golang.org/toolchain@v0.0.1-go1.9rc2.windows-amd64/src/runtime/malloc.go (about) 1 // Copyright 2014 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 // Memory allocator. 6 // 7 // This was originally based on tcmalloc, but has diverged quite a bit. 8 // http://goog-perftools.sourceforge.net/doc/tcmalloc.html 9 10 // The main allocator works in runs of pages. 11 // Small allocation sizes (up to and including 32 kB) are 12 // rounded to one of about 70 size classes, each of which 13 // has its own free set of objects of exactly that size. 14 // Any free page of memory can be split into a set of objects 15 // of one size class, which are then managed using a free bitmap. 16 // 17 // The allocator's data structures are: 18 // 19 // fixalloc: a free-list allocator for fixed-size off-heap objects, 20 // used to manage storage used by the allocator. 21 // mheap: the malloc heap, managed at page (8192-byte) granularity. 22 // mspan: a run of pages managed by the mheap. 23 // mcentral: collects all spans of a given size class. 24 // mcache: a per-P cache of mspans with free space. 25 // mstats: allocation statistics. 26 // 27 // Allocating a small object proceeds up a hierarchy of caches: 28 // 29 // 1. Round the size up to one of the small size classes 30 // and look in the corresponding mspan in this P's mcache. 31 // Scan the mspan's free bitmap to find a free slot. 32 // If there is a free slot, allocate it. 33 // This can all be done without acquiring a lock. 34 // 35 // 2. If the mspan has no free slots, obtain a new mspan 36 // from the mcentral's list of mspans of the required size 37 // class that have free space. 38 // Obtaining a whole span amortizes the cost of locking 39 // the mcentral. 40 // 41 // 3. If the mcentral's mspan list is empty, obtain a run 42 // of pages from the mheap to use for the mspan. 43 // 44 // 4. If the mheap is empty or has no page runs large enough, 45 // allocate a new group of pages (at least 1MB) from the 46 // operating system. Allocating a large run of pages 47 // amortizes the cost of talking to the operating system. 48 // 49 // Sweeping an mspan and freeing objects on it proceeds up a similar 50 // hierarchy: 51 // 52 // 1. If the mspan is being swept in response to allocation, it 53 // is returned to the mcache to satisfy the allocation. 54 // 55 // 2. Otherwise, if the mspan still has allocated objects in it, 56 // it is placed on the mcentral free list for the mspan's size 57 // class. 58 // 59 // 3. Otherwise, if all objects in the mspan are free, the mspan 60 // is now "idle", so it is returned to the mheap and no longer 61 // has a size class. 62 // This may coalesce it with adjacent idle mspans. 63 // 64 // 4. If an mspan remains idle for long enough, return its pages 65 // to the operating system. 66 // 67 // Allocating and freeing a large object uses the mheap 68 // directly, bypassing the mcache and mcentral. 69 // 70 // Free object slots in an mspan are zeroed only if mspan.needzero is 71 // false. If needzero is true, objects are zeroed as they are 72 // allocated. There are various benefits to delaying zeroing this way: 73 // 74 // 1. Stack frame allocation can avoid zeroing altogether. 75 // 76 // 2. It exhibits better temporal locality, since the program is 77 // probably about to write to the memory. 78 // 79 // 3. We don't zero pages that never get reused. 80 81 package runtime 82 83 import ( 84 "runtime/internal/sys" 85 "unsafe" 86 ) 87 88 const ( 89 debugMalloc = false 90 91 maxTinySize = _TinySize 92 tinySizeClass = _TinySizeClass 93 maxSmallSize = _MaxSmallSize 94 95 pageShift = _PageShift 96 pageSize = _PageSize 97 pageMask = _PageMask 98 // By construction, single page spans of the smallest object class 99 // have the most objects per span. 100 maxObjsPerSpan = pageSize / 8 101 102 mSpanInUse = _MSpanInUse 103 104 concurrentSweep = _ConcurrentSweep 105 106 _PageSize = 1 << _PageShift 107 _PageMask = _PageSize - 1 108 109 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems 110 _64bit = 1 << (^uintptr(0) >> 63) / 2 111 112 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go. 113 _TinySize = 16 114 _TinySizeClass = int8(2) 115 116 _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc 117 _MaxMHeapList = 1 << (20 - _PageShift) // Maximum page length for fixed-size list in MHeap. 