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 #ifdef LISP_FEATURE_X86
129 /* Should we check code objects for fixup errors after they are transported? */
130 boolean check_code_fixups = 0;
133 /* Should we check that newly allocated regions are zero filled? */
134 boolean gencgc_zero_check = 0;
136 /* Should we check that the free space is zero filled? */
137 boolean gencgc_enable_verify_zero_fill = 0;
139 /* Should we check that free pages are zero filled during gc_free_heap
140 * called after Lisp PURIFY? */
141 boolean gencgc_zero_check_during_free_heap = 0;
143 /* When loading a core, don't do a full scan of the memory for the
144 * memory region boundaries. (Set to true by coreparse.c if the core
145 * contained a pagetable entry).
147 boolean gencgc_partial_pickup = 0;
149 /* If defined, free pages are read-protected to ensure that nothing
153 /* #define READ_PROTECT_FREE_PAGES */
157 * GC structures and variables
160 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
161 os_vm_size_t bytes_allocated = 0;
162 os_vm_size_t auto_gc_trigger = 0;
164 /* the source and destination generations. These are set before a GC starts
166 generation_index_t from_space;
167 generation_index_t new_space;
169 /* Set to 1 when in GC */
170 boolean gc_active_p = 0;
172 /* should the GC be conservative on stack. If false (only right before
173 * saving a core), don't scan the stack / mark pages dont_move. */
174 static boolean conservative_stack = 1;
176 /* An array of page structures is allocated on gc initialization.
177 * This helps to quickly map between an address and its page structure.
178 * page_table_pages is set from the size of the dynamic space. */
179 page_index_t page_table_pages;
180 struct page *page_table;
182 static inline boolean page_allocated_p(page_index_t page) {
183 return (page_table[page].allocated != FREE_PAGE_FLAG);
186 static inline boolean page_no_region_p(page_index_t page) {
187 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
190 static inline boolean page_allocated_no_region_p(page_index_t page) {
191 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
192 && page_no_region_p(page));
195 static inline boolean page_free_p(page_index_t page) {
196 return (page_table[page].allocated == FREE_PAGE_FLAG);
199 static inline boolean page_boxed_p(page_index_t page) {
200 return (page_table[page].allocated & BOXED_PAGE_FLAG);
203 static inline boolean code_page_p(page_index_t page) {
204 /* This is used by the conservative pinning logic to determine if
205 * a page can contain code objects. Ideally, we'd be able to
206 * check the page allocation flag to see if it is CODE_PAGE_FLAG,
207 * but this turns out not to be reliable (in fact, badly
208 * unreliable) at the moment. On the upside, all code objects are
209 * boxed objects, so we can simply re-use the boxed_page_p() logic
210 * for a tighter result than merely "is this page allocated". */
212 return (page_table[page].allocated & CODE_PAGE_FLAG) == CODE_PAGE_FLAG;
214 return page_boxed_p(page);
218 static inline boolean page_boxed_no_region_p(page_index_t page) {
219 return page_boxed_p(page) && page_no_region_p(page);
222 static inline boolean page_unboxed_p(page_index_t page) {
223 /* Both flags set == boxed code page */
224 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
225 && !page_boxed_p(page));
228 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
229 return (page_boxed_no_region_p(page)
230 && (page_table[page].bytes_used != 0)
231 && !page_table[page].dont_move
232 && (page_table[page].gen == generation));
235 /* To map addresses to page structures the address of the first page
237 void *heap_base = NULL;
239 /* Calculate the start address for the given page number. */
241 page_address(page_index_t page_num)
243 return (heap_base + (page_num * GENCGC_CARD_BYTES));
246 /* Calculate the address where the allocation region associated with
247 * the page starts. */
249 page_scan_start(page_index_t page_index)
251 return page_address(page_index)-page_table[page_index].scan_start_offset;
254 /* True if the page starts a contiguous block. */
255 static inline boolean
256 page_starts_contiguous_block_p(page_index_t page_index)
258 return page_table[page_index].scan_start_offset == 0;
261 /* True if the page is the last page in a contiguous block. */
262 static inline boolean
263 page_ends_contiguous_block_p(page_index_t page_index, generation_index_t gen)
265 return (/* page doesn't fill block */
266 (page_table[page_index].bytes_used < GENCGC_CARD_BYTES)
267 /* page is last allocated page */
268 || ((page_index + 1) >= last_free_page)
270 || page_free_p(page_index + 1)
271 /* next page contains no data */
272 || (page_table[page_index + 1].bytes_used == 0)
273 /* next page is in different generation */
274 || (page_table[page_index + 1].gen != gen)
275 /* next page starts its own contiguous block */
276 || (page_starts_contiguous_block_p(page_index + 1)));
279 /* Find the page index within the page_table for the given
280 * address. Return -1 on failure. */
282 find_page_index(void *addr)
284 if (addr >= heap_base) {
285 page_index_t index = ((pointer_sized_uint_t)addr -
286 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
287 if (index < page_table_pages)
294 npage_bytes(page_index_t npages)
296 gc_assert(npages>=0);
297 return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
300 /* Check that X is a higher address than Y and return offset from Y to
302 static inline os_vm_size_t
303 void_diff(void *x, void *y)
306 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
309 /* a structure to hold the state of a generation
311 * CAUTION: If you modify this, make sure to touch up the alien
312 * definition in src/code/gc.lisp accordingly. ...or better yes,
313 * deal with the FIXME there...
317 /* the first page that gc_alloc() checks on its next call */
318 page_index_t alloc_start_page;
320 /* the first page that gc_alloc_unboxed() checks on its next call */
321 page_index_t alloc_unboxed_start_page;
323 /* the first page that gc_alloc_large (boxed) considers on its next
324 * call. (Although it always allocates after the boxed_region.) */
325 page_index_t alloc_large_start_page;
327 /* the first page that gc_alloc_large (unboxed) considers on its
328 * next call. (Although it always allocates after the
329 * current_unboxed_region.) */
330 page_index_t alloc_large_unboxed_start_page;
332 /* the bytes allocated to this generation */
333 os_vm_size_t bytes_allocated;
335 /* the number of bytes at which to trigger a GC */
336 os_vm_size_t gc_trigger;
338 /* to calculate a new level for gc_trigger */
339 os_vm_size_t bytes_consed_between_gc;
341 /* the number of GCs since the last raise */
344 /* the number of GCs to run on the generations before raising objects to the
346 int number_of_gcs_before_promotion;
348 /* the cumulative sum of the bytes allocated to this generation. It is
349 * cleared after a GC on this generations, and update before new
350 * objects are added from a GC of a younger generation. Dividing by
351 * the bytes_allocated will give the average age of the memory in
352 * this generation since its last GC. */
353 os_vm_size_t cum_sum_bytes_allocated;
355 /* a minimum average memory age before a GC will occur helps
356 * prevent a GC when a large number of new live objects have been
357 * added, in which case a GC could be a waste of time */
358 double minimum_age_before_gc;
361 /* an array of generation structures. There needs to be one more
362 * generation structure than actual generations as the oldest
363 * generation is temporarily raised then lowered. */
364 struct generation generations[NUM_GENERATIONS];
366 /* the oldest generation that is will currently be GCed by default.
367 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
369 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
371 * Setting this to 0 effectively disables the generational nature of
372 * the GC. In some applications generational GC may not be useful
373 * because there are no long-lived objects.
375 * An intermediate value could be handy after moving long-lived data
376 * into an older generation so an unnecessary GC of this long-lived
377 * data can be avoided. */
378 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
380 /* The maximum free page in the heap is maintained and used to update
381 * ALLOCATION_POINTER which is used by the room function to limit its
382 * search of the heap. XX Gencgc obviously needs to be better
383 * integrated with the Lisp code. */
384 page_index_t last_free_page;
386 #ifdef LISP_FEATURE_SB_THREAD
387 /* This lock is to prevent multiple threads from simultaneously
388 * allocating new regions which overlap each other. Note that the
389 * majority of GC is single-threaded, but alloc() may be called from
390 * >1 thread at a time and must be thread-safe. This lock must be
391 * seized before all accesses to generations[] or to parts of
392 * page_table[] that other threads may want to see */
393 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
394 /* This lock is used to protect non-thread-local allocation. */
395 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
398 extern os_vm_size_t gencgc_release_granularity;
399 os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
401 extern os_vm_size_t gencgc_alloc_granularity;
402 os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
406 * miscellaneous heap functions
409 /* Count the number of pages which are write-protected within the
410 * given generation. */
412 count_write_protect_generation_pages(generation_index_t generation)
414 page_index_t i, count = 0;
416 for (i = 0; i < last_free_page; i++)
417 if (page_allocated_p(i)
418 && (page_table[i].gen == generation)
419 && (page_table[i].write_protected == 1))
424 /* Count the number of pages within the given generation. */
426 count_generation_pages(generation_index_t generation)
429 page_index_t count = 0;
431 for (i = 0; i < last_free_page; i++)
432 if (page_allocated_p(i)
433 && (page_table[i].gen == generation))
440 count_dont_move_pages(void)
443 page_index_t count = 0;
444 for (i = 0; i < last_free_page; i++) {
445 if (page_allocated_p(i)
446 && (page_table[i].dont_move != 0)) {
454 /* Work through the pages and add up the number of bytes used for the
455 * given generation. */
457 count_generation_bytes_allocated (generation_index_t gen)
460 os_vm_size_t result = 0;
461 for (i = 0; i < last_free_page; i++) {
462 if (page_allocated_p(i)
463 && (page_table[i].gen == gen))
464 result += page_table[i].bytes_used;
469 /* Return the average age of the memory in a generation. */
471 generation_average_age(generation_index_t gen)
473 if (generations[gen].bytes_allocated == 0)
477 ((double)generations[gen].cum_sum_bytes_allocated)
478 / ((double)generations[gen].bytes_allocated);
482 write_generation_stats(FILE *file)
484 generation_index_t i;
486 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
487 #define FPU_STATE_SIZE 27
488 int fpu_state[FPU_STATE_SIZE];
489 #elif defined(LISP_FEATURE_PPC)
490 #define FPU_STATE_SIZE 32
491 long long fpu_state[FPU_STATE_SIZE];
492 #elif defined(LISP_FEATURE_SPARC)
494 * 32 (single-precision) FP registers, and the FP state register.
495 * But Sparc V9 has 32 double-precision registers (equivalent to 64
496 * single-precision, but can't be accessed), so we leave enough room
499 #define FPU_STATE_SIZE (((32 + 32 + 1) + 1)/2)
500 long long fpu_state[FPU_STATE_SIZE];
503 /* This code uses the FP instructions which may be set up for Lisp
504 * so they need to be saved and reset for C. */
507 /* Print the heap stats. */
509 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
511 for (i = 0; i < SCRATCH_GENERATION; i++) {
513 page_index_t boxed_cnt = 0;
514 page_index_t unboxed_cnt = 0;
515 page_index_t large_boxed_cnt = 0;
516 page_index_t large_unboxed_cnt = 0;
517 page_index_t pinned_cnt=0;
519 for (j = 0; j < last_free_page; j++)
520 if (page_table[j].gen == i) {
522 /* Count the number of boxed pages within the given
524 if (page_boxed_p(j)) {
525 if (page_table[j].large_object)
530 if(page_table[j].dont_move) pinned_cnt++;
531 /* Count the number of unboxed pages within the given
533 if (page_unboxed_p(j)) {
534 if (page_table[j].large_object)
541 gc_assert(generations[i].bytes_allocated
542 == count_generation_bytes_allocated(i));
544 " %1d: %5ld %5ld %5ld %5ld",
546 generations[i].alloc_start_page,
547 generations[i].alloc_unboxed_start_page,
548 generations[i].alloc_large_start_page,
549 generations[i].alloc_large_unboxed_start_page);
551 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
552 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
553 boxed_cnt, unboxed_cnt, large_boxed_cnt,
554 large_unboxed_cnt, pinned_cnt);
559 " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
560 generations[i].bytes_allocated,
561 (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
562 generations[i].gc_trigger,
563 count_write_protect_generation_pages(i),
564 generations[i].num_gc,
565 generation_average_age(i));
567 fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
568 fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);
570 fpu_restore(fpu_state);
574 write_heap_exhaustion_report(FILE *file, long available, long requested,
575 struct thread *thread)
578 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
579 gc_active_p ? "garbage collection" : "allocation",
582 write_generation_stats(file);
583 fprintf(file, "GC control variables:\n");
584 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
585 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
586 (SymbolValue(GC_PENDING, thread) == T) ?
587 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
588 "false" : "in progress"));
589 #ifdef LISP_FEATURE_SB_THREAD
590 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
591 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
596 print_generation_stats(void)
598 write_generation_stats(stderr);
601 extern char* gc_logfile;
602 char * gc_logfile = NULL;
605 log_generation_stats(char *logfile, char *header)
608 FILE * log = fopen(logfile, "a");
610 fprintf(log, "%s\n", header);
611 write_generation_stats(log);
614 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
621 report_heap_exhaustion(long available, long requested, struct thread *th)
624 FILE * log = fopen(gc_logfile, "a");
626 write_heap_exhaustion_report(log, available, requested, th);
629 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
633 /* Always to stderr as well. */
634 write_heap_exhaustion_report(stderr, available, requested, th);
638 #if defined(LISP_FEATURE_X86)
639 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
642 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
643 * if zeroing it ourselves, i.e. in practice give the memory back to the
644 * OS. Generally done after a large GC.
646 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
648 void *addr = page_address(start), *new_addr;
649 os_vm_size_t length = npage_bytes(1+end-start);
654 gc_assert(length >= gencgc_release_granularity);
655 gc_assert((length % gencgc_release_granularity) == 0);
657 os_invalidate(addr, length);
658 new_addr = os_validate(addr, length);
659 if (new_addr == NULL || new_addr != addr) {
660 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
664 for (i = start; i <= end; i++) {
665 page_table[i].need_to_zero = 0;
669 /* Zero the pages from START to END (inclusive). Generally done just after
670 * a new region has been allocated.
673 zero_pages(page_index_t start, page_index_t end) {
677 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
678 fast_bzero(page_address(start), npage_bytes(1+end-start));
680 bzero(page_address(start), npage_bytes(1+end-start));
686 zero_and_mark_pages(page_index_t start, page_index_t end) {
689 zero_pages(start, end);
690 for (i = start; i <= end; i++)
691 page_table[i].need_to_zero = 0;
694 /* Zero the pages from START to END (inclusive), except for those
695 * pages that are known to already zeroed. Mark all pages in the
696 * ranges as non-zeroed.
699 zero_dirty_pages(page_index_t start, page_index_t end) {
702 for (i = start; i <= end; i++) {
703 if (!page_table[i].need_to_zero) continue;
704 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
709 for (i = start; i <= end; i++) {
710 page_table[i].need_to_zero = 1;
716 * To support quick and inline allocation, regions of memory can be
717 * allocated and then allocated from with just a free pointer and a
718 * check against an end address.
720 * Since objects can be allocated to spaces with different properties
721 * e.g. boxed/unboxed, generation, ages; there may need to be many
722 * allocation regions.
724 * Each allocation region may start within a partly used page. Many
725 * features of memory use are noted on a page wise basis, e.g. the
726 * generation; so if a region starts within an existing allocated page
727 * it must be consistent with this page.
729 * During the scavenging of the newspace, objects will be transported
730 * into an allocation region, and pointers updated to point to this
731 * allocation region. It is possible that these pointers will be
732 * scavenged again before the allocation region is closed, e.g. due to
733 * trans_list which jumps all over the place to cleanup the list. It
734 * is important to be able to determine properties of all objects
735 * pointed to when scavenging, e.g to detect pointers to the oldspace.
