Source file src/runtime/malloc.go
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 in-use 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's 60 // pages are returned to the mheap and the mspan is now dead. 61 // 62 // Allocating and freeing a large object uses the mheap 63 // directly, bypassing the mcache and mcentral. 64 // 65 // If mspan.needzero is false, then free object slots in the mspan are 66 // already zeroed. Otherwise if needzero is true, objects are zeroed as 67 // they are allocated. There are various benefits to delaying zeroing 68 // this way: 69 // 70 // 1. Stack frame allocation can avoid zeroing altogether. 71 // 72 // 2. It exhibits better temporal locality, since the program is 73 // probably about to write to the memory. 74 // 75 // 3. We don't zero pages that never get reused. 76 77 // Virtual memory layout 78 // 79 // The heap consists of a set of arenas, which are 64MB on 64-bit and 80 // 4MB on 32-bit (heapArenaBytes). Each arena's start address is also 81 // aligned to the arena size. 82 // 83 // Each arena has an associated heapArena object that stores the 84 // metadata for that arena: the heap bitmap for all words in the arena 85 // and the span map for all pages in the arena. heapArena objects are 86 // themselves allocated off-heap. 87 // 88 // Since arenas are aligned, the address space can be viewed as a 89 // series of arena frames. The arena map (mheap_.arenas) maps from 90 // arena frame number to *heapArena, or nil for parts of the address 91 // space not backed by the Go heap. The arena map is structured as a 92 // two-level array consisting of a "L1" arena map and many "L2" arena 93 // maps; however, since arenas are large, on many architectures, the 94 // arena map consists of a single, large L2 map. 95 // 96 // The arena map covers the entire possible address space, allowing 97 // the Go heap to use any part of the address space. The allocator 98 // attempts to keep arenas contiguous so that large spans (and hence 99 // large objects) can cross arenas. 100 101 package runtime 102 103 import ( 104 "internal/goarch" 105 "internal/goos" 106 "internal/runtime/atomic" 107 "runtime/internal/math" 108 "runtime/internal/sys" 109 "unsafe" 110 ) 111 112 const ( 113 maxTinySize = _TinySize 114 tinySizeClass = _TinySizeClass 115 maxSmallSize = _MaxSmallSize 116 117 pageShift = _PageShift 118 pageSize = _PageSize 119 120 _PageSize = 1 << _PageShift 121 _PageMask = _PageSize - 1 122 123 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems 124 _64bit = 1 << (^uintptr(0) >> 63) / 2 125 126 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go. 127 _TinySize = 16 128 _TinySizeClass = int8(2) 129 130 _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc 131 132 // Per-P, per order stack segment cache size. 133 _StackCacheSize = 32 * 1024 134 135 // Number of orders that get caching. Order 0 is FixedStack 136 // and each successive order is twice as large. 137 // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks 138 // will be allocated directly. 139 // Since FixedStack is different on different systems, we 140 // must vary NumStackOrders to keep the same maximum cached size. 141 // OS | FixedStack | NumStackOrders 142 // -----------------+------------+--------------- 143 // linux/darwin/bsd | 2KB | 4 144 // windows/32 | 4KB | 3 145 // windows/64 | 8KB | 2 146 // plan9 | 4KB | 3 147 _NumStackOrders = 4 - goarch.PtrSize/4*goos.IsWindows - 1*goos.IsPlan9 148 149 // heapAddrBits is the number of bits in a heap address. On 150 // amd64, addresses are sign-extended beyond heapAddrBits. On 151 // other arches, they are zero-extended. 152 // 153 // On most 64-bit platforms, we limit this to 48 bits based on a 154 // combination of hardware and OS limitations. 155 // 156 // amd64 hardware limits addresses to 48 bits, sign-extended 157 // to 64 bits. Addresses where the top 16 bits are not either 158 // all 0 or all 1 are "non-canonical" and invalid. Because of 159 // these "negative" addresses, we offset addresses by 1<<47 160 // (arenaBaseOffset) on amd64 before computing indexes into 161 // the heap arenas index. In 2017, amd64 hardware added 162 // support for 57 bit addresses; however, currently only Linux 163 // supports this extension and the kernel will never choose an 164 // address above 1<<47 unless mmap is called with a hint 165 // address above 1<<47 (which we never do). 