Storage.c 34 KB
Newer Older
1
/* -----------------------------------------------------------------------------
2
 *
3
 * (c) The GHC Team, 1998-2004
4 5 6 7 8
 *
 * Storage manager front end
 *
 * ---------------------------------------------------------------------------*/

9
#include "PosixSource.h"
10 11 12 13 14 15
#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "Stats.h"
#include "Hooks.h"
#include "BlockAlloc.h"
16
#include "MBlock.h"
17
#include "Weak.h"
18
#include "Sanity.h"
19
#include "Arena.h"
20 21
#include "OSThreads.h"
#include "Capability.h"
22
#include "Storage.h"
23
#include "Schedule.h"
24
#include "RetainerProfile.h"	// for counting memory blocks (memInventory)
25
#include "OSMem.h"
26

27 28 29
#include <stdlib.h>
#include <string.h>

30
/* 
31
 * All these globals require sm_mutex to access in THREADED_RTS mode.
32
 */
33
StgClosure    *caf_list         = NULL;
34 35
StgClosure    *revertible_caf_list = NULL;
rtsBool       keepCAFs;
36 37

bdescr *small_alloc_list;	/* allocate()d small objects */
38
bdescr *pinned_object_block;    /* allocate pinned objects into this block */
39 40 41 42 43 44
nat alloc_blocks;		/* number of allocate()d blocks since GC */
nat alloc_blocks_lim;		/* approximate limit on alloc_blocks */

StgPtr alloc_Hp    = NULL;	/* next free byte in small_alloc_list */
StgPtr alloc_HpLim = NULL;	/* end of block at small_alloc_list   */

45 46 47 48
generation *generations = NULL;	/* all the generations */
generation *g0		= NULL; /* generation 0, for convenience */
generation *oldest_gen  = NULL; /* oldest generation, for convenience */
step *g0s0 		= NULL; /* generation 0, step 0, for convenience */
49

50
ullong total_allocated = 0;	/* total memory allocated during run */
51

52
nat n_nurseries         = 0;    /* == RtsFlags.ParFlags.nNodes, convenience */
53
step *nurseries         = NULL; /* array of nurseries, >1 only if THREADED_RTS */
54

55
#ifdef THREADED_RTS
56 57 58 59
/*
 * Storage manager mutex:  protects all the above state from
 * simultaneous access by two STG threads.
 */
60
Mutex sm_mutex;
61 62 63 64
/*
 * This mutex is used by atomicModifyMutVar# only
 */
Mutex atomic_modify_mutvar_mutex;
65 66
#endif

67

68 69 70 71 72 73 74
/*
 * Forward references
 */
static void *stgAllocForGMP   (size_t size_in_bytes);
static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
static void  stgDeallocForGMP (void *ptr, size_t size);

75 76 77 78 79 80
static void
initStep (step *stp, int g, int s)
{
    stp->no = s;
    stp->blocks = NULL;
    stp->n_blocks = 0;
81 82
    stp->old_blocks = NULL;
    stp->n_old_blocks = 0;
83 84 85 86 87
    stp->gen = &generations[g];
    stp->gen_no = g;
    stp->hp = NULL;
    stp->hpLim = NULL;
    stp->hp_bd = NULL;
88 89
    stp->scavd_hp = NULL;
    stp->scavd_hpLim = NULL;
90 91 92 93 94 95 96 97 98 99 100
    stp->scan = NULL;
    stp->scan_bd = NULL;
    stp->large_objects = NULL;
    stp->n_large_blocks = 0;
    stp->new_large_objects = NULL;
    stp->scavenged_large_objects = NULL;
    stp->n_scavenged_large_blocks = 0;
    stp->is_compacted = 0;
    stp->bitmap = NULL;
}

101
void
102
initStorage( void )
103
{
104
  nat g, s;
105
  generation *gen;
106

107 108 109 110 111
  if (generations != NULL) {
      // multi-init protection
      return;
  }

sof's avatar
sof committed
112 113 114 115 116 117 118
  /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
   * doing something reasonable.
   */
  ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
  ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
  ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
  
119 120
  if (RtsFlags.GcFlags.maxHeapSize != 0 &&
      RtsFlags.GcFlags.heapSizeSuggestion > 
121
      RtsFlags.GcFlags.maxHeapSize) {
122
    RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
123 124
  }

125 126 127
  if (RtsFlags.GcFlags.maxHeapSize != 0 &&
      RtsFlags.GcFlags.minAllocAreaSize > 
      RtsFlags.GcFlags.maxHeapSize) {
128
      errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
129 130 131
      exit(1);
  }

