github.com/eun/go@v0.0.0-20170811110501-92cfd07a6cfd/src/runtime/mbitmap.go (about) 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: type and heap bitmaps. 6 // 7 // Stack, data, and bss bitmaps 8 // 9 // Stack frames and global variables in the data and bss sections are described 10 // by 1-bit bitmaps in which 0 means uninteresting and 1 means live pointer 11 // to be visited during GC. The bits in each byte are consumed starting with 12 // the low bit: 1<<0, 1<<1, and so on. 13 // 14 // Heap bitmap 15 // 16 // The allocated heap comes from a subset of the memory in the range [start, used), 17 // where start == mheap_.arena_start and used == mheap_.arena_used. 18 // The heap bitmap comprises 2 bits for each pointer-sized word in that range, 19 // stored in bytes indexed backward in memory from start. 20 // That is, the byte at address start-1 holds the 2-bit entries for the four words 21 // start through start+3*ptrSize, the byte at start-2 holds the entries for 22 // start+4*ptrSize through start+7*ptrSize, and so on. 23 // 24 // In each 2-bit entry, the lower bit holds the same information as in the 1-bit 25 // bitmaps: 0 means uninteresting and 1 means live pointer to be visited during GC. 26 // The meaning of the high bit depends on the position of the word being described 27 // in its allocated object. In all words *except* the second word, the 28 // high bit indicates that the object is still being described. In 29 // these words, if a bit pair with a high bit 0 is encountered, the 30 // low bit can also be assumed to be 0, and the object description is 31 // over. This 00 is called the ``dead'' encoding: it signals that the 32 // rest of the words in the object are uninteresting to the garbage 33 // collector. 34 // 35 // In the second word, the high bit is the GC ``checkmarked'' bit (see below). 36 // 37 // The 2-bit entries are split when written into the byte, so that the top half 38 // of the byte contains 4 high bits and the bottom half contains 4 low (pointer) 39 // bits. 40 // This form allows a copy from the 1-bit to the 4-bit form to keep the 41 // pointer bits contiguous, instead of having to space them out. 42 // 43 // The code makes use of the fact that the zero value for a heap bitmap 44 // has no live pointer bit set and is (depending on position), not used, 45 // not checkmarked, and is the dead encoding. 46 // These properties must be preserved when modifying the encoding. 47 // 48 // The bitmap for noscan spans is not maintained. Code must ensure 49 // that an object is scannable before consulting its bitmap by 50 // checking either the noscan bit in the span or by consulting its 51 // type's information. 52 // 53 // Checkmarks 54 // 55 // In a concurrent garbage collector, one worries about failing to mark 56 // a live object due to mutations without write barriers or bugs in the 57 // collector implementation. As a sanity check, the GC has a 'checkmark' 58 // mode that retraverses the object graph with the world stopped, to make 59 // sure that everything that should be marked is marked. 60 // In checkmark mode, in the heap bitmap, the high bit of the 2-bit entry 61 // for the second word of the object holds the checkmark bit. 62 // When not in checkmark mode, this bit is set to 1. 63 // 64 // The smallest possible allocation is 8 bytes. On a 32-bit machine, that 65 // means every allocated object has two words, so there is room for the 66 // checkmark bit. On a 64-bit machine, however, the 8-byte allocation is 67 // just one word, so the second bit pair is not available for encoding the 68 // checkmark. However, because non-pointer allocations are combined 69 // into larger 16-byte (maxTinySize) allocations, a plain 8-byte allocation 70 // must be a pointer, so the type bit in the first word is not actually needed. 71 // It is still used in general, except in checkmark the type bit is repurposed 72 // as the checkmark bit and then reinitialized (to 1) as the type bit when 73 // finished. 74 // 75 76 package runtime 77 78 import ( 79 "runtime/internal/atomic" 80 "runtime/internal/sys" 81 "unsafe" 82 ) 83 84 const ( 85 bitPointer = 1 << 0 86 bitScan = 1 << 4 87 88 heapBitsShift = 1 // shift offset between successive bitPointer or bitScan entries 89 heapBitmapScale = sys.PtrSize * (8 / 2) // number of data bytes described by one heap bitmap byte 90 91 // all scan/pointer bits in a byte 92 bitScanAll = bitScan | bitScan<<heapBitsShift | bitScan<<(2*heapBitsShift) | bitScan<<(3*heapBitsShift) 93 bitPointerAll = bitPointer | bitPointer<<heapBitsShift | bitPointer<<(2*heapBitsShift) | bitPointer<<(3*heapBitsShift) 94 ) 95 96 // addb returns the byte pointer p+n. 97 //go:nowritebarrier 98 //go:nosplit 99 func addb(p *byte, n uintptr) *byte { 100 // Note: wrote out full expression instead of calling add(p, n) 101 // to reduce the number of temporaries generated by the 102 // compiler for this trivial expression during inlining. 103 return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + n)) 104 } 105 106 // subtractb returns the byte pointer p-n. 107 // subtractb is typically used when traversing the pointer tables referred to by hbits 108 // which are arranged in reverse order. 109 //go:nowritebarrier 110 //go:nosplit 111 func subtractb(p *byte, n uintptr) *byte { 112 // Note: wrote out full expression instead of calling add(p, -n) 113 // to reduce the number of temporaries generated by the 114 // compiler for this trivial expression during inlining. 115 return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - n)) 116 } 117 118 // add1 returns the byte pointer p+1. 119 //go:nowritebarrier 120 //go:nosplit 121 func add1(p *byte) *byte { 122 // Note: wrote out full expression instead of calling addb(p, 1) 123 // to reduce the number of temporaries generated by the 124 // compiler for this trivial expression during inlining. 125 return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + 1)) 126 } 127 128 // subtract1 returns the byte pointer p-1. 129 // subtract1 is typically used when traversing the pointer tables referred to by hbits 130 // which are arranged in reverse order. 131 //go:nowritebarrier 132 // 133 // nosplit because it is used during write barriers and must not be preempted. 134 //go:nosplit 135 func subtract1(p *byte) *byte { 136 // Note: wrote out full expression instead of calling subtractb(p, 1) 137 // to reduce the number of temporaries generated by the 138 // compiler for this trivial expression during inlining. 139 return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - 1)) 140 } 141 142 // mapBits maps any additional bitmap memory needed for the new arena memory. 143 // 144 // Don't call this directly. Call mheap.setArenaUsed. 145 // 146 //go:nowritebarrier 147 func (h *mheap) mapBits(arena_used uintptr) { 148 // Caller has added extra mappings to the arena. 149 // Add extra mappings of bitmap words as needed. 150 // We allocate extra bitmap pieces in chunks of bitmapChunk. 151 const bitmapChunk = 8192 152 153 n := (arena_used - mheap_.arena_start) / heapBitmapScale 154 n = round(n, bitmapChunk) 155 n = round(n, physPageSize) 156 if h.bitmap_mapped >= n { 157 return 158 } 159 160 sysMap(unsafe.Pointer(h.bitmap-n), n-h.bitmap_mapped, h.arena_reserved, &memstats.gc_sys) 161 h.bitmap_mapped = n 162 } 163 164 // heapBits provides access to the bitmap bits for a single heap word. 165 // The methods on heapBits take value receivers so that the compiler 166 // can more easily inline calls to those methods and registerize the 167 // struct fields independently. 168 type heapBits struct { 169 bitp *uint8 170 shift uint32 171 } 172 173 // markBits provides access to the mark bit for an object in the heap. 174 // bytep points to the byte holding the mark bit. 175 // mask is a byte with a single bit set that can be &ed with *bytep 176 // to see if the bit has been set. 177 // *m.byte&m.mask != 0 indicates the mark bit is set. 178 // index can be used along with span information to generate 179 // the address of the object in the heap. 180 // We maintain one set of mark bits for allocation and one for 181 // marking purposes. 182 type markBits struct { 183 bytep *uint8 184 mask uint8 185 index uintptr 186 } 187 188 //go:nosplit 189 func (s *mspan) allocBitsForIndex(allocBitIndex uintptr) markBits { 190 bytep, mask := s.allocBits.bitp(allocBitIndex) 191 return markBits{bytep, mask, allocBitIndex} 192 } 193 194 // refillaCache takes 8 bytes s.allocBits starting at whichByte 195 // and negates them so that ctz (count trailing zeros) instructions 196 // can be used. It then places these 8 bytes into the cached 64 bit 197 // s.allocCache. 