2 * GENerational Conservative Garbage Collector for SBCL
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
17 * For a review of garbage collection techniques (e.g. generational
18 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
19 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
20 * had been accepted for _ACM Computing Surveys_ and was available
21 * as a PostScript preprint through
22 * <http://www.cs.utexas.edu/users/oops/papers.html>
24 * <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
37 #include "interrupt.h"
42 #include "gc-internal.h"
45 #include "genesis/vector.h"
46 #include "genesis/weak-pointer.h"
47 #include "genesis/fdefn.h"
48 #include "genesis/simple-fun.h"
50 #include "genesis/hash-table.h"
51 #include "genesis/instance.h"
52 #include "genesis/layout.h"
54 #if defined(LUTEX_WIDETAG)
55 #include "pthread-lutex.h"
58 /* forward declarations */
59 page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
67 /* Generations 0-5 are normal collected generations, 6 is only used as
68 * scratch space by the collector, and should never get collected.
71 HIGHEST_NORMAL_GENERATION = 5,
72 PSEUDO_STATIC_GENERATION,
77 /* Should we use page protection to help avoid the scavenging of pages
78 * that don't have pointers to younger generations? */
79 boolean enable_page_protection = 1;
81 /* the minimum size (in bytes) for a large object*/
82 long large_object_size = 4 * PAGE_BYTES;
89 /* the verbosity level. All non-error messages are disabled at level 0;
90 * and only a few rare messages are printed at level 1. */
92 boolean gencgc_verbose = 1;
94 boolean gencgc_verbose = 0;
97 /* FIXME: At some point enable the various error-checking things below
98 * and see what they say. */
100 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
101 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
103 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
105 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
106 boolean pre_verify_gen_0 = 0;
108 /* Should we check for bad pointers after gc_free_heap is called
109 * from Lisp PURIFY? */
110 boolean verify_after_free_heap = 0;
112 /* Should we print a note when code objects are found in the dynamic space
113 * during a heap verify? */
114 boolean verify_dynamic_code_check = 0;
116 /* Should we check code objects for fixup errors after they are transported? */
117 boolean check_code_fixups = 0;
119 /* Should we check that newly allocated regions are zero filled? */
120 boolean gencgc_zero_check = 0;
122 /* Should we check that the free space is zero filled? */
123 boolean gencgc_enable_verify_zero_fill = 0;
125 /* Should we check that free pages are zero filled during gc_free_heap
126 * called after Lisp PURIFY? */
127 boolean gencgc_zero_check_during_free_heap = 0;
129 /* When loading a core, don't do a full scan of the memory for the
130 * memory region boundaries. (Set to true by coreparse.c if the core
131 * contained a pagetable entry).
133 boolean gencgc_partial_pickup = 0;
135 /* If defined, free pages are read-protected to ensure that nothing
139 /* #define READ_PROTECT_FREE_PAGES */
143 * GC structures and variables
146 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
147 unsigned long bytes_allocated = 0;
148 unsigned long auto_gc_trigger = 0;
150 /* the source and destination generations. These are set before a GC starts
152 generation_index_t from_space;
153 generation_index_t new_space;
155 /* Set to 1 when in GC */
156 boolean gc_active_p = 0;
158 /* should the GC be conservative on stack. If false (only right before
159 * saving a core), don't scan the stack / mark pages dont_move. */
160 static boolean conservative_stack = 1;
162 /* An array of page structures is allocated on gc initialization.
163 * This helps quickly map between an address its page structure.
164 * page_table_pages is set from the size of the dynamic space. */
165 page_index_t page_table_pages;
166 struct page *page_table;
168 /* To map addresses to page structures the address of the first page
170 static void *heap_base = NULL;
172 /* Calculate the start address for the given page number. */
174 page_address(page_index_t page_num)
176 return (heap_base + (page_num * PAGE_BYTES));
179 /* Calculate the address where the allocation region associated with the page starts. */
181 page_region_start(page_index_t page_index)
183 return page_address(page_index)+page_table[page_index].first_object_offset;
186 /* Find the page index within the page_table for the given
187 * address. Return -1 on failure. */
189 find_page_index(void *addr)
191 page_index_t index = addr-heap_base;
194 index = ((unsigned long)index)/PAGE_BYTES;
195 if (index < page_table_pages)
202 /* a structure to hold the state of a generation */
205 /* the first page that gc_alloc() checks on its next call */
206 page_index_t alloc_start_page;
208 /* the first page that gc_alloc_unboxed() checks on its next call */
209 page_index_t alloc_unboxed_start_page;
211 /* the first page that gc_alloc_large (boxed) considers on its next
212 * call. (Although it always allocates after the boxed_region.) */
213 page_index_t alloc_large_start_page;
215 /* the first page that gc_alloc_large (unboxed) considers on its
216 * next call. (Although it always allocates after the
217 * current_unboxed_region.) */
218 page_index_t alloc_large_unboxed_start_page;
220 /* the bytes allocated to this generation */
221 long bytes_allocated;
223 /* the number of bytes at which to trigger a GC */
226 /* to calculate a new level for gc_trigger */
227 long bytes_consed_between_gc;
229 /* the number of GCs since the last raise */
232 /* the average age after which a GC will raise objects to the
236 /* the cumulative sum of the bytes allocated to this generation. It is
237 * cleared after a GC on this generations, and update before new
238 * objects are added from a GC of a younger generation. Dividing by
239 * the bytes_allocated will give the average age of the memory in
240 * this generation since its last GC. */
241 long cum_sum_bytes_allocated;
243 /* a minimum average memory age before a GC will occur helps
244 * prevent a GC when a large number of new live objects have been
245 * added, in which case a GC could be a waste of time */
246 double min_av_mem_age;
248 /* A linked list of lutex structures in this generation, used for
249 * implementing lutex finalization. */
251 struct lutex *lutexes;
257 /* an array of generation structures. There needs to be one more
258 * generation structure than actual generations as the oldest
259 * generation is temporarily raised then lowered. */
260 struct generation generations[NUM_GENERATIONS];
262 /* the oldest generation that is will currently be GCed by default.
263 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
265 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
267 * Setting this to 0 effectively disables the generational nature of
268 * the GC. In some applications generational GC may not be useful
269 * because there are no long-lived objects.
271 * An intermediate value could be handy after moving long-lived data
272 * into an older generation so an unnecessary GC of this long-lived
273 * data can be avoided. */
274 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
276 /* The maximum free page in the heap is maintained and used to update
277 * ALLOCATION_POINTER which is used by the room function to limit its
278 * search of the heap. XX Gencgc obviously needs to be better
279 * integrated with the Lisp code. */
280 page_index_t last_free_page;
282 /* This lock is to prevent multiple threads from simultaneously
283 * allocating new regions which overlap each other. Note that the
284 * majority of GC is single-threaded, but alloc() may be called from
285 * >1 thread at a time and must be thread-safe. This lock must be
286 * seized before all accesses to generations[] or to parts of
287 * page_table[] that other threads may want to see */
289 #ifdef LISP_FEATURE_SB_THREAD
290 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
295 * miscellaneous heap functions
298 /* Count the number of pages which are write-protected within the
299 * given generation. */
301 count_write_protect_generation_pages(generation_index_t generation)
306 for (i = 0; i < last_free_page; i++)
307 if ((page_table[i].allocated != FREE_PAGE_FLAG)
308 && (page_table[i].gen == generation)
309 && (page_table[i].write_protected == 1))
314 /* Count the number of pages within the given generation. */
316 count_generation_pages(generation_index_t generation)
321 for (i = 0; i < last_free_page; i++)
322 if ((page_table[i].allocated != FREE_PAGE_FLAG)
323 && (page_table[i].gen == generation))
330 count_dont_move_pages(void)
334 for (i = 0; i < last_free_page; i++) {
335 if ((page_table[i].allocated != FREE_PAGE_FLAG)
336 && (page_table[i].dont_move != 0)) {
344 /* Work through the pages and add up the number of bytes used for the
345 * given generation. */
347 count_generation_bytes_allocated (generation_index_t gen)
351 for (i = 0; i < last_free_page; i++) {
352 if ((page_table[i].allocated != FREE_PAGE_FLAG)
353 && (page_table[i].gen == gen))
354 result += page_table[i].bytes_used;
359 /* Return the average age of the memory in a generation. */
361 gen_av_mem_age(generation_index_t gen)
363 if (generations[gen].bytes_allocated == 0)
367 ((double)generations[gen].cum_sum_bytes_allocated)
368 / ((double)generations[gen].bytes_allocated);
371 /* The verbose argument controls how much to print: 0 for normal
372 * level of detail; 1 for debugging. */
374 print_generation_stats(int verbose) /* FIXME: should take FILE argument */
376 generation_index_t i, gens;
378 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
379 #define FPU_STATE_SIZE 27
380 int fpu_state[FPU_STATE_SIZE];
381 #elif defined(LISP_FEATURE_PPC)
382 #define FPU_STATE_SIZE 32
383 long long fpu_state[FPU_STATE_SIZE];
386 /* This code uses the FP instructions which may be set up for Lisp
387 * so they need to be saved and reset for C. */
390 /* highest generation to print */
392 gens = SCRATCH_GENERATION;
394 gens = PSEUDO_STATIC_GENERATION;
396 /* Print the heap stats. */
398 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
400 for (i = 0; i < gens; i++) {
403 long unboxed_cnt = 0;
404 long large_boxed_cnt = 0;
405 long large_unboxed_cnt = 0;
408 for (j = 0; j < last_free_page; j++)
409 if (page_table[j].gen == i) {
411 /* Count the number of boxed pages within the given
413 if (page_table[j].allocated & BOXED_PAGE_FLAG) {
414 if (page_table[j].large_object)
419 if(page_table[j].dont_move) pinned_cnt++;
420 /* Count the number of unboxed pages within the given
422 if (page_table[j].allocated & UNBOXED_PAGE_FLAG) {
423 if (page_table[j].large_object)
430 gc_assert(generations[i].bytes_allocated
431 == count_generation_bytes_allocated(i));
433 " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
435 generations[i].alloc_start_page,
436 generations[i].alloc_unboxed_start_page,
437 generations[i].alloc_large_start_page,
438 generations[i].alloc_large_unboxed_start_page,
444 generations[i].bytes_allocated,
445 (count_generation_pages(i)*PAGE_BYTES - generations[i].bytes_allocated),
446 generations[i].gc_trigger,
447 count_write_protect_generation_pages(i),
448 generations[i].num_gc,
451 fprintf(stderr," Total bytes allocated=%ld\n", bytes_allocated);
453 fpu_restore(fpu_state);
457 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
458 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
461 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
462 * if zeroing it ourselves, i.e. in practice give the memory back to the
463 * OS. Generally done after a large GC.
465 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
467 void *addr = (void *) page_address(start), *new_addr;
468 size_t length = PAGE_BYTES*(1+end-start);
473 os_invalidate(addr, length);
474 new_addr = os_validate(addr, length);
475 if (new_addr == NULL || new_addr != addr) {
476 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x", start, new_addr);
479 for (i = start; i <= end; i++) {
480 page_table[i].need_to_zero = 0;
484 /* Zero the pages from START to END (inclusive). Generally done just after
485 * a new region has been allocated.
488 zero_pages(page_index_t start, page_index_t end) {
492 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
493 fast_bzero(page_address(start), PAGE_BYTES*(1+end-start));
495 bzero(page_address(start), PAGE_BYTES*(1+end-start));
500 /* Zero the pages from START to END (inclusive), except for those
501 * pages that are known to already zeroed. Mark all pages in the
502 * ranges as non-zeroed.
505 zero_dirty_pages(page_index_t start, page_index_t end) {
508 for (i = start; i <= end; i++) {
509 if (page_table[i].need_to_zero == 1) {
510 zero_pages(start, end);
515 for (i = start; i <= end; i++) {
516 page_table[i].need_to_zero = 1;
522 * To support quick and inline allocation, regions of memory can be
523 * allocated and then allocated from with just a free pointer and a
524 * check against an end address.
526 * Since objects can be allocated to spaces with different properties
527 * e.g. boxed/unboxed, generation, ages; there may need to be many
528 * allocation regions.
530 * Each allocation region may start within a partly used page. Many
531 * features of memory use are noted on a page wise basis, e.g. the
532 * generation; so if a region starts within an existing allocated page
533 * it must be consistent with this page.
535 * During the scavenging of the newspace, objects will be transported
536 * into an allocation region, and pointers updated to point to this
537 * allocation region. It is possible that these pointers will be
538 * scavenged again before the allocation region is closed, e.g. due to
539 * trans_list which jumps all over the place to cleanup the list. It
540 * is important to be able to determine properties of all objects
541 * pointed to when scavenging, e.g to detect pointers to the oldspace.
542 * Thus it's important that the allocation regions have the correct
543 * properties set when allocated, and not just set when closed. The
544 * region allocation routines return regions with the specified
545 * properties, and grab all the pages, setting their properties
546 * appropriately, except that the amount used is not known.
548 * These regions are used to support quicker allocation using just a
549 * free pointer. The actual space used by the region is not reflected
550 * in the pages tables until it is closed. It can't be scavenged until
553 * When finished with the region it should be closed, which will
554 * update the page tables for the actual space used returning unused
555 * space. Further it may be noted in the new regions which is
556 * necessary when scavenging the newspace.
558 * Large objects may be allocated directly without an allocation
559 * region, the page tables are updated immediately.
561 * Unboxed objects don't contain pointers to other objects and so
562 * don't need scavenging. Further they can't contain pointers to
563 * younger generations so WP is not needed. By allocating pages to
564 * unboxed objects the whole page never needs scavenging or
565 * write-protecting. */
567 /* We are only using two regions at present. Both are for the current
568 * newspace generation. */
569 struct alloc_region boxed_region;
570 struct alloc_region unboxed_region;
572 /* The generation currently being allocated to. */
573 static generation_index_t gc_alloc_generation;
575 /* Find a new region with room for at least the given number of bytes.
577 * It starts looking at the current generation's alloc_start_page. So
578 * may pick up from the previous region if there is enough space. This
579 * keeps the allocation contiguous when scavenging the newspace.
581 * The alloc_region should have been closed by a call to
582 * gc_alloc_update_page_tables(), and will thus be in an empty state.
584 * To assist the scavenging functions write-protected pages are not
585 * used. Free pages should not be write-protected.
587 * It is critical to the conservative GC that the start of regions be
588 * known. To help achieve this only small regions are allocated at a
591 * During scavenging, pointers may be found to within the current
592 * region and the page generation must be set so that pointers to the
593 * from space can be recognized. Therefore the generation of pages in
594 * the region are set to gc_alloc_generation. To prevent another
595 * allocation call using the same pages, all the pages in the region
596 * are allocated, although they will initially be empty.
599 gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region)
601 page_index_t first_page;
602 page_index_t last_page;
609 "/alloc_new_region for %d bytes from gen %d\n",
610 nbytes, gc_alloc_generation));
613 /* Check that the region is in a reset state. */
614 gc_assert((alloc_region->first_page == 0)
615 && (alloc_region->last_page == -1)
616 && (alloc_region->free_pointer == alloc_region->end_addr));
617 ret = thread_mutex_lock(&free_pages_lock);
621 generations[gc_alloc_generation].alloc_unboxed_start_page;
624 generations[gc_alloc_generation].alloc_start_page;
626 last_page=gc_find_freeish_pages(&first_page,nbytes,unboxed);
627 bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
628 + PAGE_BYTES*(last_page-first_page);
630 /* Set up the alloc_region. */
631 alloc_region->first_page = first_page;
632 alloc_region->last_page = last_page;
633 alloc_region->start_addr = page_table[first_page].bytes_used
634 + page_address(first_page);
635 alloc_region->free_pointer = alloc_region->start_addr;
636 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
638 /* Set up the pages. */
640 /* The first page may have already been in use. */
641 if (page_table[first_page].bytes_used == 0) {
643 page_table[first_page].allocated = UNBOXED_PAGE_FLAG;
645 page_table[first_page].allocated = BOXED_PAGE_FLAG;
646 page_table[first_page].gen = gc_alloc_generation;
647 page_table[first_page].large_object = 0;
648 page_table[first_page].first_object_offset = 0;
652 gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG);
654 gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG);
655 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
657 gc_assert(page_table[first_page].gen == gc_alloc_generation);
658 gc_assert(page_table[first_page].large_object == 0);
660 for (i = first_page+1; i <= last_page; i++) {
662 page_table[i].allocated = UNBOXED_PAGE_FLAG;
664 page_table[i].allocated = BOXED_PAGE_FLAG;
665 page_table[i].gen = gc_alloc_generation;
666 page_table[i].large_object = 0;
667 /* This may not be necessary for unboxed regions (think it was
669 page_table[i].first_object_offset =
670 alloc_region->start_addr - page_address(i);
671 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
673 /* Bump up last_free_page. */
674 if (last_page+1 > last_free_page) {
675 last_free_page = last_page+1;
676 /* do we only want to call this on special occasions? like for boxed_region? */
677 set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
679 ret = thread_mutex_unlock(&free_pages_lock);
682 #ifdef READ_PROTECT_FREE_PAGES
683 os_protect(page_address(first_page),
684 PAGE_BYTES*(1+last_page-first_page),
688 /* If the first page was only partial, don't check whether it's
689 * zeroed (it won't be) and don't zero it (since the parts that
690 * we're interested in are guaranteed to be zeroed).