132 133
  initBlockAllocator();
  
134
#if defined(THREADED_RTS)
sof's avatar
sof committed
135
  initMutex(&sm_mutex);
136
  initMutex(&atomic_modify_mutvar_mutex);
sof's avatar
sof committed
137 138
#endif

139 140
  ACQUIRE_SM_LOCK;

141 142
  /* allocate generation info array */
  generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations 
143
					     * sizeof(struct generation_),
144 145
					     "initStorage: gens");

146
  /* Initialise all generations */
147
  for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
148 149
    gen = &generations[g];
    gen->no = g;
150
    gen->mut_list = allocBlock();
151 152
    gen->collections = 0;
    gen->failed_promotions = 0;
153
    gen->max_blocks = 0;
154 155
  }

156 157 158 159 160 161 162 163 164 165 166
  /* A couple of convenience pointers */
  g0 = &generations[0];
  oldest_gen = &generations[RtsFlags.GcFlags.generations-1];

  /* Allocate step structures in each generation */
  if (RtsFlags.GcFlags.generations > 1) {
    /* Only for multiple-generations */

    /* Oldest generation: one step */
    oldest_gen->n_steps = 1;
    oldest_gen->steps = 
167
      stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
168 169 170

    /* set up all except the oldest generation with 2 steps */
    for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
171 172
      generations[g].n_steps = RtsFlags.GcFlags.steps;
      generations[g].steps  = 
173
	stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
174
			"initStorage: steps");
175 176 177 178 179
    }
    
  } else {
    /* single generation, i.e. a two-space collector */
    g0->n_steps = 1;
180
    g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
181 182
  }

183
#ifdef THREADED_RTS
184
  n_nurseries = n_capabilities;
185 186 187 188 189 190 191
  nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
			      "initStorage: nurseries");
#else
  n_nurseries = 1;
  nurseries = g0->steps; // just share nurseries[0] with g0s0
#endif  

192 193
  /* Initialise all steps */
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
194
    for (s = 0; s < generations[g].n_steps; s++) {
195
	initStep(&generations[g].steps[s], g, s);
196 197 198
    }
  }
  
199
#ifdef THREADED_RTS
200 201 202 203 204
  for (s = 0; s < n_nurseries; s++) {
      initStep(&nurseries[s], 0, s);
  }
#endif
  
205 206
  /* Set up the destination pointers in each younger gen. step */
  for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
207 208
    for (s = 0; s < generations[g].n_steps-1; s++) {
      generations[g].steps[s].to = &generations[g].steps[s+1];
209
    }
210
    generations[g].steps[s].to = &generations[g+1].steps[0];
211
  }
212
  oldest_gen->steps[0].to = &oldest_gen->steps[0];
213
  
214
#ifdef THREADED_RTS
215 216 217 218 219 220
  for (s = 0; s < n_nurseries; s++) {
      nurseries[s].to = generations[0].steps[0].to;
  }
#endif
  
  /* The oldest generation has one step. */
221
  if (RtsFlags.GcFlags.compact) {
222
      if (RtsFlags.GcFlags.generations == 1) {
223
	  errorBelch("WARNING: compaction is incompatible with -G1; disabled");
224 225 226
      } else {
	  oldest_gen->steps[0].is_compacted = 1;
      }
227
  }
228

229
#ifdef THREADED_RTS
230
  if (RtsFlags.GcFlags.generations == 1) {
231
      errorBelch("-G1 is incompatible with -threaded");
232
      stg_exit(EXIT_FAILURE);
233 234
  }
#endif
235 236 237 238

  /* generation 0 is special: that's the nursery */
  generations[0].max_blocks = 0;

239 240 241 242 243 244
  /* G0S0: the allocation area.  Policy: keep the allocation area
   * small to begin with, even if we have a large suggested heap
   * size.  Reason: we're going to do a major collection first, and we
   * don't want it to be a big one.  This vague idea is borne out by 
   * rigorous experimental evidence.
   */
245 246 247
  g0s0 = &generations[0].steps[0];

  allocNurseries();
248 249 250

  weak_ptr_list = NULL;
  caf_list = NULL;
251
  revertible_caf_list = NULL;
252 253 254 255 256 257 258 259
   
  /* initialise the allocate() interface */
  small_alloc_list = NULL;
  alloc_blocks = 0;
  alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;

  /* Tell GNU multi-precision pkg about our custom alloc functions */
  mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
260

