2 * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved.
3 * Copyright (c) 2001 by Hewlett-Packard Company. All rights reserved.
5 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
6 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
8 * Permission is hereby granted to use or copy this program
9 * for any purpose, provided the above notices are retained on all copies.
10 * Permission to modify the code and to distribute modified code is granted,
11 * provided the above notices are retained, and a notice that the code was
12 * modified is included with the above copyright notice.
16 /* Private declarations of GC marker data structures and macros */
19 * Declarations of mark stack. Needed by marker and client supplied mark
20 * routines. Transitively include gc_priv.h.
26 # include "private/config.h"
33 #if defined(KEEP_BACK_PTRS) || defined(PRINT_BLACK_LIST)
38 # include "../gc_mark.h"
45 /* The real declarations of the following is in gc_priv.h, so that */
46 /* we can avoid scanning the following table. */
48 mark_proc GC_mark_procs[MAX_MARK_PROCS];
51 #ifndef MARK_DESCR_OFFSET
52 # define MARK_DESCR_OFFSET sizeof(word)
56 * Mark descriptor stuff that should remain private for now, mostly
57 * because it's hard to export WORDSZ without including gcconfig.h.
59 #define BITMAP_BITS (WORDSZ - GC_DS_TAG_BITS)
61 (GC_mark_procs[((descr) >> GC_DS_TAG_BITS) & (GC_MAX_MARK_PROCS-1)])
63 ((descr) >> (GC_DS_TAG_BITS + GC_LOG_MAX_MARK_PROCS))
65 (((word)1 << (WORDSZ - GC_DS_TAG_BITS - GC_LOG_MAX_MARK_PROCS)) - 1)
67 GC_EXTERN unsigned GC_n_mark_procs;
69 /* Number of mark stack entries to discard on overflow. */
70 #define GC_MARK_STACK_DISCARDS (INITIAL_MARK_STACK_SIZE/8)
72 typedef struct GC_ms_entry {
73 ptr_t mse_start; /* First word of object, word aligned. */
74 GC_word mse_descr; /* Descriptor; low order two bits are tags, */
75 /* as described in gc_mark.h. */
78 GC_EXTERN size_t GC_mark_stack_size;
80 GC_EXTERN mse * GC_mark_stack_limit;
83 GC_EXTERN mse * volatile GC_mark_stack_top;
85 GC_EXTERN mse * GC_mark_stack_top;
88 GC_EXTERN mse * GC_mark_stack;
92 * Allow multiple threads to participate in the marking process.
93 * This works roughly as follows:
94 * The main mark stack never shrinks, but it can grow.
96 * The initiating threads holds the GC lock, and sets GC_help_wanted.
99 * 1) update helper_count (while holding mark_lock.)
100 * 2) allocate a local mark stack
102 * 3) Steal a global mark stack entry by atomically replacing
103 * its descriptor with 0.
104 * 4) Copy it to the local stack.
105 * 5) Mark on the local stack until it is empty, or
106 * it may be profitable to copy it back.
107 * 6) If necessary, copy local stack to global one,
109 * 7) Stop when the global mark stack is empty.
110 * 8) decrement helper_count (holding mark_lock).
112 * This is an experiment to see if we can do something along the lines
113 * of the University of Tokyo SGC in a less intrusive, though probably
114 * also less performant, way.
117 /* GC_mark_stack_top is protected by mark lock. */
120 * GC_notify_all_marker() is used when GC_help_wanted is first set,
121 * when the last helper becomes inactive,
122 * when something is added to the global mark stack, and just after
123 * GC_mark_no is incremented.
124 * This could be split into multiple CVs (and probably should be to
125 * scale to really large numbers of processors.)
