X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fgencgc.c;h=910b913b4c1d93f44a55447e1083bf2d88f8070f;hb=ab9c6bbaaa409e815a1c9696885c9621b429aed6;hp=c9d529cb2ac9c611b2e45d5c7d4e85a3b37d5304;hpb=b3f99a6af5bfdaa3090ec244bd3348e279ebbbaf;p=sbcl.git diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c index c9d529c..910b913 100644 --- a/src/runtime/gencgc.c +++ b/src/runtime/gencgc.c @@ -41,6 +41,7 @@ #include "gc.h" #include "gc-internal.h" #include "thread.h" +#include "pseudo-atomic.h" #include "alloc.h" #include "genesis/vector.h" #include "genesis/weak-pointer.h" @@ -57,7 +58,7 @@ /* forward declarations */ page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes, - int unboxed); + int page_type_flag); /* @@ -68,9 +69,7 @@ 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 }; @@ -88,7 +87,7 @@ 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; @@ -165,6 +164,48 @@ static boolean conservative_stack = 1; page_index_t page_table_pages; struct page *page_table; +static inline boolean page_allocated_p(page_index_t page) { + return (page_table[page].allocated != FREE_PAGE_FLAG); +} + +static inline boolean page_no_region_p(page_index_t page) { + return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG); +} + +static inline boolean page_allocated_no_region_p(page_index_t page) { + return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG)) + && page_no_region_p(page)); +} + +static inline boolean page_free_p(page_index_t page) { + return (page_table[page].allocated == FREE_PAGE_FLAG); +} + +static inline boolean page_boxed_p(page_index_t page) { + return (page_table[page].allocated & BOXED_PAGE_FLAG); +} + +static inline boolean code_page_p(page_index_t page) { + return (page_table[page].allocated & CODE_PAGE_FLAG); +} + +static inline boolean page_boxed_no_region_p(page_index_t page) { + return page_boxed_p(page) && page_no_region_p(page); +} + +static inline boolean page_unboxed_p(page_index_t page) { + /* Both flags set == boxed code page */ + return ((page_table[page].allocated & UNBOXED_PAGE_FLAG) + && !page_boxed_p(page)); +} + +static inline boolean protect_page_p(page_index_t page, generation_index_t generation) { + return (page_boxed_no_region_p(page) + && (page_table[page].bytes_used != 0) + && !page_table[page].dont_move + && (page_table[page].gen == generation)); +} + /* To map addresses to page structures the address of the first page * is needed. */ static void *heap_base = NULL; @@ -214,7 +255,12 @@ 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 */ @@ -244,9 +290,9 @@ 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 @@ -258,7 +304,7 @@ 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. */ @@ -294,15 +340,16 @@ generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION; * integrated with the Lisp code. */ page_index_t last_free_page; +#ifdef LISP_FEATURE_SB_THREAD /* This lock is to prevent multiple threads from simultaneously * allocating new regions which overlap each other. Note that the * majority of GC is single-threaded, but alloc() may be called from * >1 thread at a time and must be thread-safe. This lock must be * seized before all accesses to generations[] or to parts of * page_table[] that other threads may want to see */ - -#ifdef LISP_FEATURE_SB_THREAD static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER; +/* This lock is used to protect non-thread-local allocation. */ +static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER; #endif @@ -319,7 +366,7 @@ count_write_protect_generation_pages(generation_index_t generation) unsigned long count = 0; for (i = 0; i < last_free_page; i++) - if ((page_table[i].allocated != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].gen == generation) && (page_table[i].write_protected == 1)) count++; @@ -334,20 +381,20 @@ count_generation_pages(generation_index_t generation) long count = 0; for (i = 0; i < last_free_page; i++) - if ((page_table[i].allocated != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].gen == generation)) count++; return count; } -#ifdef QSHOW +#if QSHOW static long count_dont_move_pages(void) { page_index_t i; long count = 0; for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].dont_move != 0)) { ++count; } @@ -364,7 +411,7 @@ count_generation_bytes_allocated (generation_index_t gen) page_index_t i; unsigned long result = 0; for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].gen == gen)) result += page_table[i].bytes_used; } @@ -372,8 +419,8 @@ 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; @@ -385,10 +432,10 @@ gen_av_mem_age(generation_index_t gen) /* 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 +print_generation_stats() /* FIXME: should take FILE argument, or construct a string */ { - generation_index_t i, gens; + generation_index_t i; #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) #define FPU_STATE_SIZE 27 @@ -402,17 +449,11 @@ 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, " 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; @@ -425,7 +466,7 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ /* Count the number of boxed pages within the given * generation. */ - if (page_table[j].allocated & BOXED_PAGE_FLAG) { + if (page_boxed_p(j)) { if (page_table[j].large_object) large_boxed_cnt++; else @@ -434,7 +475,7 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ if(page_table[j].dont_move) pinned_cnt++; /* Count the number of unboxed pages within the given * generation. */ - if (page_table[j].allocated & UNBOXED_PAGE_FLAG) { + if (page_unboxed_p(j)) { if (page_table[j].large_object) large_unboxed_cnt++; else @@ -462,9 +503,10 @@ 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=%ld\n", bytes_allocated); + fprintf(stderr," Total bytes allocated = %lu\n", bytes_allocated); + fprintf(stderr," Dynamic-space-size bytes = %u\n", dynamic_space_size); fpu_restore(fpu_state); } @@ -589,6 +631,55 @@ struct alloc_region unboxed_region; /* The generation currently being allocated to. */ static generation_index_t gc_alloc_generation; +static inline page_index_t +generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large) +{ + if (large) { + if (UNBOXED_PAGE_FLAG == page_type_flag) { + return generations[generation].alloc_large_unboxed_start_page; + } else if (BOXED_PAGE_FLAG & page_type_flag) { + /* Both code and data. */ + return generations[generation].alloc_large_start_page; + } else { + lose("bad page type flag: %d", page_type_flag); + } + } else { + if (UNBOXED_PAGE_FLAG == page_type_flag) { + return generations[generation].alloc_unboxed_start_page; + } else if (BOXED_PAGE_FLAG & page_type_flag) { + /* Both code and data. */ + return generations[generation].alloc_start_page; + } else { + lose("bad page_type_flag: %d", page_type_flag); + } + } +} + +static inline void +set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large, + page_index_t page) +{ + if (large) { + if (UNBOXED_PAGE_FLAG == page_type_flag) { + generations[generation].alloc_large_unboxed_start_page = page; + } else if (BOXED_PAGE_FLAG & page_type_flag) { + /* Both code and data. */ + generations[generation].alloc_large_start_page = page; + } else { + lose("bad page type flag: %d", page_type_flag); + } + } else { + if (UNBOXED_PAGE_FLAG == page_type_flag) { + generations[generation].alloc_unboxed_start_page = page; + } else if (BOXED_PAGE_FLAG & page_type_flag) { + /* Both code and data. */ + generations[generation].alloc_start_page = page; + } else { + lose("bad page type flag: %d", page_type_flag); + } + } +} + /* Find a new region with room for at least the given number of bytes. * * It starts looking at the current generation's alloc_start_page. So @@ -613,7 +704,7 @@ static generation_index_t gc_alloc_generation; * are allocated, although they will initially be empty. */ static void -gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) +gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_region) { page_index_t first_page; page_index_t last_page; @@ -633,14 +724,8 @@ gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) && (alloc_region->free_pointer == alloc_region->end_addr)); ret = thread_mutex_lock(&free_pages_lock); gc_assert(ret == 0); - if (unboxed) { - first_page = - generations[gc_alloc_generation].alloc_unboxed_start_page; - } else { - first_page = - generations[gc_alloc_generation].alloc_start_page; - } - last_page=gc_find_freeish_pages(&first_page,nbytes,unboxed); + 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) + npage_bytes(last_page-first_page); @@ -656,29 +741,20 @@ gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) /* The first page may have already been in use. */ if (page_table[first_page].bytes_used == 0) { - if (unboxed) - page_table[first_page].allocated = UNBOXED_PAGE_FLAG; - else - page_table[first_page].allocated = BOXED_PAGE_FLAG; + page_table[first_page].allocated = page_type_flag; page_table[first_page].gen = gc_alloc_generation; page_table[first_page].large_object = 0; page_table[first_page].region_start_offset = 0; } - if (unboxed) - gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG); - else - gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG); + gc_assert(page_table[first_page].allocated == page_type_flag); page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG; gc_assert(page_table[first_page].gen == gc_alloc_generation); gc_assert(page_table[first_page].large_object == 0); for (i = first_page+1; i <= last_page; i++) { - if (unboxed) - page_table[i].allocated = UNBOXED_PAGE_FLAG; - else - page_table[i].allocated = BOXED_PAGE_FLAG; + page_table[i].allocated = page_type_flag; page_table[i].gen = gc_alloc_generation; page_table[i].large_object = 0; /* This may not be necessary for unboxed regions (think it was @@ -829,7 +905,7 @@ add_new_area(page_index_t first_page, size_t offset, size_t size) * it is safe to try to re-update the page table of this reset * alloc_region. */ void -gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) +gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region) { int more; page_index_t first_page; @@ -869,10 +945,7 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) gc_assert(page_table[first_page].region_start_offset == 0); page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG); - if (unboxed) - gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG); - else - gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG); + gc_assert(page_table[first_page].allocated & page_type_flag); gc_assert(page_table[first_page].gen == gc_alloc_generation); gc_assert(page_table[first_page].large_object == 0); @@ -896,10 +969,7 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) * region, and set the bytes_used. */ while (more) { page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG); - if (unboxed) - gc_assert(page_table[next_page].allocated==UNBOXED_PAGE_FLAG); - else - gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG); + gc_assert(page_table[next_page].allocated & page_type_flag); gc_assert(page_table[next_page].bytes_used == 0); gc_assert(page_table[next_page].gen == gc_alloc_generation); gc_assert(page_table[next_page].large_object == 0); @@ -930,14 +1000,10 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) /* Set the generations alloc restart page to the last page of * the region. */ - if (unboxed) - generations[gc_alloc_generation].alloc_unboxed_start_page = - next_page-1; - else - generations[gc_alloc_generation].alloc_start_page = next_page-1; + set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1); /* Add the region to the new_areas if requested. */ - if (!unboxed) + if (BOXED_PAGE_FLAG & page_type_flag) add_new_area(first_page,orig_first_page_bytes_used, region_size); /* @@ -971,38 +1037,30 @@ static inline void *gc_quick_alloc(long nbytes); /* Allocate a possibly large object. */ void * -gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) +gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region) { page_index_t first_page; page_index_t last_page; int orig_first_page_bytes_used; long byte_cnt; int more; - long bytes_used; + unsigned long bytes_used; page_index_t next_page; int ret; ret = thread_mutex_lock(&free_pages_lock); gc_assert(ret == 0); - if (unboxed) { - first_page = - generations[gc_alloc_generation].alloc_large_unboxed_start_page; - } else { - first_page = generations[gc_alloc_generation].alloc_large_start_page; - } + first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1); if (first_page <= alloc_region->last_page) { first_page = alloc_region->last_page+1; } - last_page=gc_find_freeish_pages(&first_page,nbytes,unboxed); + last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag); gc_assert(first_page > alloc_region->last_page); - if (unboxed) - generations[gc_alloc_generation].alloc_large_unboxed_start_page = - last_page; - else - generations[gc_alloc_generation].alloc_large_start_page = last_page; + + set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page); /* Set up the pages. */ orig_first_page_bytes_used = page_table[first_page].bytes_used; @@ -1010,19 +1068,13 @@ gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) /* If the first page was free then set up the gen, and * region_start_offset. */ if (page_table[first_page].bytes_used == 0) { - if (unboxed) - page_table[first_page].allocated = UNBOXED_PAGE_FLAG; - else - page_table[first_page].allocated = BOXED_PAGE_FLAG; + page_table[first_page].allocated = page_type_flag; page_table[first_page].gen = gc_alloc_generation; page_table[first_page].region_start_offset = 0; page_table[first_page].large_object = 1; } - if (unboxed) - gc_assert(page_table[first_page].allocated == UNBOXED_PAGE_FLAG); - else - gc_assert(page_table[first_page].allocated == BOXED_PAGE_FLAG); + gc_assert(page_table[first_page].allocated == page_type_flag); gc_assert(page_table[first_page].gen == gc_alloc_generation); gc_assert(page_table[first_page].large_object == 1); @@ -1044,12 +1096,9 @@ gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) * region_start_offset pointer to the start of the region, and set * the bytes_used. */ while (more) { - gc_assert(page_table[next_page].allocated == FREE_PAGE_FLAG); + gc_assert(page_free_p(next_page)); gc_assert(page_table[next_page].bytes_used == 0); - if (unboxed) - page_table[next_page].allocated = UNBOXED_PAGE_FLAG; - else - page_table[next_page].allocated = BOXED_PAGE_FLAG; + page_table[next_page].allocated = page_type_flag; page_table[next_page].gen = gc_alloc_generation; page_table[next_page].large_object = 1; @@ -1076,7 +1125,7 @@ gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) generations[gc_alloc_generation].bytes_allocated += nbytes; /* Add the region to the new_areas if requested. */ - if (!unboxed) + if (BOXED_PAGE_FLAG & page_type_flag) add_new_area(first_page,orig_first_page_bytes_used,nbytes); /* Bump up last_free_page */ @@ -1103,6 +1152,7 @@ static page_index_t gencgc_alloc_start_page = -1; void gc_heap_exhausted_error_or_lose (long available, long requested) { + struct thread *thread = arch_os_get_current_thread(); /* Write basic information before doing anything else: if we don't * call to lisp this is a must, and even if we do there is always * the danger that we bounce back here before the error has been @@ -1111,25 +1161,38 @@ gc_heap_exhausted_error_or_lose (long available, long requested) fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n", gc_active_p ? "garbage collection" : "allocation", available, requested); - 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. - */ - struct thread *thread = arch_os_get_current_thread(); - print_generation_stats(1); + print_generation_stats(); fprintf(stderr, "GC control variables:\n"); - fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n", + fprintf(stderr, " *GC-INHIBIT* = %s\n *GC-PENDING* = %s\n", SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true", - SymbolValue(GC_PENDING,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", + fprintf(stderr, " *STOP-FOR-GC-PENDING* = %s\n", SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true"); #endif + 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. + */ lose("Heap exhausted, game over."); } else { /* FIXME: assert free_pages_lock held */ (void)thread_mutex_unlock(&free_pages_lock); + gc_assert(get_pseudo_atomic_atomic(thread)); + clear_pseudo_atomic_atomic(thread); + if (get_pseudo_atomic_interrupted(thread)) + do_pending_interrupt(); + /* Another issue is that signalling HEAP-EXHAUSTED error leads + * to running user code at arbitrary places, even in a + * WITHOUT-INTERRUPTS which may lead to a deadlock without + * running out of the heap. So at this point all bets are + * off. */ + if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) + corruption_warning_and_maybe_lose + ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS."); funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR), alloc_number(available), alloc_number(requested)); lose("HEAP-EXHAUSTED-ERROR fell through"); @@ -1137,7 +1200,8 @@ gc_heap_exhausted_error_or_lose (long available, long requested) } page_index_t -gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed) +gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, + int page_type_flag) { page_index_t first_page, last_page; page_index_t restart_page = *restart_page_ptr; @@ -1150,7 +1214,8 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed) restart_page = gencgc_alloc_start_page; } - if (nbytes>=PAGE_BYTES) { + 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 @@ -1158,17 +1223,18 