2 * GENerational Conservative Garbage Collector for SBCL
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
17 * For a review of garbage collection techniques (e.g. generational
18 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
19 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
20 * had been accepted for _ACM Computing Surveys_ and was available
21 * as a PostScript preprint through
22 * <http://www.cs.utexas.edu/users/oops/papers.html>
24 * <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
37 #include "interrupt.h"
42 #include "gc-internal.h"
44 #include "pseudo-atomic.h"
46 #include "genesis/vector.h"
47 #include "genesis/weak-pointer.h"
48 #include "genesis/fdefn.h"
49 #include "genesis/simple-fun.h"
51 #include "genesis/hash-table.h"
52 #include "genesis/instance.h"
53 #include "genesis/layout.h"
55 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
56 #include "genesis/cons.h"
59 /* forward declarations */
60 page_index_t gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes,
68 /* Generations 0-5 are normal collected generations, 6 is only used as
69 * scratch space by the collector, and should never get collected.
72 SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
76 /* Should we use page protection to help avoid the scavenging of pages
77 * that don't have pointers to younger generations? */
78 boolean enable_page_protection = 1;
80 /* the minimum size (in bytes) for a large object*/
81 #if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES)
82 long large_object_size = 4 * GENCGC_ALLOC_GRANULARITY;
83 #elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY)
84 long large_object_size = 4 * GENCGC_CARD_BYTES;
86 long large_object_size = 4 * PAGE_BYTES;
94 /* the verbosity level. All non-error messages are disabled at level 0;
95 * and only a few rare messages are printed at level 1. */
97 boolean gencgc_verbose = 1;
99 boolean gencgc_verbose = 0;
102 /* FIXME: At some point enable the various error-checking things below
103 * and see what they say. */
105 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
106 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
108 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
110 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
111 boolean pre_verify_gen_0 = 0;
113 /* Should we check for bad pointers after gc_free_heap is called
114 * from Lisp PURIFY? */
115 boolean verify_after_free_heap = 0;
117 /* Should we print a note when code objects are found in the dynamic space
118 * during a heap verify? */
119 boolean verify_dynamic_code_check = 0;
121 /* Should we check code objects for fixup errors after they are transported? */
122 boolean check_code_fixups = 0;
124 /* Should we check that newly allocated regions are zero filled? */
125 boolean gencgc_zero_check = 0;
127 /* Should we check that the free space is zero filled? */
128 boolean gencgc_enable_verify_zero_fill = 0;
130 /* Should we check that free pages are zero filled during gc_free_heap
131 * called after Lisp PURIFY? */
132 boolean gencgc_zero_check_during_free_heap = 0;
134 /* When loading a core, don't do a full scan of the memory for the
135 * memory region boundaries. (Set to true by coreparse.c if the core
136 * contained a pagetable entry).
138 boolean gencgc_partial_pickup = 0;
140 /* If defined, free pages are read-protected to ensure that nothing
144 /* #define READ_PROTECT_FREE_PAGES */
148 * GC structures and variables
151 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
152 os_vm_size_t bytes_allocated = 0;
153 os_vm_size_t auto_gc_trigger = 0;
155 /* the source and destination generations. These are set before a GC starts
157 generation_index_t from_space;
158 generation_index_t new_space;
160 /* Set to 1 when in GC */
161 boolean gc_active_p = 0;
163 /* should the GC be conservative on stack. If false (only right before
164 * saving a core), don't scan the stack / mark pages dont_move. */
165 static boolean conservative_stack = 1;
167 /* An array of page structures is allocated on gc initialization.
168 * This helps quickly map between an address its page structure.
169 * page_table_pages is set from the size of the dynamic space. */
170 page_index_t page_table_pages;
171 struct page *page_table;
173 static inline boolean page_allocated_p(page_index_t page) {
174 return (page_table[page].allocated != FREE_PAGE_FLAG);
177 static inline boolean page_no_region_p(page_index_t page) {
178 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
181 static inline boolean page_allocated_no_region_p(page_index_t page) {
182 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
183 && page_no_region_p(page));
186 static inline boolean page_free_p(page_index_t page) {
187 return (page_table[page].allocated == FREE_PAGE_FLAG);
190 static inline boolean page_boxed_p(page_index_t page) {
191 return (page_table[page].allocated & BOXED_PAGE_FLAG);
194 static inline boolean code_page_p(page_index_t page) {
195 return (page_table[page].allocated & CODE_PAGE_FLAG);
198 static inline boolean page_boxed_no_region_p(page_index_t page) {
199 return page_boxed_p(page) && page_no_region_p(page);
202 static inline boolean page_unboxed_p(page_index_t page) {
203 /* Both flags set == boxed code page */
204 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
205 && !page_boxed_p(page));
208 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
209 return (page_boxed_no_region_p(page)
210 && (page_table[page].bytes_used != 0)
211 && !page_table[page].dont_move
212 && (page_table[page].gen == generation));
215 /* To map addresses to page structures the address of the first page
217 static void *heap_base = NULL;
219 /* Calculate the start address for the given page number. */
221 page_address(page_index_t page_num)
223 return (heap_base + (page_num * GENCGC_CARD_BYTES));
226 /* Calculate the address where the allocation region associated with
227 * the page starts. */
229 page_region_start(page_index_t page_index)
231 return page_address(page_index)-page_table[page_index].region_start_offset;
234 /* Find the page index within the page_table for the given
235 * address. Return -1 on failure. */
237 find_page_index(void *addr)
239 if (addr >= heap_base) {
240 page_index_t index = ((pointer_sized_uint_t)addr -
241 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
242 if (index < page_table_pages)
249 npage_bytes(page_index_t npages)
251 gc_assert(npages>=0);
252 return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
255 /* Check that X is a higher address than Y and return offset from Y to
257 static inline os_vm_size_t
258 void_diff(void *x, void *y)
261 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
264 /* a structure to hold the state of a generation
266 * CAUTION: If you modify this, make sure to touch up the alien
267 * definition in src/code/gc.lisp accordingly. ...or better yes,
268 * deal with the FIXME there...
272 /* the first page that gc_alloc() checks on its next call */
273 page_index_t alloc_start_page;
275 /* the first page that gc_alloc_unboxed() checks on its next call */
276 page_index_t alloc_unboxed_start_page;
278 /* the first page that gc_alloc_large (boxed) considers on its next
279 * call. (Although it always allocates after the boxed_region.) */
280 page_index_t alloc_large_start_page;
282 /* the first page that gc_alloc_large (unboxed) considers on its
283 * next call. (Although it always allocates after the
284 * current_unboxed_region.) */
285 page_index_t alloc_large_unboxed_start_page;
287 /* the bytes allocated to this generation */
288 os_vm_size_t bytes_allocated;
290 /* the number of bytes at which to trigger a GC */
291 os_vm_size_t gc_trigger;
293 /* to calculate a new level for gc_trigger */
294 os_vm_size_t bytes_consed_between_gc;
296 /* the number of GCs since the last raise */
299 /* the number of GCs to run on the generations before raising objects to the
301 int number_of_gcs_before_promotion;
303 /* the cumulative sum of the bytes allocated to this generation. It is
304 * cleared after a GC on this generations, and update before new
305 * objects are added from a GC of a younger generation. Dividing by
306 * the bytes_allocated will give the average age of the memory in
307 * this generation since its last GC. */
308 os_vm_size_t cum_sum_bytes_allocated;
310 /* a minimum average memory age before a GC will occur helps
311 * prevent a GC when a large number of new live objects have been
312 * added, in which case a GC could be a waste of time */
313 double minimum_age_before_gc;
316 /* an array of generation structures. There needs to be one more
317 * generation structure than actual generations as the oldest
318 * generation is temporarily raised then lowered. */
319 struct generation generations[NUM_GENERATIONS];
321 /* the oldest generation that is will currently be GCed by default.
322 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
324 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
326 * Setting this to 0 effectively disables the generational nature of
327 * the GC. In some applications generational GC may not be useful
328 * because there are no long-lived objects.
330 * An intermediate value could be handy after moving long-lived data
331 * into an older generation so an unnecessary GC of this long-lived
332 * data can be avoided. */
333 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
335 /* The maximum free page in the heap is maintained and used to update
336 * ALLOCATION_POINTER which is used by the room function to limit its
337 * search of the heap. XX Gencgc obviously needs to be better
338 * integrated with the Lisp code. */
339 page_index_t last_free_page;
341 #ifdef LISP_FEATURE_SB_THREAD
342 /* This lock is to prevent multiple threads from simultaneously
343 * allocating new regions which overlap each other. Note that the
344 * majority of GC is single-threaded, but alloc() may be called from
345 * >1 thread at a time and must be thread-safe. This lock must be
346 * seized before all accesses to generations[] or to parts of
347 * page_table[] that other threads may want to see */
348 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
349 /* This lock is used to protect non-thread-local allocation. */
350 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
353 extern os_vm_size_t gencgc_release_granularity;
354 os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
356 extern os_vm_size_t gencgc_alloc_granularity;
357 os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
361 * miscellaneous heap functions
364 /* Count the number of pages which are write-protected within the
365 * given generation. */
367 count_write_protect_generation_pages(generation_index_t generation)
369 page_index_t i, count = 0;
371 for (i = 0; i < last_free_page; i++)
372 if (page_allocated_p(i)
373 && (page_table[i].gen == generation)
374 && (page_table[i].write_protected == 1))
379 /* Count the number of pages within the given generation. */
381 count_generation_pages(generation_index_t generation)
384 page_index_t count = 0;
386 for (i = 0; i < last_free_page; i++)
387 if (page_allocated_p(i)
388 && (page_table[i].gen == generation))
395 count_dont_move_pages(void)
398 page_index_t count = 0;
399 for (i = 0; i < last_free_page; i++) {
400 if (page_allocated_p(i)
401 && (page_table[i].dont_move != 0)) {
409 /* Work through the pages and add up the number of bytes used for the
410 * given generation. */
412 count_generation_bytes_allocated (generation_index_t gen)
415 os_vm_size_t result = 0;
416 for (i = 0; i < last_free_page; i++) {
417 if (page_allocated_p(i)
418 && (page_table[i].gen == gen))
419 result += page_table[i].bytes_used;
424 /* Return the average age of the memory in a generation. */
426 generation_average_age(generation_index_t gen)
428 if (generations[gen].bytes_allocated == 0)
432 ((double)generations[gen].cum_sum_bytes_allocated)
433 / ((double)generations[gen].bytes_allocated);
437 write_generation_stats(FILE *file)
439 generation_index_t i;
441 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
442 #define FPU_STATE_SIZE 27
443 int fpu_state[FPU_STATE_SIZE];
444 #elif defined(LISP_FEATURE_PPC)
445 #define FPU_STATE_SIZE 32
446 long long fpu_state[FPU_STATE_SIZE];
449 /* This code uses the FP instructions which may be set up for Lisp
450 * so they need to be saved and reset for C. */
453 /* Print the heap stats. */
455 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
457 for (i = 0; i < SCRATCH_GENERATION; i++) {
459 page_index_t boxed_cnt = 0;
460 page_index_t unboxed_cnt = 0;
461 page_index_t large_boxed_cnt = 0;
462 page_index_t large_unboxed_cnt = 0;
463 page_index_t pinned_cnt=0;
465 for (j = 0; j < last_free_page; j++)
466 if (page_table[j].gen == i) {
468 /* Count the number of boxed pages within the given
470 if (page_boxed_p(j)) {
471 if (page_table[j].large_object)
476 if(page_table[j].dont_move) pinned_cnt++;
477 /* Count the number of unboxed pages within the given
479 if (page_unboxed_p(j)) {
480 if (page_table[j].large_object)
487 gc_assert(generations[i].bytes_allocated
488 == count_generation_bytes_allocated(i));
490 " %1d: %5ld %5ld %5ld %5ld",
492 generations[i].alloc_start_page,
493 generations[i].alloc_unboxed_start_page,
494 generations[i].alloc_large_start_page,
495 generations[i].alloc_large_unboxed_start_page);
497 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
498 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
499 boxed_cnt, unboxed_cnt, large_boxed_cnt,
500 large_unboxed_cnt, pinned_cnt);
505 " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
506 generations[i].bytes_allocated,
507 (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
508 generations[i].gc_trigger,
509 count_write_protect_generation_pages(i),
510 generations[i].num_gc,
511 generation_average_age(i));
513 fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
514 fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);
516 fpu_restore(fpu_state);
520 write_heap_exhaustion_report(FILE *file, long available, long requested,
521 struct thread *thread)
524 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
525 gc_active_p ? "garbage collection" : "allocation",
528 write_generation_stats(file);
529 fprintf(file, "GC control variables:\n");
530 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
531 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
532 (SymbolValue(GC_PENDING, thread) == T) ?
533 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
534 "false" : "in progress"));
535 #ifdef LISP_FEATURE_SB_THREAD
536 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
537 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
542 print_generation_stats(void)
544 write_generation_stats(stderr);
547 extern char* gc_logfile;
548 char * gc_logfile = NULL;
551 log_generation_stats(char *logfile, char *header)
554 FILE * log = fopen(logfile, "a");
556 fprintf(log, "%s\n", header);
557 write_generation_stats(log);
560 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
567 report_heap_exhaustion(long available, long requested, struct thread *th)
570 FILE * log = fopen(gc_logfile, "a");
572 write_heap_exhaustion_report(log, available, requested, th);
575 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
579 /* Always to stderr as well. */
580 write_heap_exhaustion_report(stderr, available, requested, th);
584 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
585 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
588 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
589 * if zeroing it ourselves, i.e. in practice give the memory back to the
590 * OS. Generally done after a large GC.
592 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
594 void *addr = page_address(start), *new_addr;
595 os_vm_size_t length = npage_bytes(1+end-start);
600 gc_assert(length >= gencgc_release_granularity);
601 gc_assert((length % gencgc_release_granularity) == 0);
603 os_invalidate(addr, length);
604 new_addr = os_validate(addr, length);
605 if (new_addr == NULL || new_addr != addr) {
606 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
610 for (i = start; i <= end; i++) {
611 page_table[i].need_to_zero = 0;
615 /* Zero the pages from START to END (inclusive). Generally done just after
616 * a new region has been allocated.
619 zero_pages(page_index_t start, page_index_t end) {
623 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
624 fast_bzero(page_address(start), npage_bytes(1+end-start));
626 bzero(page_address(start), npage_bytes(1+end-start));
632 zero_and_mark_pages(page_index_t start, page_index_t end) {
635 zero_pages(start, end);
636 for (i = start; i <= end; i++)
637 page_table[i].need_to_zero = 0;
640 /* Zero the pages from START to END (inclusive), except for those
641 * pages that are known to already zeroed. Mark all pages in the
642 * ranges as non-zeroed.
645 zero_dirty_pages(page_index_t start, page_index_t end) {
648 for (i = start; i <= end; i++) {
649 if (!page_table[i].need_to_zero) continue;
650 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
655 for (i = start; i <= end; i++) {
656 page_table[i].need_to_zero = 1;
662 * To support quick and inline allocation, regions of memory can be
663 * allocated and then allocated from with just a free pointer and a
664 * check against an end address.
666 * Since objects can be allocated to spaces with different properties
667 * e.g. boxed/unboxed, generation, ages; there may need to be many
668 * allocation regions.
670 * Each allocation region may start within a partly used page. Many
671 * features of memory use are noted on a page wise basis, e.g. the
672 * generation; so if a region starts within an existing allocated page
673 * it must be consistent with this page.
