDETAG:
2388 case DOUBLE_FLOAT_WIDETAG:
2389 #ifdef LONG_FLOAT_WIDETAG
2390 case LONG_FLOAT_WIDETAG:
2392 case SIMPLE_BASE_STRING_WIDETAG:
2393 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2394 case SIMPLE_CHARACTER_STRING_WIDETAG:
2396 case SIMPLE_BIT_VECTOR_WIDETAG:
2397 case SIMPLE_ARRAY_NIL_WIDETAG:
2398 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2399 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2400 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2401 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2402 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2403 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2404 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2405 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2407 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2408 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2409 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2410 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2412 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2413 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2415 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2416 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2418 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2419 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2421 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2422 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2424 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2425 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2427 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2428 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2430 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2431 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2433 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2434 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2436 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2437 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2438 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2439 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2441 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2442 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2444 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2445 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2447 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2448 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2451 case WEAK_POINTER_WIDETAG:
2452 #ifdef LUTEX_WIDETAG
2458 if (gencgc_verbose) {
2461 pointer, start_addr, *start_addr));
2467 if (gencgc_verbose) {
2470 pointer, start_addr, *start_addr));
2479 /* Used by the debugger to validate possibly bogus pointers before
2480 * calling MAKE-LISP-OBJ on them.
2482 * FIXME: We would like to make this perfect, because if the debugger
2483 * constructs a reference to a bugs lisp object, and it ends up in a
2484 * location scavenged by the GC all hell breaks loose.
2486 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2487 * and return true for all valid pointers, this could actually be eager
2488 * and lie about a few pointers without bad results... but that should
2489 * be reflected in the name.
2492 valid_lisp_pointer_p(lispobj *pointer)
2495 if (((start=search_dynamic_space(pointer))!=NULL) ||
2496 ((start=search_static_space(pointer))!=NULL) ||
2497 ((start=search_read_only_space(pointer))!=NULL))
2498 return looks_like_valid_lisp_pointer_p(pointer, start);
2503 /* Is there any possibility that pointer is a valid Lisp object
2504 * reference, and/or something else (e.g. subroutine call return
2505 * address) which should prevent us from moving the referred-to thing?
2506 * This is called from preserve_pointers() */
2508 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2510 lispobj *start_addr;
2512 /* Find the object start address. */
2513 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2517 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2520 /* Adjust large bignum and vector objects. This will adjust the
2521 * allocated region if the size has shrunk, and move unboxed objects
2522 * into unboxed pages. The pages are not promoted here, and the
2523 * promoted region is not added to the new_regions; this is really
2524 * only designed to be called from preserve_pointer(). Shouldn't fail
2525 * if this is missed, just may delay the moving of objects to unboxed
2526 * pages, and the freeing of pages. */
2528 maybe_adjust_large_object(lispobj *where)
2530 page_index_t first_page;
2531 page_index_t next_page;
2534 unsigned long remaining_bytes;
2535 unsigned long bytes_freed;
2536 unsigned long old_bytes_used;
2540 /* Check whether it's a vector or bignum object. */
2541 switch (widetag_of(where[0])) {
2542 case SIMPLE_VECTOR_WIDETAG:
2543 boxed = BOXED_PAGE_FLAG;
2545 case BIGNUM_WIDETAG:
2546 case SIMPLE_BASE_STRING_WIDETAG:
2547 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2548 case SIMPLE_CHARACTER_STRING_WIDETAG:
2550 case SIMPLE_BIT_VECTOR_WIDETAG:
2551 case SIMPLE_ARRAY_NIL_WIDETAG:
2552 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2553 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2554 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2555 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2556 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2557 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2558 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2559 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2561 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2562 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2563 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2564 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2566 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2567 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2569 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2570 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2572 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2573 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2575 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2576 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2578 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2579 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2581 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2582 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2584 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2585 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2587 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2588 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2590 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2591 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2592 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2593 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2595 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2596 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2598 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2599 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2601 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2602 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2604 boxed = UNBOXED_PAGE_FLAG;
2610 /* Find its current size. */
2611 nwords = (sizetab[widetag_of(where[0])])(where);
2613 first_page = find_page_index((void *)where);
2614 gc_assert(first_page >= 0);
2616 /* Note: Any page write-protection must be removed, else a later
2617 * scavenge_newspace may incorrectly not scavenge these pages.
2618 * This would not be necessary if they are added to the new areas,
2619 * but lets do it for them all (they'll probably be written
2622 gc_assert(page_table[first_page].region_start_offset == 0);
2624 next_page = first_page;
2625 remaining_bytes = nwords*N_WORD_BYTES;
2626 while (remaining_bytes > PAGE_BYTES) {
2627 gc_assert(page_table[next_page].gen == from_space);
2628 gc_assert(page_allocated_no_region_p(next_page));
2629 gc_assert(page_table[next_page].large_object);
2630 gc_assert(page_table[next_page].region_start_offset ==
2631 npage_bytes(next_page-first_page));
2632 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
2634 page_table[next_page].allocated = boxed;
2636 /* Shouldn't be write-protected at this stage. Essential that the
2638 gc_assert(!page_table[next_page].write_protected);
2639 remaining_bytes -= PAGE_BYTES;
2643 /* Now only one page remains, but the object may have shrunk so
2644 * there may be more unused pages which will be freed. */
2646 /* Object may have shrunk but shouldn't have grown - check. */
2647 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2649 page_table[next_page].allocated = boxed;
2650 gc_assert(page_table[next_page].allocated ==
2651 page_table[first_page].allocated);
2653 /* Adjust the bytes_used. */
2654 old_bytes_used = page_table[next_page].bytes_used;
2655 page_table[next_page].bytes_used = remaining_bytes;
2657 bytes_freed = old_bytes_used - remaining_bytes;
2659 /* Free any remaining pages; needs care. */
2661 while ((old_bytes_used == PAGE_BYTES) &&
2662 (page_table[next_page].gen == from_space) &&
2663 page_allocated_no_region_p(next_page) &&
2664 page_table[next_page].large_object &&
2665 (page_table[next_page].region_start_offset ==
2666 npage_bytes(next_page - first_page))) {
2667 /* It checks out OK, free the page. We don't need to both zeroing
2668 * pages as this should have been done before shrinking the
2669 * object. These pages shouldn't be write protected as they
2670 * should be zero filled. */
2671 gc_assert(page_table[next_page].write_protected == 0);
2673 old_bytes_used = page_table[next_page].bytes_used;
2674 page_table[next_page].allocated = FREE_PAGE_FLAG;
2675 page_table[next_page].bytes_used = 0;
2676 bytes_freed += old_bytes_used;
2680 if ((bytes_freed > 0) && gencgc_verbose) {
2682 "/maybe_adjust_large_object() freed %d\n",
2686 generations[from_space].bytes_allocated -= bytes_freed;
2687 bytes_allocated -= bytes_freed;
2692 /* Take a possible pointer to a Lisp object and mark its page in the
2693 * page_table so that it will not be relocated during a GC.
2695 * This involves locating the page it points to, then backing up to
2696 * the start of its region, then marking all pages dont_move from there
2697 * up to the first page that's not full or has a different generation
2699 * It is assumed that all the page static flags have been cleared at
2700 * the start of a GC.
2702 * It is also assumed that the current gc_alloc() region has been
2703 * flushed and the tables updated. */
2706 preserve_pointer(void *addr)
2708 page_index_t addr_page_index = find_page_index(addr);
2709 page_index_t first_page;
2711 unsigned int region_allocation;
2713 /* quick check 1: Address is quite likely to have been invalid. */
2714 if ((addr_page_index == -1)
2715 || page_free_p(addr_page_index)
2716 || (page_table[addr_page_index].bytes_used == 0)
2717 || (page_table[addr_page_index].gen != from_space)
2718 /* Skip if already marked dont_move. */
2719 || (page_table[addr_page_index].dont_move != 0))
2721 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2722 /* (Now that we know that addr_page_index is in range, it's
2723 * safe to index into page_table[] with it.) */
2724 region_allocation = page_table[addr_page_index].allocated;
2726 /* quick check 2: Check the offset within the page.
2729 if (((unsigned long)addr & (PAGE_BYTES - 1)) >
2730 page_table[addr_page_index].bytes_used)
2733 /* Filter out anything which can't be a pointer to a Lisp object
2734 * (or, as a special case which also requires dont_move, a return
2735 * address referring to something in a CodeObject). This is
2736 * expensive but important, since it vastly reduces the
2737 * probability that random garbage will be bogusly interpreted as
2738 * a pointer which prevents a page from moving. */
2739 if (!(code_page_p(addr_page_index)
2740 || (is_lisp_pointer((lispobj)addr) &&
2741 possibly_valid_dynamic_space_pointer(addr))))
2744 /* Find the beginning of the region. Note that there may be
2745 * objects in the region preceding the one that we were passed a
2746 * pointer to: if this is the case, we will write-protect all the
2747 * previous objects' pages too. */
2750 /* I think this'd work just as well, but without the assertions.