118 _HeapAllocChunk = 1 << 20 // Chunk size for heap growth 119 120 // Per-P, per order stack segment cache size. 121 _StackCacheSize = 32 * 1024 122 123 // Number of orders that get caching. Order 0 is FixedStack 124 // and each successive order is twice as large. 125 // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks 126 // will be allocated directly. 127 // Since FixedStack is different on different systems, we 128 // must vary NumStackOrders to keep the same maximum cached size. 129 // OS | FixedStack | NumStackOrders 130 // -----------------+------------+--------------- 131 // linux/darwin/bsd | 2KB | 4 132 // windows/32 | 4KB | 3 133 // windows/64 | 8KB | 2 134 // plan9 | 4KB | 3 135 _NumStackOrders = 4 - sys.PtrSize/4*sys.GoosWindows - 1*sys.GoosPlan9 136 137 // Number of bits in page to span calculations (4k pages). 138 // On Windows 64-bit we limit the arena to 32GB or 35 bits. 139 // Windows counts memory used by page table into committed memory 140 // of the process, so we can't reserve too much memory. 141 // See https://golang.org/issue/5402 and https://golang.org/issue/5236. 142 // On other 64-bit platforms, we limit the arena to 512GB, or 39 bits. 143 // On 32-bit, we don't bother limiting anything, so we use the full 32-bit address. 144 // The only exception is mips32 which only has access to low 2GB of virtual memory. 145 // On Darwin/arm64, we cannot reserve more than ~5GB of virtual memory, 146 // but as most devices have less than 4GB of physical memory anyway, we 147 // try to be conservative here, and only ask for a 2GB heap. 148 _MHeapMap_TotalBits = (_64bit*sys.GoosWindows)*35 + (_64bit*(1-sys.GoosWindows)*(1-sys.GoosDarwin*sys.GoarchArm64))*39 + sys.GoosDarwin*sys.GoarchArm64*31 + (1-_64bit)*(32-(sys.GoarchMips+sys.GoarchMipsle)) 149 _MHeapMap_Bits = _MHeapMap_TotalBits - _PageShift 150 151 // _MaxMem is the maximum heap arena size minus 1. 152 // 153 // On 32-bit, this is also the maximum heap pointer value, 154 // since the arena starts at address 0. 155 _MaxMem = 1<<_MHeapMap_TotalBits - 1 156 157 // Max number of threads to run garbage collection. 158 // 2, 3, and 4 are all plausible maximums depending 159 // on the hardware details of the machine. The garbage 160 // collector scales well to 32 cpus. 161 _MaxGcproc = 32 162 163 // minLegalPointer is the smallest possible legal pointer. 164 // This is the smallest possible architectural page size, 165 // since we assume that the first page is never mapped. 166 // 167 // This should agree with minZeroPage in the compiler. 168 minLegalPointer uintptr = 4096 169 ) 170 171 // physPageSize is the size in bytes of the OS's physical pages. 172 // Mapping and unmapping operations must be done at multiples of 173 // physPageSize. 174 // 175 // This must be set by the OS init code (typically in osinit) before 176 // mallocinit. 177 var physPageSize uintptr 178 179 // OS-defined helpers: 180 // 181 // sysAlloc obtains a large chunk of zeroed memory from the 182 // operating system, typically on the order of a hundred kilobytes 183 // or a megabyte. 184 // NOTE: sysAlloc returns OS-aligned memory, but the heap allocator 185 // may use larger alignment, so the caller must be careful to realign the 186 // memory obtained by sysAlloc. 187 // 188 // SysUnused notifies the operating system that the contents 189 // of the memory region are no longer needed and can be reused 190 // for other purposes. 191 // SysUsed notifies the operating system that the contents 192 // of the memory region are needed again. 193 // 194 // SysFree returns it unconditionally; this is only used if 195 // an out-of-memory error has been detected midway through 196 // an allocation. It is okay if SysFree is a no-op. 197 // 198 // SysReserve reserves address space without allocating memory. 199 // If the pointer passed to it is non-nil, the caller wants the 200 // reservation there, but SysReserve can still choose another 201 // location if that one is unavailable. On some systems and in some 202 // cases SysReserve will simply check that the address space is 203 // available and not actually reserve it. If SysReserve returns 204 // non-nil, it sets *reserved to true if the address space is 205 // reserved, false if it has merely been checked. 206 // NOTE: SysReserve returns OS-aligned memory, but the heap allocator 207 // may use larger alignment, so the caller must be careful to realign the 208 // memory obtained by sysAlloc. 