261
  IF_DEBUG(gc, statDescribeGens());
262 263

  RELEASE_SM_LOCK;
264 265
}

266 267 268
void
exitStorage (void)
{
269
    stat_exit(calcAllocated());
Simon Marlow's avatar
Simon Marlow committed
270 271 272 273 274
}

void
freeStorage (void)
{
275
    freeAllMBlocks();
276 277
}

278 279
/* -----------------------------------------------------------------------------
   CAF management.
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316

   The entry code for every CAF does the following:
     
      - builds a CAF_BLACKHOLE in the heap
      - pushes an update frame pointing to the CAF_BLACKHOLE
      - invokes UPD_CAF(), which:
          - calls newCaf, below
	  - updates the CAF with a static indirection to the CAF_BLACKHOLE
      
   Why do we build a BLACKHOLE in the heap rather than just updating
   the thunk directly?  It's so that we only need one kind of update
   frame - otherwise we'd need a static version of the update frame too.

   newCaf() does the following:
       
      - it puts the CAF on the oldest generation's mut-once list.
        This is so that we can treat the CAF as a root when collecting
	younger generations.

   For GHCI, we have additional requirements when dealing with CAFs:

      - we must *retain* all dynamically-loaded CAFs ever entered,
        just in case we need them again.
      - we must be able to *revert* CAFs that have been evaluated, to
        their pre-evaluated form.

      To do this, we use an additional CAF list.  When newCaf() is
      called on a dynamically-loaded CAF, we add it to the CAF list
      instead of the old-generation mutable list, and save away its
      old info pointer (in caf->saved_info) for later reversion.

      To revert all the CAFs, we traverse the CAF list and reset the
      info pointer to caf->saved_info, then throw away the CAF list.
      (see GC.c:revertCAFs()).

      -- SDM 29/1/01

317 318
   -------------------------------------------------------------------------- */

319 320 321
void
newCAF(StgClosure* caf)
{
sof's avatar
sof committed
322
  ACQUIRE_SM_LOCK;
323

324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
  if(keepCAFs)
  {
    // HACK:
    // If we are in GHCi _and_ we are using dynamic libraries,
    // then we can't redirect newCAF calls to newDynCAF (see below),
    // so we make newCAF behave almost like newDynCAF.
    // The dynamic libraries might be used by both the interpreted
    // program and GHCi itself, so they must not be reverted.
    // This also means that in GHCi with dynamic libraries, CAFs are not
    // garbage collected. If this turns out to be a problem, we could
    // do another hack here and do an address range test on caf to figure
    // out whether it is from a dynamic library.
    ((StgIndStatic *)caf)->saved_info  = (StgInfoTable *)caf->header.info;
    ((StgIndStatic *)caf)->static_link = caf_list;
    caf_list = caf;
  }
  else
  {
    /* Put this CAF on the mutable list for the old generation.
    * This is a HACK - the IND_STATIC closure doesn't really have
    * a mut_link field, but we pretend it has - in fact we re-use
    * the STATIC_LINK field for the time being, because when we
    * come to do a major GC we won't need the mut_link field
    * any more and can use it as a STATIC_LINK.
    */
    ((StgIndStatic *)caf)->saved_info = NULL;
    recordMutableGen(caf, oldest_gen);
  }
  
sof's avatar
sof committed
353
  RELEASE_SM_LOCK;
354 355 356 357

#ifdef PAR
  /* If we are PAR or DIST then  we never forget a CAF */
  { globalAddr *newGA;
358
    //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
359 360 361
    newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
    ASSERT(newGA);
  } 
sof's avatar
sof committed
362
#endif /* PAR */
363 364
}

365 366 367 368
// An alternate version of newCaf which is used for dynamically loaded
// object code in GHCi.  In this case we want to retain *all* CAFs in
// the object code, because they might be demanded at any time from an
// expression evaluated on the command line.
369 370
// Also, GHCi might want to revert CAFs, so we add these to the
// revertible_caf_list.
371 372 373 374 375 376 377 378 379
//
// The linker hackily arranges that references to newCaf from dynamic
// code end up pointing to newDynCAF.
void
newDynCAF(StgClosure *caf)
{
    ACQUIRE_SM_LOCK;

    ((StgIndStatic *)caf)->saved_info  = (StgInfoTable *)caf->header.info;
380 381
    ((StgIndStatic *)caf)->static_link = revertible_caf_list;
    revertible_caf_list = caf;
382 383 384 385

    RELEASE_SM_LOCK;
}

386 387 388 389
/* -----------------------------------------------------------------------------
   Nursery management.
   -------------------------------------------------------------------------- */