675 * During the scavenging of the newspace, objects will be transported
676 * into an allocation region, and pointers updated to point to this
677 * allocation region. It is possible that these pointers will be
678 * scavenged again before the allocation region is closed, e.g. due to
679 * trans_list which jumps all over the place to cleanup the list. It
680 * is important to be able to determine properties of all objects
681 * pointed to when scavenging, e.g to detect pointers to the oldspace.
682 * Thus it's important that the allocation regions have the correct
683 * properties set when allocated, and not just set when closed. The
684 * region allocation routines return regions with the specified
685 * properties, and grab all the pages, setting their properties
686 * appropriately, except that the amount used is not known.
688 * These regions are used to support quicker allocation using just a
689 * free pointer. The actual space used by the region is not reflected
690 * in the pages tables until it is closed. It can't be scavenged until
693 * When finished with the region it should be closed, which will
694 * update the page tables for the actual space used returning unused
695 * space. Further it may be noted in the new regions which is
696 * necessary when scavenging the newspace.
698 * Large objects may be allocated directly without an allocation
699 * region, the page tables are updated immediately.
701 * Unboxed objects don't contain pointers to other objects and so
702 * don't need scavenging. Further they can't contain pointers to
703 * younger generations so WP is not needed. By allocating pages to
704 * unboxed objects the whole page never needs scavenging or
705 * write-protecting. */
707 /* We are only using two regions at present. Both are for the current
708 * newspace generation. */
709 struct alloc_region boxed_region;
710 struct alloc_region unboxed_region;
712 /* The generation currently being allocated to. */
713 static generation_index_t gc_alloc_generation;
715 static inline page_index_t
716 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
719 if (UNBOXED_PAGE_FLAG == page_type_flag) {
720 return generations[generation].alloc_large_unboxed_start_page;
721 } else if (BOXED_PAGE_FLAG & page_type_flag) {
722 /* Both code and data. */
723 return generations[generation].alloc_large_start_page;
725 lose("bad page type flag: %d", page_type_flag);
728 if (UNBOXED_PAGE_FLAG == page_type_flag) {
729 return generations[generation].alloc_unboxed_start_page;
730 } else if (BOXED_PAGE_FLAG & page_type_flag) {
731 /* Both code and data. */
732 return generations[generation].alloc_start_page;
734 lose("bad page_type_flag: %d", page_type_flag);
740 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
744 if (UNBOXED_PAGE_FLAG == page_type_flag) {
745 generations[generation].alloc_large_unboxed_start_page = page;
746 } else if (BOXED_PAGE_FLAG & page_type_flag) {
747 /* Both code and data. */
748 generations[generation].alloc_large_start_page = page;
750 lose("bad page type flag: %d", page_type_flag);
753 if (UNBOXED_PAGE_FLAG == page_type_flag) {
754 generations[generation].alloc_unboxed_start_page = page;
755 } else if (BOXED_PAGE_FLAG & page_type_flag) {
756 /* Both code and data. */
757 generations[generation].alloc_start_page = page;
759 lose("bad page type flag: %d", page_type_flag);
764 /* Find a new region with room for at least the given number of bytes.
766 * It starts looking at the current generation's alloc_start_page. So
767 * may pick up from the previous region if there is enough space. This
768 * keeps the allocation contiguous when scavenging the newspace.
770 * The alloc_region should have been closed by a call to
771 * gc_alloc_update_page_tables(), and will thus be in an empty state.
773 * To assist the scavenging functions write-protected pages are not
774 * used. Free pages should not be write-protected.
776 * It is critical to the conservative GC that the start of regions be
777 * known. To help achieve this only small regions are allocated at a
780 * During scavenging, pointers may be found to within the current
781 * region and the page generation must be set so that pointers to the
782 * from space can be recognized. Therefore the generation of pages in
783 * the region are set to gc_alloc_generation. To prevent another
784 * allocation call using the same pages, all the pages in the region
785 * are allocated, although they will initially be empty.
788 gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
790 page_index_t first_page;
791 page_index_t last_page;
792 os_vm_size_t bytes_found;
798 "/alloc_new_region for %d bytes from gen %d\n",
799 nbytes, gc_alloc_generation));
802 /* Check that the region is in a reset state. */
803 gc_assert((alloc_region->first_page == 0)
804 && (alloc_region->last_page == -1)
805 && (alloc_region->free_pointer == alloc_region->end_addr));
806 ret = thread_mutex_lock(&free_pages_lock);
808 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
809 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
810 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
811 + npage_bytes(last_page-first_page);
813 /* Set up the alloc_region. */
814 alloc_region->first_page = first_page;
815 alloc_region->last_page = last_page;
816 alloc_region->start_addr = page_table[first_page].bytes_used
817 + page_address(first_page);
818 alloc_region->free_pointer = alloc_region->start_addr;
819 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
821 /* Set up the pages. */
823 /* The first page may have already been in use. */
824 if (page_table[first_page].bytes_used == 0) {
825 page_table[first_page].allocated = page_type_flag;
826 page_table[first_page].gen = gc_alloc_generation;
827 page_table[first_page].large_object = 0;
828 page_table[first_page].region_start_offset = 0;
831 gc_assert(page_table[first_page].allocated == page_type_flag);
832 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
834 gc_assert(page_table[first_page].gen == gc_alloc_generation);
835 gc_assert(page_table[first_page].large_object == 0);
837 for (i = first_page+1; i <= last_page; i++) {
838 page_table[i].allocated = page_type_flag;
839 page_table[i].gen = gc_alloc_generation;
840 page_table[i].large_object = 0;
841 /* This may not be necessary for unboxed regions (think it was
843 page_table[i].region_start_offset =
844 void_diff(page_address(i),alloc_region->start_addr);
845 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
847 /* Bump up last_free_page. */
848 if (last_page+1 > last_free_page) {
849 last_free_page = last_page+1;
850 /* do we only want to call this on special occasions? like for
852 set_alloc_pointer((lispobj)page_address(last_free_page));
854 ret = thread_mutex_unlock(&free_pages_lock);
857 #ifdef READ_PROTECT_FREE_PAGES
858 os_protect(page_address(first_page),
859 npage_bytes(1+last_page-first_page),
863 /* If the first page was only partial, don't check whether it's
864 * zeroed (it won't be) and don't zero it (since the parts that
865 * we're interested in are guaranteed to be zeroed).
867 if (page_table[first_page].bytes_used) {
871 zero_dirty_pages(first_page, last_page);
873 /* we can do this after releasing free_pages_lock */
874 if (gencgc_zero_check) {
876 for (p = (long *)alloc_region->start_addr;
877 p < (long *)alloc_region->end_addr; p++) {
879 /* KLUDGE: It would be nice to use %lx and explicit casts
880 * (long) in code like this, so that it is less likely to
881 * break randomly when running on a machine with different
882 * word sizes. -- WHN 19991129 */
883 lose("The new region at %x is not zero (start=%p, end=%p).\n",
884 p, alloc_region->start_addr, alloc_region->end_addr);
890 /* If the record_new_objects flag is 2 then all new regions created
893 * If it's 1 then then it is only recorded if the first page of the
894 * current region is <= new_areas_ignore_page. This helps avoid
895 * unnecessary recording when doing full scavenge pass.
897 * The new_object structure holds the page, byte offset, and size of
898 * new regions of objects. Each new area is placed in the array of
899 * these structures pointer to by new_areas. new_areas_index holds the
900 * offset into new_areas.
902 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
903 * later code must detect this and handle it, probably by doing a full
904 * scavenge of a generation. */
905 #define NUM_NEW_AREAS 512
906 static int record_new_objects = 0;
907 static page_index_t new_areas_ignore_page;
913 static struct new_area (*new_areas)[];
914 static long new_areas_index;
917 /* Add a new area to new_areas. */
919 add_new_area(page_index_t first_page, size_t offset, size_t size)
921 unsigned long new_area_start,c;
924 /* Ignore if full. */
925 if (new_areas_index >= NUM_NEW_AREAS)
928 switch (record_new_objects) {
932 if (first_page > new_areas_ignore_page)
941 new_area_start = npage_bytes(first_page) + offset;
943 /* Search backwards for a prior area that this follows from. If
944 found this will save adding a new area. */
945 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
946 unsigned long area_end =
947 npage_bytes((*new_areas)[i].page)
948 + (*new_areas)[i].offset
949 + (*new_areas)[i].size;
951 "/add_new_area S1 %d %d %d %d\n",
952 i, c, new_area_start, area_end));*/
953 if (new_area_start == area_end) {
955 "/adding to [%d] %d %d %d with %d %d %d:\n",
957 (*new_areas)[i].page,
958 (*new_areas)[i].offset,
959 (*new_areas)[i].size,
963 (*new_areas)[i].size += size;
968 (*new_areas)[new_areas_index].page = first_page;
969 (*new_areas)[new_areas_index].offset = offset;
970 (*new_areas)[new_areas_index].size = size;
972 "/new_area %d page %d offset %d size %d\n",
973 new_areas_index, first_page, offset, size));*/
976 /* Note the max new_areas used. */
977 if (new_areas_index > max_new_areas)
978 max_new_areas = new_areas_index;
981 /* Update the tables for the alloc_region. The region may be added to
984 * When done the alloc_region is set up so that the next quick alloc
985 * will fail safely and thus a new region will be allocated. Further
986 * it is safe to try to re-update the page table of this reset
989 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
992 page_index_t first_page;
993 page_index_t next_page;
994 unsigned long bytes_used;
995 unsigned long orig_first_page_bytes_used;
996 unsigned long region_size;
997 unsigned long byte_cnt;
1001 first_page = alloc_region->first_page;
1003 /* Catch an unused alloc_region. */
1004 if ((first_page == 0) && (alloc_region->last_page == -1))
1007 next_page = first_page+1;
1009 ret = thread_mutex_lock(&free_pages_lock);
1010 gc_assert(ret == 0);
1011 if (alloc_region->free_pointer != alloc_region->start_addr) {
1012 /* some bytes were allocated in the region */
1013 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1015 gc_assert(alloc_region->start_addr ==
1016 (page_address(first_page)
1017 + page_table[first_page].bytes_used));
1019 /* All the pages used need to be updated */
1021 /* Update the first page. */
1023 /* If the page was free then set up the gen, and
1024 * region_start_offset. */
1025 if (page_table[first_page].bytes_used == 0)
1026 gc_assert(page_table[first_page].region_start_offset == 0);
1027 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1029 gc_assert(page_table[first_page].allocated & page_type_flag);
1030 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1031 gc_assert(page_table[first_page].large_object == 0);
1035 /* Calculate the number of bytes used in this page. This is not
1036 * always the number of new bytes, unless it was free. */
1038 if ((bytes_used = void_diff(alloc_region->free_pointer,
1039 page_address(first_page)))
1040 >GENCGC_CARD_BYTES) {
1041 bytes_used = GENCGC_CARD_BYTES;
1044 page_table[first_page].bytes_used = bytes_used;
1045 byte_cnt += bytes_used;
1048 /* All the rest of the pages should be free. We need to set
1049 * their region_start_offset pointer to the start of the
1050 * region, and set the bytes_used. */
1052 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1053 gc_assert(page_table[next_page].allocated & page_type_flag);
1054 gc_assert(page_table[next_page].bytes_used == 0);
1055 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1056 gc_assert(page_table[next_page].large_object == 0);
1058 gc_assert(page_table[next_page].region_start_offset ==
1059 void_diff(page_address(next_page),
1060 alloc_region->start_addr));
1062 /* Calculate the number of bytes used in this page. */
1064 if ((bytes_used = void_diff(alloc_region->free_pointer,
1065 page_address(next_page)))>GENCGC_CARD_BYTES) {
1066 bytes_used = GENCGC_CARD_BYTES;
1069 page_table[next_page].bytes_used = bytes_used;
1070 byte_cnt += bytes_used;
1075 region_size = void_diff(alloc_region->free_pointer,
1076 alloc_region->start_addr);
1077 bytes_allocated += region_size;
1078 generations[gc_alloc_generation].bytes_allocated += region_size;
1080 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1082 /* Set the generations alloc restart page to the last page of
1084 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1086 /* Add the region to the new_areas if requested. */
1087 if (BOXED_PAGE_FLAG & page_type_flag)
1088 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1092 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1094 gc_alloc_generation));
1097 /* There are no bytes allocated. Unallocate the first_page if
1098 * there are 0 bytes_used. */
1099 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1100 if (page_table[first_page].bytes_used == 0)
1101 page_table[first_page].allocated = FREE_PAGE_FLAG;
1104 /* Unallocate any unused pages. */
1105 while (next_page <= alloc_region->last_page) {
1106 gc_assert(page_table[next_page].bytes_used == 0);
1107 page_table[next_page].allocated = FREE_PAGE_FLAG;
1110 ret = thread_mutex_unlock(&free_pages_lock);
1111 gc_assert(ret == 0);
1113 /* alloc_region is per-thread, we're ok to do this unlocked */
1114 gc_set_region_empty(alloc_region);
1117 static inline void *gc_quick_alloc(long nbytes);
1119 /* Allocate a possibly large object. */
1121 gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region)
1123 page_index_t first_page;
1124 page_index_t last_page;
1125 int orig_first_page_bytes_used;
1128 unsigned long bytes_used;
1129 page_index_t next_page;
1132 ret = thread_mutex_lock(&free_pages_lock);
1133 gc_assert(ret == 0);
1135 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1136 if (first_page <= alloc_region->last_page) {
1137 first_page = alloc_region->last_page+1;
1140 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1142 gc_assert(first_page > alloc_region->last_page);
1144 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1146 /* Set up the pages. */
1147 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1149 /* If the first page was free then set up the gen, and
1150 * region_start_offset. */
1151 if (page_table[first_page].bytes_used == 0) {
1152 page_table[first_page].allocated = page_type_flag;
1153 page_table[first_page].gen = gc_alloc_generation;
1154 page_table[first_page].region_start_offset = 0;
1155 page_table[first_page].large_object = 1;
1158 gc_assert(page_table[first_page].allocated == page_type_flag);
1159 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1160 gc_assert(page_table[first_page].large_object == 1);
1164 /* Calc. the number of bytes used in this page. This is not
1165 * always the number of new bytes, unless it was free. */
1167 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1168 bytes_used = GENCGC_CARD_BYTES;
1171 page_table[first_page].bytes_used = bytes_used;
1172 byte_cnt += bytes_used;
1174 next_page = first_page+1;
1176 /* All the rest of the pages should be free. We need to set their
1177 * region_start_offset pointer to the start of the region, and set
1178 * the bytes_used. */
1180 gc_assert(page_free_p(next_page));
1181 gc_assert(page_table[next_page].bytes_used == 0);
1182 page_table[next_page].allocated = page_type_flag;
1183 page_table[next_page].gen = gc_alloc_generation;
1184 page_table[next_page].large_object = 1;
1186 page_table[next_page].region_start_offset =
1187 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1189 /* Calculate the number of bytes used in this page. */
1191 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1192 if (bytes_used > GENCGC_CARD_BYTES) {
1193 bytes_used = GENCGC_CARD_BYTES;
1196 page_table[next_page].bytes_used = bytes_used;
1197 page_table[next_page].write_protected=0;
1198 page_table[next_page].dont_move=0;
1199 byte_cnt += bytes_used;
1203 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1205 bytes_allocated += nbytes;
1206 generations[gc_alloc_generation].bytes_allocated += nbytes;
1208 /* Add the region to the new_areas if requested. */
1209 if (BOXED_PAGE_FLAG & page_type_flag)
1210 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1212 /* Bump up last_free_page */
1213 if (last_page+1 > last_free_page) {
1214 last_free_page = last_page+1;
1215 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1217 ret = thread_mutex_unlock(&free_pages_lock);
1218 gc_assert(ret == 0);
1220 #ifdef READ_PROTECT_FREE_PAGES
1221 os_protect(page_address(first_page),
1222 npage_bytes(1+last_page-first_page),
1226 zero_dirty_pages(first_page, last_page);
1228 return page_address(first_page);
1231 static page_index_t gencgc_alloc_start_page = -1;
1234 gc_heap_exhausted_error_or_lose (long available, long requested)
1236 struct thread *thread = arch_os_get_current_thread();
1237 /* Write basic information before doing anything else: if we don't
1238 * call to lisp this is a must, and even if we do there is always
1239 * the danger that we bounce back here before the error has been
1240 * handled, or indeed even printed.
