=================================================================== RCS file: /home/cvs/OpenXM_contrib2/asir2000/gc/mallocx.c,v retrieving revision 1.2 retrieving revision 1.3 diff -u -p -r1.2 -r1.3 --- OpenXM_contrib2/asir2000/gc/mallocx.c 2000/12/01 09:26:11 1.2 +++ OpenXM_contrib2/asir2000/gc/mallocx.c 2001/04/20 07:39:19 1.3 @@ -2,6 +2,7 @@ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved. * Copyright (c) 1996 by Silicon Graphics. All rights reserved. + * Copyright (c) 2000 by Hewlett-Packard Company. All rights reserved. * * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED * OR IMPLIED. ANY USE IS AT YOUR OWN RISK. @@ -21,7 +22,7 @@ */ #include -#include "gc_priv.h" +#include "private/gc_priv.h" extern ptr_t GC_clear_stack(); /* in misc.c, behaves like identity */ void GC_extend_size_map(); /* in misc.c. */ @@ -30,69 +31,179 @@ GC_bool GC_alloc_reclaim_list(); /* in malloc.c */ /* Some externally visible but unadvertised variables to allow access to */ /* free lists from inlined allocators without including gc_priv.h */ /* or introducing dependencies on internal data structure layouts. */ -ptr_t * CONST GC_objfreelist_ptr = GC_objfreelist; -ptr_t * CONST GC_aobjfreelist_ptr = GC_aobjfreelist; -ptr_t * CONST GC_uobjfreelist_ptr = GC_uobjfreelist; +ptr_t * GC_CONST GC_objfreelist_ptr = GC_objfreelist; +ptr_t * GC_CONST GC_aobjfreelist_ptr = GC_aobjfreelist; +ptr_t * GC_CONST GC_uobjfreelist_ptr = GC_uobjfreelist; # ifdef ATOMIC_UNCOLLECTABLE - ptr_t * CONST GC_auobjfreelist_ptr = GC_auobjfreelist; + ptr_t * GC_CONST GC_auobjfreelist_ptr = GC_auobjfreelist; # endif -/* Allocate a composite object of size n bytes. The caller guarantees */ -/* that pointers past the first page are not relevant. Caller holds */ -/* allocation lock. */ -ptr_t GC_generic_malloc_inner_ignore_off_page(lb, k) -register size_t lb; -register int k; + +GC_PTR GC_generic_or_special_malloc(lb,knd) +word lb; +int knd; { - register struct hblk * h; - register word n_blocks; - register word lw; - register ptr_t op; + switch(knd) { +# ifdef STUBBORN_ALLOC + case STUBBORN: + return(GC_malloc_stubborn((size_t)lb)); +# endif + case PTRFREE: + return(GC_malloc_atomic((size_t)lb)); + case NORMAL: + return(GC_malloc((size_t)lb)); + case UNCOLLECTABLE: + return(GC_malloc_uncollectable((size_t)lb)); +# ifdef ATOMIC_UNCOLLECTABLE + case AUNCOLLECTABLE: + return(GC_malloc_atomic_uncollectable((size_t)lb)); +# endif /* ATOMIC_UNCOLLECTABLE */ + default: + return(GC_generic_malloc(lb,knd)); + } +} - if (lb <= HBLKSIZE) - return(GC_generic_malloc_inner((word)lb, k)); - n_blocks = divHBLKSZ(ADD_SLOP(lb) + HDR_BYTES + HBLKSIZE-1); - if (!GC_is_initialized) GC_init_inner(); - /* Do our share of marking work */ - if(GC_incremental && !GC_dont_gc) - GC_collect_a_little_inner((int)n_blocks); - lw = ROUNDED_UP_WORDS(lb); - h = GC_allochblk(lw, k, IGNORE_OFF_PAGE); -# ifdef USE_MUNMAP - if (0 == h) { - GC_merge_unmapped(); - h = GC_allochblk(lw, k, IGNORE_OFF_PAGE); - } -# endif - while (0 == h && GC_collect_or_expand(n_blocks, TRUE)) { - h = GC_allochblk(lw, k, IGNORE_OFF_PAGE); + +/* Change the size of the block pointed to by p to contain at least */ +/* lb bytes. The object may be (and quite likely will be) moved. */ +/* The kind (e.g. atomic) is the