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_scan_start(page_index_t page_index)
238 return page_address(page_index)-page_table[page_index].scan_start_offset;
241 /* True if the page starts a contiguous block. */
242 static inline boolean
243 page_starts_contiguous_block_p(page_index_t page_index)
245 return page_table[page_index].scan_start_offset == 0;
248 /* Find the page index within the page_table for the given
249 * address. Return -1 on failure. */
251 find_page_index(void *addr)
253 if (addr >= heap_base) {
254 page_index_t index = ((pointer_sized_uint_t)addr -
255 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
256 if (index < page_table_pages)
263 npage_bytes(page_index_t npages)
265 gc_assert(npages>=0);
266 return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
269 /* Check that X is a higher address than Y and return offset from Y to
271 static inline os_vm_size_t
272 void_diff(void *x, void *y)
275 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
278 /* a structure to hold the state of a generation
280 * CAUTION: If you modify this, make sure to touch up the alien
281 * definition in src/code/gc.lisp accordingly. ...or better yes,
282 * deal with the FIXME there...
286 /* the first page that gc_alloc() checks on its next call */
287 page_index_t alloc_start_page;
289 /* the first page that gc_alloc_unboxed() checks on its next call */
290 page_index_t alloc_unboxed_start_page;
292 /* the first page that gc_alloc_large (boxed) considers on its next
293 * call. (Although it always allocates after the boxed_region.) */
294 page_index_t alloc_large_start_page;
296 /* the first page that gc_alloc_large (unboxed) considers on its
297 * next call. (Although it always allocates after the
298 * current_unboxed_region.) */
299 page_index_t alloc_large_unboxed_start_page;
301 /* the bytes allocated to this generation */
302 os_vm_size_t bytes_allocated;
304 /* the number of bytes at which to trigger a GC */
305 os_vm_size_t gc_trigger;
307 /* to calculate a new level for gc_trigger */
308 os_vm_size_t bytes_consed_between_gc;
310 /* the number of GCs since the last raise */
313 /* the number of GCs to run on the generations before raising objects to the
315 int number_of_gcs_before_promotion;
317 /* the cumulative sum of the bytes allocated to this generation. It is
318 * cleared after a GC on this generations, and update before new
319 * objects are added from a GC of a younger generation. Dividing by
320 * the bytes_allocated will give the average age of the memory in
321 * this generation since its last GC. */
322 os_vm_size_t cum_sum_bytes_allocated;
324 /* a minimum average memory age before a GC will occur helps
325 * prevent a GC when a large number of new live objects have been
326 * added, in which case a GC could be a waste of time */
327 double minimum_age_before_gc;
330 /* an array of generation structures. There needs to be one more
331 * generation structure than actual generations as the oldest
332 * generation is temporarily raised then lowered. */
333 struct generation generations[NUM_GENERATIONS];
335 /* the oldest generation that is will currently be GCed by default.
336 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
338 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
340 * Setting this to 0 effectively disables the generational nature of
341 * the GC. In some applications generational GC may not be useful
342 * because there are no long-lived objects.
344 * An intermediate value could be handy after moving long-lived data
345 * into an older generation so an unnecessary GC of this long-lived
346 * data can be avoided. */
347 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
349 /* The maximum free page in the heap is maintained and used to update
350 * ALLOCATION_POINTER which is used by the room function to limit its
351 * search of the heap. XX Gencgc obviously needs to be better
352 * integrated with the Lisp code. */
353 page_index_t last_free_page;
355 #ifdef LISP_FEATURE_SB_THREAD
356 /* This lock is to prevent multiple threads from simultaneously
357 * allocating new regions which overlap each other. Note that the
358 * majority of GC is single-threaded, but alloc() may be called from
359 * >1 thread at a time and must be thread-safe. This lock must be
360 * seized before all accesses to generations[] or to parts of
361 * page_table[] that other threads may want to see */
362 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
363 /* This lock is used to protect non-thread-local allocation. */
364 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
367 extern os_vm_size_t gencgc_release_granularity;
368 os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
370 extern os_vm_size_t gencgc_alloc_granularity;
371 os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
375 * miscellaneous heap functions
378 /* Count the number of pages which are write-protected within the
379 * given generation. */
381 count_write_protect_generation_pages(generation_index_t generation)
383 page_index_t i, count = 0;
385 for (i = 0; i < last_free_page; i++)
386 if (page_allocated_p(i)
387 && (page_table[i].gen == generation)
388 && (page_table[i].write_protected == 1))
393 /* Count the number of pages within the given generation. */
395 count_generation_pages(generation_index_t generation)
398 page_index_t count = 0;
400 for (i = 0; i < last_free_page; i++)
401 if (page_allocated_p(i)
402 && (page_table[i].gen == generation))
409 count_dont_move_pages(void)
412 page_index_t count = 0;
413 for (i = 0; i < last_free_page; i++) {
414 if (page_allocated_p(i)
415 && (page_table[i].dont_move != 0)) {
423 /* Work through the pages and add up the number of bytes used for the
424 * given generation. */
426 count_generation_bytes_allocated (generation_index_t gen)
429 os_vm_size_t result = 0;
430 for (i = 0; i < last_free_page; i++) {
431 if (page_allocated_p(i)
432 && (page_table[i].gen == gen))
433 result += page_table[i].bytes_used;
438 /* Return the average age of the memory in a generation. */
440 generation_average_age(generation_index_t gen)
442 if (generations[gen].bytes_allocated == 0)
446 ((double)generations[gen].cum_sum_bytes_allocated)
447 / ((double)generations[gen].bytes_allocated);
451 write_generation_stats(FILE *file)
453 generation_index_t i;
455 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
456 #define FPU_STATE_SIZE 27
457 int fpu_state[FPU_STATE_SIZE];
458 #elif defined(LISP_FEATURE_PPC)
459 #define FPU_STATE_SIZE 32
460 long long fpu_state[FPU_STATE_SIZE];
461 #elif defined(LISP_FEATURE_SPARC)
463 * 32 (single-precision) FP registers, and the FP state register.
464 * But Sparc V9 has 32 double-precision registers (equivalent to 64
465 * single-precision, but can't be accessed), so we leave enough room
468 #define FPU_STATE_SIZE (((32 + 32 + 1) + 1)/2)
469 long long fpu_state[FPU_STATE_SIZE];
472 /* This code uses the FP instructions which may be set up for Lisp
473 * so they need to be saved and reset for C. */
476 /* Print the heap stats. */
478 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
480 for (i = 0; i < SCRATCH_GENERATION; i++) {
482 page_index_t boxed_cnt = 0;
483 page_index_t unboxed_cnt = 0;
484 page_index_t large_boxed_cnt = 0;
485 page_index_t large_unboxed_cnt = 0;
486 page_index_t pinned_cnt=0;
488 for (j = 0; j < last_free_page; j++)
489 if (page_table[j].gen == i) {
491 /* Count the number of boxed pages within the given
493 if (page_boxed_p(j)) {
494 if (page_table[j].large_object)
499 if(page_table[j].dont_move) pinned_cnt++;
500 /* Count the number of unboxed pages within the given
502 if (page_unboxed_p(j)) {
503 if (page_table[j].large_object)
510 gc_assert(generations[i].bytes_allocated
511 == count_generation_bytes_allocated(i));
513 " %1d: %5ld %5ld %5ld %5ld",
515 generations[i].alloc_start_page,
516 generations[i].alloc_unboxed_start_page,
517 generations[i].alloc_large_start_page,
518 generations[i].alloc_large_unboxed_start_page);
520 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
521 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
522 boxed_cnt, unboxed_cnt, large_boxed_cnt,
523 large_unboxed_cnt, pinned_cnt);
528 " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
529 generations[i].bytes_allocated,
530 (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
531 generations[i].gc_trigger,
532 count_write_protect_generation_pages(i),
533 generations[i].num_gc,
534 generation_average_age(i));
536 fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
537 fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);
539 fpu_restore(fpu_state);
543 write_heap_exhaustion_report(FILE *file, long available, long requested,
544 struct thread *thread)
547 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
548 gc_active_p ? "garbage collection" : "allocation",
551 write_generation_stats(file);
552 fprintf(file, "GC control variables:\n");
553 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
554 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
555 (SymbolValue(GC_PENDING, thread) == T) ?
556 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
557 "false" : "in progress"));
558 #ifdef LISP_FEATURE_SB_THREAD
559 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
560 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
565 print_generation_stats(void)
567 write_generation_stats(stderr);
570 extern char* gc_logfile;
571 char * gc_logfile = NULL;
574 log_generation_stats(char *logfile, char *header)
577 FILE * log = fopen(logfile, "a");
579 fprintf(log, "%s\n", header);
580 write_generation_stats(log);
583 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
590 report_heap_exhaustion(long available, long requested, struct thread *th)
593 FILE * log = fopen(gc_logfile, "a");
595 write_heap_exhaustion_report(log, available, requested, th);
598 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
602 /* Always to stderr as well. */
603 write_heap_exhaustion_report(stderr, available, requested, th);
607 #if defined(LISP_FEATURE_X86)
608 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
611 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
612 * if zeroing it ourselves, i.e. in practice give the memory back to the
613 * OS. Generally done after a large GC.
615 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
617 void *addr = page_address(start), *new_addr;
618 os_vm_size_t length = npage_bytes(1+end-start);
623 gc_assert(length >= gencgc_release_granularity);
624 gc_assert((length % gencgc_release_granularity) == 0);
626 os_invalidate(addr, length);
627 new_addr = os_validate(addr, length);
628 if (new_addr == NULL || new_addr != addr) {
629 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
633 for (i = start; i <= end; i++) {
634 page_table[i].need_to_zero = 0;
638 /* Zero the pages from START to END (inclusive). Generally done just after
639 * a new region has been allocated.
642 zero_pages(page_index_t start, page_index_t end) {
646 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
647 fast_bzero(page_address(start), npage_bytes(1+end-start));
649 bzero(page_address(start), npage_bytes(1+end-start));
655 zero_and_mark_pages(page_index_t start, page_index_t end) {
658 zero_pages(start, end);
659 for (i = start; i <= end; i++)
660 page_table[i].need_to_zero = 0;
663 /* Zero the pages from START to END (inclusive), except for those
664 * pages that are known to already zeroed. Mark all pages in the
665 * ranges as non-zeroed.
668 zero_dirty_pages(page_index_t start, page_index_t end) {
671 for (i = start; i <= end; i++) {
672 if (!page_table[i].need_to_zero) continue;
673 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
678 for (i = start; i <= end; i++) {
679 page_table[i].need_to_zero = 1;
685 * To support quick and inline allocation, regions of memory can be
686 * allocated and then allocated from with just a free pointer and a
687 * check against an end address.
689 * Since objects can be allocated to spaces with different properties
690 * e.g. boxed/unboxed, generation, ages; there may need to be many
691 * allocation regions.
693 * Each allocation region may start within a partly used page. Many
694 * features of memory use are noted on a page wise basis, e.g. the
695 * generation; so if a region starts within an existing allocated page
696 * it must be consistent with this page.
698 * During the scavenging of the newspace, objects will be transported
699 * into an allocation region, and pointers updated to point to this
700 * allocation region. It is possible that these pointers will be
701 * scavenged again before the allocation region is closed, e.g. due to
702 * trans_list which jumps all over the place to cleanup the list. It
703 * is important to be able to determine properties of all objects
704 * pointed to when scavenging, e.g to detect pointers to the oldspace.
705 * Thus it's important that the allocation regions have the correct
706 * properties set when allocated, and not just set when closed. The
707 * region allocation routines return regions with the specified
708 * properties, and grab all the pages, setting their properties
709 * appropriately, except that the amount used is not known.
711 * These regions are used to support quicker allocation using just a
712 * free pointer. The actual space used by the region is not reflected
713 * in the pages tables until it is closed. It can't be scavenged until
716 * When finished with the region it should be closed, which will
717 * update the page tables for the actual space used returning unused
718 * space. Further it may be noted in the new regions which is
719 * necessary when scavenging the newspace.
721 * Large objects may be allocated directly without an allocation
722 * region, the page tables are updated immediately.
724 * Unboxed objects don't contain pointers to other objects and so
725 * don't need scavenging. Further they can't contain pointers to
726 * younger generations so WP is not needed. By allocating pages to
727 * unboxed objects the whole page never needs scavenging or
728 * write-protecting. */
730 /* We are only using two regions at present. Both are for the current
731 * newspace generation. */
732 struct alloc_region boxed_region;
733 struct alloc_region unboxed_region;
735 /* The generation currently being allocated to. */
736 static generation_index_t gc_alloc_generation;
738 static inline page_index_t
739 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
742 if (UNBOXED_PAGE_FLAG == page_type_flag) {
743 return generations[generation].alloc_large_unboxed_start_page;
744 } else if (BOXED_PAGE_FLAG & page_type_flag) {
745 /* Both code and data. */
746 return generations[generation].alloc_large_start_page;
748 lose("bad page type flag: %d", page_type_flag);
751 if (UNBOXED_PAGE_FLAG == page_type_flag) {
752 return generations[generation].alloc_unboxed_start_page;
753 } else if (BOXED_PAGE_FLAG & page_type_flag) {
754 /* Both code and data. */
755 return generations[generation].alloc_start_page;
757 lose("bad page_type_flag: %d", page_type_flag);
763 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
767 if (UNBOXED_PAGE_FLAG == page_type_flag) {
768 generations[generation].alloc_large_unboxed_start_page = page;
769 } else if (BOXED_PAGE_FLAG & page_type_flag) {
770 /* Both code and data. */
771 generations[generation].alloc_large_start_page = page;
773 lose("bad page type flag: %d", page_type_flag);
776 if (UNBOXED_PAGE_FLAG == page_type_flag) {
777 generations[generation].alloc_unboxed_start_page = page;
778 } else if (BOXED_PAGE_FLAG & page_type_flag) {
779 /* Both code and data. */
780 generations[generation].alloc_start_page = page;
782 lose("bad page type flag: %d", page_type_flag);
787 /* Find a new region with room for at least the given number of bytes.
789 * It starts looking at the current generation's alloc_start_page. So
790 * may pick up from the previous region if there is enough space. This
791 * keeps the allocation contiguous when scavenging the newspace.
793 * The alloc_region should have been closed by a call to
794 * gc_alloc_update_page_tables(), and will thus be in an empty state.
796 * To assist the scavenging functions write-protected pages are not
797 * used. Free pages should not be write-protected.
799 * It is critical to the conservative GC that the start of regions be
800 * known. To help achieve this only small regions are allocated at a
803 * During scavenging, pointers may be found to within the current
804 * region and the page generation must be set so that pointers to the
805 * from space can be recognized. Therefore the generation of pages in
806 * the region are set to gc_alloc_generation. To prevent another
807 * allocation call using the same pages, all the pages in the region
808 * are allocated, although they will initially be empty.
