2 * GENerational Conservative Garbage Collector for SBCL
6 * This software is part of the SBCL system. See the README file for
9 * This software is derived from the CMU CL system, which was
10 * written at Carnegie Mellon University and released into the
11 * public domain. The software is in the public domain and is
12 * provided with absolutely no warranty. See the COPYING and CREDITS
13 * files for more information.
17 * For a review of garbage collection techniques (e.g. generational
18 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
19 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
20 * had been accepted for _ACM Computing Surveys_ and was available
21 * as a PostScript preprint through
22 * <http://www.cs.utexas.edu/users/oops/papers.html>
24 * <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
32 #if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD)
33 #include "pthreads_win32.h"
41 #include "interrupt.h"
46 #include "gc-internal.h"
48 #include "pseudo-atomic.h"
50 #include "genesis/vector.h"
51 #include "genesis/weak-pointer.h"
52 #include "genesis/fdefn.h"
53 #include "genesis/simple-fun.h"
55 #include "genesis/hash-table.h"
56 #include "genesis/instance.h"
57 #include "genesis/layout.h"
59 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
60 #include "genesis/cons.h"
63 /* forward declarations */
64 page_index_t gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t nbytes,
72 /* Generations 0-5 are normal collected generations, 6 is only used as
73 * scratch space by the collector, and should never get collected.
76 SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
80 /* Should we use page protection to help avoid the scavenging of pages
81 * that don't have pointers to younger generations? */
82 boolean enable_page_protection = 1;
84 /* the minimum size (in bytes) for a large object*/
85 #if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES)
86 os_vm_size_t large_object_size = 4 * GENCGC_ALLOC_GRANULARITY;
87 #elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY)
88 os_vm_size_t large_object_size = 4 * GENCGC_CARD_BYTES;
90 os_vm_size_t large_object_size = 4 * PAGE_BYTES;
93 /* Largest allocation seen since last GC. */
94 os_vm_size_t large_allocation = 0;
101 /* the verbosity level. All non-error messages are disabled at level 0;
102 * and only a few rare messages are printed at level 1. */
104 boolean gencgc_verbose = 1;
106 boolean gencgc_verbose = 0;
109 /* FIXME: At some point enable the various error-checking things below
110 * and see what they say. */
112 /* We hunt for pointers to old-space, when GCing generations >= verify_gen.
113 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
115 generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;
117 /* Should we do a pre-scan verify of generation 0 before it's GCed? */
118 boolean pre_verify_gen_0 = 0;
120 /* Should we check for bad pointers after gc_free_heap is called
121 * from Lisp PURIFY? */
122 boolean verify_after_free_heap = 0;
124 /* Should we print a note when code objects are found in the dynamic space
125 * during a heap verify? */
126 boolean verify_dynamic_code_check = 0;
128 /* Should we check code objects for fixup errors after they are transported? */
129 boolean check_code_fixups = 0;
131 /* Should we check that newly allocated regions are zero filled? */
132 boolean gencgc_zero_check = 0;
134 /* Should we check that the free space is zero filled? */
135 boolean gencgc_enable_verify_zero_fill = 0;
137 /* Should we check that free pages are zero filled during gc_free_heap
138 * called after Lisp PURIFY? */
139 boolean gencgc_zero_check_during_free_heap = 0;
141 /* When loading a core, don't do a full scan of the memory for the
142 * memory region boundaries. (Set to true by coreparse.c if the core
143 * contained a pagetable entry).
145 boolean gencgc_partial_pickup = 0;
147 /* If defined, free pages are read-protected to ensure that nothing
151 /* #define READ_PROTECT_FREE_PAGES */
155 * GC structures and variables
158 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
159 os_vm_size_t bytes_allocated = 0;
160 os_vm_size_t auto_gc_trigger = 0;
162 /* the source and destination generations. These are set before a GC starts
164 generation_index_t from_space;
165 generation_index_t new_space;
167 /* Set to 1 when in GC */
168 boolean gc_active_p = 0;
170 /* should the GC be conservative on stack. If false (only right before
171 * saving a core), don't scan the stack / mark pages dont_move. */
172 static boolean conservative_stack = 1;
174 /* An array of page structures is allocated on gc initialization.
175 * This helps to quickly map between an address and its page structure.
176 * page_table_pages is set from the size of the dynamic space. */
177 page_index_t page_table_pages;
178 struct page *page_table;
180 static inline boolean page_allocated_p(page_index_t page) {
181 return (page_table[page].allocated != FREE_PAGE_FLAG);
184 static inline boolean page_no_region_p(page_index_t page) {
185 return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
188 static inline boolean page_allocated_no_region_p(page_index_t page) {
189 return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
190 && page_no_region_p(page));
193 static inline boolean page_free_p(page_index_t page) {
194 return (page_table[page].allocated == FREE_PAGE_FLAG);
197 static inline boolean page_boxed_p(page_index_t page) {
198 return (page_table[page].allocated & BOXED_PAGE_FLAG);
201 static inline boolean code_page_p(page_index_t page) {
202 return (page_table[page].allocated & CODE_PAGE_FLAG);
205 static inline boolean page_boxed_no_region_p(page_index_t page) {
206 return page_boxed_p(page) && page_no_region_p(page);
209 static inline boolean page_unboxed_p(page_index_t page) {
210 /* Both flags set == boxed code page */
211 return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
212 && !page_boxed_p(page));
215 static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
216 return (page_boxed_no_region_p(page)
217 && (page_table[page].bytes_used != 0)
218 && !page_table[page].dont_move
219 && (page_table[page].gen == generation));
222 /* To map addresses to page structures the address of the first page
224 void *heap_base = NULL;
226 /* Calculate the start address for the given page number. */
228 page_address(page_index_t page_num)
230 return (heap_base + (page_num * GENCGC_CARD_BYTES));
233 /* Calculate the address where the allocation region associated with
234 * the page starts. */
236 page_scan_start(page_index_t page_index)
238 return page_address(page_index)-page_table[page_index].scan_start_offset;
241 /* True if the page starts a contiguous block. */
242 static inline boolean
243 page_starts_contiguous_block_p(page_index_t page_index)
245 return page_table[page_index].scan_start_offset == 0;
248 /* True if the page is the last page in a contiguous block. */
249 static inline boolean
250 page_ends_contiguous_block_p(page_index_t page_index, generation_index_t gen)
252 return (/* page doesn't fill block */
253 (page_table[page_index].bytes_used < GENCGC_CARD_BYTES)
254 /* page is last allocated page */
255 || ((page_index + 1) >= last_free_page)
257 || page_free_p(page_index + 1)
258 /* next page contains no data */
259 || (page_table[page_index + 1].bytes_used == 0)
260 /* next page is in different generation */
261 || (page_table[page_index + 1].gen != gen)
262 /* next page starts its own contiguous block */
263 || (page_starts_contiguous_block_p(page_index + 1)));
266 /* Find the page index within the page_table for the given
267 * address. Return -1 on failure. */
269 find_page_index(void *addr)
271 if (addr >= heap_base) {
272 page_index_t index = ((pointer_sized_uint_t)addr -
273 (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES;
274 if (index < page_table_pages)
281 npage_bytes(page_index_t npages)
283 gc_assert(npages>=0);
284 return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
287 /* Check that X is a higher address than Y and return offset from Y to
289 static inline os_vm_size_t
290 void_diff(void *x, void *y)
293 return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
296 /* a structure to hold the state of a generation
298 * CAUTION: If you modify this, make sure to touch up the alien
299 * definition in src/code/gc.lisp accordingly. ...or better yes,
300 * deal with the FIXME there...
304 /* the first page that gc_alloc() checks on its next call */
305 page_index_t alloc_start_page;
307 /* the first page that gc_alloc_unboxed() checks on its next call */
308 page_index_t alloc_unboxed_start_page;
310 /* the first page that gc_alloc_large (boxed) considers on its next
311 * call. (Although it always allocates after the boxed_region.) */
312 page_index_t alloc_large_start_page;
314 /* the first page that gc_alloc_large (unboxed) considers on its
315 * next call. (Although it always allocates after the
316 * current_unboxed_region.) */
317 page_index_t alloc_large_unboxed_start_page;
319 /* the bytes allocated to this generation */
320 os_vm_size_t bytes_allocated;
322 /* the number of bytes at which to trigger a GC */
323 os_vm_size_t gc_trigger;
325 /* to calculate a new level for gc_trigger */
326 os_vm_size_t bytes_consed_between_gc;
328 /* the number of GCs since the last raise */
331 /* the number of GCs to run on the generations before raising objects to the
333 int number_of_gcs_before_promotion;
335 /* the cumulative sum of the bytes allocated to this generation. It is
336 * cleared after a GC on this generations, and update before new
337 * objects are added from a GC of a younger generation. Dividing by
338 * the bytes_allocated will give the average age of the memory in
339 * this generation since its last GC. */
340 os_vm_size_t cum_sum_bytes_allocated;
342 /* a minimum average memory age before a GC will occur helps
343 * prevent a GC when a large number of new live objects have been
344 * added, in which case a GC could be a waste of time */
345 double minimum_age_before_gc;
348 /* an array of generation structures. There needs to be one more
349 * generation structure than actual generations as the oldest
350 * generation is temporarily raised then lowered. */
351 struct generation generations[NUM_GENERATIONS];
353 /* the oldest generation that is will currently be GCed by default.
354 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
356 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
358 * Setting this to 0 effectively disables the generational nature of
359 * the GC. In some applications generational GC may not be useful
360 * because there are no long-lived objects.
362 * An intermediate value could be handy after moving long-lived data
363 * into an older generation so an unnecessary GC of this long-lived
364 * data can be avoided. */
365 generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;
367 /* The maximum free page in the heap is maintained and used to update
368 * ALLOCATION_POINTER which is used by the room function to limit its
369 * search of the heap. XX Gencgc obviously needs to be better
370 * integrated with the Lisp code. */
371 page_index_t last_free_page;
373 #ifdef LISP_FEATURE_SB_THREAD
374 /* This lock is to prevent multiple threads from simultaneously
375 * allocating new regions which overlap each other. Note that the
376 * majority of GC is single-threaded, but alloc() may be called from
377 * >1 thread at a time and must be thread-safe. This lock must be
378 * seized before all accesses to generations[] or to parts of
379 * page_table[] that other threads may want to see */
380 static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
381 /* This lock is used to protect non-thread-local allocation. */
382 static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
385 extern os_vm_size_t gencgc_release_granularity;
386 os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;
388 extern os_vm_size_t gencgc_alloc_granularity;
389 os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;
393 * miscellaneous heap functions
396 /* Count the number of pages which are write-protected within the
397 * given generation. */
399 count_write_protect_generation_pages(generation_index_t generation)
401 page_index_t i, count = 0;
403 for (i = 0; i < last_free_page; i++)
404 if (page_allocated_p(i)
405 && (page_table[i].gen == generation)
406 && (page_table[i].write_protected == 1))
411 /* Count the number of pages within the given generation. */
413 count_generation_pages(generation_index_t generation)
416 page_index_t count = 0;
418 for (i = 0; i < last_free_page; i++)
419 if (page_allocated_p(i)
420 && (page_table[i].gen == generation))
427 count_dont_move_pages(void)
430 page_index_t count = 0;
431 for (i = 0; i < last_free_page; i++) {
432 if (page_allocated_p(i)
433 && (page_table[i].dont_move != 0)) {
441 /* Work through the pages and add up the number of bytes used for the
442 * given generation. */
444 count_generation_bytes_allocated (generation_index_t gen)
447 os_vm_size_t result = 0;
448 for (i = 0; i < last_free_page; i++) {
449 if (page_allocated_p(i)
450 && (page_table[i].gen == gen))
451 result += page_table[i].bytes_used;
456 /* Return the average age of the memory in a generation. */
458 generation_average_age(generation_index_t gen)
460 if (generations[gen].bytes_allocated == 0)
464 ((double)generations[gen].cum_sum_bytes_allocated)
465 / ((double)generations[gen].bytes_allocated);
469 write_generation_stats(FILE *file)
471 generation_index_t i;
473 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
474 #define FPU_STATE_SIZE 27
475 int fpu_state[FPU_STATE_SIZE];
476 #elif defined(LISP_FEATURE_PPC)
477 #define FPU_STATE_SIZE 32
478 long long fpu_state[FPU_STATE_SIZE];
479 #elif defined(LISP_FEATURE_SPARC)
481 * 32 (single-precision) FP registers, and the FP state register.
482 * But Sparc V9 has 32 double-precision registers (equivalent to 64
483 * single-precision, but can't be accessed), so we leave enough room
486 #define FPU_STATE_SIZE (((32 + 32 + 1) + 1)/2)
487 long long fpu_state[FPU_STATE_SIZE];
490 /* This code uses the FP instructions which may be set up for Lisp
491 * so they need to be saved and reset for C. */
494 /* Print the heap stats. */
496 " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
498 for (i = 0; i < SCRATCH_GENERATION; i++) {
500 page_index_t boxed_cnt = 0;
501 page_index_t unboxed_cnt = 0;
502 page_index_t large_boxed_cnt = 0;
503 page_index_t large_unboxed_cnt = 0;
504 page_index_t pinned_cnt=0;
506 for (j = 0; j < last_free_page; j++)
507 if (page_table[j].gen == i) {
509 /* Count the number of boxed pages within the given
511 if (page_boxed_p(j)) {
512 if (page_table[j].large_object)
517 if(page_table[j].dont_move) pinned_cnt++;
518 /* Count the number of unboxed pages within the given
520 if (page_unboxed_p(j)) {
521 if (page_table[j].large_object)
528 gc_assert(generations[i].bytes_allocated
529 == count_generation_bytes_allocated(i));
531 " %1d: %5ld %5ld %5ld %5ld",
533 generations[i].alloc_start_page,
534 generations[i].alloc_unboxed_start_page,
535 generations[i].alloc_large_start_page,
536 generations[i].alloc_large_unboxed_start_page);
538 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
539 " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
540 boxed_cnt, unboxed_cnt, large_boxed_cnt,
541 large_unboxed_cnt, pinned_cnt);
546 " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
547 generations[i].bytes_allocated,
548 (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
549 generations[i].gc_trigger,
550 count_write_protect_generation_pages(i),
551 generations[i].num_gc,
552 generation_average_age(i));
554 fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
555 fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);
557 fpu_restore(fpu_state);
561 write_heap_exhaustion_report(FILE *file, long available, long requested,
562 struct thread *thread)
565 "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
566 gc_active_p ? "garbage collection" : "allocation",
569 write_generation_stats(file);
570 fprintf(file, "GC control variables:\n");
571 fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n",
572 SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
573 (SymbolValue(GC_PENDING, thread) == T) ?
574 "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
575 "false" : "in progress"));
576 #ifdef LISP_FEATURE_SB_THREAD
577 fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n",
578 SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
583 print_generation_stats(void)
585 write_generation_stats(stderr);
588 extern char* gc_logfile;
589 char * gc_logfile = NULL;
592 log_generation_stats(char *logfile, char *header)
595 FILE * log = fopen(logfile, "a");
597 fprintf(log, "%s\n", header);
598 write_generation_stats(log);
601 fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
608 report_heap_exhaustion(long available, long requested, struct thread *th)
611 FILE * log = fopen(gc_logfile, "a");
613 write_heap_exhaustion_report(log, available, requested, th);
616 fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
620 /* Always to stderr as well. */
621 write_heap_exhaustion_report(stderr, available, requested, th);
625 #if defined(LISP_FEATURE_X86)
626 void fast_bzero(void*, size_t); /* in <arch>-assem.S */
629 /* Zero the pages from START to END (inclusive), but use mmap/munmap instead
630 * if zeroing it ourselves, i.e. in practice give the memory back to the
631 * OS. Generally done after a large GC.
633 void zero_pages_with_mmap(page_index_t start, page_index_t end) {
635 void *addr = page_address(start), *new_addr;
636 os_vm_size_t length = npage_bytes(1+end-start);
641 gc_assert(length >= gencgc_release_granularity);
642 gc_assert((length % gencgc_release_granularity) == 0);
644 os_invalidate(addr, length);
645 new_addr = os_validate(addr, length);
646 if (new_addr == NULL || new_addr != addr) {
647 lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
651 for (i = start; i <= end; i++) {
652 page_table[i].need_to_zero = 0;
656 /* Zero the pages from START to END (inclusive). Generally done just after
657 * a new region has been allocated.
660 zero_pages(page_index_t start, page_index_t end) {
664 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
665 fast_bzero(page_address(start), npage_bytes(1+end-start));
667 bzero(page_address(start), npage_bytes(1+end-start));
673 zero_and_mark_pages(page_index_t start, page_index_t end) {
676 zero_pages(start, end);
677 for (i = start; i <= end; i++)
678 page_table[i].need_to_zero = 0;
681 /* Zero the pages from START to END (inclusive), except for those
682 * pages that are known to already zeroed. Mark all pages in the
683 * ranges as non-zeroed.
