1 /* 2 * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved. 3 * Copyright (c) 2001 by Hewlett-Packard Company. All rights reserved. 4 * 5 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED 6 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK. 7 * 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. 13 * 14 */ 15 16 /* Private declarations of GC marker data structures and macros */ 17 18 /* 19 * Declarations of mark stack. Needed by marker and client supplied mark 20 * routines. Transitively include gc_priv.h. 21 * (Note that gc_priv.h should not be included before this, since this 22 * includes dbg_mlc.h, which wants to include gc_priv.h AFTER defining 23 * I_HIDE_POINTERS.) 24 */ 25 #ifndef GC_PMARK_H 26 # define GC_PMARK_H 27 28 # if defined(KEEP_BACK_PTRS) || defined(PRINT_BLACK_LIST) 29 # include "dbg_mlc.h" 30 # endif 31 # ifndef GC_MARK_H 32 # include "../gc_mark.h" 33 # endif 34 # ifndef GC_PRIVATE_H 35 # include "gc_priv.h" 36 # endif 37 38 /* The real declarations of the following is in gc_priv.h, so that */ 39 /* we can avoid scanning the following table. */ 40 /* 41 extern mark_proc GC_mark_procs[MAX_MARK_PROCS]; 42 */ 43 44 /* 45 * Mark descriptor stuff that should remain private for now, mostly 46 * because it's hard to export WORDSZ without including gcconfig.h. 47 */ 48 # define BITMAP_BITS (WORDSZ - GC_DS_TAG_BITS) 49 # define PROC(descr) \ 50 (GC_mark_procs[((descr) >> GC_DS_TAG_BITS) & (GC_MAX_MARK_PROCS-1)]) 51 # define ENV(descr) \ 52 ((descr) >> (GC_DS_TAG_BITS + GC_LOG_MAX_MARK_PROCS)) 53 # define MAX_ENV \ 54 (((word)1 << (WORDSZ - GC_DS_TAG_BITS - GC_LOG_MAX_MARK_PROCS)) - 1) 55 56 57 extern unsigned GC_n_mark_procs; 58 59 /* Number of mark stack entries to discard on overflow. */ 60 #define GC_MARK_STACK_DISCARDS (INITIAL_MARK_STACK_SIZE/8) 61 62 typedef struct GC_ms_entry { 63 ptr_t mse_start; /* First word of object */ 64 GC_word mse_descr; /* Descriptor; low order two bits are tags, */ 65 /* identifying the upper 30 bits as one of the */ 66 /* following: */ 67 } mse; 68 69 extern size_t GC_mark_stack_size; 70 71 extern mse * GC_mark_stack_limit; 72 73 #ifdef PARALLEL_MARK 74 extern mse * volatile GC_mark_stack_top; 75 #else 76 extern mse * GC_mark_stack_top; 77 #endif 78 79 extern mse * GC_mark_stack; 80 81 #ifdef PARALLEL_MARK 82 /* 83 * Allow multiple threads to participate in the marking process. 84 * This works roughly as follows: 85 * The main mark stack never shrinks, but it can grow. 86 * 87 * The initiating threads holds the GC lock, and sets GC_help_wanted. 88 * 89 * Other threads: 90 * 1) update helper_count (while holding mark_lock.) 91 * 2) allocate a local mark stack 92 * repeatedly: 93 * 3) Steal a global mark stack entry by atomically replacing 94 * its descriptor with 0. 95 * 4) Copy it to the local stack. 96 * 5) Mark on the local stack until it is empty, or 97 * it may be profitable to copy it back. 98 * 6) If necessary, copy local stack to global one, 99 * holding mark lock. 100 * 7) Stop when the global mark stack is empty. 101 * 8) decrement helper_count (holding mark_lock). 102 * 103 * This is an experiment to see if we can do something along the lines 104 * of the University of Tokyo SGC in a less intrusive, though probably 105 * also less performant, way. 106 */ 107 void GC_do_parallel_mark(); 108 /* inititate parallel marking. */ 109 110 extern GC_bool GC_help_wanted; /* Protected by mark lock */ 111 extern unsigned GC_helper_count; /* Number of running helpers. */ 112 /* Protected by mark lock */ 113 extern unsigned GC_active_count; /* Number of active helpers. */ 114 /* Protected by mark lock */ 115 /* May increase and decrease */ 116 /* within each mark cycle. But */ 117 /* once it returns to 0, it */ 118 /* stays zero for the cycle. */ 119 /* GC_mark_stack_top is also protected by mark lock. */ 120 /* 121 * GC_notify_all_marker() is used when GC_help_wanted is first set, 122 * when the last helper becomes inactive, 123 * when something is added to the global mark stack, and just after 124 * GC_mark_no is incremented. 