libgo/runtime/mgc0.c

   1 // Copyright 2009 The Go Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style
   3 // license that can be found in the LICENSE file.
   4
   5 // Garbage collector.
   6
   7 #include <unistd.h>
   8
   9 #include "runtime.h"
  10 #include "arch.h"
  11 #include "malloc.h"
  12
  13 #ifdef USING_SPLIT_STACK
  14
  15 extern void * __splitstack_find (void *, void *, size_t *, void **, void **,
  16                                  void **);
  17
  18 extern void * __splitstack_find_context (void *context[10], size_t *, void **,
  19                                          void **, void **);
  20
  21 #endif
  22
  23 enum {
  24         Debug = 0,
  25         PtrSize = sizeof(void*),
  26         DebugMark = 0,  // run second pass to check mark
  27
  28         // Four bits per word (see #defines below).
  29         wordsPerBitmapWord = sizeof(void*)*8/4,
  30         bitShift = sizeof(void*)*8/4,
  31 };
  32
  33 // Bits in per-word bitmap.
  34 // #defines because enum might not be able to hold the values.
  35 //
  36 // Each word in the bitmap describes wordsPerBitmapWord words
  37 // of heap memory.  There are 4 bitmap bits dedicated to each heap word,
  38 // so on a 64-bit system there is one bitmap word per 16 heap words.
  39 // The bits in the word are packed together by type first, then by
  40 // heap location, so each 64-bit bitmap word consists of, from top to bottom,
  41 // the 16 bitSpecial bits for the corresponding heap words, then the 16 bitMarked bits,
  42 // then the 16 bitNoPointers/bitBlockBoundary bits, then the 16 bitAllocated bits.
  43 // This layout makes it easier to iterate over the bits of a given type.
  44 //
  45 // The bitmap starts at mheap.arena_start and extends *backward* from
  46 // there.  On a 64-bit system the off'th word in the arena is tracked by
  47 // the off/16+1'th word before mheap.arena_start.  (On a 32-bit system,
  48 // the only difference is that the divisor is 8.)
  49 //
  50 // To pull out the bits corresponding to a given pointer p, we use:
  51 //
  52 //      off = p - (uintptr*)mheap.arena_start;  // word offset
  53 //      b = (uintptr*)mheap.arena_start - off/wordsPerBitmapWord - 1;
  54 //      shift = off % wordsPerBitmapWord
  55 //      bits = *b >> shift;
  56 //      /* then test bits & bitAllocated, bits & bitMarked, etc. */
  57 //
  58 #define bitAllocated            ((uintptr)1<<(bitShift*0))
  59 #define bitNoPointers           ((uintptr)1<<(bitShift*1))      /* when bitAllocated is set */
  60 #define bitMarked               ((uintptr)1<<(bitShift*2))      /* when bitAllocated is set */
  61 #define bitSpecial              ((uintptr)1<<(bitShift*3))      /* when bitAllocated is set - has finalizer or being profiled */
  62 #define bitBlockBoundary        ((uintptr)1<<(bitShift*1))      /* when bitAllocated is NOT set */
  63
  64 #define bitMask (bitBlockBoundary | bitAllocated | bitMarked | bitSpecial)
  65
  66 // Holding worldsema grants an M the right to try to stop the world.
  67 // The procedure is:
  68 //
  69 //      runtime_semacquire(&runtime_worldsema);
  70 //      m->gcing = 1;
  71 //      runtime_stoptheworld();
  72 //
  73 //      ... do stuff ...
  74 //
  75 //      m->gcing = 0;
  76 //      runtime_semrelease(&runtime_worldsema);
  77 //      runtime_starttheworld();
  78 //
  79 uint32 runtime_worldsema = 1;
  80
  81 // TODO: Make these per-M.
  82 static uint64 nhandoff;
  83
  84 static int32 gctrace;
  85
  86 typedef struct Workbuf Workbuf;
  87 struct Workbuf
  88 {
  89         Workbuf *next;
  90         uintptr nobj;
  91         byte *obj[512-2];
  92 };
  93
  94 typedef struct Finalizer Finalizer;
  95 struct Finalizer
  96 {
  97         void (*fn)(void*);
  98         void *arg;
  99         const struct __go_func_type *ft;
 100 };
 101
 102 typedef struct FinBlock FinBlock;
 103 struct FinBlock
 104 {
 105         FinBlock *alllink;
 106         FinBlock *next;
 107         int32 cnt;
 108         int32 cap;
 109         Finalizer fin[1];
 110 };
 111
 112 static G *fing;
 113 static FinBlock *finq; // list of finalizers that are to be executed
 114 static FinBlock *finc; // cache of free blocks
 115 static FinBlock *allfin; // list of all blocks
 116 static Lock finlock;
 117 static int32 fingwait;
 118
 119 static void runfinq(void*);
 120 static Workbuf* getempty(Workbuf*);
 121 static Workbuf* getfull(Workbuf*);
 122 static void     putempty(Workbuf*);
 123 static Workbuf* handoff(Workbuf*);
 124
 125 static struct {
 126         Lock fmu;
 127         Workbuf *full;
 128         Lock emu;
 129         Workbuf *empty;
 130         uint32  nproc;
 131         volatile uint32 nwait;
 132         volatile uint32 ndone;
 133         Note    alldone;
 134         Lock    markgate;
 135         Lock    sweepgate;
 136         MSpan   *spans;
 137
 138         Lock;
 139         byte    *chunk;
 140         uintptr nchunk;
 141 } work;
 142
 143 // scanblock scans a block of n bytes starting at pointer b for references
 144 // to other objects, scanning any it finds recursively until there are no
 145 // unscanned objects left.  Instead of using an explicit recursion, it keeps
 146 // a work list in the Workbuf* structures and loops in the main function
 147 // body.  Keeping an explicit work list is easier on the stack allocator and
 148 // more efficient.
 149 static void
 150 scanblock(byte *b, int64 n)
 151 {
 152         byte *obj, *arena_start, *arena_used, *p;
 153         void **vp;
 154         uintptr size, *bitp, bits, shift, i, j, x, xbits, off, nobj, nproc;
 155         MSpan *s;
 156         PageID k;
 157         void **wp;
 158         Workbuf *wbuf;
 159         bool keepworking;
 160
 161         if((int64)(uintptr)n != n || n < 0) {
 162                 runtime_printf("scanblock %p %D\n", b, n);
 163                 runtime_throw("scanblock");
 164         }
 165
 166         // Memory arena parameters.