166 // 167 // arm64 hardware (as of ARMv8) limits user addresses to 48 168 // bits, in the range [0, 1<<48). 169 // 170 // ppc64, mips64, and s390x support arbitrary 64 bit addresses 171 // in hardware. On Linux, Go leans on stricter OS limits. Based 172 // on Linux's processor.h, the user address space is limited as 173 // follows on 64-bit architectures: 174 // 175 // Architecture Name Maximum Value (exclusive) 176 // --------------------------------------------------------------------- 177 // amd64 TASK_SIZE_MAX 0x007ffffffff000 (47 bit addresses) 178 // arm64 TASK_SIZE_64 0x01000000000000 (48 bit addresses) 179 // ppc64{,le} TASK_SIZE_USER64 0x00400000000000 (46 bit addresses) 180 // mips64{,le} TASK_SIZE64 0x00010000000000 (40 bit addresses) 181 // s390x TASK_SIZE 1<<64 (64 bit addresses) 182 // 183 // These limits may increase over time, but are currently at 184 // most 48 bits except on s390x. On all architectures, Linux 185 // starts placing mmap'd regions at addresses that are 186 // significantly below 48 bits, so even if it's possible to 187 // exceed Go's 48 bit limit, it's extremely unlikely in 188 // practice. 189 // 190 // On 32-bit platforms, we accept the full 32-bit address 191 // space because doing so is cheap. 192 // mips32 only has access to the low 2GB of virtual memory, so 193 // we further limit it to 31 bits. 194 // 195 // On ios/arm64, although 64-bit pointers are presumably 196 // available, pointers are truncated to 33 bits in iOS <14. 197 // Furthermore, only the top 4 GiB of the address space are 198 // actually available to the application. In iOS >=14, more 199 // of the address space is available, and the OS can now 200 // provide addresses outside of those 33 bits. Pick 40 bits 201 // as a reasonable balance between address space usage by the 202 // page allocator, and flexibility for what mmap'd regions 203 // we'll accept for the heap. We can't just move to the full 204 // 48 bits because this uses too much address space for older 205 // iOS versions. 206 // TODO(mknyszek): Once iOS <14 is deprecated, promote ios/arm64 207 // to a 48-bit address space like every other arm64 platform. 208 // 209 // WebAssembly currently has a limit of 4GB linear memory. 210 heapAddrBits = (_64bit*(1-goarch.IsWasm)*(1-goos.IsIos*goarch.IsArm64))*48 + (1-_64bit+goarch.IsWasm)*(32-(goarch.IsMips+goarch.IsMipsle)) + 40*goos.IsIos*goarch.IsArm64 211 212 // maxAlloc is the maximum size of an allocation. On 64-bit, 213 // it's theoretically possible to allocate 1<<heapAddrBits bytes. On 214 // 32-bit, however, this is one less than 1<<32 because the 215 // number of bytes in the address space doesn't actually fit 216 // in a uintptr. 217 maxAlloc = (1 << heapAddrBits) - (1-_64bit)*1 218 219 // The number of bits in a heap address, the size of heap 220 // arenas, and the L1 and L2 arena map sizes are related by 221 // 222 // (1 << addr bits) = arena size * L1 entries * L2 entries 223 // 224 // Currently, we balance these as follows: 225 // 226 // Platform Addr bits Arena size L1 entries L2 entries 227 // -------------- --------- ---------- ---------- ----------- 228 // */64-bit 48 64MB 1 4M (32MB) 229 // windows/64-bit 48 4MB 64 1M (8MB) 230 // ios/arm64 33 4MB 1 2048 (8KB) 231 // */32-bit 32 4MB 1 1024 (4KB) 232 // */mips(le) 31 4MB 1 512 (2KB) 233 234 // heapArenaBytes is the size of a heap arena. The heap 235 // consists of mappings of size heapArenaBytes, aligned to 236 // heapArenaBytes. The initial heap mapping is one arena. 237 // 238 // This is currently 64MB on 64-bit non-Windows and 4MB on 239 // 32-bit and on Windows. We use smaller arenas on Windows 240 // because all committed memory is charged to the process, 241 // even if it's not touched. Hence, for processes with small 242 // heaps, the mapped arena space needs to be commensurate. 243 // This is particularly important with the race detector, 244 // since it significantly amplifies the cost of committed 245 // memory. 