692 if (page_table[first_page].bytes_used) {
696 zero_dirty_pages(first_page, last_page);
698 /* we can do this after releasing free_pages_lock */
699 if (gencgc_zero_check) {
701 for (p = (long *)alloc_region->start_addr;
702 p < (long *)alloc_region->end_addr; p++) {
704 /* KLUDGE: It would be nice to use %lx and explicit casts
705 * (long) in code like this, so that it is less likely to
706 * break randomly when running on a machine with different
707 * word sizes. -- WHN 19991129 */
708 lose("The new region at %x is not zero (start=%p, end=%p).\n",
709 p, alloc_region->start_addr, alloc_region->end_addr);
715 /* If the record_new_objects flag is 2 then all new regions created
718 * If it's 1 then then it is only recorded if the first page of the
719 * current region is <= new_areas_ignore_page. This helps avoid
720 * unnecessary recording when doing full scavenge pass.
722 * The new_object structure holds the page, byte offset, and size of
723 * new regions of objects. Each new area is placed in the array of
724 * these structures pointer to by new_areas. new_areas_index holds the
725 * offset into new_areas.
727 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
728 * later code must detect this and handle it, probably by doing a full
729 * scavenge of a generation. */
730 #define NUM_NEW_AREAS 512
731 static int record_new_objects = 0;
732 static page_index_t new_areas_ignore_page;
738 static struct new_area (*new_areas)[];
739 static long new_areas_index;
742 /* Add a new area to new_areas. */
744 add_new_area(page_index_t first_page, long offset, long size)
746 unsigned long new_area_start,c;
749 /* Ignore if full. */
750 if (new_areas_index >= NUM_NEW_AREAS)
753 switch (record_new_objects) {
757 if (first_page > new_areas_ignore_page)
766 new_area_start = PAGE_BYTES*first_page + offset;
768 /* Search backwards for a prior area that this follows from. If
769 found this will save adding a new area. */
770 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
771 unsigned long area_end =
772 PAGE_BYTES*((*new_areas)[i].page)
773 + (*new_areas)[i].offset
774 + (*new_areas)[i].size;
776 "/add_new_area S1 %d %d %d %d\n",
777 i, c, new_area_start, area_end));*/
778 if (new_area_start == area_end) {
780 "/adding to [%d] %d %d %d with %d %d %d:\n",
782 (*new_areas)[i].page,
783 (*new_areas)[i].offset,
784 (*new_areas)[i].size,
788 (*new_areas)[i].size += size;
793 (*new_areas)[new_areas_index].page = first_page;
794 (*new_areas)[new_areas_index].offset = offset;
795 (*new_areas)[new_areas_index].size = size;
797 "/new_area %d page %d offset %d size %d\n",
798 new_areas_index, first_page, offset, size));*/
801 /* Note the max new_areas used. */
802 if (new_areas_index > max_new_areas)
803 max_new_areas = new_areas_index;
806 /* Update the tables for the alloc_region. The region may be added to
809 * When done the alloc_region is set up so that the next quick alloc
810 * will fail safely and thus a new region will be allocated. Further
811 * it is safe to try to re-update the page table of this reset
814 gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region)
817 page_index_t first_page;
818 page_index_t next_page;
820 long orig_first_page_bytes_used;
826 first_page = alloc_region->first_page;
828 /* Catch an unused alloc_region. */
829 if ((first_page == 0) && (alloc_region->last_page == -1))
832 next_page = first_page+1;
834 ret = thread_mutex_lock(&free_pages_lock);
836 if (alloc_region->free_pointer != alloc_region->start_addr) {
837 /* some bytes were allocated in the region */
838 orig_first_page_bytes_used = page_table[first_page].bytes_used;
840 gc_assert(alloc_region->start_addr == (page_address(first_page) + page_table[first_page].bytes_used));
842 /* All the pages used need to be updated */
844 /* Update the first page. */
846 /* If the page was free then set up the gen, and
847 * first_object_offset. */
848 if (page_table[first_page].bytes_used == 0)
849 gc_assert(page_table[first_page].first_object_offset == 0);
850 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
853 gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG);
855 gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG);
856 gc_assert(page_table[first_page].gen == gc_alloc_generation);
857 gc_assert(page_table[first_page].large_object == 0);
861 /* Calculate the number of bytes used in this page. This is not
862 * always the number of new bytes, unless it was free. */
864 if ((bytes_used = (alloc_region->free_pointer - page_address(first_page)))>PAGE_BYTES) {
865 bytes_used = PAGE_BYTES;
868 page_table[first_page].bytes_used = bytes_used;
869 byte_cnt += bytes_used;
872 /* All the rest of the pages should be free. We need to set their
873 * first_object_offset pointer to the start of the region, and set
876 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
878 gc_assert(page_table[next_page].allocated==UNBOXED_PAGE_FLAG);
880 gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG);
881 gc_assert(page_table[next_page].bytes_used == 0);
882 gc_assert(page_table[next_page].gen == gc_alloc_generation);
883 gc_assert(page_table[next_page].large_object == 0);
885 gc_assert(page_table[next_page].first_object_offset ==
886 alloc_region->start_addr - page_address(next_page));
888 /* Calculate the number of bytes used in this page. */
890 if ((bytes_used = (alloc_region->free_pointer
891 - page_address(next_page)))>PAGE_BYTES) {
892 bytes_used = PAGE_BYTES;
895 page_table[next_page].bytes_used = bytes_used;
896 byte_cnt += bytes_used;
901 region_size = alloc_region->free_pointer - alloc_region->start_addr;
902 bytes_allocated += region_size;
903 generations[gc_alloc_generation].bytes_allocated += region_size;
905 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
907 /* Set the generations alloc restart page to the last page of
910 generations[gc_alloc_generation].alloc_unboxed_start_page =
913 generations[gc_alloc_generation].alloc_start_page = next_page-1;
915 /* Add the region to the new_areas if requested. */
917 add_new_area(first_page,orig_first_page_bytes_used, region_size);
921 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
923 gc_alloc_generation));
926 /* There are no bytes allocated. Unallocate the first_page if
927 * there are 0 bytes_used. */
928 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
929 if (page_table[first_page].bytes_used == 0)
930 page_table[first_page].allocated = FREE_PAGE_FLAG;
933 /* Unallocate any unused pages. */
934 while (next_page <= alloc_region->last_page) {
935 gc_assert(page_table[next_page].bytes_used == 0);
936 page_table[next_page].allocated = FREE_PAGE_FLAG;
939 ret = thread_mutex_unlock(&free_pages_lock);
942 /* alloc_region is per-thread, we're ok to do this unlocked */
943 gc_set_region_empty(alloc_region);
946 static inline void *gc_quick_alloc(long nbytes);
948 /* Allocate a possibly large object. */
950 gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region)
952 page_index_t first_page;
953 page_index_t last_page;
954 int orig_first_page_bytes_used;
958 page_index_t next_page;
961 ret = thread_mutex_lock(&free_pages_lock);
966 generations[gc_alloc_generation].alloc_large_unboxed_start_page;
968 first_page = generations[gc_alloc_generation].alloc_large_start_page;
970 if (first_page <= alloc_region->last_page) {
971 first_page = alloc_region->last_page+1;
974 last_page=gc_find_freeish_pages(&first_page,nbytes,unboxed);
976 gc_assert(first_page > alloc_region->last_page);
978 generations[gc_alloc_generation].alloc_large_unboxed_start_page =
981 generations[gc_alloc_generation].alloc_large_start_page = last_page;
983 /* Set up the pages. */
984 orig_first_page_bytes_used = page_table[first_page].bytes_used;
986 /* If the first page was free then set up the gen, and
987 * first_object_offset. */
988 if (page_table[first_page].bytes_used == 0) {
990 page_table[first_page].allocated = UNBOXED_PAGE_FLAG;
992 page_table[first_page].allocated = BOXED_PAGE_FLAG;
993 page_table[first_page].gen = gc_alloc_generation;
994 page_table[first_page].first_object_offset = 0;
995 page_table[first_page].large_object = 1;
999 gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG);
1001 gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG);
1002 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1003 gc_assert(page_table[first_page].large_object == 1);
1007 /* Calc. the number of bytes used in this page. This is not
1008 * always the number of new bytes, unless it was free. */
1010 if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
1011 bytes_used = PAGE_BYTES;
1014 page_table[first_page].bytes_used = bytes_used;
1015 byte_cnt += bytes_used;
1017 next_page = first_page+1;
1019 /* All the rest of the pages should be free. We need to set their
1020 * first_object_offset pointer to the start of the region, and
1021 * set the bytes_used. */
1023 gc_assert(page_table[next_page].allocated == FREE_PAGE_FLAG);
1024 gc_assert(page_table[next_page].bytes_used == 0);
1026 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1028 page_table[next_page].allocated = BOXED_PAGE_FLAG;
1029 page_table[next_page].gen = gc_alloc_generation;
1030 page_table[next_page].large_object = 1;
1032 page_table[next_page].first_object_offset =
1033 orig_first_page_bytes_used - PAGE_BYTES*(next_page-first_page);
1035 /* Calculate the number of bytes used in this page. */
1037 if ((bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt) > PAGE_BYTES) {
1038 bytes_used = PAGE_BYTES;
1041 page_table[next_page].bytes_used = bytes_used;
1042 page_table[next_page].write_protected=0;
1043 page_table[next_page].dont_move=0;
1044 byte_cnt += bytes_used;
1048 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1050 bytes_allocated += nbytes;
1051 generations[gc_alloc_generation].bytes_allocated += nbytes;
1053 /* Add the region to the new_areas if requested. */
1055 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1057 /* Bump up last_free_page */
1058 if (last_page+1 > last_free_page) {
1059 last_free_page = last_page+1;
1060 set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
1062 ret = thread_mutex_unlock(&free_pages_lock);
1063 gc_assert(ret == 0);
1065 #ifdef READ_PROTECT_FREE_PAGES
1066 os_protect(page_address(first_page),
1067 PAGE_BYTES*(1+last_page-first_page),
1071 zero_dirty_pages(first_page, last_page);
1073 return page_address(first_page);
1076 static page_index_t gencgc_alloc_start_page = -1;
1079 gc_heap_exhausted_error_or_lose (long available, long requested)
1081 /* Write basic information before doing anything else: if we don't
1082 * call to lisp this is a must, and even if we do there is always
1083 * the danger that we bounce back here before the error has been
1084 * handled, or indeed even printed.
1086 fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
1087 gc_active_p ? "garbage collection" : "allocation", available, requested);
1088 if (gc_active_p || (available == 0)) {
1089 /* If we are in GC, or totally out of memory there is no way
1090 * to sanely transfer control to the lisp-side of things.
1092 struct thread *thread = arch_os_get_current_thread();
1093 print_generation_stats(1);
1094 fprintf(stderr, "GC control variables:\n");
1095 fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
1096 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
1097 SymbolValue(GC_PENDING,thread)==NIL ? "false" : "true");
1098 #ifdef LISP_FEATURE_SB_THREAD
1099 fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n",
1100 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
1102 lose("Heap exhausted, game over.");
1105 /* FIXME: assert free_pages_lock held */
1106 (void)thread_mutex_unlock(&free_pages_lock);
1107 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1108 alloc_number(available), alloc_number(requested));
1109 lose("HEAP-EXHAUSTED-ERROR fell through");
1114 gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed)
1116 page_index_t first_page, last_page;
1117 page_index_t restart_page = *restart_page_ptr;
1118 long bytes_found = 0;
1119 long most_bytes_found = 0;
1120 /* FIXME: assert(free_pages_lock is held); */
1122 /* Toggled by gc_and_save for heap compaction, normally -1. */
1123 if (gencgc_alloc_start_page != -1) {
1124 restart_page = gencgc_alloc_start_page;
1127 if (nbytes>=PAGE_BYTES) {
1128 /* Search for a contiguous free space of at least nbytes,
1129 * aligned on a page boundary. The page-alignment is strictly
1130 * speaking needed only for objects at least large_object_size
1133 first_page = restart_page;
1134 while ((first_page < page_table_pages) &&
1135 (page_table[first_page].allocated != FREE_PAGE_FLAG))
1138 last_page = first_page;
1139 bytes_found = PAGE_BYTES;
1140 while ((bytes_found < nbytes) &&
1141 (last_page < (page_table_pages-1)) &&
1142 (page_table[last_page+1].allocated == FREE_PAGE_FLAG)) {
1144 bytes_found += PAGE_BYTES;
1145 gc_assert(page_table[last_page].write_protected == 0);
1147 if (bytes_found > most_bytes_found)
1148 most_bytes_found = bytes_found;
1149 restart_page = last_page + 1;
1150 } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
1153 /* Search for a page with at least nbytes of space. We prefer
1154 * not to split small objects on multiple pages, to reduce the
1155 * number of contiguous allocation regions spaning multiple
1156 * pages: this helps avoid excessive conservativism. */
1157 first_page = restart_page;
1158 while (first_page < page_table_pages) {
1159 if (page_table[first_page].allocated == FREE_PAGE_FLAG)
1161 bytes_found = PAGE_BYTES;
1164 else if ((page_table[first_page].allocated ==
1165 (unboxed ? UNBOXED_PAGE_FLAG : BOXED_PAGE_FLAG)) &&
1166 (page_table[first_page].large_object == 0) &&
1167 (page_table[first_page].gen == gc_alloc_generation) &&
1168 (page_table[first_page].write_protected == 0) &&
1169 (page_table[first_page].dont_move == 0))
1171 bytes_found = PAGE_BYTES - page_table[first_page].bytes_used;
1172 if (bytes_found > most_bytes_found)
1173 most_bytes_found = bytes_found;
1174 if (bytes_found >= nbytes)
1179 last_page = first_page;
1180 restart_page = first_page + 1;
1183 /* Check for a failure */
1184 if (bytes_found < nbytes) {
1185 gc_assert(restart_page >= page_table_pages);
1186 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1189 gc_assert(page_table[first_page].write_protected == 0);
1191 *restart_page_ptr = first_page;
1195 /* Allocate bytes. All the rest of the special-purpose allocation
1196 * functions will eventually call this */
1199 gc_alloc_with_region(long nbytes,int unboxed_p, struct alloc_region *my_region,
1202 void *new_free_pointer;
1204 if (nbytes>=large_object_size)
1205 return gc_alloc_large(nbytes,unboxed_p,my_region);
1207 /* Check whether there is room in the current alloc region. */
1208 new_free_pointer = my_region->free_pointer + nbytes;
1210 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1211 my_region->free_pointer, new_free_pointer); */
1213 if (new_free_pointer <= my_region->end_addr) {
1214 /* If so then allocate from the current alloc region. */
1215 void *new_obj = my_region->free_pointer;
1216 my_region->free_pointer = new_free_pointer;
1218 /* Unless a `quick' alloc was requested, check whether the
1219 alloc region is almost empty. */
1221 (my_region->end_addr - my_region->free_pointer) <= 32) {
1222 /* If so, finished with the current region. */
1223 gc_alloc_update_page_tables(unboxed_p, my_region);
1224 /* Set up a new region. */
1225 gc_alloc_new_region(32 /*bytes*/, unboxed_p, my_region);
1228 return((void *)new_obj);
1231 /* Else not enough free space in the current region: retry with a
1234 gc_alloc_update_page_tables(unboxed_p, my_region);
1235 gc_alloc_new_region(nbytes, unboxed_p, my_region);
1236 return gc_alloc_with_region(nbytes,unboxed_p,my_region,0);
1239 /* these are only used during GC: all allocation from the mutator calls
1240 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1244 gc_general_alloc(long nbytes,int unboxed_p,int quick_p)
1246 struct alloc_region *my_region =
1247 unboxed_p ? &unboxed_region : &boxed_region;
1248 return gc_alloc_with_region(nbytes,unboxed_p, my_region,quick_p);
1251 static inline void *
1252 gc_quick_alloc(long nbytes)
1254 return gc_general_alloc(nbytes,ALLOC_BOXED,ALLOC_QUICK);
1257 static inline void *
1258 gc_quick_alloc_large(long nbytes)
1260 return gc_general_alloc(nbytes,ALLOC_BOXED,ALLOC_QUICK);
1263 static inline void *
1264 gc_alloc_unboxed(long nbytes)
1266 return gc_general_alloc(nbytes,ALLOC_UNBOXED,0);
1269 static inline void *
1270 gc_quick_alloc_unboxed(long nbytes)
1272 return gc_general_alloc(nbytes,ALLOC_UNBOXED,ALLOC_QUICK);
1275 static inline void *
1276 gc_quick_alloc_large_unboxed(long nbytes)
1278 return gc_general_alloc(nbytes,ALLOC_UNBOXED,ALLOC_QUICK);
1282 * scavenging/transporting routines derived from gc.c in CMU CL ca. 18b
1285 extern long (*scavtab[256])(lispobj *where, lispobj object);
1286 extern lispobj (*transother[256])(lispobj object);
1287 extern long (*sizetab[256])(lispobj *where);
1289 /* Copy a large boxed object. If the object is in a large object
1290 * region then it is simply promoted, else it is copied. If it's large
1291 * enough then it's copied to a large object region.