209 // 210 // SysMap maps previously reserved address space for use. 211 // The reserved argument is true if the address space was really 212 // reserved, not merely checked. 213 // 214 // SysFault marks a (already sysAlloc'd) region to fault 215 // if accessed. Used only for debugging the runtime. 216 217 func mallocinit() { 218 if class_to_size[_TinySizeClass] != _TinySize { 219 throw("bad TinySizeClass") 220 } 221 222 testdefersizes() 223 224 // Copy class sizes out for statistics table. 225 for i := range class_to_size { 226 memstats.by_size[i].size = uint32(class_to_size[i]) 227 } 228 229 // Check physPageSize. 230 if physPageSize == 0 { 231 // The OS init code failed to fetch the physical page size. 232 throw("failed to get system page size") 233 } 234 if physPageSize < minPhysPageSize { 235 print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n") 236 throw("bad system page size") 237 } 238 if physPageSize&(physPageSize-1) != 0 { 239 print("system page size (", physPageSize, ") must be a power of 2\n") 240 throw("bad system page size") 241 } 242 243 // The auxiliary regions start at p and are laid out in the 244 // following order: spans, bitmap, arena. 245 var p, pSize uintptr 246 var reserved bool 247 248 // The spans array holds one *mspan per _PageSize of arena. 249 var spansSize uintptr = (_MaxMem + 1) / _PageSize * sys.PtrSize 250 spansSize = round(spansSize, _PageSize) 251 // The bitmap holds 2 bits per word of arena. 252 var bitmapSize uintptr = (_MaxMem + 1) / (sys.PtrSize * 8 / 2) 253 bitmapSize = round(bitmapSize, _PageSize) 254 255 // Set up the allocation arena, a contiguous area of memory where 256 // allocated data will be found. 257 if sys.PtrSize == 8 { 258 // On a 64-bit machine, allocate from a single contiguous reservation. 259 // 512 GB (MaxMem) should be big enough for now. 260 // 261 // The code will work with the reservation at any address, but ask 262 // SysReserve to use 0x0000XXc000000000 if possible (XX=00...7f). 263 // Allocating a 512 GB region takes away 39 bits, and the amd64 264 // doesn't let us choose the top 17 bits, so that leaves the 9 bits 265 // in the middle of 0x00c0 for us to choose. Choosing 0x00c0 means 266 // that the valid memory addresses will begin 0x00c0, 0x00c1, ..., 0x00df. 267 // In little-endian, that's c0 00, c1 00, ..., df 00. None of those are valid 268 // UTF-8 sequences, and they are otherwise as far away from 269 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0 270 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors 271 // on OS X during thread allocations. 0x00c0 causes conflicts with 272 // AddressSanitizer which reserves all memory up to 0x0100. 273 // These choices are both for debuggability and to reduce the 274 // odds of a conservative garbage collector (as is still used in gccgo) 275 // not collecting memory because some non-pointer block of memory 276 // had a bit pattern that matched a memory address. 277 // 278 // Actually we reserve 544 GB (because the bitmap ends up being 32 GB) 279 // but it hardly matters: e0 00 is not valid UTF-8 either. 280 // 281 // If this fails we fall back to the 32 bit memory mechanism 282 // 283 // However, on arm64, we ignore all this advice above and slam the 284 // allocation at 0x40 << 32 because when using 4k pages with 3-level 285 // translation buffers, the user address space is limited to 39 bits 286 // On darwin/arm64, the address space is even smaller. 287 arenaSize := round(_MaxMem, _PageSize) 288 pSize = bitmapSize + spansSize + arenaSize + _PageSize 289 for i := 0; i <= 0x7f; i++ { 290 switch { 291 case GOARCH == "arm64" && GOOS == "darwin": 292 p = uintptr(i)<<40 | uintptrMask&(0x0013<<28) 293 case GOARCH == "arm64": 294 p = uintptr(i)<<40 | uintptrMask&(0x0040<<32) 295 default: 296 p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32) 297 } 298 p = uintptr(sysReserve(unsafe.Pointer(p), pSize, &reserved)) 299 if p != 0 { 300 break 301 } 302 } 303 } 304 305 if p == 0 { 306 // On a 32-bit machine, we can't typically get away 307 // with a giant virtual address space reservation. 308 // Instead we map the memory information bitmap 309 // immediately after the data segment, large enough 310 // to handle the entire 4GB address space (256 MB), 311 // along with a reservation for an initial arena. 312 // When that gets used up, we'll start asking the kernel 313 // for any memory anywhere. 