X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fgencgc.c;h=1d143c862ac30846a95d763ceb611f32a11623f3;hb=5a942de4c3645c0c63a262eb7b825a9843e96372;hp=27ad418732f8b2abf4e81566e0cd974fcd467662;hpb=a7a9b1029e8b9e45a5b66d62e161cc476cb7b60c;p=sbcl.git diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c index 27ad418..1d143c8 100644 --- a/src/runtime/gencgc.c +++ b/src/runtime/gencgc.c @@ -57,7 +57,7 @@ #endif /* forward declarations */ -page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes, +page_index_t gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int page_type_flag); @@ -79,13 +79,16 @@ boolean enable_page_protection = 1; /* the minimum size (in bytes) for a large object*/ #if (GENCGC_ALLOC_GRANULARITY >= PAGE_BYTES) && (GENCGC_ALLOC_GRANULARITY >= GENCGC_CARD_BYTES) -long large_object_size = 4 * GENCGC_ALLOC_GRANULARITY; +os_vm_size_t large_object_size = 4 * GENCGC_ALLOC_GRANULARITY; #elif (GENCGC_CARD_BYTES >= PAGE_BYTES) && (GENCGC_CARD_BYTES >= GENCGC_ALLOC_GRANULARITY) -long large_object_size = 4 * GENCGC_CARD_BYTES; +os_vm_size_t large_object_size = 4 * GENCGC_CARD_BYTES; #else -long large_object_size = 4 * PAGE_BYTES; +os_vm_size_t large_object_size = 4 * PAGE_BYTES; #endif +/* Largest allocation seen since last GC. */ +os_vm_size_t large_allocation = 0; + /* * debugging @@ -149,8 +152,8 @@ boolean gencgc_partial_pickup = 0; */ /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */ -unsigned long bytes_allocated = 0; -unsigned long auto_gc_trigger = 0; +os_vm_size_t bytes_allocated = 0; +os_vm_size_t auto_gc_trigger = 0; /* the source and destination generations. These are set before a GC starts * scavenging. */ @@ -245,17 +248,17 @@ find_page_index(void *addr) return (-1); } -static size_t -npage_bytes(long npages) +static os_vm_size_t +npage_bytes(page_index_t npages) { gc_assert(npages>=0); - return ((unsigned long)npages)*GENCGC_CARD_BYTES; + return ((os_vm_size_t)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) +static inline os_vm_size_t +void_diff(void *x, void *y) { gc_assert(x >= y); return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y; @@ -285,13 +288,13 @@ struct generation { page_index_t alloc_large_unboxed_start_page; /* the bytes allocated to this generation */ - unsigned long bytes_allocated; + os_vm_size_t bytes_allocated; /* the number of bytes at which to trigger a GC */ - unsigned long gc_trigger; + os_vm_size_t gc_trigger; /* to calculate a new level for gc_trigger */ - unsigned long bytes_consed_between_gc; + os_vm_size_t bytes_consed_between_gc; /* the number of GCs since the last raise */ int num_gc; @@ -305,7 +308,7 @@ struct generation { * 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. */ - unsigned long cum_sum_bytes_allocated; + os_vm_size_t 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 @@ -350,11 +353,11 @@ static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER; static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER; #endif -extern unsigned long gencgc_release_granularity; -unsigned long gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY; +extern os_vm_size_t gencgc_release_granularity; +os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY; -extern unsigned long gencgc_alloc_granularity; -unsigned long gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY; +extern os_vm_size_t gencgc_alloc_granularity; +os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY; /* @@ -363,11 +366,10 @@ unsigned long gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY; /* Count the number of pages which are write-protected within the * given generation. */ -static long +static page_index_t count_write_protect_generation_pages(generation_index_t generation) { - page_index_t i; - unsigned long count = 0; + page_index_t i, count = 0; for (i = 0; i < last_free_page; i++) if (page_allocated_p(i) @@ -378,11 +380,11 @@ count_write_protect_generation_pages(generation_index_t generation) } /* Count the number of pages within the given generation. */ -static long +static page_index_t count_generation_pages(generation_index_t generation) { page_index_t i; - long count = 0; + page_index_t count = 0; for (i = 0; i < last_free_page; i++) if (page_allocated_p(i) @@ -392,11 +394,11 @@ count_generation_pages(generation_index_t generation) } #if QSHOW -static long +static page_index_t count_dont_move_pages(void) { page_index_t i; - long count = 0; + page_index_t count = 0; for (i = 0; i < last_free_page; i++) { if (page_allocated_p(i) && (page_table[i].dont_move != 0)) { @@ -409,11 +411,11 @@ count_dont_move_pages(void) /* Work through the pages and add up the number of bytes used for the * given generation. */ -static unsigned long +static os_vm_size_t count_generation_bytes_allocated (generation_index_t gen) { page_index_t i; - unsigned long result = 0; + os_vm_size_t result = 0; for (i = 0; i < last_free_page; i++) { if (page_allocated_p(i) && (page_table[i].gen == gen)) @@ -457,11 +459,11 @@ write_generation_stats(FILE *file) 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; + page_index_t boxed_cnt = 0; + page_index_t unboxed_cnt = 0; + page_index_t large_boxed_cnt = 0; + page_index_t large_unboxed_cnt = 0; + page_index_t pinned_cnt=0; for (j = 0; j < last_free_page; j++) if (page_table[j].gen == i) { @@ -488,27 +490,31 @@ write_generation_stats(FILE *file) gc_assert(generations[i].bytes_allocated == count_generation_bytes_allocated(i)); fprintf(file, - " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n", + " %1d: %5ld %5ld %5ld %5ld", i, 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].alloc_large_unboxed_start_page); + fprintf(file, + " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT + " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT, + boxed_cnt, unboxed_cnt, large_boxed_cnt, + large_unboxed_cnt, pinned_cnt); + fprintf(file, + " %8"OS_VM_SIZE_FMT + " %5"OS_VM_SIZE_FMT + " %8"OS_VM_SIZE_FMT + " %4"PAGE_INDEX_FMT" %3d %7.4f\n", generations[i].bytes_allocated, - (npage_bytes(count_generation_pages(i)) - - generations[i].bytes_allocated), + (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated), generations[i].gc_trigger, count_write_protect_generation_pages(i), generations[i].num_gc, generation_average_age(i)); } - fprintf(file," Total bytes allocated = %lu\n", bytes_allocated); - fprintf(file," Dynamic-space-size bytes = %lu\n", (unsigned long)dynamic_space_size); + fprintf(file," Total bytes allocated = %"OS_VM_SIZE_FMT"\n", bytes_allocated); + fprintf(file," Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size); fpu_restore(fpu_state); } @@ -587,9 +593,9 @@ void fast_bzero(void*, size_t); /* in -assem.S */ * OS. Generally done after a large GC. */ void zero_pages_with_mmap(page_index_t start, page_index_t end) { - int i; + page_index_t i; void *addr = page_address(start), *new_addr; - size_t length = npage_bytes(1+end-start); + os_vm_size_t length = npage_bytes(1+end-start); if (start > end) return; @@ -786,7 +792,7 @@ gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_ { page_index_t first_page; page_index_t last_page; - unsigned long bytes_found; + os_vm_size_t bytes_found; page_index_t i; int ret; @@ -869,15 +875,11 @@ gc_alloc_new_region(long nbytes, int page_type_flag, struct alloc_region *alloc_ /* we can do this after releasing free_pages_lock */ if (gencgc_zero_check) { - long *p; - for (p = (long *)alloc_region->start_addr; - p < (long *)alloc_region->end_addr; p++) { + word_t *p; + for (p = (word_t *)alloc_region->start_addr; + p < (word_t *)alloc_region->end_addr; p++) { if (*p != 0) { - /* KLUDGE: It would be nice to use %lx and explicit casts - * (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 (start=%p, end=%p).\n", + lose("The new region is not zero at %p (start=%p, end=%p).