same as that of the old. */ +/* Shrinking of large blocks is not implemented well. */ +# ifdef __STDC__ + GC_PTR GC_realloc(GC_PTR p, size_t lb) +# else + GC_PTR GC_realloc(p,lb) + GC_PTR p; + size_t lb; +# endif +{ +register struct hblk * h; +register hdr * hhdr; +register word sz; /* Current size in bytes */ +register word orig_sz; /* Original sz in bytes */ +int obj_kind; + + if (p == 0) return(GC_malloc(lb)); /* Required by ANSI */ + h = HBLKPTR(p); + hhdr = HDR(h); + sz = hhdr -> hb_sz; + obj_kind = hhdr -> hb_obj_kind; + sz = WORDS_TO_BYTES(sz); + orig_sz = sz; + + if (sz > MAXOBJBYTES) { + /* Round it up to the next whole heap block */ + register word descr; + + sz = (sz+HBLKSIZE-1) & (~HBLKMASK); + hhdr -> hb_sz = BYTES_TO_WORDS(sz); + descr = GC_obj_kinds[obj_kind].ok_descriptor; + if (GC_obj_kinds[obj_kind].ok_relocate_descr) descr += sz; + hhdr -> hb_descr = descr; + if (IS_UNCOLLECTABLE(obj_kind)) GC_non_gc_bytes += (sz - orig_sz); + /* Extra area is already cleared by GC_alloc_large_and_clear. */ } - if (h == 0) { - op = 0; + if (ADD_SLOP(lb) <= sz) { + if (lb >= (sz >> 1)) { +# ifdef STUBBORN_ALLOC + if (obj_kind == STUBBORN) GC_change_stubborn(p); +# endif + if (orig_sz > lb) { + /* Clear unneeded part of object to avoid bogus pointer */ + /* tracing. */ + /* Safe for stubborn objects. */ + BZERO(((ptr_t)p) + lb, orig_sz - lb); + } + return(p); + } else { + /* shrink */ + GC_PTR result = + GC_generic_or_special_malloc((word)lb, obj_kind); + + if (result == 0) return(0); + /* Could also return original object. But this */ + /* gives the client warning of imminent disaster. */ + BCOPY(p, result, lb); +# ifndef IGNORE_FREE + GC_free(p); +# endif + return(result); + } } else { - op = (ptr_t) (h -> hb_body); - GC_words_wasted += BYTES_TO_WORDS(n_blocks * HBLKSIZE) - lw; + /* grow */ + GC_PTR result = + GC_generic_or_special_malloc((word)lb, obj_kind); + + if (result == 0) return(0); + BCOPY(p, result, sz); +# ifndef IGNORE_FREE + GC_free(p); +# endif + return(result); } - GC_words_allocd += lw; - return((ptr_t)op); } +# if defined(REDIRECT_MALLOC) || defined(REDIRECT_REALLOC) +# ifdef __STDC__ + GC_PTR realloc(GC_PTR p, size_t lb) +# else + GC_PTR realloc(p,lb) + GC_PTR p; + size_t lb; +# endif + { +# ifdef REDIRECT_REALLOC + return(REDIRECT_REALLOC(p, lb)); +# else + return(GC_realloc(p, lb)); +# endif + } +# endif /* REDIRECT_MALLOC */ + + +/* The same thing, except caller does not hold allocation lock. */ +/* We avoid holding allocation lock while we clear memory. */ ptr_t GC_generic_malloc_ignore_off_page(lb, k) register size_t lb; register int k; { register ptr_t result; + word lw; + word n_blocks; + GC_bool init; DCL_LOCK_STATE; + if (SMALL_OBJ(lb)) + return(GC_generic_malloc((word)lb, k)); + lw = ROUNDED_UP_WORDS(lb); + n_blocks = OBJ_SZ_TO_BLOCKS(lw); + init = GC_obj_kinds[k].ok_init; GC_INVOKE_FINALIZERS(); DISABLE_SIGNALS(); LOCK(); - result = GC_generic_malloc_inner_ignore_off_page(lb,k); + result = (ptr_t)GC_alloc_large(lw, k, IGNORE_OFF_PAGE); + if (0 != result) { + if (GC_debugging_started) { + BZERO(result, n_blocks * HBLKSIZE); + } else { +# ifdef THREADS + /* Clear any memory that might be used for GC descriptors */ + /* before we release the lock. */ + ((word *)result)[0] = 0; + ((word *)result)[1] = 0; + ((word *)result)[lw-1] = 0; + ((word *)result)[lw-2] = 0; +# endif + } + } + GC_words_allocd += lw; UNLOCK(); ENABLE_SIGNALS(); if (0 == result) { return((*GC_oom_fn)(lb)); } else { + if (init & !GC_debugging_started) { + BZERO(result, n_blocks * HBLKSIZE); + } return(result); } } @@ -185,6 +296,24 @@ DCL_LOCK_STATE; } #if defined(THREADS) && !defined(SRC_M3) + +extern signed_word GC_mem_found; /* Protected by GC lock. */ + +#ifdef PARALLEL_MARK +volatile signed_word GC_words_allocd_tmp = 0; + /* Number of words of memory allocated since */ + /* we released the GC lock. Instead of */ + /* reacquiring the GC lock just to add this in, */ + /* we add it in the next time we reacquire */ + /* the lock. (Atomically adding it doesn't */ + /* work, since we would have to atomically */ + /* update it in GC_malloc, which is too */ + /* expensive. */ +#endif /* PARALLEL_MARK */ + +/* See reclaim.c: */ +extern ptr_t GC_reclaim_generic(); + /* Return a list of 1 or more objects of the indicated size, linked */ /* through the first word in the object. This has the advantage that */ /* it acquires the allocation lock only once, and may greatly reduce */ @@ -200,12 +329,19 @@ register word lb; register int k; { ptr_t op; -register ptr_t p; +ptr_t p; ptr_t *opp; word lw; -register word my_words_allocd; +word my_words_allocd = 0; +struct obj_kind * ok = &(GC_obj_kinds[k]); DCL_LOCK_STATE; +# if defined(GATHERSTATS) || defined(PARALLEL_MARK) +# define COUNT_ARG , &my_words_allocd +# else +# define COUNT_ARG +# define NEED_TO_COUNT +# endif if (!SMALL_OBJ(lb)) { op = GC_generic_malloc(lb, k); if(0 != op) obj_link(op) = 0; @@ -215,40 +351,142 @@ DCL_LOCK_STATE; GC_INVOKE_FINALIZERS(); DISABLE_SIGNALS(); LOCK(); - opp = &(GC_obj_kinds[k].ok_freelist[lw]); - if( (op = *opp) == 0 ) { - if (!GC_is_initialized) { - GC_init_inner(); - } - op = GC_clear_stack(GC_allocobj(lw, k)); - if (op == 0) { - UNLOCK(); - ENABLE_SIGNALS(); - op = (*GC_oom_fn)(lb); - if(0 != op) obj_link(op) = 0; - return(op); - } + if (!GC_is_initialized) GC_init_inner(); + /* First see if we can reclaim a page of objects waiting to be */ + /* reclaimed. */ + { + struct hblk ** rlh = ok -> ok_reclaim_list; + struct hblk * hbp; + hdr * hhdr; + + rlh += lw; + while ((hbp = *rlh) != 0) { + hhdr = HDR(hbp); + *rlh = hhdr -> hb_next; +# ifdef PARALLEL_MARK + { + signed_word my_words_allocd_tmp = GC_words_allocd_tmp; + + GC_ASSERT(my_words_allocd_tmp >= 0); + /* We only decrement it while holding the GC lock. */ + /* Thus we can't accidentally adjust it down in more */ + /* than one thread simultaneously. */ + if (my_words_allocd_tmp != 0) { + (void)GC_atomic_add( + (volatile GC_word *)(&GC_words_allocd_tmp), + (GC_word)(-my_words_allocd_tmp)); + GC_words_allocd += my_words_allocd_tmp; + } + } + GC_acquire_mark_lock(); + ++ GC_fl_builder_count; + UNLOCK(); + ENABLE_SIGNALS(); + GC_release_mark_lock(); +# endif + op = GC_reclaim_generic(hbp, hhdr, lw, + ok -> ok_init, 0 COUNT_ARG); + if (op != 0) { +# ifdef NEED_TO_COUNT + /* We are neither gathering statistics, nor marking in */ + /* parallel. Thus GC_reclaim_generic doesn't count */ + /* for us. */ + for (p = op; p != 0; p = obj_link(p)) { + my_words_allocd += lw; + } +# endif +# if defined(GATHERSTATS) + /* We