1293 * Vectors may have shrunk. If the object is not copied the space
1294 * needs to be reclaimed, and the page_tables corrected. */
1296 copy_large_object(lispobj object, long nwords)
1300 page_index_t first_page;
1302 gc_assert(is_lisp_pointer(object));
1303 gc_assert(from_space_p(object));
1304 gc_assert((nwords & 0x01) == 0);
1307 /* Check whether it's in a large object region. */
1308 first_page = find_page_index((void *)object);
1309 gc_assert(first_page >= 0);
1311 if (page_table[first_page].large_object) {
1313 /* Promote the object. */
1315 long remaining_bytes;
1316 page_index_t next_page;
1318 long old_bytes_used;
1320 /* Note: Any page write-protection must be removed, else a
1321 * later scavenge_newspace may incorrectly not scavenge these
1322 * pages. This would not be necessary if they are added to the
1323 * new areas, but let's do it for them all (they'll probably
1324 * be written anyway?). */
1326 gc_assert(page_table[first_page].first_object_offset == 0);
1328 next_page = first_page;
1329 remaining_bytes = nwords*N_WORD_BYTES;
1330 while (remaining_bytes > PAGE_BYTES) {
1331 gc_assert(page_table[next_page].gen == from_space);
1332 gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG);
1333 gc_assert(page_table[next_page].large_object);
1334 gc_assert(page_table[next_page].first_object_offset==
1335 -PAGE_BYTES*(next_page-first_page));
1336 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1338 page_table[next_page].gen = new_space;
1340 /* Remove any write-protection. We should be able to rely
1341 * on the write-protect flag to avoid redundant calls. */
1342 if (page_table[next_page].write_protected) {
1343 os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL);
1344 page_table[next_page].write_protected = 0;
1346 remaining_bytes -= PAGE_BYTES;
1350 /* Now only one page remains, but the object may have shrunk
1351 * so there may be more unused pages which will be freed. */
1353 /* The object may have shrunk but shouldn't have grown. */
1354 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1356 page_table[next_page].gen = new_space;
1357 gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG);
1359 /* Adjust the bytes_used. */
1360 old_bytes_used = page_table[next_page].bytes_used;
1361 page_table[next_page].bytes_used = remaining_bytes;
1363 bytes_freed = old_bytes_used - remaining_bytes;
1365 /* Free any remaining pages; needs care. */
1367 while ((old_bytes_used == PAGE_BYTES) &&
1368 (page_table[next_page].gen == from_space) &&
1369 (page_table[next_page].allocated == BOXED_PAGE_FLAG) &&
1370 page_table[next_page].large_object &&
1371 (page_table[next_page].first_object_offset ==
1372 -(next_page - first_page)*PAGE_BYTES)) {
1373 /* Checks out OK, free the page. Don't need to bother zeroing
1374 * pages as this should have been done before shrinking the
1375 * object. These pages shouldn't be write-protected as they
1376 * should be zero filled. */
1377 gc_assert(page_table[next_page].write_protected == 0);
1379 old_bytes_used = page_table[next_page].bytes_used;
1380 page_table[next_page].allocated = FREE_PAGE_FLAG;
1381 page_table[next_page].bytes_used = 0;
1382 bytes_freed += old_bytes_used;
1386 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords +
1388 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1389 bytes_allocated -= bytes_freed;
1391 /* Add the region to the new_areas if requested. */
1392 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1396 /* Get tag of object. */
1397 tag = lowtag_of(object);
1399 /* Allocate space. */
1400 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1402 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1404 /* Return Lisp pointer of new object. */
1405 return ((lispobj) new) | tag;
1409 /* to copy unboxed objects */
1411 copy_unboxed_object(lispobj object, long nwords)
1416 gc_assert(is_lisp_pointer(object));
1417 gc_assert(from_space_p(object));
1418 gc_assert((nwords & 0x01) == 0);
1420 /* Get tag of object. */
1421 tag = lowtag_of(object);
1423 /* Allocate space. */
1424 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1426 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1428 /* Return Lisp pointer of new object. */
1429 return ((lispobj) new) | tag;
1432 /* to copy large unboxed objects
1434 * If the object is in a large object region then it is simply
1435 * promoted, else it is copied. If it's large enough then it's copied
1436 * to a large object region.
1438 * Bignums and vectors may have shrunk. If the object is not copied
1439 * the space needs to be reclaimed, and the page_tables corrected.
1441 * KLUDGE: There's a lot of cut-and-paste duplication between this
1442 * function and copy_large_object(..). -- WHN 20000619 */
1444 copy_large_unboxed_object(lispobj object, long nwords)
1448 page_index_t first_page;
1450 gc_assert(is_lisp_pointer(object));
1451 gc_assert(from_space_p(object));
1452 gc_assert((nwords & 0x01) == 0);
1454 if ((nwords > 1024*1024) && gencgc_verbose)
1455 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n", nwords*N_WORD_BYTES));
1457 /* Check whether it's a large object. */
1458 first_page = find_page_index((void *)object);
1459 gc_assert(first_page >= 0);
1461 if (page_table[first_page].large_object) {
1462 /* Promote the object. Note: Unboxed objects may have been
1463 * allocated to a BOXED region so it may be necessary to
1464 * change the region to UNBOXED. */
1465 long remaining_bytes;
1466 page_index_t next_page;
1468 long old_bytes_used;
1470 gc_assert(page_table[first_page].first_object_offset == 0);
1472 next_page = first_page;
1473 remaining_bytes = nwords*N_WORD_BYTES;
1474 while (remaining_bytes > PAGE_BYTES) {
1475 gc_assert(page_table[next_page].gen == from_space);
1476 gc_assert((page_table[next_page].allocated == UNBOXED_PAGE_FLAG)
1477 || (page_table[next_page].allocated == BOXED_PAGE_FLAG));
1478 gc_assert(page_table[next_page].large_object);
1479 gc_assert(page_table[next_page].first_object_offset==
1480 -PAGE_BYTES*(next_page-first_page));
1481 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
1483 page_table[next_page].gen = new_space;
1484 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1485 remaining_bytes -= PAGE_BYTES;
1489 /* Now only one page remains, but the object may have shrunk so
1490 * there may be more unused pages which will be freed. */
1492 /* Object may have shrunk but shouldn't have grown - check. */
1493 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1495 page_table[next_page].gen = new_space;
1496 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1498 /* Adjust the bytes_used. */
1499 old_bytes_used = page_table[next_page].bytes_used;
1500 page_table[next_page].bytes_used = remaining_bytes;
1502 bytes_freed = old_bytes_used - remaining_bytes;
1504 /* Free any remaining pages; needs care. */
1506 while ((old_bytes_used == PAGE_BYTES) &&
1507 (page_table[next_page].gen == from_space) &&
1508 ((page_table[next_page].allocated == UNBOXED_PAGE_FLAG)
1509 || (page_table[next_page].allocated == BOXED_PAGE_FLAG)) &&
1510 page_table[next_page].large_object &&
1511 (page_table[next_page].first_object_offset ==
1512 -(next_page - first_page)*PAGE_BYTES)) {
1513 /* Checks out OK, free the page. Don't need to both zeroing
1514 * pages as this should have been done before shrinking the
1515 * object. These pages shouldn't be write-protected, even if
1516 * boxed they should be zero filled. */
1517 gc_assert(page_table[next_page].write_protected == 0);
1519 old_bytes_used = page_table[next_page].bytes_used;
1520 page_table[next_page].allocated = FREE_PAGE_FLAG;
1521 page_table[next_page].bytes_used = 0;
1522 bytes_freed += old_bytes_used;
1526 if ((bytes_freed > 0) && gencgc_verbose)
1528 "/copy_large_unboxed bytes_freed=%d\n",
1531 generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES + bytes_freed;
1532 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1533 bytes_allocated -= bytes_freed;
1538 /* Get tag of object. */
1539 tag = lowtag_of(object);
1541 /* Allocate space. */
1542 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1544 /* Copy the object. */
1545 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1547 /* Return Lisp pointer of new object. */
1548 return ((lispobj) new) | tag;
1557 * code and code-related objects
1560 static lispobj trans_fun_header(lispobj object);
1561 static lispobj trans_boxed(lispobj object);
1564 /* Scan a x86 compiled code object, looking for possible fixups that
1565 * have been missed after a move.
1567 * Two types of fixups are needed:
1568 * 1. Absolute fixups to within the code object.
1569 * 2. Relative fixups to outside the code object.
1571 * Currently only absolute fixups to the constant vector, or to the
1572 * code area are checked. */
1574 sniff_code_object(struct code *code, unsigned long displacement)
1576 #ifdef LISP_FEATURE_X86
1577 long nheader_words, ncode_words, nwords;
1579 void *constants_start_addr = NULL, *constants_end_addr;
1580 void *code_start_addr, *code_end_addr;
1581 int fixup_found = 0;
1583 if (!check_code_fixups)
1586 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1588 ncode_words = fixnum_value(code->code_size);
1589 nheader_words = HeaderValue(*(lispobj *)code);
1590 nwords = ncode_words + nheader_words;
1592 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1593 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1594 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1595 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1597 /* Work through the unboxed code. */
1598 for (p = code_start_addr; p < code_end_addr; p++) {
1599 void *data = *(void **)p;
1600 unsigned d1 = *((unsigned char *)p - 1);
1601 unsigned d2 = *((unsigned char *)p - 2);
1602 unsigned d3 = *((unsigned char *)p - 3);
1603 unsigned d4 = *((unsigned char *)p - 4);
1605 unsigned d5 = *((unsigned char *)p - 5);
1606 unsigned d6 = *((unsigned char *)p - 6);
1609 /* Check for code references. */
1610 /* Check for a 32 bit word that looks like an absolute
1611 reference to within the code adea of the code object. */
1612 if ((data >= (code_start_addr-displacement))
1613 && (data < (code_end_addr-displacement))) {
1614 /* function header */
1616 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) == (unsigned)code)) {
1617 /* Skip the function header */
1621 /* the case of PUSH imm32 */
1625 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1626 p, d6, d5, d4, d3, d2, d1, data));
1627 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1629 /* the case of MOV [reg-8],imm32 */
1631 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1632 || d2==0x45 || d2==0x46 || d2==0x47)
1636 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1637 p, d6, d5, d4, d3, d2, d1, data));
1638 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1640 /* the case of LEA reg,[disp32] */
1641 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1644 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1645 p, d6, d5, d4, d3, d2, d1, data));
1646 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1650 /* Check for constant references. */
1651 /* Check for a 32 bit word that looks like an absolute
1652 reference to within the constant vector. Constant references
1654 if ((data >= (constants_start_addr-displacement))
1655 && (data < (constants_end_addr-displacement))
1656 && (((unsigned)data & 0x3) == 0)) {
1661 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1662 p, d6, d5, d4, d3, d2, d1, data));
1663 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1666 /* the case of MOV m32,EAX */
1670 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1671 p, d6, d5, d4, d3, d2, d1, data));
1672 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1675 /* the case of CMP m32,imm32 */
1676 if ((d1 == 0x3d) && (d2 == 0x81)) {
1679 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1680 p, d6, d5, d4, d3, d2, d1, data));
1682 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1685 /* Check for a mod=00, r/m=101 byte. */
1686 if ((d1 & 0xc7) == 5) {
1691 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1692 p, d6, d5, d4, d3, d2, d1, data));
1693 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1695 /* the case of CMP reg32,m32 */
1699 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1700 p, d6, d5, d4, d3, d2, d1, data));
1701 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1703 /* the case of MOV m32,reg32 */
1707 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1708 p, d6, d5, d4, d3, d2, d1, data));
1709 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1711 /* the case of MOV reg32,m32 */
1715 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1716 p, d6, d5, d4, d3, d2, d1, data));
1717 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1719 /* the case of LEA reg32,m32 */
1723 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1724 p, d6, d5, d4, d3, d2, d1, data));
1725 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1731 /* If anything was found, print some information on the code
1735 "/compiled code object at %x: header words = %d, code words = %d\n",
1736 code, nheader_words, ncode_words));
1738 "/const start = %x, end = %x\n",
1739 constants_start_addr, constants_end_addr));
1741 "/code start = %x, end = %x\n",
1742 code_start_addr, code_end_addr));
1748 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1750 /* x86-64 uses pc-relative addressing instead of this kludge */
1751 #ifndef LISP_FEATURE_X86_64
1752 long nheader_words, ncode_words, nwords;
1753 void *constants_start_addr, *constants_end_addr;
1754 void *code_start_addr, *code_end_addr;
1755 lispobj fixups = NIL;
1756 unsigned long displacement = (unsigned long)new_code - (unsigned long)old_code;
1757 struct vector *fixups_vector;
1759 ncode_words = fixnum_value(new_code->code_size);
1760 nheader_words = HeaderValue(*(lispobj *)new_code);
1761 nwords = ncode_words + nheader_words;
1763 "/compiled code object at %x: header words = %d, code words = %d\n",
1764 new_code, nheader_words, ncode_words)); */
1765 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1766 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1767 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1768 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1771 "/const start = %x, end = %x\n",
1772 constants_start_addr,constants_end_addr));
1774 "/code start = %x; end = %x\n",
1775 code_start_addr,code_end_addr));
1778 /* The first constant should be a pointer to the fixups for this
1779 code objects. Check. */
1780 fixups = new_code->constants[0];
1782 /* It will be 0 or the unbound-marker if there are no fixups (as
1783 * will be the case if the code object has been purified, for
1784 * example) and will be an other pointer if it is valid. */
1785 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1786 !is_lisp_pointer(fixups)) {
1787 /* Check for possible errors. */
1788 if (check_code_fixups)
1789 sniff_code_object(new_code, displacement);
1794 fixups_vector = (struct vector *)native_pointer(fixups);
1796 /* Could be pointing to a forwarding pointer. */
1797 /* FIXME is this always in from_space? if so, could replace this code with
1798 * forwarding_pointer_p/forwarding_pointer_value */
1799 if (is_lisp_pointer(fixups) &&
1800 (find_page_index((void*)fixups_vector) != -1) &&
1801 (fixups_vector->header == 0x01)) {
1802 /* If so, then follow it. */
1803 /*SHOW("following pointer to a forwarding pointer");*/
1804 fixups_vector = (struct vector *)native_pointer((lispobj)fixups_vector->length);
1807 /*SHOW("got fixups");*/
1809 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1810 /* Got the fixups for the code block. Now work through the vector,
1811 and apply a fixup at each address. */
1812 long length = fixnum_value(fixups_vector->length);
1814 for (i = 0; i < length; i++) {
1815 unsigned long offset = fixups_vector->data[i];
1816 /* Now check the current value of offset. */
1817 unsigned long old_value =
1818 *(unsigned long *)((unsigned long)code_start_addr + offset);
1820 /* If it's within the old_code object then it must be an
1821 * absolute fixup (relative ones are not saved) */
1822 if ((old_value >= (unsigned long)old_code)
1823 && (old_value < ((unsigned long)old_code + nwords*N_WORD_BYTES)))
1824 /* So add the dispacement. */
1825 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1826 old_value + displacement;
1828 /* It is outside the old code object so it must be a
1829 * relative fixup (absolute fixups are not saved). So
1830 * subtract the displacement. */
1831 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1832 old_value - displacement;
1835 /* This used to just print a note to stderr, but a bogus fixup seems to
1836 * indicate real heap corruption, so a hard hailure is in order. */
1837 lose("fixup vector %p has a bad widetag: %d\n", fixups_vector, widetag_of(fixups_vector->header));
1840 /* Check for possible errors. */
1841 if (check_code_fixups) {
1842 sniff_code_object(new_code,displacement);
1849 trans_boxed_large(lispobj object)
1852 unsigned long length;
1854 gc_assert(is_lisp_pointer(object));
1856 header = *((lispobj *) native_pointer(object));
1857 length = HeaderValue(header) + 1;
1858 length = CEILING(length, 2);
1860 return copy_large_object(object, length);
1863 /* Doesn't seem to be used, delete it after the grace period. */
1866 trans_unboxed_large(lispobj object)
1869 unsigned long length;
1871 gc_assert(is_lisp_pointer(object));
1873 header = *((lispobj *) native_pointer(object));
1874 length = HeaderValue(header) + 1;
1875 length = CEILING(length, 2);
1877 return copy_large_unboxed_object(object, length);
1883 * Lutexes. Using the normal finalization machinery for finalizing
1884 * lutexes is tricky, since the finalization depends on working lutexes.
