X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fgencgc.c;h=c1fc6c85b92acfacd729273a17418a3dc97e1d60;hb=0bce9320787ec158000514620f3b7122642861b1;hp=b9bd86279cb3adf242c494720a85c4553573bda2;hpb=64eccd1724e5f1e638bfc574f7f376427ee3dcb1;p=sbcl.git diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c index b9bd862..c1fc6c8 100644 --- a/src/runtime/gencgc.c +++ b/src/runtime/gencgc.c @@ -55,6 +55,9 @@ #if defined(LUTEX_WIDETAG) #include "pthread-lutex.h" #endif +#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) +#include "genesis/cons.h" +#endif /* forward declarations */ page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes, @@ -69,9 +72,7 @@ page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes, * scratch space by the collector, and should never get collected. */ enum { - HIGHEST_NORMAL_GENERATION = 5, - PSEUDO_STATIC_GENERATION, - SCRATCH_GENERATION, + SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1, NUM_GENERATIONS }; @@ -80,7 +81,13 @@ enum { boolean enable_page_protection = 1; /* the minimum size (in bytes) for a large object*/ +#if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES) +long large_object_size = 4 * GENCGC_ALLOC_GRANULARITY; +#elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY) +long large_object_size = 4 * GENCGC_CARD_BYTES; +#else long large_object_size = 4 * PAGE_BYTES; +#endif /* @@ -89,7 +96,7 @@ long large_object_size = 4 * PAGE_BYTES; /* the verbosity level. All non-error messages are disabled at level 0; * and only a few rare messages are printed at level 1. */ -#ifdef QSHOW +#if QSHOW boolean gencgc_verbose = 1; #else boolean gencgc_verbose = 0; @@ -216,7 +223,7 @@ static void *heap_base = NULL; inline void * page_address(page_index_t page_num) { - return (heap_base + (page_num * PAGE_BYTES)); + return (heap_base + (page_num * GENCGC_CARD_BYTES)); } /* Calculate the address where the allocation region associated with @@ -234,7 +241,7 @@ find_page_index(void *addr) { if (addr >= heap_base) { page_index_t index = ((pointer_sized_uint_t)addr - - (pointer_sized_uint_t)heap_base) / PAGE_BYTES; + (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES; if (index < page_table_pages) return (index); } @@ -245,7 +252,7 @@ static size_t npage_bytes(long npages) { gc_assert(npages>=0); - return ((unsigned long)npages)*PAGE_BYTES; + return ((unsigned long)npages)*GENCGC_CARD_BYTES; } /* Check that X is a higher address than Y and return offset from Y to @@ -257,7 +264,12 @@ size_t void_diff(void *x, void *y) return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y; } -/* a structure to hold the state of a generation */ +/* a structure to hold the state of a generation + * + * CAUTION: If you modify this, make sure to touch up the alien + * definition in src/code/gc.lisp accordingly. ...or better yes, + * deal with the FIXME there... + */ struct generation { /* the first page that gc_alloc() checks on its next call */ @@ -287,9 +299,9 @@ struct generation { /* the number of GCs since the last raise */ int num_gc; - /* the average age after which a GC will raise objects to the + /* the number of GCs to run on the generations before raising objects to the * next generation */ - int trigger_age; + int number_of_gcs_before_promotion; /* the cumulative sum of the bytes allocated to this generation. It is * cleared after a GC on this generations, and update before new @@ -301,7 +313,7 @@ struct generation { /* a minimum average memory age before a GC will occur helps * prevent a GC when a large number of new live objects have been * added, in which case a GC could be a waste of time */ - double min_av_mem_age; + double minimum_age_before_gc; /* A linked list of lutex structures in this generation, used for * implementing lutex finalization. */ @@ -349,6 +361,12 @@ static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER; static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER; #endif +extern unsigned long gencgc_release_granularity; +unsigned long gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY; + +extern unsigned long gencgc_alloc_granularity; +unsigned long gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY; + /* * miscellaneous heap functions @@ -384,7 +402,7 @@ count_generation_pages(generation_index_t generation) return count; } -#ifdef QSHOW +#if QSHOW static long count_dont_move_pages(void) { @@ -416,8 +434,8 @@ count_generation_bytes_allocated (generation_index_t gen) } /* Return the average age of the memory in a generation. */ -static double -gen_av_mem_age(generation_index_t gen) +extern double +generation_average_age(generation_index_t gen) { if (generations[gen].bytes_allocated == 0) return 0.0; @@ -427,12 +445,10 @@ gen_av_mem_age(generation_index_t gen) / ((double)generations[gen].bytes_allocated); } -/* The verbose argument controls how much to print: 0 for normal - * level of detail; 1 for debugging. */ -static void -print_generation_stats(int verbose) /* FIXME: should take FILE argument */ +extern void +write_generation_stats(FILE *file) { - generation_index_t i, gens; + generation_index_t i; #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) #define FPU_STATE_SIZE 27 @@ -446,17 +462,11 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ * so they need to be saved and reset for C. */ fpu_save(fpu_state); - /* highest generation to print */ - if (verbose) - gens = SCRATCH_GENERATION; - else - gens = PSEUDO_STATIC_GENERATION; - /* Print the heap stats. */ - fprintf(stderr, + fprintf(file, " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n"); - for (i = 0; i < gens; i++) { + for (i = 0; i < SCRATCH_GENERATION; i++) { page_index_t j; long boxed_cnt = 0; long unboxed_cnt = 0; @@ -488,7 +498,7 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ gc_assert(generations[i].bytes_allocated == count_generation_bytes_allocated(i)); - fprintf(stderr, + fprintf(file, " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n", i, generations[i].alloc_start_page, @@ -506,13 +516,77 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ generations[i].gc_trigger, count_write_protect_generation_pages(i), generations[i].num_gc, - gen_av_mem_age(i)); + generation_average_age(i)); } - fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated); - fprintf(stderr," Dynamic-space-size bytes = %u\n", dynamic_space_size); + fprintf(file," Total bytes allocated = %lu\n", bytes_allocated); + fprintf(file," Dynamic-space-size bytes = %lu\n", (unsigned long)dynamic_space_size); fpu_restore(fpu_state); } + +extern void +write_heap_exhaustion_report(FILE *file, long available, long requested, + struct thread *thread) +{ + fprintf(file, + "Heap exhausted during %s: %ld bytes available, %ld requested.