198 func (s *mspan) refillAllocCache(whichByte uintptr) { 199 bytes := (*[8]uint8)(unsafe.Pointer(s.allocBits.bytep(whichByte))) 200 aCache := uint64(0) 201 aCache |= uint64(bytes[0]) 202 aCache |= uint64(bytes[1]) << (1 * 8) 203 aCache |= uint64(bytes[2]) << (2 * 8) 204 aCache |= uint64(bytes[3]) << (3 * 8) 205 aCache |= uint64(bytes[4]) << (4 * 8) 206 aCache |= uint64(bytes[5]) << (5 * 8) 207 aCache |= uint64(bytes[6]) << (6 * 8) 208 aCache |= uint64(bytes[7]) << (7 * 8) 209 s.allocCache = ^aCache 210 } 211 212 // nextFreeIndex returns the index of the next free object in s at 213 // or after s.freeindex. 214 // There are hardware instructions that can be used to make this 215 // faster if profiling warrants it. 216 func (s *mspan) nextFreeIndex() uintptr { 217 sfreeindex := s.freeindex 218 snelems := s.nelems 219 if sfreeindex == snelems { 220 return sfreeindex 221 } 222 if sfreeindex > snelems { 223 throw("s.freeindex > s.nelems") 224 } 225 226 aCache := s.allocCache 227 228 bitIndex := sys.Ctz64(aCache) 229 for bitIndex == 64 { 230 // Move index to start of next cached bits. 231 sfreeindex = (sfreeindex + 64) &^ (64 - 1) 232 if sfreeindex >= snelems { 233 s.freeindex = snelems 234 return snelems 235 } 236 whichByte := sfreeindex / 8 237 // Refill s.allocCache with the next 64 alloc bits. 238 s.refillAllocCache(whichByte) 239 aCache = s.allocCache 240 bitIndex = sys.Ctz64(aCache) 241 // nothing available in cached bits 242 // grab the next 8 bytes and try again. 243 } 244 result := sfreeindex + uintptr(bitIndex) 245 if result >= snelems { 246 s.freeindex = snelems 247 return snelems 248 } 249 250 s.allocCache >>= uint(bitIndex + 1) 251 sfreeindex = result + 1 252 253 if sfreeindex%64 == 0 && sfreeindex != snelems { 254 // We just incremented s.freeindex so it isn't 0. 255 // As each 1 in s.allocCache was encountered and used for allocation 256 // it was shifted away. At this point s.allocCache contains all 0s. 257 // Refill s.allocCache so that it corresponds 258 // to the bits at s.allocBits starting at s.freeindex. 259 whichByte := sfreeindex / 8 260 s.refillAllocCache(whichByte) 261 } 262 s.freeindex = sfreeindex 263 return result 264 } 265 266 // isFree returns whether the index'th object in s is unallocated. 267 func (s *mspan) isFree(index uintptr) bool { 268 if index < s.freeindex { 269 return false 270 } 271 bytep, mask := s.allocBits.bitp(index) 272 return *bytep&mask == 0 273 } 274 275 func (s *mspan) objIndex(p uintptr) uintptr { 276 byteOffset := p - s.base() 277 if byteOffset == 0 { 278 return 0 279 } 280 if s.baseMask != 0 { 281 // s.baseMask is 0, elemsize is a power of two, so shift by s.divShift 282 return byteOffset >> s.divShift 283 } 284 return uintptr(((uint64(byteOffset) >> s.divShift) * uint64(s.divMul)) >> s.divShift2) 285 } 286 287 func markBitsForAddr(p uintptr) markBits { 288 s := spanOf(p) 289 objIndex := s.objIndex(p) 290 return s.markBitsForIndex(objIndex) 291 } 292 293 func (s *mspan) markBitsForIndex(objIndex uintptr) markBits { 294 bytep, mask := s.gcmarkBits.bitp(objIndex) 295 return markBits{bytep, mask, objIndex} 296 } 297 298 func (s *mspan) markBitsForBase() markBits { 299 return markBits{(*uint8)(s.gcmarkBits), uint8(1), 0} 300 } 301 302 // isMarked reports whether mark bit m is set. 303 func (m markBits) isMarked() bool { 304 return *m.bytep&m.mask != 0 305 } 306 307 // setMarked sets the marked bit in the markbits, atomically. Some compilers 308 // are not able to inline atomic.Or8 function so if it appears as a hot spot consider 309 // inlining it manually. 310 func (m markBits) setMarked() { 311 // Might be racing with other updates, so use atomic update always. 312 // We used to be clever here and use a non-atomic update in certain 313 // cases, but it's not worth the risk. 314 atomic.Or8(m.bytep, m.mask) 315 } 316 317 // setMarkedNonAtomic sets the marked bit in the markbits, non-atomically. 318 func (m markBits) setMarkedNonAtomic() { 319 *m.bytep |= m.mask 320 } 321 322 // clearMarked clears the marked bit in the markbits, atomically. 323 func (m markBits) clearMarked() { 324 // Might be racing with other updates, so use atomic update always. 325 // We used to be clever here and use a non-atomic update in certain 326 // cases, but it's not worth the risk. 327 atomic.And8(m.bytep, ^m.mask) 328 } 329 330 // markBitsForSpan returns the markBits for the span base address base. 331 func markBitsForSpan(base uintptr) (mbits markBits) { 332 if base < mheap_.arena_start || base >= mheap_.arena_used { 333 throw("markBitsForSpan: base out of range") 334 } 335 mbits = markBitsForAddr(base) 336 if mbits.mask != 1 { 337 throw("markBitsForSpan: unaligned start") 338 } 339 return mbits 340 } 341 342 // advance advances the markBits to the next object in the span. 343 func (m *markBits) advance() { 344 if m.mask == 1<<7 { 345 m.bytep = (*uint8)(unsafe.Pointer(uintptr(unsafe.Pointer(m.bytep)) + 1)) 346 m.mask = 1 347 } else { 348 m.mask = m.mask << 1 349 } 350 m.index++ 351 } 352 353 // heapBitsForAddr returns the heapBits for the address addr. 354 // The caller must have already checked that addr is in the range [mheap_.arena_start, mheap_.arena_used). 355 // 356 // nosplit because it is used during write barriers and must not be preempted. 357 //go:nosplit 358 func heapBitsForAddr(addr uintptr) heapBits { 359 // 2 bits per work, 4 pairs per byte, and a mask is hard coded. 360 off := (addr - mheap_.arena_start) / sys.PtrSize 361 return heapBits{(*uint8)(unsafe.Pointer(mheap_.bitmap - off/4 - 1)), uint32(off & 3)} 362 } 363 364 // heapBitsForSpan returns the heapBits for the span base address base. 365 func heapBitsForSpan(base uintptr) (hbits heapBits) { 366 if base < mheap_.arena_start || base >= mheap_.arena_used { 367 print("runtime: base ", hex(base), " not in range [", hex(mheap_.arena_start), ",", hex(mheap_.arena_used), ")\n") 368 throw("heapBitsForSpan: base out of range") 369 } 370 return heapBitsForAddr(base) 371 } 372 373 // heapBitsForObject returns the base address for the heap object 374 // containing the address p, the heapBits for base, 375 // the object's span, and of the index of the object in s. 376 // If p does not point into a heap object, 377 // return base == 0 378 // otherwise return the base of the object. 379 // 380 // refBase and refOff optionally give the base address of the object 381 // in which the pointer p was found and the byte offset at which it 382 // was found. These are used for error reporting. 383 func heapBitsForObject(p, refBase, refOff uintptr) (base uintptr, hbits heapBits, s *mspan, objIndex uintptr) { 384 arenaStart := mheap_.arena_start 385 if p < arenaStart || p >= mheap_.arena_used { 386 return 387 } 388 off := p - arenaStart 389 idx := off >> _PageShift 390 // p points into the heap, but possibly to the middle of an object. 391 // Consult the span table to find the block beginning. 392 s = mheap_.spans[idx] 393 if s == nil || p < s.base() || p >= s.limit || s.state != mSpanInUse { 394 if s == nil || s.state == _MSpanManual { 395 // If s is nil, the virtual address has never been part of the heap. 396 // This pointer may be to some mmap'd region, so we allow it. 397 // Pointers into stacks are also ok, the runtime manages these explicitly. 398 return 399 } 400 401 // The following ensures that we are rigorous about what data 402 // structures hold valid pointers. 403 if debug.invalidptr != 0 { 404 // Typically this indicates an incorrect use 405 // of unsafe or cgo to store a bad pointer in 406 // the Go heap. It may also indicate a runtime 407 // bug. 408 // 409 // TODO(austin): We could be more aggressive 410 // and detect pointers to unallocated objects 411 // in allocated spans. 412 printlock() 413 print("runtime: pointer ", hex(p)) 414 if s.state != mSpanInUse { 415 print(" to unallocated span") 416 } else { 417 print(" to unused region of span") 418 } 419 print(" idx=", hex(idx), " span.base()=", hex(s.base()), " span.limit=", hex(s.limit), " span.state=", s.state, "\n") 420 if refBase != 0 { 421 print("runtime: found in object at *(", hex(refBase), "+", hex(refOff), ")\n") 422 gcDumpObject("object", refBase, refOff) 423 } 424 getg().m.traceback = 2 425 throw("found bad pointer in Go heap (incorrect use of unsafe or cgo?)") 426 } 427 return 428 } 429 // If this span holds object of a power of 2 size, just mask off the bits to 430 // the interior of the object. Otherwise use the size to get the base. 431 if s.baseMask != 0 { 432 // optimize for power of 2 sized objects. 