1242 report_heap_exhaustion(available, requested, thread);
1243 if (gc_active_p || (available == 0)) {
1244 /* If we are in GC, or totally out of memory there is no way
1245 * to sanely transfer control to the lisp-side of things.
1247 lose("Heap exhausted, game over.");
1250 /* FIXME: assert free_pages_lock held */
1251 (void)thread_mutex_unlock(&free_pages_lock);
1252 gc_assert(get_pseudo_atomic_atomic(thread));
1253 clear_pseudo_atomic_atomic(thread);
1254 if (get_pseudo_atomic_interrupted(thread))
1255 do_pending_interrupt();
1256 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1257 * to running user code at arbitrary places, even in a
1258 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1259 * running out of the heap. So at this point all bets are
1261 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1262 corruption_warning_and_maybe_lose
1263 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1264 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1265 alloc_number(available), alloc_number(requested));
1266 lose("HEAP-EXHAUSTED-ERROR fell through");
1271 gc_find_freeish_pages(page_index_t *restart_page_ptr, long bytes,
1274 page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
1275 page_index_t first_page, last_page, restart_page = *restart_page_ptr;
1276 os_vm_size_t nbytes = bytes;
1277 os_vm_size_t nbytes_goal = nbytes;
1278 os_vm_size_t bytes_found = 0;
1279 os_vm_size_t most_bytes_found = 0;
1280 boolean small_object = nbytes < GENCGC_CARD_BYTES;
1281 /* FIXME: assert(free_pages_lock is held); */
1283 if (nbytes_goal < gencgc_alloc_granularity)
1284 nbytes_goal = gencgc_alloc_granularity;
1286 /* Toggled by gc_and_save for heap compaction, normally -1. */
1287 if (gencgc_alloc_start_page != -1) {
1288 restart_page = gencgc_alloc_start_page;
1291 /* FIXME: This is on bytes instead of nbytes pending cleanup of
1292 * long from the interface. */
1293 gc_assert(bytes>=0);
1294 /* Search for a page with at least nbytes of space. We prefer
1295 * not to split small objects on multiple pages, to reduce the
1296 * number of contiguous allocation regions spaning multiple
1297 * pages: this helps avoid excessive conservativism.
1299 * For other objects, we guarantee that they start on their own
1302 first_page = restart_page;
1303 while (first_page < page_table_pages) {
1305 if (page_free_p(first_page)) {
1306 gc_assert(0 == page_table[first_page].bytes_used);
1307 bytes_found = GENCGC_CARD_BYTES;
1308 } else if (small_object &&
1309 (page_table[first_page].allocated == page_type_flag) &&
1310 (page_table[first_page].large_object == 0) &&
1311 (page_table[first_page].gen == gc_alloc_generation) &&
1312 (page_table[first_page].write_protected == 0) &&
1313 (page_table[first_page].dont_move == 0)) {
1314 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1315 if (bytes_found < nbytes) {
1316 if (bytes_found > most_bytes_found)
1317 most_bytes_found = bytes_found;
1326 gc_assert(page_table[first_page].write_protected == 0);
1327 for (last_page = first_page+1;
1328 ((last_page < page_table_pages) &&
1329 page_free_p(last_page) &&
1330 (bytes_found < nbytes_goal));
1332 bytes_found += GENCGC_CARD_BYTES;
1333 gc_assert(0 == page_table[last_page].bytes_used);
1334 gc_assert(0 == page_table[last_page].write_protected);
1337 if (bytes_found > most_bytes_found) {
1338 most_bytes_found = bytes_found;
1339 most_bytes_found_from = first_page;
1340 most_bytes_found_to = last_page;
1342 if (bytes_found >= nbytes_goal)
1345 first_page = last_page;
1348 bytes_found = most_bytes_found;
1349 restart_page = first_page + 1;
1351 /* Check for a failure */
1352 if (bytes_found < nbytes) {
1353 gc_assert(restart_page >= page_table_pages);
1354 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1357 gc_assert(most_bytes_found_to);
1358 *restart_page_ptr = most_bytes_found_from;
1359 return most_bytes_found_to-1;
1362 /* Allocate bytes. All the rest of the special-purpose allocation
1363 * functions will eventually call this */
1366 gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region,
1369 void *new_free_pointer;
1371 if (nbytes>=large_object_size)
1372 return gc_alloc_large(nbytes, page_type_flag, my_region);
1374 /* Check whether there is room in the current alloc region. */
1375 new_free_pointer = my_region->free_pointer + nbytes;
1377 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1378 my_region->free_pointer, new_free_pointer); */
1380 if (new_free_pointer <= my_region->end_addr) {
1381 /* If so then allocate from the current alloc region. */
1382 void *new_obj = my_region->free_pointer;
1383 my_region->free_pointer = new_free_pointer;
1385 /* Unless a `quick' alloc was requested, check whether the
1386 alloc region is almost empty. */
1388 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1389 /* If so, finished with the current region. */
1390 gc_alloc_update_page_tables(page_type_flag, my_region);
1391 /* Set up a new region. */
1392 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1395 return((void *)new_obj);
1398 /* Else not enough free space in the current region: retry with a
1401 gc_alloc_update_page_tables(page_type_flag, my_region);
1402 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1403 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1406 /* these are only used during GC: all allocation from the mutator calls
1407 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1410 static inline void *
1411 gc_quick_alloc(long nbytes)
1413 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1416 static inline void *
1417 gc_quick_alloc_large(long nbytes)
1419 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK);
1422 static inline void *
1423 gc_alloc_unboxed(long nbytes)
1425 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1428 static inline void *
1429 gc_quick_alloc_unboxed(long nbytes)
1431 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1434 static inline void *
1435 gc_quick_alloc_large_unboxed(long nbytes)
1437 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1441 /* Copy a large boxed object. If the object is in a large object
1442 * region then it is simply promoted, else it is copied. If it's large
1443 * enough then it's copied to a large object region.
1445 * Vectors may have shrunk. If the object is not copied the space
1446 * needs to be reclaimed, and the page_tables corrected. */
1448 copy_large_object(lispobj object, long nwords)
1452 page_index_t first_page;
1454 gc_assert(is_lisp_pointer(object));
1455 gc_assert(from_space_p(object));
1456 gc_assert((nwords & 0x01) == 0);
1459 /* Check whether it's in a large object region. */
1460 first_page = find_page_index((void *)object);
1461 gc_assert(first_page >= 0);
1463 if (page_table[first_page].large_object) {
1465 /* Promote the object. */
1467 unsigned long remaining_bytes;
1468 page_index_t next_page;
1469 unsigned long bytes_freed;
1470 unsigned long old_bytes_used;
1472 /* Note: Any page write-protection must be removed, else a
1473 * later scavenge_newspace may incorrectly not scavenge these
1474 * pages. This would not be necessary if they are added to the
1475 * new areas, but let's do it for them all (they'll probably
1476 * be written anyway?). */
1478 gc_assert(page_table[first_page].region_start_offset == 0);
1480 next_page = first_page;
1481 remaining_bytes = nwords*N_WORD_BYTES;
1482 while (remaining_bytes > GENCGC_CARD_BYTES) {
1483 gc_assert(page_table[next_page].gen == from_space);
1484 gc_assert(page_boxed_p(next_page));
1485 gc_assert(page_table[next_page].large_object);
1486 gc_assert(page_table[next_page].region_start_offset ==
1487 npage_bytes(next_page-first_page));
1488 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1489 /* Should have been unprotected by unprotect_oldspace(). */
1490 gc_assert(page_table[next_page].write_protected == 0);
1492 page_table[next_page].gen = new_space;
1494 remaining_bytes -= GENCGC_CARD_BYTES;
1498 /* Now only one page remains, but the object may have shrunk
1499 * so there may be more unused pages which will be freed. */
1501 /* The object may have shrunk but shouldn't have grown. */
1502 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1504 page_table[next_page].gen = new_space;
1505 gc_assert(page_boxed_p(next_page));
1507 /* Adjust the bytes_used. */
1508 old_bytes_used = page_table[next_page].bytes_used;
1509 page_table[next_page].bytes_used = remaining_bytes;
1511 bytes_freed = old_bytes_used - remaining_bytes;
1513 /* Free any remaining pages; needs care. */
1515 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1516 (page_table[next_page].gen == from_space) &&
1517 page_boxed_p(next_page) &&
1518 page_table[next_page].large_object &&
1519 (page_table[next_page].region_start_offset ==
1520 npage_bytes(next_page - first_page))) {
1521 /* Checks out OK, free the page. Don't need to bother zeroing
1522 * pages as this should have been done before shrinking the
1523 * object. These pages shouldn't be write-protected as they
1524 * should be zero filled. */
1525 gc_assert(page_table[next_page].write_protected == 0);
1527 old_bytes_used = page_table[next_page].bytes_used;
1528 page_table[next_page].allocated = FREE_PAGE_FLAG;
1529 page_table[next_page].bytes_used = 0;
1530 bytes_freed += old_bytes_used;
1534 generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords
1536 generations[new_space].bytes_allocated += N_WORD_BYTES*nwords;
1537 bytes_allocated -= bytes_freed;
1539 /* Add the region to the new_areas if requested. */
1540 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1544 /* Get tag of object. */
1545 tag = lowtag_of(object);
1547 /* Allocate space. */
1548 new = gc_quick_alloc_large(nwords*N_WORD_BYTES);
1550 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1552 /* Return Lisp pointer of new object. */
1553 return ((lispobj) new) | tag;
1557 /* to copy unboxed objects */
1559 copy_unboxed_object(lispobj object, long nwords)
1564 gc_assert(is_lisp_pointer(object));
1565 gc_assert(from_space_p(object));
1566 gc_assert((nwords & 0x01) == 0);
1568 /* Get tag of object. */
1569 tag = lowtag_of(object);
1571 /* Allocate space. */
1572 new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES);
1574 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1576 /* Return Lisp pointer of new object. */
1577 return ((lispobj) new) | tag;
1580 /* to copy large unboxed objects
1582 * If the object is in a large object region then it is simply
1583 * promoted, else it is copied. If it's large enough then it's copied
1584 * to a large object region.
1586 * Bignums and vectors may have shrunk. If the object is not copied
1587 * the space needs to be reclaimed, and the page_tables corrected.
1589 * KLUDGE: There's a lot of cut-and-paste duplication between this
1590 * function and copy_large_object(..). -- WHN 20000619 */
1592 copy_large_unboxed_object(lispobj object, long nwords)
1596 page_index_t first_page;
1598 gc_assert(is_lisp_pointer(object));
1599 gc_assert(from_space_p(object));
1600 gc_assert((nwords & 0x01) == 0);
1602 if ((nwords > 1024*1024) && gencgc_verbose) {
1603 FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n",
1604 nwords*N_WORD_BYTES));
1607 /* Check whether it's a large object. */
1608 first_page = find_page_index((void *)object);
1609 gc_assert(first_page >= 0);
1611 if (page_table[first_page].large_object) {
1612 /* Promote the object. Note: Unboxed objects may have been
1613 * allocated to a BOXED region so it may be necessary to
1614 * change the region to UNBOXED. */
1615 unsigned long remaining_bytes;
1616 page_index_t next_page;
1617 unsigned long bytes_freed;
1618 unsigned long old_bytes_used;
1620 gc_assert(page_table[first_page].region_start_offset == 0);
1622 next_page = first_page;
1623 remaining_bytes = nwords*N_WORD_BYTES;
1624 while (remaining_bytes > GENCGC_CARD_BYTES) {
1625 gc_assert(page_table[next_page].gen == from_space);
1626 gc_assert(page_allocated_no_region_p(next_page));
1627 gc_assert(page_table[next_page].large_object);
1628 gc_assert(page_table[next_page].region_start_offset ==
1629 npage_bytes(next_page-first_page));
1630 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1632 page_table[next_page].gen = new_space;
1633 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1634 remaining_bytes -= GENCGC_CARD_BYTES;
1638 /* Now only one page remains, but the object may have shrunk so
1639 * there may be more unused pages which will be freed. */
1641 /* Object may have shrunk but shouldn't have grown - check. */
1642 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1644 page_table[next_page].gen = new_space;
1645 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1647 /* Adjust the bytes_used. */
1648 old_bytes_used = page_table[next_page].bytes_used;
1649 page_table[next_page].bytes_used = remaining_bytes;
1651 bytes_freed = old_bytes_used - remaining_bytes;
1653 /* Free any remaining pages; needs care. */
1655 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1656 (page_table[next_page].gen == from_space) &&
1657 page_allocated_no_region_p(next_page) &&
1658 page_table[next_page].large_object &&
1659 (page_table[next_page].region_start_offset ==
1660 npage_bytes(next_page - first_page))) {
1661 /* Checks out OK, free the page. Don't need to both zeroing
1662 * pages as this should have been done before shrinking the
1663 * object. These pages shouldn't be write-protected, even if
1664 * boxed they should be zero filled. */
1665 gc_assert(page_table[next_page].write_protected == 0);
1667 old_bytes_used = page_table[next_page].bytes_used;
1668 page_table[next_page].allocated = FREE_PAGE_FLAG;
1669 page_table[next_page].bytes_used = 0;
1670 bytes_freed += old_bytes_used;
1674 if ((bytes_freed > 0) && gencgc_verbose) {
1676 "/copy_large_unboxed bytes_freed=%d\n",
1680 generations[from_space].bytes_allocated -=
1681 nwords*N_WORD_BYTES + bytes_freed;
1682 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1683 bytes_allocated -= bytes_freed;
1688 /* Get tag of object. */
1689 tag = lowtag_of(object);
1691 /* Allocate space. */
1692 new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES);
1694 /* Copy the object. */
1695 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1697 /* Return Lisp pointer of new object. */
1698 return ((lispobj) new) | tag;
1707 * code and code-related objects
1710 static lispobj trans_fun_header(lispobj object);
1711 static lispobj trans_boxed(lispobj object);
1714 /* Scan a x86 compiled code object, looking for possible fixups that
1715 * have been missed after a move.
1717 * Two types of fixups are needed:
1718 * 1. Absolute fixups to within the code object.
1719 * 2. Relative fixups to outside the code object.