\n", + gc_active_p ? "garbage collection" : "allocation", + available, + requested); + write_generation_stats(file); + fprintf(file, "GC control variables:\n"); + fprintf(file, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n", + SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true", + (SymbolValue(GC_PENDING, thread) == T) ? + "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ? + "false" : "in progress")); +#ifdef LISP_FEATURE_SB_THREAD + fprintf(file, " *STOP-FOR-GC-PENDING* = %s\n", + SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true"); +#endif +} + +extern void +print_generation_stats(void) +{ + write_generation_stats(stderr); +} + +extern char* gc_logfile; +char * gc_logfile = NULL; + +extern void +log_generation_stats(char *logfile, char *header) +{ + if (logfile) { + FILE * log = fopen(logfile, "a"); + if (log) { + fprintf(log, "%s\n", header); + write_generation_stats(log); + fclose(log); + } else { + fprintf(stderr, "Could not open gc logfile: %s\n", logfile); + fflush(stderr); + } + } +} + +extern void +report_heap_exhaustion(long available, long requested, struct thread *th) +{ + if (gc_logfile) { + FILE * log = fopen(gc_logfile, "a"); + if (log) { + write_heap_exhaustion_report(log, available, requested, th); + fclose(log); + } else { + fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile); + fflush(stderr); + } + } + /* Always to stderr as well. */ + write_heap_exhaustion_report(stderr, available, requested, th); +} #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) @@ -531,6 +605,9 @@ void zero_pages_with_mmap(page_index_t start, page_index_t end) { if (start > end) return; + gc_assert(length >= gencgc_release_granularity); + gc_assert((length % gencgc_release_granularity) == 0); + os_invalidate(addr, length); new_addr = os_validate(addr, length); if (new_addr == NULL || new_addr != addr) { @@ -559,19 +636,28 @@ zero_pages(page_index_t start, page_index_t end) { } +static void +zero_and_mark_pages(page_index_t start, page_index_t end) { + page_index_t i; + + zero_pages(start, end); + for (i = start; i <= end; i++) + page_table[i].need_to_zero = 0; +} + /* Zero the pages from START to END (inclusive), except for those * pages that are known to already zeroed. Mark all pages in the * ranges as non-zeroed. */ static void zero_dirty_pages(page_index_t start, page_index_t end) { - page_index_t i; + page_index_t i, j; for (i = start; i <= end; i++) { - if (page_table[i].need_to_zero == 1) { - zero_pages(start, end); - break; - } + if (!page_table[i].need_to_zero) continue; + for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++); + zero_pages(i, j-1); + i = j; } for (i = start; i <= end; i++) { @@ -729,7 +815,7 @@ gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_ gc_assert(ret == 0); first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0); last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag); - bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used) + bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used) + npage_bytes(last_page-first_page); /* Set up the alloc_region. */ @@ -959,8 +1045,8 @@ gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_regio more = 0; if ((bytes_used = void_diff(alloc_region->free_pointer, page_address(first_page))) - >PAGE_BYTES) { - bytes_used = PAGE_BYTES; + >GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[first_page].bytes_used = bytes_used; @@ -984,8 +1070,8 @@ gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_regio /* Calculate the number of bytes used in this page. */ more = 0; if ((bytes_used = void_diff(alloc_region->free_pointer, - page_address(next_page)))>PAGE_BYTES) { - bytes_used = PAGE_BYTES; + page_address(next_page)))>GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[next_page].bytes_used = bytes_used; @@ -1086,8 +1172,8 @@ gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_regio /* Calc. the number of bytes used in this page. This is not * always the number of new bytes, unless it was free. */ more = 0; - if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) { - bytes_used = PAGE_BYTES; + if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[first_page].bytes_used = bytes_used; @@ -1111,8 +1197,8 @@ gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_regio /* Calculate the number of bytes used in this page. */ more = 0; bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt; - if (bytes_used > PAGE_BYTES) { - bytes_used = PAGE_BYTES; + if (bytes_used > GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[next_page].bytes_used = bytes_used; @@ -1161,24 +1247,11 @@ gc_heap_exhausted_error_or_lose (long available, long requested) * the danger that we bounce back here before the error has been * handled, or indeed even printed. */ - fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n", - gc_active_p ? "garbage collection" : "allocation", - available, requested); + report_heap_exhaustion(available, requested, thread); if (gc_active_p || (available == 0)) { /* If we are in GC, or totally out of memory there is no way * to sanely transfer control to the lisp-side of things. */ - print_generation_stats(1); - fprintf(stderr, "GC control variables:\n"); - fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n", - SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true", - (SymbolValue(GC_PENDING, thread) == T) ? - "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ? - "false" : "in progress")); -#ifdef LISP_FEATURE_SB_THREAD - fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n", - SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true"); -#endif lose("Heap exhausted, game over."); } else { @@ -1208,84 +1281,87 