811 gc_alloc_new_region(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
813 page_index_t first_page;
814 page_index_t last_page;
815 os_vm_size_t bytes_found;
821 "/alloc_new_region for %d bytes from gen %d\n",
822 nbytes, gc_alloc_generation));
825 /* Check that the region is in a reset state. */
826 gc_assert((alloc_region->first_page == 0)
827 && (alloc_region->last_page == -1)
828 && (alloc_region->free_pointer == alloc_region->end_addr));
829 ret = thread_mutex_lock(&free_pages_lock);
831 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
832 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
833 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
834 + npage_bytes(last_page-first_page);
836 /* Set up the alloc_region. */
837 alloc_region->first_page = first_page;
838 alloc_region->last_page = last_page;
839 alloc_region->start_addr = page_table[first_page].bytes_used
840 + page_address(first_page);
841 alloc_region->free_pointer = alloc_region->start_addr;
842 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
844 /* Set up the pages. */
846 /* The first page may have already been in use. */
847 if (page_table[first_page].bytes_used == 0) {
848 page_table[first_page].allocated = page_type_flag;
849 page_table[first_page].gen = gc_alloc_generation;
850 page_table[first_page].large_object = 0;
851 page_table[first_page].scan_start_offset = 0;
854 gc_assert(page_table[first_page].allocated == page_type_flag);
855 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
857 gc_assert(page_table[first_page].gen == gc_alloc_generation);
858 gc_assert(page_table[first_page].large_object == 0);
860 for (i = first_page+1; i <= last_page; i++) {
861 page_table[i].allocated = page_type_flag;
862 page_table[i].gen = gc_alloc_generation;
863 page_table[i].large_object = 0;
864 /* This may not be necessary for unboxed regions (think it was
866 page_table[i].scan_start_offset =
867 void_diff(page_address(i),alloc_region->start_addr);
868 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
870 /* Bump up last_free_page. */
871 if (last_page+1 > last_free_page) {
872 last_free_page = last_page+1;
873 /* do we only want to call this on special occasions? like for
875 set_alloc_pointer((lispobj)page_address(last_free_page));
877 ret = thread_mutex_unlock(&free_pages_lock);
880 #ifdef READ_PROTECT_FREE_PAGES
881 os_protect(page_address(first_page),
882 npage_bytes(1+last_page-first_page),
886 /* If the first page was only partial, don't check whether it's
887 * zeroed (it won't be) and don't zero it (since the parts that
888 * we're interested in are guaranteed to be zeroed).
890 if (page_table[first_page].bytes_used) {
894 zero_dirty_pages(first_page, last_page);
896 /* we can do this after releasing free_pages_lock */
897 if (gencgc_zero_check) {
899 for (p = (word_t *)alloc_region->start_addr;
900 p < (word_t *)alloc_region->end_addr; p++) {
902 lose("The new region is not zero at %p (start=%p, end=%p).\n",
903 p, alloc_region->start_addr, alloc_region->end_addr);
909 /* If the record_new_objects flag is 2 then all new regions created
912 * If it's 1 then then it is only recorded if the first page of the
913 * current region is <= new_areas_ignore_page. This helps avoid
914 * unnecessary recording when doing full scavenge pass.
916 * The new_object structure holds the page, byte offset, and size of
917 * new regions of objects. Each new area is placed in the array of
918 * these structures pointer to by new_areas. new_areas_index holds the
919 * offset into new_areas.
921 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
922 * later code must detect this and handle it, probably by doing a full
923 * scavenge of a generation. */
924 #define NUM_NEW_AREAS 512
925 static int record_new_objects = 0;
926 static page_index_t new_areas_ignore_page;
932 static struct new_area (*new_areas)[];
933 static size_t new_areas_index;
934 size_t max_new_areas;
936 /* Add a new area to new_areas. */
938 add_new_area(page_index_t first_page, size_t offset, size_t size)
940 size_t new_area_start, c;
943 /* Ignore if full. */
944 if (new_areas_index >= NUM_NEW_AREAS)
947 switch (record_new_objects) {
951 if (first_page > new_areas_ignore_page)
960 new_area_start = npage_bytes(first_page) + offset;
962 /* Search backwards for a prior area that this follows from. If
963 found this will save adding a new area. */
964 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
966 npage_bytes((*new_areas)[i].page)
967 + (*new_areas)[i].offset
968 + (*new_areas)[i].size;
970 "/add_new_area S1 %d %d %d %d\n",
971 i, c, new_area_start, area_end));*/
972 if (new_area_start == area_end) {
974 "/adding to [%d] %d %d %d with %d %d %d:\n",
976 (*new_areas)[i].page,
977 (*new_areas)[i].offset,
978 (*new_areas)[i].size,
982 (*new_areas)[i].size += size;
987 (*new_areas)[new_areas_index].page = first_page;
988 (*new_areas)[new_areas_index].offset = offset;
989 (*new_areas)[new_areas_index].size = size;
991 "/new_area %d page %d offset %d size %d\n",
992 new_areas_index, first_page, offset, size));*/
995 /* Note the max new_areas used. */
996 if (new_areas_index > max_new_areas)
997 max_new_areas = new_areas_index;
1000 /* Update the tables for the alloc_region. The region may be added to
1003 * When done the alloc_region is set up so that the next quick alloc
1004 * will fail safely and thus a new region will be allocated. Further
1005 * it is safe to try to re-update the page table of this reset
1008 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
1011 page_index_t first_page;
1012 page_index_t next_page;
1013 os_vm_size_t bytes_used;
1014 os_vm_size_t region_size;
1015 os_vm_size_t byte_cnt;
1016 page_bytes_t orig_first_page_bytes_used;
1020 first_page = alloc_region->first_page;
1022 /* Catch an unused alloc_region. */
1023 if ((first_page == 0) && (alloc_region->last_page == -1))
1026 next_page = first_page+1;
1028 ret = thread_mutex_lock(&free_pages_lock);
1029 gc_assert(ret == 0);
1030 if (alloc_region->free_pointer != alloc_region->start_addr) {
1031 /* some bytes were allocated in the region */
1032 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1034 gc_assert(alloc_region->start_addr ==
1035 (page_address(first_page)
1036 + page_table[first_page].bytes_used));
1038 /* All the pages used need to be updated */
1040 /* Update the first page. */
1042 /* If the page was free then set up the gen, and
1043 * scan_start_offset. */
1044 if (page_table[first_page].bytes_used == 0)
1045 gc_assert(page_starts_contiguous_block_p(first_page));
1046 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1048 gc_assert(page_table[first_page].allocated & page_type_flag);
1049 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1050 gc_assert(page_table[first_page].large_object == 0);
1054 /* Calculate the number of bytes used in this page. This is not
1055 * always the number of new bytes, unless it was free. */
1057 if ((bytes_used = void_diff(alloc_region->free_pointer,
1058 page_address(first_page)))
1059 >GENCGC_CARD_BYTES) {
1060 bytes_used = GENCGC_CARD_BYTES;
1063 page_table[first_page].bytes_used = bytes_used;
1064 byte_cnt += bytes_used;
1067 /* All the rest of the pages should be free. We need to set
1068 * their scan_start_offset pointer to the start of the
1069 * region, and set the bytes_used. */
1071 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1072 gc_assert(page_table[next_page].allocated & page_type_flag);
1073 gc_assert(page_table[next_page].bytes_used == 0);
1074 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1075 gc_assert(page_table[next_page].large_object == 0);
1077 gc_assert(page_table[next_page].scan_start_offset ==
1078 void_diff(page_address(next_page),
1079 alloc_region->start_addr));
1081 /* Calculate the number of bytes used in this page. */
1083 if ((bytes_used = void_diff(alloc_region->free_pointer,
1084 page_address(next_page)))>GENCGC_CARD_BYTES) {
1085 bytes_used = GENCGC_CARD_BYTES;
1088 page_table[next_page].bytes_used = bytes_used;
1089 byte_cnt += bytes_used;
1094 region_size = void_diff(alloc_region->free_pointer,
1095 alloc_region->start_addr);
1096 bytes_allocated += region_size;
1097 generations[gc_alloc_generation].bytes_allocated += region_size;
1099 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1101 /* Set the generations alloc restart page to the last page of
1103 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1105 /* Add the region to the new_areas if requested. */
1106 if (BOXED_PAGE_FLAG & page_type_flag)
1107 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1111 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1113 gc_alloc_generation));
1116 /* There are no bytes allocated. Unallocate the first_page if
1117 * there are 0 bytes_used. */
1118 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1119 if (page_table[first_page].bytes_used == 0)
1120 page_table[first_page].allocated = FREE_PAGE_FLAG;
1123 /* Unallocate any unused pages. */
1124 while (next_page <= alloc_region->last_page) {
1125 gc_assert(page_table[next_page].bytes_used == 0);
1126 page_table[next_page].allocated = FREE_PAGE_FLAG;
1129 ret = thread_mutex_unlock(&free_pages_lock);
1130 gc_assert(ret == 0);
1132 /* alloc_region is per-thread, we're ok to do this unlocked */
1133 gc_set_region_empty(alloc_region);
1136 static inline void *gc_quick_alloc(word_t nbytes);
1138 /* Allocate a possibly large object. */
1140 gc_alloc_large(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
1143 page_index_t first_page, next_page, last_page;
1144 page_bytes_t orig_first_page_bytes_used;
1145 os_vm_size_t byte_cnt;
1146 os_vm_size_t bytes_used;
1149 ret = thread_mutex_lock(&free_pages_lock);
1150 gc_assert(ret == 0);
1152 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1153 if (first_page <= alloc_region->last_page) {
1154 first_page = alloc_region->last_page+1;
1157 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1159 gc_assert(first_page > alloc_region->last_page);
1161 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1163 /* Set up the pages. */
1164 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1166 /* If the first page was free then set up the gen, and
1167 * scan_start_offset. */
1168 if (page_table[first_page].bytes_used == 0) {
1169 page_table[first_page].allocated = page_type_flag;
1170 page_table[first_page].gen = gc_alloc_generation;
1171 page_table[first_page].scan_start_offset = 0;
1172 page_table[first_page].large_object = 1;
1175 gc_assert(page_table[first_page].allocated == page_type_flag);
1176 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1177 gc_assert(page_table[first_page].large_object == 1);
1181 /* Calc. the number of bytes used in this page. This is not
1182 * always the number of new bytes, unless it was free. */
1184 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1185 bytes_used = GENCGC_CARD_BYTES;
1188 page_table[first_page].bytes_used = bytes_used;
1189 byte_cnt += bytes_used;
1191 next_page = first_page+1;
1193 /* All the rest of the pages should be free. We need to set their
1194 * scan_start_offset pointer to the start of the region, and set
1195 * the bytes_used. */
1197 gc_assert(page_free_p(next_page));
1198 gc_assert(page_table[next_page].bytes_used == 0);
1199 page_table[next_page].allocated = page_type_flag;
1200 page_table[next_page].gen = gc_alloc_generation;
1201 page_table[next_page].large_object = 1;
1203 page_table[next_page].scan_start_offset =
1204 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1206 /* Calculate the number of bytes used in this page. */
1208 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1209 if (bytes_used > GENCGC_CARD_BYTES) {
1210 bytes_used = GENCGC_CARD_BYTES;
1213 page_table[next_page].bytes_used = bytes_used;
1214 page_table[next_page].write_protected=0;
1215 page_table[next_page].dont_move=0;
1216 byte_cnt += bytes_used;
1220 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1222 bytes_allocated += nbytes;
1223 generations[gc_alloc_generation].bytes_allocated += nbytes;
1225 /* Add the region to the new_areas if requested. */
1226 if (BOXED_PAGE_FLAG & page_type_flag)
1227 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1229 /* Bump up last_free_page */
1230 if (last_page+1 > last_free_page) {
1231 last_free_page = last_page+1;
1232 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1234 ret = thread_mutex_unlock(&free_pages_lock);
1235 gc_assert(ret == 0);
1237 #ifdef READ_PROTECT_FREE_PAGES
1238 os_protect(page_address(first_page),
1239 npage_bytes(1+last_page-first_page),
1243 zero_dirty_pages(first_page, last_page);
1245 return page_address(first_page);
1248 static page_index_t gencgc_alloc_start_page = -1;
1251 gc_heap_exhausted_error_or_lose (sword_t available, sword_t requested)
1253 struct thread *thread = arch_os_get_current_thread();
1254 /* Write basic information before doing anything else: if we don't
1255 * call to lisp this is a must, and even if we do there is always
1256 * the danger that we bounce back here before the error has been
1257 * handled, or indeed even printed.
1259 report_heap_exhaustion(available, requested, thread);
1260 if (gc_active_p || (available == 0)) {
1261 /* If we are in GC, or totally out of memory there is no way
1262 * to sanely transfer control to the lisp-side of things.
1264 lose("Heap exhausted, game over.");
1267 /* FIXME: assert free_pages_lock held */
1268 (void)thread_mutex_unlock(&free_pages_lock);
1269 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
1270 gc_assert(get_pseudo_atomic_atomic(thread));
1271 clear_pseudo_atomic_atomic(thread);
1272 if (get_pseudo_atomic_interrupted(thread))
1273 do_pending_interrupt();
1275 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1276 * to running user code at arbitrary places, even in a
1277 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1278 * running out of the heap. So at this point all bets are
1280 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1281 corruption_warning_and_maybe_lose
1282 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1283 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1284 alloc_number(available), alloc_number(requested));
1285 lose("HEAP-EXHAUSTED-ERROR fell through");
1290 gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t bytes,
1293 page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
1294 page_index_t first_page, last_page, restart_page = *restart_page_ptr;
1295 os_vm_size_t nbytes = bytes;
1296 os_vm_size_t nbytes_goal = nbytes;
1297 os_vm_size_t bytes_found = 0;
1298 os_vm_size_t most_bytes_found = 0;
1299 boolean small_object = nbytes < GENCGC_CARD_BYTES;
1300 /* FIXME: assert(free_pages_lock is held); */
1302 if (nbytes_goal < gencgc_alloc_granularity)
1303 nbytes_goal = gencgc_alloc_granularity;
1305 /* Toggled by gc_and_save for heap compaction, normally -1. */
1306 if (gencgc_alloc_start_page != -1) {
1307 restart_page = gencgc_alloc_start_page;
1310 /* FIXME: This is on bytes instead of nbytes pending cleanup of
1311 * long from the interface. */
1312 gc_assert(bytes>=0);
1313 /* Search for a page with at least nbytes of space. We prefer
1314 * not to split small objects on multiple pages, to reduce the
1315 * number of contiguous allocation regions spaning multiple
1316 * pages: this helps avoid excessive conservativism.