736 * Thus it's important that the allocation regions have the correct
737 * properties set when allocated, and not just set when closed. The
738 * region allocation routines return regions with the specified
739 * properties, and grab all the pages, setting their properties
740 * appropriately, except that the amount used is not known.
742 * These regions are used to support quicker allocation using just a
743 * free pointer. The actual space used by the region is not reflected
744 * in the pages tables until it is closed. It can't be scavenged until
747 * When finished with the region it should be closed, which will
748 * update the page tables for the actual space used returning unused
749 * space. Further it may be noted in the new regions which is
750 * necessary when scavenging the newspace.
752 * Large objects may be allocated directly without an allocation
753 * region, the page tables are updated immediately.
755 * Unboxed objects don't contain pointers to other objects and so
756 * don't need scavenging. Further they can't contain pointers to
757 * younger generations so WP is not needed. By allocating pages to
758 * unboxed objects the whole page never needs scavenging or
759 * write-protecting. */
761 /* We are only using two regions at present. Both are for the current
762 * newspace generation. */
763 struct alloc_region boxed_region;
764 struct alloc_region unboxed_region;
766 /* The generation currently being allocated to. */
767 static generation_index_t gc_alloc_generation;
769 static inline page_index_t
770 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
773 if (UNBOXED_PAGE_FLAG == page_type_flag) {
774 return generations[generation].alloc_large_unboxed_start_page;
775 } else if (BOXED_PAGE_FLAG & page_type_flag) {
776 /* Both code and data. */
777 return generations[generation].alloc_large_start_page;
779 lose("bad page type flag: %d", page_type_flag);
782 if (UNBOXED_PAGE_FLAG == page_type_flag) {
783 return generations[generation].alloc_unboxed_start_page;
784 } else if (BOXED_PAGE_FLAG & page_type_flag) {
785 /* Both code and data. */
786 return generations[generation].alloc_start_page;
788 lose("bad page_type_flag: %d", page_type_flag);
794 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
798 if (UNBOXED_PAGE_FLAG == page_type_flag) {
799 generations[generation].alloc_large_unboxed_start_page = page;
800 } else if (BOXED_PAGE_FLAG & page_type_flag) {
801 /* Both code and data. */
802 generations[generation].alloc_large_start_page = page;
804 lose("bad page type flag: %d", page_type_flag);
807 if (UNBOXED_PAGE_FLAG == page_type_flag) {
808 generations[generation].alloc_unboxed_start_page = page;
809 } else if (BOXED_PAGE_FLAG & page_type_flag) {
810 /* Both code and data. */
811 generations[generation].alloc_start_page = page;
813 lose("bad page type flag: %d", page_type_flag);
818 /* Find a new region with room for at least the given number of bytes.
820 * It starts looking at the current generation's alloc_start_page. So
821 * may pick up from the previous region if there is enough space. This
822 * keeps the allocation contiguous when scavenging the newspace.
824 * The alloc_region should have been closed by a call to
825 * gc_alloc_update_page_tables(), and will thus be in an empty state.
827 * To assist the scavenging functions write-protected pages are not
828 * used. Free pages should not be write-protected.
830 * It is critical to the conservative GC that the start of regions be
831 * known. To help achieve this only small regions are allocated at a
834 * During scavenging, pointers may be found to within the current
835 * region and the page generation must be set so that pointers to the
836 * from space can be recognized. Therefore the generation of pages in
837 * the region are set to gc_alloc_generation. To prevent another
838 * allocation call using the same pages, all the pages in the region
839 * are allocated, although they will initially be empty.
842 gc_alloc_new_region(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
844 page_index_t first_page;
845 page_index_t last_page;
846 os_vm_size_t bytes_found;
852 "/alloc_new_region for %d bytes from gen %d\n",
853 nbytes, gc_alloc_generation));
856 /* Check that the region is in a reset state. */
857 gc_assert((alloc_region->first_page == 0)
858 && (alloc_region->last_page == -1)
859 && (alloc_region->free_pointer == alloc_region->end_addr));
860 ret = thread_mutex_lock(&free_pages_lock);
862 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
863 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
864 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
865 + npage_bytes(last_page-first_page);
867 /* Set up the alloc_region. */
868 alloc_region->first_page = first_page;
869 alloc_region->last_page = last_page;
870 alloc_region->start_addr = page_table[first_page].bytes_used
871 + page_address(first_page);
872 alloc_region->free_pointer = alloc_region->start_addr;
873 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
875 /* Set up the pages. */
877 /* The first page may have already been in use. */
878 if (page_table[first_page].bytes_used == 0) {
879 page_table[first_page].allocated = page_type_flag;
880 page_table[first_page].gen = gc_alloc_generation;
881 page_table[first_page].large_object = 0;
882 page_table[first_page].scan_start_offset = 0;
885 gc_assert(page_table[first_page].allocated == page_type_flag);
886 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
888 gc_assert(page_table[first_page].gen == gc_alloc_generation);
889 gc_assert(page_table[first_page].large_object == 0);
891 for (i = first_page+1; i <= last_page; i++) {
892 page_table[i].allocated = page_type_flag;
893 page_table[i].gen = gc_alloc_generation;
894 page_table[i].large_object = 0;
895 /* This may not be necessary for unboxed regions (think it was
897 page_table[i].scan_start_offset =
898 void_diff(page_address(i),alloc_region->start_addr);
899 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
901 /* Bump up last_free_page. */
902 if (last_page+1 > last_free_page) {
903 last_free_page = last_page+1;
904 /* do we only want to call this on special occasions? like for
906 set_alloc_pointer((lispobj)page_address(last_free_page));
908 ret = thread_mutex_unlock(&free_pages_lock);
911 #ifdef READ_PROTECT_FREE_PAGES
912 os_protect(page_address(first_page),
913 npage_bytes(1+last_page-first_page),
917 /* If the first page was only partial, don't check whether it's
918 * zeroed (it won't be) and don't zero it (since the parts that
919 * we're interested in are guaranteed to be zeroed).
921 if (page_table[first_page].bytes_used) {
925 zero_dirty_pages(first_page, last_page);
927 /* we can do this after releasing free_pages_lock */
928 if (gencgc_zero_check) {
930 for (p = (word_t *)alloc_region->start_addr;
931 p < (word_t *)alloc_region->end_addr; p++) {
933 lose("The new region is not zero at %p (start=%p, end=%p).\n",
934 p, alloc_region->start_addr, alloc_region->end_addr);
940 /* If the record_new_objects flag is 2 then all new regions created
943 * If it's 1 then then it is only recorded if the first page of the
944 * current region is <= new_areas_ignore_page. This helps avoid
945 * unnecessary recording when doing full scavenge pass.
947 * The new_object structure holds the page, byte offset, and size of
948 * new regions of objects. Each new area is placed in the array of
949 * these structures pointer to by new_areas. new_areas_index holds the
950 * offset into new_areas.
952 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
953 * later code must detect this and handle it, probably by doing a full
954 * scavenge of a generation. */
955 #define NUM_NEW_AREAS 512
956 static int record_new_objects = 0;
957 static page_index_t new_areas_ignore_page;
963 static struct new_area (*new_areas)[];
964 static size_t new_areas_index;
965 size_t max_new_areas;
967 /* Add a new area to new_areas. */
969 add_new_area(page_index_t first_page, size_t offset, size_t size)
971 size_t new_area_start, c;
974 /* Ignore if full. */
975 if (new_areas_index >= NUM_NEW_AREAS)
978 switch (record_new_objects) {
982 if (first_page > new_areas_ignore_page)
991 new_area_start = npage_bytes(first_page) + offset;
993 /* Search backwards for a prior area that this follows from. If
994 found this will save adding a new area. */
995 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
997 npage_bytes((*new_areas)[i].page)
998 + (*new_areas)[i].offset
999 + (*new_areas)[i].size;
1001 "/add_new_area S1 %d %d %d %d\n",
1002 i, c, new_area_start, area_end));*/
1003 if (new_area_start == area_end) {
1005 "/adding to [%d] %d %d %d with %d %d %d:\n",
1007 (*new_areas)[i].page,
1008 (*new_areas)[i].offset,
1009 (*new_areas)[i].size,
1013 (*new_areas)[i].size += size;
1018 (*new_areas)[new_areas_index].page = first_page;
1019 (*new_areas)[new_areas_index].offset = offset;
1020 (*new_areas)[new_areas_index].size = size;
1022 "/new_area %d page %d offset %d size %d\n",
1023 new_areas_index, first_page, offset, size));*/
1026 /* Note the max new_areas used. */
1027 if (new_areas_index > max_new_areas)
1028 max_new_areas = new_areas_index;
1031 /* Update the tables for the alloc_region. The region may be added to
1034 * When done the alloc_region is set up so that the next quick alloc
1035 * will fail safely and thus a new region will be allocated. Further
1036 * it is safe to try to re-update the page table of this reset
1039 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
1042 page_index_t first_page;
1043 page_index_t next_page;
1044 os_vm_size_t bytes_used;
1045 os_vm_size_t region_size;
1046 os_vm_size_t byte_cnt;
1047 page_bytes_t orig_first_page_bytes_used;
1051 first_page = alloc_region->first_page;
1053 /* Catch an unused alloc_region. */
1054 if ((first_page == 0) && (alloc_region->last_page == -1))
1057 next_page = first_page+1;
1059 ret = thread_mutex_lock(&free_pages_lock);
1060 gc_assert(ret == 0);
1061 if (alloc_region->free_pointer != alloc_region->start_addr) {
1062 /* some bytes were allocated in the region */
1063 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1065 gc_assert(alloc_region->start_addr ==
1066 (page_address(first_page)
1067 + page_table[first_page].bytes_used));
1069 /* All the pages used need to be updated */
1071 /* Update the first page. */
1073 /* If the page was free then set up the gen, and
1074 * scan_start_offset. */
1075 if (page_table[first_page].bytes_used == 0)
1076 gc_assert(page_starts_contiguous_block_p(first_page));
1077 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1079 gc_assert(page_table[first_page].allocated & page_type_flag);
1080 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1081 gc_assert(page_table[first_page].large_object == 0);
1085 /* Calculate the number of bytes used in this page. This is not
1086 * always the number of new bytes, unless it was free. */
1088 if ((bytes_used = void_diff(alloc_region->free_pointer,
1089 page_address(first_page)))
1090 >GENCGC_CARD_BYTES) {
1091 bytes_used = GENCGC_CARD_BYTES;
1094 page_table[first_page].bytes_used = bytes_used;
1095 byte_cnt += bytes_used;
1098 /* All the rest of the pages should be free. We need to set
1099 * their scan_start_offset pointer to the start of the
1100 * region, and set the bytes_used. */
1102 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1103 gc_assert(page_table[next_page].allocated & page_type_flag);
1104 gc_assert(page_table[next_page].bytes_used == 0);
1105 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1106 gc_assert(page_table[next_page].large_object == 0);
1108 gc_assert(page_table[next_page].scan_start_offset ==
1109 void_diff(page_address(next_page),
1110 alloc_region->start_addr));
1112 /* Calculate the number of bytes used in this page. */
1114 if ((bytes_used = void_diff(alloc_region->free_pointer,
1115 page_address(next_page)))>GENCGC_CARD_BYTES) {
1116 bytes_used = GENCGC_CARD_BYTES;
1119 page_table[next_page].bytes_used = bytes_used;
1120 byte_cnt += bytes_used;
1125 region_size = void_diff(alloc_region->free_pointer,
1126 alloc_region->start_addr);
1127 bytes_allocated += region_size;
1128 generations[gc_alloc_generation].bytes_allocated += region_size;
1130 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1132 /* Set the generations alloc restart page to the last page of
1134 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1136 /* Add the region to the new_areas if requested. */
1137 if (BOXED_PAGE_FLAG & page_type_flag)
1138 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1142 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1144 gc_alloc_generation));
1147 /* There are no bytes allocated. Unallocate the first_page if
1148 * there are 0 bytes_used. */
1149 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1150 if (page_table[first_page].bytes_used == 0)
1151 page_table[first_page].allocated = FREE_PAGE_FLAG;
1154 /* Unallocate any unused pages. */
1155 while (next_page <= alloc_region->last_page) {
1156 gc_assert(page_table[next_page].bytes_used == 0);
1157 page_table[next_page].allocated = FREE_PAGE_FLAG;
1160 ret = thread_mutex_unlock(&free_pages_lock);
1161 gc_assert(ret == 0);
1163 /* alloc_region is per-thread, we're ok to do this unlocked */
1164 gc_set_region_empty(alloc_region);
1167 static inline void *gc_quick_alloc(word_t nbytes);
1169 /* Allocate a possibly large object. */
1171 gc_alloc_large(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
1174 page_index_t first_page, next_page, last_page;
1175 page_bytes_t orig_first_page_bytes_used;
1176 os_vm_size_t byte_cnt;
1177 os_vm_size_t bytes_used;
1180 ret = thread_mutex_lock(&free_pages_lock);
1181 gc_assert(ret == 0);
1183 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1184 if (first_page <= alloc_region->last_page) {
1185 first_page = alloc_region->last_page+1;
1188 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1190 gc_assert(first_page > alloc_region->last_page);
1192 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1194 /* Set up the pages. */
1195 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1197 /* If the first page was free then set up the gen, and
1198 * scan_start_offset. */
1199 if (page_table[first_page].bytes_used == 0) {
1200 page_table[first_page].allocated = page_type_flag;
1201 page_table[first_page].gen = gc_alloc_generation;
1202 page_table[first_page].scan_start_offset = 0;
1203 page_table[first_page].large_object = 1;
1206 gc_assert(page_table[first_page].allocated == page_type_flag);
1207 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1208 gc_assert(page_table[first_page].large_object == 1);
1212 /* Calc. the number of bytes used in this page. This is not
1213 * always the number of new bytes, unless it was free. */
1215 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1216 bytes_used = GENCGC_CARD_BYTES;
1219 page_table[first_page].bytes_used = bytes_used;
1220 byte_cnt += bytes_used;
1222 next_page = first_page+1;
1224 /* All the rest of the pages should be free. We need to set their
1225 * scan_start_offset pointer to the start of the region, and set
1226 * the bytes_used. */
1228 gc_assert(page_free_p(next_page));
1229 gc_assert(page_table[next_page].bytes_used == 0);
1230 page_table[next_page].allocated = page_type_flag;
1231 page_table[next_page].gen = gc_alloc_generation;
1232 page_table[next_page].large_object = 1;
1234 page_table[next_page].scan_start_offset =
1235 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1237 /* Calculate the number of bytes used in this page. */
1239 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1240 if (bytes_used > GENCGC_CARD_BYTES) {
1241 bytes_used = GENCGC_CARD_BYTES;
1244 page_table[next_page].bytes_used = bytes_used;
1245 page_table[next_page].write_protected=0;
1246 page_table[next_page].dont_move=0;
1247 byte_cnt += bytes_used;
1251 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1253 bytes_allocated += nbytes;
1254 generations[gc_alloc_generation].bytes_allocated += nbytes;
1256 /* Add the region to the new_areas if requested. */
1257 if (BOXED_PAGE_FLAG & page_type_flag)
1258 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1260 /* Bump up last_free_page */
1261 if (last_page+1 > last_free_page) {
1262 last_free_page = last_page+1;
1263 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1265 ret = thread_mutex_unlock(&free_pages_lock);
1266 gc_assert(ret == 0);
1268 #ifdef READ_PROTECT_FREE_PAGES
1269 os_protect(page_address(first_page),
1270 npage_bytes(1+last_page-first_page),
1274 zero_dirty_pages(first_page, last_page);
1276 return page_address(first_page);
1279 static page_index_t gencgc_alloc_start_page = -1;
1282 gc_heap_exhausted_error_or_lose (sword_t available, sword_t requested)
1284 struct thread *thread = arch_os_get_current_thread();
1285 /* Write basic information before doing anything else: if we don't
1286 * call to lisp this is a must, and even if we do there is always
1287 * the danger that we bounce back here before the error has been
1288 * handled, or indeed even printed.