\n", p, alloc_region->start_addr, alloc_region->end_addr); } } @@ -908,15 +910,14 @@ struct new_area { size_t size; }; static struct new_area (*new_areas)[]; -static long new_areas_index; -long max_new_areas; +static size_t new_areas_index; +size_t max_new_areas; /* Add a new area to new_areas. */ static void add_new_area(page_index_t first_page, size_t offset, size_t size) { - unsigned long new_area_start,c; - long i; + size_t new_area_start, c, i; /* Ignore if full. */ if (new_areas_index >= NUM_NEW_AREAS) @@ -940,7 +941,7 @@ add_new_area(page_index_t first_page, size_t offset, size_t size) /* 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 long area_end = + size_t area_end = npage_bytes((*new_areas)[i].page) + (*new_areas)[i].offset + (*new_areas)[i].size; @@ -985,13 +986,13 @@ add_new_area(page_index_t first_page, size_t offset, size_t size) void gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region) { - int more; + boolean 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; + os_vm_size_t bytes_used; + os_vm_size_t region_size; + os_vm_size_t byte_cnt; + page_bytes_t orig_first_page_bytes_used; int ret; @@ -1117,13 +1118,11 @@ static inline void *gc_quick_alloc(long nbytes); void * gc_alloc_large(long nbytes, int page_type_flag, struct alloc_region *alloc_region) { - page_index_t first_page; - page_index_t last_page; - int orig_first_page_bytes_used; - long byte_cnt; - int more; - unsigned long bytes_used; - page_index_t next_page; + boolean more; + page_index_t first_page, next_page, last_page; + page_bytes_t orig_first_page_bytes_used; + os_vm_size_t byte_cnt; + os_vm_size_t bytes_used; int ret; ret = thread_mutex_lock(&free_pages_lock); @@ -1265,27 +1264,29 @@ gc_heap_exhausted_error_or_lose (long available, long requested) } page_index_t -gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, +gc_find_freeish_pages(page_index_t *restart_page_ptr, long bytes, int page_type_flag) { - 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; + page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0; + page_index_t first_page, last_page, restart_page = *restart_page_ptr; + os_vm_size_t nbytes = bytes; + os_vm_size_t nbytes_goal = nbytes; + os_vm_size_t bytes_found = 0; + os_vm_size_t most_bytes_found = 0; + boolean small_object = nbytes < GENCGC_CARD_BYTES; /* FIXME: assert(free_pages_lock is held); */ if (nbytes_goal < gencgc_alloc_granularity) - nbytes_goal = gencgc_alloc_granularity; + nbytes_goal = gencgc_alloc_granularity; /* Toggled by gc_and_save for heap compaction, normally -1. */ if (gencgc_alloc_start_page != -1) { restart_page = gencgc_alloc_start_page; } - gc_assert(nbytes>=0); + /* FIXME: This is on bytes instead of nbytes pending cleanup of + * long from the interface. */ + gc_assert(bytes>=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 @@ -1349,6 +1350,7 @@ gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes); } + gc_assert(most_bytes_found_to); *restart_page_ptr = most_bytes_found_from; return most_bytes_found_to-1; } @@ -1408,12 +1410,6 @@ gc_quick_alloc(long nbytes) } static inline void * -gc_quick_alloc_large(long nbytes) -{ - return gc_general_alloc(nbytes, BOXED_PAGE_FLAG ,ALLOC_QUICK); -} - -static inline void * gc_alloc_unboxed(long nbytes) { return gc_general_alloc(nbytes, UNBOXED_PAGE_FLAG, 0); @@ -1424,166 +1420,15 @@ gc_quick_alloc_unboxed(long nbytes) { 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, UNBOXED_PAGE_FLAG, ALLOC_QUICK); -} - -/* 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 - * enough then it's copied to a large object region. - * - * Vectors may have shrunk. If the object is not copied the space - * needs to be reclaimed, and the page_tables corrected. */ -lispobj -copy_large_object(lispobj object, long nwords) -{ - int tag; - lispobj *new; - page_index_t first_page; - - gc_assert(is_lisp_pointer(object)); - gc_assert(from_space_p(object)); - gc_assert((nwords & 0x01) == 0); - - - /* Check whether it's in a large object region. */ - first_page = find_page_index((void *)object); - gc_assert(first_page >= 0); - - if (page_table[first_page].large_object) { - - /* Promote the object. */ - - 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 - * pages. This would not be necessary if they are added to the - * new areas, but let's do it for them all (they'll probably - * be written anyway?). */ - - gc_assert(page_table[first_page].region_start_offset == 0); - - next_page = first_page; - remaining_bytes = nwords*N_WORD_BYTES; - while (remaining_bytes > GENCGC_CARD_BYTES) { - gc_assert(page_table[next_page].gen == from_space); - gc_assert(page_boxed_p(next_page)); - gc_assert(page_table[next_page].large_object); - gc_assert(page_table[next_page].region_start_offset == - npage_bytes(next_page-first_page)); - gc_assert(page_table[next_page].bytes_used == 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; - - remaining_bytes -= GENCGC_CARD_BYTES; - next_page++; - } - - /* Now only one page remains, but the object may have shrunk - * so there may be more unused pages which will be freed. */ - - /* The object may have shrunk but shouldn't have grown. */ - gc_assert(page_table[next_page].bytes_used >= remaining_bytes); - - page_table[next_page].gen = new_space; - gc_assert(page_boxed_p(next_page)); - - /* Adjust the bytes_used. */ - old_bytes_used = page_table[next_page].bytes_used; - page_table[next_page].bytes_used = remaining_bytes; - - bytes_freed = old_bytes_used - remaining_bytes; - - /* Free any remaining pages; needs care. */ - next_page++; - while ((old_bytes_used == GENCGC_CARD_BYTES) && - (page_table[next_page].gen == from_space) && - page_boxed_p(next_page) && - page_table[next_page].large_object && - (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 - * should be zero filled. */ - gc_assert(page_table[next_page].write_protected == 0); - - old_bytes_used = page_table[next_page].bytes_used; - page_table[next_page].allocated = FREE_PAGE_FLAG; - page_table[next_page].bytes_used = 0; - bytes_freed += old_bytes_used; - next_page++; - } - - generations[from_space].bytes_allocated -= N_WORD_BYTES*nwords - + bytes_freed; - generations[new_space].bytes_allocated += N_WORD_BYTES*nwords; - bytes_allocated -= bytes_freed; - - /* Add the region to the new_areas if requested. */ - add_new_area(first_page,0,nwords*N_WORD_BYTES); - - return(object); - } else { - /* Get tag of object. */ - tag = lowtag_of(object); - - /* Allocate space. */ - new = gc_quick_alloc_large(nwords*N_WORD_BYTES); - - memcpy(new,native_pointer(object),nwords*N_WORD_BYTES); - - /* Return Lisp pointer of new object. */ - return ((lispobj) new) | tag; - } -} - -/* to copy unboxed objects */ -lispobj -copy_unboxed_object(lispobj object, long nwords) -{ - long tag; - lispobj *new; - - gc_assert(is_lisp_pointer(object)); - gc_assert(from_space_p(object)); - gc_assert((nwords & 0x01) == 0); - - /* Get tag of object. */ - tag = lowtag_of(object); - - /* Allocate space. */ - new = gc_quick_alloc_unboxed(nwords*N_WORD_BYTES); - - memcpy(new,native_pointer(object),nwords*N_WORD_BYTES); - - /* Return Lisp pointer of new object. */ - return ((lispobj) new) | tag; -} - -/* to copy large unboxed objects - * - * If the object is in a large object region then it is simply - * promoted, else it is copied. If it's large enough then it's copied - * to a large object region. +/* Copy a large object. If the object is in a large object region then + * it is simply promoted, else it is copied. If it's large enough then + * it's copied to a large object region. * * Bignums and vectors may have shrunk. If the object is not copied - * the space needs to be reclaimed, and the page_tables corrected. - * - * KLUDGE: There's a lot of cut-and-paste duplication between this - * function and copy_large_object(..). -- WHN 20000619 */ -lispobj -copy_large_unboxed_object(lispobj object, long nwords) + * the space needs to be reclaimed, and the page_tables corrected. */ +static lispobj +general_copy_large_object(lispobj object, long nwords, boolean boxedp) { int tag; lispobj *new; @@ -1594,7 +1439,7 @@ copy_large_unboxed_object(lispobj object, long nwords) gc_assert((nwords & 0x01) == 0); if ((nwords > 1024*1024) && gencgc_verbose) { - FSHOW((stderr, "/copy_large_unboxed_object: %d bytes\n", + FSHOW((stderr, "/general_copy_large_object: %d bytes\n", nwords*N_WORD_BYTES)); } @@ -1606,25 +1451,42 @@ 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. */ - unsigned long remaining_bytes; + os_vm_size_t remaining_bytes; + os_vm_size_t bytes_freed; page_index_t next_page; - unsigned long bytes_freed; - unsigned long old_bytes_used; + page_bytes_t old_bytes_used; - gc_assert(page_table[first_page].region_start_offset == 0); + /* FIXME: This comment is somewhat stale. + * + * Note: Any page write-protection must be removed, else a + * later scavenge_newspace may incorrectly not scavenge these + * pages. This