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>.
32 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD)
33 #include "pthreads_win32.h"
41 #include "interrupt.h"
46 #include "gc-internal.h"
48 #include "pseudo-atomic.h"
50 #include "genesis/vector.h"
51 #include "genesis/weak-pointer.h"
52 #include "genesis/fdefn.h"
53 #include "genesis/simple-fun.h"
55 #include "genesis/hash-table.h"
56 #include "genesis/instance.h"
57 #include "genesis/layout.h"
59 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
60 #include "genesis/cons.h"
63 /* forward declarations */
64 page_index_t gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t nbytes,
72 /* Generations 0-5 are normal collected generations, 6 is only used as
73 * scratch space by the collector, and should never get collected.
76 SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
80 /* Should we use page protection to help avoid the scavenging of pages
81 * that don't have pointers to younger generations? */
82 boolean enable_page_protection = 1;
84 /* the minimum size (in bytes) for a large object*/
85 #if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES)
86 os_vm_size_t large_object_size = 4 * GENCGC_ALLOC_GRANULARITY;
87 #elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY)
88 os_vm_size_t large_object_size = 4 * GENCGC_CARD_BYTES;
90 os_vm_size_t large_object_size = 4 * PAGE_BYTES;
93 /* Largest allocation seen since last GC. */
94 os_vm_size_t large_allocation = 0;
101 /* the verbosity level. All non-error messages are disabled at level 0;
102 * and only a few rare messages are printed at level 1. */
104 boolean gencgc_verbose = 1;
106 boolean gencgc_verbose = 0;
109 /* FIXME: At some point enable the various error-checking things below
110 * and see what they say. */
112 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
113 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
115 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
117 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
118 boolean pre_verify_gen_0 = 0;
120 /* Should we check for bad pointers after gc_free_heap is called
121 * from Lisp PURIFY? */
122 boolean verify_after_free_heap = 0;
124 /* Should we print a note when code objects are found in the dynamic space
125 * during a heap verify? */
126 boolean verify_dynamic_code_check = 0;
128 /* Should we check code objects for fixup errors after they are transported? */
129 boolean check_code_fixups = 0;
131 /* Should we check that newly allocated regions are zero filled? */
132 boolean gencgc_zero_check = 0;
134 /* Should we check that the free space is zero filled? */
135 boolean gencgc_enable_verify_zero_fill = 0;
137 /* Should we check that free pages are zero filled during gc_free_heap
138 * called after Lisp PURIFY? */
139 boolean gencgc_zero_check_during_free_heap = 0;
141 /* When loading a core, don't do a full scan of the memory for the
142 * memory region boundaries. (Set to true by coreparse.c if the core
143 * contained a pagetable entry).
145 boolean gencgc_partial_pickup = 0;
147 /* If defined, free pages are read-protected to ensure that nothing
151 /* #define READ_PROTECT_FREE_PAGES */
155 * GC structures and variables
158 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
159 os_vm_size_t bytes_allocated = 0;
160 os_vm_size_t auto_gc_trigger = 0;
162 /* the source and destination generations. These are set before a GC starts
164 generation_index_t from_space;
165 generation_index_t new_space;
167 /* Set to 1 when in GC */
168 boolean gc_active_p = 0;
170 /* should the GC be conservative on stack. If false (only right before
171 * saving a core), don't scan the stack / mark pages dont_move. */
172 static boolean conservative_stack = 1;
174 /* An array of page structures is allocated on gc initialization.
175 * This helps to quickly map between an address and its page structure.
176 * page_table_pages is set from the size of the dynamic space. */
177 page_index_t page_table_pages;
178 struct page *page_table;
180 static inline boolean page_allocated_p(page_index_t page) {
181 return (page_table[page].allocated != FREE_PAGE_FLAG);
184 static inline boolean page_no_region_p(page_index_t page) {
185 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
188 static inline boolean page_allocated_no_region_p(page_index_t page) {
189 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
190 && page_no_region_p(page));
193 static inline boolean page_free_p(page_index_t page) {
194 return (page_table[page].allocated == FREE_PAGE_FLAG);
197 static inline boolean page_boxed_p(page_index_t page) {
198 return (page_table[page].allocated & BOXED_PAGE_FLAG);
201 static inline boolean code_page_p(page_index_t page) {
202 return (page_table[page].allocated & CODE_PAGE_FLAG);
205 static inline boolean page_boxed_no_region_p(page_index_t page) {
206 return page_boxed_p(page) && page_no_region_p(page);
209 static inline boolean page_unboxed_p(page_index_t page) {
210 /* Both flags set == boxed code page */
211 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
212 && !page_boxed_p(page));
215 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
216 return (page_boxed_no_region_p(page)
217 && (page_table[page].bytes_used != 0)
218 && !page_table[page].dont_move
219 && (page_table[page].gen == generation));
222 /* To map addresses to page structures the address of the first page
224 void *heap_base = NULL;
226 /* Calculate the start address for the given page number. */
228 page_address(page_index_t page_num)
230 return (heap_base + (page_num * GENCGC_CARD_BYTES));
233 /* Calculate the address where the allocation region associated with
234 * the page starts. */
236 page_region_start(page_index_t page_index)
238 return page_address(page_index)-page_table[page_index].region_start_offset;
241 /* Find the page index within the page_table for the given
242 * address. Return -1 on failure. */
244 find_page_index(void *addr)
246 if (addr >= heap_base) {
247 page_index_t index = ((pointer_sized_uint_t)addr -
248 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
249 if (index < page_table_pages)
256 npage_bytes(page_index_t npages)
258 gc_assert(npages>=0);
259 return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
262 /* Check that X is a higher address than Y and return offset from Y to
264 static inline os_vm_size_t
265 void_diff(void *x, void *y)
268 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
271 /* a structure to hold the state of a generation
273 * CAUTION: If you modify this, make sure to touch up the alien
274 * definition in src/code/gc.lisp accordingly. ...or better yes,
275 * deal with the FIXME there...
279 /* the first page that gc_alloc() checks on its next call */
280 page_index_t alloc_start_page;
282 /* the first page that gc_alloc_unboxed() checks on its next call */
283 page_index_t alloc_unboxed_start_page;
285 /* the first page that gc_alloc_large (boxed) considers on its next
286 * call. (Although it always allocates after the boxed_region.) */
287 page_index_t alloc_large_start_page;
289 /* the first page that gc_alloc_large (unboxed) considers on its
290 * next call. (Although it always allocates after the
291 * current_unboxed_region.) */
292 page_index_t alloc_large_unboxed_start_page;
294 /* the bytes allocated to this generation */
295 os_vm_size_t bytes_allocated;
297 /* the number of bytes at which to trigger a GC */
298 os_vm_size_t gc_trigger;
300 /* to calculate a new level for gc_trigger */
301 os_vm_size_t bytes_consed_between_gc;
303 /* the number of GCs since the last raise */
306 /* the number of GCs to run on the generations before raising objects to the
308 int number_of_gcs_before_promotion;
310 /* the cumulative sum of the bytes allocated to this generation. It is
311 * cleared after a GC on this generations, and update before new
312 * objects are added from a GC of a younger generation. Dividing by
313 * the bytes_allocated will give the average age of the memory in
314 * this generation since its last GC. */
315 os_vm_size_t cum_sum_bytes_allocated;
317 /* a minimum average memory age before a GC will occur helps
318 * prevent a GC when a large number of new live objects have been
319 * added, in which case a GC could be a waste of time */
320 double minimum_age_before_gc;
323 /* an array of generation structures. There needs to be one more
324 * generation structure than actual generations as the oldest
325 * generation is temporarily raised then lowered. */
326 struct generation generations[NUM_GENERATIONS];
328 /* the oldest generation that is will currently be GCed by default.
329 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
331 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
333 * Setting this to 0 effectively disables the generational nature of
334 * the GC. In some applications generational GC may not be useful
335 * because there are no long-lived objects.
337 * An intermediate value could be handy after moving long-lived data
338 * into an older generation so an unnecessary GC of this long-lived
339 * data can be avoided. */
340 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
342 /* The maximum free page in the heap is maintained and used to update
343 * ALLOCATION_POINTER which is used by the room function to limit its
344 * search of the heap. XX Gencgc obviously needs to be better
345 * integrated with the Lisp code. */
346 page_index_t last_free_page;
348 #ifdef LISP_FEATURE_SB_THREAD
349 /* This lock is to prevent multiple threads from simultaneously
350 * allocating new regions which overlap each other. Note that the
351 * majority of GC is single-threaded, but alloc() may be called from
352 * >1 thread at a time and must be thread-safe. This lock must be
353 * seized before all accesses to generations[] or to parts of
354 * page_table[] that other threads may want to see */
355 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
356 /* This lock is used to protect non-thread-local allocation. */
357 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
360 extern os_vm_size_t gencgc_release_granularity;
361 os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
363 extern os_vm_size_t gencgc_alloc_granularity;
364 os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
368 * miscellaneous heap functions
371 /* Count the number of pages which are write-protected within the
372 * given generation. */
374 count_write_protect_generation_pages(generation_index_t generation)
376 page_index_t i, count = 0;
378 for (i = 0; i < last_free_page; i++)
379 if (page_allocated_p(i)
380 && (page_table[i].gen == generation)
381 && (page_table[i].write_protected == 1))
386 /* Count the number of pages within the given generation. */
388 count_generation_pages(generation_index_t generation)
391 page_index_t count = 0;
393 for (i = 0; i < last_free_page; i++)
394 if (page_allocated_p(i)
395 && (page_table[i].gen == generation))
402 count_dont_move_pages(void)
405 page_index_t count = 0;
406 for (i = 0; i < last_free_page; i++) {
407 if (page_allocated_p(i)
408 && (page_table[i].dont_move != 0)) {
416 /* Work through the pages and add up the number of bytes used for the
417 * given generation. */
419 count_generation_bytes_allocated (generation_index_t gen)
422 os_vm_size_t result = 0;
423 for (i = 0; i < last_free_page; i++) {
424 if (page_allocated_p(i)
425 && (page_table[i].gen == gen))
426 result += page_table[i].bytes_used;
431 /* Return the average age of the memory in a generation. */
433 generation_average_age(generation_index_t gen)
435 if (generations[gen].bytes_allocated == 0)
439 ((double)generations[gen].cum_sum_bytes_allocated)
440 / ((double)generations[gen].bytes_allocated);
444 write_generation_stats(FILE *file)
446 generation_index_t i;
448 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
449 #define FPU_STATE_SIZE 27
450 int fpu_state[FPU_STATE_SIZE];
451 #elif defined(LISP_FEATURE_PPC)
452 #define FPU_STATE_SIZE 32
453 long long fpu_state[FPU_STATE_SIZE];
454 #elif defined(LISP_FEATURE_SPARC)
456 * 32 (single-precision) FP registers, and the FP state register.
457 * But Sparc V9 has 32 double-precision registers (equivalent to 64
458 * single-precision, but can't be accessed), so we leave enough room
461 #define FPU_STATE_SIZE (((32 + 32 + 1) + 1)/2)
462 long long fpu_state[FPU_STATE_SIZE];
465 /* This code uses the FP instructions which may be set up for Lisp
466 * so they need to be saved and reset for C. */
469 /* Print the heap stats. */
471 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
473 for (i = 0; i < SCRATCH_GENERATION; i++) {
475 page_index_t boxed_cnt = 0;
476 page_index_t unboxed_cnt = 0;
477 page_index_t large_boxed_cnt = 0;
478 page_index_t large_unboxed_cnt = 0;
479 page_index_t pinned_cnt=0;
481 for (j = 0; j < last_free_page; j++)
482 if (page_table[j].gen == i) {
484 /* Count the number of boxed pages within the given
486 if (page_boxed_p(j)) {
487 if (page_table[j].large_object)
492 if(page_table[j].dont_move) pinned_cnt++;
493 /* Count the number of unboxed pages within the given
495 if (page_unboxed_p(j)) {
496 if (page_table[j].large_object)
503 gc_assert(generations[i].bytes_allocated
504 == count_generation_bytes_allocated(i));
506 " %1d: %5ld %5ld %5ld %5ld",
508 generations[i].alloc_start_page,
509 generations[i].alloc_unboxed_start_page,
510 generations[i].alloc_large_start_page,
511 generations[i].alloc_large_unboxed_start_page);
513 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
514 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
515 boxed_cnt, unboxed_cnt, large_boxed_cnt,
516 large_unboxed_cnt, pinned_cnt);
521 " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
522 generations[i].bytes_allocated,
523 (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
524 generations[i].gc_trigger,
525 count_write_protect_generation_pages(i),
526 generations[i].num_gc,
527 generation_average_age(i));
529 fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
530 fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);
532 fpu_restore(fpu_state);
536 write_heap_exhaustion_report(FILE *file, long available, long requested,
537 struct thread *thread)
540 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
541 gc_active_p ? "garbage collection" : "allocation",
544 write_generation_stats(file);
545 fprintf(file, "GC control variables:\n");
546 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
547 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
548 (SymbolValue(GC_PENDING, thread) == T) ?
549 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
550 "false" : "in progress"));
551 #ifdef LISP_FEATURE_SB_THREAD
552 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
553 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
558 print_generation_stats(void)
560 write_generation_stats(stderr);
563 extern char* gc_logfile;
564 char * gc_logfile = NULL;
567 log_generation_stats(char *logfile, char *header)
570 FILE * log = fopen(logfile, "a");
572 fprintf(log, "%s\n", header);
573 write_generation_stats(log);
576 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
583 report_heap_exhaustion(long available, long requested, struct thread *th)
586 FILE * log = fopen(gc_logfile, "a");
588 write_heap_exhaustion_report(log, available, requested, th);
591 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
595 /* Always to stderr as well. */
596 write_heap_exhaustion_report(stderr, available, requested, th);
600 #if defined(LISP_FEATURE_X86)
601 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
604 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
605 * if zeroing it ourselves, i.e. in practice give the memory back to the
606 * OS. Generally done after a large GC.
608 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
610 void *addr = page_address(start), *new_addr;
611 os_vm_size_t length = npage_bytes(1+end-start);
616 gc_assert(length >= gencgc_release_granularity);
617 gc_assert((length % gencgc_release_granularity) == 0);
619 os_invalidate(addr, length);
620 new_addr = os_validate(addr, length);
621 if (new_addr == NULL || new_addr != addr) {
622 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
626 for (i = start; i <= end; i++) {
627 page_table[i].need_to_zero = 0;
631 /* Zero the pages from START to END (inclusive). Generally done just after
632 * a new region has been allocated.
635 zero_pages(page_index_t start, page_index_t end) {
639 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
640 fast_bzero(page_address(start), npage_bytes(1+end-start));
642 bzero(page_address(start), npage_bytes(1+end-start));
648 zero_and_mark_pages(page_index_t start, page_index_t end) {
651 zero_pages(start, end);
652 for (i = start; i <= end; i++)
653 page_table[i].need_to_zero = 0;
656 /* Zero the pages from START to END (inclusive), except for those
657 * pages that are known to already zeroed. Mark all pages in the
658 * ranges as non-zeroed.
661 zero_dirty_pages(page_index_t start, page_index_t end) {
664 for (i = start; i <= end; i++) {
665 if (!page_table[i].need_to_zero) continue;
666 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
671 for (i = start; i <= end; i++) {
672 page_table[i].need_to_zero = 1;
678 * To support quick and inline allocation, regions of memory can be
679 * allocated and then allocated from with just a free pointer and a
680 * check against an end address.
682 * Since objects can be allocated to spaces with different properties
683 * e.g. boxed/unboxed, generation, ages; there may need to be many
684 * allocation regions.
686 * Each allocation region may start within a partly used page. Many
687 * features of memory use are noted on a page wise basis, e.g. the
688 * generation; so if a region starts within an existing allocated page
689 * it must be consistent with this page.