433 base = s.base() 434 base = base + (p-base)&uintptr(s.baseMask) 435 objIndex = (base - s.base()) >> s.divShift 436 // base = p & s.baseMask is faster for small spans, 437 // but doesn't work for large spans. 438 // Overall, it's faster to use the more general computation above. 439 } else { 440 base = s.base() 441 if p-base >= s.elemsize { 442 // n := (p - base) / s.elemsize, using division by multiplication 443 objIndex = uintptr(p-base) >> s.divShift * uintptr(s.divMul) >> s.divShift2 444 base += objIndex * s.elemsize 445 } 446 } 447 // Now that we know the actual base, compute heapBits to return to caller. 448 hbits = heapBitsForAddr(base) 449 return 450 } 451 452 // next returns the heapBits describing the next pointer-sized word in memory. 453 // That is, if h describes address p, h.next() describes p+ptrSize. 454 // Note that next does not modify h. The caller must record the result. 455 // 456 // nosplit because it is used during write barriers and must not be preempted. 457 //go:nosplit 458 func (h heapBits) next() heapBits { 459 if h.shift < 3*heapBitsShift { 460 return heapBits{h.bitp, h.shift + heapBitsShift} 461 } 462 return heapBits{subtract1(h.bitp), 0} 463 } 464 465 // forward returns the heapBits describing n pointer-sized words ahead of h in memory. 466 // That is, if h describes address p, h.forward(n) describes p+n*ptrSize. 467 // h.forward(1) is equivalent to h.next(), just slower. 468 // Note that forward does not modify h. The caller must record the result. 469 // bits returns the heap bits for the current word. 470 func (h heapBits) forward(n uintptr) heapBits { 471 n += uintptr(h.shift) / heapBitsShift 472 return heapBits{subtractb(h.bitp, n/4), uint32(n%4) * heapBitsShift} 473 } 474 475 // The caller can test morePointers and isPointer by &-ing with bitScan and bitPointer. 476 // The result includes in its higher bits the bits for subsequent words 477 // described by the same bitmap byte. 478 func (h heapBits) bits() uint32 { 479 // The (shift & 31) eliminates a test and conditional branch 480 // from the generated code. 481 return uint32(*h.bitp) >> (h.shift & 31) 482 } 483 484 // morePointers returns true if this word and all remaining words in this object 485 // are scalars. 486 // h must not describe the second word of the object. 487 func (h heapBits) morePointers() bool { 488 return h.bits()&bitScan != 0 489 } 490 491 // isPointer reports whether the heap bits describe a pointer word. 492 // 493 // nosplit because it is used during write barriers and must not be preempted. 494 //go:nosplit 495 func (h heapBits) isPointer() bool { 496 return h.bits()&bitPointer != 0 497 } 498 499 // isCheckmarked reports whether the heap bits have the checkmarked bit set. 500 // It must be told how large the object at h is, because the encoding of the 501 // checkmark bit varies by size. 502 // h must describe the initial word of the object. 503 func (h heapBits) isCheckmarked(size uintptr) bool { 504 if size == sys.PtrSize { 505 return (*h.bitp>>h.shift)&bitPointer != 0 506 } 507 // All multiword objects are 2-word aligned, 508 // so we know that the initial word's 2-bit pair 509 // and the second word's 2-bit pair are in the 510 // same heap bitmap byte, *h.bitp. 511 return (*h.bitp>>(heapBitsShift+h.shift))&bitScan != 0 512 } 513 514 // setCheckmarked sets the checkmarked bit. 515 // It must be told how large the object at h is, because the encoding of the 516 // checkmark bit varies by size. 517 // h must describe the initial word of the object. 518 func (h heapBits) setCheckmarked(size uintptr) { 519 if size == sys.PtrSize { 520 atomic.Or8(h.bitp, bitPointer<<h.shift) 521 return 522 } 523 atomic.Or8(h.bitp, bitScan<<(heapBitsShift+h.shift)) 524 } 525 526 // bulkBarrierPreWrite executes writebarrierptr_prewrite1 527 // for every pointer slot in the memory range [src, src+size), 528 // using pointer/scalar information from [dst, dst+size). 529 // This executes the write barriers necessary before a memmove. 530 // src, dst, and size must be pointer-aligned. 531 // The range [dst, dst+size) must lie within a single object. 532 // 533 // As a special case, src == 0 indicates that this is being used for a 534 // memclr. bulkBarrierPreWrite will pass 0 for the src of each write 535 // barrier. 536 // 537 // Callers should call bulkBarrierPreWrite immediately before 538 // calling memmove(dst, src, size). This function is marked nosplit 539 // to avoid being preempted; the GC must not stop the goroutine 540 // between the memmove and the execution of the barriers. 541 // The caller is also responsible for cgo pointer checks if this 542 // may be writing Go pointers into non-Go memory. 543 // 544 // The pointer bitmap is not maintained for allocations containing 545 // no pointers at all; any caller of bulkBarrierPreWrite must first 546 // make sure the underlying allocation contains pointers, usually 547 // by checking typ.kind&kindNoPointers. 548 // 549 //go:nosplit 550 func bulkBarrierPreWrite(dst, src, size uintptr) { 551 if (dst|src|size)&(sys.PtrSize-1) != 0 { 552 throw("bulkBarrierPreWrite: unaligned arguments") 553 } 554 if !writeBarrier.needed { 555 return 556 } 557 if !inheap(dst) { 558 gp := getg().m.curg 559 if gp != nil && gp.stack.lo <= dst && dst < gp.stack.hi { 560 // Destination is our own stack. No need for barriers. 561 return 562 } 563 564 // If dst is a global, use the data or BSS bitmaps to 565 // execute write barriers. 566 for _, datap := range activeModules() { 567 if datap.data <= dst && dst < datap.edata { 568 bulkBarrierBitmap(dst, src, size, dst-datap.data, datap.gcdatamask.bytedata) 569 return 570 } 571 } 572 for _, datap := range activeModules() { 573 if datap.bss <= dst && dst < datap.ebss { 574 bulkBarrierBitmap(dst, src, size, dst-datap.bss, datap.gcbssmask.bytedata) 575 return 576 } 577 } 578 return 579 } 580 581 h := heapBitsForAddr(dst) 582 if src == 0 { 583 for i := uintptr(0); i < size; i += sys.PtrSize { 584 if h.isPointer() { 585 dstx := (*uintptr)(unsafe.Pointer(dst + i)) 586 writebarrierptr_prewrite1(dstx, 0) 587 } 588 h = h.next() 589 } 590 } else { 591 for i := uintptr(0); i < size; i += sys.PtrSize { 592 if h.isPointer() { 593 dstx := (*uintptr)(unsafe.Pointer(dst + i)) 594 srcx := (*uintptr)(unsafe.Pointer(src + i)) 595 writebarrierptr_prewrite1(dstx, *srcx) 596 } 597 h = h.next() 598 } 599 } 600 } 601 602 // bulkBarrierBitmap executes write barriers for copying from [src, 603 // src+size) to [dst, dst+size) using a 1-bit pointer bitmap. src is 604 // assumed to start maskOffset bytes into the data covered by the 605 // bitmap in bits (which may not be a multiple of 8). 606 // 607 // This is used by bulkBarrierPreWrite for writes to data and BSS. 608 // 609 //go:nosplit 610 func bulkBarrierBitmap(dst, src, size, maskOffset uintptr, bits *uint8) { 611 word := maskOffset / sys.PtrSize 612 bits = addb(bits, word/8) 613 mask := uint8(1) << (word % 8) 614 615 for i := uintptr(0); i < size; i += sys.PtrSize { 616 if mask == 0 { 617 bits = addb(bits, 1) 618 if *bits == 0 { 619 // Skip 8 words. 620 i += 7 * sys.PtrSize 621 continue 622 } 623 mask = 1 624 } 625 if *bits&mask != 0 { 626 dstx := (*uintptr)(unsafe.Pointer(dst + i)) 627 if src == 0 { 628 writebarrierptr_prewrite1(dstx, 0) 629 } else { 630 srcx := (*uintptr)(unsafe.Pointer(src + i)) 631 writebarrierptr_prewrite1(dstx, *srcx) 632 } 633 } 634 mask <<= 1 635 } 636 } 637 638 // typeBitsBulkBarrier executes writebarrierptr_prewrite for every 639 // pointer that would be copied from [src, src+size) to [dst, 640 // dst+size) by a memmove using the type bitmap to locate those 641 // pointer slots. 642 // 643 // The type typ must correspond exactly to [src, src+size) and [dst, dst+size). 644 // dst, src, and size must be pointer-aligned. 645 // The type typ must have a plain bitmap, not a GC program. 646 // The only use of this function is in channel sends, and the 647 // 64 kB channel element limit takes care of this for us. 648 // 649 // Must not be preempted because it typically runs right before memmove, 650 // and the GC must observe them as an atomic action. 