would not be necessary if they are added to the + * new areas, but let's do it for them all (they'll probably + * be written anyway?). */ + gc_assert(page_table[first_page].region_start_offset == 0); next_page = first_page; remaining_bytes = nwords*N_WORD_BYTES; + while (remaining_bytes > GENCGC_CARD_BYTES) { gc_assert(page_table[next_page].gen == from_space); - gc_assert(page_allocated_no_region_p(next_page)); gc_assert(page_table[next_page].large_object); gc_assert(page_table[next_page].region_start_offset == npage_bytes(next_page-first_page)); gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES); - + /* Should have been unprotected by unprotect_oldspace() + * for boxed objects, and after promotion unboxed ones + * should not be on protected pages at all. */ + gc_assert(!page_table[next_page].write_protected); + + if (boxedp) + gc_assert(page_boxed_p(next_page)); + else { + gc_assert(page_allocated_no_region_p(next_page)); + page_table[next_page].allocated = UNBOXED_PAGE_FLAG; + } page_table[next_page].gen = new_space; - page_table[next_page].allocated = UNBOXED_PAGE_FLAG; + remaining_bytes -= GENCGC_CARD_BYTES; next_page++; } @@ -1636,7 +1498,11 @@ copy_large_unboxed_object(lispobj object, long nwords) gc_assert(page_table[next_page].bytes_used >= remaining_bytes); page_table[next_page].gen = new_space; - page_table[next_page].allocated = UNBOXED_PAGE_FLAG; + + if (boxedp) + gc_assert(page_boxed_p(next_page)); + else + page_table[next_page].allocated = UNBOXED_PAGE_FLAG; /* Adjust the bytes_used. */ old_bytes_used = page_table[next_page].bytes_used; @@ -1648,7 +1514,14 @@ copy_large_unboxed_object(lispobj object, long nwords) next_page++; while ((old_bytes_used == GENCGC_CARD_BYTES) && (page_table[next_page].gen == from_space) && - page_allocated_no_region_p(next_page) && + /* FIXME: It is not obvious to me why this is necessary + * as a loop condition: it seems to me that the + * region_start_offset test should be sufficient, but + * experimentally that is not the case. --NS + * 2011-11-28 */ + (boxedp ? + page_boxed_p(next_page) : + page_allocated_no_region_p(next_page)) && page_table[next_page].large_object && (page_table[next_page].region_start_offset == npage_bytes(next_page - first_page))) { @@ -1667,23 +1540,29 @@ copy_large_unboxed_object(lispobj object, long nwords) if ((bytes_freed > 0) && gencgc_verbose) { FSHOW((stderr, - "/copy_large_unboxed bytes_freed=%d\n", + "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\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; + /* Add the region to the new_areas if requested. */ + if (boxedp) + add_new_area(first_page,0,nwords*N_WORD_BYTES); + return(object); - } - else { + + } else { /* Get tag of object. */ tag = lowtag_of(object); /* Allocate space. */ - new = gc_quick_alloc_large_unboxed(nwords*N_WORD_BYTES); + new = gc_general_alloc(nwords*N_WORD_BYTES, + (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG), + ALLOC_QUICK); /* Copy the object. */ memcpy(new,native_pointer(object),nwords*N_WORD_BYTES); @@ -1693,8 +1572,24 @@ copy_large_unboxed_object(lispobj object, long nwords) } } +lispobj +copy_large_object(lispobj object, long nwords) +{ + return general_copy_large_object(object, nwords, 1); +} +lispobj +copy_large_unboxed_object(lispobj object, long nwords) +{ + return general_copy_large_object(object, nwords, 0); +} +/* to copy unboxed objects */ +lispobj +copy_unboxed_object(lispobj object, long nwords) +{ + return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG); +} /* @@ -1715,13 +1610,13 @@ 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 long displacement) +sniff_code_object(struct code *code, os_vm_size_t displacement) { #ifdef LISP_FEATURE_X86 long nheader_words, ncode_words, nwords; - void *p; - void *constants_start_addr = NULL, *constants_end_addr; - void *code_start_addr, *code_end_addr; + os_vm_address_t constants_start_addr = NULL, constants_end_addr, p; + os_vm_address_t code_start_addr, code_end_addr; + os_vm_address_t code_addr = (os_vm_address_t)code; int fixup_found = 0; if (!check_code_fixups) @@ -1733,10 +1628,10 @@ sniff_code_object(struct code *code, unsigned long