 167         arena_start = runtime_mheap.arena_start;
 168         arena_used = runtime_mheap.arena_used;
 169         nproc = work.nproc;
 170
 171         wbuf = nil;  // current work buffer
 172         wp = nil;  // storage for next queued pointer (write pointer)
 173         nobj = 0;  // number of queued objects
 174
 175         // Scanblock helpers pass b==nil.
 176         // The main proc needs to return to make more
 177         // calls to scanblock.  But if work.nproc==1 then
 178         // might as well process blocks as soon as we
 179         // have them.
 180         keepworking = b == nil || work.nproc == 1;
 181
 182         // Align b to a word boundary.
 183         off = (uintptr)b & (PtrSize-1);
 184         if(off != 0) {
 185                 b += PtrSize - off;
 186                 n -= PtrSize - off;
 187         }
 188
 189         for(;;) {
 190                 // Each iteration scans the block b of length n, queueing pointers in
 191                 // the work buffer.
 192                 if(Debug > 1)
 193                         runtime_printf("scanblock %p %D\n", b, n);
 194
 195                 vp = (void**)b;
 196                 n >>= (2+PtrSize/8);  /* n /= PtrSize (4 or 8) */
 197                 for(i=0; i<(uintptr)n; i++) {
 198                         obj = (byte*)vp[i];
 199
 200                         // Words outside the arena cannot be pointers.
 201                         if((byte*)obj < arena_start || (byte*)obj >= arena_used)
 202                                 continue;
 203
 204                         // obj may be a pointer to a live object.
 205                         // Try to find the beginning of the object.
 206
 207                         // Round down to word boundary.
 208                         obj = (void*)((uintptr)obj & ~((uintptr)PtrSize-1));
 209
 210                         // Find bits for this word.
 211                         off = (uintptr*)obj - (uintptr*)arena_start;
 212                         bitp = (uintptr*)arena_start - off/wordsPerBitmapWord - 1;
 213                         shift = off % wordsPerBitmapWord;
 214                         xbits = *bitp;
 215                         bits = xbits >> shift;
 216
 217                         // Pointing at the beginning of a block?
 218                         if((bits & (bitAllocated|bitBlockBoundary)) != 0)
 219                                 goto found;
 220
 221                         // Pointing just past the beginning?
 222                         // Scan backward a little to find a block boundary.
 223                         for(j=shift; j-->0; ) {
 224                                 if(((xbits>>j) & (bitAllocated|bitBlockBoundary)) != 0) {
 225                                         obj = (byte*)obj - (shift-j)*PtrSize;
 226                                         shift = j;
 227                                         bits = xbits>>shift;
 228                                         goto found;
 229                                 }
 230                         }
 231
 232                         // Otherwise consult span table to find beginning.
 233                         // (Manually inlined copy of MHeap_LookupMaybe.)
 234                         k = (uintptr)obj>>PageShift;
 235                         x = k;
 236                         if(sizeof(void*) == 8)
 237                                 x -= (uintptr)arena_start>>PageShift;
 238                         s = runtime_mheap.map[x];
 239                         if(s == nil || k < s->start || k - s->start >= s->npages || s->state != MSpanInUse)
 240                                 continue;
 241                         p =  (byte*)((uintptr)s->start<<PageShift);
 242                         if(s->sizeclass == 0) {
 243                                 obj = p;
 244                         } else {
 245                                 if((byte*)obj >= (byte*)s->limit)
 246                                         continue;
 247                                 size = runtime_class_to_size[s->sizeclass];
 248                                 int32 i = ((byte*)obj - p)/size;
 249                                 obj = p+i*size;
 250                         }
 251
 252                         // Now that we know the object header, reload bits.
 253                         off = (uintptr*)obj - (uintptr*)arena_start;
 254                         bitp = (uintptr*)arena_start - off/wordsPerBitmapWord - 1;
 255                         shift = off % wordsPerBitmapWord;
 256                         xbits = *bitp;
 257                         bits = xbits >> shift;
 258
 259                 found:
 260                         // Now we have bits, bitp, and shift correct for
 261                         // obj pointing at the base of the object.
 262                         // Only care about allocated and not marked.
 263                         if((bits & (bitAllocated|bitMarked)) != bitAllocated)
 264                                 continue;
 265                         if(nproc == 1)
 266                                 *bitp |= bitMarked<<shift;
 267                         else {
 268                                 for(;;) {
 269                                         x = *bitp;
 270                                         if(x & (bitMarked<<shift))
 271                                                 goto continue_obj;
 272                                         if(runtime_casp((void**)bitp, (void*)x, (void*)(x|(bitMarked<<shift))))
 273                                                 break;
 274                                 }
 275                         }
 276
 277                         // If object has no pointers, don't need to scan further.
 278                         if((bits & bitNoPointers) != 0)
 279                                 continue;
 280
 281                         // If another proc wants a pointer, give it some.
 282                         if(nobj > 4 && work.nwait > 0 && work.full == nil) {
 283                                 wbuf->nobj = nobj;
 284                                 wbuf = handoff(wbuf);
 285                                 nobj = wbuf->nobj;
 286                                 wp = (void**)(wbuf->obj + nobj);
 287                         }
 288
 289                         // If buffer is full, get a new one.
 290                         if(wbuf == nil || nobj >= nelem(wbuf->obj)) {
 291                                 if(wbuf != nil)
 292                                         wbuf->nobj = nobj;
 293                                 wbuf = getempty(wbuf);
 294                                 wp = (void**)(wbuf->obj);
 295                                 nobj = 0;
 296                         }
 297                         *wp++ = obj;
 298                         nobj++;
 299                 continue_obj:;
 300                 }
 301
 302                 // Done scanning [b, b+n).  Prepare for the next iteration of
 303                 // the loop by setting b and n to the parameters for the next block.
 304
 305                 // Fetch b from the work buffer.
 306                 if(nobj == 0) {
 307                         if(!keepworking) {
 308                                 putempty(wbuf);
 309                                 return;
 310                         }
 311                         // Emptied our buffer: refill.
