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(int verbose) /* FIXME: should take FILE argument */
435 generation_index_t i, gens;
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 /* highest generation to print */
451 gens = SCRATCH_GENERATION;
453 gens = PSEUDO_STATIC_GENERATION;
455 /* Print the heap stats. */
457 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
459 for (i = 0; i < gens; i++) {
462 long unboxed_cnt = 0;
463 long large_boxed_cnt = 0;
464 long large_unboxed_cnt = 0;
467 for (j = 0; j < last_free_page; j++)
468 if (page_table[j].gen == i) {
470 /* Count the number of boxed pages within the given
472 if (page_boxed_p(j)) {
473 if (page_table[j].large_object)
478 if(page_table[j].dont_move) pinned_cnt++;
479 /* Count the number of unboxed pages within the given
481 if (page_unboxed_p(j)) {
482 if (page_table[j].large_object)
489 gc_assert(generations[i].bytes_allocated
490 == count_generation_bytes_allocated(i));
492 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
494 generations[i].alloc_start_page,
495 generations[i].alloc_unboxed_start_page,
496 generations[i].alloc_large_start_page,
497 generations[i].alloc_large_unboxed_start_page,
503 generations[i].bytes_allocated,
504 (npage_bytes(count_generation_pages(i))
505 - generations[i].bytes_allocated),
506 generations[i].gc_trigger,
507 count_write_protect_generation_pages(i),
508 generations[i].num_gc,
511 fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated);
512 fprintf(stderr," Dynamic-space-size bytes = %u\n", dynamic_space_size);
514 fpu_restore(fpu_state);
518 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
519 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
522 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
523 * if zeroing it ourselves, i.e. in practice give the memory back to the
524 * OS. Generally done after a large GC.
526 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
528 void *addr = page_address(start), *new_addr;
529 size_t length = npage_bytes(1+end-start);
534 os_invalidate(addr, length);
535 new_addr = os_validate(addr, length);
536 if (new_addr == NULL || new_addr != addr) {
537 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
541 for (i = start; i <= end; i++) {
542 page_table[i].need_to_zero = 0;
546 /* Zero the pages from START to END (inclusive). Generally done just after
547 * a new region has been allocated.
550 zero_pages(page_index_t start, page_index_t end) {
554 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
555 fast_bzero(page_address(start), npage_bytes(1+end-start));
557 bzero(page_address(start), npage_bytes(1+end-start));
562 /* Zero the pages from START to END (inclusive), except for those
563 * pages that are known to already zeroed. Mark all pages in the
564 * ranges as non-zeroed.
567 zero_dirty_pages(page_index_t start, page_index_t end) {
570 for (i = start; i <= end; i++) {
571 if (page_table[i].need_to_zero == 1) {
572 zero_pages(start, end);
577 for (i = start; i <= end; i++) {
578 page_table[i].need_to_zero = 1;
584 * To support quick and inline allocation, regions of memory can be
585 * allocated and then allocated from with just a free pointer and a
586 * check against an end address.
588 * Since objects can be allocated to spaces with different properties
589 * e.g. boxed/unboxed, generation, ages; there may need to be many
590 * allocation regions.
592 * Each allocation region may start within a partly used page. Many
593 * features of memory use are noted on a page wise basis, e.g. the
594 * generation; so if a region starts within an existing allocated page
595 * it must be consistent with this page.
597 * During the scavenging of the newspace, objects will be transported
598 * into an allocation region, and pointers updated to point to this
599 * allocation region. It is possible that these pointers will be
600 * scavenged again before the allocation region is closed, e.g. due to
601 * trans_list which jumps all over the place to cleanup the list. It
602 * is important to be able to determine properties of all objects
603 * pointed to when scavenging, e.g to detect pointers to the oldspace.
604 * Thus it's important that the allocation regions have the correct
605 * properties set when allocated, and not just set when closed. The
606 * region allocation routines return regions with the specified
607 * properties, and grab all the pages, setting their properties
608 * appropriately, except that the amount used is not known.
610 * These regions are used to support quicker allocation using just a
611 * free pointer. The actual space used by the region is not reflected
612 * in the pages tables until it is closed. It can't be scavenged until
615 * When finished with the region it should be closed, which will
616 * update the page tables for the actual space used returning unused
617 * space. Further it may be noted in the new regions which is
618 * necessary when scavenging the newspace.
620 * Large objects may be allocated directly without an allocation
621 * region, the page tables are updated immediately.
623 * Unboxed objects don't contain pointers to other objects and so
624 * don't need scavenging. Further they can't contain pointers to
625 * younger generations so WP is not needed. By allocating pages to
626 * unboxed objects the whole page never needs scavenging or
627 * write-protecting. */
629 /* We are only using two regions at present. Both are for the current
630 * newspace generation. */
631 struct alloc_region boxed_region;
632 struct alloc_region unboxed_region;
634 /* The generation currently being allocated to. */
635 static generation_index_t gc_alloc_generation;
637 static inline page_index_t
638 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
641 if (UNBOXED_PAGE_FLAG == page_type_flag) {
642 return generations[generation].alloc_large_unboxed_start_page;
643 } else if (BOXED_PAGE_FLAG & page_type_flag) {
644 /* Both code and data. */
645 return generations[generation].alloc_large_start_page;
647 lose("bad page type flag: %d", page_type_flag);
650 if (UNBOXED_PAGE_FLAG == page_type_flag) {
651 return generations[generation].alloc_unboxed_start_page;
652 } else if (BOXED_PAGE_FLAG & page_type_flag) {
653 /* Both code and data. */
654 return generations[generation].alloc_start_page;
656 lose("bad page_type_flag: %d", page_type_flag);
662 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
666 if (UNBOXED_PAGE_FLAG == page_type_flag) {
667 generations[generation].alloc_large_unboxed_start_page = page;
668 } else if (BOXED_PAGE_FLAG & page_type_flag) {
669 /* Both code and data. */
670 generations[generation].alloc_large_start_page = page;
672 lose("bad page type flag: %d", page_type_flag);
675 if (UNBOXED_PAGE_FLAG == page_type_flag) {
676 generations[generation].alloc_unboxed_start_page = page;
677 } else if (BOXED_PAGE_FLAG & page_type_flag) {
678 /* Both code and data. */
679 generations[generation].alloc_start_page = page;
681 lose("bad page type flag: %d", page_type_flag);
686 /* Find a new region with room for at least the given number of bytes.
688 * It starts looking at the current generation's alloc_start_page. So
689 * may pick up from the previous region if there is enough space. This
690 * keeps the allocation contiguous when scavenging the newspace.
692 * The alloc_region should have been closed by a call to
693 * gc_alloc_update_page_tables(), and will thus be in an empty state.
695 * To assist the scavenging functions write-protected pages are not
696 * used. Free pages should not be write-protected.
698 * It is critical to the conservative GC that the start of regions be
699 * known. To help achieve this only small regions are allocated at a
702 * During scavenging, pointers may be found to within the current
703 * region and the page generation must be set so that pointers to the
704 * from space can be recognized. Therefore the generation of pages in
705 * the region are set to gc_alloc_generation. To prevent another
706 * allocation call using the same pages, all the pages in the region
707 * are allocated, although they will initially be empty.
710 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
712 page_index_t first_page;
713 page_index_t last_page;
714 unsigned long bytes_found;
720 "/alloc_new_region for %d bytes from gen %d\n",
721 nbytes, gc_alloc_generation));
724 /* Check that the region is in a reset state. */
725 gc_assert((alloc_region->first_page == 0)
726 && (alloc_region->last_page == -1)
727 && (alloc_region->free_pointer == alloc_region->end_addr));
728 ret = thread_mutex_lock(&free_pages_lock);
730 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
731 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
732 bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
733 + npage_bytes(last_page-first_page);
735 /* Set up the alloc_region. */
736 alloc_region->first_page = first_page;
737 alloc_region->last_page = last_page;
738 alloc_region->start_addr = page_table[first_page].bytes_used
739 + page_address(first_page);
740 alloc_region->free_pointer = alloc_region->start_addr;
741 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
743 /* Set up the pages. */
745 /* The first page may have already been in use. */
746 if (page_table[first_page].bytes_used == 0) {
747 page_table[first_page].allocated = page_type_flag;
748 page_table[first_page].gen = gc_alloc_generation;
749 page_table[first_page].large_object = 0;
750 page_table[first_page].region_start_offset = 0;
753 gc_assert(page_table[first_page].allocated == page_type_flag);
754 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
756 gc_assert(page_table[first_page].gen == gc_alloc_generation);
757 gc_assert(page_table[first_page].large_object == 0);
759 for (i = first_page+1; i <= last_page; i++) {
760 page_table[i].allocated = page_type_flag;
761 page_table[i].gen = gc_alloc_generation;
762 page_table[i].large_object = 0;
763 /* This may not be necessary for unboxed regions (think it was
765 page_table[i].region_start_offset =
766 void_diff(page_address(i),alloc_region->start_addr);
767 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
769 /* Bump up last_free_page. */
770 if (last_page+1 > last_free_page) {
771 last_free_page = last_page+1;
772 /* do we only want to call this on special occasions? like for
774 set_alloc_pointer((lispobj)page_address(last_free_page));
776 ret = thread_mutex_unlock(&free_pages_lock);
779 #ifdef READ_PROTECT_FREE_PAGES
780 os_protect(page_address(first_page),
781 npage_bytes(1+last_page-first_page),
785 /* If the first page was only partial, don't check whether it's
786 * zeroed (it won't be) and don't zero it (since the parts that
787 * we're interested in are guaranteed to be zeroed).
789 if (page_table[first_page].bytes_used) {
793 zero_dirty_pages(first_page, last_page);
795 /* we can do this after releasing free_pages_lock */
796 if (gencgc_zero_check) {
798 for (p = (long *)alloc_region->start_addr;
799 p < (long *)alloc_region->end_addr; p++) {
801 /* KLUDGE: It would be nice to use %lx and explicit casts
802 * (long) in code like this, so that it is less likely to
803 * break randomly when running on a machine with different
804 * word sizes. -- WHN 19991129 */
805 lose("The new region at %x is not zero (start=%p, end=%p).\n",
806 p, alloc_region->start_addr, alloc_region->end_addr);
812 /* If the record_new_objects flag is 2 then all new regions created
815 * If it's 1 then then it is only recorded if the first page of the
816 * current region is <= new_areas_ignore_page. This helps avoid
817 * unnecessary recording when doing full scavenge pass.
819 * The new_object structure holds the page, byte offset, and size of
820 * new regions of objects. Each new area is placed in the array of
821 * these structures pointer to by new_areas. new_areas_index holds the
822 * offset into new_areas.
