boehm-gc/include/private/gc_pmark.h

   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 # include "gc.h"        /* For configuration */
  29
  30 # if defined(KEEP_BACK_PTRS) || defined(PRINT_BLACK_LIST)
  31 #   include "dbg_mlc.h"
  32 # endif
  33 # ifndef GC_MARK_H
  34 #   include "../gc_mark.h"
  35 # endif
  36 # ifndef GC_PRIVATE_H
  37 #   include "gc_priv.h"
  38 # endif
  39
  40 /* The real declarations of the following is in gc_priv.h, so that      */
  41 /* we can avoid scanning the following table.                           */
  42 /*
  43 extern mark_proc GC_mark_procs[MAX_MARK_PROCS];
  44 */
  45
  46 /*
  47  * Mark descriptor stuff that should remain private for now, mostly
  48  * because it's hard to export WORDSZ without including gcconfig.h.
  49  */
  50 # define BITMAP_BITS (WORDSZ - GC_DS_TAG_BITS)
  51 # define PROC(descr) \
  52         (GC_mark_procs[((descr) >> GC_DS_TAG_BITS) & (GC_MAX_MARK_PROCS-1)])
  53 # define ENV(descr) \
  54         ((descr) >> (GC_DS_TAG_BITS + GC_LOG_MAX_MARK_PROCS))
  55 # define MAX_ENV \
  56         (((word)1 << (WORDSZ - GC_DS_TAG_BITS - GC_LOG_MAX_MARK_PROCS)) - 1)
  57
  58
  59 extern word GC_n_mark_procs;
  60
  61 /* Number of mark stack entries to discard on overflow. */
  62 #define GC_MARK_STACK_DISCARDS (INITIAL_MARK_STACK_SIZE/8)
  63
  64 typedef struct GC_ms_entry {
  65     GC_word * mse_start;   /* First word of object */
  66     GC_word mse_descr;  /* Descriptor; low order two bits are tags,     */
  67                         /* identifying the upper 30 bits as one of the  */
  68                         /* following:                                   */
  69 } mse;
  70
  71 extern word GC_mark_stack_size;
  72
  73 extern mse * GC_mark_stack_limit;
  74
  75 #ifdef PARALLEL_MARK
  76   extern mse * VOLATILE GC_mark_stack_top;
  77 #else
  78   extern mse * GC_mark_stack_top;
  79 #endif
  80
  81 extern mse * GC_mark_stack;
  82
  83 #ifdef PARALLEL_MARK
  84     /*
  85      * Allow multiple threads to participate in the marking process.
  86      * This works roughly as follows:
  87      *  The main mark stack never shrinks, but it can grow.
  88      *
  89      *  The initiating threads holds the GC lock, and sets GC_help_wanted.
  90      *
  91      *  Other threads:
  92      *     1) update helper_count (while holding mark_lock.)
  93      *     2) allocate a local mark stack
  94      *     repeatedly:
  95      *          3) Steal a global mark stack entry by atomically replacing
  96      *             its descriptor with 0.
  97      *          4) Copy it to the local stack.
  98      *          5) Mark on the local stack until it is empty, or
  99      *             it may be profitable to copy it back.
 100      *          6) If necessary, copy local stack to global one,
 101      *             holding mark lock.
 102      *    7) Stop when the global mark stack is empty.
 103      *    8) decrement helper_count (holding mark_lock).
 104      *
 105      * This is an experiment to see if we can do something along the lines
 106      * of the University of Tokyo SGC in a less intrusive, though probably
 107      * also less performant, way.
 108      */
 109     void GC_do_parallel_mark();
 110                 /* inititate parallel marking.  */
 111
 112     extern GC_bool GC_help_wanted;      /* Protected by mark lock       */
 113     extern unsigned GC_helper_count;    /* Number of running helpers.   */
 114                                         /* Protected by mark lock       */
 115     extern unsigned GC_active_count;    /* Number of active helpers.    */
 116                                         /* Protected by mark lock       */
 117                                         /* May increase and decrease    */
 118                                         /* within each mark cycle.  But */
 119                                         /* once it returns to 0, it     */
 120                                         /* stays zero for the cycle.    */
 121     /* GC_mark_stack_top is also protected by mark lock.        */
 122     extern mse * VOLATILE GC_first_nonempty;
 123                                         /* Lowest entry on mark stack   */
 124                                         /* that may be nonempty.        */
 125                                         /* Updated only by initiating   */
 126                                         /* thread.                      */
 127     /*
 128      * GC_notify_all_marker() is used when GC_help_wanted is first set,
 129      * when the last helper becomes inactive,
 130      * when something is added to the global mark stack, and just after
 131      * GC_mark_no is incremented.