651 // 652 //go:nosplit 653 func typeBitsBulkBarrier(typ *_type, dst, src, size uintptr) { 654 if typ == nil { 655 throw("runtime: typeBitsBulkBarrier without type") 656 } 657 if typ.size != size { 658 println("runtime: typeBitsBulkBarrier with type ", typ.string(), " of size ", typ.size, " but memory size", size) 659 throw("runtime: invalid typeBitsBulkBarrier") 660 } 661 if typ.kind&kindGCProg != 0 { 662 println("runtime: typeBitsBulkBarrier with type ", typ.string(), " with GC prog") 663 throw("runtime: invalid typeBitsBulkBarrier") 664 } 665 if !writeBarrier.needed { 666 return 667 } 668 ptrmask := typ.gcdata 669 var bits uint32 670 for i := uintptr(0); i < typ.ptrdata; i += sys.PtrSize { 671 if i&(sys.PtrSize*8-1) == 0 { 672 bits = uint32(*ptrmask) 673 ptrmask = addb(ptrmask, 1) 674 } else { 675 bits = bits >> 1 676 } 677 if bits&1 != 0 { 678 dstx := (*uintptr)(unsafe.Pointer(dst + i)) 679 srcx := (*uintptr)(unsafe.Pointer(src + i)) 680 writebarrierptr_prewrite(dstx, *srcx) 681 } 682 } 683 } 684 685 // The methods operating on spans all require that h has been returned 686 // by heapBitsForSpan and that size, n, total are the span layout description 687 // returned by the mspan's layout method. 688 // If total > size*n, it means that there is extra leftover memory in the span, 689 // usually due to rounding. 690 // 691 // TODO(rsc): Perhaps introduce a different heapBitsSpan type. 692 693 // initSpan initializes the heap bitmap for a span. 694 // It clears all checkmark bits. 695 // If this is a span of pointer-sized objects, it initializes all 696 // words to pointer/scan. 697 // Otherwise, it initializes all words to scalar/dead. 698 func (h heapBits) initSpan(s *mspan) { 699 size, n, total := s.layout() 700 701 // Init the markbit structures 702 s.freeindex = 0 703 s.allocCache = ^uint64(0) // all 1s indicating all free. 704 s.nelems = n 705 s.allocBits = nil 706 s.gcmarkBits = nil 707 s.gcmarkBits = newMarkBits(s.nelems) 708 s.allocBits = newAllocBits(s.nelems) 709 710 // Clear bits corresponding to objects. 711 if total%heapBitmapScale != 0 { 712 throw("initSpan: unaligned length") 713 } 714 nbyte := total / heapBitmapScale 715 if sys.PtrSize == 8 && size == sys.PtrSize { 716 end := h.bitp 717 bitp := subtractb(end, nbyte-1) 718 for { 719 *bitp = bitPointerAll | bitScanAll 720 if bitp == end { 721 break 722 } 723 bitp = add1(bitp) 724 } 725 return 726 } 727 memclrNoHeapPointers(unsafe.Pointer(subtractb(h.bitp, nbyte-1)), nbyte) 728 } 729 730 // initCheckmarkSpan initializes a span for being checkmarked. 731 // It clears the checkmark bits, which are set to 1 in normal operation. 732 func (h heapBits) initCheckmarkSpan(size, n, total uintptr) { 733 // The ptrSize == 8 is a compile-time constant false on 32-bit and eliminates this code entirely. 734 if sys.PtrSize == 8 && size == sys.PtrSize { 735 // Checkmark bit is type bit, bottom bit of every 2-bit entry. 736 // Only possible on 64-bit system, since minimum size is 8. 737 // Must clear type bit (checkmark bit) of every word. 738 // The type bit is the lower of every two-bit pair. 739 bitp := h.bitp 740 for i := uintptr(0); i < n; i += 4 { 741 *bitp &^= bitPointerAll 742 bitp = subtract1(bitp) 743 } 744 return 745 } 746 for i := uintptr(0); i < n; i++ { 747 *h.bitp &^= bitScan << (heapBitsShift + h.shift) 748 h = h.forward(size / sys.PtrSize) 749 } 750 } 751 752 // clearCheckmarkSpan undoes all the checkmarking in a span. 753 // The actual checkmark bits are ignored, so the only work to do 754 // is to fix the pointer bits. (Pointer bits are ignored by scanobject 755 // but consulted by typedmemmove.) 756 func (h heapBits) clearCheckmarkSpan(size, n, total uintptr) { 757 // The ptrSize == 8 is a compile-time constant false on 32-bit and eliminates this code entirely. 758 if sys.PtrSize == 8 && size == sys.PtrSize { 759 // Checkmark bit is type bit, bottom bit of every 2-bit entry. 760 // Only possible on 64-bit system, since minimum size is 8. 761 // Must clear type bit (checkmark bit) of every word. 762 // The type bit is the lower of every two-bit pair. 763 bitp := h.bitp 764 for i := uintptr(0); i < n; i += 4 { 765 *bitp |= bitPointerAll 766 bitp = subtract1(bitp) 767 } 768 } 769 } 770 771 // oneBitCount is indexed by byte and produces the 772 // number of 1 bits in that byte. For example 128 has 1 bit set 773 // and oneBitCount[128] will holds 1. 774 var oneBitCount = [256]uint8{ 775 0, 1, 1, 2, 1, 2, 2, 3, 776 1, 2, 2, 3, 2, 3, 3, 4, 777 1, 2, 2, 3, 2, 3, 3, 4, 778 2, 3, 3, 4, 3, 4, 4, 5, 779 1, 2, 2, 3, 2, 3, 3, 4, 780 2, 3, 3, 4, 3, 4, 4, 5, 781 2, 3, 3, 4, 3, 4, 4, 5, 782 3, 4, 4, 5, 4, 5, 5, 6, 783 1, 2, 2, 3, 2, 3, 3, 4, 784 2, 3, 3, 4, 3, 4, 4, 5, 785 2, 3, 3, 4, 3, 4, 4, 5, 786 3, 4, 4, 5, 4, 5, 5, 6, 787 2, 3, 3, 4, 3, 4, 4, 5, 788 3, 4, 4, 5, 4, 5, 5, 6, 789 3, 4, 4, 5, 4, 5, 5, 6, 790 4, 5, 5, 6, 5, 6, 6, 7, 791 1, 2, 2, 3, 2, 3, 3, 4, 792 2, 3, 3, 4, 3, 4, 4, 5, 793 2, 3, 3, 4, 3, 4, 4, 5, 794 3, 4, 4, 5, 4, 5, 5, 6, 795 2, 3, 3, 4, 3, 4, 4, 5, 796 3, 4, 4, 5, 4, 5, 5, 6, 797 3, 4, 4, 5, 4, 5, 5, 6, 798 4, 5, 5, 6, 5, 6, 6, 7, 799 2, 3, 3, 4, 3, 4, 4, 5, 800 3, 4, 4, 5, 4, 5, 5, 6, 801 3, 4, 4, 5, 4, 5, 5, 6, 802 4, 5, 5, 6, 5, 6, 6, 7, 803 3, 4, 4, 5, 4, 5, 5, 6, 804 4, 5, 5, 6, 5, 6, 6, 7, 805 4, 5, 5, 6, 5, 6, 6, 7, 806 5, 6, 6, 7, 6, 7, 7, 8} 807 808 // countAlloc returns the number of objects allocated in span s by 809 // scanning the allocation bitmap. 810 // TODO:(rlh) Use popcount intrinsic. 811 func (s *mspan) countAlloc() int { 812 count := 0 813 maxIndex := s.nelems / 8 814 for i := uintptr(0); i < maxIndex; i++ { 815 mrkBits := *s.gcmarkBits.bytep(i) 816 count += int(oneBitCount[mrkBits]) 817 } 818 if bitsInLastByte := s.nelems % 8; bitsInLastByte != 0 { 819 mrkBits := *s.gcmarkBits.bytep(maxIndex) 820 mask := uint8((1 << bitsInLastByte) - 1) 821 bits := mrkBits & mask 822 count += int(oneBitCount[bits]) 823 } 824 return count 825 } 826 827 // heapBitsSetType records that the new allocation [x, x+size) 828 // holds in [x, x+dataSize) one or more values of type typ. 829 // (The number of values is given by dataSize / typ.size.) 830 // If dataSize < size, the fragment [x+dataSize, x+size) is 831 // recorded as non-pointer data. 832 // It is known that the type has pointers somewhere; 833 // malloc does not call heapBitsSetType when there are no pointers, 834 // because all free objects are marked as noscan during 835 // heapBitsSweepSpan. 836 // 837 // There can only be one allocation from a given span active at a time, 838 // and the bitmap for a span always falls on byte boundaries, 839 // so there are no write-write races for access to the heap bitmap. 840 // Hence, heapBitsSetType can access the bitmap without atomics. 841 // 842 // There can be read-write races between heapBitsSetType and things 843 // that read the heap bitmap like scanobject. However, since 844 // heapBitsSetType is only used for objects that have not yet been 845 // made reachable, readers will ignore bits being modified by this 846 // function. This does mean this function cannot transiently modify 847 // bits that belong to neighboring objects. Also, on weakly-ordered 848 // machines, callers must execute a store/store (publication) barrier 849 // between calling this function and making the object reachable. 850 func heapBitsSetType(x, size, dataSize uintptr, typ *_type) { 851 const doubleCheck = false // slow but helpful; enable to test modifications to this code 852 853 // dataSize is always size rounded up to the next malloc size class, 854 // except in the case of allocating a defer block, in which case 855 // size is sizeof(_defer{}) (at least 6 words) and dataSize may be 856 // arbitrarily larger. 857 // 858 // The checks for size == sys.PtrSize and size == 2*sys.PtrSize can therefore 859 // assume that dataSize == size without checking it explicitly. 860 861 if sys.PtrSize == 8 && size == sys.PtrSize { 862 // It's one word and it has pointers, it must be a pointer. 863 // Since all allocated one-word objects are pointers 864 // (non-pointers are aggregated into tinySize allocations), 865 // initSpan sets the pointer bits for us. Nothing to do here. 866 if doubleCheck { 867 h := heapBitsForAddr(x) 868 if !h.isPointer() { 869 throw("heapBitsSetType: pointer bit missing") 870 } 871 if !h.morePointers() { 872 throw("heapBitsSetType: scan bit missing") 873 } 874 } 875 return 876 } 877 878 h := heapBitsForAddr(x) 879 ptrmask := typ.gcdata // start of 1-bit pointer mask (or GC program, handled below) 880 881 // Heap bitmap bits for 2-word object are only 4 bits, 882 // so also shared with objects next to it. 883 // This is called out as a special case primarily for 32-bit systems, 884 // so that on 32-bit systems the code below can assume all objects 885 // are 4-word aligned (because they're all 16-byte aligned). 