691 * During the scavenging of the newspace, objects will be transported
692 * into an allocation region, and pointers updated to point to this
693 * allocation region. It is possible that these pointers will be
694 * scavenged again before the allocation region is closed, e.g. due to
695 * trans_list which jumps all over the place to cleanup the list. It
696 * is important to be able to determine properties of all objects
697 * pointed to when scavenging, e.g to detect pointers to the oldspace.
698 * Thus it's important that the allocation regions have the correct
699 * properties set when allocated, and not just set when closed. The
700 * region allocation routines return regions with the specified
701 * properties, and grab all the pages, setting their properties
702 * appropriately, except that the amount used is not known.
704 * These regions are used to support quicker allocation using just a
705 * free pointer. The actual space used by the region is not reflected
706 * in the pages tables until it is closed. It can't be scavenged until
709 * When finished with the region it should be closed, which will
710 * update the page tables for the actual space used returning unused
711 * space. Further it may be noted in the new regions which is
712 * necessary when scavenging the newspace.
714 * Large objects may be allocated directly without an allocation
715 * region, the page tables are updated immediately.
717 * Unboxed objects don't contain pointers to other objects and so
718 * don't need scavenging. Further they can't contain pointers to
719 * younger generations so WP is not needed. By allocating pages to
720 * unboxed objects the whole page never needs scavenging or
721 * write-protecting. */
723 /* We are only using two regions at present. Both are for the current
724 * newspace generation. */
725 struct alloc_region boxed_region;
726 struct alloc_region unboxed_region;
728 /* The generation currently being allocated to. */
729 static generation_index_t gc_alloc_generation;
731 static inline page_index_t
732 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
735 if (UNBOXED_PAGE_FLAG == page_type_flag) {
736 return generations[generation].alloc_large_unboxed_start_page;
737 } else if (BOXED_PAGE_FLAG & page_type_flag) {
738 /* Both code and data. */
739 return generations[generation].alloc_large_start_page;
741 lose("bad page type flag: %d", page_type_flag);
744 if (UNBOXED_PAGE_FLAG == page_type_flag) {
745 return generations[generation].alloc_unboxed_start_page;
746 } else if (BOXED_PAGE_FLAG & page_type_flag) {
747 /* Both code and data. */
748 return generations[generation].alloc_start_page;
750 lose("bad page_type_flag: %d", page_type_flag);
756 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
760 if (UNBOXED_PAGE_FLAG == page_type_flag) {
761 generations[generation].alloc_large_unboxed_start_page = page;
762 } else if (BOXED_PAGE_FLAG & page_type_flag) {
763 /* Both code and data. */
764 generations[generation].alloc_large_start_page = page;
766 lose("bad page type flag: %d", page_type_flag);
769 if (UNBOXED_PAGE_FLAG == page_type_flag) {
770 generations[generation].alloc_unboxed_start_page = page;
771 } else if (BOXED_PAGE_FLAG & page_type_flag) {
772 /* Both code and data. */
773 generations[generation].alloc_start_page = page;
775 lose("bad page type flag: %d", page_type_flag);
780 /* Find a new region with room for at least the given number of bytes.
782 * It starts looking at the current generation's alloc_start_page. So
783 * may pick up from the previous region if there is enough space. This
784 * keeps the allocation contiguous when scavenging the newspace.
786 * The alloc_region should have been closed by a call to
787 * gc_alloc_update_page_tables(), and will thus be in an empty state.
789 * To assist the scavenging functions write-protected pages are not
790 * used. Free pages should not be write-protected.
792 * It is critical to the conservative GC that the start of regions be
793 * known. To help achieve this only small regions are allocated at a
796 * During scavenging, pointers may be found to within the current
797 * region and the page generation must be set so that pointers to the
798 * from space can be recognized. Therefore the generation of pages in
799 * the region are set to gc_alloc_generation. To prevent another
800 * allocation call using the same pages, all the pages in the region
801 * are allocated, although they will initially be empty.
804 gc_alloc_new_region(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
806 page_index_t first_page;
807 page_index_t last_page;
808 os_vm_size_t bytes_found;
814 "/alloc_new_region for %d bytes from gen %d\n",
815 nbytes, gc_alloc_generation));
818 /* Check that the region is in a reset state. */
819 gc_assert((alloc_region->first_page == 0)
820 && (alloc_region->last_page == -1)
821 && (alloc_region->free_pointer == alloc_region->end_addr));
822 ret = thread_mutex_lock(&free_pages_lock);
824 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
825 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
826 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
827 + npage_bytes(last_page-first_page);
829 /* Set up the alloc_region. */
830 alloc_region->first_page = first_page;
831 alloc_region->last_page = last_page;
832 alloc_region->start_addr = page_table[first_page].bytes_used
833 + page_address(first_page);
834 alloc_region->free_pointer = alloc_region->start_addr;
835 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
837 /* Set up the pages. */
839 /* The first page may have already been in use. */
840 if (page_table[first_page].bytes_used == 0) {
841 page_table[first_page].allocated = page_type_flag;
842 page_table[first_page].gen = gc_alloc_generation;
843 page_table[first_page].large_object = 0;
844 page_table[first_page].region_start_offset = 0;
847 gc_assert(page_table[first_page].allocated == page_type_flag);
848 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
850 gc_assert(page_table[first_page].gen == gc_alloc_generation);
851 gc_assert(page_table[first_page].large_object == 0);
853 for (i = first_page+1; i <= last_page; i++) {
854 page_table[i].allocated = page_type_flag;
855 page_table[i].gen = gc_alloc_generation;
856 page_table[i].large_object = 0;
857 /* This may not be necessary for unboxed regions (think it was
859 page_table[i].region_start_offset =
860 void_diff(page_address(i),alloc_region->start_addr);
861 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
863 /* Bump up last_free_page. */
864 if (last_page+1 > last_free_page) {
865 last_free_page = last_page+1;
866 /* do we only want to call this on special occasions? like for
868 set_alloc_pointer((lispobj)page_address(last_free_page));
870 ret = thread_mutex_unlock(&free_pages_lock);
873 #ifdef READ_PROTECT_FREE_PAGES
874 os_protect(page_address(first_page),
875 npage_bytes(1+last_page-first_page),
879 /* If the first page was only partial, don't check whether it's
880 * zeroed (it won't be) and don't zero it (since the parts that
881 * we're interested in are guaranteed to be zeroed).
883 if (page_table[first_page].bytes_used) {
887 zero_dirty_pages(first_page, last_page);
889 /* we can do this after releasing free_pages_lock */
890 if (gencgc_zero_check) {
892 for (p = (word_t *)alloc_region->start_addr;
893 p < (word_t *)alloc_region->end_addr; p++) {
895 lose("The new region is not zero at %p (start=%p, end=%p).\n",
896 p, alloc_region->start_addr, alloc_region->end_addr);
902 /* If the record_new_objects flag is 2 then all new regions created
905 * If it's 1 then then it is only recorded if the first page of the
906 * current region is <= new_areas_ignore_page. This helps avoid
907 * unnecessary recording when doing full scavenge pass.
909 * The new_object structure holds the page, byte offset, and size of
910 * new regions of objects. Each new area is placed in the array of
911 * these structures pointer to by new_areas. new_areas_index holds the
912 * offset into new_areas.
914 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
915 * later code must detect this and handle it, probably by doing a full
916 * scavenge of a generation. */
917 #define NUM_NEW_AREAS 512
918 static int record_new_objects = 0;
919 static page_index_t new_areas_ignore_page;
925 static struct new_area (*new_areas)[];
926 static size_t new_areas_index;
927 size_t max_new_areas;
929 /* Add a new area to new_areas. */
931 add_new_area(page_index_t first_page, size_t offset, size_t size)
933 size_t new_area_start, c;
936 /* Ignore if full. */
937 if (new_areas_index >= NUM_NEW_AREAS)
940 switch (record_new_objects) {
944 if (first_page > new_areas_ignore_page)
953 new_area_start = npage_bytes(first_page) + offset;
955 /* Search backwards for a prior area that this follows from. If
956 found this will save adding a new area. */
957 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
959 npage_bytes((*new_areas)[i].page)
960 + (*new_areas)[i].offset
961 + (*new_areas)[i].size;
963 "/add_new_area S1 %d %d %d %d\n",
964 i, c, new_area_start, area_end));*/
965 if (new_area_start == area_end) {
967 "/adding to [%d] %d %d %d with %d %d %d:\n",
969 (*new_areas)[i].page,
970 (*new_areas)[i].offset,
971 (*new_areas)[i].size,
975 (*new_areas)[i].size += size;
980 (*new_areas)[new_areas_index].page = first_page;
981 (*new_areas)[new_areas_index].offset = offset;
982 (*new_areas)[new_areas_index].size = size;
984 "/new_area %d page %d offset %d size %d\n",
985 new_areas_index, first_page, offset, size));*/
988 /* Note the max new_areas used. */
989 if (new_areas_index > max_new_areas)
990 max_new_areas = new_areas_index;
993 /* Update the tables for the alloc_region. The region may be added to
996 * When done the alloc_region is set up so that the next quick alloc
997 * will fail safely and thus a new region will be allocated. Further
998 * it is safe to try to re-update the page table of this reset
1001 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
1004 page_index_t first_page;
1005 page_index_t next_page;
1006 os_vm_size_t bytes_used;
1007 os_vm_size_t region_size;
1008 os_vm_size_t byte_cnt;
1009 page_bytes_t orig_first_page_bytes_used;
1013 first_page = alloc_region->first_page;
1015 /* Catch an unused alloc_region. */
1016 if ((first_page == 0) && (alloc_region->last_page == -1))
1019 next_page = first_page+1;
1021 ret = thread_mutex_lock(&free_pages_lock);
1022 gc_assert(ret == 0);
1023 if (alloc_region->free_pointer != alloc_region->start_addr) {
1024 /* some bytes were allocated in the region */
1025 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1027 gc_assert(alloc_region->start_addr ==
1028 (page_address(first_page)
1029 + page_table[first_page].bytes_used));
1031 /* All the pages used need to be updated */
1033 /* Update the first page. */
1035 /* If the page was free then set up the gen, and
1036 * region_start_offset. */
1037 if (page_table[first_page].bytes_used == 0)
1038 gc_assert(page_table[first_page].region_start_offset == 0);
1039 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1041 gc_assert(page_table[first_page].allocated & page_type_flag);
1042 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1043 gc_assert(page_table[first_page].large_object == 0);
1047 /* Calculate the number of bytes used in this page. This is not
1048 * always the number of new bytes, unless it was free. */
1050 if ((bytes_used = void_diff(alloc_region->free_pointer,
1051 page_address(first_page)))
1052 >GENCGC_CARD_BYTES) {
1053 bytes_used = GENCGC_CARD_BYTES;
1056 page_table[first_page].bytes_used = bytes_used;
1057 byte_cnt += bytes_used;
1060 /* All the rest of the pages should be free. We need to set
1061 * their region_start_offset pointer to the start of the
1062 * region, and set the bytes_used. */
1064 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1065 gc_assert(page_table[next_page].allocated & page_type_flag);
1066 gc_assert(page_table[next_page].bytes_used == 0);
1067 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1068 gc_assert(page_table[next_page].large_object == 0);
1070 gc_assert(page_table[next_page].region_start_offset ==
1071 void_diff(page_address(next_page),
1072 alloc_region->start_addr));
1074 /* Calculate the number of bytes used in this page. */
1076 if ((bytes_used = void_diff(alloc_region->free_pointer,
1077 page_address(next_page)))>GENCGC_CARD_BYTES) {
1078 bytes_used = GENCGC_CARD_BYTES;
1081 page_table[next_page].bytes_used = bytes_used;
1082 byte_cnt += bytes_used;
1087 region_size = void_diff(alloc_region->free_pointer,
1088 alloc_region->start_addr);
1089 bytes_allocated += region_size;
1090 generations[gc_alloc_generation].bytes_allocated += region_size;
1092 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1094 /* Set the generations alloc restart page to the last page of
1096 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1098 /* Add the region to the new_areas if requested. */
1099 if (BOXED_PAGE_FLAG & page_type_flag)
1100 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1104 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1106 gc_alloc_generation));
1109 /* There are no bytes allocated. Unallocate the first_page if
1110 * there are 0 bytes_used. */
1111 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1112 if (page_table[first_page].bytes_used == 0)
1113 page_table[first_page].allocated = FREE_PAGE_FLAG;
1116 /* Unallocate any unused pages. */
1117 while (next_page <= alloc_region->last_page) {
1118 gc_assert(page_table[next_page].bytes_used == 0);
1119 page_table[next_page].allocated = FREE_PAGE_FLAG;
1122 ret = thread_mutex_unlock(&free_pages_lock);
1123 gc_assert(ret == 0);
1125 /* alloc_region is per-thread, we're ok to do this unlocked */
1126 gc_set_region_empty(alloc_region);
1129 static inline void *gc_quick_alloc(word_t nbytes);
1131 /* Allocate a possibly large object. */
1133 gc_alloc_large(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
1136 page_index_t first_page, next_page, last_page;
1137 page_bytes_t orig_first_page_bytes_used;
1138 os_vm_size_t byte_cnt;
1139 os_vm_size_t bytes_used;
1142 ret = thread_mutex_lock(&free_pages_lock);
1143 gc_assert(ret == 0);
1145 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1146 if (first_page <= alloc_region->last_page) {
1147 first_page = alloc_region->last_page+1;
1150 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1152 gc_assert(first_page > alloc_region->last_page);
1154 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1156 /* Set up the pages. */
1157 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1159 /* If the first page was free then set up the gen, and
1160 * region_start_offset. */
1161 if (page_table[first_page].bytes_used == 0) {
1162 page_table[first_page].allocated = page_type_flag;
1163 page_table[first_page].gen = gc_alloc_generation;
1164 page_table[first_page].region_start_offset = 0;
1165 page_table[first_page].large_object = 1;
1168 gc_assert(page_table[first_page].allocated == page_type_flag);
1169 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1170 gc_assert(page_table[first_page].large_object == 1);
1174 /* Calc. the number of bytes used in this page. This is not
1175 * always the number of new bytes, unless it was free. */
1177 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1178 bytes_used = GENCGC_CARD_BYTES;
1181 page_table[first_page].bytes_used = bytes_used;
1182 byte_cnt += bytes_used;
1184 next_page = first_page+1;
1186 /* All the rest of the pages should be free. We need to set their
1187 * region_start_offset pointer to the start of the region, and set
1188 * the bytes_used. */
1190 gc_assert(page_free_p(next_page));
1191 gc_assert(page_table[next_page].bytes_used == 0);
1192 page_table[next_page].allocated = page_type_flag;
1193 page_table[next_page].gen = gc_alloc_generation;
1194 page_table[next_page].large_object = 1;
1196 page_table[next_page].region_start_offset =
1197 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1199 /* Calculate the number of bytes used in this page. */
1201 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1202 if (bytes_used > GENCGC_CARD_BYTES) {
1203 bytes_used = GENCGC_CARD_BYTES;
1206 page_table[next_page].bytes_used = bytes_used;
1207 page_table[next_page].write_protected=0;
1208 page_table[next_page].dont_move=0;
1209 byte_cnt += bytes_used;
1213 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1215 bytes_allocated += nbytes;
1216 generations[gc_alloc_generation].bytes_allocated += nbytes;
1218 /* Add the region to the new_areas if requested. */
1219 if (BOXED_PAGE_FLAG & page_type_flag)
1220 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1222 /* Bump up last_free_page */
1223 if (last_page+1 > last_free_page) {
1224 last_free_page = last_page+1;
1225 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1227 ret = thread_mutex_unlock(&free_pages_lock);
1228 gc_assert(ret == 0);
1230 #ifdef READ_PROTECT_FREE_PAGES
1231 os_protect(page_address(first_page),
1232 npage_bytes(1+last_page-first_page),
1236 zero_dirty_pages(first_page, last_page);
1238 return page_address(first_page);
1241 static page_index_t gencgc_alloc_start_page = -1;
1244 gc_heap_exhausted_error_or_lose (sword_t available, sword_t requested)
1246 struct thread *thread = arch_os_get_current_thread();
1247 /* Write basic information before doing anything else: if we don't
1248 * call to lisp this is a must, and even if we do there is always
1249 * the danger that we bounce back here before the error has been
1250 * handled, or indeed even printed.