1721 * Currently only absolute fixups to the constant vector, or to the
1722 * code area are checked. */
1724 sniff_code_object(struct code *code, unsigned long displacement)
1726 #ifdef LISP_FEATURE_X86
1727 long nheader_words, ncode_words, nwords;
1729 void *constants_start_addr = NULL, *constants_end_addr;
1730 void *code_start_addr, *code_end_addr;
1731 int fixup_found = 0;
1733 if (!check_code_fixups)
1736 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1738 ncode_words = fixnum_value(code->code_size);
1739 nheader_words = HeaderValue(*(lispobj *)code);
1740 nwords = ncode_words + nheader_words;
1742 constants_start_addr = (void *)code + 5*N_WORD_BYTES;
1743 constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
1744 code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
1745 code_end_addr = (void *)code + nwords*N_WORD_BYTES;
1747 /* Work through the unboxed code. */
1748 for (p = code_start_addr; p < code_end_addr; p++) {
1749 void *data = *(void **)p;
1750 unsigned d1 = *((unsigned char *)p - 1);
1751 unsigned d2 = *((unsigned char *)p - 2);
1752 unsigned d3 = *((unsigned char *)p - 3);
1753 unsigned d4 = *((unsigned char *)p - 4);
1755 unsigned d5 = *((unsigned char *)p - 5);
1756 unsigned d6 = *((unsigned char *)p - 6);
1759 /* Check for code references. */
1760 /* Check for a 32 bit word that looks like an absolute
1761 reference to within the code adea of the code object. */
1762 if ((data >= (code_start_addr-displacement))
1763 && (data < (code_end_addr-displacement))) {
1764 /* function header */
1766 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1768 /* Skip the function header */
1772 /* the case of PUSH imm32 */
1776 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1777 p, d6, d5, d4, d3, d2, d1, data));
1778 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1780 /* the case of MOV [reg-8],imm32 */
1782 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1783 || d2==0x45 || d2==0x46 || d2==0x47)
1787 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1788 p, d6, d5, d4, d3, d2, d1, data));
1789 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1791 /* the case of LEA reg,[disp32] */
1792 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1795 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1796 p, d6, d5, d4, d3, d2, d1, data));
1797 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1801 /* Check for constant references. */
1802 /* Check for a 32 bit word that looks like an absolute
1803 reference to within the constant vector. Constant references
1805 if ((data >= (constants_start_addr-displacement))
1806 && (data < (constants_end_addr-displacement))
1807 && (((unsigned)data & 0x3) == 0)) {
1812 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1813 p, d6, d5, d4, d3, d2, d1, data));
1814 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1817 /* the case of MOV m32,EAX */
1821 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1822 p, d6, d5, d4, d3, d2, d1, data));
1823 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1826 /* the case of CMP m32,imm32 */
1827 if ((d1 == 0x3d) && (d2 == 0x81)) {
1830 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1831 p, d6, d5, d4, d3, d2, d1, data));
1833 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1836 /* Check for a mod=00, r/m=101 byte. */
1837 if ((d1 & 0xc7) == 5) {
1842 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1843 p, d6, d5, d4, d3, d2, d1, data));
1844 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1846 /* the case of CMP reg32,m32 */
1850 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1851 p, d6, d5, d4, d3, d2, d1, data));
1852 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1854 /* the case of MOV m32,reg32 */
1858 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1859 p, d6, d5, d4, d3, d2, d1, data));
1860 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1862 /* the case of MOV reg32,m32 */
1866 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1867 p, d6, d5, d4, d3, d2, d1, data));
1868 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1870 /* the case of LEA reg32,m32 */
1874 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1875 p, d6, d5, d4, d3, d2, d1, data));
1876 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1882 /* If anything was found, print some information on the code
1886 "/compiled code object at %x: header words = %d, code words = %d\n",
1887 code, nheader_words, ncode_words));
1889 "/const start = %x, end = %x\n",
1890 constants_start_addr, constants_end_addr));
1892 "/code start = %x, end = %x\n",
1893 code_start_addr, code_end_addr));
1899 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1901 /* x86-64 uses pc-relative addressing instead of this kludge */
1902 #ifndef LISP_FEATURE_X86_64
1903 long nheader_words, ncode_words, nwords;
1904 void *constants_start_addr, *constants_end_addr;
1905 void *code_start_addr, *code_end_addr;
1906 lispobj fixups = NIL;
1907 unsigned long displacement =
1908 (unsigned long)new_code - (unsigned long)old_code;
1909 struct vector *fixups_vector;
1911 ncode_words = fixnum_value(new_code->code_size);
1912 nheader_words = HeaderValue(*(lispobj *)new_code);
1913 nwords = ncode_words + nheader_words;
1915 "/compiled code object at %x: header words = %d, code words = %d\n",
1916 new_code, nheader_words, ncode_words)); */
1917 constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
1918 constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1919 code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
1920 code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
1923 "/const start = %x, end = %x\n",
1924 constants_start_addr,constants_end_addr));
1926 "/code start = %x; end = %x\n",
1927 code_start_addr,code_end_addr));
1930 /* The first constant should be a pointer to the fixups for this
1931 code objects. Check. */
1932 fixups = new_code->constants[0];
1934 /* It will be 0 or the unbound-marker if there are no fixups (as
1935 * will be the case if the code object has been purified, for
1936 * example) and will be an other pointer if it is valid. */
1937 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1938 !is_lisp_pointer(fixups)) {
1939 /* Check for possible errors. */
1940 if (check_code_fixups)
1941 sniff_code_object(new_code, displacement);
1946 fixups_vector = (struct vector *)native_pointer(fixups);
1948 /* Could be pointing to a forwarding pointer. */
1949 /* FIXME is this always in from_space? if so, could replace this code with
1950 * forwarding_pointer_p/forwarding_pointer_value */
1951 if (is_lisp_pointer(fixups) &&
1952 (find_page_index((void*)fixups_vector) != -1) &&
1953 (fixups_vector->header == 0x01)) {
1954 /* If so, then follow it. */
1955 /*SHOW("following pointer to a forwarding pointer");*/
1957 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1960 /*SHOW("got fixups");*/
1962 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1963 /* Got the fixups for the code block. Now work through the vector,
1964 and apply a fixup at each address. */
1965 long length = fixnum_value(fixups_vector->length);
1967 for (i = 0; i < length; i++) {
1968 unsigned long offset = fixups_vector->data[i];
1969 /* Now check the current value of offset. */
1970 unsigned long old_value =
1971 *(unsigned long *)((unsigned long)code_start_addr + offset);
1973 /* If it's within the old_code object then it must be an
1974 * absolute fixup (relative ones are not saved) */
1975 if ((old_value >= (unsigned long)old_code)
1976 && (old_value < ((unsigned long)old_code
1977 + nwords*N_WORD_BYTES)))
1978 /* So add the dispacement. */
1979 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1980 old_value + displacement;
1982 /* It is outside the old code object so it must be a
1983 * relative fixup (absolute fixups are not saved). So
1984 * subtract the displacement. */
1985 *(unsigned long *)((unsigned long)code_start_addr + offset) =
1986 old_value - displacement;
1989 /* This used to just print a note to stderr, but a bogus fixup seems to
1990 * indicate real heap corruption, so a hard hailure is in order. */
1991 lose("fixup vector %p has a bad widetag: %d\n",
1992 fixups_vector, widetag_of(fixups_vector->header));
1995 /* Check for possible errors. */
1996 if (check_code_fixups) {
1997 sniff_code_object(new_code,displacement);
2004 trans_boxed_large(lispobj object)
2007 unsigned long length;
2009 gc_assert(is_lisp_pointer(object));
2011 header = *((lispobj *) native_pointer(object));
2012 length = HeaderValue(header) + 1;
2013 length = CEILING(length, 2);
2015 return copy_large_object(object, length);
2018 /* Doesn't seem to be used, delete it after the grace period. */
2021 trans_unboxed_large(lispobj object)
2024 unsigned long length;
2026 gc_assert(is_lisp_pointer(object));
2028 header = *((lispobj *) native_pointer(object));
2029 length = HeaderValue(header) + 1;
2030 length = CEILING(length, 2);
2032 return copy_large_unboxed_object(object, length);
2040 /* XX This is a hack adapted from cgc.c. These don't work too
2041 * efficiently with the gencgc as a list of the weak pointers is
2042 * maintained within the objects which causes writes to the pages. A
2043 * limited attempt is made to avoid unnecessary writes, but this needs
2045 #define WEAK_POINTER_NWORDS \
2046 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
2049 scav_weak_pointer(lispobj *where, lispobj object)
2051 /* Since we overwrite the 'next' field, we have to make
2052 * sure not to do so for pointers already in the list.
2053 * Instead of searching the list of weak_pointers each
2054 * time, we ensure that next is always NULL when the weak
2055 * pointer isn't in the list, and not NULL otherwise.
2056 * Since we can't use NULL to denote end of list, we
2057 * use a pointer back to the same weak_pointer.
2059 struct weak_pointer * wp = (struct weak_pointer*)where;
2061 if (NULL == wp->next) {
2062 wp->next = weak_pointers;
2064 if (NULL == wp->next)
2068 /* Do not let GC scavenge the value slot of the weak pointer.
2069 * (That is why it is a weak pointer.) */
2071 return WEAK_POINTER_NWORDS;
2076 search_read_only_space(void *pointer)
2078 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2079 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2080 if ((pointer < (void *)start) || (pointer >= (void *)end))
2082 return (gc_search_space(start,
2083 (((lispobj *)pointer)+2)-start,
2084 (lispobj *) pointer));
2088 search_static_space(void *pointer)
2090 lispobj *start = (lispobj *)STATIC_SPACE_START;
2091 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2092 if ((pointer < (void *)start) || (pointer >= (void *)end))
2094 return (gc_search_space(start,
2095 (((lispobj *)pointer)+2)-start,
2096 (lispobj *) pointer));
2099 /* a faster version for searching the dynamic space. This will work even
2100 * if the object is in a current allocation region. */
2102 search_dynamic_space(void *pointer)
2104 page_index_t page_index = find_page_index(pointer);
2107 /* The address may be invalid, so do some checks. */
2108 if ((page_index == -1) || page_free_p(page_index))
2110 start = (lispobj *)page_region_start(page_index);
2111 return (gc_search_space(start,
2112 (((lispobj *)pointer)+2)-start,
2113 (lispobj *)pointer));
2116 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2118 /* Is there any possibility that pointer is a valid Lisp object
2119 * reference, and/or something else (e.g. subroutine call return
2120 * address) which should prevent us from moving the referred-to thing?
2121 * This is called from preserve_pointers() */
2123 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2125 lispobj *start_addr;
2127 /* Find the object start address. */
2128 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2132 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2135 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2137 /* Adjust large bignum and vector objects. This will adjust the
2138 * allocated region if the size has shrunk, and move unboxed objects
2139 * into unboxed pages. The pages are not promoted here, and the
2140 * promoted region is not added to the new_regions; this is really
2141 * only designed to be called from preserve_pointer(). Shouldn't fail
2142 * if this is missed, just may delay the moving of objects to unboxed
2143 * pages, and the freeing of pages. */
2145 maybe_adjust_large_object(lispobj *where)
2147 page_index_t first_page;
2148 page_index_t next_page;
2151 unsigned long remaining_bytes;
2152 unsigned long bytes_freed;
2153 unsigned long old_bytes_used;
2157 /* Check whether it's a vector or bignum object. */
2158 switch (widetag_of(where[0])) {
2159 case SIMPLE_VECTOR_WIDETAG:
2160 boxed = BOXED_PAGE_FLAG;
2162 case BIGNUM_WIDETAG:
2163 case SIMPLE_BASE_STRING_WIDETAG:
2164 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2165 case SIMPLE_CHARACTER_STRING_WIDETAG:
2167 case SIMPLE_BIT_VECTOR_WIDETAG:
2168 case SIMPLE_ARRAY_NIL_WIDETAG:
2169 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2170 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2171 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2172 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2173 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2174 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2176 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2178 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2179 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2180 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2181 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2183 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2184 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2186 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2187 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2189 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2190 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2193 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2195 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2196 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2198 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2199 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2201 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2202 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2203 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2204 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2206 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2207 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2209 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2210 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2212 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2213 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2215 boxed = UNBOXED_PAGE_FLAG;
2221 /* Find its current size. */
2222 nwords = (sizetab[widetag_of(where[0])])(where);
2224 first_page = find_page_index((void *)where);
2225 gc_assert(first_page >= 0);
2227 /* Note: Any page write-protection must be removed, else a later
2228 * scavenge_newspace may incorrectly not scavenge these pages.
2229 * This would not be necessary if they are added to the new areas,
2230 * but lets do it for them all (they'll probably be written
2233 gc_assert(page_table[first_page].region_start_offset == 0);
2235 next_page = first_page;
2236 remaining_bytes = nwords*N_WORD_BYTES;
2237 while (remaining_bytes > GENCGC_CARD_BYTES) {
2238 gc_assert(page_table[next_page].gen == from_space);
2239 gc_assert(page_allocated_no_region_p(next_page));
2240 gc_assert(page_table[next_page].large_object);
2241 gc_assert(page_table[next_page].region_start_offset ==
2242 npage_bytes(next_page-first_page));
2243 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2245 page_table[next_page].allocated = boxed;
2247 /* Shouldn't be write-protected at this stage. Essential that the
2249 gc_assert(!page_table[next_page].write_protected);
2250 remaining_bytes -= GENCGC_CARD_BYTES;
2254 /* Now only one page remains, but the object may have shrunk so
2255 * there may be more unused pages which will be freed. */
2257 /* Object may have shrunk but shouldn't have grown - check. */
2258 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2260 page_table[next_page].allocated = boxed;
2261 gc_assert(page_table[next_page].allocated ==
2262 page_table[first_page].allocated);
2264 /* Adjust the bytes_used. */
2265 old_bytes_used = page_table[next_page].bytes_used;
2266 page_table[next_page].bytes_used = remaining_bytes;
2268 bytes_freed = old_bytes_used - remaining_bytes;
2270 /* Free any remaining pages; needs care. */
2272 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2273 (page_table[next_page].gen == from_space) &&
2274 page_allocated_no_region_p(next_page) &&
2275 page_table[next_page].large_object &&
2276 (page_table[next_page].region_start_offset ==
2277 npage_bytes(next_page - first_page))) {
2278 /* It checks out OK, free the page. We don't need to both zeroing
2279 * pages as this should have been done before shrinking the
2280 * object. These pages shouldn't be write protected as they
2281 * should be zero filled. */
2282 gc_assert(page_table[next_page].write_protected == 0);
2284 old_bytes_used = page_table[next_page].bytes_used;
2285 page_table[next_page].allocated = FREE_PAGE_FLAG;
2286 page_table[next_page].bytes_used = 0;
2287 bytes_freed += old_bytes_used;
2291 if ((bytes_freed > 0) && gencgc_verbose) {
2293 "/maybe_adjust_large_object() freed %d\n",
2297 generations[from_space].bytes_allocated -= bytes_freed;
2298 bytes_allocated -= bytes_freed;
2303 /* Take a possible pointer to a Lisp object and mark its page in the
2304 * page_table so that it will not be relocated during a GC.
2306 * This involves locating the page it points to, then backing up to
2307 * the start of its region, then marking all pages dont_move from there
2308 * up to the first page that's not full or has a different generation
2310 * It is assumed that all the page static flags have been cleared at
2311 * the start of a GC.