390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
static bdescr *
allocNursery (step *stp, bdescr *tail, nat blocks)
{
    bdescr *bd;
    nat i;

    // Allocate a nursery: we allocate fresh blocks one at a time and
    // cons them on to the front of the list, not forgetting to update
    // the back pointer on the tail of the list to point to the new block.
    for (i=0; i < blocks; i++) {
	// @LDV profiling
	/*
	  processNursery() in LdvProfile.c assumes that every block group in
	  the nursery contains only a single block. So, if a block group is
	  given multiple blocks, change processNursery() accordingly.
	*/
	bd = allocBlock();
	bd->link = tail;
	// double-link the nursery: we might need to insert blocks
	if (tail != NULL) {
	    tail->u.back = bd;
	}
	bd->step = stp;
	bd->gen_no = 0;
	bd->flags = 0;
	bd->free = bd->start;
	tail = bd;
    }
    tail->u.back = NULL;
    return tail;
}

static void
assignNurseriesToCapabilities (void)
{
425
#ifdef THREADED_RTS
426
    nat i;
sof's avatar
sof committed
427

428 429 430
    for (i = 0; i < n_nurseries; i++) {
	capabilities[i].r.rNursery        = &nurseries[i];
	capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
431
	capabilities[i].r.rCurrentAlloc   = NULL;
432
    }
433
#else /* THREADED_RTS */
434 435
    MainCapability.r.rNursery        = &nurseries[0];
    MainCapability.r.rCurrentNursery = nurseries[0].blocks;
436
    MainCapability.r.rCurrentAlloc   = NULL;
437 438
#endif
}
439 440 441 442 443 444 445 446 447 448 449

void
allocNurseries( void )
{ 
    nat i;

    for (i = 0; i < n_nurseries; i++) {
	nurseries[i].blocks = 
	    allocNursery(&nurseries[i], NULL, 
			 RtsFlags.GcFlags.minAllocAreaSize);
	nurseries[i].n_blocks    = RtsFlags.GcFlags.minAllocAreaSize;
450 451
	nurseries[i].old_blocks   = NULL;
	nurseries[i].n_old_blocks = 0;
452 453 454 455
	/* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */
    }
    assignNurseriesToCapabilities();
}
456 457 458 459
      
void
resetNurseries( void )
{
460 461 462
    nat i;
    bdescr *bd;
    step *stp;
463

464 465 466 467 468 469 470 471
    for (i = 0; i < n_nurseries; i++) {
	stp = &nurseries[i];
	for (bd = stp->blocks; bd; bd = bd->link) {
	    bd->free = bd->start;
	    ASSERT(bd->gen_no == 0);
	    ASSERT(bd->step == stp);
	    IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
	}
472
    }
473
    assignNurseriesToCapabilities();
474 475
}

476 477
lnat
countNurseryBlocks (void)
478
{
479 480
    nat i;
    lnat blocks = 0;
481

482 483
    for (i = 0; i < n_nurseries; i++) {
	blocks += nurseries[i].n_blocks;
484
    }
485
    return blocks;
486 487
}

488 489
static void
resizeNursery ( step *stp, nat blocks )
490 491
{
  bdescr *bd;
492
  nat nursery_blocks;
493

494 495
  nursery_blocks = stp->n_blocks;
  if (nursery_blocks == blocks) return;
496

497
  if (nursery_blocks < blocks) {
498
    IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n", 
499
			 blocks));
500
    stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
501 502 503 504
  } 
  else {
    bdescr *next_bd;
    
505
    IF_DEBUG(gc, debugBelch("Decreasing size of nursery to %d blocks\n", 
506
			 blocks));
507

508
    bd = stp->blocks;
509 510 511 512 513 514
    while (nursery_blocks > blocks) {
	next_bd = bd->link;
	next_bd->u.back = NULL;
	nursery_blocks -= bd->blocks; // might be a large block
	freeGroup(bd);
	bd = next_bd;
515
    }
516
    stp->blocks = bd;
517 518 519
    // might have gone just under, by freeing a large block, so make
    // up the difference.
    if (nursery_blocks < blocks) {
520
	stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
521
    }
522 523
  }
  
524 525 526
  stp->n_blocks = blocks;
  ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
}
527

528 529 530 531
// 
// Resize each of the nurseries to the specified size.
//
void
532
resizeNurseriesFixed (nat blocks)
533 534 535 536 537
{
    nat i;
    for (i = 0; i < n_nurseries; i++) {
	resizeNursery(&nurseries[i], blocks);
    }
538 539
}