246 heapArenaBytes = 1 << logHeapArenaBytes 247 248 heapArenaWords = heapArenaBytes / goarch.PtrSize 249 250 // logHeapArenaBytes is log_2 of heapArenaBytes. For clarity, 251 // prefer using heapArenaBytes where possible (we need the 252 // constant to compute some other constants). 253 logHeapArenaBytes = (6+20)*(_64bit*(1-goos.IsWindows)*(1-goarch.IsWasm)*(1-goos.IsIos*goarch.IsArm64)) + (2+20)*(_64bit*goos.IsWindows) + (2+20)*(1-_64bit) + (2+20)*goarch.IsWasm + (2+20)*goos.IsIos*goarch.IsArm64 254 255 // heapArenaBitmapWords is the size of each heap arena's bitmap in uintptrs. 256 heapArenaBitmapWords = heapArenaWords / (8 * goarch.PtrSize) 257 258 pagesPerArena = heapArenaBytes / pageSize 259 260 // arenaL1Bits is the number of bits of the arena number 261 // covered by the first level arena map. 262 // 263 // This number should be small, since the first level arena 264 // map requires PtrSize*(1<<arenaL1Bits) of space in the 265 // binary's BSS. It can be zero, in which case the first level 266 // index is effectively unused. There is a performance benefit 267 // to this, since the generated code can be more efficient, 268 // but comes at the cost of having a large L2 mapping. 269 // 270 // We use the L1 map on 64-bit Windows because the arena size 271 // is small, but the address space is still 48 bits, and 272 // there's a high cost to having a large L2. 273 arenaL1Bits = 6 * (_64bit * goos.IsWindows) 274 275 // arenaL2Bits is the number of bits of the arena number 276 // covered by the second level arena index. 277 // 278 // The size of each arena map allocation is proportional to 279 // 1<<arenaL2Bits, so it's important that this not be too 280 // large. 48 bits leads to 32MB arena index allocations, which 281 // is about the practical threshold. 282 arenaL2Bits = heapAddrBits - logHeapArenaBytes - arenaL1Bits 283 284 // arenaL1Shift is the number of bits to shift an arena frame 285 // number by to compute an index into the first level arena map. 286 arenaL1Shift = arenaL2Bits 287 288 // arenaBits is the total bits in a combined arena map index. 289 // This is split between the index into the L1 arena map and 290 // the L2 arena map. 291 arenaBits = arenaL1Bits + arenaL2Bits 292 293 // arenaBaseOffset is the pointer value that corresponds to 294 // index 0 in the heap arena map. 295 // 296 // On amd64, the address space is 48 bits, sign extended to 64 297 // bits. This offset lets us handle "negative" addresses (or 298 // high addresses if viewed as unsigned). 299 // 300 // On aix/ppc64, this offset allows to keep the heapAddrBits to 301 // 48. Otherwise, it would be 60 in order to handle mmap addresses 302 // (in range 0x0a00000000000000 - 0x0afffffffffffff). But in this 303 // case, the memory reserved in (s *pageAlloc).init for chunks 304 // is causing important slowdowns. 305 // 306 // On other platforms, the user address space is contiguous 307 // and starts at 0, so no offset is necessary. 308 arenaBaseOffset = 0xffff800000000000*goarch.IsAmd64 + 0x0a00000000000000*goos.IsAix 309 // A typed version of this constant that will make it into DWARF (for viewcore). 310 arenaBaseOffsetUintptr = uintptr(arenaBaseOffset) 311 312 // Max number of threads to run garbage collection. 313 // 2, 3, and 4 are all plausible maximums depending 314 // on the hardware details of the machine. The garbage 315 // collector scales well to 32 cpus. 316 _MaxGcproc = 32 317 318 // minLegalPointer is the smallest possible legal pointer. 319 // This is the smallest possible architectural page size, 320 // since we assume that the first page is never mapped. 321 // 322 // This should agree with minZeroPage in the compiler. 323 minLegalPointer uintptr = 4096 324 325 // minHeapForMetadataHugePages sets a threshold on when certain kinds of 326 // heap metadata, currently the arenas map L2 entries and page alloc bitmap 327 // mappings, are allowed to be backed by huge pages. If the heap goal ever 328 // exceeds this threshold, then huge pages are enabled. 329 // 330 // These numbers are chosen with the assumption that huge pages are on the 331 // order of a few MiB in size. 332 // 333 // The kind of metadata this applies to has a very low overhead when compared 334 // to address space used, but their constant overheads for small heaps would 335 // be very high if they were to be backed by huge pages (e.g. a few MiB makes 336 // a huge difference for an 8 MiB heap, but barely any difference for a 1 GiB 337 // heap). The benefit of huge pages is also not worth it for small heaps, 338 // because only a very, very small part of the metadata is used for small heaps. 