314 315 // We want to start the arena low, but if we're linked 316 // against C code, it's possible global constructors 317 // have called malloc and adjusted the process' brk. 318 // Query the brk so we can avoid trying to map the 319 // arena over it (which will cause the kernel to put 320 // the arena somewhere else, likely at a high 321 // address). 322 procBrk := sbrk0() 323 324 // If we fail to allocate, try again with a smaller arena. 325 // This is necessary on Android L where we share a process 326 // with ART, which reserves virtual memory aggressively. 327 // In the worst case, fall back to a 0-sized initial arena, 328 // in the hope that subsequent reservations will succeed. 329 arenaSizes := []uintptr{ 330 512 << 20, 331 256 << 20, 332 128 << 20, 333 0, 334 } 335 336 for _, arenaSize := range arenaSizes { 337 // SysReserve treats the address we ask for, end, as a hint, 338 // not as an absolute requirement. If we ask for the end 339 // of the data segment but the operating system requires 340 // a little more space before we can start allocating, it will 341 // give out a slightly higher pointer. Except QEMU, which 342 // is buggy, as usual: it won't adjust the pointer upward. 343 // So adjust it upward a little bit ourselves: 1/4 MB to get 344 // away from the running binary image and then round up 345 // to a MB boundary. 346 p = round(firstmoduledata.end+(1<<18), 1<<20) 347 pSize = bitmapSize + spansSize + arenaSize + _PageSize 348 if p <= procBrk && procBrk < p+pSize { 349 // Move the start above the brk, 350 // leaving some room for future brk 351 // expansion. 352 p = round(procBrk+(1<<20), 1<<20) 353 } 354 p = uintptr(sysReserve(unsafe.Pointer(p), pSize, &reserved)) 355 if p != 0 { 356 break 357 } 358 } 359 if p == 0 { 360 throw("runtime: cannot reserve arena virtual address space") 361 } 362 } 363 364 // PageSize can be larger than OS definition of page size, 365 // so SysReserve can give us a PageSize-unaligned pointer. 366 // To overcome this we ask for PageSize more and round up the pointer. 367 p1 := round(p, _PageSize) 368 pSize -= p1 - p 369 370 spansStart := p1 371 p1 += spansSize 372 mheap_.bitmap = p1 + bitmapSize 373 p1 += bitmapSize 374 if sys.PtrSize == 4 { 375 // Set arena_start such that we can accept memory 376 // reservations located anywhere in the 4GB virtual space. 377 mheap_.arena_start = 0 378 } else { 379 mheap_.arena_start = p1 380 } 381 mheap_.arena_end = p + pSize 382 mheap_.arena_used = p1 383 mheap_.arena_alloc = p1 384 mheap_.arena_reserved = reserved 385 386 if mheap_.arena_start&(_PageSize-1) != 0 { 387 println("bad pagesize", hex(p), hex(p1), hex(spansSize), hex(bitmapSize), hex(_PageSize), "start", hex(mheap_.arena_start)) 388 throw("misrounded allocation in mallocinit") 389 } 390 391 // Initialize the rest of the allocator. 392 mheap_.init(spansStart, spansSize) 393 _g_ := getg() 394 _g_.m.mcache = allocmcache() 395 } 396 397 // sysAlloc allocates the next n bytes from the heap arena. The 398 // returned pointer is always _PageSize aligned and between 399 // h.arena_start and h.arena_end. sysAlloc returns nil on failure. 400 // There is no corresponding free function. 401 func (h *mheap) sysAlloc(n uintptr) unsafe.Pointer { 402 // strandLimit is the maximum number of bytes to strand from 403 // the current arena block. If we would need to strand more 404 // than this, we fall back to sysAlloc'ing just enough for 405 // this allocation. 406 const strandLimit = 16 << 20 407 408 if n > h.arena_end-h.arena_alloc { 409 // If we haven't grown the arena to _MaxMem yet, try 410 // to reserve some more address space. 411 p_size := round(n+_PageSize, 256<<20) 412 new_end := h.arena_end + p_size // Careful: can overflow 413 if h.arena_end <= new_end && new_end-h.arena_start-1 <= _MaxMem { 414 // TODO: It would be bad if part of the arena 415 // is reserved and part is not. 416 var reserved bool 417 p := uintptr(sysReserve(unsafe.Pointer(h.arena_end), p_size, &reserved)) 418 if p == 0 { 419 // TODO: Try smaller reservation 420 // growths in case we're in a crowded 421 // 32-bit address space. 422 goto reservationFailed 423 } 424 // p can be just about anywhere in the address 425 // space, including before arena_end. 426 if p == h.arena_end { 427 // The new block is contiguous with 428 // the current block. Extend the 429 // current arena block. 