1290 report_heap_exhaustion(available, requested, thread);
1291 if (gc_active_p || (available == 0)) {
1292 /* If we are in GC, or totally out of memory there is no way
1293 * to sanely transfer control to the lisp-side of things.
1295 lose("Heap exhausted, game over.");
1298 /* FIXME: assert free_pages_lock held */
1299 (void)thread_mutex_unlock(&free_pages_lock);
1300 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
1301 gc_assert(get_pseudo_atomic_atomic(thread));
1302 clear_pseudo_atomic_atomic(thread);
1303 if (get_pseudo_atomic_interrupted(thread))
1304 do_pending_interrupt();
1306 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1307 * to running user code at arbitrary places, even in a
1308 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1309 * running out of the heap. So at this point all bets are
1311 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1312 corruption_warning_and_maybe_lose
1313 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1314 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1315 alloc_number(available), alloc_number(requested));
1316 lose("HEAP-EXHAUSTED-ERROR fell through");
1321 gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t bytes,
1324 page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
1325 page_index_t first_page, last_page, restart_page = *restart_page_ptr;
1326 os_vm_size_t nbytes = bytes;
1327 os_vm_size_t nbytes_goal = nbytes;
1328 os_vm_size_t bytes_found = 0;
1329 os_vm_size_t most_bytes_found = 0;
1330 boolean small_object = nbytes < GENCGC_CARD_BYTES;
1331 /* FIXME: assert(free_pages_lock is held); */
1333 if (nbytes_goal < gencgc_alloc_granularity)
1334 nbytes_goal = gencgc_alloc_granularity;
1336 /* Toggled by gc_and_save for heap compaction, normally -1. */
1337 if (gencgc_alloc_start_page != -1) {
1338 restart_page = gencgc_alloc_start_page;
1341 /* FIXME: This is on bytes instead of nbytes pending cleanup of
1342 * long from the interface. */
1343 gc_assert(bytes>=0);
1344 /* Search for a page with at least nbytes of space. We prefer
1345 * not to split small objects on multiple pages, to reduce the
1346 * number of contiguous allocation regions spaning multiple
1347 * pages: this helps avoid excessive conservativism.
1349 * For other objects, we guarantee that they start on their own
1352 first_page = restart_page;
1353 while (first_page < page_table_pages) {
1355 if (page_free_p(first_page)) {
1356 gc_assert(0 == page_table[first_page].bytes_used);
1357 bytes_found = GENCGC_CARD_BYTES;
1358 } else if (small_object &&
1359 (page_table[first_page].allocated == page_type_flag) &&
1360 (page_table[first_page].large_object == 0) &&
1361 (page_table[first_page].gen == gc_alloc_generation) &&
1362 (page_table[first_page].write_protected == 0) &&
1363 (page_table[first_page].dont_move == 0)) {
1364 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1365 if (bytes_found < nbytes) {
1366 if (bytes_found > most_bytes_found)
1367 most_bytes_found = bytes_found;
1376 gc_assert(page_table[first_page].write_protected == 0);
1377 for (last_page = first_page+1;
1378 ((last_page < page_table_pages) &&
1379 page_free_p(last_page) &&
1380 (bytes_found < nbytes_goal));
1382 bytes_found += GENCGC_CARD_BYTES;
1383 gc_assert(0 == page_table[last_page].bytes_used);
1384 gc_assert(0 == page_table[last_page].write_protected);
1387 if (bytes_found > most_bytes_found) {
1388 most_bytes_found = bytes_found;
1389 most_bytes_found_from = first_page;
1390 most_bytes_found_to = last_page;
1392 if (bytes_found >= nbytes_goal)
1395 first_page = last_page;
1398 bytes_found = most_bytes_found;
1399 restart_page = first_page + 1;
1401 /* Check for a failure */
1402 if (bytes_found < nbytes) {
1403 gc_assert(restart_page >= page_table_pages);
1404 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1407 gc_assert(most_bytes_found_to);
1408 *restart_page_ptr = most_bytes_found_from;
1409 return most_bytes_found_to-1;
1412 /* Allocate bytes. All the rest of the special-purpose allocation
1413 * functions will eventually call this */
1416 gc_alloc_with_region(sword_t nbytes,int page_type_flag, struct alloc_region *my_region,
1419 void *new_free_pointer;
1421 if (nbytes>=large_object_size)
1422 return gc_alloc_large(nbytes, page_type_flag, my_region);
1424 /* Check whether there is room in the current alloc region. */
1425 new_free_pointer = my_region->free_pointer + nbytes;
1427 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1428 my_region->free_pointer, new_free_pointer); */
1430 if (new_free_pointer <= my_region->end_addr) {
1431 /* If so then allocate from the current alloc region. */
1432 void *new_obj = my_region->free_pointer;
1433 my_region->free_pointer = new_free_pointer;
1435 /* Unless a `quick' alloc was requested, check whether the
1436 alloc region is almost empty. */
1438 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1439 /* If so, finished with the current region. */
1440 gc_alloc_update_page_tables(page_type_flag, my_region);
1441 /* Set up a new region. */
1442 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1445 return((void *)new_obj);
1448 /* Else not enough free space in the current region: retry with a
1451 gc_alloc_update_page_tables(page_type_flag, my_region);
1452 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1453 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1456 /* these are only used during GC: all allocation from the mutator calls
1457 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1460 static inline void *
1461 gc_quick_alloc(word_t nbytes)
1463 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1466 static inline void *
1467 gc_alloc_unboxed(word_t nbytes)
1469 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1472 static inline void *
1473 gc_quick_alloc_unboxed(word_t nbytes)
1475 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1478 /* Copy a large object. If the object is in a large object region then
1479 * it is simply promoted, else it is copied. If it's large enough then
1480 * it's copied to a large object region.
1482 * Bignums and vectors may have shrunk. If the object is not copied
1483 * the space needs to be reclaimed, and the page_tables corrected. */
1485 general_copy_large_object(lispobj object, word_t nwords, boolean boxedp)
1489 page_index_t first_page;
1491 gc_assert(is_lisp_pointer(object));
1492 gc_assert(from_space_p(object));
1493 gc_assert((nwords & 0x01) == 0);
1495 if ((nwords > 1024*1024) && gencgc_verbose) {
1496 FSHOW((stderr, "/general_copy_large_object: %d bytes\n",
1497 nwords*N_WORD_BYTES));
1500 /* Check whether it's a large object. */
1501 first_page = find_page_index((void *)object);
1502 gc_assert(first_page >= 0);
1504 if (page_table[first_page].large_object) {
1505 /* Promote the object. Note: Unboxed objects may have been
1506 * allocated to a BOXED region so it may be necessary to
1507 * change the region to UNBOXED. */
1508 os_vm_size_t remaining_bytes;
1509 os_vm_size_t bytes_freed;
1510 page_index_t next_page;
1511 page_bytes_t old_bytes_used;
1513 /* FIXME: This comment is somewhat stale.
1515 * Note: Any page write-protection must be removed, else a
1516 * later scavenge_newspace may incorrectly not scavenge these
1517 * pages. This would not be necessary if they are added to the
1518 * new areas, but let's do it for them all (they'll probably
1519 * be written anyway?). */
1521 gc_assert(page_starts_contiguous_block_p(first_page));
1522 next_page = first_page;
1523 remaining_bytes = nwords*N_WORD_BYTES;
1525 while (remaining_bytes > GENCGC_CARD_BYTES) {
1526 gc_assert(page_table[next_page].gen == from_space);
1527 gc_assert(page_table[next_page].large_object);
1528 gc_assert(page_table[next_page].scan_start_offset ==
1529 npage_bytes(next_page-first_page));
1530 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1531 /* Should have been unprotected by unprotect_oldspace()
1532 * for boxed objects, and after promotion unboxed ones
1533 * should not be on protected pages at all. */
1534 gc_assert(!page_table[next_page].write_protected);
1537 gc_assert(page_boxed_p(next_page));
1539 gc_assert(page_allocated_no_region_p(next_page));
1540 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1542 page_table[next_page].gen = new_space;
1544 remaining_bytes -= GENCGC_CARD_BYTES;
1548 /* Now only one page remains, but the object may have shrunk so
1549 * there may be more unused pages which will be freed. */
1551 /* Object may have shrunk but shouldn't have grown - check. */
1552 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1554 page_table[next_page].gen = new_space;
1557 gc_assert(page_boxed_p(next_page));
1559 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1561 /* Adjust the bytes_used. */
1562 old_bytes_used = page_table[next_page].bytes_used;
1563 page_table[next_page].bytes_used = remaining_bytes;
1565 bytes_freed = old_bytes_used - remaining_bytes;
1567 /* Free any remaining pages; needs care. */
1569 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1570 (page_table[next_page].gen == from_space) &&
1571 /* FIXME: It is not obvious to me why this is necessary
1572 * as a loop condition: it seems to me that the
1573 * scan_start_offset test should be sufficient, but
1574 * experimentally that is not the case. --NS
1577 page_boxed_p(next_page) :
1578 page_allocated_no_region_p(next_page)) &&
1579 page_table[next_page].large_object &&
1580 (page_table[next_page].scan_start_offset ==
1581 npage_bytes(next_page - first_page))) {
1582 /* Checks out OK, free the page. Don't need to both zeroing
1583 * pages as this should have been done before shrinking the
1584 * object. These pages shouldn't be write-protected, even if
1585 * boxed they should be zero filled. */
1586 gc_assert(page_table[next_page].write_protected == 0);
1588 old_bytes_used = page_table[next_page].bytes_used;
1589 page_table[next_page].allocated = FREE_PAGE_FLAG;
1590 page_table[next_page].bytes_used = 0;
1591 bytes_freed += old_bytes_used;
1595 if ((bytes_freed > 0) && gencgc_verbose) {
1597 "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\n",
1601 generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES
1603 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1604 bytes_allocated -= bytes_freed;
1606 /* Add the region to the new_areas if requested. */
1608 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1613 /* Get tag of object. */
1614 tag = lowtag_of(object);
1616 /* Allocate space. */
1617 new = gc_general_alloc(nwords*N_WORD_BYTES,
1618 (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG),
1621 /* Copy the object. */
1622 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1624 /* Return Lisp pointer of new object. */
1625 return ((lispobj) new) | tag;
1630 copy_large_object(lispobj object, sword_t nwords)
1632 return general_copy_large_object(object, nwords, 1);
1636 copy_large_unboxed_object(lispobj object, sword_t nwords)
1638 return general_copy_large_object(object, nwords, 0);
1641 /* to copy unboxed objects */
1643 copy_unboxed_object(lispobj object, sword_t nwords)
1645 return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG);
1650 * code and code-related objects
1653 static lispobj trans_fun_header(lispobj object);
1654 static lispobj trans_boxed(lispobj object);
1657 /* Scan a x86 compiled code object, looking for possible fixups that
1658 * have been missed after a move.
1660 * Two types of fixups are needed:
1661 * 1. Absolute fixups to within the code object.
1662 * 2. Relative fixups to outside the code object.
1664 * Currently only absolute fixups to the constant vector, or to the
1665 * code area are checked. */
1666 #ifdef LISP_FEATURE_X86
1668 sniff_code_object(struct code *code, os_vm_size_t displacement)
1670 sword_t nheader_words, ncode_words, nwords;
1671 os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
1672 os_vm_address_t code_start_addr, code_end_addr;
1673 os_vm_address_t code_addr = (os_vm_address_t)code;
1674 int fixup_found = 0;
1676 if (!check_code_fixups)
1679 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1681 ncode_words = fixnum_value(code->code_size);
1682 nheader_words = HeaderValue(*(lispobj *)code);
1683 nwords = ncode_words + nheader_words;
1685 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1686 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1687 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1688 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1690 /* Work through the unboxed code. */
1691 for (p = code_start_addr; p < code_end_addr; p++) {
1692 void *data = *(void **)p;
1693 unsigned d1 = *((unsigned char *)p - 1);
1694 unsigned d2 = *((unsigned char *)p - 2);
1695 unsigned d3 = *((unsigned char *)p - 3);
1696 unsigned d4 = *((unsigned char *)p - 4);
1698 unsigned d5 = *((unsigned char *)p - 5);
1699 unsigned d6 = *((unsigned char *)p - 6);
1702 /* Check for code references. */
1703 /* Check for a 32 bit word that looks like an absolute
1704 reference to within the code adea of the code object. */
1705 if ((data >= (void*)(code_start_addr-displacement))
1706 && (data < (void*)(code_end_addr-displacement))) {
1707 /* function header */
1709 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1711 /* Skip the function header */
1715 /* the case of PUSH imm32 */
1719 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1720 p, d6, d5, d4, d3, d2, d1, data));
1721 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1723 /* the case of MOV [reg-8],imm32 */
1725 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1726 || d2==0x45 || d2==0x46 || d2==0x47)
1730 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1731 p, d6, d5, d4, d3, d2, d1, data));
1732 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1734 /* the case of LEA reg,[disp32] */
1735 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1738 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1739 p, d6, d5, d4, d3, d2, d1, data));
1740 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1744 /* Check for constant references. */
1745 /* Check for a 32 bit word that looks like an absolute
1746 reference to within the constant vector. Constant references
1748 if ((data >= (void*)(constants_start_addr-displacement))
1749 && (data < (void*)(constants_end_addr-displacement))
1750 && (((unsigned)data & 0x3) == 0)) {
1755 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1756 p, d6, d5, d4, d3, d2, d1, data));
1757 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1760 /* the case of MOV m32,EAX */
1764 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1765 p, d6, d5, d4, d3, d2, d1, data));
1766 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1769 /* the case of CMP m32,imm32 */
1770 if ((d1 == 0x3d) && (d2 == 0x81)) {
1773 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1774 p, d6, d5, d4, d3, d2, d1, data));
1776 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1779 /* Check for a mod=00, r/m=101 byte. */
1780 if ((d1 & 0xc7) == 5) {
1785 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1786 p, d6, d5, d4, d3, d2, d1, data));
1787 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1789 /* the case of CMP reg32,m32 */
1793 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1794 p, d6, d5, d4, d3, d2, d1, data));
1795 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1797 /* the case of MOV m32,reg32 */
1801 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1802 p, d6, d5, d4, d3, d2, d1, data));
1803 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1805 /* the case of MOV reg32,m32 */
1809 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1810 p, d6, d5, d4, d3, d2, d1, data));
1811 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1813 /* the case of LEA reg32,m32 */
1817 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1818 p, d6, d5, d4, d3, d2, d1, data));
1819 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1825 /* If anything was found, print some information on the code
1829 "/compiled code object at %x: header words = %d, code words = %d\n",
1830 code, nheader_words, ncode_words));
1832 "/const start = %x, end = %x\n",
1833 constants_start_addr, constants_end_addr));
1835 "/code start = %x, end = %x\n",
1836 code_start_addr, code_end_addr));
1841 #ifdef LISP_FEATURE_X86
1843 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1845 sword_t nheader_words, ncode_words, nwords;
1846 os_vm_address_t constants_start_addr, constants_end_addr;
1847 os_vm_address_t code_start_addr, code_end_addr;
1848 os_vm_address_t code_addr = (os_vm_address_t)new_code;
1849 os_vm_address_t old_addr = (os_vm_address_t)old_code;
1850 os_vm_size_t displacement = code_addr - old_addr;
1851 lispobj fixups = NIL;
1852 struct vector *fixups_vector;
1854 ncode_words = fixnum_value(new_code->code_size);
1855 nheader_words = HeaderValue(*(lispobj *)new_code);
1856 nwords = ncode_words + nheader_words;
1858 "/compiled code object at %x: header words = %d, code words = %d\n",
1859 new_code, nheader_words, ncode_words)); */
1860 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1861 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1862 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1863 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1866 "/const start = %x, end = %x\n",
1867 constants_start_addr,constants_end_addr));
1869 "/code start = %x; end = %x\n",
1870 code_start_addr,code_end_addr));
1873 /* The first constant should be a pointer to the fixups for this
1874 code objects. Check. */
1875 fixups = new_code->constants[0];
1877 /* It will be 0 or the unbound-marker if there are no fixups (as
1878 * will be the case if the code object has been purified, for
1879 * example) and will be an other pointer if it is valid. */
1880 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1881 !is_lisp_pointer(fixups)) {
1882 /* Check for possible errors. */
1883 if (check_code_fixups)
1884 sniff_code_object(new_code, displacement);
1889 fixups_vector = (struct vector *)native_pointer(fixups);
1891 /* Could be pointing to a forwarding pointer. */
1892 /* FIXME is this always in from_space? if so, could replace this code with
1893 * forwarding_pointer_p/forwarding_pointer_value */
1894 if (is_lisp_pointer(fixups) &&
1895 (find_page_index((void*)fixups_vector) != -1) &&
1896 (fixups_vector->header == 0x01)) {
1897 /* If so, then follow it. */
1898 /*SHOW("following pointer to a forwarding pointer");*/
1900 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1903 /*SHOW("got fixups");*/
1905 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1906 /* Got the fixups for the code block. Now work through the vector,
1907 and apply a fixup at each address. */
1908 sword_t length = fixnum_value(fixups_vector->length);
1910 for (i = 0; i < length; i++) {
1911 long offset = fixups_vector->data[i];
1912 /* Now check the current value of offset. */
1913 os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);
1915 /* If it's within the old_code object then it must be an
1916 * absolute fixup (relative ones are not saved) */
1917 if ((old_value >= old_addr)
1918 && (old_value < (old_addr + nwords*N_WORD_BYTES)))
1919 /* So add the dispacement. */
1920 *(os_vm_address_t *)(code_start_addr + offset) =
1921 old_value + displacement;
1923 /* It is outside the old code object so it must be a
1924 * relative fixup (absolute fixups are not saved). So
1925 * subtract the displacement. */
1926 *(os_vm_address_t *)(code_start_addr + offset) =
1927 old_value - displacement;
1930 /* This used to just print a note to stderr, but a bogus fixup seems to
1931 * indicate real heap corruption, so a hard hailure is in order. */
1932 lose("fixup vector %p has a bad widetag: %d\n",
1933 fixups_vector, widetag_of(fixups_vector->header));
1936 /* Check for possible errors. */
1937 if (check_code_fixups) {
1938 sniff_code_object(new_code,displacement);
1944 trans_boxed_large(lispobj object)
1949 gc_assert(is_lisp_pointer(object));
1951 header = *((lispobj *) native_pointer(object));
1952 length = HeaderValue(header) + 1;
1953 length = CEILING(length, 2);
1955 return copy_large_object(object, length);
1958 /* Doesn't seem to be used, delete it after the grace period. */
1961 trans_unboxed_large(lispobj object)
1966 gc_assert(is_lisp_pointer(object));
1968 header = *((lispobj *) native_pointer(object));
1969 length = HeaderValue(header) + 1;
1970 length = CEILING(length, 2);
1972 return copy_large_unboxed_object(object, length);
1980 /* XX This is a hack adapted from cgc.c. These don't work too
1981 * efficiently with the gencgc as a list of the weak pointers is
1982 * maintained within the objects which causes writes to the pages. A
1983 * limited attempt is made to avoid unnecessary writes, but this needs
1985 #define WEAK_POINTER_NWORDS \
1986 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
1989 scav_weak_pointer(lispobj *where, lispobj object)
1991 /* Since we overwrite the 'next' field, we have to make
1992 * sure not to do so for pointers already in the list.