339 // 340 // N.B. If the heap goal exceeds the threshold then shrinks to a very small size 341 // again, then huge pages will still be enabled for this mapping. The reason is that 342 // there's no point unless we're also returning the physical memory for these 343 // metadata mappings back to the OS. That would be quite complex to do in general 344 // as the heap is likely fragmented after a reduction in heap size. 345 minHeapForMetadataHugePages = 1 << 30 346 ) 347 348 // physPageSize is the size in bytes of the OS's physical pages. 349 // Mapping and unmapping operations must be done at multiples of 350 // physPageSize. 351 // 352 // This must be set by the OS init code (typically in osinit) before 353 // mallocinit. 354 var physPageSize uintptr 355 356 // physHugePageSize is the size in bytes of the OS's default physical huge 357 // page size whose allocation is opaque to the application. It is assumed 358 // and verified to be a power of two. 359 // 360 // If set, this must be set by the OS init code (typically in osinit) before 361 // mallocinit. However, setting it at all is optional, and leaving the default 362 // value is always safe (though potentially less efficient). 363 // 364 // Since physHugePageSize is always assumed to be a power of two, 365 // physHugePageShift is defined as physHugePageSize == 1 << physHugePageShift. 366 // The purpose of physHugePageShift is to avoid doing divisions in 367 // performance critical functions. 368 var ( 369 physHugePageSize uintptr 370 physHugePageShift uint 371 ) 372 373 func mallocinit() { 374 if class_to_size[_TinySizeClass] != _TinySize { 375 throw("bad TinySizeClass") 376 } 377 378 if heapArenaBitmapWords&(heapArenaBitmapWords-1) != 0 { 379 // heapBits expects modular arithmetic on bitmap 380 // addresses to work. 381 throw("heapArenaBitmapWords not a power of 2") 382 } 383 384 // Check physPageSize. 385 if physPageSize == 0 { 386 // The OS init code failed to fetch the physical page size. 387 throw("failed to get system page size") 388 } 389 if physPageSize > maxPhysPageSize { 390 print("system page size (", physPageSize, ") is larger than maximum page size (", maxPhysPageSize, ")\n") 391 throw("bad system page size") 392 } 393 if physPageSize < minPhysPageSize { 394 print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n") 395 throw("bad system page size") 396 } 397 if physPageSize&(physPageSize-1) != 0 { 398 print("system page size (", physPageSize, ") must be a power of 2\n") 399 throw("bad system page size") 400 } 401 if physHugePageSize&(physHugePageSize-1) != 0 { 402 print("system huge page size (", physHugePageSize, ") must be a power of 2\n") 403 throw("bad system huge page size") 404 } 405 if physHugePageSize > maxPhysHugePageSize { 406 // physHugePageSize is greater than the maximum supported huge page size. 407 // Don't throw here, like in the other cases, since a system configured 408 // in this way isn't wrong, we just don't have the code to support them. 409 // Instead, silently set the huge page size to zero. 410 physHugePageSize = 0 411 } 412 if physHugePageSize != 0 { 413 // Since physHugePageSize is a power of 2, it suffices to increase 414 // physHugePageShift until 1<<physHugePageShift == physHugePageSize. 415 for 1<<physHugePageShift != physHugePageSize { 416 physHugePageShift++ 417 } 418 } 419 if pagesPerArena%pagesPerSpanRoot != 0 { 420 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerSpanRoot (", pagesPerSpanRoot, ")\n") 421 throw("bad pagesPerSpanRoot") 422 } 423 if pagesPerArena%pagesPerReclaimerChunk != 0 { 424 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerReclaimerChunk (", pagesPerReclaimerChunk, ")\n") 425 throw("bad pagesPerReclaimerChunk") 426 } 427 // Check that the minimum size (exclusive) for a malloc header is also 428 // a size class boundary. This is important to making sure checks align 429 // across different parts of the runtime. 