430 h.arena_end = new_end 431 h.arena_reserved = reserved 432 } else if h.arena_start <= p && p+p_size-h.arena_start-1 <= _MaxMem && h.arena_end-h.arena_alloc < strandLimit { 433 // We were able to reserve more memory 434 // within the arena space, but it's 435 // not contiguous with our previous 436 // reservation. It could be before or 437 // after our current arena_used. 438 // 439 // Keep everything page-aligned. 440 // Our pages are bigger than hardware pages. 441 h.arena_end = p + p_size 442 p = round(p, _PageSize) 443 h.arena_alloc = p 444 h.arena_reserved = reserved 445 } else { 446 // We got a mapping, but either 447 // 448 // 1) It's not in the arena, so we 449 // can't use it. (This should never 450 // happen on 32-bit.) 451 // 452 // 2) We would need to discard too 453 // much of our current arena block to 454 // use it. 455 // 456 // We haven't added this allocation to 457 // the stats, so subtract it from a 458 // fake stat (but avoid underflow). 459 // 460 // We'll fall back to a small sysAlloc. 461 stat := uint64(p_size) 462 sysFree(unsafe.Pointer(p), p_size, &stat) 463 } 464 } 465 } 466 467 if n <= h.arena_end-h.arena_alloc { 468 // Keep taking from our reservation. 469 p := h.arena_alloc 470 sysMap(unsafe.Pointer(p), n, h.arena_reserved, &memstats.heap_sys) 471 h.arena_alloc += n 472 if h.arena_alloc > h.arena_used { 473 h.setArenaUsed(h.arena_alloc, true) 474 } 475 476 if p&(_PageSize-1) != 0 { 477 throw("misrounded allocation in MHeap_SysAlloc") 478 } 479 return unsafe.Pointer(p) 480 } 481 482 reservationFailed: 483 // If using 64-bit, our reservation is all we have. 484 if sys.PtrSize != 4 { 485 return nil 486 } 487 488 // On 32-bit, once the reservation is gone we can 489 // try to get memory at a location chosen by the OS. 490 p_size := round(n, _PageSize) + _PageSize 491 p := uintptr(sysAlloc(p_size, &memstats.heap_sys)) 492 if p == 0 { 493 return nil 494 } 495 496 if p < h.arena_start || p+p_size-h.arena_start > _MaxMem { 497 // This shouldn't be possible because _MaxMem is the 498 // whole address space on 32-bit. 499 top := uint64(h.arena_start) + _MaxMem 500 print("runtime: memory allocated by OS (", hex(p), ") not in usable range [", hex(h.arena_start), ",", hex(top), ")\n") 501 sysFree(unsafe.Pointer(p), p_size, &memstats.heap_sys) 502 return nil 503 } 504 505 p += -p & (_PageSize - 1) 506 if p+n > h.arena_used { 507 h.setArenaUsed(p+n, true) 508 } 509 510 if p&(_PageSize-1) != 0 { 511 throw("misrounded allocation in MHeap_SysAlloc") 512 } 513 return unsafe.Pointer(p) 514 } 515 516 // base address for all 0-byte allocations 517 var zerobase uintptr 518 519 // nextFreeFast returns the next free object if one is quickly available. 520 // Otherwise it returns 0. 521 func nextFreeFast(s *mspan) gclinkptr { 522 theBit := sys.Ctz64(s.allocCache) // Is there a free object in the allocCache? 523 if theBit < 64 { 524 result := s.freeindex + uintptr(theBit) 525 if result < s.nelems { 526 freeidx := result + 1 527 if freeidx%64 == 0 && freeidx != s.nelems { 528 return 0 529 } 530 s.allocCache >>= uint(theBit + 1) 531 s.freeindex = freeidx 532 v := gclinkptr(result*s.elemsize + s.base()) 533 s.allocCount++ 534 return v 535 } 536 } 537 return 0 538 } 539 540 // nextFree returns the next free object from the cached span if one is available. 541 // Otherwise it refills the cache with a span with an available object and 542 // returns that object along with a flag indicating that this was a heavy 543 // weight allocation. If it is a heavy weight allocation the caller must 544 // determine whether a new GC cycle needs to be started or if the GC is active 545 // whether this goroutine needs to assist the GC. 546 func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) { 547 s = c.alloc[spc] 548 shouldhelpgc = false 549 freeIndex := s.nextFreeIndex() 550 if freeIndex == s.nelems { 551 // The span is full. 552 if uintptr(s.allocCount) != s.nelems { 553 println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 554 throw("s.allocCount != s.nelems && freeIndex == s.nelems") 555 } 556 systemstack(func() { 557 c.refill(spc) 558 }) 559 shouldhelpgc = true 560 s = c.alloc[spc] 561 562 freeIndex = s.nextFreeIndex() 563 } 564 565 if freeIndex >= s.nelems { 566 throw("freeIndex is not valid") 567 } 568 569 v = gclinkptr(freeIndex*s.elemsize + s.base()) 570 s.allocCount++ 571 if uintptr(s.allocCount) > s.nelems { 572 println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 573 throw("s.allocCount > s.nelems") 574 } 575 return 576 } 577 578 // Allocate an object of size bytes. 