125 * This could be split into multiple CVs (and probably should be to 126 * scale to really large numbers of processors.) 127 */ 128 #endif /* PARALLEL_MARK */ 129 130 /* Return a pointer to within 1st page of object. */ 131 /* Set *new_hdr_p to corr. hdr. */ 132 ptr_t GC_find_start(ptr_t current, hdr *hhdr, hdr **new_hdr_p); 133 134 mse * GC_signal_mark_stack_overflow(mse *msp); 135 136 /* Push the object obj with corresponding heap block header hhdr onto */ 137 /* the mark stack. */ 138 # define PUSH_OBJ(obj, hhdr, mark_stack_top, mark_stack_limit) \ 139 { \ 140 register word _descr = (hhdr) -> hb_descr; \ 141 \ 142 if (_descr != 0) { \ 143 mark_stack_top++; \ 144 if (mark_stack_top >= mark_stack_limit) { \ 145 mark_stack_top = GC_signal_mark_stack_overflow(mark_stack_top); \ 146 } \ 147 mark_stack_top -> mse_start = (obj); \ 148 mark_stack_top -> mse_descr = _descr; \ 149 } \ 150 } 151 152 /* Push the contents of current onto the mark stack if it is a valid */ 153 /* ptr to a currently unmarked object. Mark it. */ 154 /* If we assumed a standard-conforming compiler, we could probably */ 155 /* generate the exit_label transparently. */ 156 # define PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \ 157 source, exit_label) \ 158 { \ 159 hdr * my_hhdr; \ 160 \ 161 HC_GET_HDR(current, my_hhdr, source, exit_label); \ 162 PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \ 163 source, exit_label, my_hhdr, TRUE); \ 164 exit_label: ; \ 165 } 166 167 /* Set mark bit, exit if it was already set. */ 168 169 # ifdef USE_MARK_BITS 170 # ifdef PARALLEL_MARK 171 /* The following may fail to exit even if the bit was already set. */ 172 /* For our uses, that's benign: */ 173 # define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \ 174 { \ 175 if (!(*(addr) & (mask))) { \ 176 AO_or((AO_t *)(addr), (mask); \ 177 } else { \ 178 goto label; \ 179 } \ 180 } 181 # else 182 # define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \ 183 { \ 184 word old = *(addr); \ 185 word my_bits = (bits); \ 186 if (old & my_bits) goto exit_label; \ 187 *(addr) = (old | my_bits); \ 188 } 189 # endif /* !PARALLEL_MARK */ 190 # define SET_MARK_BIT_EXIT_IF_SET(hhdr,bit_no,exit_label) \ 191 { \ 192 word * mark_word_addr = hhdr -> hb_marks + divWORDSZ(bit_no); \ 193 \ 194 OR_WORD_EXIT_IF_SET(mark_word_addr, (word)1 << modWORDSZ(bit_no), \ 195 exit_label); \ 196 } 197 # endif 198 199 200 #ifdef USE_MARK_BYTES 201 # if defined(I386) && defined(__GNUC__) 202 # define LONG_MULT(hprod, lprod, x, y) { \ 203 asm("mull %2" : "=a"(lprod), "=d"(hprod) : "g"(y), "0"(x)); \ 204 } 205 # else /* No in-line X86 assembly code */ 206 # define LONG_MULT(hprod, lprod, x, y) { \ 207 unsigned long long prod = (unsigned long long)x \ 208 * (unsigned long long)y; \ 209 hprod = prod >> 32; \ 210 lprod = (unsigned32)prod; \ 211 } 212 # endif 213 214 /* There is a race here, and we may set */ 215 /* the bit twice in the concurrent case. This can result in the */ 216 /* object being pushed twice. But that's only a performance issue. */ 217 # define SET_MARK_BIT_EXIT_IF_SET(hhdr,bit_no,exit_label) \ 218 { \ 219 