2751 * -dan 2004.01.01 */
2752 first_page = find_page_index(page_region_start(addr_page_index))
2754 first_page = addr_page_index;
2755 while (page_table[first_page].region_start_offset != 0) {
2757 /* Do some checks. */
2758 gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
2759 gc_assert(page_table[first_page].gen == from_space);
2760 gc_assert(page_table[first_page].allocated == region_allocation);
2764 /* Adjust any large objects before promotion as they won't be
2765 * copied after promotion. */
2766 if (page_table[first_page].large_object) {
2767 maybe_adjust_large_object(page_address(first_page));
2768 /* If a large object has shrunk then addr may now point to a
2769 * free area in which case it's ignored here. Note it gets
2770 * through the valid pointer test above because the tail looks
2772 if (page_free_p(addr_page_index)
2773 || (page_table[addr_page_index].bytes_used == 0)
2774 /* Check the offset within the page. */
2775 || (((unsigned long)addr & (PAGE_BYTES - 1))
2776 > page_table[addr_page_index].bytes_used)) {
2778 "weird? ignore ptr 0x%x to freed area of large object\n",
2782 /* It may have moved to unboxed pages. */
2783 region_allocation = page_table[first_page].allocated;
2786 /* Now work forward until the end of this contiguous area is found,
2787 * marking all pages as dont_move. */
2788 for (i = first_page; ;i++) {
2789 gc_assert(page_table[i].allocated == region_allocation);
2791 /* Mark the page static. */
2792 page_table[i].dont_move = 1;
2794 /* Move the page to the new_space. XX I'd rather not do this
2795 * but the GC logic is not quite able to copy with the static
2796 * pages remaining in the from space. This also requires the
2797 * generation bytes_allocated counters be updated. */
2798 page_table[i].gen = new_space;
2799 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2800 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2802 /* It is essential that the pages are not write protected as
2803 * they may have pointers into the old-space which need
2804 * scavenging. They shouldn't be write protected at this
2806 gc_assert(!page_table[i].write_protected);
2808 /* Check whether this is the last page in this contiguous block.. */
2809 if ((page_table[i].bytes_used < PAGE_BYTES)
2810 /* ..or it is PAGE_BYTES and is the last in the block */
2812 || (page_table[i+1].bytes_used == 0) /* next page free */
2813 || (page_table[i+1].gen != from_space) /* diff. gen */
2814 || (page_table[i+1].region_start_offset == 0))
2818 /* Check that the page is now static. */
2819 gc_assert(page_table[addr_page_index].dont_move != 0);
2822 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2825 /* If the given page is not write-protected, then scan it for pointers
2826 * to younger generations or the top temp. generation, if no
2827 * suspicious pointers are found then the page is write-protected.
2829 * Care is taken to check for pointers to the current gc_alloc()
2830 * region if it is a younger generation or the temp. generation. This
2831 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2832 * the gc_alloc_generation does not need to be checked as this is only
2833 * called from scavenge_generation() when the gc_alloc generation is
2834 * younger, so it just checks if there is a pointer to the current
2837 * We return 1 if the page was write-protected, else 0. */
2839 update_page_write_prot(page_index_t page)
2841 generation_index_t gen = page_table[page].gen;
2844 void **page_addr = (void **)page_address(page);
2845 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2847 /* Shouldn't be a free page. */
2848 gc_assert(page_allocated_p(page));
2849 gc_assert(page_table[page].bytes_used != 0);
2851 /* Skip if it's already write-protected, pinned, or unboxed */
2852 if (page_table[page].write_protected
2853 /* FIXME: What's the reason for not write-protecting pinned pages? */
2854 || page_table[page].dont_move
2855 || page_unboxed_p(page))
2858 /* Scan the page for pointers to younger generations or the
2859 * top temp. generation. */
2861 for (j = 0; j < num_words; j++) {
2862 void *ptr = *(page_addr+j);
2863 page_index_t index = find_page_index(ptr);
2865 /* Check that it's in the dynamic space */
2867 if (/* Does it point to a younger or the temp. generation? */
2868 (page_allocated_p(index)
2869 && (page_table[index].bytes_used != 0)
2870 && ((page_table[index].gen < gen)
2871 || (page_table[index].gen == SCRATCH_GENERATION)))
2873 /* Or does it point within a current gc_alloc() region? */
2874 || ((boxed_region.start_addr <= ptr)
2875 && (ptr <= boxed_region.free_pointer))
2876 || ((unboxed_region.start_addr <= ptr)
2877 && (ptr <= unboxed_region.free_pointer))) {
2884 /* Write-protect the page. */
2885 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2887 os_protect((void *)page_addr,
2889 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2891 /* Note the page as protected in the page tables. */
2892 page_table[page].write_protected = 1;
2898 /* Scavenge all generations from FROM to TO, inclusive, except for
2899 * new_space which needs special handling, as new objects may be
2900 * added which are not checked here - use scavenge_newspace generation.
2902 * Write-protected pages should not have any pointers to the
2903 * from_space so do need scavenging; thus write-protected pages are
2904 * not always scavenged. There is some code to check that these pages
2905 * are not written; but to check fully the write-protected pages need
2906 * to be scavenged by disabling the code to skip them.
2908 * Under the current scheme when a generation is GCed the younger
2909 * generations will be empty. So, when a generation is being GCed it
2910 * is only necessary to scavenge the older generations for pointers
2911 * not the younger. So a page that does not have pointers to younger
2912 * generations does not need to be scavenged.
2914 * The write-protection can be used to note pages that don't have
2915 * pointers to younger pages. But pages can be written without having
2916 * pointers to younger generations. After the pages are scavenged here
2917 * they can be scanned for pointers to younger generations and if
2918 * there are none the page can be write-protected.
2920 * One complication is when the newspace is the top temp. generation.
2922 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2923 * that none were written, which they shouldn't be as they should have
2924 * no pointers to younger generations. This breaks down for weak
2925 * pointers as the objects contain a link to the next and are written
2926 * if a weak pointer is scavenged. Still it's a useful check. */
2928 scavenge_generations(generation_index_t from, generation_index_t to)
2935 /* Clear the write_protected_cleared flags on all pages. */
2936 for (i = 0; i < page_table_pages; i++)
2937 page_table[i].write_protected_cleared = 0;
2940 for (i = 0; i < last_free_page; i++) {
2941 generation_index_t generation = page_table[i].gen;
2943 && (page_table[i].bytes_used != 0)
2944 && (generation != new_space)
2945 && (generation >= from)
2946 && (generation <= to)) {
2947 page_index_t last_page,j;
2948 int write_protected=1;
2950 /* This should be the start of a region */
2951 gc_assert(page_table[i].region_start_offset == 0);
2953 /* Now work forward until the end of the region */
2954 for (last_page = i; ; last_page++) {
2956 write_protected && page_table[last_page].write_protected;
2957 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2958 /* Or it is PAGE_BYTES and is the last in the block */
2959 || (!page_boxed_p(last_page+1))
2960 || (page_table[last_page+1].bytes_used == 0)
2961 || (page_table[last_page+1].gen != generation)
2962 || (page_table[last_page+1].region_start_offset == 0))
2965 if (!write_protected) {
2966 scavenge(page_address(i),
2967 ((unsigned long)(page_table[last_page].bytes_used
2968 + npage_bytes(last_page-i)))
2971 /* Now scan the pages and write protect those that
2972 * don't have pointers to younger generations. */
2973 if (enable_page_protection) {
2974 for (j = i; j <= last_page; j++) {
2975 num_wp += update_page_write_prot(j);
2978 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2980 "/write protected %d pages within generation %d\n",
2981 num_wp, generation));
2989 /* Check that none of the write_protected pages in this generation
2990 * have been written to. */
2991 for (i = 0; i < page_table_pages; i++) {
2992 if (page_allocated_p(i)
2993 && (page_table[i].bytes_used != 0)
2994 && (page_table[i].gen == generation)
2995 && (page_table[i].write_protected_cleared != 0)) {
2996 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2998 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
2999 page_table[i].bytes_used,
3000 page_table[i].region_start_offset,
3001 page_table[i].dont_move));
3002 lose("write to protected page %d in scavenge_generation()\n", i);
3009 /* Scavenge a newspace generation. As it is scavenged new objects may
3010 * be allocated to it; these will also need to be scavenged. This
3011 * repeats until there are no more objects unscavenged in the
3012 * newspace generation.
3014 * To help improve the efficiency, areas written are recorded by
3015 * gc_alloc() and only these scavenged. Sometimes a little more will be
3016 * scavenged, but this causes no harm. An easy check is done that the
3017 * scavenged bytes equals the number allocated in the previous
3020 * Write-protected pages are not scanned except if they are marked
3021 * dont_move in which case they may have been promoted and still have
3022 * pointers to the from space.