1885 * So we track the lutexes in the GC and finalize them manually.
1888 #if defined(LUTEX_WIDETAG)
1891 * Start tracking LUTEX in the GC, by adding it to the linked list of
1892 * lutexes in the nursery generation. The caller is responsible for
1893 * locking, and GCs must be inhibited until the registration is
1897 gencgc_register_lutex (struct lutex *lutex) {
1898 int index = find_page_index(lutex);
1899 generation_index_t gen;
1902 /* This lutex is in static space, so we don't need to worry about
1908 gen = page_table[index].gen;
1910 gc_assert(gen >= 0);
1911 gc_assert(gen < NUM_GENERATIONS);
1913 head = generations[gen].lutexes;
1920 generations[gen].lutexes = lutex;
1924 * Stop tracking LUTEX in the GC by removing it from the appropriate
1925 * linked lists. This will only be called during GC, so no locking is
1929 gencgc_unregister_lutex (struct lutex *lutex) {
1931 lutex->prev->next = lutex->next;
1933 generations[lutex->gen].lutexes = lutex->next;
1937 lutex->next->prev = lutex->prev;
1946 * Mark all lutexes in generation GEN as not live.
1949 unmark_lutexes (generation_index_t gen) {
1950 struct lutex *lutex = generations[gen].lutexes;
1954 lutex = lutex->next;
1959 * Finalize all lutexes in generation GEN that have not been marked live.
1962 reap_lutexes (generation_index_t gen) {
1963 struct lutex *lutex = generations[gen].lutexes;
1966 struct lutex *next = lutex->next;
1968 lutex_destroy((tagged_lutex_t) lutex);
1969 gencgc_unregister_lutex(lutex);
1976 * Mark LUTEX as live.
1979 mark_lutex (lispobj tagged_lutex) {
1980 struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
1986 * Move all lutexes in generation FROM to generation TO.
1989 move_lutexes (generation_index_t from, generation_index_t to) {
1990 struct lutex *tail = generations[from].lutexes;
1992 /* Nothing to move */
1996 /* Change the generation of the lutexes in FROM. */
1997 while (tail->next) {
2003 /* Link the last lutex in the FROM list to the start of the TO list */
2004 tail->next = generations[to].lutexes;
2006 /* And vice versa */
2007 if (generations[to].lutexes) {
2008 generations[to].lutexes->prev = tail;
2011 /* And update the generations structures to match this */
2012 generations[to].lutexes = generations[from].lutexes;
2013 generations[from].lutexes = NULL;
2017 scav_lutex(lispobj *where, lispobj object)
2019 mark_lutex((lispobj) where);
2021 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2025 trans_lutex(lispobj object)
2027 struct lutex *lutex = (struct lutex *) native_pointer(object);
2029 size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2030 gc_assert(is_lisp_pointer(object));
2031 copied = copy_object(object, words);
2033 /* Update the links, since the lutex moved in memory. */
2035 lutex->next->prev = (struct lutex *) native_pointer(copied);
2039 lutex->prev->next = (struct lutex *) native_pointer(copied);
2041 generations[lutex->gen].lutexes =
2042 (struct lutex *) native_pointer(copied);
2049 size_lutex(lispobj *where)
2051 return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
2053 #endif /* LUTEX_WIDETAG */
2060 /* XX This is a hack adapted from cgc.c. These don't work too
2061 * efficiently with the gencgc as a list of the weak pointers is
2062 * maintained within the objects which causes writes to the pages. A
2063 * limited attempt is made to avoid unnecessary writes, but this needs
2065 #define WEAK_POINTER_NWORDS \
2066 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2069 scav_weak_pointer(lispobj *where, lispobj object)
2071 /* Since we overwrite the 'next' field, we have to make
2072 * sure not to do so for pointers already in the list.
2073 * Instead of searching the list of weak_pointers each
2074 * time, we ensure that next is always NULL when the weak
2075 * pointer isn't in the list, and not NULL otherwise.
2076 * Since we can't use NULL to denote end of list, we
2077 * use a pointer back to the same weak_pointer.
2079 struct weak_pointer * wp = (struct weak_pointer*)where;
2081 if (NULL == wp->next) {
2082 wp->next = weak_pointers;
2084 if (NULL == wp->next)
2088 /* Do not let GC scavenge the value slot of the weak pointer.
2089 * (That is why it is a weak pointer.) */
2091 return WEAK_POINTER_NWORDS;
2096 search_read_only_space(void *pointer)
2098 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2099 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2100 if ((pointer < (void *)start) || (pointer >= (void *)end))
2102 return (gc_search_space(start,
2103 (((lispobj *)pointer)+2)-start,
2104 (lispobj *) pointer));
2108 search_static_space(void *pointer)
2110 lispobj *start = (lispobj *)STATIC_SPACE_START;
2111 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2112 if ((pointer < (void *)start) || (pointer >= (void *)end))
2114 return (gc_search_space(start,
2115 (((lispobj *)pointer)+2)-start,
2116 (lispobj *) pointer));
2119 /* a faster version for searching the dynamic space. This will work even
2120 * if the object is in a current allocation region. */
2122 search_dynamic_space(void *pointer)
2124 page_index_t page_index = find_page_index(pointer);
2127 /* The address may be invalid, so do some checks. */
2128 if ((page_index == -1) ||
2129 (page_table[page_index].allocated == FREE_PAGE_FLAG))
2131 start = (lispobj *)page_region_start(page_index);
2132 return (gc_search_space(start,
2133 (((lispobj *)pointer)+2)-start,
2134 (lispobj *)pointer));
2137 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2139 /* Helper for valid_lisp_pointer_p and
2140 * possibly_valid_dynamic_space_pointer.
2142 * pointer is the pointer to validate, and start_addr is the address
2143 * of the enclosing object.
2146 looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr)
2148 /* We need to allow raw pointers into Code objects for return
2149 * addresses. This will also pick up pointers to functions in code
2151 if (widetag_of(*start_addr) == CODE_HEADER_WIDETAG)
2152 /* XXX could do some further checks here */
2155 if (!is_lisp_pointer((lispobj)pointer)) {
2159 /* Check that the object pointed to is consistent with the pointer
2161 switch (lowtag_of((lispobj)pointer)) {
2162 case FUN_POINTER_LOWTAG:
2163 /* Start_addr should be the enclosing code object, or a closure
2165 switch (widetag_of(*start_addr)) {
2166 case CODE_HEADER_WIDETAG:
2167 /* This case is probably caught above. */
2169 case CLOSURE_HEADER_WIDETAG:
2170 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2171 if ((unsigned long)pointer !=
2172 ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) {
2176 pointer, start_addr, *start_addr));
2184 pointer, start_addr, *start_addr));
2188 case LIST_POINTER_LOWTAG:
2189 if ((unsigned long)pointer !=
2190 ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) {
2194 pointer, start_addr, *start_addr));
2197 /* Is it plausible cons? */
2198 if ((is_lisp_pointer(start_addr[0]) || is_lisp_immediate(start_addr[0])) &&
2199 (is_lisp_pointer(start_addr[1]) || is_lisp_immediate(start_addr[1])))
2205 pointer, start_addr, *start_addr));
2208 case INSTANCE_POINTER_LOWTAG:
2209 if ((unsigned long)pointer !=
2210 ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) {
2214 pointer, start_addr, *start_addr));
2217 if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) {
2221 pointer, start_addr, *start_addr));
2225 case OTHER_POINTER_LOWTAG:
2226 if ((unsigned long)pointer !=
2227 ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) {
2231 pointer, start_addr, *start_addr));
2234 /* Is it plausible? Not a cons. XXX should check the headers. */
2235 if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) {
2239 pointer, start_addr, *start_addr));
2242 switch (widetag_of(start_addr[0])) {
2243 case UNBOUND_MARKER_WIDETAG:
2244 case NO_TLS_VALUE_MARKER_WIDETAG:
2245 case CHARACTER_WIDETAG:
2246 #if N_WORD_BITS == 64
2247 case SINGLE_FLOAT_WIDETAG:
2252 pointer, start_addr, *start_addr));
2255 /* only pointed to by function pointers? */
2256 case CLOSURE_HEADER_WIDETAG:
2257 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2261 pointer, start_addr, *start_addr));
2264 case INSTANCE_HEADER_WIDETAG:
2268 pointer, start_addr, *start_addr));
2271 /* the valid other immediate pointer objects */
2272 case SIMPLE_VECTOR_WIDETAG:
2274 case COMPLEX_WIDETAG:
2275 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
2276 case COMPLEX_SINGLE_FLOAT_WIDETAG:
2278 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
2279 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
2281 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
2282 case COMPLEX_LONG_FLOAT_WIDETAG:
2284 case SIMPLE_ARRAY_WIDETAG:
2285 case COMPLEX_BASE_STRING_WIDETAG:
2286 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2287 case COMPLEX_CHARACTER_STRING_WIDETAG:
2289 case COMPLEX_VECTOR_NIL_WIDETAG:
2290 case COMPLEX_BIT_VECTOR_WIDETAG:
2291 case COMPLEX_VECTOR_WIDETAG:
2292 case COMPLEX_ARRAY_WIDETAG:
2293 case VALUE_CELL_HEADER_WIDETAG:
2294 case SYMBOL_HEADER_WIDETAG:
2296 case CODE_HEADER_WIDETAG:
2297 case BIGNUM_WIDETAG:
2298 #if N_WORD_BITS != 64
2299 case SINGLE_FLOAT_WIDETAG:
2301 case DOUBLE_FLOAT_WIDETAG:
2302 #ifdef LONG_FLOAT_WIDETAG
2303 case LONG_FLOAT_WIDETAG:
2305 case SIMPLE_BASE_STRING_WIDETAG:
2306 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2307 case SIMPLE_CHARACTER_STRING_WIDETAG:
2309 case SIMPLE_BIT_VECTOR_WIDETAG:
2310 case SIMPLE_ARRAY_NIL_WIDETAG:
2311 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2312 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2313 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2314 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2315 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2316 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2317 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2318 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2320 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2321 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2322 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2323 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2325 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2326 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2328 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2329 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2331 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2332 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2334 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2335 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2337 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2338 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2340 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2341 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2343 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2344 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2346 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2347 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2349 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2350 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2351 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2352 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2354 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2355 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2357 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2358 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2360 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2361 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2364 case WEAK_POINTER_WIDETAG:
2365 #ifdef LUTEX_WIDETAG
2374 pointer, start_addr, *start_addr));
2382 pointer, start_addr, *start_addr));
2390 /* Used by the debugger to validate possibly bogus pointers before
2391 * calling MAKE-LISP-OBJ on them.
2393 * FIXME: We would like to make this perfect, because if the debugger
2394 * constructs a reference to a bugs lisp object, and it ends up in a
2395 * location scavenged by the GC all hell breaks loose.
2397 * Whereas possibly_valid_dynamic_space_pointer has to be conservative
2398 * and return true for all valid pointers, this could actually be eager
2399 * and lie about a few pointers without bad results... but that should
2400 * be reflected in the name.
2403 valid_lisp_pointer_p(lispobj *pointer)
2406 if (((start=search_dynamic_space(pointer))!=NULL) ||
2407 ((start=search_static_space(pointer))!=NULL) ||
2408 ((start=search_read_only_space(pointer))!=NULL))
2409 return looks_like_valid_lisp_pointer_p(pointer, start);
2414 /* Is there any possibility that pointer is a valid Lisp object
2415 * reference, and/or something else (e.g. subroutine call return
2416 * address) which should prevent us from moving the referred-to thing?
2417 * This is called from preserve_pointers() */
2419 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2421 lispobj *start_addr;
2423 /* Find the object start address. */
2424 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2428 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2431 /* Adjust large bignum and vector objects. This will adjust the
2432 * allocated region if the size has shrunk, and move unboxed objects
2433 * into unboxed pages. The pages are not promoted here, and the
2434 * promoted region is not added to the new_regions; this is really
2435 * only designed to be called from preserve_pointer(). Shouldn't fail
2436 * if this is missed, just may delay the moving of objects to unboxed
2437 * pages, and the freeing of pages. */
2439 maybe_adjust_large_object(lispobj *where)
2441 page_index_t first_page;
2442 page_index_t next_page;
2445 long remaining_bytes;
2447 long old_bytes_used;
2451 /* Check whether it's a vector or bignum object. */
2452 switch (widetag_of(where[0])) {
2453 case SIMPLE_VECTOR_WIDETAG:
2454 boxed = BOXED_PAGE_FLAG;
2456 case BIGNUM_WIDETAG:
2457 case SIMPLE_BASE_STRING_WIDETAG:
2458 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2459 case SIMPLE_CHARACTER_STRING_WIDETAG:
2461 case SIMPLE_BIT_VECTOR_WIDETAG:
2462 case SIMPLE_ARRAY_NIL_WIDETAG:
2463 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2464 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2465 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2466 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2467 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2468 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2469 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
2470 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
2472 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2473 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2474 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
2475 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
2477 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2478 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2480 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2481 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2483 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2484 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2486 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2487 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2489 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
2490 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
2492 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2493 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2495 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
2496 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
2498 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2499 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2501 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2502 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2503 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2504 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2506 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2507 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2509 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2510 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2512 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2513 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2515 boxed = UNBOXED_PAGE_FLAG;
2521 /* Find its current size. */
2522 nwords = (sizetab[widetag_of(where[0])])(where);
2524 first_page = find_page_index((void *)where);
2525 gc_assert(first_page >= 0);
2527 /* Note: Any page write-protection must be removed, else a later
2528 * scavenge_newspace may incorrectly not scavenge these pages.
2529 * This would not be necessary if they are added to the new areas,
2530 * but lets do it for them all (they'll probably be written
2533 gc_assert(page_table[first_page].first_object_offset == 0);
2535 next_page = first_page;
2536 remaining_bytes = nwords*N_WORD_BYTES;
2537 while (remaining_bytes > PAGE_BYTES) {
2538 gc_assert(page_table[next_page].gen == from_space);
2539 gc_assert((page_table[next_page].allocated == BOXED_PAGE_FLAG)
2540 || (page_table[next_page].allocated == UNBOXED_PAGE_FLAG));
2541 gc_assert(page_table[next_page].large_object);
2542 gc_assert(page_table[next_page].first_object_offset ==
2543 -PAGE_BYTES*(next_page-first_page));
2544 gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
2546 page_table[next_page].allocated = boxed;
2548 /* Shouldn't be write-protected at this stage. Essential that the
2550 gc_assert(!page_table[next_page].write_protected);
2551 remaining_bytes -= PAGE_BYTES;
2555 /* Now only one page remains, but the object may have shrunk so
2556 * there may be more unused pages which will be freed. */
2558 /* Object may have shrunk but shouldn't have grown - check. */
2559 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2561 page_table[next_page].allocated = boxed;
2562 gc_assert(page_table[next_page].allocated ==
2563 page_table[first_page].allocated);
2565 /* Adjust the bytes_used. */
2566 old_bytes_used = page_table[next_page].bytes_used;
2567 page_table[next_page].bytes_used = remaining_bytes;
2569 bytes_freed = old_bytes_used - remaining_bytes;
2571 /* Free any remaining pages; needs care. */
2573 while ((old_bytes_used == PAGE_BYTES) &&
2574 (page_table[next_page].gen == from_space) &&
2575 ((page_table[next_page].allocated == UNBOXED_PAGE_FLAG)
2576 || (page_table[next_page].allocated == BOXED_PAGE_FLAG)) &&
2577 page_table[next_page].large_object &&
2578 (page_table[next_page].first_object_offset ==
2579 -(next_page - first_page)*PAGE_BYTES)) {
2580 /* It checks out OK, free the page. We don't need to both zeroing
2581 * pages as this should have been done before shrinking the
2582 * object. These pages shouldn't be write protected as they
2583 * should be zero filled. */
2584 gc_assert(page_table[next_page].write_protected == 0);
2586 old_bytes_used = page_table[next_page].bytes_used;
2587 page_table[next_page].allocated = FREE_PAGE_FLAG;
2588 page_table[next_page].bytes_used = 0;
2589 bytes_freed += old_bytes_used;
2593 if ((bytes_freed > 0) && gencgc_verbose) {
2595 "/maybe_adjust_large_object() freed %d\n",
2599 generations[from_space].bytes_allocated -= bytes_freed;
2600 bytes_allocated -= bytes_freed;
2605 /* Take a possible pointer to a Lisp object and mark its page in the
2606 * page_table so that it will not be relocated during a GC.