540 541 542 543 544 545 546 547 548 549 550
// 
// Resize the nurseries to the total specified size.
//
void
resizeNurseries (nat blocks)
{
    // If there are multiple nurseries, then we just divide the number
    // of available blocks between them.
    resizeNurseriesFixed(blocks / n_nurseries);
}

551 552 553 554 555 556 557 558 559
/* -----------------------------------------------------------------------------
   The allocate() interface

   allocate(n) always succeeds, and returns a chunk of memory n words
   long.  n can be larger than the size of a block if necessary, in
   which case a contiguous block group will be allocated.
   -------------------------------------------------------------------------- */

StgPtr
560
allocate( nat n )
561
{
562 563
    bdescr *bd;
    StgPtr p;
564

565
    ACQUIRE_SM_LOCK;
566

567 568
    TICK_ALLOC_HEAP_NOCTR(n);
    CCS_ALLOC(CCCS,n);
569

570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598
    /* big allocation (>LARGE_OBJECT_THRESHOLD) */
    /* ToDo: allocate directly into generation 1 */
    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
	nat req_blocks =  (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
	bd = allocGroup(req_blocks);
	dbl_link_onto(bd, &g0s0->large_objects);
	g0s0->n_large_blocks += req_blocks;
	bd->gen_no  = 0;
	bd->step = g0s0;
	bd->flags = BF_LARGE;
	bd->free = bd->start + n;
	alloc_blocks += req_blocks;
	RELEASE_SM_LOCK;
	return bd->start;
	
	/* small allocation (<LARGE_OBJECT_THRESHOLD) */
    } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
	if (small_alloc_list) {
	    small_alloc_list->free = alloc_Hp;
	}
	bd = allocBlock();
	bd->link = small_alloc_list;
	small_alloc_list = bd;
	bd->gen_no = 0;
	bd->step = g0s0;
	bd->flags = 0;
	alloc_Hp = bd->start;
	alloc_HpLim = bd->start + BLOCK_SIZE_W;
	alloc_blocks++;
599
    }
600 601 602 603 604
    
    p = alloc_Hp;
    alloc_Hp += n;
    RELEASE_SM_LOCK;
    return p;
605 606
}

607 608
lnat
allocated_bytes( void )
609
{
610 611 612 613 614 615 616 617 618
    lnat allocated;

    allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
    if (pinned_object_block != NULL) {
	allocated -= (pinned_object_block->start + BLOCK_SIZE_W) - 
	    pinned_object_block->free;
    }
	
    return allocated;
619 620
}

621 622 623 624 625 626 627 628 629 630
void
tidyAllocateLists (void)
{
    if (small_alloc_list != NULL) {
	ASSERT(alloc_Hp >= small_alloc_list->start && 
	       alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
	small_alloc_list->free = alloc_Hp;
    }
}

631 632 633 634 635 636 637 638 639 640 641 642 643 644
/* -----------------------------------------------------------------------------
   allocateLocal()

   This allocates memory in the current thread - it is intended for
   use primarily from STG-land where we have a Capability.  It is
   better than allocate() because it doesn't require taking the
   sm_mutex lock in the common case.

   Memory is allocated directly from the nursery if possible (but not
   from the current nursery block, so as not to interfere with
   Hp/HpLim).
   -------------------------------------------------------------------------- */

StgPtr
645
allocateLocal (Capability *cap, nat n)
646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671
{
    bdescr *bd;
    StgPtr p;

    TICK_ALLOC_HEAP_NOCTR(n);
    CCS_ALLOC(CCCS,n);
    
    /* big allocation (>LARGE_OBJECT_THRESHOLD) */
    /* ToDo: allocate directly into generation 1 */
    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
	nat req_blocks =  (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
	ACQUIRE_SM_LOCK;
	bd = allocGroup(req_blocks);
	dbl_link_onto(bd, &g0s0->large_objects);
	g0s0->n_large_blocks += req_blocks;
	bd->gen_no  = 0;
	bd->step = g0s0;
	bd->flags = BF_LARGE;
	bd->free = bd->start + n;
	alloc_blocks += req_blocks;
	RELEASE_SM_LOCK;
	return bd->start;
	
	/* small allocation (<LARGE_OBJECT_THRESHOLD) */
    } else {

672
	bd = cap->r.rCurrentAlloc;
673 674 675 676 677
	if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {

	    // The CurrentAlloc block is full, we need to find another
	    // one.  First, we try taking the next block from the
	    // nursery:
678
	    bd = cap->r.rCurrentNursery->link;
679 680 681 682 683 684

	    if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
		// The nursery is empty, or the next block is already
		// full: allocate a fresh block (we can't fail here).
		ACQUIRE_SM_LOCK;
		bd = allocBlock();
685
		cap->r.rNursery->n_blocks++;
686 687
		RELEASE_SM_LOCK;
		bd->gen_no = 0;
688
		bd->step = cap->r.rNursery;
689 690 691 692 693
		bd->flags = 0;
	    } else {
		// we have a block in the nursery: take it and put
		// it at the *front* of the nursery list, and use it
		// to allocate() from.
694
		cap->r.rCurrentNursery->link = bd->link;
695
		if (bd->link != NULL) {
696
		    bd->link->u.back = cap->r.rCurrentNursery;
697
		}
698
	    }
699 700 701
	    dbl_link_onto(bd, &cap->r.rNursery->blocks);
	    cap->r.rCurrentAlloc = bd;
	    IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
702 703 704 705 706 707 708
	}
    }
    p = bd->free;
    bd->free += n;
    return p;
}

709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
/* ---------------------------------------------------------------------------
   Allocate a fixed/pinned object.

   We allocate small pinned objects into a single block, allocating a
   new block when the current one overflows.  The block is chained
   onto the large_object_list of generation 0 step 0.

   NOTE: The GC can't in general handle pinned objects.  This
   interface is only safe to use for ByteArrays, which have no
   pointers and don't require scavenging.  It works because the
   block's descriptor has the BF_LARGE flag set, so the block is
   treated as a large object and chained onto various lists, rather
   than the individual objects being copied.  However, when it comes
   to scavenge the block, the GC will only scavenge the first object.
   The reason is that the GC can't linearly scan a block of pinned
   objects at the moment (doing so would require using the
   mostly-copying techniques).  But since we're restricting ourselves
   to pinned ByteArrays, not scavenging is ok.

   This function is called by newPinnedByteArray# which immediately
   fills the allocated memory with a MutableByteArray#.
   ------------------------------------------------------------------------- */

StgPtr
allocatePinned( nat n )
{
    StgPtr p;
    bdescr *bd = pinned_object_block;

    // If the request is for a large object, then allocate()
    // will give us a pinned object anyway.
    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
	return allocate(n);
    }

sof's avatar
sof committed
744 745 746 747 748
    ACQUIRE_SM_LOCK;
    
    TICK_ALLOC_HEAP_NOCTR(n);
    CCS_ALLOC(CCCS,n);

749 750 751
    // we always return 8-byte aligned memory.  bd->free must be
    // 8-byte aligned to begin with, so we just round up n to
    // the nearest multiple of 8 bytes.
752 753 754
    if (sizeof(StgWord) == 4) {
	n = (n+1) & ~1;
    }
755

756 757 758 759 760 761 762
    // If we don't have a block of pinned objects yet, or the current
    // one isn't large enough to hold the new object, allocate a new one.
    if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
	pinned_object_block = bd = allocBlock();
	dbl_link_onto(bd, &g0s0->large_objects);
	bd->gen_no = 0;
	bd->step   = g0s0;
763
	bd->flags  = BF_PINNED | BF_LARGE;
764 765 766 767 768 769
	bd->free   = bd->start;
	alloc_blocks++;
    }

    p = bd->free;
    bd->free += n;
sof's avatar
sof committed
770
    RELEASE_SM_LOCK;
771 772 773
    return p;
}

774 775 776 777 778 779 780 781
/* -----------------------------------------------------------------------------
   This is the write barrier for MUT_VARs, a.k.a. IORefs.  A
   MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
   is.  When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
   and is put on the mutable list.
   -------------------------------------------------------------------------- */

void
782
dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
783
{
784
    Capability *cap = regTableToCapability(reg);
785
    bdescr *bd;
786 787
    if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
	p->header.info = &stg_MUT_VAR_DIRTY_info;
Simon Marlow's avatar
Simon Marlow committed
788
	bd = Bdescr((StgPtr)p);
789
	if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
790 791 792
    }
}

793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809
/* -----------------------------------------------------------------------------
   Allocation functions for GMP.