3024 * Write-protected pages could potentially be written by alloc however
3025 * to avoid having to handle re-scavenging of write-protected pages
3026 * gc_alloc() does not write to write-protected pages.
3028 * New areas of objects allocated are recorded alternatively in the two
3029 * new_areas arrays below. */
3030 static struct new_area new_areas_1[NUM_NEW_AREAS];
3031 static struct new_area new_areas_2[NUM_NEW_AREAS];
3033 /* Do one full scan of the new space generation. This is not enough to
3034 * complete the job as new objects may be added to the generation in
3035 * the process which are not scavenged. */
3037 scavenge_newspace_generation_one_scan(generation_index_t generation)
3042 "/starting one full scan of newspace generation %d\n",
3044 for (i = 0; i < last_free_page; i++) {
3045 /* Note that this skips over open regions when it encounters them. */
3047 && (page_table[i].bytes_used != 0)
3048 && (page_table[i].gen == generation)
3049 && ((page_table[i].write_protected == 0)
3050 /* (This may be redundant as write_protected is now
3051 * cleared before promotion.) */
3052 || (page_table[i].dont_move == 1))) {
3053 page_index_t last_page;
3056 /* The scavenge will start at the region_start_offset of
3059 * We need to find the full extent of this contiguous
3060 * block in case objects span pages.
3062 * Now work forward until the end of this contiguous area
3063 * is found. A small area is preferred as there is a
3064 * better chance of its pages being write-protected. */
3065 for (last_page = i; ;last_page++) {
3066 /* If all pages are write-protected and movable,
3067 * then no need to scavenge */
3068 all_wp=all_wp && page_table[last_page].write_protected &&
3069 !page_table[last_page].dont_move;
3071 /* Check whether this is the last page in this
3072 * contiguous block */
3073 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3074 /* Or it is PAGE_BYTES and is the last in the block */
3075 || (!page_boxed_p(last_page+1))
3076 || (page_table[last_page+1].bytes_used == 0)
3077 || (page_table[last_page+1].gen != generation)
3078 || (page_table[last_page+1].region_start_offset == 0))
3082 /* Do a limited check for write-protected pages. */
3084 long nwords = (((unsigned long)
3085 (page_table[last_page].bytes_used
3086 + npage_bytes(last_page-i)
3087 + page_table[i].region_start_offset))
3089 new_areas_ignore_page = last_page;
3091 scavenge(page_region_start(i), nwords);
3098 "/done with one full scan of newspace generation %d\n",
3102 /* Do a complete scavenge of the newspace generation. */
3104 scavenge_newspace_generation(generation_index_t generation)
3108 /* the new_areas array currently being written to by gc_alloc() */
3109 struct new_area (*current_new_areas)[] = &new_areas_1;
3110 long current_new_areas_index;
3112 /* the new_areas created by the previous scavenge cycle */
3113 struct new_area (*previous_new_areas)[] = NULL;
3114 long previous_new_areas_index;
3116 /* Flush the current regions updating the tables. */
3117 gc_alloc_update_all_page_tables();
3119 /* Turn on the recording of new areas by gc_alloc(). */
3120 new_areas = current_new_areas;
3121 new_areas_index = 0;
3123 /* Don't need to record new areas that get scavenged anyway during
3124 * scavenge_newspace_generation_one_scan. */
3125 record_new_objects = 1;
3127 /* Start with a full scavenge. */
3128 scavenge_newspace_generation_one_scan(generation);
3130 /* Record all new areas now. */
3131 record_new_objects = 2;
3133 /* Give a chance to weak hash tables to make other objects live.
3134 * FIXME: The algorithm implemented here for weak hash table gcing
3135 * is O(W^2+N) as Bruno Haible warns in
3136 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3137 * see "Implementation 2". */
3138 scav_weak_hash_tables();
3140 /* Flush the current regions updating the tables. */
3141 gc_alloc_update_all_page_tables();
3143 /* Grab new_areas_index. */
3144 current_new_areas_index = new_areas_index;
3147 "The first scan is finished; current_new_areas_index=%d.\n",
3148 current_new_areas_index));*/
3150 while (current_new_areas_index > 0) {
3151 /* Move the current to the previous new areas */
3152 previous_new_areas = current_new_areas;
3153 previous_new_areas_index = current_new_areas_index;
3155 /* Scavenge all the areas in previous new areas. Any new areas
3156 * allocated are saved in current_new_areas. */
3158 /* Allocate an array for current_new_areas; alternating between
3159 * new_areas_1 and 2 */
3160 if (previous_new_areas == &new_areas_1)
3161 current_new_areas = &new_areas_2;
3163 current_new_areas = &new_areas_1;
3165 /* Set up for gc_alloc(). */
3166 new_areas = current_new_areas;
3167 new_areas_index = 0;
3169 /* Check whether previous_new_areas had overflowed. */
3170 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3172 /* New areas of objects allocated have been lost so need to do a
3173 * full scan to be sure! If this becomes a problem try
3174 * increasing NUM_NEW_AREAS. */
3175 if (gencgc_verbose) {
3176 SHOW("new_areas overflow, doing full scavenge");
3179 /* Don't need to record new areas that get scavenged
3180 * anyway during scavenge_newspace_generation_one_scan. */
3181 record_new_objects = 1;
3183 scavenge_newspace_generation_one_scan(generation);
3185 /* Record all new areas now. */
3186 record_new_objects = 2;
3188 scav_weak_hash_tables();
3190 /* Flush the current regions updating the tables. */
3191 gc_alloc_update_all_page_tables();
3195 /* Work through previous_new_areas. */
3196 for (i = 0; i < previous_new_areas_index; i++) {
3197 page_index_t page = (*previous_new_areas)[i].page;
3198 size_t offset = (*previous_new_areas)[i].offset;
3199 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3200 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3201 scavenge(page_address(page)+offset, size);
3204 scav_weak_hash_tables();
3206 /* Flush the current regions updating the tables. */
3207 gc_alloc_update_all_page_tables();
3210 current_new_areas_index = new_areas_index;
3213 "The re-scan has finished; current_new_areas_index=%d.\n",
3214 current_new_areas_index));*/
3217 /* Turn off recording of areas allocated by gc_alloc(). */
3218 record_new_objects = 0;
3221 /* Check that none of the write_protected pages in this generation
3222 * have been written to. */
3223 for (i = 0; i < page_table_pages; i++) {
3224 if (page_allocated_p(i)
3225 && (page_table[i].bytes_used != 0)
3226 && (page_table[i].gen == generation)
3227 && (page_table[i].write_protected_cleared != 0)
3228 && (page_table[i].dont_move == 0)) {
3229 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3230 i, generation, page_table[i].dont_move);
3236 /* Un-write-protect all the pages in from_space. This is done at the
3237 * start of a GC else there may be many page faults while scavenging
3238 * the newspace (I've seen drive the system time to 99%). These pages
3239 * would need to be unprotected anyway before unmapping in
3240 * free_oldspace; not sure what effect this has on paging.. */
3242 unprotect_oldspace(void)
3246 for (i = 0; i < last_free_page; i++) {
3247 if (page_allocated_p(i)
3248 && (page_table[i].bytes_used != 0)
3249 && (page_table[i].gen == from_space)) {
3252 page_start = (void *)page_address(i);
3254 /* Remove any write-protection. We should be able to rely
3255 * on the write-protect flag to avoid redundant calls. */
3256 if (page_table[i].write_protected) {
3257 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3258 page_table[i].write_protected = 0;
3264 /* Work through all the pages and free any in from_space. This
3265 * assumes that all objects have been copied or promoted to an older
3266 * generation. Bytes_allocated and the generation bytes_allocated
3267 * counter are updated. The number of bytes freed is returned. */
3268 static unsigned long
3271 unsigned long bytes_freed = 0;
3272 page_index_t first_page, last_page;
3277 /* Find a first page for the next region of pages. */
3278 while ((first_page < last_free_page)
3279 && (page_free_p(first_page)
3280 || (page_table[first_page].bytes_used == 0)
3281 || (page_table[first_page].gen != from_space)))
3284 if (first_page >= last_free_page)
3287 /* Find the last page of this region. */
3288 last_page = first_page;
3291 /* Free the page. */
3292 bytes_freed += page_table[last_page].bytes_used;
3293 generations[page_table[last_page].gen].bytes_allocated -=
3294 page_table[last_page].bytes_used;
3295 page_table[last_page].allocated = FREE_PAGE_FLAG;
3296 page_table[last_page].bytes_used = 0;
3298 /* Remove any write-protection. We should be able to rely
3299 * on the write-protect flag to avoid redundant calls. */
3301 void *page_start = (void *)page_address(last_page);
3303 if (page_table[last_page].write_protected) {
3304 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3305 page_table[last_page].write_protected = 0;
3310 while ((last_page < last_free_page)
3311 && page_allocated_p(last_page)
3312 && (page_table[last_page].bytes_used != 0)
3313 && (page_table[last_page].gen == from_space));
3315 #ifdef READ_PROTECT_FREE_PAGES
3316 os_protect(page_address(first_page),
3317 npage_bytes(last_page-first_page),
3320 first_page = last_page;
3321 } while (first_page < last_free_page);
3323 bytes_allocated -= bytes_freed;
3328 /* Print some information about a pointer at the given address. */
3330 print_ptr(lispobj *addr)
3332 /* If addr is in the dynamic space then out the page information. */
3333 page_index_t pi1 = find_page_index((void*)addr);
3336 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
3337 (unsigned long) addr,
3339 page_table[pi1].allocated,
3340 page_table[pi1].gen,
3341 page_table[pi1].bytes_used,
3342 page_table[pi1].region_start_offset,
3343 page_table[pi1].dont_move);
3344 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3358 verify_space(lispobj *start, size_t words)
3360 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3361 int is_in_readonly_space =
3362 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3363 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3367 lispobj thing = *(lispobj*)start;
3369 if (is_lisp_pointer(thing)) {
3370 page_index_t page_index = find_page_index((void*)thing);
3371 long to_readonly_space =
3372 (READ_ONLY_SPACE_START <= thing &&
3373 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3374 long to_static_space =
3375 (STATIC_SPACE_START <= thing &&
3376 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3378 /* Does it point to the dynamic space? */
3379 if (page_index != -1) {
3380 /* If it's within the dynamic space it should point to a used
3381 * page. XX Could check the offset too. */
3382 if (page_allocated_p(page_index)
3383 && (page_table[page_index].bytes_used == 0))
3384 lose ("Ptr %x @ %x sees free page.\n", thing, start);
3385 /* Check that it doesn't point to a forwarding pointer! */
3386 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3387 lose("Ptr %x @ %x sees forwarding ptr.\n", thing, start);
3389 /* Check that its not in the RO space as it would then be a
3390 * pointer from the RO to the dynamic space. */
3391 if (is_in_readonly_space) {
3392 lose("ptr to dynamic space %x from RO space %x\n",
3395 /* Does it point to a plausible object? This check slows
3396 * it down a lot (so it's commented out).