2608 * This involves locating the page it points to, then backing up to
2609 * the start of its region, then marking all pages dont_move from there
2610 * up to the first page that's not full or has a different generation
2612 * It is assumed that all the page static flags have been cleared at
2613 * the start of a GC.
2615 * It is also assumed that the current gc_alloc() region has been
2616 * flushed and the tables updated. */
2619 preserve_pointer(void *addr)
2621 page_index_t addr_page_index = find_page_index(addr);
2622 page_index_t first_page;
2624 unsigned int region_allocation;
2626 /* quick check 1: Address is quite likely to have been invalid. */
2627 if ((addr_page_index == -1)
2628 || (page_table[addr_page_index].allocated == FREE_PAGE_FLAG)
2629 || (page_table[addr_page_index].bytes_used == 0)
2630 || (page_table[addr_page_index].gen != from_space)
2631 /* Skip if already marked dont_move. */
2632 || (page_table[addr_page_index].dont_move != 0))
2634 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2635 /* (Now that we know that addr_page_index is in range, it's
2636 * safe to index into page_table[] with it.) */
2637 region_allocation = page_table[addr_page_index].allocated;
2639 /* quick check 2: Check the offset within the page.
2642 if (((unsigned long)addr & (PAGE_BYTES - 1)) > page_table[addr_page_index].bytes_used)
2645 /* Filter out anything which can't be a pointer to a Lisp object
2646 * (or, as a special case which also requires dont_move, a return
2647 * address referring to something in a CodeObject). This is
2648 * expensive but important, since it vastly reduces the
2649 * probability that random garbage will be bogusly interpreted as
2650 * a pointer which prevents a page from moving. */
2651 if (!(possibly_valid_dynamic_space_pointer(addr)))
2654 /* Find the beginning of the region. Note that there may be
2655 * objects in the region preceding the one that we were passed a
2656 * pointer to: if this is the case, we will write-protect all the
2657 * previous objects' pages too. */
2660 /* I think this'd work just as well, but without the assertions.
2661 * -dan 2004.01.01 */
2662 first_page = find_page_index(page_region_start(addr_page_index))
2664 first_page = addr_page_index;
2665 while (page_table[first_page].first_object_offset != 0) {
2667 /* Do some checks. */
2668 gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
2669 gc_assert(page_table[first_page].gen == from_space);
2670 gc_assert(page_table[first_page].allocated == region_allocation);
2674 /* Adjust any large objects before promotion as they won't be
2675 * copied after promotion. */
2676 if (page_table[first_page].large_object) {
2677 maybe_adjust_large_object(page_address(first_page));
2678 /* If a large object has shrunk then addr may now point to a
2679 * free area in which case it's ignored here. Note it gets
2680 * through the valid pointer test above because the tail looks
2682 if ((page_table[addr_page_index].allocated == FREE_PAGE_FLAG)
2683 || (page_table[addr_page_index].bytes_used == 0)
2684 /* Check the offset within the page. */
2685 || (((unsigned long)addr & (PAGE_BYTES - 1))
2686 > page_table[addr_page_index].bytes_used)) {
2688 "weird? ignore ptr 0x%x to freed area of large object\n",
2692 /* It may have moved to unboxed pages. */
2693 region_allocation = page_table[first_page].allocated;
2696 /* Now work forward until the end of this contiguous area is found,
2697 * marking all pages as dont_move. */
2698 for (i = first_page; ;i++) {
2699 gc_assert(page_table[i].allocated == region_allocation);
2701 /* Mark the page static. */
2702 page_table[i].dont_move = 1;
2704 /* Move the page to the new_space. XX I'd rather not do this
2705 * but the GC logic is not quite able to copy with the static
2706 * pages remaining in the from space. This also requires the
2707 * generation bytes_allocated counters be updated. */
2708 page_table[i].gen = new_space;
2709 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2710 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2712 /* It is essential that the pages are not write protected as
2713 * they may have pointers into the old-space which need
2714 * scavenging. They shouldn't be write protected at this
2716 gc_assert(!page_table[i].write_protected);
2718 /* Check whether this is the last page in this contiguous block.. */
2719 if ((page_table[i].bytes_used < PAGE_BYTES)
2720 /* ..or it is PAGE_BYTES and is the last in the block */
2721 || (page_table[i+1].allocated == FREE_PAGE_FLAG)
2722 || (page_table[i+1].bytes_used == 0) /* next page free */
2723 || (page_table[i+1].gen != from_space) /* diff. gen */
2724 || (page_table[i+1].first_object_offset == 0))
2728 /* Check that the page is now static. */
2729 gc_assert(page_table[addr_page_index].dont_move != 0);
2732 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2735 /* If the given page is not write-protected, then scan it for pointers
2736 * to younger generations or the top temp. generation, if no
2737 * suspicious pointers are found then the page is write-protected.
2739 * Care is taken to check for pointers to the current gc_alloc()
2740 * region if it is a younger generation or the temp. generation. This
2741 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2742 * the gc_alloc_generation does not need to be checked as this is only
2743 * called from scavenge_generation() when the gc_alloc generation is
2744 * younger, so it just checks if there is a pointer to the current
2747 * We return 1 if the page was write-protected, else 0. */
2749 update_page_write_prot(page_index_t page)
2751 generation_index_t gen = page_table[page].gen;
2754 void **page_addr = (void **)page_address(page);
2755 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2757 /* Shouldn't be a free page. */
2758 gc_assert(page_table[page].allocated != FREE_PAGE_FLAG);
2759 gc_assert(page_table[page].bytes_used != 0);
2761 /* Skip if it's already write-protected, pinned, or unboxed */
2762 if (page_table[page].write_protected
2763 /* FIXME: What's the reason for not write-protecting pinned pages? */
2764 || page_table[page].dont_move
2765 || (page_table[page].allocated & UNBOXED_PAGE_FLAG))
2768 /* Scan the page for pointers to younger generations or the
2769 * top temp. generation. */
2771 for (j = 0; j < num_words; j++) {
2772 void *ptr = *(page_addr+j);
2773 page_index_t index = find_page_index(ptr);
2775 /* Check that it's in the dynamic space */
2777 if (/* Does it point to a younger or the temp. generation? */
2778 ((page_table[index].allocated != FREE_PAGE_FLAG)
2779 && (page_table[index].bytes_used != 0)
2780 && ((page_table[index].gen < gen)
2781 || (page_table[index].gen == SCRATCH_GENERATION)))
2783 /* Or does it point within a current gc_alloc() region? */
2784 || ((boxed_region.start_addr <= ptr)
2785 && (ptr <= boxed_region.free_pointer))
2786 || ((unboxed_region.start_addr <= ptr)
2787 && (ptr <= unboxed_region.free_pointer))) {
2794 /* Write-protect the page. */
2795 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2797 os_protect((void *)page_addr,
2799 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2801 /* Note the page as protected in the page tables. */
2802 page_table[page].write_protected = 1;
2808 /* Scavenge all generations from FROM to TO, inclusive, except for
2809 * new_space which needs special handling, as new objects may be
2810 * added which are not checked here - use scavenge_newspace generation.
2812 * Write-protected pages should not have any pointers to the
2813 * from_space so do need scavenging; thus write-protected pages are
2814 * not always scavenged. There is some code to check that these pages
2815 * are not written; but to check fully the write-protected pages need
2816 * to be scavenged by disabling the code to skip them.
2818 * Under the current scheme when a generation is GCed the younger
2819 * generations will be empty. So, when a generation is being GCed it
2820 * is only necessary to scavenge the older generations for pointers
2821 * not the younger. So a page that does not have pointers to younger
2822 * generations does not need to be scavenged.
2824 * The write-protection can be used to note pages that don't have
2825 * pointers to younger pages. But pages can be written without having
2826 * pointers to younger generations. After the pages are scavenged here
2827 * they can be scanned for pointers to younger generations and if
2828 * there are none the page can be write-protected.
2830 * One complication is when the newspace is the top temp. generation.
2832 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2833 * that none were written, which they shouldn't be as they should have
2834 * no pointers to younger generations. This breaks down for weak
2835 * pointers as the objects contain a link to the next and are written
2836 * if a weak pointer is scavenged. Still it's a useful check. */
2838 scavenge_generations(generation_index_t from, generation_index_t to)
2845 /* Clear the write_protected_cleared flags on all pages. */
2846 for (i = 0; i < page_table_pages; i++)
2847 page_table[i].write_protected_cleared = 0;
2850 for (i = 0; i < last_free_page; i++) {
2851 generation_index_t generation = page_table[i].gen;
2852 if ((page_table[i].allocated & BOXED_PAGE_FLAG)
2853 && (page_table[i].bytes_used != 0)
2854 && (generation != new_space)
2855 && (generation >= from)
2856 && (generation <= to)) {
2857 page_index_t last_page,j;
2858 int write_protected=1;
2860 /* This should be the start of a region */
2861 gc_assert(page_table[i].first_object_offset == 0);
2863 /* Now work forward until the end of the region */
2864 for (last_page = i; ; last_page++) {
2866 write_protected && page_table[last_page].write_protected;
2867 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2868 /* Or it is PAGE_BYTES and is the last in the block */
2869 || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG))
2870 || (page_table[last_page+1].bytes_used == 0)
2871 || (page_table[last_page+1].gen != generation)
2872 || (page_table[last_page+1].first_object_offset == 0))
2875 if (!write_protected) {
2876 scavenge(page_address(i),
2877 (page_table[last_page].bytes_used +
2878 (last_page-i)*PAGE_BYTES)/N_WORD_BYTES);
2880 /* Now scan the pages and write protect those that
2881 * don't have pointers to younger generations. */
2882 if (enable_page_protection) {
2883 for (j = i; j <= last_page; j++) {
2884 num_wp += update_page_write_prot(j);
2887 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2889 "/write protected %d pages within generation %d\n",
2890 num_wp, generation));
2898 /* Check that none of the write_protected pages in this generation
2899 * have been written to. */
2900 for (i = 0; i < page_table_pages; i++) {
2901 if ((page_table[i].allocation != FREE_PAGE_FLAG)
2902 && (page_table[i].bytes_used != 0)
2903 && (page_table[i].gen == generation)
2904 && (page_table[i].write_protected_cleared != 0)) {
2905 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2907 "/page bytes_used=%d first_object_offset=%d dont_move=%d\n",
2908 page_table[i].bytes_used,
2909 page_table[i].first_object_offset,
2910 page_table[i].dont_move));
2911 lose("write to protected page %d in scavenge_generation()\n", i);
2918 /* Scavenge a newspace generation. As it is scavenged new objects may
2919 * be allocated to it; these will also need to be scavenged. This
2920 * repeats until there are no more objects unscavenged in the
2921 * newspace generation.
2923 * To help improve the efficiency, areas written are recorded by
2924 * gc_alloc() and only these scavenged. Sometimes a little more will be
2925 * scavenged, but this causes no harm. An easy check is done that the
2926 * scavenged bytes equals the number allocated in the previous
2929 * Write-protected pages are not scanned except if they are marked
2930 * dont_move in which case they may have been promoted and still have
2931 * pointers to the from space.
2933 * Write-protected pages could potentially be written by alloc however
2934 * to avoid having to handle re-scavenging of write-protected pages
2935 * gc_alloc() does not write to write-protected pages.
2937 * New areas of objects allocated are recorded alternatively in the two
2938 * new_areas arrays below. */
2939 static struct new_area new_areas_1[NUM_NEW_AREAS];
2940 static struct new_area new_areas_2[NUM_NEW_AREAS];
2942 /* Do one full scan of the new space generation. This is not enough to
2943 * complete the job as new objects may be added to the generation in
2944 * the process which are not scavenged. */
2946 scavenge_newspace_generation_one_scan(generation_index_t generation)
2951 "/starting one full scan of newspace generation %d\n",
2953 for (i = 0; i < last_free_page; i++) {
2954 /* Note that this skips over open regions when it encounters them. */
2955 if ((page_table[i].allocated & BOXED_PAGE_FLAG)
2956 && (page_table[i].bytes_used != 0)
2957 && (page_table[i].gen == generation)
2958 && ((page_table[i].write_protected == 0)
2959 /* (This may be redundant as write_protected is now
2960 * cleared before promotion.) */
2961 || (page_table[i].dont_move == 1))) {
2962 page_index_t last_page;
2965 /* The scavenge will start at the first_object_offset of page i.
2967 * We need to find the full extent of this contiguous
2968 * block in case objects span pages.
2970 * Now work forward until the end of this contiguous area
2971 * is found. A small area is preferred as there is a
2972 * better chance of its pages being write-protected. */
2973 for (last_page = i; ;last_page++) {
2974 /* If all pages are write-protected and movable,
2975 * then no need to scavenge */
2976 all_wp=all_wp && page_table[last_page].write_protected &&
2977 !page_table[last_page].dont_move;
2979 /* Check whether this is the last page in this
2980 * contiguous block */
2981 if ((page_table[last_page].bytes_used < PAGE_BYTES)
2982 /* Or it is PAGE_BYTES and is the last in the block */
2983 || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG))
2984 || (page_table[last_page+1].bytes_used == 0)
2985 || (page_table[last_page+1].gen != generation)
2986 || (page_table[last_page+1].first_object_offset == 0))
2990 /* Do a limited check for write-protected pages. */
2994 size = (page_table[last_page].bytes_used
2995 + (last_page-i)*PAGE_BYTES
2996 - page_table[i].first_object_offset)/N_WORD_BYTES;
2997 new_areas_ignore_page = last_page;
2999 scavenge(page_region_start(i), size);
3006 "/done with one full scan of newspace generation %d\n",
3010 /* Do a complete scavenge of the newspace generation. */
3012 scavenge_newspace_generation(generation_index_t generation)
3016 /* the new_areas array currently being written to by gc_alloc() */
3017 struct new_area (*current_new_areas)[] = &new_areas_1;
3018 long current_new_areas_index;
3020 /* the new_areas created by the previous scavenge cycle */
3021 struct new_area (*previous_new_areas)[] = NULL;
3022 long previous_new_areas_index;
3024 /* Flush the current regions updating the tables. */
3025 gc_alloc_update_all_page_tables();
3027 /* Turn on the recording of new areas by gc_alloc(). */
3028 new_areas = current_new_areas;
3029 new_areas_index = 0;
3031 /* Don't need to record new areas that get scavenged anyway during
3032 * scavenge_newspace_generation_one_scan. */
3033 record_new_objects = 1;
3035 /* Start with a full scavenge. */
3036 scavenge_newspace_generation_one_scan(generation);
3038 /* Record all new areas now. */
3039 record_new_objects = 2;
3041 /* Give a chance to weak hash tables to make other objects live.