824 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
825 * later code must detect this and handle it, probably by doing a full
826 * scavenge of a generation. */
827 #define NUM_NEW_AREAS 512
828 static int record_new_objects = 0;
829 static page_index_t new_areas_ignore_page;
835 static struct new_area (*new_areas)[];
836 static long new_areas_index;
839 /* Add a new area to new_areas. */
841 add_new_area(page_index_t first_page, size_t offset, size_t size)
843 unsigned long new_area_start,c;
846 /* Ignore if full. */
847 if (new_areas_index >= NUM_NEW_AREAS)
850 switch (record_new_objects) {
854 if (first_page > new_areas_ignore_page)
863 new_area_start = npage_bytes(first_page) + offset;
865 /* Search backwards for a prior area that this follows from. If
866 found this will save adding a new area. */
867 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
868 unsigned long area_end =
869 npage_bytes((*new_areas)[i].page)
870 + (*new_areas)[i].offset
871 + (*new_areas)[i].size;
873 "/add_new_area S1 %d %d %d %d\n",
874 i, c, new_area_start, area_end));*/
875 if (new_area_start == area_end) {
877 "/adding to [%d] %d %d %d with %d %d %d:\n",
879 (*new_areas)[i].page,
880 (*new_areas)[i].offset,
881 (*new_areas)[i].size,
885 (*new_areas)[i].size += size;
890 (*new_areas)[new_areas_index].page = first_page;
891 (*new_areas)[new_areas_index].offset = offset;
892 (*new_areas)[new_areas_index].size = size;
894 "/new_area %d page %d offset %d size %d\n",
895 new_areas_index, first_page, offset, size));*/
898 /* Note the max new_areas used. */
899 if (new_areas_index > max_new_areas)
900 max_new_areas = new_areas_index;
903 /* Update the tables for the alloc_region. The region may be added to
906 * When done the alloc_region is set up so that the next quick alloc
907 * will fail safely and thus a new region will be allocated. Further
908 * it is safe to try to re-update the page table of this reset
911 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
914 page_index_t first_page;
915 page_index_t next_page;
916 unsigned long bytes_used;
917 unsigned long orig_first_page_bytes_used;
918 unsigned long region_size;
919 unsigned long byte_cnt;
923 first_page = alloc_region->first_page;
925 /* Catch an unused alloc_region. */
926 if ((first_page == 0) && (alloc_region->last_page == -1))
929 next_page = first_page+1;
931 ret = thread_mutex_lock(&free_pages_lock);
933 if (alloc_region->free_pointer != alloc_region->start_addr) {
934 /* some bytes were allocated in the region */
935 orig_first_page_bytes_used = page_table[first_page].bytes_used;
937 gc_assert(alloc_region->start_addr ==
938 (page_address(first_page)
939 + page_table[first_page].bytes_used));
941 /* All the pages used need to be updated */
943 /* Update the first page. */
945 /* If the page was free then set up the gen, and
946 * region_start_offset. */
947 if (page_table[first_page].bytes_used == 0)
948 gc_assert(page_table[first_page].region_start_offset == 0);
949 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
951 gc_assert(page_table[first_page].allocated & page_type_flag);
952 gc_assert(page_table[first_page].gen == gc_alloc_generation);
953 gc_assert(page_table[first_page].large_object == 0);
957 /* Calculate the number of bytes used in this page. This is not
958 * always the number of new bytes, unless it was free. */
960 if ((bytes_used = void_diff(alloc_region->free_pointer,
961 page_address(first_page)))
963 bytes_used = PAGE_BYTES;
966 page_table[first_page].bytes_used = bytes_used;
967 byte_cnt += bytes_used;
970 /* All the rest of the pages should be free. We need to set
971 * their region_start_offset pointer to the start of the
972 * region, and set the bytes_used. */
974 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
975 gc_assert(page_table[next_page].allocated & page_type_flag);
976 gc_assert(page_table[next_page].bytes_used == 0);
977 gc_assert(page_table[next_page].gen == gc_alloc_generation);
978 gc_assert(page_table[next_page].large_object == 0);
980 gc_assert(page_table[next_page].region_start_offset ==
981 void_diff(page_address(next_page),
982 alloc_region->start_addr));
984 /* Calculate the number of bytes used in this page. */
986 if ((bytes_used = void_diff(alloc_region->free_pointer,
987 page_address(next_page)))>PAGE_BYTES) {
988 bytes_used = PAGE_BYTES;
991 page_table[next_page].bytes_used = bytes_used;
992 byte_cnt += bytes_used;
997 region_size = void_diff(alloc_region->free_pointer,
998 alloc_region->start_addr);
999 bytes_allocated += region_size;
1000 generations[gc_alloc_generation].bytes_allocated += region_size;
1002 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1004 /* Set the generations alloc restart page to the last page of
1006 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1008 /* Add the region to the new_areas if requested. */
1009 if (BOXED_PAGE_FLAG & page_type_flag)
1010 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1014 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1016 gc_alloc_generation));
1019 /* There are no bytes allocated. Unallocate the first_page if
1020 * there are 0 bytes_used. */
1021 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1022 if (page_table[first_page].bytes_used == 0)
1023 page_table[first_page].allocated = FREE_PAGE_FLAG;
1026 /* Unallocate any unused pages. */
1027 while (next_page <= alloc_region->last_page) {
1028 gc_assert(page_table[next_page].bytes_used == 0);
1029 page_table[next_page].allocated = FREE_PAGE_FLAG;
1032 ret = thread_mutex_unlock(&free_pages_lock);
1033 gc_assert(ret == 0);
1035 /* alloc_region is per-thread, we're ok to do this unlocked */
1036 gc_set_region_empty(alloc_region);
1039 static inline void *gc_quick_alloc(long nbytes);
1041 /* Allocate a possibly large object. */
1043 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1045 page_index_t first_page;
1046 page_index_t last_page;
1047 int orig_first_page_bytes_used;
1050 unsigned long bytes_used;
1051 page_index_t next_page;
1054 ret = thread_mutex_lock(&free_pages_lock);
1055 gc_assert(ret == 0);
1057 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1058 if (first_page <= alloc_region->last_page) {
1059 first_page = alloc_region->last_page+1;
1062 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1064 gc_assert(first_page > alloc_region->last_page);
1066 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1068 /* Set up the pages. */
1069 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1071 /* If the first page was free then set up the gen, and
1072 * region_start_offset. */
1073 if (page_table[first_page].bytes_used == 0) {
1074 page_table[first_page].allocated = page_type_flag;
1075 page_table[first_page].gen = gc_alloc_generation;
1076 page_table[first_page].region_start_offset = 0;
1077 page_table[first_page].large_object = 1;
1080 gc_assert(page_table[first_page].allocated == page_type_flag);
1081 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1082 gc_assert(page_table[first_page].large_object == 1);
1086 /* Calc. the number of bytes used in this page. This is not
1087 * always the number of new bytes, unless it was free. */
1089 if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
1090 bytes_used = PAGE_BYTES;
1093 page_table[first_page].bytes_used = bytes_used;
1094 byte_cnt += bytes_used;
1096 next_page = first_page+1;
1098 /* All the rest of the pages should be free. We need to set their
1099 * region_start_offset pointer to the start of the region, and set
1100 * the bytes_used. */
1102 gc_assert(page_free_p(next_page));
1103 gc_assert(page_table[next_page].bytes_used == 0);
1104 page_table[next_page].allocated = page_type_flag;
1105 page_table[next_page].gen = gc_alloc_generation;
1106 page_table[next_page].large_object = 1;
1108 page_table[next_page].region_start_offset =
1109 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1111 /* Calculate the number of bytes used in this page. */
1113 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1114 if (bytes_used > PAGE_BYTES) {
1115 bytes_used = PAGE_BYTES;
1118 page_table[next_page].bytes_used = bytes_used;
1119 page_table[next_page].write_protected=0;
1120 page_table[next_page].dont_move=0;
1121 byte_cnt += bytes_used;
1125 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1127 bytes_allocated += nbytes;
1128 generations[gc_alloc_generation].bytes_allocated += nbytes;
1130 /* Add the region to the new_areas if requested. */
1131 if (BOXED_PAGE_FLAG & page_type_flag)
1132 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1134 /* Bump up last_free_page */
1135 if (last_page+1 > last_free_page) {
1136 last_free_page = last_page+1;
1137 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1139 ret = thread_mutex_unlock(&free_pages_lock);
1140 gc_assert(ret == 0);
1142 #ifdef READ_PROTECT_FREE_PAGES
1143 os_protect(page_address(first_page),
1144 npage_bytes(1+last_page-first_page),
1148 zero_dirty_pages(first_page, last_page);
1150 return page_address(first_page);
1153 static page_index_t gencgc_alloc_start_page = -1;
1156 gc_heap_exhausted_error_or_lose (long available, long requested)
1158 struct thread *thread = arch_os_get_current_thread();
1159 /* Write basic information before doing anything else: if we don't
1160 * call to lisp this is a must, and even if we do there is always
1161 * the danger that we bounce back here before the error has been
1162 * handled, or indeed even printed.
1164 fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
1165 gc_active_p ? "garbage collection" : "allocation",
1166 available, requested);
1167 if (gc_active_p || (available == 0)) {
1168 /* If we are in GC, or totally out of memory there is no way
1169 * to sanely transfer control to the lisp-side of things.
1171 print_generation_stats(1);
1172 fprintf(stderr, "GC control variables:\n");
1173 fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
1174 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
1175 SymbolValue(GC_PENDING,thread)==NIL ? "false" : "true");
1176 #ifdef LISP_FEATURE_SB_THREAD
1177 fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n",
1178 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
1180 lose("Heap exhausted, game over.");
1183 /* FIXME: assert free_pages_lock held */
1184 (void)thread_mutex_unlock(&free_pages_lock);
1185 gc_assert(get_pseudo_atomic_atomic(thread));
1186 clear_pseudo_atomic_atomic(thread);
1187 if (get_pseudo_atomic_interrupted(thread))
1188 do_pending_interrupt();
1189 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1190 * to running user code at arbitrary places, even in a
1191 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1192 * running out of the heap. So at this point all bets are
1194 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1195 corruption_warning_and_maybe_lose
1196 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1197 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1198 alloc_number(available), alloc_number(requested));
1199 lose("HEAP-EXHAUSTED-ERROR fell through");
1204 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes,
1207 page_index_t first_page, last_page;
1208 page_index_t restart_page = *restart_page_ptr;
1209 long bytes_found = 0;
1210 long most_bytes_found = 0;
1211 /* FIXME: assert(free_pages_lock is held); */
1213 /* Toggled by gc_and_save for heap compaction, normally -1. */
1214 if (gencgc_alloc_start_page != -1) {
1215 restart_page = gencgc_alloc_start_page;
1218 gc_assert(nbytes>=0);
1219 if (((unsigned long)nbytes)>=PAGE_BYTES) {
1220 /* Search for a contiguous free space of at least nbytes,
1221 * aligned on a page boundary. The page-alignment is strictly
1222 * speaking needed only for objects at least large_object_size
1225 first_page = restart_page;
1226 while ((first_page < page_table_pages) &&
1227 page_allocated_p(first_page))
1230 last_page = first_page;
1231 bytes_found = PAGE_BYTES;
1232 while ((bytes_found < nbytes) &&
1233 (last_page < (page_table_pages-1)) &&
1234 page_free_p(last_page+1)) {
1236 bytes_found += PAGE_BYTES;
1237 gc_assert(0 == page_table[last_page].bytes_used);
1238 gc_assert(0 == page_table[last_page].write_protected);
1240 if (bytes_found > most_bytes_found)
1241 most_bytes_found = bytes_found;
1242 restart_page = last_page + 1;
1243 } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
1246 /* Search for a page with at least nbytes of space. We prefer
1247 * not to split small objects on multiple pages, to reduce the
1248 * number of contiguous allocation regions spaning multiple
1249 * pages: this helps avoid excessive conservativism. */
1250 first_page = restart_page;
1251 while (first_page < page_table_pages) {
1252 if (page_free_p(first_page))
1254 gc_assert(0 == page_table[first_page].bytes_used);
1255 bytes_found = PAGE_BYTES;
1258 else if ((page_table[first_page].allocated == page_type_flag) &&
1259 (page_table[first_page].large_object == 0) &&
1260 (page_table[first_page].gen == gc_alloc_generation) &&
1261 (page_table[first_page].write_protected == 0) &&
1262 (page_table[first_page].dont_move == 0))
1264 bytes_found = PAGE_BYTES
1265 - page_table[first_page].bytes_used;
1266 if (bytes_found > most_bytes_found)
1267 most_bytes_found = bytes_found;
1268 if (bytes_found >= nbytes)
1273 last_page = first_page;
1274 restart_page = first_page + 1;
1277 /* Check for a failure */
1278 if (bytes_found < nbytes) {
1279 gc_assert(restart_page >= page_table_pages);
1280 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1283 gc_assert(page_table[first_page].write_protected == 0);
1285 *restart_page_ptr = first_page;
1289 /* Allocate bytes. All the rest of the special-purpose allocation
1290 * functions will eventually call this */
1293 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1296 void *new_free_pointer;
1298 if (nbytes>=large_object_size)
1299 return gc_alloc_large(nbytes, page_type_flag, my_region);
1301 /* Check whether there is room in the current alloc region. */
1302 new_free_pointer = my_region->free_pointer + nbytes;
1304 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1305 my_region->free_pointer, new_free_pointer); */
1307 if (new_free_pointer <= my_region->end_addr) {
1308 /* If so then allocate from the current alloc region. */
1309 void *new_obj = my_region->free_pointer;
1310 my_region->free_pointer = new_free_pointer;
1312 /* Unless a `quick' alloc was requested, check whether the
1313 alloc region is almost empty. */
1315 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1316 /* If so, finished with the current region. */
1317 gc_alloc_update_page_tables(page_type_flag, my_region);
1318 /* Set up a new region. */
1319 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1322 return((void *)new_obj);
1325 /* Else not enough free space in the current region: retry with a
1328 gc_alloc_update_page_tables(page_type_flag, my_region);
1329 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1330 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1333 /* these are only used during GC: all allocation from the mutator calls
1334 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1337 static inline void *
1338 gc_quick_alloc(long nbytes)
1340 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1343 static inline void *
1344 gc_quick_alloc_large(long nbytes)
1346 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1349 static inline void *
1350 gc_alloc_unboxed(long nbytes)
1352 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1355 static inline void *
1356 gc_quick_alloc_unboxed(long nbytes)
1358 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1361 static inline void *
1362 gc_quick_alloc_large_unboxed(long nbytes)
1364 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1368 /* Copy a large boxed object. If the object is in a large object
1369 * region then it is simply promoted, else it is copied. If it's large
1370 * enough then it's copied to a large object region.
1372 * Vectors may have shrunk. If the object is not copied the space
1373 * needs to be reclaimed, and the page_tables corrected. */
1375 copy_large_object(lispobj object, long nwords)
1379 page_index_t first_page;
1381 gc_assert(is_lisp_pointer(object));
1382 gc_assert(from_space_p(object));
1383 gc_assert((nwords & 0x01) == 0);
1386 /* Check whether it's in a large object region. */
1387 first_page = find_page_index((void *)object);
1388 gc_assert(first_page >= 0);
1390 if (page_table[first_page].large_object) {
1392 /* Promote the object. */
1394 unsigned long remaining_bytes;
1395 page_index_t next_page;
1396 unsigned long bytes_freed;
1397 unsigned long old_bytes_used;
1399 /* Note: Any page write-protection must be removed, else a
1400 * later scavenge_newspace may incorrectly not scavenge these
1401 * pages. This would not be necessary if they are added to the
1402 * new areas, but let's do it for them all (they'll probably
1403 * be written anyway?). */
1405 gc_assert(page_table[first_page].region_start_offset == 0);
1407 next_page = first_page;
1408 remaining_bytes = nwords*N_WORD_BYTES;
1409 while (remaining_bytes > PAGE_BYTES) {
1410 gc_assert(page_table[next_page].gen == from_space);
1411 gc_assert(page_boxed_p(next_page));
1412 gc_assert(page_table[next_page].large_object);
1413 gc_assert(page_table[next_page].region_start_offset ==
1414 npage_bytes(next_page-first_page));
1415 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1417 page_table[next_page].gen = new_space;
1419 /* Remove any write-protection. We should be able to rely
1420 * on the write-protect flag to avoid redundant calls. */
1421 if (page_table[next_page].write_protected) {
1422 os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL);
1423 page_table[next_page].write_protected = 0;
1425 remaining_bytes -= PAGE_BYTES;
1429 /* Now only one page remains, but the object may have shrunk
1430 * so there may be more unused pages which will be freed. */
1432 /* The object may have shrunk but shouldn't have grown. */
1433 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1435 page_table[next_page].gen = new_space;
1436 gc_assert(page_boxed_p(next_page));
1438 /* Adjust the bytes_used. */
1439 old_bytes_used = page_table[next_page].bytes_used;
1440 page_table[next_page].bytes_used = remaining_bytes;
1442 bytes_freed = old_bytes_used - remaining_bytes;
1444 /* Free any remaining pages; needs care. */
1446 while ((old_bytes_used == PAGE_BYTES) &&
1447 (page_table[next_page].gen == from_space) &&
1448 page_boxed_p(next_page) &&
1449 page_table[next_page].large_object &&
1450 (page_table[next_page].region_start_offset ==
1451 npage_bytes(next_page - first_page))) {
1452 /* Checks out OK, free the page. Don't need to bother zeroing
1453 * pages as this should have been done before shrinking the
1454 * object. These pages shouldn't be write-protected as they
1455 * should be zero filled. */
1456 gc_assert(page_table[next_page].write_protected == 0);
1458 old_bytes_used = page_table[next_page].bytes_used;
1459 page_table[next_page].allocated = FREE_PAGE_FLAG;
1460 page_table[next_page].bytes_used = 0;
1461 bytes_freed += old_bytes_used;
1465 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1467 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1468 bytes_allocated -= bytes_freed;
1470 /* Add the region to the new_areas if requested. */
1471 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1475 /* Get tag of object. */
1476 tag = lowtag_of(object);
1478 /* Allocate space. */
1479 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1481 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1483 /* Return Lisp pointer of new object. */
1484 return ((lispobj) new) | tag;
1488 /* to copy unboxed objects */
1490 copy_unboxed_object(lispobj object, long nwords)
1495 gc_assert(is_lisp_pointer(object));
1496 gc_assert(from_space_p(object));
1497 gc_assert((nwords & 0x01) == 0);
1499 /* Get tag of object. */
1500 tag = lowtag_of(object);
1502 /* Allocate space. */
1503 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1505 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1507 /* Return Lisp pointer of new object. */
1508 return ((lispobj) new) | tag;
1511 /* to copy large unboxed objects
1513 * If the object is in a large object region then it is simply
1514 * promoted, else it is copied. If it's large enough then it's copied
1515 * to a large object region.
1517 * Bignums and vectors may have shrunk. If the object is not copied
1518 * the space needs to be reclaimed, and the page_tables corrected.
