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