3042 * FIXME: The algorithm implemented here for weak hash table gcing
3043 * is O(W^2+N) as Bruno Haible warns in
3044 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
3045 * see "Implementation 2". */
3046 scav_weak_hash_tables();
3048 /* Flush the current regions updating the tables. */
3049 gc_alloc_update_all_page_tables();
3051 /* Grab new_areas_index. */
3052 current_new_areas_index = new_areas_index;
3055 "The first scan is finished; current_new_areas_index=%d.\n",
3056 current_new_areas_index));*/
3058 while (current_new_areas_index > 0) {
3059 /* Move the current to the previous new areas */
3060 previous_new_areas = current_new_areas;
3061 previous_new_areas_index = current_new_areas_index;
3063 /* Scavenge all the areas in previous new areas. Any new areas
3064 * allocated are saved in current_new_areas. */
3066 /* Allocate an array for current_new_areas; alternating between
3067 * new_areas_1 and 2 */
3068 if (previous_new_areas == &new_areas_1)
3069 current_new_areas = &new_areas_2;
3071 current_new_areas = &new_areas_1;
3073 /* Set up for gc_alloc(). */
3074 new_areas = current_new_areas;
3075 new_areas_index = 0;
3077 /* Check whether previous_new_areas had overflowed. */
3078 if (previous_new_areas_index >= NUM_NEW_AREAS) {
3080 /* New areas of objects allocated have been lost so need to do a
3081 * full scan to be sure! If this becomes a problem try
3082 * increasing NUM_NEW_AREAS. */
3084 SHOW("new_areas overflow, doing full scavenge");
3086 /* Don't need to record new areas that get scavenged
3087 * anyway during scavenge_newspace_generation_one_scan. */
3088 record_new_objects = 1;
3090 scavenge_newspace_generation_one_scan(generation);
3092 /* Record all new areas now. */
3093 record_new_objects = 2;
3095 scav_weak_hash_tables();
3097 /* Flush the current regions updating the tables. */
3098 gc_alloc_update_all_page_tables();
3102 /* Work through previous_new_areas. */
3103 for (i = 0; i < previous_new_areas_index; i++) {
3104 long page = (*previous_new_areas)[i].page;
3105 long offset = (*previous_new_areas)[i].offset;
3106 long size = (*previous_new_areas)[i].size / N_WORD_BYTES;
3107 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
3108 scavenge(page_address(page)+offset, size);
3111 scav_weak_hash_tables();
3113 /* Flush the current regions updating the tables. */
3114 gc_alloc_update_all_page_tables();
3117 current_new_areas_index = new_areas_index;
3120 "The re-scan has finished; current_new_areas_index=%d.\n",
3121 current_new_areas_index));*/
3124 /* Turn off recording of areas allocated by gc_alloc(). */
3125 record_new_objects = 0;
3128 /* Check that none of the write_protected pages in this generation
3129 * have been written to. */
3130 for (i = 0; i < page_table_pages; i++) {
3131 if ((page_table[i].allocation != FREE_PAGE_FLAG)
3132 && (page_table[i].bytes_used != 0)
3133 && (page_table[i].gen == generation)
3134 && (page_table[i].write_protected_cleared != 0)
3135 && (page_table[i].dont_move == 0)) {
3136 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
3137 i, generation, page_table[i].dont_move);
3143 /* Un-write-protect all the pages in from_space. This is done at the
3144 * start of a GC else there may be many page faults while scavenging
3145 * the newspace (I've seen drive the system time to 99%). These pages
3146 * would need to be unprotected anyway before unmapping in
3147 * free_oldspace; not sure what effect this has on paging.. */
3149 unprotect_oldspace(void)
3153 for (i = 0; i < last_free_page; i++) {
3154 if ((page_table[i].allocated != FREE_PAGE_FLAG)
3155 && (page_table[i].bytes_used != 0)
3156 && (page_table[i].gen == from_space)) {
3159 page_start = (void *)page_address(i);
3161 /* Remove any write-protection. We should be able to rely
3162 * on the write-protect flag to avoid redundant calls. */
3163 if (page_table[i].write_protected) {
3164 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3165 page_table[i].write_protected = 0;
3171 /* Work through all the pages and free any in from_space. This
3172 * assumes that all objects have been copied or promoted to an older
3173 * generation. Bytes_allocated and the generation bytes_allocated
3174 * counter are updated. The number of bytes freed is returned. */
3178 long bytes_freed = 0;
3179 page_index_t first_page, last_page;
3184 /* Find a first page for the next region of pages. */
3185 while ((first_page < last_free_page)
3186 && ((page_table[first_page].allocated == FREE_PAGE_FLAG)
3187 || (page_table[first_page].bytes_used == 0)
3188 || (page_table[first_page].gen != from_space)))
3191 if (first_page >= last_free_page)
3194 /* Find the last page of this region. */
3195 last_page = first_page;
3198 /* Free the page. */
3199 bytes_freed += page_table[last_page].bytes_used;
3200 generations[page_table[last_page].gen].bytes_allocated -=
3201 page_table[last_page].bytes_used;
3202 page_table[last_page].allocated = FREE_PAGE_FLAG;
3203 page_table[last_page].bytes_used = 0;
3205 /* Remove any write-protection. We should be able to rely
3206 * on the write-protect flag to avoid redundant calls. */
3208 void *page_start = (void *)page_address(last_page);
3210 if (page_table[last_page].write_protected) {
3211 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
3212 page_table[last_page].write_protected = 0;
3217 while ((last_page < last_free_page)
3218 && (page_table[last_page].allocated != FREE_PAGE_FLAG)
3219 && (page_table[last_page].bytes_used != 0)
3220 && (page_table[last_page].gen == from_space));
3222 #ifdef READ_PROTECT_FREE_PAGES
3223 os_protect(page_address(first_page),
3224 PAGE_BYTES*(last_page-first_page),
3227 first_page = last_page;
3228 } while (first_page < last_free_page);
3230 bytes_allocated -= bytes_freed;
3235 /* Print some information about a pointer at the given address. */
3237 print_ptr(lispobj *addr)
3239 /* If addr is in the dynamic space then out the page information. */
3240 page_index_t pi1 = find_page_index((void*)addr);
3243 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %d dont_move %d\n",
3244 (unsigned long) addr,
3246 page_table[pi1].allocated,
3247 page_table[pi1].gen,
3248 page_table[pi1].bytes_used,
3249 page_table[pi1].first_object_offset,
3250 page_table[pi1].dont_move);
3251 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
3265 verify_space(lispobj *start, size_t words)
3267 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3268 int is_in_readonly_space =
3269 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3270 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3274 lispobj thing = *(lispobj*)start;
3276 if (is_lisp_pointer(thing)) {
3277 page_index_t page_index = find_page_index((void*)thing);
3278 long to_readonly_space =
3279 (READ_ONLY_SPACE_START <= thing &&
3280 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3281 long to_static_space =
3282 (STATIC_SPACE_START <= thing &&
3283 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3285 /* Does it point to the dynamic space? */
3286 if (page_index != -1) {
3287 /* If it's within the dynamic space it should point to a used
3288 * page. XX Could check the offset too. */
3289 if ((page_table[page_index].allocated != FREE_PAGE_FLAG)
3290 && (page_table[page_index].bytes_used == 0))
3291 lose ("Ptr %x @ %x sees free page.\n", thing, start);
3292 /* Check that it doesn't point to a forwarding pointer! */
3293 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3294 lose("Ptr %x @ %x sees forwarding ptr.\n", thing, start);
3296 /* Check that its not in the RO space as it would then be a
3297 * pointer from the RO to the dynamic space. */
3298 if (is_in_readonly_space) {
3299 lose("ptr to dynamic space %x from RO space %x\n",
3302 /* Does it point to a plausible object? This check slows
3303 * it down a lot (so it's commented out).
3305 * "a lot" is serious: it ate 50 minutes cpu time on
3306 * my duron 950 before I came back from lunch and
3309 * FIXME: Add a variable to enable this
3312 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3313 lose("ptr %x to invalid object %x\n", thing, start);
3317 /* Verify that it points to another valid space. */
3318 if (!to_readonly_space && !to_static_space) {
3319 lose("Ptr %x @ %x sees junk.\n", thing, start);
3323 if (!(fixnump(thing))) {
3325 switch(widetag_of(*start)) {
3328 case SIMPLE_VECTOR_WIDETAG:
3330 case COMPLEX_WIDETAG:
3331 case SIMPLE_ARRAY_WIDETAG:
3332 case COMPLEX_BASE_STRING_WIDETAG:
3333 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3334 case COMPLEX_CHARACTER_STRING_WIDETAG:
3336 case COMPLEX_VECTOR_NIL_WIDETAG:
3337 case COMPLEX_BIT_VECTOR_WIDETAG:
3338 case COMPLEX_VECTOR_WIDETAG:
3339 case COMPLEX_ARRAY_WIDETAG:
3340 case CLOSURE_HEADER_WIDETAG:
3341 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3342 case VALUE_CELL_HEADER_WIDETAG:
3343 case SYMBOL_HEADER_WIDETAG:
3344 case CHARACTER_WIDETAG:
3345 #if N_WORD_BITS == 64
3346 case SINGLE_FLOAT_WIDETAG:
3348 case UNBOUND_MARKER_WIDETAG:
3353 case INSTANCE_HEADER_WIDETAG:
3356 long ntotal = HeaderValue(thing);
3357 lispobj layout = ((struct instance *)start)->slots[0];
3362 nuntagged = ((struct layout *)native_pointer(layout))->n_untagged_slots;
3363 verify_space(start + 1, ntotal - fixnum_value(nuntagged));
3367 case CODE_HEADER_WIDETAG:
3369 lispobj object = *start;
3371 long nheader_words, ncode_words, nwords;
3373 struct simple_fun *fheaderp;
3375 code = (struct code *) start;
3377 /* Check that it's not in the dynamic space.
3378 * FIXME: Isn't is supposed to be OK for code
3379 * objects to be in the dynamic space these days? */
3380 if (is_in_dynamic_space
3381 /* It's ok if it's byte compiled code. The trace
3382 * table offset will be a fixnum if it's x86
3383 * compiled code - check.
3385 * FIXME: #^#@@! lack of abstraction here..
3386 * This line can probably go away now that
3387 * there's no byte compiler, but I've got
3388 * too much to worry about right now to try
3389 * to make sure. -- WHN 2001-10-06 */
3390 && fixnump(code->trace_table_offset)
3391 /* Only when enabled */
3392 && verify_dynamic_code_check) {
3394 "/code object at %x in the dynamic space\n",
3398 ncode_words = fixnum_value(code->code_size);
3399 nheader_words = HeaderValue(object);
3400 nwords = ncode_words + nheader_words;
3401 nwords = CEILING(nwords, 2);
3402 /* Scavenge the boxed section of the code data block */
3403 verify_space(start + 1, nheader_words - 1);
3405 /* Scavenge the boxed section of each function
3406 * object in the code data block. */
3407 fheaderl = code->entry_points;
3408 while (fheaderl != NIL) {
3410 (struct simple_fun *) native_pointer(fheaderl);
3411 gc_assert(widetag_of(fheaderp->header) == SIMPLE_FUN_HEADER_WIDETAG);
3412 verify_space(&fheaderp->name, 1);
3413 verify_space(&fheaderp->arglist, 1);
3414 verify_space(&fheaderp->type, 1);
3415 fheaderl = fheaderp->next;
3421 /* unboxed objects */
3422 case BIGNUM_WIDETAG:
3423 #if N_WORD_BITS != 64
3424 case SINGLE_FLOAT_WIDETAG:
3426 case DOUBLE_FLOAT_WIDETAG:
3427 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3428 case LONG_FLOAT_WIDETAG:
3430 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3431 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3433 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3434 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3436 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3437 case COMPLEX_LONG_FLOAT_WIDETAG:
3439 case SIMPLE_BASE_STRING_WIDETAG:
3440 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3441 case SIMPLE_CHARACTER_STRING_WIDETAG:
3443 case SIMPLE_BIT_VECTOR_WIDETAG:
3444 case SIMPLE_ARRAY_NIL_WIDETAG:
3445 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3446 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3447 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3448 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3449 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3450 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3451 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG
3452 case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
3454 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3455 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3456 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG
3457 case SIMPLE_ARRAY_UNSIGNED_BYTE_60_WIDETAG:
3459 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3460 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3462 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3463 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3465 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3466 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3468 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3469 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3471 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG
3472 case SIMPLE_ARRAY_SIGNED_BYTE_30_WIDETAG:
3474 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3475 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3477 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG
3478 case SIMPLE_ARRAY_SIGNED_BYTE_61_WIDETAG:
3480 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3481 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3483 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3484 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3485 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3486 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3488 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3489 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3491 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3492 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3494 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3495 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3498 case WEAK_POINTER_WIDETAG:
3499 #ifdef LUTEX_WIDETAG
3502 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3503 case NO_TLS_VALUE_MARKER_WIDETAG:
3505 count = (sizetab[widetag_of(*start)])(start);
3509 lose("Unhandled widetag 0x%x at 0x%x\n", widetag_of(*start), start);
3521 /* FIXME: It would be nice to make names consistent so that
3522 * foo_size meant size *in* *bytes* instead of size in some
3523 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3524 * Some counts of lispobjs are called foo_count; it might be good
3525 * to grep for all foo_size and rename the appropriate ones to
3527 long read_only_space_size =
3528 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3529 - (lispobj*)READ_ONLY_SPACE_START;
3530 long static_space_size =
3531 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3532 - (lispobj*)STATIC_SPACE_START;
3534 for_each_thread(th) {
3535 long binding_stack_size =
3536 (lispobj*)get_binding_stack_pointer(th)
3537 - (lispobj*)th->binding_stack_start;
3538 verify_space(th->binding_stack_start, binding_stack_size);
3540 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3541 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3545 verify_generation(generation_index_t generation)
3549 for (i = 0; i < last_free_page; i++) {
3550 if ((page_table[i].allocated != FREE_PAGE_FLAG)
3551 && (page_table[i].bytes_used != 0)
3552 && (page_table[i].gen == generation)) {
3553 page_index_t last_page;
3554 int region_allocation = page_table[i].allocated;
3556 /* This should be the start of a contiguous block */
3557 gc_assert(page_table[i].first_object_offset == 0);
3559 /* Need to find the full extent of this contiguous block in case
3560 objects span pages. */
3562 /* Now work forward until the end of this contiguous area is
3564 for (last_page = i; ;last_page++)
3565 /* Check whether this is the last page in this contiguous
3567 if ((page_table[last_page].bytes_used < PAGE_BYTES)
3568 /* Or it is PAGE_BYTES and is the last in the block */
3569 || (page_table[last_page+1].allocated != region_allocation)
3570 || (page_table[last_page+1].bytes_used == 0)
3571 || (page_table[last_page+1].gen != generation)
3572 || (page_table[last_page+1].first_object_offset == 0))
3575 verify_space(page_address(i), (page_table[last_page].bytes_used
3576 + (last_page-i)*PAGE_BYTES)/N_WORD_BYTES);
3582 /* Check that all the free space is zero filled. */
3584 verify_zero_fill(void)
3588 for (page = 0; page < last_free_page; page++) {
3589 if (page_table[page].allocated == FREE_PAGE_FLAG) {
3590 /* The whole page should be zero filled. */
3591 long *start_addr = (long *)page_address(page);
3594 for (i = 0; i < size; i++) {
3595 if (start_addr[i] != 0) {
3596 lose("free page not zero at %x\n", start_addr + i);
3600 long free_bytes = PAGE_BYTES - page_table[page].bytes_used;
3601 if (free_bytes > 0) {
3602 long *start_addr = (long *)((unsigned long)page_address(page)
3603 + page_table[page].bytes_used);
3604 long size = free_bytes / N_WORD_BYTES;
3606 for (i = 0; i < size; i++) {
3607 if (start_addr[i] != 0) {
3608 lose("free region not zero at %x\n", start_addr + i);
3616 /* External entry point for verify_zero_fill */
3618 gencgc_verify_zero_fill(void)
3620 /* Flush the alloc regions updating the tables. */
3621 gc_alloc_update_all_page_tables();
3622 SHOW("verifying zero fill");
3627 verify_dynamic_space(void)
3629 generation_index_t i;
3631 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3632 verify_generation(i);
3634 if (gencgc_enable_verify_zero_fill)
3638 /* Write-protect all the dynamic boxed pages in the given generation. */
3640 write_protect_generation_pages(generation_index_t generation)
3644 gc_assert(generation < SCRATCH_GENERATION);
3646 for (start = 0; start < last_free_page; start++) {
3647 if ((page_table[start].allocated == BOXED_PAGE_FLAG)
3648 && (page_table[start].bytes_used != 0)
3649 && !page_table[start].dont_move
3650 && (page_table[start].gen == generation)) {
3654 /* Note the page as protected in the page tables. */
3655 page_table[start].write_protected = 1;
3657 for (last = start + 1; last < last_free_page; last++) {
3658 if ((page_table[last].allocated != BOXED_PAGE_FLAG)
3659 || (page_table[last].bytes_used == 0)
3660 || page_table[last].dont_move
3661 || (page_table[last].gen != generation))
3663 page_table[last].write_protected = 1;
3666 page_start = (void *)page_address(start);
3668 os_protect(page_start,
3669 PAGE_BYTES * (last - start),
3670 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3676 if (gencgc_verbose > 1) {
3678 "/write protected %d of %d pages in generation %d\n",
3679 count_write_protect_generation_pages(generation),
3680 count_generation_pages(generation),
3685 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3688 scavenge_control_stack()
3690 unsigned long control_stack_size;
3692 /* This is going to be a big problem when we try to port threads
3694 struct thread *th = arch_os_get_current_thread();
3695 lispobj *control_stack =
3696 (lispobj *)(th->control_stack_start);
3698 control_stack_size = current_control_stack_pointer - control_stack;
3699 scavenge(control_stack, control_stack_size);
3702 /* Scavenging Interrupt Contexts */
3704 static int boxed_registers[] = BOXED_REGISTERS;
3707 scavenge_interrupt_context(os_context_t * context)
3713 unsigned long lip_offset;
3714 int lip_register_pair;
3716 unsigned long pc_code_offset;
3718 #ifdef ARCH_HAS_LINK_REGISTER
3719 unsigned long lr_code_offset;
3721 #ifdef ARCH_HAS_NPC_REGISTER
3722 unsigned long npc_code_offset;
3726 /* Find the LIP's register pair and calculate it's offset */
3727 /* before we scavenge the context. */
3730 * I (RLT) think this is trying to find the boxed register that is
3731 * closest to the LIP address, without going past it. Usually, it's
3732 * reg_CODE or reg_LRA. But sometimes, nothing can be found.