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed) do { first_page = restart_page; while ((first_page < page_table_pages) && - (page_table[first_page].allocated != FREE_PAGE_FLAG)) + page_allocated_p(first_page)) first_page++; last_page = first_page; bytes_found = PAGE_BYTES; while ((bytes_found < nbytes) && (last_page < (page_table_pages-1)) && - (page_table[last_page+1].allocated == FREE_PAGE_FLAG)) { + page_free_p(last_page+1)) { last_page++; bytes_found += PAGE_BYTES; - gc_assert(page_table[last_page].write_protected == 0); + 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; @@ -1182,13 +1248,13 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed) * pages: this helps avoid excessive conservativism. */ first_page = restart_page; while (first_page < page_table_pages) { - if (page_table[first_page].allocated == FREE_PAGE_FLAG) + 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 == - (unboxed ? UNBOXED_PAGE_FLAG : BOXED_PAGE_FLAG)) && + 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) && @@ -1223,13 +1289,13 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed) * functions will eventually call this */ void * -gc_alloc_with_region(long nbytes,int unboxed_p, struct alloc_region *my_region, +gc_alloc_with_region(long nbytes,int page_type_flag, struct alloc_region *my_region, int quick_p) { void *new_free_pointer; if (nbytes>=large_object_size) - return gc_alloc_large(nbytes,unboxed_p,my_region); + return gc_alloc_large(nbytes, page_type_flag, my_region); /* Check whether there is room in the current alloc region. */ new_free_pointer = my_region->free_pointer + nbytes; @@ -1247,9 +1313,9 @@ gc_alloc_with_region(long nbytes,int unboxed_p, struct alloc_region *my_region, if (!quick_p && void_diff(my_region->end_addr,my_region->free_pointer) <= 32) { /* If so, finished with the current region. */ - gc_alloc_update_page_tables(unboxed_p, my_region); + gc_alloc_update_page_tables(page_type_flag, my_region); /* Set up a new region. */ - gc_alloc_new_region(32 /*bytes*/, unboxed_p, my_region); + gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region); } return((void *)new_obj); @@ -1258,51 +1324,43 @@ gc_alloc_with_region(long nbytes,int unboxed_p, struct alloc_region *my_region, /* Else not enough free space in the current region: retry with a * new region. */ - gc_alloc_update_page_tables(unboxed_p, my_region); - gc_alloc_new_region(nbytes, unboxed_p, my_region); - return gc_alloc_with_region(nbytes,unboxed_p,my_region,0); + gc_alloc_update_page_tables(page_type_flag, my_region); + gc_alloc_new_region(nbytes, page_type_flag, my_region); + return gc_alloc_with_region(nbytes, page_type_flag, my_region,0); } /* these are only used during GC: all allocation from the mutator calls * alloc() -> gc_alloc_with_region() with the appropriate per-thread * region */ -void * -gc_general_alloc(long nbytes,int unboxed_p,int quick_p) -{ - struct alloc_region *my_region = - unboxed_p ? &unboxed_region : &boxed_region; - return gc_alloc_with_region(nbytes,unboxed_p, my_region,quick_p); -} - static inline void * gc_quick_alloc(long nbytes) { - return gc_general_alloc(nbytes,ALLOC_BOXED,ALLOC_QUICK); + return gc_general_alloc(nbytes, BOXED_PAGE_FLAG, ALLOC_QUICK); } static inline void * gc_quick_alloc_large(long nbytes) { - return gc_general_alloc(nbytes,ALLOC_BOXED,ALLOC_QUICK); + return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK); } static inline void * gc_alloc_unboxed(long nbytes) { - return gc_general_alloc(nbytes,ALLOC_UNBOXED,0); + return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0); } static inline void * gc_quick_alloc_unboxed(long nbytes) { - return gc_general_alloc(nbytes,ALLOC_UNBOXED,ALLOC_QUICK); + return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK); } static inline void * gc_quick_alloc_large_unboxed(long nbytes) { - return gc_general_alloc(nbytes,ALLOC_UNBOXED,ALLOC_QUICK); + return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, ALLOC_QUICK); } @@ -1349,20 +1407,16 @@ copy_large_object(lispobj object, long nwords) remaining_bytes = nwords*N_WORD_BYTES; while (remaining_bytes > PAGE_BYTES) { gc_assert(page_table[next_page].gen == from_space); - gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG); + 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); + /* 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; next_page++; } @@ -1374,7 +1428,7 @@ copy_large_object(lispobj object, long nwords) gc_assert(page_table[next_page].bytes_used >= remaining_bytes); page_table[next_page].gen = new_space; - gc_assert(page_table[next_page].allocated == BOXED_PAGE_FLAG); + gc_assert(page_boxed_p(next_page)); /* Adjust the bytes_used. */ old_bytes_used = page_table[next_page].bytes_used; @@ -1386,7 +1440,7 @@ copy_large_object(lispobj object, long nwords) next_page++; while ((old_bytes_used == PAGE_BYTES) && (page_table[next_page].gen == from_space) && - (page_table[next_page].allocated == BOXED_PAGE_FLAG) && + page_boxed_p(next_page) && page_table[next_page].large_object && (page_table[next_page].region_start_offset == npage_bytes(next_page - first_page))) { @@ -1471,9 +1525,10 @@ copy_large_unboxed_object(lispobj object, long nwords) gc_assert(from_space_p(object)); gc_assert((nwords & 0x01) == 0); - if ((nwords > 1024*1024) && gencgc_verbose) + if ((nwords > 1024*1024) && gencgc_verbose) { FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n", nwords*N_WORD_BYTES)); + } /* Check whether it's a large object. */ first_page = find_page_index((void *)object); @@ -1494,8 +1549,7 @@ copy_large_unboxed_object(lispobj object, long nwords) remaining_bytes = nwords*N_WORD_BYTES; while (remaining_bytes > PAGE_BYTES) { gc_assert(page_table[next_page].gen == from_space); - gc_assert((page_table[next_page].allocated == UNBOXED_PAGE_FLAG) - || (page_table[next_page].allocated == BOXED_PAGE_FLAG)); + 