686 zero_dirty_pages(page_index_t start, page_index_t end) {
689 for (i = start; i <= end; i++) {
690 if (!page_table[i].need_to_zero) continue;
691 for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
696 for (i = start; i <= end; i++) {
697 page_table[i].need_to_zero = 1;
703 * To support quick and inline allocation, regions of memory can be
704 * allocated and then allocated from with just a free pointer and a
705 * check against an end address.
707 * Since objects can be allocated to spaces with different properties
708 * e.g. boxed/unboxed, generation, ages; there may need to be many
709 * allocation regions.
711 * Each allocation region may start within a partly used page. Many
712 * features of memory use are noted on a page wise basis, e.g. the
713 * generation; so if a region starts within an existing allocated page
714 * it must be consistent with this page.
716 * During the scavenging of the newspace, objects will be transported
717 * into an allocation region, and pointers updated to point to this
718 * allocation region. It is possible that these pointers will be
719 * scavenged again before the allocation region is closed, e.g. due to
720 * trans_list which jumps all over the place to cleanup the list. It
721 * is important to be able to determine properties of all objects
722 * pointed to when scavenging, e.g to detect pointers to the oldspace.
723 * Thus it's important that the allocation regions have the correct
724 * properties set when allocated, and not just set when closed. The
725 * region allocation routines return regions with the specified
726 * properties, and grab all the pages, setting their properties
727 * appropriately, except that the amount used is not known.
729 * These regions are used to support quicker allocation using just a
730 * free pointer. The actual space used by the region is not reflected
731 * in the pages tables until it is closed. It can't be scavenged until
734 * When finished with the region it should be closed, which will
735 * update the page tables for the actual space used returning unused
736 * space. Further it may be noted in the new regions which is
737 * necessary when scavenging the newspace.
739 * Large objects may be allocated directly without an allocation
740 * region, the page tables are updated immediately.
742 * Unboxed objects don't contain pointers to other objects and so
743 * don't need scavenging. Further they can't contain pointers to
744 * younger generations so WP is not needed. By allocating pages to
745 * unboxed objects the whole page never needs scavenging or
746 * write-protecting. */
748 /* We are only using two regions at present. Both are for the current
749 * newspace generation. */
750 struct alloc_region boxed_region;
751 struct alloc_region unboxed_region;
753 /* The generation currently being allocated to. */
754 static generation_index_t gc_alloc_generation;
756 static inline page_index_t
757 generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
760 if (UNBOXED_PAGE_FLAG == page_type_flag) {
761 return generations[generation].alloc_large_unboxed_start_page;
762 } else if (BOXED_PAGE_FLAG & page_type_flag) {
763 /* Both code and data. */
764 return generations[generation].alloc_large_start_page;
766 lose("bad page type flag: %d", page_type_flag);
769 if (UNBOXED_PAGE_FLAG == page_type_flag) {
770 return generations[generation].alloc_unboxed_start_page;
771 } else if (BOXED_PAGE_FLAG & page_type_flag) {
772 /* Both code and data. */
773 return generations[generation].alloc_start_page;
775 lose("bad page_type_flag: %d", page_type_flag);
781 set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
785 if (UNBOXED_PAGE_FLAG == page_type_flag) {
786 generations[generation].alloc_large_unboxed_start_page = page;
787 } else if (BOXED_PAGE_FLAG & page_type_flag) {
788 /* Both code and data. */
789 generations[generation].alloc_large_start_page = page;
791 lose("bad page type flag: %d", page_type_flag);
794 if (UNBOXED_PAGE_FLAG == page_type_flag) {
795 generations[generation].alloc_unboxed_start_page = page;
796 } else if (BOXED_PAGE_FLAG & page_type_flag) {
797 /* Both code and data. */
798 generations[generation].alloc_start_page = page;
800 lose("bad page type flag: %d", page_type_flag);
805 /* Find a new region with room for at least the given number of bytes.
807 * It starts looking at the current generation's alloc_start_page. So
808 * may pick up from the previous region if there is enough space. This
809 * keeps the allocation contiguous when scavenging the newspace.
811 * The alloc_region should have been closed by a call to
812 * gc_alloc_update_page_tables(), and will thus be in an empty state.
814 * To assist the scavenging functions write-protected pages are not
815 * used. Free pages should not be write-protected.
817 * It is critical to the conservative GC that the start of regions be
818 * known. To help achieve this only small regions are allocated at a
821 * During scavenging, pointers may be found to within the current
822 * region and the page generation must be set so that pointers to the
823 * from space can be recognized. Therefore the generation of pages in
824 * the region are set to gc_alloc_generation. To prevent another
825 * allocation call using the same pages, all the pages in the region
826 * are allocated, although they will initially be empty.
829 gc_alloc_new_region(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
831 page_index_t first_page;
832 page_index_t last_page;
833 os_vm_size_t bytes_found;
839 "/alloc_new_region for %d bytes from gen %d\n",
840 nbytes, gc_alloc_generation));
843 /* Check that the region is in a reset state. */
844 gc_assert((alloc_region->first_page == 0)
845 && (alloc_region->last_page == -1)
846 && (alloc_region->free_pointer == alloc_region->end_addr));
847 ret = thread_mutex_lock(&free_pages_lock);
849 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
850 last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
851 bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
852 + npage_bytes(last_page-first_page);
854 /* Set up the alloc_region. */
855 alloc_region->first_page = first_page;
856 alloc_region->last_page = last_page;
857 alloc_region->start_addr = page_table[first_page].bytes_used
858 + page_address(first_page);
859 alloc_region->free_pointer = alloc_region->start_addr;
860 alloc_region->end_addr = alloc_region->start_addr + bytes_found;
862 /* Set up the pages. */
864 /* The first page may have already been in use. */
865 if (page_table[first_page].bytes_used == 0) {
866 page_table[first_page].allocated = page_type_flag;
867 page_table[first_page].gen = gc_alloc_generation;
868 page_table[first_page].large_object = 0;
869 page_table[first_page].scan_start_offset = 0;
872 gc_assert(page_table[first_page].allocated == page_type_flag);
873 page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;
875 gc_assert(page_table[first_page].gen == gc_alloc_generation);
876 gc_assert(page_table[first_page].large_object == 0);
878 for (i = first_page+1; i <= last_page; i++) {
879 page_table[i].allocated = page_type_flag;
880 page_table[i].gen = gc_alloc_generation;
881 page_table[i].large_object = 0;
882 /* This may not be necessary for unboxed regions (think it was
884 page_table[i].scan_start_offset =
885 void_diff(page_address(i),alloc_region->start_addr);
886 page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
888 /* Bump up last_free_page. */
889 if (last_page+1 > last_free_page) {
890 last_free_page = last_page+1;
891 /* do we only want to call this on special occasions? like for
893 set_alloc_pointer((lispobj)page_address(last_free_page));
895 ret = thread_mutex_unlock(&free_pages_lock);
898 #ifdef READ_PROTECT_FREE_PAGES
899 os_protect(page_address(first_page),
900 npage_bytes(1+last_page-first_page),
904 /* If the first page was only partial, don't check whether it's
905 * zeroed (it won't be) and don't zero it (since the parts that
906 * we're interested in are guaranteed to be zeroed).
908 if (page_table[first_page].bytes_used) {
912 zero_dirty_pages(first_page, last_page);
914 /* we can do this after releasing free_pages_lock */
915 if (gencgc_zero_check) {
917 for (p = (word_t *)alloc_region->start_addr;
918 p < (word_t *)alloc_region->end_addr; p++) {
920 lose("The new region is not zero at %p (start=%p, end=%p).\n",
921 p, alloc_region->start_addr, alloc_region->end_addr);
927 /* If the record_new_objects flag is 2 then all new regions created
930 * If it's 1 then then it is only recorded if the first page of the
931 * current region is <= new_areas_ignore_page. This helps avoid
932 * unnecessary recording when doing full scavenge pass.
934 * The new_object structure holds the page, byte offset, and size of
935 * new regions of objects. Each new area is placed in the array of
936 * these structures pointer to by new_areas. new_areas_index holds the
937 * offset into new_areas.
939 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
940 * later code must detect this and handle it, probably by doing a full
941 * scavenge of a generation. */
942 #define NUM_NEW_AREAS 512
943 static int record_new_objects = 0;
944 static page_index_t new_areas_ignore_page;
950 static struct new_area (*new_areas)[];
951 static size_t new_areas_index;
952 size_t max_new_areas;
954 /* Add a new area to new_areas. */
956 add_new_area(page_index_t first_page, size_t offset, size_t size)
958 size_t new_area_start, c;
961 /* Ignore if full. */
962 if (new_areas_index >= NUM_NEW_AREAS)
965 switch (record_new_objects) {
969 if (first_page > new_areas_ignore_page)
978 new_area_start = npage_bytes(first_page) + offset;
980 /* Search backwards for a prior area that this follows from. If
981 found this will save adding a new area. */
982 for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
984 npage_bytes((*new_areas)[i].page)
985 + (*new_areas)[i].offset
986 + (*new_areas)[i].size;
988 "/add_new_area S1 %d %d %d %d\n",
989 i, c, new_area_start, area_end));*/
990 if (new_area_start == area_end) {
992 "/adding to [%d] %d %d %d with %d %d %d:\n",
994 (*new_areas)[i].page,
995 (*new_areas)[i].offset,
996 (*new_areas)[i].size,
1000 (*new_areas)[i].size += size;
1005 (*new_areas)[new_areas_index].page = first_page;
1006 (*new_areas)[new_areas_index].offset = offset;
1007 (*new_areas)[new_areas_index].size = size;
1009 "/new_area %d page %d offset %d size %d\n",
1010 new_areas_index, first_page, offset, size));*/
1013 /* Note the max new_areas used. */
1014 if (new_areas_index > max_new_areas)
1015 max_new_areas = new_areas_index;
1018 /* Update the tables for the alloc_region. The region may be added to
1021 * When done the alloc_region is set up so that the next quick alloc
1022 * will fail safely and thus a new region will be allocated. Further
1023 * it is safe to try to re-update the page table of this reset
1026 gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
1029 page_index_t first_page;
1030 page_index_t next_page;
1031 os_vm_size_t bytes_used;
1032 os_vm_size_t region_size;
1033 os_vm_size_t byte_cnt;
1034 page_bytes_t orig_first_page_bytes_used;
1038 first_page = alloc_region->first_page;
1040 /* Catch an unused alloc_region. */
1041 if ((first_page == 0) && (alloc_region->last_page == -1))
1044 next_page = first_page+1;
1046 ret = thread_mutex_lock(&free_pages_lock);
1047 gc_assert(ret == 0);
1048 if (alloc_region->free_pointer != alloc_region->start_addr) {
1049 /* some bytes were allocated in the region */
1050 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1052 gc_assert(alloc_region->start_addr ==
1053 (page_address(first_page)
1054 + page_table[first_page].bytes_used));
1056 /* All the pages used need to be updated */
1058 /* Update the first page. */
1060 /* If the page was free then set up the gen, and
1061 * scan_start_offset. */
1062 if (page_table[first_page].bytes_used == 0)
1063 gc_assert(page_starts_contiguous_block_p(first_page));
1064 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1066 gc_assert(page_table[first_page].allocated & page_type_flag);
1067 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1068 gc_assert(page_table[first_page].large_object == 0);
1072 /* Calculate the number of bytes used in this page. This is not
1073 * always the number of new bytes, unless it was free. */
1075 if ((bytes_used = void_diff(alloc_region->free_pointer,
1076 page_address(first_page)))
1077 >GENCGC_CARD_BYTES) {
1078 bytes_used = GENCGC_CARD_BYTES;
1081 page_table[first_page].bytes_used = bytes_used;
1082 byte_cnt += bytes_used;
1085 /* All the rest of the pages should be free. We need to set
1086 * their scan_start_offset pointer to the start of the
1087 * region, and set the bytes_used. */
1089 page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1090 gc_assert(page_table[next_page].allocated & page_type_flag);
1091 gc_assert(page_table[next_page].bytes_used == 0);
1092 gc_assert(page_table[next_page].gen == gc_alloc_generation);
1093 gc_assert(page_table[next_page].large_object == 0);
1095 gc_assert(page_table[next_page].scan_start_offset ==
1096 void_diff(page_address(next_page),
1097 alloc_region->start_addr));
1099 /* Calculate the number of bytes used in this page. */
1101 if ((bytes_used = void_diff(alloc_region->free_pointer,
1102 page_address(next_page)))>GENCGC_CARD_BYTES) {
1103 bytes_used = GENCGC_CARD_BYTES;
1106 page_table[next_page].bytes_used = bytes_used;
1107 byte_cnt += bytes_used;
1112 region_size = void_diff(alloc_region->free_pointer,
1113 alloc_region->start_addr);
1114 bytes_allocated += region_size;
1115 generations[gc_alloc_generation].bytes_allocated += region_size;
1117 gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);
1119 /* Set the generations alloc restart page to the last page of
1121 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);
1123 /* Add the region to the new_areas if requested. */
1124 if (BOXED_PAGE_FLAG & page_type_flag)
1125 add_new_area(first_page,orig_first_page_bytes_used, region_size);
1129 "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
1131 gc_alloc_generation));
1134 /* There are no bytes allocated. Unallocate the first_page if
1135 * there are 0 bytes_used. */
1136 page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
1137 if (page_table[first_page].bytes_used == 0)
1138 page_table[first_page].allocated = FREE_PAGE_FLAG;
1141 /* Unallocate any unused pages. */
1142 while (next_page <= alloc_region->last_page) {
1143 gc_assert(page_table[next_page].bytes_used == 0);
1144 page_table[next_page].allocated = FREE_PAGE_FLAG;
1147 ret = thread_mutex_unlock(&free_pages_lock);
1148 gc_assert(ret == 0);
1150 /* alloc_region is per-thread, we're ok to do this unlocked */
1151 gc_set_region_empty(alloc_region);
1154 static inline void *gc_quick_alloc(word_t nbytes);
1156 /* Allocate a possibly large object. */
1158 gc_alloc_large(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
1161 page_index_t first_page, next_page, last_page;
1162 page_bytes_t orig_first_page_bytes_used;
1163 os_vm_size_t byte_cnt;
1164 os_vm_size_t bytes_used;
1167 ret = thread_mutex_lock(&free_pages_lock);
1168 gc_assert(ret == 0);
1170 first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
1171 if (first_page <= alloc_region->last_page) {
1172 first_page = alloc_region->last_page+1;
1175 last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);
1177 gc_assert(first_page > alloc_region->last_page);
1179 set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);
1181 /* Set up the pages. */
1182 orig_first_page_bytes_used = page_table[first_page].bytes_used;
1184 /* If the first page was free then set up the gen, and
1185 * scan_start_offset. */
1186 if (page_table[first_page].bytes_used == 0) {
1187 page_table[first_page].allocated = page_type_flag;
1188 page_table[first_page].gen = gc_alloc_generation;
1189 page_table[first_page].scan_start_offset = 0;
1190 page_table[first_page].large_object = 1;
1193 gc_assert(page_table[first_page].allocated == page_type_flag);
1194 gc_assert(page_table[first_page].gen == gc_alloc_generation);
1195 gc_assert(page_table[first_page].large_object == 1);
1199 /* Calc. the number of bytes used in this page. This is not
1200 * always the number of new bytes, unless it was free. */
1202 if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
1203 bytes_used = GENCGC_CARD_BYTES;
1206 page_table[first_page].bytes_used = bytes_used;
1207 byte_cnt += bytes_used;
1209 next_page = first_page+1;
1211 /* All the rest of the pages should be free. We need to set their
1212 * scan_start_offset pointer to the start of the region, and set
1213 * the bytes_used. */
1215 gc_assert(page_free_p(next_page));
1216 gc_assert(page_table[next_page].bytes_used == 0);
1217 page_table[next_page].allocated = page_type_flag;
1218 page_table[next_page].gen = gc_alloc_generation;
1219 page_table[next_page].large_object = 1;
1221 page_table[next_page].scan_start_offset =
1222 npage_bytes(next_page-first_page) - orig_first_page_bytes_used;
1224 /* Calculate the number of bytes used in this page. */
1226 bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
1227 if (bytes_used > GENCGC_CARD_BYTES) {
1228 bytes_used = GENCGC_CARD_BYTES;
1231 page_table[next_page].bytes_used = bytes_used;
1232 page_table[next_page].write_protected=0;
1233 page_table[next_page].dont_move=0;
1234 byte_cnt += bytes_used;
1238 gc_assert((byte_cnt-orig_first_page_bytes_used) == nbytes);
1240 bytes_allocated += nbytes;
1241 generations[gc_alloc_generation].bytes_allocated += nbytes;
1243 /* Add the region to the new_areas if requested. */
1244 if (BOXED_PAGE_FLAG & page_type_flag)
1245 add_new_area(first_page,orig_first_page_bytes_used,nbytes);
1247 /* Bump up last_free_page */
1248 if (last_page+1 > last_free_page) {
1249 last_free_page = last_page+1;
1250 set_alloc_pointer((lispobj)(page_address(last_free_page)));
1252 ret = thread_mutex_unlock(&free_pages_lock);
1253 gc_assert(ret == 0);
1255 #ifdef READ_PROTECT_FREE_PAGES
1256 os_protect(page_address(first_page),
1257 npage_bytes(1+last_page-first_page),
1261 zero_dirty_pages(first_page, last_page);
1263 return page_address(first_page);
1266 static page_index_t gencgc_alloc_start_page = -1;
1269 gc_heap_exhausted_error_or_lose (sword_t available, sword_t requested)
1271 struct thread *thread = arch_os_get_current_thread();
1272 /* Write basic information before doing anything else: if we don't
1273 * call to lisp this is a must, and even if we do there is always
1274 * the danger that we bounce back here before the error has been
1275 * handled, or indeed even printed.