1993 * Instead of searching the list of weak_pointers each
1994 * time, we ensure that next is always NULL when the weak
1995 * pointer isn't in the list, and not NULL otherwise.
1996 * Since we can't use NULL to denote end of list, we
1997 * use a pointer back to the same weak_pointer.
1999 struct weak_pointer * wp = (struct weak_pointer*)where;
2001 if (NULL == wp->next) {
2002 wp->next = weak_pointers;
2004 if (NULL == wp->next)
2008 /* Do not let GC scavenge the value slot of the weak pointer.
2009 * (That is why it is a weak pointer.) */
2011 return WEAK_POINTER_NWORDS;
2016 search_read_only_space(void *pointer)
2018 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2019 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2020 if ((pointer < (void *)start) || (pointer >= (void *)end))
2022 return (gc_search_space(start,
2023 (((lispobj *)pointer)+2)-start,
2024 (lispobj *) pointer));
2028 search_static_space(void *pointer)
2030 lispobj *start = (lispobj *)STATIC_SPACE_START;
2031 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2032 if ((pointer < (void *)start) || (pointer >= (void *)end))
2034 return (gc_search_space(start,
2035 (((lispobj *)pointer)+2)-start,
2036 (lispobj *) pointer));
2039 /* a faster version for searching the dynamic space. This will work even
2040 * if the object is in a current allocation region. */
2042 search_dynamic_space(void *pointer)
2044 page_index_t page_index = find_page_index(pointer);
2047 /* The address may be invalid, so do some checks. */
2048 if ((page_index == -1) || page_free_p(page_index))
2050 start = (lispobj *)page_scan_start(page_index);
2051 return (gc_search_space(start,
2052 (((lispobj *)pointer)+2)-start,
2053 (lispobj *)pointer));
2056 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2058 /* Is there any possibility that pointer is a valid Lisp object
2059 * reference, and/or something else (e.g. subroutine call return
2060 * address) which should prevent us from moving the referred-to thing?
2061 * This is called from preserve_pointers() */
2063 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2065 lispobj *start_addr;
2067 /* Find the object start address. */
2068 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2072 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2075 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2078 valid_conservative_root_p(void *addr, page_index_t addr_page_index)
2080 #ifdef GENCGC_IS_PRECISE
2081 /* If we're in precise gencgc (non-x86oid as of this writing) then
2082 * we are only called on valid object pointers in the first place,
2083 * so we just have to do a bounds-check against the heap, a
2084 * generation check, and the already-pinned check. */
2085 if ((addr_page_index == -1)
2086 || (page_table[addr_page_index].gen != from_space)
2087 || (page_table[addr_page_index].dont_move != 0))
2090 /* quick check 1: Address is quite likely to have been invalid. */
2091 if ((addr_page_index == -1)
2092 || page_free_p(addr_page_index)
2093 || (page_table[addr_page_index].bytes_used == 0)
2094 || (page_table[addr_page_index].gen != from_space)
2095 /* Skip if already marked dont_move. */
2096 || (page_table[addr_page_index].dont_move != 0))
2098 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2100 /* quick check 2: Check the offset within the page.
2103 if (((uword_t)addr & (GENCGC_CARD_BYTES - 1)) >
2104 page_table[addr_page_index].bytes_used)
2107 /* Filter out anything which can't be a pointer to a Lisp object
2108 * (or, as a special case which also requires dont_move, a return
2109 * address referring to something in a CodeObject). This is
2110 * expensive but important, since it vastly reduces the
2111 * probability that random garbage will be bogusly interpreted as
2112 * a pointer which prevents a page from moving. */
2113 if (!(code_page_p(addr_page_index)
2114 || (is_lisp_pointer((lispobj)addr) &&
2115 possibly_valid_dynamic_space_pointer(addr))))
2122 /* Adjust large bignum and vector objects. This will adjust the
2123 * allocated region if the size has shrunk, and move unboxed objects
2124 * into unboxed pages. The pages are not promoted here, and the
2125 * promoted region is not added to the new_regions; this is really
2126 * only designed to be called from preserve_pointer(). Shouldn't fail
2127 * if this is missed, just may delay the moving of objects to unboxed
2128 * pages, and the freeing of pages. */
2130 maybe_adjust_large_object(lispobj *where)
2132 page_index_t first_page;
2133 page_index_t next_page;
2136 uword_t remaining_bytes;
2137 uword_t bytes_freed;
2138 uword_t old_bytes_used;
2142 /* Check whether it's a vector or bignum object. */
2143 switch (widetag_of(where[0])) {
2144 case SIMPLE_VECTOR_WIDETAG:
2145 boxed = BOXED_PAGE_FLAG;
2147 case BIGNUM_WIDETAG:
2148 case SIMPLE_BASE_STRING_WIDETAG:
2149 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2150 case SIMPLE_CHARACTER_STRING_WIDETAG:
2152 case SIMPLE_BIT_VECTOR_WIDETAG:
2153 case SIMPLE_ARRAY_NIL_WIDETAG:
2154 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2155 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2156 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2157 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2158 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2159 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2161 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2163 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2164 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2165 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2166 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2168 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2169 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2171 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2172 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2174 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2175 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2178 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2180 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2181 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2183 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2184 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2186 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2187 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2188 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2189 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2191 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2192 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2194 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2195 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2197 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2198 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2200 boxed = UNBOXED_PAGE_FLAG;
2206 /* Find its current size. */
2207 nwords = (sizetab[widetag_of(where[0])])(where);
2209 first_page = find_page_index((void *)where);
2210 gc_assert(first_page >= 0);
2212 /* Note: Any page write-protection must be removed, else a later
2213 * scavenge_newspace may incorrectly not scavenge these pages.
2214 * This would not be necessary if they are added to the new areas,
2215 * but lets do it for them all (they'll probably be written
2218 gc_assert(page_starts_contiguous_block_p(first_page));
2220 next_page = first_page;
2221 remaining_bytes = nwords*N_WORD_BYTES;
2222 while (remaining_bytes > GENCGC_CARD_BYTES) {
2223 gc_assert(page_table[next_page].gen == from_space);
2224 gc_assert(page_allocated_no_region_p(next_page));
2225 gc_assert(page_table[next_page].large_object);
2226 gc_assert(page_table[next_page].scan_start_offset ==
2227 npage_bytes(next_page-first_page));
2228 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2230 page_table[next_page].allocated = boxed;
2232 /* Shouldn't be write-protected at this stage. Essential that the
2234 gc_assert(!page_table[next_page].write_protected);
2235 remaining_bytes -= GENCGC_CARD_BYTES;
2239 /* Now only one page remains, but the object may have shrunk so
2240 * there may be more unused pages which will be freed. */
2242 /* Object may have shrunk but shouldn't have grown - check. */
2243 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2245 page_table[next_page].allocated = boxed;
2246 gc_assert(page_table[next_page].allocated ==
2247 page_table[first_page].allocated);
2249 /* Adjust the bytes_used. */
2250 old_bytes_used = page_table[next_page].bytes_used;
2251 page_table[next_page].bytes_used = remaining_bytes;
2253 bytes_freed = old_bytes_used - remaining_bytes;
2255 /* Free any remaining pages; needs care. */
2257 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2258 (page_table[next_page].gen == from_space) &&
2259 page_allocated_no_region_p(next_page) &&
2260 page_table[next_page].large_object &&
2261 (page_table[next_page].scan_start_offset ==
2262 npage_bytes(next_page - first_page))) {
2263 /* It checks out OK, free the page. We don't need to both zeroing
2264 * pages as this should have been done before shrinking the
2265 * object. These pages shouldn't be write protected as they
2266 * should be zero filled. */
2267 gc_assert(page_table[next_page].write_protected == 0);
2269 old_bytes_used = page_table[next_page].bytes_used;
2270 page_table[next_page].allocated = FREE_PAGE_FLAG;
2271 page_table[next_page].bytes_used = 0;
2272 bytes_freed += old_bytes_used;
2276 if ((bytes_freed > 0) && gencgc_verbose) {
2278 "/maybe_adjust_large_object() freed %d\n",
2282 generations[from_space].bytes_allocated -= bytes_freed;
2283 bytes_allocated -= bytes_freed;
2288 /* Take a possible pointer to a Lisp object and mark its page in the
2289 * page_table so that it will not be relocated during a GC.
2291 * This involves locating the page it points to, then backing up to
2292 * the start of its region, then marking all pages dont_move from there
2293 * up to the first page that's not full or has a different generation
2295 * It is assumed that all the page static flags have been cleared at
2296 * the start of a GC.
2298 * It is also assumed that the current gc_alloc() region has been
2299 * flushed and the tables updated. */
2302 preserve_pointer(void *addr)
2304 page_index_t addr_page_index = find_page_index(addr);
2305 page_index_t first_page;
2307 unsigned int region_allocation;
2309 if (!valid_conservative_root_p(addr, addr_page_index))
2312 /* (Now that we know that addr_page_index is in range, it's
2313 * safe to index into page_table[] with it.) */
2314 region_allocation = page_table[addr_page_index].allocated;
2316 /* Find the beginning of the region. Note that there may be
2317 * objects in the region preceding the one that we were passed a
2318 * pointer to: if this is the case, we will write-protect all the
2319 * previous objects' pages too. */
2322 /* I think this'd work just as well, but without the assertions.
2323 * -dan 2004.01.01 */
2324 first_page = find_page_index(page_scan_start(addr_page_index))
2326 first_page = addr_page_index;
2327 while (!page_starts_contiguous_block_p(first_page)) {
2329 /* Do some checks. */
2330 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2331 gc_assert(page_table[first_page].gen == from_space);
2332 gc_assert(page_table[first_page].allocated == region_allocation);
2336 /* Adjust any large objects before promotion as they won't be
2337 * copied after promotion. */
2338 if (page_table[first_page].large_object) {
2339 /* Large objects (specifically vectors and bignums) can
2340 * shrink, leaving a "tail" of zeroed space, which appears to
2341 * the filter above as a seris of valid conses, both car and
2342 * cdr of which contain the fixnum zero, but will be
2343 * deallocated when the GC shrinks the large object region to
2344 * fit the object within. We allow raw pointers within code
2345 * space, but for boxed and unboxed space we do not, nor do
2346 * pointers to within a non-code object appear valid above. A
2347 * cons cell will never merit allocation to a large object
2348 * page, so pick them off now, before we try to adjust the
2350 if ((lowtag_of((lispobj)addr) == LIST_POINTER_LOWTAG) &&
2351 !code_page_p(first_page)) {
2354 maybe_adjust_large_object(page_address(first_page));
2355 /* It may have moved to unboxed pages. */
2356 region_allocation = page_table[first_page].allocated;
2359 /* Now work forward until the end of this contiguous area is found,
2360 * marking all pages as dont_move. */
2361 for (i = first_page; ;i++) {
2362 gc_assert(page_table[i].allocated == region_allocation);
2364 /* Mark the page static. */
2365 page_table[i].dont_move = 1;
2367 /* It is essential that the pages are not write protected as
2368 * they may have pointers into the old-space which need
2369 * scavenging. They shouldn't be write protected at this
2371 gc_assert(!page_table[i].write_protected);
2373 /* Check whether this is the last page in this contiguous block.. */
2374 if (page_ends_contiguous_block_p(i, from_space))
2378 /* Check that the page is now static. */
2379 gc_assert(page_table[addr_page_index].dont_move != 0);
2382 /* If the given page is not write-protected, then scan it for pointers
2383 * to younger generations or the top temp. generation, if no
2384 * suspicious pointers are found then the page is write-protected.