2313 * It is also assumed that the current gc_alloc() region has been
2314 * flushed and the tables updated. */
2317 preserve_pointer(void *addr)
2319 page_index_t addr_page_index = find_page_index(addr);
2320 page_index_t first_page;
2322 unsigned int region_allocation;
2324 /* quick check 1: Address is quite likely to have been invalid. */
2325 if ((addr_page_index == -1)
2326 || page_free_p(addr_page_index)
2327 || (page_table[addr_page_index].bytes_used == 0)
2328 || (page_table[addr_page_index].gen != from_space)
2329 /* Skip if already marked dont_move. */
2330 || (page_table[addr_page_index].dont_move != 0))
2332 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2333 /* (Now that we know that addr_page_index is in range, it's
2334 * safe to index into page_table[] with it.) */
2335 region_allocation = page_table[addr_page_index].allocated;
2337 /* quick check 2: Check the offset within the page.
2340 if (((unsigned long)addr & (GENCGC_CARD_BYTES - 1)) >
2341 page_table[addr_page_index].bytes_used)
2344 /* Filter out anything which can't be a pointer to a Lisp object
2345 * (or, as a special case which also requires dont_move, a return
2346 * address referring to something in a CodeObject). This is
2347 * expensive but important, since it vastly reduces the
2348 * probability that random garbage will be bogusly interpreted as
2349 * a pointer which prevents a page from moving.
2351 * This only needs to happen on x86oids, where this is used for
2352 * conservative roots. Non-x86oid systems only ever call this
2353 * function on known-valid lisp objects. */
2354 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2355 if (!(code_page_p(addr_page_index)
2356 || (is_lisp_pointer((lispobj)addr) &&
2357 possibly_valid_dynamic_space_pointer(addr))))
2361 /* Find the beginning of the region. Note that there may be
2362 * objects in the region preceding the one that we were passed a
2363 * pointer to: if this is the case, we will write-protect all the
2364 * previous objects' pages too. */
2367 /* I think this'd work just as well, but without the assertions.
2368 * -dan 2004.01.01 */
2369 first_page = find_page_index(page_region_start(addr_page_index))
2371 first_page = addr_page_index;
2372 while (page_table[first_page].region_start_offset != 0) {
2374 /* Do some checks. */
2375 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2376 gc_assert(page_table[first_page].gen == from_space);
2377 gc_assert(page_table[first_page].allocated == region_allocation);
2381 /* Adjust any large objects before promotion as they won't be
2382 * copied after promotion. */
2383 if (page_table[first_page].large_object) {
2384 maybe_adjust_large_object(page_address(first_page));
2385 /* If a large object has shrunk then addr may now point to a
2386 * free area in which case it's ignored here. Note it gets
2387 * through the valid pointer test above because the tail looks
2389 if (page_free_p(addr_page_index)
2390 || (page_table[addr_page_index].bytes_used == 0)
2391 /* Check the offset within the page. */
2392 || (((unsigned long)addr & (GENCGC_CARD_BYTES - 1))
2393 > page_table[addr_page_index].bytes_used)) {
2395 "weird? ignore ptr 0x%x to freed area of large object\n",
2399 /* It may have moved to unboxed pages. */
2400 region_allocation = page_table[first_page].allocated;
2403 /* Now work forward until the end of this contiguous area is found,
2404 * marking all pages as dont_move. */
2405 for (i = first_page; ;i++) {
2406 gc_assert(page_table[i].allocated == region_allocation);
2408 /* Mark the page static. */
2409 page_table[i].dont_move = 1;
2411 /* Move the page to the new_space. XX I'd rather not do this
2412 * but the GC logic is not quite able to copy with the static
2413 * pages remaining in the from space. This also requires the
2414 * generation bytes_allocated counters be updated. */
2415 page_table[i].gen = new_space;
2416 generations[new_space].bytes_allocated += page_table[i].bytes_used;
2417 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
2419 /* It is essential that the pages are not write protected as
2420 * they may have pointers into the old-space which need
2421 * scavenging. They shouldn't be write protected at this
2423 gc_assert(!page_table[i].write_protected);
2425 /* Check whether this is the last page in this contiguous block.. */
2426 if ((page_table[i].bytes_used < GENCGC_CARD_BYTES)
2427 /* ..or it is CARD_BYTES and is the last in the block */
2429 || (page_table[i+1].bytes_used == 0) /* next page free */
2430 || (page_table[i+1].gen != from_space) /* diff. gen */
2431 || (page_table[i+1].region_start_offset == 0))
2435 /* Check that the page is now static. */
2436 gc_assert(page_table[addr_page_index].dont_move != 0);
2439 /* If the given page is not write-protected, then scan it for pointers
2440 * to younger generations or the top temp. generation, if no
2441 * suspicious pointers are found then the page is write-protected.
2443 * Care is taken to check for pointers to the current gc_alloc()
2444 * region if it is a younger generation or the temp. generation. This
2445 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2446 * the gc_alloc_generation does not need to be checked as this is only
2447 * called from scavenge_generation() when the gc_alloc generation is
2448 * younger, so it just checks if there is a pointer to the current
2451 * We return 1 if the page was write-protected, else 0. */
2453 update_page_write_prot(page_index_t page)
2455 generation_index_t gen = page_table[page].gen;
2458 void **page_addr = (void **)page_address(page);
2459 long num_words = page_table[page].bytes_used / N_WORD_BYTES;
2461 /* Shouldn't be a free page. */
2462 gc_assert(page_allocated_p(page));
2463 gc_assert(page_table[page].bytes_used != 0);
2465 /* Skip if it's already write-protected, pinned, or unboxed */
2466 if (page_table[page].write_protected
2467 /* FIXME: What's the reason for not write-protecting pinned pages? */
2468 || page_table[page].dont_move
2469 || page_unboxed_p(page))
2472 /* Scan the page for pointers to younger generations or the
2473 * top temp. generation. */
2475 for (j = 0; j < num_words; j++) {
2476 void *ptr = *(page_addr+j);
2477 page_index_t index = find_page_index(ptr);
2479 /* Check that it's in the dynamic space */
2481 if (/* Does it point to a younger or the temp. generation? */
2482 (page_allocated_p(index)
2483 && (page_table[index].bytes_used != 0)
2484 && ((page_table[index].gen < gen)
2485 || (page_table[index].gen == SCRATCH_GENERATION)))
2487 /* Or does it point within a current gc_alloc() region? */
2488 || ((boxed_region.start_addr <= ptr)
2489 && (ptr <= boxed_region.free_pointer))
2490 || ((unboxed_region.start_addr <= ptr)
2491 && (ptr <= unboxed_region.free_pointer))) {
2498 /* Write-protect the page. */
2499 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2501 os_protect((void *)page_addr,
2503 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2505 /* Note the page as protected in the page tables. */
2506 page_table[page].write_protected = 1;
2512 /* Scavenge all generations from FROM to TO, inclusive, except for
2513 * new_space which needs special handling, as new objects may be
2514 * added which are not checked here - use scavenge_newspace generation.
2516 * Write-protected pages should not have any pointers to the
2517 * from_space so do need scavenging; thus write-protected pages are
2518 * not always scavenged. There is some code to check that these pages
2519 * are not written; but to check fully the write-protected pages need
2520 * to be scavenged by disabling the code to skip them.
2522 * Under the current scheme when a generation is GCed the younger
2523 * generations will be empty. So, when a generation is being GCed it
2524 * is only necessary to scavenge the older generations for pointers
2525 * not the younger. So a page that does not have pointers to younger
2526 * generations does not need to be scavenged.
2528 * The write-protection can be used to note pages that don't have
2529 * pointers to younger pages. But pages can be written without having
2530 * pointers to younger generations. After the pages are scavenged here
2531 * they can be scanned for pointers to younger generations and if
2532 * there are none the page can be write-protected.
2534 * One complication is when the newspace is the top temp. generation.
2536 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2537 * that none were written, which they shouldn't be as they should have
2538 * no pointers to younger generations. This breaks down for weak
2539 * pointers as the objects contain a link to the next and are written
2540 * if a weak pointer is scavenged. Still it's a useful check. */
2542 scavenge_generations(generation_index_t from, generation_index_t to)
2545 page_index_t num_wp = 0;
2549 /* Clear the write_protected_cleared flags on all pages. */
2550 for (i = 0; i < page_table_pages; i++)
2551 page_table[i].write_protected_cleared = 0;
2554 for (i = 0; i < last_free_page; i++) {
2555 generation_index_t generation = page_table[i].gen;
2557 && (page_table[i].bytes_used != 0)
2558 && (generation != new_space)
2559 && (generation >= from)
2560 && (generation <= to)) {
2561 page_index_t last_page,j;
2562 int write_protected=1;
2564 /* This should be the start of a region */
2565 gc_assert(page_table[i].region_start_offset == 0);
2567 /* Now work forward until the end of the region */
2568 for (last_page = i; ; last_page++) {
2570 write_protected && page_table[last_page].write_protected;
2571 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2572 /* Or it is CARD_BYTES and is the last in the block */
2573 || (!page_boxed_p(last_page+1))
2574 || (page_table[last_page+1].bytes_used == 0)
2575 || (page_table[last_page+1].gen != generation)
2576 || (page_table[last_page+1].region_start_offset == 0))
2579 if (!write_protected) {
2580 scavenge(page_address(i),
2581 ((unsigned long)(page_table[last_page].bytes_used
2582 + npage_bytes(last_page-i)))
2585 /* Now scan the pages and write protect those that
2586 * don't have pointers to younger generations. */
2587 if (enable_page_protection) {
2588 for (j = i; j <= last_page; j++) {
2589 num_wp += update_page_write_prot(j);
2592 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2594 "/write protected %d pages within generation %d\n",
2595 num_wp, generation));
2603 /* Check that none of the write_protected pages in this generation
2604 * have been written to. */
2605 for (i = 0; i < page_table_pages; i++) {
2606 if (page_allocated_p(i)
2607 && (page_table[i].bytes_used != 0)
2608 && (page_table[i].gen == generation)
2609 && (page_table[i].write_protected_cleared != 0)) {
2610 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2612 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
2613 page_table[i].bytes_used,
2614 page_table[i].region_start_offset,
2615 page_table[i].dont_move));
2616 lose("write to protected page %d in scavenge_generation()\n", i);
2623 /* Scavenge a newspace generation. As it is scavenged new objects may
2624 * be allocated to it; these will also need to be scavenged. This
2625 * repeats until there are no more objects unscavenged in the
2626 * newspace generation.
2628 * To help improve the efficiency, areas written are recorded by
2629 * gc_alloc() and only these scavenged. Sometimes a little more will be
2630 * scavenged, but this causes no harm. An easy check is done that the
2631 * scavenged bytes equals the number allocated in the previous
2634 * Write-protected pages are not scanned except if they are marked
2635 * dont_move in which case they may have been promoted and still have
2636 * pointers to the from space.
2638 * Write-protected pages could potentially be written by alloc however
2639 * to avoid having to handle re-scavenging of write-protected pages
2640 * gc_alloc() does not write to write-protected pages.
2642 * New areas of objects allocated are recorded alternatively in the two
2643 * new_areas arrays below. */
2644 static struct new_area new_areas_1[NUM_NEW_AREAS];
2645 static struct new_area new_areas_2[NUM_NEW_AREAS];
2647 /* Do one full scan of the new space generation. This is not enough to
2648 * complete the job as new objects may be added to the generation in
2649 * the process which are not scavenged. */
2651 scavenge_newspace_generation_one_scan(generation_index_t generation)
2656 "/starting one full scan of newspace generation %d\n",
2658 for (i = 0; i < last_free_page; i++) {
2659 /* Note that this skips over open regions when it encounters them. */
2661 && (page_table[i].bytes_used != 0)
2662 && (page_table[i].gen == generation)
2663 && ((page_table[i].write_protected == 0)
2664 /* (This may be redundant as write_protected is now
2665 * cleared before promotion.) */
2666 || (page_table[i].dont_move == 1))) {
2667 page_index_t last_page;
2670 /* The scavenge will start at the region_start_offset of
2673 * We need to find the full extent of this contiguous
2674 * block in case objects span pages.
2676 * Now work forward until the end of this contiguous area
2677 * is found. A small area is preferred as there is a
2678 * better chance of its pages being write-protected. */
2679 for (last_page = i; ;last_page++) {
2680 /* If all pages are write-protected and movable,
2681 * then no need to scavenge */
2682 all_wp=all_wp && page_table[last_page].write_protected &&
2683 !page_table[last_page].dont_move;
2685 /* Check whether this is the last page in this
2686 * contiguous block */
2687 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2688 /* Or it is CARD_BYTES and is the last in the block */
2689 || (!page_boxed_p(last_page+1))
2690 || (page_table[last_page+1].bytes_used == 0)
2691 || (page_table[last_page+1].gen != generation)
2692 || (page_table[last_page+1].region_start_offset == 0))
2696 /* Do a limited check for write-protected pages. */
2698 long nwords = (((unsigned long)
2699 (page_table[last_page].bytes_used
2700 + npage_bytes(last_page-i)
2701 + page_table[i].region_start_offset))
2703 new_areas_ignore_page = last_page;
2705 scavenge(page_region_start(i), nwords);
2712 "/done with one full scan of newspace generation %d\n",
2716 /* Do a complete scavenge of the newspace generation. */
2718 scavenge_newspace_generation(generation_index_t generation)
2722 /* the new_areas array currently being written to by gc_alloc() */
2723 struct new_area (*current_new_areas)[] = &new_areas_1;
2724 long current_new_areas_index;
2726 /* the new_areas created by the previous scavenge cycle */
2727 struct new_area (*previous_new_areas)[] = NULL;
2728 long previous_new_areas_index;
2730 /* Flush the current regions updating the tables. */
2731 gc_alloc_update_all_page_tables();
2733 /* Turn on the recording of new areas by gc_alloc(). */
2734 new_areas = current_new_areas;
2735 new_areas_index = 0;
2737 /* Don't need to record new areas that get scavenged anyway during
2738 * scavenge_newspace_generation_one_scan. */
2739 record_new_objects = 1;
2741 /* Start with a full scavenge. */
2742 scavenge_newspace_generation_one_scan(generation);
2744 /* Record all new areas now. */
2745 record_new_objects = 2;
2747 /* Give a chance to weak hash tables to make other objects live.