1520 * KLUDGE: There's a lot of cut-and-paste duplication between this
1521 * function and copy_large_object(..). -- WHN 20000619 */
1523 copy_large_unboxed_object(lispobj object, long nwords)
1527 page_index_t first_page;
1529 gc_assert(is_lisp_pointer(object));
1530 gc_assert(from_space_p(object));
1531 gc_assert((nwords & 0x01) == 0);
1533 if ((nwords > 1024*1024) && gencgc_verbose) {
1534 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1535 nwords*N_WORD_BYTES));
1538 /* Check whether it's a large object. */
1539 first_page = find_page_index((void *)object);
1540 gc_assert(first_page >= 0);
1542 if (page_table[first_page].large_object) {
1543 /* Promote the object. Note: Unboxed objects may have been
1544 * allocated to a BOXED region so it may be necessary to
1545 * change the region to UNBOXED. */
1546 unsigned long remaining_bytes;
1547 page_index_t next_page;
1548 unsigned long bytes_freed;
1549 unsigned long old_bytes_used;
1551 gc_assert(page_table[first_page].region_start_offset == 0);
1553 next_page = first_page;
1554 remaining_bytes = nwords*N_WORD_BYTES;
1555 while (remaining_bytes > PAGE_BYTES) {
1556 gc_assert(page_table[next_page].gen == from_space);
1557 gc_assert(page_allocated_no_region_p(next_page));
1558 gc_assert(page_table[next_page].large_object);
1559 gc_assert(page_table[next_page].region_start_offset ==
1560 npage_bytes(next_page-first_page));
1561 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1563 page_table[next_page].gen = new_space;
1564 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1565 remaining_bytes -= PAGE_BYTES;
1569 /* Now only one page remains, but the object may have shrunk so
1570 * there may be more unused pages which will be freed. */
1572 /* Object may have shrunk but shouldn't have grown - check. */
1573 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1575 page_table[next_page].gen = new_space;
1576 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1578 /* Adjust the bytes_used. */
1579 old_bytes_used = page_table[next_page].bytes_used;
1580 page_table[next_page].bytes_used = remaining_bytes;
1582 bytes_freed = old_bytes_used - remaining_bytes;
1584 /* Free any remaining pages; needs care. */
1586 while ((old_bytes_used == PAGE_BYTES) &&
1587 (page_table[next_page].gen == from_space) &&
1588 page_allocated_no_region_p(next_page) &&
1589 page_table[next_page].large_object &&
1590 (page_table[next_page].region_start_offset ==
1591 npage_bytes(next_page - first_page))) {
1592 /* Checks out OK, free the page. Don't need to both zeroing
1593 * pages as this should have been done before shrinking the
1594 * object. These pages shouldn't be write-protected, even if
1595 * boxed they should be zero filled. */
1596 gc_assert(page_table[next_page].write_protected == 0);
1598 old_bytes_used = page_table[next_page].bytes_used;
1599 page_table[next_page].allocated = FREE_PAGE_FLAG;
1600 page_table[next_page].bytes_used = 0;
1601 bytes_freed += old_bytes_used;
1605 if ((bytes_freed > 0) && gencgc_verbose) {
1607 "/copy_large_unboxed bytes_freed=%d\n",
1611 generations[from_space].bytes_allocated -=
1612 nwords*N_WORD_BYTES + bytes_freed;
1613 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1614 bytes_allocated -= bytes_freed;
1619 /* Get tag of object. */
1620 tag = lowtag_of(object);
1622 /* Allocate space. */
1623 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1625 /* Copy the object. */
1626 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1628 /* Return Lisp pointer of new object. */
1629 return ((lispobj) new) | tag;
1638 * code and code-related objects
1641 static lispobj trans_fun_header(lispobj object);
1642 static lispobj trans_boxed(lispobj object);
1645 /* Scan a x86 compiled code object, looking for possible fixups that
1646 * have been missed after a move.
1648 * Two types of fixups are needed:
1649 * 1. Absolute fixups to within the code object.
1650 * 2. Relative fixups to outside the code object.
1652 * Currently only absolute fixups to the constant vector, or to the
1653 * code area are checked. */
1655 sniff_code_object(struct code *code, unsigned long displacement)
1657 #ifdef LISP_FEATURE_X86
1658 long nheader_words, ncode_words, nwords;
1660 void *constants_start_addr = NULL, *constants_end_addr;
1661 void *code_start_addr, *code_end_addr;
1662 int fixup_found = 0;
1664 if (!check_code_fixups)
1667 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1669 ncode_words = fixnum_value(code->code_size);
1670 nheader_words = HeaderValue(*(lispobj *)code);
1671 nwords = ncode_words + nheader_words;
1673 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1674 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1675 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1676 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1678 /* Work through the unboxed code. */
1679 for (p = code_start_addr; p < code_end_addr; p++) {
1680 void *data = *(void **)p;
1681 unsigned d1 = *((unsigned char *)p - 1);
1682 unsigned d2 = *((unsigned char *)p - 2);
1683 unsigned d3 = *((unsigned char *)p - 3);
1684 unsigned d4 = *((unsigned char *)p - 4);
1686 unsigned d5 = *((unsigned char *)p - 5);
1687 unsigned d6 = *((unsigned char *)p - 6);
1690 /* Check for code references. */
1691 /* Check for a 32 bit word that looks like an absolute
1692 reference to within the code adea of the code object. */
1693 if ((data >= (code_start_addr-displacement))
1694 && (data < (code_end_addr-displacement))) {
1695 /* function header */
1697 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1699 /* Skip the function header */
1703 /* the case of PUSH imm32 */
1707 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1708 p, d6, d5, d4, d3, d2, d1, data));
1709 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1711 /* the case of MOV [reg-8],imm32 */
1713 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1714 || d2==0x45 || d2==0x46 || d2==0x47)
1718 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1719 p, d6, d5, d4, d3, d2, d1, data));
1720 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1722 /* the case of LEA reg,[disp32] */
1723 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1726 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1727 p, d6, d5, d4, d3, d2, d1, data));
1728 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1732 /* Check for constant references. */
1733 /* Check for a 32 bit word that looks like an absolute
1734 reference to within the constant vector. Constant references
1736 if ((data >= (constants_start_addr-displacement))
1737 && (data < (constants_end_addr-displacement))
1738 && (((unsigned)data & 0x3) == 0)) {
1743 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1744 p, d6, d5, d4, d3, d2, d1, data));
1745 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1748 /* the case of MOV m32,EAX */
1752 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1753 p, d6, d5, d4, d3, d2, d1, data));
1754 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1757 /* the case of CMP m32,imm32 */
1758 if ((d1 == 0x3d) && (d2 == 0x81)) {
1761 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1762 p, d6, d5, d4, d3, d2, d1, data));
1764 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1767 /* Check for a mod=00, r/m=101 byte. */
1768 if ((d1 & 0xc7) == 5) {
1773 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1774 p, d6, d5, d4, d3, d2, d1, data));
1775 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1777 /* the case of CMP reg32,m32 */
1781 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1782 p, d6, d5, d4, d3, d2, d1, data));
1783 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1785 /* the case of MOV m32,reg32 */
1789 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1790 p, d6, d5, d4, d3, d2, d1, data));
1791 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1793 /* the case of MOV reg32,m32 */
1797 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1798 p, d6, d5, d4, d3, d2, d1, data));
1799 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1801 /* the case of LEA reg32,m32 */
1805 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1806 p, d6, d5, d4, d3, d2, d1, data));
1807 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1813 /* If anything was found, print some information on the code
1817 "/compiled code object at %x: header words = %d, code words = %d\n",
1818 code, nheader_words, ncode_words));
1820 "/const start = %x, end = %x\n",
1821 constants_start_addr, constants_end_addr));
1823 "/code start = %x, end = %x\n",
1824 code_start_addr, code_end_addr));
1830 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1832 /* x86-64 uses pc-relative addressing instead of this kludge */
1833 #ifndef LISP_FEATURE_X86_64
1834 long nheader_words, ncode_words, nwords;
1835 void *constants_start_addr, *constants_end_addr;
1836 void *code_start_addr, *code_end_addr;
1837 lispobj fixups = NIL;
1838 unsigned long displacement =
1839 (unsigned long)new_code - (unsigned long)old_code;
1840 struct vector *fixups_vector;
1842 ncode_words = fixnum_value(new_code->code_size);
1843 nheader_words = HeaderValue(*(lispobj *)new_code);
1844 nwords = ncode_words + nheader_words;
1846 "/compiled code object at %x: header words = %d, code words = %d\n",
1847 new_code, nheader_words, ncode_words)); */
1848 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1849 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1850 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1851 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1854 "/const start = %x, end = %x\n",
1855 constants_start_addr,constants_end_addr));
1857 "/code start = %x; end = %x\n",
1858 code_start_addr,code_end_addr));
1861 /* The first constant should be a pointer to the fixups for this
1862 code objects. Check. */
1863 fixups = new_code->constants[0];
1865 /* It will be 0 or the unbound-marker if there are no fixups (as
1866 * will be the case if the code object has been purified, for
1867 * example) and will be an other pointer if it is valid. */
1868 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1869 !is_lisp_pointer(fixups)) {
1870 /* Check for possible errors. */
1871 if (check_code_fixups)
1872 sniff_code_object(new_code, displacement);
1877 fixups_vector = (struct vector *)native_pointer(fixups);
1879 /* Could be pointing to a forwarding pointer. */
1880 /* FIXME is this always in from_space? if so, could replace this code with
1881 * forwarding_pointer_p/forwarding_pointer_value */
1882 if (is_lisp_pointer(fixups) &&
1883 (find_page_index((void*)fixups_vector) != -1) &&
1884 (fixups_vector->header == 0x01)) {
1885 /* If so, then follow it. */
1886 /*SHOW("following pointer to a forwarding pointer");*/
1888 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1891 /*SHOW("got fixups");*/
1893 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1894 /* Got the fixups for the code block. Now work through the vector,
1895 and apply a fixup at each address. */
1896 long length = fixnum_value(fixups_vector->length);
1898 for (i = 0; i < length; i++) {
1899 unsigned long offset = fixups_vector->data[i];
1900 /* Now check the current value of offset. */
1901 unsigned long old_value =
1902 *(unsigned long *)((unsigned long)code_start_addr + offset);
1904 /* If it's within the old_code object then it must be an
1905 * absolute fixup (relative ones are not saved) */
1906 if ((old_value >= (unsigned long)old_code)
1907 && (old_value < ((unsigned long)old_code
1908 + nwords*N_WORD_BYTES)))
1909 /* So add the dispacement. */
1910 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1911 old_value + displacement;
1913 /* It is outside the old code object so it must be a
1914 * relative fixup (absolute fixups are not saved). So
1915 * subtract the displacement. */
1916 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1917 old_value - displacement;
1920 /* This used to just print a note to stderr, but a bogus fixup seems to
1921 * indicate real heap corruption, so a hard hailure is in order. */
1922 lose("fixup vector %p has a bad widetag: %d\n",
1923 fixups_vector, widetag_of(fixups_vector->header));
1926 /* Check for possible errors. */
1927 if (check_code_fixups) {
1928 sniff_code_object(new_code,displacement);
1935 trans_boxed_large(lispobj object)
1938 unsigned long length;
1940 gc_assert(is_lisp_pointer(object));
1942 header = *((lispobj *) native_pointer(object));
1943 length = HeaderValue(header) + 1;
1944 length = CEILING(length, 2);
1946 return copy_large_object(object, length);
1949 /* Doesn't seem to be used, delete it after the grace period. */
1952 trans_unboxed_large(lispobj object)
1955 unsigned long length;
1957 gc_assert(is_lisp_pointer(object));
1959 header = *((lispobj *) native_pointer(object));
1960 length = HeaderValue(header) + 1;
1961 length = CEILING(length, 2);
1963 return copy_large_unboxed_object(object, length);
1969 * Lutexes. Using the normal finalization machinery for finalizing
1970 * lutexes is tricky, since the finalization depends on working lutexes.
1971 * So we track the lutexes in the GC and finalize them manually.
1974 #if defined(LUTEX_WIDETAG)
1977 * Start tracking LUTEX in the GC, by adding it to the linked list of
1978 * lutexes in the nursery generation. The caller is responsible for
1979 * locking, and GCs must be inhibited until the registration is
1983 gencgc_register_lutex (struct lutex *lutex) {
1984 int index = find_page_index(lutex);
1985 generation_index_t gen;
1988 /* This lutex is in static space, so we don't need to worry about
1994 gen = page_table[index].gen;
1996 gc_assert(gen >= 0);
1997 gc_assert(gen < NUM_GENERATIONS);
1999 head = generations[gen].lutexes;
2006 generations[gen].lutexes = lutex;
2010 * Stop tracking LUTEX in the GC by removing it from the appropriate
2011 * linked lists. This will only be called during GC, so no locking is
2015 gencgc_unregister_lutex (struct lutex *lutex) {
2017 lutex->prev->next = lutex->next;
2019 generations[lutex->gen].lutexes = lutex->next;
2023 lutex->next->prev = lutex->prev;
2032 * Mark all lutexes in generation GEN as not live.
2035 unmark_lutexes (generation_index_t gen) {
2036 struct lutex *lutex = generations[gen].lutexes;
2040 lutex = lutex->next;
2045 * Finalize all lutexes in generation GEN that have not been marked live.
2048 reap_lutexes (generation_index_t gen) {
2049 struct lutex *lutex = generations[gen].lutexes;
2052 struct lutex *next = lutex->next;
2054 lutex_destroy((tagged_lutex_t) lutex);
2055 gencgc_unregister_lutex(lutex);
2062 * Mark LUTEX as live.
2065 mark_lutex (lispobj tagged_lutex) {
2066 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
2072 * Move all lutexes in generation FROM to generation TO.
2075 move_lutexes (generation_index_t from, generation_index_t to) {
2076 struct lutex *tail = generations[from].lutexes;
2078 /* Nothing to move */
2082 /* Change the generation of the lutexes in FROM. */
2083 while (tail->next) {
2089 /* Link the last lutex in the FROM list to the start of the TO list */
2090 tail->next = generations[to].lutexes;
2092 /* And vice versa */
2093 if (generations[to].lutexes) {
2094 generations[to].lutexes->prev = tail;
2097 /* And update the generations structures to match this */
2098 generations[to].lutexes = generations[from].lutexes;
2099 generations[from].lutexes = NULL;
2103 scav_lutex(lispobj *where, lispobj object)
2105 mark_lutex((lispobj) where);
2107 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2111 trans_lutex(lispobj object)
2113 struct lutex *lutex = (struct lutex *) native_pointer(object);
2115 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2116 gc_assert(is_lisp_pointer(object));
2117 copied = copy_object(object, words);
2119 /* Update the links, since the lutex moved in memory. */
2121 lutex->next->prev = (struct lutex *) native_pointer(copied);
2125 lutex->prev->next = (struct lutex *) native_pointer(copied);
2127 generations[lutex->gen].lutexes =
2128 (struct lutex *) native_pointer(copied);
2135 size_lutex(lispobj *where)
2137 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2139 #endif /* LUTEX_WIDETAG */
2146 /* XX This is a hack adapted from cgc.c. These don't work too
2147 * efficiently with the gencgc as a list of the weak pointers is
2148 * maintained within the objects which causes writes to the pages. A
2149 * limited attempt is made to avoid unnecessary writes, but this needs
2151 #define WEAK_POINTER_NWORDS \
2152 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2155 scav_weak_pointer(lispobj *where, lispobj object)
2157 /* Since we overwrite the 'next' field, we have to make
2158 * sure not to do so for pointers already in the list.