displacement) nheader_words = HeaderValue(*(lispobj *)code); nwords = ncode_words + nheader_words; - constants_start_addr = (void *)code + 5*N_WORD_BYTES; - constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES; - code_start_addr = (void *)code + nheader_words*N_WORD_BYTES; - code_end_addr = (void *)code + nwords*N_WORD_BYTES; + constants_start_addr = code_addr + 5*N_WORD_BYTES; + constants_end_addr = code_addr + nheader_words*N_WORD_BYTES; + code_start_addr = code_addr + nheader_words*N_WORD_BYTES; + code_end_addr = code_addr + nwords*N_WORD_BYTES; /* Work through the unboxed code. */ for (p = code_start_addr; p < code_end_addr; p++) { @@ -1753,8 +1648,8 @@ sniff_code_object(struct code *code, unsigned long displacement) /* Check for code references. */ /* Check for a 32 bit word that looks like an absolute reference to within the code adea of the code object. */ - if ((data >= (code_start_addr-displacement)) - && (data < (code_end_addr-displacement))) { + if ((data >= (void*)(code_start_addr-displacement)) + && (data < (void*)(code_end_addr-displacement))) { /* function header */ if ((d4 == 0x5e) && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) == @@ -1796,8 +1691,8 @@ sniff_code_object(struct code *code, unsigned long displacement) /* Check for a 32 bit word that looks like an absolute reference to within the constant vector. Constant references will be aligned. */ - if ((data >= (constants_start_addr-displacement)) - && (data < (constants_end_addr-displacement)) + if ((data >= (void*)(constants_start_addr-displacement)) + && (data < (void*)(constants_end_addr-displacement)) && (((unsigned)data & 0x3) == 0)) { /* Mov eax,m32 */ if (d1 == 0xa1) { @@ -1895,11 +1790,12 @@ 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; + os_vm_address_t constants_start_addr, constants_end_addr; + os_vm_address_t code_start_addr, code_end_addr; + os_vm_address_t code_addr = (os_vm_address_t)new_code; + os_vm_address_t old_addr = (os_vm_address_t)old_code; + os_vm_size_t displacement = code_addr - old_addr; lispobj fixups = NIL; - unsigned long displacement = - (unsigned long)new_code - (unsigned long)old_code; struct vector *fixups_vector; ncode_words = fixnum_value(new_code->code_size); @@ -1908,10 +1804,10 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code) /* FSHOW((stderr, "/compiled code object at %x: header words = %d, code words = %d\n", new_code, nheader_words, ncode_words)); */ - constants_start_addr = (void *)new_code + 5*N_WORD_BYTES; - constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES; - code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES; - code_end_addr = (void *)new_code + nwords*N_WORD_BYTES; + constants_start_addr = code_addr + 5*N_WORD_BYTES; + constants_end_addr = code_addr + nheader_words*N_WORD_BYTES; + code_start_addr = code_addr + nheader_words*N_WORD_BYTES; + code_end_addr = code_addr + nwords*N_WORD_BYTES; /* FSHOW((stderr, "/const start = %x, end = %x\n", @@ -1959,24 +1855,22 @@ 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 long offset = fixups_vector->data[i]; + long offset = fixups_vector->data[i]; /* Now check the current value of offset. */ - unsigned long old_value = - *(unsigned long *)((unsigned long)code_start_addr + offset); + os_vm_address_t old_value = *(os_vm_address_t *)(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 long)old_code) - && (old_value < ((unsigned long)old_code - + nwords*N_WORD_BYTES))) + if ((old_value >= old_addr) + && (old_value < (old_addr + nwords*N_WORD_BYTES))) /* So add the dispacement. */ - *(unsigned long *)((unsigned long)code_start_addr + offset) = + *(os_vm_address_t *)(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 long *)((unsigned long)code_start_addr + offset) = + *(os_vm_address_t *)(code_start_addr + offset) = old_value - displacement; } } else { @@ -2536,7 +2430,7 @@ static void scavenge_generations(generation_index_t from, generation_index_t to) { page_index_t i; - int num_wp = 0; + page_index_t num_wp = 0; #define SC_GEN_CK 0 #if SC_GEN_CK @@ -2711,15 +2605,15 @@ scavenge_newspace_generation_one_scan(generation_index_t generation) static void scavenge_newspace_generation(generation_index_t