886 if size == 2*sys.PtrSize { 887 if typ.size == sys.PtrSize { 888 // We're allocating a block big enough to hold two pointers. 889 // On 64-bit, that means the actual object must be two pointers, 890 // or else we'd have used the one-pointer-sized block. 891 // On 32-bit, however, this is the 8-byte block, the smallest one. 892 // So it could be that we're allocating one pointer and this was 893 // just the smallest block available. Distinguish by checking dataSize. 894 // (In general the number of instances of typ being allocated is 895 // dataSize/typ.size.) 896 if sys.PtrSize == 4 && dataSize == sys.PtrSize { 897 // 1 pointer object. On 32-bit machines clear the bit for the 898 // unused second word. 899 *h.bitp &^= (bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << h.shift 900 *h.bitp |= (bitPointer | bitScan) << h.shift 901 } else { 902 // 2-element slice of pointer. 903 *h.bitp |= (bitPointer | bitScan | bitPointer<<heapBitsShift) << h.shift 904 } 905 return 906 } 907 // Otherwise typ.size must be 2*sys.PtrSize, 908 // and typ.kind&kindGCProg == 0. 909 if doubleCheck { 910 if typ.size != 2*sys.PtrSize || typ.kind&kindGCProg != 0 { 911 print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, " gcprog=", typ.kind&kindGCProg != 0, "\n") 912 throw("heapBitsSetType") 913 } 914 } 915 b := uint32(*ptrmask) 916 hb := (b & 3) | bitScan 917 // bitPointer == 1, bitScan is 1 << 4, heapBitsShift is 1. 918 // 110011 is shifted h.shift and complemented. 919 // This clears out the bits that are about to be 920 // ored into *h.hbitp in the next instructions. 921 *h.bitp &^= (bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << h.shift 922 *h.bitp |= uint8(hb << h.shift) 923 return 924 } 925 926 // Copy from 1-bit ptrmask into 2-bit bitmap. 927 // The basic approach is to use a single uintptr as a bit buffer, 928 // alternating between reloading the buffer and writing bitmap bytes. 929 // In general, one load can supply two bitmap byte writes. 930 // This is a lot of lines of code, but it compiles into relatively few 931 // machine instructions. 932 933 var ( 934 // Ptrmask input. 935 p *byte // last ptrmask byte read 936 b uintptr // ptrmask bits already loaded 937 nb uintptr // number of bits in b at next read 938 endp *byte // final ptrmask byte to read (then repeat) 939 endnb uintptr // number of valid bits in *endp 940 pbits uintptr // alternate source of bits 941 942 // Heap bitmap output. 943 w uintptr // words processed 944 nw uintptr // number of words to process 945 hbitp *byte // next heap bitmap byte to write 946 hb uintptr // bits being prepared for *hbitp 947 ) 948 949 hbitp = h.bitp 950 951 // Handle GC program. Delayed until this part of the code 952 // so that we can use the same double-checking mechanism 953 // as the 1-bit case. Nothing above could have encountered 954 // GC programs: the cases were all too small. 955 if typ.kind&kindGCProg != 0 { 956 heapBitsSetTypeGCProg(h, typ.ptrdata, typ.size, dataSize, size, addb(typ.gcdata, 4)) 957 if doubleCheck { 958 // Double-check the heap bits written by GC program 959 // by running the GC program to create a 1-bit pointer mask 960 // and then jumping to the double-check code below. 961 // This doesn't catch bugs shared between the 1-bit and 4-bit 962 // GC program execution, but it does catch mistakes specific 963 // to just one of those and bugs in heapBitsSetTypeGCProg's 964 // implementation of arrays. 965 lock(&debugPtrmask.lock) 966 if debugPtrmask.data == nil { 967 debugPtrmask.data = (*byte)(persistentalloc(1<<20, 1, &memstats.other_sys)) 968 } 969 ptrmask = debugPtrmask.data 970 runGCProg(addb(typ.gcdata, 4), nil, ptrmask, 1) 971 goto Phase4 972 } 973 return 974 } 975 976 // Note about sizes: 977 // 978 // typ.size is the number of words in the object, 979 // and typ.ptrdata is the number of words in the prefix 980 // of the object that contains pointers. That is, the final 981 // typ.size - typ.ptrdata words contain no pointers. 982 // This allows optimization of a common pattern where 983 // an object has a small header followed by a large scalar 984 // buffer. If we know the pointers are over, we don't have 985 // to scan the buffer's heap bitmap at all. 986 // The 1-bit ptrmasks are sized to contain only bits for 987 // the typ.ptrdata prefix, zero padded out to a full byte 988 // of bitmap. This code sets nw (below) so that heap bitmap 989 // bits are only written for the typ.ptrdata prefix; if there is 990 // more room in the allocated object, the next heap bitmap 991 // entry is a 00, indicating that there are no more pointers 992 // to scan. So only the ptrmask for the ptrdata bytes is needed. 993 // 994 // Replicated copies are not as nice: if there is an array of 995 // objects with scalar tails, all but the last tail does have to 996 // be initialized, because there is no way to say "skip forward". 997 // However, because of the possibility of a repeated type with 998 // size not a multiple of 4 pointers (one heap bitmap byte), 999 // the code already must handle the last ptrmask byte specially 1000 // by treating it as containing only the bits for endnb pointers, 1001 // where endnb <= 4. We represent large scalar tails that must 1002 // be expanded in the replication by setting endnb larger than 4. 1003 // This will have the effect of reading many bits out of b, 1004 // but once the real bits are shifted out, b will supply as many 1005 // zero bits as we try to read, which is exactly what we need. 1006 1007 p = ptrmask 1008 if typ.size < dataSize { 1009 // Filling in bits for an array of typ. 1010 // Set up for repetition of ptrmask during main loop. 1011 // Note that ptrmask describes only a prefix of 1012 const maxBits = sys.PtrSize*8 - 7 1013 if typ.ptrdata/sys.PtrSize <= maxBits { 1014 // Entire ptrmask fits in uintptr with room for a byte fragment. 1015 // Load into pbits and never read from ptrmask again. 1016 // This is especially important when the ptrmask has 1017 // fewer than 8 bits in it; otherwise the reload in the middle 1018 // of the Phase 2 loop would itself need to loop to gather 1019 // at least 8 bits. 1020 1021 // Accumulate ptrmask into b. 1022 // ptrmask is sized to describe only typ.ptrdata, but we record 1023 // it as describing typ.size bytes, since all the high bits are zero. 1024 nb = typ.ptrdata / sys.PtrSize 1025 for i := uintptr(0); i < nb; i += 8 { 1026 b |= uintptr(*p) << i 1027 p = add1(p) 1028 } 1029 nb = typ.size / sys.PtrSize 1030 1031 // Replicate ptrmask to fill entire pbits uintptr. 1032 // Doubling and truncating is fewer steps than 1033 // iterating by nb each time. (nb could be 1.) 1034 // Since we loaded typ.ptrdata/sys.PtrSize bits 1035 // but are pretending to have typ.size/sys.PtrSize, 1036 // there might be no replication necessary/possible. 1037 pbits = b 1038 endnb = nb 1039 if nb+nb <= maxBits { 1040 for endnb <= sys.PtrSize*8 { 1041 pbits |= pbits << endnb 1042 endnb += endnb 1043 } 1044 // Truncate to a multiple of original ptrmask. 1045 // Because nb+nb <= maxBits, nb fits in a byte. 1046 // Byte division is cheaper than uintptr division. 1047 endnb = uintptr(maxBits/byte(nb)) * nb 1048 pbits &= 1<<endnb - 1 1049 b = pbits 1050 nb = endnb 1051 } 1052 1053 // Clear p and endp as sentinel for using pbits. 1054 // Checked during Phase 2 loop. 1055 p = nil 1056 endp = nil 1057 } else { 1058 // Ptrmask is larger. Read it multiple times. 1059 n := (typ.ptrdata/sys.PtrSize+7)/8 - 1 1060 endp = addb(ptrmask, n) 1061 endnb = typ.size/sys.PtrSize - n*8 1062 } 1063 } 1064 if p != nil { 1065 b = uintptr(*p) 1066 p = add1(p) 1067 nb = 8 1068 } 1069 1070 if typ.size == dataSize { 1071 // Single entry: can stop once we reach the non-pointer data. 1072 nw = typ.ptrdata / sys.PtrSize 1073 } else { 1074 // Repeated instances of typ in an array. 1075 // Have to process first N-1 entries in full, but can stop 1076 // once we reach the non-pointer data in the final entry. 1077 nw = ((dataSize/typ.size-1)*typ.size + typ.ptrdata) / sys.PtrSize 1078 } 1079 if nw == 0 { 1080 // No pointers! Caller was supposed to check. 