1277 report_heap_exhaustion(available, requested, thread);
1278 if (gc_active_p || (available == 0)) {
1279 /* If we are in GC, or totally out of memory there is no way
1280 * to sanely transfer control to the lisp-side of things.
1282 lose("Heap exhausted, game over.");
1285 /* FIXME: assert free_pages_lock held */
1286 (void)thread_mutex_unlock(&free_pages_lock);
1287 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
1288 gc_assert(get_pseudo_atomic_atomic(thread));
1289 clear_pseudo_atomic_atomic(thread);
1290 if (get_pseudo_atomic_interrupted(thread))
1291 do_pending_interrupt();
1293 /* Another issue is that signalling HEAP-EXHAUSTED error leads
1294 * to running user code at arbitrary places, even in a
1295 * WITHOUT-INTERRUPTS which may lead to a deadlock without
1296 * running out of the heap. So at this point all bets are
1298 if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
1299 corruption_warning_and_maybe_lose
1300 ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
1301 funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR),
1302 alloc_number(available), alloc_number(requested));
1303 lose("HEAP-EXHAUSTED-ERROR fell through");
1308 gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t bytes,
1311 page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
1312 page_index_t first_page, last_page, restart_page = *restart_page_ptr;
1313 os_vm_size_t nbytes = bytes;
1314 os_vm_size_t nbytes_goal = nbytes;
1315 os_vm_size_t bytes_found = 0;
1316 os_vm_size_t most_bytes_found = 0;
1317 boolean small_object = nbytes < GENCGC_CARD_BYTES;
1318 /* FIXME: assert(free_pages_lock is held); */
1320 if (nbytes_goal < gencgc_alloc_granularity)
1321 nbytes_goal = gencgc_alloc_granularity;
1323 /* Toggled by gc_and_save for heap compaction, normally -1. */
1324 if (gencgc_alloc_start_page != -1) {
1325 restart_page = gencgc_alloc_start_page;
1328 /* FIXME: This is on bytes instead of nbytes pending cleanup of
1329 * long from the interface. */
1330 gc_assert(bytes>=0);
1331 /* Search for a page with at least nbytes of space. We prefer
1332 * not to split small objects on multiple pages, to reduce the
1333 * number of contiguous allocation regions spaning multiple
1334 * pages: this helps avoid excessive conservativism.
1336 * For other objects, we guarantee that they start on their own
1339 first_page = restart_page;
1340 while (first_page < page_table_pages) {
1342 if (page_free_p(first_page)) {
1343 gc_assert(0 == page_table[first_page].bytes_used);
1344 bytes_found = GENCGC_CARD_BYTES;
1345 } else if (small_object &&
1346 (page_table[first_page].allocated == page_type_flag) &&
1347 (page_table[first_page].large_object == 0) &&
1348 (page_table[first_page].gen == gc_alloc_generation) &&
1349 (page_table[first_page].write_protected == 0) &&
1350 (page_table[first_page].dont_move == 0)) {
1351 bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
1352 if (bytes_found < nbytes) {
1353 if (bytes_found > most_bytes_found)
1354 most_bytes_found = bytes_found;
1363 gc_assert(page_table[first_page].write_protected == 0);
1364 for (last_page = first_page+1;
1365 ((last_page < page_table_pages) &&
1366 page_free_p(last_page) &&
1367 (bytes_found < nbytes_goal));
1369 bytes_found += GENCGC_CARD_BYTES;
1370 gc_assert(0 == page_table[last_page].bytes_used);
1371 gc_assert(0 == page_table[last_page].write_protected);
1374 if (bytes_found > most_bytes_found) {
1375 most_bytes_found = bytes_found;
1376 most_bytes_found_from = first_page;
1377 most_bytes_found_to = last_page;
1379 if (bytes_found >= nbytes_goal)
1382 first_page = last_page;
1385 bytes_found = most_bytes_found;
1386 restart_page = first_page + 1;
1388 /* Check for a failure */
1389 if (bytes_found < nbytes) {
1390 gc_assert(restart_page >= page_table_pages);
1391 gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
1394 gc_assert(most_bytes_found_to);
1395 *restart_page_ptr = most_bytes_found_from;
1396 return most_bytes_found_to-1;
1399 /* Allocate bytes. All the rest of the special-purpose allocation
1400 * functions will eventually call this */
1403 gc_alloc_with_region(sword_t nbytes,int page_type_flag, struct alloc_region *my_region,
1406 void *new_free_pointer;
1408 if (nbytes>=large_object_size)
1409 return gc_alloc_large(nbytes, page_type_flag, my_region);
1411 /* Check whether there is room in the current alloc region. */
1412 new_free_pointer = my_region->free_pointer + nbytes;
1414 /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
1415 my_region->free_pointer, new_free_pointer); */
1417 if (new_free_pointer <= my_region->end_addr) {
1418 /* If so then allocate from the current alloc region. */
1419 void *new_obj = my_region->free_pointer;
1420 my_region->free_pointer = new_free_pointer;
1422 /* Unless a `quick' alloc was requested, check whether the
1423 alloc region is almost empty. */
1425 void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
1426 /* If so, finished with the current region. */
1427 gc_alloc_update_page_tables(page_type_flag, my_region);
1428 /* Set up a new region. */
1429 gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
1432 return((void *)new_obj);
1435 /* Else not enough free space in the current region: retry with a
1438 gc_alloc_update_page_tables(page_type_flag, my_region);
1439 gc_alloc_new_region(nbytes, page_type_flag, my_region);
1440 return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
1443 /* these are only used during GC: all allocation from the mutator calls
1444 * alloc() -> gc_alloc_with_region() with the appropriate per-thread
1447 static inline void *
1448 gc_quick_alloc(word_t nbytes)
1450 return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK);
1453 static inline void *
1454 gc_alloc_unboxed(word_t nbytes)
1456 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0);
1459 static inline void *
1460 gc_quick_alloc_unboxed(word_t nbytes)
1462 return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK);
1465 /* Copy a large object. If the object is in a large object region then
1466 * it is simply promoted, else it is copied. If it's large enough then
1467 * it's copied to a large object region.
1469 * Bignums and vectors may have shrunk. If the object is not copied
1470 * the space needs to be reclaimed, and the page_tables corrected. */
1472 general_copy_large_object(lispobj object, word_t nwords, boolean boxedp)
1476 page_index_t first_page;
1478 gc_assert(is_lisp_pointer(object));
1479 gc_assert(from_space_p(object));
1480 gc_assert((nwords & 0x01) == 0);
1482 if ((nwords > 1024*1024) && gencgc_verbose) {
1483 FSHOW((stderr, "/general_copy_large_object: %d bytes\n",
1484 nwords*N_WORD_BYTES));
1487 /* Check whether it's a large object. */
1488 first_page = find_page_index((void *)object);
1489 gc_assert(first_page >= 0);
1491 if (page_table[first_page].large_object) {
1492 /* Promote the object. Note: Unboxed objects may have been
1493 * allocated to a BOXED region so it may be necessary to
1494 * change the region to UNBOXED. */
1495 os_vm_size_t remaining_bytes;
1496 os_vm_size_t bytes_freed;
1497 page_index_t next_page;
1498 page_bytes_t old_bytes_used;
1500 /* FIXME: This comment is somewhat stale.
1502 * Note: Any page write-protection must be removed, else a
1503 * later scavenge_newspace may incorrectly not scavenge these
1504 * pages. This would not be necessary if they are added to the
1505 * new areas, but let's do it for them all (they'll probably
1506 * be written anyway?). */
1508 gc_assert(page_starts_contiguous_block_p(first_page));
1509 next_page = first_page;
1510 remaining_bytes = nwords*N_WORD_BYTES;
1512 while (remaining_bytes > GENCGC_CARD_BYTES) {
1513 gc_assert(page_table[next_page].gen == from_space);
1514 gc_assert(page_table[next_page].large_object);
1515 gc_assert(page_table[next_page].scan_start_offset ==
1516 npage_bytes(next_page-first_page));
1517 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
1518 /* Should have been unprotected by unprotect_oldspace()
1519 * for boxed objects, and after promotion unboxed ones
1520 * should not be on protected pages at all. */
1521 gc_assert(!page_table[next_page].write_protected);
1524 gc_assert(page_boxed_p(next_page));
1526 gc_assert(page_allocated_no_region_p(next_page));
1527 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1529 page_table[next_page].gen = new_space;
1531 remaining_bytes -= GENCGC_CARD_BYTES;
1535 /* Now only one page remains, but the object may have shrunk so
1536 * there may be more unused pages which will be freed. */
1538 /* Object may have shrunk but shouldn't have grown - check. */
1539 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
1541 page_table[next_page].gen = new_space;
1544 gc_assert(page_boxed_p(next_page));
1546 page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
1548 /* Adjust the bytes_used. */
1549 old_bytes_used = page_table[next_page].bytes_used;
1550 page_table[next_page].bytes_used = remaining_bytes;
1552 bytes_freed = old_bytes_used - remaining_bytes;
1554 /* Free any remaining pages; needs care. */
1556 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
1557 (page_table[next_page].gen == from_space) &&
1558 /* FIXME: It is not obvious to me why this is necessary
1559 * as a loop condition: it seems to me that the
1560 * scan_start_offset test should be sufficient, but
1561 * experimentally that is not the case. --NS
1564 page_boxed_p(next_page) :
1565 page_allocated_no_region_p(next_page)) &&
1566 page_table[next_page].large_object &&
1567 (page_table[next_page].scan_start_offset ==
1568 npage_bytes(next_page - first_page))) {
1569 /* Checks out OK, free the page. Don't need to both zeroing
1570 * pages as this should have been done before shrinking the
1571 * object. These pages shouldn't be write-protected, even if
1572 * boxed they should be zero filled. */
1573 gc_assert(page_table[next_page].write_protected == 0);
1575 old_bytes_used = page_table[next_page].bytes_used;
1576 page_table[next_page].allocated = FREE_PAGE_FLAG;
1577 page_table[next_page].bytes_used = 0;
1578 bytes_freed += old_bytes_used;
1582 if ((bytes_freed > 0) && gencgc_verbose) {
1584 "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\n",
1588 generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES
1590 generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
1591 bytes_allocated -= bytes_freed;
1593 /* Add the region to the new_areas if requested. */
1595 add_new_area(first_page,0,nwords*N_WORD_BYTES);
1600 /* Get tag of object. */
1601 tag = lowtag_of(object);
1603 /* Allocate space. */
1604 new = gc_general_alloc(nwords*N_WORD_BYTES,
1605 (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG),
1608 /* Copy the object. */
1609 memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);
1611 /* Return Lisp pointer of new object. */
1612 return ((lispobj) new) | tag;
1617 copy_large_object(lispobj object, sword_t nwords)
1619 return general_copy_large_object(object, nwords, 1);
1623 copy_large_unboxed_object(lispobj object, sword_t nwords)
1625 return general_copy_large_object(object, nwords, 0);
1628 /* to copy unboxed objects */
1630 copy_unboxed_object(lispobj object, sword_t nwords)
1632 return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG);
1637 * code and code-related objects
1640 static lispobj trans_fun_header(lispobj object);
1641 static lispobj trans_boxed(lispobj object);
1644 /* Scan a x86 compiled code object, looking for possible fixups that
1645 * have been missed after a move.
1647 * Two types of fixups are needed:
1648 * 1. Absolute fixups to within the code object.
1649 * 2. Relative fixups to outside the code object.