char * mark_byte_addr = (char *)hhdr -> hb_marks + (bit_no); \ 220 char mark_byte = *mark_byte_addr; \ 221 \ 222 if (mark_byte) goto exit_label; \ 223 *mark_byte_addr = 1; \ 224 } 225 #endif /* USE_MARK_BYTES */ 226 227 #ifdef PARALLEL_MARK 228 # define INCR_MARKS(hhdr) \ 229 AO_store(&(hhdr -> hb_n_marks), AO_load(&(hhdr -> hb_n_marks))+1); 230 #else 231 # define INCR_MARKS(hhdr) ++(hhdr -> hb_n_marks) 232 #endif 233 234 #ifdef ENABLE_TRACE 235 # define TRACE(source, cmd) \ 236 if (GC_trace_addr != 0 && (ptr_t)(source) == GC_trace_addr) cmd 237 # define TRACE_TARGET(target, cmd) \ 238 if (GC_trace_addr != 0 && (target) == *(ptr_t *)GC_trace_addr) cmd 239 #else 240 # define TRACE(source, cmd) 241 # define TRACE_TARGET(source, cmd) 242 #endif 243 /* If the mark bit corresponding to current is not set, set it, and */ 244 /* push the contents of the object on the mark stack. Current points */ 245 /* to the bginning of the object. We rely on the fact that the */ 246 /* preceding header calculation will succeed for a pointer past the */ 247 /* forst page of an object, only if it is in fact a valid pointer */ 248 /* to the object. Thus we can omit the otherwise necessary tests */ 249 /* here. Note in particular tha the "displ" value is the displacement */ 250 /* from the beggining of the heap block, which may itself be in the */ 251 /* interior of a large object. */ 252 #ifdef MARK_BIT_PER_GRANULE 253 # define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \ 254 source, exit_label, hhdr, do_offset_check) \ 255 { \ 256 size_t displ = HBLKDISPL(current); /* Displacement in block; in bytes. */\ 257 /* displ is always within range. If current doesn't point to */ \ 258 /* first block, then we are in the all_interior_pointers case, and */ \ 259 /* it is safe to use any displacement value. */ \ 260 size_t gran_displ = BYTES_TO_GRANULES(displ); \ 261 size_t gran_offset = hhdr -> hb_map[gran_displ]; \ 262 size_t byte_offset = displ & (GRANULE_BYTES - 1); \ 263 ptr_t base = current; \ 264 /* The following always fails for large block references. */ \ 265 if (EXPECT((gran_offset | byte_offset) != 0, FALSE)) { \ 266 if (hhdr -> hb_large_block) { \ 267 /* gran_offset is bogus. */ \ 268 size_t obj_displ; \ 269 base = (ptr_t)(hhdr -> hb_block); \ 270 obj_displ = (ptr_t)(current) - base; \ 271 if (obj_displ != displ) { \ 272 GC_ASSERT(obj_displ < hhdr -> hb_sz); \ 273 /* Must be in all_interior_pointer case, not first block */ \ 274 /* already did validity check on cache miss. */ \ 275 ; \ 276 } else { \ 277 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \ 278 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \ 279 goto exit_label; \ 280 } \ 281 } \ 282 gran_displ = 0; \ 283 GC_ASSERT(hhdr -> hb_sz > HBLKSIZE || \ 284 hhdr -> hb_block == HBLKPTR(current)); \ 285 GC_ASSERT((ptr_t)(hhdr -> hb_block) <= (ptr_t) current); \ 286 } else { \ 287 size_t obj_displ = GRANULES_TO_BYTES(gran_offset) \ 288 + byte_offset; \ 289 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \ 290 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \ 291 goto exit_label; \ 292 } \ 293 gran_displ -= gran_offset; \ 294 base -= obj_displ; \ 295 } \ 296 } \ 297 GC_ASSERT(hhdr == GC_find_header(base)); \ 298 GC_ASSERT(gran_displ % BYTES_TO_GRANULES(hhdr -> hb_sz) == 0); \ 299 TRACE(source, GC_log_printf("GC:%d: passed validity tests\n",GC_gc_no)); \ 300 SET_MARK_BIT_EXIT_IF_SET(hhdr, gran_displ, exit_label); \ 301 TRACE(source, GC_log_printf("GC:%d: previously unmarked\n",GC_gc_no)); \ 302 TRACE_TARGET(base, \ 303 GC_log_printf("GC:%d: marking %p from %p