2386 * Care is taken to check for pointers to the current gc_alloc()
2387 * region if it is a younger generation or the temp. generation. This
2388 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2389 * the gc_alloc_generation does not need to be checked as this is only
2390 * called from scavenge_generation() when the gc_alloc generation is
2391 * younger, so it just checks if there is a pointer to the current
2394 * We return 1 if the page was write-protected, else 0. */
2396 update_page_write_prot(page_index_t page)
2398 generation_index_t gen = page_table[page].gen;
2401 void **page_addr = (void **)page_address(page);
2402 sword_t num_words = page_table[page].bytes_used / N_WORD_BYTES;
2404 /* Shouldn't be a free page. */
2405 gc_assert(page_allocated_p(page));
2406 gc_assert(page_table[page].bytes_used != 0);
2408 /* Skip if it's already write-protected, pinned, or unboxed */
2409 if (page_table[page].write_protected
2410 /* FIXME: What's the reason for not write-protecting pinned pages? */
2411 || page_table[page].dont_move
2412 || page_unboxed_p(page))
2415 /* Scan the page for pointers to younger generations or the
2416 * top temp. generation. */
2418 for (j = 0; j < num_words; j++) {
2419 void *ptr = *(page_addr+j);
2420 page_index_t index = find_page_index(ptr);
2422 /* Check that it's in the dynamic space */
2424 if (/* Does it point to a younger or the temp. generation? */
2425 (page_allocated_p(index)
2426 && (page_table[index].bytes_used != 0)
2427 && ((page_table[index].gen < gen)
2428 || (page_table[index].gen == SCRATCH_GENERATION)))
2430 /* Or does it point within a current gc_alloc() region? */
2431 || ((boxed_region.start_addr <= ptr)
2432 && (ptr <= boxed_region.free_pointer))
2433 || ((unboxed_region.start_addr <= ptr)
2434 && (ptr <= unboxed_region.free_pointer))) {
2441 /* Write-protect the page. */
2442 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2444 os_protect((void *)page_addr,
2446 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2448 /* Note the page as protected in the page tables. */
2449 page_table[page].write_protected = 1;
2455 /* Scavenge all generations from FROM to TO, inclusive, except for
2456 * new_space which needs special handling, as new objects may be
2457 * added which are not checked here - use scavenge_newspace generation.
2459 * Write-protected pages should not have any pointers to the
2460 * from_space so do need scavenging; thus write-protected pages are
2461 * not always scavenged. There is some code to check that these pages
2462 * are not written; but to check fully the write-protected pages need
2463 * to be scavenged by disabling the code to skip them.
2465 * Under the current scheme when a generation is GCed the younger
2466 * generations will be empty. So, when a generation is being GCed it
2467 * is only necessary to scavenge the older generations for pointers
2468 * not the younger. So a page that does not have pointers to younger
2469 * generations does not need to be scavenged.
2471 * The write-protection can be used to note pages that don't have
2472 * pointers to younger pages. But pages can be written without having
2473 * pointers to younger generations. After the pages are scavenged here
2474 * they can be scanned for pointers to younger generations and if
2475 * there are none the page can be write-protected.
2477 * One complication is when the newspace is the top temp. generation.
2479 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2480 * that none were written, which they shouldn't be as they should have
2481 * no pointers to younger generations. This breaks down for weak
2482 * pointers as the objects contain a link to the next and are written
2483 * if a weak pointer is scavenged. Still it's a useful check. */
2485 scavenge_generations(generation_index_t from, generation_index_t to)
2488 page_index_t num_wp = 0;
2492 /* Clear the write_protected_cleared flags on all pages. */
2493 for (i = 0; i < page_table_pages; i++)
2494 page_table[i].write_protected_cleared = 0;
2497 for (i = 0; i < last_free_page; i++) {
2498 generation_index_t generation = page_table[i].gen;
2500 && (page_table[i].bytes_used != 0)
2501 && (generation != new_space)
2502 && (generation >= from)
2503 && (generation <= to)) {
2504 page_index_t last_page,j;
2505 int write_protected=1;
2507 /* This should be the start of a region */
2508 gc_assert(page_starts_contiguous_block_p(i));
2510 /* Now work forward until the end of the region */
2511 for (last_page = i; ; last_page++) {
2513 write_protected && page_table[last_page].write_protected;
2514 if (page_ends_contiguous_block_p(last_page, generation))
2517 if (!write_protected) {
2518 scavenge(page_address(i),
2519 ((uword_t)(page_table[last_page].bytes_used
2520 + npage_bytes(last_page-i)))
2523 /* Now scan the pages and write protect those that
2524 * don't have pointers to younger generations. */
2525 if (enable_page_protection) {
2526 for (j = i; j <= last_page; j++) {
2527 num_wp += update_page_write_prot(j);
2530 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2532 "/write protected %d pages within generation %d\n",
2533 num_wp, generation));
2541 /* Check that none of the write_protected pages in this generation
2542 * have been written to. */
2543 for (i = 0; i < page_table_pages; i++) {
2544 if (page_allocated_p(i)
2545 && (page_table[i].bytes_used != 0)
2546 && (page_table[i].gen == generation)
2547 && (page_table[i].write_protected_cleared != 0)) {
2548 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2550 "/page bytes_used=%d scan_start_offset=%lu dont_move=%d\n",
2551 page_table[i].bytes_used,
2552 page_table[i].scan_start_offset,
2553 page_table[i].dont_move));
2554 lose("write to protected page %d in scavenge_generation()\n", i);
2561 /* Scavenge a newspace generation. As it is scavenged new objects may
2562 * be allocated to it; these will also need to be scavenged. This
2563 * repeats until there are no more objects unscavenged in the
2564 * newspace generation.
2566 * To help improve the efficiency, areas written are recorded by
2567 * gc_alloc() and only these scavenged. Sometimes a little more will be
2568 * scavenged, but this causes no harm. An easy check is done that the
2569 * scavenged bytes equals the number allocated in the previous
2572 * Write-protected pages are not scanned except if they are marked
2573 * dont_move in which case they may have been promoted and still have
2574 * pointers to the from space.
2576 * Write-protected pages could potentially be written by alloc however
2577 * to avoid having to handle re-scavenging of write-protected pages
2578 * gc_alloc() does not write to write-protected pages.
2580 * New areas of objects allocated are recorded alternatively in the two
2581 * new_areas arrays below. */
2582 static struct new_area new_areas_1[NUM_NEW_AREAS];
2583 static struct new_area new_areas_2[NUM_NEW_AREAS];
2585 /* Do one full scan of the new space generation. This is not enough to
2586 * complete the job as new objects may be added to the generation in
2587 * the process which are not scavenged. */
2589 scavenge_newspace_generation_one_scan(generation_index_t generation)
2594 "/starting one full scan of newspace generation %d\n",
2596 for (i = 0; i < last_free_page; i++) {
2597 /* Note that this skips over open regions when it encounters them. */
2599 && (page_table[i].bytes_used != 0)
2600 && (page_table[i].gen == generation)
2601 && ((page_table[i].write_protected == 0)
2602 /* (This may be redundant as write_protected is now
2603 * cleared before promotion.) */
2604 || (page_table[i].dont_move == 1))) {
2605 page_index_t last_page;
2608 /* The scavenge will start at the scan_start_offset of
2611 * We need to find the full extent of this contiguous
2612 * block in case objects span pages.
2614 * Now work forward until the end of this contiguous area
2615 * is found. A small area is preferred as there is a
2616 * better chance of its pages being write-protected. */
2617 for (last_page = i; ;last_page++) {
2618 /* If all pages are write-protected and movable,
2619 * then no need to scavenge */
2620 all_wp=all_wp && page_table[last_page].write_protected &&
2621 !page_table[last_page].dont_move;
2623 /* Check whether this is the last page in this
2624 * contiguous block */
2625 if (page_ends_contiguous_block_p(last_page, generation))
2629 /* Do a limited check for write-protected pages. */
2631 sword_t nwords = (((uword_t)
2632 (page_table[last_page].bytes_used
2633 + npage_bytes(last_page-i)
2634 + page_table[i].scan_start_offset))
2636 new_areas_ignore_page = last_page;
2638 scavenge(page_scan_start(i), nwords);
2645 "/done with one full scan of newspace generation %d\n",
2649 /* Do a complete scavenge of the newspace generation. */
2651 scavenge_newspace_generation(generation_index_t generation)
2655 /* the new_areas array currently being written to by gc_alloc() */
2656 struct new_area (*current_new_areas)[] = &new_areas_1;
2657 size_t current_new_areas_index;
2659 /* the new_areas created by the previous scavenge cycle */
2660 struct new_area (*previous_new_areas)[] = NULL;
2661 size_t previous_new_areas_index;
2663 /* Flush the current regions updating the tables. */
2664 gc_alloc_update_all_page_tables();
2666 /* Turn on the recording of new areas by gc_alloc(). */
2667 new_areas = current_new_areas;
2668 new_areas_index = 0;
2670 /* Don't need to record new areas that get scavenged anyway during
2671 * scavenge_newspace_generation_one_scan. */
2672 record_new_objects = 1;
2674 /* Start with a full scavenge. */
2675 scavenge_newspace_generation_one_scan(generation);
2677 /* Record all new areas now. */
2678 record_new_objects = 2;
2680 /* Give a chance to weak hash tables to make other objects live.
2681 * FIXME: The algorithm implemented here for weak hash table gcing
2682 * is O(W^2+N) as Bruno Haible warns in
2683 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
2684 * see "Implementation 2". */
2685 scav_weak_hash_tables();
2687 /* Flush the current regions updating the tables. */
2688 gc_alloc_update_all_page_tables();
2690 /* Grab new_areas_index. */
2691 current_new_areas_index = new_areas_index;
2694 "The first scan is finished; current_new_areas_index=%d.\n",
2695 current_new_areas_index));*/
2697 while (current_new_areas_index > 0) {
2698 /* Move the current to the previous new areas */
2699 previous_new_areas = current_new_areas;
2700 previous_new_areas_index = current_new_areas_index;
2702 /* Scavenge all the areas in previous new areas. Any new areas
2703 * allocated are saved in current_new_areas. */
2705 /* Allocate an array for current_new_areas; alternating between
2706 * new_areas_1 and 2 */
2707 if (previous_new_areas == &new_areas_1)
2708 current_new_areas = &new_areas_2;
2710 current_new_areas = &new_areas_1;
2712 /* Set up for gc_alloc(). */
2713 new_areas = current_new_areas;
2714 new_areas_index = 0;
2716 /* Check whether previous_new_areas had overflowed. */
2717 if (previous_new_areas_index >= NUM_NEW_AREAS) {
2719 /* New areas of objects allocated have been lost so need to do a
2720 * full scan to be sure! If this becomes a problem try
2721 * increasing NUM_NEW_AREAS. */
2722 if (gencgc_verbose) {
2723 SHOW("new_areas overflow, doing full scavenge");
2726 /* Don't need to record new areas that get scavenged
2727 * anyway during scavenge_newspace_generation_one_scan. */
2728 record_new_objects = 1;
2730 scavenge_newspace_generation_one_scan(generation);
2732 /* Record all new areas now. */
2733 record_new_objects = 2;
2735 scav_weak_hash_tables();
2737 /* Flush the current regions updating the tables. */
2738 gc_alloc_update_all_page_tables();
2742 /* Work through previous_new_areas. */
2743 for (i = 0; i < previous_new_areas_index; i++) {
2744 page_index_t page = (*previous_new_areas)[i].page;
2745 size_t offset = (*previous_new_areas)[i].offset;
2746 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
2747 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
2748 scavenge(page_address(page)+offset, size);
2751 scav_weak_hash_tables();
2753 /* Flush the current regions updating the tables. */
2754 gc_alloc_update_all_page_tables();
2757 current_new_areas_index = new_areas_index;
2760 "The re-scan has finished; current_new_areas_index=%d.\n",
2761 current_new_areas_index));*/
2764 /* Turn off recording of areas allocated by gc_alloc(). */
2765 record_new_objects = 0;
2770 /* Check that none of the write_protected pages in this generation
2771 * have been written to. */
2772 for (i = 0; i < page_table_pages; i++) {
2773 if (page_allocated_p(i)
2774 && (page_table[i].bytes_used != 0)
2775 && (page_table[i].gen == generation)
2776 && (page_table[i].write_protected_cleared != 0)
2777 && (page_table[i].dont_move == 0)) {
2778 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
2779 i, generation, page_table[i].dont_move);
2786 /* Un-write-protect all the pages in from_space. This is done at the
2787 * start of a GC else there may be many page faults while scavenging
2788 * the newspace (I've seen drive the system time to 99%). These pages
2789 * would need to be unprotected anyway before unmapping in
2790 * free_oldspace; not sure what effect this has on paging.. */
2792 unprotect_oldspace(void)
2795 void *region_addr = 0;
2796 void *page_addr = 0;
2797 uword_t region_bytes = 0;
2799 for (i = 0; i < last_free_page; i++) {
2800 if (page_allocated_p(i)
2801 && (page_table[i].bytes_used != 0)
2802 && (page_table[i].gen == from_space)) {
2804 /* Remove any write-protection. We should be able to rely
2805 * on the write-protect flag to avoid redundant calls. */
2806 if (page_table[i].write_protected) {
2807 page_table[i].write_protected = 0;
2808 page_addr = page_address(i);
2811 region_addr = page_addr;
2812 region_bytes = GENCGC_CARD_BYTES;
2813 } else if (region_addr + region_bytes == page_addr) {
2814 /* Region continue. */
2815 region_bytes += GENCGC_CARD_BYTES;
2817 /* Unprotect previous region. */
2818 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2819 /* First page in new region. */
2820 region_addr = page_addr;
2821 region_bytes = GENCGC_CARD_BYTES;
2827 /* Unprotect last region. */
2828 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2832 /* Work through all the pages and free any in from_space. This
2833 * assumes that all objects have been copied or promoted to an older
2834 * generation. Bytes_allocated and the generation bytes_allocated
2835 * counter are updated. The number of bytes freed is returned. */
2839 uword_t bytes_freed = 0;
2840 page_index_t first_page, last_page;
2845 /* Find a first page for the next region of pages. */
2846 while ((first_page < last_free_page)
2847 && (page_free_p(first_page)
2848 || (page_table[first_page].bytes_used == 0)
2849 || (page_table[first_page].gen != from_space)))
2852 if (first_page >= last_free_page)
2855 /* Find the last page of this region. */
2856 last_page = first_page;
2859 /* Free the page. */
2860 bytes_freed += page_table[last_page].bytes_used;
2861 generations[page_table[last_page].gen].bytes_allocated -=
2862 page_table[last_page].bytes_used;
2863 page_table[last_page].allocated = FREE_PAGE_FLAG;
2864 page_table[last_page].bytes_used = 0;
2865 /* Should already be unprotected by unprotect_oldspace(). */
2866 gc_assert(!page_table[last_page].write_protected);
2869 while ((last_page < last_free_page)
2870 && page_allocated_p(last_page)
2871 && (page_table[last_page].bytes_used != 0)
2872 && (page_table[last_page].gen == from_space));
2874 #ifdef READ_PROTECT_FREE_PAGES
2875 os_protect(page_address(first_page),
2876 npage_bytes(last_page-first_page),
2879 first_page = last_page;
2880 } while (first_page < last_free_page);
2882 bytes_allocated -= bytes_freed;
2887 /* Print some information about a pointer at the given address. */
2889 print_ptr(lispobj *addr)
2891 /* If addr is in the dynamic space then out the page information. */
2892 page_index_t pi1 = find_page_index((void*)addr);
2895 fprintf(stderr," %p: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
2898 page_table[pi1].allocated,
2899 page_table[pi1].gen,
2900 page_table[pi1].bytes_used,
2901 page_table[pi1].scan_start_offset,
2902 page_table[pi1].dont_move);
2903 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
2917 is_in_stack_space(lispobj ptr)
2919 /* For space verification: Pointers can be valid if they point
2920 * to a thread stack space. This would be faster if the thread
2921 * structures had page-table entries as if they were part of
2922 * the heap space. */
2924 for_each_thread(th) {
2925 if ((th->control_stack_start <= (lispobj *)ptr) &&
2926 (th->control_stack_end >= (lispobj *)ptr)) {
2934 verify_space(lispobj *start, size_t words)
2936 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
2937 int is_in_readonly_space =
2938 (READ_ONLY_SPACE_START <= (uword_t)start &&
2939 (uword_t)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2943 lispobj thing = *(lispobj*)start;
2945 if (is_lisp_pointer(thing)) {
2946 page_index_t page_index = find_page_index((void*)thing);
2947 sword_t to_readonly_space =
2948 (READ_ONLY_SPACE_START <= thing &&
2949 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2950 sword_t to_static_space =
2951 (STATIC_SPACE_START <= thing &&
2952 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
2954 /* Does it point to the dynamic space? */
2955 if (page_index != -1) {
2956 /* If it's within the dynamic space it should point to a used
2957 * page. XX Could check the offset too. */
2958 if (page_allocated_p(page_index)
2959 && (page_table[page_index].bytes_used == 0))
2960 lose ("Ptr %p @ %p sees free page.\n", thing, start);
2961 /* Check that it doesn't point to a forwarding pointer! */
2962 if (*((lispobj *)native_pointer(thing)) == 0x01) {
2963 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
2965 /* Check that its not in the RO space as it would then be a
2966 * pointer from the RO to the dynamic space. */
2967 if (is_in_readonly_space) {
2968 lose("ptr to dynamic space %p from RO space %x\n",
2971 /* Does it point to a plausible object? This check slows
2972 * it down a lot (so it's commented out).