 132      * This could be split into multiple CVs (and probably should be to
 133      * scale to really large numbers of processors.)
 134      */
 135 #endif /* PARALLEL_MARK */
 136
 137 /* Return a pointer to within 1st page of object.       */
 138 /* Set *new_hdr_p to corr. hdr.                         */
 139 #ifdef __STDC__
 140   ptr_t GC_find_start(ptr_t current, hdr *hhdr, hdr **new_hdr_p);
 141 #else
 142   ptr_t GC_find_start();
 143 #endif
 144
 145 mse * GC_signal_mark_stack_overflow GC_PROTO((mse *msp));
 146
 147 # ifdef GATHERSTATS
 148 #   define ADD_TO_ATOMIC(sz) GC_atomic_in_use += (sz)
 149 #   define ADD_TO_COMPOSITE(sz) GC_composite_in_use += (sz)
 150 # else
 151 #   define ADD_TO_ATOMIC(sz)
 152 #   define ADD_TO_COMPOSITE(sz)
 153 # endif
 154
 155 /* Push the object obj with corresponding heap block header hhdr onto   */
 156 /* the mark stack.                                                      */
 157 # define PUSH_OBJ(obj, hhdr, mark_stack_top, mark_stack_limit) \
 158 { \
 159     register word _descr = (hhdr) -> hb_descr; \
 160         \
 161     if (_descr == 0) { \
 162         ADD_TO_ATOMIC((hhdr) -> hb_sz); \
 163     } else { \
 164         ADD_TO_COMPOSITE((hhdr) -> hb_sz); \
 165         mark_stack_top++; \
 166         if (mark_stack_top >= mark_stack_limit) { \
 167           mark_stack_top = GC_signal_mark_stack_overflow(mark_stack_top); \
 168         } \
 169         mark_stack_top -> mse_start = (obj); \
 170         mark_stack_top -> mse_descr = _descr; \
 171     } \
 172 }
 173
 174 /* Push the contents of current onto the mark stack if it is a valid    */
 175 /* ptr to a currently unmarked object.  Mark it.                        */
 176 /* If we assumed a standard-conforming compiler, we could probably      */
 177 /* generate the exit_label transparently.                               */
 178 # define PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \
 179                        source, exit_label) \
 180 { \
 181     hdr * my_hhdr; \
 182     ptr_t my_current = current; \
 183  \
 184     GET_HDR(my_current, my_hhdr); \
 185     if (IS_FORWARDING_ADDR_OR_NIL(my_hhdr)) { \
 186          hdr * new_hdr = GC_invalid_header; \
 187          my_current = GC_find_start(my_current, my_hhdr, &new_hdr); \
 188          my_hhdr = new_hdr; \
 189     } \
 190     PUSH_CONTENTS_HDR(my_current, mark_stack_top, mark_stack_limit, \
 191                   source, exit_label, my_hhdr); \
 192 exit_label: ; \
 193 }
 194
 195 /* As above, but use header cache for header lookup.    */
 196 # define HC_PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \
 197                        source, exit_label) \
 198 { \
 199     hdr * my_hhdr; \
 200     ptr_t my_current = current; \
 201  \
 202     HC_GET_HDR(my_current, my_hhdr, source); \
 203     PUSH_CONTENTS_HDR(my_current, mark_stack_top, mark_stack_limit, \
 204                   source, exit_label, my_hhdr); \
 205 exit_label: ; \
 206 }
 207
 208 /* Set mark bit, exit if it was already set.    */
 209
 210 # ifdef USE_MARK_BYTES
 211     /* Unlike the mark bit case, there is a race here, and we may set   */
 212     /* the bit twice in the concurrent case.  This can result in the    */
 213     /* object being pushed twice.  But that's only a performance issue. */
 214 #   define SET_MARK_BIT_EXIT_IF_SET(hhdr,displ,exit_label) \
 215     { \
 216         register VOLATILE char * mark_byte_addr = \
 217                                 hhdr -> hb_marks + ((displ) >> 1); \
 218         register char mark_byte = *mark_byte_addr; \
 219           \
 220         if (mark_byte) goto exit_label; \
 221         *mark_byte_addr = 1;  \
 222     }
 223 # else
 224 #   define SET_MARK_BIT_EXIT_IF_SET(hhdr,displ,exit_label) \
 225     { \