instead\n", GC_gc_no, \ 304 base, source)); \ 305 INCR_MARKS(hhdr); \ 306 GC_STORE_BACK_PTR((ptr_t)source, base); \ 307 PUSH_OBJ(base, hhdr, mark_stack_top, mark_stack_limit); \ 308 } 309 #endif /* MARK_BIT_PER_GRANULE */ 310 311 #ifdef MARK_BIT_PER_OBJ 312 # define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \ 313 source, exit_label, hhdr, do_offset_check) \ 314 { \ 315 size_t displ = HBLKDISPL(current); /* Displacement in block; in bytes. */\ 316 unsigned32 low_prod, high_prod, offset_fraction; \ 317 unsigned32 inv_sz = hhdr -> hb_inv_sz; \ 318 ptr_t base = current; \ 319 LONG_MULT(high_prod, low_prod, displ, inv_sz); \ 320 /* product is > and within sz_in_bytes of displ * sz_in_bytes * 2**32 */ \ 321 if (EXPECT(low_prod >> 16 != 0, FALSE)) { \ 322 FIXME: fails if offset is a multiple of HBLKSIZE which becomes 0 \ 323 if (inv_sz == LARGE_INV_SZ) { \ 324 size_t obj_displ; \ 325 base = (ptr_t)(hhdr -> hb_block); \ 326 obj_displ = (ptr_t)(current) - base; \ 327 if (obj_displ != displ) { \ 328 GC_ASSERT(obj_displ < hhdr -> hb_sz); \ 329 /* Must be in all_interior_pointer case, not first block */ \ 330 /* already did validity check on cache miss. */ \ 331 ; \ 332 } else { \ 333 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \ 334 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \ 335 goto exit_label; \ 336 } \ 337 } \ 338 GC_ASSERT(hhdr -> hb_sz > HBLKSIZE || \ 339 hhdr -> hb_block == HBLKPTR(current)); \ 340 GC_ASSERT((ptr_t)(hhdr -> hb_block) < (ptr_t) current); \ 341 } else { \ 342 /* Accurate enough if HBLKSIZE <= 2**15. */ \ 343 GC_ASSERT(HBLKSIZE <= (1 << 15)); \ 344 size_t obj_displ = (((low_prod >> 16) + 1) * (hhdr -> hb_sz)) >> 16; \ 345 if (do_offset_check && !GC_valid_offsets[obj_displ]) { \ 346 GC_ADD_TO_BLACK_LIST_NORMAL(current, source); \ 347 goto exit_label; \ 348 } \ 349 base -= obj_displ; \ 350 } \ 351 } \ 352 /* May get here for pointer to start of block not at */ \ 353 /* beginning of object. If so, it's valid, and we're fine. */ \ 354 GC_ASSERT(high_prod >= 0 && high_prod <= HBLK_OBJS(hhdr -> hb_sz)); \ 355 TRACE(source, GC_log_printf("GC:%d: passed validity tests\n",GC_gc_no)); \ 356 SET_MARK_BIT_EXIT_IF_SET(hhdr, high_prod, exit_label); \ 357 TRACE(source, GC_log_printf("GC:%d: previously unmarked\n",GC_gc_no)); \ 358 TRACE_TARGET(base, \ 359 GC_log_printf("GC:%d: marking %p from %p instead\n", GC_gc_no, \ 360 base, source)); \ 361 INCR_MARKS(hhdr); \ 362 GC_STORE_BACK_PTR((ptr_t)source, base); \ 363 PUSH_OBJ(base, hhdr, mark_stack_top, mark_stack_limit); \ 364 } 365 #endif /* MARK_BIT_PER_OBJ */ 366 367 #if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS) 368 # define PUSH_ONE_CHECKED_STACK(p, source) \ 369 GC_mark_and_push_stack(p, (ptr_t)(source)) 370 #else 371 # define PUSH_ONE_CHECKED_STACK(p, source) \ 372 GC_mark_and_push_stack(p) 373 #endif 374 375 /* 376 * Push a single value onto mark stack. Mark from the object pointed to by p. 377 * Invoke FIXUP_POINTER(p) before any further processing. 378 * P is considered valid even if it is an interior pointer. 379 * Previously marked objects are not pushed. Hence we make progress even 380 * if the mark stack overflows. 381 */ 382 383 # if NEED_FIXUP_POINTER 384 /* Try both the raw version and the fixed up one. */ 385 # define GC_PUSH_ONE_STACK(p, source) \ 386 if ((p) >= (ptr_t)GC_least_plausible_heap_addr \ 387 && (p) < (ptr_t)GC_greatest_plausible_heap_addr) { \ 388 PUSH_ONE_CHECKED_STACK(p, source); \ 389 } \ 390 FIXUP_POINTER(p); \ 391 if ((p) >= (ptr_t)GC_least_plausible_heap_addr \ 392 && (p) < (ptr_t)GC_greatest_plausible_heap_addr) { \ 393 PUSH_ONE_CHECKED_STACK(p, source); \ 394 } 395 # else /* !NEED_FIXUP_POINTER */ 396 # define GC_PUSH_ONE_STACK(p, source) \ 397 if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr \ 398 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) { \ 399 PUSH_ONE_CHECKED_STACK(p, source); \ 400 } 401 # endif 402 403 404 /* 405 * As above, but interior pointer recognition as for 406 * normal heap pointers. 