X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fgencgc.c;h=d5ea2335388f51f0869e3ee53526bd5bc38d98a4;hb=e8571be6d533b80768bdae4e3e15316e4faa22fa;hp=a5904dfde55773dcbdd238b9bc211d171902b25b;hpb=40e3ba03d0e1b824e4d1ae75d74246b975b70964;p=sbcl.git diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c index a5904df..d5ea233 100644 --- a/src/runtime/gencgc.c +++ b/src/runtime/gencgc.c @@ -1,5 +1,5 @@ /* - * GENerational Conservative Garbage Collector for SBCL x86 + * GENerational Conservative Garbage Collector for SBCL */ /* @@ -24,6 +24,7 @@ * . */ +#include #include #include #include @@ -37,73 +38,71 @@ #include "validate.h" #include "lispregs.h" #include "arch.h" -#include "fixnump.h" #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" +#include "genesis/fdefn.h" #include "genesis/simple-fun.h" +#include "save.h" #include "genesis/hash-table.h" +#include "genesis/instance.h" +#include "genesis/layout.h" +#include "gencgc.h" +#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 */ -long gc_find_freeish_pages(long *restart_page_ptr, long nbytes, int unboxed); -static void gencgc_pickup_dynamic(void); +page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes, + int page_type_flag); /* * GC parameters */ -/* the number of actual generations. (The number of 'struct - * generation' objects is one more than this, because one object - * serves as scratch when GC'ing.) */ -#define NUM_GENERATIONS 6 +/* Generations 0-5 are normal collected generations, 6 is only used as + * scratch space by the collector, and should never get collected. + */ +enum { + SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1, + NUM_GENERATIONS +}; /* Should we use page protection to help avoid the scavenging of pages * that don't have pointers to younger generations? */ boolean enable_page_protection = 1; -/* Should we unmap a page and re-mmap it to have it zero filled? */ -#if defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__NetBSD__) || defined(__sun) -/* comment from cmucl-2.4.8: This can waste a lot of swap on FreeBSD - * so don't unmap there. - * - * The CMU CL comment didn't specify a version, but was probably an - * old version of FreeBSD (pre-4.0), so this might no longer be true. - * OTOH, if it is true, this behavior might exist on OpenBSD too, so - * for now we don't unmap there either. -- WHN 2001-04-07 */ -/* Apparently this flag is required to be 0 for SunOS/x86, as there - * are reports of heap corruption otherwise. */ -boolean gencgc_unmap_zero = 0; -#else -boolean gencgc_unmap_zero = 1; -#endif - /* the minimum size (in bytes) for a large object*/ -unsigned large_object_size = 4 * PAGE_BYTES; +long large_object_size = 4 * GENCGC_ALLOC_GRANULARITY; /* * debugging */ - - /* 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 -unsigned gencgc_verbose = 1; +#if QSHOW +boolean gencgc_verbose = 1; #else -unsigned gencgc_verbose = 0; +boolean gencgc_verbose = 0; #endif /* FIXME: At some point enable the various error-checking things below * and see what they say. */ /* We hunt for pointers to old-space, when GCing generations >= verify_gen. - * Set verify_gens to NUM_GENERATIONS to disable this kind of check. */ -int verify_gens = NUM_GENERATIONS; + * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of + * check. */ +generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1; /* Should we do a pre-scan verify of generation 0 before it's GCed? */ boolean pre_verify_gen_0 = 0; @@ -128,6 +127,19 @@ boolean gencgc_enable_verify_zero_fill = 0; /* Should we check that free pages are zero filled during gc_free_heap * called after Lisp PURIFY? */ boolean gencgc_zero_check_during_free_heap = 0; + +/* When loading a core, don't do a full scan of the memory for the + * memory region boundaries. (Set to true by coreparse.c if the core + * contained a pagetable entry). + */ +boolean gencgc_partial_pickup = 0; + +/* If defined, free pages are read-protected to ensure that nothing + * accesses them. + */ + +/* #define READ_PROTECT_FREE_PAGES */ + /* * GC structures and variables @@ -135,113 +147,186 @@ boolean gencgc_zero_check_during_free_heap = 0; /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */ unsigned long bytes_allocated = 0; -extern unsigned long bytes_consed_between_gcs; /* gc-common.c */ unsigned long auto_gc_trigger = 0; /* the source and destination generations. These are set before a GC starts * scavenging. */ -long from_space; -long new_space; +generation_index_t from_space; +generation_index_t new_space; + +/* Set to 1 when in GC */ +boolean gc_active_p = 0; +/* should the GC be conservative on stack. If false (only right before + * saving a core), don't scan the stack / mark pages dont_move. */ +static boolean conservative_stack = 1; -/* An array of page structures is statically allocated. +/* An array of page structures is allocated on gc initialization. * This helps quickly map between an address its page structure. - * NUM_PAGES is set from the size of the dynamic space. */ -struct page page_table[NUM_PAGES]; + * page_table_pages is set from the size of the dynamic space. */ +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; -#if N_WORD_BITS == 32 - #define SIMPLE_ARRAY_WORD_WIDETAG SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG -#elif N_WORD_BITS == 64 - #define SIMPLE_ARRAY_WORD_WIDETAG SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG -#endif - /* Calculate the start address for the given page number. */ inline void * -page_address(long page_num) +page_address(page_index_t page_num) +{ + return (heap_base + (page_num * GENCGC_CARD_BYTES)); +} + +/* Calculate the address where the allocation region associated with + * the page starts. */ +static inline void * +page_region_start(page_index_t page_index) { - return (heap_base + (page_num * PAGE_BYTES)); + return page_address(page_index)-page_table[page_index].region_start_offset; } /* Find the page index within the page_table for the given * address. Return -1 on failure. */ -inline long +inline page_index_t find_page_index(void *addr) { - long index = addr-heap_base; - - if (index >= 0) { - index = ((unsigned long)index)/PAGE_BYTES; - if (index < NUM_PAGES) + if (addr >= heap_base) { + page_index_t index = ((pointer_sized_uint_t)addr - + (pointer_sized_uint_t)heap_base) / GENCGC_CARD_BYTES; + if (index < page_table_pages) return (index); } - return (-1); } -/* a structure to hold the state of a generation */ +static size_t +npage_bytes(long npages) +{ + gc_assert(npages>=0); + return ((unsigned long)npages)*GENCGC_CARD_BYTES; +} + +/* Check that X is a higher address than Y and return offset from Y to + * X in bytes. */ +static inline +size_t void_diff(void *x, void *y) +{ + gc_assert(x >= y); + return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y; +} + +/* 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 */ - long alloc_start_page; + page_index_t alloc_start_page; /* the first page that gc_alloc_unboxed() checks on its next call */ - long alloc_unboxed_start_page; + page_index_t alloc_unboxed_start_page; /* the first page that gc_alloc_large (boxed) considers on its next * call. (Although it always allocates after the boxed_region.) */ - long alloc_large_start_page; + page_index_t alloc_large_start_page; /* the first page that gc_alloc_large (unboxed) considers on its * next call. (Although it always allocates after the * current_unboxed_region.) */ - long alloc_large_unboxed_start_page; + page_index_t alloc_large_unboxed_start_page; /* the bytes allocated to this generation */ - long bytes_allocated; + unsigned long bytes_allocated; /* the number of bytes at which to trigger a GC */ - long gc_trigger; + unsigned long gc_trigger; /* to calculate a new level for gc_trigger */ - long bytes_consed_between_gc; + unsigned long bytes_consed_between_gc; /* 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 * objects are added from a GC of a younger generation. Dividing by * the bytes_allocated will give the average age of the memory in * this generation since its last GC. */ - long cum_sum_bytes_allocated; + unsigned long cum_sum_bytes_allocated; /* 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. */ +#ifdef LUTEX_WIDETAG + struct lutex *lutexes; +#else + void *lutexes; +#endif }; -/* the number of actual generations. (The number of 'struct - * generation' objects is one more than this, because one object - * serves as scratch when GC'ing.) */ -#define NUM_GENERATIONS 6 /* an array of generation structures. There needs to be one more * generation structure than actual generations as the oldest * generation is temporarily raised then lowered. */ -struct generation generations[NUM_GENERATIONS+1]; +struct generation generations[NUM_GENERATIONS]; /* the oldest generation that is will currently be GCed by default. - * Valid values are: 0, 1, ... (NUM_GENERATIONS-1) + * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION * - * The default of (NUM_GENERATIONS-1) enables GC on all generations. + * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations. * * Setting this to 0 effectively disables the generational nature of * the GC. In some applications generational GC may not be useful @@ -250,22 +335,31 @@ struct generation generations[NUM_GENERATIONS+1]; * An intermediate value could be handy after moving long-lived data * into an older generation so an unnecessary GC of this long-lived * data can be avoided. */ -unsigned int gencgc_oldest_gen_to_gc = NUM_GENERATIONS-1; +generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION; /* The maximum free page in the heap is maintained and used to update * ALLOCATION_POINTER which is used by the room function to limit its * search of the heap. XX Gencgc obviously needs to be better * integrated with the Lisp code. */ -static long last_free_page; +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 */ +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 + +extern unsigned long gencgc_release_granularity; +unsigned long gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY; -static lispobj free_pages_lock=0; +extern unsigned long gencgc_alloc_granularity; +unsigned long gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY; /* @@ -275,13 +369,13 @@ static lispobj free_pages_lock=0; /* Count the number of pages which are write-protected within the * given generation. */ static long -count_write_protect_generation_pages(int generation) +count_write_protect_generation_pages(generation_index_t generation) { - long i; - long count = 0; + page_index_t i; + 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++; @@ -290,26 +384,27 @@ count_write_protect_generation_pages(int generation) /* Count the number of pages within the given generation. */ static long -count_generation_pages(int generation) +count_generation_pages(generation_index_t generation) { - long i; + page_index_t i; long count = 0; for (i = 0; i < last_free_page; i++) - if ((page_table[i].allocated != 0) + if (page_allocated_p(i) && (page_table[i].gen == generation)) count++; return count; } -#ifdef QSHOW +#if QSHOW static long count_dont_move_pages(void) { - long i; + page_index_t i; long count = 0; for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated != 0) && (page_table[i].dont_move != 0)) { + if (page_allocated_p(i) + && (page_table[i].dont_move != 0)) { ++count; } } @@ -319,21 +414,22 @@ count_dont_move_pages(void) /* Work through the pages and add up the number of bytes used for the * given generation. */ -static long -count_generation_bytes_allocated (int gen) +static unsigned long +count_generation_bytes_allocated (generation_index_t gen) { - long i; - long result = 0; + page_index_t i; + unsigned long result = 0; for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated != 0) && (page_table[i].gen == gen)) + if (page_allocated_p(i) + && (page_table[i].gen == gen)) result += page_table[i].bytes_used; } return result; } /* Return the average age of the memory in a generation. */ -static double -gen_av_mem_age(int gen) +extern double +generation_average_age(generation_index_t gen) { if (generations[gen].bytes_allocated == 0) return 0.0; @@ -343,44 +439,41 @@ gen_av_mem_age(int gen) / ((double)generations[gen].bytes_allocated); } -void fpu_save(int *); /* defined in x86-assem.S */ -void fpu_restore(int *); /* defined in x86-assem.S */ -/* 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) { - int i, gens; - int fpu_state[27]; + generation_index_t i; + +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) +#define