1651 * Currently only absolute fixups to the constant vector, or to the
1652 * code area are checked. */
1654 sniff_code_object(struct code *code, os_vm_size_t displacement)
1656 #ifdef LISP_FEATURE_X86
1657 sword_t nheader_words, ncode_words, nwords;
1658 os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
1659 os_vm_address_t code_start_addr, code_end_addr;
1660 os_vm_address_t code_addr = (os_vm_address_t)code;
1661 int fixup_found = 0;
1663 if (!check_code_fixups)
1666 FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));
1668 ncode_words = fixnum_value(code->code_size);
1669 nheader_words = HeaderValue(*(lispobj *)code);
1670 nwords = ncode_words + nheader_words;
1672 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1673 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1674 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1675 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1677 /* Work through the unboxed code. */
1678 for (p = code_start_addr; p < code_end_addr; p++) {
1679 void *data = *(void **)p;
1680 unsigned d1 = *((unsigned char *)p - 1);
1681 unsigned d2 = *((unsigned char *)p - 2);
1682 unsigned d3 = *((unsigned char *)p - 3);
1683 unsigned d4 = *((unsigned char *)p - 4);
1685 unsigned d5 = *((unsigned char *)p - 5);
1686 unsigned d6 = *((unsigned char *)p - 6);
1689 /* Check for code references. */
1690 /* Check for a 32 bit word that looks like an absolute
1691 reference to within the code adea of the code object. */
1692 if ((data >= (void*)(code_start_addr-displacement))
1693 && (data < (void*)(code_end_addr-displacement))) {
1694 /* function header */
1696 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
1698 /* Skip the function header */
1702 /* the case of PUSH imm32 */
1706 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1707 p, d6, d5, d4, d3, d2, d1, data));
1708 FSHOW((stderr, "/PUSH $0x%.8x\n", data));
1710 /* the case of MOV [reg-8],imm32 */
1712 && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
1713 || d2==0x45 || d2==0x46 || d2==0x47)
1717 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1718 p, d6, d5, d4, d3, d2, d1, data));
1719 FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
1721 /* the case of LEA reg,[disp32] */
1722 if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
1725 "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1726 p, d6, d5, d4, d3, d2, d1, data));
1727 FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
1731 /* Check for constant references. */
1732 /* Check for a 32 bit word that looks like an absolute
1733 reference to within the constant vector. Constant references
1735 if ((data >= (void*)(constants_start_addr-displacement))
1736 && (data < (void*)(constants_end_addr-displacement))
1737 && (((unsigned)data & 0x3) == 0)) {
1742 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1743 p, d6, d5, d4, d3, d2, d1, data));
1744 FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
1747 /* the case of MOV m32,EAX */
1751 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1752 p, d6, d5, d4, d3, d2, d1, data));
1753 FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
1756 /* the case of CMP m32,imm32 */
1757 if ((d1 == 0x3d) && (d2 == 0x81)) {
1760 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1761 p, d6, d5, d4, d3, d2, d1, data));
1763 FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
1766 /* Check for a mod=00, r/m=101 byte. */
1767 if ((d1 & 0xc7) == 5) {
1772 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1773 p, d6, d5, d4, d3, d2, d1, data));
1774 FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
1776 /* the case of CMP reg32,m32 */
1780 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1781 p, d6, d5, d4, d3, d2, d1, data));
1782 FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
1784 /* the case of MOV m32,reg32 */
1788 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1789 p, d6, d5, d4, d3, d2, d1, data));
1790 FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
1792 /* the case of MOV reg32,m32 */
1796 "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1797 p, d6, d5, d4, d3, d2, d1, data));
1798 FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
1800 /* the case of LEA reg32,m32 */
1804 "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
1805 p, d6, d5, d4, d3, d2, d1, data));
1806 FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
1812 /* If anything was found, print some information on the code
1816 "/compiled code object at %x: header words = %d, code words = %d\n",
1817 code, nheader_words, ncode_words));
1819 "/const start = %x, end = %x\n",
1820 constants_start_addr, constants_end_addr));
1822 "/code start = %x, end = %x\n",
1823 code_start_addr, code_end_addr));
1829 gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
1831 /* x86-64 uses pc-relative addressing instead of this kludge */
1832 #ifndef LISP_FEATURE_X86_64
1833 sword_t nheader_words, ncode_words, nwords;
1834 os_vm_address_t constants_start_addr, constants_end_addr;
1835 os_vm_address_t code_start_addr, code_end_addr;
1836 os_vm_address_t code_addr = (os_vm_address_t)new_code;
1837 os_vm_address_t old_addr = (os_vm_address_t)old_code;
1838 os_vm_size_t displacement = code_addr - old_addr;
1839 lispobj fixups = NIL;
1840 struct vector *fixups_vector;
1842 ncode_words = fixnum_value(new_code->code_size);
1843 nheader_words = HeaderValue(*(lispobj *)new_code);
1844 nwords = ncode_words + nheader_words;
1846 "/compiled code object at %x: header words = %d, code words = %d\n",
1847 new_code, nheader_words, ncode_words)); */
1848 constants_start_addr = code_addr + 5*N_WORD_BYTES;
1849 constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
1850 code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
1851 code_end_addr = code_addr + nwords*N_WORD_BYTES;
1854 "/const start = %x, end = %x\n",
1855 constants_start_addr,constants_end_addr));
1857 "/code start = %x; end = %x\n",
1858 code_start_addr,code_end_addr));
1861 /* The first constant should be a pointer to the fixups for this
1862 code objects. Check. */
1863 fixups = new_code->constants[0];
1865 /* It will be 0 or the unbound-marker if there are no fixups (as
1866 * will be the case if the code object has been purified, for
1867 * example) and will be an other pointer if it is valid. */
1868 if ((fixups == 0) || (fixups == UNBOUND_MARKER_WIDETAG) ||
1869 !is_lisp_pointer(fixups)) {
1870 /* Check for possible errors. */
1871 if (check_code_fixups)
1872 sniff_code_object(new_code, displacement);
1877 fixups_vector = (struct vector *)native_pointer(fixups);
1879 /* Could be pointing to a forwarding pointer. */
1880 /* FIXME is this always in from_space? if so, could replace this code with
1881 * forwarding_pointer_p/forwarding_pointer_value */
1882 if (is_lisp_pointer(fixups) &&
1883 (find_page_index((void*)fixups_vector) != -1) &&
1884 (fixups_vector->header == 0x01)) {
1885 /* If so, then follow it. */
1886 /*SHOW("following pointer to a forwarding pointer");*/
1888 (struct vector *)native_pointer((lispobj)fixups_vector->length);
1891 /*SHOW("got fixups");*/
1893 if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
1894 /* Got the fixups for the code block. Now work through the vector,
1895 and apply a fixup at each address. */
1896 sword_t length = fixnum_value(fixups_vector->length);
1898 for (i = 0; i < length; i++) {
1899 long offset = fixups_vector->data[i];
1900 /* Now check the current value of offset. */
1901 os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);
1903 /* If it's within the old_code object then it must be an
1904 * absolute fixup (relative ones are not saved) */
1905 if ((old_value >= old_addr)
1906 && (old_value < (old_addr + nwords*N_WORD_BYTES)))
1907 /* So add the dispacement. */
1908 *(os_vm_address_t *)(code_start_addr + offset) =
1909 old_value + displacement;
1911 /* It is outside the old code object so it must be a
1912 * relative fixup (absolute fixups are not saved). So
1913 * subtract the displacement. */
1914 *(os_vm_address_t *)(code_start_addr + offset) =
1915 old_value - displacement;
1918 /* This used to just print a note to stderr, but a bogus fixup seems to
1919 * indicate real heap corruption, so a hard hailure is in order. */
1920 lose("fixup vector %p has a bad widetag: %d\n",
1921 fixups_vector, widetag_of(fixups_vector->header));
1924 /* Check for possible errors. */
1925 if (check_code_fixups) {
1926 sniff_code_object(new_code,displacement);
1933 trans_boxed_large(lispobj object)
1938 gc_assert(is_lisp_pointer(object));
1940 header = *((lispobj *) native_pointer(object));
1941 length = HeaderValue(header) + 1;
1942 length = CEILING(length, 2);
1944 return copy_large_object(object, length);
1947 /* Doesn't seem to be used, delete it after the grace period. */
1950 trans_unboxed_large(lispobj object)
1955 gc_assert(is_lisp_pointer(object));
1957 header = *((lispobj *) native_pointer(object));
1958 length = HeaderValue(header) + 1;
1959 length = CEILING(length, 2);
1961 return copy_large_unboxed_object(object, length);
1969 /* XX This is a hack adapted from cgc.c. These don't work too
1970 * efficiently with the gencgc as a list of the weak pointers is
1971 * maintained within the objects which causes writes to the pages. A
1972 * limited attempt is made to avoid unnecessary writes, but this needs
1974 #define WEAK_POINTER_NWORDS \
1975 CEILING((sizeof(struct weak_pointer) / sizeof(lispobj)), 2)
1978 scav_weak_pointer(lispobj *where, lispobj object)
1980 /* Since we overwrite the 'next' field, we have to make
1981 * sure not to do so for pointers already in the list.
1982 * Instead of searching the list of weak_pointers each
1983 * time, we ensure that next is always NULL when the weak
1984 * pointer isn't in the list, and not NULL otherwise.
1985 * Since we can't use NULL to denote end of list, we
1986 * use a pointer back to the same weak_pointer.
1988 struct weak_pointer * wp = (struct weak_pointer*)where;
1990 if (NULL == wp->next) {
1991 wp->next = weak_pointers;
1993 if (NULL == wp->next)
1997 /* Do not let GC scavenge the value slot of the weak pointer.
1998 * (That is why it is a weak pointer.) */
2000 return WEAK_POINTER_NWORDS;
2005 search_read_only_space(void *pointer)
2007 lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
2008 lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
2009 if ((pointer < (void *)start) || (pointer >= (void *)end))
2011 return (gc_search_space(start,
2012 (((lispobj *)pointer)+2)-start,
2013 (lispobj *) pointer));
2017 search_static_space(void *pointer)
2019 lispobj *start = (lispobj *)STATIC_SPACE_START;
2020 lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
2021 if ((pointer < (void *)start) || (pointer >= (void *)end))
2023 return (gc_search_space(start,
2024 (((lispobj *)pointer)+2)-start,
2025 (lispobj *) pointer));
2028 /* a faster version for searching the dynamic space. This will work even
2029 * if the object is in a current allocation region. */
2031 search_dynamic_space(void *pointer)
2033 page_index_t page_index = find_page_index(pointer);
2036 /* The address may be invalid, so do some checks. */
2037 if ((page_index == -1) || page_free_p(page_index))
2039 start = (lispobj *)page_scan_start(page_index);
2040 return (gc_search_space(start,
2041 (((lispobj *)pointer)+2)-start,
2042 (lispobj *)pointer));
2045 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2047 /* Is there any possibility that pointer is a valid Lisp object
2048 * reference, and/or something else (e.g. subroutine call return
2049 * address) which should prevent us from moving the referred-to thing?
2050 * This is called from preserve_pointers() */
2052 possibly_valid_dynamic_space_pointer(lispobj *pointer)
2054 lispobj *start_addr;
2056 /* Find the object start address. */
2057 if ((start_addr = search_dynamic_space(pointer)) == NULL) {
2061 return looks_like_valid_lisp_pointer_p(pointer, start_addr);
2064 #endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2066 /* Adjust large bignum and vector objects. This will adjust the
2067 * allocated region if the size has shrunk, and move unboxed objects
2068 * into unboxed pages. The pages are not promoted here, and the
2069 * promoted region is not added to the new_regions; this is really
2070 * only designed to be called from preserve_pointer(). Shouldn't fail
2071 * if this is missed, just may delay the moving of objects to unboxed
2072 * pages, and the freeing of pages. */
2074 maybe_adjust_large_object(lispobj *where)
2076 page_index_t first_page;
2077 page_index_t next_page;
2080 uword_t remaining_bytes;
2081 uword_t bytes_freed;
2082 uword_t old_bytes_used;
2086 /* Check whether it's a vector or bignum object. */
2087 switch (widetag_of(where[0])) {
2088 case SIMPLE_VECTOR_WIDETAG:
2089 boxed = BOXED_PAGE_FLAG;
2091 case BIGNUM_WIDETAG:
2092 case SIMPLE_BASE_STRING_WIDETAG:
2093 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
2094 case SIMPLE_CHARACTER_STRING_WIDETAG:
2096 case SIMPLE_BIT_VECTOR_WIDETAG:
2097 case SIMPLE_ARRAY_NIL_WIDETAG:
2098 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
2099 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
2100 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
2101 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
2102 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
2103 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
2105 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
2107 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
2108 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
2109 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
2110 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
2112 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
2113 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
2115 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
2116 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
2118 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
2119 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
2122 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
2124 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
2125 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
2127 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
2128 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
2130 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
2131 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
2132 #ifdef SIMPLE_ARRAY_LONG_FLOAT_WIDETAG
2133 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
2135 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
2136 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
2138 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
2139 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
2141 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
2142 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
2144 boxed = UNBOXED_PAGE_FLAG;
2150 /* Find its current size. */
2151 nwords = (sizetab[widetag_of(where[0])])(where);
2153 first_page = find_page_index((void *)where);
2154 gc_assert(first_page >= 0);
2156 /* Note: Any page write-protection must be removed, else a later
2157 * scavenge_newspace may incorrectly not scavenge these pages.
2158 * This would not be necessary if they are added to the new areas,
2159 * but lets do it for them all (they'll probably be written
2162 gc_assert(page_starts_contiguous_block_p(first_page));
2164 next_page = first_page;
2165 remaining_bytes = nwords*N_WORD_BYTES;
2166 while (remaining_bytes > GENCGC_CARD_BYTES) {
2167 gc_assert(page_table[next_page].gen == from_space);
2168 gc_assert(page_allocated_no_region_p(next_page));
2169 gc_assert(page_table[next_page].large_object);
2170 gc_assert(page_table[next_page].scan_start_offset ==
2171 npage_bytes(next_page-first_page));
2172 gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
2174 page_table[next_page].allocated = boxed;
2176 /* Shouldn't be write-protected at this stage. Essential that the
2178 gc_assert(!page_table[next_page].write_protected);
2179 remaining_bytes -= GENCGC_CARD_BYTES;
2183 /* Now only one page remains, but the object may have shrunk so
2184 * there may be more unused pages which will be freed. */
2186 /* Object may have shrunk but shouldn't have grown - check. */
2187 gc_assert(page_table[next_page].bytes_used >= remaining_bytes);
2189 page_table[next_page].allocated = boxed;
2190 gc_assert(page_table[next_page].allocated ==
2191 page_table[first_page].allocated);
2193 /* Adjust the bytes_used. */
2194 old_bytes_used = page_table[next_page].bytes_used;
2195 page_table[next_page].bytes_used = remaining_bytes;
2197 bytes_freed = old_bytes_used - remaining_bytes;
2199 /* Free any remaining pages; needs care. */
2201 while ((old_bytes_used == GENCGC_CARD_BYTES) &&
2202 (page_table[next_page].gen == from_space) &&
2203 page_allocated_no_region_p(next_page) &&
2204 page_table[next_page].large_object &&
2205 (page_table[next_page].scan_start_offset ==
2206 npage_bytes(next_page - first_page))) {
2207 /* It checks out OK, free the page. We don't need to both zeroing
2208 * pages as this should have been done before shrinking the
2209 * object. These pages shouldn't be write protected as they
2210 * should be zero filled. */
2211 gc_assert(page_table[next_page].write_protected == 0);
2213 old_bytes_used = page_table[next_page].bytes_used;
2214 page_table[next_page].allocated = FREE_PAGE_FLAG;
2215 page_table[next_page].bytes_used = 0;
2216 bytes_freed += old_bytes_used;
2220 if ((bytes_freed > 0) && gencgc_verbose) {
2222 "/maybe_adjust_large_object() freed %d\n",
2226 generations[from_space].bytes_allocated -= bytes_freed;
2227 bytes_allocated -= bytes_freed;
2232 /* Take a possible pointer to a Lisp object and mark its page in the
2233 * page_table so that it will not be relocated during a GC.
2235 * This involves locating the page it points to, then backing up to
2236 * the start of its region, then marking all pages dont_move from there
2237 * up to the first page that's not full or has a different generation
2239 * It is assumed that all the page static flags have been cleared at
2240 * the start of a GC.
2242 * It is also assumed that the current gc_alloc() region has been
2243 * flushed and the tables updated. */
2246 preserve_pointer(void *addr)
2248 page_index_t addr_page_index = find_page_index(addr);
2249 page_index_t first_page;
2251 unsigned int region_allocation;
2253 /* quick check 1: Address is quite likely to have been invalid. */
2254 if ((addr_page_index == -1)
2255 || page_free_p(addr_page_index)
2256 || (page_table[addr_page_index].bytes_used == 0)
2257 || (page_table[addr_page_index].gen != from_space)
2258 /* Skip if already marked dont_move. */
2259 || (page_table[addr_page_index].dont_move != 0))
2261 gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));
2262 /* (Now that we know that addr_page_index is in range, it's
2263 * safe to index into page_table[] with it.) */
2264 region_allocation = page_table[addr_page_index].allocated;
2266 /* quick check 2: Check the offset within the page.
2269 if (((uword_t)addr & (GENCGC_CARD_BYTES - 1)) >
2270 page_table[addr_page_index].bytes_used)
2273 /* Filter out anything which can't be a pointer to a Lisp object
2274 * (or, as a special case which also requires dont_move, a return
2275 * address referring to something in a CodeObject). This is
2276 * expensive but important, since it vastly reduces the
2277 * probability that random garbage will be bogusly interpreted as
2278 * a pointer which prevents a page from moving.
2280 * This only needs to happen on x86oids, where this is used for
2281 * conservative roots. Non-x86oid systems only ever call this
2282 * function on known-valid lisp objects. */
2283 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
2284 if (!(code_page_p(addr_page_index)
2285 || (is_lisp_pointer((lispobj)addr) &&
2286 possibly_valid_dynamic_space_pointer(addr))))
2290 /* Find the beginning of the region. Note that there may be
2291 * objects in the region preceding the one that we were passed a
2292 * pointer to: if this is the case, we will write-protect all the
2293 * previous objects' pages too. */
2296 /* I think this'd work just as well, but without the assertions.
2297 * -dan 2004.01.01 */
2298 first_page = find_page_index(page_scan_start(addr_page_index))
2300 first_page = addr_page_index;
2301 while (!page_starts_contiguous_block_p(first_page)) {
2303 /* Do some checks. */
2304 gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
2305 gc_assert(page_table[first_page].gen == from_space);
2306 gc_assert(page_table[first_page].allocated == region_allocation);
2310 /* Adjust any large objects before promotion as they won't be
2311 * copied after promotion. */
2312 if (page_table[first_page].large_object) {
2313 maybe_adjust_large_object(page_address(first_page));
2314 /* If a large object has shrunk then addr may now point to a
2315 * free area in which case it's ignored here. Note it gets
2316 * through the valid pointer test above because the tail looks
2318 if (page_free_p(addr_page_index)
2319 || (page_table[addr_page_index].bytes_used == 0)
2320 /* Check the offset within the page. */
2321 || (((uword_t)addr & (GENCGC_CARD_BYTES - 1))
2322 > page_table[addr_page_index].bytes_used)) {
2324 "weird? ignore ptr 0x%x to freed area of large object\n",
2328 /* It may have moved to unboxed pages. */
2329 region_allocation = page_table[first_page].allocated;
2332 /* Now work forward until the end of this contiguous area is found,
2333 * marking all pages as dont_move. */
2334 for (i = first_page; ;i++) {
2335 gc_assert(page_table[i].allocated == region_allocation);
2337 /* Mark the page static. */
2338 page_table[i].dont_move = 1;
2340 /* It is essential that the pages are not write protected as
2341 * they may have pointers into the old-space which need
2342 * scavenging. They shouldn't be write protected at this
2344 gc_assert(!page_table[i].write_protected);
2346 /* Check whether this is the last page in this contiguous block.. */
2347 if (page_ends_contiguous_block_p(i, from_space))
2351 /* Check that the page is now static. */
2352 gc_assert(page_table[addr_page_index].dont_move != 0);
2355 /* If the given page is not write-protected, then scan it for pointers
2356 * to younger generations or the top temp. generation, if no
2357 * suspicious pointers are found then the page is write-protected.