2748 * FIXME: The algorithm implemented here for weak hash table gcing
2749 * is O(W^2+N) as Bruno Haible warns in
2750 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
2751 * see "Implementation 2". */
2752 scav_weak_hash_tables();
2754 /* Flush the current regions updating the tables. */
2755 gc_alloc_update_all_page_tables();
2757 /* Grab new_areas_index. */
2758 current_new_areas_index = new_areas_index;
2761 "The first scan is finished; current_new_areas_index=%d.\n",
2762 current_new_areas_index));*/
2764 while (current_new_areas_index > 0) {
2765 /* Move the current to the previous new areas */
2766 previous_new_areas = current_new_areas;
2767 previous_new_areas_index = current_new_areas_index;
2769 /* Scavenge all the areas in previous new areas. Any new areas
2770 * allocated are saved in current_new_areas. */
2772 /* Allocate an array for current_new_areas; alternating between
2773 * new_areas_1 and 2 */
2774 if (previous_new_areas == &new_areas_1)
2775 current_new_areas = &new_areas_2;
2777 current_new_areas = &new_areas_1;
2779 /* Set up for gc_alloc(). */
2780 new_areas = current_new_areas;
2781 new_areas_index = 0;
2783 /* Check whether previous_new_areas had overflowed. */
2784 if (previous_new_areas_index >= NUM_NEW_AREAS) {
2786 /* New areas of objects allocated have been lost so need to do a
2787 * full scan to be sure! If this becomes a problem try
2788 * increasing NUM_NEW_AREAS. */
2789 if (gencgc_verbose) {
2790 SHOW("new_areas overflow, doing full scavenge");
2793 /* Don't need to record new areas that get scavenged
2794 * anyway during scavenge_newspace_generation_one_scan. */
2795 record_new_objects = 1;
2797 scavenge_newspace_generation_one_scan(generation);
2799 /* Record all new areas now. */
2800 record_new_objects = 2;
2802 scav_weak_hash_tables();
2804 /* Flush the current regions updating the tables. */
2805 gc_alloc_update_all_page_tables();
2809 /* Work through previous_new_areas. */
2810 for (i = 0; i < previous_new_areas_index; i++) {
2811 page_index_t page = (*previous_new_areas)[i].page;
2812 size_t offset = (*previous_new_areas)[i].offset;
2813 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
2814 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
2815 scavenge(page_address(page)+offset, size);
2818 scav_weak_hash_tables();
2820 /* Flush the current regions updating the tables. */
2821 gc_alloc_update_all_page_tables();
2824 current_new_areas_index = new_areas_index;
2827 "The re-scan has finished; current_new_areas_index=%d.\n",
2828 current_new_areas_index));*/
2831 /* Turn off recording of areas allocated by gc_alloc(). */
2832 record_new_objects = 0;
2837 /* Check that none of the write_protected pages in this generation
2838 * have been written to. */
2839 for (i = 0; i < page_table_pages; i++) {
2840 if (page_allocated_p(i)
2841 && (page_table[i].bytes_used != 0)
2842 && (page_table[i].gen == generation)
2843 && (page_table[i].write_protected_cleared != 0)
2844 && (page_table[i].dont_move == 0)) {
2845 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
2846 i, generation, page_table[i].dont_move);
2853 /* Un-write-protect all the pages in from_space. This is done at the
2854 * start of a GC else there may be many page faults while scavenging
2855 * the newspace (I've seen drive the system time to 99%). These pages
2856 * would need to be unprotected anyway before unmapping in
2857 * free_oldspace; not sure what effect this has on paging.. */
2859 unprotect_oldspace(void)
2862 void *region_addr = 0;
2863 void *page_addr = 0;
2864 unsigned long region_bytes = 0;
2866 for (i = 0; i < last_free_page; i++) {
2867 if (page_allocated_p(i)
2868 && (page_table[i].bytes_used != 0)
2869 && (page_table[i].gen == from_space)) {
2871 /* Remove any write-protection. We should be able to rely
2872 * on the write-protect flag to avoid redundant calls. */
2873 if (page_table[i].write_protected) {
2874 page_table[i].write_protected = 0;
2875 page_addr = page_address(i);
2878 region_addr = page_addr;
2879 region_bytes = GENCGC_CARD_BYTES;
2880 } else if (region_addr + region_bytes == page_addr) {
2881 /* Region continue. */
2882 region_bytes += GENCGC_CARD_BYTES;
2884 /* Unprotect previous region. */
2885 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2886 /* First page in new region. */
2887 region_addr = page_addr;
2888 region_bytes = GENCGC_CARD_BYTES;
2894 /* Unprotect last region. */
2895 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2899 /* Work through all the pages and free any in from_space. This
2900 * assumes that all objects have been copied or promoted to an older
2901 * generation. Bytes_allocated and the generation bytes_allocated
2902 * counter are updated. The number of bytes freed is returned. */
2903 static unsigned long
2906 unsigned long bytes_freed = 0;
2907 page_index_t first_page, last_page;
2912 /* Find a first page for the next region of pages. */
2913 while ((first_page < last_free_page)
2914 && (page_free_p(first_page)
2915 || (page_table[first_page].bytes_used == 0)
2916 || (page_table[first_page].gen != from_space)))
2919 if (first_page >= last_free_page)
2922 /* Find the last page of this region. */
2923 last_page = first_page;
2926 /* Free the page. */
2927 bytes_freed += page_table[last_page].bytes_used;
2928 generations[page_table[last_page].gen].bytes_allocated -=
2929 page_table[last_page].bytes_used;
2930 page_table[last_page].allocated = FREE_PAGE_FLAG;
2931 page_table[last_page].bytes_used = 0;
2932 /* Should already be unprotected by unprotect_oldspace(). */
2933 gc_assert(!page_table[last_page].write_protected);
2936 while ((last_page < last_free_page)
2937 && page_allocated_p(last_page)
2938 && (page_table[last_page].bytes_used != 0)
2939 && (page_table[last_page].gen == from_space));
2941 #ifdef READ_PROTECT_FREE_PAGES
2942 os_protect(page_address(first_page),
2943 npage_bytes(last_page-first_page),
2946 first_page = last_page;
2947 } while (first_page < last_free_page);
2949 bytes_allocated -= bytes_freed;
2954 /* Print some information about a pointer at the given address. */
2956 print_ptr(lispobj *addr)
2958 /* If addr is in the dynamic space then out the page information. */
2959 page_index_t pi1 = find_page_index((void*)addr);
2962 fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
2963 (unsigned long) addr,
2965 page_table[pi1].allocated,
2966 page_table[pi1].gen,
2967 page_table[pi1].bytes_used,
2968 page_table[pi1].region_start_offset,
2969 page_table[pi1].dont_move);
2970 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
2984 is_in_stack_space(lispobj ptr)
2986 /* For space verification: Pointers can be valid if they point
2987 * to a thread stack space. This would be faster if the thread
2988 * structures had page-table entries as if they were part of
2989 * the heap space. */
2991 for_each_thread(th) {
2992 if ((th->control_stack_start <= (lispobj *)ptr) &&
2993 (th->control_stack_end >= (lispobj *)ptr)) {
3001 verify_space(lispobj *start, size_t words)
3003 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
3004 int is_in_readonly_space =
3005 (READ_ONLY_SPACE_START <= (unsigned long)start &&
3006 (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3010 lispobj thing = *(lispobj*)start;
3012 if (is_lisp_pointer(thing)) {
3013 page_index_t page_index = find_page_index((void*)thing);
3014 long to_readonly_space =
3015 (READ_ONLY_SPACE_START <= thing &&
3016 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
3017 long to_static_space =
3018 (STATIC_SPACE_START <= thing &&
3019 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
3021 /* Does it point to the dynamic space? */
3022 if (page_index != -1) {
3023 /* If it's within the dynamic space it should point to a used
3024 * page. XX Could check the offset too. */
3025 if (page_allocated_p(page_index)
3026 && (page_table[page_index].bytes_used == 0))
3027 lose ("Ptr %p @ %p sees free page.\n", thing, start);
3028 /* Check that it doesn't point to a forwarding pointer! */
3029 if (*((lispobj *)native_pointer(thing)) == 0x01) {
3030 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
3032 /* Check that its not in the RO space as it would then be a
3033 * pointer from the RO to the dynamic space. */
3034 if (is_in_readonly_space) {
3035 lose("ptr to dynamic space %p from RO space %x\n",
3038 /* Does it point to a plausible object? This check slows
3039 * it down a lot (so it's commented out).
3041 * "a lot" is serious: it ate 50 minutes cpu time on
3042 * my duron 950 before I came back from lunch and
3045 * FIXME: Add a variable to enable this
3048 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
3049 lose("ptr %p to invalid object %p\n", thing, start);
3053 extern void funcallable_instance_tramp;
3054 /* Verify that it points to another valid space. */
3055 if (!to_readonly_space && !to_static_space
3056 && (thing != (lispobj)&funcallable_instance_tramp)
3057 && !is_in_stack_space(thing)) {
3058 lose("Ptr %p @ %p sees junk.\n", thing, start);
3062 if (!(fixnump(thing))) {
3064 switch(widetag_of(*start)) {
3067 case SIMPLE_VECTOR_WIDETAG:
3069 case COMPLEX_WIDETAG:
3070 case SIMPLE_ARRAY_WIDETAG:
3071 case COMPLEX_BASE_STRING_WIDETAG:
3072 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3073 case COMPLEX_CHARACTER_STRING_WIDETAG:
3075 case COMPLEX_VECTOR_NIL_WIDETAG:
3076 case COMPLEX_BIT_VECTOR_WIDETAG:
3077 case COMPLEX_VECTOR_WIDETAG:
3078 case COMPLEX_ARRAY_WIDETAG:
3079 case CLOSURE_HEADER_WIDETAG:
3080 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3081 case VALUE_CELL_HEADER_WIDETAG:
3082 case SYMBOL_HEADER_WIDETAG:
3083 case CHARACTER_WIDETAG:
3084 #if N_WORD_BITS == 64
3085 case SINGLE_FLOAT_WIDETAG:
3087 case UNBOUND_MARKER_WIDETAG:
3092 case INSTANCE_HEADER_WIDETAG:
3095 long ntotal = HeaderValue(thing);
3096 lispobj layout = ((struct instance *)start)->slots[0];
3101 nuntagged = ((struct layout *)
3102 native_pointer(layout))->n_untagged_slots;
3103 verify_space(start + 1,
3104 ntotal - fixnum_value(nuntagged));
3108 case CODE_HEADER_WIDETAG:
3110 lispobj object = *start;
3112 long nheader_words, ncode_words, nwords;
3114 struct simple_fun *fheaderp;
3116 code = (struct code *) start;
3118 /* Check that it's not in the dynamic space.
3119 * FIXME: Isn't is supposed to be OK for code
3120 * objects to be in the dynamic space these days? */
3121 if (is_in_dynamic_space
3122 /* It's ok if it's byte compiled code. The trace
3123 * table offset will be a fixnum if it's x86
3124 * compiled code - check.
3126 * FIXME: #^#@@! lack of abstraction here..
3127 * This line can probably go away now that
3128 * there's no byte compiler, but I've got
3129 * too much to worry about right now to try
3130 * to make sure. -- WHN 2001-10-06 */
3131 && fixnump(code->trace_table_offset)
3132 /* Only when enabled */
3133 && verify_dynamic_code_check) {
3135 "/code object at %p in the dynamic space\n",
3139 ncode_words = fixnum_value(code->code_size);
3140 nheader_words = HeaderValue(object);
3141 nwords = ncode_words + nheader_words;
3142 nwords = CEILING(nwords, 2);
3143 /* Scavenge the boxed section of the code data block */
3144 verify_space(start + 1, nheader_words - 1);
3146 /* Scavenge the boxed section of each function
3147 * object in the code data block. */
3148 fheaderl = code->entry_points;
3149 while (fheaderl != NIL) {
3151 (struct simple_fun *) native_pointer(fheaderl);
3152 gc_assert(widetag_of(fheaderp->header) ==
3153 SIMPLE_FUN_HEADER_WIDETAG);
3154 verify_space(&fheaderp->name, 1);
3155 verify_space(&fheaderp->arglist, 1);
3156 verify_space(&fheaderp->type, 1);
3157 fheaderl = fheaderp->next;
3163 /* unboxed objects */
3164 case BIGNUM_WIDETAG:
3165 #if N_WORD_BITS != 64
3166 case SINGLE_FLOAT_WIDETAG:
3168 case DOUBLE_FLOAT_WIDETAG:
3169 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3170 case LONG_FLOAT_WIDETAG:
3172 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3173 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3175 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3176 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3178 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3179 case COMPLEX_LONG_FLOAT_WIDETAG:
3181 case SIMPLE_BASE_STRING_WIDETAG:
3182 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3183 case SIMPLE_CHARACTER_STRING_WIDETAG:
3185 case SIMPLE_BIT_VECTOR_WIDETAG:
3186 case SIMPLE_ARRAY_NIL_WIDETAG:
3187 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3188 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3189 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3190 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3191 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3192 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3194 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3196 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3197 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3198 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3199 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3201 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3202 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3204 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3205 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3207 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3208 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3211 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3213 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3214 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3216 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3217 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3219 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3220 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3221 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3222 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3224 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3225 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3227 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3228 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3230 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3231 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3234 case WEAK_POINTER_WIDETAG:
3235 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3236 case NO_TLS_VALUE_MARKER_WIDETAG:
3238 count = (sizetab[widetag_of(*start)])(start);
3242 lose("Unhandled widetag %p at %p\n",
3243 widetag_of(*start), start);
3255 /* FIXME: It would be nice to make names consistent so that
3256 * foo_size meant size *in* *bytes* instead of size in some
3257 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3258 * Some counts of lispobjs are called foo_count; it might be good
3259 * to grep for all foo_size and rename the appropriate ones to
3261 long read_only_space_size =
3262 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3263 - (lispobj*)READ_ONLY_SPACE_START;
3264 long static_space_size =
3265 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3266 - (lispobj*)STATIC_SPACE_START;
3268 for_each_thread(th) {
3269 long binding_stack_size =
3270 (lispobj*)get_binding_stack_pointer(th)
3271 - (lispobj*)th->binding_stack_start;
3272 verify_space(th->binding_stack_start, binding_stack_size);
3274 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3275 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3279 verify_generation(generation_index_t generation)
3283 for (i = 0; i < last_free_page; i++) {
3284 if (page_allocated_p(i)
3285 && (page_table[i].bytes_used != 0)
3286 && (page_table[i].gen == generation)) {
3287 page_index_t last_page;
3288 int region_allocation = page_table[i].allocated;
3290 /* This should be the start of a contiguous block */
3291 gc_assert(page_table[i].region_start_offset == 0);
3293 /* Need to find the full extent of this contiguous block in case
3294 objects span pages. */
3296 /* Now work forward until the end of this contiguous area is
3298 for (last_page = i; ;last_page++)
3299 /* Check whether this is the last page in this contiguous
3301 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3302 /* Or it is CARD_BYTES and is the last in the block */
3303 || (page_table[last_page+1].allocated != region_allocation)
3304 || (page_table[last_page+1].bytes_used == 0)
3305 || (page_table[last_page+1].gen != generation)
3306 || (page_table[last_page+1].region_start_offset == 0))
3309 verify_space(page_address(i),
3311 (page_table[last_page].bytes_used
3312 + npage_bytes(last_page-i)))
3319 /* Check that all the free space is zero filled. */
3321 verify_zero_fill(void)
3325 for (page = 0; page < last_free_page; page++) {
3326 if (page_free_p(page)) {
3327 /* The whole page should be zero filled. */
3328 long *start_addr = (long *)page_address(page);
3331 for (i = 0; i < size; i++) {
3332 if (start_addr[i] != 0) {
3333 lose("free page not zero at %x\n", start_addr + i);
3337 long free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3338 if (free_bytes > 0) {
3339 long *start_addr = (long *)((unsigned long)page_address(page)
3340 + page_table[page].bytes_used);
3341 long size = free_bytes / N_WORD_BYTES;
3343 for (i = 0; i < size; i++) {
3344 if (start_addr[i] != 0) {
3345 lose("free region not zero at %x\n", start_addr + i);
3353 /* External entry point for verify_zero_fill */
3355 gencgc_verify_zero_fill(void)
3357 /* Flush the alloc regions updating the tables. */
3358 gc_alloc_update_all_page_tables();
3359 SHOW("verifying zero fill");
3364 verify_dynamic_space(void)
3366 generation_index_t i;
3368 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3369 verify_generation(i);
3371 if (gencgc_enable_verify_zero_fill)
3375 /* Write-protect all the dynamic boxed pages in the given generation. */
3377 write_protect_generation_pages(generation_index_t generation)
3381 gc_assert(generation < SCRATCH_GENERATION);
3383 for (start = 0; start < last_free_page; start++) {
3384 if (protect_page_p(start, generation)) {
3388 /* Note the page as protected in the page tables. */
3389 page_table[start].write_protected = 1;
3391 for (last = start + 1; last < last_free_page; last++) {
3392 if (!protect_page_p(last, generation))
3394 page_table[last].write_protected = 1;
3397 page_start = (void *)page_address(start);
3399 os_protect(page_start,
3400 npage_bytes(last - start),
3401 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3407 if (gencgc_verbose > 1) {
3409 "/write protected %d of %d pages in generation %d\n",
3410 count_write_protect_generation_pages(generation),
3411 count_generation_pages(generation),
3416 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3418 scavenge_control_stack(struct thread *th)
3420 lispobj *control_stack =
3421 (lispobj *)(th->control_stack_start);
3422 unsigned long control_stack_size =
3423 access_control_stack_pointer(th) - control_stack;
3425 scavenge(control_stack, control_stack_size);
3429 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3431 preserve_context_registers (os_context_t *c)
3434 /* On Darwin the signal context isn't a contiguous block of memory,
3435 * so just preserve_pointering its contents won't be sufficient.