2159 * Instead of searching the list of weak_pointers each
2160 * time, we ensure that next is always NULL when the weak
2161 * pointer isn't in the list, and not NULL otherwise.
2162 * Since we can't use NULL to denote end of list, we
2163 * use a pointer back to the same weak_pointer.
2165 struct weak_pointer * wp = (struct weak_pointer*)where;
2167 if (NULL == wp->next) {
2168 wp->next = weak_pointers;
2170 if (NULL == wp->next)
2174 /* Do not let GC scavenge the value slot of the weak pointer.
2175 * (That is why it is a weak pointer.) */
2177 return WEAK_POINTER_NWORDS;
2182 search_read_only_space(void *pointer)
2184 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2185 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2186 if ((pointer < (void *)start) || (pointer >= (void *)end))
2188 return (gc_search_space(start,
2189 (((lispobj *)pointer)+2)-start,
2190 (lispobj *) pointer));
2194 search_static_space(void *pointer)
2196 lispobj *start = (lispobj *)STATIC_SPACE_START;
2197 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2198 if ((pointer < (void *)start) || (pointer >= (void *)end))
2200 return (gc_search_space(start,
2201 (((lispobj *)pointer)+2)-start,
2202 (lispobj *) pointer));
2205 /* a faster version for searching the dynamic space. This will work even
2206 * if the object is in a current allocation region. */
2208 search_dynamic_space(void *pointer)
2210 page_index_t page_index = find_page_index(pointer);
2213 /* The address may be invalid, so do some checks. */
2214 if ((page_index == -1) || page_free_p(page_index))
2216 start = (lispobj *)page_region_start(page_index);
2217 return (gc_search_space(start,
2218 (((lispobj *)pointer)+2)-start,
2219 (lispobj *)pointer));
2222 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2224 /* Helper for valid_lisp_pointer_p and
2225 * possibly_valid_dynamic_space_pointer.
2227 * pointer is the pointer to validate, and start_addr is the address
2228 * of the enclosing object.
2231 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2233 if (!is_lisp_pointer((lispobj)pointer)) {
2237 /* Check that the object pointed to is consistent with the pointer
2239 switch (lowtag_of((lispobj)pointer)) {
2240 case FUN_POINTER_LOWTAG:
2241 /* Start_addr should be the enclosing code object, or a closure
2243 switch (widetag_of(*start_addr)) {
2244 case CODE_HEADER_WIDETAG:
2245 /* This case is probably caught above. */
2247 case CLOSURE_HEADER_WIDETAG:
2248 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2249 if ((unsigned long)pointer !=
2250 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2251 if (gencgc_verbose) {
2254 pointer, start_addr, *start_addr));
2260 if (gencgc_verbose) {
2263 pointer, start_addr, *start_addr));
2268 case LIST_POINTER_LOWTAG:
2269 if ((unsigned long)pointer !=
2270 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2271 if (gencgc_verbose) {
2274 pointer, start_addr, *start_addr));
2278 /* Is it plausible cons? */
2279 if ((is_lisp_pointer(start_addr[0]) ||
2280 is_lisp_immediate(start_addr[0])) &&
2281 (is_lisp_pointer(start_addr[1]) ||
2282 is_lisp_immediate(start_addr[1])))
2285 if (gencgc_verbose) {
2288 pointer, start_addr, *start_addr));
2292 case INSTANCE_POINTER_LOWTAG:
2293 if ((unsigned long)pointer !=
2294 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2295 if (gencgc_verbose) {
2298 pointer, start_addr, *start_addr));
2302 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2303 if (gencgc_verbose) {
2306 pointer, start_addr, *start_addr));
2311 case OTHER_POINTER_LOWTAG:
2312 if ((unsigned long)pointer !=
2313 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2314 if (gencgc_verbose) {
2317 pointer, start_addr, *start_addr));
2321 /* Is it plausible? Not a cons. XXX should check the headers. */
2322 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2323 if (gencgc_verbose) {
2326 pointer, start_addr, *start_addr));
2330 switch (widetag_of(start_addr[0])) {
2331 case UNBOUND_MARKER_WIDETAG:
2332 case NO_TLS_VALUE_MARKER_WIDETAG:
2333 case CHARACTER_WIDETAG:
2334 #if N_WORD_BITS == 64
2335 case SINGLE_FLOAT_WIDETAG:
2337 if (gencgc_verbose) {
2340 pointer, start_addr, *start_addr));
2344 /* only pointed to by function pointers? */
2345 case CLOSURE_HEADER_WIDETAG:
2346 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2347 if (gencgc_verbose) {
2350 pointer, start_addr, *start_addr));
2354 case INSTANCE_HEADER_WIDETAG:
2355 if (gencgc_verbose) {
2358 pointer, start_addr, *start_addr));
2362 /* the valid other immediate pointer objects */
2363 case SIMPLE_VECTOR_WIDETAG:
2365 case COMPLEX_WIDETAG:
2366 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2367 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2369 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2370 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2372 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2373 case COMPLEX_LONG_FLOAT_WIDETAG:
2375 case SIMPLE_ARRAY_WIDETAG:
2376 case COMPLEX_BASE_STRING_WIDETAG:
2377 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2378 case COMPLEX_CHARACTER_STRING_WIDETAG:
2380 case COMPLEX_VECTOR_NIL_WIDETAG:
2381 case COMPLEX_BIT_VECTOR_WIDETAG:
2382 case COMPLEX_VECTOR_WIDETAG:
2383 case COMPLEX_ARRAY_WIDETAG:
2384 case VALUE_CELL_HEADER_WIDETAG:
2385 case SYMBOL_HEADER_WIDETAG:
2387 case CODE_HEADER_WIDETAG:
2388 case BIGNUM_WIDETAG:
2389 #if N_WORD_BITS != 64
2390 case SINGLE_FLOAT_WIDETAG:
2392 case DOUBLE_FLOAT_WIDETAG:
2393 #ifdef LONG_FLOAT_WIDETAG
2394 case LONG_FLOAT_WIDETAG:
2396 case SIMPLE_BASE_STRING_WIDETAG:
2397 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2398 case SIMPLE_CHARACTER_STRING_WIDETAG:
2400 case SIMPLE_BIT_VECTOR_WIDETAG:
2401 case SIMPLE_ARRAY_NIL_WIDETAG:
2402 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2403 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2404 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2405 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2406 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2407 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2408 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2409 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2411 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2412 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2413 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2414 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2416 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2417 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2419 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2420 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2422 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2423 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2425 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2426 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2428 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2429 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2431 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2432 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2434 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2435 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2437 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2438 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2440 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2441 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2442 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2443 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2445 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2446 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2448 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2449 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2451 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2452 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2455 case WEAK_POINTER_WIDETAG:
2456 #ifdef LUTEX_WIDETAG
2462 if (gencgc_verbose) {
2465 pointer, start_addr, *start_addr));
2471 if (gencgc_verbose) {
2474 pointer, start_addr, *start_addr));
2483 /* Used by the debugger to validate possibly bogus pointers before
2484 * calling MAKE-LISP-OBJ on them.
2486 * FIXME: We would like to make this perfect, because if the debugger
2487 * constructs a reference to a bugs lisp object, and it ends up in a
2488 * location scavenged by the GC all hell breaks loose.
2490 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2491 * and return true for all valid pointers, this could actually be eager
2492 * and lie about a few pointers without bad results... but that should
2493 * be reflected in the name.
2496 valid_lisp_pointer_p(lispobj *pointer)
2499 if (((start=search_dynamic_space(pointer))!=NULL) ||
2500 ((start=search_static_space(pointer))!=NULL) ||
2501 ((start=search_read_only_space(pointer))!=NULL))
2502 return looks_like_valid_lisp_pointer_p(pointer, start);
2507 /* Is there any possibility that pointer is a valid Lisp object
2508 * reference, and/or something else (e.g. subroutine call return
2509 * address) which should prevent us from moving the referred-to thing?
2510 * This is called from preserve_pointers() */
2512 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2514 lispobj *start_addr;
2516 /* Find the object start address. */
2517 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2521 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2524 /* Adjust large bignum and vector objects. This will adjust the
2525 * allocated region if the size has shrunk, and move unboxed objects
2526 * into unboxed pages. The pages are not promoted here, and the
2527 * promoted region is not added to the new_regions; this is really
2528 * only designed to be called from preserve_pointer(). Shouldn't fail
2529 * if this is missed, just may delay the moving of objects to unboxed
2530 * pages, and the freeing of pages. */
2532 maybe_adjust_large_object(lispobj *where)
2534 page_index_t first_page;
2535 page_index_t next_page;
2538 unsigned long remaining_bytes;
2539 unsigned long bytes_freed;
2540 unsigned long old_bytes_used;
2544 /* Check whether it's a vector or bignum object. */
2545 switch (widetag_of(where[0])) {
2546 case SIMPLE_VECTOR_WIDETAG:
2547 boxed = BOXED_PAGE_FLAG;
2549 case BIGNUM_WIDETAG:
2550 case SIMPLE_BASE_STRING_WIDETAG:
2551 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2552 case SIMPLE_CHARACTER_STRING_WIDETAG:
2554 case SIMPLE_BIT_VECTOR_WIDETAG:
2555 case SIMPLE_ARRAY_NIL_WIDETAG:
2556 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2557 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2558 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2559 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2560 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2561 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2562 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2563 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2565 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2566 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2567 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2568 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2570 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2571 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2573 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2574 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2576 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2577 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2579 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2580 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2582 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2583 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2585 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2586 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2588 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2589 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2591 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2592 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2594 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2595 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2596 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2597 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2599 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2600 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2602 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2603 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2605 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2606 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2608 boxed = UNBOXED_PAGE_FLAG;
2614 /* Find its current size. */
2615 nwords = (sizetab[widetag_of(where[0])])(where);
2617 first_page = find_page_index((void *)where);
2618 gc_assert(first_page >= 0);
2620 /* Note: Any page write-protection must be removed, else a later
2621 * scavenge_newspace may incorrectly not scavenge these pages.
2622 * This would not be necessary if they are added to the new areas,
2623 * but lets do it for them all (they'll probably be written
2626 gc_assert(page_table[first_page].region_start_offset == 0);
2628 next_page = first_page;
2629 remaining_bytes = nwords*N_WORD_BYTES;
2630 while (remaining_bytes > PAGE_BYTES) {
2631 gc_assert(page_table[next_page].gen == from_space);
2632 gc_assert(page_allocated_no_region_p(next_page));
2633 gc_assert(page_table[next_page].large_object);
2634 gc_assert(page_table[next_page].region_start_offset ==
2635 npage_bytes(next_page-first_page));
2636 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
2638 page_table[next_page].allocated = boxed;
2640 /* Shouldn't be write-protected at this stage. Essential that the
2642 gc_assert(!page_table[next_page].write_protected);
2643 remaining_bytes -= PAGE_BYTES;
2647 /* Now only one page remains, but the object may have shrunk so
2648 * there may be more unused pages which will be freed. */
2650 /* Object may have shrunk but shouldn't have grown - check. */
2651 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2653 page_table[next_page].allocated = boxed;
2654 gc_assert(page_table[next_page].allocated ==
2655 page_table[first_page].allocated);
2657 /* Adjust the bytes_used. */
2658 old_bytes_used = page_table[next_page].bytes_used;
2659 page_table[next_page].bytes_used = remaining_bytes;
2661 bytes_freed = old_bytes_used - remaining_bytes;
2663 /* Free any remaining pages; needs care. */
2665 while ((old_bytes_used == PAGE_BYTES) &&
2666 (page_table[next_page].gen == from_space) &&
2667 page_allocated_no_region_p(next_page) &&
2668 page_table[next_page].large_object &&
2669 (page_table[next_page].region_start_offset ==
2670 npage_bytes(next_page - first_page))) {
2671 /* It checks out OK, free the page. We don't need to both zeroing
2672 * pages as this should have been done before shrinking the
2673 * object. These pages shouldn't be write protected as they
2674 * should be zero filled. */
2675 gc_assert(page_table[next_page].write_protected == 0);
2677 old_bytes_used = page_table[next_page].bytes_used;
2678 page_table[next_page].allocated = FREE_PAGE_FLAG;
2679 page_table[next_page].bytes_used = 0;
2680 bytes_freed += old_bytes_used;
2684 if ((bytes_freed > 0) && gencgc_verbose) {
2686 "/maybe_adjust_large_object() freed %d\n",
2690 generations[from_space].bytes_allocated -= bytes_freed;
2691 bytes_allocated -= bytes_freed;
2696 /* Take a possible pointer to a Lisp object and mark its page in the
2697 * page_table so that it will not be relocated during a GC.
2699 * This involves locating the page it points to, then backing up to
2700 * the start of its region, then marking all pages dont_move from there
2701 * up to the first page that's not full or has a different generation
2703 * It is assumed that all the page static flags have been cleared at
2704 * the start of a GC.
2706 * It is also assumed that the current gc_alloc() region has been
2707 * flushed and the tables updated. */
2710 preserve_pointer(void *addr)
2712 page_index_t addr_page_index = find_page_index(addr);
2713 page_index_t first_page;
2715 unsigned int region_allocation;
2717 /* quick check 1: Address is quite likely to have been invalid. */
2718 if ((addr_page_index == -1)
2719 || page_free_p(addr_page_index)
2720 || (page_table[addr_page_index].bytes_used == 0)
2721 || (page_table[addr_page_index].gen != from_space)
2722 /* Skip if already marked dont_move. */
2723 || (page_table[addr_page_index].dont_move != 0))
2725 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2726 /* (Now that we know that addr_page_index is in range, it's
2727 * safe to index into page_table[] with it.) */
2728 region_allocation = page_table[addr_page_index].allocated;
2730 /* quick check 2: Check the offset within the page.
2733 if (((unsigned long)addr & (PAGE_BYTES - 1)) >
2734 page_table[addr_page_index].bytes_used)
2737 /* Filter out anything which can't be a pointer to a Lisp object
2738 * (or, as a special case which also requires dont_move, a return
2739 * address referring to something in a CodeObject). This is
2740 * expensive but important, since it vastly reduces the
2741 * probability that random garbage will be bogusly interpreted as
2742 * a pointer which prevents a page from moving. */
2743 if (!(code_page_p(addr_page_index)
2744 || (is_lisp_pointer((lispobj)addr) &&
2745 possibly_valid_dynamic_space_pointer(addr))))
2748 /* Find the beginning of the region. Note that there may be
2749 * objects in the region preceding the one that we were passed a
2750 * pointer to: if this is the case, we will write-protect all the
2751 * previous objects' pages too. */
2754 /* I think this'd work just as well, but without the assertions.