FPU_STATE_SIZE 27 + int fpu_state[FPU_STATE_SIZE]; +#elif defined(LISP_FEATURE_PPC) +#define FPU_STATE_SIZE 32 + long long fpu_state[FPU_STATE_SIZE]; +#endif /* This code uses the FP instructions which may be set up for Lisp * so they need to be saved and reset for C. */ fpu_save(fpu_state); - /* number of generations to print */ - if (verbose) - gens = NUM_GENERATIONS+1; - else - gens = NUM_GENERATIONS; - /* Print the heap stats. */ - fprintf(stderr, - " Gen Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n"); + 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++) { - int j; - int boxed_cnt = 0; - int unboxed_cnt = 0; - int large_boxed_cnt = 0; - int large_unboxed_cnt = 0; - int pinned_cnt=0; + for (i = 0; i < SCRATCH_GENERATION; i++) { + page_index_t j; + long boxed_cnt = 0; + long unboxed_cnt = 0; + long large_boxed_cnt = 0; + long large_unboxed_cnt = 0; + long pinned_cnt=0; for (j = 0; j < last_free_page; j++) if (page_table[j].gen == i) { /* 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 @@ -389,7 +482,7 @@ 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 @@ -399,27 +492,164 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */ gc_assert(generations[i].bytes_allocated == count_generation_bytes_allocated(i)); - fprintf(stderr, - " %1d: %5d %5d %5d %5d %5d %8ld %5ld %8ld %4ld %3d %7.4f\n", + fprintf(file, + " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n", i, - boxed_cnt, unboxed_cnt, large_boxed_cnt, large_unboxed_cnt, + generations[i].alloc_start_page, + generations[i].alloc_unboxed_start_page, + generations[i].alloc_large_start_page, + generations[i].alloc_large_unboxed_start_page, + boxed_cnt, + unboxed_cnt, + large_boxed_cnt, + large_unboxed_cnt, pinned_cnt, generations[i].bytes_allocated, - (count_generation_pages(i)*PAGE_BYTES + (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated), 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(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); +} -/* - * allocation routines + +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) +void fast_bzero(void*, size_t); /* in -assem.S */ +#endif + +/* Zero the pages from START to END (inclusive), but use mmap/munmap instead + * if zeroing it ourselves, i.e. in practice give the memory back to the + * OS. Generally done after a large GC. + */ +void zero_pages_with_mmap(page_index_t start, page_index_t end) { + int i; + void *addr = page_address(start), *new_addr; + size_t length = npage_bytes(1+end-start); + + 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) { + lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x", + start, new_addr); + } + + for (i = start; i <= end; i++) { + page_table[i].need_to_zero = 0; + } +} + +/* Zero the pages from START to END (inclusive). Generally done just after + * a new region has been allocated. */ +static void +zero_pages(page_index_t start, page_index_t end) { + if (start > end) + return; + +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + fast_bzero(page_address(start), npage_bytes(1+end-start)); +#else + bzero(page_address(start), npage_bytes(1+end-start)); +#endif + +} + +/* 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, j; + + for (i = start; i <= end; i++) { + 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++) { + page_table[i].need_to_zero = 1; + } +} + /* * To support quick and inline allocation, regions of memory can be @@ -473,7 +703,56 @@ struct alloc_region boxed_region; struct alloc_region unboxed_region; /* The generation currently being allocated to. */ -static int gc_alloc_generation; +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. * @@ -499,12 +778,13 @@ static int 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) { - long first_page; - long last_page; - long bytes_found; - long i; + page_index_t first_page; + page_index_t last_page; + unsigned long bytes_found; + page_index_t i; + int ret; /* FSHOW((stderr, @@ -516,17 +796,12 @@ gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) gc_assert((alloc_region->first_page == 0) && (alloc_region->last_page == -1) && (alloc_region->free_pointer == alloc_region->end_addr)); - get_spinlock(&free_pages_lock,(long) alloc_region); - 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); - bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used) - + PAGE_BYTES*(last_page-first_page); + ret = thread_mutex_lock(&free_pages_lock); + 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=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used) + + npage_bytes(last_page-first_page); /* Set up the alloc_region. */ alloc_region->first_page = first_page; @@ -540,45 +815,53 @@ 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].first_object_offset = 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 * broken before!) */ - page_table[i].first_object_offset = - alloc_region->start_addr - page_address(i); + page_table[i].region_start_offset = + void_diff(page_address(i),alloc_region->start_addr); page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ; } /* Bump up last_free_page. */ if (last_page+1 > last_free_page) { last_free_page = last_page+1; - SetSymbolValue(ALLOCATION_POINTER, - (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES), - 0); + /* do we only want to call this on special occasions? like for + * boxed_region? */ + set_alloc_pointer((lispobj)page_address(last_free_page)); + } + ret = thread_mutex_unlock(&free_pages_lock); + gc_assert(ret == 0); + +#ifdef READ_PROTECT_FREE_PAGES + os_protect(page_address(first_page), + npage_bytes(1+last_page-first_page), + OS_VM_PROT_ALL); +#endif + + /* If the first page was only partial, don't check whether it's + * zeroed (it won't be) and don't zero it (since the parts that + * we're interested in are guaranteed to be zeroed). + */ + if (page_table[first_page].bytes_used) { + first_page++; } - release_spinlock(&free_pages_lock); + + zero_dirty_pages(first_page, last_page); /* we can do this after releasing free_pages_lock */ if (gencgc_zero_check) { @@ -590,13 +873,13 @@ gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) * (long) in code like this, so that it is less likely to * break randomly when running on a machine with different * word sizes. -- WHN 19991129 */ - lose("The new region at %x is not zero.", p); + lose("The new region at %x is not zero (start=%p, end=%p).\n", + p, alloc_region->start_addr, alloc_region->end_addr); } + } } } -} - /* If the record_new_objects flag is 2 then all new regions created * are recorded. * @@ -614,11 +897,11 @@ gc_alloc_new_region(long nbytes, int unboxed, struct alloc_region *alloc_region) * scavenge of a generation. */ #define NUM_NEW_AREAS 512 static int record_new_objects = 0; -static long new_areas_ignore_page; +static page_index_t new_areas_ignore_page; struct new_area { - long page; - long offset; - long size; + page_index_t page; + size_t offset; + size_t size; }; static struct new_area (*new_areas)[]; static long new_areas_index; @@ -626,9 +909,9 @@ long max_new_areas; /* Add a new area to new_areas. */ static void -add_new_area(long first_page, long offset, long size) +add_new_area(page_index_t first_page, size_t offset, size_t size) { - unsigned new_area_start,c; + unsigned long new_area_start,c; long i; /* Ignore if full. */ @@ -648,13 +931,13 @@ add_new_area(long first_page, long offset, long size) gc_abort(); } - new_area_start = PAGE_BYTES*first_page + offset; + new_area_start = npage_bytes(first_page) + offset; /* Search backwards for a prior area that this follows from. If found this will save adding a new area. */ for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) { - unsigned area_end = - PAGE_BYTES*((*new_areas)[i].page) + unsigned long area_end = + npage_bytes((*new_areas)[i].page) + (*new_areas)[i].offset + (*new_areas)[i].size; /*FSHOW((stderr, @@ -696,15 +979,16 @@ add_new_area(long first_page, long offset, long 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) { - long more; - long first_page; - long next_page; - long bytes_used; - long orig_first_page_bytes_used; - long region_size; - long byte_cnt; + int more; + page_index_t first_page; + page_index_t next_page; + unsigned long bytes_used; + unsigned long orig_first_page_bytes_used; + unsigned long region_size; + unsigned long byte_cnt; + int ret; first_page = alloc_region->first_page; @@ -715,27 +999,27 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) next_page = first_page+1; - get_spinlock(&free_pages_lock,(long) alloc_region); + ret = thread_mutex_lock(&free_pages_lock); + gc_assert(ret == 0); if (alloc_region->free_pointer != alloc_region->start_addr) { /* some bytes were allocated in the region */ orig_first_page_bytes_used = page_table[first_page].bytes_used; - gc_assert(alloc_region->start_addr == (page_address(first_page) + page_table[first_page].bytes_used)); + gc_assert(alloc_region->start_addr == + (page_address(first_page) + + page_table[first_page].bytes_used)); /* All the pages used need to be updated */ /* Update the first page. */ /* If the page was free then set up the gen, and - * first_object_offset. */ + * region_start_offset. */ if (page_table[first_page].bytes_used == 0) - gc_assert(page_table[first_page].first_object_offset == 0); + 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); @@ -744,35 +1028,35 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) /* Calculate 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 = (alloc_region->free_pointer - page_address(first_page)))>PAGE_BYTES) { - bytes_used = PAGE_BYTES; + if ((bytes_used = void_diff(alloc_region->free_pointer, + page_address(first_page))) + >GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[first_page].bytes_used = bytes_used; byte_cnt += bytes_used; - /* All the rest of the pages should be free. We need to set their - * first_object_offset pointer to the start of the region, and set - * the bytes_used. */ + /* All the rest of the pages should be free. We need to set + * their region_start_offset pointer to the start of the + * 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); - gc_assert(page_table[next_page].first_object_offset == - alloc_region->start_addr - page_address(next_page)); + gc_assert(page_table[next_page].region_start_offset == + void_diff(page_address(next_page), + alloc_region->start_addr)); /* Calculate the number of bytes used in this page. */ more = 0; - if ((bytes_used = (alloc_region->free_pointer - - page_address(next_page)))>PAGE_BYTES) { - bytes_used = PAGE_BYTES; + if ((bytes_used = void_diff(alloc_region->free_pointer, + page_address(next_page)))>GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[next_page].bytes_used = bytes_used; @@ -781,7 +1065,8 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) next_page++; } - region_size = alloc_region->free_pointer - alloc_region->start_addr; + region_size = void_diff(alloc_region->free_pointer, + alloc_region->start_addr); bytes_allocated += region_size; generations[gc_alloc_generation].bytes_allocated += region_size; @@ -789,14 +1074,10 @@ 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); /* @@ -819,7 +1100,9 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) page_table[next_page].allocated = FREE_PAGE_FLAG; next_page++; } - release_spinlock(&free_pages_lock); + ret = thread_mutex_unlock(&free_pages_lock); + gc_assert(ret == 0); + /* alloc_region is per-thread, we're ok to do this unlocked */ gc_set_region_empty(alloc_region); } @@ -828,56 +1111,44 @@ 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) { - long first_page; - long last_page; - long orig_first_page_bytes_used; + page_index_t first_page; + page_index_t last_page; + int orig_first_page_bytes_used; long byte_cnt; - long more; - long bytes_used; - long next_page; + int more; + unsigned long bytes_used; + page_index_t next_page; + int ret; - get_spinlock(&free_pages_lock,(long) alloc_region); + 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; /* If the first page was free then set up the gen, and - * first_object_offset. */ + * 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].first_object_offset = 0; + 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); @@ -886,8 +1157,8 @@ gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) /* 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; @@ -896,25 +1167,23 @@ gc_alloc_large(long nbytes, int unboxed, struct alloc_region *alloc_region) next_page = first_page+1; /* All the rest of the pages should be free. We need to set their - * first_object_offset pointer to the start of the