1081 println("runtime: invalid type ", typ.string()) 1082 throw("heapBitsSetType: called with non-pointer type") 1083 return 1084 } 1085 if nw < 2 { 1086 // Must write at least 2 words, because the "no scan" 1087 // encoding doesn't take effect until the third word. 1088 nw = 2 1089 } 1090 1091 // Phase 1: Special case for leading byte (shift==0) or half-byte (shift==4). 1092 // The leading byte is special because it contains the bits for word 1, 1093 // which does not have the scan bit set. 1094 // The leading half-byte is special because it's a half a byte, 1095 // so we have to be careful with the bits already there. 1096 switch { 1097 default: 1098 throw("heapBitsSetType: unexpected shift") 1099 1100 case h.shift == 0: 1101 // Ptrmask and heap bitmap are aligned. 1102 // Handle first byte of bitmap specially. 1103 // 1104 // The first byte we write out covers the first four 1105 // words of the object. The scan/dead bit on the first 1106 // word must be set to scan since there are pointers 1107 // somewhere in the object. The scan/dead bit on the 1108 // second word is the checkmark, so we don't set it. 1109 // In all following words, we set the scan/dead 1110 // appropriately to indicate that the object contains 1111 // to the next 2-bit entry in the bitmap. 1112 // 1113 // TODO: It doesn't matter if we set the checkmark, so 1114 // maybe this case isn't needed any more. 1115 hb = b & bitPointerAll 1116 hb |= bitScan | bitScan<<(2*heapBitsShift) | bitScan<<(3*heapBitsShift) 1117 if w += 4; w >= nw { 1118 goto Phase3 1119 } 1120 *hbitp = uint8(hb) 1121 hbitp = subtract1(hbitp) 1122 b >>= 4 1123 nb -= 4 1124 1125 case sys.PtrSize == 8 && h.shift == 2: 1126 // Ptrmask and heap bitmap are misaligned. 1127 // The bits for the first two words are in a byte shared 1128 // with another object, so we must be careful with the bits 1129 // already there. 1130 // We took care of 1-word and 2-word objects above, 1131 // so this is at least a 6-word object. 1132 hb = (b & (bitPointer | bitPointer<<heapBitsShift)) << (2 * heapBitsShift) 1133 // This is not noscan, so set the scan bit in the 1134 // first word. 1135 hb |= bitScan << (2 * heapBitsShift) 1136 b >>= 2 1137 nb -= 2 1138 // Note: no bitScan for second word because that's 1139 // the checkmark. 1140 *hbitp &^= uint8((bitPointer | bitScan | (bitPointer << heapBitsShift)) << (2 * heapBitsShift)) 1141 *hbitp |= uint8(hb) 1142 hbitp = subtract1(hbitp) 1143 if w += 2; w >= nw { 1144 // We know that there is more data, because we handled 2-word objects above. 1145 // This must be at least a 6-word object. If we're out of pointer words, 1146 // mark no scan in next bitmap byte and finish. 1147 hb = 0 1148 w += 4 1149 goto Phase3 1150 } 1151 } 1152 1153 // Phase 2: Full bytes in bitmap, up to but not including write to last byte (full or partial) in bitmap. 1154 // The loop computes the bits for that last write but does not execute the write; 1155 // it leaves the bits in hb for processing by phase 3. 1156 // To avoid repeated adjustment of nb, we subtract out the 4 bits we're going to 1157 // use in the first half of the loop right now, and then we only adjust nb explicitly 1158 // if the 8 bits used by each iteration isn't balanced by 8 bits loaded mid-loop. 1159 nb -= 4 1160 for { 1161 // Emit bitmap byte. 1162 // b has at least nb+4 bits, with one exception: 1163 // if w+4 >= nw, then b has only nw-w bits, 1164 // but we'll stop at the break and then truncate 1165 // appropriately in Phase 3. 1166 hb = b & bitPointerAll 1167 hb |= bitScanAll 1168 if w += 4; w >= nw { 1169 break 1170 } 1171 *hbitp = uint8(hb) 1172 hbitp = subtract1(hbitp) 1173 b >>= 4 1174 1175 // Load more bits. b has nb right now. 1176 if p != endp { 1177 // Fast path: keep reading from ptrmask. 1178 // nb unmodified: we just loaded 8 bits, 1179 // and the next iteration will consume 8 bits, 1180 // leaving us with the same nb the next time we're here. 1181 if nb < 8 { 1182 b |= uintptr(*p) << nb 1183 p = add1(p) 1184 } else { 1185 // Reduce the number of bits in b. 1186 // This is important if we skipped 1187 // over a scalar tail, since nb could 1188 // be larger than the bit width of b. 1189 nb -= 8 1190 } 1191 } else if p == nil { 1192 // Almost as fast path: track bit count and refill from pbits. 1193 // For short repetitions. 1194 if nb < 8 { 1195 b |= pbits << nb 1196 nb += endnb 1197 } 1198 nb -= 8 // for next iteration 1199 } else { 1200 // Slow path: reached end of ptrmask. 1201 // Process final partial byte and rewind to start. 1202 b |= uintptr(*p) << nb 1203 nb += endnb 1204 if nb < 8 { 1205 b |= uintptr(*ptrmask) << nb 1206 p = add1(ptrmask) 1207 } else { 1208 nb -= 8 1209 p = ptrmask 1210 } 1211 } 1212 1213 // Emit bitmap byte. 1214 hb = b & bitPointerAll 1215 hb |= bitScanAll 1216 if w += 4; w >= nw { 1217 break 1218 } 1219 *hbitp = uint8(hb) 1220 hbitp = subtract1(hbitp) 1221 b >>= 4 1222 } 1223 1224 Phase3: 1225 // Phase 3: Write last byte or partial byte and zero the rest of the bitmap entries. 1226 if w > nw { 1227 // Counting the 4 entries in hb not yet written to memory, 1228 // there are more entries than possible pointer slots. 1229 // Discard the excess entries (can't be more than 3). 1230 mask := uintptr(1)<<(4-(w-nw)) - 1 1231 hb &= mask | mask<<4 // apply mask to both pointer bits and scan bits 1232 } 1233 1234 // Change nw from counting possibly-pointer words to total words in allocation. 1235 nw = size / sys.PtrSize 1236 1237 // Write whole bitmap bytes. 1238 // The first is hb, the rest are zero. 1239 if w <= nw { 1240 *hbitp = uint8(hb) 1241 hbitp = subtract1(hbitp) 1242 hb = 0 // for possible final half-byte below 1243 for w += 4; w <= nw; w += 4 { 1244 *hbitp = 0 1245 hbitp = subtract1(hbitp) 1246 } 1247 } 1248 1249 // Write final partial bitmap byte if any. 1250 // We know w > nw, or else we'd still be in the loop above. 1251 // It can be bigger only due to the 4 entries in hb that it counts. 1252 // If w == nw+4 then there's nothing left to do: we wrote all nw entries 1253 // and can discard the 4 sitting in hb. 1254 // But if w == nw+2, we need to write first two in hb. 1255 // The byte is shared with the next object, so be careful with 1256 // existing bits. 1257 if w == nw+2 { 1258 *hbitp = *hbitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | uint8(hb) 1259 } 1260 1261 Phase4: 1262 // Phase 4: all done, but perhaps double check. 1263 if doubleCheck { 1264 end := heapBitsForAddr(x + size) 1265 if typ.kind&kindGCProg == 0 && (hbitp != end.bitp || (w == nw+2) != (end.shift == 2)) { 1266 println("ended at wrong bitmap byte for", typ.string(), "x", dataSize/typ.size) 1267 print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n") 1268 print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n") 1269 h0 := heapBitsForAddr(x) 1270 print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n") 1271 print("ended at hbitp=", hbitp, " but next starts at bitp=", end.bitp, " shift=", end.shift, "\n") 1272 throw("bad heapBitsSetType") 1273 } 1274 1275 // Double-check that bits to be written were written correctly. 1276 // Does not check that other bits were not written, unfortunately. 1277 h := heapBitsForAddr(x) 1278 nptr := typ.ptrdata / sys.PtrSize 1279 ndata := typ.size / sys.PtrSize 1280 count := dataSize / typ.size 1281 totalptr := ((count-1)*typ.size + typ.ptrdata) / sys.PtrSize 1282 for i := uintptr(0); i < size/sys.PtrSize; i++ { 1283 j := i % ndata 1284 var have, want uint8 1285 have = (*h.bitp >> h.shift) & (bitPointer | bitScan) 1286 if i >= totalptr { 1287 want = 0 // deadmarker 1288 if typ.kind&kindGCProg != 0 && i < (totalptr+3)/4*4 { 1289 want = bitScan 1290 } 1291 } else { 1292 if j < nptr && (*addb(ptrmask, j/8)>>(j%8))&1 != 0 { 1293 want |= bitPointer 1294 } 1295 if i != 1 { 1296 want |= bitScan 1297 } else { 1298 have &^= bitScan 1299 } 1300 } 1301 if have != want { 1302 println("mismatch writing bits for", typ.string(), "x", dataSize/typ.size) 1303 print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n") 1304 print("kindGCProg=", typ.kind&kindGCProg != 0, "\n") 1305 print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n") 1306 h0 := heapBitsForAddr(x) 1307 print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n") 1308 print("current bits h.bitp=", h.bitp, " h.shift=", h.shift, " *h.bitp=", hex(*h.bitp), "\n") 1309 print("ptrmask=", ptrmask, " p=", p, " endp=", endp, " endnb=", endnb, " pbits=", hex(pbits), " b=", hex(b), " nb=", nb, "\n") 1310 println("at word", i, "offset", i*sys.PtrSize, "have", have, "want", want) 1311 if typ.kind&kindGCProg != 0 { 1312 println("GC program:") 1313 dumpGCProg(addb(typ.gcdata, 4)) 1314 } 1315 throw("bad heapBitsSetType") 1316 } 1317 h = h.next() 1318 } 1319 if ptrmask == debugPtrmask.data { 1320 unlock(&debugPtrmask.lock) 1321 } 1322 } 1323 } 1324 1325 var debugPtrmask struct { 1326 lock mutex 1327 data *byte 1328 } 1329 1330 // heapBitsSetTypeGCProg implements heapBitsSetType using a GC program. 