127 #endif /* PARALLEL_MARK */
129 GC_INNER mse * GC_signal_mark_stack_overflow(mse *msp);
131 /* Push the object obj with corresponding heap block header hhdr onto */
132 /* the mark stack. */
133 #define PUSH_OBJ(obj, hhdr, mark_stack_top, mark_stack_limit) \
135 register word _descr = (hhdr) -> hb_descr; \
136 GC_ASSERT(!HBLK_IS_FREE(hhdr)); \
139 if (mark_stack_top >= mark_stack_limit) { \
140 mark_stack_top = GC_signal_mark_stack_overflow(mark_stack_top); \
142 mark_stack_top -> mse_start = (obj); \
143 mark_stack_top -> mse_descr = _descr; \
147 /* Push the contents of current onto the mark stack if it is a valid */
148 /* ptr to a currently unmarked object. Mark it. */
149 /* If we assumed a standard-conforming compiler, we could probably */
150 /* generate the exit_label transparently. */
151 #define PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \
152 source, exit_label) \
155 HC_GET_HDR(current, my_hhdr, source, exit_label); \
156 PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \
157 source, exit_label, my_hhdr, TRUE); \
161 /* Set mark bit, exit if it was already set. */
162 #ifdef USE_MARK_BYTES
163 /* There is a race here, and we may set */
164 /* the bit twice in the concurrent case. This can result in the */
165 /* object being pushed twice. But that's only a performance issue. */
166 # define SET_MARK_BIT_EXIT_IF_SET(hhdr,bit_no,exit_label) \
168 char * mark_byte_addr = (char *)hhdr -> hb_marks + (bit_no); \
169 if (*mark_byte_addr) goto exit_label; \
170 *mark_byte_addr = 1; \
173 # ifdef PARALLEL_MARK
174 /* This is used only if we explicitly set USE_MARK_BITS. */
175 /* The following may fail to exit even if the bit was already set. */
176 /* For our uses, that's benign: */
177 # define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \
179 if (!(*(addr) & (bits))) { \
180 AO_or((AO_t *)(addr), (bits)); \
186 # define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \
188 word old = *(addr); \
189 word my_bits = (bits); \
190 if (old & my_bits) goto exit_label; \
191 *(addr) = (old | my_bits); \
193 # endif /* !PARALLEL_MARK */
194 # define SET_MARK_BIT_EXIT_IF_SET(hhdr,bit_no,exit_label) \
196 word * mark_word_addr = hhdr -> hb_marks + divWORDSZ(bit_no); \
197 OR_WORD_EXIT_IF_SET(mark_word_addr, (word)1 << modWORDSZ(bit_no), \
200 #endif /* !USE_MARK_BYTES */
203 # define INCR_MARKS(hhdr) \
204 AO_store(&hhdr->hb_n_marks, AO_load(&hhdr->hb_n_marks) + 1)
206 # define INCR_MARKS(hhdr) (void)(++hhdr->hb_n_marks)
210 # define TRACE(source, cmd) \
211 if (GC_trace_addr != 0 && (ptr_t)(source) == GC_trace_addr) cmd
212 # define TRACE_TARGET(target, cmd) \
213 if (GC_trace_addr != 0 && (target) == *(ptr_t *)GC_trace_addr) cmd
215 # define TRACE(source, cmd)
216 # define TRACE_TARGET(source, cmd)
219 #if defined(I386) && defined(__GNUC__)
220 # define LONG_MULT(hprod, lprod, x, y) { \
221 __asm__ __volatile__("mull %2" : "=a"(lprod), "=d"(hprod) \
225 # define LONG_MULT(hprod, lprod, x, y) { \
226 unsigned long long prod = (unsigned long long)(x) \
227 * (unsigned long long)(y); \
228 hprod = prod >> 32; \
229 lprod = (unsigned32)prod; \
233 /* If the mark bit corresponding to current is not set, set it, and */
234 /* push the contents of the object on the mark stack. Current points */
235 /* to the beginning of the object. We rely on the fact that the */
236 /* preceding header calculation will succeed for a pointer past the */
237 /* first page of an object, only if it is in fact a valid pointer */
238 /* to the object. Thus we can omit the otherwise necessary tests */
239 /* here. Note in particular that the "displ" value is the displacement */
240 /* from the beginning of the heap block, which may itself be in the */
241 /* interior of a large object. */
242 #ifdef MARK_BIT_PER_GRANULE
243 # define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \
244 source, exit_label, hhdr, do_offset_check) \
246 size_t displ = HBLKDISPL(current); /* Displacement in block; in bytes. */\
247 /* displ is always within range. If current doesn't point to */ \
248 /* first block, then we are in the all_interior_pointers case, and */ \
249 /* it is safe to use any displacement value. */ \
250 size_t gran_displ = BYTES_TO_GRANULES(displ); \