region, and - * set the bytes_used. */ + * 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; - page_table[next_page].first_object_offset = - orig_first_page_bytes_used - PAGE_BYTES*(next_page-first_page); + page_table[next_page].region_start_offset = + npage_bytes(next_page-first_page) - orig_first_page_bytes_used; /* Calculate the number of bytes used in this page. */ more = 0; - if ((bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt) > PAGE_BYTES) { - bytes_used = PAGE_BYTES; + bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt; + if (bytes_used > GENCGC_CARD_BYTES) { + bytes_used = GENCGC_CARD_BYTES; more = 1; } page_table[next_page].bytes_used = bytes_used; @@ -930,111 +1199,167 @@ 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 */ if (last_page+1 > last_free_page) { last_free_page = last_page+1; - SetSymbolValue(ALLOCATION_POINTER, - (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0); + set_alloc_pointer((lispobj)(page_address(last_free_page))); } - release_spinlock(&free_pages_lock); + ret = thread_mutex_unlock(&free_pages_lock); + gc_assert(ret == 0); + +#ifdef READ_PROTECT_FREE_PAGES + os_protect(page_address(first_page), + npage_bytes(1+last_page-first_page), + OS_VM_PROT_ALL); +#endif + + zero_dirty_pages(first_page, last_page); - return((void *)(page_address(first_page)+orig_first_page_bytes_used)); + return page_address(first_page); } -long -gc_find_freeish_pages(long *restart_page_ptr, long nbytes, int unboxed) +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 + * handled, or indeed even printed. + */ + 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. + */ + 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"); + } +} + +page_index_t +gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, + int page_type_flag) { - long first_page; - long last_page; - long region_size; - long restart_page=*restart_page_ptr; - long bytes_found; - long num_pages; - long large_p=(nbytes>=large_object_size); - gc_assert(free_pages_lock); - - /* Search for a contiguous free space of at least nbytes. If it's - * a large object then align it on a page boundary by searching - * for a free page. */ + 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; + } - do { - first_page = restart_page; - if (large_p) - while ((first_page < NUM_PAGES) - && (page_table[first_page].allocated != FREE_PAGE_FLAG)) - first_page++; - else - while (first_page < NUM_PAGES) { - if(page_table[first_page].allocated == FREE_PAGE_FLAG) - break; - if((page_table[first_page].allocated == - (unboxed ? UNBOXED_PAGE_FLAG : BOXED_PAGE_FLAG)) && + gc_assert(nbytes>=0); + /* 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].bytes_used < (PAGE_BYTES-32)) && (page_table[first_page].write_protected == 0) && (page_table[first_page].dont_move == 0)) { - break; - } + 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++; + continue; } - - if (first_page >= NUM_PAGES) { - fprintf(stderr, - "Argh! gc_find_free_space failed (first_page), nbytes=%ld.\n", - nbytes); - print_generation_stats(1); - lose(NULL); + } else { + first_page++; + continue; } 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); + } - last_page = first_page; - bytes_found = PAGE_BYTES - page_table[first_page].bytes_used; - num_pages = 1; - while (((bytes_found < nbytes) - || (!large_p && (num_pages < 2))) - && (last_page < (NUM_PAGES-1)) - && (page_table[last_page+1].allocated == FREE_PAGE_FLAG)) { - last_page++; - num_pages++; - bytes_found += PAGE_BYTES; - gc_assert(page_table[last_page].write_protected == 0); + 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; - region_size = (PAGE_BYTES - page_table[first_page].bytes_used) - + PAGE_BYTES*(last_page-first_page); + first_page = last_page; + } - gc_assert(bytes_found == region_size); - restart_page = last_page + 1; - } while ((restart_page < NUM_PAGES) && (bytes_found < nbytes)); + bytes_found = most_bytes_found; + restart_page = first_page + 1; /* Check for a failure */ - if ((restart_page >= NUM_PAGES) && (bytes_found < nbytes)) { - fprintf(stderr, - "Argh! gc_find_freeish_pages failed (restart_page), nbytes=%ld.\n", - nbytes); - print_generation_stats(1); - lose(NULL); + if (bytes_found < nbytes) { + gc_assert(restart_page >= page_table_pages); + gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes); } - *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 * 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); + if (nbytes>=large_object_size) + 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; @@ -1050,11 +1375,11 @@ gc_alloc_with_region(long nbytes,int unboxed_p, struct alloc_region *my_region, /* Unless a `quick' alloc was requested, check whether the alloc region is almost empty. */ if (!quick_p && - (my_region->end_addr - my_region->free_pointer) <= 32) { + 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); @@ -1063,60 +1388,45 @@ 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); } -/* - * scavenging/transporting routines derived from gc.c in CMU CL ca. 18b - */ - -extern long (*scavtab[256])(lispobj *where, lispobj object); -extern lispobj (*transother[256])(lispobj object); -extern long (*sizetab[256])(lispobj *where); /* Copy a large boxed object. If the object is in a large object * region then it is simply promoted, else it is copied. If it's large @@ -1129,7 +1439,7 @@ copy_large_object(lispobj object, long nwords) { int tag; lispobj *new; - long first_page; + page_index_t first_page; gc_assert(is_lisp_pointer(object)); gc_assert(from_space_p(object)); @@ -1144,10 +1454,10 @@ copy_large_object(lispobj object, long nwords) /* Promote the object. */ - long remaining_bytes; - long next_page; - long bytes_freed; - long old_bytes_used; + unsigned long remaining_bytes; + page_index_t next_page; + unsigned long bytes_freed; + unsigned long old_bytes_used; /* Note: Any page write-protection must be removed, else a * later scavenge_newspace may incorrectly not scavenge these @@ -1155,27 +1465,23 @@ copy_large_object(lispobj object, long nwords) * new areas, but let's do it for them all (they'll probably * be written anyway?). */ - gc_assert(page_table[first_page].first_object_offset == 0); + gc_assert(page_table[first_page].region_start_offset == 0); 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_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].first_object_offset== - -PAGE_BYTES*(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].region_start_offset == + npage_bytes(next_page-first_page)); + 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++; } @@ -1186,7 +1492,7 @@ 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; @@ -1196,12 +1502,12 @@ 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_table[next_page].allocated == BOXED_PAGE_FLAG) && + page_boxed_p(next_page) && page_table[next_page].large_object && - (page_table[next_page].first_object_offset == - -(next_page - first_page)*PAGE_BYTES)) { + (page_table[next_page].region_start_offset == + npage_bytes(next_page - first_page))) { /* Checks out OK, free the page. Don't need to bother zeroing * pages as this should have been done before shrinking the * object. These pages shouldn't be write-protected as they @@ -1215,8 +1521,8 @@ copy_large_object(lispobj object, long nwords) next_page++; } - generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords + - bytes_freed; + generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords + + bytes_freed; generations[new_space].bytes_allocated += N_WORD_BYTES*nwords; bytes_allocated -= bytes_freed; @@ -1277,14 +1583,16 @@ copy_large_unboxed_object(lispobj object, long nwords) { int tag; lispobj *new; - long first_page; + page_index_t first_page; gc_assert(is_lisp_pointer(object)); gc_assert(from_space_p(object)); gc_assert((nwords & 0x01) == 0); - if ((nwords > 1024*1024) && gencgc_verbose) - FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n", nwords*N_WORD_BYTES)); + 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); @@ -1294,27 +1602,26 @@ copy_large_unboxed_object(lispobj object, long nwords) /* Promote the object. Note: Unboxed objects may have been * allocated to a BOXED region so it may be necessary to * change the region to UNBOXED. */ - long remaining_bytes; - long next_page; - long bytes_freed; - long old_bytes_used; + unsigned long remaining_bytes; + page_index_t next_page; + unsigned long bytes_freed; + unsigned long old_bytes_used; - gc_assert(page_table[first_page].first_object_offset == 0); + gc_assert(page_table[first_page].region_start_offset == 0); 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_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].first_object_offset== - -PAGE_BYTES*(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].region_start_offset == + npage_bytes(next_page-first_page)); + 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++; } @@ -1335,13 +1642,12 @@ 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_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].first_object_offset == - -(next_page - first_page)*PAGE_BYTES)) { + (page_table[next_page].region_start_offset == + npage_bytes(next_page - first_page))) { /* Checks out OK, free the page. Don't need to both zeroing * pages as this should have been done before shrinking the * object. These pages shouldn't be write-protected, even if @@ -1355,12 +1661,14 @@ 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; + generations[from_space].bytes_allocated -= + nwords*N_WORD_BYTES + bytes_freed; generations[new_space].bytes_allocated += nwords*N_WORD_BYTES; bytes_allocated -= bytes_freed; @@ -1403,17 +1711,20 @@ static lispobj trans_boxed(lispobj object); * Currently only absolute fixups to the constant vector, or to the * code area are checked. */ void -sniff_code_object(struct code *code, unsigned displacement) +sniff_code_object(struct code *code, unsigned long displacement) { +#ifdef LISP_FEATURE_X86 long nheader_words, ncode_words, nwords; void *p; - void *constants_start_addr, *constants_end_addr; + void *constants_start_addr = NULL, *constants_end_addr; void *code_start_addr, *code_end_addr; int fixup_found = 0; if (!check_code_fixups) return; + FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement)); + ncode_words = fixnum_value(code->code_size); nheader_words = HeaderValue(*(lispobj *)code); nwords = ncode_words + nheader_words; @@ -1430,7 +1741,7 @@ sniff_code_object(struct code *code, unsigned 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 @@ -1442,7 +1753,8 @@ sniff_code_object(struct code *code, unsigned displacement) && (data < (code_end_addr-displacement))) { /* function header */ if ((d4 == 0x5e) - && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) == (unsigned)code)) { + && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) == + (unsigned)code)) { /* Skip the function header */ p += 6*4 - 4 - 1; continue; @@ -1570,16 +1882,20 @@ sniff_code_object(struct code *code, unsigned displacement) "/code start = %x, end = %x\n", code_start_addr, code_end_addr)); } +#endif } void gencgc_apply_code_fixups(struct code *old_code, struct code *new_code) { +/* x86-64 uses pc-relative addressing instead of this kludge */ +#ifndef LISP_FEATURE_X86_64 long nheader_words, ncode_words, nwords; void *constants_start_addr, *constants_end_addr; void *code_start_addr, *code_end_addr; lispobj fixups = NIL; - unsigned displacement = (unsigned)new_code - (unsigned)old_code; + unsigned long displacement = + (unsigned long)new_code - (unsigned long)old_code; struct vector *fixups_vector; ncode_words = fixnum_value(new_code->code_size); @@ -1627,7 +1943,8 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code) (fixups_vector->header == 0x01)) { /* If so, then follow it. */ /*SHOW("following pointer to a forwarding pointer");*/ - fixups_vector = (struct vector *)native_pointer((lispobj)fixups_vector->length); + fixups_vector = + (struct vector *)native_pointer((lispobj)fixups_vector->length); } /*SHOW("got fixups");*/ @@ -1638,33 +1955,38 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code) long length = fixnum_value(fixups_vector->length); long i; for (i = 0; i < length; i++) { - unsigned offset = fixups_vector->data[i]; + unsigned long offset = fixups_vector->data[i]; /* Now check the current value of offset. */ - unsigned old_value = - *(unsigned *)((unsigned)code_start_addr + offset); + unsigned long old_value = + *(unsigned long *)((unsigned long)code_start_addr + offset); /* If it's within the old_code object then it must be an * absolute fixup (relative ones are not saved) */ - if ((old_value >= (unsigned)old_code) - && (old_value < ((unsigned)old_code + nwords*N_WORD_BYTES))) + if ((old_value >= (unsigned long)old_code) + && (old_value < ((unsigned long)old_code + + nwords*N_WORD_BYTES))) /* So add the dispacement. */ - *(unsigned *)((unsigned)code_start_addr + offset) = + *(unsigned long *)((unsigned long)code_start_addr + offset) = old_value + displacement; else /* It is outside the old code object so it must be a * relative fixup (absolute fixups are not saved). So * subtract the displacement. */ - *(unsigned *)((unsigned)code_start_addr + offset) = + *(unsigned long *)((unsigned long)code_start_addr + offset) = old_value - displacement; } } else { - fprintf(stderr, "widetag of fixup vector is %d\n", widetag_of(fixups_vector->header)); + /* This used to just print a note to stderr, but a bogus fixup seems to + * indicate real heap corruption, so a hard hailure is in order. */ + lose("fixup vector %p has a bad widetag: %d\n", + fixups_vector, widetag_of(fixups_vector->header)); } /* Check for possible errors. */ if (check_code_fixups) { sniff_code_object(new_code,displacement); } +#endif } @@ -1691,7 +2013,6 @@ trans_unboxed_large(lispobj object) lispobj header; unsigned long length; - gc_assert(is_lisp_pointer(object)); header = *((lispobj *) native_pointer(object)); @@ -1704,219 +2025,177 @@ trans_unboxed_large(lispobj object) /* - * vector-like objects + * Lutexes. Using the normal finalization machinery for finalizing + * lutexes is tricky, since the finalization depends on working lutexes. + * So we track the lutexes in the GC and finalize them manually. */ +#if defined(LUTEX_WIDETAG) -/* FIXME: What does this mean? */ -int gencgc_hash = 1; +/* + * Start tracking LUTEX in the GC, by adding it to the linked list of + * lutexes in the nursery generation. The caller is responsible for + * locking, and GCs must be inhibited until the registration is + * complete. + */ +void +gencgc_register_lutex (struct lutex *lutex) { + int index = find_page_index(lutex); + generation_index_t gen; + struct lutex *head; -static long -scav_vector(lispobj *where, lispobj object) -{ - unsigned long kv_length; - lispobj *kv_vector; - unsigned long length = 0; /* (0 = dummy to stop GCC warning) */ - struct hash_table *hash_table; - lispobj empty_symbol; - unsigned long *index_vector = NULL; /* (NULL = dummy to stop GCC warning) */ - unsigned long *next_vector = NULL; /* (NULL = dummy to stop GCC warning) */ - unsigned long *hash_vector = NULL; /* (NULL = dummy to stop GCC warning) */ - lispobj weak_p_obj; - unsigned next_vector_length = 0; - - /* FIXME: A comment explaining this would be nice. It looks as - * though SB-VM:VECTOR-VALID-HASHING-SUBTYPE is set for EQ-based - * hash tables in the Lisp HASH-TABLE code, and nowhere else. */ - if (HeaderValue(object) != subtype_VectorValidHashing) - return 1; + /* This lutex is in static space, so we don't need to worry about + * finalizing it. + */ + if (index == -1) + return; - if (!gencgc_hash) { - /* This is set for backward compatibility. FIXME: Do we need - * this any more? */ - *where = - (subtype_VectorMustRehash<= 0); + gc_assert(gen < NUM_GENERATIONS); - /* Scavenge element 0, which may be a hash-table structure. */ - scavenge(where+2, 1); - if (!is_lisp_pointer(where[2])) { - lose("no pointer at %x in hash table", where[2]); - } - hash_table = (struct hash_table *)native_pointer(where[2]); - /*FSHOW((stderr,"/hash_table = %x\n", hash_table));*/ - if (widetag_of(hash_table->header) != INSTANCE_HEADER_WIDETAG) { - lose("hash table not instance (%x at %x)", - hash_table->header, - hash_table); - } + head = generations[gen].lutexes; - /* Scavenge element 1, which should be some internal symbol that - * the hash table code reserves for marking empty slots. */ - scavenge(where+3, 1); - if (!is_lisp_pointer(where[3])) { - lose("not empty-hash-table-slot symbol pointer: %x", where[3]); - } - empty_symbol = where[3]; - /* fprintf(stderr,"* empty_symbol = %x\n", empty_symbol);*/ - if (widetag_of(*(lispobj *)native_pointer(empty_symbol)) != - SYMBOL_HEADER_WIDETAG) { - lose("not a symbol where empty-hash-table-slot symbol expected: %x", - *(lispobj *)native_pointer(empty_symbol)); - } + lutex->gen = gen; + lutex->next = head; + lutex->prev = NULL; + if (head) + head->prev = lutex; + generations[gen].lutexes = lutex; +} - /* Scavenge hash table, which will fix the positions of the other - * needed objects. */ - scavenge((lispobj *)hash_table, - sizeof(struct hash_table) / sizeof(lispobj)); +/* + * Stop tracking LUTEX in the GC by removing it from the appropriate + * linked lists. This will only be called during GC, so no locking is + * needed. + */ +void +gencgc_unregister_lutex (struct lutex *lutex) { + if (lutex->prev) { + lutex->prev->next = lutex->next; + } else { + generations[lutex->gen].lutexes = lutex->next; + } - /* Cross-check the kv_vector. */ - if (where != (lispobj *)native_pointer(hash_table->table)) { - lose("hash_table table!=this table %x", hash_table->table); + if (lutex->next) { + lutex->next->prev = lutex->prev; } - /* WEAK-P */ - weak_p_obj = hash_table->weak_p; + lutex->next = NULL; + lutex->prev = NULL; + lutex->gen = -1; +} - /* index vector */ - { - lispobj index_vector_obj = hash_table->index_vector; - - if (is_lisp_pointer(index_vector_obj) && - (widetag_of(*(lispobj *)native_pointer(index_vector_obj)) == - SIMPLE_ARRAY_WORD_WIDETAG)) { - index_vector = - ((unsigned long *)native_pointer(index_vector_obj)) + 2; - /*FSHOW((stderr, "/index_vector = %x\n",index_vector));*/ - length = fixnum_value(((lispobj *)native_pointer(index_vector_obj))[1]); - /*FSHOW((stderr, "/length = %d\n", length));*/ - } else { - lose("invalid index_vector %x", index_vector_obj); - } +/* + * Mark all lutexes in generation GEN as not live. + */ +static void +unmark_lutexes (generation_index_t gen) { + struct lutex *lutex = generations[gen].lutexes; + + while (lutex) { + lutex->live = 0; + lutex = lutex->next; } +} - /* next vector */ - { - lispobj next_vector_obj = hash_table->next_vector; - - if (is_lisp_pointer(next_vector_obj) && - (widetag_of(*(lispobj *)native_pointer(next_vector_obj)) == - SIMPLE_ARRAY_WORD_WIDETAG)) { - next_vector = ((unsigned long *)native_pointer(next_vector_obj)) + 2; - /*FSHOW((stderr, "/next_vector = %x\n", next_vector));*/ - next_vector_length = fixnum_value(((lispobj *)native_pointer(next_vector_obj))[1]); - /*FSHOW((stderr, "/next_vector_length = %d\n", next_vector_length));*/ - } else { - lose("invalid next_vector %x", next_vector_obj); +/* + * Finalize all lutexes in generation GEN that have not been marked live. + */ +static void +reap_lutexes (generation_index_t gen) { + struct lutex *lutex = generations[gen].lutexes; + + while (lutex) { + struct lutex *next = lutex->next; + if (!lutex->live) { + lutex_destroy((tagged_lutex_t) lutex); + gencgc_unregister_lutex(lutex); } + lutex = next; } +} - /* maybe hash vector */ - { - lispobj hash_vector_obj = hash_table->hash_vector; - - if (is_lisp_pointer(hash_vector_obj) && - (widetag_of(*(lispobj *)native_pointer(hash_vector_obj)) == - SIMPLE_ARRAY_WORD_WIDETAG)){ - hash_vector = - ((unsigned long *)native_pointer(hash_vector_obj)) + 2; - /*FSHOW((stderr, "/hash_vector = %x\n", hash_vector));*/ - gc_assert(fixnum_value(((lispobj *)native_pointer(hash_vector_obj))[1]) - == next_vector_length); - } else { - hash_vector = NULL; - /*FSHOW((stderr, "/no hash_vector: %x\n", hash_vector_obj));*/ - } +/* + * Mark LUTEX as live. + */ +static void +mark_lutex (lispobj tagged_lutex) { + struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex); + + lutex->live = 1; +} + +/* + * Move all lutexes in generation FROM to generation TO. + */ +static void +move_lutexes (generation_index_t from, generation_index_t to) { + struct lutex *tail = generations[from].lutexes; + + /* Nothing to move */ + if (!tail) + return; + + /* Change the generation of the lutexes in FROM. */ + while (tail->next) { + tail->gen = to; + tail = tail->next; } + tail->gen = to; - /* These lengths could be different as the index_vector can be a - * different length from the others, a larger index_vector could help - * reduce collisions. */ - gc_assert(next_vector_length*2 == kv_length); + /* Link the last lutex in the FROM list to the start of the TO list */ + tail->next = generations[to].lutexes; - /* now all set up.. */ + /* And vice versa */ + if (generations[to].lutexes) { + generations[to].lutexes->prev = tail; + } - /* Work through the KV vector. */ - { - long i; - for (i = 1; i < next_vector_length; i++) { - lispobj old_key = kv_vector[2*i]; + /* And update the generations structures to match this */ + generations[to].lutexes = generations[from].lutexes; + generations[from].lutexes = NULL; +} -#if N_WORD_BITS == 32 - unsigned long old_index = (old_key & 0x1fffffff)%length; -#elif N_WORD_BITS == 64 - unsigned long old_index = (old_key & 0x1fffffffffffffff)%length; -#endif +static long +scav_lutex(lispobj *where, lispobj object) +{ + mark_lutex((lispobj) where); - /* Scavenge the key and value. */ - scavenge(&kv_vector[2*i],2); + return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2); +} - /* Check whether the key has moved and is EQ based. */ - { - lispobj new_key = kv_vector[2*i]; -#if N_WORD_BITS == 32 - unsigned long new_index = (new_key & 0x1fffffff)%length; -#elif N_WORD_BITS == 64 - unsigned long new_index = (new_key & 0x1fffffffffffffff)%length; -#endif +static lispobj +trans_lutex(lispobj object) +{ + struct lutex *lutex = (struct lutex *) native_pointer(object); + lispobj copied; + size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2); + gc_assert(is_lisp_pointer(object)); + copied = copy_object(object, words); - if ((old_index != new_index) && - ((!hash_vector) || - (hash_vector[i] == MAGIC_HASH_VECTOR_VALUE)) && - ((new_key != empty_symbol) || - (kv_vector[2*i] != empty_symbol))) { - - /*FSHOW((stderr, - "* EQ key %d moved from %x to %x; index %d to %d\n", - i, old_key, new_key, old_index, new_index));*/ - - if (index_vector[old_index] != 0) { - /*FSHOW((stderr, "/P1 %d\n", index_vector[old_index]));*/ - - /* Unlink the key from the old_index chain. */ - if (index_vector[old_index] == i) { - /*FSHOW((stderr, "/P2a %d\n", next_vector[i]));*/ - index_vector[old_index] = next_vector[i]; - /* Link it into the needing rehash chain. */ - next_vector[i] = fixnum_value(hash_table->needing_rehash); - hash_table->needing_rehash = make_fixnum(i); - /*SHOW("P2");*/ - } else { - unsigned prior = index_vector[old_index]; - unsigned next = next_vector[prior]; - - /*FSHOW((stderr, "/P3a %d %d\n", prior, next));*/ - - while (next != 0) { - /*FSHOW((stderr, "/P3b %d %d\n", prior, next));*/ - if (next == i) { - /* Unlink it. */ - next_vector[prior] = next_vector[next]; - /* Link it into the needing rehash - * chain. */ - next_vector[next] = - fixnum_value(hash_table->needing_rehash); - hash_table->needing_rehash = make_fixnum(next); - /*SHOW("/P3");*/ - break; - } - prior = next; - next = next_vector[next]; - } - } - } - } - } - } + /* Update the links, since the lutex moved in memory. */ + if (lutex->next) { + lutex->next->prev = (struct lutex *) native_pointer(copied); + } + + if (lutex->prev) { + lutex->prev->next = (struct lutex *) native_pointer(copied); + } else { + generations[lutex->gen].lutexes = + (struct lutex *) native_pointer(copied); } - return (CEILING(kv_length + 2, 2)); + + return copied; } +static long +size_lutex(lispobj *where) +{ + return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2); +} +#endif /* LUTEX_WIDETAG */ /* @@ -1934,29 +2213,21 @@ scav_vector(lispobj *where, lispobj object) static long scav_weak_pointer(lispobj *where, lispobj object) { - struct weak_pointer *wp = weak_pointers; - /* Push the weak pointer onto the list of weak pointers. - * Do I have to watch for duplicates? Originally this was - * part of trans_weak_pointer but that didn't work in the - * case where the WP was in a promoted region. + /* Since we overwrite the 'next' field, we have to make + * sure not to do so for pointers already in the list. + * Instead of searching the list of weak_pointers each + * time, we ensure that next is always NULL when the weak + * pointer isn't in the list, and not NULL otherwise. + * Since we can't use NULL to denote end of list, we + * use a pointer back to the same weak_pointer. */ + struct weak_pointer * wp = (struct weak_pointer*)where; - /* Check whether it's already in the list. */ - while (wp != NULL) { - if (wp == (struct weak_pointer*)where) { - break; - } - wp = wp->next; - } - if (wp == NULL) { - /* Add it to the start of the list. */ - wp = (struct weak_pointer*)where; - if (wp->next != weak_pointers) { - wp->next = weak_pointers; - } else { - /*SHOW("avoided write to weak pointer");*/ - } + if (NULL == wp->next) { + wp->next = weak_pointers; weak_pointers = wp; + if (NULL == wp->next) + wp->next = wp; } /* Do not let GC scavenge the value slot of the weak pointer. @@ -1995,170 +2266,175 @@ search_static_space(void *pointer) lispobj * search_dynamic_space(void *pointer) { - long page_index = find_page_index(pointer); + page_index_t page_index = find_page_index(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 *)((void *)page_address(page_index) - + page_table[page_index].first_object_offset); + start = (lispobj *)page_region_start(page_index); return (gc_search_space(start, (((lispobj *)pointer)+2)-start, (lispobj *)pointer)); } -/* 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? - * This is called from preserve_pointers() */ +/* Helper for valid_lisp_pointer_p and + * possibly_valid_dynamic_space_pointer. + * + * pointer is the pointer to validate, and start_addr is the address + * of the enclosing object. + */ static int -possibly_valid_dynamic_space_pointer(lispobj *pointer) +looks_like_valid_lisp_pointer_p(lispobj *pointer, lispobj *start_addr) { - lispobj *start_addr; - - /* Find the object start address. */ - if ((start_addr = search_dynamic_space(pointer)) == NULL) { - return 0; - } - - /* 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 it's not a return address then it needs to be a valid Lisp - * pointer. */ if (!is_lisp_pointer((lispobj)pointer)) { return 0; } /* Check that the object pointed to is consistent with the pointer - * low tag. - */ + * low tag. */ switch (lowtag_of((lispobj)pointer)) { case FUN_POINTER_LOWTAG: /* Start_addr should be the enclosing code object, or a closure * 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)pointer != - ((unsigned)start_addr+FUN_POINTER_LOWTAG)) { - if (gencgc_verbose) + if ((unsigned long)pointer != + ((unsigned long)start_addr+FUN_POINTER_LOWTAG)) { + 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)pointer != - ((unsigned)start_addr+LIST_POINTER_LOWTAG)) { - if (gencgc_verbose) + if ((unsigned long)pointer != + ((unsigned long)start_addr+LIST_POINTER_LOWTAG)) { + if (gencgc_verbose) { FSHOW((stderr, "/Wl1: %x %x %x\n", pointer, start_addr, *start_addr)); + } return 0; } /* Is it plausible cons? */ - if ((is_lisp_pointer(start_addr[0]) - || (fixnump(start_addr[0])) - || (widetag_of(start_addr[0]) == CHARACTER_WIDETAG) -#if N_WORD_BITS == 64 - || (widetag_of(start_addr[0]) == SINGLE_FLOAT_WIDETAG) -#endif - || (widetag_of(start_addr[0]) == UNBOUND_MARKER_WIDETAG)) - && (is_lisp_pointer(start_addr[1]) - || (fixnump(start_addr[1])) - || (widetag_of(start_addr[1]) == CHARACTER_WIDETAG) -#if N_WORD_BITS == 64 - || (widetag_of(start_addr[1]) == SINGLE_FLOAT_WIDETAG) -#endif - || (widetag_of(start_addr[1]) == UNBOUND_MARKER_WIDETAG))) + if ((is_lisp_pointer(start_addr[0]) || + is_lisp_immediate(start_addr[0])) && + (is_lisp_pointer(start_addr[1]) || + 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)pointer != - ((unsigned)start_addr+INSTANCE_POINTER_LOWTAG)) { - if (gencgc_verbose) + if ((unsigned long)pointer != + ((unsigned long)start_addr+INSTANCE_POINTER_LOWTAG)) { + 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 ((unsigned)pointer != - ((int)start_addr+OTHER_POINTER_LOWTAG)) { - if (gencgc_verbose) + +#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) { 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])) { case UNBOUND_MARKER_WIDETAG: + case NO_TLS_VALUE_MARKER_WIDETAG: case CHARACTER_WIDETAG: #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 */ @@ -2255,28 +2531,78 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer) #endif case SAP_WIDETAG: case WEAK_POINTER_WIDETAG: +#ifdef LUTEX_WIDETAG + case LUTEX_WIDETAG: +#endif 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; + } + + /* looks good */ + return 1; +} + +/* Used by the debugger to validate possibly bogus pointers before + * calling MAKE-LISP-OBJ on them. + * + * FIXME: We would like to make this perfect, because if the debugger + * constructs a reference to a bugs lisp object, and it ends up in a + * location scavenged by the GC all hell breaks loose. + * + * Whereas possibly_valid_dynamic_space_pointer has to be conservative + * and return true for all valid pointers, this could actually be eager + * and lie about a few pointers without bad results... but that should + * be reflected in the name. + */ +int +valid_lisp_pointer_p(lispobj *pointer) +{ + lispobj *start; + if (((start=search_dynamic_space(pointer))!=NULL) || + ((start=search_static_space(pointer))!=NULL) || + ((start=search_read_only_space(pointer))!=NULL)) + return looks_like_valid_lisp_pointer_p(pointer, start); + else + 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? + * This is called from preserve_pointers() */ +static int +possibly_valid_dynamic_space_pointer(lispobj *pointer) +{ + lispobj *start_addr; + + /* Find the object start address. */ + if ((start_addr = search_dynamic_space(pointer)) == NULL) { return 0; } - /* looks good */ - return 1; + 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 @@ -2287,13 +2613,13 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer) static void maybe_adjust_large_object(lispobj *where) { - long first_page; + page_index_t first_page; + page_index_t next_page; long nwords; - long remaining_bytes; - long next_page; - long bytes_freed; - long old_bytes_used; + unsigned long remaining_bytes; + unsigned long bytes_freed; + unsigned long old_bytes_used; int boxed; @@ -2379,25 +2705,24 @@ maybe_adjust_large_object(lispobj *where) * but lets do it for them all (they'll probably be written * anyway?). */ - gc_assert(page_table[first_page].first_object_offset == 0); + gc_assert(page_table[first_page].region_start_offset == 0); 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_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].first_object_offset == - -PAGE_BYTES*(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == PAGE_BYTES); + gc_assert(page_table[next_page].region_start_offset == + npage_bytes(next_page-first_page)); + 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++; } @@ -2419,13 +2744,12 @@ 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_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].first_object_offset == - -(next_page - first_page)*PAGE_BYTES)) { + (page_table[next_page].region_start_offset == + npage_bytes(next_page - first_page))) { /* It checks out OK, free the page. We don't need to both zeroing * pages as this should have been done before shrinking the * object. These pages shouldn't be write protected as they @@ -2463,17 +2787,18 @@ maybe_adjust_large_object(lispobj *where) * * It is also assumed that the current gc_alloc() region has been * flushed and the tables updated. */ + static void preserve_pointer(void *addr) { - long addr_page_index = find_page_index(addr); - long first_page; - long i; - unsigned region_allocation; + page_index_t addr_page_index = find_page_index(addr); + page_index_t first_page; + page_index_t i; + unsigned int region_allocation; /* 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. */ @@ -2487,7 +2812,8 @@ preserve_pointer(void *addr) /* quick check 2: Check the offset within the page. * */ - if (((unsigned)addr & (PAGE_BYTES - 1)) > page_table[addr_page_index].bytes_used) + if (((unsigned long)addr & (GENCGC_CARD_BYTES - 1)) > + page_table[addr_page_index].bytes_used) return; /* Filter out anything which can't be a pointer to a Lisp object @@ -2495,9 +2821,17 @@ 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. */ - if (!(possibly_valid_dynamic_space_pointer(addr))) + * 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 @@ -2507,15 +2841,13 @@ preserve_pointer(void *addr) #if 0 /* I think this'd work just as well, but without the assertions. * -dan 2004.01.01 */ - first_page= - find_page_index(page_address(addr_page_index)+ - page_table[addr_page_index].first_object_offset); + first_page = find_page_index(page_region_start(addr_page_index)) #else first_page = addr_page_index; - while (page_table[first_page].first_object_offset != 0) { + 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); } @@ -2529,10 +2861,10 @@ 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)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", @@ -2566,12 +2898,12 @@ 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 */ - || (page_table[i+1].allocated == FREE_PAGE_FLAG) + 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 */ - || (page_table[i+1].first_object_offset == 0)) + || (page_table[i+1].region_start_offset == 0)) break; } @@ -2593,22 +2925,23 @@ preserve_pointer(void *addr) * * We return 1 if the page was write-protected, else 0. */ static int -update_page_write_prot(long page) +update_page_write_prot(page_index_t page) { - int gen = page_table[page].gen; + generation_index_t gen = page_table[page].gen; long j; int wp_it = 1; void **page_addr = (void **)page_address(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 @@ -2616,15 +2949,15 @@ update_page_write_prot(long page) for (j = 0; j < num_words; j++) { void *ptr = *(page_addr+j); - long index = find_page_index(ptr); + page_index_t index = find_page_index(ptr); /* 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 == NUM_GENERATIONS))) + || (page_table[index].gen == SCRATCH_GENERATION))) /* Or does it point within a current gc_alloc() region? */ || ((boxed_region.start_addr <= ptr) @@ -2641,7 +2974,7 @@ update_page_write_prot(long 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. */ @@ -2651,11 +2984,9 @@ update_page_write_prot(long page) return (wp_it); } -/* Scavenge a generation. - * - * This will not resolve all pointers when generation is the new - * space, as new objects may be added which are not checked here - use - * scavenge_newspace generation. +/* Scavenge all generations from FROM to TO, inclusive, except for + * new_space which needs special handling, as new objects may be + * added which are not checked here - use scavenge_newspace generation. * * Write-protected pages should not have any pointers to the * from_space so do need scavenging; thus write-protected pages are @@ -2683,44 +3014,48 @@ update_page_write_prot(long page) * pointers as the objects contain a link to the next and are written * if a weak pointer is scavenged. Still it's a useful check. */ static void -scavenge_generation(int generation) +scavenge_generations(generation_index_t from, generation_index_t to) { - long i; + page_index_t i; int num_wp = 0; #define SC_GEN_CK 0 #if SC_GEN_CK /* Clear the write_protected_cleared flags on all pages. */ - for (i = 0; i < NUM_PAGES; i++) + for (i = 0; i < page_table_pages; i++) page_table[i].write_protected_cleared = 0; #endif for (i = 0; i < last_free_page; i++) { - if ((page_table[i].allocated & BOXED_PAGE_FLAG) + generation_index_t generation = page_table[i].gen; + if (page_boxed_p(i) && (page_table[i].bytes_used != 0) - && (page_table[i].gen == generation)) { - long last_page,j; + && (generation != new_space) + && (generation >= from) + && (generation <= to)) { + page_index_t last_page,j; int write_protected=1; /* This should be the start of a region */ - gc_assert(page_table[i].first_object_offset == 0); + gc_assert(page_table[i].region_start_offset == 0); /* Now work forward until the end of the region */ 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 */ - || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG)) + 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) - || (page_table[last_page+1].first_object_offset == 0)) + || (page_table[last_page+1].region_start_offset == 0)) break; } if (!write_protected) { scavenge(page_address(i), - (page_table[last_page].bytes_used + - (last_page-i)*PAGE_BYTES)/N_WORD_BYTES); + ((unsigned long)(page_table[last_page].bytes_used + + npage_bytes(last_page-i))) + /N_WORD_BYTES); /* Now scan the pages and write protect those that * don't have pointers to younger generations. */ @@ -2729,31 +3064,31 @@ scavenge_generation(int generation) num_wp += update_page_write_prot(j); } } + if ((gencgc_verbose > 1) && (num_wp != 0)) { + FSHOW((stderr, + "/write protected %d pages within generation %d\n", + num_wp, generation)); + } } i = last_page; } } - if ((gencgc_verbose > 1) && (num_wp != 0)) { - FSHOW((stderr, - "/write protected %d pages within generation %d\n", - num_wp, generation)); - } #if SC_GEN_CK /* Check that none of the write_protected pages in this generation * have been written to. */ - for (i = 0; i < NUM_PAGES; i++) { - if ((page_table[i].allocation != FREE_PAGE_FLAG) + for (i = 0; i < page_table_pages; i++) { + if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation) && (page_table[i].write_protected_cleared != 0)) { FSHOW((stderr, "/scavenge_generation() %d\n", generation)); FSHOW((stderr, - "/page bytes_used=%d first_object_offset=%d dont_move=%d\n", + "/page bytes_used=%d region_start_offset=%lu dont_move=%d\n", page_table[i].bytes_used, - page_table[i].first_object_offset, + page_table[i].region_start_offset, page_table[i].dont_move)); - lose("write to protected page %d in scavenge_generation()", i); + lose("write to protected page %d in scavenge_generation()\n", i); } } #endif @@ -2788,26 +3123,27 @@ static struct new_area new_areas_2[NUM_NEW_AREAS]; * complete the job as new objects may be added to the generation in * the process which are not scavenged. */ static void -scavenge_newspace_generation_one_scan(int generation) +scavenge_newspace_generation_one_scan(generation_index_t generation) { - long i; + page_index_t i; FSHOW((stderr, "/starting one full scan of newspace generation %d\n", 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) /* (This may be redundant as write_protected is now * cleared before promotion.) */ || (page_table[i].dont_move == 1))) { - long last_page; + page_index_t last_page; int all_wp=1; - /* The scavenge will start at the first_object_offset of page i. + /* The scavenge will start at the region_start_offset of + * page i. * * We need to find the full extent of this contiguous * block in case objects span pages. @@ -2823,27 +3159,25 @@ scavenge_newspace_generation_one_scan(int 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 */ - || (!(page_table[last_page+1].allocated & BOXED_PAGE_FLAG)) + 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) - || (page_table[last_page+1].first_object_offset == 0)) + || (page_table[last_page+1].region_start_offset == 0)) break; } /* Do a limited check for write-protected pages. */ if (!all_wp) { - long size; - - size = (page_table[last_page].bytes_used - + (last_page-i)*PAGE_BYTES - - page_table[i].first_object_offset)/N_WORD_BYTES; + long nwords = (((unsigned long) + (page_table[last_page].bytes_used + + npage_bytes(last_page-i) + + page_table[i].region_start_offset)) + / N_WORD_BYTES); new_areas_ignore_page = last_page; - scavenge(page_address(i) + - page_table[i].first_object_offset, - size); + scavenge(page_region_start(i), nwords); } i = last_page; @@ -2856,7 +3190,7 @@ scavenge_newspace_generation_one_scan(int generation) /* Do a complete scavenge of the newspace generation. */ static void -scavenge_newspace_generation(int generation) +scavenge_newspace_generation(generation_index_t generation) { long i; @@ -2885,6 +3219,13 @@ scavenge_newspace_generation(int generation) /* Record all new areas now. */ record_new_objects = 2; + /* Give a chance to weak hash tables to make other objects live. + * FIXME: The algorithm implemented here for weak hash table gcing + * is O(W^2+N) as Bruno Haible warns in + * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html + * see "Implementation 2". */ + scav_weak_hash_tables(); + /* Flush the current regions updating the tables. */ gc_alloc_update_all_page_tables(); @@ -2920,11 +3261,12 @@ scavenge_newspace_generation(int 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 scavenge anyway - * during scavenge_newspace_generation_one_scan. */ + /* Don't need to record new areas that get scavenged + * anyway during scavenge_newspace_generation_one_scan. */ record_new_objects = 1; scavenge_newspace_generation_one_scan(generation); @@ -2932,6 +3274,8 @@ scavenge_newspace_generation(int generation) /* Record all new areas now. */ record_new_objects = 2; + scav_weak_hash_tables(); + /* Flush the current regions updating the tables. */ gc_alloc_update_all_page_tables(); @@ -2939,13 +3283,15 @@ scavenge_newspace_generation(int generation) /* Work through previous_new_areas. */ for (i = 0; i < previous_new_areas_index; i++) { - long page = (*previous_new_areas)[i].page; - long offset = (*previous_new_areas)[i].offset; - long size = (*previous_new_areas)[i].size / N_WORD_BYTES; + page_index_t page = (*previous_new_areas)[i].page; + size_t offset = (*previous_new_areas)[i].offset; + size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES; gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0); scavenge(page_address(page)+offset, size); } + scav_weak_hash_tables(); + /* Flush the current regions updating the tables. */ gc_alloc_update_all_page_tables(); } @@ -2963,13 +3309,13 @@ scavenge_newspace_generation(int generation) #if SC_NS_GEN_CK /* Check that none of the write_protected pages in this generation * have been written to. */ - for (i = 0; i < NUM_PAGES; i++) { - if ((page_table[i].allocation != FREE_PAGE_FLAG) + for (i = 0; i < page_table_pages; i++) { + if (page_allocated_p(i) && (page_table[i].bytes_used != 0) && (page_table[i].gen == generation) && (page_table[i].write_protected_cleared != 0) && (page_table[i].dont_move == 0)) { - lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d", + lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n", i, generation, page_table[i].dont_move); } } @@ -2984,42 +3330,60 @@ scavenge_newspace_generation(int generation) static void unprotect_oldspace(void) { - long i; + 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 = 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 * assumes that all objects have been copied or promoted to an older * generation. Bytes_allocated and the generation bytes_allocated * counter are updated. The number of bytes freed is returned. */ -static long +static unsigned long free_oldspace(void) { - long bytes_freed = 0; - long first_page, last_page; + unsigned long bytes_freed = 0; + page_index_t first_page, last_page; first_page = 0; 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++; @@ -3037,49 +3401,21 @@ 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)); - /* Zero pages from first_page to (last_page-1). - * - * FIXME: Why not use os_zero(..) function instead of - * hand-coding this again? (Check other gencgc_unmap_zero - * stuff too. */ - if (gencgc_unmap_zero) { - void *page_start, *addr; - - page_start = (void *)page_address(first_page); - - os_invalidate(page_start, PAGE_BYTES*(last_page-first_page)); - addr = os_validate(page_start, PAGE_BYTES*(last_page-first_page)); - if (addr == NULL || addr != page_start) { - lose("free_oldspace: page moved, 0x%08x ==> 0x%08x",page_start, - addr); - } - } else { - long *page_start; - - page_start = (long *)page_address(first_page); - memset(page_start, 0,PAGE_BYTES*(last_page-first_page)); - } - +#ifdef READ_PROTECT_FREE_PAGES + os_protect(page_address(first_page), + npage_bytes(last_page-first_page), + OS_VM_PROT_NONE); +#endif first_page = last_page; - } while (first_page < last_free_page); bytes_allocated -= bytes_freed; @@ -3092,16 +3428,16 @@ static void print_ptr(lispobj *addr) { /* If addr is in the dynamic space then out the page information. */ - long pi1 = find_page_index((void*)addr); + page_index_t pi1 = find_page_index((void*)addr); if (pi1 != -1) - fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %d dont_move %d\n", + fprintf(stderr," %x: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n", (unsigned long) addr, pi1, page_table[pi1].allocated, page_table[pi1].gen, page_table[pi1].bytes_used, - page_table[pi1].first_object_offset, + page_table[pi1].region_start_offset, page_table[pi1].dont_move); fprintf(stderr," %x %x %x %x (%x) %x %x %x %x\n", *(addr-4), @@ -3116,22 +3452,37 @@ print_ptr(lispobj *addr) } #endif -extern long undefined_tramp; +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) { int is_in_dynamic_space = (find_page_index((void*)start) != -1); int is_in_readonly_space = - (READ_ONLY_SPACE_START <= (unsigned)start && - (unsigned)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)); + (READ_ONLY_SPACE_START <= (unsigned long)start && + (unsigned long)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)); while (words > 0) { size_t count = 1; lispobj thing = *(lispobj*)start; if (is_lisp_pointer(thing)) { - long page_index = find_page_index((void*)thing); + page_index_t page_index = find_page_index((void*)thing); long to_readonly_space = (READ_ONLY_SPACE_START <= thing && thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)); @@ -3143,17 +3494,17 @@ 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.", 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.", 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", + lose("ptr to dynamic space %p from RO space %x\n", thing, start); } /* Does it point to a plausible object? This check slows @@ -3167,14 +3518,16 @@ verify_space(lispobj *start, size_t words) * dynamically. */ /* if (!possibly_valid_dynamic_space_pointer((lispobj *)thing)) { - lose("ptr %x to invalid object %x", 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 - && (thing != (unsigned)&undefined_tramp)) { - lose("Ptr %x @ %x sees junk.", thing, start); + && (thing != (lispobj)&funcallable_instance_tramp) + && !is_in_stack_space(thing)) { + lose("Ptr %p @ %p sees junk.\n", thing, start); } } } else { @@ -3204,11 +3557,26 @@ verify_space(lispobj *start, size_t words) case SINGLE_FLOAT_WIDETAG: #endif case UNBOUND_MARKER_WIDETAG: - case INSTANCE_HEADER_WIDETAG: case FDEFN_WIDETAG: count = 1; break; + case INSTANCE_HEADER_WIDETAG: + { + lispobj nuntagged; + long ntotal = HeaderValue(thing); + lispobj layout = ((struct instance *)start)->slots[0]; + if (!layout) { + count = 1; + break; + } + nuntagged = ((struct layout *) + native_pointer(layout))->n_untagged_slots; + verify_space(start + 1, + ntotal - fixnum_value(nuntagged)); + count = ntotal + 1; + break; + } case CODE_HEADER_WIDETAG: { lispobj object = *start; @@ -3236,7 +3604,7 @@ 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)); } @@ -3253,7 +3621,8 @@ verify_space(lispobj *start, size_t words) while (fheaderl != NIL) { fheaderp = (struct simple_fun *) native_pointer(fheaderl); - gc_assert(widetag_of(fheaderp->header) == SIMPLE_FUN_HEADER_WIDETAG); + gc_assert(widetag_of(fheaderp->header) == + SIMPLE_FUN_HEADER_WIDETAG); verify_space(&fheaderp->name, 1); verify_space(&fheaderp->arglist, 1); verify_space(&fheaderp->type, 1); @@ -3341,11 +3710,18 @@ verify_space(lispobj *start, size_t words) #endif case SAP_WIDETAG: case WEAK_POINTER_WIDETAG: +#ifdef LUTEX_WIDETAG + case LUTEX_WIDETAG: +#endif +#ifdef NO_TLS_VALUE_MARKER_WIDETAG + case NO_TLS_VALUE_MARKER_WIDETAG: +#endif count = (sizetab[widetag_of(*start)])(start); break; default: - gc_abort(); + lose("Unhandled widetag %p at %p\n", + widetag_of(*start), start); } } } @@ -3372,7 +3748,7 @@ verify_gc(void) struct thread *th; for_each_thread(th) { long binding_stack_size = - (lispobj*)SymbolValue(BINDING_STACK_POINTER,th) + (lispobj*)get_binding_stack_pointer(th) - (lispobj*)th->binding_stack_start; verify_space(th->binding_stack_start, binding_stack_size); } @@ -3381,19 +3757,19 @@ verify_gc(void) } static void -verify_generation(int generation) +verify_generation(generation_index_t generation) { - int i; + 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)) { - long last_page; + page_index_t last_page; int region_allocation = page_table[i].allocated; /* This should be the start of a contiguous block */ - gc_assert(page_table[i].first_object_offset == 0); + gc_assert(page_table[i].region_start_offset == 0); /* Need to find the full extent of this contiguous block in case objects span pages. */ @@ -3403,16 +3779,19 @@ verify_generation(int 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) - || (page_table[last_page+1].first_object_offset == 0)) + || (page_table[last_page+1].region_start_offset == 0)) break; - verify_space(page_address(i), (page_table[last_page].bytes_used - + (last_page-i)*PAGE_BYTES)/N_WORD_BYTES); + verify_space(page_address(i), + ((unsigned long) + (page_table[last_page].bytes_used + + npage_bytes(last_page-i))) + / N_WORD_BYTES); i = last_page; } } @@ -3422,29 +3801,29 @@ verify_generation(int generation) static void verify_zero_fill(void) { - long page; + 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; long i; for (i = 0; i < size; i++) { if (start_addr[i] != 0) { - lose("free page not zero at %x", start_addr + i); + lose("free page not zero at %x\n", start_addr + i); } } } 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)page_address(page) + long *start_addr = (long *)((unsigned long)page_address(page) + page_table[page].bytes_used); long size = free_bytes / N_WORD_BYTES; long i; for (i = 0; i < size; i++) { if (start_addr[i] != 0) { - lose("free region not zero at %x", start_addr + i); + lose("free region not zero at %x\n", start_addr + i); } } } @@ -3465,9 +3844,9 @@ gencgc_verify_zero_fill(void) static void verify_dynamic_space(void) { - long i; + generation_index_t i; - for (i = 0; i < NUM_GENERATIONS; i++) + for (i = 0; i <= HIGHEST_NORMAL_GENERATION; i++) verify_generation(i); if (gencgc_enable_verify_zero_fill) @@ -3476,28 +3855,35 @@ verify_dynamic_space(void) /* Write-protect all the dynamic boxed pages in the given generation. */ static void -write_protect_generation_pages(int generation) +write_protect_generation_pages(generation_index_t generation) { - long i; + page_index_t start; - gc_assert(generation < NUM_GENERATIONS); + gc_assert(generation < SCRATCH_GENERATION); - for (i = 0; i < last_free_page; i++) - if ((page_table[i].allocated == BOXED_PAGE_FLAG) - && (page_table[i].bytes_used != 0) - && !page_table[i].dont_move - && (page_table[i].gen == generation)) { + for (start = 0; start < last_free_page; start++) { + if (protect_page_p(start, generation)) { void *page_start; + page_index_t last; + + /* Note the page as protected in the page tables. */ + page_table[start].write_protected = 1; + + for (last = start + 1; last < last_free_page; last++) { + if (!protect_page_p(last, generation)) + break; + page_table[last].write_protected = 1; + } - page_start = (void *)page_address(i); + page_start = (void *)page_address(start); os_protect(page_start, - PAGE_BYTES, + npage_bytes(last - start), OS_VM_PROT_READ | OS_VM_PROT_EXECUTE); - /* Note the page as protected in the page tables. */ - page_table[i].write_protected = 1; + start = last; } + } if (gencgc_verbose > 1) { FSHOW((stderr, @@ -3508,29 +3894,93 @@ write_protect_generation_pages(int generation) } } +#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) +static void +scavenge_control_stack(struct thread *th) +{ + lispobj *control_stack = + (lispobj *)(th->control_stack_start); + unsigned long control_stack_size = + access_control_stack_pointer(th) - control_stack; + + scavenge(control_stack, control_stack_size); +} +#endif + +#if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)) +static void +preserve_context_registers (os_context_t *c) +{ + void **ptr; + /* On Darwin the signal context isn't a contiguous block of memory, + * so just preserve_pointering its contents won't be sufficient. + */ +#if defined(LISP_FEATURE_DARWIN) +#if defined LISP_FEATURE_X86 + preserve_pointer((void*)*os_context_register_addr(c,reg_EAX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_ECX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_EDX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_EBX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_ESI)); + preserve_pointer((void*)*os_context_register_addr(c,reg_EDI)); + preserve_pointer((void*)*os_context_pc_addr(c)); +#elif defined LISP_FEATURE_X86_64 + preserve_pointer((void*)*os_context_register_addr(c,reg_RAX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_RCX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_RDX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_RBX)); + preserve_pointer((void*)*os_context_register_addr(c,reg_RSI)); + preserve_pointer((void*)*os_context_register_addr(c,reg_RDI)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R8)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R9)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R10)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R11)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R12)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R13)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R14)); + preserve_pointer((void*)*os_context_register_addr(c,reg_R15)); + preserve_pointer((void*)*os_context_pc_addr(c)); +#else + #error "preserve_context_registers needs to be tweaked for non-x86 Darwin" +#endif +#endif + for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) { + preserve_pointer(*ptr); + } +} +#endif + /* Garbage collect a generation. If raise is 0 then the remains of the * generation are not raised to the next generation. */ static void -garbage_collect_generation(int generation, int raise) +garbage_collect_generation(generation_index_t generation, int raise) { unsigned long bytes_freed; - unsigned long i; + page_index_t i; unsigned long static_space_size; struct thread *th; - gc_assert(generation <= (NUM_GENERATIONS-1)); + + gc_assert(generation <= HIGHEST_NORMAL_GENERATION); /* The oldest generation can't be raised. */ - gc_assert((generation != (NUM_GENERATIONS-1)) || (raise == 0)); + gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0)); + + /* Check if weak hash tables were processed in the previous GC. */ + gc_assert(weak_hash_tables == NULL); /* Initialize the weak pointer list. */ weak_pointers = NULL; +#ifdef LUTEX_WIDETAG + unmark_lutexes(generation); +#endif + /* When a generation is not being raised it is transported to a * temporary generation (NUM_GENERATIONS), and lowered when * done. Set up this new generation. There should be no pages * allocated to it yet. */ if (!raise) { - gc_assert(generations[NUM_GENERATIONS].bytes_allocated == 0); + gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0); } /* Set the global src and dest. generations */ @@ -3538,7 +3988,7 @@ garbage_collect_generation(int generation, int raise) if (raise) new_space = generation+1; else - new_space = NUM_GENERATIONS; + new_space = SCRATCH_GENERATION; /* Change to a new space for allocation, resetting the alloc_start_page */ gc_alloc_generation = new_space; @@ -3575,74 +4025,107 @@ garbage_collect_generation(int generation, int raise) /* we assume that none of the preceding applies to the thread that * initiates GC. If you ever call GC from inside an altstack * handler, you will lose. */ - for_each_thread(th) { - void **ptr; - void **esp=(void **)-1; + +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + /* And if we're saving a core, there's no point in being conservative. */ + if (conservative_stack) { + for_each_thread(th) { + void **ptr; + void **esp=(void **)-1; #ifdef LISP_FEATURE_SB_THREAD - long i,free; - if(th==arch_os_get_current_thread()) { - /* Somebody is going to burn in hell for this, but casting - * it in two steps shuts gcc up about strict aliasing. */ - esp = (void **)((void *)&raise); - } else { - void **esp1; - free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th)); - for(i=free-1;i>=0;i--) { - os_context_t *c=th->interrupt_contexts[i]; - esp1 = (void **) *os_context_register_addr(c,reg_SP); - if (esp1>=(void **)th->control_stack_start && - esp1<(void **)th->control_stack_end) { - if(esp1=(void **)c; ptr--) { - preserve_pointer(*ptr); + long i,free; + if(th==arch_os_get_current_thread()) { + /* Somebody is going to burn in hell for this, but casting + * it in two steps shuts gcc up about strict aliasing. */ + esp = (void **)((void *)&raise); + } else { + void **esp1; + free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th)); + for(i=free-1;i>=0;i--) { + os_context_t *c=th->interrupt_contexts[i]; + esp1 = (void **) *os_context_register_addr(c,reg_SP); + if (esp1>=(void **)th->control_stack_start && + esp1<(void **)th->control_stack_end) { + if(esp1control_stack_end; ptr > esp; ptr--) { - preserve_pointer(*ptr); + for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp; ptr--) { + preserve_pointer(*ptr); + } + } + } +#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, "/non-movable pages due to conservative pointers = %d (%d bytes)\n", num_dont_move_pages, - num_dont_move_pages * PAGE_BYTES); + npage_bytes(num_dont_move_pages)); } #endif /* Scavenge all the rest of the roots. */ +#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) + /* + * If not x86, we need to scavenge the interrupt context(s) and the + * 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 * care to avoid SIG_DFL and SIG_IGN. */ - for_each_thread(th) { - struct interrupt_data *data=th->interrupt_data; for (i = 0; i < NSIG; i++) { - union interrupt_handler handler = data->interrupt_handlers[i]; + union interrupt_handler handler = interrupt_handlers[i]; if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) && !ARE_SAME_HANDLER(handler.c, SIG_DFL)) { - scavenge((lispobj *)(data->interrupt_handlers + i), 1); - } + scavenge((lispobj *)(interrupt_handlers + i), 1); } } /* Scavenge the binding stacks. */ - { - struct thread *th; - for_each_thread(th) { - long len= (lispobj *)SymbolValue(BINDING_STACK_POINTER,th) - - th->binding_stack_start; - scavenge((lispobj *) th->binding_stack_start,len); + { + struct thread *th; + for_each_thread(th) { + long len= (lispobj *)get_binding_stack_pointer(th) - + th->binding_stack_start; + scavenge((lispobj *) th->binding_stack_start,len); #ifdef LISP_FEATURE_SB_THREAD - /* do the tls as well */ - len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) - - (sizeof (struct thread))/(sizeof (lispobj)); - scavenge((lispobj *) (th+1),len); + /* do the tls as well */ + len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) - + (sizeof (struct thread))/(sizeof (lispobj)); + scavenge((lispobj *) (th+1),len); #endif } } @@ -3680,11 +4163,7 @@ garbage_collect_generation(int generation, int raise) /* All generations but the generation being GCed need to be * scavenged. The new_space generation needs special handling as * objects may be moved in - it is handled separately below. */ - for (i = 0; i < NUM_GENERATIONS; i++) { - if ((i != generation) && (i != new_space)) { - scavenge_generation(i); - } - } + scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION); /* Finally scavenge the new_space generation. Keep going until no * more objects are moved into the new generation */ @@ -3713,12 +4192,13 @@ garbage_collect_generation(int generation, int raise) bytes_allocated = bytes_allocated - old_bytes_allocated; if (bytes_allocated != 0) { - lose("Rescan of new_space allocated %d more bytes.", + lose("Rescan of new_space allocated %d more bytes.\n", bytes_allocated); } } #endif + scan_weak_hash_tables(); scan_weak_pointers(); /* Flush the current regions, updating the tables. */ @@ -3732,12 +4212,12 @@ garbage_collect_generation(int generation, int raise) if (!raise) { for (i = 0; i < last_free_page; i++) if ((page_table[i].bytes_used != 0) - && (page_table[i].gen == NUM_GENERATIONS)) + && (page_table[i].gen == SCRATCH_GENERATION)) page_table[i].gen = generation; gc_assert(generations[generation].bytes_allocated == 0); generations[generation].bytes_allocated = - generations[NUM_GENERATIONS].bytes_allocated; - generations[NUM_GENERATIONS].bytes_allocated = 0; + generations[SCRATCH_GENERATION].bytes_allocated; + generations[SCRATCH_GENERATION].bytes_allocated = 0; } /* Reset the alloc_start_page for generation. */ @@ -3747,8 +4227,9 @@ garbage_collect_generation(int 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(); } @@ -3762,27 +4243,98 @@ garbage_collect_generation(int generation, int raise) generations[generation].num_gc = 0; else ++generations[generation].num_gc; + +#ifdef LUTEX_WIDETAG + reap_lutexes(generation); + if (raise) + move_lutexes(generation, generation+1); +#endif } /* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */ long -update_x86_dynamic_space_free_pointer(void) +update_dynamic_space_free_pointer(void) { - long last_page = -1; - long i; + 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; - SetSymbolValue(ALLOCATION_POINTER, - (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0); + set_alloc_pointer((lispobj)(page_address(last_free_page))); return 0; /* dummy value: return something ... */ } +static void +remap_page_range (page_index_t from, page_index_t to, int forcibly) +{ + /* There's a mysterious Solaris/x86 problem with using mmap + * tricks for memory zeroing. See sbcl-devel thread + * "Re: patch: standalone executable redux". + * + * Since pages don't have to be zeroed ahead of time, only do + * so when called from purify. + */ +#if defined(LISP_FEATURE_SUNOS) + if (forcibly) + zero_pages(from, to); +#else + page_index_t aligned_from, aligned_end, end = to+1; + + 