3398 * "a lot" is serious: it ate 50 minutes cpu time on
3399 * my duron 950 before I came back from lunch and
3402 * FIXME: Add a variable to enable this
3405 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3406 lose("ptr %x to invalid object %x\n", thing, start);
3410 /* Verify that it points to another valid space. */
3411 if (!to_readonly_space && !to_static_space) {
3412 lose("Ptr %x @ %x sees junk.\n", thing, start);
3416 if (!(fixnump(thing))) {
3418 switch(widetag_of(*start)) {
3421 case SIMPLE_VECTOR_WIDETAG:
3423 case COMPLEX_WIDETAG:
3424 case SIMPLE_ARRAY_WIDETAG:
3425 case COMPLEX_BASE_STRING_WIDETAG:
3426 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3427 case COMPLEX_CHARACTER_STRING_WIDETAG:
3429 case COMPLEX_VECTOR_NIL_WIDETAG:
3430 case COMPLEX_BIT_VECTOR_WIDETAG:
3431 case COMPLEX_VECTOR_WIDETAG:
3432 case COMPLEX_ARRAY_WIDETAG:
3433 case CLOSURE_HEADER_WIDETAG:
3434 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3435 case VALUE_CELL_HEADER_WIDETAG:
3436 case SYMBOL_HEADER_WIDETAG:
3437 case CHARACTER_WIDETAG:
3438 #if N_WORD_BITS == 64
3439 case SINGLE_FLOAT_WIDETAG:
3441 case UNBOUND_MARKER_WIDETAG:
3446 case INSTANCE_HEADER_WIDETAG:
3449 long ntotal = HeaderValue(thing);
3450 lispobj layout = ((struct instance *)start)->slots[0];
3455 nuntagged = ((struct layout *)
3456 native_pointer(layout))->n_untagged_slots;
3457 verify_space(start + 1,
3458 ntotal - fixnum_value(nuntagged));
3462 case CODE_HEADER_WIDETAG:
3464 lispobj object = *start;
3466 long nheader_words, ncode_words, nwords;
3468 struct simple_fun *fheaderp;
3470 code = (struct code *) start;
3472 /* Check that it's not in the dynamic space.
3473 * FIXME: Isn't is supposed to be OK for code
3474 * objects to be in the dynamic space these days? */
3475 if (is_in_dynamic_space
3476 /* It's ok if it's byte compiled code. The trace
3477 * table offset will be a fixnum if it's x86
3478 * compiled code - check.
3480 * FIXME: #^#@@! lack of abstraction here..
3481 * This line can probably go away now that
3482 * there's no byte compiler, but I've got
3483 * too much to worry about right now to try
3484 * to make sure. -- WHN 2001-10-06 */
3485 && fixnump(code->trace_table_offset)
3486 /* Only when enabled */
3487 && verify_dynamic_code_check) {
3489 "/code object at %x in the dynamic space\n",
3493 ncode_words = fixnum_value(code->code_size);
3494 nheader_words = HeaderValue(object);
3495 nwords = ncode_words + nheader_words;
3496 nwords = CEILING(nwords, 2);
3497 /* Scavenge the boxed section of the code data block */
3498 verify_space(start + 1, nheader_words - 1);
3500 /* Scavenge the boxed section of each function
3501 * object in the code data block. */
3502 fheaderl = code->entry_points;
3503 while (fheaderl != NIL) {
3505 (struct simple_fun *) native_pointer(fheaderl);
3506 gc_assert(widetag_of(fheaderp->header) ==
3507 SIMPLE_FUN_HEADER_WIDETAG);
3508 verify_space(&fheaderp->name, 1);
3509 verify_space(&fheaderp->arglist, 1);
3510 verify_space(&fheaderp->type, 1);
3511 fheaderl = fheaderp->next;
3517 /* unboxed objects */
3518 case BIGNUM_WIDETAG:
3519 #if N_WORD_BITS != 64
3520 case SINGLE_FLOAT_WIDETAG:
3522 case DOUBLE_FLOAT_WIDETAG:
3523 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3524 case LONG_FLOAT_WIDETAG:
3526 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3527 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3529 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3530 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3532 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3533 case COMPLEX_LONG_FLOAT_WIDETAG:
3535 case SIMPLE_BASE_STRING_WIDETAG:
3536 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3537 case SIMPLE_CHARACTER_STRING_WIDETAG:
3539 case SIMPLE_BIT_VECTOR_WIDETAG:
3540 case SIMPLE_ARRAY_NIL_WIDETAG:
3541 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3542 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3543 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3544 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3545 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3546 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3547 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
3548 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
3550 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3551 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3552 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
3553 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
3555 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3556 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3558 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3559 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3561 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3562 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3564 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3565 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3567 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
3568 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
3570 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3571 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3573 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
3574 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
3576 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3577 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3579 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3580 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3581 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3582 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3584 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3585 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3587 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3588 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3590 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3591 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3594 case WEAK_POINTER_WIDETAG:
3595 #ifdef LUTEX_WIDETAG
3598 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3599 case NO_TLS_VALUE_MARKER_WIDETAG:
3601 count = (sizetab[widetag_of(*start)])(start);
3605 lose("Unhandled widetag 0x%x at 0x%x\n",
3606 widetag_of(*start), start);
3618 /* FIXME: It would be nice to make names consistent so that
3619 * foo_size meant size *in* *bytes* instead of size in some
3620 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3621 * Some counts of lispobjs are called foo_count; it might be good
3622 * to grep for all foo_size and rename the appropriate ones to
3624 long read_only_space_size =
3625 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3626 - (lispobj*)READ_ONLY_SPACE_START;
3627 long static_space_size =
3628 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3629 - (lispobj*)STATIC_SPACE_START;
3631 for_each_thread(th) {
3632 long binding_stack_size =
3633 (lispobj*)get_binding_stack_pointer(th)
3634 - (lispobj*)th->binding_stack_start;
3635 verify_space(th->binding_stack_start, binding_stack_size);
3637 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3638 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3642 verify_generation(generation_index_t generation)
3646 for (i = 0; i < last_free_page; i++) {
3647 if (page_allocated_p(i)
3648 && (page_table[i].bytes_used != 0)
3649 && (page_table[i].gen == generation)) {
3650 page_index_t last_page;
3651 int region_allocation = page_table[i].allocated;
3653 /* This should be the start of a contiguous block */
3654 gc_assert(page_table[i].region_start_offset == 0);
3656 /* Need to find the full extent of this contiguous block in case
3657 objects span pages. */
3659 /* Now work forward until the end of this contiguous area is
3661 for (last_page = i; ;last_page++)
3662 /* Check whether this is the last page in this contiguous
3664 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3665 /* Or it is PAGE_BYTES and is the last in the block */
3666 || (page_table[last_page+1].allocated != region_allocation)
3667 || (page_table[last_page+1].bytes_used == 0)
3668 || (page_table[last_page+1].gen != generation)
3669 || (page_table[last_page+1].region_start_offset == 0))
3672 verify_space(page_address(i),
3674 (page_table[last_page].bytes_used
3675 + npage_bytes(last_page-i)))
3682 /* Check that all the free space is zero filled. */
3684 verify_zero_fill(void)
3688 for (page = 0; page < last_free_page; page++) {
3689 if (page_free_p(page)) {
3690 /* The whole page should be zero filled. */
3691 long *start_addr = (long *)page_address(page);
3694 for (i = 0; i < size; i++) {
3695 if (start_addr[i] != 0) {
3696 lose("free page not zero at %x\n", start_addr + i);
3700 long free_bytes = PAGE_BYTES - page_table[page].bytes_used;
3701 if (free_bytes > 0) {
3702 long *start_addr = (long *)((unsigned long)page_address(page)
3703 + page_table[page].bytes_used);
3704 long size = free_bytes / N_WORD_BYTES;
3706 for (i = 0; i < size; i++) {
3707 if (start_addr[i] != 0) {
3708 lose("free region not zero at %x\n", start_addr + i);
3716 /* External entry point for verify_zero_fill */
3718 gencgc_verify_zero_fill(void)
3720 /* Flush the alloc regions updating the tables. */
3721 gc_alloc_update_all_page_tables();