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, { page_index_t first_page, last_page; page_index_t restart_page = *restart_page_ptr; + long nbytes_goal = nbytes; long bytes_found = 0; long most_bytes_found = 0; + page_index_t most_bytes_found_from, most_bytes_found_to; + int small_object = nbytes < GENCGC_CARD_BYTES; /* FIXME: assert(free_pages_lock is held); */ + if (nbytes_goal < gencgc_alloc_granularity) + nbytes_goal = gencgc_alloc_granularity; + /* Toggled by gc_and_save for heap compaction, normally -1. */ if (gencgc_alloc_start_page != -1) { restart_page = gencgc_alloc_start_page; } gc_assert(nbytes>=0); - if (((unsigned long)nbytes)>=PAGE_BYTES) { - /* Search for a contiguous free space of at least nbytes, - * aligned on a page boundary. The page-alignment is strictly - * speaking needed only for objects at least large_object_size - * bytes in size. */ - do { - first_page = restart_page; - while ((first_page < page_table_pages) && - page_allocated_p(first_page)) + /* Search for a page with at least nbytes of space. We prefer + * not to split small objects on multiple pages, to reduce the + * number of contiguous allocation regions spaning multiple + * pages: this helps avoid excessive conservativism. + * + * For other objects, we guarantee that they start on their own + * page boundary. + */ + first_page = restart_page; + while (first_page < page_table_pages) { + bytes_found = 0; + if (page_free_p(first_page)) { + gc_assert(0 == page_table[first_page].bytes_used); + bytes_found = GENCGC_CARD_BYTES; + } else if (small_object && + (page_table[first_page].allocated == page_type_flag) && + (page_table[first_page].large_object == 0) && + (page_table[first_page].gen == gc_alloc_generation) && + (page_table[first_page].write_protected == 0) && + (page_table[first_page].dont_move == 0)) { + bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used; + if (bytes_found < nbytes) { + if (bytes_found > most_bytes_found) + most_bytes_found = bytes_found; first_page++; - - last_page = first_page; - bytes_found = PAGE_BYTES; - while ((bytes_found < nbytes) && - (last_page < (page_table_pages-1)) && - page_free_p(last_page+1)) { - last_page++; - bytes_found += PAGE_BYTES; - gc_assert(0 == page_table[last_page].bytes_used); - gc_assert(0 == page_table[last_page].write_protected); + continue; } - if (bytes_found > most_bytes_found) - most_bytes_found = bytes_found; - restart_page = last_page + 1; - } while ((restart_page < page_table_pages) && (bytes_found < nbytes)); - - } else { - /* Search for a page with at least nbytes of space. We prefer - * not to split small objects on multiple pages, to reduce the - * number of contiguous allocation regions spaning multiple - * pages: this helps avoid excessive conservativism. */ - first_page = restart_page; - while (first_page < page_table_pages) { - if (page_free_p(first_page)) - { - gc_assert(0 == page_table[first_page].bytes_used); - bytes_found = PAGE_BYTES; - break; - } - else if ((page_table[first_page].allocated == page_type_flag) && - (page_table[first_page].large_object == 0) && - (page_table[first_page].gen == gc_alloc_generation) && - (page_table[first_page].write_protected == 0) && - (page_table[first_page].dont_move == 0)) - { - bytes_found = PAGE_BYTES - - page_table[first_page].bytes_used; - if (bytes_found > most_bytes_found) - most_bytes_found = bytes_found; - if (bytes_found >= nbytes) - break; - } + } else { first_page++; + continue; } - last_page = first_page; - restart_page = first_page + 1; + + gc_assert(page_table[first_page].write_protected == 0); + for (last_page = first_page+1; + ((last_page < page_table_pages) && + page_free_p(last_page) && + (bytes_found < nbytes_goal)); + last_page++) { + bytes_found += GENCGC_CARD_BYTES; + gc_assert(0 == page_table[last_page].bytes_used); + gc_assert(0 == page_table[last_page].write_protected); + } + + if (bytes_found > most_bytes_found) { + most_bytes_found = bytes_found; + most_bytes_found_from = first_page; + most_bytes_found_to = last_page; + } + if (bytes_found >= nbytes_goal) + break; + + first_page = last_page; } + bytes_found = most_bytes_found; + restart_page = first_page + 1; + /* Check for a failure */ if (bytes_found < nbytes) { gc_assert(restart_page >= page_table_pages); gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes); } - gc_assert(page_table[first_page].write_protected == 0); - - *restart_page_ptr = first_page; - return last_page; + *restart_page_ptr = most_bytes_found_from; + return most_bytes_found_to-1; } /* Allocate bytes. All the rest of the special-purpose allocation @@ -1408,23 +1484,19 @@ copy_large_object(lispobj object, long nwords) next_page = first_page; remaining_bytes = nwords*N_WORD_BYTES; - while (remaining_bytes > PAGE_BYTES) { + while (remaining_bytes > GENCGC_CARD_BYTES) { gc_assert(page_table[next_page].gen == from_space); gc_assert(page_boxed_p(next_page)); gc_assert(page_table[next_page].large_object); gc_assert(page_table[next_page].region_start_offset == npage_bytes(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES); + /* Should have been unprotected by unprotect_oldspace(). */ + gc_assert(page_table[next_page].write_protected == 0); page_table[next_page].gen = new_space; - /* Remove any write-protection. We should be able to rely - * on the write-protect flag to avoid redundant calls. */ - if (page_table[next_page].write_protected) { - os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL); - page_table[next_page].write_protected = 0; - } - remaining_bytes -= PAGE_BYTES; + remaining_bytes -= GENCGC_CARD_BYTES; next_page++; } @@ -1445,7 +1517,7 @@ copy_large_object(lispobj object, long nwords) /* Free any remaining pages; needs care. */ next_page++; - while ((old_bytes_used == PAGE_BYTES) && + while ((old_bytes_used == GENCGC_CARD_BYTES) && (page_table[next_page].gen == from_space) && page_boxed_p(next_page) && page_table[next_page].large_object && @@ -1554,17 +1626,17 @@ copy_large_unboxed_object(lispobj object, long nwords) next_page = first_page; remaining_bytes = nwords*N_WORD_BYTES; - while (remaining_bytes > PAGE_BYTES) { + while (remaining_bytes > GENCGC_CARD_BYTES) { gc_assert(page_table[next_page].gen == from_space); gc_assert(page_allocated_no_region_p(next_page)); gc_assert(page_table[next_page].large_object); gc_assert(page_table[next_page].region_start_offset == npage_bytes(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES); page_table[next_page].gen = new_space; page_table[next_page].allocated = UNBOXED_PAGE_FLAG; - remaining_bytes -= PAGE_BYTES; + remaining_bytes -= GENCGC_CARD_BYTES; next_page++; } @@ -1585,7 +1657,7 @@ copy_large_unboxed_object(lispobj object, long nwords) /* Free any remaining pages; needs care. */ next_page++; - while ((old_bytes_used == PAGE_BYTES) && + while ((old_bytes_used == GENCGC_CARD_BYTES) && (page_table[next_page].gen == from_space) && page_allocated_no_region_p(next_page) && page_table[next_page].large_object && @@ -1684,7 +1756,7 @@ sniff_code_object(struct code *code, unsigned long displacement) unsigned d2 = *((unsigned char *)p - 2); unsigned d3 = *((unsigned char *)p - 3); unsigned d4 = *((unsigned char *)p - 4); -#ifdef QSHOW +#if QSHOW unsigned d5 = *((unsigned char *)p - 5); unsigned d6 = *((unsigned char *)p - 6); #endif @@ -2221,8 +2293,6 @@ search_dynamic_space(void *pointer) (lispobj *)pointer)); } -#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) - /* Helper for valid_lisp_pointer_p and * possibly_valid_dynamic_space_pointer. * @@ -2244,8 +2314,12 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) * header. */ switch (widetag_of(*start_addr)) { case CODE_HEADER_WIDETAG: - /* This case is probably caught above. */ - break; + /* Make sure we actually point to a function in the code object, + * as opposed to a random point there. */ + if (SIMPLE_FUN_HEADER_WIDETAG==widetag_of(*(pointer-FUN_POINTER_LOWTAG))) + return 1; + else + return 0; case CLOSURE_HEADER_WIDETAG: case FUNCALLABLE_INSTANCE_HEADER_WIDETAG: if ((unsigned long)pointer != @@ -2311,6 +2385,23 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) } break; case OTHER_POINTER_LOWTAG: + +#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) + /* The all-architecture test below is good as far as it goes, + * but an LRA object is similar to a FUN-POINTER: It is + * embedded within a CODE-OBJECT pointed to by start_addr, and + * cannot be found by simply walking the heap, therefore we + * need to check for it. -- AB, 2010-Jun-04 */ + if ((widetag_of(start_addr[0]) == CODE_HEADER_WIDETAG)) { + lispobj *potential_lra = + (lispobj *)(((unsigned long)pointer) - OTHER_POINTER_LOWTAG); + if ((widetag_of(potential_lra[0]) == RETURN_PC_HEADER_WIDETAG) && + ((potential_lra - HeaderValue(potential_lra[0])) == start_addr)) { + return 1; /* It's as good as we can verify. */ + } + } +#endif + if ((unsigned long)pointer != ((unsigned long)start_addr+OTHER_POINTER_LOWTAG)) { if (gencgc_verbose) { @@ -2506,6 +2597,8 @@ valid_lisp_pointer_p(lispobj *pointer) return 0; } +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + /* Is there any possibility that pointer is a valid Lisp object * reference, and/or something else (e.g. subroutine call return * address) which should prevent us from moving the referred-to thing? @@ -2523,6 +2616,8 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer) return looks_like_valid_lisp_pointer_p(pointer, start_addr); } +#endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + /* Adjust large bignum and vector objects. This will adjust the * allocated region if the size has shrunk, and move unboxed objects * into unboxed pages. The pages are not promoted here, and the @@ -2629,20 +2724,20 @@ maybe_adjust_large_object(lispobj *where) next_page = first_page; remaining_bytes = nwords*N_WORD_BYTES; - while (remaining_bytes > PAGE_BYTES) { + while (remaining_bytes > GENCGC_CARD_BYTES) { gc_assert(page_table[next_page].gen == from_space); gc_assert(page_allocated_no_region_p(next_page)); gc_assert(page_table[next_page].large_object); gc_assert(page_table[next_page].region_start_offset == npage_bytes(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES); page_table[next_page].allocated = boxed; /* Shouldn't be write-protected at this stage. Essential that the * pages aren't. */ gc_assert(!page_table[next_page].write_protected); - remaining_bytes -= PAGE_BYTES; + remaining_bytes -= GENCGC_CARD_BYTES; next_page++; } @@ -2664,7 +2759,7 @@ maybe_adjust_large_object(lispobj *where) /* Free any remaining pages; needs care. */ next_page++; - while ((old_bytes_used == PAGE_BYTES) && + while ((old_bytes_used == GENCGC_CARD_BYTES) && (page_table[next_page].gen == from_space) && page_allocated_no_region_p(next_page) && page_table[next_page].large_object && @@ -2732,7 +2827,7 @@ preserve_pointer(void *addr) /* quick check 2: Check the offset within the page. * */ - if (((unsigned long)addr & (PAGE_BYTES - 1)) > + if (((unsigned long)addr & (GENCGC_CARD_BYTES - 1)) > page_table[addr_page_index].bytes_used) return; @@ -2741,11 +2836,17 @@ preserve_pointer(void *addr) * address referring to something in a CodeObject). This is * expensive but important, since it vastly reduces the * probability that random garbage will be bogusly interpreted as - * a pointer which prevents a page from moving. */ + * a pointer which prevents a page from moving. + * + * This only needs to happen on x86oids, where this is used for + * conservative roots. Non-x86oid systems only ever call this + * function on known-valid lisp objects. */ +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) if (!