1252 report_heap_exhaustion(available, requested, thread);
1253 if (gc_active_p || (available == 0)) {
1254 /* If we are in GC, or totally out of memory there is no way
1255 * to sanely transfer control to the lisp-side of things.
1257 lose("Heap exhausted, game over.");
1260 /* FIXME: assert free_pages_lock held */
1261 (void)thread_mutex_unlock(&free_pages_lock);
1262 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
1263 gc_assert(get_pseudo_atomic_atomic(thread));
1264 clear_pseudo_atomic_atomic(thread);
1265 if (get_pseudo_atomic_interrupted(thread))
1266 do_pending_interrupt();
1268 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1269 * to running user code at arbitrary places, even in a
1270 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1271 * running out of the heap. So at this point all bets are
1273 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1274 corruption_warning_and_maybe_lose
1275 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1276 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1277 alloc_number(available), alloc_number(requested));
1278 lose("HEAP-EXHAUSTED-ERROR fell through");
1283 gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t bytes,
1286 page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
1287 page_index_t first_page, last_page, restart_page = *restart_page_ptr;
1288 os_vm_size_t nbytes = bytes;
1289 os_vm_size_t nbytes_goal = nbytes;
1290 os_vm_size_t bytes_found = 0;
1291 os_vm_size_t most_bytes_found = 0;
1292 boolean small_object = nbytes < GENCGC_CARD_BYTES;
1293 /* FIXME: assert(free_pages_lock is held); */
1295 if (nbytes_goal < gencgc_alloc_granularity)
1296 nbytes_goal = gencgc_alloc_granularity;
1298 /* Toggled by gc_and_save for heap compaction, normally -1. */
1299 if (gencgc_alloc_start_page != -1) {
1300 restart_page = gencgc_alloc_start_page;
1303 /* FIXME: This is on bytes instead of nbytes pending cleanup of
1304 * long from the interface. */
1305 gc_assert(bytes>=0);
1306 /* Search for a page with at least nbytes of space. We prefer
1307 * not to split small objects on multiple pages, to reduce the
1308 * number of contiguous allocation regions spaning multiple
1309 * pages: this helps avoid excessive conservativism.
1311 * For other objects, we guarantee that they start on their own
1314 first_page = restart_page;
1315 while (first_page < page_table_pages) {
1317 if (page_free_p(first_page)) {
1318 gc_assert(0 == page_table[first_page].bytes_used);
1319 bytes_found = GENCGC_CARD_BYTES;
1320 } else if (small_object &&
1321 (page_table[first_page].allocated == page_type_flag) &&
1322 (page_table[first_page].large_object == 0) &&
1323 (page_table[first_page].gen == gc_alloc_generation) &&
1324 (page_table[first_page].write_protected == 0) &&
1325 (page_table[first_page].dont_move == 0)) {
1326 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1327 if (bytes_found < nbytes) {
1328 if (bytes_found > most_bytes_found)
1329 most_bytes_found = bytes_found;
1338 gc_assert(page_table[first_page].write_protected == 0);
1339 for (last_page = first_page+1;
1340 ((last_page < page_table_pages) &&
1341 page_free_p(last_page) &&
1342 (bytes_found < nbytes_goal));
1344 bytes_found += GENCGC_CARD_BYTES;
1345 gc_assert(0 == page_table[last_page].bytes_used);
1346 gc_assert(0 == page_table[last_page].write_protected);
1349 if (bytes_found > most_bytes_found) {
1350 most_bytes_found = bytes_found;
1351 most_bytes_found_from = first_page;
1352 most_bytes_found_to = last_page;
1354 if (bytes_found >= nbytes_goal)
1357 first_page = last_page;
1360 bytes_found = most_bytes_found;
1361 restart_page = first_page + 1;
1363 /* Check for a failure */
1364 if (bytes_found < nbytes) {
1365 gc_assert(restart_page >= page_table_pages);
1366 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1369 gc_assert(most_bytes_found_to);
1370 *restart_page_ptr = most_bytes_found_from;
1371 return most_bytes_found_to-1;
1374 /* Allocate bytes. All the rest of the special-purpose allocation
1375 * functions will eventually call this */
1378 gc_alloc_with_region(sword_t nbytes,int page_type_flag, struct alloc_region *my_region,
1381 void *new_free_pointer;
1383 if (nbytes>=large_object_size)
1384 return gc_alloc_large(nbytes, page_type_flag, my_region);
1386 /* Check whether there is room in the current alloc region. */
1387 new_free_pointer = my_region->free_pointer + nbytes;
1389 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1390 my_region->free_pointer, new_free_pointer); */
1392 if (new_free_pointer <= my_region->end_addr) {
1393 /* If so then allocate from the current alloc region. */
1394 void *new_obj = my_region->free_pointer;
1395 my_region->free_pointer = new_free_pointer;
1397 /* Unless a `quick' alloc was requested, check whether the
1398 alloc region is almost empty. */
1400 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1401 /* If so, finished with the current region. */
1402 gc_alloc_update_page_tables(page_type_flag, my_region);
1403 /* Set up a new region. */
1404 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1407 return((void *)new_obj);
1410 /* Else not enough free space in the current region: retry with a
1413 gc_alloc_update_page_tables(page_type_flag, my_region);
1414 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1415 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1418 /* these are only used during GC: all allocation from the mutator calls
1419 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1422 static inline void *
1423 gc_quick_alloc(word_t nbytes)
1425 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1428 static inline void *
1429 gc_alloc_unboxed(word_t nbytes)
1431 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1434 static inline void *
1435 gc_quick_alloc_unboxed(word_t nbytes)
1437 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1440 /* Copy a large object. If the object is in a large object region then
1441 * it is simply promoted, else it is copied. If it's large enough then
1442 * it's copied to a large object region.
1444 * Bignums and vectors may have shrunk. If the object is not copied
1445 * the space needs to be reclaimed, and the page_tables corrected. */
1447 general_copy_large_object(lispobj object, word_t nwords, boolean boxedp)
1451 page_index_t first_page;
1453 gc_assert(is_lisp_pointer(object));
1454 gc_assert(from_space_p(object));
1455 gc_assert((nwords & 0x01) == 0);
1457 if ((nwords > 1024*1024) && gencgc_verbose) {
1458 FSHOW((stderr, "/general_copy_large_object: %d bytes\n",
1459 nwords*N_WORD_BYTES));
1462 /* Check whether it's a large object. */
1463 first_page = find_page_index((void *)object);
1464 gc_assert(first_page >= 0);
1466 if (page_table[first_page].large_object) {
1467 /* Promote the object. Note: Unboxed objects may have been
1468 * allocated to a BOXED region so it may be necessary to
1469 * change the region to UNBOXED. */
1470 os_vm_size_t remaining_bytes;
1471 os_vm_size_t bytes_freed;
1472 page_index_t next_page;
1473 page_bytes_t old_bytes_used;
1475 /* FIXME: This comment is somewhat stale.
1477 * Note: Any page write-protection must be removed, else a
1478 * later scavenge_newspace may incorrectly not scavenge these
1479 * pages. This would not be necessary if they are added to the
1480 * new areas, but let's do it for them all (they'll probably
1481 * be written anyway?). */
1483 gc_assert(page_table[first_page].region_start_offset == 0);
1484 next_page = first_page;
1485 remaining_bytes = nwords*N_WORD_BYTES;
1487 while (remaining_bytes > GENCGC_CARD_BYTES) {
1488 gc_assert(page_table[next_page].gen == from_space);
1489 gc_assert(page_table[next_page].large_object);
1490 gc_assert(page_table[next_page].region_start_offset ==
1491 npage_bytes(next_page-first_page));
1492 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1493 /* Should have been unprotected by unprotect_oldspace()
1494 * for boxed objects, and after promotion unboxed ones
1495 * should not be on protected pages at all. */
1496 gc_assert(!page_table[next_page].write_protected);
1499 gc_assert(page_boxed_p(next_page));
1501 gc_assert(page_allocated_no_region_p(next_page));
1502 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1504 page_table[next_page].gen = new_space;
1506 remaining_bytes -= GENCGC_CARD_BYTES;
1510 /* Now only one page remains, but the object may have shrunk so
1511 * there may be more unused pages which will be freed. */
1513 /* Object may have shrunk but shouldn't have grown - check. */
1514 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1516 page_table[next_page].gen = new_space;
1519 gc_assert(page_boxed_p(next_page));
1521 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1523 /* Adjust the bytes_used. */
1524 old_bytes_used = page_table[next_page].bytes_used;
1525 page_table[next_page].bytes_used = remaining_bytes;
1527 bytes_freed = old_bytes_used - remaining_bytes;
1529 /* Free any remaining pages; needs care. */
1531 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1532 (page_table[next_page].gen == from_space) &&
1533 /* FIXME: It is not obvious to me why this is necessary
1534 * as a loop condition: it seems to me that the
1535 * region_start_offset test should be sufficient, but
1536 * experimentally that is not the case. --NS
1539 page_boxed_p(next_page) :
1540 page_allocated_no_region_p(next_page)) &&
1541 page_table[next_page].large_object &&
1542 (page_table[next_page].region_start_offset ==
1543 npage_bytes(next_page - first_page))) {
1544 /* Checks out OK, free the page. Don't need to both zeroing
1545 * pages as this should have been done before shrinking the
1546 * object. These pages shouldn't be write-protected, even if
1547 * boxed they should be zero filled. */
1548 gc_assert(page_table[next_page].write_protected == 0);
1550 old_bytes_used = page_table[next_page].bytes_used;
1551 page_table[next_page].allocated = FREE_PAGE_FLAG;
1552 page_table[next_page].bytes_used = 0;
1553 bytes_freed += old_bytes_used;
1557 if ((bytes_freed > 0) && gencgc_verbose) {
1559 "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\n",
1563 generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES
1565 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1566 bytes_allocated -= bytes_freed;
1568 /* Add the region to the new_areas if requested. */
1570 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1575 /* Get tag of object. */
1576 tag = lowtag_of(object);
1578 /* Allocate space. */
1579 new = gc_general_alloc(nwords*N_WORD_BYTES,
1580 (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG),
1583 /* Copy the object. */
1584 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1586 /* Return Lisp pointer of new object. */
1587 return ((lispobj) new) | tag;
1592 copy_large_object(lispobj object, sword_t nwords)
1594 return general_copy_large_object(object, nwords, 1);
1598 copy_large_unboxed_object(lispobj object, sword_t nwords)
1600 return general_copy_large_object(object, nwords, 0);
1603 /* to copy unboxed objects */
1605 copy_unboxed_object(lispobj object, sword_t nwords)
1607 return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG);
1612 * code and code-related objects
1615 static lispobj trans_fun_header(lispobj object);
1616 static lispobj trans_boxed(lispobj object);
1619 /* Scan a x86 compiled code object, looking for possible fixups that
1620 * have been missed after a move.
1622 * Two types of fixups are needed:
1623 * 1. Absolute fixups to within the code object.
1624 * 2. Relative fixups to outside the code object.
1626 * Currently only absolute fixups to the constant vector, or to the
1627 * code area are checked. */
1629 sniff_code_object(struct code *code, os_vm_size_t displacement)
1631 #ifdef LISP_FEATURE_X86
1632 sword_t nheader_words, ncode_words, nwords;
1633 os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
1634 os_vm_address_t code_start_addr, code_end_addr;
1635 os_vm_address_t code_addr = (os_vm_address_t)code;
1636 int fixup_found = 0;
1638 if (!check_code_fixups)
1641 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1643 ncode_words = fixnum_value(code->code_size);
1644 nheader_words = HeaderValue(*(lispobj *)code);
1645 nwords = ncode_words + nheader_words;
1647 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1648 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1649 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1650 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1652 /* Work through the unboxed code. */
1653 for (p = code_start_addr; p < code_end_addr; p++) {
1654 void *data = *(void **)p;
1655 unsigned d1 = *((unsigned char *)p - 1);
1656 unsigned d2 = *((unsigned char *)p - 2);
1657 unsigned d3 = *((unsigned char *)p - 3);
1658 unsigned d4 = *((unsigned char *)p - 4);
1660 unsigned d5 = *((unsigned char *)p - 5);
1661 unsigned d6 = *((unsigned char *)p - 6);
1664 /* Check for code references. */
1665 /* Check for a 32 bit word that looks like an absolute
1666 reference to within the code adea of the code object. */
1667 if ((data >= (void*)(code_start_addr-displacement))
1668 && (data < (void*)(code_end_addr-displacement))) {
1669 /* function header */
1671 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1673 /* Skip the function header */
1677 /* the case of PUSH imm32 */
1681 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1682 p, d6, d5, d4, d3, d2, d1, data));
1683 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1685 /* the case of MOV [reg-8],imm32 */
1687 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1688 || d2==0x45 || d2==0x46 || d2==0x47)
1692 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1693 p, d6, d5, d4, d3, d2, d1, data));
1694 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1696 /* the case of LEA reg,[disp32] */
1697 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1700 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1701 p, d6, d5, d4, d3, d2, d1, data));
1702 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1706 /* Check for constant references. */
1707 /* Check for a 32 bit word that looks like an absolute
1708 reference to within the constant vector. Constant references
1710 if ((data >= (void*)(constants_start_addr-displacement))
1711 && (data < (void*)(constants_end_addr-displacement))
1712 && (((unsigned)data & 0x3) == 0)) {
1717 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1718 p, d6, d5, d4, d3, d2, d1, data));
1719 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1722 /* the case of MOV m32,EAX */
1726 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1727 p, d6, d5, d4, d3, d2, d1, data));
1728 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1731 /* the case of CMP m32,imm32 */
1732 if ((d1 == 0x3d) && (d2 == 0x81)) {
1735 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1736 p, d6, d5, d4, d3, d2, d1, data));
1738 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1741 /* Check for a mod=00, r/m=101 byte. */
1742 if ((d1 & 0xc7) == 5) {
1747 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1748 p, d6, d5, d4, d3, d2, d1, data));
1749 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1751 /* the case of CMP reg32,m32 */
1755 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1756 p, d6, d5, d4, d3, d2, d1, data));
1757 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1759 /* the case of MOV m32,reg32 */
1763 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1764 p, d6, d5, d4, d3, d2, d1, data));
1765 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1767 /* the case of MOV reg32,m32 */
1771 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1772 p, d6, d5, d4, d3, d2, d1, data));
1773 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1775 /* the case of LEA reg32,m32 */
1779 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1780 p, d6, d5, d4, d3, d2, d1, data));
1781 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1787 /* If anything was found, print some information on the code
1791 "/compiled code object at %x: header words = %d, code words = %d\n",
1792 code, nheader_words, ncode_words));
1794 "/const start = %x, end = %x\n",
1795 constants_start_addr, constants_end_addr));
1797 "/code start = %x, end = %x\n",
1798 code_start_addr, code_end_addr));
1804 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1806 /* x86-64 uses pc-relative addressing instead of this kludge */
1807 #ifndef LISP_FEATURE_X86_64
1808 sword_t nheader_words, ncode_words, nwords;
1809 os_vm_address_t constants_start_addr, constants_end_addr;
1810 os_vm_address_t code_start_addr, code_end_addr;
1811 os_vm_address_t code_addr = (os_vm_address_t)new_code;
1812 os_vm_address_t old_addr = (os_vm_address_t)old_code;
1813 os_vm_size_t displacement = code_addr - old_addr;
1814 lispobj fixups = NIL;
1815 struct vector *fixups_vector;
1817 ncode_words = fixnum_value(new_code->code_size);
1818 nheader_words = HeaderValue(*(lispobj *)new_code);
1819 nwords = ncode_words + nheader_words;
1821 "/compiled code object at %x: header words = %d, code words = %d\n",
1822 new_code, nheader_words, ncode_words)); */
1823 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1824 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1825 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1826 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1829 "/const start = %x, end = %x\n",
1830 constants_start_addr,constants_end_addr));
1832 "/code start = %x; end = %x\n",
1833 code_start_addr,code_end_addr));
1836 /* The first constant should be a pointer to the fixups for this
1837 code objects. Check. */
1838 fixups = new_code->constants[0];
1840 /* It will be 0 or the unbound-marker if there are no fixups (as
1841 * will be the case if the code object has been purified, for
1842 * example) and will be an other pointer if it is valid. */
1843 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1844 !is_lisp_pointer(fixups)) {
1845 /* Check for possible errors. */
1846 if (check_code_fixups)
1847 sniff_code_object(new_code, displacement);
1852 fixups_vector = (struct vector *)native_pointer(fixups);
1854 /* Could be pointing to a forwarding pointer. */
1855 /* FIXME is this always in from_space? if so, could replace this code with
1856 * forwarding_pointer_p/forwarding_pointer_value */
1857 if (is_lisp_pointer(fixups) &&
1858 (find_page_index((void*)fixups_vector) != -1) &&
1859 (fixups_vector->header == 0x01)) {
1860 /* If so, then follow it. */
1861 /*SHOW("following pointer to a forwarding pointer");*/
1863 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1866 /*SHOW("got fixups");*/
1868 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1869 /* Got the fixups for the code block. Now work through the vector,
1870 and apply a fixup at each address. */
1871 sword_t length = fixnum_value(fixups_vector->length);
1873 for (i = 0; i < length; i++) {
1874 long offset = fixups_vector->data[i];
1875 /* Now check the current value of offset. */
1876 os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);
1878 /* If it's within the old_code object then it must be an
1879 * absolute fixup (relative ones are not saved) */
1880 if ((old_value >= old_addr)
1881 && (old_value < (old_addr + nwords*N_WORD_BYTES)))
1882 /* So add the dispacement. */
1883 *(os_vm_address_t *)(code_start_addr + offset) =
1884 old_value + displacement;
1886 /* It is outside the old code object so it must be a
1887 * relative fixup (absolute fixups are not saved). So
1888 * subtract the displacement. */
1889 *(os_vm_address_t *)(code_start_addr + offset) =
1890 old_value - displacement;
1893 /* This used to just print a note to stderr, but a bogus fixup seems to
1894 * indicate real heap corruption, so a hard hailure is in order. */
1895 lose("fixup vector %p has a bad widetag: %d\n",
1896 fixups_vector, widetag_of(fixups_vector->header));
1899 /* Check for possible errors. */
1900 if (check_code_fixups) {
1901 sniff_code_object(new_code,displacement);
1908 trans_boxed_large(lispobj object)
1913 gc_assert(is_lisp_pointer(object));
1915 header = *((lispobj *) native_pointer(object));
1916 length = HeaderValue(header) + 1;
1917 length = CEILING(length, 2);
1919 return copy_large_object(object, length);
1922 /* Doesn't seem to be used, delete it after the grace period. */
1925 trans_unboxed_large(lispobj object)
1930 gc_assert(is_lisp_pointer(object));
1932 header = *((lispobj *) native_pointer(object));
1933 length = HeaderValue(header) + 1;
1934 length = CEILING(length, 2);
1936 return copy_large_unboxed_object(object, length);
1944 /* XX This is a hack adapted from cgc.c. These don't work too
1945 * efficiently with the gencgc as a list of the weak pointers is
1946 * maintained within the objects which causes writes to the pages. A
1947 * limited attempt is made to avoid unnecessary writes, but this needs
1949 #define WEAK_POINTER_NWORDS \
1950 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
1953 scav_weak_pointer(lispobj *where, lispobj object)
1955 /* Since we overwrite the 'next' field, we have to make
1956 * sure not to do so for pointers already in the list.