3722 SHOW("verifying zero fill");
3727 verify_dynamic_space(void)
3729 generation_index_t i;
3731 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3732 verify_generation(i);
3734 if (gencgc_enable_verify_zero_fill)
3738 /* Write-protect all the dynamic boxed pages in the given generation. */
3740 write_protect_generation_pages(generation_index_t generation)
3744 gc_assert(generation < SCRATCH_GENERATION);
3746 for (start = 0; start < last_free_page; start++) {
3747 if (protect_page_p(start, generation)) {
3751 /* Note the page as protected in the page tables. */
3752 page_table[start].write_protected = 1;
3754 for (last = start + 1; last < last_free_page; last++) {
3755 if (!protect_page_p(last, generation))
3757 page_table[last].write_protected = 1;
3760 page_start = (void *)page_address(start);
3762 os_protect(page_start,
3763 npage_bytes(last - start),
3764 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3770 if (gencgc_verbose > 1) {
3772 "/write protected %d of %d pages in generation %d\n",
3773 count_write_protect_generation_pages(generation),
3774 count_generation_pages(generation),
3779 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3782 scavenge_control_stack()
3784 unsigned long control_stack_size;
3786 /* This is going to be a big problem when we try to port threads
3788 struct thread *th = arch_os_get_current_thread();
3789 lispobj *control_stack =
3790 (lispobj *)(th->control_stack_start);
3792 control_stack_size = current_control_stack_pointer - control_stack;
3793 scavenge(control_stack, control_stack_size);
3796 /* Scavenging Interrupt Contexts */
3798 static int boxed_registers[] = BOXED_REGISTERS;
3801 scavenge_interrupt_context(os_context_t * context)
3807 unsigned long lip_offset;
3808 int lip_register_pair;
3810 unsigned long pc_code_offset;
3812 #ifdef ARCH_HAS_LINK_REGISTER
3813 unsigned long lr_code_offset;
3815 #ifdef ARCH_HAS_NPC_REGISTER
3816 unsigned long npc_code_offset;
3820 /* Find the LIP's register pair and calculate it's offset */
3821 /* before we scavenge the context. */
3824 * I (RLT) think this is trying to find the boxed register that is
3825 * closest to the LIP address, without going past it. Usually, it's
3826 * reg_CODE or reg_LRA. But sometimes, nothing can be found.
3828 lip = *os_context_register_addr(context, reg_LIP);
3829 lip_offset = 0x7FFFFFFF;
3830 lip_register_pair = -1;
3831 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3836 index = boxed_registers[i];
3837 reg = *os_context_register_addr(context, index);
3838 if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
3840 if (offset < lip_offset) {
3841 lip_offset = offset;
3842 lip_register_pair = index;
3846 #endif /* reg_LIP */
3848 /* Compute the PC's offset from the start of the CODE */
3850 pc_code_offset = *os_context_pc_addr(context)
3851 - *os_context_register_addr(context, reg_CODE);
3852 #ifdef ARCH_HAS_NPC_REGISTER
3853 npc_code_offset = *os_context_npc_addr(context)
3854 - *os_context_register_addr(context, reg_CODE);
3855 #endif /* ARCH_HAS_NPC_REGISTER */
3857 #ifdef ARCH_HAS_LINK_REGISTER
3859 *os_context_lr_addr(context) -
3860 *os_context_register_addr(context, reg_CODE);
3863 /* Scanvenge all boxed registers in the context. */
3864 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3868 index = boxed_registers[i];
3869 foo = *os_context_register_addr(context, index);
3871 *os_context_register_addr(context, index) = foo;
3873 scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
3880 * But what happens if lip_register_pair is -1?
3881 * *os_context_register_addr on Solaris (see
3882 * solaris_register_address in solaris-os.c) will return
3883 * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
3884 * that what we really want? My guess is that that is not what we
3885 * want, so if lip_register_pair is -1, we don't touch reg_LIP at
3886 * all. But maybe it doesn't really matter if LIP is trashed?
3888 if (lip_register_pair >= 0) {
3889 *os_context_register_addr(context, reg_LIP) =
3890 *os_context_register_addr(context, lip_register_pair)
3893 #endif /* reg_LIP */
3895 /* Fix the PC if it was in from space */
3896 if (from_space_p(*os_context_pc_addr(context)))
3897 *os_context_pc_addr(context) =
3898 *os_context_register_addr(context, reg_CODE) + pc_code_offset;
3900 #ifdef ARCH_HAS_LINK_REGISTER
3901 /* Fix the LR ditto; important if we're being called from
3902 * an assembly routine that expects to return using blr, otherwise
3904 if (from_space_p(*os_context_lr_addr(context)))
3905 *os_context_lr_addr(context) =
3906 *os_context_register_addr(context, reg_CODE) + lr_code_offset;
3909 #ifdef ARCH_HAS_NPC_REGISTER
3910 if (from_space_p(*os_context_npc_addr(context)))
3911 *os_context_npc_addr(context) =
3912 *os_context_register_addr(context, reg_CODE) + npc_code_offset;
3913 #endif /* ARCH_HAS_NPC_REGISTER */
3917 scavenge_interrupt_contexts(void)
3920 os_context_t *context;
3922 struct thread *th=arch_os_get_current_thread();
3924 index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
3926 #if defined(DEBUG_PRINT_CONTEXT_INDEX)
3927 printf("Number of active contexts: %d\n", index);
3930 for (i = 0; i < index; i++) {
3931 context = th->interrupt_contexts[i];
3932 scavenge_interrupt_context(context);
3938 #if defined(LISP_FEATURE_SB_THREAD)
3940 preserve_context_registers (os_context_t *c)
3943 /* On Darwin the signal context isn't a contiguous block of memory,
3944 * so just preserve_pointering its contents won't be sufficient.
3946 #if defined(LISP_FEATURE_DARWIN)
3947 #if defined LISP_FEATURE_X86
3948 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3949 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3950 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3951 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3952 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3953 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3954 preserve_pointer((void*)*os_context_pc_addr(c));
3955 #elif defined LISP_FEATURE_X86_64
3956 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3957 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3958 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3959 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3960 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3961 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3962 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3963 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3964 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3965 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3966 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3967 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3968 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3969 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3970 preserve_pointer((void*)*os_context_pc_addr(c));
3972 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3975 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3976 preserve_pointer(*ptr);
3981 /* Garbage collect a generation. If raise is 0 then the remains of the
3982 * generation are not raised to the next generation. */
3984 garbage_collect_generation(generation_index_t generation, int raise)
3986 unsigned long bytes_freed;
3988 unsigned long static_space_size;
3989 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3992 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3994 /* The oldest generation can't be raised. */
3995 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3997 /* Check if weak hash tables were processed in the previous GC. */
3998 gc_assert(weak_hash_tables == NULL);
4000 /* Initialize the weak pointer list. */
4001 weak_pointers = NULL;
4003 #ifdef LUTEX_WIDETAG
4004 unmark_lutexes(generation);
4007 /* When a generation is not being raised it is transported to a
4008 * temporary generation (NUM_GENERATIONS), and lowered when
4009 * done. Set up this new generation. There should be no pages
4010 * allocated to it yet. */
4012 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
4015 /* Set the global src and dest. generations */
4016 from_space = generation;
4018 new_space = generation+1;
4020 new_space = SCRATCH_GENERATION;
4022 /* Change to a new space for allocation, resetting the alloc_start_page */
4023 gc_alloc_generation = new_space;
4024 generations[new_space].alloc_start_page = 0;
4025 generations[new_space].alloc_unboxed_start_page = 0;
4026 generations[new_space].alloc_large_start_page = 0;
4027 generations[new_space].alloc_large_unboxed_start_page = 0;
4029 /* Before any pointers are preserved, the dont_move flags on the
4030 * pages need to be cleared. */
4031 for (i = 0; i < last_free_page; i++)
4032 if(page_table[i].gen==from_space)
4033 page_table[i].dont_move = 0;
4035 /* Un-write-protect the old-space pages. This is essential for the
4036 * promoted pages as they may contain pointers into the old-space
4037 * which need to be scavenged. It also helps avoid unnecessary page
4038 * faults as forwarding pointers are written into them. They need to
4039 * be un-protected anyway before unmapping later. */
4040 unprotect_oldspace();
4042 /* Scavenge the stacks' conservative roots. */
4044 /* there are potentially two stacks for each thread: the main
4045 * stack, which may contain Lisp pointers, and the alternate stack.