(code_page_p(addr_page_index) || (is_lisp_pointer((lispobj)addr) && possibly_valid_dynamic_space_pointer(addr)))) return; +#endif /* Find the beginning of the region. Note that there may be * objects in the region preceding the one that we were passed a @@ -2761,7 +2862,7 @@ preserve_pointer(void *addr) while (page_table[first_page].region_start_offset != 0) { --first_page; /* Do some checks. */ - gc_assert(page_table[first_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES); gc_assert(page_table[first_page].gen == from_space); gc_assert(page_table[first_page].allocated == region_allocation); } @@ -2778,7 +2879,7 @@ preserve_pointer(void *addr) if (page_free_p(addr_page_index) || (page_table[addr_page_index].bytes_used == 0) /* Check the offset within the page. */ - || (((unsigned long)addr & (PAGE_BYTES - 1)) + || (((unsigned long)addr & (GENCGC_CARD_BYTES - 1)) > page_table[addr_page_index].bytes_used)) { FSHOW((stderr, "weird? ignore ptr 0x%x to freed area of large object\n", @@ -2812,8 +2913,8 @@ preserve_pointer(void *addr) gc_assert(!page_table[i].write_protected); /* Check whether this is the last page in this contiguous block.. */ - if ((page_table[i].bytes_used < PAGE_BYTES) - /* ..or it is PAGE_BYTES and is the last in the block */ + if ((page_table[i].bytes_used < GENCGC_CARD_BYTES) + /* ..or it is CARD_BYTES and is the last in the block */ || page_free_p(i+1) || (page_table[i+1].bytes_used == 0) /* next page free */ || (page_table[i+1].gen != from_space) /* diff. gen */ @@ -2824,9 +2925,6 @@ preserve_pointer(void *addr) /* Check that the page is now static. */ gc_assert(page_table[addr_page_index].dont_move != 0); } - -#endif // defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) - /* If the given page is not write-protected, then scan it for pointers * to younger generations or the top temp. generation, if no @@ -2891,7 +2989,7 @@ update_page_write_prot(page_index_t page) /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/ os_protect((void *)page_addr, - PAGE_BYTES, + GENCGC_CARD_BYTES, OS_VM_PROT_READ|OS_VM_PROT_EXECUTE); /* Note the page as protected in the page tables. */ @@ -2960,8 +3058,8 @@ scavenge_generations(generation_index_t from, generation_index_t to) for (last_page = i; ; last_page++) { write_protected = write_protected && page_table[last_page].write_protected; - if ((page_table[last_page].bytes_used < PAGE_BYTES) - /* Or it is PAGE_BYTES and is the last in the block */ + if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES) + /* Or it is CARD_BYTES and is the last in the block */ || (!page_boxed_p(last_page+1)) || (page_table[last_page+1].bytes_used == 0) || (page_table[last_page+1].gen != generation) @@ -3076,8 +3174,8 @@ scavenge_newspace_generation_one_scan(generation_index_t generation) /* Check whether this is the last page in this * contiguous block */ - if ((page_table[last_page].bytes_used < PAGE_BYTES) - /* Or it is PAGE_BYTES and is the last in the block */ + if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES) + /* Or it is CARD_BYTES and is the last in the block */ || (!page_boxed_p(last_page+1)) || (page_table[last_page+1].bytes_used == 0) || (page_table[last_page+1].gen != generation) @@ -3248,23 +3346,41 @@ static void unprotect_oldspace(void) { page_index_t i; + void *region_addr = 0; + void *page_addr = 0; + unsigned long region_bytes = 0; for (i = 0; i < last_free_page; i++) { if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == from_space)) { - void *page_start; - - page_start = (void *)page_address(i); /* Remove any write-protection. We should be able to rely * on the write-protect flag to avoid redundant calls. */ if (page_table[i].write_protected) { - os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL); page_table[i].write_protected = 0; + page_addr = page_address(i); + if (!region_addr) { + /* First region. */ + region_addr = page_addr; + region_bytes = GENCGC_CARD_BYTES; + } else if (region_addr + region_bytes == page_addr) { + /* Region continue. */ + region_bytes += GENCGC_CARD_BYTES; + } else { + /* Unprotect previous region. */ + os_protect(region_addr, region_bytes, OS_VM_PROT_ALL); + /* First page in new region. */ + region_addr = page_addr; + region_bytes = GENCGC_CARD_BYTES; + } } } } + if (region_addr) { + /* Unprotect last region. */ + os_protect(region_addr, region_bytes, OS_VM_PROT_ALL); + } } /* Work through all the pages and free any in from_space. This @@ -3300,17 +3416,8 @@ free_oldspace(void) page_table[last_page].bytes_used; page_table[last_page].allocated = FREE_PAGE_FLAG; page_table[last_page].bytes_used = 0; - - /* Remove any write-protection. We should be able to rely - * on the write-protect flag to avoid redundant calls. */ - { - void *page_start = (void *)page_address(last_page); - - if (page_table[last_page].write_protected) { - os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL); - page_table[last_page].write_protected = 0; - } - } + /* Should already be unprotected by unprotect_oldspace(). */ + gc_assert(!page_table[last_page].write_protected); last_page++; } while ((last_page < last_free_page) @@ -3360,6 +3467,23 @@ print_ptr(lispobj *addr) } #endif +static int +is_in_stack_space(lispobj ptr) +{ + /* For space verification: Pointers can be valid if they point + * to a thread stack space. This would be faster if the thread + * structures had page-table entries as if they were part of + * the heap space. */ + struct thread *th; + for_each_thread(th) { + if ((th->control_stack_start <= (lispobj *)ptr) && + (th->control_stack_end >= (lispobj *)ptr)) { + return 1; + } + } + return 0; +} + static void verify_space(lispobj *start, size_t words) { @@ -3387,15 +3511,15 @@ verify_space(lispobj *start, size_t words) * page. XX Could check the offset too. */ if (page_allocated_p(page_index) && (page_table[page_index].bytes_used == 0)) - lose ("Ptr %x @ %x sees free page.\n", thing, start); + lose ("Ptr %p @ %p sees free page.\n", thing, start); /* Check that it doesn't point to a forwarding pointer! */ if (*((lispobj *)native_pointer(thing)) == 0x01) { - lose("Ptr %x @ %x sees forwarding ptr.\n", thing, start); + lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start); } /* Check that its not in the RO space as it would then be a * pointer from the RO to the dynamic space. */ if (is_in_readonly_space) { - lose("ptr to dynamic space %x from RO space %x\n", + lose("ptr to dynamic space %p from RO space %x\n", thing, start); } /* Does it point to a plausible object? This check slows @@ -3409,13 +3533,16 @@ verify_space(lispobj *start, size_t words) * dynamically. */ /* if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) { - lose("ptr %x to invalid object %x\n", thing, start); + lose("ptr %p to invalid object %p\n", thing, start); } */ } else { + extern void funcallable_instance_tramp; /* Verify that it points to another valid space. */ - if (!to_readonly_space && !to_static_space) { - lose("Ptr %x @ %x sees junk.\n", thing, start); + if (!to_readonly_space && !to_static_space + && (thing != (lispobj)&funcallable_instance_tramp) + && !is_in_stack_space(thing)) { + lose("Ptr %p @ %p sees junk.