579 // Small objects are allocated from the per-P cache's free lists. 580 // Large objects (> 32 kB) are allocated straight from the heap. 581 func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer { 582 if gcphase == _GCmarktermination { 583 throw("mallocgc called with gcphase == _GCmarktermination") 584 } 585 586 if size == 0 { 587 return unsafe.Pointer(&zerobase) 588 } 589 590 if debug.sbrk != 0 { 591 align := uintptr(16) 592 if typ != nil { 593 align = uintptr(typ.align) 594 } 595 return persistentalloc(size, align, &memstats.other_sys) 596 } 597 598 // assistG is the G to charge for this allocation, or nil if 599 // GC is not currently active. 600 var assistG *g 601 if gcBlackenEnabled != 0 { 602 // Charge the current user G for this allocation. 603 assistG = getg() 604 if assistG.m.curg != nil { 605 assistG = assistG.m.curg 606 } 607 // Charge the allocation against the G. We'll account 608 // for internal fragmentation at the end of mallocgc. 609 assistG.gcAssistBytes -= int64(size) 610 611 if assistG.gcAssistBytes < 0 { 612 // This G is in debt. Assist the GC to correct 613 // this before allocating. This must happen 614 // before disabling preemption. 615 gcAssistAlloc(assistG) 616 } 617 } 618 619 // Set mp.mallocing to keep from being preempted by GC. 620 mp := acquirem() 621 if mp.mallocing != 0 { 622 throw("malloc deadlock") 623 } 624 if mp.gsignal == getg() { 625 throw("malloc during signal") 626 } 627 mp.mallocing = 1 628 629 shouldhelpgc := false 630 dataSize := size 631 c := gomcache() 632 var x unsafe.Pointer 633 noscan := typ == nil || typ.kind&kindNoPointers != 0 634 if size <= maxSmallSize { 635 if noscan && size < maxTinySize { 636 // Tiny allocator. 637 // 638 // Tiny allocator combines several tiny allocation requests 639 // into a single memory block. The resulting memory block 640 // is freed when all subobjects are unreachable. The subobjects 641 // must be noscan (don't have pointers), this ensures that 642 // the amount of potentially wasted memory is bounded. 643 // 644 // Size of the memory block used for combining (maxTinySize) is tunable. 645 // Current setting is 16 bytes, which relates to 2x worst case memory 646 // wastage (when all but one subobjects are unreachable). 647 // 8 bytes would result in no wastage at all, but provides less 648 // opportunities for combining. 649 // 32 bytes provides more opportunities for combining, 650 // but can lead to 4x worst case wastage. 651 // The best case winning is 8x regardless of block size. 652 // 653 // Objects obtained from tiny allocator must not be freed explicitly. 654 // So when an object will be freed explicitly, we ensure that 655 // its size >= maxTinySize. 656 // 657 // SetFinalizer has a special case for objects potentially coming 658 // from tiny allocator, it such case it allows to set finalizers 659 // for an inner byte of a memory block. 660 // 661 // The main targets of tiny allocator are small strings and 662 // standalone escaping variables. On a json benchmark 663 // the allocator reduces number of allocations by ~12% and 664 // reduces heap size by ~20%. 665 off := c.tinyoffset 666 // Align tiny pointer for required (conservative) alignment. 667 if size&7 == 0 { 668 off = round(off, 8) 669 } else if size&3 == 0 { 670 off = round(off, 4) 671 } else if size&1 == 0 { 672 off = round(off, 2) 673 } 674 if off+size <= maxTinySize && c.tiny != 0 { 675 // The object fits into existing tiny block. 676 x = unsafe.Pointer(c.tiny + off) 677 c.tinyoffset = off + size 678 c.local_tinyallocs++ 679 mp.mallocing = 0 680 releasem(mp) 681 return x 682 } 683 // Allocate a new maxTinySize block. 684 span := c.alloc[tinySpanClass] 685 v := nextFreeFast(span) 686 if v == 0 { 687 v, _, shouldhelpgc = c.nextFree(tinySpanClass) 688 } 689 x = unsafe.Pointer(v) 690 (*[2]uint64)(x)[0] = 0 691 (*[2]uint64)(x)[1] = 0 692 // See if we need to replace the existing tiny block with the new one 693 // based on amount of remaining free space. 