2974 * "a lot" is serious: it ate 50 minutes cpu time on
2975 * my duron 950 before I came back from lunch and
2978 * FIXME: Add a variable to enable this
2981 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
2982 lose("ptr %p to invalid object %p\n", thing, start);
2986 extern void funcallable_instance_tramp;
2987 /* Verify that it points to another valid space. */
2988 if (!to_readonly_space && !to_static_space
2989 && (thing != (lispobj)&funcallable_instance_tramp)
2990 && !is_in_stack_space(thing)) {
2991 lose("Ptr %p @ %p sees junk.\n", thing, start);
2995 if (!(fixnump(thing))) {
2997 switch(widetag_of(*start)) {
3000 case SIMPLE_VECTOR_WIDETAG:
3002 case COMPLEX_WIDETAG:
3003 case SIMPLE_ARRAY_WIDETAG:
3004 case COMPLEX_BASE_STRING_WIDETAG:
3005 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
3006 case COMPLEX_CHARACTER_STRING_WIDETAG:
3008 case COMPLEX_VECTOR_NIL_WIDETAG:
3009 case COMPLEX_BIT_VECTOR_WIDETAG:
3010 case COMPLEX_VECTOR_WIDETAG:
3011 case COMPLEX_ARRAY_WIDETAG:
3012 case CLOSURE_HEADER_WIDETAG:
3013 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
3014 case VALUE_CELL_HEADER_WIDETAG:
3015 case SYMBOL_HEADER_WIDETAG:
3016 case CHARACTER_WIDETAG:
3017 #if N_WORD_BITS == 64
3018 case SINGLE_FLOAT_WIDETAG:
3020 case UNBOUND_MARKER_WIDETAG:
3025 case INSTANCE_HEADER_WIDETAG:
3028 sword_t ntotal = HeaderValue(thing);
3029 lispobj layout = ((struct instance *)start)->slots[0];
3034 nuntagged = ((struct layout *)
3035 native_pointer(layout))->n_untagged_slots;
3036 verify_space(start + 1,
3037 ntotal - fixnum_value(nuntagged));
3041 case CODE_HEADER_WIDETAG:
3043 lispobj object = *start;
3045 sword_t nheader_words, ncode_words, nwords;
3047 struct simple_fun *fheaderp;
3049 code = (struct code *) start;
3051 /* Check that it's not in the dynamic space.
3052 * FIXME: Isn't is supposed to be OK for code
3053 * objects to be in the dynamic space these days? */
3054 if (is_in_dynamic_space
3055 /* It's ok if it's byte compiled code. The trace
3056 * table offset will be a fixnum if it's x86
3057 * compiled code - check.
3059 * FIXME: #^#@@! lack of abstraction here..
3060 * This line can probably go away now that
3061 * there's no byte compiler, but I've got
3062 * too much to worry about right now to try
3063 * to make sure. -- WHN 2001-10-06 */
3064 && fixnump(code->trace_table_offset)
3065 /* Only when enabled */
3066 && verify_dynamic_code_check) {
3068 "/code object at %p in the dynamic space\n",
3072 ncode_words = fixnum_value(code->code_size);
3073 nheader_words = HeaderValue(object);
3074 nwords = ncode_words + nheader_words;
3075 nwords = CEILING(nwords, 2);
3076 /* Scavenge the boxed section of the code data block */
3077 verify_space(start + 1, nheader_words - 1);
3079 /* Scavenge the boxed section of each function
3080 * object in the code data block. */
3081 fheaderl = code->entry_points;
3082 while (fheaderl != NIL) {
3084 (struct simple_fun *) native_pointer(fheaderl);
3085 gc_assert(widetag_of(fheaderp->header) ==
3086 SIMPLE_FUN_HEADER_WIDETAG);
3087 verify_space(&fheaderp->name, 1);
3088 verify_space(&fheaderp->arglist, 1);
3089 verify_space(&fheaderp->type, 1);
3090 fheaderl = fheaderp->next;
3096 /* unboxed objects */
3097 case BIGNUM_WIDETAG:
3098 #if N_WORD_BITS != 64
3099 case SINGLE_FLOAT_WIDETAG:
3101 case DOUBLE_FLOAT_WIDETAG:
3102 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3103 case LONG_FLOAT_WIDETAG:
3105 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3106 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3108 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3109 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3111 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3112 case COMPLEX_LONG_FLOAT_WIDETAG:
3114 #ifdef SIMD_PACK_WIDETAG
3115 case SIMD_PACK_WIDETAG:
3117 case SIMPLE_BASE_STRING_WIDETAG:
3118 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3119 case SIMPLE_CHARACTER_STRING_WIDETAG:
3121 case SIMPLE_BIT_VECTOR_WIDETAG:
3122 case SIMPLE_ARRAY_NIL_WIDETAG:
3123 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3124 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3125 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3126 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3127 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3128 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3130 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3132 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3133 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3134 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3135 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3137 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3138 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3140 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3141 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3143 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3144 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3147 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3149 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3150 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3152 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3153 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3155 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3156 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3157 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3158 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3160 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3161 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3163 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3164 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3166 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3167 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3170 case WEAK_POINTER_WIDETAG:
3171 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3172 case NO_TLS_VALUE_MARKER_WIDETAG:
3174 count = (sizetab[widetag_of(*start)])(start);
3178 lose("Unhandled widetag %p at %p\n",
3179 widetag_of(*start), start);
3191 /* FIXME: It would be nice to make names consistent so that
3192 * foo_size meant size *in* *bytes* instead of size in some
3193 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3194 * Some counts of lispobjs are called foo_count; it might be good
3195 * to grep for all foo_size and rename the appropriate ones to
3197 sword_t read_only_space_size =
3198 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3199 - (lispobj*)READ_ONLY_SPACE_START;
3200 sword_t static_space_size =
3201 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3202 - (lispobj*)STATIC_SPACE_START;
3204 for_each_thread(th) {
3205 sword_t binding_stack_size =
3206 (lispobj*)get_binding_stack_pointer(th)
3207 - (lispobj*)th->binding_stack_start;
3208 verify_space(th->binding_stack_start, binding_stack_size);
3210 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3211 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3215 verify_generation(generation_index_t generation)
3219 for (i = 0; i < last_free_page; i++) {
3220 if (page_allocated_p(i)
3221 && (page_table[i].bytes_used != 0)
3222 && (page_table[i].gen == generation)) {
3223 page_index_t last_page;
3225 /* This should be the start of a contiguous block */
3226 gc_assert(page_starts_contiguous_block_p(i));
3228 /* Need to find the full extent of this contiguous block in case
3229 objects span pages. */
3231 /* Now work forward until the end of this contiguous area is
3233 for (last_page = i; ;last_page++)
3234 /* Check whether this is the last page in this contiguous
3236 if (page_ends_contiguous_block_p(last_page, generation))
3239 verify_space(page_address(i),
3241 (page_table[last_page].bytes_used
3242 + npage_bytes(last_page-i)))
3249 /* Check that all the free space is zero filled. */
3251 verify_zero_fill(void)
3255 for (page = 0; page < last_free_page; page++) {
3256 if (page_free_p(page)) {
3257 /* The whole page should be zero filled. */
3258 sword_t *start_addr = (sword_t *)page_address(page);
3259 sword_t size = 1024;
3261 for (i = 0; i < size; i++) {
3262 if (start_addr[i] != 0) {
3263 lose("free page not zero at %x\n", start_addr + i);
3267 sword_t free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3268 if (free_bytes > 0) {
3269 sword_t *start_addr = (sword_t *)((uword_t)page_address(page)
3270 + page_table[page].bytes_used);
3271 sword_t size = free_bytes / N_WORD_BYTES;
3273 for (i = 0; i < size; i++) {
3274 if (start_addr[i] != 0) {
3275 lose("free region not zero at %x\n", start_addr + i);
3283 /* External entry point for verify_zero_fill */
3285 gencgc_verify_zero_fill(void)
3287 /* Flush the alloc regions updating the tables. */
3288 gc_alloc_update_all_page_tables();
3289 SHOW("verifying zero fill");
3294 verify_dynamic_space(void)
3296 generation_index_t i;
3298 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3299 verify_generation(i);
3301 if (gencgc_enable_verify_zero_fill)
3305 /* Write-protect all the dynamic boxed pages in the given generation. */
3307 write_protect_generation_pages(generation_index_t generation)
3311 gc_assert(generation < SCRATCH_GENERATION);
3313 for (start = 0; start < last_free_page; start++) {
3314 if (protect_page_p(start, generation)) {
3318 /* Note the page as protected in the page tables. */
3319 page_table[start].write_protected = 1;
3321 for (last = start + 1; last < last_free_page; last++) {
3322 if (!protect_page_p(last, generation))
3324 page_table[last].write_protected = 1;
3327 page_start = (void *)page_address(start);
3329 os_protect(page_start,
3330 npage_bytes(last - start),
3331 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3337 if (gencgc_verbose > 1) {
3339 "/write protected %d of %d pages in generation %d\n",
3340 count_write_protect_generation_pages(generation),
3341 count_generation_pages(generation),
3346 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3348 preserve_context_registers (os_context_t *c)
3351 /* On Darwin the signal context isn't a contiguous block of memory,
3352 * so just preserve_pointering its contents won't be sufficient.
3354 #if defined(LISP_FEATURE_DARWIN)||defined(LISP_FEATURE_WIN32)
3355 #if defined LISP_FEATURE_X86
3356 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3357 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3358 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3359 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3360 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3361 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3362 preserve_pointer((void*)*os_context_pc_addr(c));
3363 #elif defined LISP_FEATURE_X86_64
3364 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3365 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3366 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3367 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3368 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3369 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3370 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3371 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3372 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3373 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3374 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3375 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3376 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3377 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3378 preserve_pointer((void*)*os_context_pc_addr(c));
3380 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3383 #if !defined(LISP_FEATURE_WIN32)
3384 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3385 preserve_pointer(*ptr);
3392 move_pinned_pages_to_newspace()
3396 /* scavenge() will evacuate all oldspace pages, but no newspace
3397 * pages. Pinned pages are precisely those pages which must not
3398 * be evacuated, so move them to newspace directly. */
3400 for (i = 0; i < last_free_page; i++) {
3401 if (page_table[i].dont_move &&
3402 /* dont_move is cleared lazily, so validate the space as well. */
3403 page_table[i].gen == from_space) {
3404 page_table[i].gen = new_space;
3405 /* And since we're moving the pages wholesale, also adjust
3406 * the generation allocation counters. */
3407 generations[new_space].bytes_allocated += page_table[i].bytes_used;
3408 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
3413 /* Garbage collect a generation. If raise is 0 then the remains of the
3414 * generation are not raised to the next generation. */
3416 garbage_collect_generation(generation_index_t generation, int raise)
3418 uword_t bytes_freed;
3420 uword_t static_space_size;
3423 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3425 /* The oldest generation can't be raised. */
3426 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3428 /* Check if weak hash tables were processed in the previous GC. */
3429 gc_assert(weak_hash_tables == NULL);
3431 /* Initialize the weak pointer list. */
3432 weak_pointers = NULL;
3434 /* When a generation is not being raised it is transported to a
3435 * temporary generation (NUM_GENERATIONS), and lowered when
3436 * done. Set up this new generation. There should be no pages
3437 * allocated to it yet. */
3439 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3442 /* Set the global src and dest. generations */
3443 from_space = generation;
3445 new_space = generation+1;
3447 new_space = SCRATCH_GENERATION;
3449 /* Change to a new space for allocation, resetting the alloc_start_page */
3450 gc_alloc_generation = new_space;
3451 generations[new_space].alloc_start_page = 0;
3452 generations[new_space].alloc_unboxed_start_page = 0;
3453 generations[new_space].alloc_large_start_page = 0;
3454 generations[new_space].alloc_large_unboxed_start_page = 0;
3456 /* Before any pointers are preserved, the dont_move flags on the
3457 * pages need to be cleared. */
3458 for (i = 0; i < last_free_page; i++)
3459 if(page_table[i].gen==from_space)
3460 page_table[i].dont_move = 0;
3462 /* Un-write-protect the old-space pages. This is essential for the
3463 * promoted pages as they may contain pointers into the old-space
3464 * which need to be scavenged. It also helps avoid unnecessary page
3465 * faults as forwarding pointers are written into them. They need to
3466 * be un-protected anyway before unmapping later. */
3467 unprotect_oldspace();
3469 /* Scavenge the stacks' conservative roots. */
3471 /* there are potentially two stacks for each thread: the main
3472 * stack, which may contain Lisp pointers, and the alternate stack.