\n", thing, start); } } } else { @@ -3492,7 +3619,7 @@ verify_space(lispobj *start, size_t words) /* Only when enabled */ && verify_dynamic_code_check) { FSHOW((stderr, - "/code object at %x in the dynamic space\n", + "/code object at %p in the dynamic space\n", start)); } @@ -3608,7 +3735,7 @@ verify_space(lispobj *start, size_t words) break; default: - lose("Unhandled widetag 0x%x at 0x%x\n", + lose("Unhandled widetag %p at %p\n", widetag_of(*start), start); } } @@ -3667,8 +3794,8 @@ verify_generation(generation_index_t generation) for (last_page = i; ;last_page++) /* Check whether this is the last page in this contiguous * block. */ - if ((page_table[last_page].bytes_used < PAGE_BYTES) - /* Or it is PAGE_BYTES and is the last in the block */ + if ((page_table[last_page].bytes_used < GENCGC_CARD_BYTES) + /* Or it is CARD_BYTES and is the last in the block */ || (page_table[last_page+1].allocated != region_allocation) || (page_table[last_page+1].bytes_used == 0) || (page_table[last_page+1].gen != generation) @@ -3703,7 +3830,7 @@ verify_zero_fill(void) } } } else { - long free_bytes = PAGE_BYTES - page_table[page].bytes_used; + long free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used; if (free_bytes > 0) { long *start_addr = (long *)((unsigned long)page_address(page) + page_table[page].bytes_used); @@ -3783,165 +3910,19 @@ write_protect_generation_pages(generation_index_t generation) } #if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) - static void -scavenge_control_stack() +scavenge_control_stack(struct thread *th) { - unsigned long control_stack_size; - - /* This is going to be a big problem when we try to port threads - * to PPC... CLH */ - struct thread *th = arch_os_get_current_thread(); lispobj *control_stack = (lispobj *)(th->control_stack_start); + unsigned long control_stack_size = + access_control_stack_pointer(th) - control_stack; - control_stack_size = current_control_stack_pointer - control_stack; scavenge(control_stack, control_stack_size); } - -/* Scavenging Interrupt Contexts */ - -static int boxed_registers[] = BOXED_REGISTERS; - -static void -scavenge_interrupt_context(os_context_t * context) -{ - int i; - -#ifdef reg_LIP - unsigned long lip; - unsigned long lip_offset; - int lip_register_pair; -#endif - unsigned long pc_code_offset; - -#ifdef ARCH_HAS_LINK_REGISTER - unsigned long lr_code_offset; -#endif -#ifdef ARCH_HAS_NPC_REGISTER - unsigned long npc_code_offset; -#endif - -#ifdef reg_LIP - /* Find the LIP's register pair and calculate it's offset */ - /* before we scavenge the context. */ - - /* - * I (RLT) think this is trying to find the boxed register that is - * closest to the LIP address, without going past it. Usually, it's - * reg_CODE or reg_LRA. But sometimes, nothing can be found. - */ - lip = *os_context_register_addr(context, reg_LIP); - lip_offset = 0x7FFFFFFF; - lip_register_pair = -1; - for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) { - unsigned long reg; - long offset; - int index; - - index = boxed_registers[i]; - reg = *os_context_register_addr(context, index); - if ((reg & ~((1L<uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is - * that what we really want? My guess is that that is not what we - * want, so if lip_register_pair is -1, we don't touch reg_LIP at - * all. But maybe it doesn't really matter if LIP is trashed? - */ - if (lip_register_pair >= 0) { - *os_context_register_addr(context, reg_LIP) = - *os_context_register_addr(context, lip_register_pair) - + lip_offset; - } -#endif /* reg_LIP */ - - /* Fix the PC if it was in from space */ - if (from_space_p(*os_context_pc_addr(context))) - *os_context_pc_addr(context) = - *os_context_register_addr(context, reg_CODE) + pc_code_offset; - -#ifdef ARCH_HAS_LINK_REGISTER - /* Fix the LR ditto; important if we're being called from - * an assembly routine that expects to return using blr, otherwise - * harmless */ - if (from_space_p(*os_context_lr_addr(context))) - *os_context_lr_addr(context) = - *os_context_register_addr(context, reg_CODE) + lr_code_offset; -#endif - -#ifdef ARCH_HAS_NPC_REGISTER - if (from_space_p(*os_context_npc_addr(context))) - *os_context_npc_addr(context) = - *os_context_register_addr(context, reg_CODE) + npc_code_offset; -#endif /* ARCH_HAS_NPC_REGISTER */ -} - -void -scavenge_interrupt_contexts(void) -{ - int i, index; - os_context_t *context; - - struct thread *th=arch_os_get_current_thread(); - - index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0)); - -#if defined(DEBUG_PRINT_CONTEXT_INDEX) - printf("Number of active contexts: %d\n", index); -#endif - - for (i = 0; i < index; i++) { - context = th->interrupt_contexts[i]; - scavenge_interrupt_context(context); - } -} - #endif -#if defined(LISP_FEATURE_SB_THREAD) +#if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)) static void preserve_context_registers (os_context_t *c) { @@ -3992,9 +3973,8 @@ garbage_collect_generation(generation_index_t generation, int raise) unsigned long bytes_freed; page_index_t i; unsigned long static_space_size; -#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) struct thread *th; -#endif + gc_assert(generation <= HIGHEST_NORMAL_GENERATION); /* The oldest generation can't be raised. */ @@ -4094,9 +4074,22 @@ garbage_collect_generation(generation_index_t generation, int raise) } } } +#else + /* Non-x86oid systems don't have "conservative roots" as such, but + * the same mechanism is used for objects pinned for use by alien + * code. */ + for_each_thread(th) { + lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th); + while (pin_list != NIL) { + struct cons *list_entry = + (struct cons *)native_pointer(pin_list); + preserve_pointer(list_entry->car); + pin_list = list_entry->cdr; + } + } #endif -#ifdef QSHOW +#if QSHOW if (gencgc_verbose > 1) { long num_dont_move_pages = count_dont_move_pages(); fprintf(stderr, @@ -4113,8 +4106,18 @@ garbage_collect_generation(generation_index_t generation, int raise) * If not x86, we need to scavenge the interrupt context(s) and the * control stack. */ - scavenge_interrupt_contexts(); - scavenge_control_stack(); + { + struct