generation) { - long i; + size_t i; /* the new_areas array currently being written to by gc_alloc() */ struct new_area (*current_new_areas)[] = &new_areas_1; - long current_new_areas_index; + size_t current_new_areas_index; /* the new_areas created by the previous scavenge cycle */ struct new_area (*previous_new_areas)[] = NULL; - long previous_new_areas_index; + size_t previous_new_areas_index; /* Flush the current regions updating the tables. */ gc_alloc_update_all_page_tables(); @@ -2826,16 +2720,19 @@ scavenge_newspace_generation(generation_index_t generation) record_new_objects = 0; #if SC_NS_GEN_CK - /* Check that none of the write_protected pages in this generation - * have been written to. */ - for (i = 0; i < page_table_pages; i++) { - if (page_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\n", - i, generation, page_table[i].dont_move); + { + page_index_t i; + /* Check that none of the write_protected pages in this generation + * have been written to. */ + for (i = 0; i < page_table_pages; i++) { + if (page_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\n", + i, generation, page_table[i].dont_move); + } } } #endif @@ -2950,8 +2847,8 @@ print_ptr(lispobj *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 %lu dont_move %d\n", - (unsigned long) addr, + fprintf(stderr," %p: page %d alloc %d gen %d bytes_used %d offset %lu dont_move %d\n", + addr, pi1, page_table[pi1].allocated, page_table[pi1].gen, @@ -3404,19 +3301,6 @@ write_protect_generation_pages(generation_index_t 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) @@ -3686,8 +3570,8 @@ garbage_collect_generation(generation_index_t generation, int raise) /* As a check re-scavenge the newspace once; no new objects should * be found. */ { - long old_bytes_allocated = bytes_allocated; - long bytes_allocated; + os_vm_size_t old_bytes_allocated = bytes_allocated; + os_vm_size_t bytes_allocated; /* Start with a full scavenge. */ scavenge_newspace_generation_one_scan(new_space); @@ -3836,7 +3720,7 @@ void collect_garbage(generation_index_t last_gen) { generation_index_t gen = 0, i; - int raise; + int raise, more = 0; int gen_to_wp; /* The largest value of last_free_page seen since the time * remap_free_pages was called. */ @@ -3869,13 +3753,23 @@ collect_garbage(generation_index_t last_gen) do { /* Collect the generation. */ - if (gen >= gencgc_oldest_gen_to_gc) { - /* Never raise the oldest generation. */ + if (more || (gen >= gencgc_oldest_gen_to_gc)) { + /* Never raise the oldest generation. Never raise the extra generation + * collected due to more-flag. */ raise = 0; + more = 0; } else { raise = (gen < last_gen) || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion); + /* If we would not normally raise this one, but we're + * running low on space in comparison to the object-sizes + * we've been seeing, raise it and collect the next one + * too. */ + if (!raise && gen == last_gen) { + more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated); + raise = more; + } } if (gencgc_verbose > 1) { @@ -3908,8 +3802,8 @@ collect_garbage(generation_index_t last_gen) gen++; } while ((gen <= gencgc_oldest_gen_to_gc) && ((gen < last_gen) - || ((gen <= gencgc_oldest_gen_to_gc) - && raise + || more + || (raise && (generations[gen].bytes_allocated > generations[gen].gc_trigger) && (generation_average_age(gen) @@ -3951,9 +3845,15 @@ collect_garbage(generation_index_t last_gen) update_dynamic_space_free_pointer(); - auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs; + /* Update auto_gc_trigger. Make sure we trigger the next GC before + * running out of heap! */ + if (bytes_consed_between_gcs >= dynamic_space_size - bytes_allocated) + auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs; + else + auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2; + if(gencgc_verbose) - fprintf(stderr,"Next gc when %ld bytes have been consed\n", + fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n", auto_gc_trigger); /* If we did a big GC (arbitrarily defined as gen > 1), release memory @@ -3967,6 +3867,7 @@ collect_garbage(generation_index_t last_gen) } gc_active_p = 0; + large_allocation = 