   These all use the allocate() interface - we can't have any garbage
   collection going on during a gmp operation, so we use allocate()
   which always succeeds.  The gmp operations which might need to
   allocate will ask the storage manager (via doYouWantToGC()) whether
   a garbage collection is required, in case we get into a loop doing
   only allocate() style allocation.
   -------------------------------------------------------------------------- */

static void *
stgAllocForGMP (size_t size_in_bytes)
{
  StgArrWords* arr;
  nat data_size_in_words, total_size_in_words;
  
810 811
  /* round up to a whole number of words */
  data_size_in_words  = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
812 813 814
  total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
  
  /* allocate and fill it in. */
815
#if defined(THREADED_RTS)
816
  arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
817
#else
818
  arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
819
#endif
820
  SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
821 822
  
  /* and return a ptr to the goods inside the array */
823
  return arr->payload;
824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846
}

static void *
stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
{
    void *new_stuff_ptr = stgAllocForGMP(new_size);
    nat i = 0;
    char *p = (char *) ptr;
    char *q = (char *) new_stuff_ptr;

    for (; i < old_size; i++, p++, q++) {
	*q = *p;
    }

    return(new_stuff_ptr);
}

static void
stgDeallocForGMP (void *ptr STG_UNUSED, 
		  size_t size STG_UNUSED)
{
    /* easy for us: the garbage collector does the dealloc'n */
}
847

848
/* -----------------------------------------------------------------------------
849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
 * Stats and stuff
 * -------------------------------------------------------------------------- */

/* -----------------------------------------------------------------------------
 * calcAllocated()
 *
 * Approximate how much we've allocated: number of blocks in the
 * nursery + blocks allocated via allocate() - unused nusery blocks.
 * This leaves a little slop at the end of each block, and doesn't
 * take into account large objects (ToDo).
 * -------------------------------------------------------------------------- */

lnat
calcAllocated( void )
{
  nat allocated;
  bdescr *bd;

867
  allocated = allocated_bytes();
868
  allocated += countNurseryBlocks() * BLOCK_SIZE_W;
869
  
870
  {
871
#ifdef THREADED_RTS
872
  nat i;
873 874
  for (i = 0; i < n_nurseries; i++) {
      Capability *cap;
875
      for ( bd = capabilities[i].r.rCurrentNursery->link; 
876 877 878 879 880 881 882 883 884
	    bd != NULL; bd = bd->link ) {
	  allocated -= BLOCK_SIZE_W;
      }
      cap = &capabilities[i];
      if (cap->r.rCurrentNursery->free < 
	  cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
	  allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
	      - cap->r.rCurrentNursery->free;
      }
885
  }
886
#else
887
  bdescr *current_nursery = MainCapability.r.rCurrentNursery;
888 889

  for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
890
      allocated -= BLOCK_SIZE_W;
891 892
  }
  if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
893 894
      allocated -= (current_nursery->start + BLOCK_SIZE_W)
	  - current_nursery->free;
895 896
  }
#endif
897
  }
898

899
  total_allocated += allocated;
900 901
  return allocated;
}  
902 903 904 905 906 907 908 909 910

/* Approximate the amount of live data in the heap.  To be called just
 * after garbage collection (see GarbageCollect()).
 */
extern lnat 
calcLive(void)
{
  nat g, s;
  lnat live = 0;
911
  step *stp;
912 913

  if (RtsFlags.GcFlags.generations == 1) {
914
    live = (g0s0->n_blocks - 1) * BLOCK_SIZE_W + 
915
      ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
916
    return live;
917 918 919 920 921
  }

  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
    for (s = 0; s < generations[g].n_steps; s++) {
      /* approximate amount of live data (doesn't take into account slop
922 923
       * at end of each block).
       */
924 925 926
      if (g == 0 && s == 0) { 
	  continue; 
      }
927
      stp = &generations[g].steps[s];
928
      live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
929 930 931 932
      if (stp->hp_bd != NULL) {
	  live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start) 
	      / sizeof(W_);
      }
933 934 935
      if (stp->scavd_hp != NULL) {
	  live -= (P_)(BLOCK_ROUND_UP(stp->scavd_hp)) - stp->scavd_hp;
      }
936 937 938 939 940 941 942 943 944 945 946 947 948 949 950
    }
  }
  return live;
}

/* Approximate the number of blocks that will be needed at the next
 * garbage collection.
 *
 * Assume: all data currently live will remain live.  Steps that will
 * be collected next time will therefore need twice as many blocks
 * since all the data will be copied.
 */
extern lnat 
calcNeeded(void)
{
951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967
    lnat needed = 0;
    nat g, s;
    step *stp;
    
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
	for (s = 0; s < generations[g].n_steps; s++) {
	    if (g == 0 && s == 0) { continue; }
	    stp = &generations[g].steps[s];
	    if (generations[g].steps[0].n_blocks +
		generations[g].steps[0].n_large_blocks 
		> generations[g].max_blocks
		&& stp->is_compacted == 0) {
		needed += 2 * stp->n_blocks;
	    } else {
		needed += stp->n_blocks;
	    }
	}
968
    }
969
    return needed;
970 971
}

972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
/* ----------------------------------------------------------------------------
   Executable memory

   Executable memory must be managed separately from non-executable
   memory.  Most OSs these days require you to jump through hoops to
   dynamically allocate executable memory, due to various security
   measures.