694 if size < c.tinyoffset || c.tiny == 0 { 695 c.tiny = uintptr(x) 696 c.tinyoffset = size 697 } 698 size = maxTinySize 699 } else { 700 var sizeclass uint8 701 if size <= smallSizeMax-8 { 702 sizeclass = size_to_class8[(size+smallSizeDiv-1)/smallSizeDiv] 703 } else { 704 sizeclass = size_to_class128[(size-smallSizeMax+largeSizeDiv-1)/largeSizeDiv] 705 } 706 size = uintptr(class_to_size[sizeclass]) 707 spc := makeSpanClass(sizeclass, noscan) 708 span := c.alloc[spc] 709 v := nextFreeFast(span) 710 if v == 0 { 711 v, span, shouldhelpgc = c.nextFree(spc) 712 } 713 x = unsafe.Pointer(v) 714 if needzero && span.needzero != 0 { 715 memclrNoHeapPointers(unsafe.Pointer(v), size) 716 } 717 } 718 } else { 719 var s *mspan 720 shouldhelpgc = true 721 systemstack(func() { 722 s = largeAlloc(size, needzero, noscan) 723 }) 724 s.freeindex = 1 725 s.allocCount = 1 726 x = unsafe.Pointer(s.base()) 727 size = s.elemsize 728 } 729 730 var scanSize uintptr 731 if !noscan { 732 // If allocating a defer+arg block, now that we've picked a malloc size 733 // large enough to hold everything, cut the "asked for" size down to 734 // just the defer header, so that the GC bitmap will record the arg block 735 // as containing nothing at all (as if it were unused space at the end of 736 // a malloc block caused by size rounding). 737 // The defer arg areas are scanned as part of scanstack. 738 if typ == deferType { 739 dataSize = unsafe.Sizeof(_defer{}) 740 } 741 heapBitsSetType(uintptr(x), size, dataSize, typ) 742 if dataSize > typ.size { 743 // Array allocation. If there are any 744 // pointers, GC has to scan to the last 745 // element. 746 if typ.ptrdata != 0 { 747 scanSize = dataSize - typ.size + typ.ptrdata 748 } 749 } else { 750 scanSize = typ.ptrdata 751 } 752 c.local_scan += scanSize 753 } 754 755 // Ensure that the stores above that initialize x to 756 // type-safe memory and set the heap bits occur before 757 // the caller can make x observable to the garbage 758 // collector. Otherwise, on weakly ordered machines, 759 // the garbage collector could follow a pointer to x, 760 // but see uninitialized memory or stale heap bits. 761 publicationBarrier() 762 763 // Allocate black during GC. 764 // All slots hold nil so no scanning is needed. 765 // This may be racing with GC so do it atomically if there can be 766 // a race marking the bit. 767 if gcphase != _GCoff { 768 gcmarknewobject(uintptr(x), size, scanSize) 769 } 770 771 if raceenabled { 772 racemalloc(x, size) 773 } 774 775 if msanenabled { 776 msanmalloc(x, size) 777 } 778 779 mp.mallocing = 0 780 releasem(mp) 781 782 if debug.allocfreetrace != 0 { 783 tracealloc(x, size, typ) 784 } 785 786 if rate := MemProfileRate; rate > 0 { 787 if size < uintptr(rate) && int32(size) < c.next_sample { 788 c.next_sample -= int32(size) 789 } else { 790 mp := acquirem() 791 profilealloc(mp, x, size) 792 releasem(mp) 793 } 794 } 795 796 if assistG != nil { 797 // Account for internal fragmentation in the assist 798 // debt now that we know it. 799 assistG.gcAssistBytes -= int64(size - dataSize) 800 } 801 802 if shouldhelpgc { 803 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 804 gcStart(gcBackgroundMode, t) 805 } 806 } 807 808 return x 809 } 810 811 func largeAlloc(size uintptr, needzero bool, noscan bool) *mspan { 812 // print("largeAlloc size=", size, "\n") 813 814 if size+_PageSize < size { 815 throw("out of memory") 816 } 817 npages := size >> _PageShift 818 if size&_PageMask != 0 { 819 npages++ 820 } 821 822 // Deduct credit for this span allocation and sweep if 823 // necessary. mHeap_Alloc will also sweep npages, so this only 824 // pays the debt down to npage pages. 825 deductSweepCredit(npages*_PageSize, npages) 826 827 s := mheap_.alloc(npages, makeSpanClass(0, noscan), true, needzero) 828 if s == nil { 829 throw("out of memory") 830 } 831 s.limit = s.base() + size 832 heapBitsForSpan(s.base()).initSpan(s) 833 return s 834 } 835 836 // implementation of new builtin 837 // compiler (both frontend and SSA backend) knows the signature 838 // of this function 839 func newobject(typ *_type) unsafe.Pointer { 840 return mallocgc(typ.size, typ, true) 841 } 842 843 //go:linkname reflect_unsafe_New reflect.unsafe_New 844 func reflect_unsafe_New(typ *_type) unsafe.Pointer { 845 return newobject(typ) 846 } 847 848 // newarray allocates an array of n elements of type typ. 849 func newarray(typ *_type, n int) unsafe.Pointer { 850 if n < 0 || uintptr(n) > maxSliceCap(typ.size) { 851 panic(plainError("runtime: allocation size out of range")) 852 } 853 return mallocgc(typ.size*uintptr(n), typ, true) 854 } 855 856 //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray 857 func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer { 858 return newarray(typ, n) 859 } 860 861 func profilealloc(mp *m, x unsafe.Pointer, size uintptr) { 862 mp.mcache.next_sample = nextSample() 863 mProf_Malloc(x, size) 864 } 865 866 // nextSample returns the next sampling point for heap profiling. 