2359 * Care is taken to check for pointers to the current gc_alloc()
2360 * region if it is a younger generation or the temp. generation. This
2361 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
2362 * the gc_alloc_generation does not need to be checked as this is only
2363 * called from scavenge_generation() when the gc_alloc generation is
2364 * younger, so it just checks if there is a pointer to the current
2367 * We return 1 if the page was write-protected, else 0. */
2369 update_page_write_prot(page_index_t page)
2371 generation_index_t gen = page_table[page].gen;
2374 void **page_addr = (void **)page_address(page);
2375 sword_t num_words = page_table[page].bytes_used / N_WORD_BYTES;
2377 /* Shouldn't be a free page. */
2378 gc_assert(page_allocated_p(page));
2379 gc_assert(page_table[page].bytes_used != 0);
2381 /* Skip if it's already write-protected, pinned, or unboxed */
2382 if (page_table[page].write_protected
2383 /* FIXME: What's the reason for not write-protecting pinned pages? */
2384 || page_table[page].dont_move
2385 || page_unboxed_p(page))
2388 /* Scan the page for pointers to younger generations or the
2389 * top temp. generation. */
2391 for (j = 0; j < num_words; j++) {
2392 void *ptr = *(page_addr+j);
2393 page_index_t index = find_page_index(ptr);
2395 /* Check that it's in the dynamic space */
2397 if (/* Does it point to a younger or the temp. generation? */
2398 (page_allocated_p(index)
2399 && (page_table[index].bytes_used != 0)
2400 && ((page_table[index].gen < gen)
2401 || (page_table[index].gen == SCRATCH_GENERATION)))
2403 /* Or does it point within a current gc_alloc() region? */
2404 || ((boxed_region.start_addr <= ptr)
2405 && (ptr <= boxed_region.free_pointer))
2406 || ((unboxed_region.start_addr <= ptr)
2407 && (ptr <= unboxed_region.free_pointer))) {
2414 /* Write-protect the page. */
2415 /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
2417 os_protect((void *)page_addr,
2419 OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
2421 /* Note the page as protected in the page tables. */
2422 page_table[page].write_protected = 1;
2428 /* Scavenge all generations from FROM to TO, inclusive, except for
2429 * new_space which needs special handling, as new objects may be
2430 * added which are not checked here - use scavenge_newspace generation.
2432 * Write-protected pages should not have any pointers to the
2433 * from_space so do need scavenging; thus write-protected pages are
2434 * not always scavenged. There is some code to check that these pages
2435 * are not written; but to check fully the write-protected pages need
2436 * to be scavenged by disabling the code to skip them.
2438 * Under the current scheme when a generation is GCed the younger
2439 * generations will be empty. So, when a generation is being GCed it
2440 * is only necessary to scavenge the older generations for pointers
2441 * not the younger. So a page that does not have pointers to younger
2442 * generations does not need to be scavenged.
2444 * The write-protection can be used to note pages that don't have
2445 * pointers to younger pages. But pages can be written without having
2446 * pointers to younger generations. After the pages are scavenged here
2447 * they can be scanned for pointers to younger generations and if
2448 * there are none the page can be write-protected.
2450 * One complication is when the newspace is the top temp. generation.
2452 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
2453 * that none were written, which they shouldn't be as they should have
2454 * no pointers to younger generations. This breaks down for weak
2455 * pointers as the objects contain a link to the next and are written
2456 * if a weak pointer is scavenged. Still it's a useful check. */
2458 scavenge_generations(generation_index_t from, generation_index_t to)
2461 page_index_t num_wp = 0;
2465 /* Clear the write_protected_cleared flags on all pages. */
2466 for (i = 0; i < page_table_pages; i++)
2467 page_table[i].write_protected_cleared = 0;
2470 for (i = 0; i < last_free_page; i++) {
2471 generation_index_t generation = page_table[i].gen;
2473 && (page_table[i].bytes_used != 0)
2474 && (generation != new_space)
2475 && (generation >= from)
2476 && (generation <= to)) {
2477 page_index_t last_page,j;
2478 int write_protected=1;
2480 /* This should be the start of a region */
2481 gc_assert(page_starts_contiguous_block_p(i));
2483 /* Now work forward until the end of the region */
2484 for (last_page = i; ; last_page++) {
2486 write_protected && page_table[last_page].write_protected;
2487 if (page_ends_contiguous_block_p(last_page, generation))
2490 if (!write_protected) {
2491 scavenge(page_address(i),
2492 ((uword_t)(page_table[last_page].bytes_used
2493 + npage_bytes(last_page-i)))
2496 /* Now scan the pages and write protect those that
2497 * don't have pointers to younger generations. */
2498 if (enable_page_protection) {
2499 for (j = i; j <= last_page; j++) {
2500 num_wp += update_page_write_prot(j);
2503 if ((gencgc_verbose > 1) && (num_wp != 0)) {
2505 "/write protected %d pages within generation %d\n",
2506 num_wp, generation));
2514 /* Check that none of the write_protected pages in this generation
2515 * have been written to. */
2516 for (i = 0; i < page_table_pages; i++) {
2517 if (page_allocated_p(i)
2518 && (page_table[i].bytes_used != 0)
2519 && (page_table[i].gen == generation)
2520 && (page_table[i].write_protected_cleared != 0)) {
2521 FSHOW((stderr, "/scavenge_generation() %d\n", generation));
2523 "/page bytes_used=%d scan_start_offset=%lu dont_move=%d\n",
2524 page_table[i].bytes_used,
2525 page_table[i].scan_start_offset,
2526 page_table[i].dont_move));
2527 lose("write to protected page %d in scavenge_generation()\n", i);
2534 /* Scavenge a newspace generation. As it is scavenged new objects may
2535 * be allocated to it; these will also need to be scavenged. This
2536 * repeats until there are no more objects unscavenged in the
2537 * newspace generation.
2539 * To help improve the efficiency, areas written are recorded by
2540 * gc_alloc() and only these scavenged. Sometimes a little more will be
2541 * scavenged, but this causes no harm. An easy check is done that the
2542 * scavenged bytes equals the number allocated in the previous
2545 * Write-protected pages are not scanned except if they are marked
2546 * dont_move in which case they may have been promoted and still have
2547 * pointers to the from space.
2549 * Write-protected pages could potentially be written by alloc however
2550 * to avoid having to handle re-scavenging of write-protected pages
2551 * gc_alloc() does not write to write-protected pages.
2553 * New areas of objects allocated are recorded alternatively in the two
2554 * new_areas arrays below. */
2555 static struct new_area new_areas_1[NUM_NEW_AREAS];
2556 static struct new_area new_areas_2[NUM_NEW_AREAS];
2558 /* Do one full scan of the new space generation. This is not enough to
2559 * complete the job as new objects may be added to the generation in
2560 * the process which are not scavenged. */
2562 scavenge_newspace_generation_one_scan(generation_index_t generation)
2567 "/starting one full scan of newspace generation %d\n",
2569 for (i = 0; i < last_free_page; i++) {
2570 /* Note that this skips over open regions when it encounters them. */
2572 && (page_table[i].bytes_used != 0)
2573 && (page_table[i].gen == generation)
2574 && ((page_table[i].write_protected == 0)
2575 /* (This may be redundant as write_protected is now
2576 * cleared before promotion.) */
2577 || (page_table[i].dont_move == 1))) {
2578 page_index_t last_page;
2581 /* The scavenge will start at the scan_start_offset of
2584 * We need to find the full extent of this contiguous
2585 * block in case objects span pages.
2587 * Now work forward until the end of this contiguous area
2588 * is found. A small area is preferred as there is a
2589 * better chance of its pages being write-protected. */
2590 for (last_page = i; ;last_page++) {
2591 /* If all pages are write-protected and movable,
2592 * then no need to scavenge */
2593 all_wp=all_wp && page_table[last_page].write_protected &&
2594 !page_table[last_page].dont_move;
2596 /* Check whether this is the last page in this
2597 * contiguous block */
2598 if (page_ends_contiguous_block_p(last_page, generation))
2602 /* Do a limited check for write-protected pages. */
2604 sword_t nwords = (((uword_t)
2605 (page_table[last_page].bytes_used
2606 + npage_bytes(last_page-i)
2607 + page_table[i].scan_start_offset))
2609 new_areas_ignore_page = last_page;
2611 scavenge(page_scan_start(i), nwords);
2618 "/done with one full scan of newspace generation %d\n",
2622 /* Do a complete scavenge of the newspace generation. */
2624 scavenge_newspace_generation(generation_index_t generation)
2628 /* the new_areas array currently being written to by gc_alloc() */
2629 struct new_area (*current_new_areas)[] = &new_areas_1;
2630 size_t current_new_areas_index;
2632 /* the new_areas created by the previous scavenge cycle */
2633 struct new_area (*previous_new_areas)[] = NULL;
2634 size_t previous_new_areas_index;
2636 /* Flush the current regions updating the tables. */
2637 gc_alloc_update_all_page_tables();
2639 /* Turn on the recording of new areas by gc_alloc(). */
2640 new_areas = current_new_areas;
2641 new_areas_index = 0;
2643 /* Don't need to record new areas that get scavenged anyway during
2644 * scavenge_newspace_generation_one_scan. */
2645 record_new_objects = 1;
2647 /* Start with a full scavenge. */
2648 scavenge_newspace_generation_one_scan(generation);
2650 /* Record all new areas now. */
2651 record_new_objects = 2;
2653 /* Give a chance to weak hash tables to make other objects live.
2654 * FIXME: The algorithm implemented here for weak hash table gcing
2655 * is O(W^2+N) as Bruno Haible warns in
2656 * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
2657 * see "Implementation 2". */
2658 scav_weak_hash_tables();
2660 /* Flush the current regions updating the tables. */
2661 gc_alloc_update_all_page_tables();
2663 /* Grab new_areas_index. */
2664 current_new_areas_index = new_areas_index;
2667 "The first scan is finished; current_new_areas_index=%d.\n",
2668 current_new_areas_index));*/
2670 while (current_new_areas_index > 0) {
2671 /* Move the current to the previous new areas */
2672 previous_new_areas = current_new_areas;
2673 previous_new_areas_index = current_new_areas_index;
2675 /* Scavenge all the areas in previous new areas. Any new areas
2676 * allocated are saved in current_new_areas. */
2678 /* Allocate an array for current_new_areas; alternating between
2679 * new_areas_1 and 2 */
2680 if (previous_new_areas == &new_areas_1)
2681 current_new_areas = &new_areas_2;
2683 current_new_areas = &new_areas_1;
2685 /* Set up for gc_alloc(). */
2686 new_areas = current_new_areas;
2687 new_areas_index = 0;
2689 /* Check whether previous_new_areas had overflowed. */
2690 if (previous_new_areas_index >= NUM_NEW_AREAS) {
2692 /* New areas of objects allocated have been lost so need to do a
2693 * full scan to be sure! If this becomes a problem try
2694 * increasing NUM_NEW_AREAS. */
2695 if (gencgc_verbose) {
2696 SHOW("new_areas overflow, doing full scavenge");
2699 /* Don't need to record new areas that get scavenged
2700 * anyway during scavenge_newspace_generation_one_scan. */
2701 record_new_objects = 1;
2703 scavenge_newspace_generation_one_scan(generation);
2705 /* Record all new areas now. */
2706 record_new_objects = 2;
2708 scav_weak_hash_tables();
2710 /* Flush the current regions updating the tables. */
2711 gc_alloc_update_all_page_tables();
2715 /* Work through previous_new_areas. */
2716 for (i = 0; i < previous_new_areas_index; i++) {
2717 page_index_t page = (*previous_new_areas)[i].page;
2718 size_t offset = (*previous_new_areas)[i].offset;
2719 size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
2720 gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
2721 scavenge(page_address(page)+offset, size);
2724 scav_weak_hash_tables();
2726 /* Flush the current regions updating the tables. */
2727 gc_alloc_update_all_page_tables();
2730 current_new_areas_index = new_areas_index;
2733 "The re-scan has finished; current_new_areas_index=%d.\n",
2734 current_new_areas_index));*/
2737 /* Turn off recording of areas allocated by gc_alloc(). */
2738 record_new_objects = 0;
2743 /* Check that none of the write_protected pages in this generation
2744 * have been written to. */
2745 for (i = 0; i < page_table_pages; i++) {
2746 if (page_allocated_p(i)
2747 && (page_table[i].bytes_used != 0)
2748 && (page_table[i].gen == generation)
2749 && (page_table[i].write_protected_cleared != 0)
2750 && (page_table[i].dont_move == 0)) {
2751 lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
2752 i, generation, page_table[i].dont_move);
2759 /* Un-write-protect all the pages in from_space. This is done at the
2760 * start of a GC else there may be many page faults while scavenging
2761 * the newspace (I've seen drive the system time to 99%). These pages
2762 * would need to be unprotected anyway before unmapping in
2763 * free_oldspace; not sure what effect this has on paging.. */
2765 unprotect_oldspace(void)
2768 void *region_addr = 0;
2769 void *page_addr = 0;
2770 uword_t region_bytes = 0;
2772 for (i = 0; i < last_free_page; i++) {
2773 if (page_allocated_p(i)
2774 && (page_table[i].bytes_used != 0)
2775 && (page_table[i].gen == from_space)) {
2777 /* Remove any write-protection. We should be able to rely
2778 * on the write-protect flag to avoid redundant calls. */
2779 if (page_table[i].write_protected) {
2780 page_table[i].write_protected = 0;
2781 page_addr = page_address(i);
2784 region_addr = page_addr;
2785 region_bytes = GENCGC_CARD_BYTES;
2786 } else if (region_addr + region_bytes == page_addr) {
2787 /* Region continue. */
2788 region_bytes += GENCGC_CARD_BYTES;
2790 /* Unprotect previous region. */
2791 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2792 /* First page in new region. */
2793 region_addr = page_addr;
2794 region_bytes = GENCGC_CARD_BYTES;
2800 /* Unprotect last region. */
2801 os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
2805 /* Work through all the pages and free any in from_space. This
2806 * assumes that all objects have been copied or promoted to an older
2807 * generation. Bytes_allocated and the generation bytes_allocated
2808 * counter are updated. The number of bytes freed is returned. */
2812 uword_t bytes_freed = 0;
2813 page_index_t first_page, last_page;
2818 /* Find a first page for the next region of pages. */
2819 while ((first_page < last_free_page)
2820 && (page_free_p(first_page)
2821 || (page_table[first_page].bytes_used == 0)
2822 || (page_table[first_page].gen != from_space)))
2825 if (first_page >= last_free_page)
2828 /* Find the last page of this region. */
2829 last_page = first_page;
2832 /* Free the page. */
2833 bytes_freed += page_table[last_page].bytes_used;
2834 generations[page_table[last_page].gen].bytes_allocated -=
2835 page_table[last_page].bytes_used;
2836 page_table[last_page].allocated = FREE_PAGE_FLAG;
2837 page_table[last_page].bytes_used = 0;
2838 /* Should already be unprotected by unprotect_oldspace(). */
2839 gc_assert(!page_table[last_page].write_protected);
2842 while ((last_page < last_free_page)
2843 && page_allocated_p(last_page)
2844 && (page_table[last_page].bytes_used != 0)
2845 && (page_table[last_page].gen == from_space));
2847 #ifdef READ_PROTECT_FREE_PAGES
2848 os_protect(page_address(first_page),
2849 npage_bytes(last_page-first_page),
2852 first_page = last_page;
2853 } while (first_page < last_free_page);
2855 bytes_allocated -= bytes_freed;
2860 /* Print some information about a pointer at the given address. */
2862 print_ptr(lispobj *addr)
2864 /* If addr is in the dynamic space then out the page information. */
2865 page_index_t pi1 = find_page_index((void*)addr);
2868 fprintf(stderr," %p: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n",
2871 page_table[pi1].allocated,
2872 page_table[pi1].gen,
2873 page_table[pi1].bytes_used,
2874 page_table[pi1].scan_start_offset,
2875 page_table[pi1].dont_move);
2876 fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n",
2890 is_in_stack_space(lispobj ptr)
2892 /* For space verification: Pointers can be valid if they point
2893 * to a thread stack space. This would be faster if the thread
2894 * structures had page-table entries as if they were part of
2895 * the heap space. */
2897 for_each_thread(th) {
2898 if ((th->control_stack_start <= (lispobj *)ptr) &&
2899 (th->control_stack_end >= (lispobj *)ptr)) {
2907 verify_space(lispobj *start, size_t words)
2909 int is_in_dynamic_space = (find_page_index((void*)start) != -1);
2910 int is_in_readonly_space =
2911 (READ_ONLY_SPACE_START <= (uword_t)start &&
2912 (uword_t)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2916 lispobj thing = *(lispobj*)start;
2918 if (is_lisp_pointer(thing)) {
2919 page_index_t page_index = find_page_index((void*)thing);
2920 sword_t to_readonly_space =
2921 (READ_ONLY_SPACE_START <= thing &&
2922 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
2923 sword_t to_static_space =
2924 (STATIC_SPACE_START <= thing &&
2925 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
2927 /* Does it point to the dynamic space? */
2928 if (page_index != -1) {
2929 /* If it's within the dynamic space it should point to a used
2930 * page. XX Could check the offset too. */
2931 if (page_allocated_p(page_index)
2932 && (page_table[page_index].bytes_used == 0))
2933 lose ("Ptr %p @ %p sees free page.\n", thing, start);
2934 /* Check that it doesn't point to a forwarding pointer! */
2935 if (*((lispobj *)native_pointer(thing)) == 0x01) {
2936 lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
2938 /* Check that its not in the RO space as it would then be a
2939 * pointer from the RO to the dynamic space. */
2940 if (is_in_readonly_space) {
2941 lose("ptr to dynamic space %p from RO space %x\n",
2944 /* Does it point to a plausible object? This check slows
2945 * it down a lot (so it's commented out).