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)); @@ -1526,8 +1580,7 @@ copy_large_unboxed_object(lispobj object, long nwords) next_page++; while ((old_bytes_used == PAGE_BYTES) && (page_table[next_page].gen == from_space) && - ((page_table[next_page].allocated == UNBOXED_PAGE_FLAG) - || (page_table[next_page].allocated == BOXED_PAGE_FLAG)) && + page_allocated_no_region_p(next_page) && page_table[next_page].large_object && (page_table[next_page].region_start_offset == npage_bytes(next_page - first_page))) { @@ -1544,10 +1597,11 @@ copy_large_unboxed_object(lispobj object, long nwords) next_page++; } - if ((bytes_freed > 0) && gencgc_verbose) + if ((bytes_freed > 0) && gencgc_verbose) { FSHOW((stderr, "/copy_large_unboxed bytes_freed=%d\n", bytes_freed)); + } generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES + bytes_freed; @@ -1623,7 +1677,7 @@ 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 @@ -2152,8 +2206,7 @@ search_dynamic_space(void *pointer) lispobj *start; /* The address may be invalid, so do some checks. */ - if ((page_index == -1) || - (page_table[page_index].allocated == FREE_PAGE_FLAG)) + if ((page_index == -1) || page_free_p(page_index)) return NULL; start = (lispobj *)page_region_start(page_index); return (gc_search_space(start, @@ -2161,8 +2214,6 @@ 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. * @@ -2172,13 +2223,6 @@ search_dynamic_space(void *pointer) static int looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) { - /* We need to allow raw pointers into Code objects for return - * addresses. This will also pick up pointers to functions in code - * objects. */ - if (widetag_of(*start_addr) == CODE_HEADER_WIDETAG) - /* XXX could do some further checks here */ - return 1; - if (!is_lisp_pointer((lispobj)pointer)) { return 0; } @@ -2197,28 +2241,31 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) case FUNCALLABLE_INSTANCE_HEADER_WIDETAG: if ((unsigned long)pointer != ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wf2: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } break; default: - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wf3: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } break; case LIST_POINTER_LOWTAG: if ((unsigned long)pointer != ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wl1: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } /* Is it plausible cons? */ @@ -2228,44 +2275,66 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) is_lisp_immediate(start_addr[1]))) break; else { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wl2: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } case INSTANCE_POINTER_LOWTAG: if ((unsigned long)pointer != ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wi1: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } if (widetag_of(start_addr[0]) != INSTANCE_HEADER_WIDETAG) { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wi2: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } 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) + if (gencgc_verbose) { FSHOW((stderr, "/Wo1: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } /* Is it plausible? Not a cons. XXX should check the headers. */ if (is_lisp_pointer(start_addr[0]) || ((start_addr[0] & 3) == 0)) { - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wo2: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } switch (widetag_of(start_addr[0])) { @@ -2275,26 +2344,29 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) #if N_WORD_BITS == 64 case SINGLE_FLOAT_WIDETAG: #endif - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "*Wo3: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; /* only pointed to by function pointers? */ case CLOSURE_HEADER_WIDETAG: case FUNCALLABLE_INSTANCE_HEADER_WIDETAG: - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "*Wo4: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; case INSTANCE_HEADER_WIDETAG: - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "*Wo5: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; /* the valid other immediate pointer objects */ @@ -2397,18 +2469,20 @@ looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) break; default: - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "/Wo6: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } break; default: - if (gencgc_verbose) + if (gencgc_verbose) { FSHOW((stderr, "*W?: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } @@ -2440,6 +2514,8 @@ 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? @@ -2565,8 +2641,7 @@ maybe_adjust_large_object(lispobj *where) remaining_bytes = nwords*N_WORD_BYTES; while (remaining_bytes > PAGE_BYTES) { gc_assert(page_table[next_page].gen == from_space); - gc_assert((page_table[next_page].allocated == BOXED_PAGE_FLAG) - || (page_table[next_page].allocated == UNBOXED_PAGE_FLAG)); + 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)); @@ -2601,8 +2676,7 @@ maybe_adjust_large_object(lispobj *where) next_page++; while ((old_bytes_used == PAGE_BYTES) && (page_table[next_page].gen == from_space) && - ((page_table[next_page].allocated == UNBOXED_PAGE_FLAG) - || (page_table[next_page].allocated == BOXED_PAGE_FLAG)) && + page_allocated_no_region_p(next_page) && page_table[next_page].large_object && (page_table[next_page].region_start_offset == npage_bytes(next_page - first_page))) { @@ -2654,7 +2728,7 @@ preserve_pointer(void *addr) /* quick check 1: Address is quite likely to have been invalid. */ if ((addr_page_index == -1) - || (page_table[addr_page_index].allocated == FREE_PAGE_FLAG) + || page_free_p(addr_page_index) || (page_table[addr_page_index].bytes_used == 0) || (page_table[addr_page_index].gen != from_space) /* Skip if already marked dont_move. */ @@ -2678,7 +2752,9 @@ preserve_pointer(void *addr) * 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. */ - if (!