1331 // progSize is the size of the memory described by the program. 1332 // elemSize is the size of the element that the GC program describes (a prefix of). 1333 // dataSize is the total size of the intended data, a multiple of elemSize. 1334 // allocSize is the total size of the allocated memory. 1335 // 1336 // GC programs are only used for large allocations. 1337 // heapBitsSetType requires that allocSize is a multiple of 4 words, 1338 // so that the relevant bitmap bytes are not shared with surrounding 1339 // objects. 1340 func heapBitsSetTypeGCProg(h heapBits, progSize, elemSize, dataSize, allocSize uintptr, prog *byte) { 1341 if sys.PtrSize == 8 && allocSize%(4*sys.PtrSize) != 0 { 1342 // Alignment will be wrong. 1343 throw("heapBitsSetTypeGCProg: small allocation") 1344 } 1345 var totalBits uintptr 1346 if elemSize == dataSize { 1347 totalBits = runGCProg(prog, nil, h.bitp, 2) 1348 if totalBits*sys.PtrSize != progSize { 1349 println("runtime: heapBitsSetTypeGCProg: total bits", totalBits, "but progSize", progSize) 1350 throw("heapBitsSetTypeGCProg: unexpected bit count") 1351 } 1352 } else { 1353 count := dataSize / elemSize 1354 1355 // Piece together program trailer to run after prog that does: 1356 // literal(0) 1357 // repeat(1, elemSize-progSize-1) // zeros to fill element size 1358 // repeat(elemSize, count-1) // repeat that element for count 1359 // This zero-pads the data remaining in the first element and then 1360 // repeats that first element to fill the array. 1361 var trailer [40]byte // 3 varints (max 10 each) + some bytes 1362 i := 0 1363 if n := elemSize/sys.PtrSize - progSize/sys.PtrSize; n > 0 { 1364 // literal(0) 1365 trailer[i] = 0x01 1366 i++ 1367 trailer[i] = 0 1368 i++ 1369 if n > 1 { 1370 // repeat(1, n-1) 1371 trailer[i] = 0x81 1372 i++ 1373 n-- 1374 for ; n >= 0x80; n >>= 7 { 1375 trailer[i] = byte(n | 0x80) 1376 i++ 1377 } 1378 trailer[i] = byte(n) 1379 i++ 1380 } 1381 } 1382 // repeat(elemSize/ptrSize, count-1) 1383 trailer[i] = 0x80 1384 i++ 1385 n := elemSize / sys.PtrSize 1386 for ; n >= 0x80; n >>= 7 { 1387 trailer[i] = byte(n | 0x80) 1388 i++ 1389 } 1390 trailer[i] = byte(n) 1391 i++ 1392 n = count - 1 1393 for ; n >= 0x80; n >>= 7 { 1394 trailer[i] = byte(n | 0x80) 1395 i++ 1396 } 1397 trailer[i] = byte(n) 1398 i++ 1399 trailer[i] = 0 1400 i++ 1401 1402 runGCProg(prog, &trailer[0], h.bitp, 2) 1403 1404 // Even though we filled in the full array just now, 1405 // record that we only filled in up to the ptrdata of the 1406 // last element. This will cause the code below to 1407 // memclr the dead section of the final array element, 1408 // so that scanobject can stop early in the final element. 1409 totalBits = (elemSize*(count-1) + progSize) / sys.PtrSize 1410 } 1411 endProg := unsafe.Pointer(subtractb(h.bitp, (totalBits+3)/4)) 1412 endAlloc := unsafe.Pointer(subtractb(h.bitp, allocSize/heapBitmapScale)) 1413 memclrNoHeapPointers(add(endAlloc, 1), uintptr(endProg)-uintptr(endAlloc)) 1414 } 1415 1416 // progToPointerMask returns the 1-bit pointer mask output by the GC program prog. 1417 // size the size of the region described by prog, in bytes. 1418 // The resulting bitvector will have no more than size/sys.PtrSize bits. 1419 func progToPointerMask(prog *byte, size uintptr) bitvector { 1420 n := (size/sys.PtrSize + 7) / 8 1421 x := (*[1 << 30]byte)(persistentalloc(n+1, 1, &memstats.buckhash_sys))[:n+1] 1422 x[len(x)-1] = 0xa1 // overflow check sentinel 1423 n = runGCProg(prog, nil, &x[0], 1) 1424 if x[len(x)-1] != 0xa1 { 1425 throw("progToPointerMask: overflow") 1426 } 1427 return bitvector{int32(n), &x[0]} 1428 } 1429 1430 // Packed GC pointer bitmaps, aka GC programs. 1431 // 1432 // For large types containing arrays, the type information has a 1433 // natural repetition that can be encoded to save space in the 1434 // binary and in the memory representation of the type information. 1435 // 1436 // The encoding is a simple Lempel-Ziv style bytecode machine 1437 // with the following instructions: 1438 // 1439 // 00000000: stop 1440 // 0nnnnnnn: emit n bits copied from the next (n+7)/8 bytes 1441 // 10000000 n c: repeat the previous n bits c times; n, c are varints 1442 // 1nnnnnnn c: repeat the previous n bits c times; c is a varint 1443 1444 // runGCProg executes the GC program prog, and then trailer if non-nil, 1445 // writing to dst with entries of the given size. 1446 // If size == 1, dst is a 1-bit pointer mask laid out moving forward from dst. 1447 // If size == 2, dst is the 2-bit heap bitmap, and writes move backward 1448 // starting at dst (because the heap bitmap does). In this case, the caller guarantees 1449 // that only whole bytes in dst need to be written. 1450 // 1451 // runGCProg returns the number of 1- or 2-bit entries written to memory. 1452 func runGCProg(prog, trailer, dst *byte, size int) uintptr { 1453 dstStart := dst 1454 1455 // Bits waiting to be written to memory. 1456 var bits uintptr 1457 var nbits uintptr 1458 1459 p := prog 1460 Run: 1461 for { 1462 // Flush accumulated full bytes. 1463 // The rest of the loop assumes that nbits <= 7. 1464 for ; nbits >= 8; nbits -= 8 { 1465 if size == 1 { 1466 *dst = uint8(bits) 1467 dst = add1(dst) 1468 bits >>= 8 1469 } else { 1470 v := bits&bitPointerAll | bitScanAll 1471 *dst = uint8(v) 1472 dst = subtract1(dst) 1473 bits >>= 4 1474 v = bits&bitPointerAll | bitScanAll 1475 *dst = uint8(v) 1476 dst = subtract1(dst) 1477 bits >>= 4 1478 } 1479 } 1480 1481 // Process one instruction. 1482 inst := uintptr(*p) 1483 p = add1(p) 1484 n := inst & 0x7F 1485 if inst&0x80 == 0 { 1486 // Literal bits; n == 0 means end of program. 1487 if n == 0 { 1488 // Program is over; continue in trailer if present. 1489 if trailer != nil { 1490 //println("trailer") 1491 p = trailer 1492 trailer = nil 1493 continue 1494 } 1495 //println("done") 1496 break Run 1497 } 1498 //println("lit", n, dst) 1499 nbyte := n / 8 1500 for i := uintptr(0); i < nbyte; i++ { 1501 bits |= uintptr(*p) << nbits 1502 p = add1(p) 1503 if size == 1 { 1504 *dst = uint8(bits) 1505 dst = add1(dst) 1506 bits >>= 8 1507 } else { 1508 v := bits&0xf | bitScanAll 1509 *dst = uint8(v) 1510 dst = subtract1(dst) 1511 bits >>= 4 1512 v = bits&0xf | bitScanAll 1513 *dst = uint8(v) 1514 dst = subtract1(dst) 1515 bits >>= 4 1516 } 1517 } 1518 if n %= 8; n > 0 { 1519 bits |= uintptr(*p) << nbits 1520 p = add1(p) 1521 nbits += n 1522 } 1523 continue Run 1524 } 1525 1526 // Repeat. If n == 0, it is encoded in a varint in the next bytes. 1527 if n == 0 { 1528 for off := uint(0); ; off += 7 { 1529 x := uintptr(*p) 1530 p = add1(p) 1531 n |= (x & 0x7F) << off 1532 if x&0x80 == 0 { 1533 break 1534 } 1535 } 1536 } 1537 1538 // Count is encoded in a varint in the next bytes. 1539 c := uintptr(0) 1540 for off := uint(0); ; off += 7 { 1541 x := uintptr(*p) 1542 p = add1(p) 1543 c |= (x & 0x7F) << off 1544 if x&0x80 == 0 { 1545 break 1546 } 1547 } 1548 c *= n // now total number of bits to copy 1549 1550 // If the number of bits being repeated is small, load them 1551 // into a register and use that register for the entire loop 1552 // instead of repeatedly reading from memory. 1553 // Handling fewer than 8 bits here makes the general loop simpler. 1554 // The cutoff is sys.PtrSize*8 - 7 to guarantee that when we add 1555 // the pattern to a bit buffer holding at most 7 bits (a partial byte) 1556 // it will not overflow. 1557 src := dst 1558 const maxBits = sys.PtrSize*8 - 7 1559 if n <= maxBits { 1560 // Start with bits in output buffer. 1561 pattern := bits 1562 npattern := nbits 1563 1564 // If we need more bits, fetch them from memory. 1565 if size == 1 { 1566 src = subtract1(src) 1567 for npattern < n { 1568 pattern <<= 8 1569 pattern |= uintptr(*src) 1570 src = subtract1(src) 1571 npattern += 8 1572 } 1573 } else { 1574 src = add1(src) 1575 for npattern < n { 1576 pattern <<= 4 1577 pattern |= uintptr(*src) & 0xf 1578 src = add1(src) 1579 npattern += 4 1580 } 1581 } 1582 1583 // We started with the whole bit output buffer, 1584 // and then we loaded bits from whole bytes. 