2755 * -dan 2004.01.01 */
2756 first_page = find_page_index(page_region_start(addr_page_index))
2758 first_page = addr_page_index;
2759 while (page_table[first_page].region_start_offset != 0) {
2761 /* Do some checks. */
2762 gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
2763 gc_assert(page_table[first_page].gen == from_space);
2764 gc_assert(page_table[first_page].allocated == region_allocation);
2768 /* Adjust any large objects before promotion as they won't be
2769 * copied after promotion. */
2770 if (page_table[first_page].large_object) {
2771 maybe_adjust_large_object(page_address(first_page));
2772 /* If a large object has shrunk then addr may now point to a
2773 * free area in which case it's ignored here. Note it gets
2774 * through the valid pointer test above because the tail looks
2776 if (page_free_p(addr_page_index)
2777 || (page_table[addr_page_index].bytes_used == 0)
2778 /* Check the offset within the page. */
2779 || (((unsigned long)addr & (PAGE_BYTES - 1))
2780 > page_table[addr_page_index].bytes_used)) {
2782 "weird? ignore ptr 0x%x to freed area of large object\n",
2786 /* It may have moved to unboxed pages. */
2787 region_allocation = page_table[first_page].allocated;
2790 /* Now work forward until the end of this contiguous area is found,
2791 * marking all pages as dont_move. */
2792 for (i = first_page; ;i++) {
2793 gc_assert(page_table[i].allocated == region_allocation);
2795 /* Mark the page static. */
2796 page_table[i].dont_move = 1;
2798 /* Move the page to the new_space. XX I'd rather not do this
2799 * but the GC logic is not quite able to copy with the static
2800 * pages remaining in the from space. This also requires the
2801 * generation bytes_allocated counters be updated. */
2802 page_table[i].gen = new_space;
2803 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2804 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2806 /* It is essential that the pages are not write protected as
2807 * they may have pointers into the old-space which need
2808 * scavenging. They shouldn't be write protected at this
2810 gc_assert(!page_table[i].write_protected);
2812 /* Check whether this is the last page in this contiguous block.. */
2813 if ((page_table[i].bytes_used < PAGE_BYTES)
2814 /* ..or it is PAGE_BYTES and is the last in the block */
2816 || (page_table[i+1].bytes_used == 0) /* next page free */
2817 || (page_table[i+1].gen != from_space) /* diff. gen */
2818 || (page_table[i+1].region_start_offset == 0))
2822 /* Check that the page is now static. */
2823 gc_assert(page_table[addr_page_index].dont_move != 0);
2826 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2829 /* If the given page is not write-protected, then scan it for pointers
2830 * to younger generations or the top temp. generation, if no
2831 * suspicious pointers are found then the page is write-protected.
2833 * Care is taken to check for pointers to the current gc_alloc()
2834 * region if it is a younger generation or the temp. generation. This
2835 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2836 * the gc_alloc_generation does not need to be checked as this is only
2837 * called from scavenge_generation() when the gc_alloc generation is
2838 * younger, so it just checks if there is a pointer to the current
2841 * We return 1 if the page was write-protected, else 0. */
2843 update_page_write_prot(page_index_t page)
2845 generation_index_t gen = page_table[page].gen;
2848 void **page_addr = (void **)page_address(page);
2849 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2851 /* Shouldn't be a free page. */
2852 gc_assert(page_allocated_p(page));
2853 gc_assert(page_table[page].bytes_used != 0);
2855 /* Skip if it's already write-protected, pinned, or unboxed */
2856 if (page_table[page].write_protected
2857 /* FIXME: What's the reason for not write-protecting pinned pages? */
2858 || page_table[page].dont_move
2859 || page_unboxed_p(page))
2862 /* Scan the page for pointers to younger generations or the
2863 * top temp. generation. */
2865 for (j = 0; j < num_words; j++) {
2866 void *ptr = *(page_addr+j);
2867 page_index_t index = find_page_index(ptr);
2869 /* Check that it's in the dynamic space */
2871 if (/* Does it point to a younger or the temp. generation? */
2872 (page_allocated_p(index)
2873 && (page_table[index].bytes_used != 0)
2874 && ((page_table[index].gen < gen)
2875 || (page_table[index].gen == SCRATCH_GENERATION)))
2877 /* Or does it point within a current gc_alloc() region? */
2878 || ((boxed_region.start_addr <= ptr)
2879 && (ptr <= boxed_region.free_pointer))
2880 || ((unboxed_region.start_addr <= ptr)
2881 && (ptr <= unboxed_region.free_pointer))) {
2888 /* Write-protect the page. */
2889 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2891 os_protect((void *)page_addr,
2893 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2895 /* Note the page as protected in the page tables. */
2896 page_table[page].write_protected = 1;
2902 /* Scavenge all generations from FROM to TO, inclusive, except for
2903 * new_space which needs special handling, as new objects may be
2904 * added which are not checked here - use scavenge_newspace generation.
2906 * Write-protected pages should not have any pointers to the
2907 * from_space so do need scavenging; thus write-protected pages are
2908 * not always scavenged. There is some code to check that these pages
2909 * are not written; but to check fully the write-protected pages need
2910 * to be scavenged by disabling the code to skip them.
2912 * Under the current scheme when a generation is GCed the younger
2913 * generations will be empty. So, when a generation is being GCed it
2914 * is only necessary to scavenge the older generations for pointers
2915 * not the younger. So a page that does not have pointers to younger
2916 * generations does not need to be scavenged.
2918 * The write-protection can be used to note pages that don't have
2919 * pointers to younger pages. But pages can be written without having
2920 * pointers to younger generations. After the pages are scavenged here
2921 * they can be scanned for pointers to younger generations and if
2922 * there are none the page can be write-protected.
2924 * One complication is when the newspace is the top temp. generation.
2926 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2927 * that none were written, which they shouldn't be as they should have
2928 * no pointers to younger generations. This breaks down for weak
2929 * pointers as the objects contain a link to the next and are written
2930 * if a weak pointer is scavenged. Still it's a useful check. */
2932 scavenge_generations(generation_index_t from, generation_index_t to)
2939 /* Clear the write_protected_cleared flags on all pages. */
2940 for (i = 0; i < page_table_pages; i++)
2941 page_table[i].write_protected_cleared = 0;
2944 for (i = 0; i < last_free_page; i++) {
2945 generation_index_t generation = page_table[i].gen;
2947 && (page_table[i].bytes_used != 0)
2948 && (generation != new_space)
2949 && (generation >= from)
2950 && (generation <= to)) {
2951 page_index_t last_page,j;
2952 int write_protected=1;
2954 /* This should be the start of a region */
2955 gc_assert(page_table[i].region_start_offset == 0);
2957 /* Now work forward until the end of the region */
2958 for (last_page = i; ; last_page++) {
2960 write_protected && page_table[last_page].write_protected;
2961 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2962 /* Or it is PAGE_BYTES and is the last in the block */
2963 || (!page_boxed_p(last_page+1))
2964 || (page_table[last_page+1].bytes_used == 0)
2965 || (page_table[last_page+1].gen != generation)
2966 || (page_table[last_page+1].region_start_offset == 0))
2969 if (!write_protected) {
2970 scavenge(page_address(i),
2971 ((unsigned long)(page_table[last_page].bytes_used
2972 + npage_bytes(last_page-i)))
2975 /* Now scan the pages and write protect those that
2976 * don't have pointers to younger generations. */
2977 if (enable_page_protection) {
2978 for (j = i; j <= last_page; j++) {
2979 num_wp += update_page_write_prot(j);
2982 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2984 "/write protected %d pages within generation %d\n",
2985 num_wp, generation));
2993 /* Check that none of the write_protected pages in this generation
2994 * have been written to. */
2995 for (i = 0; i < page_table_pages; i++) {
2996 if (page_allocated_p(i)
2997 && (page_table[i].bytes_used != 0)
2998 && (page_table[i].gen == generation)
2999 && (page_table[i].write_protected_cleared != 0)) {
3000 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
3002 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
3003 page_table[i].bytes_used,
3004 page_table[i].region_start_offset,
3005 page_table[i].dont_move));
3006 lose("write to protected page %d in scavenge_generation()\n", i);
3013 /* Scavenge a newspace generation. As it is scavenged new objects may
3014 * be allocated to it; these will also need to be scavenged. This
3015 * repeats until there are no more objects unscavenged in the
3016 * newspace generation.
3018 * To help improve the efficiency, areas written are recorded by
3019 * gc_alloc() and only these scavenged. Sometimes a little more will be
3020 * scavenged, but this causes no harm. An easy check is done that the
3021 * scavenged bytes equals the number allocated in the previous
3024 * Write-protected pages are not scanned except if they are marked
3025 * dont_move in which case they may have been promoted and still have
3026 * pointers to the from space.
3028 * Write-protected pages could potentially be written by alloc however
3029 * to avoid having to handle re-scavenging of write-protected pages
3030 * gc_alloc() does not write to write-protected pages.
3032 * New areas of objects allocated are recorded alternatively in the two
3033 * new_areas arrays below. */
3034 static struct new_area new_areas_1[NUM_NEW_AREAS];
3035 static struct new_area new_areas_2[NUM_NEW_AREAS];
3037 /* Do one full scan of the new space generation. This is not enough to
3038 * complete the job as new objects may be added to the generation in
3039 * the process which are not scavenged. */
3041 scavenge_newspace_generation_one_scan(generation_index_t generation)
3046 "/starting one full scan of newspace generation %d\n",
3048 for (i = 0; i < last_free_page; i++) {
3049 /* Note that this skips over open regions when it encounters them. */
3051 && (page_table[i].bytes_used != 0)
3052 && (page_table[i].gen == generation)
3053 && ((page_table[i].write_protected == 0)
3054 /* (This may be redundant as write_protected is now
3055 * cleared before promotion.) */
3056 || (page_table[i].dont_move == 1))) {
3057 page_index_t last_page;
3060 /* The scavenge will start at the region_start_offset of
3063 * We need to find the full extent of this contiguous
3064 * block in case objects span pages.
3066 * Now work forward until the end of this contiguous area
3067 * is found. A small area is preferred as there is a
3068 * better chance of its pages being write-protected. */
3069 for (last_page = i; ;last_page++) {
3070 /* If all pages are write-protected and movable,
3071 * then no need to scavenge */
3072 all_wp=all_wp && page_table[last_page].write_protected &&
3073 !page_table[last_page].dont_move;
3075 /* Check whether this is the last page in this
3076 * contiguous block */
3077 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3078 /* Or it is PAGE_BYTES and is the last in the block */
3079 || (!page_boxed_p(last_page+1))
3080 || (page_table[last_page+1].bytes_used == 0)
3081 || (page_table[last_page+1].gen != generation)
3082 || (page_table[last_page+1].region_start_offset == 0))
3086 /* Do a limited check for write-protected pages. */
3088 long nwords = (((unsigned long)
3089 (page_table[last_page].bytes_used
3090 + npage_bytes(last_page-i)
3091 + page_table[i].region_start_offset))
3093 new_areas_ignore_page = last_page;
3095 scavenge(page_region_start(i), nwords);
3102 "/done with one full scan of newspace generation %d\n",
3106 /* Do a complete scavenge of the newspace generation. */
3108 scavenge_newspace_generation(generation_index_t generation)
3112 /* the new_areas array currently being written to by gc_alloc() */
3113 struct new_area (*current_new_areas)[] = &new_areas_1;
3114 long current_new_areas_index;
3116 /* the new_areas created by the previous scavenge cycle */
3117 struct new_area (*previous_new_areas)[] = NULL;
3118 long previous_new_areas_index;
3120 /* Flush the current regions updating the tables. */
3121 gc_alloc_update_all_page_tables();
3123 /* Turn on the recording of new areas by gc_alloc(). */
3124 new_areas = current_new_areas;
3125 new_areas_index = 0;
3127 /* Don't need to record new areas that get scavenged anyway during
3128 * scavenge_newspace_generation_one_scan. */
3129 record_new_objects = 1;
3131 /* Start with a full scavenge. */
3132 scavenge_newspace_generation_one_scan(generation);
3134 /* Record all new areas now. */
3135 record_new_objects = 2;
3137 /* Give a chance to weak hash tables to make other objects live.