 312                         wbuf = getfull(wbuf);
 313                         if(wbuf == nil)
 314                                 return;
 315                         nobj = wbuf->nobj;
 316                         wp = (void**)(wbuf->obj + wbuf->nobj);
 317                 }
 318                 b = *--wp;
 319                 nobj--;
 320
 321                 // Ask span about size class.
 322                 // (Manually inlined copy of MHeap_Lookup.)
 323                 x = (uintptr)b>>PageShift;
 324                 if(sizeof(void*) == 8)
 325                         x -= (uintptr)arena_start>>PageShift;
 326                 s = runtime_mheap.map[x];
 327                 if(s->sizeclass == 0)
 328                         n = s->npages<<PageShift;
 329                 else
 330                         n = runtime_class_to_size[s->sizeclass];
 331         }
 332 }
 333
 334 // debug_scanblock is the debug copy of scanblock.
 335 // it is simpler, slower, single-threaded, recursive,
 336 // and uses bitSpecial as the mark bit.
 337 static void
 338 debug_scanblock(byte *b, int64 n)
 339 {
 340         byte *obj, *p;
 341         void **vp;
 342         uintptr size, *bitp, bits, shift, i, xbits, off;
 343         MSpan *s;
 344
 345         if(!DebugMark)
 346                 runtime_throw("debug_scanblock without DebugMark");
 347
 348         if((int64)(uintptr)n != n || n < 0) {
 349                 runtime_printf("debug_scanblock %p %D\n", b, n);
 350                 runtime_throw("debug_scanblock");
 351         }
 352
 353         // Align b to a word boundary.
 354         off = (uintptr)b & (PtrSize-1);
 355         if(off != 0) {
 356                 b += PtrSize - off;
 357                 n -= PtrSize - off;
 358         }
 359
 360         vp = (void**)b;
 361         n /= PtrSize;
 362         for(i=0; i<(uintptr)n; i++) {
 363                 obj = (byte*)vp[i];
 364
 365                 // Words outside the arena cannot be pointers.
 366                 if((byte*)obj < runtime_mheap.arena_start || (byte*)obj >= runtime_mheap.arena_used)
 367                         continue;
 368
 369                 // Round down to word boundary.
 370                 obj = (void*)((uintptr)obj & ~((uintptr)PtrSize-1));
 371
 372                 // Consult span table to find beginning.
 373                 s = runtime_MHeap_LookupMaybe(&runtime_mheap, obj);
 374                 if(s == nil)
 375                         continue;
 376
 377
 378                 p =  (byte*)((uintptr)s->start<<PageShift);
 379                 if(s->sizeclass == 0) {
 380                         obj = p;
 381                         size = (uintptr)s->npages<<PageShift;
 382                 } else {
 383                         if((byte*)obj >= (byte*)s->limit)
 384                                 continue;
 385                         size = runtime_class_to_size[s->sizeclass];
 386                         int32 i = ((byte*)obj - p)/size;
 387                         obj = p+i*size;
 388                 }
 389
 390                 // Now that we know the object header, reload bits.
 391                 off = (uintptr*)obj - (uintptr*)runtime_mheap.arena_start;
 392                 bitp = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
 393                 shift = off % wordsPerBitmapWord;
 394                 xbits = *bitp;
 395                 bits = xbits >> shift;
 396
 397                 // Now we have bits, bitp, and shift correct for
 398                 // obj pointing at the base of the object.
 399                 // If not allocated or already marked, done.
 400                 if((bits & bitAllocated) == 0 || (bits & bitSpecial) != 0)  // NOTE: bitSpecial not bitMarked
 401                         continue;
 402                 *bitp |= bitSpecial<<shift;
 403                 if(!(bits & bitMarked))
 404                         runtime_printf("found unmarked block %p in %p\n", obj, vp+i);
 405
 406                 // If object has no pointers, don't need to scan further.
 407                 if((bits & bitNoPointers) != 0)
 408                         continue;
 409
 410                 debug_scanblock(obj, size);
 411         }
 412 }
 413
 414 // Get an empty work buffer off the work.empty list,
 415 // allocating new buffers as needed.
 416 static Workbuf*
 417 getempty(Workbuf *b)
 418 {
 419         if(work.nproc == 1) {
 420                 // Put b on full list.
 421                 if(b != nil) {
 422                         b->next = work.full;
 423                         work.full = b;
 424                 }
 425                 // Grab from empty list if possible.
 426                 b = work.empty;
 427                 if(b != nil) {
 428                         work.empty = b->next;
 429                         goto haveb;
 430                 }
 431         } else {
 432                 // Put b on full list.
 433                 if(b != nil) {
 434                         runtime_lock(&work.fmu);
 435                         b->next = work.full;
 436                         work.full = b;
 437                         runtime_unlock(&work.fmu);
 438                 }
 439                 // Grab from empty list if possible.
 440                 runtime_lock(&work.emu);
 441                 b = work.empty;
 442                 if(b != nil)
 443                         work.empty = b->next;
 444                 runtime_unlock(&work.emu);
 445                 if(b != nil)
 446                         goto haveb;
 447         }
 448
 449         // Need to allocate.
 450         runtime_lock(&work);
 451         if(work.nchunk < sizeof *b) {
 452                 work.nchunk = 1<<20;
 453                 work.chunk = runtime_SysAlloc(work.nchunk);
 454         }
 455         b = (Workbuf*)work.chunk;
 456         work.chunk += sizeof *b;
 457         work.nchunk -= sizeof *b;
 458         runtime_unlock(&work);
 459
 460 haveb:
 461         b->nobj = 0;
 462         return b;
 463 }
 464
 465 static void
 466 putempty(Workbuf *b)
 467 {
 468         if(b == nil)
 469                 return;
 470
 471         if(work.nproc == 1) {
 472                 b->next = work.empty;
 473                 work.empty = b;
 474                 return;
 475         }
 476
 477         runtime_lock(&work.emu);
 478         b->next = work.empty;
 479         work.empty = b;
 480         runtime_unlock(&work.emu);
 481 }
 482
 483 // Get a full work buffer off the work.full list, or return nil.
 484 static Workbuf*
 485 getfull(Workbuf *b)
 486 {
 487         int32 i;
 488         Workbuf *b1;
 489
 490         if(work.nproc == 1) {
 491                 // Put b on empty list.