(possibly_valid_dynamic_space_pointer(addr))) + if (!(code_page_p(addr_page_index) + || (is_lisp_pointer((lispobj)addr) && + possibly_valid_dynamic_space_pointer(addr)))) return; /* Find the beginning of the region. Note that there may be @@ -2709,7 +2785,7 @@ preserve_pointer(void *addr) * free area in which case it's ignored here. Note it gets * through the valid pointer test above because the tail looks * like conses. */ - if ((page_table[addr_page_index].allocated == FREE_PAGE_FLAG) + 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)) @@ -2748,7 +2824,7 @@ preserve_pointer(void *addr) /* 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 */ - || (page_table[i+1].allocated == FREE_PAGE_FLAG) + || page_free_p(i+1) || (page_table[i+1].bytes_used == 0) /* next page free */ || (page_table[i+1].gen != from_space) /* diff. gen */ || (page_table[i+1].region_start_offset == 0)) @@ -2785,14 +2861,14 @@ update_page_write_prot(page_index_t page) long num_words = page_table[page].bytes_used / N_WORD_BYTES; /* Shouldn't be a free page. */ - gc_assert(page_table[page].allocated != FREE_PAGE_FLAG); + gc_assert(page_allocated_p(page)); gc_assert(page_table[page].bytes_used != 0); /* Skip if it's already write-protected, pinned, or unboxed */ if (page_table[page].write_protected /* FIXME: What's the reason for not write-protecting pinned pages? */ || page_table[page].dont_move - || (page_table[page].allocated & UNBOXED_PAGE_FLAG)) + || page_unboxed_p(page)) return (0); /* Scan the page for pointers to younger generations or the @@ -2805,7 +2881,7 @@ update_page_write_prot(page_index_t page) /* Check that it's in the dynamic space */ if (index != -1) if (/* Does it point to a younger or the temp. generation? */ - ((page_table[index].allocated != FREE_PAGE_FLAG) + (page_allocated_p(index) && (page_table[index].bytes_used != 0) && ((page_table[index].gen < gen) || (page_table[index].gen == SCRATCH_GENERATION))) @@ -2879,7 +2955,7 @@ scavenge_generations(generation_index_t from, generation_index_t to) for (i = 0; i < last_free_page; i++) { generation_index_t generation = page_table[i].gen; - if ((page_table[i].allocated & BOXED_PAGE_FLAG) + if (page_boxed_p(i) && (page_table[i].bytes_used != 0) && (generation != new_space) && (generation >= from) @@ -2896,7 +2972,7 @@ scavenge_generations(generation_index_t from, generation_index_t to) 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 */ - || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG)) + || (!page_boxed_p(last_page+1)) || (page_table[last_page+1].bytes_used == 0) || (page_table[last_page+1].gen != generation) || (page_table[last_page+1].region_start_offset == 0)) @@ -2929,7 +3005,7 @@ scavenge_generations(generation_index_t from, generation_index_t to) /* Check that none of the write_protected pages in this generation * have been written to. */ for (i = 0; i < page_table_pages; i++) { - if ((page_table[i].allocation != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation) && (page_table[i].write_protected_cleared != 0)) { @@ -2983,7 +3059,7 @@ scavenge_newspace_generation_one_scan(generation_index_t generation) generation)); for (i = 0; i < last_free_page; i++) { /* Note that this skips over open regions when it encounters them. */ - if ((page_table[i].allocated & BOXED_PAGE_FLAG) + if (page_boxed_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation) && ((page_table[i].write_protected == 0) @@ -3012,7 +3088,7 @@ scavenge_newspace_generation_one_scan(generation_index_t generation) * contiguous block */ if ((page_table[last_page].bytes_used < PAGE_BYTES) /* Or it is PAGE_BYTES and is the last in the block */ - || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG)) + || (!page_boxed_p(last_page+1)) || (page_table[last_page+1].bytes_used == 0) || (page_table[last_page+1].gen != generation) || (page_table[last_page+1].region_start_offset == 0)) @@ -3112,8 +3188,9 @@ scavenge_newspace_generation(generation_index_t generation) /* New areas of objects allocated have been lost so need to do a * full scan to be sure! If this becomes a problem try * increasing NUM_NEW_AREAS. */ - if (gencgc_verbose) + if (gencgc_verbose) { SHOW("new_areas overflow, doing full scavenge"); + } /* Don't need to record new areas that get scavenged * anyway during scavenge_newspace_generation_one_scan. */ @@ -3160,7 +3237,7 @@ scavenge_newspace_generation(generation_index_t generation) /* Check that none of the write_protected pages in this generation * have been written to. */ for (i = 0; i < page_table_pages; i++) { - if ((page_table[i].allocation != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation) && (page_table[i].write_protected_cleared != 0) @@ -3181,23 +3258,41 @@ 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_table[i].allocated != FREE_PAGE_FLAG) + 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 = PAGE_BYTES; + } else if (region_addr + region_bytes == page_addr) { + /* Region continue. */ + region_bytes += PAGE_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 = PAGE_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 @@ -3215,7 +3310,7 @@ free_oldspace(void) do { /* Find a first page for the next region of pages. */ while ((first_page < last_free_page) - && ((page_table[first_page].allocated == FREE_PAGE_FLAG) + && (page_free_p(first_page) || (page_table[first_page].bytes_used == 0) || (page_table[first_page].gen != from_space))) first_page++; @@ -3233,21 +3328,12 @@ 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) - && (page_table[last_page].allocated != FREE_PAGE_FLAG) + && page_allocated_p(last_page) && (page_table[last_page].bytes_used != 0) && (page_table[last_page].gen == from_space)); @@ -3293,6 +3379,23 @@ 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) { @@ -3318,17 +3421,17 @@ verify_space(lispobj *start, size_t words) if (page_index != -1) { /* If it's