also reclaimed memory, so we need to adjust */ + /* that count. */ + /* This should be atomic, so the results may be */ + /* inaccurate. */ + GC_mem_found += my_words_allocd; +# endif +# ifdef PARALLEL_MARK + (void)GC_atomic_add( + (volatile GC_word *)(&GC_words_allocd_tmp), + (GC_word)(my_words_allocd)); + GC_acquire_mark_lock(); + -- GC_fl_builder_count; + if (GC_fl_builder_count == 0) GC_notify_all_builder(); + GC_release_mark_lock(); + return GC_clear_stack(op); +# else + GC_words_allocd += my_words_allocd; + goto out; +# endif + } +# ifdef PARALLEL_MARK + GC_acquire_mark_lock(); + -- GC_fl_builder_count; + if (GC_fl_builder_count == 0) GC_notify_all_builder(); + GC_release_mark_lock(); + DISABLE_SIGNALS(); + LOCK(); + /* GC lock is needed for reclaim list access. We */ + /* must decrement fl_builder_count before reaquiring GC */ + /* lock. Hopefully this path is rare. */ +# endif + } } - *opp = 0; - my_words_allocd = 0; - for (p = op; p != 0; p = obj_link(p)) { - my_words_allocd += lw; - if (my_words_allocd >= BODY_SZ) { + /* Next try to use prefix of global free list if there is one. */ + /* We don't refill it, but we need to use it up before allocating */ + /* a new block ourselves. */ + opp = &(GC_obj_kinds[k].ok_freelist[lw]); + if ( (op = *opp) != 0 ) { + *opp = 0; + my_words_allocd = 0; + for (p = op; p != 0; p = obj_link(p)) { + my_words_allocd += lw; + if (my_words_allocd >= BODY_SZ) { *opp = obj_link(p); obj_link(p) = 0; break; + } } + GC_words_allocd += my_words_allocd; + goto out; + } + /* Next try to allocate a new block worth of objects of this size. */ + { + struct hblk *h = GC_allochblk(lw, k, 0); + if (h != 0) { + if (IS_UNCOLLECTABLE(k)) GC_set_hdr_marks(HDR(h)); + GC_words_allocd += BYTES_TO_WORDS(HBLKSIZE) + - BYTES_TO_WORDS(HBLKSIZE) % lw; +# ifdef PARALLEL_MARK + GC_acquire_mark_lock(); + ++ GC_fl_builder_count; + UNLOCK(); + ENABLE_SIGNALS(); + GC_release_mark_lock(); +# endif + + op = GC_build_fl(h, lw, ok -> ok_init, 0); +# ifdef PARALLEL_MARK + GC_acquire_mark_lock(); + -- GC_fl_builder_count; + if (GC_fl_builder_count == 0) GC_notify_all_builder(); + GC_release_mark_lock(); + return GC_clear_stack(op); +# else + goto out; +# endif + } } - GC_words_allocd += my_words_allocd; -out: + /* As a last attempt, try allocating a single object. Note that */ + /* this may trigger a collection or expand the heap. */ + op = GC_generic_malloc_inner(lb, k); + if (0 != op) obj_link(op) = 0; + + out: UNLOCK(); ENABLE_SIGNALS(); - return(op); - + return(GC_clear_stack(op)); } -void * GC_malloc_many(size_t lb) +GC_PTR GC_malloc_many(size_t lb) { return(GC_generic_malloc_many(lb, NORMAL)); } @@ -272,11 +510,9 @@ DCL_LOCK_STATE; if( SMALL_OBJ(lb) ) { # ifdef MERGE_SIZES -# ifdef ADD_BYTE_AT_END - if (lb != 0) lb--; + if (EXTRA_BYTES != 0 && lb != 0) lb--; /* We don't need the extra byte, since this won't be */ /* collected anyway. */ -# endif lw = GC_size_map[lb]; # else lw = ALIGNED_WORDS(lb); @@ -338,11 +574,9 @@ DCL_LOCK_STATE; if( SMALL_OBJ(lb) ) { # ifdef MERGE_SIZES -# ifdef ADD_BYTE_AT_END - if (lb != 0) lb--; + if (EXTRA_BYTES != 0 && lb != 0) lb--; /* We don't need the extra byte, since this won't be */ /* collected anyway. */ -# endif lw = GC_size_map[lb]; # else lw = ALIGNED_WORDS(lb);