1957 * Instead of searching the list of weak_pointers each
1958 * time, we ensure that next is always NULL when the weak
1959 * pointer isn't in the list, and not NULL otherwise.
1960 * Since we can't use NULL to denote end of list, we
1961 * use a pointer back to the same weak_pointer.
1963 struct weak_pointer * wp = (struct weak_pointer*)where;
1965 if (NULL == wp->next) {
1966 wp->next = weak_pointers;
1968 if (NULL == wp->next)
1972 /* Do not let GC scavenge the value slot of the weak pointer.
1973 * (That is why it is a weak pointer.) */
1975 return WEAK_POINTER_NWORDS;
1980 search_read_only_space(void *pointer)
1982 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
1983 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
1984 if ((pointer < (void *)start) || (pointer >= (void *)end))
1986 return (gc_search_space(start,
1987 (((lispobj *)pointer)+2)-start,
1988 (lispobj *) pointer));
1992 search_static_space(void *pointer)
1994 lispobj *start = (lispobj *)STATIC_SPACE_START;
1995 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
1996 if ((pointer < (void *)start) || (pointer >= (void *)end))
1998 return (gc_search_space(start,
1999 (((lispobj *)pointer)+2)-start,
2000 (lispobj *) pointer));
2003 /* a faster version for searching the dynamic space. This will work even
2004 * if the object is in a current allocation region. */
2006 search_dynamic_space(void *pointer)
2008 page_index_t page_index = find_page_index(pointer);
2011 /* The address may be invalid, so do some checks. */
2012 if ((page_index == -1) || page_free_p(page_index))
2014 start = (lispobj *)page_region_start(page_index);
2015 return (gc_search_space(start,
2016 (((lispobj *)pointer)+2)-start,
2017 (lispobj *)pointer));
2020 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2022 /* Is there any possibility that pointer is a valid Lisp object
2023 * reference, and/or something else (e.g. subroutine call return
2024 * address) which should prevent us from moving the referred-to thing?
2025 * This is called from preserve_pointers() */
2027 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2029 lispobj *start_addr;
2031 /* Find the object start address. */
2032 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2036 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2039 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2041 /* Adjust large bignum and vector objects. This will adjust the
2042 * allocated region if the size has shrunk, and move unboxed objects
2043 * into unboxed pages. The pages are not promoted here, and the
2044 * promoted region is not added to the new_regions; this is really
2045 * only designed to be called from preserve_pointer(). Shouldn't fail
2046 * if this is missed, just may delay the moving of objects to unboxed
2047 * pages, and the freeing of pages. */
2049 maybe_adjust_large_object(lispobj *where)
2051 page_index_t first_page;
2052 page_index_t next_page;
2055 uword_t remaining_bytes;
2056 uword_t bytes_freed;
2057 uword_t old_bytes_used;
2061 /* Check whether it's a vector or bignum object. */
2062 switch (widetag_of(where[0])) {
2063 case SIMPLE_VECTOR_WIDETAG:
2064 boxed = BOXED_PAGE_FLAG;
2066 case BIGNUM_WIDETAG:
2067 case SIMPLE_BASE_STRING_WIDETAG:
2068 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2069 case SIMPLE_CHARACTER_STRING_WIDETAG:
2071 case SIMPLE_BIT_VECTOR_WIDETAG:
2072 case SIMPLE_ARRAY_NIL_WIDETAG:
2073 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2074 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2075 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2076 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2077 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2078 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2080 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2082 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2083 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2084 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2085 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2087 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2088 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2090 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2091 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2093 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2094 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2097 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2099 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2100 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2102 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2103 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2105 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2106 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2107 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2108 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2110 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2111 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2113 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2114 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2116 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2117 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2119 boxed = UNBOXED_PAGE_FLAG;
2125 /* Find its current size. */
2126 nwords = (sizetab[widetag_of(where[0])])(where);
2128 first_page = find_page_index((void *)where);
2129 gc_assert(first_page >= 0);
2131 /* Note: Any page write-protection must be removed, else a later
2132 * scavenge_newspace may incorrectly not scavenge these pages.
2133 * This would not be necessary if they are added to the new areas,
2134 * but lets do it for them all (they'll probably be written
2137 gc_assert(page_table[first_page].region_start_offset == 0);
2139 next_page = first_page;
2140 remaining_bytes = nwords*N_WORD_BYTES;
2141 while (remaining_bytes > GENCGC_CARD_BYTES) {
2142 gc_assert(page_table[next_page].gen == from_space);
2143 gc_assert(page_allocated_no_region_p(next_page));
2144 gc_assert(page_table[next_page].large_object);
2145 gc_assert(page_table[next_page].region_start_offset ==
2146 npage_bytes(next_page-first_page));
2147 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2149 page_table[next_page].allocated = boxed;
2151 /* Shouldn't be write-protected at this stage. Essential that the
2153 gc_assert(!page_table[next_page].write_protected);
2154 remaining_bytes -= GENCGC_CARD_BYTES;
2158 /* Now only one page remains, but the object may have shrunk so
2159 * there may be more unused pages which will be freed. */
2161 /* Object may have shrunk but shouldn't have grown - check. */
2162 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2164 page_table[next_page].allocated = boxed;
2165 gc_assert(page_table[next_page].allocated ==
2166 page_table[first_page].allocated);
2168 /* Adjust the bytes_used. */
2169 old_bytes_used = page_table[next_page].bytes_used;
2170 page_table[next_page].bytes_used = remaining_bytes;
2172 bytes_freed = old_bytes_used - remaining_bytes;
2174 /* Free any remaining pages; needs care. */
2176 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2177 (page_table[next_page].gen == from_space) &&
2178 page_allocated_no_region_p(next_page) &&
2179 page_table[next_page].large_object &&
2180 (page_table[next_page].region_start_offset ==
2181 npage_bytes(next_page - first_page))) {
2182 /* It checks out OK, free the page. We don't need to both zeroing
2183 * pages as this should have been done before shrinking the
2184 * object. These pages shouldn't be write protected as they
2185 * should be zero filled. */
2186 gc_assert(page_table[next_page].write_protected == 0);
2188 old_bytes_used = page_table[next_page].bytes_used;
2189 page_table[next_page].allocated = FREE_PAGE_FLAG;
2190 page_table[next_page].bytes_used = 0;
2191 bytes_freed += old_bytes_used;
2195 if ((bytes_freed > 0) && gencgc_verbose) {
2197 "/maybe_adjust_large_object() freed %d\n",
2201 generations[from_space].bytes_allocated -= bytes_freed;
2202 bytes_allocated -= bytes_freed;
2207 /* Take a possible pointer to a Lisp object and mark its page in the
2208 * page_table so that it will not be relocated during a GC.
2210 * This involves locating the page it points to, then backing up to
2211 * the start of its region, then marking all pages dont_move from there
2212 * up to the first page that's not full or has a different generation
2214 * It is assumed that all the page static flags have been cleared at
2215 * the start of a GC.
2217 * It is also assumed that the current gc_alloc() region has been
2218 * flushed and the tables updated. */
2221 preserve_pointer(void *addr)
2223 page_index_t addr_page_index = find_page_index(addr);
2224 page_index_t first_page;
2226 unsigned int region_allocation;
2228 /* quick check 1: Address is quite likely to have been invalid. */
2229 if ((addr_page_index == -1)
2230 || page_free_p(addr_page_index)
2231 || (page_table[addr_page_index].bytes_used == 0)
2232 || (page_table[addr_page_index].gen != from_space)
2233 /* Skip if already marked dont_move. */
2234 || (page_table[addr_page_index].dont_move != 0))
2236 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2237 /* (Now that we know that addr_page_index is in range, it's
2238 * safe to index into page_table[] with it.) */
2239 region_allocation = page_table[addr_page_index].allocated;
2241 /* quick check 2: Check the offset within the page.
2244 if (((uword_t)addr & (GENCGC_CARD_BYTES - 1)) >
2245 page_table[addr_page_index].bytes_used)
2248 /* Filter out anything which can't be a pointer to a Lisp object
2249 * (or, as a special case which also requires dont_move, a return
2250 * address referring to something in a CodeObject). This is
2251 * expensive but important, since it vastly reduces the
2252 * probability that random garbage will be bogusly interpreted as
2253 * a pointer which prevents a page from moving.
2255 * This only needs to happen on x86oids, where this is used for
2256 * conservative roots. Non-x86oid systems only ever call this
2257 * function on known-valid lisp objects. */
2258 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2259 if (!(code_page_p(addr_page_index)
2260 || (is_lisp_pointer((lispobj)addr) &&
2261 possibly_valid_dynamic_space_pointer(addr))))
2265 /* Find the beginning of the region. Note that there may be
2266 * objects in the region preceding the one that we were passed a
2267 * pointer to: if this is the case, we will write-protect all the
2268 * previous objects' pages too. */
2271 /* I think this'd work just as well, but without the assertions.
2272 * -dan 2004.01.01 */
2273 first_page = find_page_index(page_region_start(addr_page_index))
2275 first_page = addr_page_index;
2276 while (page_table[first_page].region_start_offset != 0) {
2278 /* Do some checks. */
2279 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2280 gc_assert(page_table[first_page].gen == from_space);
2281 gc_assert(page_table[first_page].allocated == region_allocation);
2285 /* Adjust any large objects before promotion as they won't be
2286 * copied after promotion. */
2287 if (page_table[first_page].large_object) {
2288 maybe_adjust_large_object(page_address(first_page));
2289 /* If a large object has shrunk then addr may now point to a
2290 * free area in which case it's ignored here. Note it gets
2291 * through the valid pointer test above because the tail looks
2293 if (page_free_p(addr_page_index)
2294 || (page_table[addr_page_index].bytes_used == 0)
2295 /* Check the offset within the page. */
2296 || (((uword_t)addr & (GENCGC_CARD_BYTES - 1))
2297 > page_table[addr_page_index].bytes_used)) {
2299 "weird? ignore ptr 0x%x to freed area of large object\n",
2303 /* It may have moved to unboxed pages. */
2304 region_allocation = page_table[first_page].allocated;
2307 /* Now work forward until the end of this contiguous area is found,
2308 * marking all pages as dont_move. */
2309 for (i = first_page; ;i++) {
2310 gc_assert(page_table[i].allocated == region_allocation);
2312 /* Mark the page static. */
2313 page_table[i].dont_move = 1;
2315 /* It is essential that the pages are not write protected as
2316 * they may have pointers into the old-space which need
2317 * scavenging. They shouldn't be write protected at this
2319 gc_assert(!page_table[i].write_protected);
2321 /* Check whether this is the last page in this contiguous block.. */
2322 if ((page_table[i].bytes_used < GENCGC_CARD_BYTES)
2323 /* ..or it is CARD_BYTES and is the last in the block */
2325 || (page_table[i+1].bytes_used == 0) /* next page free */
2326 || (page_table[i+1].gen != from_space) /* diff. gen */
2327 || (page_table[i+1].region_start_offset == 0))
2331 /* Check that the page is now static. */
2332 gc_assert(page_table[addr_page_index].dont_move != 0);
2335 /* If the given page is not write-protected, then scan it for pointers
2336 * to younger generations or the top temp. generation, if no
2337 * suspicious pointers are found then the page is write-protected.