2947 * "a lot" is serious: it ate 50 minutes cpu time on
2948 * my duron 950 before I came back from lunch and
2951 * FIXME: Add a variable to enable this
2954 if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) {
2955 lose("ptr %p to invalid object %p\n", thing, start);
2959 extern void funcallable_instance_tramp;
2960 /* Verify that it points to another valid space. */
2961 if (!to_readonly_space && !to_static_space
2962 && (thing != (lispobj)&funcallable_instance_tramp)
2963 && !is_in_stack_space(thing)) {
2964 lose("Ptr %p @ %p sees junk.\n", thing, start);
2968 if (!(fixnump(thing))) {
2970 switch(widetag_of(*start)) {
2973 case SIMPLE_VECTOR_WIDETAG:
2975 case COMPLEX_WIDETAG:
2976 case SIMPLE_ARRAY_WIDETAG:
2977 case COMPLEX_BASE_STRING_WIDETAG:
2978 #ifdef COMPLEX_CHARACTER_STRING_WIDETAG
2979 case COMPLEX_CHARACTER_STRING_WIDETAG:
2981 case COMPLEX_VECTOR_NIL_WIDETAG:
2982 case COMPLEX_BIT_VECTOR_WIDETAG:
2983 case COMPLEX_VECTOR_WIDETAG:
2984 case COMPLEX_ARRAY_WIDETAG:
2985 case CLOSURE_HEADER_WIDETAG:
2986 case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
2987 case VALUE_CELL_HEADER_WIDETAG:
2988 case SYMBOL_HEADER_WIDETAG:
2989 case CHARACTER_WIDETAG:
2990 #if N_WORD_BITS == 64
2991 case SINGLE_FLOAT_WIDETAG:
2993 case UNBOUND_MARKER_WIDETAG:
2998 case INSTANCE_HEADER_WIDETAG:
3001 sword_t ntotal = HeaderValue(thing);
3002 lispobj layout = ((struct instance *)start)->slots[0];
3007 nuntagged = ((struct layout *)
3008 native_pointer(layout))->n_untagged_slots;
3009 verify_space(start + 1,
3010 ntotal - fixnum_value(nuntagged));
3014 case CODE_HEADER_WIDETAG:
3016 lispobj object = *start;
3018 sword_t nheader_words, ncode_words, nwords;
3020 struct simple_fun *fheaderp;
3022 code = (struct code *) start;
3024 /* Check that it's not in the dynamic space.
3025 * FIXME: Isn't is supposed to be OK for code
3026 * objects to be in the dynamic space these days? */
3027 if (is_in_dynamic_space
3028 /* It's ok if it's byte compiled code. The trace
3029 * table offset will be a fixnum if it's x86
3030 * compiled code - check.
3032 * FIXME: #^#@@! lack of abstraction here..
3033 * This line can probably go away now that
3034 * there's no byte compiler, but I've got
3035 * too much to worry about right now to try
3036 * to make sure. -- WHN 2001-10-06 */
3037 && fixnump(code->trace_table_offset)
3038 /* Only when enabled */
3039 && verify_dynamic_code_check) {
3041 "/code object at %p in the dynamic space\n",
3045 ncode_words = fixnum_value(code->code_size);
3046 nheader_words = HeaderValue(object);
3047 nwords = ncode_words + nheader_words;
3048 nwords = CEILING(nwords, 2);
3049 /* Scavenge the boxed section of the code data block */
3050 verify_space(start + 1, nheader_words - 1);
3052 /* Scavenge the boxed section of each function
3053 * object in the code data block. */
3054 fheaderl = code->entry_points;
3055 while (fheaderl != NIL) {
3057 (struct simple_fun *) native_pointer(fheaderl);
3058 gc_assert(widetag_of(fheaderp->header) ==
3059 SIMPLE_FUN_HEADER_WIDETAG);
3060 verify_space(&fheaderp->name, 1);
3061 verify_space(&fheaderp->arglist, 1);
3062 verify_space(&fheaderp->type, 1);
3063 fheaderl = fheaderp->next;
3069 /* unboxed objects */
3070 case BIGNUM_WIDETAG:
3071 #if N_WORD_BITS != 64
3072 case SINGLE_FLOAT_WIDETAG:
3074 case DOUBLE_FLOAT_WIDETAG:
3075 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3076 case LONG_FLOAT_WIDETAG:
3078 #ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
3079 case COMPLEX_SINGLE_FLOAT_WIDETAG:
3081 #ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
3082 case COMPLEX_DOUBLE_FLOAT_WIDETAG:
3084 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
3085 case COMPLEX_LONG_FLOAT_WIDETAG:
3087 case SIMPLE_BASE_STRING_WIDETAG:
3088 #ifdef SIMPLE_CHARACTER_STRING_WIDETAG
3089 case SIMPLE_CHARACTER_STRING_WIDETAG:
3091 case SIMPLE_BIT_VECTOR_WIDETAG:
3092 case SIMPLE_ARRAY_NIL_WIDETAG:
3093 case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
3094 case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
3095 case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
3096 case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
3097 case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
3098 case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
3100 case SIMPLE_ARRAY_UNSIGNED_FIXNUM_WIDETAG:
3102 case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
3103 case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
3104 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG
3105 case SIMPLE_ARRAY_UNSIGNED_BYTE_63_WIDETAG:
3107 #ifdef SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
3108 case SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG:
3110 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
3111 case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
3113 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG
3114 case SIMPLE_ARRAY_SIGNED_BYTE_16_WIDETAG:
3117 case SIMPLE_ARRAY_FIXNUM_WIDETAG:
3119 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG
3120 case SIMPLE_ARRAY_SIGNED_BYTE_32_WIDETAG:
3122 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG
3123 case SIMPLE_ARRAY_SIGNED_BYTE_64_WIDETAG:
3125 case SIMPLE_ARRAY_SINGLE_FLOAT_WIDETAG:
3126 case SIMPLE_ARRAY_DOUBLE_FLOAT_WIDETAG:
3127 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3128 case SIMPLE_ARRAY_LONG_FLOAT_WIDETAG:
3130 #ifdef SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG
3131 case SIMPLE_ARRAY_COMPLEX_SINGLE_FLOAT_WIDETAG:
3133 #ifdef SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG
3134 case SIMPLE_ARRAY_COMPLEX_DOUBLE_FLOAT_WIDETAG:
3136 #ifdef SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG
3137 case SIMPLE_ARRAY_COMPLEX_LONG_FLOAT_WIDETAG:
3140 case WEAK_POINTER_WIDETAG:
3141 #ifdef NO_TLS_VALUE_MARKER_WIDETAG
3142 case NO_TLS_VALUE_MARKER_WIDETAG:
3144 count = (sizetab[widetag_of(*start)])(start);
3148 lose("Unhandled widetag %p at %p\n",
3149 widetag_of(*start), start);
3161 /* FIXME: It would be nice to make names consistent so that
3162 * foo_size meant size *in* *bytes* instead of size in some
3163 * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
3164 * Some counts of lispobjs are called foo_count; it might be good
3165 * to grep for all foo_size and rename the appropriate ones to
3167 sword_t read_only_space_size =
3168 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
3169 - (lispobj*)READ_ONLY_SPACE_START;
3170 sword_t static_space_size =
3171 (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
3172 - (lispobj*)STATIC_SPACE_START;
3174 for_each_thread(th) {
3175 sword_t binding_stack_size =
3176 (lispobj*)get_binding_stack_pointer(th)
3177 - (lispobj*)th->binding_stack_start;
3178 verify_space(th->binding_stack_start, binding_stack_size);
3180 verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
3181 verify_space((lispobj*)STATIC_SPACE_START , static_space_size);
3185 verify_generation(generation_index_t generation)
3189 for (i = 0; i < last_free_page; i++) {
3190 if (page_allocated_p(i)
3191 && (page_table[i].bytes_used != 0)
3192 && (page_table[i].gen == generation)) {
3193 page_index_t last_page;
3194 int region_allocation = page_table[i].allocated;
3196 /* This should be the start of a contiguous block */
3197 gc_assert(page_starts_contiguous_block_p(i));
3199 /* Need to find the full extent of this contiguous block in case
3200 objects span pages. */
3202 /* Now work forward until the end of this contiguous area is
3204 for (last_page = i; ;last_page++)
3205 /* Check whether this is the last page in this contiguous
3207 if (page_ends_contiguous_block_p(last_page, generation))
3210 verify_space(page_address(i),
3212 (page_table[last_page].bytes_used
3213 + npage_bytes(last_page-i)))
3220 /* Check that all the free space is zero filled. */
3222 verify_zero_fill(void)
3226 for (page = 0; page < last_free_page; page++) {
3227 if (page_free_p(page)) {
3228 /* The whole page should be zero filled. */
3229 sword_t *start_addr = (sword_t *)page_address(page);
3230 sword_t size = 1024;
3232 for (i = 0; i < size; i++) {
3233 if (start_addr[i] != 0) {
3234 lose("free page not zero at %x\n", start_addr + i);
3238 sword_t free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
3239 if (free_bytes > 0) {
3240 sword_t *start_addr = (sword_t *)((uword_t)page_address(page)
3241 + page_table[page].bytes_used);
3242 sword_t size = free_bytes / N_WORD_BYTES;
3244 for (i = 0; i < size; i++) {
3245 if (start_addr[i] != 0) {
3246 lose("free region not zero at %x\n", start_addr + i);
3254 /* External entry point for verify_zero_fill */
3256 gencgc_verify_zero_fill(void)
3258 /* Flush the alloc regions updating the tables. */
3259 gc_alloc_update_all_page_tables();
3260 SHOW("verifying zero fill");
3265 verify_dynamic_space(void)
3267 generation_index_t i;
3269 for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++)
3270 verify_generation(i);
3272 if (gencgc_enable_verify_zero_fill)
3276 /* Write-protect all the dynamic boxed pages in the given generation. */
3278 write_protect_generation_pages(generation_index_t generation)
3282 gc_assert(generation < SCRATCH_GENERATION);
3284 for (start = 0; start < last_free_page; start++) {
3285 if (protect_page_p(start, generation)) {
3289 /* Note the page as protected in the page tables. */
3290 page_table[start].write_protected = 1;
3292 for (last = start + 1; last < last_free_page; last++) {
3293 if (!protect_page_p(last, generation))
3295 page_table[last].write_protected = 1;
3298 page_start = (void *)page_address(start);
3300 os_protect(page_start,
3301 npage_bytes(last - start),
3302 OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
3308 if (gencgc_verbose > 1) {
3310 "/write protected %d of %d pages in generation %d\n",
3311 count_write_protect_generation_pages(generation),
3312 count_generation_pages(generation),
3317 #if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
3319 preserve_context_registers (os_context_t *c)
3322 /* On Darwin the signal context isn't a contiguous block of memory,
3323 * so just preserve_pointering its contents won't be sufficient.
3325 #if defined(LISP_FEATURE_DARWIN)||defined(LISP_FEATURE_WIN32)
3326 #if defined LISP_FEATURE_X86
3327 preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
3328 preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
3329 preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
3330 preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
3331 preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
3332 preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
3333 preserve_pointer((void*)*os_context_pc_addr(c));
3334 #elif defined LISP_FEATURE_X86_64
3335 preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
3336 preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
3337 preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
3338 preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
3339 preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
3340 preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
3341 preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
3342 preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
3343 preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
3344 preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
3345 preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
3346 preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
3347 preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
3348 preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
3349 preserve_pointer((void*)*os_context_pc_addr(c));
3351 #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
3354 #if !defined(LISP_FEATURE_WIN32)
3355 for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
3356 preserve_pointer(*ptr);
3363 move_pinned_pages_to_newspace()
3367 /* scavenge() will evacuate all oldspace pages, but no newspace
3368 * pages. Pinned pages are precisely those pages which must not
3369 * be evacuated, so move them to newspace directly. */
3371 for (i = 0; i < last_free_page; i++) {
3372 if (page_table[i].dont_move &&
3373 /* dont_move is cleared lazily, so validate the space as well. */
3374 page_table[i].gen == from_space) {
3375 page_table[i].gen = new_space;
3376 /* And since we're moving the pages wholesale, also adjust
3377 * the generation allocation counters. */
3378 generations[new_space].bytes_allocated += page_table[i].bytes_used;
3379 generations[from_space].bytes_allocated -= page_table[i].bytes_used;
3384 /* Garbage collect a generation. If raise is 0 then the remains of the
3385 * generation are not raised to the next generation. */
3387 garbage_collect_generation(generation_index_t generation, int raise)
3389 uword_t bytes_freed;
3391 uword_t static_space_size;
3394 gc_assert(generation <= HIGHEST_NORMAL_GENERATION);
3396 /* The oldest generation can't be raised. */
3397 gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
3399 /* Check if weak hash tables were processed in the previous GC. */
3400 gc_assert(weak_hash_tables == NULL);
3402 /* Initialize the weak pointer list. */
3403 weak_pointers = NULL;
3405 /* When a generation is not being raised it is transported to a
3406 * temporary generation (NUM_GENERATIONS), and lowered when
3407 * done. Set up this new generation. There should be no pages
3408 * allocated to it yet. */
3410 gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
3413 /* Set the global src and dest. generations */
3414 from_space = generation;
3416 new_space = generation+1;
3418 new_space = SCRATCH_GENERATION;
3420 /* Change to a new space for allocation, resetting the alloc_start_page */
3421 gc_alloc_generation = new_space;
3422 generations[new_space].alloc_start_page = 0;
3423 generations[new_space].alloc_unboxed_start_page = 0;
3424 generations[new_space].alloc_large_start_page = 0;
3425 generations[new_space].alloc_large_unboxed_start_page = 0;
3427 /* Before any pointers are preserved, the dont_move flags on the
3428 * pages need to be cleared. */
3429 for (i = 0; i < last_free_page; i++)
3430 if(page_table[i].gen==from_space)
3431 page_table[i].dont_move = 0;
3433 /* Un-write-protect the old-space pages. This is essential for the
3434 * promoted pages as they may contain pointers into the old-space
3435 * which need to be scavenged. It also helps avoid unnecessary page
3436 * faults as forwarding pointers are written into them. They need to
3437 * be un-protected anyway before unmapping later. */
3438 unprotect_oldspace();
3440 /* Scavenge the stacks' conservative roots. */
3442 /* there are potentially two stacks for each thread: the main
3443 * stack, which may contain Lisp pointers, and the alternate stack.