 492                 if(b != nil) {
 493                         b->next = work.empty;
 494                         work.empty = b;
 495                 }
 496                 // Grab from full list if possible.
 497                 // Since work.nproc==1, no one else is
 498                 // going to give us work.
 499                 b = work.full;
 500                 if(b != nil)
 501                         work.full = b->next;
 502                 return b;
 503         }
 504
 505         putempty(b);
 506
 507         // Grab buffer from full list if possible.
 508         for(;;) {
 509                 b1 = work.full;
 510                 if(b1 == nil)
 511                         break;
 512                 runtime_lock(&work.fmu);
 513                 if(work.full != nil) {
 514                         b1 = work.full;
 515                         work.full = b1->next;
 516                         runtime_unlock(&work.fmu);
 517                         return b1;
 518                 }
 519                 runtime_unlock(&work.fmu);
 520         }
 521
 522         runtime_xadd(&work.nwait, +1);
 523         for(i=0;; i++) {
 524                 b1 = work.full;
 525                 if(b1 != nil) {
 526                         runtime_lock(&work.fmu);
 527                         if(work.full != nil) {
 528                                 runtime_xadd(&work.nwait, -1);
 529                                 b1 = work.full;
 530                                 work.full = b1->next;
 531                                 runtime_unlock(&work.fmu);
 532                                 return b1;
 533                         }
 534                         runtime_unlock(&work.fmu);
 535                         continue;
 536                 }
 537                 if(work.nwait == work.nproc)
 538                         return nil;
 539                 if(i < 10)
 540                         runtime_procyield(20);
 541                 else if(i < 20)
 542                         runtime_osyield();
 543                 else
 544                         runtime_usleep(100);
 545         }
 546 }
 547
 548 static Workbuf*
 549 handoff(Workbuf *b)
 550 {
 551         int32 n;
 552         Workbuf *b1;
 553
 554         // Make new buffer with half of b's pointers.
 555         b1 = getempty(nil);
 556         n = b->nobj/2;
 557         b->nobj -= n;
 558         b1->nobj = n;
 559         runtime_memmove(b1->obj, b->obj+b->nobj, n*sizeof b1->obj[0]);
 560         nhandoff += n;
 561
 562         // Put b on full list - let first half of b get stolen.
 563         runtime_lock(&work.fmu);
 564         b->next = work.full;
 565         work.full = b;
 566         runtime_unlock(&work.fmu);
 567
 568         return b1;
 569 }
 570
 571 // Scanstack calls scanblock on each of gp's stack segments.
 572 static void
 573 scanstack(void (*scanblock)(byte*, int64), G *gp)
 574 {
 575 #ifdef USING_SPLIT_STACK
 576         M *mp;
 577         void* sp;
 578         size_t spsize;
 579         void* next_segment;
 580         void* next_sp;
 581         void* initial_sp;
 582
 583         if(gp == runtime_g()) {
 584                 // Scanning our own stack.
 585                 sp = __splitstack_find(nil, nil, &spsize, &next_segment,
 586                                        &next_sp, &initial_sp);
 587         } else if((mp = gp->m) != nil && mp->helpgc) {
 588                 // gchelper's stack is in active use and has no interesting pointers.
 589                 return;
 590         } else {
 591                 // Scanning another goroutine's stack.
 592                 // The goroutine is usually asleep (the world is stopped).
 593
 594                 // The exception is that if the goroutine is about to enter or might
 595                 // have just exited a system call, it may be executing code such
 596                 // as schedlock and may have needed to start a new stack segment.
 597                 // Use the stack segment and stack pointer at the time of
 598                 // the system call instead, since that won't change underfoot.
 599                 if(gp->gcstack != nil) {
 600                         sp = gp->gcstack;
 601                         spsize = gp->gcstack_size;
 602                         next_segment = gp->gcnext_segment;
 603                         next_sp = gp->gcnext_sp;
 604                         initial_sp = gp->gcinitial_sp;
 605                 } else {
 606                         sp = __splitstack_find_context(&gp->stack_context[0],
 607                                                        &spsize, &next_segment,
 608                                                        &next_sp, &initial_sp);
 609                 }
 610         }
 611         if(sp != nil) {
 612                 scanblock(sp, spsize);
 613                 while((sp = __splitstack_find(next_segment, next_sp,
 614                                               &spsize, &next_segment,
 615                                               &next_sp, &initial_sp)) != nil)
 616                         scanblock(sp, spsize);
 617         }
 618 #else
 619         M *mp;
 620         byte* bottom;
 621         byte* top;
 622
 623         if(gp == runtime_g()) {
 624                 // Scanning our own stack.
 625                 bottom = (byte*)&gp;
 626         } else if((mp = gp->m) != nil && mp->helpgc) {
 627                 // gchelper's stack is in active use and has no interesting pointers.
 628                 return;
 629         } else {
 630                 // Scanning another goroutine's stack.
 631                 // The goroutine is usually asleep (the world is stopped).
 632                 bottom = (byte*)gp->gcnext_sp;
 633                 if(bottom == nil)
 634                         return;
 635         }
 636         top = (byte*)gp->gcinitial_sp + gp->gcstack_size;
 637         if(top > bottom)
 638                 scanblock(bottom, top - bottom);
 639         else
 640                 scanblock(top, bottom - top);
 641 #endif
 642 }
 643
 644 // Markfin calls scanblock on the blocks that have finalizers:
 645 // the things pointed at cannot be freed until the finalizers have run.
 646 static void
 647 markfin(void *v)
 648 {
 649         uintptr size;
 650
 651         size = 0;
 652         if(!runtime_mlookup(v, (byte**)&v, &size, nil) || !runtime_blockspecial(v))
 653                 runtime_throw("mark - finalizer inconsistency");
 654
 655         // do not mark the finalizer block itself.  just mark the things it points at.
 656         scanblock(v, size);
 657 }
 658
 659 static struct root_list* roots;
 660
 661 void
 662 __go_register_gc_roots (struct root_list* r)
 663 {
 664         // FIXME: This needs locking if multiple goroutines can call
 665         // dlopen simultaneously.