4046 * We don't ever run Lisp code on the altstack, but it may
4047 * host a sigcontext with lisp objects in it */
4049 /* what we need to do: (1) find the stack pointer for the main
4050 * stack; scavenge it (2) find the interrupt context on the
4051 * alternate stack that might contain lisp values, and scavenge
4054 /* we assume that none of the preceding applies to the thread that
4055 * initiates GC. If you ever call GC from inside an altstack
4056 * handler, you will lose. */
4058 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4059 /* And if we're saving a core, there's no point in being conservative. */
4060 if (conservative_stack) {
4061 for_each_thread(th) {
4063 void **esp=(void **)-1;
4064 #ifdef LISP_FEATURE_SB_THREAD
4066 if(th==arch_os_get_current_thread()) {
4067 /* Somebody is going to burn in hell for this, but casting
4068 * it in two steps shuts gcc up about strict aliasing. */
4069 esp = (void **)((void *)&raise);
4072 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
4073 for(i=free-1;i>=0;i--) {
4074 os_context_t *c=th->interrupt_contexts[i];
4075 esp1 = (void **) *os_context_register_addr(c,reg_SP);
4076 if (esp1>=(void **)th->control_stack_start &&
4077 esp1<(void **)th->control_stack_end) {
4078 if(esp1<esp) esp=esp1;
4079 preserve_context_registers(c);
4084 esp = (void **)((void *)&raise);
4086 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
4087 preserve_pointer(*ptr);
4094 if (gencgc_verbose > 1) {
4095 long num_dont_move_pages = count_dont_move_pages();
4097 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
4098 num_dont_move_pages,
4099 npage_bytes(num_dont_move_pages));
4103 /* Scavenge all the rest of the roots. */
4105 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4107 * If not x86, we need to scavenge the interrupt context(s) and the
4110 scavenge_interrupt_contexts();
4111 scavenge_control_stack();
4114 /* Scavenge the Lisp functions of the interrupt handlers, taking
4115 * care to avoid SIG_DFL and SIG_IGN. */
4116 for (i = 0; i < NSIG; i++) {
4117 union interrupt_handler handler = interrupt_handlers[i];
4118 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4119 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4120 scavenge((lispobj *)(interrupt_handlers + i), 1);
4123 /* Scavenge the binding stacks. */
4126 for_each_thread(th) {
4127 long len= (lispobj *)get_binding_stack_pointer(th) -
4128 th->binding_stack_start;
4129 scavenge((lispobj *) th->binding_stack_start,len);
4130 #ifdef LISP_FEATURE_SB_THREAD
4131 /* do the tls as well */
4132 len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
4133 (sizeof (struct thread))/(sizeof (lispobj));
4134 scavenge((lispobj *) (th+1),len);
4139 /* The original CMU CL code had scavenge-read-only-space code
4140 * controlled by the Lisp-level variable
4141 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4142 * wasn't documented under what circumstances it was useful or
4143 * safe to turn it on, so it's been turned off in SBCL. If you
4144 * want/need this functionality, and can test and document it,
4145 * please submit a patch. */
4147 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4148 unsigned long read_only_space_size =
4149 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4150 (lispobj*)READ_ONLY_SPACE_START;
4152 "/scavenge read only space: %d bytes\n",
4153 read_only_space_size * sizeof(lispobj)));
4154 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4158 /* Scavenge static space. */
4160 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4161 (lispobj *)STATIC_SPACE_START;
4162 if (gencgc_verbose > 1) {
4164 "/scavenge static space: %d bytes\n",
4165 static_space_size * sizeof(lispobj)));
4167 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4169 /* All generations but the generation being GCed need to be
4170 * scavenged. The new_space generation needs special handling as
4171 * objects may be moved in - it is handled separately below. */
4172 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4174 /* Finally scavenge the new_space generation. Keep going until no
4175 * more objects are moved into the new generation */
4176 scavenge_newspace_generation(new_space);
4178 /* FIXME: I tried reenabling this check when debugging unrelated
4179 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4180 * Since the current GC code seems to work well, I'm guessing that
4181 * this debugging code is just stale, but I haven't tried to
4182 * figure it out. It should be figured out and then either made to
4183 * work or just deleted. */
4184 #define RESCAN_CHECK 0
4186 /* As a check re-scavenge the newspace once; no new objects should
4189 long old_bytes_allocated = bytes_allocated;
4190 long bytes_allocated;
4192 /* Start with a full scavenge. */
4193 scavenge_newspace_generation_one_scan(new_space);
4195 /* Flush the current regions, updating the tables. */
4196 gc_alloc_update_all_page_tables();
4198 bytes_allocated = bytes_allocated - old_bytes_allocated;
4200 if (bytes_allocated != 0) {
4201 lose("Rescan of new_space allocated %d more bytes.\n",
4207 scan_weak_hash_tables();
4208 scan_weak_pointers();
4210 /* Flush the current regions, updating the tables. */
4211 gc_alloc_update_all_page_tables();
4213 /* Free the pages in oldspace, but not those marked dont_move. */
4214 bytes_freed = free_oldspace();
4216 /* If the GC is not raising the age then lower the generation back
4217 * to its normal generation number */
4219 for (i = 0; i < last_free_page; i++)
4220 if ((page_table[i].bytes_used != 0)
4221 && (page_table[i].gen == SCRATCH_GENERATION))
4222 page_table[i].gen = generation;
4223 gc_assert(generations[generation].bytes_allocated == 0);
4224 generations[generation].bytes_allocated =
4225 generations[SCRATCH_GENERATION].bytes_allocated;
4226 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4229 /* Reset the alloc_start_page for generation. */
4230 generations[generation].alloc_start_page = 0;
4231 generations[generation].alloc_unboxed_start_page = 0;
4232 generations[generation].alloc_large_start_page = 0;
4233 generations[generation].alloc_large_unboxed_start_page = 0;
4235 if (generation >= verify_gens) {
4236 if (gencgc_verbose) {
4240 verify_dynamic_space();
4243 /* Set the new gc trigger for the GCed generation. */
4244 generations[generation].gc_trigger =
4245 generations[generation].bytes_allocated
4246 + generations[generation].bytes_consed_between_gc;
4249 generations[generation].num_gc = 0;
4251 ++generations[generation].num_gc;
4253 #ifdef LUTEX_WIDETAG
4254 reap_lutexes(generation);
4256 move_lutexes(generation, generation+1);
4260 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4262 update_dynamic_space_free_pointer(void)
4264 page_index_t last_page = -1, i;
4266 for (i = 0; i < last_free_page; i++)
4267 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
4270 last_free_page = last_page+1;
4272 set_alloc_pointer((lispobj)(page_address(last_free_page)));
4273 return 0; /* dummy value: return something ... */
4277 remap_free_pages (page_index_t from, page_index_t to)
4279 page_index_t first_page, last_page;
4281 for (first_page = from; first_page <= to; first_page++) {
4282 if (page_allocated_p(first_page) ||
4283 (page_table[first_page].need_to_zero == 0)) {
4287 last_page = first_page + 1;
4288 while (page_free_p(last_page) &&
4290 (page_table[last_page].need_to_zero == 1)) {
4294 /* There's a mysterious Solaris/x86 problem with using mmap
4295 * tricks for memory zeroing. See sbcl-devel thread
4296 * "Re: patch: standalone executable redux".