thread *th; + for_each_thread(th) { + scavenge_interrupt_contexts(th); + scavenge_control_stack(th); + } + + /* Scrub the unscavenged control stack space, so that we can't run + * into any stale pointers in a later GC (this is done by the + * stop-for-gc handler in the other threads). */ + scrub_control_stack(); + } #endif /* Scavenge the Lisp functions of the interrupt handlers, taking @@ -4280,32 +4283,55 @@ update_dynamic_space_free_pointer(void) } static void -remap_free_pages (page_index_t from, page_index_t to) +remap_page_range (page_index_t from, page_index_t to) { - page_index_t first_page, last_page; + /* There's a mysterious Solaris/x86 problem with using mmap + * tricks for memory zeroing. See sbcl-devel thread + * "Re: patch: standalone executable redux". + */ +#if defined(LISP_FEATURE_SUNOS) + zero_and_mark_pages(from, to); +#else + const page_index_t + release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES, + release_mask = release_granularity-1, + end = to+1, + aligned_from = (from+release_mask)&~release_mask, + aligned_end = (end&~release_mask); + + if (aligned_from < aligned_end) { + zero_pages_with_mmap(aligned_from, aligned_end-1); + if (aligned_from != from) + zero_and_mark_pages(from, aligned_from-1); + if (aligned_end != end) + zero_and_mark_pages(aligned_end, end-1); + } else { + zero_and_mark_pages(from, to); + } +#endif +} + +static void +remap_free_pages (page_index_t from, page_index_t to, int forcibly) +{ + page_index_t first_page, last_page, + first_aligned_page, last_aligned_page; + + if (forcibly) + return remap_page_range(from, to); for (first_page = from; first_page <= to; first_page++) { if (page_allocated_p(first_page) || - (page_table[first_page].need_to_zero == 0)) { + (page_table[first_page].need_to_zero == 0)) continue; - } last_page = first_page + 1; while (page_free_p(last_page) && - (last_page < to) && - (page_table[last_page].need_to_zero == 1)) { + (last_page <= to) && + (page_table[last_page].need_to_zero == 1)) last_page++; - } - /* There's a mysterious Solaris/x86 problem with using mmap - * tricks for memory zeroing. See sbcl-devel thread - * "Re: patch: standalone executable redux". - */ -#if defined(LISP_FEATURE_SUNOS) - zero_pages(first_page, last_page-1); -#else - zero_pages_with_mmap(first_page, last_page-1); -#endif + remap_page_range(first_page, last_page-1); first_page = last_page; } @@ -4332,6 +4358,7 @@ collect_garbage(generation_index_t last_gen) static page_index_t high_water_mark = 0; FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen)); + log_generation_stats(gc_logfile, "=== GC Start ==="); gc_active_p = 1; @@ -4352,7 +4379,7 @@ collect_garbage(generation_index_t last_gen) } if (gencgc_verbose > 1) - print_generation_stats(0); + print_generation_stats(); do { /* Collect the generation. */ @@ -4363,7 +4390,7 @@ collect_garbage(generation_index_t last_gen) } else { raise = (gen < last_gen) - || (generations[gen].num_gc >= generations[gen].trigger_age); + || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion); } if (gencgc_verbose > 1) { @@ -4390,7 +4417,7 @@ collect_garbage(generation_index_t last_gen) if (gencgc_verbose > 1) { FSHOW((stderr, "GC of generation %d finished:\n", gen)); - print_generation_stats(0); + print_generation_stats(); } gen++; @@ -4400,8 +4427,8 @@ collect_garbage(generation_index_t last_gen) && raise && (generations[gen].bytes_allocated > generations[gen].gc_trigger) - && (gen_av_mem_age(gen) - > generations[gen].min_av_mem_age)))); + && (generation_average_age(gen) + > generations[gen].minimum_age_before_gc)))); /* Now if gen-1 was raised all generations before gen are empty. * If it wasn't raised then all generations before gen-1 are empty. @@ -4450,12 +4477,13 @@ collect_garbage(generation_index_t last_gen) if (gen > small_generation_limit) { if (last_free_page > high_water_mark) high_water_mark = last_free_page; - remap_free_pages(0, high_water_mark); + remap_free_pages(0, high_water_mark, 0); high_water_mark = 0; } gc_active_p = 0; + log_generation_stats(gc_logfile, "=== GC End ==="); SHOW("returning from collect_garbage"); } @@ -4467,7 +4495,7 @@ collect_garbage(generation_index_t last_gen) void gc_free_heap(void) { - page_index_t page; + page_index_t page, last_page; if (gencgc_verbose > 1) { SHOW("entering gc_free_heap"); @@ -4477,33 +4505,25 @@ gc_free_heap(void) /* Skip free pages which should already be zero filled. */ if (page_allocated_p(page)) { void *page_start, *addr; - - /* Mark the page free. The other slots are assumed invalid - * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it - * should not be write-protected -- except that the - * generation is used for the current region but it sets - * that up. */ - page_table[page].allocated = FREE_PAGE_FLAG; - page_table[page].bytes_used = 0; + for (last_page = page; + (last_page < page_table_pages) && page_allocated_p(last_page); + last_page++) { + /* Mark the page free. The other slots are assumed invalid + * when it is a FREE_PAGE_FLAG and bytes_used is 0 and it + * should not be write-protected -- except that the + * generation is used for the current region but it sets + * that up. */ + page_table[page].allocated = FREE_PAGE_FLAG; + page_table[page].bytes_used = 0; + page_table[page].write_protected = 0; + } #ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure * about this change. */ - /* Zero the page. */ page_start = (void *)page_address(page); - - /* First, remove any write-protection. */ - os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL); - page_table[page].write_protected = 0; - - os_invalidate(page_start,PAGE_BYTES); - addr = os_validate(page_start,PAGE_BYTES); - if (addr == NULL || addr != page_start) { - lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x\n", - page_start, - addr); - } -#else - page_table[page].write_protected = 0; + os_protect(page_start, npage_bytes(last_page-page), OS_VM_PROT_ALL); + remap_free_pages(page, last_page-1, 1); + page = last_page-1; #endif } else if (gencgc_zero_check_during_free_heap) { /* Double-check that the page is zero filled. */ @@ -4512,7 +4532,7 @@ gc_free_heap(void) gc_assert(page_free_p(page)); gc_assert(page_table[page].bytes_used == 