2339 * Care is taken to check for pointers to the current gc_alloc()
2340 * region if it is a younger generation or the temp. generation. This
2341 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2342 * the gc_alloc_generation does not need to be checked as this is only
2343 * called from scavenge_generation() when the gc_alloc generation is
2344 * younger, so it just checks if there is a pointer to the current
2347 * We return 1 if the page was write-protected, else 0. */
2349 update_page_write_prot(page_index_t page)
2351 generation_index_t gen = page_table[page].gen;
2354 void **page_addr = (void **)page_address(page);
2355 sword_t num_words = page_table[page].bytes_used / N_WORD_BYTES;
2357 /* Shouldn't be a free page. */
2358 gc_assert(page_allocated_p(page));
2359 gc_assert(page_table[page].bytes_used != 0);
2361 /* Skip if it's already write-protected, pinned, or unboxed */
2362 if (page_table[page].write_protected
2363 /* FIXME: What's the reason for not write-protecting pinned pages? */
2364 || page_table[page].dont_move
2365 || page_unboxed_p(page))
2368 /* Scan the page for pointers to younger generations or the
2369 * top temp. generation. */
2371 for (j = 0; j < num_words; j++) {
2372 void *ptr = *(page_addr+j);
2373 page_index_t index = find_page_index(ptr);
2375 /* Check that it's in the dynamic space */
2377 if (/* Does it point to a younger or the temp. generation? */
2378 (page_allocated_p(index)
2379 && (page_table[index].bytes_used != 0)
2380 && ((page_table[index].gen < gen)
2381 || (page_table[index].gen == SCRATCH_GENERATION)))
2383 /* Or does it point within a current gc_alloc() region? */
2384 || ((boxed_region.start_addr <= ptr)
2385 && (ptr <= boxed_region.free_pointer))
2386 || ((unboxed_region.start_addr <= ptr)
2387 && (ptr <= unboxed_region.free_pointer))) {
2394 /* Write-protect the page. */
2395 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2397 os_protect((void *)page_addr,
2399 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2401 /* Note the page as protected in the page tables. */
2402 page_table[page].write_protected = 1;
2408 /* Scavenge all generations from FROM to TO, inclusive, except for
2409 * new_space which needs special handling, as new objects may be
2410 * added which are not checked here - use scavenge_newspace generation.
2412 * Write-protected pages should not have any pointers to the
2413 * from_space so do need scavenging; thus write-protected pages are
2414 * not always scavenged. There is some code to check that these pages
2415 * are not written; but to check fully the write-protected pages need
2416 * to be scavenged by disabling the code to skip them.
2418 * Under the current scheme when a generation is GCed the younger
2419 * generations will be empty. So, when a generation is being GCed it
2420 * is only necessary to scavenge the older generations for pointers
2421 * not the younger. So a page that does not have pointers to younger
2422 * generations does not need to be scavenged.
2424 * The write-protection can be used to note pages that don't have
2425 * pointers to younger pages. But pages can be written without having
2426 * pointers to younger generations. After the pages are scavenged here
2427 * they can be scanned for pointers to younger generations and if
2428 * there are none the page can be write-protected.
2430 * One complication is when the newspace is the top temp. generation.
2432 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2433 * that none were written, which they shouldn't be as they should have
2434 * no pointers to younger generations. This breaks down for weak
2435 * pointers as the objects contain a link to the next and are written
2436 * if a weak pointer is scavenged. Still it's a useful check. */
2438 scavenge_generations(generation_index_t from, generation_index_t to)
2441 page_index_t num_wp = 0;
2445 /* Clear the write_protected_cleared flags on all pages. */
2446 for (i = 0; i < page_table_pages; i++)
2447 page_table[i].write_protected_cleared = 0;
2450 for (i = 0; i < last_free_page; i++) {
2451 generation_index_t generation = page_table[i].gen;
2453 && (page_table[i].bytes_used != 0)
2454 && (generation != new_space)
2455 && (generation >= from)
2456 && (generation <= to)) {
2457 page_index_t last_page,j;
2458 int write_protected=1;
2460 /* This should be the start of a region */
2461 gc_assert(page_table[i].region_start_offset == 0);
2463 /* Now work forward until the end of the region */
2464 for (last_page = i; ; last_page++) {
2466 write_protected && page_table[last_page].write_protected;
2467 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2468 /* Or it is CARD_BYTES and is the last in the block */
2469 || (!page_boxed_p(last_page+1))
2470 || (page_table[last_page+1].bytes_used == 0)
2471 || (page_table[last_page+1].gen != generation)
2472 || (page_table[last_page+1].region_start_offset == 0))
2475 if (!write_protected) {
2476 scavenge(page_address(i),
2477 ((uword_t)(page_table[last_page].bytes_used
2478 + npage_bytes(last_page-i)))
2481 /* Now scan the pages and write protect those that
2482 * don't have pointers to younger generations. */
2483 if (enable_page_protection) {
2484 for (j = i; j <= last_page; j++) {
2485 num_wp += update_page_write_prot(j);
2488 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2490 "/write protected %d pages within generation %d\n",
2491 num_wp, generation));
2499 /* Check that none of the write_protected pages in this generation
2500 * have been written to. */
2501 for (i = 0; i < page_table_pages; i++) {
2502 if (page_allocated_p(i)
2503 && (page_table[i].bytes_used != 0)
2504 && (page_table[i].gen == generation)
2505 && (page_table[i].write_protected_cleared != 0)) {
2506 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2508 "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n",
2509 page_table[i].bytes_used,
2510 page_table[i].region_start_offset,
2511 page_table[i].dont_move));
2512 lose("write to protected page %d in scavenge_generation()\n", i);
2519 /* Scavenge a newspace generation. As it is scavenged new objects may
2520 * be allocated to it; these will also need to be scavenged. This
2521 * repeats until there are no more objects unscavenged in the
2522 * newspace generation.
2524 * To help improve the efficiency, areas written are recorded by
2525 * gc_alloc() and only these scavenged. Sometimes a little more will be
2526 * scavenged, but this causes no harm. An easy check is done that the
2527 * scavenged bytes equals the number allocated in the previous
2530 * Write-protected pages are not scanned except if they are marked
2531 * dont_move in which case they may have been promoted and still have
2532 * pointers to the from space.
2534 * Write-protected pages could potentially be written by alloc however
2535 * to avoid having to handle re-scavenging of write-protected pages
2536 * gc_alloc() does not write to write-protected pages.
2538 * New areas of objects allocated are recorded alternatively in the two
2539 * new_areas arrays below. */
2540 static struct new_area new_areas_1[NUM_NEW_AREAS];
2541 static struct new_area new_areas_2[NUM_NEW_AREAS];
2543 /* Do one full scan of the new space generation. This is not enough to
2544 * complete the job as new objects may be added to the generation in
2545 * the process which are not scavenged. */
2547 scavenge_newspace_generation_one_scan(generation_index_t generation)
2552 "/starting one full scan of newspace generation %d\n",
2554 for (i = 0; i < last_free_page; i++) {
2555 /* Note that this skips over open regions when it encounters them. */
2557 && (page_table[i].bytes_used != 0)
2558 && (page_table[i].gen == generation)
2559 && ((page_table[i].write_protected == 0)
2560 /* (This may be redundant as write_protected is now
2561 * cleared before promotion.) */
2562 || (page_table[i].dont_move == 1))) {
2563 page_index_t last_page;
2566 /* The scavenge will start at the region_start_offset of
2569 * We need to find the full extent of this contiguous
2570 * block in case objects span pages.
2572 * Now work forward until the end of this contiguous area
2573 * is found. A small area is preferred as there is a
2574 * better chance of its pages being write-protected. */
2575 for (last_page = i; ;last_page++) {
2576 /* If all pages are write-protected and movable,
2577 * then no need to scavenge */
2578 all_wp=all_wp && page_table[last_page].write_protected &&
2579 !page_table[last_page].dont_move;
2581 /* Check whether this is the last page in this
2582 * contiguous block */
2583 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2584 /* Or it is CARD_BYTES and is the last in the block */
2585 || (!page_boxed_p(last_page+1))
2586 || (page_table[last_page+1].bytes_used == 0)
2587 || (page_table[last_page+1].gen != generation)
2588 || (page_table[last_page+1].region_start_offset == 0))
2592 /* Do a limited check for write-protected pages. */
2594 sword_t nwords = (((uword_t)
2595 (page_table[last_page].bytes_used
2596 + npage_bytes(last_page-i)
2597 + page_table[i].region_start_offset))
2599 new_areas_ignore_page = last_page;
2601 scavenge(page_region_start(i), nwords);
2608 "/done with one full scan of newspace generation %d\n",
2612 /* Do a complete scavenge of the newspace generation. */
2614 scavenge_newspace_generation(generation_index_t generation)
2618 /* the new_areas array currently being written to by gc_alloc() */
2619 struct new_area (*current_new_areas)[] = &new_areas_1;
2620 size_t current_new_areas_index;
2622 /* the new_areas created by the previous scavenge cycle */
2623 struct new_area (*previous_new_areas)[] = NULL;
2624 size_t previous_new_areas_index;
2626 /* Flush the current regions updating the tables. */
2627 gc_alloc_update_all_page_tables();
2629 /* Turn on the recording of new areas by gc_alloc(). */
2630 new_areas = current_new_areas;
2631 new_areas_index = 0;
2633 /* Don't need to record new areas that get scavenged anyway during
2634 * scavenge_newspace_generation_one_scan. */
2635 record_new_objects = 1;
2637 /* Start with a full scavenge. */
2638 scavenge_newspace_generation_one_scan(generation);
2640 /* Record all new areas now. */
2641 record_new_objects = 2;
2643 /* Give a chance to weak hash tables to make other objects live.
2644 * FIXME: The algorithm implemented here for weak hash table gcing
2645 * is O(W^2+N) as Bruno Haible warns in
2646 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
2647 * see "Implementation 2". */
2648 scav_weak_hash_tables();
2650 /* Flush the current regions updating the tables. */
2651 gc_alloc_update_all_page_tables();
2653 /* Grab new_areas_index. */
2654 current_new_areas_index = new_areas_index;
2657 "The first scan is finished; current_new_areas_index=%d.\n",
2658 current_new_areas_index));*/
2660 while (current_new_areas_index > 0) {
2661 /* Move the current to the previous new areas */
2662 previous_new_areas = current_new_areas;
2663 previous_new_areas_index = current_new_areas_index;
2665 /* Scavenge all the areas in previous new areas. Any new areas
2666 * allocated are saved in current_new_areas. */
2668 /* Allocate an array for current_new_areas; alternating between
2669 * new_areas_1 and 2 */
2670 if (previous_new_areas == &new_areas_1)
2671 current_new_areas = &new_areas_2;
2673 current_new_areas = &new_areas_1;
2675 /* Set up for gc_alloc(). */
2676 new_areas = current_new_areas;
2677 new_areas_index = 0;
2679 /* Check whether previous_new_areas had overflowed. */
2680 if (previous_new_areas_index >= NUM_NEW_AREAS) {
2682 /* New areas of objects allocated have been lost so need to do a
2683 * full scan to be sure! If this becomes a problem try
2684 * increasing NUM_NEW_AREAS. */
2685 if (gencgc_verbose) {
2686 SHOW("new_areas overflow, doing full scavenge");
2689 /* Don't need to record new areas that get scavenged
2690 * anyway during scavenge_newspace_generation_one_scan. */
2691 record_new_objects = 1;
2693 scavenge_newspace_generation_one_scan(generation);
2695 /* Record all new areas now. */
2696 record_new_objects = 2;
2698 scav_weak_hash_tables();
2700 /* Flush the current regions updating the tables. */
2701 gc_alloc_update_all_page_tables();
2705 /* Work through previous_new_areas. */
2706 for (i = 0; i < previous_new_areas_index; i++) {
2707 page_index_t page = (*previous_new_areas)[i].page;
2708 size_t offset = (*previous_new_areas)[i].offset;
2709 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
2710 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
2711 scavenge(page_address(page)+offset, size);
2714 scav_weak_hash_tables();
2716 /* Flush the current regions updating the tables. */
2717 gc_alloc_update_all_page_tables();
2720 current_new_areas_index = new_areas_index;
2723 "The re-scan has finished; current_new_areas_index=%d.\n",
2724 current_new_areas_index));*/
2727 /* Turn off recording of areas allocated by gc_alloc(). */
2728 record_new_objects = 0;
2733 /* Check that none of the write_protected pages in this generation
2734 * have been written to. */
2735 for (i = 0; i < page_table_pages; i++) {
2736 if (page_allocated_p(i)
2737 && (page_table[i].bytes_used != 0)
2738 && (page_table[i].gen == generation)
2739 && (page_table[i].write_protected_cleared != 0)
2740 && (page_table[i].dont_move == 0)) {
2741 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
2742 i, generation, page_table[i].dont_move);
2749 /* Un-write-protect all the pages in from_space. This is done at the
2750 * start of a GC else there may be many page faults while scavenging
2751 * the newspace (I've seen drive the system time to 99%). These pages
2752 * would need to be unprotected anyway before unmapping in
2753 * free_oldspace; not sure what effect this has on paging.. */
2755 unprotect_oldspace(void)
2758 void *region_addr = 0;
2759 void *page_addr = 0;
2760 uword_t region_bytes = 0;
2762 for (i = 0; i < last_free_page; i++) {
2763 if (page_allocated_p(i)
2764 && (page_table[i].bytes_used != 0)
2765 && (page_table[i].gen == from_space)) {
2767 /* Remove any write-protection. We should be able to rely
2768 * on the write-protect flag to avoid redundant calls. */
2769 if (page_table[i].write_protected) {
2770 page_table[i].write_protected = 0;
2771 page_addr = page_address(i);
2774 region_addr = page_addr;
2775 region_bytes = GENCGC_CARD_BYTES;
2776 } else if (region_addr + region_bytes == page_addr) {
2777 /* Region continue. */
2778 region_bytes += GENCGC_CARD_BYTES;
2780 /* Unprotect previous region. */
2781 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2782 /* First page in new region. */
2783 region_addr = page_addr;
2784 region_bytes = GENCGC_CARD_BYTES;
2790 /* Unprotect last region. */
2791 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2795 /* Work through all the pages and free any in from_space. This
2796 * assumes that all objects have been copied or promoted to an older
2797 * generation. Bytes_allocated and the generation bytes_allocated
2798 * counter are updated. The number of bytes freed is returned. */
2802 uword_t bytes_freed = 0;
2803 page_index_t first_page, last_page;
2808 /* Find a first page for the next region of pages. */
2809 while ((first_page < last_free_page)
2810 && (page_free_p(first_page)
2811 || (page_table[first_page].bytes_used == 0)
2812 || (page_table[first_page].gen != from_space)))
2815 if (first_page >= last_free_page)
2818 /* Find the last page of this region. */
2819 last_page = first_page;
2822 /* Free the page. */
2823 bytes_freed += page_table[last_page].bytes_used;
2824 generations[page_table[last_page].gen].bytes_allocated -=
2825 page_table[last_page].bytes_used;
2826 page_table[last_page].allocated = FREE_PAGE_FLAG;
2827 page_table[last_page].bytes_used = 0;
2828 /* Should already be unprotected by unprotect_oldspace(). */
2829 gc_assert(!page_table[last_page].write_protected);
2832 while ((last_page < last_free_page)
2833 && page_allocated_p(last_page)
2834 && (page_table[last_page].bytes_used != 0)
2835 && (page_table[last_page].gen == from_space));
2837 #ifdef READ_PROTECT_FREE_PAGES
2838 os_protect(page_address(first_page),
2839 npage_bytes(last_page-first_page),
2842 first_page = last_page;
2843 } while (first_page < last_free_page);
2845 bytes_allocated -= bytes_freed;
2850 /* Print some information about a pointer at the given address. */
2852 print_ptr(lispobj *addr)
2854 /* If addr is in the dynamic space then out the page information. */
2855 page_index_t pi1 = find_page_index((void*)addr);
2858 fprintf(stderr," %p: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
2861 page_table[pi1].allocated,
2862 page_table[pi1].gen,
2863 page_table[pi1].bytes_used,
2864 page_table[pi1].region_start_offset,
2865 page_table[pi1].dont_move);
2866 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
2880 is_in_stack_space(lispobj ptr)
2882 /* For space verification: Pointers can be valid if they point
2883 * to a thread stack space. This would be faster if the thread
2884 * structures had page-table entries as if they were part of
2885 * the heap space. */
2887 for_each_thread(th) {
2888 if ((th->control_stack_start <= (lispobj *)ptr) &&
2889 (th->control_stack_end >= (lispobj *)ptr)) {
2897 verify_space(lispobj *start, size_t words)
2899 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
2900 int is_in_readonly_space =
2901 (READ_ONLY_SPACE_START <= (uword_t)start &&
2902 (uword_t)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2906 lispobj thing = *(lispobj*)start;
2908 if (is_lisp_pointer(thing)) {
2909 page_index_t page_index = find_page_index((void*)thing);
2910 sword_t to_readonly_space =
2911 (READ_ONLY_SPACE_START <= thing &&
2912 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2913 sword_t to_static_space =
2914 (STATIC_SPACE_START <= thing &&
2915 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
2917 /* Does it point to the dynamic space? */
2918 if (page_index != -1) {
2919 /* If it's within the dynamic space it should point to a used
2920 * page. XX Could check the offset too. */
2921 if (page_allocated_p(page_index)
2922 && (page_table[page_index].bytes_used == 0))
2923 lose ("Ptr %p @ %p sees free page.\n", thing, start);
2924 /* Check that it doesn't point to a forwarding pointer! */
2925 if (*((lispobj *)native_pointer(thing)) == 0x01) {
2926 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
2928 /* Check that its not in the RO space as it would then be a
2929 * pointer from the RO to the dynamic space. */
2930 if (is_in_readonly_space) {
2931 lose("ptr to dynamic space %p from RO space %x\n",
2934 /* Does it point to a plausible object? This check slows
2935 * it down a lot (so it's commented out).