4298 #if defined(LISP_FEATURE_SUNOS)
4299 zero_pages(first_page, last_page-1);
4301 zero_pages_with_mmap(first_page, last_page-1);
4304 first_page = last_page;
4308 generation_index_t small_generation_limit = 1;
4310 /* GC all generations newer than last_gen, raising the objects in each
4311 * to the next older generation - we finish when all generations below
4312 * last_gen are empty. Then if last_gen is due for a GC, or if
4313 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4314 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4316 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4317 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4319 collect_garbage(generation_index_t last_gen)
4321 generation_index_t gen = 0, i;
4324 /* The largest value of last_free_page seen since the time
4325 * remap_free_pages was called. */
4326 static page_index_t high_water_mark = 0;
4328 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4332 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4334 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4339 /* Flush the alloc regions updating the tables. */
4340 gc_alloc_update_all_page_tables();
4342 /* Verify the new objects created by Lisp code. */
4343 if (pre_verify_gen_0) {
4344 FSHOW((stderr, "pre-checking generation 0\n"));
4345 verify_generation(0);
4348 if (gencgc_verbose > 1)
4349 print_generation_stats();
4352 /* Collect the generation. */
4354 if (gen >= gencgc_oldest_gen_to_gc) {
4355 /* Never raise the oldest generation. */
4360 || (generations[gen].num_gc >= generations[gen].trigger_age);
4363 if (gencgc_verbose > 1) {
4365 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4368 generations[gen].bytes_allocated,
4369 generations[gen].gc_trigger,
4370 generations[gen].num_gc));
4373 /* If an older generation is being filled, then update its
4376 generations[gen+1].cum_sum_bytes_allocated +=
4377 generations[gen+1].bytes_allocated;
4380 garbage_collect_generation(gen, raise);
4382 /* Reset the memory age cum_sum. */
4383 generations[gen].cum_sum_bytes_allocated = 0;
4385 if (gencgc_verbose > 1) {
4386 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4387 print_generation_stats();
4391 } while ((gen <= gencgc_oldest_gen_to_gc)
4392 && ((gen < last_gen)
4393 || ((gen <= gencgc_oldest_gen_to_gc)
4395 && (generations[gen].bytes_allocated
4396 > generations[gen].gc_trigger)
4397 && (gen_av_mem_age(gen)
4398 > generations[gen].min_av_mem_age))));
4400 /* Now if gen-1 was raised all generations before gen are empty.
4401 * If it wasn't raised then all generations before gen-1 are empty.
4403 * Now objects within this gen's pages cannot point to younger
4404 * generations unless they are written to. This can be exploited
4405 * by write-protecting the pages of gen; then when younger
4406 * generations are GCed only the pages which have been written
4411 gen_to_wp = gen - 1;
4413 /* There's not much point in WPing pages in generation 0 as it is
4414 * never scavenged (except promoted pages). */
4415 if ((gen_to_wp > 0) && enable_page_protection) {
4416 /* Check that they are all empty. */
4417 for (i = 0; i < gen_to_wp; i++) {
4418 if (generations[i].bytes_allocated)
4419 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4422 write_protect_generation_pages(gen_to_wp);
4425 /* Set gc_alloc() back to generation 0. The current regions should
4426 * be flushed after the above GCs. */
4427 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4428 gc_alloc_generation = 0;
4430 /* Save the high-water mark before updating last_free_page */
4431 if (last_free_page > high_water_mark)
4432 high_water_mark = last_free_page;
4434 update_dynamic_space_free_pointer();
4436 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4438 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4441 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4444 if (gen > small_generation_limit) {
4445 if (last_free_page > high_water_mark)
4446 high_water_mark = last_free_page;
4447 remap_free_pages(0, high_water_mark);
4448 high_water_mark = 0;
4453 SHOW("returning from collect_garbage");
4456 /* This is called by Lisp PURIFY when it is finished. All live objects
4457 * will have been moved to the RO and Static heaps. The dynamic space
4458 * will need a full re-initialization. We don't bother having Lisp
4459 * PURIFY flush the current gc_alloc() region, as the page_tables are
4460 * re-initialized, and every page is zeroed to be sure. */
4466 if (gencgc_verbose > 1) {
4467 SHOW("entering gc_free_heap");
4470 for (page = 0; page < page_table_pages; page++) {
4471 /* Skip free pages which should already be zero filled. */
4472 if (page_allocated_p(page)) {
4473 void *page_start, *addr;
4475 /* Mark the page free. The other slots are assumed invalid
4476 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4477 * should not be write-protected -- except that the
4478 * generation is used for the current region but it sets
4480 page_table[page].allocated = FREE_PAGE_FLAG;
4481 page_table[page].bytes_used = 0;
4483 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4484 * about this change. */
4485 /* Zero the page. */
4486 page_start = (void *)page_address(page);
4488 /* First, remove any write-protection. */
4489 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
4490 page_table[page].write_protected = 0;
4492 os_invalidate(page_start,PAGE_BYTES);
4493 addr = os_validate(page_start,PAGE_BYTES);
4494 if (addr == NULL || addr != page_start) {
4495 lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n",
4500 page_table[page].write_protected = 0;
4502 } else if (gencgc_zero_check_during_free_heap) {
4503 /* Double-check that the page is zero filled. */
4506 gc_assert(page_free_p(page));
4507 gc_assert(page_table[page].bytes_used == 0);
4508 page_start = (long *)page_address(page);
4509 for (i=0; i<1024; i++) {
4510 if (page_start[i] != 0) {
4511 lose("free region not zero at %x\n", page_start + i);
4517 bytes_allocated = 0;
4519 /* Initialize the generations. */
4520 for (page = 0; page < NUM_GENERATIONS; page++) {
4521 generations[page].alloc_start_page = 0;
4522 generations[page].alloc_unboxed_start_page = 0;
4523 generations[page].alloc_large_start_page = 0;
4524 generations[page].alloc_large_unboxed_start_page = 0;
4525 generations[page].bytes_allocated = 0;
4526 generations[page].gc_trigger = 2000000;
4527 generations[page].num_gc = 0;
4528 generations[page].cum_sum_bytes_allocated = 0;
4529 generations[page].lutexes = NULL;
4532 if (gencgc_verbose > 1)
4533 print_generation_stats();
4535 /* Initialize gc_alloc(). */
4536 gc_alloc_generation = 0;
4538 gc_set_region_empty(&boxed_region);
4539 gc_set_region_empty(&unboxed_region);
4542 set_alloc_pointer((lispobj)((char *)heap_base));
4544 if (verify_after_free_heap) {
4545 /* Check whether purify has left any bad pointers. */
4546 FSHOW((stderr, "checking after free_heap\n"));
4556 /* Compute the number of pages needed for the dynamic space.
4557 * Dynamic space size should be aligned on page size. */
4558 page_table_pages = dynamic_space_size/PAGE_BYTES;
4559 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4561 page_table = calloc(page_table_pages, sizeof(struct page));
4562 gc_assert(page_table);
4565 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4566 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4568 #ifdef LUTEX_WIDETAG
4569 scavtab[LUTEX_WIDETAG] = scav_lutex;
4570 transother[LUTEX_WIDETAG] = trans_lutex;
4571 sizetab[LUTEX_WIDETAG] = size_lutex;
4574 heap_base = (void*)DYNAMIC_SPACE_START;
4576 /* Initialize each page structure. */
4577 for (i = 0; i < page_table_pages; i++) {
4578 /* Initialize all pages as free. */
4579 page_table[i].allocated = FREE_PAGE_FLAG;
4580 page_table[i].bytes_used = 0;
4582 /* Pages are not write-protected at startup. */
4583 page_table[i].write_protected = 0;
4586 bytes_allocated = 0;
4588 /* Initialize the generations.
4590 * FIXME: very similar to code in gc_free_heap(), should be shared */
4591 for (i = 0; i < NUM_GENERATIONS; i++) {
4592 generations[i].alloc_start_page = 0;
4593 generations[i].alloc_unboxed_start_page = 0;
4594 generations[i].alloc_large_start_page = 0;
4595 generations[i].alloc_large_unboxed_start_page = 0;
4596 generations[i].bytes_allocated = 0;
4597 generations[i].gc_trigger = 2000000;
4598 generations[i].num_gc = 0;
4599 generations[i].cum_sum_bytes_allocated = 0;
4600 /* the tune-able parameters */
4601 generations[i].bytes_consed_between_gc = 2000000;
4602 generations[i].trigger_age = 1;
4603 generations[i].min_av_mem_age = 0.75;
4604 generations[i].lutexes = NULL;
4607 /* Initialize gc_alloc. */
4608 gc_alloc_generation = 0;
4609 gc_set_region_empty(&boxed_region);
4610 gc_set_region_empty(&unboxed_region);
4615 /* Pick up the dynamic space from after a core load.
4617 * The ALLOCATION_POINTER points to the end of the dynamic space.