 226         register word * mark_word_addr = hhdr -> hb_marks + divWORDSZ(displ); \
 227           \
 228         OR_WORD_EXIT_IF_SET(mark_word_addr, (word)1 << modWORDSZ(displ), \
 229                             exit_label); \
 230     }
 231 # endif /* USE_MARK_BYTES */
 232
 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.  For a small      */
 235 /* object we assume that current is the (possibly interior) pointer     */
 236 /* to the object.  For large objects we assume that current points      */
 237 /* to somewhere inside the first page of the object.  If                */
 238 /* GC_all_interior_pointers is set, it may have been previously         */
 239 /* adjusted to make that true.                                          */
 240 # define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \
 241                            source, exit_label, hhdr) \
 242 { \
 243     int displ;  /* Displacement in block; first bytes, then words */ \
 244     int map_entry; \
 245     \
 246     displ = HBLKDISPL(current); \
 247     map_entry = MAP_ENTRY((hhdr -> hb_map), displ); \
 248     displ = BYTES_TO_WORDS(displ); \
 249     if (map_entry > CPP_MAX_OFFSET) { \
 250         if (map_entry == OFFSET_TOO_BIG) { \
 251           map_entry = displ % (hhdr -> hb_sz); \
 252           displ -= map_entry; \
 253           if (displ + (hhdr -> hb_sz) > BYTES_TO_WORDS(HBLKSIZE)) { \
 254             GC_ADD_TO_BLACK_LIST_NORMAL((word)current, source); \
 255             goto exit_label; \
 256           } \
 257         } else { \
 258           GC_ADD_TO_BLACK_LIST_NORMAL((word)current, source); goto exit_label; \
 259         } \
 260     } else { \
 261         displ -= map_entry; \
 262     } \
 263     GC_ASSERT(displ >= 0 && displ < MARK_BITS_PER_HBLK); \
 264     SET_MARK_BIT_EXIT_IF_SET(hhdr, displ, exit_label); \
 265     GC_STORE_BACK_PTR((ptr_t)source, (ptr_t)HBLKPTR(current) \
 266                                       + WORDS_TO_BYTES(displ)); \
 267     PUSH_OBJ(((word *)(HBLKPTR(current)) + displ), hhdr, \
 268              mark_stack_top, mark_stack_limit) \
 269 }
 270
 271 #if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
 272 #   define PUSH_ONE_CHECKED_STACK(p, source) \
 273         GC_mark_and_push_stack(p, (ptr_t)(source))
 274 #else
 275 #   define PUSH_ONE_CHECKED_STACK(p, source) \
 276         GC_mark_and_push_stack(p)
 277 #endif
 278
 279 /*
 280  * Push a single value onto mark stack. Mark from the object pointed to by p.
 281  * Invoke FIXUP_POINTER(p) before any further processing.
 282  * P is considered valid even if it is an interior pointer.
 283  * Previously marked objects are not pushed.  Hence we make progress even
 284  * if the mark stack overflows.
 285  */
 286
 287 # if NEED_FIXUP_POINTER
 288     /* Try both the raw version and the fixed up one.   */
 289 #   define GC_PUSH_ONE_STACK(p, source) \
 290       if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr     \
 291          && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {      \
 292          PUSH_ONE_CHECKED_STACK(p, source);     \
 293       } \
 294       FIXUP_POINTER(p); \
 295       if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr     \
 296          && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {      \
 297          PUSH_ONE_CHECKED_STACK(p, source);     \
 298       }
 299 # else /* !NEED_FIXUP_POINTER */
 300 #   define GC_PUSH_ONE_STACK(p, source) \
 301       if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr     \
 302          && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {      \
 303          PUSH_ONE_CHECKED_STACK(p, source);     \
 304       }
 305 # endif
 306
 307
 308 /*
 309  * As above, but interior pointer recognition as for
 310  * normal for heap pointers.