1585 // Either way, we might now have too many instead of too few. 1586 // Discard the extra. 1587 if npattern > n { 1588 pattern >>= npattern - n 1589 npattern = n 1590 } 1591 1592 // Replicate pattern to at most maxBits. 1593 if npattern == 1 { 1594 // One bit being repeated. 1595 // If the bit is 1, make the pattern all 1s. 1596 // If the bit is 0, the pattern is already all 0s, 1597 // but we can claim that the number of bits 1598 // in the word is equal to the number we need (c), 1599 // because right shift of bits will zero fill. 1600 if pattern == 1 { 1601 pattern = 1<<maxBits - 1 1602 npattern = maxBits 1603 } else { 1604 npattern = c 1605 } 1606 } else { 1607 b := pattern 1608 nb := npattern 1609 if nb+nb <= maxBits { 1610 // Double pattern until the whole uintptr is filled. 1611 for nb <= sys.PtrSize*8 { 1612 b |= b << nb 1613 nb += nb 1614 } 1615 // Trim away incomplete copy of original pattern in high bits. 1616 // TODO(rsc): Replace with table lookup or loop on systems without divide? 1617 nb = maxBits / npattern * npattern 1618 b &= 1<<nb - 1 1619 pattern = b 1620 npattern = nb 1621 } 1622 } 1623 1624 // Add pattern to bit buffer and flush bit buffer, c/npattern times. 1625 // Since pattern contains >8 bits, there will be full bytes to flush 1626 // on each iteration. 1627 for ; c >= npattern; c -= npattern { 1628 bits |= pattern << nbits 1629 nbits += npattern 1630 if size == 1 { 1631 for nbits >= 8 { 1632 *dst = uint8(bits) 1633 dst = add1(dst) 1634 bits >>= 8 1635 nbits -= 8 1636 } 1637 } else { 1638 for nbits >= 4 { 1639 *dst = uint8(bits&0xf | bitScanAll) 1640 dst = subtract1(dst) 1641 bits >>= 4 1642 nbits -= 4 1643 } 1644 } 1645 } 1646 1647 // Add final fragment to bit buffer. 1648 if c > 0 { 1649 pattern &= 1<<c - 1 1650 bits |= pattern << nbits 1651 nbits += c 1652 } 1653 continue Run 1654 } 1655 1656 // Repeat; n too large to fit in a register. 1657 // Since nbits <= 7, we know the first few bytes of repeated data 1658 // are already written to memory. 1659 off := n - nbits // n > nbits because n > maxBits and nbits <= 7 1660 if size == 1 { 1661 // Leading src fragment. 1662 src = subtractb(src, (off+7)/8) 1663 if frag := off & 7; frag != 0 { 1664 bits |= uintptr(*src) >> (8 - frag) << nbits 1665 src = add1(src) 1666 nbits += frag 1667 c -= frag 1668 } 1669 // Main loop: load one byte, write another. 1670 // The bits are rotating through the bit buffer. 1671 for i := c / 8; i > 0; i-- { 1672 bits |= uintptr(*src) << nbits 1673 src = add1(src) 1674 *dst = uint8(bits) 1675 dst = add1(dst) 1676 bits >>= 8 1677 } 1678 // Final src fragment. 1679 if c %= 8; c > 0 { 1680 bits |= (uintptr(*src) & (1<<c - 1)) << nbits 1681 nbits += c 1682 } 1683 } else { 1684 // Leading src fragment. 1685 src = addb(src, (off+3)/4) 1686 if frag := off & 3; frag != 0 { 1687 bits |= (uintptr(*src) & 0xf) >> (4 - frag) << nbits 1688 src = subtract1(src) 1689 nbits += frag 1690 c -= frag 1691 } 1692 // Main loop: load one byte, write another. 1693 // The bits are rotating through the bit buffer. 1694 for i := c / 4; i > 0; i-- { 1695 bits |= (uintptr(*src) & 0xf) << nbits 1696 src = subtract1(src) 1697 *dst = uint8(bits&0xf | bitScanAll) 1698 dst = subtract1(dst) 1699 bits >>= 4 1700 } 1701 // Final src fragment. 1702 if c %= 4; c > 0 { 1703 bits |= (uintptr(*src) & (1<<c - 1)) << nbits 1704 nbits += c 1705 } 1706 } 1707 } 1708 1709 // Write any final bits out, using full-byte writes, even for the final byte. 1710 var totalBits uintptr 1711 if size == 1 { 1712 totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*8 + nbits 1713 nbits += -nbits & 7 1714 for ; nbits > 0; nbits -= 8 { 1715 *dst = uint8(bits) 1716 dst = add1(dst) 1717 bits >>= 8 1718 } 1719 } else { 1720 totalBits = (uintptr(unsafe.Pointer(dstStart))-uintptr(unsafe.Pointer(dst)))*4 + nbits 1721 nbits += -nbits & 3 1722 for ; nbits > 0; nbits -= 4 { 1723 v := bits&0xf | bitScanAll 1724 *dst = uint8(v) 1725 dst = subtract1(dst) 1726 bits >>= 4 1727 } 1728 } 1729 return totalBits 1730 } 1731 1732 func dumpGCProg(p *byte) { 1733 nptr := 0 1734 for { 1735 x := *p 1736 p = add1(p) 1737 if x == 0 { 1738 print("\t", nptr, " end\n") 1739 break 1740 } 1741 if x&0x80 == 0 { 1742 print("\t", nptr, " lit ", x, ":") 1743 n := int(x+7) / 8 1744 for i := 0; i < n; i++ { 1745 print(" ", hex(*p)) 1746 p = add1(p) 1747 } 1748 print("\n") 1749 nptr += int(x) 1750 } else { 1751 nbit := int(x &^ 0x80) 1752 if nbit == 0 { 1753 for nb := uint(0); ; nb += 7 { 1754 x := *p 1755 p = add1(p) 1756 nbit |= int(x&0x7f) << nb 1757 if x&0x80 == 0 { 1758 break 1759 } 1760 } 1761 } 1762 count := 0 1763 for nb := uint(0); ; nb += 7 { 1764 x := *p 1765 p = add1(p) 1766 count |= int(x&0x7f) << nb 1767 if x&0x80 == 0 { 1768 break 1769 } 1770 } 1771 print("\t", nptr, " repeat ", nbit, " × ", count, "\n") 1772 nptr += nbit * count 1773 } 1774 } 1775 } 1776 1777 // Testing. 1778 1779 func getgcmaskcb(frame *stkframe, ctxt unsafe.Pointer) bool { 1780 target := (*stkframe)(ctxt) 1781 if frame.sp <= target.sp && target.sp < frame.varp { 1782 *target = *frame 1783 return false 1784 } 1785 return true 1786 } 1787 1788 // gcbits returns the GC type info for x, for testing. 1789 // The result is the bitmap entries (0 or 1), one entry per byte. 1790 //go:linkname reflect_gcbits reflect.gcbits 1791 func reflect_gcbits(x interface{}) []byte { 1792 ret := getgcmask(x) 1793 typ := (*ptrtype)(unsafe.Pointer(efaceOf(&x)._type)).elem 1794 nptr := typ.ptrdata / sys.PtrSize 1795 for uintptr(len(ret)) > nptr && ret[len(ret)-1] == 0 { 1796 ret = ret[:len(ret)-1] 1797 } 1798 return ret 1799 } 1800 1801 // Returns GC type info for object p for testing. 1802 func getgcmask(ep interface{}) (mask []byte) { 1803 e := *efaceOf(&ep) 1804 p := e.data 1805 t := e._type 1806 // data or bss 1807 for _, datap := range activeModules() { 1808 // data 1809 if datap.data <= uintptr(p) && uintptr(p) < datap.edata { 1810 bitmap := datap.gcdatamask.bytedata 1811 n := (*ptrtype)(unsafe.Pointer(t)).elem.size 1812 mask = make([]byte, n/sys.PtrSize) 1813 for i := uintptr(0); i < n; i += sys.PtrSize { 1814 off := (uintptr(p) + i - datap.data) / sys.PtrSize 1815 mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1 1816 } 1817 return 1818 } 1819 1820 // bss 1821 if datap.bss <= uintptr(p) && uintptr(p) < datap.ebss { 1822 bitmap := datap.gcbssmask.bytedata 1823 n := (*ptrtype)(unsafe.Pointer(t)).elem.size 1824 mask = make([]byte, n/sys.PtrSize) 1825 for i := uintptr(0); i < n; i += sys.PtrSize { 1826 off := (uintptr(p) + i - datap.bss) / sys.PtrSize 1827 mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1 1828 } 1829 return 1830 } 1831 } 1832 1833 // heap 1834 var n uintptr 1835 var base uintptr 1836 if mlookup(uintptr(p), &base, &n, nil) != 0 { 1837 mask = make([]byte, n/sys.PtrSize) 1838 for i := uintptr(0); i < n; i += sys.PtrSize { 1839 hbits := heapBitsForAddr(base + i) 1840 if hbits.isPointer() { 1841 mask[i/sys.PtrSize] = 1 1842 } 1843 if i != 1*sys.PtrSize && !hbits.morePointers() { 1844 mask = mask[:i/sys.PtrSize] 1845 break 1846 } 1847 } 1848 return 1849 } 1850 1851 // stack 1852 if _g_ := getg(); _g_.m.curg.stack.lo <= uintptr(p) && uintptr(p) < _g_.m.curg.stack.hi { 1853 var frame stkframe 1854 frame.sp = uintptr(p) 1855 _g_ := getg() 1856 gentraceback(_g_.m.curg.sched.pc, _g_.m.curg.sched.sp, 0, _g_.m.curg, 0, nil, 1000, getgcmaskcb, noescape(unsafe.Pointer(&frame)), 0) 1857 if frame.fn.valid() { 1858 f := frame.fn 1859 targetpc := frame.continpc 1860 if targetpc == 0 { 1861 return 1862 } 1863 if targetpc != f.entry { 1864 targetpc-- 1865 } 1866 pcdata := pcdatavalue(f, _PCDATA_StackMapIndex, targetpc, nil) 1867 if pcdata == -1 { 1868 return 1869 } 1870 stkmap := (*stackmap)(funcdata(f, _FUNCDATA_LocalsPointerMaps)) 1871 if stkmap == nil || stkmap.n <= 0 { 1872 return 1873 } 1874 bv := stackmapdata(stkmap, pcdata) 1875 size := uintptr(bv.n) * sys.PtrSize 1876 n := (*ptrtype)(unsafe.Pointer(t)).elem.size 1877 mask = make([]byte, n/sys.PtrSize) 1878 for i := uintptr(0); i < n; i += sys.PtrSize { 1879 bitmap := bv.bytedata 1880 off := (uintptr(p) + i - frame.varp + size) / sys.PtrSize 1881 mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1 1882 } 1883 } 1884 return 1885 } 1886 1887 // otherwise, not something the GC knows about. 1888 // possibly read-only data, like malloc(0). 1889 // must not have pointers 1890 return 1891 }