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 (forcibly) { + if (aligned_from != from) + zero_pages(from, aligned_from-1); + if (aligned_end != end) + zero_pages(aligned_end, end-1); + } + } else if (forcibly) + zero_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, 1); + + /* See comment above about mysterious failures on Solaris/x86. + */ +#if !defined(LISP_FEATURE_SUNOS) + for (first_page = from; first_page <= to; first_page++) { + if (page_allocated_p(first_page) || + (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++; + + remap_page_range(first_page, last_page-1, 0); + + first_page = last_page; + } +#endif +} + +generation_index_t small_generation_limit = 1; + /* GC all generations newer than last_gen, raising the objects in each * to the next older generation - we finish when all generations below * last_gen are empty. Then if last_gen is due for a GC, or if @@ -3791,18 +4343,22 @@ update_x86_dynamic_space_free_pointer(void) * * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */ - void -collect_garbage(unsigned last_gen) +collect_garbage(generation_index_t last_gen) { - int gen = 0; + generation_index_t gen = 0, i; int raise; int gen_to_wp; - long i; + /* The largest value of last_free_page seen since the time + * remap_free_pages was called. */ + static page_index_t high_water_mark = 0; FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen)); + log_generation_stats(gc_logfile, "=== GC Start ==="); - if (last_gen > NUM_GENERATIONS) { + gc_active_p = 1; + + if (last_gen > HIGHEST_NORMAL_GENERATION+1) { FSHOW((stderr, "/collect_garbage: last_gen = %d, doing a level 0 GC\n", last_gen)); @@ -3819,7 +4375,7 @@ collect_garbage(unsigned last_gen) } if (gencgc_verbose > 1) - print_generation_stats(0); + print_generation_stats(); do { /* Collect the generation. */ @@ -3830,7 +4386,7 @@ collect_garbage(unsigned 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) { @@ -3857,7 +4413,7 @@ collect_garbage(unsigned last_gen) if (gencgc_verbose > 1) { FSHOW((stderr, "GC of generation %d finished:\n", gen)); - print_generation_stats(0); + print_generation_stats(); } gen++; @@ -3867,8 +4423,8 @@ collect_garbage(unsigned 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. @@ -3889,7 +4445,7 @@ collect_garbage(unsigned last_gen) /* Check that they are all empty. */ for (i = 0; i < gen_to_wp; i++) { if (generations[i].bytes_allocated) - lose("trying to write-protect gen. %d when gen. %d nonempty", + lose("trying to write-protect gen. %d when gen. %d nonempty\n", gen_to_wp, i); } write_protect_generation_pages(gen_to_wp); @@ -3900,11 +4456,30 @@ collect_garbage(unsigned last_gen) gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0); gc_alloc_generation = 0; - update_x86_dynamic_space_free_pointer(); + /* Save the high-water mark before updating last_free_page */ + if (last_free_page > high_water_mark) + high_water_mark = last_free_page; + + update_dynamic_space_free_pointer(); + auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs; if(gencgc_verbose) fprintf(stderr,"Next gc when %ld bytes have been consed\n", auto_gc_trigger); + + /* If we did a big GC (arbitrarily defined as gen > 1), release memory + * back to the OS. + */ + 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, 0); + high_water_mark = 0; + } + + gc_active_p = 0; + + log_generation_stats(gc_logfile, "=== GC End ==="); SHOW("returning from collect_garbage"); } @@ -3916,47 +4491,46 @@ collect_garbage(unsigned last_gen) void gc_free_heap(void) { - long page; + page_index_t page, last_page; - if (gencgc_verbose > 1) + if (gencgc_verbose > 1) { SHOW("entering gc_free_heap"); + } - for (page = 0; page < NUM_PAGES; page++) { + 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; + 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; + } - /* 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; - - /* Zero the page. */ +#ifndef LISP_FEATURE_WIN32 /* Pages already zeroed on win32? Not sure + * about this change. */ 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", - page_start, - addr); - } + 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. */ - long *page_start, i; - gc_assert(page_table[page].allocated == FREE_PAGE_FLAG); + long *page_start; + page_index_t i; + 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; @@ -3986,12 +4561,11 @@ gc_free_heap(void) gc_set_region_empty(&unboxed_region); last_free_page = 0; - SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base),0); + set_alloc_pointer((lispobj)((char *)heap_base)); if (verify_after_free_heap) { /* Check whether purify has left any bad pointers. */ - if (gencgc_verbose) - SHOW("checking after free_heap\n"); + FSHOW((stderr, "checking after free_heap\n")); verify_gc(); } } @@ -3999,23 +4573,64 @@ gc_free_heap(void) void gc_init(void) { - long i; + page_index_t i; + + /* 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/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); gc_init_tables(); - scavtab[SIMPLE_VECTOR_WIDETAG] = scav_vector; scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer; transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large; - heap_base = (void*)DYNAMIC_SPACE_START; +#ifdef LUTEX_WIDETAG + scavtab[LUTEX_WIDETAG] = scav_lutex; + transother[LUTEX_WIDETAG] = trans_lutex; + sizetab[LUTEX_WIDETAG] = size_lutex; +#endif - /* Initialize each page structure. */ - for (i = 0; i < NUM_PAGES; i++) { - /* Initialize all pages as free. */ - page_table[i].allocated = FREE_PAGE_FLAG; - page_table[i].bytes_used = 0; + heap_base = (void*)DYNAMIC_SPACE_START; - /* 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; @@ -4034,8 +4649,9 @@ 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; } /* Initialize gc_alloc. */ @@ -4044,7 +4660,6 @@ gc_init(void) gc_set_region_empty(&unboxed_region); last_free_page = 0; - } /* Pick up the dynamic space from after a core load. @@ -4055,37 +4670,57 @@ gc_init(void) static void gencgc_pickup_dynamic(void) { - long page = 0; - long alloc_ptr = SymbolValue(ALLOCATION_POINTER,0); + page_index_t page = 0; + 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 = 0; - page_table[page].bytes_used = PAGE_BYTES; - page_table[page].large_object = 0; - - first=gc_search_space(prev,(ptr+2)-prev,ptr); - if(ptr == first) prev=ptr; - page_table[page].first_object_offset = - (void *)prev - page_address(page); + + 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; + page_table[page].region_start_offset = + page_address(page) - (void *)prev; + } page++; - } while ((long)page_address(page) < alloc_ptr); + } while (page_address(page) < alloc_ptr); + +#ifdef LUTEX_WIDETAG + /* Lutexes have been registered in generation 0 by coreparse, and + * need to be moved to the right one manually. + */ + move_lutexes(0, PSEUDO_STATIC_GENERATION); +#endif - generations[0].bytes_allocated = PAGE_BYTES*page; - bytes_allocated = PAGE_BYTES*page; + last_free_page = page; -} + generations[gen].bytes_allocated = npage_bytes(page); + bytes_allocated = npage_bytes(page); + gc_alloc_update_all_page_tables(); + write_protect_generation_pages(gen); +} void gc_initialize_pointers(void) { gencgc_pickup_dynamic(); } - - /* alloc(..) is the external interface for memory allocation. It @@ -4100,41 +4735,24 @@ 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. */ -char * -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; +#ifndef LISP_FEATURE_WIN32 + lispobj alloc_signal; #endif void *new_obj; void *new_free_pointer; + gc_assert(nbytes>0); + /* Check for alignment allocation problems. */ - gc_assert((((unsigned)region->free_pointer & LOWTAG_MASK) == 0) + 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(!SymbolValue(PSEUDO_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(SymbolValue(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; @@ -4144,11 +4762,10 @@ 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(fixnum_value(SymbolValue(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. */ @@ -4156,20 +4773,79 @@ 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) - arch_set_pseudo_atomic_interrupted(0); + 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, 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); + raise(SIGPROF); + } else { + SetSymbolValue(ALLOC_SIGNAL, + alloc_signal - (1 << N_FIXNUM_TAG_BITS), + thread); } } - new_obj = gc_alloc_with_region(nbytes,0,region,0); +#endif + 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); +void unhandled_sigmemoryfault(void* addr); /* Depending on which OS we're running under, different signals might * be raised for a violation of write protection in the heap. This @@ -4184,9 +4860,9 @@ void unhandled_sigmemoryfault(void); int gencgc_handle_wp_violation(void* fault_addr) { - long page_index = find_page_index(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 @@ -4196,15 +4872,18 @@ gencgc_handle_wp_violation(void* fault_addr) /* It can be helpful to be able to put a breakpoint on this * case to help diagnose low-level problems. */ - unhandled_sigmemoryfault(); + unhandled_sigmemoryfault(fault_addr); /* not within the dynamic space -- not our responsibility */ 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 { @@ -4215,8 +4894,12 @@ gencgc_handle_wp_violation(void* fault_addr) * does this test after the first one has already set wp=0 */ if(page_table[page_index].write_protected_cleared != 1) - lose("fault in heap page not marked as write-protected"); + lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n", + 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; } @@ -4226,7 +4909,7 @@ gencgc_handle_wp_violation(void* fault_addr) * are about to let Lisp deal with it. It's basically just a * convenient place to set a gdb breakpoint. */ void -unhandled_sigmemoryfault() +unhandled_sigmemoryfault(void *addr) {} void gc_alloc_update_all_page_tables(void) @@ -4234,10 +4917,11 @@ 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 gc_set_region_empty(struct alloc_region *region) { @@ -4247,3 +4931,93 @@ gc_set_region_empty(struct alloc_region *region) region->free_pointer = page_address(0); region->end_addr = page_address(0); } + +static void +zero_all_free_pages() +{ + page_index_t i; + + for (i = 0; i < last_free_page; i++) { + if (page_free_p(i)) { +#ifdef READ_PROTECT_FREE_PAGES + os_protect(page_address(i), + GENCGC_CARD_BYTES, + OS_VM_PROT_ALL); +#endif + zero_pages(i, i); + } + } +} + +/* Things to do before doing a final GC before saving a core (without + * purify). + * + * + Pages in large_object pages aren't moved by the GC, so we need to + * unset that flag from all pages. + * + The pseudo-static generation isn't normally collected, but it seems + * reasonable to collect it at least when saving a core. So move the + * pages to a normal generation. + */ +static void +prepare_for_final_gc () +{ + page_index_t i; + for (i = 0; i < last_free_page; i++) { + page_table[i].large_object = 0; + if (page_table[i].gen == PSEUDO_STATIC_GENERATION) { + int used = page_table[i].bytes_used; + page_table[i].gen = HIGHEST_NORMAL_GENERATION; + generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used; + generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used; + } + } +} + + +/* Do a non-conservative GC, and then save a core with the initial + * function being set to the value of the static symbol + * SB!VM:RESTART-LISP-FUNCTION */ +void +gc_and_save(char *filename, boolean prepend_runtime, + boolean save_runtime_options) +{ + FILE *file; + void *runtime_bytes = NULL; + size_t runtime_size; + + file = prepare_to_save(filename, prepend_runtime, &runtime_bytes, + &runtime_size); + if (file == NULL) + return; + + conservative_stack = 0; + + /* The filename might come from Lisp, and be moved by the now + * non-conservative GC. */ + filename = strdup(filename); + + /* Collect twice: once into relatively high memory, and then back + * into low memory. This compacts the retained data into the lower + * pages, minimizing the size of the core file. + */ + prepare_for_final_gc(); + gencgc_alloc_start_page = last_free_page; + collect_garbage(HIGHEST_NORMAL_GENERATION+1); + + prepare_for_final_gc(); + gencgc_alloc_start_page = -1; + collect_garbage(HIGHEST_NORMAL_GENERATION+1); + + if (prepend_runtime) + save_runtime_to_filehandle(file, runtime_bytes, runtime_size); + + /* 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, 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 + * going to be rather unsatisfactory too... */ + lose("Attempt to save core after non-conservative GC failed.\n"); +}