251 size_t gran_offset = hhdr -> hb_map[gran_displ]; \
252 size_t byte_offset = displ & (GRANULE_BYTES - 1); \
253 ptr_t base = current; \
254 /* The following always fails for large block references. */ \
255 if (EXPECT((gran_offset | byte_offset) != 0, FALSE)) { \
256 if (hhdr -> hb_large_block) { \
257 /* gran_offset is bogus. */ \
259 base = (ptr_t)(hhdr -> hb_block); \
260 obj_displ = (ptr_t)(current) - base; \
261 if (obj_displ != displ) { \
262 GC_ASSERT(obj_displ < hhdr -> hb_sz); \
263 /* Must be in all_interior_pointer case, not first block */ \
264 /* already did validity check on cache miss. */ \
266 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \
267 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \
272 GC_ASSERT(hhdr -> hb_sz > HBLKSIZE || \
273 hhdr -> hb_block == HBLKPTR(current)); \
274 GC_ASSERT((ptr_t)(hhdr -> hb_block) <= (ptr_t) current); \
276 size_t obj_displ = GRANULES_TO_BYTES(gran_offset) \
278 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \
279 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \
282 gran_displ -= gran_offset; \
286 GC_ASSERT(hhdr == GC_find_header(base)); \
287 GC_ASSERT(gran_displ % BYTES_TO_GRANULES(hhdr -> hb_sz) == 0); \
288 TRACE(source, GC_log_printf("GC:%u: passed validity tests\n", \
289 (unsigned)GC_gc_no)); \
290 SET_MARK_BIT_EXIT_IF_SET(hhdr, gran_displ, exit_label); \
291 TRACE(source, GC_log_printf("GC:%u: previously unmarked\n", \
292 (unsigned)GC_gc_no)); \
294 GC_log_printf("GC:%u: marking %p from %p instead\n", \
295 (unsigned)GC_gc_no, base, source)); \
297 GC_STORE_BACK_PTR((ptr_t)source, base); \
298 PUSH_OBJ(base, hhdr, mark_stack_top, mark_stack_limit); \
300 #endif /* MARK_BIT_PER_GRANULE */
302 #ifdef MARK_BIT_PER_OBJ
303 # define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \
304 source, exit_label, hhdr, do_offset_check) \
306 size_t displ = HBLKDISPL(current); /* Displacement in block; in bytes. */\
307 unsigned32 low_prod, high_prod; \
308 unsigned32 inv_sz = hhdr -> hb_inv_sz; \
309 ptr_t base = current; \
310 LONG_MULT(high_prod, low_prod, displ, inv_sz); \
311 /* product is > and within sz_in_bytes of displ * sz_in_bytes * 2**32 */ \
312 if (EXPECT(low_prod >> 16 != 0, FALSE)) { \
313 FIXME: fails if offset is a multiple of HBLKSIZE which becomes 0 \
314 if (inv_sz == LARGE_INV_SZ) { \
316 base = (ptr_t)(hhdr -> hb_block); \
317 obj_displ = (ptr_t)(current) - base; \
318 if (obj_displ != displ) { \
319 GC_ASSERT(obj_displ < hhdr -> hb_sz); \
320 /* Must be in all_interior_pointer case, not first block */ \
321 /* already did validity check on cache miss. */ \
323 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \
324 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \
328 GC_ASSERT(hhdr -> hb_sz > HBLKSIZE || \
329 hhdr -> hb_block == HBLKPTR(current)); \
330 GC_ASSERT((ptr_t)(hhdr -> hb_block) < (ptr_t) current); \
332 /* Accurate enough if HBLKSIZE <= 2**15. */ \
333 GC_STATIC_ASSERT(HBLKSIZE <= (1 << 15)); \
334 size_t obj_displ = (((low_prod >> 16) + 1) * (hhdr->hb_sz)) >> 16; \
335 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \
336 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \
342 /* May get here for pointer to start of block not at */ \
343 /* beginning of object. If so, it's valid, and we're fine. */ \
344 GC_ASSERT(high_prod >= 0 && high_prod <= HBLK_OBJS(hhdr -> hb_sz)); \
345 TRACE(source, GC_log_printf("GC:%u: passed validity tests\n", \
346 (unsigned)GC_gc_no)); \
347 SET_MARK_BIT_EXIT_IF_SET(hhdr, high_prod, exit_label); \
348 TRACE(source, GC_log_printf("GC:%u: previously unmarked\n", \
349 (unsigned)GC_gc_no)); \
351 GC_log_printf("GC:%u: marking %p from %p instead\n", \
352 (unsigned)GC_gc_no, base, source)); \
354 GC_STORE_BACK_PTR((ptr_t)source, base); \
355 PUSH_OBJ(base, hhdr, mark_stack_top, mark_stack_limit); \
357 #endif /* MARK_BIT_PER_OBJ */
359 #if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
360 # define PUSH_ONE_CHECKED_STACK(p, source) \
361 GC_mark_and_push_stack((ptr_t)(p), (ptr_t)(source))
363 # define PUSH_ONE_CHECKED_STACK(p, source) \
364 GC_mark_and_push_stack((ptr_t)(p))
368 * Push a single value onto mark stack. Mark from the object pointed to by p.