3437 #if defined(LISP_FEATURE_DARWIN)
3438 #if defined LISP_FEATURE_X86
3439 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3440 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3441 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3442 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3443 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3444 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3445 preserve_pointer((void*)*os_context_pc_addr(c));
3446 #elif defined LISP_FEATURE_X86_64
3447 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3448 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3449 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3450 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3451 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3452 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3453 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3454 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3455 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3456 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3457 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3458 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3459 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3460 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3461 preserve_pointer((void*)*os_context_pc_addr(c));
3463 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3466 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3467 preserve_pointer(*ptr);
3472 /* Garbage collect a generation. If raise is 0 then the remains of the
3473 * generation are not raised to the next generation. */
3475 garbage_collect_generation(generation_index_t generation, int raise)
3477 unsigned long bytes_freed;
3479 unsigned long static_space_size;
3482 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3484 /* The oldest generation can't be raised. */
3485 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3487 /* Check if weak hash tables were processed in the previous GC. */
3488 gc_assert(weak_hash_tables == NULL);
3490 /* Initialize the weak pointer list. */
3491 weak_pointers = NULL;
3493 /* When a generation is not being raised it is transported to a
3494 * temporary generation (NUM_GENERATIONS), and lowered when
3495 * done. Set up this new generation. There should be no pages
3496 * allocated to it yet. */
3498 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3501 /* Set the global src and dest. generations */
3502 from_space = generation;
3504 new_space = generation+1;
3506 new_space = SCRATCH_GENERATION;
3508 /* Change to a new space for allocation, resetting the alloc_start_page */
3509 gc_alloc_generation = new_space;
3510 generations[new_space].alloc_start_page = 0;
3511 generations[new_space].alloc_unboxed_start_page = 0;
3512 generations[new_space].alloc_large_start_page = 0;
3513 generations[new_space].alloc_large_unboxed_start_page = 0;
3515 /* Before any pointers are preserved, the dont_move flags on the
3516 * pages need to be cleared. */
3517 for (i = 0; i < last_free_page; i++)
3518 if(page_table[i].gen==from_space)
3519 page_table[i].dont_move = 0;
3521 /* Un-write-protect the old-space pages. This is essential for the
3522 * promoted pages as they may contain pointers into the old-space
3523 * which need to be scavenged. It also helps avoid unnecessary page
3524 * faults as forwarding pointers are written into them. They need to
3525 * be un-protected anyway before unmapping later. */
3526 unprotect_oldspace();
3528 /* Scavenge the stacks' conservative roots. */
3530 /* there are potentially two stacks for each thread: the main
3531 * stack, which may contain Lisp pointers, and the alternate stack.
3532 * We don't ever run Lisp code on the altstack, but it may
3533 * host a sigcontext with lisp objects in it */
3535 /* what we need to do: (1) find the stack pointer for the main
3536 * stack; scavenge it (2) find the interrupt context on the
3537 * alternate stack that might contain lisp values, and scavenge
3540 /* we assume that none of the preceding applies to the thread that
3541 * initiates GC. If you ever call GC from inside an altstack
3542 * handler, you will lose. */
3544 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3545 /* And if we're saving a core, there's no point in being conservative. */
3546 if (conservative_stack) {
3547 for_each_thread(th) {
3549 void **esp=(void **)-1;
3550 #ifdef LISP_FEATURE_SB_THREAD
3552 if(th==arch_os_get_current_thread()) {
3553 /* Somebody is going to burn in hell for this, but casting
3554 * it in two steps shuts gcc up about strict aliasing. */
3555 esp = (void **)((void *)&raise);
3558 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3559 for(i=free-1;i>=0;i--) {
3560 os_context_t *c=th->interrupt_contexts[i];
3561 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3562 if (esp1>=(void **)th->control_stack_start &&
3563 esp1<(void **)th->control_stack_end) {
3564 if(esp1<esp) esp=esp1;
3565 preserve_context_registers(c);
3570 esp = (void **)((void *)&raise);
3572 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3573 preserve_pointer(*ptr);
3578 /* Non-x86oid systems don't have "conservative roots" as such, but
3579 * the same mechanism is used for objects pinned for use by alien
3581 for_each_thread(th) {
3582 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
3583 while (pin_list != NIL) {
3584 struct cons *list_entry =
3585 (struct cons *)native_pointer(pin_list);
3586 preserve_pointer(list_entry->car);
3587 pin_list = list_entry->cdr;
3593 if (gencgc_verbose > 1) {
3594 long num_dont_move_pages = count_dont_move_pages();
3596 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3597 num_dont_move_pages,
3598 npage_bytes(num_dont_move_pages));
3602 /* Scavenge all the rest of the roots. */
3604 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3606 * If not x86, we need to scavenge the interrupt context(s) and the
3611 for_each_thread(th) {
3612 scavenge_interrupt_contexts(th);
3613 scavenge_control_stack(th);
3616 /* Scrub the unscavenged control stack space, so that we can't run
3617 * into any stale pointers in a later GC (this is done by the
3618 * stop-for-gc handler in the other threads). */
3619 scrub_control_stack();
3623 /* Scavenge the Lisp functions of the interrupt handlers, taking
3624 * care to avoid SIG_DFL and SIG_IGN. */
3625 for (i = 0; i < NSIG; i++) {
3626 union interrupt_handler handler = interrupt_handlers[i];
3627 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
3628 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
3629 scavenge((lispobj *)(interrupt_handlers + i), 1);
3632 /* Scavenge the binding stacks. */
3635 for_each_thread(th) {
3636 long len= (lispobj *)get_binding_stack_pointer(th) -
3637 th->binding_stack_start;
3638 scavenge((lispobj *) th->binding_stack_start,len);
3639 #ifdef LISP_FEATURE_SB_THREAD
3640 /* do the tls as well */
3641 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
3642 (sizeof (struct thread))/(sizeof (lispobj));
3643 scavenge((lispobj *) (th+1),len);
3648 /* The original CMU CL code had scavenge-read-only-space code
3649 * controlled by the Lisp-level variable
3650 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
3651 * wasn't documented under what circumstances it was useful or
3652 * safe to turn it on, so it's been turned off in SBCL. If you
3653 * want/need this functionality, and can test and document it,
3654 * please submit a patch. */
3656 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
3657 unsigned long read_only_space_size =
3658 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
3659 (lispobj*)READ_ONLY_SPACE_START;
3661 "/scavenge read only space: %d bytes\n",
3662 read_only_space_size * sizeof(lispobj)));
3663 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
3667 /* Scavenge static space. */
3669 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
3670 (lispobj *)STATIC_SPACE_START;
3671 if (gencgc_verbose > 1) {
3673 "/scavenge static space: %d bytes\n",
3674 static_space_size * sizeof(lispobj)));
3676 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
3678 /* All generations but the generation being GCed need to be
3679 * scavenged. The new_space generation needs special handling as
3680 * objects may be moved in - it is handled separately below. */
3681 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
3683 /* Finally scavenge the new_space generation. Keep going until no
3684 * more objects are moved into the new generation */
3685 scavenge_newspace_generation(new_space);
3687 /* FIXME: I tried reenabling this check when debugging unrelated
3688 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
3689 * Since the current GC code seems to work well, I'm guessing that
3690 * this debugging code is just stale, but I haven't tried to
3691 * figure it out. It should be figured out and then either made to
3692 * work or just deleted. */
3693 #define RESCAN_CHECK 0
3695 /* As a check re-scavenge the newspace once; no new objects should
3698 os_vm_size_t old_bytes_allocated = bytes_allocated;
3699 os_vm_size_t bytes_allocated;
3701 /* Start with a full scavenge. */
3702 scavenge_newspace_generation_one_scan(new_space);
3704 /* Flush the current regions, updating the tables. */
3705 gc_alloc_update_all_page_tables();
3707 bytes_allocated = bytes_allocated - old_bytes_allocated;
3709 if (bytes_allocated != 0) {
3710 lose("Rescan of new_space allocated %d more bytes.\n",
3716 scan_weak_hash_tables();
3717 scan_weak_pointers();
3719 /* Flush the current regions, updating the tables. */
3720 gc_alloc_update_all_page_tables();
3722 /* Free the pages in oldspace, but not those marked dont_move. */
3723 bytes_freed = free_oldspace();
3725 /* If the GC is not raising the age then lower the generation back
3726 * to its normal generation number */
3728 for (i = 0; i < last_free_page; i++)
3729 if ((page_table[i].bytes_used != 0)
3730 && (page_table[i].gen == SCRATCH_GENERATION))
3731 page_table[i].gen = generation;
3732 gc_assert(generations[generation].bytes_allocated == 0);
3733 generations[generation].bytes_allocated =
3734 generations[SCRATCH_GENERATION].bytes_allocated;
3735 generations[SCRATCH_GENERATION].bytes_allocated = 0;
3738 /* Reset the alloc_start_page for generation. */
3739 generations[generation].alloc_start_page = 0;
3740 generations[generation].alloc_unboxed_start_page = 0;
3741 generations[generation].alloc_large_start_page = 0;
3742 generations[generation].alloc_large_unboxed_start_page = 0;
3744 if (generation >= verify_gens) {
3745 if (gencgc_verbose) {
3749 verify_dynamic_space();
3752 /* Set the new gc trigger for the GCed generation. */
3753 generations[generation].gc_trigger =
3754 generations[generation].bytes_allocated
3755 + generations[generation].bytes_consed_between_gc;
3758 generations[generation].num_gc = 0;
3760 ++generations[generation].num_gc;
3764 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
3766 update_dynamic_space_free_pointer(void)
3768 page_index_t last_page = -1, i;
3770 for (i = 0; i < last_free_page; i++)
3771 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
3774 last_free_page = last_page+1;
3776 set_alloc_pointer((lispobj)(page_address(last_free_page)));
3777 return 0; /* dummy value: return something ... */
3781 remap_page_range (page_index_t from, page_index_t to)
3783 /* There's a mysterious Solaris/x86 problem with using mmap
3784 * tricks for memory zeroing. See sbcl-devel thread
3785 * "Re: patch: standalone executable redux".
3787 #if defined(LISP_FEATURE_SUNOS)
3788 zero_and_mark_pages(from, to);
3791 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
3792 release_mask = release_granularity-1,
3794 aligned_from = (from+release_mask)&~release_mask,
3795 aligned_end = (end&~release_mask);
3797 if (aligned_from < aligned_end) {
3798 zero_pages_with_mmap(aligned_from, aligned_end-1);
3799 if (aligned_from != from)
3800 zero_and_mark_pages(from, aligned_from-1);
3801 if (aligned_end != end)
3802 zero_and_mark_pages(aligned_end, end-1);
3804 zero_and_mark_pages(from, to);
3810 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
3812 page_index_t first_page, last_page;
3815 return remap_page_range(from, to);
3817 for (first_page = from; first_page <= to; first_page++) {
3818 if (page_allocated_p(first_page) ||
3819 (page_table[first_page].need_to_zero == 0))
3822 last_page = first_page + 1;
3823 while (page_free_p(last_page) &&
3824 (last_page <= to) &&
3825 (page_table[last_page].need_to_zero == 1))
3828 remap_page_range(first_page, last_page-1);
3830 first_page = last_page;
3834 generation_index_t small_generation_limit = 1;
3836 /* GC all generations newer than last_gen, raising the objects in each
3837 * to the next older generation - we finish when all generations below
3838 * last_gen are empty. Then if last_gen is due for a GC, or if
3839 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
3840 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
3842 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
3843 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
3845 collect_garbage(generation_index_t last_gen)
3847 generation_index_t gen = 0, i;
3850 /* The largest value of last_free_page seen since the time
3851 * remap_free_pages was called. */
3852 static page_index_t high_water_mark = 0;
3854 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
3855 log_generation_stats(gc_logfile, "=== GC Start ===");
3859 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
3861 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
3866 /* Flush the alloc regions updating the tables. */
3867 gc_alloc_update_all_page_tables();
3869 /* Verify the new objects created by Lisp code. */
3870 if (pre_verify_gen_0) {
3871 FSHOW((stderr, "pre-checking generation 0\n"));
3872 verify_generation(0);
3875 if (gencgc_verbose > 1)
3876 print_generation_stats();
3879 /* Collect the generation. */
3881 if (gen >= gencgc_oldest_gen_to_gc) {
3882 /* Never raise the oldest generation. */
3887 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
3890 if (gencgc_verbose > 1) {
3892 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
3895 generations[gen].bytes_allocated,
3896 generations[gen].gc_trigger,
3897 generations[gen].num_gc));
3900 /* If an older generation is being filled, then update its
3903 generations[gen+1].cum_sum_bytes_allocated +=
3904 generations[gen+1].bytes_allocated;
3907 garbage_collect_generation(gen, raise);
3909 /* Reset the memory age cum_sum. */
3910 generations[gen].cum_sum_bytes_allocated = 0;
3912 if (gencgc_verbose > 1) {
3913 FSHOW((stderr, "GC of generation %d finished:\n", gen));
3914 print_generation_stats();
3918 } while ((gen <= gencgc_oldest_gen_to_gc)
3919 && ((gen < last_gen)
3920 || ((gen <= gencgc_oldest_gen_to_gc)
3922 && (generations[gen].bytes_allocated
3923 > generations[gen].gc_trigger)
3924 && (generation_average_age(gen)
3925 > generations[gen].minimum_age_before_gc))));
3927 /* Now if gen-1 was raised all generations before gen are empty.
3928 * If it wasn't raised then all generations before gen-1 are empty.