1318 * For other objects, we guarantee that they start on their own
1321 first_page = restart_page;
1322 while (first_page < page_table_pages) {
1324 if (page_free_p(first_page)) {
1325 gc_assert(0 == page_table[first_page].bytes_used);
1326 bytes_found = GENCGC_CARD_BYTES;
1327 } else if (small_object &&
1328 (page_table[first_page].allocated == page_type_flag) &&
1329 (page_table[first_page].large_object == 0) &&
1330 (page_table[first_page].gen == gc_alloc_generation) &&
1331 (page_table[first_page].write_protected == 0) &&
1332 (page_table[first_page].dont_move == 0)) {
1333 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1334 if (bytes_found < nbytes) {
1335 if (bytes_found > most_bytes_found)
1336 most_bytes_found = bytes_found;
1345 gc_assert(page_table[first_page].write_protected == 0);
1346 for (last_page = first_page+1;
1347 ((last_page < page_table_pages) &&
1348 page_free_p(last_page) &&
1349 (bytes_found < nbytes_goal));
1351 bytes_found += GENCGC_CARD_BYTES;
1352 gc_assert(0 == page_table[last_page].bytes_used);
1353 gc_assert(0 == page_table[last_page].write_protected);
1356 if (bytes_found > most_bytes_found) {
1357 most_bytes_found = bytes_found;
1358 most_bytes_found_from = first_page;
1359 most_bytes_found_to = last_page;
1361 if (bytes_found >= nbytes_goal)
1364 first_page = last_page;
1367 bytes_found = most_bytes_found;
1368 restart_page = first_page + 1;
1370 /* Check for a failure */
1371 if (bytes_found < nbytes) {
1372 gc_assert(restart_page >= page_table_pages);
1373 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1376 gc_assert(most_bytes_found_to);
1377 *restart_page_ptr = most_bytes_found_from;
1378 return most_bytes_found_to-1;
1381 /* Allocate bytes. All the rest of the special-purpose allocation
1382 * functions will eventually call this */
1385 gc_alloc_with_region(sword_t nbytes,int page_type_flag, struct alloc_region *my_region,
1388 void *new_free_pointer;
1390 if (nbytes>=large_object_size)
1391 return gc_alloc_large(nbytes, page_type_flag, my_region);
1393 /* Check whether there is room in the current alloc region. */
1394 new_free_pointer = my_region->free_pointer + nbytes;
1396 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1397 my_region->free_pointer, new_free_pointer); */
1399 if (new_free_pointer <= my_region->end_addr) {
1400 /* If so then allocate from the current alloc region. */
1401 void *new_obj = my_region->free_pointer;
1402 my_region->free_pointer = new_free_pointer;
1404 /* Unless a `quick' alloc was requested, check whether the
1405 alloc region is almost empty. */
1407 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1408 /* If so, finished with the current region. */
1409 gc_alloc_update_page_tables(page_type_flag, my_region);
1410 /* Set up a new region. */
1411 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1414 return((void *)new_obj);
1417 /* Else not enough free space in the current region: retry with a
1420 gc_alloc_update_page_tables(page_type_flag, my_region);
1421 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1422 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1425 /* these are only used during GC: all allocation from the mutator calls
1426 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1429 static inline void *
1430 gc_quick_alloc(word_t nbytes)
1432 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1435 static inline void *
1436 gc_alloc_unboxed(word_t nbytes)
1438 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1441 static inline void *
1442 gc_quick_alloc_unboxed(word_t nbytes)
1444 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1447 /* Copy a large object. If the object is in a large object region then
1448 * it is simply promoted, else it is copied. If it's large enough then
1449 * it's copied to a large object region.
1451 * Bignums and vectors may have shrunk. If the object is not copied
1452 * the space needs to be reclaimed, and the page_tables corrected. */
1454 general_copy_large_object(lispobj object, word_t nwords, boolean boxedp)
1458 page_index_t first_page;
1460 gc_assert(is_lisp_pointer(object));
1461 gc_assert(from_space_p(object));
1462 gc_assert((nwords & 0x01) == 0);
1464 if ((nwords > 1024*1024) && gencgc_verbose) {
1465 FSHOW((stderr, "/general_copy_large_object: %d bytes\n",
1466 nwords*N_WORD_BYTES));
1469 /* Check whether it's a large object. */
1470 first_page = find_page_index((void *)object);
1471 gc_assert(first_page >= 0);
1473 if (page_table[first_page].large_object) {
1474 /* Promote the object. Note: Unboxed objects may have been
1475 * allocated to a BOXED region so it may be necessary to
1476 * change the region to UNBOXED. */
1477 os_vm_size_t remaining_bytes;
1478 os_vm_size_t bytes_freed;
1479 page_index_t next_page;
1480 page_bytes_t old_bytes_used;
1482 /* FIXME: This comment is somewhat stale.
1484 * Note: Any page write-protection must be removed, else a
1485 * later scavenge_newspace may incorrectly not scavenge these
1486 * pages. This would not be necessary if they are added to the
1487 * new areas, but let's do it for them all (they'll probably
1488 * be written anyway?). */
1490 gc_assert(page_starts_contiguous_block_p(first_page));
1491 next_page = first_page;
1492 remaining_bytes = nwords*N_WORD_BYTES;
1494 while (remaining_bytes > GENCGC_CARD_BYTES) {
1495 gc_assert(page_table[next_page].gen == from_space);
1496 gc_assert(page_table[next_page].large_object);
1497 gc_assert(page_table[next_page].scan_start_offset ==
1498 npage_bytes(next_page-first_page));
1499 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1500 /* Should have been unprotected by unprotect_oldspace()
1501 * for boxed objects, and after promotion unboxed ones
1502 * should not be on protected pages at all. */
1503 gc_assert(!page_table[next_page].write_protected);
1506 gc_assert(page_boxed_p(next_page));
1508 gc_assert(page_allocated_no_region_p(next_page));
1509 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1511 page_table[next_page].gen = new_space;
1513 remaining_bytes -= GENCGC_CARD_BYTES;
1517 /* Now only one page remains, but the object may have shrunk so
1518 * there may be more unused pages which will be freed. */
1520 /* Object may have shrunk but shouldn't have grown - check. */
1521 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1523 page_table[next_page].gen = new_space;
1526 gc_assert(page_boxed_p(next_page));
1528 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1530 /* Adjust the bytes_used. */
1531 old_bytes_used = page_table[next_page].bytes_used;
1532 page_table[next_page].bytes_used = remaining_bytes;
1534 bytes_freed = old_bytes_used - remaining_bytes;
1536 /* Free any remaining pages; needs care. */
1538 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1539 (page_table[next_page].gen == from_space) &&
1540 /* FIXME: It is not obvious to me why this is necessary
1541 * as a loop condition: it seems to me that the
1542 * scan_start_offset test should be sufficient, but
1543 * experimentally that is not the case. --NS
1546 page_boxed_p(next_page) :
1547 page_allocated_no_region_p(next_page)) &&
1548 page_table[next_page].large_object &&
1549 (page_table[next_page].scan_start_offset ==
1550 npage_bytes(next_page - first_page))) {
1551 /* Checks out OK, free the page. Don't need to both zeroing
1552 * pages as this should have been done before shrinking the
1553 * object. These pages shouldn't be write-protected, even if
1554 * boxed they should be zero filled. */
1555 gc_assert(page_table[next_page].write_protected == 0);
1557 old_bytes_used = page_table[next_page].bytes_used;
1558 page_table[next_page].allocated = FREE_PAGE_FLAG;
1559 page_table[next_page].bytes_used = 0;
1560 bytes_freed += old_bytes_used;
1564 if ((bytes_freed > 0) && gencgc_verbose) {
1566 "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\n",
1570 generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES
1572 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1573 bytes_allocated -= bytes_freed;
1575 /* Add the region to the new_areas if requested. */
1577 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1582 /* Get tag of object. */
1583 tag = lowtag_of(object);
1585 /* Allocate space. */
1586 new = gc_general_alloc(nwords*N_WORD_BYTES,
1587 (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG),
1590 /* Copy the object. */
1591 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1593 /* Return Lisp pointer of new object. */
1594 return ((lispobj) new) | tag;
1599 copy_large_object(lispobj object, sword_t nwords)
1601 return general_copy_large_object(object, nwords, 1);
1605 copy_large_unboxed_object(lispobj object, sword_t nwords)
1607 return general_copy_large_object(object, nwords, 0);
1610 /* to copy unboxed objects */
1612 copy_unboxed_object(lispobj object, sword_t nwords)
1614 return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG);
1619 * code and code-related objects
1622 static lispobj trans_fun_header(lispobj object);
1623 static lispobj trans_boxed(lispobj object);
1626 /* Scan a x86 compiled code object, looking for possible fixups that
1627 * have been missed after a move.
1629 * Two types of fixups are needed:
1630 * 1. Absolute fixups to within the code object.
1631 * 2. Relative fixups to outside the code object.
1633 * Currently only absolute fixups to the constant vector, or to the
1634 * code area are checked. */
1636 sniff_code_object(struct code *code, os_vm_size_t displacement)
1638 #ifdef LISP_FEATURE_X86
1639 sword_t nheader_words, ncode_words, nwords;
1640 os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
1641 os_vm_address_t code_start_addr, code_end_addr;
1642 os_vm_address_t code_addr = (os_vm_address_t)code;
1643 int fixup_found = 0;
1645 if (!check_code_fixups)
1648 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1650 ncode_words = fixnum_value(code->code_size);
1651 nheader_words = HeaderValue(*(lispobj *)code);
1652 nwords = ncode_words + nheader_words;
1654 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1655 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1656 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1657 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1659 /* Work through the unboxed code. */
1660 for (p = code_start_addr; p < code_end_addr; p++) {
1661 void *data = *(void **)p;
1662 unsigned d1 = *((unsigned char *)p - 1);
1663 unsigned d2 = *((unsigned char *)p - 2);
1664 unsigned d3 = *((unsigned char *)p - 3);
1665 unsigned d4 = *((unsigned char *)p - 4);
1667 unsigned d5 = *((unsigned char *)p - 5);
1668 unsigned d6 = *((unsigned char *)p - 6);
1671 /* Check for code references. */
1672 /* Check for a 32 bit word that looks like an absolute
1673 reference to within the code adea of the code object. */
1674 if ((data >= (void*)(code_start_addr-displacement))
1675 && (data < (void*)(code_end_addr-displacement))) {
1676 /* function header */
1678 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1680 /* Skip the function header */
1684 /* the case of PUSH imm32 */
1688 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1689 p, d6, d5, d4, d3, d2, d1, data));
1690 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1692 /* the case of MOV [reg-8],imm32 */
1694 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1695 || d2==0x45 || d2==0x46 || d2==0x47)
1699 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1700 p, d6, d5, d4, d3, d2, d1, data));
1701 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1703 /* the case of LEA reg,[disp32] */
1704 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1707 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1708 p, d6, d5, d4, d3, d2, d1, data));
1709 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1713 /* Check for constant references. */
1714 /* Check for a 32 bit word that looks like an absolute
1715 reference to within the constant vector. Constant references
1717 if ((data >= (void*)(constants_start_addr-displacement))
1718 && (data < (void*)(constants_end_addr-displacement))
1719 && (((unsigned)data & 0x3) == 0)) {
1724 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1725 p, d6, d5, d4, d3, d2, d1, data));
1726 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1729 /* the case of MOV m32,EAX */
1733 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1734 p, d6, d5, d4, d3, d2, d1, data));
1735 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1738 /* the case of CMP m32,imm32 */
1739 if ((d1 == 0x3d) && (d2 == 0x81)) {
1742 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1743 p, d6, d5, d4, d3, d2, d1, data));
1745 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1748 /* Check for a mod=00, r/m=101 byte. */
1749 if ((d1 & 0xc7) == 5) {
1754 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1755 p, d6, d5, d4, d3, d2, d1, data));
1756 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1758 /* the case of CMP reg32,m32 */
1762 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1763 p, d6, d5, d4, d3, d2, d1, data));
1764 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1766 /* the case of MOV m32,reg32 */
1770 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1771 p, d6, d5, d4, d3, d2, d1, data));
1772 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1774 /* the case of MOV reg32,m32 */
1778 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1779 p, d6, d5, d4, d3, d2, d1, data));
1780 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1782 /* the case of LEA reg32,m32 */
1786 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1787 p, d6, d5, d4, d3, d2, d1, data));
1788 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1794 /* If anything was found, print some information on the code
1798 "/compiled code object at %x: header words = %d, code words = %d\n",
1799 code, nheader_words, ncode_words));
1801 "/const start = %x, end = %x\n",
1802 constants_start_addr, constants_end_addr));
1804 "/code start = %x, end = %x\n",
1805 code_start_addr, code_end_addr));
1811 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1813 /* x86-64 uses pc-relative addressing instead of this kludge */
1814 #ifndef LISP_FEATURE_X86_64
1815 sword_t nheader_words, ncode_words, nwords;
1816 os_vm_address_t constants_start_addr, constants_end_addr;
1817 os_vm_address_t code_start_addr, code_end_addr;
1818 os_vm_address_t code_addr = (os_vm_address_t)new_code;
1819 os_vm_address_t old_addr = (os_vm_address_t)old_code;
1820 os_vm_size_t displacement = code_addr - old_addr;
1821 lispobj fixups = NIL;
1822 struct vector *fixups_vector;
1824 ncode_words = fixnum_value(new_code->code_size);
1825 nheader_words = HeaderValue(*(lispobj *)new_code);
1826 nwords = ncode_words + nheader_words;
1828 "/compiled code object at %x: header words = %d, code words = %d\n",
1829 new_code, nheader_words, ncode_words)); */
1830 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1831 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1832 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1833 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1836 "/const start = %x, end = %x\n",
1837 constants_start_addr,constants_end_addr));
1839 "/code start = %x; end = %x\n",
1840 code_start_addr,code_end_addr));
1843 /* The first constant should be a pointer to the fixups for this
1844 code objects. Check. */
1845 fixups = new_code->constants[0];
1847 /* It will be 0 or the unbound-marker if there are no fixups (as
1848 * will be the case if the code object has been purified, for
1849 * example) and will be an other pointer if it is valid. */
1850 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1851 !is_lisp_pointer(fixups)) {
1852 /* Check for possible errors. */
1853 if (check_code_fixups)
1854 sniff_code_object(new_code, displacement);
1859 fixups_vector = (struct vector *)native_pointer(fixups);
1861 /* Could be pointing to a forwarding pointer. */
1862 /* FIXME is this always in from_space? if so, could replace this code with
1863 * forwarding_pointer_p/forwarding_pointer_value */
1864 if (is_lisp_pointer(fixups) &&
1865 (find_page_index((void*)fixups_vector) != -1) &&
1866 (fixups_vector->header == 0x01)) {
1867 /* If so, then follow it. */
1868 /*SHOW("following pointer to a forwarding pointer");*/
1870 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1873 /*SHOW("got fixups");*/
1875 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1876 /* Got the fixups for the code block. Now work through the vector,
1877 and apply a fixup at each address. */
1878 sword_t length = fixnum_value(fixups_vector->length);
1880 for (i = 0; i < length; i++) {
1881 long offset = fixups_vector->data[i];
1882 /* Now check the current value of offset. */
1883 os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);
1885 /* If it's within the old_code object then it must be an
1886 * absolute fixup (relative ones are not saved) */
1887 if ((old_value >= old_addr)
1888 && (old_value < (old_addr + nwords*N_WORD_BYTES)))
1889 /* So add the dispacement. */
1890 *(os_vm_address_t *)(code_start_addr + offset) =
1891 old_value + displacement;
1893 /* It is outside the old code object so it must be a
1894 * relative fixup (absolute fixups are not saved). So
1895 * subtract the displacement. */
1896 *(os_vm_address_t *)(code_start_addr + offset) =
1897 old_value - displacement;
1900 /* This used to just print a note to stderr, but a bogus fixup seems to
1901 * indicate real heap corruption, so a hard hailure is in order. */
1902 lose("fixup vector %p has a bad widetag: %d\n",
1903 fixups_vector, widetag_of(fixups_vector->header));
1906 /* Check for possible errors. */
1907 if (check_code_fixups) {
1908 sniff_code_object(new_code,displacement);
1915 trans_boxed_large(lispobj object)
1920 gc_assert(is_lisp_pointer(object));
1922 header = *((lispobj *) native_pointer(object));
1923 length = HeaderValue(header) + 1;
1924 length = CEILING(length, 2);
1926 return copy_large_object(object, length);
1929 /* Doesn't seem to be used, delete it after the grace period. */
1932 trans_unboxed_large(lispobj object)
1937 gc_assert(is_lisp_pointer(object));
1939 header = *((lispobj *) native_pointer(object));
1940 length = HeaderValue(header) + 1;
1941 length = CEILING(length, 2);
1943 return copy_large_unboxed_object(object, length);
1951 /* XX This is a hack adapted from cgc.c. These don't work too
1952 * efficiently with the gencgc as a list of the weak pointers is
1953 * maintained within the objects which causes writes to the pages. A
1954 * limited attempt is made to avoid unnecessary writes, but this needs
1956 #define WEAK_POINTER_NWORDS \
1957 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
1960 scav_weak_pointer(lispobj *where, lispobj object)
1962 /* Since we overwrite the 'next' field, we have to make
1963 * sure not to do so for pointers already in the list.