3138 * FIXME: The algorithm implemented here for weak hash table gcing
3139 * is O(W^2+N) as Bruno Haible warns in
3140 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3141 * see "Implementation 2". */
3142 scav_weak_hash_tables();
3144 /* Flush the current regions updating the tables. */
3145 gc_alloc_update_all_page_tables();
3147 /* Grab new_areas_index. */
3148 current_new_areas_index = new_areas_index;
3151 "The first scan is finished; current_new_areas_index=%d.\n",
3152 current_new_areas_index));*/
3154 while (current_new_areas_index > 0) {
3155 /* Move the current to the previous new areas */
3156 previous_new_areas = current_new_areas;
3157 previous_new_areas_index = current_new_areas_index;
3159 /* Scavenge all the areas in previous new areas. Any new areas
3160 * allocated are saved in current_new_areas. */
3162 /* Allocate an array for current_new_areas; alternating between
3163 * new_areas_1 and 2 */
3164 if (previous_new_areas == &new_areas_1)
3165 current_new_areas = &new_areas_2;
3167 current_new_areas = &new_areas_1;
3169 /* Set up for gc_alloc(). */
3170 new_areas = current_new_areas;
3171 new_areas_index = 0;
3173 /* Check whether previous_new_areas had overflowed. */
3174 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3176 /* New areas of objects allocated have been lost so need to do a
3177 * full scan to be sure! If this becomes a problem try
3178 * increasing NUM_NEW_AREAS. */
3179 if (gencgc_verbose) {
3180 SHOW("new_areas overflow, doing full scavenge");
3183 /* Don't need to record new areas that get scavenged
3184 * anyway during scavenge_newspace_generation_one_scan. */
3185 record_new_objects = 1;
3187 scavenge_newspace_generation_one_scan(generation);
3189 /* Record all new areas now. */
3190 record_new_objects = 2;
3192 scav_weak_hash_tables();
3194 /* Flush the current regions updating the tables. */
3195 gc_alloc_update_all_page_tables();
3199 /* Work through previous_new_areas. */
3200 for (i = 0; i < previous_new_areas_index; i++) {
3201 page_index_t page = (*previous_new_areas)[i].page;
3202 size_t offset = (*previous_new_areas)[i].offset;
3203 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3204 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3205 scavenge(page_address(page)+offset, size);
3208 scav_weak_hash_tables();
3210 /* Flush the current regions updating the tables. */
3211 gc_alloc_update_all_page_tables();
3214 current_new_areas_index = new_areas_index;
3217 "The re-scan has finished; current_new_areas_index=%d.\n",
3218 current_new_areas_index));*/
3221 /* Turn off recording of areas allocated by gc_alloc(). */
3222 record_new_objects = 0;
3225 /* Check that none of the write_protected pages in this generation
3226 * have been written to. */
3227 for (i = 0; i < page_table_pages; i++) {
3228 if (page_allocated_p(i)
3229 && (page_table[i].bytes_used != 0)
3230 && (page_table[i].gen == generation)
3231 && (page_table[i].write_protected_cleared != 0)
3232 && (page_table[i].dont_move == 0)) {
3233 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3234 i, generation, page_table[i].dont_move);
3240 /* Un-write-protect all the pages in from_space. This is done at the
3241 * start of a GC else there may be many page faults while scavenging
3242 * the newspace (I've seen drive the system time to 99%). These pages
3243 * would need to be unprotected anyway before unmapping in
3244 * free_oldspace; not sure what effect this has on paging.. */
3246 unprotect_oldspace(void)
3250 for (i = 0; i < last_free_page; i++) {
3251 if (page_allocated_p(i)
3252 && (page_table[i].bytes_used != 0)
3253 && (page_table[i].gen == from_space)) {
3256 page_start = (void *)page_address(i);
3258 /* Remove any write-protection. We should be able to rely
3259 * on the write-protect flag to avoid redundant calls. */
3260 if (page_table[i].write_protected) {
3261 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3262 page_table[i].write_protected = 0;
3268 /* Work through all the pages and free any in from_space. This
3269 * assumes that all objects have been copied or promoted to an older
3270 * generation. Bytes_allocated and the generation bytes_allocated
3271 * counter are updated. The number of bytes freed is returned. */
3272 static unsigned long
3275 unsigned long bytes_freed = 0;
3276 page_index_t first_page, last_page;
3281 /* Find a first page for the next region of pages. */
3282 while ((first_page < last_free_page)
3283 && (page_free_p(first_page)
3284 || (page_table[first_page].bytes_used == 0)
3285 || (page_table[first_page].gen != from_space)))
3288 if (first_page >= last_free_page)
3291 /* Find the last page of this region. */
3292 last_page = first_page;
3295 /* Free the page. */
3296 bytes_freed += page_table[last_page].bytes_used;
3297 generations[page_table[last_page].gen].bytes_allocated -=
3298 page_table[last_page].bytes_used;
3299 page_table[last_page].allocated = FREE_PAGE_FLAG;
3300 page_table[last_page].bytes_used = 0;
3302 /* Remove any write-protection. We should be able to rely
3303 * on the write-protect flag to avoid redundant calls. */
3305 void *page_start = (void *)page_address(last_page);
3307 if (page_table[last_page].write_protected) {
3308 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3309 page_table[last_page].write_protected = 0;
3314 while ((last_page < last_free_page)
3315 && page_allocated_p(last_page)
3316 && (page_table[last_page].bytes_used != 0)
3317 && (page_table[last_page].gen == from_space));
3319 #ifdef READ_PROTECT_FREE_PAGES
3320 os_protect(page_address(first_page),
3321 npage_bytes(last_page-first_page),
3324 first_page = last_page;
3325 } while (first_page < last_free_page);
3327 bytes_allocated -= bytes_freed;
3332 /* Print some information about a pointer at the given address. */
3334 print_ptr(lispobj *addr)
3336 /* If addr is in the dynamic space then out the page information. */
3337 page_index_t pi1 = find_page_index((void*)addr);
3340 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
3341 (unsigned long) addr,
3343 page_table[pi1].allocated,
3344 page_table[pi1].gen,
3345 page_table[pi1].bytes_used,
3346 page_table[pi1].region_start_offset,
3347 page_table[pi1].dont_move);
3348 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3362 verify_space(lispobj *start, size_t words)
3364 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3365 int is_in_readonly_space =
3366 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3367 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3371 lispobj thing = *(lispobj*)start;
3373 if (is_lisp_pointer(thing)) {
3374 page_index_t page_index = find_page_index((void*)thing);
3375 long to_readonly_space =
3376 (READ_ONLY_SPACE_START <= thing &&
3377 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3378 long to_static_space =
3379 (STATIC_SPACE_START <= thing &&
3380 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3382 /* Does it point to the dynamic space? */
3383 if (page_index != -1) {
3384 /* If it's within the dynamic space it should point to a used
3385 * page. XX Could check the offset too. */
3386 if (page_allocated_p(page_index)
3387 && (page_table[page_index].bytes_used == 0))
3388 lose ("Ptr %x @ %x sees free page.\n", thing, start);
3389 /* Check that it doesn't point to a forwarding pointer! */
3390 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3391 lose("Ptr %x @ %x sees forwarding ptr.\n", thing, start);
3393 /* Check that its not in the RO space as it would then be a
3394 * pointer from the RO to the dynamic space. */
3395 if (is_in_readonly_space) {
3396 lose("ptr to dynamic space %x from RO space %x\n",
3399 /* Does it point to a plausible object? This check slows
3400 * it down a lot (so it's commented out).
3402 * "a lot" is serious: it ate 50 minutes cpu time on
3403 * my duron 950 before I came back from lunch and
3406 * FIXME: Add a variable to enable this
3409 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3410 lose("ptr %x to invalid object %x\n", thing, start);
3414 /* Verify that it points to another valid space. */
3415 if (!to_readonly_space && !to_static_space) {
3416 lose("Ptr %x @ %x sees junk.\n", thing, start);
3420 if (!(fixnump(thing))) {
3422 switch(widetag_of(*start)) {
3425 case SIMPLE_VECTOR_WIDETAG:
3427 case COMPLEX_WIDETAG:
3428 case SIMPLE_ARRAY_WIDETAG:
3429 case COMPLEX_BASE_STRING_WIDETAG:
3430 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3431 case COMPLEX_CHARACTER_STRING_WIDETAG:
3433 case COMPLEX_VECTOR_NIL_WIDETAG:
3434 case COMPLEX_BIT_VECTOR_WIDETAG:
3435 case COMPLEX_VECTOR_WIDETAG:
3436 case COMPLEX_ARRAY_WIDETAG:
3437 case CLOSURE_HEADER_WIDETAG:
3438 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3439 case VALUE_CELL_HEADER_WIDETAG:
3440 case SYMBOL_HEADER_WIDETAG:
3441 case CHARACTER_WIDETAG:
3442 #if N_WORD_BITS == 64
3443 case SINGLE_FLOAT_WIDETAG:
3445 case UNBOUND_MARKER_WIDETAG:
3450 case INSTANCE_HEADER_WIDETAG:
3453 long ntotal = HeaderValue(thing);
3454 lispobj layout = ((struct instance *)start)->slots[0];
3459 nuntagged = ((struct layout *)
3460 native_pointer(layout))->n_untagged_slots;
3461 verify_space(start + 1,
3462 ntotal - fixnum_value(nuntagged));
3466 case CODE_HEADER_WIDETAG:
3468 lispobj object = *start;
3470 long nheader_words, ncode_words, nwords;
3472 struct simple_fun *fheaderp;
3474 code = (struct code *) start;
3476 /* Check that it's not in the dynamic space.
3477 * FIXME: Isn't is supposed to be OK for code
3478 * objects to be in the dynamic space these days? */
3479 if (is_in_dynamic_space
3480 /* It's ok if it's byte compiled code. The trace
3481 * table offset will be a fixnum if it's x86
3482 * compiled code - check.
3484 * FIXME: #^#@@! lack of abstraction here..
3485 * This line can probably go away now that
3486 * there's no byte compiler, but I've got
3487 * too much to worry about right now to try
3488 * to make sure. -- WHN 2001-10-06 */
3489 && fixnump(code->trace_table_offset)
3490 /* Only when enabled */
3491 && verify_dynamic_code_check) {
3493 "/code object at %x in the dynamic space\n",
3497 ncode_words = fixnum_value(code->code_size);
3498 nheader_words = HeaderValue(object);
3499 nwords = ncode_words + nheader_words;
3500 nwords = CEILING(nwords, 2);
3501 /* Scavenge the boxed section of the code data block */
3502 verify_space(start + 1, nheader_words - 1);
3504 /* Scavenge the boxed section of each function
3505 * object in the code data block. */
3506 fheaderl = code->entry_points;
3507 while (fheaderl != NIL) {
3509 (struct simple_fun *) native_pointer(fheaderl);
3510 gc_assert(widetag_of(fheaderp->header) ==
3511 SIMPLE_FUN_HEADER_WIDETAG);
3512 verify_space(&fheaderp->name, 1);
3513 verify_space(&fheaderp->arglist, 1);
3514 verify_space(&fheaderp->type, 1);
3515 fheaderl = fheaderp->next;
3521 /* unboxed objects */
3522 case BIGNUM_WIDETAG:
3523 #if N_WORD_BITS != 64
3524 case SINGLE_FLOAT_WIDETAG:
3526 case DOUBLE_FLOAT_WIDETAG:
3527 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3528 case LONG_FLOAT_WIDETAG:
3530 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3531 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3533 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3534 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3536 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3537 case COMPLEX_LONG_FLOAT_WIDETAG:
3539 case SIMPLE_BASE_STRING_WIDETAG:
3540 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3541 case SIMPLE_CHARACTER_STRING_WIDETAG:
3543 case SIMPLE_BIT_VECTOR_WIDETAG:
3544 case SIMPLE_ARRAY_NIL_WIDETAG:
3545 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3546 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3547 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3548 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3549 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3550 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3551 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
3552 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
3554 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3555 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3556 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
3557 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
3559 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3560 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3562 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3563 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3565 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3566 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3568 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3569 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3571 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
3572 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
3574 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3575 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3577 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
3578 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
3580 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3581 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3583 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3584 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3585 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3586 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3588 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3589 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3591 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3592 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3594 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3595 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3598 case WEAK_POINTER_WIDETAG:
3599 #ifdef LUTEX_WIDETAG
3602 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3603 case NO_TLS_VALUE_MARKER_WIDETAG:
3605 count = (sizetab[widetag_of(*start)])(start);
3609 lose("Unhandled widetag 0x%x at 0x%x\n",
3610 widetag_of(*start), start);
3622 /* FIXME: It would be nice to make names consistent so that
3623 * foo_size meant size *in* *bytes* instead of size in some
3624 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3625 * Some counts of lispobjs are called foo_count; it might be good
3626 * to grep for all foo_size and rename the appropriate ones to
3628 long read_only_space_size =
3629 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3630 - (lispobj*)READ_ONLY_SPACE_START;
3631 long static_space_size =
3632 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3633 - (lispobj*)STATIC_SPACE_START;
3635 for_each_thread(th) {
3636 long binding_stack_size =
3637 (lispobj*)get_binding_stack_pointer(th)
3638 - (lispobj*)th->binding_stack_start;
3639 verify_space(th->binding_stack_start, binding_stack_size);
3641 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3642 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3646 verify_generation(generation_index_t generation)
3650 for (i = 0; i < last_free_page; i++) {
3651 if (page_allocated_p(i)
3652 && (page_table[i].bytes_used != 0)
3653 && (page_table[i].gen == generation)) {
3654 page_index_t last_page;
3655 int region_allocation = page_table[i].allocated;
3657 /* This should be the start of a contiguous block */
3658 gc_assert(page_table[i].region_start_offset == 0);
3660 /* Need to find the full extent of this contiguous block in case
3661 objects span pages. */
3663 /* Now work forward until the end of this contiguous area is
3665 for (last_page = i; ;last_page++)
3666 /* Check whether this is the last page in this contiguous
3668 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3669 /* Or it is PAGE_BYTES and is the last in the block */
3670 || (page_table[last_page+1].allocated != region_allocation)
3671 || (page_table[last_page+1].bytes_used == 0)
3672 || (page_table[last_page+1].gen != generation)
3673 || (page_table[last_page+1].region_start_offset == 0))
3676 verify_space(page_address(i),
3678 (page_table[last_page].bytes_used
3679 + npage_bytes(last_page-i)))
3686 /* Check that all the free space is zero filled. */
3688 verify_zero_fill(void)
3692 for (page = 0; page < last_free_page; page++) {
3693 if (page_free_p(page)) {
3694 /* The whole page should be zero filled. */
3695 long *start_addr = (long *)page_address(page);
3698 for (i = 0; i < size; i++) {
3699 if (start_addr[i] != 0) {
3700 lose("free page not zero at %x\n", start_addr + i);
3704 long free_bytes = PAGE_BYTES - page_table[page].bytes_used;
3705 if (free_bytes > 0) {
3706 long *start_addr = (long *)((unsigned long)page_address(page)
3707 + page_table[page].bytes_used);
3708 long size = free_bytes / N_WORD_BYTES;
3710 for (i = 0; i < size; i++) {
3711 if (start_addr[i] != 0) {
3712 lose("free region not zero at %x\n", start_addr + i);
3720 /* External entry point for verify_zero_fill */
3722 gencgc_verify_zero_fill(void)
3724 /* Flush the alloc regions updating the tables. */
3725 gc_alloc_update_all_page_tables();
3726 SHOW("verifying zero fill");
3731 verify_dynamic_space(void)
3733 generation_index_t i;
3735 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3736 verify_generation(i);
3738 if (gencgc_enable_verify_zero_fill)
3742 /* Write-protect all the dynamic boxed pages in the given generation. */
3744 write_protect_generation_pages(generation_index_t generation)
3748 gc_assert(generation < SCRATCH_GENERATION);
3750 for (start = 0; start < last_free_page; start++) {
3751 if (protect_page_p(start, generation)) {
3755 /* Note the page as protected in the page tables. */
3756 page_table[start].write_protected = 1;
3758 for (last = start + 1; last < last_free_page; last++) {
3759 if (!protect_page_p(last, generation))
3761 page_table[last].write_protected = 1;
3764 page_start = (void *)page_address(start);
3766 os_protect(page_start,
3767 npage_bytes(last - start),
3768 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3774 if (gencgc_verbose > 1) {
3776 "/write protected %d of %d pages in generation %d\n",
3777 count_write_protect_generation_pages(generation),
3778 count_generation_pages(generation),
3783 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3786 scavenge_control_stack()
3788 unsigned long control_stack_size;
3790 /* This is going to be a big problem when we try to port threads
3792 struct thread *th = arch_os_get_current_thread();
3793 lispobj *control_stack =
3794 (lispobj *)(th->control_stack_start);
3796 control_stack_size = current_control_stack_pointer - control_stack;
3797 scavenge(control_stack, control_stack_size);
3800 /* Scavenging Interrupt Contexts */
3802 static int boxed_registers[] = BOXED_REGISTERS;
3805 scavenge_interrupt_context(os_context_t * context)
3811 unsigned long lip_offset;
3812 int lip_register_pair;
3814 unsigned long pc_code_offset;
3816 #ifdef ARCH_HAS_LINK_REGISTER
3817 unsigned long lr_code_offset;
3819 #ifdef ARCH_HAS_NPC_REGISTER
3820 unsigned long npc_code_offset;
3824 /* Find the LIP's register pair and calculate it's offset */
3825 /* before we scavenge the context. */
3828 * I (RLT) think this is trying to find the boxed register that is
3829 * closest to the LIP address, without going past it. Usually, it's
3830 * reg_CODE or reg_LRA. But sometimes, nothing can be found.