2937 * "a lot" is serious: it ate 50 minutes cpu time on
2938 * my duron 950 before I came back from lunch and
2941 * FIXME: Add a variable to enable this
2944 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
2945 lose("ptr %p to invalid object %p\n", thing, start);
2949 extern void funcallable_instance_tramp;
2950 /* Verify that it points to another valid space. */
2951 if (!to_readonly_space && !to_static_space
2952 && (thing != (lispobj)&funcallable_instance_tramp)
2953 && !is_in_stack_space(thing)) {
2954 lose("Ptr %p @ %p sees junk.\n", thing, start);
2958 if (!(fixnump(thing))) {
2960 switch(widetag_of(*start)) {
2963 case SIMPLE_VECTOR_WIDETAG:
2965 case COMPLEX_WIDETAG:
2966 case SIMPLE_ARRAY_WIDETAG:
2967 case COMPLEX_BASE_STRING_WIDETAG:
2968 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2969 case COMPLEX_CHARACTER_STRING_WIDETAG:
2971 case COMPLEX_VECTOR_NIL_WIDETAG:
2972 case COMPLEX_BIT_VECTOR_WIDETAG:
2973 case COMPLEX_VECTOR_WIDETAG:
2974 case COMPLEX_ARRAY_WIDETAG:
2975 case CLOSURE_HEADER_WIDETAG:
2976 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2977 case VALUE_CELL_HEADER_WIDETAG:
2978 case SYMBOL_HEADER_WIDETAG:
2979 case CHARACTER_WIDETAG:
2980 #if N_WORD_BITS == 64
2981 case SINGLE_FLOAT_WIDETAG:
2983 case UNBOUND_MARKER_WIDETAG:
2988 case INSTANCE_HEADER_WIDETAG:
2991 sword_t ntotal = HeaderValue(thing);
2992 lispobj layout = ((struct instance *)start)->slots[0];
2997 nuntagged = ((struct layout *)
2998 native_pointer(layout))->n_untagged_slots;
2999 verify_space(start + 1,
3000 ntotal - fixnum_value(nuntagged));
3004 case CODE_HEADER_WIDETAG:
3006 lispobj object = *start;
3008 sword_t nheader_words, ncode_words, nwords;
3010 struct simple_fun *fheaderp;
3012 code = (struct code *) start;
3014 /* Check that it's not in the dynamic space.
3015 * FIXME: Isn't is supposed to be OK for code
3016 * objects to be in the dynamic space these days? */
3017 if (is_in_dynamic_space
3018 /* It's ok if it's byte compiled code. The trace
3019 * table offset will be a fixnum if it's x86
3020 * compiled code - check.
3022 * FIXME: #^#@@! lack of abstraction here..
3023 * This line can probably go away now that
3024 * there's no byte compiler, but I've got
3025 * too much to worry about right now to try
3026 * to make sure. -- WHN 2001-10-06 */
3027 && fixnump(code->trace_table_offset)
3028 /* Only when enabled */
3029 && verify_dynamic_code_check) {
3031 "/code object at %p in the dynamic space\n",
3035 ncode_words = fixnum_value(code->code_size);
3036 nheader_words = HeaderValue(object);
3037 nwords = ncode_words + nheader_words;
3038 nwords = CEILING(nwords, 2);
3039 /* Scavenge the boxed section of the code data block */
3040 verify_space(start + 1, nheader_words - 1);
3042 /* Scavenge the boxed section of each function
3043 * object in the code data block. */
3044 fheaderl = code->entry_points;
3045 while (fheaderl != NIL) {
3047 (struct simple_fun *) native_pointer(fheaderl);
3048 gc_assert(widetag_of(fheaderp->header) ==
3049 SIMPLE_FUN_HEADER_WIDETAG);
3050 verify_space(&fheaderp->name, 1);
3051 verify_space(&fheaderp->arglist, 1);
3052 verify_space(&fheaderp->type, 1);
3053 fheaderl = fheaderp->next;
3059 /* unboxed objects */
3060 case BIGNUM_WIDETAG:
3061 #if N_WORD_BITS != 64
3062 case SINGLE_FLOAT_WIDETAG:
3064 case DOUBLE_FLOAT_WIDETAG:
3065 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3066 case LONG_FLOAT_WIDETAG:
3068 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3069 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3071 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3072 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3074 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3075 case COMPLEX_LONG_FLOAT_WIDETAG:
3077 case SIMPLE_BASE_STRING_WIDETAG:
3078 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3079 case SIMPLE_CHARACTER_STRING_WIDETAG:
3081 case SIMPLE_BIT_VECTOR_WIDETAG:
3082 case SIMPLE_ARRAY_NIL_WIDETAG:
3083 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3084 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3085 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3086 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3087 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3088 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3090 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3092 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3093 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3094 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3095 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3097 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3098 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3100 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3101 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3103 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3104 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3107 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3109 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3110 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3112 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3113 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3115 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3116 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3117 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3118 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3120 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3121 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3123 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3124 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3126 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3127 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3130 case WEAK_POINTER_WIDETAG:
3131 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3132 case NO_TLS_VALUE_MARKER_WIDETAG:
3134 count = (sizetab[widetag_of(*start)])(start);
3138 lose("Unhandled widetag %p at %p\n",
3139 widetag_of(*start), start);
3151 /* FIXME: It would be nice to make names consistent so that
3152 * foo_size meant size *in* *bytes* instead of size in some
3153 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3154 * Some counts of lispobjs are called foo_count; it might be good
3155 * to grep for all foo_size and rename the appropriate ones to
3157 sword_t read_only_space_size =
3158 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3159 - (lispobj*)READ_ONLY_SPACE_START;
3160 sword_t static_space_size =
3161 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3162 - (lispobj*)STATIC_SPACE_START;
3164 for_each_thread(th) {
3165 sword_t binding_stack_size =
3166 (lispobj*)get_binding_stack_pointer(th)
3167 - (lispobj*)th->binding_stack_start;
3168 verify_space(th->binding_stack_start, binding_stack_size);
3170 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3171 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3175 verify_generation(generation_index_t generation)
3179 for (i = 0; i < last_free_page; i++) {
3180 if (page_allocated_p(i)
3181 && (page_table[i].bytes_used != 0)
3182 && (page_table[i].gen == generation)) {
3183 page_index_t last_page;
3184 int region_allocation = page_table[i].allocated;
3186 /* This should be the start of a contiguous block */
3187 gc_assert(page_table[i].region_start_offset == 0);
3189 /* Need to find the full extent of this contiguous block in case
3190 objects span pages. */
3192 /* Now work forward until the end of this contiguous area is
3194 for (last_page = i; ;last_page++)
3195 /* Check whether this is the last page in this contiguous
3197 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3198 /* Or it is CARD_BYTES and is the last in the block */
3199 || (page_table[last_page+1].allocated != region_allocation)
3200 || (page_table[last_page+1].bytes_used == 0)
3201 || (page_table[last_page+1].gen != generation)
3202 || (page_table[last_page+1].region_start_offset == 0))
3205 verify_space(page_address(i),
3207 (page_table[last_page].bytes_used
3208 + npage_bytes(last_page-i)))
3215 /* Check that all the free space is zero filled. */
3217 verify_zero_fill(void)
3221 for (page = 0; page < last_free_page; page++) {
3222 if (page_free_p(page)) {
3223 /* The whole page should be zero filled. */
3224 sword_t *start_addr = (sword_t *)page_address(page);
3225 sword_t size = 1024;
3227 for (i = 0; i < size; i++) {
3228 if (start_addr[i] != 0) {
3229 lose("free page not zero at %x\n", start_addr + i);
3233 sword_t free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3234 if (free_bytes > 0) {
3235 sword_t *start_addr = (sword_t *)((uword_t)page_address(page)
3236 + page_table[page].bytes_used);
3237 sword_t size = free_bytes / N_WORD_BYTES;
3239 for (i = 0; i < size; i++) {
3240 if (start_addr[i] != 0) {
3241 lose("free region not zero at %x\n", start_addr + i);
3249 /* External entry point for verify_zero_fill */
3251 gencgc_verify_zero_fill(void)
3253 /* Flush the alloc regions updating the tables. */
3254 gc_alloc_update_all_page_tables();
3255 SHOW("verifying zero fill");
3260 verify_dynamic_space(void)
3262 generation_index_t i;
3264 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3265 verify_generation(i);
3267 if (gencgc_enable_verify_zero_fill)
3271 /* Write-protect all the dynamic boxed pages in the given generation. */
3273 write_protect_generation_pages(generation_index_t generation)
3277 gc_assert(generation < SCRATCH_GENERATION);
3279 for (start = 0; start < last_free_page; start++) {
3280 if (protect_page_p(start, generation)) {
3284 /* Note the page as protected in the page tables. */
3285 page_table[start].write_protected = 1;
3287 for (last = start + 1; last < last_free_page; last++) {
3288 if (!protect_page_p(last, generation))
3290 page_table[last].write_protected = 1;
3293 page_start = (void *)page_address(start);
3295 os_protect(page_start,
3296 npage_bytes(last - start),
3297 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3303 if (gencgc_verbose > 1) {
3305 "/write protected %d of %d pages in generation %d\n",
3306 count_write_protect_generation_pages(generation),
3307 count_generation_pages(generation),
3312 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3314 preserve_context_registers (os_context_t *c)
3317 /* On Darwin the signal context isn't a contiguous block of memory,
3318 * so just preserve_pointering its contents won't be sufficient.
3320 #if defined(LISP_FEATURE_DARWIN)||defined(LISP_FEATURE_WIN32)
3321 #if defined LISP_FEATURE_X86
3322 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3323 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3324 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3325 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3326 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3327 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3328 preserve_pointer((void*)*os_context_pc_addr(c));
3329 #elif defined LISP_FEATURE_X86_64
3330 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3331 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3332 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3333 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3334 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3335 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3336 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3337 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3338 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3339 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3340 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3341 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3342 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3343 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3344 preserve_pointer((void*)*os_context_pc_addr(c));
3346 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3349 #if !defined(LISP_FEATURE_WIN32)
3350 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3351 preserve_pointer(*ptr);
3358 move_pinned_pages_to_newspace()
3362 /* scavenge() will evacuate all oldspace pages, but no newspace
3363 * pages. Pinned pages are precisely those pages which must not
3364 * be evacuated, so move them to newspace directly. */
3366 for (i = 0; i < last_free_page; i++) {
3367 if (page_table[i].dont_move &&
3368 /* dont_move is cleared lazily, so validate the space as well. */
3369 page_table[i].gen == from_space) {
3370 page_table[i].gen = new_space;
3371 /* And since we're moving the pages wholesale, also adjust
3372 * the generation allocation counters. */
3373 generations[new_space].bytes_allocated += page_table[i].bytes_used;
3374 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
3379 /* Garbage collect a generation. If raise is 0 then the remains of the
3380 * generation are not raised to the next generation. */
3382 garbage_collect_generation(generation_index_t generation, int raise)
3384 uword_t bytes_freed;
3386 uword_t static_space_size;
3389 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3391 /* The oldest generation can't be raised. */
3392 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3394 /* Check if weak hash tables were processed in the previous GC. */
3395 gc_assert(weak_hash_tables == NULL);
3397 /* Initialize the weak pointer list. */
3398 weak_pointers = NULL;
3400 /* When a generation is not being raised it is transported to a
3401 * temporary generation (NUM_GENERATIONS), and lowered when
3402 * done. Set up this new generation. There should be no pages
3403 * allocated to it yet. */
3405 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3408 /* Set the global src and dest. generations */
3409 from_space = generation;
3411 new_space = generation+1;
3413 new_space = SCRATCH_GENERATION;
3415 /* Change to a new space for allocation, resetting the alloc_start_page */
3416 gc_alloc_generation = new_space;
3417 generations[new_space].alloc_start_page = 0;
3418 generations[new_space].alloc_unboxed_start_page = 0;
3419 generations[new_space].alloc_large_start_page = 0;
3420 generations[new_space].alloc_large_unboxed_start_page = 0;
3422 /* Before any pointers are preserved, the dont_move flags on the
3423 * pages need to be cleared. */
3424 for (i = 0; i < last_free_page; i++)
3425 if(page_table[i].gen==from_space)
3426 page_table[i].dont_move = 0;
3428 /* Un-write-protect the old-space pages. This is essential for the
3429 * promoted pages as they may contain pointers into the old-space
3430 * which need to be scavenged. It also helps avoid unnecessary page
3431 * faults as forwarding pointers are written into them. They need to
3432 * be un-protected anyway before unmapping later. */
3433 unprotect_oldspace();
3435 /* Scavenge the stacks' conservative roots. */
3437 /* there are potentially two stacks for each thread: the main
3438 * stack, which may contain Lisp pointers, and the alternate stack.