 311  */
 312 # define GC_PUSH_ONE_HEAP(p,source) \
 313     FIXUP_POINTER(p); \
 314     if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr       \
 315          && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {      \
 316             GC_mark_stack_top = GC_mark_and_push( \
 317                             (GC_PTR)(p), GC_mark_stack_top, \
 318                             GC_mark_stack_limit, (GC_PTR *)(source)); \
 319     }
 320
 321 /* Mark starting at mark stack entry top (incl.) down to        */
 322 /* mark stack entry bottom (incl.).  Stop after performing      */
 323 /* about one page worth of work.  Return the new mark stack     */
 324 /* top entry.                                                   */
 325 mse * GC_mark_from GC_PROTO((mse * top, mse * bottom, mse *limit));
 326
 327 #define MARK_FROM_MARK_STACK() \
 328         GC_mark_stack_top = GC_mark_from(GC_mark_stack_top, \
 329                                          GC_mark_stack, \
 330                                          GC_mark_stack + GC_mark_stack_size);
 331
 332 /*
 333  * Mark from one finalizable object using the specified
 334  * mark proc. May not mark the object pointed to by
 335  * real_ptr. That is the job of the caller, if appropriate
 336  */
 337 # define GC_MARK_FO(real_ptr, mark_proc) \
 338 { \
 339     (*(mark_proc))(real_ptr); \
 340     while (!GC_mark_stack_empty()) MARK_FROM_MARK_STACK(); \
 341     if (GC_mark_state != MS_NONE) { \
 342         GC_set_mark_bit(real_ptr); \
 343         while (!GC_mark_some((ptr_t)0)) {} \
 344     } \
 345 }
 346
 347 extern GC_bool GC_mark_stack_too_small;
 348                                 /* We need a larger mark stack.  May be */
 349                                 /* set by client supplied mark routines.*/
 350
 351 typedef int mark_state_t;       /* Current state of marking, as follows:*/
 352                                 /* Used to remember where we are during */
 353                                 /* concurrent marking.                  */
 354
 355                                 /* We say something is dirty if it was  */
 356                                 /* written since the last time we       */
 357                                 /* retrieved dirty bits.  We say it's   */
 358                                 /* grungy if it was marked dirty in the */
 359                                 /* last set of bits we retrieved.       */
 360
 361                                 /* Invariant I: all roots and marked    */
 362                                 /* objects p are either dirty, or point */
 363                                 /* to objects q that are either marked  */
 364                                 /* or a pointer to q appears in a range */
 365                                 /* on the mark stack.                   */
 366
 367 # define MS_NONE 0              /* No marking in progress. I holds.     */
 368                                 /* Mark stack is empty.                 */
 369
 370 # define MS_PUSH_RESCUERS 1     /* Rescuing objects are currently       */
 371                                 /* being pushed.  I holds, except       */
 372                                 /* that grungy roots may point to       */
 373                                 /* unmarked objects, as may marked      */
 374                                 /* grungy objects above scan_ptr.       */
 375
 376 # define MS_PUSH_UNCOLLECTABLE 2
 377                                 /* I holds, except that marked          */
 378                                 /* uncollectable objects above scan_ptr */
 379                                 /* may point to unmarked objects.       */
 380                                 /* Roots may point to unmarked objects  */
 381
 382 # define MS_ROOTS_PUSHED 3      /* I holds, mark stack may be nonempty  */
 383
 384 # define MS_PARTIALLY_INVALID 4 /* I may not hold, e.g. because of M.S. */
 385                                 /* overflow.  However marked heap       */
 386                                 /* objects below scan_ptr point to      */
 387                                 /* marked or stacked objects.           */
 388
 389 # define MS_INVALID 5           /* I may not hold.                      */
 390
 391 extern mark_state_t GC_mark_state;
 392
 393 #endif  /* GC_PMARK_H */
 394