407 */ 408 # define GC_PUSH_ONE_HEAP(p,source) \ 409 FIXUP_POINTER(p); \ 410 if ((p) >= (ptr_t)GC_least_plausible_heap_addr \ 411 && (p) < (ptr_t)GC_greatest_plausible_heap_addr) { \ 412 GC_mark_stack_top = GC_mark_and_push( \ 413 (void *)(p), GC_mark_stack_top, \ 414 GC_mark_stack_limit, (void * *)(source)); \ 415 } 416 417 /* Mark starting at mark stack entry top (incl.) down to */ 418 /* mark stack entry bottom (incl.). Stop after performing */ 419 /* about one page worth of work. Return the new mark stack */ 420 /* top entry. */ 421 mse * GC_mark_from(mse * top, mse * bottom, mse *limit); 422 423 #define MARK_FROM_MARK_STACK() \ 424 GC_mark_stack_top = GC_mark_from(GC_mark_stack_top, \ 425 GC_mark_stack, \ 426 GC_mark_stack + GC_mark_stack_size); 427 428 /* 429 * Mark from one finalizable object using the specified 430 * mark proc. May not mark the object pointed to by 431 * real_ptr. That is the job of the caller, if appropriate. 432 * Note that this is called with the mutator running, but 433 * with us holding the allocation lock. This is safe only if the 434 * mutator needs tha allocation lock to reveal hidden pointers. 435 * FIXME: Why do we need the GC_mark_state test below? 436 */ 437 # define GC_MARK_FO(real_ptr, mark_proc) \ 438 { \ 439 (*(mark_proc))(real_ptr); \ 440 while (!GC_mark_stack_empty()) MARK_FROM_MARK_STACK(); \ 441 if (GC_mark_state != MS_NONE) { \ 442 GC_set_mark_bit(real_ptr); \ 443 while (!GC_mark_some((ptr_t)0)) {} \ 444 } \ 445 } 446 447 extern GC_bool GC_mark_stack_too_small; 448 /* We need a larger mark stack. May be */ 449 /* set by client supplied mark routines.*/ 450 451 typedef int mark_state_t; /* Current state of marking, as follows:*/ 452 /* Used to remember where we are during */ 453 /* concurrent marking. */ 454 455 /* We say something is dirty if it was */ 456 /* written since the last time we */ 457 /* retrieved dirty bits. We say it's */ 458 /* grungy if it was marked dirty in the */ 459 /* last set of bits we retrieved. */ 460 461 /* Invariant I: all roots and marked */ 462 /* objects p are either dirty, or point */ 463 /* to objects q that are either marked */ 464 /* or a pointer to q appears in a range */ 465 /* on the mark stack. */ 466 467 # define MS_NONE 0 /* No marking in progress. I holds. */ 468 /* Mark stack is empty. */ 469 470 # define MS_PUSH_RESCUERS 1 /* Rescuing objects are currently */ 471 /* being pushed. I holds, except */ 472 /* that grungy roots may point to */ 473 /* unmarked objects, as may marked */ 474 /* grungy objects above scan_ptr. */ 475 476 # define MS_PUSH_UNCOLLECTABLE 2 477 /* I holds, except that marked */ 478 /* uncollectable objects above scan_ptr */ 479 /* may point to unmarked objects. */ 480 /* Roots may point to unmarked objects */ 481 482 # define MS_ROOTS_PUSHED 3 /* I holds, mark stack may be nonempty */ 483 484 # define MS_PARTIALLY_INVALID 4 /* I may not hold, e.g. because of M.S. */ 485 /* overflow. However marked heap */ 486 /* objects below scan_ptr point to */ 487 /* marked or stacked objects. */ 488 489 # define MS_INVALID 5 /* I may not hold. */ 490 491 extern mark_state_t GC_mark_state; 492 493 #endif /* GC_PMARK_H */ 494 495