3473 * We don't ever run Lisp code on the altstack, but it may
3474 * host a sigcontext with lisp objects in it */
3476 /* what we need to do: (1) find the stack pointer for the main
3477 * stack; scavenge it (2) find the interrupt context on the
3478 * alternate stack that might contain lisp values, and scavenge
3481 /* we assume that none of the preceding applies to the thread that
3482 * initiates GC. If you ever call GC from inside an altstack
3483 * handler, you will lose. */
3485 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3486 /* And if we're saving a core, there's no point in being conservative. */
3487 if (conservative_stack) {
3488 for_each_thread(th) {
3490 void **esp=(void **)-1;
3491 if (th->state == STATE_DEAD)
3493 # if defined(LISP_FEATURE_SB_SAFEPOINT)
3494 /* Conservative collect_garbage is always invoked with a
3495 * foreign C call or an interrupt handler on top of every
3496 * existing thread, so the stored SP in each thread
3497 * structure is valid, no matter which thread we are looking
3498 * at. For threads that were running Lisp code, the pitstop
3499 * and edge functions maintain this value within the
3500 * interrupt or exception handler. */
3501 esp = os_get_csp(th);
3502 assert_on_stack(th, esp);
3504 /* In addition to pointers on the stack, also preserve the
3505 * return PC, the only value from the context that we need
3506 * in addition to the SP. The return PC gets saved by the
3507 * foreign call wrapper, and removed from the control stack
3508 * into a register. */
3509 preserve_pointer(th->pc_around_foreign_call);
3511 /* And on platforms with interrupts: scavenge ctx registers. */
3513 /* Disabled on Windows, because it does not have an explicit
3514 * stack of `interrupt_contexts'. The reported CSP has been
3515 * chosen so that the current context on the stack is
3516 * covered by the stack scan. See also set_csp_from_context(). */
3517 # ifndef LISP_FEATURE_WIN32
3518 if (th != arch_os_get_current_thread()) {
3519 long k = fixnum_value(
3520 SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3522 preserve_context_registers(th->interrupt_contexts[--k]);
3525 # elif defined(LISP_FEATURE_SB_THREAD)
3527 if(th==arch_os_get_current_thread()) {
3528 /* Somebody is going to burn in hell for this, but casting
3529 * it in two steps shuts gcc up about strict aliasing. */
3530 esp = (void **)((void *)&raise);
3533 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3534 for(i=free-1;i>=0;i--) {
3535 os_context_t *c=th->interrupt_contexts[i];
3536 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3537 if (esp1>=(void **)th->control_stack_start &&
3538 esp1<(void **)th->control_stack_end) {
3539 if(esp1<esp) esp=esp1;
3540 preserve_context_registers(c);
3545 esp = (void **)((void *)&raise);
3547 if (!esp || esp == (void*) -1)
3548 lose("garbage_collect: no SP known for thread %x (OS %x)",
3550 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3551 preserve_pointer(*ptr);
3556 /* Non-x86oid systems don't have "conservative roots" as such, but
3557 * the same mechanism is used for objects pinned for use by alien
3559 for_each_thread(th) {
3560 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
3561 while (pin_list != NIL) {
3562 struct cons *list_entry =
3563 (struct cons *)native_pointer(pin_list);
3564 preserve_pointer(list_entry->car);
3565 pin_list = list_entry->cdr;
3571 if (gencgc_verbose > 1) {
3572 sword_t num_dont_move_pages = count_dont_move_pages();
3574 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3575 num_dont_move_pages,
3576 npage_bytes(num_dont_move_pages));
3580 /* Now that all of the pinned (dont_move) pages are known, and
3581 * before we start to scavenge (and thus relocate) objects,
3582 * relocate the pinned pages to newspace, so that the scavenger
3583 * will not attempt to relocate their contents. */
3584 move_pinned_pages_to_newspace();
3586 /* Scavenge all the rest of the roots. */
3588 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3590 * If not x86, we need to scavenge the interrupt context(s) and the
3595 for_each_thread(th) {
3596 scavenge_interrupt_contexts(th);
3597 scavenge_control_stack(th);
3600 # ifdef LISP_FEATURE_SB_SAFEPOINT
3601 /* In this case, scrub all stacks right here from the GCing thread
3602 * instead of doing what the comment below says. Suboptimal, but
3605 scrub_thread_control_stack(th);
3607 /* Scrub the unscavenged control stack space, so that we can't run
3608 * into any stale pointers in a later GC (this is done by the
3609 * stop-for-gc handler in the other threads). */
3610 scrub_control_stack();
3615 /* Scavenge the Lisp functions of the interrupt handlers, taking
3616 * care to avoid SIG_DFL and SIG_IGN. */
3617 for (i = 0; i < NSIG; i++) {
3618 union interrupt_handler handler = interrupt_handlers[i];
3619 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
3620 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
3621 scavenge((lispobj *)(interrupt_handlers + i), 1);
3624 /* Scavenge the binding stacks. */
3627 for_each_thread(th) {
3628 sword_t len= (lispobj *)get_binding_stack_pointer(th) -
3629 th->binding_stack_start;
3630 scavenge((lispobj *) th->binding_stack_start,len);
3631 #ifdef LISP_FEATURE_SB_THREAD
3632 /* do the tls as well */
3633 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
3634 (sizeof (struct thread))/(sizeof (lispobj));
3635 scavenge((lispobj *) (th+1),len);
3640 /* The original CMU CL code had scavenge-read-only-space code
3641 * controlled by the Lisp-level variable
3642 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
3643 * wasn't documented under what circumstances it was useful or
3644 * safe to turn it on, so it's been turned off in SBCL. If you
3645 * want/need this functionality, and can test and document it,
3646 * please submit a patch. */
3648 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
3649 uword_t read_only_space_size =
3650 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
3651 (lispobj*)READ_ONLY_SPACE_START;
3653 "/scavenge read only space: %d bytes\n",
3654 read_only_space_size * sizeof(lispobj)));
3655 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
3659 /* Scavenge static space. */
3661 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
3662 (lispobj *)STATIC_SPACE_START;
3663 if (gencgc_verbose > 1) {
3665 "/scavenge static space: %d bytes\n",
3666 static_space_size * sizeof(lispobj)));
3668 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
3670 /* All generations but the generation being GCed need to be
3671 * scavenged. The new_space generation needs special handling as
3672 * objects may be moved in - it is handled separately below. */
3673 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
3675 /* Finally scavenge the new_space generation. Keep going until no
3676 * more objects are moved into the new generation */
3677 scavenge_newspace_generation(new_space);
3679 /* FIXME: I tried reenabling this check when debugging unrelated
3680 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
3681 * Since the current GC code seems to work well, I'm guessing that
3682 * this debugging code is just stale, but I haven't tried to
3683 * figure it out. It should be figured out and then either made to
3684 * work or just deleted. */
3685 #define RESCAN_CHECK 0
3687 /* As a check re-scavenge the newspace once; no new objects should
3690 os_vm_size_t old_bytes_allocated = bytes_allocated;
3691 os_vm_size_t bytes_allocated;
3693 /* Start with a full scavenge. */
3694 scavenge_newspace_generation_one_scan(new_space);
3696 /* Flush the current regions, updating the tables. */
3697 gc_alloc_update_all_page_tables();
3699 bytes_allocated = bytes_allocated - old_bytes_allocated;
3701 if (bytes_allocated != 0) {
3702 lose("Rescan of new_space allocated %d more bytes.\n",
3708 scan_weak_hash_tables();
3709 scan_weak_pointers();
3711 /* Flush the current regions, updating the tables. */
3712 gc_alloc_update_all_page_tables();
3714 /* Free the pages in oldspace, but not those marked dont_move. */
3715 bytes_freed = free_oldspace();
3717 /* If the GC is not raising the age then lower the generation back
3718 * to its normal generation number */
3720 for (i = 0; i < last_free_page; i++)
3721 if ((page_table[i].bytes_used != 0)
3722 && (page_table[i].gen == SCRATCH_GENERATION))
3723 page_table[i].gen = generation;
3724 gc_assert(generations[generation].bytes_allocated == 0);
3725 generations[generation].bytes_allocated =
3726 generations[SCRATCH_GENERATION].bytes_allocated;
3727 generations[SCRATCH_GENERATION].bytes_allocated = 0;
3730 /* Reset the alloc_start_page for generation. */
3731 generations[generation].alloc_start_page = 0;
3732 generations[generation].alloc_unboxed_start_page = 0;
3733 generations[generation].alloc_large_start_page = 0;
3734 generations[generation].alloc_large_unboxed_start_page = 0;
3736 if (generation >= verify_gens) {
3737 if (gencgc_verbose) {
3741 verify_dynamic_space();
3744 /* Set the new gc trigger for the GCed generation. */
3745 generations[generation].gc_trigger =
3746 generations[generation].bytes_allocated
3747 + generations[generation].bytes_consed_between_gc;
3750 generations[generation].num_gc = 0;
3752 ++generations[generation].num_gc;
3756 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
3758 update_dynamic_space_free_pointer(void)
3760 page_index_t last_page = -1, i;
3762 for (i = 0; i < last_free_page; i++)
3763 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
3766 last_free_page = last_page+1;
3768 set_alloc_pointer((lispobj)(page_address(last_free_page)));
3769 return 0; /* dummy value: return something ... */
3773 remap_page_range (page_index_t from, page_index_t to)
3775 /* There's a mysterious Solaris/x86 problem with using mmap
3776 * tricks for memory zeroing. See sbcl-devel thread
3777 * "Re: patch: standalone executable redux".
3779 #if defined(LISP_FEATURE_SUNOS)
3780 zero_and_mark_pages(from, to);
3783 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
3784 release_mask = release_granularity-1,
3786 aligned_from = (from+release_mask)&~release_mask,
3787 aligned_end = (end&~release_mask);
3789 if (aligned_from < aligned_end) {
3790 zero_pages_with_mmap(aligned_from, aligned_end-1);
3791 if (aligned_from != from)
3792 zero_and_mark_pages(from, aligned_from-1);
3793 if (aligned_end != end)
3794 zero_and_mark_pages(aligned_end, end-1);
3796 zero_and_mark_pages(from, to);
3802 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
3804 page_index_t first_page, last_page;
3807 return remap_page_range(from, to);
3809 for (first_page = from; first_page <= to; first_page++) {
3810 if (page_allocated_p(first_page) ||
3811 (page_table[first_page].need_to_zero == 0))
3814 last_page = first_page + 1;
3815 while (page_free_p(last_page) &&
3816 (last_page <= to) &&
3817 (page_table[last_page].need_to_zero == 1))
3820 remap_page_range(first_page, last_page-1);
3822 first_page = last_page;
3826 generation_index_t small_generation_limit = 1;
3828 /* GC all generations newer than last_gen, raising the objects in each
3829 * to the next older generation - we finish when all generations below
3830 * last_gen are empty. Then if last_gen is due for a GC, or if
3831 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
3832 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
3834 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
3835 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
3837 collect_garbage(generation_index_t last_gen)
3839 generation_index_t gen = 0, i;
3840 int raise, more = 0;
3842 /* The largest value of last_free_page seen since the time
3843 * remap_free_pages was called. */
3844 static page_index_t high_water_mark = 0;
3846 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
3847 log_generation_stats(gc_logfile, "=== GC Start ===");
3851 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
3853 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
3858 /* Flush the alloc regions updating the tables. */
3859 gc_alloc_update_all_page_tables();
3861 /* Verify the new objects created by Lisp code. */
3862 if (pre_verify_gen_0) {
3863 FSHOW((stderr, "pre-checking generation 0\n"));
3864 verify_generation(0);
3867 if (gencgc_verbose > 1)
3868 print_generation_stats();
3871 /* Collect the generation. */
3873 if (more || (gen >= gencgc_oldest_gen_to_gc)) {
3874 /* Never raise the oldest generation. Never raise the extra generation
3875 * collected due to more-flag. */
3881 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
3882 /* If we would not normally raise this one, but we're
3883 * running low on space in comparison to the object-sizes
3884 * we've been seeing, raise it and collect the next one
3886 if (!raise && gen == last_gen) {
3887 more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated);
3892 if (gencgc_verbose > 1) {
3894 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
3897 generations[gen].bytes_allocated,
3898 generations[gen].gc_trigger,
3899 generations[gen].num_gc));
3902 /* If an older generation is being filled, then update its
3905 generations[gen+1].cum_sum_bytes_allocated +=
3906 generations[gen+1].bytes_allocated;
3909 garbage_collect_generation(gen, raise);
3911 /* Reset the memory age cum_sum. */
3912 generations[gen].cum_sum_bytes_allocated = 0;
3914 if (gencgc_verbose > 1) {
3915 FSHOW((stderr, "GC of generation %d finished:\n", gen));
3916 print_generation_stats();
3920 } while ((gen <= gencgc_oldest_gen_to_gc)
3921 && ((gen < last_gen)
3924 && (generations[gen].bytes_allocated
3925 > generations[gen].gc_trigger)
3926 && (generation_average_age(gen)
3927 > generations[gen].minimum_age_before_gc))));
3929 /* Now if gen-1 was raised all generations before gen are empty.
3930 * If it wasn't raised then all generations before gen-1 are empty.
3932 * Now objects within this gen's pages cannot point to younger
3933 * generations unless they are written to. This can be exploited
3934 * by write-protecting the pages of gen; then when younger
3935 * generations are GCed only the pages which have been written
3940 gen_to_wp = gen - 1;
3942 /* There's not much point in WPing pages in generation 0 as it is
3943 * never scavenged (except promoted pages). */
3944 if ((gen_to_wp > 0) && enable_page_protection) {
3945 /* Check that they are all empty. */
3946 for (i = 0; i < gen_to_wp; i++) {
3947 if (generations[i].bytes_allocated)
3948 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
3951 write_protect_generation_pages(gen_to_wp);
3954 /* Set gc_alloc() back to generation 0. The current regions should
3955 * be flushed after the above GCs. */
3956 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
3957 gc_alloc_generation = 0;
3959 /* Save the high-water mark before updating last_free_page */
3960 if (last_free_page > high_water_mark)
3961 high_water_mark = last_free_page;
3963 update_dynamic_space_free_pointer();
3965 /* Update auto_gc_trigger. Make sure we trigger the next GC before
3966 * running out of heap! */
3967 if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
3968 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
3970 auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;
3973 fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n",
3976 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
3979 if (gen > small_generation_limit) {
3980 if (last_free_page > high_water_mark)
3981 high_water_mark = last_free_page;
3982 remap_free_pages(0, high_water_mark, 0);
3983 high_water_mark = 0;
3987 large_allocation = 0;
3989 log_generation_stats(gc_logfile, "=== GC End ===");
3990 SHOW("returning from collect_garbage");
3993 /* This is called by Lisp PURIFY when it is finished. All live objects
3994 * will have been moved to the RO and Static heaps. The dynamic space
3995 * will need a full re-initialization. We don't bother having Lisp
3996 * PURIFY flush the current gc_alloc() region, as the page_tables are
3997 * re-initialized, and every page is zeroed to be sure. */
4001 page_index_t page, last_page;
4003 if (gencgc_verbose > 1) {
4004 SHOW("entering gc_free_heap");
4007 for (page = 0; page < page_table_pages; page++) {
4008 /* Skip free pages which should already be zero filled. */
4009 if (page_allocated_p(page)) {
4011 for (last_page = page;
4012 (last_page < page_table_pages) && page_allocated_p(last_page);
4014 /* Mark the page free. The other slots are assumed invalid
4015 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
4016 * should not be write-protected -- except that the
4017 * generation is used for the current region but it sets
4019 page_table[page].allocated = FREE_PAGE_FLAG;
4020 page_table[page].bytes_used = 0;
4021 page_table[page].write_protected = 0;
4024 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
4025 * about this change. */
4026 page_start = (void *)page_address(page);
4027 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
4028 remap_free_pages(page, last_page-1, 1);
4031 } else if (gencgc_zero_check_during_free_heap) {
4032 /* Double-check that the page is zero filled. */
4033 sword_t *page_start;
4035 gc_assert(page_free_p(page));
4036 gc_assert(page_table[page].bytes_used == 0);
4037 page_start = (sword_t *)page_address(page);
4038 for (i=0; i<GENCGC_CARD_BYTES/sizeof(sword_t); i++) {
4039 if (page_start[i] != 0) {
4040 lose("free region not zero at %x\n", page_start + i);
4046 bytes_allocated = 0;
4048 /* Initialize the generations. */
4049 for (page = 0; page < NUM_GENERATIONS; page++) {
4050 generations[page].alloc_start_page = 0;
4051 generations[page].alloc_unboxed_start_page = 0;
4052 generations[page].alloc_large_start_page = 0;
4053 generations[page].alloc_large_unboxed_start_page = 0;
4054 generations[page].bytes_allocated = 0;
4055 generations[page].gc_trigger = 2000000;
4056 generations[page].num_gc = 0;
4057 generations[page].cum_sum_bytes_allocated = 0;
4060 if (gencgc_verbose > 1)
4061 print_generation_stats();
4063 /* Initialize gc_alloc(). */
4064 gc_alloc_generation = 0;
4066 gc_set_region_empty(&boxed_region);
4067 gc_set_region_empty(&unboxed_region);
4070 set_alloc_pointer((lispobj)((char *)heap_base));
4072 if (verify_after_free_heap) {
4073 /* Check whether purify has left any bad pointers. */
4074 FSHOW((stderr, "checking after free_heap\n"));
4084 #if defined(LISP_FEATURE_SB_SAFEPOINT)
4088 /* Compute the number of pages needed for the dynamic space.