1964 * Instead of searching the list of weak_pointers each
1965 * time, we ensure that next is always NULL when the weak
1966 * pointer isn't in the list, and not NULL otherwise.
1967 * Since we can't use NULL to denote end of list, we
1968 * use a pointer back to the same weak_pointer.
1970 struct weak_pointer * wp = (struct weak_pointer*)where;
1972 if (NULL == wp->next) {
1973 wp->next = weak_pointers;
1975 if (NULL == wp->next)
1979 /* Do not let GC scavenge the value slot of the weak pointer.
1980 * (That is why it is a weak pointer.) */
1982 return WEAK_POINTER_NWORDS;
1987 search_read_only_space(void *pointer)
1989 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
1990 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
1991 if ((pointer < (void *)start) || (pointer >= (void *)end))
1993 return (gc_search_space(start,
1994 (((lispobj *)pointer)+2)-start,
1995 (lispobj *) pointer));
1999 search_static_space(void *pointer)
2001 lispobj *start = (lispobj *)STATIC_SPACE_START;
2002 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2003 if ((pointer < (void *)start) || (pointer >= (void *)end))
2005 return (gc_search_space(start,
2006 (((lispobj *)pointer)+2)-start,
2007 (lispobj *) pointer));
2010 /* a faster version for searching the dynamic space. This will work even
2011 * if the object is in a current allocation region. */
2013 search_dynamic_space(void *pointer)
2015 page_index_t page_index = find_page_index(pointer);
2018 /* The address may be invalid, so do some checks. */
2019 if ((page_index == -1) || page_free_p(page_index))
2021 start = (lispobj *)page_scan_start(page_index);
2022 return (gc_search_space(start,
2023 (((lispobj *)pointer)+2)-start,
2024 (lispobj *)pointer));
2027 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2029 /* Is there any possibility that pointer is a valid Lisp object
2030 * reference, and/or something else (e.g. subroutine call return
2031 * address) which should prevent us from moving the referred-to thing?
2032 * This is called from preserve_pointers() */
2034 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2036 lispobj *start_addr;
2038 /* Find the object start address. */
2039 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2043 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2046 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2048 /* Adjust large bignum and vector objects. This will adjust the
2049 * allocated region if the size has shrunk, and move unboxed objects
2050 * into unboxed pages. The pages are not promoted here, and the
2051 * promoted region is not added to the new_regions; this is really
2052 * only designed to be called from preserve_pointer(). Shouldn't fail
2053 * if this is missed, just may delay the moving of objects to unboxed
2054 * pages, and the freeing of pages. */
2056 maybe_adjust_large_object(lispobj *where)
2058 page_index_t first_page;
2059 page_index_t next_page;
2062 uword_t remaining_bytes;
2063 uword_t bytes_freed;
2064 uword_t old_bytes_used;
2068 /* Check whether it's a vector or bignum object. */
2069 switch (widetag_of(where[0])) {
2070 case SIMPLE_VECTOR_WIDETAG:
2071 boxed = BOXED_PAGE_FLAG;
2073 case BIGNUM_WIDETAG:
2074 case SIMPLE_BASE_STRING_WIDETAG:
2075 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2076 case SIMPLE_CHARACTER_STRING_WIDETAG:
2078 case SIMPLE_BIT_VECTOR_WIDETAG:
2079 case SIMPLE_ARRAY_NIL_WIDETAG:
2080 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2081 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2082 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2083 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2084 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2085 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2087 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2089 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2090 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2091 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2092 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2094 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2095 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2097 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2098 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2100 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2101 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2104 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2106 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2107 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2109 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2110 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2112 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2113 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2114 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2115 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2117 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2118 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2120 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2121 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2123 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2124 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2126 boxed = UNBOXED_PAGE_FLAG;
2132 /* Find its current size. */
2133 nwords = (sizetab[widetag_of(where[0])])(where);
2135 first_page = find_page_index((void *)where);
2136 gc_assert(first_page >= 0);
2138 /* Note: Any page write-protection must be removed, else a later
2139 * scavenge_newspace may incorrectly not scavenge these pages.
2140 * This would not be necessary if they are added to the new areas,
2141 * but lets do it for them all (they'll probably be written
2144 gc_assert(page_starts_contiguous_block_p(first_page));
2146 next_page = first_page;
2147 remaining_bytes = nwords*N_WORD_BYTES;
2148 while (remaining_bytes > GENCGC_CARD_BYTES) {
2149 gc_assert(page_table[next_page].gen == from_space);
2150 gc_assert(page_allocated_no_region_p(next_page));
2151 gc_assert(page_table[next_page].large_object);
2152 gc_assert(page_table[next_page].scan_start_offset ==
2153 npage_bytes(next_page-first_page));
2154 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2156 page_table[next_page].allocated = boxed;
2158 /* Shouldn't be write-protected at this stage. Essential that the
2160 gc_assert(!page_table[next_page].write_protected);
2161 remaining_bytes -= GENCGC_CARD_BYTES;
2165 /* Now only one page remains, but the object may have shrunk so
2166 * there may be more unused pages which will be freed. */
2168 /* Object may have shrunk but shouldn't have grown - check. */
2169 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2171 page_table[next_page].allocated = boxed;
2172 gc_assert(page_table[next_page].allocated ==
2173 page_table[first_page].allocated);
2175 /* Adjust the bytes_used. */
2176 old_bytes_used = page_table[next_page].bytes_used;
2177 page_table[next_page].bytes_used = remaining_bytes;
2179 bytes_freed = old_bytes_used - remaining_bytes;
2181 /* Free any remaining pages; needs care. */
2183 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2184 (page_table[next_page].gen == from_space) &&
2185 page_allocated_no_region_p(next_page) &&
2186 page_table[next_page].large_object &&
2187 (page_table[next_page].scan_start_offset ==
2188 npage_bytes(next_page - first_page))) {
2189 /* It checks out OK, free the page. We don't need to both zeroing
2190 * pages as this should have been done before shrinking the
2191 * object. These pages shouldn't be write protected as they
2192 * should be zero filled. */
2193 gc_assert(page_table[next_page].write_protected == 0);
2195 old_bytes_used = page_table[next_page].bytes_used;
2196 page_table[next_page].allocated = FREE_PAGE_FLAG;
2197 page_table[next_page].bytes_used = 0;
2198 bytes_freed += old_bytes_used;
2202 if ((bytes_freed > 0) && gencgc_verbose) {
2204 "/maybe_adjust_large_object() freed %d\n",
2208 generations[from_space].bytes_allocated -= bytes_freed;
2209 bytes_allocated -= bytes_freed;
2214 /* Take a possible pointer to a Lisp object and mark its page in the
2215 * page_table so that it will not be relocated during a GC.
2217 * This involves locating the page it points to, then backing up to
2218 * the start of its region, then marking all pages dont_move from there
2219 * up to the first page that's not full or has a different generation
2221 * It is assumed that all the page static flags have been cleared at
2222 * the start of a GC.
2224 * It is also assumed that the current gc_alloc() region has been
2225 * flushed and the tables updated. */
2228 preserve_pointer(void *addr)
2230 page_index_t addr_page_index = find_page_index(addr);
2231 page_index_t first_page;
2233 unsigned int region_allocation;
2235 /* quick check 1: Address is quite likely to have been invalid. */
2236 if ((addr_page_index == -1)
2237 || page_free_p(addr_page_index)
2238 || (page_table[addr_page_index].bytes_used == 0)
2239 || (page_table[addr_page_index].gen != from_space)
2240 /* Skip if already marked dont_move. */
2241 || (page_table[addr_page_index].dont_move != 0))
2243 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2244 /* (Now that we know that addr_page_index is in range, it's
2245 * safe to index into page_table[] with it.) */
2246 region_allocation = page_table[addr_page_index].allocated;
2248 /* quick check 2: Check the offset within the page.
2251 if (((uword_t)addr & (GENCGC_CARD_BYTES - 1)) >
2252 page_table[addr_page_index].bytes_used)
2255 /* Filter out anything which can't be a pointer to a Lisp object
2256 * (or, as a special case which also requires dont_move, a return
2257 * address referring to something in a CodeObject). This is
2258 * expensive but important, since it vastly reduces the
2259 * probability that random garbage will be bogusly interpreted as
2260 * a pointer which prevents a page from moving.
2262 * This only needs to happen on x86oids, where this is used for
2263 * conservative roots. Non-x86oid systems only ever call this
2264 * function on known-valid lisp objects. */
2265 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2266 if (!(code_page_p(addr_page_index)
2267 || (is_lisp_pointer((lispobj)addr) &&
2268 possibly_valid_dynamic_space_pointer(addr))))
2272 /* Find the beginning of the region. Note that there may be
2273 * objects in the region preceding the one that we were passed a
2274 * pointer to: if this is the case, we will write-protect all the
2275 * previous objects' pages too. */
2278 /* I think this'd work just as well, but without the assertions.
2279 * -dan 2004.01.01 */
2280 first_page = find_page_index(page_scan_start(addr_page_index))
2282 first_page = addr_page_index;
2283 while (!page_starts_contiguous_block_p(first_page)) {
2285 /* Do some checks. */
2286 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2287 gc_assert(page_table[first_page].gen == from_space);
2288 gc_assert(page_table[first_page].allocated == region_allocation);
2292 /* Adjust any large objects before promotion as they won't be
2293 * copied after promotion. */
2294 if (page_table[first_page].large_object) {
2295 maybe_adjust_large_object(page_address(first_page));
2296 /* If a large object has shrunk then addr may now point to a
2297 * free area in which case it's ignored here. Note it gets
2298 * through the valid pointer test above because the tail looks
2300 if (page_free_p(addr_page_index)
2301 || (page_table[addr_page_index].bytes_used == 0)
2302 /* Check the offset within the page. */
2303 || (((uword_t)addr & (GENCGC_CARD_BYTES - 1))
2304 > page_table[addr_page_index].bytes_used)) {
2306 "weird? ignore ptr 0x%x to freed area of large object\n",
2310 /* It may have moved to unboxed pages. */
2311 region_allocation = page_table[first_page].allocated;
2314 /* Now work forward until the end of this contiguous area is found,
2315 * marking all pages as dont_move. */
2316 for (i = first_page; ;i++) {
2317 gc_assert(page_table[i].allocated == region_allocation);
2319 /* Mark the page static. */
2320 page_table[i].dont_move = 1;
2322 /* It is essential that the pages are not write protected as
2323 * they may have pointers into the old-space which need
2324 * scavenging. They shouldn't be write protected at this
2326 gc_assert(!page_table[i].write_protected);
2328 /* Check whether this is the last page in this contiguous block.. */
2329 if ((page_table[i].bytes_used < GENCGC_CARD_BYTES)
2330 /* ..or it is CARD_BYTES and is the last in the block */
2332 || (page_table[i+1].bytes_used == 0) /* next page free */
2333 || (page_table[i+1].gen != from_space) /* diff. gen */
2334 || (page_starts_contiguous_block_p(i+1)))
2338 /* Check that the page is now static. */
2339 gc_assert(page_table[addr_page_index].dont_move != 0);
2342 /* If the given page is not write-protected, then scan it for pointers
2343 * to younger generations or the top temp. generation, if no
2344 * suspicious pointers are found then the page is write-protected.
2346 * Care is taken to check for pointers to the current gc_alloc()
2347 * region if it is a younger generation or the temp. generation. This
2348 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2349 * the gc_alloc_generation does not need to be checked as this is only
2350 * called from scavenge_generation() when the gc_alloc generation is
2351 * younger, so it just checks if there is a pointer to the current
2354 * We return 1 if the page was write-protected, else 0. */
2356 update_page_write_prot(page_index_t page)
2358 generation_index_t gen = page_table[page].gen;
2361 void **page_addr = (void **)page_address(page);
2362 sword_t num_words = page_table[page].bytes_used / N_WORD_BYTES;
2364 /* Shouldn't be a free page. */
2365 gc_assert(page_allocated_p(page));
2366 gc_assert(page_table[page].bytes_used != 0);
2368 /* Skip if it's already write-protected, pinned, or unboxed */
2369 if (page_table[page].write_protected
2370 /* FIXME: What's the reason for not write-protecting pinned pages? */
2371 || page_table[page].dont_move
2372 || page_unboxed_p(page))
2375 /* Scan the page for pointers to younger generations or the
2376 * top temp. generation. */
2378 for (j = 0; j < num_words; j++) {
2379 void *ptr = *(page_addr+j);
2380 page_index_t index = find_page_index(ptr);
2382 /* Check that it's in the dynamic space */
2384 if (/* Does it point to a younger or the temp. generation? */
2385 (page_allocated_p(index)
2386 && (page_table[index].bytes_used != 0)
2387 && ((page_table[index].gen < gen)
2388 || (page_table[index].gen == SCRATCH_GENERATION)))
2390 /* Or does it point within a current gc_alloc() region? */
2391 || ((boxed_region.start_addr <= ptr)
2392 && (ptr <= boxed_region.free_pointer))
2393 || ((unboxed_region.start_addr <= ptr)
2394 && (ptr <= unboxed_region.free_pointer))) {
2401 /* Write-protect the page. */
2402 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2404 os_protect((void *)page_addr,
2406 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2408 /* Note the page as protected in the page tables. */
2409 page_table[page].write_protected = 1;
2415 /* Scavenge all generations from FROM to TO, inclusive, except for
2416 * new_space which needs special handling, as new objects may be
2417 * added which are not checked here - use scavenge_newspace generation.
2419 * Write-protected pages should not have any pointers to the
2420 * from_space so do need scavenging; thus write-protected pages are
2421 * not always scavenged. There is some code to check that these pages
2422 * are not written; but to check fully the write-protected pages need
2423 * to be scavenged by disabling the code to skip them.
2425 * Under the current scheme when a generation is GCed the younger
2426 * generations will be empty. So, when a generation is being GCed it
2427 * is only necessary to scavenge the older generations for pointers
2428 * not the younger. So a page that does not have pointers to younger
2429 * generations does not need to be scavenged.