 666         r->next = roots;
 667         roots = r;
 668 }
 669
 670 static void
 671 debug_markfin(void *v)
 672 {
 673         uintptr size;
 674
 675         if(!runtime_mlookup(v, (byte**)&v, &size, nil))
 676                 runtime_throw("debug_mark - finalizer inconsistency");
 677         debug_scanblock(v, size);
 678 }
 679
 680 // Mark
 681 static void
 682 mark(void (*scan)(byte*, int64))
 683 {
 684         struct root_list *pl;
 685         G *gp;
 686         FinBlock *fb;
 687
 688         // mark data+bss.
 689         for(pl = roots; pl != nil; pl = pl->next) {
 690                 struct root* pr = &pl->roots[0];
 691                 while(1) {
 692                         void *decl = pr->decl;
 693                         if(decl == nil)
 694                                 break;
 695                         scanblock(decl, pr->size);
 696                         pr++;
 697                 }
 698         }
 699
 700         scan((byte*)&runtime_m0, sizeof runtime_m0);
 701         scan((byte*)&runtime_g0, sizeof runtime_g0);
 702         scan((byte*)&runtime_allg, sizeof runtime_allg);
 703         scan((byte*)&runtime_allm, sizeof runtime_allm);
 704         runtime_MProf_Mark(scan);
 705         runtime_time_scan(scan);
 706         runtime_trampoline_scan(scan);
 707
 708         // mark stacks
 709         for(gp=runtime_allg; gp!=nil; gp=gp->alllink) {
 710                 switch(gp->status){
 711                 default:
 712                         runtime_printf("unexpected G.status %d\n", gp->status);
 713                         runtime_throw("mark - bad status");
 714                 case Gdead:
 715                         break;
 716                 case Grunning:
 717                         if(gp != runtime_g())
 718                                 runtime_throw("mark - world not stopped");
 719                         scanstack(scan, gp);
 720                         break;
 721                 case Grunnable:
 722                 case Gsyscall:
 723                 case Gwaiting:
 724                         scanstack(scan, gp);
 725                         break;
 726                 }
 727         }
 728
 729         // mark things pointed at by objects with finalizers
 730         if(scan == debug_scanblock)
 731                 runtime_walkfintab(debug_markfin, scan);
 732         else
 733                 runtime_walkfintab(markfin, scan);
 734
 735         for(fb=allfin; fb; fb=fb->alllink)
 736                 scanblock((byte*)fb->fin, fb->cnt*sizeof(fb->fin[0]));
 737
 738         // in multiproc mode, join in the queued work.
 739         scan(nil, 0);
 740 }
 741
 742 static bool
 743 handlespecial(byte *p, uintptr size)
 744 {
 745         void (*fn)(void*);
 746         const struct __go_func_type *ft;
 747         FinBlock *block;
 748         Finalizer *f;
 749
 750         if(!runtime_getfinalizer(p, true, &fn, &ft)) {
 751                 runtime_setblockspecial(p, false);
 752                 runtime_MProf_Free(p, size);
 753                 return false;
 754         }
 755
 756         runtime_lock(&finlock);
 757         if(finq == nil || finq->cnt == finq->cap) {
 758                 if(finc == nil) {
 759                         finc = runtime_SysAlloc(PageSize);
 760                         finc->cap = (PageSize - sizeof(FinBlock)) / sizeof(Finalizer) + 1;
 761                         finc->alllink = allfin;
 762                         allfin = finc;
 763                 }
 764                 block = finc;
 765                 finc = block->next;
 766                 block->next = finq;
 767                 finq = block;
 768         }
 769         f = &finq->fin[finq->cnt];
 770         finq->cnt++;
 771         f->fn = fn;
 772         f->ft = ft;
 773         f->arg = p;
 774         runtime_unlock(&finlock);
 775         return true;
 776 }
 777
 778 // Sweep frees or collects finalizers for blocks not marked in the mark phase.
 779 // It clears the mark bits in preparation for the next GC round.
 780 static void
 781 sweep(void)
 782 {
 783         M *m;
 784         MSpan *s;
 785         int32 cl, n, npages;
 786         uintptr size;
 787         byte *p;
 788         MCache *c;
 789         byte *arena_start;
 790         int64 now;
 791
 792         m = runtime_m();
 793         arena_start = runtime_mheap.arena_start;
 794         now = runtime_nanotime();
 795
 796         for(;;) {
 797                 s = work.spans;
 798                 if(s == nil)
 799                         break;
 800                 if(!runtime_casp(&work.spans, s, s->allnext))
 801                         continue;
 802
 803                 // Stamp newly unused spans. The scavenger will use that
 804                 // info to potentially give back some pages to the OS.
 805                 if(s->state == MSpanFree && s->unusedsince == 0)
 806                         s->unusedsince = now;
 807
 808                 if(s->state != MSpanInUse)
 809                         continue;
 810
 811                 p = (byte*)(s->start << PageShift);
 812                 cl = s->sizeclass;
 813                 if(cl == 0) {
 814                         size = s->npages<<PageShift;
 815                         n = 1;
 816                 } else {
 817                         // Chunk full of small blocks.
 818                         size = runtime_class_to_size[cl];
 819                         npages = runtime_class_to_allocnpages[cl];
 820                         n = (npages << PageShift) / size;
 821                 }
 822
 823                 // Sweep through n objects of given size starting at p.
 824                 // This thread owns the span now, so it can manipulate
 825                 // the block bitmap without atomic operations.
 826                 for(; n > 0; n--, p += size) {
 827                         uintptr off, *bitp, shift, bits;
 828
 829                         off = (uintptr*)p - (uintptr*)arena_start;
 830                         bitp = (uintptr*)arena_start - off/wordsPerBitmapWord - 1;
 831                         shift = off % wordsPerBitmapWord;
 832                         bits = *bitp>>shift;
 833
 834                         if((bits & bitAllocated) == 0)
 835                                 continue;
 836
 837                         if((bits & bitMarked) != 0) {
 838                                 if(DebugMark) {
 839                                         if(!(bits & bitSpecial))
 840                                                 runtime_printf("found spurious mark on %p\n", p);
 841                                         *bitp &= ~(bitSpecial<<shift);
 842                                 }
 843                                 *bitp &= ~(bitMarked<<shift);
 844                                 continue;
 845                         }
 846
 847                         // Special means it has a finalizer or is being profiled.
 848                         // In DebugMark mode, the bit has been coopted so
 849                         // we have to assume all blocks are special.