3444 * We don't ever run Lisp code on the altstack, but it may
3445 * host a sigcontext with lisp objects in it */
3447 /* what we need to do: (1) find the stack pointer for the main
3448 * stack; scavenge it (2) find the interrupt context on the
3449 * alternate stack that might contain lisp values, and scavenge
3452 /* we assume that none of the preceding applies to the thread that
3453 * initiates GC. If you ever call GC from inside an altstack
3454 * handler, you will lose. */
3456 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
3457 /* And if we're saving a core, there's no point in being conservative. */
3458 if (conservative_stack) {
3459 for_each_thread(th) {
3461 void **esp=(void **)-1;
3462 if (th->state == STATE_DEAD)
3464 # if defined(LISP_FEATURE_SB_SAFEPOINT)
3465 /* Conservative collect_garbage is always invoked with a
3466 * foreign C call or an interrupt handler on top of every
3467 * existing thread, so the stored SP in each thread
3468 * structure is valid, no matter which thread we are looking
3469 * at. For threads that were running Lisp code, the pitstop
3470 * and edge functions maintain this value within the
3471 * interrupt or exception handler. */
3472 esp = os_get_csp(th);
3473 assert_on_stack(th, esp);
3475 /* In addition to pointers on the stack, also preserve the
3476 * return PC, the only value from the context that we need
3477 * in addition to the SP. The return PC gets saved by the
3478 * foreign call wrapper, and removed from the control stack
3479 * into a register. */
3480 preserve_pointer(th->pc_around_foreign_call);
3482 /* And on platforms with interrupts: scavenge ctx registers. */
3484 /* Disabled on Windows, because it does not have an explicit
3485 * stack of `interrupt_contexts'. The reported CSP has been
3486 * chosen so that the current context on the stack is
3487 * covered by the stack scan. See also set_csp_from_context(). */
3488 # ifndef LISP_FEATURE_WIN32
3489 if (th != arch_os_get_current_thread()) {
3490 long k = fixnum_value(
3491 SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3493 preserve_context_registers(th->interrupt_contexts[--k]);
3496 # elif defined(LISP_FEATURE_SB_THREAD)
3498 if(th==arch_os_get_current_thread()) {
3499 /* Somebody is going to burn in hell for this, but casting
3500 * it in two steps shuts gcc up about strict aliasing. */
3501 esp = (void **)((void *)&raise);
3504 free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
3505 for(i=free-1;i>=0;i--) {
3506 os_context_t *c=th->interrupt_contexts[i];
3507 esp1 = (void **) *os_context_register_addr(c,reg_SP);
3508 if (esp1>=(void **)th->control_stack_start &&
3509 esp1<(void **)th->control_stack_end) {
3510 if(esp1<esp) esp=esp1;
3511 preserve_context_registers(c);
3516 esp = (void **)((void *)&raise);
3518 if (!esp || esp == (void*) -1)
3519 lose("garbage_collect: no SP known for thread %x (OS %x)",
3521 for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) {
3522 preserve_pointer(*ptr);
3527 /* Non-x86oid systems don't have "conservative roots" as such, but
3528 * the same mechanism is used for objects pinned for use by alien
3530 for_each_thread(th) {
3531 lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
3532 while (pin_list != NIL) {
3533 struct cons *list_entry =
3534 (struct cons *)native_pointer(pin_list);
3535 preserve_pointer(list_entry->car);
3536 pin_list = list_entry->cdr;
3542 if (gencgc_verbose > 1) {
3543 sword_t num_dont_move_pages = count_dont_move_pages();
3545 "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
3546 num_dont_move_pages,
3547 npage_bytes(num_dont_move_pages));
3551 /* Now that all of the pinned (dont_move) pages are known, and
3552 * before we start to scavenge (and thus relocate) objects,
3553 * relocate the pinned pages to newspace, so that the scavenger
3554 * will not attempt to relocate their contents. */
3555 move_pinned_pages_to_newspace();
3557 /* Scavenge all the rest of the roots. */
3559 #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
3561 * If not x86, we need to scavenge the interrupt context(s) and the
3566 for_each_thread(th) {
3567 scavenge_interrupt_contexts(th);
3568 scavenge_control_stack(th);
3571 # ifdef LISP_FEATURE_SB_SAFEPOINT
3572 /* In this case, scrub all stacks right here from the GCing thread
3573 * instead of doing what the comment below says. Suboptimal, but
3576 scrub_thread_control_stack(th);
3578 /* Scrub the unscavenged control stack space, so that we can't run
3579 * into any stale pointers in a later GC (this is done by the
3580 * stop-for-gc handler in the other threads). */
3581 scrub_control_stack();
3586 /* Scavenge the Lisp functions of the interrupt handlers, taking
3587 * care to avoid SIG_DFL and SIG_IGN. */
3588 for (i = 0; i < NSIG; i++) {
3589 union interrupt_handler handler = interrupt_handlers[i];
3590 if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
3591 !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
3592 scavenge((lispobj *)(interrupt_handlers + i), 1);
3595 /* Scavenge the binding stacks. */
3598 for_each_thread(th) {
3599 sword_t len= (lispobj *)get_binding_stack_pointer(th) -
3600 th->binding_stack_start;
3601 scavenge((lispobj *) th->binding_stack_start,len);
3602 #ifdef LISP_FEATURE_SB_THREAD
3603 /* do the tls as well */
3604 len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
3605 (sizeof (struct thread))/(sizeof (lispobj));
3606 scavenge((lispobj *) (th+1),len);
3611 /* The original CMU CL code had scavenge-read-only-space code
3612 * controlled by the Lisp-level variable
3613 * *SCAVENGE-READ-ONLY-SPACE*. It was disabled by default, and it
3614 * wasn't documented under what circumstances it was useful or
3615 * safe to turn it on, so it's been turned off in SBCL. If you
3616 * want/need this functionality, and can test and document it,
3617 * please submit a patch. */
3619 if (SymbolValue(SCAVENGE_READ_ONLY_SPACE) != NIL) {
3620 uword_t read_only_space_size =
3621 (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) -
3622 (lispobj*)READ_ONLY_SPACE_START;
3624 "/scavenge read only space: %d bytes\n",
3625 read_only_space_size * sizeof(lispobj)));
3626 scavenge( (lispobj *) READ_ONLY_SPACE_START, read_only_space_size);
3630 /* Scavenge static space. */
3632 (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
3633 (lispobj *)STATIC_SPACE_START;
3634 if (gencgc_verbose > 1) {
3636 "/scavenge static space: %d bytes\n",
3637 static_space_size * sizeof(lispobj)));
3639 scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);
3641 /* All generations but the generation being GCed need to be
3642 * scavenged. The new_space generation needs special handling as
3643 * objects may be moved in - it is handled separately below. */
3644 scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);
3646 /* Finally scavenge the new_space generation. Keep going until no
3647 * more objects are moved into the new generation */
3648 scavenge_newspace_generation(new_space);
3650 /* FIXME: I tried reenabling this check when debugging unrelated
3651 * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
3652 * Since the current GC code seems to work well, I'm guessing that
3653 * this debugging code is just stale, but I haven't tried to
3654 * figure it out. It should be figured out and then either made to
3655 * work or just deleted. */
3656 #define RESCAN_CHECK 0
3658 /* As a check re-scavenge the newspace once; no new objects should
3661 os_vm_size_t old_bytes_allocated = bytes_allocated;
3662 os_vm_size_t bytes_allocated;
3664 /* Start with a full scavenge. */
3665 scavenge_newspace_generation_one_scan(new_space);
3667 /* Flush the current regions, updating the tables. */
3668 gc_alloc_update_all_page_tables();
3670 bytes_allocated = bytes_allocated - old_bytes_allocated;
3672 if (bytes_allocated != 0) {
3673 lose("Rescan of new_space allocated %d more bytes.\n",
3679 scan_weak_hash_tables();
3680 scan_weak_pointers();
3682 /* Flush the current regions, updating the tables. */
3683 gc_alloc_update_all_page_tables();
3685 /* Free the pages in oldspace, but not those marked dont_move. */
3686 bytes_freed = free_oldspace();
3688 /* If the GC is not raising the age then lower the generation back
3689 * to its normal generation number */
3691 for (i = 0; i < last_free_page; i++)
3692 if ((page_table[i].bytes_used != 0)
3693 && (page_table[i].gen == SCRATCH_GENERATION))
3694 page_table[i].gen = generation;
3695 gc_assert(generations[generation].bytes_allocated == 0);
3696 generations[generation].bytes_allocated =
3697 generations[SCRATCH_GENERATION].bytes_allocated;
3698 generations[SCRATCH_GENERATION].bytes_allocated = 0;
3701 /* Reset the alloc_start_page for generation. */
3702 generations[generation].alloc_start_page = 0;
3703 generations[generation].alloc_unboxed_start_page = 0;
3704 generations[generation].alloc_large_start_page = 0;
3705 generations[generation].alloc_large_unboxed_start_page = 0;
3707 if (generation >= verify_gens) {
3708 if (gencgc_verbose) {
3712 verify_dynamic_space();
3715 /* Set the new gc trigger for the GCed generation. */
3716 generations[generation].gc_trigger =
3717 generations[generation].bytes_allocated
3718 + generations[generation].bytes_consed_between_gc;
3721 generations[generation].num_gc = 0;
3723 ++generations[generation].num_gc;
3727 /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
3729 update_dynamic_space_free_pointer(void)
3731 page_index_t last_page = -1, i;
3733 for (i = 0; i < last_free_page; i++)
3734 if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
3737 last_free_page = last_page+1;
3739 set_alloc_pointer((lispobj)(page_address(last_free_page)));
3740 return 0; /* dummy value: return something ... */
3744 remap_page_range (page_index_t from, page_index_t to)
3746 /* There's a mysterious Solaris/x86 problem with using mmap
3747 * tricks for memory zeroing. See sbcl-devel thread
3748 * "Re: patch: standalone executable redux".
3750 #if defined(LISP_FEATURE_SUNOS)
3751 zero_and_mark_pages(from, to);
3754 release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
3755 release_mask = release_granularity-1,
3757 aligned_from = (from+release_mask)&~release_mask,
3758 aligned_end = (end&~release_mask);
3760 if (aligned_from < aligned_end) {
3761 zero_pages_with_mmap(aligned_from, aligned_end-1);
3762 if (aligned_from != from)
3763 zero_and_mark_pages(from, aligned_from-1);
3764 if (aligned_end != end)
3765 zero_and_mark_pages(aligned_end, end-1);
3767 zero_and_mark_pages(from, to);
3773 remap_free_pages (page_index_t from, page_index_t to, int forcibly)
3775 page_index_t first_page, last_page;
3778 return remap_page_range(from, to);
3780 for (first_page = from; first_page <= to; first_page++) {
3781 if (page_allocated_p(first_page) ||
3782 (page_table[first_page].need_to_zero == 0))
3785 last_page = first_page + 1;
3786 while (page_free_p(last_page) &&
3787 (last_page <= to) &&
3788 (page_table[last_page].need_to_zero == 1))
3791 remap_page_range(first_page, last_page-1);
3793 first_page = last_page;
3797 generation_index_t small_generation_limit = 1;
3799 /* GC all generations newer than last_gen, raising the objects in each
3800 * to the next older generation - we finish when all generations below
3801 * last_gen are empty. Then if last_gen is due for a GC, or if
3802 * last_gen==NUM_GENERATIONS (the scratch generation? eh?) we GC that
3803 * too. The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
3805 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
3806 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
3808 collect_garbage(generation_index_t last_gen)
3810 generation_index_t gen = 0, i;
3811 int raise, more = 0;
3813 /* The largest value of last_free_page seen since the time
3814 * remap_free_pages was called. */
3815 static page_index_t high_water_mark = 0;
3817 FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
3818 log_generation_stats(gc_logfile, "=== GC Start ===");
3822 if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
3824 "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
3829 /* Flush the alloc regions updating the tables. */
3830 gc_alloc_update_all_page_tables();
3832 /* Verify the new objects created by Lisp code. */
3833 if (pre_verify_gen_0) {
3834 FSHOW((stderr, "pre-checking generation 0\n"));
3835 verify_generation(0);
3838 if (gencgc_verbose > 1)
3839 print_generation_stats();
3842 /* Collect the generation. */
3844 if (more || (gen >= gencgc_oldest_gen_to_gc)) {
3845 /* Never raise the oldest generation. Never raise the extra generation
3846 * collected due to more-flag. */
3852 || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
3853 /* If we would not normally raise this one, but we're
3854 * running low on space in comparison to the object-sizes
3855 * we've been seeing, raise it and collect the next one
3857 if (!raise && gen == last_gen) {
3858 more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated);
3863 if (gencgc_verbose > 1) {
3865 "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
3868 generations[gen].bytes_allocated,
3869 generations[gen].gc_trigger,
3870 generations[gen].num_gc));
3873 /* If an older generation is being filled, then update its
3876 generations[gen+1].cum_sum_bytes_allocated +=
3877 generations[gen+1].bytes_allocated;
3880 garbage_collect_generation(gen, raise);
3882 /* Reset the memory age cum_sum. */
3883 generations[gen].cum_sum_bytes_allocated = 0;
3885 if (gencgc_verbose > 1) {
3886 FSHOW((stderr, "GC of generation %d finished:\n", gen));
3887 print_generation_stats();
3891 } while ((gen <= gencgc_oldest_gen_to_gc)
3892 && ((gen < last_gen)
3895 && (generations[gen].bytes_allocated
3896 > generations[gen].gc_trigger)
3897 && (generation_average_age(gen)
3898 > generations[gen].minimum_age_before_gc))));
3900 /* Now if gen-1 was raised all generations before gen are empty.
3901 * If it wasn't raised then all generations before gen-1 are empty.
3903 * Now objects within this gen's pages cannot point to younger
3904 * generations unless they are written to. This can be exploited
3905 * by write-protecting the pages of gen; then when younger
3906 * generations are GCed only the pages which have been written
3911 gen_to_wp = gen - 1;
3913 /* There's not much point in WPing pages in generation 0 as it is
3914 * never scavenged (except promoted pages). */
3915 if ((gen_to_wp > 0) && enable_page_protection) {
3916 /* Check that they are all empty. */
3917 for (i = 0; i < gen_to_wp; i++) {
3918 if (generations[i].bytes_allocated)
3919 lose("trying to write-protect gen. %d when gen. %d nonempty\n",
3922 write_protect_generation_pages(gen_to_wp);
3925 /* Set gc_alloc() back to generation 0. The current regions should
3926 * be flushed after the above GCs. */
3927 gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
3928 gc_alloc_generation = 0;
3930 /* Save the high-water mark before updating last_free_page */
3931 if (last_free_page > high_water_mark)
3932 high_water_mark = last_free_page;
3934 update_dynamic_space_free_pointer();
3936 /* Update auto_gc_trigger. Make sure we trigger the next GC before
3937 * running out of heap! */
3938 if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
3939 auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
3941 auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;
3944 fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n",
3947 /* If we did a big GC (arbitrarily defined as gen > 1), release memory
3950 if (gen > small_generation_limit) {
3951 if (last_free_page > high_water_mark)
3952 high_water_mark = last_free_page;
3953 remap_free_pages(0, high_water_mark, 0);
3954 high_water_mark = 0;
3958 large_allocation = 0;
3960 log_generation_stats(gc_logfile, "=== GC End ===");
3961 SHOW("returning from collect_garbage");
3964 /* This is called by Lisp PURIFY when it is finished. All live objects
3965 * will have been moved to the RO and Static heaps. The dynamic space
3966 * will need a full re-initialization. We don't bother having Lisp
3967 * PURIFY flush the current gc_alloc() region, as the page_tables are
3968 * re-initialized, and every page is zeroed to be sure. */
3972 page_index_t page, last_page;
3974 if (gencgc_verbose > 1) {
3975 SHOW("entering gc_free_heap");
3978 for (page = 0; page < page_table_pages; page++) {
3979 /* Skip free pages which should already be zero filled. */
3980 if (page_allocated_p(page)) {
3982 for (last_page = page;
3983 (last_page < page_table_pages) && page_allocated_p(last_page);
3985 /* Mark the page free. The other slots are assumed invalid
3986 * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it
3987 * should not be write-protected -- except that the
3988 * generation is used for the current region but it sets
3990 page_table[page].allocated = FREE_PAGE_FLAG;
3991 page_table[page].bytes_used = 0;
3992 page_table[page].write_protected = 0;
3995 #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure
3996 * about this change. */
3997 page_start = (void *)page_address(page);
3998 os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL);
3999 remap_free_pages(page, last_page-1, 1);
4002 } else if (gencgc_zero_check_during_free_heap) {
4003 /* Double-check that the page is zero filled. */
4004 sword_t *page_start;
4006 gc_assert(page_free_p(page));
4007 gc_assert(page_table[page].bytes_used == 0);
4008 page_start = (sword_t *)page_address(page);
4009 for (i=0; i<GENCGC_CARD_BYTES/sizeof(sword_t); i++) {
4010 if (page_start[i] != 0) {
4011 lose("free region not zero at %x\n", page_start + i);
4017 bytes_allocated = 0;
4019 /* Initialize the generations. */
4020 for (page = 0; page < NUM_GENERATIONS; page++) {
4021 generations[page].alloc_start_page = 0;
4022 generations[page].alloc_unboxed_start_page = 0;
4023 generations[page].alloc_large_start_page = 0;
4024 generations[page].alloc_large_unboxed_start_page = 0;
4025 generations[page].bytes_allocated = 0;
4026 generations[page].gc_trigger = 2000000;
4027 generations[page].num_gc = 0;
4028 generations[page].cum_sum_bytes_allocated = 0;
4031 if (gencgc_verbose > 1)
4032 print_generation_stats();
4034 /* Initialize gc_alloc(). */
4035 gc_alloc_generation = 0;
4037 gc_set_region_empty(&boxed_region);
4038 gc_set_region_empty(&unboxed_region);
4041 set_alloc_pointer((lispobj)((char *)heap_base));
4043 if (verify_after_free_heap) {
4044 /* Check whether purify has left any bad pointers. */
4045 FSHOW((stderr, "checking after free_heap\n"));
4055 #if defined(LISP_FEATURE_SB_SAFEPOINT)
4059 /* Compute the number of pages needed for the dynamic space.