2431 * The write-protection can be used to note pages that don't have
2432 * pointers to younger pages. But pages can be written without having
2433 * pointers to younger generations. After the pages are scavenged here
2434 * they can be scanned for pointers to younger generations and if
2435 * there are none the page can be write-protected.
2437 * One complication is when the newspace is the top temp. generation.
2439 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2440 * that none were written, which they shouldn't be as they should have
2441 * no pointers to younger generations. This breaks down for weak
2442 * pointers as the objects contain a link to the next and are written
2443 * if a weak pointer is scavenged. Still it's a useful check. */
2445 scavenge_generations(generation_index_t from, generation_index_t to)
2448 page_index_t num_wp = 0;
2452 /* Clear the write_protected_cleared flags on all pages. */
2453 for (i = 0; i < page_table_pages; i++)
2454 page_table[i].write_protected_cleared = 0;
2457 for (i = 0; i < last_free_page; i++) {
2458 generation_index_t generation = page_table[i].gen;
2460 && (page_table[i].bytes_used != 0)
2461 && (generation != new_space)
2462 && (generation >= from)
2463 && (generation <= to)) {
2464 page_index_t last_page,j;
2465 int write_protected=1;
2467 /* This should be the start of a region */
2468 gc_assert(page_starts_contiguous_block_p(i));
2470 /* Now work forward until the end of the region */
2471 for (last_page = i; ; last_page++) {
2473 write_protected && page_table[last_page].write_protected;
2474 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2475 /* Or it is CARD_BYTES and is the last in the block */
2476 || (!page_boxed_p(last_page+1))
2477 || (page_table[last_page+1].bytes_used == 0)
2478 || (page_table[last_page+1].gen != generation)
2479 || (page_starts_contiguous_block_p(last_page+1)))
2482 if (!write_protected) {
2483 scavenge(page_address(i),
2484 ((uword_t)(page_table[last_page].bytes_used
2485 + npage_bytes(last_page-i)))
2488 /* Now scan the pages and write protect those that
2489 * don't have pointers to younger generations. */
2490 if (enable_page_protection) {
2491 for (j = i; j <= last_page; j++) {
2492 num_wp += update_page_write_prot(j);
2495 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2497 "/write protected %d pages within generation %d\n",
2498 num_wp, generation));
2506 /* Check that none of the write_protected pages in this generation
2507 * have been written to. */
2508 for (i = 0; i < page_table_pages; i++) {
2509 if (page_allocated_p(i)
2510 && (page_table[i].bytes_used != 0)
2511 && (page_table[i].gen == generation)
2512 && (page_table[i].write_protected_cleared != 0)) {
2513 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2515 "/page bytes_used=%d scan_start_offset=%lu dont_move=%d\n",
2516 page_table[i].bytes_used,
2517 page_table[i].scan_start_offset,
2518 page_table[i].dont_move));
2519 lose("write to protected page %d in scavenge_generation()\n", i);
2526 /* Scavenge a newspace generation. As it is scavenged new objects may
2527 * be allocated to it; these will also need to be scavenged. This
2528 * repeats until there are no more objects unscavenged in the
2529 * newspace generation.
2531 * To help improve the efficiency, areas written are recorded by
2532 * gc_alloc() and only these scavenged. Sometimes a little more will be
2533 * scavenged, but this causes no harm. An easy check is done that the
2534 * scavenged bytes equals the number allocated in the previous
2537 * Write-protected pages are not scanned except if they are marked
2538 * dont_move in which case they may have been promoted and still have
2539 * pointers to the from space.
2541 * Write-protected pages could potentially be written by alloc however
2542 * to avoid having to handle re-scavenging of write-protected pages
2543 * gc_alloc() does not write to write-protected pages.
2545 * New areas of objects allocated are recorded alternatively in the two
2546 * new_areas arrays below. */
2547 static struct new_area new_areas_1[NUM_NEW_AREAS];
2548 static struct new_area new_areas_2[NUM_NEW_AREAS];
2550 /* Do one full scan of the new space generation. This is not enough to
2551 * complete the job as new objects may be added to the generation in
2552 * the process which are not scavenged. */
2554 scavenge_newspace_generation_one_scan(generation_index_t generation)
2559 "/starting one full scan of newspace generation %d\n",
2561 for (i = 0; i < last_free_page; i++) {
2562 /* Note that this skips over open regions when it encounters them. */
2564 && (page_table[i].bytes_used != 0)
2565 && (page_table[i].gen == generation)
2566 && ((page_table[i].write_protected == 0)
2567 /* (This may be redundant as write_protected is now
2568 * cleared before promotion.) */
2569 || (page_table[i].dont_move == 1))) {
2570 page_index_t last_page;
2573 /* The scavenge will start at the scan_start_offset of
2576 * We need to find the full extent of this contiguous
2577 * block in case objects span pages.
2579 * Now work forward until the end of this contiguous area
2580 * is found. A small area is preferred as there is a
2581 * better chance of its pages being write-protected. */
2582 for (last_page = i; ;last_page++) {
2583 /* If all pages are write-protected and movable,
2584 * then no need to scavenge */
2585 all_wp=all_wp && page_table[last_page].write_protected &&
2586 !page_table[last_page].dont_move;
2588 /* Check whether this is the last page in this
2589 * contiguous block */
2590 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
2591 /* Or it is CARD_BYTES and is the last in the block */
2592 || (!page_boxed_p(last_page+1))
2593 || (page_table[last_page+1].bytes_used == 0)
2594 || (page_table[last_page+1].gen != generation)
2595 || (page_starts_contiguous_block_p(last_page+1)))
2599 /* Do a limited check for write-protected pages. */
2601 sword_t nwords = (((uword_t)
2602 (page_table[last_page].bytes_used
2603 + npage_bytes(last_page-i)
2604 + page_table[i].scan_start_offset))
2606 new_areas_ignore_page = last_page;
2608 scavenge(page_scan_start(i), nwords);
2615 "/done with one full scan of newspace generation %d\n",
2619 /* Do a complete scavenge of the newspace generation. */
2621 scavenge_newspace_generation(generation_index_t generation)
2625 /* the new_areas array currently being written to by gc_alloc() */
2626 struct new_area (*current_new_areas)[] = &new_areas_1;
2627 size_t current_new_areas_index;
2629 /* the new_areas created by the previous scavenge cycle */
2630 struct new_area (*previous_new_areas)[] = NULL;
2631 size_t previous_new_areas_index;
2633 /* Flush the current regions updating the tables. */
2634 gc_alloc_update_all_page_tables();
2636 /* Turn on the recording of new areas by gc_alloc(). */
2637 new_areas = current_new_areas;
2638 new_areas_index = 0;
2640 /* Don't need to record new areas that get scavenged anyway during
2641 * scavenge_newspace_generation_one_scan. */
2642 record_new_objects = 1;
2644 /* Start with a full scavenge. */
2645 scavenge_newspace_generation_one_scan(generation);
2647 /* Record all new areas now. */
2648 record_new_objects = 2;
2650 /* Give a chance to weak hash tables to make other objects live.
2651 * FIXME: The algorithm implemented here for weak hash table gcing
2652 * is O(W^2+N) as Bruno Haible warns in
2653 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
2654 * see "Implementation 2". */
2655 scav_weak_hash_tables();
2657 /* Flush the current regions updating the tables. */
2658 gc_alloc_update_all_page_tables();
2660 /* Grab new_areas_index. */
2661 current_new_areas_index = new_areas_index;
2664 "The first scan is finished; current_new_areas_index=%d.\n",
2665 current_new_areas_index));*/
2667 while (current_new_areas_index > 0) {
2668 /* Move the current to the previous new areas */
2669 previous_new_areas = current_new_areas;
2670 previous_new_areas_index = current_new_areas_index;
2672 /* Scavenge all the areas in previous new areas. Any new areas
2673 * allocated are saved in current_new_areas. */
2675 /* Allocate an array for current_new_areas; alternating between
2676 * new_areas_1 and 2 */
2677 if (previous_new_areas == &new_areas_1)
2678 current_new_areas = &new_areas_2;
2680 current_new_areas = &new_areas_1;
2682 /* Set up for gc_alloc(). */
2683 new_areas = current_new_areas;
2684 new_areas_index = 0;
2686 /* Check whether previous_new_areas had overflowed. */
2687 if (previous_new_areas_index >= NUM_NEW_AREAS) {
2689 /* New areas of objects allocated have been lost so need to do a
2690 * full scan to be sure! If this becomes a problem try
2691 * increasing NUM_NEW_AREAS. */
2692 if (gencgc_verbose) {
2693 SHOW("new_areas overflow, doing full scavenge");
2696 /* Don't need to record new areas that get scavenged
2697 * anyway during scavenge_newspace_generation_one_scan. */
2698 record_new_objects = 1;
2700 scavenge_newspace_generation_one_scan(generation);
2702 /* Record all new areas now. */
2703 record_new_objects = 2;
2705 scav_weak_hash_tables();
2707 /* Flush the current regions updating the tables. */
2708 gc_alloc_update_all_page_tables();
2712 /* Work through previous_new_areas. */
2713 for (i = 0; i < previous_new_areas_index; i++) {
2714 page_index_t page = (*previous_new_areas)[i].page;
2715 size_t offset = (*previous_new_areas)[i].offset;
2716 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
2717 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
2718 scavenge(page_address(page)+offset, size);
2721 scav_weak_hash_tables();
2723 /* Flush the current regions updating the tables. */
2724 gc_alloc_update_all_page_tables();
2727 current_new_areas_index = new_areas_index;
2730 "The re-scan has finished; current_new_areas_index=%d.\n",
2731 current_new_areas_index));*/
2734 /* Turn off recording of areas allocated by gc_alloc(). */
2735 record_new_objects = 0;
2740 /* Check that none of the write_protected pages in this generation
2741 * have been written to. */
2742 for (i = 0; i < page_table_pages; i++) {
2743 if (page_allocated_p(i)
2744 && (page_table[i].bytes_used != 0)
2745 && (page_table[i].gen == generation)
2746 && (page_table[i].write_protected_cleared != 0)
2747 && (page_table[i].dont_move == 0)) {
2748 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
2749 i, generation, page_table[i].dont_move);
2756 /* Un-write-protect all the pages in from_space. This is done at the
2757 * start of a GC else there may be many page faults while scavenging
2758 * the newspace (I've seen drive the system time to 99%). These pages
2759 * would need to be unprotected anyway before unmapping in
2760 * free_oldspace; not sure what effect this has on paging.. */
2762 unprotect_oldspace(void)
2765 void *region_addr = 0;
2766 void *page_addr = 0;
2767 uword_t region_bytes = 0;
2769 for (i = 0; i < last_free_page; i++) {
2770 if (page_allocated_p(i)
2771 && (page_table[i].bytes_used != 0)
2772 && (page_table[i].gen == from_space)) {
2774 /* Remove any write-protection. We should be able to rely
2775 * on the write-protect flag to avoid redundant calls. */
2776 if (page_table[i].write_protected) {
2777 page_table[i].write_protected = 0;
2778 page_addr = page_address(i);
2781 region_addr = page_addr;
2782 region_bytes = GENCGC_CARD_BYTES;
2783 } else if (region_addr + region_bytes == page_addr) {
2784 /* Region continue. */
2785 region_bytes += GENCGC_CARD_BYTES;
2787 /* Unprotect previous region. */
2788 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2789 /* First page in new region. */
2790 region_addr = page_addr;
2791 region_bytes = GENCGC_CARD_BYTES;
2797 /* Unprotect last region. */
2798 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2802 /* Work through all the pages and free any in from_space. This
2803 * assumes that all objects have been copied or promoted to an older
2804 * generation. Bytes_allocated and the generation bytes_allocated
2805 * counter are updated. The number of bytes freed is returned. */
2809 uword_t bytes_freed = 0;
2810 page_index_t first_page, last_page;
2815 /* Find a first page for the next region of pages. */
2816 while ((first_page < last_free_page)
2817 && (page_free_p(first_page)
2818 || (page_table[first_page].bytes_used == 0)
2819 || (page_table[first_page].gen != from_space)))
2822 if (first_page >= last_free_page)
2825 /* Find the last page of this region. */
2826 last_page = first_page;
2829 /* Free the page. */
2830 bytes_freed += page_table[last_page].bytes_used;
2831 generations[page_table[last_page].gen].bytes_allocated -=
2832 page_table[last_page].bytes_used;
2833 page_table[last_page].allocated = FREE_PAGE_FLAG;
2834 page_table[last_page].bytes_used = 0;
2835 /* Should already be unprotected by unprotect_oldspace(). */
2836 gc_assert(!page_table[last_page].write_protected);
2839 while ((last_page < last_free_page)
2840 && page_allocated_p(last_page)
2841 && (page_table[last_page].bytes_used != 0)
2842 && (page_table[last_page].gen == from_space));
2844 #ifdef READ_PROTECT_FREE_PAGES
2845 os_protect(page_address(first_page),
2846 npage_bytes(last_page-first_page),
2849 first_page = last_page;
2850 } while (first_page < last_free_page);
2852 bytes_allocated -= bytes_freed;
2857 /* Print some information about a pointer at the given address. */
2859 print_ptr(lispobj *addr)
2861 /* If addr is in the dynamic space then out the page information. */
2862 page_index_t pi1 = find_page_index((void*)addr);
2865 fprintf(stderr," %p: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
2868 page_table[pi1].allocated,
2869 page_table[pi1].gen,
2870 page_table[pi1].bytes_used,
2871 page_table[pi1].scan_start_offset,
2872 page_table[pi1].dont_move);
2873 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
2887 is_in_stack_space(lispobj ptr)
2889 /* For space verification: Pointers can be valid if they point
2890 * to a thread stack space. This would be faster if the thread
2891 * structures had page-table entries as if they were part of
2892 * the heap space. */
2894 for_each_thread(th) {
2895 if ((th->control_stack_start <= (lispobj *)ptr) &&
2896 (th->control_stack_end >= (lispobj *)ptr)) {
2904 verify_space(lispobj *start, size_t words)
2906 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
2907 int is_in_readonly_space =
2908 (READ_ONLY_SPACE_START <= (uword_t)start &&
2909 (uword_t)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2913 lispobj thing = *(lispobj*)start;
2915 if (is_lisp_pointer(thing)) {
2916 page_index_t page_index = find_page_index((void*)thing);
2917 sword_t to_readonly_space =
2918 (READ_ONLY_SPACE_START <= thing &&
2919 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2920 sword_t to_static_space =
2921 (STATIC_SPACE_START <= thing &&
2922 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
2924 /* Does it point to the dynamic space? */
2925 if (page_index != -1) {
2926 /* If it's within the dynamic space it should point to a used
2927 * page. XX Could check the offset too. */
2928 if (page_allocated_p(page_index)
2929 && (page_table[page_index].bytes_used == 0))
2930 lose ("Ptr %p @ %p sees free page.\n", thing, start);
2931 /* Check that it doesn't point to a forwarding pointer! */
2932 if (*((lispobj *)native_pointer(thing)) == 0x01) {
2933 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
2935 /* Check that its not in the RO space as it would then be a
2936 * pointer from the RO to the dynamic space. */
2937 if (is_in_readonly_space) {
2938 lose("ptr to dynamic space %p from RO space %x\n",
2941 /* Does it point to a plausible object? This check slows
2942 * it down a lot (so it's commented out).