 850                         if(DebugMark || (bits & bitSpecial) != 0) {
 851                                 if(handlespecial(p, size))
 852                                         continue;
 853                         }
 854
 855                         // Mark freed; restore block boundary bit.
 856                         *bitp = (*bitp & ~(bitMask<<shift)) | (bitBlockBoundary<<shift);
 857
 858                         c = m->mcache;
 859                         if(s->sizeclass == 0) {
 860                                 // Free large span.
 861                                 runtime_unmarkspan(p, 1<<PageShift);
 862                                 *(uintptr*)p = 1;       // needs zeroing
 863                                 runtime_MHeap_Free(&runtime_mheap, s, 1);
 864                         } else {
 865                                 // Free small object.
 866                                 if(size > sizeof(uintptr))
 867                                         ((uintptr*)p)[1] = 1;   // mark as "needs to be zeroed"
 868                                 c->local_by_size[s->sizeclass].nfree++;
 869                                 runtime_MCache_Free(c, p, s->sizeclass, size);
 870                         }
 871                         c->local_alloc -= size;
 872                         c->local_nfree++;
 873                 }
 874         }
 875 }
 876
 877 void
 878 runtime_gchelper(void)
 879 {
 880         // Wait until main proc is ready for mark help.
 881         runtime_lock(&work.markgate);
 882         runtime_unlock(&work.markgate);
 883         scanblock(nil, 0);
 884
 885         // Wait until main proc is ready for sweep help.
 886         runtime_lock(&work.sweepgate);
 887         runtime_unlock(&work.sweepgate);
 888         sweep();
 889
 890         if(runtime_xadd(&work.ndone, +1) == work.nproc-1)
 891                 runtime_notewakeup(&work.alldone);
 892 }
 893
 894 // Initialized from $GOGC.  GOGC=off means no gc.
 895 //
 896 // Next gc is after we've allocated an extra amount of
 897 // memory proportional to the amount already in use.
 898 // If gcpercent=100 and we're using 4M, we'll gc again
 899 // when we get to 8M.  This keeps the gc cost in linear
 900 // proportion to the allocation cost.  Adjusting gcpercent
 901 // just changes the linear constant (and also the amount of
 902 // extra memory used).
 903 static int32 gcpercent = -2;
 904
 905 static void
 906 stealcache(void)
 907 {
 908         M *m;
 909
 910         for(m=runtime_allm; m; m=m->alllink)
 911                 runtime_MCache_ReleaseAll(m->mcache);
 912 }
 913
 914 static void
 915 cachestats(void)
 916 {
 917         M *m;
 918         MCache *c;
 919         uint32 i;
 920         uint64 stacks_inuse;
 921         uint64 stacks_sys;
 922
 923         stacks_inuse = 0;
 924         stacks_sys = runtime_stacks_sys;
 925         for(m=runtime_allm; m; m=m->alllink) {
 926                 runtime_purgecachedstats(m);
 927                 // stacks_inuse += m->stackalloc->inuse;
 928                 // stacks_sys += m->stackalloc->sys;
 929                 c = m->mcache;
 930                 for(i=0; i<nelem(c->local_by_size); i++) {
 931                         mstats.by_size[i].nmalloc += c->local_by_size[i].nmalloc;
 932                         c->local_by_size[i].nmalloc = 0;
 933                         mstats.by_size[i].nfree += c->local_by_size[i].nfree;
 934                         c->local_by_size[i].nfree = 0;
 935                 }
 936         }
 937         mstats.stacks_inuse = stacks_inuse;
 938         mstats.stacks_sys = stacks_sys;
 939 }
 940
 941 void
 942 runtime_gc(int32 force)
 943 {
 944         M *m;
 945         int64 t0, t1, t2, t3;
 946         uint64 heap0, heap1, obj0, obj1;
 947         const byte *p;
 948         bool extra;
 949
 950         // Make sure all registers are saved on stack so that
 951         // scanstack sees them.
 952         __builtin_unwind_init();
 953
 954         // The gc is turned off (via enablegc) until
 955         // the bootstrap has completed.
 956         // Also, malloc gets called in the guts
 957         // of a number of libraries that might be
 958         // holding locks.  To avoid priority inversion
 959         // problems, don't bother trying to run gc
 960         // while holding a lock.  The next mallocgc
 961         // without a lock will do the gc instead.
 962         m = runtime_m();
 963         if(!mstats.enablegc || m->locks > 0 || runtime_panicking)
 964                 return;
 965
 966         if(gcpercent == -2) {   // first time through
 967                 p = runtime_getenv("GOGC");
 968                 if(p == nil || p[0] == '\0')
 969                         gcpercent = 100;
 970                 else if(runtime_strcmp((const char*)p, "off") == 0)
 971                         gcpercent = -1;
 972                 else
 973                         gcpercent = runtime_atoi(p);
 974
 975                 p = runtime_getenv("GOGCTRACE");
 976                 if(p != nil)
 977                         gctrace = runtime_atoi(p);
 978         }
 979         if(gcpercent < 0)
 980                 return;
 981
 982         runtime_semacquire(&runtime_worldsema);
 983         if(!force && mstats.heap_alloc < mstats.next_gc) {
 984                 runtime_semrelease(&runtime_worldsema);
 985                 return;
 986         }
 987
 988         t0 = runtime_nanotime();
 989         nhandoff = 0;
 990
 991         m->gcing = 1;
 992         runtime_stoptheworld();
 993
 994         cachestats();
 995         heap0 = mstats.heap_alloc;
 996         obj0 = mstats.nmalloc - mstats.nfree;
 997
 998         runtime_lock(&work.markgate);
 999         runtime_lock(&work.sweepgate);
1000
1001         extra = false;
1002         work.nproc = 1;
1003         if(runtime_gomaxprocs > 1 && runtime_ncpu > 1) {
1004                 runtime_noteclear(&work.alldone);
1005                 work.nproc += runtime_helpgc(&extra);
1006         }
1007         work.nwait = 0;
1008         work.ndone = 0;
1009
1010         runtime_unlock(&work.markgate);  // let the helpers in
1011         mark(scanblock);
1012         if(DebugMark)
1013                 mark(debug_scanblock);
1014         t1 = runtime_nanotime();
1015
1016         work.spans = runtime_mheap.allspans;
1017         runtime_unlock(&work.sweepgate);  // let the helpers in
1018         sweep();