3439 * We don't ever run Lisp code on the altstack, but it may
3440 * host a sigcontext with lisp objects in it */
3442 /* what we need to do: (1) find the stack pointer for the main
3443 * stack; scavenge it (2) find the interrupt context on the
3444 * alternate stack that might contain lisp values, and scavenge
3447 /* we assume that none of the preceding applies to the thread that
3448 * initiates GC. If you ever call GC from inside an altstack
3449 * handler, you will lose. */
3451 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3452 /* And if we're saving a core, there's no point in being conservative. */
3453 if (conservative_stack) {
3454 for_each_thread(th) {
3456 void **esp=(void **)-1;
3457 if (th->state == STATE_DEAD)
3459 # if defined(LISP_FEATURE_SB_SAFEPOINT)
3460 /* Conservative collect_garbage is always invoked with a
3461 * foreign C call or an interrupt handler on top of every
3462 * existing thread, so the stored SP in each thread
3463 * structure is valid, no matter which thread we are looking
3464 * at. For threads that were running Lisp code, the pitstop
3465 * and edge functions maintain this value within the
3466 * interrupt or exception handler. */
3467 esp = os_get_csp(th);
3468 assert_on_stack(th, esp);
3470 /* In addition to pointers on the stack, also preserve the
3471 * return PC, the only value from the context that we need
3472 * in addition to the SP. The return PC gets saved by the
3473 * foreign call wrapper, and removed from the control stack
3474 * into a register. */
3475 preserve_pointer(th->pc_around_foreign_call);
3477 /* And on platforms with interrupts: scavenge ctx registers. */
3479 /* Disabled on Windows, because it does not have an explicit
3480 * stack of `interrupt_contexts'. The reported CSP has been
3481 * chosen so that the current context on the stack is
3482 * covered by the stack scan. See also set_csp_from_context(). */
3483 # ifndef LISP_FEATURE_WIN32
3484 if (th != arch_os_get_current_thread()) {
3485 long k = fixnum_value(
3486 SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3488 preserve_context_registers(th->interrupt_contexts[--k]);
3491 # elif defined(LISP_FEATURE_SB_THREAD)
3493 if(th==arch_os_get_current_thread()) {
3494 /* Somebody is going to burn in hell for this, but casting
3495 * it in two steps shuts gcc up about strict aliasing. */
3496 esp = (void **)((void *)&raise);
3499 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3500 for(i=free-1;i>=0;i--) {
3501 os_context_t *c=th->interrupt_contexts[i];
3502 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3503 if (esp1>=(void **)th->control_stack_start &&
3504 esp1<(void **)th->control_stack_end) {
3505 if(esp1<esp) esp=esp1;
3506 preserve_context_registers(c);
3511 esp = (void **)((void *)&raise);
3513 if (!esp || esp == (void*) -1)
3514 lose("garbage_collect: no SP known for thread %x (OS %x)",
3516 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3517 preserve_pointer(*ptr);
3522 /* Non-x86oid systems don't have "conservative roots" as such, but
3523 * the same mechanism is used for objects pinned for use by alien
3525 for_each_thread(th) {
3526 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
3527 while (pin_list != NIL) {
3528 struct cons *list_entry =
3529 (struct cons *)native_pointer(pin_list);
3530 preserve_pointer(list_entry->car);
3531 pin_list = list_entry->cdr;
3537 if (gencgc_verbose > 1) {
3538 sword_t num_dont_move_pages = count_dont_move_pages();
3540 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3541 num_dont_move_pages,
3542 npage_bytes(num_dont_move_pages));
3546 /* Now that all of the pinned (dont_move) pages are known, and
3547 * before we start to scavenge (and thus relocate) objects,
3548 * relocate the pinned pages to newspace, so that the scavenger
3549 * will not attempt to relocate their contents. */
3550 move_pinned_pages_to_newspace();
3552 /* Scavenge all the rest of the roots. */
3554 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3556 * If not x86, we need to scavenge the interrupt context(s) and the
3561 for_each_thread(th) {
3562 scavenge_interrupt_contexts(th);
3563 scavenge_control_stack(th);
3566 # ifdef LISP_FEATURE_SB_SAFEPOINT
3567 /* In this case, scrub all stacks right here from the GCing thread
3568 * instead of doing what the comment below says. Suboptimal, but
3571 scrub_thread_control_stack(th);
3573 /* Scrub the unscavenged control stack space, so that we can't run
3574 * into any stale pointers in a later GC (this is done by the
3575 * stop-for-gc handler in the other threads). */
3576 scrub_control_stack();
3581 /* Scavenge the Lisp functions of the interrupt handlers, taking
3582 * care to avoid SIG_DFL and SIG_IGN. */
3583 for (i = 0; i < NSIG; i++) {
3584 union interrupt_handler handler = interrupt_handlers[i];
3585 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
3586 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
3587 scavenge((lispobj *)(interrupt_handlers + i), 1);
3590 /* Scavenge the binding stacks. */
3593 for_each_thread(th) {
3594 sword_t len= (lispobj *)get_binding_stack_pointer(th) -
3595 th->binding_stack_start;
3596 scavenge((lispobj *) th->binding_stack_start,len);
3597 #ifdef LISP_FEATURE_SB_THREAD
3598 /* do the tls as well */
3599 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
3600 (sizeof (struct thread))/(sizeof (lispobj));
3601 scavenge((lispobj *) (th+1),len);
3606 /* The original CMU CL code had scavenge-read-only-space code
3607 * controlled by the Lisp-level variable
3608 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
3609 * wasn't documented under what circumstances it was useful or
3610 * safe to turn it on, so it's been turned off in SBCL. If you
3611 * want/need this functionality, and can test and document it,
3612 * please submit a patch. */
3614 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
3615 uword_t read_only_space_size =
3616 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
3617 (lispobj*)READ_ONLY_SPACE_START;
3619 "/scavenge read only space: %d bytes\n",
3620 read_only_space_size * sizeof(lispobj)));
3621 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
3625 /* Scavenge static space. */
3627 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
3628 (lispobj *)STATIC_SPACE_START;
3629 if (gencgc_verbose > 1) {
3631 "/scavenge static space: %d bytes\n",
3632 static_space_size * sizeof(lispobj)));
3634 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
3636 /* All generations but the generation being GCed need to be
3637 * scavenged. The new_space generation needs special handling as
3638 * objects may be moved in - it is handled separately below. */
3639 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
3641 /* Finally scavenge the new_space generation. Keep going until no
3642 * more objects are moved into the new generation */
3643 scavenge_newspace_generation(new_space);
3645 /* FIXME: I tried reenabling this check when debugging unrelated
3646 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
3647 * Since the current GC code seems to work well, I'm guessing that
3648 * this debugging code is just stale, but I haven't tried to
3649 * figure it out. It should be figured out and then either made to
3650 * work or just deleted. */
3651 #define RESCAN_CHECK 0
3653 /* As a check re-scavenge the newspace once; no new objects should
3656 os_vm_size_t old_bytes_allocated = bytes_allocated;
3657 os_vm_size_t bytes_allocated;
3659 /* Start with a full scavenge. */
3660 scavenge_newspace_generation_one_scan(new_space);
3662 /* Flush the current regions, updating the tables. */
3663 gc_alloc_update_all_page_tables();
3665 bytes_allocated = bytes_allocated - old_bytes_allocated;
3667 if (bytes_allocated != 0) {
3668 lose("Rescan of new_space allocated %d more bytes.\n",
3674 scan_weak_hash_tables();
3675 scan_weak_pointers();
3677 /* Flush the current regions, updating the tables. */
3678 gc_alloc_update_all_page_tables();
3680 /* Free the pages in oldspace, but not those marked dont_move. */
3681 bytes_freed = free_oldspace();
3683 /* If the GC is not raising the age then lower the generation back
3684 * to its normal generation number */
3686 for (i = 0; i < last_free_page; i++)
3687 if ((page_table[i].bytes_used != 0)
3688 && (page_table[i].gen == SCRATCH_GENERATION))
3689 page_table[i].gen = generation;
3690 gc_assert(generations[generation].bytes_allocated == 0);
3691 generations[generation].bytes_allocated =
3692 generations[SCRATCH_GENERATION].bytes_allocated;
3693 generations[SCRATCH_GENERATION].bytes_allocated = 0;
3696 /* Reset the alloc_start_page for generation. */
3697 generations[generation].alloc_start_page = 0;
3698 generations[generation].alloc_unboxed_start_page = 0;
3699 generations[generation].alloc_large_start_page = 0;
3700 generations[generation].alloc_large_unboxed_start_page = 0;
3702 if (generation >= verify_gens) {
3703 if (gencgc_verbose) {
3707 verify_dynamic_space();
3710 /* Set the new gc trigger for the GCed generation. */
3711 generations[generation].gc_trigger =
3712 generations[generation].bytes_allocated
3713 + generations[generation].bytes_consed_between_gc;
3716 generations[generation].num_gc = 0;
3718 ++generations[generation].num_gc;
3722 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
3724 update_dynamic_space_free_pointer(void)
3726 page_index_t last_page = -1, i;
3728 for (i = 0; i < last_free_page; i++)
3729 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
3732 last_free_page = last_page+1;
3734 set_alloc_pointer((lispobj)(page_address(last_free_page)));
3735 return 0; /* dummy value: return something ... */
3739 remap_page_range (page_index_t from, page_index_t to)
3741 /* There's a mysterious Solaris/x86 problem with using mmap
3742 * tricks for memory zeroing. See sbcl-devel thread
3743 * "Re: patch: standalone executable redux".
3745 #if defined(LISP_FEATURE_SUNOS)
3746 zero_and_mark_pages(from, to);
3749 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
3750 release_mask = release_granularity-1,
3752 aligned_from = (from+release_mask)&~release_mask,
3753 aligned_end = (end&~release_mask);
3755 if (aligned_from < aligned_end) {
3756 zero_pages_with_mmap(aligned_from, aligned_end-1);
3757 if (aligned_from != from)
3758 zero_and_mark_pages(from, aligned_from-1);
3759 if (aligned_end != end)
3760 zero_and_mark_pages(aligned_end, end-1);
3762 zero_and_mark_pages(from, to);
3768 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
3770 page_index_t first_page, last_page;
3773 return remap_page_range(from, to);
3775 for (first_page = from; first_page <= to; first_page++) {
3776 if (page_allocated_p(first_page) ||
3777 (page_table[first_page].need_to_zero == 0))
3780 last_page = first_page + 1;
3781 while (page_free_p(last_page) &&
3782 (last_page <= to) &&
3783 (page_table[last_page].need_to_zero == 1))
3786 remap_page_range(first_page, last_page-1);
3788 first_page = last_page;
3792 generation_index_t small_generation_limit = 1;
3794 /* GC all generations newer than last_gen, raising the objects in each
3795 * to the next older generation - we finish when all generations below
3796 * last_gen are empty. Then if last_gen is due for a GC, or if
3797 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
3798 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
3800 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
3801 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
3803 collect_garbage(generation_index_t last_gen)
3805 generation_index_t gen = 0, i;
3806 int raise, more = 0;
3808 /* The largest value of last_free_page seen since the time
3809 * remap_free_pages was called. */
3810 static page_index_t high_water_mark = 0;
3812 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
3813 log_generation_stats(gc_logfile, "=== GC Start ===");
3817 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
3819 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
3824 /* Flush the alloc regions updating the tables. */
3825 gc_alloc_update_all_page_tables();
3827 /* Verify the new objects created by Lisp code. */
3828 if (pre_verify_gen_0) {
3829 FSHOW((stderr, "pre-checking generation 0\n"));
3830 verify_generation(0);
3833 if (gencgc_verbose > 1)
3834 print_generation_stats();
3837 /* Collect the generation. */
3839 if (more || (gen >= gencgc_oldest_gen_to_gc)) {
3840 /* Never raise the oldest generation. Never raise the extra generation
3841 * collected due to more-flag. */
3847 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
3848 /* If we would not normally raise this one, but we're
3849 * running low on space in comparison to the object-sizes
3850 * we've been seeing, raise it and collect the next one
3852 if (!raise && gen == last_gen) {
3853 more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated);
3858 if (gencgc_verbose > 1) {
3860 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
3863 generations[gen].bytes_allocated,
3864 generations[gen].gc_trigger,
3865 generations[gen].num_gc));
3868 /* If an older generation is being filled, then update its
3871 generations[gen+1].cum_sum_bytes_allocated +=
3872 generations[gen+1].bytes_allocated;
3875 garbage_collect_generation(gen, raise);
3877 /* Reset the memory age cum_sum. */
3878 generations[gen].cum_sum_bytes_allocated = 0;
3880 if (gencgc_verbose > 1) {
3881 FSHOW((stderr, "GC of generation %d finished:\n", gen));
3882 print_generation_stats();
3886 } while ((gen <= gencgc_oldest_gen_to_gc)
3887 && ((gen < last_gen)
3890 && (generations[gen].bytes_allocated
3891 > generations[gen].gc_trigger)
3892 && (generation_average_age(gen)
3893 > generations[gen].minimum_age_before_gc))));
3895 /* Now if gen-1 was raised all generations before gen are empty.
3896 * If it wasn't raised then all generations before gen-1 are empty.
3898 * Now objects within this gen's pages cannot point to younger
3899 * generations unless they are written to. This can be exploited
3900 * by write-protecting the pages of gen; then when younger
3901 * generations are GCed only the pages which have been written
3906 gen_to_wp = gen - 1;
3908 /* There's not much point in WPing pages in generation 0 as it is
3909 * never scavenged (except promoted pages). */
3910 if ((gen_to_wp > 0) && enable_page_protection) {
3911 /* Check that they are all empty. */
3912 for (i = 0; i < gen_to_wp; i++) {
3913 if (generations[i].bytes_allocated)
3914 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
3917 write_protect_generation_pages(gen_to_wp);
3920 /* Set gc_alloc() back to generation 0. The current regions should
3921 * be flushed after the above GCs. */
3922 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
3923 gc_alloc_generation = 0;
3925 /* Save the high-water mark before updating last_free_page */
3926 if (last_free_page > high_water_mark)
3927 high_water_mark = last_free_page;
3929 update_dynamic_space_free_pointer();
3931 /* Update auto_gc_trigger. Make sure we trigger the next GC before
3932 * running out of heap! */
3933 if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
3934 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
3936 auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;
3939 fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n",
3942 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
3945 if (gen > small_generation_limit) {
3946 if (last_free_page > high_water_mark)
3947 high_water_mark = last_free_page;
3948 remap_free_pages(0, high_water_mark, 0);
3949 high_water_mark = 0;
3953 large_allocation = 0;
3955 log_generation_stats(gc_logfile, "=== GC End ===");
3956 SHOW("returning from collect_garbage");
3959 /* This is called by Lisp PURIFY when it is finished. All live objects
3960 * will have been moved to the RO and Static heaps. The dynamic space
3961 * will need a full re-initialization. We don't bother having Lisp
3962 * PURIFY flush the current gc_alloc() region, as the page_tables are
3963 * re-initialized, and every page is zeroed to be sure. */
3967 page_index_t page, last_page;
3969 if (gencgc_verbose > 1) {
3970 SHOW("entering gc_free_heap");
3973 for (page = 0; page < page_table_pages; page++) {
3974 /* Skip free pages which should already be zero filled. */
3975 if (page_allocated_p(page)) {
3977 for (last_page = page;
3978 (last_page < page_table_pages) && page_allocated_p(last_page);
3980 /* Mark the page free. The other slots are assumed invalid
3981 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
3982 * should not be write-protected -- except that the
3983 * generation is used for the current region but it sets
3985 page_table[page].allocated = FREE_PAGE_FLAG;
3986 page_table[page].bytes_used = 0;
3987 page_table[page].write_protected = 0;
3990 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
3991 * about this change. */
3992 page_start = (void *)page_address(page);
3993 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
3994 remap_free_pages(page, last_page-1, 1);
3997 } else if (gencgc_zero_check_during_free_heap) {
3998 /* Double-check that the page is zero filled. */
3999 sword_t *page_start;
4001 gc_assert(page_free_p(page));
4002 gc_assert(page_table[page].bytes_used == 0);
4003 page_start = (sword_t *)page_address(page);
4004 for (i=0; i<GENCGC_CARD_BYTES/sizeof(sword_t); i++) {
4005 if (page_start[i] != 0) {
4006 lose("free region not zero at %x\n", page_start + i);
4012 bytes_allocated = 0;
4014 /* Initialize the generations. */
4015 for (page = 0; page < NUM_GENERATIONS; page++) {
4016 generations[page].alloc_start_page = 0;
4017 generations[page].alloc_unboxed_start_page = 0;
4018 generations[page].alloc_large_start_page = 0;
4019 generations[page].alloc_large_unboxed_start_page = 0;
4020 generations[page].bytes_allocated = 0;
4021 generations[page].gc_trigger = 2000000;
4022 generations[page].num_gc = 0;
4023 generations[page].cum_sum_bytes_allocated = 0;
4026 if (gencgc_verbose > 1)
4027 print_generation_stats();
4029 /* Initialize gc_alloc(). */
4030 gc_alloc_generation = 0;
4032 gc_set_region_empty(&boxed_region);
4033 gc_set_region_empty(&unboxed_region);
4036 set_alloc_pointer((lispobj)((char *)heap_base));
4038 if (verify_after_free_heap) {
4039 /* Check whether purify has left any bad pointers. */
4040 FSHOW((stderr, "checking after free_heap\n"));
4050 #if defined(LISP_FEATURE_SB_SAFEPOINT)
4054 /* Compute the number of pages needed for the dynamic space.