3734 lip = *os_context_register_addr(context, reg_LIP);
3735 lip_offset = 0x7FFFFFFF;
3736 lip_register_pair = -1;
3737 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3742 index = boxed_registers[i];
3743 reg = *os_context_register_addr(context, index);
3744 if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
3746 if (offset < lip_offset) {
3747 lip_offset = offset;
3748 lip_register_pair = index;
3752 #endif /* reg_LIP */
3754 /* Compute the PC's offset from the start of the CODE */
3756 pc_code_offset = *os_context_pc_addr(context) - *os_context_register_addr(context, reg_CODE);
3757 #ifdef ARCH_HAS_NPC_REGISTER
3758 npc_code_offset = *os_context_npc_addr(context) - *os_context_register_addr(context, reg_CODE);
3759 #endif /* ARCH_HAS_NPC_REGISTER */
3761 #ifdef ARCH_HAS_LINK_REGISTER
3763 *os_context_lr_addr(context) -
3764 *os_context_register_addr(context, reg_CODE);
3767 /* Scanvenge all boxed registers in the context. */
3768 for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
3772 index = boxed_registers[i];
3773 foo = *os_context_register_addr(context, index);
3775 *os_context_register_addr(context, index) = foo;
3777 scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
3784 * But what happens if lip_register_pair is -1? *os_context_register_addr on Solaris
3785 * (see solaris_register_address in solaris-os.c) will return
3786 * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
3787 * that what we really want? My guess is that that is not what we
3788 * want, so if lip_register_pair is -1, we don't touch reg_LIP at
3789 * all. But maybe it doesn't really matter if LIP is trashed?
3791 if (lip_register_pair >= 0) {
3792 *os_context_register_addr(context, reg_LIP) =
3793 *os_context_register_addr(context, lip_register_pair) + lip_offset;
3795 #endif /* reg_LIP */
3797 /* Fix the PC if it was in from space */
3798 if (from_space_p(*os_context_pc_addr(context)))
3799 *os_context_pc_addr(context) = *os_context_register_addr(context, reg_CODE) + pc_code_offset;
3801 #ifdef ARCH_HAS_LINK_REGISTER
3802 /* Fix the LR ditto; important if we're being called from
3803 * an assembly routine that expects to return using blr, otherwise
3805 if (from_space_p(*os_context_lr_addr(context)))
3806 *os_context_lr_addr(context) =
3807 *os_context_register_addr(context, reg_CODE) + lr_code_offset;
3810 #ifdef ARCH_HAS_NPC_REGISTER
3811 if (from_space_p(*os_context_npc_addr(context)))
3812 *os_context_npc_addr(context) = *os_context_register_addr(context, reg_CODE) + npc_code_offset;
3813 #endif /* ARCH_HAS_NPC_REGISTER */
3817 scavenge_interrupt_contexts(void)
3820 os_context_t *context;
3822 struct thread *th=arch_os_get_current_thread();
3824 index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
3826 #if defined(DEBUG_PRINT_CONTEXT_INDEX)
3827 printf("Number of active contexts: %d\n", index);
3830 for (i = 0; i < index; i++) {
3831 context = th->interrupt_contexts[i];
3832 scavenge_interrupt_context(context);
3838 #if defined(LISP_FEATURE_SB_THREAD)
3840 preserve_context_registers (os_context_t *c)
3843 /* On Darwin the signal context isn't a contiguous block of memory,
3844 * so just preserve_pointering its contents won't be sufficient.
3846 #if defined(LISP_FEATURE_DARWIN)
3847 #if defined LISP_FEATURE_X86
3848 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3849 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3850 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3851 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3852 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3853 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3854 preserve_pointer((void*)*os_context_pc_addr(c));
3855 #elif defined LISP_FEATURE_X86_64
3856 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3857 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3858 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3859 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3860 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3861 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3862 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3863 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3864 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3865 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3866 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3867 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3868 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3869 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3870 preserve_pointer((void*)*os_context_pc_addr(c));
3872 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3875 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3876 preserve_pointer(*ptr);
3881 /* Garbage collect a generation. If raise is 0 then the remains of the
3882 * generation are not raised to the next generation. */
3884 garbage_collect_generation(generation_index_t generation, int raise)
3886 unsigned long bytes_freed;
3888 unsigned long static_space_size;
3889 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3892 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3894 /* The oldest generation can't be raised. */
3895 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3897 /* Check if weak hash tables were processed in the previous GC. */
3898 gc_assert(weak_hash_tables == NULL);
3900 /* Initialize the weak pointer list. */
3901 weak_pointers = NULL;
3903 #ifdef LUTEX_WIDETAG
3904 unmark_lutexes(generation);
3907 /* When a generation is not being raised it is transported to a
3908 * temporary generation (NUM_GENERATIONS), and lowered when
3909 * done. Set up this new generation. There should be no pages
3910 * allocated to it yet. */
3912 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3915 /* Set the global src and dest. generations */
3916 from_space = generation;
3918 new_space = generation+1;
3920 new_space = SCRATCH_GENERATION;
3922 /* Change to a new space for allocation, resetting the alloc_start_page */
3923 gc_alloc_generation = new_space;
3924 generations[new_space].alloc_start_page = 0;
3925 generations[new_space].alloc_unboxed_start_page = 0;
3926 generations[new_space].alloc_large_start_page = 0;
3927 generations[new_space].alloc_large_unboxed_start_page = 0;
3929 /* Before any pointers are preserved, the dont_move flags on the
3930 * pages need to be cleared. */
3931 for (i = 0; i < last_free_page; i++)
3932 if(page_table[i].gen==from_space)
3933 page_table[i].dont_move = 0;
3935 /* Un-write-protect the old-space pages. This is essential for the
3936 * promoted pages as they may contain pointers into the old-space
3937 * which need to be scavenged. It also helps avoid unnecessary page
3938 * faults as forwarding pointers are written into them. They need to
3939 * be un-protected anyway before unmapping later. */
3940 unprotect_oldspace();
3942 /* Scavenge the stacks' conservative roots. */
3944 /* there are potentially two stacks for each thread: the main
3945 * stack, which may contain Lisp pointers, and the alternate stack.
3946 * We don't ever run Lisp code on the altstack, but it may
3947 * host a sigcontext with lisp objects in it */
3949 /* what we need to do: (1) find the stack pointer for the main
3950 * stack; scavenge it (2) find the interrupt context on the
3951 * alternate stack that might contain lisp values, and scavenge
3954 /* we assume that none of the preceding applies to the thread that
3955 * initiates GC. If you ever call GC from inside an altstack
3956 * handler, you will lose. */
3958 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3959 /* And if we're saving a core, there's no point in being conservative. */
3960 if (conservative_stack) {
3961 for_each_thread(th) {
3963 void **esp=(void **)-1;
3964 #ifdef LISP_FEATURE_SB_THREAD
3966 if(th==arch_os_get_current_thread()) {
3967 /* Somebody is going to burn in hell for this, but casting
3968 * it in two steps shuts gcc up about strict aliasing. */
3969 esp = (void **)((void *)&raise);
3972 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3973 for(i=free-1;i>=0;i--) {
3974 os_context_t *c=th->interrupt_contexts[i];
3975 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3976 if (esp1>=(void **)th->control_stack_start &&
3977 esp1<(void **)th->control_stack_end) {
3978 if(esp1<esp) esp=esp1;
3979 preserve_context_registers(c);
3984 esp = (void **)((void *)&raise);
3986 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3987 preserve_pointer(*ptr);
3994 if (gencgc_verbose > 1) {
3995 long num_dont_move_pages = count_dont_move_pages();
3997 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3998 num_dont_move_pages,
3999 num_dont_move_pages * PAGE_BYTES);
4003 /* Scavenge all the rest of the roots. */
4005 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
4007 * If not x86, we need to scavenge the interrupt context(s) and the
4010 scavenge_interrupt_contexts();
4011 scavenge_control_stack();
4014 /* Scavenge the Lisp functions of the interrupt handlers, taking
4015 * care to avoid SIG_DFL and SIG_IGN. */
4016 for (i = 0; i < NSIG; i++) {
4017 union interrupt_handler handler = interrupt_handlers[i];
4018 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
4019 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
4020 scavenge((lispobj *)(interrupt_handlers + i), 1);
4023 /* Scavenge the binding stacks. */
4026 for_each_thread(th) {
4027 long len= (lispobj *)get_binding_stack_pointer(th) -
4028 th->binding_stack_start;
4029 scavenge((lispobj *) th->binding_stack_start,len);
4030 #ifdef LISP_FEATURE_SB_THREAD
4031 /* do the tls as well */
4032 len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
4033 (sizeof (struct thread))/(sizeof (lispobj));
4034 scavenge((lispobj *) (th+1),len);
4039 /* The original CMU CL code had scavenge-read-only-space code
4040 * controlled by the Lisp-level variable
4041 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
4042 * wasn't documented under what circumstances it was useful or
4043 * safe to turn it on, so it's been turned off in SBCL. If you
4044 * want/need this functionality, and can test and document it,
4045 * please submit a patch. */
4047 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
4048 unsigned long read_only_space_size =
4049 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
4050 (lispobj*)READ_ONLY_SPACE_START;
4052 "/scavenge read only space: %d bytes\n",
4053 read_only_space_size * sizeof(lispobj)));
4054 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
4058 /* Scavenge static space. */
4060 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
4061 (lispobj *)STATIC_SPACE_START;
4062 if (gencgc_verbose > 1) {
4064 "/scavenge static space: %d bytes\n",
4065 static_space_size * sizeof(lispobj)));
4067 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
4069 /* All generations but the generation being GCed need to be
4070 * scavenged. The new_space generation needs special handling as
4071 * objects may be moved in - it is handled separately below. */
4072 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
4074 /* Finally scavenge the new_space generation. Keep going until no
4075 * more objects are moved into the new generation */
4076 scavenge_newspace_generation(new_space);
4078 /* FIXME: I tried reenabling this check when debugging unrelated
4079 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
4080 * Since the current GC code seems to work well, I'm guessing that
4081 * this debugging code is just stale, but I haven't tried to
4082 * figure it out. It should be figured out and then either made to
4083 * work or just deleted. */
4084 #define RESCAN_CHECK 0
4086 /* As a check re-scavenge the newspace once; no new objects should
4089 long old_bytes_allocated = bytes_allocated;
4090 long bytes_allocated;
4092 /* Start with a full scavenge. */
4093 scavenge_newspace_generation_one_scan(new_space);
4095 /* Flush the current regions, updating the tables. */
4096 gc_alloc_update_all_page_tables();
4098 bytes_allocated = bytes_allocated - old_bytes_allocated;
4100 if (bytes_allocated != 0) {
4101 lose("Rescan of new_space allocated %d more bytes.\n",
4107 scan_weak_hash_tables();
4108 scan_weak_pointers();
4110 /* Flush the current regions, updating the tables. */
4111 gc_alloc_update_all_page_tables();
4113 /* Free the pages in oldspace, but not those marked dont_move. */
4114 bytes_freed = free_oldspace();
4116 /* If the GC is not raising the age then lower the generation back
4117 * to its normal generation number */
4119 for (i = 0; i < last_free_page; i++)
4120 if ((page_table[i].bytes_used != 0)
4121 && (page_table[i].gen == SCRATCH_GENERATION))
4122 page_table[i].gen = generation;
4123 gc_assert(generations[generation].bytes_allocated == 0);
4124 generations[generation].bytes_allocated =
4125 generations[SCRATCH_GENERATION].bytes_allocated;
4126 generations[SCRATCH_GENERATION].bytes_allocated = 0;
4129 /* Reset the alloc_start_page for generation. */
4130 generations[generation].alloc_start_page = 0;
4131 generations[generation].alloc_unboxed_start_page = 0;
4132 generations[generation].alloc_large_start_page = 0;
4133 generations[generation].alloc_large_unboxed_start_page = 0;
4135 if (generation >= verify_gens) {
4139 verify_dynamic_space();
4142 /* Set the new gc trigger for the GCed generation. */
4143 generations[generation].gc_trigger =
4144 generations[generation].bytes_allocated
4145 + generations[generation].bytes_consed_between_gc;
4148 generations[generation].num_gc = 0;
4150 ++generations[generation].num_gc;
4152 #ifdef LUTEX_WIDETAG
4153 reap_lutexes(generation);
4155 move_lutexes(generation, generation+1);
4159 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
4161 update_dynamic_space_free_pointer(void)
4163 page_index_t last_page = -1, i;
4165 for (i = 0; i < last_free_page; i++)
4166 if ((page_table[i].allocated != FREE_PAGE_FLAG)
4167 && (page_table[i].bytes_used != 0))
4170 last_free_page = last_page+1;
4172 set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
4173 return 0; /* dummy value: return something ... */
4177 remap_free_pages (page_index_t from, page_index_t to)
4179 page_index_t first_page, last_page;
4181 for (first_page = from; first_page <= to; first_page++) {
4182 if (page_table[first_page].allocated != FREE_PAGE_FLAG ||
4183 page_table[first_page].need_to_zero == 0) {
4187 last_page = first_page + 1;
4188 while (page_table[last_page].allocated == FREE_PAGE_FLAG &&
4190 page_table[last_page].need_to_zero == 1) {
4194 /* There's a mysterious Solaris/x86 problem with using mmap
4195 * tricks for memory zeroing. See sbcl-devel thread
4196 * "Re: patch: standalone executable redux".
4198 #if defined(LISP_FEATURE_SUNOS)
4199 zero_pages(first_page, last_page-1);
4201 zero_pages_with_mmap(first_page, last_page-1);
4204 first_page = last_page;
4208 generation_index_t small_generation_limit = 1;
4210 /* GC all generations newer than last_gen, raising the objects in each
4211 * to the next older generation - we finish when all generations below
4212 * last_gen are empty. Then if last_gen is due for a GC, or if
4213 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
4214 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
4216 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
4217 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
4219 collect_garbage(generation_index_t last_gen)
4221 generation_index_t gen = 0, i;
4224 /* The largest value of last_free_page seen since the time
4225 * remap_free_pages was called. */
4226 static page_index_t high_water_mark = 0;
4228 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
4232 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
4234 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
4239 /* Flush the alloc regions updating the tables. */
4240 gc_alloc_update_all_page_tables();
4242 /* Verify the new objects created by Lisp code. */
4243 if (pre_verify_gen_0) {
4244 FSHOW((stderr, "pre-checking generation 0\n"));
4245 verify_generation(0);
4248 if (gencgc_verbose > 1)
4249 print_generation_stats(0);
4252 /* Collect the generation. */
4254 if (gen >= gencgc_oldest_gen_to_gc) {
4255 /* Never raise the oldest generation. */
4260 || (generations[gen].num_gc >= generations[gen].trigger_age);
4263 if (gencgc_verbose > 1) {
4265 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
4268 generations[gen].bytes_allocated,
4269 generations[gen].gc_trigger,
4270 generations[gen].num_gc));
4273 /* If an older generation is being filled, then update its
4276 generations[gen+1].cum_sum_bytes_allocated +=
4277 generations[gen+1].bytes_allocated;
4280 garbage_collect_generation(gen, raise);
4282 /* Reset the memory age cum_sum. */
4283 generations[gen].cum_sum_bytes_allocated = 0;
4285 if (gencgc_verbose > 1) {
4286 FSHOW((stderr, "GC of generation %d finished:\n", gen));
4287 print_generation_stats(0);
4291 } while ((gen <= gencgc_oldest_gen_to_gc)
4292 && ((gen < last_gen)
4293 || ((gen <= gencgc_oldest_gen_to_gc)
4295 && (generations[gen].bytes_allocated
4296 > generations[gen].gc_trigger)
4297 && (gen_av_mem_age(gen)
4298 > generations[gen].min_av_mem_age))));
4300 /* Now if gen-1 was raised all generations before gen are empty.