4089 * Dynamic space size should be aligned on page size. */
4090 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4091 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4093 /* Default nursery size to 5% of the total dynamic space size,
4095 bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
4096 if (bytes_consed_between_gcs < (1024*1024))
4097 bytes_consed_between_gcs = 1024*1024;
4099 /* The page_table must be allocated using "calloc" to initialize
4100 * the page structures correctly. There used to be a separate
4101 * initialization loop (now commented out; see below) but that was
4102 * unnecessary and did hurt startup time. */
4103 page_table = calloc(page_table_pages, sizeof(struct page));
4104 gc_assert(page_table);
4107 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4108 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4110 heap_base = (void*)DYNAMIC_SPACE_START;
4112 /* The page structures are initialized implicitly when page_table
4113 * is allocated with "calloc" above. Formerly we had the following
4114 * explicit initialization here (comments converted to C99 style
4115 * for readability as C's block comments don't nest):
4117 * // Initialize each page structure.
4118 * for (i = 0; i < page_table_pages; i++) {
4119 * // Initialize all pages as free.
4120 * page_table[i].allocated = FREE_PAGE_FLAG;
4121 * page_table[i].bytes_used = 0;
4123 * // Pages are not write-protected at startup.
4124 * page_table[i].write_protected = 0;
4127 * Without this loop the image starts up much faster when dynamic
4128 * space is large -- which it is on 64-bit platforms already by
4129 * default -- and when "calloc" for large arrays is implemented
4130 * using copy-on-write of a page of zeroes -- which it is at least
4131 * on Linux. In this case the pages that page_table_pages is stored
4132 * in are mapped and cleared not before the corresponding part of
4133 * dynamic space is used. For example, this saves clearing 16 MB of
4134 * memory at startup if the page size is 4 KB and the size of
4135 * dynamic space is 4 GB.
4136 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4137 * asserted below: */
4139 /* Compile time assertion: If triggered, declares an array
4140 * of dimension -1 forcing a syntax error. The intent of the
4141 * assignment is to avoid an "unused variable" warning. */
4142 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4143 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4146 bytes_allocated = 0;
4148 /* Initialize the generations.
4150 * FIXME: very similar to code in gc_free_heap(), should be shared */
4151 for (i = 0; i < NUM_GENERATIONS; i++) {
4152 generations[i].alloc_start_page = 0;
4153 generations[i].alloc_unboxed_start_page = 0;
4154 generations[i].alloc_large_start_page = 0;
4155 generations[i].alloc_large_unboxed_start_page = 0;
4156 generations[i].bytes_allocated = 0;
4157 generations[i].gc_trigger = 2000000;
4158 generations[i].num_gc = 0;
4159 generations[i].cum_sum_bytes_allocated = 0;
4160 /* the tune-able parameters */
4161 generations[i].bytes_consed_between_gc
4162 = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
4163 generations[i].number_of_gcs_before_promotion = 1;
4164 generations[i].minimum_age_before_gc = 0.75;
4167 /* Initialize gc_alloc. */
4168 gc_alloc_generation = 0;
4169 gc_set_region_empty(&boxed_region);
4170 gc_set_region_empty(&unboxed_region);
4175 /* Pick up the dynamic space from after a core load.
4177 * The ALLOCATION_POINTER points to the end of the dynamic space.
4181 gencgc_pickup_dynamic(void)
4183 page_index_t page = 0;
4184 void *alloc_ptr = (void *)get_alloc_pointer();
4185 lispobj *prev=(lispobj *)page_address(page);
4186 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4188 bytes_allocated = 0;
4191 lispobj *first,*ptr= (lispobj *)page_address(page);
4193 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4194 /* It is possible, though rare, for the saved page table
4195 * to contain free pages below alloc_ptr. */
4196 page_table[page].gen = gen;
4197 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4198 page_table[page].large_object = 0;
4199 page_table[page].write_protected = 0;
4200 page_table[page].write_protected_cleared = 0;
4201 page_table[page].dont_move = 0;
4202 page_table[page].need_to_zero = 1;
4204 bytes_allocated += GENCGC_CARD_BYTES;
4207 if (!gencgc_partial_pickup) {
4208 page_table[page].allocated = BOXED_PAGE_FLAG;
4209 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4212 page_table[page].scan_start_offset =
4213 page_address(page) - (void *)prev;
4216 } while (page_address(page) < alloc_ptr);
4218 last_free_page = page;
4220 generations[gen].bytes_allocated = bytes_allocated;
4222 gc_alloc_update_all_page_tables();
4223 write_protect_generation_pages(gen);
4227 gc_initialize_pointers(void)
4229 gencgc_pickup_dynamic();
4233 /* alloc(..) is the external interface for memory allocation. It
4234 * allocates to generation 0. It is not called from within the garbage
4235 * collector as it is only external uses that need the check for heap
4236 * size (GC trigger) and to disable the interrupts (interrupts are
4237 * always disabled during a GC).
4239 * The vops that call alloc(..) assume that the returned space is zero-filled.
4240 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4242 * The check for a GC trigger is only performed when the current
4243 * region is full, so in most cases it's not needed. */
4245 static inline lispobj *
4246 general_alloc_internal(sword_t nbytes, int page_type_flag, struct alloc_region *region,
4247 struct thread *thread)
4249 #ifndef LISP_FEATURE_WIN32
4250 lispobj alloc_signal;
4253 void *new_free_pointer;
4254 os_vm_size_t trigger_bytes = 0;
4256 gc_assert(nbytes>0);
4258 /* Check for alignment allocation problems. */
4259 gc_assert((((uword_t)region->free_pointer & LOWTAG_MASK) == 0)
4260 && ((nbytes & LOWTAG_MASK) == 0));
4262 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
4263 /* Must be inside a PA section. */
4264 gc_assert(get_pseudo_atomic_atomic(thread));
4267 if (nbytes > large_allocation)
4268 large_allocation = nbytes;
4270 /* maybe we can do this quickly ... */
4271 new_free_pointer = region->free_pointer + nbytes;
4272 if (new_free_pointer <= region->end_addr) {
4273 new_obj = (void*)(region->free_pointer);
4274 region->free_pointer = new_free_pointer;
4275 return(new_obj); /* yup */
4278 /* We don't want to count nbytes against auto_gc_trigger unless we
4279 * have to: it speeds up the tenuring of objects and slows down
4280 * allocation. However, unless we do so when allocating _very_
4281 * large objects we are in danger of exhausting the heap without
4282 * running sufficient GCs.
4284 if (nbytes >= bytes_consed_between_gcs)
4285 trigger_bytes = nbytes;
4287 /* we have to go the long way around, it seems. Check whether we
4288 * should GC in the near future
4290 if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
4291 /* Don't flood the system with interrupts if the need to gc is
4292 * already noted. This can happen for example when SUB-GC
4293 * allocates or after a gc triggered in a WITHOUT-GCING. */
4294 if (SymbolValue(GC_PENDING,thread) == NIL) {
4295 /* set things up so that GC happens when we finish the PA
4297 SetSymbolValue(GC_PENDING,T,thread);
4298 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4299 #ifdef LISP_FEATURE_SB_SAFEPOINT
4300 thread_register_gc_trigger();
4302 set_pseudo_atomic_interrupted(thread);
4303 #ifdef GENCGC_IS_PRECISE
4304 /* PPC calls alloc() from a trap or from pa_alloc(),
4305 * look up the most context if it's from a trap. */
4307 os_context_t *context =
4308 thread->interrupt_data->allocation_trap_context;
4309 maybe_save_gc_mask_and_block_deferrables
4310 (context ? os_context_sigmask_addr(context) : NULL);
4313 maybe_save_gc_mask_and_block_deferrables(NULL);
4319 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4321 #ifndef LISP_FEATURE_WIN32
4322 /* for sb-prof, and not supported on Windows yet */
4323 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4324 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4325 if ((sword_t) alloc_signal <= 0) {
4326 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4329 SetSymbolValue(ALLOC_SIGNAL,
4330 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4340 general_alloc(sword_t nbytes, int page_type_flag)
4342 struct thread *thread = arch_os_get_current_thread();
4343 /* Select correct region, and call general_alloc_internal with it.
4344 * For other then boxed allocation we must lock first, since the
4345 * region is shared. */
4346 if (BOXED_PAGE_FLAG & page_type_flag) {
4347 #ifdef LISP_FEATURE_SB_THREAD
4348 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4350 struct alloc_region *region = &boxed_region;
4352 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4353 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4355 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4356 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4357 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4360 lose("bad page type flag: %d", page_type_flag);
4364 lispobj AMD64_SYSV_ABI *
4367 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4368 struct thread *self = arch_os_get_current_thread();
4369 int was_pseudo_atomic = get_pseudo_atomic_atomic(self);
4370 if (!was_pseudo_atomic)
4371 set_pseudo_atomic_atomic(self);
4373 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4376 lispobj *result = general_alloc(nbytes, BOXED_PAGE_FLAG);
4378 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4379 if (!was_pseudo_atomic)
4380 clear_pseudo_atomic_atomic(self);
4387 * shared support for the OS-dependent signal handlers which
4388 * catch GENCGC-related write-protect violations
4390 void unhandled_sigmemoryfault(void* addr);
4392 /* Depending on which OS we're running under, different signals might
4393 * be raised for a violation of write protection in the heap. This
4394 * function factors out the common generational GC magic which needs
4395 * to invoked in this case, and should be called from whatever signal
4396 * handler is appropriate for the OS we're running under.
4398 * Return true if this signal is a normal generational GC thing that
4399 * we were able to handle, or false if it was abnormal and control
4400 * should fall through to the general SIGSEGV/SIGBUS/whatever logic.
4402 * We have two control flags for this: one causes us to ignore faults
4403 * on unprotected pages completely, and the second complains to stderr
4404 * but allows us to continue without losing.
4406 extern boolean ignore_memoryfaults_on_unprotected_pages;
4407 boolean ignore_memoryfaults_on_unprotected_pages = 0;
4409 extern boolean continue_after_memoryfault_on_unprotected_pages;
4410 boolean continue_after_memoryfault_on_unprotected_pages = 0;
4413 gencgc_handle_wp_violation(void* fault_addr)
4415 page_index_t page_index = find_page_index(fault_addr);
4418 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4419 fault_addr, page_index));
4422 /* Check whether the fault is within the dynamic space. */
4423 if (page_index == (-1)) {
4425 /* It can be helpful to be able to put a breakpoint on this
4426 * case to help diagnose low-level problems. */
4427 unhandled_sigmemoryfault(fault_addr);
4429 /* not within the dynamic space -- not our responsibility */
4434 ret = thread_mutex_lock(&free_pages_lock);
4435 gc_assert(ret == 0);
4436 if (page_table[page_index].write_protected) {
4437 /* Unprotect the page. */
4438 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4439 page_table[page_index].write_protected_cleared = 1;
4440 page_table[page_index].write_protected = 0;
4441 } else if (!ignore_memoryfaults_on_unprotected_pages) {
4442 /* The only acceptable reason for this signal on a heap
4443 * access is that GENCGC write-protected the page.
4444 * However, if two CPUs hit a wp page near-simultaneously,
4445 * we had better not have the second one lose here if it
4446 * does this test after the first one has already set wp=0
4448 if(page_table[page_index].write_protected_cleared != 1) {
4449 void lisp_backtrace(int frames);
4452 "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n"
4453 " boxed_region.first_page: %"PAGE_INDEX_FMT","
4454 " boxed_region.last_page %"PAGE_INDEX_FMT"\n"
4455 " page.scan_start_offset: %"OS_VM_SIZE_FMT"\n"
4456 " page.bytes_used: %"PAGE_BYTES_FMT"\n"
4457 " page.allocated: %d\n"
4458 " page.write_protected: %d\n"
4459 " page.write_protected_cleared: %d\n"
4460 " page.generation: %d\n",
4463 boxed_region.first_page,
4464 boxed_region.last_page,
4465 page_table[page_index].scan_start_offset,
4466 page_table[page_index].bytes_used,
4467 page_table[page_index].allocated,
4468 page_table[page_index].write_protected,
4469 page_table[page_index].write_protected_cleared,
4470 page_table[page_index].gen);
4471 if (!continue_after_memoryfault_on_unprotected_pages)
4475 ret = thread_mutex_unlock(&free_pages_lock);
4476 gc_assert(ret == 0);
4477 /* Don't worry, we can handle it. */
4481 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4482 * it's not just a case of the program hitting the write barrier, and
4483 * are about to let Lisp deal with it. It's basically just a
4484 * convenient place to set a gdb breakpoint. */
4486 unhandled_sigmemoryfault(void *addr)
4489 void gc_alloc_update_all_page_tables(void)
4491 /* Flush the alloc regions updating the tables. */
4493 for_each_thread(th) {
4494 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4495 #if defined(LISP_FEATURE_SB_SAFEPOINT_STRICTLY) && !defined(LISP_FEATURE_WIN32)
4496 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->sprof_alloc_region);
4499 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4500 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4504 gc_set_region_empty(struct alloc_region *region)
4506 region->first_page = 0;
4507 region->last_page = -1;
4508 region->start_addr = page_address(0);
4509 region->free_pointer = page_address(0);
4510 region->end_addr = page_address(0);
4514 zero_all_free_pages()
4518 for (i = 0; i < last_free_page; i++) {
4519 if (page_free_p(i)) {
4520 #ifdef READ_PROTECT_FREE_PAGES
4521 os_protect(page_address(i),
4530 /* Things to do before doing a final GC before saving a core (without
4533 * + Pages in large_object pages aren't moved by the GC, so we need to
4534 * unset that flag from all pages.
4535 * + The pseudo-static generation isn't normally collected, but it seems
4536 * reasonable to collect it at least when saving a core. So move the
4537 * pages to a normal generation.
4540 prepare_for_final_gc ()
4543 for (i = 0; i < last_free_page; i++) {
4544 page_table[i].large_object = 0;
4545 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4546 int used = page_table[i].bytes_used;
4547 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4548 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4549 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4555 /* Do a non-conservative GC, and then save a core with the initial
4556 * function being set to the value of the static symbol
4557 * SB!VM:RESTART-LISP-FUNCTION */
4559 gc_and_save(char *filename, boolean prepend_runtime,
4560 boolean save_runtime_options, boolean compressed,
4561 int compression_level, int application_type)
4564 void *runtime_bytes = NULL;
4565 size_t runtime_size;
4567 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4572 conservative_stack = 0;
4574 /* The filename might come from Lisp, and be moved by the now
4575 * non-conservative GC. */
4576 filename = strdup(filename);
4578 /* Collect twice: once into relatively high memory, and then back
4579 * into low memory. This compacts the retained data into the lower
4580 * pages, minimizing the size of the core file.
4582 prepare_for_final_gc();
4583 gencgc_alloc_start_page = last_free_page;
4584 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4586 prepare_for_final_gc();
4587 gencgc_alloc_start_page = -1;
4588 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4590 if (prepend_runtime)
4591 save_runtime_to_filehandle(file, runtime_bytes, runtime_size,
4594 /* The dumper doesn't know that pages need to be zeroed before use. */
4595 zero_all_free_pages();
4596 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4597 prepend_runtime, save_runtime_options,
4598 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
4599 /* Oops. Save still managed to fail. Since we've mangled the stack
4600 * beyond hope, there's not much we can do.
4601 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4602 * going to be rather unsatisfactory too... */
4603 lose("Attempt to save core after non-conservative GC failed.\n");