4060 * Dynamic space size should be aligned on page size. */
4061 page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
4062 gc_assert(dynamic_space_size == npage_bytes(page_table_pages));
4064 /* Default nursery size to 5% of the total dynamic space size,
4066 bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
4067 if (bytes_consed_between_gcs < (1024*1024))
4068 bytes_consed_between_gcs = 1024*1024;
4070 /* The page_table must be allocated using "calloc" to initialize
4071 * the page structures correctly. There used to be a separate
4072 * initialization loop (now commented out; see below) but that was
4073 * unnecessary and did hurt startup time. */
4074 page_table = calloc(page_table_pages, sizeof(struct page));
4075 gc_assert(page_table);
4078 scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
4079 transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
4081 heap_base = (void*)DYNAMIC_SPACE_START;
4083 /* The page structures are initialized implicitly when page_table
4084 * is allocated with "calloc" above. Formerly we had the following
4085 * explicit initialization here (comments converted to C99 style
4086 * for readability as C's block comments don't nest):
4088 * // Initialize each page structure.
4089 * for (i = 0; i < page_table_pages; i++) {
4090 * // Initialize all pages as free.
4091 * page_table[i].allocated = FREE_PAGE_FLAG;
4092 * page_table[i].bytes_used = 0;
4094 * // Pages are not write-protected at startup.
4095 * page_table[i].write_protected = 0;
4098 * Without this loop the image starts up much faster when dynamic
4099 * space is large -- which it is on 64-bit platforms already by
4100 * default -- and when "calloc" for large arrays is implemented
4101 * using copy-on-write of a page of zeroes -- which it is at least
4102 * on Linux. In this case the pages that page_table_pages is stored
4103 * in are mapped and cleared not before the corresponding part of
4104 * dynamic space is used. For example, this saves clearing 16 MB of
4105 * memory at startup if the page size is 4 KB and the size of
4106 * dynamic space is 4 GB.
4107 * FREE_PAGE_FLAG must be 0 for this to work correctly which is
4108 * asserted below: */
4110 /* Compile time assertion: If triggered, declares an array
4111 * of dimension -1 forcing a syntax error. The intent of the
4112 * assignment is to avoid an "unused variable" warning. */
4113 char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
4114 assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
4117 bytes_allocated = 0;
4119 /* Initialize the generations.
4121 * FIXME: very similar to code in gc_free_heap(), should be shared */
4122 for (i = 0; i < NUM_GENERATIONS; i++) {
4123 generations[i].alloc_start_page = 0;
4124 generations[i].alloc_unboxed_start_page = 0;
4125 generations[i].alloc_large_start_page = 0;
4126 generations[i].alloc_large_unboxed_start_page = 0;
4127 generations[i].bytes_allocated = 0;
4128 generations[i].gc_trigger = 2000000;
4129 generations[i].num_gc = 0;
4130 generations[i].cum_sum_bytes_allocated = 0;
4131 /* the tune-able parameters */
4132 generations[i].bytes_consed_between_gc
4133 = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
4134 generations[i].number_of_gcs_before_promotion = 1;
4135 generations[i].minimum_age_before_gc = 0.75;
4138 /* Initialize gc_alloc. */
4139 gc_alloc_generation = 0;
4140 gc_set_region_empty(&boxed_region);
4141 gc_set_region_empty(&unboxed_region);
4146 /* Pick up the dynamic space from after a core load.
4148 * The ALLOCATION_POINTER points to the end of the dynamic space.
4152 gencgc_pickup_dynamic(void)
4154 page_index_t page = 0;
4155 void *alloc_ptr = (void *)get_alloc_pointer();
4156 lispobj *prev=(lispobj *)page_address(page);
4157 generation_index_t gen = PSEUDO_STATIC_GENERATION;
4159 bytes_allocated = 0;
4162 lispobj *first,*ptr= (lispobj *)page_address(page);
4164 if (!gencgc_partial_pickup || page_allocated_p(page)) {
4165 /* It is possible, though rare, for the saved page table
4166 * to contain free pages below alloc_ptr. */
4167 page_table[page].gen = gen;
4168 page_table[page].bytes_used = GENCGC_CARD_BYTES;
4169 page_table[page].large_object = 0;
4170 page_table[page].write_protected = 0;
4171 page_table[page].write_protected_cleared = 0;
4172 page_table[page].dont_move = 0;
4173 page_table[page].need_to_zero = 1;
4175 bytes_allocated += GENCGC_CARD_BYTES;
4178 if (!gencgc_partial_pickup) {
4179 page_table[page].allocated = BOXED_PAGE_FLAG;
4180 first=gc_search_space(prev,(ptr+2)-prev,ptr);
4183 page_table[page].scan_start_offset =
4184 page_address(page) - (void *)prev;
4187 } while (page_address(page) < alloc_ptr);
4189 last_free_page = page;
4191 generations[gen].bytes_allocated = bytes_allocated;
4193 gc_alloc_update_all_page_tables();
4194 write_protect_generation_pages(gen);
4198 gc_initialize_pointers(void)
4200 gencgc_pickup_dynamic();
4204 /* alloc(..) is the external interface for memory allocation. It
4205 * allocates to generation 0. It is not called from within the garbage
4206 * collector as it is only external uses that need the check for heap
4207 * size (GC trigger) and to disable the interrupts (interrupts are
4208 * always disabled during a GC).
4210 * The vops that call alloc(..) assume that the returned space is zero-filled.
4211 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
4213 * The check for a GC trigger is only performed when the current
4214 * region is full, so in most cases it's not needed. */
4216 static inline lispobj *
4217 general_alloc_internal(sword_t nbytes, int page_type_flag, struct alloc_region *region,
4218 struct thread *thread)
4220 #ifndef LISP_FEATURE_WIN32
4221 lispobj alloc_signal;
4224 void *new_free_pointer;
4225 os_vm_size_t trigger_bytes = 0;
4227 gc_assert(nbytes>0);
4229 /* Check for alignment allocation problems. */
4230 gc_assert((((uword_t)region->free_pointer & LOWTAG_MASK) == 0)
4231 && ((nbytes & LOWTAG_MASK) == 0));
4233 #if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
4234 /* Must be inside a PA section. */
4235 gc_assert(get_pseudo_atomic_atomic(thread));
4238 if (nbytes > large_allocation)
4239 large_allocation = nbytes;
4241 /* maybe we can do this quickly ... */
4242 new_free_pointer = region->free_pointer + nbytes;
4243 if (new_free_pointer <= region->end_addr) {
4244 new_obj = (void*)(region->free_pointer);
4245 region->free_pointer = new_free_pointer;
4246 return(new_obj); /* yup */
4249 /* We don't want to count nbytes against auto_gc_trigger unless we
4250 * have to: it speeds up the tenuring of objects and slows down
4251 * allocation. However, unless we do so when allocating _very_
4252 * large objects we are in danger of exhausting the heap without
4253 * running sufficient GCs.
4255 if (nbytes >= bytes_consed_between_gcs)
4256 trigger_bytes = nbytes;
4258 /* we have to go the long way around, it seems. Check whether we
4259 * should GC in the near future
4261 if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
4262 /* Don't flood the system with interrupts if the need to gc is
4263 * already noted. This can happen for example when SUB-GC
4264 * allocates or after a gc triggered in a WITHOUT-GCING. */
4265 if (SymbolValue(GC_PENDING,thread) == NIL) {
4266 /* set things up so that GC happens when we finish the PA
4268 SetSymbolValue(GC_PENDING,T,thread);
4269 if (SymbolValue(GC_INHIBIT,thread) == NIL) {
4270 #ifdef LISP_FEATURE_SB_SAFEPOINT
4271 thread_register_gc_trigger();
4273 set_pseudo_atomic_interrupted(thread);
4274 #ifdef GENCGC_IS_PRECISE
4275 /* PPC calls alloc() from a trap or from pa_alloc(),
4276 * look up the most context if it's from a trap. */
4278 os_context_t *context =
4279 thread->interrupt_data->allocation_trap_context;
4280 maybe_save_gc_mask_and_block_deferrables
4281 (context ? os_context_sigmask_addr(context) : NULL);
4284 maybe_save_gc_mask_and_block_deferrables(NULL);
4290 new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
4292 #ifndef LISP_FEATURE_WIN32
4293 /* for sb-prof, and not supported on Windows yet */
4294 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
4295 if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
4296 if ((sword_t) alloc_signal <= 0) {
4297 SetSymbolValue(ALLOC_SIGNAL, T, thread);
4300 SetSymbolValue(ALLOC_SIGNAL,
4301 alloc_signal - (1 << N_FIXNUM_TAG_BITS),
4311 general_alloc(sword_t nbytes, int page_type_flag)
4313 struct thread *thread = arch_os_get_current_thread();
4314 /* Select correct region, and call general_alloc_internal with it.
4315 * For other then boxed allocation we must lock first, since the
4316 * region is shared. */
4317 if (BOXED_PAGE_FLAG & page_type_flag) {
4318 #ifdef LISP_FEATURE_SB_THREAD
4319 struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
4321 struct alloc_region *region = &boxed_region;
4323 return general_alloc_internal(nbytes, page_type_flag, region, thread);
4324 } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
4326 gc_assert(0 == thread_mutex_lock(&allocation_lock));
4327 obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
4328 gc_assert(0 == thread_mutex_unlock(&allocation_lock));
4331 lose("bad page type flag: %d", page_type_flag);
4335 lispobj AMD64_SYSV_ABI *
4338 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4339 struct thread *self = arch_os_get_current_thread();
4340 int was_pseudo_atomic = get_pseudo_atomic_atomic(self);
4341 if (!was_pseudo_atomic)
4342 set_pseudo_atomic_atomic(self);
4344 gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
4347 lispobj *result = general_alloc(nbytes, BOXED_PAGE_FLAG);
4349 #ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
4350 if (!was_pseudo_atomic)
4351 clear_pseudo_atomic_atomic(self);
4358 * shared support for the OS-dependent signal handlers which
4359 * catch GENCGC-related write-protect violations
4361 void unhandled_sigmemoryfault(void* addr);
4363 /* Depending on which OS we're running under, different signals might
4364 * be raised for a violation of write protection in the heap. This
4365 * function factors out the common generational GC magic which needs
4366 * to invoked in this case, and should be called from whatever signal
4367 * handler is appropriate for the OS we're running under.
4369 * Return true if this signal is a normal generational GC thing that
4370 * we were able to handle, or false if it was abnormal and control
4371 * should fall through to the general SIGSEGV/SIGBUS/whatever logic.
4373 * We have two control flags for this: one causes us to ignore faults
4374 * on unprotected pages completely, and the second complains to stderr
4375 * but allows us to continue without losing.
4377 extern boolean ignore_memoryfaults_on_unprotected_pages;
4378 boolean ignore_memoryfaults_on_unprotected_pages = 0;
4380 extern boolean continue_after_memoryfault_on_unprotected_pages;
4381 boolean continue_after_memoryfault_on_unprotected_pages = 0;
4384 gencgc_handle_wp_violation(void* fault_addr)
4386 page_index_t page_index = find_page_index(fault_addr);
4389 FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n",
4390 fault_addr, page_index));
4393 /* Check whether the fault is within the dynamic space. */
4394 if (page_index == (-1)) {
4396 /* It can be helpful to be able to put a breakpoint on this
4397 * case to help diagnose low-level problems. */
4398 unhandled_sigmemoryfault(fault_addr);
4400 /* not within the dynamic space -- not our responsibility */
4405 ret = thread_mutex_lock(&free_pages_lock);
4406 gc_assert(ret == 0);
4407 if (page_table[page_index].write_protected) {
4408 /* Unprotect the page. */
4409 os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
4410 page_table[page_index].write_protected_cleared = 1;
4411 page_table[page_index].write_protected = 0;
4412 } else if (!ignore_memoryfaults_on_unprotected_pages) {
4413 /* The only acceptable reason for this signal on a heap
4414 * access is that GENCGC write-protected the page.
4415 * However, if two CPUs hit a wp page near-simultaneously,
4416 * we had better not have the second one lose here if it
4417 * does this test after the first one has already set wp=0
4419 if(page_table[page_index].write_protected_cleared != 1) {
4420 void lisp_backtrace(int frames);
4423 "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n"
4424 " boxed_region.first_page: %"PAGE_INDEX_FMT","
4425 " boxed_region.last_page %"PAGE_INDEX_FMT"\n"
4426 " page.scan_start_offset: %"OS_VM_SIZE_FMT"\n"
4427 " page.bytes_used: %"PAGE_BYTES_FMT"\n"
4428 " page.allocated: %d\n"
4429 " page.write_protected: %d\n"
4430 " page.write_protected_cleared: %d\n"
4431 " page.generation: %d\n",
4434 boxed_region.first_page,
4435 boxed_region.last_page,
4436 page_table[page_index].scan_start_offset,
4437 page_table[page_index].bytes_used,
4438 page_table[page_index].allocated,
4439 page_table[page_index].write_protected,
4440 page_table[page_index].write_protected_cleared,
4441 page_table[page_index].gen);
4442 if (!continue_after_memoryfault_on_unprotected_pages)
4446 ret = thread_mutex_unlock(&free_pages_lock);
4447 gc_assert(ret == 0);
4448 /* Don't worry, we can handle it. */
4452 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
4453 * it's not just a case of the program hitting the write barrier, and
4454 * are about to let Lisp deal with it. It's basically just a
4455 * convenient place to set a gdb breakpoint. */
4457 unhandled_sigmemoryfault(void *addr)
4460 void gc_alloc_update_all_page_tables(void)
4462 /* Flush the alloc regions updating the tables. */
4464 for_each_thread(th) {
4465 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
4466 #if defined(LISP_FEATURE_SB_SAFEPOINT_STRICTLY) && !defined(LISP_FEATURE_WIN32)
4467 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->sprof_alloc_region);
4470 gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
4471 gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
4475 gc_set_region_empty(struct alloc_region *region)
4477 region->first_page = 0;
4478 region->last_page = -1;
4479 region->start_addr = page_address(0);
4480 region->free_pointer = page_address(0);
4481 region->end_addr = page_address(0);
4485 zero_all_free_pages()
4489 for (i = 0; i < last_free_page; i++) {
4490 if (page_free_p(i)) {
4491 #ifdef READ_PROTECT_FREE_PAGES
4492 os_protect(page_address(i),
4501 /* Things to do before doing a final GC before saving a core (without
4504 * + Pages in large_object pages aren't moved by the GC, so we need to
4505 * unset that flag from all pages.
4506 * + The pseudo-static generation isn't normally collected, but it seems
4507 * reasonable to collect it at least when saving a core. So move the
4508 * pages to a normal generation.
4511 prepare_for_final_gc ()
4514 for (i = 0; i < last_free_page; i++) {
4515 page_table[i].large_object = 0;
4516 if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
4517 int used = page_table[i].bytes_used;
4518 page_table[i].gen = HIGHEST_NORMAL_GENERATION;
4519 generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
4520 generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
4526 /* Do a non-conservative GC, and then save a core with the initial
4527 * function being set to the value of the static symbol
4528 * SB!VM:RESTART-LISP-FUNCTION */
4530 gc_and_save(char *filename, boolean prepend_runtime,
4531 boolean save_runtime_options,
4532 boolean compressed, int compression_level)
4535 void *runtime_bytes = NULL;
4536 size_t runtime_size;
4538 file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
4543 conservative_stack = 0;
4545 /* The filename might come from Lisp, and be moved by the now
4546 * non-conservative GC. */
4547 filename = strdup(filename);
4549 /* Collect twice: once into relatively high memory, and then back
4550 * into low memory. This compacts the retained data into the lower
4551 * pages, minimizing the size of the core file.
4553 prepare_for_final_gc();
4554 gencgc_alloc_start_page = last_free_page;
4555 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4557 prepare_for_final_gc();
4558 gencgc_alloc_start_page = -1;
4559 collect_garbage(HIGHEST_NORMAL_GENERATION+1);
4561 if (prepend_runtime)
4562 save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
4564 /* The dumper doesn't know that pages need to be zeroed before use. */
4565 zero_all_free_pages();
4566 save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
4567 prepend_runtime, save_runtime_options,
4568 compressed ? compression_level : COMPRESSION_LEVEL_NONE);
4569 /* Oops. Save still managed to fail. Since we've mangled the stack
4570 * beyond hope, there's not much we can do.
4571 * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
4572 * going to be rather unsatisfactory too... */
4573 lose("Attempt to save core after non-conservative GC failed.\n");