4301 * If it wasn't raised then all generations before gen-1 are empty.
4303 * Now objects within this gen's pages cannot point to younger
4304 * generations unless they are written to. This can be exploited
4305 * by write-protecting the pages of gen; then when younger
4306 * generations are GCed only the pages which have been written
4311 gen_to_wp = gen - 1;
4313 /* There's not much point in WPing pages in generation 0 as it is
4314 * never scavenged (except promoted pages). */
4315 if ((gen_to_wp > 0) && enable_page_protection) {
4316 /* Check that they are all empty. */
4317 for (i = 0; i < gen_to_wp; i++) {
4318 if (generations[i].bytes_allocated)
4319 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
4322 write_protect_generation_pages(gen_to_wp);
4325 /* Set gc_alloc() back to generation 0. The current regions should
4326 * be flushed after the above GCs. */
4327 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
4328 gc_alloc_generation = 0;
4330 /* Save the high-water mark before updating last_free_page */
4331 if (last_free_page > high_water_mark)
4332 high_water_mark = last_free_page;
4334 update_dynamic_space_free_pointer();
4336 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
4338 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
4341 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
4344 if (gen > small_generation_limit) {
4345 if (last_free_page > high_water_mark)
4346 high_water_mark = last_free_page;
4347 remap_free_pages(0, high_water_mark);
4348 high_water_mark = 0;
4353 SHOW("returning from collect_garbage");
4356 /* This is called by Lisp PURIFY when it is finished. All live objects
4357 * will have been moved to the RO and Static heaps. The dynamic space
4358 * will need a full re-initialization. We don't bother having Lisp
4359 * PURIFY flush the current gc_alloc() region, as the page_tables are
4360 * re-initialized, and every page is zeroed to be sure. */
4366 if (gencgc_verbose > 1)
4367 SHOW("entering gc_free_heap");
4369 for (page = 0; page < page_table_pages; page++) {
4370 /* Skip free pages which should already be zero filled. */
4371 if (page_table[page].allocated != FREE_PAGE_FLAG) {
4372 void *page_start, *addr;
4374 /* Mark the page free. The other slots are assumed invalid
4375 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4376 * should not be write-protected -- except that the
4377 * generation is used for the current region but it sets
4379 page_table[page].allocated = FREE_PAGE_FLAG;
4380 page_table[page].bytes_used = 0;
4382 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure about this change. */
4383 /* Zero the page. */
4384 page_start = (void *)page_address(page);
4386 /* First, remove any write-protection. */
4387 os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
4388 page_table[page].write_protected = 0;
4390 os_invalidate(page_start,PAGE_BYTES);
4391 addr = os_validate(page_start,PAGE_BYTES);
4392 if (addr == NULL || addr != page_start) {
4393 lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n",
4398 page_table[page].write_protected = 0;
4400 } else if (gencgc_zero_check_during_free_heap) {
4401 /* Double-check that the page is zero filled. */
4404 gc_assert(page_table[page].allocated == FREE_PAGE_FLAG);
4405 gc_assert(page_table[page].bytes_used == 0);
4406 page_start = (long *)page_address(page);
4407 for (i=0; i<1024; i++) {
4408 if (page_start[i] != 0) {
4409 lose("free region not zero at %x\n", page_start + i);
4415 bytes_allocated = 0;
4417 /* Initialize the generations. */
4418 for (page = 0; page < NUM_GENERATIONS; page++) {
4419 generations[page].alloc_start_page = 0;
4420 generations[page].alloc_unboxed_start_page = 0;
4421 generations[page].alloc_large_start_page = 0;
4422 generations[page].alloc_large_unboxed_start_page = 0;
4423 generations[page].bytes_allocated = 0;
4424 generations[page].gc_trigger = 2000000;
4425 generations[page].num_gc = 0;
4426 generations[page].cum_sum_bytes_allocated = 0;
4427 generations[page].lutexes = NULL;
4430 if (gencgc_verbose > 1)
4431 print_generation_stats(0);
4433 /* Initialize gc_alloc(). */
4434 gc_alloc_generation = 0;
4436 gc_set_region_empty(&boxed_region);
4437 gc_set_region_empty(&unboxed_region);
4440 set_alloc_pointer((lispobj)((char *)heap_base));
4442 if (verify_after_free_heap) {
4443 /* Check whether purify has left any bad pointers. */
4444 FSHOW((stderr, "checking after free_heap\n"));
4454 /* Compute the number of pages needed for the dynamic space.
4455 * Dynamic space size should be aligned on page size. */
4456 page_table_pages = dynamic_space_size/PAGE_BYTES;
4457 gc_assert(dynamic_space_size == (size_t) page_table_pages*PAGE_BYTES);
4459 page_table = calloc(page_table_pages, sizeof(struct page));
4460 gc_assert(page_table);
4463 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4464 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4466 #ifdef LUTEX_WIDETAG
4467 scavtab[LUTEX_WIDETAG] = scav_lutex;
4468 transother[LUTEX_WIDETAG] = trans_lutex;
4469 sizetab[LUTEX_WIDETAG] = size_lutex;
4472 heap_base = (void*)DYNAMIC_SPACE_START;
4474 /* Initialize each page structure. */
4475 for (i = 0; i < page_table_pages; i++) {
4476 /* Initialize all pages as free. */
4477 page_table[i].allocated = FREE_PAGE_FLAG;
4478 page_table[i].bytes_used = 0;
4480 /* Pages are not write-protected at startup. */
4481 page_table[i].write_protected = 0;
4484 bytes_allocated = 0;
4486 /* Initialize the generations.
4488 * FIXME: very similar to code in gc_free_heap(), should be shared */
4489 for (i = 0; i < NUM_GENERATIONS; i++) {
4490 generations[i].alloc_start_page = 0;
4491 generations[i].alloc_unboxed_start_page = 0;
4492 generations[i].alloc_large_start_page = 0;
4493 generations[i].alloc_large_unboxed_start_page = 0;
4494 generations[i].bytes_allocated = 0;
4495 generations[i].gc_trigger = 2000000;
4496 generations[i].num_gc = 0;
4497 generations[i].cum_sum_bytes_allocated = 0;
4498 /* the tune-able parameters */
4499 generations[i].bytes_consed_between_gc = 2000000;
4500 generations[i].trigger_age = 1;
4501 generations[i].min_av_mem_age = 0.75;
4502 generations[i].lutexes = NULL;
4505 /* Initialize gc_alloc. */
4506 gc_alloc_generation = 0;
4507 gc_set_region_empty(&boxed_region);
4508 gc_set_region_empty(&unboxed_region);
4513 /* Pick up the dynamic space from after a core load.
4515 * The ALLOCATION_POINTER points to the end of the dynamic space.
4519 gencgc_pickup_dynamic(void)
4521 page_index_t page = 0;
4522 long alloc_ptr = get_alloc_pointer();
4523 lispobj *prev=(lispobj *)page_address(page);
4524 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4527 lispobj *first,*ptr= (lispobj *)page_address(page);
4528 page_table[page].allocated = BOXED_PAGE_FLAG;
4529 page_table[page].gen = gen;
4530 page_table[page].bytes_used = PAGE_BYTES;
4531 page_table[page].large_object = 0;
4532 page_table[page].write_protected = 0;
4533 page_table[page].write_protected_cleared = 0;
4534 page_table[page].dont_move = 0;
4535 page_table[page].need_to_zero = 1;
4537 if (!gencgc_partial_pickup) {
4538 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4539 if(ptr == first) prev=ptr;
4540 page_table[page].first_object_offset =
4541 (void *)prev - page_address(page);
4544 } while ((long)page_address(page) < alloc_ptr);
4546 #ifdef LUTEX_WIDETAG
4547 /* Lutexes have been registered in generation 0 by coreparse, and
4548 * need to be moved to the right one manually.
4550 move_lutexes(0, PSEUDO_STATIC_GENERATION);
4553 last_free_page = page;
4555 generations[gen].bytes_allocated = PAGE_BYTES*page;
4556 bytes_allocated = PAGE_BYTES*page;
4558 gc_alloc_update_all_page_tables();
4559 write_protect_generation_pages(gen);
4563 gc_initialize_pointers(void)
4565 gencgc_pickup_dynamic();
4571 /* alloc(..) is the external interface for memory allocation. It
4572 * allocates to generation 0. It is not called from within the garbage
4573 * collector as it is only external uses that need the check for heap
4574 * size (GC trigger) and to disable the interrupts (interrupts are
4575 * always disabled during a GC).
4577 * The vops that call alloc(..) assume that the returned space is zero-filled.
4578 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4580 * The check for a GC trigger is only performed when the current
4581 * region is full, so in most cases it's not needed. */
4586 struct thread *thread=arch_os_get_current_thread();
4587 struct alloc_region *region=
4588 #ifdef LISP_FEATURE_SB_THREAD
4589 thread ? &(thread->alloc_region) : &boxed_region;
4593 #ifndef LISP_FEATURE_WIN32
4594 lispobj alloc_signal;
4597 void *new_free_pointer;
4599 gc_assert(nbytes>0);
4601 /* Check for alignment allocation problems. */
4602 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4603 && ((nbytes & LOWTAG_MASK) == 0));
4607 /* there are a few places in the C code that allocate data in the
4608 * heap before Lisp starts. This is before interrupts are enabled,
4609 * so we don't need to check for pseudo-atomic */
4610 #ifdef LISP_FEATURE_SB_THREAD
4611 if(!get_psuedo_atomic_atomic(th)) {
4613 fprintf(stderr, "fatal error in thread 0x%x, tid=%ld\n",
4615 __asm__("movl %fs,%0" : "=r" (fs) : );
4616 fprintf(stderr, "fs is %x, th->tls_cookie=%x \n",
4617 debug_get_fs(),th->tls_cookie);
4618 lose("If you see this message before 2004.01.31, mail details to sbcl-devel\n");
4621 gc_assert(get_pseudo_atomic_atomic(th));
4625 /* maybe we can do this quickly ... */
4626 new_free_pointer = region->free_pointer + nbytes;
4627 if (new_free_pointer <= region->end_addr) {
4628 new_obj = (void*)(region->free_pointer);
4629 region->free_pointer = new_free_pointer;
4630 return(new_obj); /* yup */
4633 /* we have to go the long way around, it seems. Check whether
4634 * we should GC in the near future
4636 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4637 gc_assert(get_pseudo_atomic_atomic(thread));
4638 /* Don't flood the system with interrupts if the need to gc is
4639 * already noted. This can happen for example when SUB-GC
4640 * allocates or after a gc triggered in a WITHOUT-GCING. */
4641 if (SymbolValue(GC_PENDING,thread) == NIL) {
4642 /* set things up so that GC happens when we finish the PA
4644 SetSymbolValue(GC_PENDING,T,thread);
4645 if (SymbolValue(GC_INHIBIT,thread) == NIL)
4646 set_pseudo_atomic_interrupted(thread);
4649 new_obj = gc_alloc_with_region(nbytes,0,region,0);
4651 #ifndef LISP_FEATURE_WIN32
4652 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4653 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4654 if ((signed long) alloc_signal <= 0) {
4655 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4656 #ifdef LISP_FEATURE_SB_THREAD
4657 kill_thread_safely(thread->os_thread, SIGPROF);
4662 SetSymbolValue(ALLOC_SIGNAL,
4663 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4673 * shared support for the OS-dependent signal handlers which
4674 * catch GENCGC-related write-protect violations
4677 void unhandled_sigmemoryfault(void* addr);
4679 /* Depending on which OS we're running under, different signals might
4680 * be raised for a violation of write protection in the heap. This
4681 * function factors out the common generational GC magic which needs
4682 * to invoked in this case, and should be called from whatever signal
4683 * handler is appropriate for the OS we're running under.
4685 * Return true if this signal is a normal generational GC thing that
4686 * we were able to handle, or false if it was abnormal and control
4687 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4690 gencgc_handle_wp_violation(void* fault_addr)
4692 page_index_t page_index = find_page_index(fault_addr);
4694 #ifdef QSHOW_SIGNALS
4695 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4696 fault_addr, page_index));
4699 /* Check whether the fault is within the dynamic space. */
4700 if (page_index == (-1)) {
4702 /* It can be helpful to be able to put a breakpoint on this
4703 * case to help diagnose low-level problems. */
4704 unhandled_sigmemoryfault(fault_addr);
4706 /* not within the dynamic space -- not our responsibility */
4710 if (page_table[page_index].write_protected) {
4711 /* Unprotect the page. */
4712 os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
4713 page_table[page_index].write_protected_cleared = 1;
4714 page_table[page_index].write_protected = 0;
4716 /* The only acceptable reason for this signal on a heap
4717 * access is that GENCGC write-protected the page.
4718 * However, if two CPUs hit a wp page near-simultaneously,
4719 * we had better not have the second one lose here if it
4720 * does this test after the first one has already set wp=0
4722 if(page_table[page_index].write_protected_cleared != 1)
4723 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4724 page_index, boxed_region.first_page, boxed_region.last_page);
4726 /* Don't worry, we can handle it. */
4730 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4731 * it's not just a case of the program hitting the write barrier, and
4732 * are about to let Lisp deal with it. It's basically just a
4733 * convenient place to set a gdb breakpoint. */
4735 unhandled_sigmemoryfault(void *addr)
4738 void gc_alloc_update_all_page_tables(void)
4740 /* Flush the alloc regions updating the tables. */
4743 gc_alloc_update_page_tables(0, &th->alloc_region);
4744 gc_alloc_update_page_tables(1, &unboxed_region);
4745 gc_alloc_update_page_tables(0, &boxed_region);
4749 gc_set_region_empty(struct alloc_region *region)
4751 region->first_page = 0;
4752 region->last_page = -1;
4753 region->start_addr = page_address(0);
4754 region->free_pointer = page_address(0);
4755 region->end_addr = page_address(0);
4759 zero_all_free_pages()
4763 for (i = 0; i < last_free_page; i++) {
4764 if (page_table[i].allocated == FREE_PAGE_FLAG) {
4765 #ifdef READ_PROTECT_FREE_PAGES
4766 os_protect(page_address(i),
4775 /* Things to do before doing a final GC before saving a core (without
4778 * + Pages in large_object pages aren't moved by the GC, so we need to
4779 * unset that flag from all pages.
4780 * + The pseudo-static generation isn't normally collected, but it seems
4781 * reasonable to collect it at least when saving a core. So move the
4782 * pages to a normal generation.
4785 prepare_for_final_gc ()
4788 for (i = 0; i < last_free_page; i++) {
4789 page_table[i].large_object = 0;
4790 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4791 int used = page_table[i].bytes_used;
4792 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4793 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4794 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4800 /* Do a non-conservative GC, and then save a core with the initial
4801 * function being set to the value of the static symbol
4802 * SB!VM:RESTART-LISP-FUNCTION */
4804 gc_and_save(char *filename, int prepend_runtime)
4807 void *runtime_bytes = NULL;
4808 size_t runtime_size;
4810 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4815 conservative_stack = 0;
4817 /* The filename might come from Lisp, and be moved by the now
4818 * non-conservative GC. */
4819 filename = strdup(filename);
4821 /* Collect twice: once into relatively high memory, and then back
4822 * into low memory. This compacts the retained data into the lower
4823 * pages, minimizing the size of the core file.
4825 prepare_for_final_gc();
4826 gencgc_alloc_start_page = last_free_page;
4827 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4829 prepare_for_final_gc();
4830 gencgc_alloc_start_page = -1;
4831 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4833 if (prepend_runtime)
4834 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4836 /* The dumper doesn't know that pages need to be zeroed before use. */
4837 zero_all_free_pages();
4838 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4840 /* Oops. Save still managed to fail. Since we've mangled the stack
4841 * beyond hope, there's not much we can do.
4842 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4843 * going to be rather unsatisfactory too... */
4844 lose("Attempt to save core after non-conservative GC failed.\n");