369 * Invoke FIXUP_POINTER(p) before any further processing.
370 * P is considered valid even if it is an interior pointer.
371 * Previously marked objects are not pushed. Hence we make progress even
372 * if the mark stack overflows.
375 #if NEED_FIXUP_POINTER
376 /* Try both the raw version and the fixed up one. */
377 # define GC_PUSH_ONE_STACK(p, source) \
378 if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr \
379 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) { \
380 PUSH_ONE_CHECKED_STACK(p, source); \
383 if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr \
384 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) { \
385 PUSH_ONE_CHECKED_STACK(p, source); \
387 #else /* !NEED_FIXUP_POINTER */
388 # define GC_PUSH_ONE_STACK(p, source) \
389 if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr \
390 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) { \
391 PUSH_ONE_CHECKED_STACK(p, source); \
397 * As above, but interior pointer recognition as for
398 * normal heap pointers.
400 #define GC_PUSH_ONE_HEAP(p,source) \
402 if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr \
403 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) { \
404 GC_mark_stack_top = GC_mark_and_push( \
405 (void *)(p), GC_mark_stack_top, \
406 GC_mark_stack_limit, (void * *)(source)); \
409 /* Mark starting at mark stack entry top (incl.) down to */
410 /* mark stack entry bottom (incl.). Stop after performing */
411 /* about one page worth of work. Return the new mark stack */
413 GC_INNER mse * GC_mark_from(mse * top, mse * bottom, mse *limit);
415 #define MARK_FROM_MARK_STACK() \
416 GC_mark_stack_top = GC_mark_from(GC_mark_stack_top, \
418 GC_mark_stack + GC_mark_stack_size);
421 * Mark from one finalizable object using the specified
422 * mark proc. May not mark the object pointed to by
423 * real_ptr. That is the job of the caller, if appropriate.
424 * Note that this is called with the mutator running, but
425 * with us holding the allocation lock. This is safe only if the
426 * mutator needs the allocation lock to reveal hidden pointers.
427 * FIXME: Why do we need the GC_mark_state test below?
429 #define GC_MARK_FO(real_ptr, mark_proc) \
431 (*(mark_proc))(real_ptr); \
432 while (!GC_mark_stack_empty()) MARK_FROM_MARK_STACK(); \
433 if (GC_mark_state != MS_NONE) { \
434 GC_set_mark_bit(real_ptr); \
435 while (!GC_mark_some((ptr_t)0)) {} \
439 GC_EXTERN GC_bool GC_mark_stack_too_small;
440 /* We need a larger mark stack. May be */
441 /* set by client supplied mark routines.*/
443 typedef int mark_state_t; /* Current state of marking, as follows:*/
444 /* Used to remember where we are during */
445 /* concurrent marking. */
447 /* We say something is dirty if it was */
448 /* written since the last time we */
449 /* retrieved dirty bits. We say it's */
450 /* grungy if it was marked dirty in the */
451 /* last set of bits we retrieved. */
453 /* Invariant I: all roots and marked */
454 /* objects p are either dirty, or point */
455 /* to objects q that are either marked */
456 /* or a pointer to q appears in a range */
457 /* on the mark stack. */
459 #define MS_NONE 0 /* No marking in progress. I holds. */
460 /* Mark stack is empty. */
462 #define MS_PUSH_RESCUERS 1 /* Rescuing objects are currently */
463 /* being pushed. I holds, except */
464 /* that grungy roots may point to */
465 /* unmarked objects, as may marked */
466 /* grungy objects above scan_ptr. */
468 #define MS_PUSH_UNCOLLECTABLE 2 /* I holds, except that marked */
469 /* uncollectable objects above scan_ptr */
470 /* may point to unmarked objects. */
471 /* Roots may point to unmarked objects */
473 #define MS_ROOTS_PUSHED 3 /* I holds, mark stack may be nonempty */
475 #define MS_PARTIALLY_INVALID 4 /* I may not hold, e.g. because of M.S. */
476 /* overflow. However marked heap */
477 /* objects below scan_ptr point to */
478 /* marked or stacked objects. */
480 #define MS_INVALID 5 /* I may not hold. */
482 GC_EXTERN mark_state_t GC_mark_state;
484 #endif /* GC_PMARK_H */