2944 * "a lot" is serious: it ate 50 minutes cpu time on
2945 * my duron 950 before I came back from lunch and
2948 * FIXME: Add a variable to enable this
2951 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
2952 lose("ptr %p to invalid object %p\n", thing, start);
2956 extern void funcallable_instance_tramp;
2957 /* Verify that it points to another valid space. */
2958 if (!to_readonly_space && !to_static_space
2959 && (thing != (lispobj)&funcallable_instance_tramp)
2960 && !is_in_stack_space(thing)) {
2961 lose("Ptr %p @ %p sees junk.\n", thing, start);
2965 if (!(fixnump(thing))) {
2967 switch(widetag_of(*start)) {
2970 case SIMPLE_VECTOR_WIDETAG:
2972 case COMPLEX_WIDETAG:
2973 case SIMPLE_ARRAY_WIDETAG:
2974 case COMPLEX_BASE_STRING_WIDETAG:
2975 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2976 case COMPLEX_CHARACTER_STRING_WIDETAG:
2978 case COMPLEX_VECTOR_NIL_WIDETAG:
2979 case COMPLEX_BIT_VECTOR_WIDETAG:
2980 case COMPLEX_VECTOR_WIDETAG:
2981 case COMPLEX_ARRAY_WIDETAG:
2982 case CLOSURE_HEADER_WIDETAG:
2983 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2984 case VALUE_CELL_HEADER_WIDETAG:
2985 case SYMBOL_HEADER_WIDETAG:
2986 case CHARACTER_WIDETAG:
2987 #if N_WORD_BITS == 64
2988 case SINGLE_FLOAT_WIDETAG:
2990 case UNBOUND_MARKER_WIDETAG:
2995 case INSTANCE_HEADER_WIDETAG:
2998 sword_t ntotal = HeaderValue(thing);
2999 lispobj layout = ((struct instance *)start)->slots[0];
3004 nuntagged = ((struct layout *)
3005 native_pointer(layout))->n_untagged_slots;
3006 verify_space(start + 1,
3007 ntotal - fixnum_value(nuntagged));
3011 case CODE_HEADER_WIDETAG:
3013 lispobj object = *start;
3015 sword_t nheader_words, ncode_words, nwords;
3017 struct simple_fun *fheaderp;
3019 code = (struct code *) start;
3021 /* Check that it's not in the dynamic space.
3022 * FIXME: Isn't is supposed to be OK for code
3023 * objects to be in the dynamic space these days? */
3024 if (is_in_dynamic_space
3025 /* It's ok if it's byte compiled code. The trace
3026 * table offset will be a fixnum if it's x86
3027 * compiled code - check.
3029 * FIXME: #^#@@! lack of abstraction here..
3030 * This line can probably go away now that
3031 * there's no byte compiler, but I've got
3032 * too much to worry about right now to try
3033 * to make sure. -- WHN 2001-10-06 */
3034 && fixnump(code->trace_table_offset)
3035 /* Only when enabled */
3036 && verify_dynamic_code_check) {
3038 "/code object at %p in the dynamic space\n",
3042 ncode_words = fixnum_value(code->code_size);
3043 nheader_words = HeaderValue(object);
3044 nwords = ncode_words + nheader_words;
3045 nwords = CEILING(nwords, 2);
3046 /* Scavenge the boxed section of the code data block */
3047 verify_space(start + 1, nheader_words - 1);
3049 /* Scavenge the boxed section of each function
3050 * object in the code data block. */
3051 fheaderl = code->entry_points;
3052 while (fheaderl != NIL) {
3054 (struct simple_fun *) native_pointer(fheaderl);
3055 gc_assert(widetag_of(fheaderp->header) ==
3056 SIMPLE_FUN_HEADER_WIDETAG);
3057 verify_space(&fheaderp->name, 1);
3058 verify_space(&fheaderp->arglist, 1);
3059 verify_space(&fheaderp->type, 1);
3060 fheaderl = fheaderp->next;
3066 /* unboxed objects */
3067 case BIGNUM_WIDETAG:
3068 #if N_WORD_BITS != 64
3069 case SINGLE_FLOAT_WIDETAG:
3071 case DOUBLE_FLOAT_WIDETAG:
3072 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3073 case LONG_FLOAT_WIDETAG:
3075 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3076 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3078 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3079 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3081 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3082 case COMPLEX_LONG_FLOAT_WIDETAG:
3084 case SIMPLE_BASE_STRING_WIDETAG:
3085 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3086 case SIMPLE_CHARACTER_STRING_WIDETAG:
3088 case SIMPLE_BIT_VECTOR_WIDETAG:
3089 case SIMPLE_ARRAY_NIL_WIDETAG:
3090 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3091 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3092 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3093 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3094 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3095 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3097 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3099 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3100 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3101 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3102 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3104 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3105 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3107 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3108 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3110 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3111 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3114 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3116 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3117 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3119 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3120 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3122 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3123 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3124 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3125 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3127 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3128 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3130 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3131 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3133 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3134 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3137 case WEAK_POINTER_WIDETAG:
3138 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3139 case NO_TLS_VALUE_MARKER_WIDETAG:
3141 count = (sizetab[widetag_of(*start)])(start);
3145 lose("Unhandled widetag %p at %p\n",
3146 widetag_of(*start), start);
3158 /* FIXME: It would be nice to make names consistent so that
3159 * foo_size meant size *in* *bytes* instead of size in some
3160 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3161 * Some counts of lispobjs are called foo_count; it might be good
3162 * to grep for all foo_size and rename the appropriate ones to
3164 sword_t read_only_space_size =
3165 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3166 - (lispobj*)READ_ONLY_SPACE_START;
3167 sword_t static_space_size =
3168 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3169 - (lispobj*)STATIC_SPACE_START;
3171 for_each_thread(th) {
3172 sword_t binding_stack_size =
3173 (lispobj*)get_binding_stack_pointer(th)
3174 - (lispobj*)th->binding_stack_start;
3175 verify_space(th->binding_stack_start, binding_stack_size);
3177 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3178 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3182 verify_generation(generation_index_t generation)
3186 for (i = 0; i < last_free_page; i++) {
3187 if (page_allocated_p(i)
3188 && (page_table[i].bytes_used != 0)
3189 && (page_table[i].gen == generation)) {
3190 page_index_t last_page;
3191 int region_allocation = page_table[i].allocated;
3193 /* This should be the start of a contiguous block */
3194 gc_assert(page_starts_contiguous_block_p(i));
3196 /* Need to find the full extent of this contiguous block in case
3197 objects span pages. */
3199 /* Now work forward until the end of this contiguous area is
3201 for (last_page = i; ;last_page++)
3202 /* Check whether this is the last page in this contiguous
3204 if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES)
3205 /* Or it is CARD_BYTES and is the last in the block */
3206 || (page_table[last_page+1].allocated != region_allocation)
3207 || (page_table[last_page+1].bytes_used == 0)
3208 || (page_table[last_page+1].gen != generation)
3209 || (page_starts_contiguous_block_p(last_page+1)))
3212 verify_space(page_address(i),
3214 (page_table[last_page].bytes_used
3215 + npage_bytes(last_page-i)))
3222 /* Check that all the free space is zero filled. */
3224 verify_zero_fill(void)
3228 for (page = 0; page < last_free_page; page++) {
3229 if (page_free_p(page)) {
3230 /* The whole page should be zero filled. */
3231 sword_t *start_addr = (sword_t *)page_address(page);
3232 sword_t size = 1024;
3234 for (i = 0; i < size; i++) {
3235 if (start_addr[i] != 0) {
3236 lose("free page not zero at %x\n", start_addr + i);
3240 sword_t free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3241 if (free_bytes > 0) {
3242 sword_t *start_addr = (sword_t *)((uword_t)page_address(page)
3243 + page_table[page].bytes_used);
3244 sword_t size = free_bytes / N_WORD_BYTES;
3246 for (i = 0; i < size; i++) {
3247 if (start_addr[i] != 0) {
3248 lose("free region not zero at %x\n", start_addr + i);
3256 /* External entry point for verify_zero_fill */
3258 gencgc_verify_zero_fill(void)
3260 /* Flush the alloc regions updating the tables. */
3261 gc_alloc_update_all_page_tables();
3262 SHOW("verifying zero fill");
3267 verify_dynamic_space(void)
3269 generation_index_t i;
3271 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3272 verify_generation(i);
3274 if (gencgc_enable_verify_zero_fill)
3278 /* Write-protect all the dynamic boxed pages in the given generation. */
3280 write_protect_generation_pages(generation_index_t generation)
3284 gc_assert(generation < SCRATCH_GENERATION);
3286 for (start = 0; start < last_free_page; start++) {
3287 if (protect_page_p(start, generation)) {
3291 /* Note the page as protected in the page tables. */
3292 page_table[start].write_protected = 1;
3294 for (last = start + 1; last < last_free_page; last++) {
3295 if (!protect_page_p(last, generation))
3297 page_table[last].write_protected = 1;
3300 page_start = (void *)page_address(start);
3302 os_protect(page_start,
3303 npage_bytes(last - start),
3304 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3310 if (gencgc_verbose > 1) {
3312 "/write protected %d of %d pages in generation %d\n",
3313 count_write_protect_generation_pages(generation),
3314 count_generation_pages(generation),
3319 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3321 preserve_context_registers (os_context_t *c)
3324 /* On Darwin the signal context isn't a contiguous block of memory,
3325 * so just preserve_pointering its contents won't be sufficient.
3327 #if defined(LISP_FEATURE_DARWIN)||defined(LISP_FEATURE_WIN32)
3328 #if defined LISP_FEATURE_X86
3329 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3330 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3331 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3332 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3333 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3334 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3335 preserve_pointer((void*)*os_context_pc_addr(c));
3336 #elif defined LISP_FEATURE_X86_64
3337 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3338 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3339 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3340 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3341 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3342 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3343 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3344 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3345 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3346 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3347 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3348 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3349 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3350 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3351 preserve_pointer((void*)*os_context_pc_addr(c));
3353 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3356 #if !defined(LISP_FEATURE_WIN32)
3357 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3358 preserve_pointer(*ptr);
3365 move_pinned_pages_to_newspace()
3369 /* scavenge() will evacuate all oldspace pages, but no newspace
3370 * pages. Pinned pages are precisely those pages which must not
3371 * be evacuated, so move them to newspace directly. */
3373 for (i = 0; i < last_free_page; i++) {
3374 if (page_table[i].dont_move &&
3375 /* dont_move is cleared lazily, so validate the space as well. */
3376 page_table[i].gen == from_space) {
3377 page_table[i].gen = new_space;
3378 /* And since we're moving the pages wholesale, also adjust
3379 * the generation allocation counters. */
3380 generations[new_space].bytes_allocated += page_table[i].bytes_used;
3381 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
3386 /* Garbage collect a generation. If raise is 0 then the remains of the
3387 * generation are not raised to the next generation. */
3389 garbage_collect_generation(generation_index_t generation, int raise)
3391 uword_t bytes_freed;
3393 uword_t static_space_size;
3396 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3398 /* The oldest generation can't be raised. */
3399 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3401 /* Check if weak hash tables were processed in the previous GC. */
3402 gc_assert(weak_hash_tables == NULL);
3404 /* Initialize the weak pointer list. */
3405 weak_pointers = NULL;
3407 /* When a generation is not being raised it is transported to a
3408 * temporary generation (NUM_GENERATIONS), and lowered when
3409 * done. Set up this new generation. There should be no pages
3410 * allocated to it yet. */
3412 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3415 /* Set the global src and dest. generations */
3416 from_space = generation;
3418 new_space = generation+1;
3420 new_space = SCRATCH_GENERATION;
3422 /* Change to a new space for allocation, resetting the alloc_start_page */
3423 gc_alloc_generation = new_space;
3424 generations[new_space].alloc_start_page = 0;
3425 generations[new_space].alloc_unboxed_start_page = 0;
3426 generations[new_space].alloc_large_start_page = 0;
3427 generations[new_space].alloc_large_unboxed_start_page = 0;
3429 /* Before any pointers are preserved, the dont_move flags on the
3430 * pages need to be cleared. */
3431 for (i = 0; i < last_free_page; i++)
3432 if(page_table[i].gen==from_space)
3433 page_table[i].dont_move = 0;
3435 /* Un-write-protect the old-space pages. This is essential for the
3436 * promoted pages as they may contain pointers into the old-space
3437 * which need to be scavenged. It also helps avoid unnecessary page
3438 * faults as forwarding pointers are written into them. They need to
3439 * be un-protected anyway before unmapping later. */
3440 unprotect_oldspace();
3442 /* Scavenge the stacks' conservative roots. */
3444 /* there are potentially two stacks for each thread: the main
3445 * stack, which may contain Lisp pointers, and the alternate stack.