867 // It produces a random variable with a geometric distribution and 868 // mean MemProfileRate. This is done by generating a uniformly 869 // distributed random number and applying the cumulative distribution 870 // function for an exponential. 871 func nextSample() int32 { 872 if GOOS == "plan9" { 873 // Plan 9 doesn't support floating point in note handler. 874 if g := getg(); g == g.m.gsignal { 875 return nextSampleNoFP() 876 } 877 } 878 879 period := MemProfileRate 880 881 // make nextSample not overflow. Maximum possible step is 882 // -ln(1/(1<<kRandomBitCount)) * period, approximately 20 * period. 883 switch { 884 case period > 0x7000000: 885 period = 0x7000000 886 case period == 0: 887 return 0 888 } 889 890 // Let m be the sample rate, 891 // the probability distribution function is m*exp(-mx), so the CDF is 892 // p = 1 - exp(-mx), so 893 // q = 1 - p == exp(-mx) 894 // log_e(q) = -mx 895 // -log_e(q)/m = x 896 // x = -log_e(q) * period 897 // x = log_2(q) * (-log_e(2)) * period ; Using log_2 for efficiency 898 const randomBitCount = 26 899 q := fastrand()%(1<<randomBitCount) + 1 900 qlog := fastlog2(float64(q)) - randomBitCount 901 if qlog > 0 { 902 qlog = 0 903 } 904 const minusLog2 = -0.6931471805599453 // -ln(2) 905 return int32(qlog*(minusLog2*float64(period))) + 1 906 } 907 908 // nextSampleNoFP is similar to nextSample, but uses older, 909 // simpler code to avoid floating point. 910 func nextSampleNoFP() int32 { 911 // Set first allocation sample size. 912 rate := MemProfileRate 913 if rate > 0x3fffffff { // make 2*rate not overflow 914 rate = 0x3fffffff 915 } 916 if rate != 0 { 917 return int32(fastrand() % uint32(2*rate)) 918 } 919 return 0 920 } 921 922 type persistentAlloc struct { 923 base unsafe.Pointer 924 off uintptr 925 } 926 927 var globalAlloc struct { 928 mutex 929 persistentAlloc 930 } 931 932 // Wrapper around sysAlloc that can allocate small chunks. 933 // There is no associated free operation. 934 // Intended for things like function/type/debug-related persistent data. 935 // If align is 0, uses default align (currently 8). 936 // The returned memory will be zeroed. 937 // 938 // Consider marking persistentalloc'd types go:notinheap. 939 func persistentalloc(size, align uintptr, sysStat *uint64) unsafe.Pointer { 940 var p unsafe.Pointer 941 systemstack(func() { 942 p = persistentalloc1(size, align, sysStat) 943 }) 944 return p 945 } 946 947 // Must run on system stack because stack growth can (re)invoke it. 948 // See issue 9174. 949 //go:systemstack 950 func persistentalloc1(size, align uintptr, sysStat *uint64) unsafe.Pointer { 951 const ( 952 chunk = 256 << 10 953 maxBlock = 64 << 10 // VM reservation granularity is 64K on windows 954 ) 955 956 if size == 0 { 957 throw("persistentalloc: size == 0") 958 } 959 if align != 0 { 960 if align&(align-1) != 0 { 961 throw("persistentalloc: align is not a power of 2") 962 } 963 if align > _PageSize { 964 throw("persistentalloc: align is too large") 965 } 966 } else { 967 align = 8 968 } 969 970 if size >= maxBlock { 971 return sysAlloc(size, sysStat) 972 } 973 974 mp := acquirem() 975 var persistent *persistentAlloc 976 if mp != nil && mp.p != 0 { 977 persistent = &mp.p.ptr().palloc 978 } else { 979 lock(&globalAlloc.mutex) 980 persistent = &globalAlloc.persistentAlloc 981 } 982 persistent.off = round(persistent.off, align) 983 if persistent.off+size > chunk || persistent.base == nil { 984 persistent.base = sysAlloc(chunk, &memstats.other_sys) 985 if persistent.base == nil { 986 if persistent == &globalAlloc.persistentAlloc { 987 unlock(&globalAlloc.mutex) 988 } 989 throw("runtime: cannot allocate memory") 990 } 991 persistent.off = 0 992 } 993 p := add(persistent.base, persistent.off) 994 persistent.off += size 995 releasem(mp) 996 if persistent == &globalAlloc.persistentAlloc { 997 unlock(&globalAlloc.mutex) 998 } 999 1000 if sysStat != &memstats.other_sys { 1001 mSysStatInc(sysStat, size) 1002 mSysStatDec(&memstats.other_sys, size) 1003 } 1004 return p 1005 }