1019         if(work.nproc > 1)
1020                 runtime_notesleep(&work.alldone);
1021         t2 = runtime_nanotime();
1022
1023         stealcache();
1024         cachestats();
1025
1026         mstats.next_gc = mstats.heap_alloc+(mstats.heap_alloc-runtime_stacks_sys)*gcpercent/100;
1027         m->gcing = 0;
1028
1029         m->locks++;     // disable gc during the mallocs in newproc
1030         if(finq != nil) {
1031                 // kick off or wake up goroutine to run queued finalizers
1032                 if(fing == nil)
1033                         fing = __go_go(runfinq, nil);
1034                 else if(fingwait) {
1035                         fingwait = 0;
1036                         runtime_ready(fing);
1037                 }
1038         }
1039         m->locks--;
1040
1041         cachestats();
1042         heap1 = mstats.heap_alloc;
1043         obj1 = mstats.nmalloc - mstats.nfree;
1044
1045         t3 = runtime_nanotime();
1046         mstats.last_gc = t3;
1047         mstats.pause_ns[mstats.numgc%nelem(mstats.pause_ns)] = t3 - t0;
1048         mstats.pause_total_ns += t3 - t0;
1049         mstats.numgc++;
1050         if(mstats.debuggc)
1051                 runtime_printf("pause %D\n", t3-t0);
1052
1053         if(gctrace) {
1054                 runtime_printf("gc%d(%d): %D+%D+%D ms, %D -> %D MB %D -> %D (%D-%D) objects\n",
1055                         mstats.numgc, work.nproc, (t1-t0)/1000000, (t2-t1)/1000000, (t3-t2)/1000000,
1056                         heap0>>20, heap1>>20, obj0, obj1,
1057                         mstats.nmalloc, mstats.nfree);
1058         }
1059
1060         runtime_MProf_GC();
1061         runtime_semrelease(&runtime_worldsema);
1062
1063         // If we could have used another helper proc, start one now,
1064         // in the hope that it will be available next time.
1065         // It would have been even better to start it before the collection,
1066         // but doing so requires allocating memory, so it's tricky to
1067         // coordinate.  This lazy approach works out in practice:
1068         // we don't mind if the first couple gc rounds don't have quite
1069         // the maximum number of procs.
1070         runtime_starttheworld(extra);
1071
1072         // give the queued finalizers, if any, a chance to run
1073         if(finq != nil)
1074                 runtime_gosched();
1075
1076         if(gctrace > 1 && !force)
1077                 runtime_gc(1);
1078 }
1079
1080 void runtime_ReadMemStats(MStats *)
1081   __asm__("runtime.ReadMemStats");
1082
1083 void
1084 runtime_ReadMemStats(MStats *stats)
1085 {
1086         M *m;
1087
1088         // Have to acquire worldsema to stop the world,
1089         // because stoptheworld can only be used by
1090         // one goroutine at a time, and there might be
1091         // a pending garbage collection already calling it.
1092         runtime_semacquire(&runtime_worldsema);
1093         m = runtime_m();
1094         m->gcing = 1;
1095         runtime_stoptheworld();
1096         cachestats();
1097         *stats = mstats;
1098         m->gcing = 0;
1099         runtime_semrelease(&runtime_worldsema);
1100         runtime_starttheworld(false);
1101 }
1102
1103 static void
1104 runfinq(void* dummy __attribute__ ((unused)))
1105 {
1106         G* gp;
1107         Finalizer *f;
1108         FinBlock *fb, *next;
1109         uint32 i;
1110
1111         gp = runtime_g();
1112         for(;;) {
1113                 // There's no need for a lock in this section
1114                 // because it only conflicts with the garbage
1115                 // collector, and the garbage collector only
1116                 // runs when everyone else is stopped, and
1117                 // runfinq only stops at the gosched() or
1118                 // during the calls in the for loop.
1119                 fb = finq;
1120                 finq = nil;
1121                 if(fb == nil) {
1122                         fingwait = 1;
1123                         gp->status = Gwaiting;
1124                         gp->waitreason = "finalizer wait";
1125                         runtime_gosched();
1126                         continue;
1127                 }
1128                 for(; fb; fb=next) {
1129                         next = fb->next;
1130                         for(i=0; i<(uint32)fb->cnt; i++) {
1131                                 void *params[1];
1132
1133                                 f = &fb->fin[i];
1134                                 params[0] = &f->arg;
1135                                 reflect_call(f->ft, (void*)f->fn, 0, 0, params, nil);
1136                                 f->fn = nil;
1137                                 f->arg = nil;
1138                         }
1139                         fb->cnt = 0;
1140                         fb->next = finc;
1141                         finc = fb;
1142                 }
1143                 runtime_gc(1);  // trigger another gc to clean up the finalized objects, if possible
1144         }
1145 }
1146
1147 // mark the block at v of size n as allocated.
1148 // If noptr is true, mark it as having no pointers.
1149 void
1150 runtime_markallocated(void *v, uintptr n, bool noptr)
1151 {
1152         uintptr *b, obits, bits, off, shift;
1153
1154         if(0)
1155                 runtime_printf("markallocated %p+%p\n", v, n);
1156
1157         if((byte*)v+n > (byte*)runtime_mheap.arena_used || (byte*)v < runtime_mheap.arena_start)
1158                 runtime_throw("markallocated: bad pointer");
1159
1160         off = (uintptr*)v - (uintptr*)runtime_mheap.arena_start;  // word offset
1161         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1162         shift = off % wordsPerBitmapWord;
1163
1164         for(;;) {
1165                 obits = *b;
1166                 bits = (obits & ~(bitMask<<shift)) | (bitAllocated<<shift);
1167                 if(noptr)
1168                         bits |= bitNoPointers<<shift;
1169                 if(runtime_singleproc) {
1170                         *b = bits;
1171                         break;
1172                 } else {
1173                         // more than one goroutine is potentially running: use atomic op
1174                         if(runtime_casp((void**)b, (void*)obits, (void*)bits))
1175                                 break;
1176                 }
1177         }
1178 }
1179
1180 // mark the block at v of size n as freed.