3446 * We don't ever run Lisp code on the altstack, but it may
3447 * host a sigcontext with lisp objects in it */
3449 /* what we need to do: (1) find the stack pointer for the main
3450 * stack; scavenge it (2) find the interrupt context on the
3451 * alternate stack that might contain lisp values, and scavenge
3454 /* we assume that none of the preceding applies to the thread that
3455 * initiates GC. If you ever call GC from inside an altstack
3456 * handler, you will lose. */
3458 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3459 /* And if we're saving a core, there's no point in being conservative. */
3460 if (conservative_stack) {
3461 for_each_thread(th) {
3463 void **esp=(void **)-1;
3464 if (th->state == STATE_DEAD)
3466 # if defined(LISP_FEATURE_SB_SAFEPOINT)
3467 /* Conservative collect_garbage is always invoked with a
3468 * foreign C call or an interrupt handler on top of every
3469 * existing thread, so the stored SP in each thread
3470 * structure is valid, no matter which thread we are looking
3471 * at. For threads that were running Lisp code, the pitstop
3472 * and edge functions maintain this value within the
3473 * interrupt or exception handler. */
3474 esp = os_get_csp(th);
3475 assert_on_stack(th, esp);
3477 /* In addition to pointers on the stack, also preserve the
3478 * return PC, the only value from the context that we need
3479 * in addition to the SP. The return PC gets saved by the
3480 * foreign call wrapper, and removed from the control stack
3481 * into a register. */
3482 preserve_pointer(th->pc_around_foreign_call);
3484 /* And on platforms with interrupts: scavenge ctx registers. */
3486 /* Disabled on Windows, because it does not have an explicit
3487 * stack of `interrupt_contexts'. The reported CSP has been
3488 * chosen so that the current context on the stack is
3489 * covered by the stack scan. See also set_csp_from_context(). */
3490 # ifndef LISP_FEATURE_WIN32
3491 if (th != arch_os_get_current_thread()) {
3492 long k = fixnum_value(
3493 SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3495 preserve_context_registers(th->interrupt_contexts[--k]);
3498 # elif defined(LISP_FEATURE_SB_THREAD)
3500 if(th==arch_os_get_current_thread()) {
3501 /* Somebody is going to burn in hell for this, but casting
3502 * it in two steps shuts gcc up about strict aliasing. */
3503 esp = (void **)((void *)&raise);
3506 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3507 for(i=free-1;i>=0;i--) {
3508 os_context_t *c=th->interrupt_contexts[i];
3509 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3510 if (esp1>=(void **)th->control_stack_start &&
3511 esp1<(void **)th->control_stack_end) {
3512 if(esp1<esp) esp=esp1;
3513 preserve_context_registers(c);
3518 esp = (void **)((void *)&raise);
3520 if (!esp || esp == (void*) -1)
3521 lose("garbage_collect: no SP known for thread %x (OS %x)",
3523 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3524 preserve_pointer(*ptr);
3529 /* Non-x86oid systems don't have "conservative roots" as such, but
3530 * the same mechanism is used for objects pinned for use by alien
3532 for_each_thread(th) {
3533 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
3534 while (pin_list != NIL) {
3535 struct cons *list_entry =
3536 (struct cons *)native_pointer(pin_list);
3537 preserve_pointer(list_entry->car);
3538 pin_list = list_entry->cdr;
3544 if (gencgc_verbose > 1) {
3545 sword_t num_dont_move_pages = count_dont_move_pages();
3547 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3548 num_dont_move_pages,
3549 npage_bytes(num_dont_move_pages));
3553 /* Now that all of the pinned (dont_move) pages are known, and
3554 * before we start to scavenge (and thus relocate) objects,
3555 * relocate the pinned pages to newspace, so that the scavenger
3556 * will not attempt to relocate their contents. */
3557 move_pinned_pages_to_newspace();
3559 /* Scavenge all the rest of the roots. */
3561 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3563 * If not x86, we need to scavenge the interrupt context(s) and the
3568 for_each_thread(th) {
3569 scavenge_interrupt_contexts(th);
3570 scavenge_control_stack(th);
3573 # ifdef LISP_FEATURE_SB_SAFEPOINT
3574 /* In this case, scrub all stacks right here from the GCing thread
3575 * instead of doing what the comment below says. Suboptimal, but
3578 scrub_thread_control_stack(th);
3580 /* Scrub the unscavenged control stack space, so that we can't run
3581 * into any stale pointers in a later GC (this is done by the
3582 * stop-for-gc handler in the other threads). */
3583 scrub_control_stack();
3588 /* Scavenge the Lisp functions of the interrupt handlers, taking
3589 * care to avoid SIG_DFL and SIG_IGN. */
3590 for (i = 0; i < NSIG; i++) {
3591 union interrupt_handler handler = interrupt_handlers[i];
3592 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
3593 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
3594 scavenge((lispobj *)(interrupt_handlers + i), 1);
3597 /* Scavenge the binding stacks. */
3600 for_each_thread(th) {
3601 sword_t len= (lispobj *)get_binding_stack_pointer(th) -
3602 th->binding_stack_start;
3603 scavenge((lispobj *) th->binding_stack_start,len);
3604 #ifdef LISP_FEATURE_SB_THREAD
3605 /* do the tls as well */
3606 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
3607 (sizeof (struct thread))/(sizeof (lispobj));
3608 scavenge((lispobj *) (th+1),len);
3613 /* The original CMU CL code had scavenge-read-only-space code
3614 * controlled by the Lisp-level variable
3615 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
3616 * wasn't documented under what circumstances it was useful or
3617 * safe to turn it on, so it's been turned off in SBCL. If you
3618 * want/need this functionality, and can test and document it,
3619 * please submit a patch. */
3621 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
3622 uword_t read_only_space_size =
3623 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
3624 (lispobj*)READ_ONLY_SPACE_START;
3626 "/scavenge read only space: %d bytes\n",
3627 read_only_space_size * sizeof(lispobj)));
3628 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
3632 /* Scavenge static space. */
3634 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
3635 (lispobj *)STATIC_SPACE_START;
3636 if (gencgc_verbose > 1) {
3638 "/scavenge static space: %d bytes\n",
3639 static_space_size * sizeof(lispobj)));
3641 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
3643 /* All generations but the generation being GCed need to be
3644 * scavenged. The new_space generation needs special handling as
3645 * objects may be moved in - it is handled separately below. */
3646 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
3648 /* Finally scavenge the new_space generation. Keep going until no
3649 * more objects are moved into the new generation */
3650 scavenge_newspace_generation(new_space);
3652 /* FIXME: I tried reenabling this check when debugging unrelated
3653 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
3654 * Since the current GC code seems to work well, I'm guessing that
3655 * this debugging code is just stale, but I haven't tried to
3656 * figure it out. It should be figured out and then either made to
3657 * work or just deleted. */
3658 #define RESCAN_CHECK 0
3660 /* As a check re-scavenge the newspace once; no new objects should
3663 os_vm_size_t old_bytes_allocated = bytes_allocated;
3664 os_vm_size_t bytes_allocated;
3666 /* Start with a full scavenge. */
3667 scavenge_newspace_generation_one_scan(new_space);
3669 /* Flush the current regions, updating the tables. */
3670 gc_alloc_update_all_page_tables();
3672 bytes_allocated = bytes_allocated - old_bytes_allocated;
3674 if (bytes_allocated != 0) {
3675 lose("Rescan of new_space allocated %d more bytes.\n",
3681 scan_weak_hash_tables();
3682 scan_weak_pointers();
3684 /* Flush the current regions, updating the tables. */
3685 gc_alloc_update_all_page_tables();
3687 /* Free the pages in oldspace, but not those marked dont_move. */
3688 bytes_freed = free_oldspace();
3690 /* If the GC is not raising the age then lower the generation back
3691 * to its normal generation number */
3693 for (i = 0; i < last_free_page; i++)
3694 if ((page_table[i].bytes_used != 0)
3695 && (page_table[i].gen == SCRATCH_GENERATION))
3696 page_table[i].gen = generation;
3697 gc_assert(generations[generation].bytes_allocated == 0);
3698 generations[generation].bytes_allocated =
3699 generations[SCRATCH_GENERATION].bytes_allocated;
3700 generations[SCRATCH_GENERATION].bytes_allocated = 0;
3703 /* Reset the alloc_start_page for generation. */
3704 generations[generation].alloc_start_page = 0;
3705 generations[generation].alloc_unboxed_start_page = 0;
3706 generations[generation].alloc_large_start_page = 0;
3707 generations[generation].alloc_large_unboxed_start_page = 0;
3709 if (generation >= verify_gens) {
3710 if (gencgc_verbose) {
3714 verify_dynamic_space();
3717 /* Set the new gc trigger for the GCed generation. */
3718 generations[generation].gc_trigger =
3719 generations[generation].bytes_allocated
3720 + generations[generation].bytes_consed_between_gc;
3723 generations[generation].num_gc = 0;
3725 ++generations[generation].num_gc;
3729 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
3731 update_dynamic_space_free_pointer(void)
3733 page_index_t last_page = -1, i;
3735 for (i = 0; i < last_free_page; i++)
3736 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
3739 last_free_page = last_page+1;
3741 set_alloc_pointer((lispobj)(page_address(last_free_page)));
3742 return 0; /* dummy value: return something ... */
3746 remap_page_range (page_index_t from, page_index_t to)
3748 /* There's a mysterious Solaris/x86 problem with using mmap
3749 * tricks for memory zeroing. See sbcl-devel thread
3750 * "Re: patch: standalone executable redux".
3752 #if defined(LISP_FEATURE_SUNOS)
3753 zero_and_mark_pages(from, to);
3756 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
3757 release_mask = release_granularity-1,
3759 aligned_from = (from+release_mask)&~release_mask,
3760 aligned_end = (end&~release_mask);
3762 if (aligned_from < aligned_end) {
3763 zero_pages_with_mmap(aligned_from, aligned_end-1);
3764 if (aligned_from != from)
3765 zero_and_mark_pages(from, aligned_from-1);
3766 if (aligned_end != end)
3767 zero_and_mark_pages(aligned_end, end-1);
3769 zero_and_mark_pages(from, to);
3775 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
3777 page_index_t first_page, last_page;
3780 return remap_page_range(from, to);
3782 for (first_page = from; first_page <= to; first_page++) {
3783 if (page_allocated_p(first_page) ||
3784 (page_table[first_page].need_to_zero == 0))
3787 last_page = first_page + 1;
3788 while (page_free_p(last_page) &&
3789 (last_page <= to) &&
3790 (page_table[last_page].need_to_zero == 1))
3793 remap_page_range(first_page, last_page-1);
3795 first_page = last_page;
3799 generation_index_t small_generation_limit = 1;
3801 /* GC all generations newer than last_gen, raising the objects in each
3802 * to the next older generation - we finish when all generations below
3803 * last_gen are empty. Then if last_gen is due for a GC, or if
3804 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
3805 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
3807 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
3808 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
3810 collect_garbage(generation_index_t last_gen)
3812 generation_index_t gen = 0, i;
3813 int raise, more = 0;
3815 /* The largest value of last_free_page seen since the time
3816 * remap_free_pages was called. */
3817 static page_index_t high_water_mark = 0;
3819 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
3820 log_generation_stats(gc_logfile, "=== GC Start ===");
3824 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
3826 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
3831 /* Flush the alloc regions updating the tables. */
3832 gc_alloc_update_all_page_tables();
3834 /* Verify the new objects created by Lisp code. */
3835 if (pre_verify_gen_0) {
3836 FSHOW((stderr, "pre-checking generation 0\n"));
3837 verify_generation(0);
3840 if (gencgc_verbose > 1)
3841 print_generation_stats();
3844 /* Collect the generation. */
3846 if (more || (gen >= gencgc_oldest_gen_to_gc)) {
3847 /* Never raise the oldest generation. Never raise the extra generation
3848 * collected due to more-flag. */
3854 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
3855 /* If we would not normally raise this one, but we're
3856 * running low on space in comparison to the object-sizes
3857 * we've been seeing, raise it and collect the next one
3859 if (!raise && gen == last_gen) {
3860 more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated);
3865 if (gencgc_verbose > 1) {
3867 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
3870 generations[gen].bytes_allocated,
3871 generations[gen].gc_trigger,
3872 generations[gen].num_gc));
3875 /* If an older generation is being filled, then update its
3878 generations[gen+1].cum_sum_bytes_allocated +=
3879 generations[gen+1].bytes_allocated;
3882 garbage_collect_generation(gen, raise);
3884 /* Reset the memory age cum_sum. */
3885 generations[gen].cum_sum_bytes_allocated = 0;
3887 if (gencgc_verbose > 1) {
3888 FSHOW((stderr, "GC of generation %d finished:\n", gen));
3889 print_generation_stats();
3893 } while ((gen <= gencgc_oldest_gen_to_gc)
3894 && ((gen < last_gen)
3897 && (generations[gen].bytes_allocated
3898 > generations[gen].gc_trigger)
3899 && (generation_average_age(gen)
3900 > generations[gen].minimum_age_before_gc))));
3902 /* Now if gen-1 was raised all generations before gen are empty.
3903 * If it wasn't raised then all generations before gen-1 are empty.
3905 * Now objects within this gen's pages cannot point to younger
3906 * generations unless they are written to. This can be exploited
3907 * by write-protecting the pages of gen; then when younger
3908 * generations are GCed only the pages which have been written
3913 gen_to_wp = gen - 1;
3915 /* There's not much point in WPing pages in generation 0 as it is
3916 * never scavenged (except promoted pages). */
3917 if ((gen_to_wp > 0) && enable_page_protection) {
3918 /* Check that they are all empty. */
3919 for (i = 0; i < gen_to_wp; i++) {
3920 if (generations[i].bytes_allocated)
3921 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
3924 write_protect_generation_pages(gen_to_wp);
3927 /* Set gc_alloc() back to generation 0. The current regions should
3928 * be flushed after the above GCs. */
3929 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
3930 gc_alloc_generation = 0;
3932 /* Save the high-water mark before updating last_free_page */
3933 if (last_free_page > high_water_mark)
3934 high_water_mark = last_free_page;
3936 update_dynamic_space_free_pointer();
3938 /* Update auto_gc_trigger. Make sure we trigger the next GC before
3939 * running out of heap! */
3940 if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
3941 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
3943 auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;
3946 fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n",
3949 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
3952 if (gen > small_generation_limit) {
3953 if (last_free_page > high_water_mark)
3954 high_water_mark = last_free_page;
3955 remap_free_pages(0, high_water_mark, 0);
3956 high_water_mark = 0;
3960 large_allocation = 0;
3962 log_generation_stats(gc_logfile, "=== GC End ===");
3963 SHOW("returning from collect_garbage");
3966 /* This is called by Lisp PURIFY when it is finished. All live objects
3967 * will have been moved to the RO and Static heaps. The dynamic space
3968 * will need a full re-initialization. We don't bother having Lisp
3969 * PURIFY flush the current gc_alloc() region, as the page_tables are
3970 * re-initialized, and every page is zeroed to be sure. */
3974 page_index_t page, last_page;
3976 if (gencgc_verbose > 1) {
3977 SHOW("entering gc_free_heap");
3980 for (page = 0; page < page_table_pages; page++) {
3981 /* Skip free pages which should already be zero filled. */
3982 if (page_allocated_p(page)) {
3984 for (last_page = page;
3985 (last_page < page_table_pages) && page_allocated_p(last_page);
3987 /* Mark the page free. The other slots are assumed invalid
3988 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
3989 * should not be write-protected -- except that the
3990 * generation is used for the current region but it sets
3992 page_table[page].allocated = FREE_PAGE_FLAG;
3993 page_table[page].bytes_used = 0;
3994 page_table[page].write_protected = 0;
3997 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
3998 * about this change. */
3999 page_start = (void *)page_address(page);
4000 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
4001 remap_free_pages(page, last_page-1, 1);
4004 } else if (gencgc_zero_check_during_free_heap) {
4005 /* Double-check that the page is zero filled. */
4006 sword_t *page_start;
4008 gc_assert(page_free_p(page));
4009 gc_assert(page_table[page].bytes_used == 0);
4010 page_start = (sword_t *)page_address(page);
4011 for (i=0; i<GENCGC_CARD_BYTES/sizeof(sword_t); i++) {
4012 if (page_start[i] != 0) {
4013 lose("free region not zero at %x\n", page_start + i);
4019 bytes_allocated = 0;
4021 /* Initialize the generations. */
4022 for (page = 0; page < NUM_GENERATIONS; page++) {
4023 generations[page].alloc_start_page = 0;
4024 generations[page].alloc_unboxed_start_page = 0;
4025 generations[page].alloc_large_start_page = 0;
4026 generations[page].alloc_large_unboxed_start_page = 0;
4027 generations[page].bytes_allocated = 0;
4028 generations[page].gc_trigger = 2000000;
4029 generations[page].num_gc = 0;
4030 generations[page].cum_sum_bytes_allocated = 0;
4033 if (gencgc_verbose > 1)
4034 print_generation_stats();
4036 /* Initialize gc_alloc(). */
4037 gc_alloc_generation = 0;
4039 gc_set_region_empty(&boxed_region);
4040 gc_set_region_empty(&unboxed_region);
4043 set_alloc_pointer((lispobj)((char *)heap_base));
4045 if (verify_after_free_heap) {
4046 /* Check whether purify has left any bad pointers. */
4047 FSHOW((stderr, "checking after free_heap\n"));
4057 #if defined(LISP_FEATURE_SB_SAFEPOINT)
4061 /* Compute the number of pages needed for the dynamic space.