4621 gencgc_pickup_dynamic(void)
4623 page_index_t page = 0;
4624 void *alloc_ptr = (void *)get_alloc_pointer();
4625 lispobj *prev=(lispobj *)page_address(page);
4626 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4628 lispobj *first,*ptr= (lispobj *)page_address(page);
4629 page_table[page].allocated = BOXED_PAGE_FLAG;
4630 page_table[page].gen = gen;
4631 page_table[page].bytes_used = PAGE_BYTES;
4632 page_table[page].large_object = 0;
4633 page_table[page].write_protected = 0;
4634 page_table[page].write_protected_cleared = 0;
4635 page_table[page].dont_move = 0;
4636 page_table[page].need_to_zero = 1;
4638 if (!gencgc_partial_pickup) {
4639 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4640 if(ptr == first) prev=ptr;
4641 page_table[page].region_start_offset =
4642 page_address(page) - (void *)prev;
4645 } while (page_address(page) < alloc_ptr);
4647 #ifdef LUTEX_WIDETAG
4648 /* Lutexes have been registered in generation 0 by coreparse, and
4649 * need to be moved to the right one manually.
4651 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4654 last_free_page = page;
4656 generations[gen].bytes_allocated = npage_bytes(page);
4657 bytes_allocated = npage_bytes(page);
4659 gc_alloc_update_all_page_tables();
4660 write_protect_generation_pages(gen);
4664 gc_initialize_pointers(void)
4666 gencgc_pickup_dynamic();
4670 /* alloc(..) is the external interface for memory allocation. It
4671 * allocates to generation 0. It is not called from within the garbage
4672 * collector as it is only external uses that need the check for heap
4673 * size (GC trigger) and to disable the interrupts (interrupts are
4674 * always disabled during a GC).
4676 * The vops that call alloc(..) assume that the returned space is zero-filled.
4677 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4679 * The check for a GC trigger is only performed when the current
4680 * region is full, so in most cases it's not needed. */
4682 static inline lispobj *
4683 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4684 struct thread *thread)
4686 #ifndef LISP_FEATURE_WIN32
4687 lispobj alloc_signal;
4690 void *new_free_pointer;
4692 gc_assert(nbytes>0);
4694 /* Check for alignment allocation problems. */
4695 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4696 && ((nbytes & LOWTAG_MASK) == 0));
4698 /* Must be inside a PA section. */
4699 gc_assert(get_pseudo_atomic_atomic(thread));
4701 /* maybe we can do this quickly ... */
4702 new_free_pointer = region->free_pointer + nbytes;
4703 if (new_free_pointer <= region->end_addr) {
4704 new_obj = (void*)(region->free_pointer);
4705 region->free_pointer = new_free_pointer;
4706 return(new_obj); /* yup */
4709 /* we have to go the long way around, it seems. Check whether we
4710 * should GC in the near future
4712 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4713 /* Don't flood the system with interrupts if the need to gc is
4714 * already noted. This can happen for example when SUB-GC
4715 * allocates or after a gc triggered in a WITHOUT-GCING. */
4716 if (SymbolValue(GC_PENDING,thread) == NIL) {
4717 /* set things up so that GC happens when we finish the PA
4719 SetSymbolValue(GC_PENDING,T,thread);
4720 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4721 set_pseudo_atomic_interrupted(thread);
4722 #ifdef LISP_FEATURE_PPC
4723 /* PPC calls alloc() from a trap or from pa_alloc(),
4724 * look up the most context if it's from a trap. */
4726 os_context_t *context =
4727 thread->interrupt_data->allocation_trap_context;
4728 maybe_save_gc_mask_and_block_deferrables
4729 (context ? os_context_sigmask_addr(context) : NULL);
4732 maybe_save_gc_mask_and_block_deferrables(NULL);
4737 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4739 #ifndef LISP_FEATURE_WIN32
4740 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4741 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4742 if ((signed long) alloc_signal <= 0) {
4743 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4746 SetSymbolValue(ALLOC_SIGNAL,
4747 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4757 general_alloc(long nbytes, int page_type_flag)
4759 struct thread *thread = arch_os_get_current_thread();
4760 /* Select correct region, and call general_alloc_internal with it.
4761 * For other then boxed allocation we must lock first, since the
4762 * region is shared. */
4763 if (BOXED_PAGE_FLAG & page_type_flag) {
4764 #ifdef LISP_FEATURE_SB_THREAD
4765 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4767 struct alloc_region *region = &boxed_region;
4769 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4770 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4772 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4773 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4774 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4777 lose("bad page type flag: %d", page_type_flag);
4784 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4785 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4789 * shared support for the OS-dependent signal handlers which
4790 * catch GENCGC-related write-protect violations
4792 void unhandled_sigmemoryfault(void* addr);
4794 /* Depending on which OS we're running under, different signals might
4795 * be raised for a violation of write protection in the heap. This
4796 * function factors out the common generational GC magic which needs
4797 * to invoked in this case, and should be called from whatever signal
4798 * handler is appropriate for the OS we're running under.
4800 * Return true if this signal is a normal generational GC thing that
4801 * we were able to handle, or false if it was abnormal and control
4802 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4805 gencgc_handle_wp_violation(void* fault_addr)
4807 page_index_t page_index = find_page_index(fault_addr);
4810 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4811 fault_addr, page_index));
4814 /* Check whether the fault is within the dynamic space. */
4815 if (page_index == (-1)) {
4817 /* It can be helpful to be able to put a breakpoint on this
4818 * case to help diagnose low-level problems. */
4819 unhandled_sigmemoryfault(fault_addr);
4821 /* not within the dynamic space -- not our responsibility */
4826 ret = thread_mutex_lock(&free_pages_lock);
4827 gc_assert(ret == 0);
4828 if (page_table[page_index].write_protected) {
4829 /* Unprotect the page. */
4830 os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
4831 page_table[page_index].write_protected_cleared = 1;
4832 page_table[page_index].write_protected = 0;
4834 /* The only acceptable reason for this signal on a heap
4835 * access is that GENCGC write-protected the page.
4836 * However, if two CPUs hit a wp page near-simultaneously,
4837 * we had better not have the second one lose here if it
4838 * does this test after the first one has already set wp=0
4840 if(page_table[page_index].write_protected_cleared != 1)
4841 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4842 page_index, boxed_region.first_page,
4843 boxed_region.last_page);
4845 ret = thread_mutex_unlock(&free_pages_lock);
4846 gc_assert(ret == 0);
4847 /* Don't worry, we can handle it. */
4851 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4852 * it's not just a case of the program hitting the write barrier, and
4853 * are about to let Lisp deal with it. It's basically just a
4854 * convenient place to set a gdb breakpoint. */
4856 unhandled_sigmemoryfault(void *addr)
4859 void gc_alloc_update_all_page_tables(void)
4861 /* Flush the alloc regions updating the tables. */
4864 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4865 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4866 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4870 gc_set_region_empty(struct alloc_region *region)
4872 region->first_page = 0;
4873 region->last_page = -1;
4874 region->start_addr = page_address(0);
4875 region->free_pointer = page_address(0);
4876 region->end_addr = page_address(0);
4880 zero_all_free_pages()
4884 for (i = 0; i < last_free_page; i++) {
4885 if (page_free_p(i)) {
4886 #ifdef READ_PROTECT_FREE_PAGES
4887 os_protect(page_address(i),
4896 /* Things to do before doing a final GC before saving a core (without
4899 * + Pages in large_object pages aren't moved by the GC, so we need to
4900 * unset that flag from all pages.
4901 * + The pseudo-static generation isn't normally collected, but it seems
4902 * reasonable to collect it at least when saving a core. So move the
4903 * pages to a normal generation.
4906 prepare_for_final_gc ()
4909 for (i = 0; i < last_free_page; i++) {
4910 page_table[i].large_object = 0;
4911 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4912 int used = page_table[i].bytes_used;
4913 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4914 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4915 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4921 /* Do a non-conservative GC, and then save a core with the initial
4922 * function being set to the value of the static symbol
4923 * SB!VM:RESTART-LISP-FUNCTION */
4925 gc_and_save(char *filename, boolean prepend_runtime,
4926 boolean save_runtime_options)
4929 void *runtime_bytes = NULL;
4930 size_t runtime_size;
4932 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4937 conservative_stack = 0;
4939 /* The filename might come from Lisp, and be moved by the now
4940 * non-conservative GC. */
4941 filename = strdup(filename);
4943 /* Collect twice: once into relatively high memory, and then back
4944 * into low memory. This compacts the retained data into the lower
4945 * pages, minimizing the size of the core file.
4947 prepare_for_final_gc();
4948 gencgc_alloc_start_page = last_free_page;
4949 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4951 prepare_for_final_gc();
4952 gencgc_alloc_start_page = -1;
4953 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4955 if (prepend_runtime)
4956 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4958 /* The dumper doesn't know that pages need to be zeroed before use. */
4959 zero_all_free_pages();
4960 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4961 prepend_runtime, save_runtime_options);
4962 /* Oops. Save still managed to fail. Since we've mangled the stack
4963 * beyond hope, there's not much we can do.
4964 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4965 * going to be rather unsatisfactory too... */
4966 lose("Attempt to save core after non-conservative GC failed.\n");