0; log_generation_stats(gc_logfile, "=== GC End ==="); SHOW("returning from collect_garbage"); @@ -3989,7 +3890,7 @@ gc_free_heap(void) for (page = 0; page < page_table_pages; page++) { /* Skip free pages which should already be zero filled. */ if (page_allocated_p(page)) { - void *page_start, *addr; + void *page_start; for (last_page = page; (last_page < page_table_pages) && page_allocated_p(last_page); last_page++) { @@ -4068,6 +3969,12 @@ gc_init(void) page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES; gc_assert(dynamic_space_size == npage_bytes(page_table_pages)); + /* Default nursery size to 5% of the total dynamic space size, + * min 1Mb. */ + bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20; + if (bytes_consed_between_gcs < (1024*1024)) + bytes_consed_between_gcs = 1024*1024; + /* 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 @@ -4130,7 +4037,7 @@ gc_init(void) generations[i].num_gc = 0; generations[i].cum_sum_bytes_allocated = 0; /* the tune-able parameters */ - generations[i].bytes_consed_between_gc = 2000000; + generations[i].bytes_consed_between_gc = bytes_consed_between_gcs; generations[i].number_of_gcs_before_promotion = 1; generations[i].minimum_age_before_gc = 0.75; } @@ -4228,6 +4135,9 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg /* Must be inside a PA section. */ gc_assert(get_pseudo_atomic_atomic(thread)); + if (nbytes > large_allocation) + large_allocation = nbytes; + /* maybe we can do this quickly ... */ new_free_pointer = region->free_pointer + nbytes; if (new_free_pointer <= region->end_addr) { @@ -4239,7 +4149,7 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg /* 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) { + if (auto_gc_trigger && bytes_allocated+nbytes > auto_gc_trigger) { /* 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. */ @@ -4329,7 +4239,17 @@ void unhandled_sigmemoryfault(void* addr); * * Return true if this signal is a normal generational GC thing that * we were able to handle, or false if it was abnormal and control - * should fall through to the general SIGSEGV/SIGBUS/whatever logic. */ + * should fall through to the general SIGSEGV/SIGBUS/whatever logic. + * + * We have two control flags for this: one causes us to ignore faults + * on unprotected pages completely, and the second complains to stderr + * but allows us to continue without losing. + */ +extern boolean ignore_memoryfaults_on_unprotected_pages; +boolean ignore_memoryfaults_on_unprotected_pages = 0; + +extern boolean continue_after_memoryfault_on_unprotected_pages; +boolean continue_after_memoryfault_on_unprotected_pages = 0; int gencgc_handle_wp_violation(void* fault_addr) @@ -4360,17 +4280,39 @@ gencgc_handle_wp_violation(void* fault_addr) 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 { + } else if (!ignore_memoryfaults_on_unprotected_pages) { /* The only acceptable reason for this signal on a heap * access is that GENCGC write-protected the page. * However, if two CPUs hit a wp page near-simultaneously, * we had better not have the second one lose here if it * 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 %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); + if(page_table[page_index].write_protected_cleared != 1) { + void lisp_backtrace(int frames); + lisp_backtrace(10); + fprintf(stderr, + "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n" + " boxed_region.first_page: %"PAGE_INDEX_FMT"," + " boxed_region.last_page %"PAGE_INDEX_FMT"\n" + " page.region_start_offset: %"OS_VM_SIZE_FMT"\n" + " page.bytes_used: %"PAGE_BYTES_FMT"\n" + " page.allocated: %d\n" + " page.write_protected: %d\n" + " page.write_protected_cleared: %d\n" + " page.generation: %d\n", + fault_addr, + page_index, + boxed_region.first_page, + boxed_region.last_page, + page_table[page_index].region_start_offset, + page_table[page_index].bytes_used, + page_table[page_index].allocated, + page_table[page_index].write_protected, + page_table[page_index].write_protected_cleared, + page_table[page_index].gen); + if (!continue_after_memoryfault_on_unprotected_pages) + lose("Feh.\n"); + } } ret = thread_mutex_unlock(&free_pages_lock); gc_assert(ret == 0);