3930 * Now objects within this gen's pages cannot point to younger
3931 * generations unless they are written to. This can be exploited
3932 * by write-protecting the pages of gen; then when younger
3933 * generations are GCed only the pages which have been written
3938 gen_to_wp = gen - 1;
3940 /* There's not much point in WPing pages in generation 0 as it is
3941 * never scavenged (except promoted pages). */
3942 if ((gen_to_wp > 0) && enable_page_protection) {
3943 /* Check that they are all empty. */
3944 for (i = 0; i < gen_to_wp; i++) {
3945 if (generations[i].bytes_allocated)
3946 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
3949 write_protect_generation_pages(gen_to_wp);
3952 /* Set gc_alloc() back to generation 0. The current regions should
3953 * be flushed after the above GCs. */
3954 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
3955 gc_alloc_generation = 0;
3957 /* Save the high-water mark before updating last_free_page */
3958 if (last_free_page > high_water_mark)
3959 high_water_mark = last_free_page;
3961 update_dynamic_space_free_pointer();
3963 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
3965 fprintf(stderr,"Next gc when %ld bytes have been consed\n",
3968 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
3971 if (gen > small_generation_limit) {
3972 if (last_free_page > high_water_mark)
3973 high_water_mark = last_free_page;
3974 remap_free_pages(0, high_water_mark, 0);
3975 high_water_mark = 0;
3980 log_generation_stats(gc_logfile, "=== GC End ===");
3981 SHOW("returning from collect_garbage");
3984 /* This is called by Lisp PURIFY when it is finished. All live objects
3985 * will have been moved to the RO and Static heaps. The dynamic space
3986 * will need a full re-initialization. We don't bother having Lisp
3987 * PURIFY flush the current gc_alloc() region, as the page_tables are
3988 * re-initialized, and every page is zeroed to be sure. */
3992 page_index_t page, last_page;
3994 if (gencgc_verbose > 1) {
3995 SHOW("entering gc_free_heap");
3998 for (page = 0; page < page_table_pages; page++) {
3999 /* Skip free pages which should already be zero filled. */
4000 if (page_allocated_p(page)) {
4002 for (last_page = page;
4003 (last_page < page_table_pages) && page_allocated_p(last_page);
4005 /* Mark the page free. The other slots are assumed invalid
4006 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4007 * should not be write-protected -- except that the
4008 * generation is used for the current region but it sets
4010 page_table[page].allocated = FREE_PAGE_FLAG;
4011 page_table[page].bytes_used = 0;
4012 page_table[page].write_protected = 0;
4015 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4016 * about this change. */
4017 page_start = (void *)page_address(page);
4018 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
4019 remap_free_pages(page, last_page-1, 1);
4022 } else if (gencgc_zero_check_during_free_heap) {
4023 /* Double-check that the page is zero filled. */
4026 gc_assert(page_free_p(page));
4027 gc_assert(page_table[page].bytes_used == 0);
4028 page_start = (long *)page_address(page);
4029 for (i=0; i<GENCGC_CARD_BYTES/sizeof(long); i++) {
4030 if (page_start[i] != 0) {
4031 lose("free region not zero at %x\n", page_start + i);
4037 bytes_allocated = 0;
4039 /* Initialize the generations. */
4040 for (page = 0; page < NUM_GENERATIONS; page++) {
4041 generations[page].alloc_start_page = 0;
4042 generations[page].alloc_unboxed_start_page = 0;
4043 generations[page].alloc_large_start_page = 0;
4044 generations[page].alloc_large_unboxed_start_page = 0;
4045 generations[page].bytes_allocated = 0;
4046 generations[page].gc_trigger = 2000000;
4047 generations[page].num_gc = 0;
4048 generations[page].cum_sum_bytes_allocated = 0;
4051 if (gencgc_verbose > 1)
4052 print_generation_stats();
4054 /* Initialize gc_alloc(). */
4055 gc_alloc_generation = 0;
4057 gc_set_region_empty(&boxed_region);
4058 gc_set_region_empty(&unboxed_region);
4061 set_alloc_pointer((lispobj)((char *)heap_base));
4063 if (verify_after_free_heap) {
4064 /* Check whether purify has left any bad pointers. */
4065 FSHOW((stderr, "checking after free_heap\n"));
4075 /* Compute the number of pages needed for the dynamic space.
4076 * Dynamic space size should be aligned on page size. */
4077 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4078 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4080 /* Default nursery size to 5% of the total dynamic space size,
4082 bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
4083 if (bytes_consed_between_gcs < (1024*1024))
4084 bytes_consed_between_gcs = 1024*1024;
4086 /* The page_table must be allocated using "calloc" to initialize
4087 * the page structures correctly. There used to be a separate
4088 * initialization loop (now commented out; see below) but that was
4089 * unnecessary and did hurt startup time. */
4090 page_table = calloc(page_table_pages, sizeof(struct page));
4091 gc_assert(page_table);
4094 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4095 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4097 heap_base = (void*)DYNAMIC_SPACE_START;
4099 /* The page structures are initialized implicitly when page_table
4100 * is allocated with "calloc" above. Formerly we had the following
4101 * explicit initialization here (comments converted to C99 style
4102 * for readability as C's block comments don't nest):
4104 * // Initialize each page structure.
4105 * for (i = 0; i < page_table_pages; i++) {
4106 * // Initialize all pages as free.
4107 * page_table[i].allocated = FREE_PAGE_FLAG;
4108 * page_table[i].bytes_used = 0;
4110 * // Pages are not write-protected at startup.
4111 * page_table[i].write_protected = 0;
4114 * Without this loop the image starts up much faster when dynamic
4115 * space is large -- which it is on 64-bit platforms already by
4116 * default -- and when "calloc" for large arrays is implemented
4117 * using copy-on-write of a page of zeroes -- which it is at least
4118 * on Linux. In this case the pages that page_table_pages is stored
4119 * in are mapped and cleared not before the corresponding part of
4120 * dynamic space is used. For example, this saves clearing 16 MB of
4121 * memory at startup if the page size is 4 KB and the size of
4122 * dynamic space is 4 GB.
4123 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4124 * asserted below: */
4126 /* Compile time assertion: If triggered, declares an array
4127 * of dimension -1 forcing a syntax error. The intent of the
4128 * assignment is to avoid an "unused variable" warning. */
4129 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4130 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4133 bytes_allocated = 0;
4135 /* Initialize the generations.
4137 * FIXME: very similar to code in gc_free_heap(), should be shared */
4138 for (i = 0; i < NUM_GENERATIONS; i++) {
4139 generations[i].alloc_start_page = 0;
4140 generations[i].alloc_unboxed_start_page = 0;
4141 generations[i].alloc_large_start_page = 0;
4142 generations[i].alloc_large_unboxed_start_page = 0;
4143 generations[i].bytes_allocated = 0;
4144 generations[i].gc_trigger = 2000000;
4145 generations[i].num_gc = 0;
4146 generations[i].cum_sum_bytes_allocated = 0;
4147 /* the tune-able parameters */
4148 generations[i].bytes_consed_between_gc = bytes_consed_between_gcs;
4149 generations[i].number_of_gcs_before_promotion = 1;
4150 generations[i].minimum_age_before_gc = 0.75;
4153 /* Initialize gc_alloc. */
4154 gc_alloc_generation = 0;
4155 gc_set_region_empty(&boxed_region);
4156 gc_set_region_empty(&unboxed_region);
4161 /* Pick up the dynamic space from after a core load.
4163 * The ALLOCATION_POINTER points to the end of the dynamic space.
4167 gencgc_pickup_dynamic(void)
4169 page_index_t page = 0;
4170 void *alloc_ptr = (void *)get_alloc_pointer();
4171 lispobj *prev=(lispobj *)page_address(page);
4172 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4174 lispobj *first,*ptr= (lispobj *)page_address(page);
4176 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4177 /* It is possible, though rare, for the saved page table
4178 * to contain free pages below alloc_ptr. */
4179 page_table[page].gen = gen;
4180 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4181 page_table[page].large_object = 0;
4182 page_table[page].write_protected = 0;
4183 page_table[page].write_protected_cleared = 0;
4184 page_table[page].dont_move = 0;
4185 page_table[page].need_to_zero = 1;
4188 if (!gencgc_partial_pickup) {
4189 page_table[page].allocated = BOXED_PAGE_FLAG;
4190 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4193 page_table[page].region_start_offset =
4194 page_address(page) - (void *)prev;
4197 } while (page_address(page) < alloc_ptr);
4199 last_free_page = page;
4201 generations[gen].bytes_allocated = npage_bytes(page);
4202 bytes_allocated = npage_bytes(page);
4204 gc_alloc_update_all_page_tables();
4205 write_protect_generation_pages(gen);
4209 gc_initialize_pointers(void)
4211 gencgc_pickup_dynamic();
4215 /* alloc(..) is the external interface for memory allocation. It
4216 * allocates to generation 0. It is not called from within the garbage
4217 * collector as it is only external uses that need the check for heap
4218 * size (GC trigger) and to disable the interrupts (interrupts are
4219 * always disabled during a GC).
4221 * The vops that call alloc(..) assume that the returned space is zero-filled.
4222 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4224 * The check for a GC trigger is only performed when the current
4225 * region is full, so in most cases it's not needed. */
4227 static inline lispobj *
4228 general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region,
4229 struct thread *thread)
4231 #ifndef LISP_FEATURE_WIN32
4232 lispobj alloc_signal;
4235 void *new_free_pointer;
4237 gc_assert(nbytes>0);
4239 /* Check for alignment allocation problems. */
4240 gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
4241 && ((nbytes & LOWTAG_MASK) == 0));
4243 /* Must be inside a PA section. */
4244 gc_assert(get_pseudo_atomic_atomic(thread));
4246 /* maybe we can do this quickly ... */
4247 new_free_pointer = region->free_pointer + nbytes;
4248 if (new_free_pointer <= region->end_addr) {
4249 new_obj = (void*)(region->free_pointer);
4250 region->free_pointer = new_free_pointer;
4251 return(new_obj); /* yup */
4254 /* we have to go the long way around, it seems. Check whether we
4255 * should GC in the near future
4257 if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
4258 /* Don't flood the system with interrupts if the need to gc is
4259 * already noted. This can happen for example when SUB-GC
4260 * allocates or after a gc triggered in a WITHOUT-GCING. */
4261 if (SymbolValue(GC_PENDING,thread) == NIL) {
4262 /* set things up so that GC happens when we finish the PA
4264 SetSymbolValue(GC_PENDING,T,thread);
4265 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4266 set_pseudo_atomic_interrupted(thread);
4267 #ifdef LISP_FEATURE_PPC
4268 /* PPC calls alloc() from a trap or from pa_alloc(),
4269 * look up the most context if it's from a trap. */
4271 os_context_t *context =
4272 thread->interrupt_data->allocation_trap_context;
4273 maybe_save_gc_mask_and_block_deferrables
4274 (context ? os_context_sigmask_addr(context) : NULL);
4277 maybe_save_gc_mask_and_block_deferrables(NULL);
4282 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4284 #ifndef LISP_FEATURE_WIN32
4285 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4286 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4287 if ((signed long) alloc_signal <= 0) {
4288 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4291 SetSymbolValue(ALLOC_SIGNAL,
4292 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4302 general_alloc(long nbytes, int page_type_flag)
4304 struct thread *thread = arch_os_get_current_thread();
4305 /* Select correct region, and call general_alloc_internal with it.
4306 * For other then boxed allocation we must lock first, since the
4307 * region is shared. */
4308 if (BOXED_PAGE_FLAG & page_type_flag) {
4309 #ifdef LISP_FEATURE_SB_THREAD
4310 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4312 struct alloc_region *region = &boxed_region;
4314 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4315 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4317 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4318 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4319 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4322 lose("bad page type flag: %d", page_type_flag);
4329 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4330 return general_alloc(nbytes, BOXED_PAGE_FLAG);
4334 * shared support for the OS-dependent signal handlers which
4335 * catch GENCGC-related write-protect violations
4337 void unhandled_sigmemoryfault(void* addr);
4339 /* Depending on which OS we're running under, different signals might
4340 * be raised for a violation of write protection in the heap. This
4341 * function factors out the common generational GC magic which needs
4342 * to invoked in this case, and should be called from whatever signal
4343 * handler is appropriate for the OS we're running under.
4345 * Return true if this signal is a normal generational GC thing that
4346 * we were able to handle, or false if it was abnormal and control
4347 * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */
4350 gencgc_handle_wp_violation(void* fault_addr)
4352 page_index_t page_index = find_page_index(fault_addr);
4355 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4356 fault_addr, page_index));
4359 /* Check whether the fault is within the dynamic space. */
4360 if (page_index == (-1)) {
4362 /* It can be helpful to be able to put a breakpoint on this
4363 * case to help diagnose low-level problems. */
4364 unhandled_sigmemoryfault(fault_addr);
4366 /* not within the dynamic space -- not our responsibility */
4371 ret = thread_mutex_lock(&free_pages_lock);
4372 gc_assert(ret == 0);
4373 if (page_table[page_index].write_protected) {
4374 /* Unprotect the page. */
4375 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4376 page_table[page_index].write_protected_cleared = 1;
4377 page_table[page_index].write_protected = 0;
4379 /* The only acceptable reason for this signal on a heap
4380 * access is that GENCGC write-protected the page.
4381 * However, if two CPUs hit a wp page near-simultaneously,
4382 * we had better not have the second one lose here if it
4383 * does this test after the first one has already set wp=0
4385 if(page_table[page_index].write_protected_cleared != 1)
4386 lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
4387 page_index, boxed_region.first_page,
4388 boxed_region.last_page);
4390 ret = thread_mutex_unlock(&free_pages_lock);
4391 gc_assert(ret == 0);
4392 /* Don't worry, we can handle it. */
4396 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4397 * it's not just a case of the program hitting the write barrier, and
4398 * are about to let Lisp deal with it. It's basically just a
4399 * convenient place to set a gdb breakpoint. */
4401 unhandled_sigmemoryfault(void *addr)
4404 void gc_alloc_update_all_page_tables(void)
4406 /* Flush the alloc regions updating the tables. */
4409 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4410 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4411 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4415 gc_set_region_empty(struct alloc_region *region)
4417 region->first_page = 0;
4418 region->last_page = -1;
4419 region->start_addr = page_address(0);
4420 region->free_pointer = page_address(0);
4421 region->end_addr = page_address(0);
4425 zero_all_free_pages()
4429 for (i = 0; i < last_free_page; i++) {
4430 if (page_free_p(i)) {
4431 #ifdef READ_PROTECT_FREE_PAGES
4432 os_protect(page_address(i),
4441 /* Things to do before doing a final GC before saving a core (without
4444 * + Pages in large_object pages aren't moved by the GC, so we need to
4445 * unset that flag from all pages.
4446 * + The pseudo-static generation isn't normally collected, but it seems
4447 * reasonable to collect it at least when saving a core. So move the
4448 * pages to a normal generation.
4451 prepare_for_final_gc ()
4454 for (i = 0; i < last_free_page; i++) {
4455 page_table[i].large_object = 0;
4456 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4457 int used = page_table[i].bytes_used;
4458 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4459 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4460 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4466 /* Do a non-conservative GC, and then save a core with the initial
4467 * function being set to the value of the static symbol
4468 * SB!VM:RESTART-LISP-FUNCTION */
4470 gc_and_save(char *filename, boolean prepend_runtime,
4471 boolean save_runtime_options,
4472 boolean compressed, int compression_level)
4475 void *runtime_bytes = NULL;
4476 size_t runtime_size;
4478 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4483 conservative_stack = 0;
4485 /* The filename might come from Lisp, and be moved by the now
4486 * non-conservative GC. */
4487 filename = strdup(filename);
4489 /* Collect twice: once into relatively high memory, and then back
4490 * into low memory. This compacts the retained data into the lower
4491 * pages, minimizing the size of the core file.
4493 prepare_for_final_gc();
4494 gencgc_alloc_start_page = last_free_page;
4495 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4497 prepare_for_final_gc();
4498 gencgc_alloc_start_page = -1;
4499 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4501 if (prepend_runtime)
4502 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4504 /* The dumper doesn't know that pages need to be zeroed before use. */
4505 zero_all_free_pages();
4506 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4507 prepend_runtime, save_runtime_options,
4508 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
4509 /* Oops. Save still managed to fail. Since we've mangled the stack
4510 * beyond hope, there's not much we can do.
4511 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4512 * going to be rather unsatisfactory too... */
4513 lose("Attempt to save core after non-conservative GC failed.\n");