3832 lip = *os_context_register_addr(context, reg_LIP);
3833 lip_offset = 0x7FFFFFFF;
3834 lip_register_pair = -1;
3835 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3840 index = boxed_registers[i];
3841 reg = *os_context_register_addr(context, index);
3842 if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
3844 if (offset < lip_offset) {
3845 lip_offset = offset;
3846 lip_register_pair = index;
3850 #endif /* reg_LIP */
3852 /* Compute the PC's offset from the start of the CODE */
3854 pc_code_offset = *os_context_pc_addr(context)
3855 - *os_context_register_addr(context, reg_CODE);
3856 #ifdef ARCH_HAS_NPC_REGISTER
3857 npc_code_offset = *os_context_npc_addr(context)
3858 - *os_context_register_addr(context, reg_CODE);
3859 #endif /* ARCH_HAS_NPC_REGISTER */
3861 #ifdef ARCH_HAS_LINK_REGISTER
3863 *os_context_lr_addr(context) -
3864 *os_context_register_addr(context, reg_CODE);
3867 /* Scanvenge all boxed registers in the context. */
3868 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3872 index = boxed_registers[i];
3873 foo = *os_context_register_addr(context, index);
3875 *os_context_register_addr(context, index) = foo;
3877 scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
3884 * But what happens if lip_register_pair is -1?
3885 * *os_context_register_addr on Solaris (see
3886 * solaris_register_address in solaris-os.c) will return
3887 * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
3888 * that what we really want? My guess is that that is not what we
3889 * want, so if lip_register_pair is -1, we don't touch reg_LIP at
3890 * all. But maybe it doesn't really matter if LIP is trashed?
3892 if (lip_register_pair >= 0) {
3893 *os_context_register_addr(context, reg_LIP) =
3894 *os_context_register_addr(context, lip_register_pair)
3897 #endif /* reg_LIP */
3899 /* Fix the PC if it was in from space */
3900 if (from_space_p(*os_context_pc_addr(context)))
3901 *os_context_pc_addr(context) =
3902 *os_context_register_addr(context, reg_CODE) + pc_code_offset;
3904 #ifdef ARCH_HAS_LINK_REGISTER
3905 /* Fix the LR ditto; important if we're being called from
3906 * an assembly routine that expects to return using blr, otherwise
3908 if (from_space_p(*os_context_lr_addr(context)))
3909 *os_context_lr_addr(context) =
3910 *os_context_register_addr(context, reg_CODE) + lr_code_offset;
3913 #ifdef ARCH_HAS_NPC_REGISTER
3914 if (from_space_p(*os_context_npc_addr(context)))
3915 *os_context_npc_addr(context) =
3916 *os_context_register_addr(context, reg_CODE) + npc_code_offset;
3917 #endif /* ARCH_HAS_NPC_REGISTER */
3921 scavenge_interrupt_contexts(void)
3924 os_context_t *context;
3926 struct thread *th=arch_os_get_current_thread();
3928 index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
3930 #if defined(DEBUG_PRINT_CONTEXT_INDEX)
3931 printf("Number of active contexts: %d\n", index);
3934 for (i = 0; i < index; i++) {
3935 context = th->interrupt_contexts[i];
3936 scavenge_interrupt_context(context);
3942 #if defined(LISP_FEATURE_SB_THREAD)
3944 preserve_context_registers (os_context_t *c)
3947 /* On Darwin the signal context isn't a contiguous block of memory,
3948 * so just preserve_pointering its contents won't be sufficient.
3950 #if defined(LISP_FEATURE_DARWIN)
3951 #if defined LISP_FEATURE_X86
3952 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3953 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3954 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3955 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3956 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3957 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3958 preserve_pointer((void*)*os_context_pc_addr(c));
3959 #elif defined LISP_FEATURE_X86_64
3960 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3961 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3962 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3963 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3964 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3965 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3966 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3967 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3968 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3969 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3970 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3971 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3972 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3973 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3974 preserve_pointer((void*)*os_context_pc_addr(c));
3976 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3979 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3980 preserve_pointer(*ptr);
3985 /* Garbage collect a generation. If raise is 0 then the remains of the
3986 * generation are not raised to the next generation. */
3988 garbage_collect_generation(generation_index_t generation, int raise)
3990 unsigned long bytes_freed;
3992 unsigned long static_space_size;
3993 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3996 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3998 /* The oldest generation can't be raised. */
3999 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
4001 /* Check if weak hash tables were processed in the previous GC. */
4002 gc_assert(weak_hash_tables == NULL);
4004 /* Initialize the weak pointer list. */
4005 weak_pointers = NULL;
4007 #ifdef LUTEX_WIDETAG
4008 unmark_lutexes(generation);
4011 /* When a generation is not being raised it is transported to a
4012 * temporary generation (NUM_GENERATIONS), and lowered when
4013 * done. Set up this new generation. There should be no pages
4014 * allocated to it yet. */
4016 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
4019 /* Set the global src and dest. generations */
4020 from_space = generation;
4022 new_space = generation+1;
4024 new_space = SCRATCH_GENERATION;
4026 /* Change to a new space for allocation, resetting the alloc_start_page */
4027 gc_alloc_generation = new_space;
4028 generations[new_space].alloc_start_page = 0;
4029 generations[new_space].alloc_unboxed_start_page = 0;
4030 generations[new_space].alloc_large_start_page = 0;
4031 generations[new_space].alloc_large_unboxed_start_page = 0;
4033 /* Before any pointers are preserved, the dont_move flags on the
4034 * pages need to be cleared. */
4035 for (i = 0; i < last_free_page; i++)
4036 if(page_table[i].gen==from_space)
4037 page_table[i].dont_move = 0;
4039 /* Un-write-protect the old-space pages. This is essential for the
4040 * promoted pages as they may contain pointers into the old-space
4041 * which need to be scavenged. It also helps avoid unnecessary page
4042 * faults as forwarding pointers are written into them. They need to
4043 * be un-protected anyway before unmapping later. */
4044 unprotect_oldspace();
4046 /* Scavenge the stacks' conservative roots. */
4048 /* there are potentially two stacks for each thread: the main
4049 * stack, which may contain Lisp pointers, and the alternate stack.
4050 * We don't ever run Lisp code on the altstack, but it may
4051 * host a sigcontext with lisp objects in it */
4053 /* what we need to do: (1) find the stack pointer for the main
4054 * stack; scavenge it (2) find the interrupt context on the
4055 * alternate stack that might contain lisp values, and scavenge
4058 /* we assume that none of the preceding applies to the thread that
4059 * initiates GC. If you ever call GC from inside an altstack
4060 * handler, you will lose. */
4062 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
4063 /* And if we're saving a core, there's no point in being conservative. */
4064 if (conservative_stack) {
4065 for_each_thread(th) {
4067 void **esp=(void **)-1;
4068 #ifdef LISP_FEATURE_SB_THREAD
4070 if(th==arch_os_get_current_thread()) {
4071 /* Somebody is going to burn in hell for this, but casting
4072 * it in two steps shuts gcc up about strict aliasing. */
4073 esp = (void **)((void *)&raise);
4076 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
4077 for(i=free-1;i>=0;i--) {
4078 os_context_t *c=th->interrupt_contexts[i];
4079 esp1 = (void **) *os_context_register_addr(c,reg_SP);
4080 if (esp1>=(void **)th->control_stack_start &&
4081 esp1<(void **)th->control_stack_end) {
4082 if(esp1<esp) esp=esp1;
4083 preserve_context_registers(c);
4088 esp = (void **)((void *)&raise);
4090 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
4091 preserve_pointer(*ptr);
4098 if (gencgc_verbose > 1) {
4099 long num_dont_move_pages = count_dont_move_pages();
4101 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
4102 num_dont_move_pages,
4103 npage_bytes(num_dont_move_pages));
4107 /* Scavenge all the rest of the roots. */
4109 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4111 * If not x86, we need to scavenge the interrupt context(s) and the
4114 scavenge_interrupt_contexts();
4115 scavenge_control_stack();
4118 /* Scavenge the Lisp functions of the interrupt handlers, taking
4119 * care to avoid SIG_DFL and SIG_IGN. */
4120 for (i = 0; i < NSIG; i++) {
4121 union interrupt_handler handler = interrupt_handlers[i];
4122 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4123 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4124 scavenge((lispobj *)(interrupt_handlers + i), 1);
4127 /* Scavenge the binding stacks. */
4130 for_each_thread(th) {
4131 long len= (lispobj *)get_binding_stack_pointer(th) -
4132 th->binding_stack_start;
4133 scavenge((lispobj *) th->binding_stack_start,len);
4134 #ifdef LISP_FEATURE_SB_THREAD
4135 /* do the tls as well */
4136 len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
4137 (sizeof (struct thread))/(sizeof (lispobj));
4138 scavenge((lispobj *) (th+1),len);
4143 /* The original CMU CL code had scavenge-read-only-space code
4144 * controlled by the Lisp-level variable
4145 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4146 * wasn't documented under what circumstances it was useful or
4147 * safe to turn it on, so it's been turned off in SBCL. If you
4148 * want/need this functionality, and can test and document it,
4149 * please submit a patch. */
4151 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4152 unsigned long read_only_space_size =
4153 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4154 (lispobj*)READ_ONLY_SPACE_START;
4156 "/scavenge read only space: %d bytes\n",
4157 read_only_space_size * sizeof(lispobj)));
4158 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4162 /* Scavenge static space. */
4164 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4165 (lispobj *)STATIC_SPACE_START;
4166 if (gencgc_verbose > 1) {
4168 "/scavenge static space: %d bytes\n",
4169 static_space_size * sizeof(lispobj)));
4171 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4173 /* All generations but the generation being GCed need to be
4174 * scavenged. The new_space generation needs special handling as
4175 * objects may be moved in - it is handled separately below. */
4176 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4178 /* Finally scavenge the new_space generation. Keep going until no
4179 * more objects are moved into the new generation */
4180 scavenge_newspace_generation(new_space);
4182 /* FIXME: I tried reenabling this check when debugging unrelated
4183 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4184 * Since the current GC code seems to work well, I'm guessing that
4185 * this debugging code is just stale, but I haven't tried to
4186 * figure it out. It should be figured out and then either made to
4187 * work or just deleted. */
4188 #define RESCAN_CHECK 0
4190 /* As a check re-scavenge the newspace once; no new objects should
4193 long old_bytes_allocated = bytes_allocated;
4194 long bytes_allocated;
4196 /* Start with a full scavenge. */
4197 scavenge_newspace_generation_one_scan(new_space);
4199 /* Flush the current regions, updating the tables. */
4200 gc_alloc_update_all_page_tables();
4202 bytes_allocated = bytes_allocated - old_bytes_allocated;
4204 if (bytes_allocated != 0) {
4205 lose("Rescan of new_space allocated %d more bytes.\n",
4211 scan_weak_hash_tables();
4212 scan_weak_pointers();
4214 /* Flush the current regions, updating the tables. */
4215 gc_alloc_update_all_page_tables();
4217 /* Free the pages in oldspace, but not those marked dont_move. */
4218 bytes_freed = free_oldspace();
4220 /* If the GC is not raising the age then lower the generation back
4221 * to its normal generation number */
4223 for (i = 0; i < last_free_page; i++)
4224 if ((page_table[i].bytes_used != 0)
4225 && (page_table[i].gen == SCRATCH_GENERATION))
4226 page_table[i].gen = generation;
4227 gc_assert(generations[generation].bytes_allocated == 0);
4228 generations[generation].bytes_allocated =
4229 generations[SCRATCH_GENERATION].bytes_allocated;
4230 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4233 /* Reset the alloc_start_page for generation. */
4234 generations[generation].alloc_start_page = 0;
4235 generations[generation].alloc_unboxed_start_page = 0;
4236 generations[generation].alloc_large_start_page = 0;
4237 generations[generation].alloc_large_unboxed_start_page = 0;
4239 if (generation >= verify_gens) {
4240 if (gencgc_verbose) {
4244 verify_dynamic_space();
4247 /* Set the new gc trigger for the GCed generation. */
4248 generations[generation].gc_trigger =
4249 generations[generation].bytes_allocated
4250 + generations[generation].bytes_consed_between_gc;
4253 generations[generation].num_gc = 0;
4255 ++generations[generation].num_gc;
4257 #ifdef LUTEX_WIDETAG
4258 reap_lutexes(generation);
4260 move_lutexes(generation, generation+1);
4264 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4266 update_dynamic_space_free_pointer(void)
4268 page_index_t last_page = -1, i;
4270 for (i = 0; i < last_free_page; i++)
4271 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
4274 last_free_page = last_page+1;
4276 set_alloc_pointer((lispobj)(page_address(last_free_page)));
4277 return 0; /* dummy value: return something ... */
4281 remap_free_pages (page_index_t from, page_index_t to)
4283 page_index_t first_page, last_page;
4285 for (first_page = from; first_page <= to; first_page++) {
4286 if (page_allocated_p(first_page) ||
4287 (page_table[first_page].need_to_zero == 0)) {
4291 last_page = first_page + 1;
4292 while (page_free_p(last_page) &&
4294 (page_table[last_page].need_to_zero == 1)) {
4298 /* There's a mysterious Solaris/x86 problem with using mmap
4299 * tricks for memory zeroing. See sbcl-devel thread
4300 * "Re: patch: standalone executable redux".
4302 #if defined(LISP_FEATURE_SUNOS)
4303 zero_pages(first_page, last_page-1);
4305 zero_pages_with_mmap(first_page, last_page-1);
4308 first_page = last_page;
4312 generation_index_t small_generation_limit = 1;
4314 /* GC all generations newer than last_gen, raising the objects in each
4315 * to the next older generation - we finish when all generations below
4316 * last_gen are empty. Then if last_gen is due for a GC, or if
4317 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4318 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4320 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4321 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4323 collect_garbage(generation_index_t last_gen)
4325 generation_index_t gen = 0, i;
4328 /* The largest value of last_free_page seen since the time
4329 * remap_free_pages was called. */
4330 static page_index_t high_water_mark = 0;
4332 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4336 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4338 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4343 /* Flush the alloc regions updating the tables. */
4344 gc_alloc_update_all_page_tables();
4346 /* Verify the new objects created by Lisp code. */
4347 if (pre_verify_gen_0) {
4348 FSHOW((stderr, "pre-checking generation 0\n"));
4349 verify_generation(0);
4352 if (gencgc_verbose > 1)
4353 print_generation_stats(0);
4356 /* Collect the generation. */
4358 if (gen >= gencgc_oldest_gen_to_gc) {
4359 /* Never raise the oldest generation. */
4364 || (generations[gen].num_gc >= generations[gen].trigger_age);
4367 if (gencgc_verbose > 1) {
4369 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4372 generations[gen].bytes_allocated,
4373 generations[gen].gc_trigger,
4374 generations[gen].num_gc));
4377 /* If an older generation is being filled, then update its
4380 generations[gen+1].cum_sum_bytes_allocated +=
4381 generations[gen+1].bytes_allocated;
4384 garbage_collect_generation(gen, raise);
4386 /* Reset the memory age cum_sum. */
4387 generations[gen].cum_sum_bytes_allocated = 0;
4389 if (gencgc_verbose > 1) {
4390 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4391 print_generation_stats(0);
4395 } while ((gen <= gencgc_oldest_gen_to_gc)
4396 && ((gen < last_gen)
4397 || ((gen <= gencgc_oldest_gen_to_gc)
4399 && (generations[gen].bytes_allocated
4400 > generations[gen].gc_trigger)
4401 && (gen_av_mem_age(gen)
4402 > generations[gen].min_av_mem_age))));
4404 /* Now if gen-1 was raised all generations before gen are empty.