   Here we provide a small memory allocator for executable memory.
   Memory is managed with a page granularity; we allocate linearly
   in the page, and when the page is emptied (all objects on the page
   are free) we free the page again, not forgetting to make it
   non-executable.
   ------------------------------------------------------------------------- */

static bdescr *exec_block;

void *allocateExec (nat bytes)
{
    void *ret;
    nat n;

    ACQUIRE_SM_LOCK;

    // round up to words.
    n  = (bytes + sizeof(W_) + 1) / sizeof(W_);

    if (n+1 > BLOCK_SIZE_W) {
	barf("allocateExec: can't handle large objects");
    }

    if (exec_block == NULL || 
	exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
	bdescr *bd;
	lnat pagesize = getPageSize();
	bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
	IF_DEBUG(gc, debugBelch("allocate exec block %p\n", bd->start));
	bd->gen_no = 0;
	bd->flags = BF_EXEC;
	bd->link = exec_block;
	if (exec_block != NULL) {
	    exec_block->u.back = bd;
	}
	bd->u.back = NULL;
	setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
	exec_block = bd;
    }
    *(exec_block->free) = n;  // store the size of this chunk
    exec_block->gen_no += n;  // gen_no stores the number of words allocated
    ret = exec_block->free + 1;
    exec_block->free += n + 1;

    RELEASE_SM_LOCK
    return ret;
}

void freeExec (void *addr)
{
    StgPtr p = (StgPtr)addr - 1;
    bdescr *bd = Bdescr((StgPtr)p);

    if ((bd->flags & BF_EXEC) == 0) {
	barf("freeExec: not executable");
    }

    if (*(StgPtr)p == 0) {
	barf("freeExec: already free?");
    }

    ACQUIRE_SM_LOCK;

    bd->gen_no -= *(StgPtr)p;
    *(StgPtr)p = 0;

    // Free the block if it is empty, but not if it is the block at
    // the head of the queue.
    if (bd->gen_no == 0 && bd != exec_block) {
	IF_DEBUG(gc, debugBelch("free exec block %p\n", bd->start));
	if (bd->u.back) {
	    bd->u.back->link = bd->link;
	} else {
	    exec_block = bd->link;
	}
	if (bd->link) {
	    bd->link->u.back = bd->u.back;
	}
	setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
	freeGroup(bd);
    }

    RELEASE_SM_LOCK
}    

1065 1066 1067 1068 1069 1070 1071 1072 1073 1074
/* -----------------------------------------------------------------------------
   Debugging

   memInventory() checks for memory leaks by counting up all the
   blocks we know about and comparing that to the number of blocks
   allegedly floating around in the system.
   -------------------------------------------------------------------------- */

#ifdef DEBUG

1075 1076 1077 1078 1079 1080 1081
static lnat
stepBlocks (step *stp)
{
    lnat total_blocks;
    bdescr *bd;

    total_blocks = stp->n_blocks;    
1082
    total_blocks += stp->n_old_blocks;
1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
    for (bd = stp->large_objects; bd; bd = bd->link) {
	total_blocks += bd->blocks;
	/* hack for megablock groups: they have an extra block or two in
	   the second and subsequent megablocks where the block
	   descriptors would normally go.
	*/
	if (bd->blocks > BLOCKS_PER_MBLOCK) {
	    total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
		* (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
	}
    }
    return total_blocks;
}

1097
void
1098 1099
memInventory(void)
{
1100
  nat g, s, i;
1101
  step *stp;
1102 1103 1104 1105
  bdescr *bd;
  lnat total_blocks = 0, free_blocks = 0;

  /* count the blocks we current have */
1106

1107
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1108 1109 1110 1111 1112
      for (i = 0; i < n_capabilities; i++) {
	  for (bd = capabilities[i].mut_lists[g]; bd != NULL; bd = bd->link) {
	      total_blocks += bd->blocks;
	  }
      }	  
1113 1114
      for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
	  total_blocks += bd->blocks;
1115
      }
simonmar's avatar