4062 * Dynamic space size should be aligned on page size. */
4063 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4064 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4066 /* Default nursery size to 5% of the total dynamic space size,
4068 bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
4069 if (bytes_consed_between_gcs < (1024*1024))
4070 bytes_consed_between_gcs = 1024*1024;
4072 /* The page_table must be allocated using "calloc" to initialize
4073 * the page structures correctly. There used to be a separate
4074 * initialization loop (now commented out; see below) but that was
4075 * unnecessary and did hurt startup time. */
4076 page_table = calloc(page_table_pages, sizeof(struct page));
4077 gc_assert(page_table);
4080 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4081 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4083 heap_base = (void*)DYNAMIC_SPACE_START;
4085 /* The page structures are initialized implicitly when page_table
4086 * is allocated with "calloc" above. Formerly we had the following
4087 * explicit initialization here (comments converted to C99 style
4088 * for readability as C's block comments don't nest):
4090 * // Initialize each page structure.
4091 * for (i = 0; i < page_table_pages; i++) {
4092 * // Initialize all pages as free.
4093 * page_table[i].allocated = FREE_PAGE_FLAG;
4094 * page_table[i].bytes_used = 0;
4096 * // Pages are not write-protected at startup.
4097 * page_table[i].write_protected = 0;
4100 * Without this loop the image starts up much faster when dynamic
4101 * space is large -- which it is on 64-bit platforms already by
4102 * default -- and when "calloc" for large arrays is implemented
4103 * using copy-on-write of a page of zeroes -- which it is at least
4104 * on Linux. In this case the pages that page_table_pages is stored
4105 * in are mapped and cleared not before the corresponding part of
4106 * dynamic space is used. For example, this saves clearing 16 MB of
4107 * memory at startup if the page size is 4 KB and the size of
4108 * dynamic space is 4 GB.
4109 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4110 * asserted below: */
4112 /* Compile time assertion: If triggered, declares an array
4113 * of dimension -1 forcing a syntax error. The intent of the
4114 * assignment is to avoid an "unused variable" warning. */
4115 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4116 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4119 bytes_allocated = 0;
4121 /* Initialize the generations.
4123 * FIXME: very similar to code in gc_free_heap(), should be shared */
4124 for (i = 0; i < NUM_GENERATIONS; i++) {
4125 generations[i].alloc_start_page = 0;
4126 generations[i].alloc_unboxed_start_page = 0;
4127 generations[i].alloc_large_start_page = 0;
4128 generations[i].alloc_large_unboxed_start_page = 0;
4129 generations[i].bytes_allocated = 0;
4130 generations[i].gc_trigger = 2000000;
4131 generations[i].num_gc = 0;
4132 generations[i].cum_sum_bytes_allocated = 0;
4133 /* the tune-able parameters */
4134 generations[i].bytes_consed_between_gc
4135 = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
4136 generations[i].number_of_gcs_before_promotion = 1;
4137 generations[i].minimum_age_before_gc = 0.75;
4140 /* Initialize gc_alloc. */
4141 gc_alloc_generation = 0;
4142 gc_set_region_empty(&boxed_region);
4143 gc_set_region_empty(&unboxed_region);
4148 /* Pick up the dynamic space from after a core load.
4150 * The ALLOCATION_POINTER points to the end of the dynamic space.
4154 gencgc_pickup_dynamic(void)
4156 page_index_t page = 0;
4157 void *alloc_ptr = (void *)get_alloc_pointer();
4158 lispobj *prev=(lispobj *)page_address(page);
4159 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4161 bytes_allocated = 0;
4164 lispobj *first,*ptr= (lispobj *)page_address(page);
4166 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4167 /* It is possible, though rare, for the saved page table
4168 * to contain free pages below alloc_ptr. */
4169 page_table[page].gen = gen;
4170 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4171 page_table[page].large_object = 0;
4172 page_table[page].write_protected = 0;
4173 page_table[page].write_protected_cleared = 0;
4174 page_table[page].dont_move = 0;
4175 page_table[page].need_to_zero = 1;
4177 bytes_allocated += GENCGC_CARD_BYTES;
4180 if (!gencgc_partial_pickup) {
4181 page_table[page].allocated = BOXED_PAGE_FLAG;
4182 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4185 page_table[page].scan_start_offset =
4186 page_address(page) - (void *)prev;
4189 } while (page_address(page) < alloc_ptr);
4191 last_free_page = page;
4193 generations[gen].bytes_allocated = bytes_allocated;
4195 gc_alloc_update_all_page_tables();
4196 write_protect_generation_pages(gen);
4200 gc_initialize_pointers(void)
4202 gencgc_pickup_dynamic();
4206 /* alloc(..) is the external interface for memory allocation. It
4207 * allocates to generation 0. It is not called from within the garbage
4208 * collector as it is only external uses that need the check for heap
4209 * size (GC trigger) and to disable the interrupts (interrupts are
4210 * always disabled during a GC).
4212 * The vops that call alloc(..) assume that the returned space is zero-filled.
4213 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4215 * The check for a GC trigger is only performed when the current
4216 * region is full, so in most cases it's not needed. */
4218 static inline lispobj *
4219 general_alloc_internal(sword_t nbytes, int page_type_flag, struct alloc_region *region,
4220 struct thread *thread)
4222 #ifndef LISP_FEATURE_WIN32
4223 lispobj alloc_signal;
4226 void *new_free_pointer;
4227 os_vm_size_t trigger_bytes = 0;
4229 gc_assert(nbytes>0);
4231 /* Check for alignment allocation problems. */
4232 gc_assert((((uword_t)region->free_pointer & LOWTAG_MASK) == 0)
4233 && ((nbytes & LOWTAG_MASK) == 0));
4235 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
4236 /* Must be inside a PA section. */
4237 gc_assert(get_pseudo_atomic_atomic(thread));
4240 if (nbytes > large_allocation)
4241 large_allocation = nbytes;
4243 /* maybe we can do this quickly ... */
4244 new_free_pointer = region->free_pointer + nbytes;
4245 if (new_free_pointer <= region->end_addr) {
4246 new_obj = (void*)(region->free_pointer);
4247 region->free_pointer = new_free_pointer;
4248 return(new_obj); /* yup */
4251 /* We don't want to count nbytes against auto_gc_trigger unless we
4252 * have to: it speeds up the tenuring of objects and slows down
4253 * allocation. However, unless we do so when allocating _very_
4254 * large objects we are in danger of exhausting the heap without
4255 * running sufficient GCs.
4257 if (nbytes >= bytes_consed_between_gcs)
4258 trigger_bytes = nbytes;
4260 /* we have to go the long way around, it seems. Check whether we
4261 * should GC in the near future
4263 if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
4264 /* Don't flood the system with interrupts if the need to gc is
4265 * already noted. This can happen for example when SUB-GC
4266 * allocates or after a gc triggered in a WITHOUT-GCING. */
4267 if (SymbolValue(GC_PENDING,thread) == NIL) {
4268 /* set things up so that GC happens when we finish the PA
4270 SetSymbolValue(GC_PENDING,T,thread);
4271 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4272 #ifdef LISP_FEATURE_SB_SAFEPOINT
4273 thread_register_gc_trigger();
4275 set_pseudo_atomic_interrupted(thread);
4276 #ifdef GENCGC_IS_PRECISE
4277 /* PPC calls alloc() from a trap or from pa_alloc(),
4278 * look up the most context if it's from a trap. */
4280 os_context_t *context =
4281 thread->interrupt_data->allocation_trap_context;
4282 maybe_save_gc_mask_and_block_deferrables
4283 (context ? os_context_sigmask_addr(context) : NULL);
4286 maybe_save_gc_mask_and_block_deferrables(NULL);
4292 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4294 #ifndef LISP_FEATURE_WIN32
4295 /* for sb-prof, and not supported on Windows yet */
4296 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4297 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4298 if ((sword_t) alloc_signal <= 0) {
4299 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4302 SetSymbolValue(ALLOC_SIGNAL,
4303 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4313 general_alloc(sword_t nbytes, int page_type_flag)
4315 struct thread *thread = arch_os_get_current_thread();
4316 /* Select correct region, and call general_alloc_internal with it.
4317 * For other then boxed allocation we must lock first, since the
4318 * region is shared. */
4319 if (BOXED_PAGE_FLAG & page_type_flag) {
4320 #ifdef LISP_FEATURE_SB_THREAD
4321 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4323 struct alloc_region *region = &boxed_region;
4325 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4326 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4328 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4329 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4330 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4333 lose("bad page type flag: %d", page_type_flag);
4337 lispobj AMD64_SYSV_ABI *
4340 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4341 struct thread *self = arch_os_get_current_thread();
4342 int was_pseudo_atomic = get_pseudo_atomic_atomic(self);
4343 if (!was_pseudo_atomic)
4344 set_pseudo_atomic_atomic(self);
4346 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4349 lispobj *result = general_alloc(nbytes, BOXED_PAGE_FLAG);
4351 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4352 if (!was_pseudo_atomic)
4353 clear_pseudo_atomic_atomic(self);
4360 * shared support for the OS-dependent signal handlers which
4361 * catch GENCGC-related write-protect violations
4363 void unhandled_sigmemoryfault(void* addr);
4365 /* Depending on which OS we're running under, different signals might
4366 * be raised for a violation of write protection in the heap. This
4367 * function factors out the common generational GC magic which needs
4368 * to invoked in this case, and should be called from whatever signal
4369 * handler is appropriate for the OS we're running under.
4371 * Return true if this signal is a normal generational GC thing that
4372 * we were able to handle, or false if it was abnormal and control
4373 * should fall through to the general SIGSEGV/SIGBUS/whatever logic.
4375 * We have two control flags for this: one causes us to ignore faults
4376 * on unprotected pages completely, and the second complains to stderr
4377 * but allows us to continue without losing.
4379 extern boolean ignore_memoryfaults_on_unprotected_pages;
4380 boolean ignore_memoryfaults_on_unprotected_pages = 0;
4382 extern boolean continue_after_memoryfault_on_unprotected_pages;
4383 boolean continue_after_memoryfault_on_unprotected_pages = 0;
4386 gencgc_handle_wp_violation(void* fault_addr)
4388 page_index_t page_index = find_page_index(fault_addr);
4391 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4392 fault_addr, page_index));
4395 /* Check whether the fault is within the dynamic space. */
4396 if (page_index == (-1)) {
4398 /* It can be helpful to be able to put a breakpoint on this
4399 * case to help diagnose low-level problems. */
4400 unhandled_sigmemoryfault(fault_addr);
4402 /* not within the dynamic space -- not our responsibility */
4407 ret = thread_mutex_lock(&free_pages_lock);
4408 gc_assert(ret == 0);
4409 if (page_table[page_index].write_protected) {
4410 /* Unprotect the page. */
4411 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4412 page_table[page_index].write_protected_cleared = 1;
4413 page_table[page_index].write_protected = 0;
4414 } else if (!ignore_memoryfaults_on_unprotected_pages) {
4415 /* The only acceptable reason for this signal on a heap
4416 * access is that GENCGC write-protected the page.
4417 * However, if two CPUs hit a wp page near-simultaneously,
4418 * we had better not have the second one lose here if it
4419 * does this test after the first one has already set wp=0
4421 if(page_table[page_index].write_protected_cleared != 1) {
4422 void lisp_backtrace(int frames);
4425 "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n"
4426 " boxed_region.first_page: %"PAGE_INDEX_FMT","
4427 " boxed_region.last_page %"PAGE_INDEX_FMT"\n"
4428 " page.scan_start_offset: %"OS_VM_SIZE_FMT"\n"
4429 " page.bytes_used: %"PAGE_BYTES_FMT"\n"
4430 " page.allocated: %d\n"
4431 " page.write_protected: %d\n"
4432 " page.write_protected_cleared: %d\n"
4433 " page.generation: %d\n",
4436 boxed_region.first_page,
4437 boxed_region.last_page,
4438 page_table[page_index].scan_start_offset,
4439 page_table[page_index].bytes_used,
4440 page_table[page_index].allocated,
4441 page_table[page_index].write_protected,
4442 page_table[page_index].write_protected_cleared,
4443 page_table[page_index].gen);
4444 if (!continue_after_memoryfault_on_unprotected_pages)
4448 ret = thread_mutex_unlock(&free_pages_lock);
4449 gc_assert(ret == 0);
4450 /* Don't worry, we can handle it. */
4454 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4455 * it's not just a case of the program hitting the write barrier, and
4456 * are about to let Lisp deal with it. It's basically just a
4457 * convenient place to set a gdb breakpoint. */
4459 unhandled_sigmemoryfault(void *addr)
4462 void gc_alloc_update_all_page_tables(void)
4464 /* Flush the alloc regions updating the tables. */
4466 for_each_thread(th) {
4467 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4468 #if defined(LISP_FEATURE_SB_SAFEPOINT_STRICTLY) && !defined(LISP_FEATURE_WIN32)
4469 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->sprof_alloc_region);
4472 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4473 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4477 gc_set_region_empty(struct alloc_region *region)
4479 region->first_page = 0;
4480 region->last_page = -1;
4481 region->start_addr = page_address(0);
4482 region->free_pointer = page_address(0);
4483 region->end_addr = page_address(0);
4487 zero_all_free_pages()
4491 for (i = 0; i < last_free_page; i++) {
4492 if (page_free_p(i)) {
4493 #ifdef READ_PROTECT_FREE_PAGES
4494 os_protect(page_address(i),
4503 /* Things to do before doing a final GC before saving a core (without
4506 * + Pages in large_object pages aren't moved by the GC, so we need to
4507 * unset that flag from all pages.
4508 * + The pseudo-static generation isn't normally collected, but it seems
4509 * reasonable to collect it at least when saving a core. So move the
4510 * pages to a normal generation.
4513 prepare_for_final_gc ()
4516 for (i = 0; i < last_free_page; i++) {
4517 page_table[i].large_object = 0;
4518 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4519 int used = page_table[i].bytes_used;
4520 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4521 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4522 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4528 /* Do a non-conservative GC, and then save a core with the initial
4529 * function being set to the value of the static symbol
4530 * SB!VM:RESTART-LISP-FUNCTION */
4532 gc_and_save(char *filename, boolean prepend_runtime,
4533 boolean save_runtime_options,
4534 boolean compressed, int compression_level)
4537 void *runtime_bytes = NULL;
4538 size_t runtime_size;
4540 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4545 conservative_stack = 0;
4547 /* The filename might come from Lisp, and be moved by the now
4548 * non-conservative GC. */
4549 filename = strdup(filename);
4551 /* Collect twice: once into relatively high memory, and then back
4552 * into low memory. This compacts the retained data into the lower
4553 * pages, minimizing the size of the core file.
4555 prepare_for_final_gc();
4556 gencgc_alloc_start_page = last_free_page;
4557 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4559 prepare_for_final_gc();
4560 gencgc_alloc_start_page = -1;
4561 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4563 if (prepend_runtime)
4564 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4566 /* The dumper doesn't know that pages need to be zeroed before use. */
4567 zero_all_free_pages();
4568 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4569 prepend_runtime, save_runtime_options,
4570 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
4571 /* Oops. Save still managed to fail. Since we've mangled the stack
4572 * beyond hope, there's not much we can do.
4573 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4574 * going to be rather unsatisfactory too... */
4575 lose("Attempt to save core after non-conservative GC failed.\n");