within the dynamic space it should point to a used * page. XX Could check the offset too. */ - if ((page_table[page_index].allocated != FREE_PAGE_FLAG) + 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 @@ -3342,13 +3445,16 @@ 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 { @@ -3425,7 +3531,7 @@ 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)); } @@ -3541,7 +3647,7 @@ 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); } } @@ -3583,7 +3689,7 @@ verify_generation(generation_index_t generation) page_index_t i; for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated != FREE_PAGE_FLAG) + if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation)) { page_index_t last_page; @@ -3625,7 +3731,7 @@ verify_zero_fill(void) page_index_t page; for (page = 0; page < last_free_page; page++) { - if (page_table[page].allocated == FREE_PAGE_FLAG) { + if (page_free_p(page)) { /* The whole page should be zero filled. */ long *start_addr = (long *)page_address(page); long size = 1024; @@ -3683,10 +3789,7 @@ write_protect_generation_pages(generation_index_t generation) gc_assert(generation < SCRATCH_GENERATION); for (start = 0; start < last_free_page; start++) { - if ((page_table[start].allocated == BOXED_PAGE_FLAG) - && (page_table[start].bytes_used != 0) - && !page_table[start].dont_move - && (page_table[start].gen == generation)) { + if (protect_page_p(start, generation)) { void *page_start; page_index_t last; @@ -3694,10 +3797,7 @@ write_protect_generation_pages(generation_index_t generation) page_table[start].write_protected = 1; for (last = start + 1; last < last_free_page; last++) { - if ((page_table[last].allocated != BOXED_PAGE_FLAG) - || (page_table[last].bytes_used == 0) - || page_table[last].dont_move - || (page_table[last].gen != generation)) + if (!protect_page_p(last, generation)) break; page_table[last].write_protected = 1; } @@ -3722,165 +3822,22 @@ 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); 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) { @@ -4035,13 +3992,13 @@ garbage_collect_generation(generation_index_t generation, int raise) } #endif -#ifdef QSHOW +#if QSHOW if (gencgc_verbose > 1) { long num_dont_move_pages = count_dont_move_pages(); fprintf(stderr, "/non-movable pages due to conservative pointers = %d (%d bytes)\n", num_dont_move_pages, - npage_bytes(num_dont_move_pages); + npage_bytes(num_dont_move_pages)); } #endif @@ -4052,8 +4009,18 @@ 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 @@ -4178,8 +4145,9 @@ garbage_collect_generation(generation_index_t generation, int raise) generations[generation].alloc_large_unboxed_start_page = 0; if (generation >= verify_gens) { - if (gencgc_verbose) + if (gencgc_verbose) { SHOW("verifying"); + } verify_gc(); verify_dynamic_space(); } @@ -4208,8 +4176,7 @@ update_dynamic_space_free_pointer(void) page_index_t last_page = -1, i; for (i = 0; i < last_free_page; i++) - if ((page_table[i].allocated != FREE_PAGE_FLAG) - && (page_table[i].bytes_used != 0)) + if (page_allocated_p(i) && (page_table[i].bytes_used != 0)) last_page = i; last_free_page = last_page+1; @@ -4224,15 +4191,15 @@ remap_free_pages (page_index_t from, page_index_t to) page_index_t first_page, last_page; for (first_page = from; first_page <= to; first_page++) { - if (page_table[first_page].allocated != FREE_PAGE_FLAG || - page_table[first_page].need_to_zero == 0) { + if (page_allocated_p(first_page) || + (page_table[first_page].need_to_zero == 0)) { continue; } last_page = first_page + 1; - while (page_table[last_page].allocated == FREE_PAGE_FLAG && - last_page < to && - page_table[last_page].need_to_zero == 1) { + while (page_free_p(last_page) && + (last_page < to) && + (page_table[last_page].need_to_zero == 1)) { last_page++; } @@ -4291,7 +4258,7 @@ collect_garbage(generation_index_t last_gen) } if (gencgc_verbose > 1) - print_generation_stats(0); + print_generation_stats(); do { /* Collect the generation. */ @@ -4302,7 +4269,7 @@ 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) { @@ -4329,7 +4296,7 @@ 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++; @@ -4339,8 +4306,8 @@ 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. @@ -4408,12 +4375,13 @@ gc_free_heap(void) { page_index_t page; - if (gencgc_verbose > 1) + if (gencgc_verbose > 1) { SHOW("entering gc_free_heap"); + } for (page = 0; page < page_table_pages; page++) { /* Skip free pages which should already be zero filled. */ - if (page_table[page].allocated != FREE_PAGE_FLAG) { + if (page_allocated_p(page)) { void *page_start, *addr; /* Mark the page free. The other slots are assumed invalid @@ -4447,7 +4415,7 @@ gc_free_heap(void) /* Double-check that the page is zero filled. */ long *page_start; page_index_t i; - gc_assert(page_table[page].allocated == FREE_PAGE_FLAG); + 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++) { @@ -4474,7 +4442,7 @@ gc_free_heap(void) } if (gencgc_verbose > 1) - print_generation_stats(0); + print_generation_stats(); /* Initialize gc_alloc(). */ gc_alloc_generation = 0; @@ -4543,8 +4511,8 @@ 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; } @@ -4568,21 +4536,26 @@ gencgc_pickup_dynamic(void) void *alloc_ptr = (void *)get_alloc_pointer(); lispobj *prev=(lispobj *)page_address(page); 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 = 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_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; } @@ -4610,8 +4583,6 @@ gc_initialize_pointers(void) { gencgc_pickup_dynamic(); } - - /* alloc(..) is the external interface for memory allocation. It @@ -4626,16 +4597,10 @@ gc_initialize_pointers(void) * The check for a GC trigger is only performed when the current * region is full, so in most cases it's not needed. */ -lispobj * -alloc(long nbytes) +static inline lispobj * +general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *region, + struct