0); page_start = (long *)page_address(page); - for (i=0; i<1024; i++) { + for (i=0; i 1) - print_generation_stats(0); + print_generation_stats(); /* Initialize gc_alloc(). */ gc_alloc_generation = 0; @@ -4561,9 +4581,13 @@ gc_init(void) /* Compute the number of pages needed for the dynamic space. * Dynamic space size should be aligned on page size. */ - page_table_pages = dynamic_space_size/PAGE_BYTES; + page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES; gc_assert(dynamic_space_size == npage_bytes(page_table_pages)); + /* The page_table must be allocated using "calloc" to initialize + * the page structures correctly. There used to be a separate + * initialization loop (now commented out; see below) but that was + * unnecessary and did hurt startup time. */ page_table = calloc(page_table_pages, sizeof(struct page)); gc_assert(page_table); @@ -4579,14 +4603,38 @@ gc_init(void) heap_base = (void*)DYNAMIC_SPACE_START; - /* Initialize each page structure. */ - for (i = 0; i < page_table_pages; i++) { - /* Initialize all pages as free. */ - page_table[i].allocated = FREE_PAGE_FLAG; - page_table[i].bytes_used = 0; - - /* Pages are not write-protected at startup. */ - page_table[i].write_protected = 0; + /* The page structures are initialized implicitly when page_table + * is allocated with "calloc" above. Formerly we had the following + * explicit initialization here (comments converted to C99 style + * for readability as C's block comments don't nest): + * + * // Initialize each page structure. + * for (i = 0; i < page_table_pages; i++) { + * // Initialize all pages as free. + * page_table[i].allocated = FREE_PAGE_FLAG; + * page_table[i].bytes_used = 0; + * + * // Pages are not write-protected at startup. + * page_table[i].write_protected = 0; + * } + * + * Without this loop the image starts up much faster when dynamic + * space is large -- which it is on 64-bit platforms already by + * default -- and when "calloc" for large arrays is implemented + * using copy-on-write of a page of zeroes -- which it is at least + * on Linux. In this case the pages that page_table_pages is stored + * in are mapped and cleared not before the corresponding part of + * dynamic space is used. For example, this saves clearing 16 MB of + * memory at startup if the page size is 4 KB and the size of + * dynamic space is 4 GB. + * FREE_PAGE_FLAG must be 0 for this to work correctly which is + * asserted below: */ + { + /* Compile time assertion: If triggered, declares an array + * of dimension -1 forcing a syntax error. The intent of the + * assignment is to avoid an "unused variable" warning. */ + char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1]; + assert_free_page_flag_0[0] = assert_free_page_flag_0[0]; } bytes_allocated = 0; @@ -4605,8 +4653,8 @@ gc_init(void) generations[i].cum_sum_bytes_allocated = 0; /* the tune-able parameters */ generations[i].bytes_consed_between_gc = 2000000; - generations[i].trigger_age = 1; - generations[i].min_av_mem_age = 0.75; + generations[i].number_of_gcs_before_promotion = 1; + generations[i].minimum_age_before_gc = 0.75; generations[i].lutexes = NULL; } @@ -4632,18 +4680,24 @@ gencgc_pickup_dynamic(void) generation_index_t gen = PSEUDO_STATIC_GENERATION; do { lispobj *first,*ptr= (lispobj *)page_address(page); - page_table[page].allocated = BOXED_PAGE_FLAG; - page_table[page].gen = gen; - page_table[page].bytes_used = PAGE_BYTES; - page_table[page].large_object = 0; - page_table[page].write_protected = 0; - page_table[page].write_protected_cleared = 0; - page_table[page].dont_move = 0; - page_table[page].need_to_zero = 1; + + if (!gencgc_partial_pickup || page_allocated_p(page)) { + /* It is possible, though rare, for the saved page table + * to contain free pages below alloc_ptr. */ + page_table[page].gen = gen; + page_table[page].bytes_used = GENCGC_CARD_BYTES; + page_table[page].large_object = 0; + page_table[page].write_protected = 0; + page_table[page].write_protected_cleared = 0; + page_table[page].dont_move = 0; + page_table[page].need_to_zero = 1; + } if (!gencgc_partial_pickup) { + page_table[page].allocated = BOXED_PAGE_FLAG; first=gc_search_space(prev,(ptr+2)-prev,ptr); - if(ptr == first) prev=ptr; + if(ptr == first) + prev=ptr; page_table[page].region_start_offset = page_address(page) - (void *)prev; } @@ -4726,10 +4780,14 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg if (SymbolValue(GC_INHIBIT,thread) == NIL) { set_pseudo_atomic_interrupted(thread); #ifdef LISP_FEATURE_PPC - /* PPC calls alloc() from a trap, look up the most - * recent one and frob that. */ - maybe_save_gc_mask_and_block_deferrables - (get_interrupt_context_for_thread(thread)); + /* PPC calls alloc() from a trap or from pa_alloc(), + * look up the most context if it's from a trap. */ + { + os_context_t *context = + thread->interrupt_data->allocation_trap_context; + maybe_save_gc_mask_and_block_deferrables + (context ? os_context_sigmask_addr(context) : NULL); + } #else maybe_save_gc_mask_and_block_deferrables(NULL); #endif @@ -4808,7 +4866,7 @@ gencgc_handle_wp_violation(void* fault_addr) { page_index_t page_index = find_page_index(fault_addr); -#ifdef QSHOW_SIGNALS +#if QSHOW_SIGNALS FSHOW((stderr, "heap WP violation? fault_addr=%x, page_index=%d\n", fault_addr, page_index)); #endif @@ -4824,9 +4882,12 @@ gencgc_handle_wp_violation(void* fault_addr) return 0; } else { + int ret; + ret = thread_mutex_lock(&free_pages_lock); + gc_assert(ret == 0); if (page_table[page_index].write_protected) { /* Unprotect the page. */ - os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL); + os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL); page_table[page_index].write_protected_cleared = 1; page_table[page_index].write_protected = 0; } else { @@ -4841,6 +4902,8 @@ gencgc_handle_wp_violation(void* fault_addr) page_index, boxed_region.first_page, boxed_region.last_page); } + ret = thread_mutex_unlock(&free_pages_lock); + gc_assert(ret == 0); /* Don't worry, we can handle it. */ return 1; } @@ -4882,7 +4945,7 @@ zero_all_free_pages() if (page_free_p(i)) { #ifdef READ_PROTECT_FREE_PAGES os_protect(page_address(i), - PAGE_BYTES, + GENCGC_CARD_BYTES, OS_VM_PROT_ALL); #endif zero_pages(i, i);