4405 * If it wasn't raised then all generations before gen-1 are empty.
4407 * Now objects within this gen's pages cannot point to younger
4408 * generations unless they are written to. This can be exploited
4409 * by write-protecting the pages of gen; then when younger
4410 * generations are GCed only the pages which have been written
4415 gen_to_wp = gen - 1;
4417 /* There's not much point in WPing pages in generation 0 as it is
4418 * never scavenged (except promoted pages). */
4419 if ((gen_to_wp > 0) && enable_page_protection) {
4420 /* Check that they are all empty. */
4421 for (i = 0; i < gen_to_wp; i++) {
4422 if (generations[i].bytes_allocated)
4423 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4426 write_protect_generation_pages(gen_to_wp);
4429 /* Set gc_alloc() back to generation 0. The current regions should
4430 * be flushed after the above GCs. */
4431 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4432 gc_alloc_generation = 0;
4434 /* Save the high-water mark before updating last_free_page */
4435 if (last_free_page > high_water_mark)
4436 high_water_mark = last_free_page;
4438 update_dynamic_space_free_pointer();
4440 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4442 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4445 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4448 if (gen > small_generation_limit) {
4449 if (last_free_page > high_water_mark)
4450 high_water_mark = last_free_page;
4451 remap_free_pages(0, high_water_mark);
4452 high_water_mark = 0;
4457 SHOW("returning from collect_garbage");
4460 /* This is called by Lisp PURIFY when it is finished. All live objects
4461 * will have been moved to the RO and Static heaps. The dynamic space
4462 * will need a full re-initialization. We don't bother having Lisp
4463 * PURIFY flush the current gc_alloc() region, as the page_tables are
4464 * re-initialized, and every page is zeroed to be sure. */
4470 if (gencgc_verbose > 1) {
4471 SHOW("entering gc_free_heap");
4474 for (page = 0; page < page_table_pages; page++) {
4475 /* Skip free pages which should already be zero filled. */
4476 if (page_allocated_p(page)) {
4477 void *page_start, *addr;
4479 /* Mark the page free. The other slots are assumed invalid
4480 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4481 * should not be write-protected -- except that the
4482 * generation is used for the current region but it sets
4484 page_table[page].allocated = FREE_PAGE_FLAG;
4485 page_table[page].bytes_used = 0;
4487 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4488 * about this change. */
4489 /* Zero the page. */
4490 page_start = (void *)page_address(page);
4492 /* First, remove any write-protection. */
4493 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
4494 page_table[page].write_protected = 0;
4496 os_invalidate(page_start,PAGE_BYTES);
4497 addr = os_validate(page_start,PAGE_BYTES);
4498 if (addr == NULL || addr != page_start) {
4499 lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n",
4504 page_table[page].write_protected = 0;
4506 } else if (gencgc_zero_check_during_free_heap) {
4507 /* Double-check that the page is zero filled. */
4510 gc_assert(page_free_p(page));
4511 gc_assert(page_table[page].bytes_used == 0);
4512 page_start = (long *)page_address(page);
4513 for (i=0; i<1024; i++) {
4514 if (page_start[i] != 0) {
4515 lose("free region not zero at %x\n", page_start + i);
4521 bytes_allocated = 0;
4523 /* Initialize the generations. */
4524 for (page = 0; page < NUM_GENERATIONS; page++) {
4525 generations[page].alloc_start_page = 0;
4526 generations[page].alloc_unboxed_start_page = 0;
4527 generations[page].alloc_large_start_page = 0;
4528 generations[page].alloc_large_unboxed_start_page = 0;
4529 generations[page].bytes_allocated = 0;
4530 generations[page].gc_trigger = 2000000;
4531 generations[page].num_gc = 0;
4532 generations[page].cum_sum_bytes_allocated = 0;
4533 generations[page].lutexes = NULL;
4536 if (gencgc_verbose > 1)
4537 print_generation_stats(0);
4539 /* Initialize gc_alloc(). */
4540 gc_alloc_generation = 0;
4542 gc_set_region_empty(&boxed_region);
4543 gc_set_region_empty(&unboxed_region);
4546 set_alloc_pointer((lispobj)((char *)heap_base));
4548 if (verify_after_free_heap) {
4549 /* Check whether purify has left any bad pointers. */
4550 FSHOW((stderr, "checking after free_heap\n"));
4560 /* Compute the number of pages needed for the dynamic space.
4561 * Dynamic space size should be aligned on page size. */
4562 page_table_pages = dynamic_space_size/PAGE_BYTES;
4563 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4565 page_table = calloc(page_table_pages, sizeof(struct page));
4566 gc_assert(page_table);
4569 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4570 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4572 #ifdef LUTEX_WIDETAG
4573 scavtab[LUTEX_WIDETAG] = scav_lutex;
4574 transother[LUTEX_WIDETAG] = trans_lutex;
4575 sizetab[LUTEX_WIDETAG] = size_lutex;
4578 heap_base = (void*)DYNAMIC_SPACE_START;
4580 /* Initialize each page structure. */
4581 for (i = 0; i < page_table_pages; i++) {
4582 /* Initialize all pages as free. */
4583 page_table[i].allocated = FREE_PAGE_FLAG;
4584 page_table[i].bytes_used = 0;
4586 /* Pages are not write-protected at startup. */
4587 page_table[i].write_protected = 0;
4590 bytes_allocated = 0;
4592 /* Initialize the generations.
4594 * FIXME: very similar to code in gc_free_heap(), should be shared */
4595 for (i = 0; i < NUM_GENERATIONS; i++) {
4596 generations[i].alloc_start_page = 0;
4597 generations[i].alloc_unboxed_start_page = 0;
4598 generations[i].alloc_large_start_page = 0;
4599 generations[i].alloc_large_unboxed_start_page = 0;
4600 generations[i].bytes_allocated = 0;
4601 generations[i].gc_trigger = 2000000;
4602 generations[i].num_gc = 0;
4603 generations[i].cum_sum_bytes_allocated = 0;
4604 /* the tune-able parameters */
4605 generations[i].bytes_consed_between_gc = 2000000;
4606 generations[i].trigger_age = 1;
4607 generations[i].min_av_mem_age = 0.75;
4608 generations[i].lutexes = NULL;
4611 /* Initialize gc_alloc. */
4612 gc_alloc_generation = 0;
4613 gc_set_region_empty(&boxed_region);
4614 gc_set_region_empty(&unboxed_region);
4619 /* Pick up the dynamic space from after a core load.
4621 * The ALLOCATION_POINTER points to the end of the dynamic space.
4625 gencgc_pickup_dynamic(void)
4627 page_index_t page = 0;
4628 void *alloc_ptr = (void *)get_alloc_pointer();
4629 lispobj *prev=(lispobj *)page_address(page);
4630 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4632 lispobj *first,*ptr= (lispobj *)page_address(page);
4633 page_table[page].allocated = BOXED_PAGE_FLAG;
4634 page_table[page].gen = gen;
4635 page_table[page].bytes_used = PAGE_BYTES;
4636 page_table[page].large_object = 0;
4637 page_table[page].write_protected = 0;
4638 page_table[page].write_protected_cleared = 0;
4639 page_table[page].dont_move = 0;
4640 page_table[page].need_to_zero = 1;
4642 if (!gencgc_partial_pickup) {
4643 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4644 if(ptr == first) prev=ptr;
4645 page_table[page].region_start_offset =
4646 page_address(page) - (void *)prev;
4649 } while (page_address(page) < alloc_ptr);
4651 #ifdef LUTEX_WIDETAG
4652 /* Lutexes have been registered in generation 0 by coreparse, and
4653 * need to be moved to the right one manually.
4655 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4658 last_free_page = page;
4660 generations[gen].bytes_allocated = npage_bytes(page);
4661 bytes_allocated = npage_bytes(page);
4663 gc_alloc_update_all_page_tables();
4664 write_protect_generation_pages(gen);
4668 gc_initialize_pointers(void)
4670 gencgc_pickup_dynamic();
4674 /* alloc(..) is the external interface for memory allocation. It
4675 * allocates to generation 0. It is not called from within the garbage
4676 * collector as it is only external uses that need the check for heap
4677 * size (GC trigger) and to disable the interrupts (interrupts are
4678 * always disabled during a GC).
4680 * The vops that call alloc(..) assume that the returned space is zero-filled.
4681 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4683 * The check for a GC trigger is only performed when the current
4684 * region is full, so in most cases it's not needed. */
4686 static inline lispobj *
4687 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4688 struct thread *thread)
4690 #ifndef LISP_FEATURE_WIN32
4691 lispobj alloc_signal;
4694 void *new_free_pointer;
4696 gc_assert(nbytes>0);
4698 /* Check for alignment allocation problems. */
4699 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4700 && ((nbytes & LOWTAG_MASK) == 0));
4702 /* Must be inside a PA section. */
4703 gc_assert(get_pseudo_atomic_atomic(thread));
4705 /* maybe we can do this quickly ... */
4706 new_free_pointer = region->free_pointer + nbytes;
4707 if (new_free_pointer <= region->end_addr) {
4708 new_obj = (void*)(region->free_pointer);
4709 region->free_pointer = new_free_pointer;
4710 return(new_obj); /* yup */
4713 /* we have to go the long way around, it seems. Check whether we
4714 * should GC in the near future
4716 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4717 /* Don't flood the system with interrupts if the need to gc is
4718 * already noted. This can happen for example when SUB-GC
4719 * allocates or after a gc triggered in a WITHOUT-GCING. */
4720 if (SymbolValue(GC_PENDING,thread) == NIL) {
4721 /* set things up so that GC happens when we finish the PA
4723 SetSymbolValue(GC_PENDING,T,thread);
4724 if (SymbolValue(GC_INHIBIT,thread) == NIL)
4725 set_pseudo_atomic_interrupted(thread);
4728 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4730 #ifndef LISP_FEATURE_WIN32
4731 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4732 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4733 if ((signed long) alloc_signal <= 0) {
4734 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4735 thread_kill(thread->os_thread, SIGPROF);
4737 SetSymbolValue(ALLOC_SIGNAL,
4738 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4748 general_alloc(long nbytes, int page_type_flag)
4750 struct thread *thread = arch_os_get_current_thread();
4751 /* Select correct region, and call general_alloc_internal with it.
4752 * For other then boxed allocation we must lock first, since the
4753 * region is shared. */
4754 if (BOXED_PAGE_FLAG & page_type_flag) {
4755 #ifdef LISP_FEATURE_SB_THREAD
4756 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4758 struct alloc_region *region = &boxed_region;
4760 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4761 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4763 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4764 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4765 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4768 lose("bad page type flag: %d", page_type_flag);
4775 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4776 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4780 * shared support for the OS-dependent signal handlers which
4781 * catch GENCGC-related write-protect violations
4783 void unhandled_sigmemoryfault(void* addr);
4785 /* Depending on which OS we're running under, different signals might
4786 * be raised for a violation of write protection in the heap. This
4787 * function factors out the common generational GC magic which needs
4788 * to invoked in this case, and should be called from whatever signal
4789 * handler is appropriate for the OS we're running under.
4791 * Return true if this signal is a normal generational GC thing that
4792 * we were able to handle, or false if it was abnormal and control
4793 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4796 gencgc_handle_wp_violation(void* fault_addr)
4798 page_index_t page_index = find_page_index(fault_addr);
4800 #ifdef QSHOW_SIGNALS
4801 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4802 fault_addr, page_index));
4805 /* Check whether the fault is within the dynamic space. */
4806 if (page_index == (-1)) {
4808 /* It can be helpful to be able to put a breakpoint on this
4809 * case to help diagnose low-level problems. */
4810 unhandled_sigmemoryfault(fault_addr);
4812 /* not within the dynamic space -- not our responsibility */
4816 if (page_table[page_index].write_protected) {
4817 /* Unprotect the page. */
4818 os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
4819 page_table[page_index].write_protected_cleared = 1;
4820 page_table[page_index].write_protected = 0;
4822 /* The only acceptable reason for this signal on a heap
4823 * access is that GENCGC write-protected the page.
4824 * However, if two CPUs hit a wp page near-simultaneously,
4825 * we had better not have the second one lose here if it
4826 * does this test after the first one has already set wp=0
4828 if(page_table[page_index].write_protected_cleared != 1)
4829 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4830 page_index, boxed_region.first_page,
4831 boxed_region.last_page);
4833 /* Don't worry, we can handle it. */
4837 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4838 * it's not just a case of the program hitting the write barrier, and
4839 * are about to let Lisp deal with it. It's basically just a
4840 * convenient place to set a gdb breakpoint. */
4842 unhandled_sigmemoryfault(void *addr)
4845 void gc_alloc_update_all_page_tables(void)
4847 /* Flush the alloc regions updating the tables. */
4850 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4851 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4852 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4856 gc_set_region_empty(struct alloc_region *region)
4858 region->first_page = 0;
4859 region->last_page = -1;
4860 region->start_addr = page_address(0);
4861 region->free_pointer = page_address(0);
4862 region->end_addr = page_address(0);
4866 zero_all_free_pages()
4870 for (i = 0; i < last_free_page; i++) {
4871 if (page_free_p(i)) {
4872 #ifdef READ_PROTECT_FREE_PAGES
4873 os_protect(page_address(i),
4882 /* Things to do before doing a final GC before saving a core (without
4885 * + Pages in large_object pages aren't moved by the GC, so we need to
4886 * unset that flag from all pages.
4887 * + The pseudo-static generation isn't normally collected, but it seems
4888 * reasonable to collect it at least when saving a core. So move the
4889 * pages to a normal generation.
4892 prepare_for_final_gc ()
4895 for (i = 0; i < last_free_page; i++) {
4896 page_table[i].large_object = 0;
4897 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4898 int used = page_table[i].bytes_used;
4899 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4900 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4901 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4907 /* Do a non-conservative GC, and then save a core with the initial
4908 * function being set to the value of the static symbol
4909 * SB!VM:RESTART-LISP-FUNCTION */
4911 gc_and_save(char *filename, boolean prepend_runtime,
4912 boolean save_runtime_options)
4915 void *runtime_bytes = NULL;
4916 size_t runtime_size;
4918 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4923 conservative_stack = 0;
4925 /* The filename might come from Lisp, and be moved by the now
4926 * non-conservative GC. */
4927 filename = strdup(filename);
4929 /* Collect twice: once into relatively high memory, and then back
4930 * into low memory. This compacts the retained data into the lower
4931 * pages, minimizing the size of the core file.
4933 prepare_for_final_gc();
4934 gencgc_alloc_start_page = last_free_page;
4935 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4937 prepare_for_final_gc();
4938 gencgc_alloc_start_page = -1;
4939 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4941 if (prepend_runtime)
4942 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4944 /* The dumper doesn't know that pages need to be zeroed before use. */
4945 zero_all_free_pages();
4946 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4947 prepend_runtime, save_runtime_options);
4948 /* Oops. Save still managed to fail. Since we've mangled the stack
4949 * beyond hope, there's not much we can do.
4950 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4951 * going to be rather unsatisfactory too... */
4952 lose("Attempt to save core after non-conservative GC failed.\n");