4055 * Dynamic space size should be aligned on page size. */
4056 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4057 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4059 /* Default nursery size to 5% of the total dynamic space size,
4061 bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
4062 if (bytes_consed_between_gcs < (1024*1024))
4063 bytes_consed_between_gcs = 1024*1024;
4065 /* The page_table must be allocated using "calloc" to initialize
4066 * the page structures correctly. There used to be a separate
4067 * initialization loop (now commented out; see below) but that was
4068 * unnecessary and did hurt startup time. */
4069 page_table = calloc(page_table_pages, sizeof(struct page));
4070 gc_assert(page_table);
4073 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4074 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4076 heap_base = (void*)DYNAMIC_SPACE_START;
4078 /* The page structures are initialized implicitly when page_table
4079 * is allocated with "calloc" above. Formerly we had the following
4080 * explicit initialization here (comments converted to C99 style
4081 * for readability as C's block comments don't nest):
4083 * // Initialize each page structure.
4084 * for (i = 0; i < page_table_pages; i++) {
4085 * // Initialize all pages as free.
4086 * page_table[i].allocated = FREE_PAGE_FLAG;
4087 * page_table[i].bytes_used = 0;
4089 * // Pages are not write-protected at startup.
4090 * page_table[i].write_protected = 0;
4093 * Without this loop the image starts up much faster when dynamic
4094 * space is large -- which it is on 64-bit platforms already by
4095 * default -- and when "calloc" for large arrays is implemented
4096 * using copy-on-write of a page of zeroes -- which it is at least
4097 * on Linux. In this case the pages that page_table_pages is stored
4098 * in are mapped and cleared not before the corresponding part of
4099 * dynamic space is used. For example, this saves clearing 16 MB of
4100 * memory at startup if the page size is 4 KB and the size of
4101 * dynamic space is 4 GB.
4102 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4103 * asserted below: */
4105 /* Compile time assertion: If triggered, declares an array
4106 * of dimension -1 forcing a syntax error. The intent of the
4107 * assignment is to avoid an "unused variable" warning. */
4108 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4109 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4112 bytes_allocated = 0;
4114 /* Initialize the generations.
4116 * FIXME: very similar to code in gc_free_heap(), should be shared */
4117 for (i = 0; i < NUM_GENERATIONS; i++) {
4118 generations[i].alloc_start_page = 0;
4119 generations[i].alloc_unboxed_start_page = 0;
4120 generations[i].alloc_large_start_page = 0;
4121 generations[i].alloc_large_unboxed_start_page = 0;
4122 generations[i].bytes_allocated = 0;
4123 generations[i].gc_trigger = 2000000;
4124 generations[i].num_gc = 0;
4125 generations[i].cum_sum_bytes_allocated = 0;
4126 /* the tune-able parameters */
4127 generations[i].bytes_consed_between_gc
4128 = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
4129 generations[i].number_of_gcs_before_promotion = 1;
4130 generations[i].minimum_age_before_gc = 0.75;
4133 /* Initialize gc_alloc. */
4134 gc_alloc_generation = 0;
4135 gc_set_region_empty(&boxed_region);
4136 gc_set_region_empty(&unboxed_region);
4141 /* Pick up the dynamic space from after a core load.
4143 * The ALLOCATION_POINTER points to the end of the dynamic space.
4147 gencgc_pickup_dynamic(void)
4149 page_index_t page = 0;
4150 void *alloc_ptr = (void *)get_alloc_pointer();
4151 lispobj *prev=(lispobj *)page_address(page);
4152 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4154 bytes_allocated = 0;
4157 lispobj *first,*ptr= (lispobj *)page_address(page);
4159 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4160 /* It is possible, though rare, for the saved page table
4161 * to contain free pages below alloc_ptr. */
4162 page_table[page].gen = gen;
4163 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4164 page_table[page].large_object = 0;
4165 page_table[page].write_protected = 0;
4166 page_table[page].write_protected_cleared = 0;
4167 page_table[page].dont_move = 0;
4168 page_table[page].need_to_zero = 1;
4170 bytes_allocated += GENCGC_CARD_BYTES;
4173 if (!gencgc_partial_pickup) {
4174 page_table[page].allocated = BOXED_PAGE_FLAG;
4175 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4178 page_table[page].region_start_offset =
4179 page_address(page) - (void *)prev;
4182 } while (page_address(page) < alloc_ptr);
4184 last_free_page = page;
4186 generations[gen].bytes_allocated = bytes_allocated;
4188 gc_alloc_update_all_page_tables();
4189 write_protect_generation_pages(gen);
4193 gc_initialize_pointers(void)
4195 gencgc_pickup_dynamic();
4199 /* alloc(..) is the external interface for memory allocation. It
4200 * allocates to generation 0. It is not called from within the garbage
4201 * collector as it is only external uses that need the check for heap
4202 * size (GC trigger) and to disable the interrupts (interrupts are
4203 * always disabled during a GC).
4205 * The vops that call alloc(..) assume that the returned space is zero-filled.
4206 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4208 * The check for a GC trigger is only performed when the current
4209 * region is full, so in most cases it's not needed. */
4211 static inline lispobj *
4212 general_alloc_internal(sword_t nbytes, int page_type_flag, struct alloc_region *region,
4213 struct thread *thread)
4215 #ifndef LISP_FEATURE_WIN32
4216 lispobj alloc_signal;
4219 void *new_free_pointer;
4220 os_vm_size_t trigger_bytes = 0;
4222 gc_assert(nbytes>0);
4224 /* Check for alignment allocation problems. */
4225 gc_assert((((uword_t)region->free_pointer & LOWTAG_MASK) == 0)
4226 && ((nbytes & LOWTAG_MASK) == 0));
4228 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
4229 /* Must be inside a PA section. */
4230 gc_assert(get_pseudo_atomic_atomic(thread));
4233 if (nbytes > large_allocation)
4234 large_allocation = nbytes;
4236 /* maybe we can do this quickly ... */
4237 new_free_pointer = region->free_pointer + nbytes;
4238 if (new_free_pointer <= region->end_addr) {
4239 new_obj = (void*)(region->free_pointer);
4240 region->free_pointer = new_free_pointer;
4241 return(new_obj); /* yup */
4244 /* We don't want to count nbytes against auto_gc_trigger unless we
4245 * have to: it speeds up the tenuring of objects and slows down
4246 * allocation. However, unless we do so when allocating _very_
4247 * large objects we are in danger of exhausting the heap without
4248 * running sufficient GCs.
4250 if (nbytes >= bytes_consed_between_gcs)
4251 trigger_bytes = nbytes;
4253 /* we have to go the long way around, it seems. Check whether we
4254 * should GC in the near future
4256 if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
4257 /* Don't flood the system with interrupts if the need to gc is
4258 * already noted. This can happen for example when SUB-GC
4259 * allocates or after a gc triggered in a WITHOUT-GCING. */
4260 if (SymbolValue(GC_PENDING,thread) == NIL) {
4261 /* set things up so that GC happens when we finish the PA
4263 SetSymbolValue(GC_PENDING,T,thread);
4264 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4265 #ifdef LISP_FEATURE_SB_SAFEPOINT
4266 thread_register_gc_trigger();
4268 set_pseudo_atomic_interrupted(thread);
4269 #ifdef GENCGC_IS_PRECISE
4270 /* PPC calls alloc() from a trap or from pa_alloc(),
4271 * look up the most context if it's from a trap. */
4273 os_context_t *context =
4274 thread->interrupt_data->allocation_trap_context;
4275 maybe_save_gc_mask_and_block_deferrables
4276 (context ? os_context_sigmask_addr(context) : NULL);
4279 maybe_save_gc_mask_and_block_deferrables(NULL);
4285 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4287 #ifndef LISP_FEATURE_WIN32
4288 /* for sb-prof, and not supported on Windows yet */
4289 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4290 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4291 if ((sword_t) alloc_signal <= 0) {
4292 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4295 SetSymbolValue(ALLOC_SIGNAL,
4296 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4306 general_alloc(sword_t nbytes, int page_type_flag)
4308 struct thread *thread = arch_os_get_current_thread();
4309 /* Select correct region, and call general_alloc_internal with it.
4310 * For other then boxed allocation we must lock first, since the
4311 * region is shared. */
4312 if (BOXED_PAGE_FLAG & page_type_flag) {
4313 #ifdef LISP_FEATURE_SB_THREAD
4314 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4316 struct alloc_region *region = &boxed_region;
4318 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4319 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4321 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4322 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4323 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4326 lose("bad page type flag: %d", page_type_flag);
4330 lispobj AMD64_SYSV_ABI *
4333 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4334 struct thread *self = arch_os_get_current_thread();
4335 int was_pseudo_atomic = get_pseudo_atomic_atomic(self);
4336 if (!was_pseudo_atomic)
4337 set_pseudo_atomic_atomic(self);
4339 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4342 lispobj *result = general_alloc(nbytes, BOXED_PAGE_FLAG);
4344 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4345 if (!was_pseudo_atomic)
4346 clear_pseudo_atomic_atomic(self);
4353 * shared support for the OS-dependent signal handlers which
4354 * catch GENCGC-related write-protect violations
4356 void unhandled_sigmemoryfault(void* addr);
4358 /* Depending on which OS we're running under, different signals might
4359 * be raised for a violation of write protection in the heap. This
4360 * function factors out the common generational GC magic which needs
4361 * to invoked in this case, and should be called from whatever signal
4362 * handler is appropriate for the OS we're running under.
4364 * Return true if this signal is a normal generational GC thing that
4365 * we were able to handle, or false if it was abnormal and control
4366 * should fall through to the general SIGSEGV/SIGBUS/whatever logic.
4368 * We have two control flags for this: one causes us to ignore faults
4369 * on unprotected pages completely, and the second complains to stderr
4370 * but allows us to continue without losing.
4372 extern boolean ignore_memoryfaults_on_unprotected_pages;
4373 boolean ignore_memoryfaults_on_unprotected_pages = 0;
4375 extern boolean continue_after_memoryfault_on_unprotected_pages;
4376 boolean continue_after_memoryfault_on_unprotected_pages = 0;
4379 gencgc_handle_wp_violation(void* fault_addr)
4381 page_index_t page_index = find_page_index(fault_addr);
4384 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4385 fault_addr, page_index));
4388 /* Check whether the fault is within the dynamic space. */
4389 if (page_index == (-1)) {
4391 /* It can be helpful to be able to put a breakpoint on this
4392 * case to help diagnose low-level problems. */
4393 unhandled_sigmemoryfault(fault_addr);
4395 /* not within the dynamic space -- not our responsibility */
4400 ret = thread_mutex_lock(&free_pages_lock);
4401 gc_assert(ret == 0);
4402 if (page_table[page_index].write_protected) {
4403 /* Unprotect the page. */
4404 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4405 page_table[page_index].write_protected_cleared = 1;
4406 page_table[page_index].write_protected = 0;
4407 } else if (!ignore_memoryfaults_on_unprotected_pages) {
4408 /* The only acceptable reason for this signal on a heap
4409 * access is that GENCGC write-protected the page.
4410 * However, if two CPUs hit a wp page near-simultaneously,
4411 * we had better not have the second one lose here if it
4412 * does this test after the first one has already set wp=0
4414 if(page_table[page_index].write_protected_cleared != 1) {
4415 void lisp_backtrace(int frames);
4418 "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n"
4419 " boxed_region.first_page: %"PAGE_INDEX_FMT","
4420 " boxed_region.last_page %"PAGE_INDEX_FMT"\n"
4421 " page.region_start_offset: %"OS_VM_SIZE_FMT"\n"
4422 " page.bytes_used: %"PAGE_BYTES_FMT"\n"
4423 " page.allocated: %d\n"
4424 " page.write_protected: %d\n"
4425 " page.write_protected_cleared: %d\n"
4426 " page.generation: %d\n",
4429 boxed_region.first_page,
4430 boxed_region.last_page,
4431 page_table[page_index].region_start_offset,
4432 page_table[page_index].bytes_used,
4433 page_table[page_index].allocated,
4434 page_table[page_index].write_protected,
4435 page_table[page_index].write_protected_cleared,
4436 page_table[page_index].gen);
4437 if (!continue_after_memoryfault_on_unprotected_pages)
4441 ret = thread_mutex_unlock(&free_pages_lock);
4442 gc_assert(ret == 0);
4443 /* Don't worry, we can handle it. */
4447 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4448 * it's not just a case of the program hitting the write barrier, and
4449 * are about to let Lisp deal with it. It's basically just a
4450 * convenient place to set a gdb breakpoint. */
4452 unhandled_sigmemoryfault(void *addr)
4455 void gc_alloc_update_all_page_tables(void)
4457 /* Flush the alloc regions updating the tables. */
4459 for_each_thread(th) {
4460 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4461 #if defined(LISP_FEATURE_SB_SAFEPOINT_STRICTLY) && !defined(LISP_FEATURE_WIN32)
4462 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->sprof_alloc_region);
4465 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4466 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4470 gc_set_region_empty(struct alloc_region *region)
4472 region->first_page = 0;
4473 region->last_page = -1;
4474 region->start_addr = page_address(0);
4475 region->free_pointer = page_address(0);
4476 region->end_addr = page_address(0);
4480 zero_all_free_pages()
4484 for (i = 0; i < last_free_page; i++) {
4485 if (page_free_p(i)) {
4486 #ifdef READ_PROTECT_FREE_PAGES
4487 os_protect(page_address(i),
4496 /* Things to do before doing a final GC before saving a core (without
4499 * + Pages in large_object pages aren't moved by the GC, so we need to
4500 * unset that flag from all pages.
4501 * + The pseudo-static generation isn't normally collected, but it seems
4502 * reasonable to collect it at least when saving a core. So move the
4503 * pages to a normal generation.
4506 prepare_for_final_gc ()
4509 for (i = 0; i < last_free_page; i++) {
4510 page_table[i].large_object = 0;
4511 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4512 int used = page_table[i].bytes_used;
4513 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4514 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4515 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4521 /* Do a non-conservative GC, and then save a core with the initial
4522 * function being set to the value of the static symbol
4523 * SB!VM:RESTART-LISP-FUNCTION */
4525 gc_and_save(char *filename, boolean prepend_runtime,
4526 boolean save_runtime_options,
4527 boolean compressed, int compression_level)
4530 void *runtime_bytes = NULL;
4531 size_t runtime_size;
4533 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4538 conservative_stack = 0;
4540 /* The filename might come from Lisp, and be moved by the now
4541 * non-conservative GC. */
4542 filename = strdup(filename);
4544 /* Collect twice: once into relatively high memory, and then back
4545 * into low memory. This compacts the retained data into the lower
4546 * pages, minimizing the size of the core file.
4548 prepare_for_final_gc();
4549 gencgc_alloc_start_page = last_free_page;
4550 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4552 prepare_for_final_gc();
4553 gencgc_alloc_start_page = -1;
4554 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4556 if (prepend_runtime)
4557 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4559 /* The dumper doesn't know that pages need to be zeroed before use. */
4560 zero_all_free_pages();
4561 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4562 prepend_runtime, save_runtime_options,
4563 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
4564 /* Oops. Save still managed to fail. Since we've mangled the stack
4565 * beyond hope, there's not much we can do.
4566 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4567 * going to be rather unsatisfactory too... */
4568 lose("Attempt to save core after non-conservative GC failed.\n");