1181 void
1182 runtime_markfreed(void *v, uintptr n)
1183 {
1184         uintptr *b, obits, bits, off, shift;
1185
1186         if(0)
1187                 runtime_printf("markallocated %p+%p\n", v, n);
1188
1189         if((byte*)v+n > (byte*)runtime_mheap.arena_used || (byte*)v < runtime_mheap.arena_start)
1190                 runtime_throw("markallocated: bad pointer");
1191
1192         off = (uintptr*)v - (uintptr*)runtime_mheap.arena_start;  // word offset
1193         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1194         shift = off % wordsPerBitmapWord;
1195
1196         for(;;) {
1197                 obits = *b;
1198                 bits = (obits & ~(bitMask<<shift)) | (bitBlockBoundary<<shift);
1199                 if(runtime_singleproc) {
1200                         *b = bits;
1201                         break;
1202                 } else {
1203                         // more than one goroutine is potentially running: use atomic op
1204                         if(runtime_casp((void**)b, (void*)obits, (void*)bits))
1205                                 break;
1206                 }
1207         }
1208 }
1209
1210 // check that the block at v of size n is marked freed.
1211 void
1212 runtime_checkfreed(void *v, uintptr n)
1213 {
1214         uintptr *b, bits, off, shift;
1215
1216         if(!runtime_checking)
1217                 return;
1218
1219         if((byte*)v+n > (byte*)runtime_mheap.arena_used || (byte*)v < runtime_mheap.arena_start)
1220                 return; // not allocated, so okay
1221
1222         off = (uintptr*)v - (uintptr*)runtime_mheap.arena_start;  // word offset
1223         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1224         shift = off % wordsPerBitmapWord;
1225
1226         bits = *b>>shift;
1227         if((bits & bitAllocated) != 0) {
1228                 runtime_printf("checkfreed %p+%p: off=%p have=%p\n",
1229                         v, n, off, bits & bitMask);
1230                 runtime_throw("checkfreed: not freed");
1231         }
1232 }
1233
1234 // mark the span of memory at v as having n blocks of the given size.
1235 // if leftover is true, there is left over space at the end of the span.
1236 void
1237 runtime_markspan(void *v, uintptr size, uintptr n, bool leftover)
1238 {
1239         uintptr *b, off, shift;
1240         byte *p;
1241
1242         if((byte*)v+size*n > (byte*)runtime_mheap.arena_used || (byte*)v < runtime_mheap.arena_start)
1243                 runtime_throw("markspan: bad pointer");
1244
1245         p = v;
1246         if(leftover)    // mark a boundary just past end of last block too
1247                 n++;
1248         for(; n-- > 0; p += size) {
1249                 // Okay to use non-atomic ops here, because we control
1250                 // the entire span, and each bitmap word has bits for only
1251                 // one span, so no other goroutines are changing these
1252                 // bitmap words.
1253                 off = (uintptr*)p - (uintptr*)runtime_mheap.arena_start;  // word offset
1254                 b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1255                 shift = off % wordsPerBitmapWord;
1256                 *b = (*b & ~(bitMask<<shift)) | (bitBlockBoundary<<shift);
1257         }
1258 }
1259
1260 // unmark the span of memory at v of length n bytes.
1261 void
1262 runtime_unmarkspan(void *v, uintptr n)
1263 {
1264         uintptr *p, *b, off;
1265
1266         if((byte*)v+n > (byte*)runtime_mheap.arena_used || (byte*)v < runtime_mheap.arena_start)
1267                 runtime_throw("markspan: bad pointer");
1268
1269         p = v;
1270         off = p - (uintptr*)runtime_mheap.arena_start;  // word offset
1271         if(off % wordsPerBitmapWord != 0)
1272                 runtime_throw("markspan: unaligned pointer");
1273         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1274         n /= PtrSize;
1275         if(n%wordsPerBitmapWord != 0)
1276                 runtime_throw("unmarkspan: unaligned length");
1277         // Okay to use non-atomic ops here, because we control
1278         // the entire span, and each bitmap word has bits for only
1279         // one span, so no other goroutines are changing these
1280         // bitmap words.
1281         n /= wordsPerBitmapWord;
1282         while(n-- > 0)
1283                 *b-- = 0;
1284 }
1285
1286 bool
1287 runtime_blockspecial(void *v)
1288 {
1289         uintptr *b, off, shift;
1290
1291         if(DebugMark)
1292                 return true;
1293
1294         off = (uintptr*)v - (uintptr*)runtime_mheap.arena_start;
1295         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1296         shift = off % wordsPerBitmapWord;
1297
1298         return (*b & (bitSpecial<<shift)) != 0;
1299 }
1300
1301 void
1302 runtime_setblockspecial(void *v, bool s)
1303 {
1304         uintptr *b, off, shift, bits, obits;
1305
1306         if(DebugMark)
1307                 return;
1308
1309         off = (uintptr*)v - (uintptr*)runtime_mheap.arena_start;
1310         b = (uintptr*)runtime_mheap.arena_start - off/wordsPerBitmapWord - 1;
1311         shift = off % wordsPerBitmapWord;
1312
1313         for(;;) {
1314                 obits = *b;
1315                 if(s)
1316                         bits = obits | (bitSpecial<<shift);
1317                 else
1318                         bits = obits & ~(bitSpecial<<shift);
1319                 if(runtime_singleproc) {
1320                         *b = bits;
1321                         break;
1322                 } else {
1323                         // more than one goroutine is potentially running: use atomic op
1324                         if(runtime_casp((void**)b, (void*)obits, (void*)bits))
1325                                 break;
1326                 }
1327         }
1328 }
1329
1330 void
1331 runtime_MHeap_MapBits(MHeap *h)
1332 {
1333         size_t page_size;
1334
1335         // Caller has added extra mappings to the arena.
1336         // Add extra mappings of bitmap words as needed.
1337         // We allocate extra bitmap pieces in chunks of bitmapChunk.
1338         enum {
1339                 bitmapChunk = 8192
1340         };
1341         uintptr n;
1342
1343         n = (h->arena_used - h->arena_start) / wordsPerBitmapWord;
1344         n = (n+bitmapChunk-1) & ~(bitmapChunk-1);
1345         if(h->bitmap_mapped >= n)
1346                 return;
1347
1348         page_size = getpagesize();
1349         n = (n+page_size-1) & ~(page_size-1);
1350
1351         runtime_SysMap(h->arena_start - n, n - h->bitmap_mapped);
1352         h->bitmap_mapped = n;
1353 }