thread *thread) { - struct thread *thread=arch_os_get_current_thread(); - struct alloc_region *region= -#ifdef LISP_FEATURE_SB_THREAD - thread ? &(thread->alloc_region) : &boxed_region; -#else - &boxed_region; -#endif #ifndef LISP_FEATURE_WIN32 lispobj alloc_signal; #endif @@ -4648,25 +4613,8 @@ alloc(long nbytes) gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0) && ((nbytes & LOWTAG_MASK) == 0)); -#if 0 - if(all_threads) - /* there are a few places in the C code that allocate data in the - * heap before Lisp starts. This is before interrupts are enabled, - * so we don't need to check for pseudo-atomic */ -#ifdef LISP_FEATURE_SB_THREAD - if(!get_psuedo_atomic_atomic(th)) { - register u32 fs; - fprintf(stderr, "fatal error in thread 0x%x, tid=%ld\n", - th,th->os_thread); - __asm__("movl %fs,%0" : "=r" (fs) : ); - fprintf(stderr, "fs is %x, th->tls_cookie=%x \n", - debug_get_fs(),th->tls_cookie); - lose("If you see this message before 2004.01.31, mail details to sbcl-devel\n"); - } -#else - gc_assert(get_pseudo_atomic_atomic(th)); -#endif -#endif + /* Must be inside a PA section. */ + gc_assert(get_pseudo_atomic_atomic(thread)); /* maybe we can do this quickly ... */ new_free_pointer = region->free_pointer + nbytes; @@ -4676,11 +4624,10 @@ alloc(long nbytes) return(new_obj); /* yup */ } - /* we have to go the long way around, it seems. Check whether - * we should GC in the near future + /* we have to go the long way around, it seems. Check whether we + * should GC in the near future */ if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) { - gc_assert(get_pseudo_atomic_atomic(thread)); /* Don't flood the system with interrupts if the need to gc is * already noted. This can happen for example when SUB-GC * allocates or after a gc triggered in a WITHOUT-GCING. */ @@ -4688,22 +4635,31 @@ alloc(long nbytes) /* set things up so that GC happens when we finish the PA * section */ SetSymbolValue(GC_PENDING,T,thread); - if (SymbolValue(GC_INHIBIT,thread) == NIL) - set_pseudo_atomic_interrupted(thread); + if (SymbolValue(GC_INHIBIT,thread) == NIL) { + set_pseudo_atomic_interrupted(thread); +#ifdef LISP_FEATURE_PPC + /* 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 + } } } - new_obj = gc_alloc_with_region(nbytes,0,region,0); + new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0); #ifndef LISP_FEATURE_WIN32 alloc_signal = SymbolValue(ALLOC_SIGNAL,thread); if ((alloc_signal & FIXNUM_TAG_MASK) == 0) { if ((signed long) alloc_signal <= 0) { SetSymbolValue(ALLOC_SIGNAL, T, thread); -#ifdef LISP_FEATURE_SB_THREAD - kill_thread_safely(thread->os_thread, SIGPROF); -#else raise(SIGPROF); -#endif } else { SetSymbolValue(ALLOC_SIGNAL, alloc_signal - (1 << N_FIXNUM_TAG_BITS), @@ -4714,12 +4670,43 @@ alloc(long nbytes) return (new_obj); } + +lispobj * +general_alloc(long nbytes, int page_type_flag) +{ + struct thread *thread = arch_os_get_current_thread(); + /* Select correct region, and call general_alloc_internal with it. + * For other then boxed allocation we must lock first, since the + * region is shared. */ + if (BOXED_PAGE_FLAG & page_type_flag) { +#ifdef LISP_FEATURE_SB_THREAD + struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region); +#else + struct alloc_region *region = &boxed_region; +#endif + return general_alloc_internal(nbytes, page_type_flag, region, thread); + } else if (UNBOXED_PAGE_FLAG == page_type_flag) { + lispobj * obj; + gc_assert(0 == thread_mutex_lock(&allocation_lock)); + obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread); + gc_assert(0 == thread_mutex_unlock(&allocation_lock)); + return obj; + } else { + lose("bad page type flag: %d", page_type_flag); + } +} + +lispobj * +alloc(long nbytes) +{ + gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread())); + return general_alloc(nbytes, BOXED_PAGE_FLAG); +} /* * shared support for the OS-dependent signal handlers which * catch GENCGC-related write-protect violations */ - void unhandled_sigmemoryfault(void* addr); /* Depending on which OS we're running under, different signals might @@ -4737,7 +4724,7 @@ 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 @@ -4753,6 +4740,9 @@ 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); @@ -4770,6 +4760,8 @@ 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; } @@ -4787,9 +4779,9 @@ void gc_alloc_update_all_page_tables(void) /* Flush the alloc regions updating the tables. */ struct thread *th; for_each_thread(th) - gc_alloc_update_page_tables(0, &th->alloc_region); - gc_alloc_update_page_tables(1, &unboxed_region); - gc_alloc_update_page_tables(0, &boxed_region); + gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region); + gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region); + gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region); } void @@ -4808,7 +4800,7 @@ zero_all_free_pages() page_index_t i; for (i = 0; i < last_free_page; i++) { - if (page_table[i].allocated == FREE_PAGE_FLAG) { + if (page_free_p(i)) { #ifdef READ_PROTECT_FREE_PAGES os_protect(page_address(i), PAGE_BYTES, @@ -4848,7 +4840,8 @@ prepare_for_final_gc () * function being set to the value of the static symbol * SB!VM:RESTART-LISP-FUNCTION */ void -gc_and_save(char *filename, int prepend_runtime) +gc_and_save(char *filename, boolean prepend_runtime, + boolean save_runtime_options) { FILE *file; void *runtime_bytes = NULL; @@ -4883,7 +4876,7 @@ gc_and_save(char *filename, int prepend_runtime) /* The dumper doesn't know that pages need to be zeroed before use. */ zero_all_free_pages(); save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0), - prepend_runtime); + prepend_runtime, save_runtime_options); /* Oops. Save still managed to fail. Since we've mangled the stack * beyond hope, there's not much we can do. * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's