430 minSizeForMallocHeaderIsSizeClass := false 431 for i := 0; i < len(class_to_size); i++ { 432 if minSizeForMallocHeader == uintptr(class_to_size[i]) { 433 minSizeForMallocHeaderIsSizeClass = true 434 break 435 } 436 } 437 if !minSizeForMallocHeaderIsSizeClass { 438 throw("min size of malloc header is not a size class boundary") 439 } 440 // Check that the pointer bitmap for all small sizes without a malloc header 441 // fits in a word. 442 if minSizeForMallocHeader/goarch.PtrSize > 8*goarch.PtrSize { 443 throw("max pointer/scan bitmap size for headerless objects is too large") 444 } 445 446 if minTagBits > taggedPointerBits { 447 throw("taggedPointerbits too small") 448 } 449 450 // Initialize the heap. 451 mheap_.init() 452 mcache0 = allocmcache() 453 lockInit(&gcBitsArenas.lock, lockRankGcBitsArenas) 454 lockInit(&profInsertLock, lockRankProfInsert) 455 lockInit(&profBlockLock, lockRankProfBlock) 456 lockInit(&profMemActiveLock, lockRankProfMemActive) 457 for i := range profMemFutureLock { 458 lockInit(&profMemFutureLock[i], lockRankProfMemFuture) 459 } 460 lockInit(&globalAlloc.mutex, lockRankGlobalAlloc) 461 462 // Create initial arena growth hints. 463 if goarch.PtrSize == 8 { 464 // On a 64-bit machine, we pick the following hints 465 // because: 466 // 467 // 1. Starting from the middle of the address space 468 // makes it easier to grow out a contiguous range 469 // without running in to some other mapping. 470 // 471 // 2. This makes Go heap addresses more easily 472 // recognizable when debugging. 473 // 474 // 3. Stack scanning in gccgo is still conservative, 475 // so it's important that addresses be distinguishable 476 // from other data. 477 // 478 // Starting at 0x00c0 means that the valid memory addresses 479 // will begin 0x00c0, 0x00c1, ... 480 // In little-endian, that's c0 00, c1 00, ... None of those are valid 481 // UTF-8 sequences, and they are otherwise as far away from 482 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0 483 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors 484 // on OS X during thread allocations. 0x00c0 causes conflicts with 485 // AddressSanitizer which reserves all memory up to 0x0100. 486 // These choices reduce the odds of a conservative garbage collector 487 // not collecting memory because some non-pointer block of memory 488 // had a bit pattern that matched a memory address. 489 // 490 // However, on arm64, we ignore all this advice above and slam the 491 // allocation at 0x40 << 32 because when using 4k pages with 3-level 492 // translation buffers, the user address space is limited to 39 bits 493 // On ios/arm64, the address space is even smaller. 494 // 495 // On AIX, mmaps starts at 0x0A00000000000000 for 64-bit. 496 // processes. 497 // 498 // Space mapped for user arenas comes immediately after the range 499 // originally reserved for the regular heap when race mode is not 500 // enabled because user arena chunks can never be used for regular heap 501 // allocations and we want to avoid fragmenting the address space. 502 // 503 // In race mode we have no choice but to just use the same hints because 504 // the race detector requires that the heap be mapped contiguously. 505 for i := 0x7f; i >= 0; i-- { 506 var p uintptr 507 switch { 508 case raceenabled: 509 // The TSAN runtime requires the heap 510 // to be in the range [0x00c000000000, 511 // 0x00e000000000). 512 p = uintptr(i)<<32 | uintptrMask&(0x00c0<<32) 513 if p >= uintptrMask&0x00e000000000 { 514 continue 515 } 516 case GOARCH == "arm64" && GOOS == "ios": 517 p = uintptr(i)<<40 | uintptrMask&(0x0013<<28) 518 case GOARCH == "arm64": 519 p = uintptr(i)<<40 | uintptrMask&(0x0040<<32) 520 case GOOS == "aix": 521 if i == 0 { 522 // We don't use addresses directly after 0x0A00000000000000 523 // to avoid collisions with others mmaps done by non-go programs. 524 continue 525 } 526 p = uintptr(i)<<40 | uintptrMask&(0xa0<<52) 527 default: 528 p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32) 529 } 530 // Switch to generating hints for user arenas if we've gone 531 // through about half the hints. In race mode, take only about 532 // a quarter; we don't have very much space to work with. 533 hintList := &mheap_.arenaHints 534 if (!raceenabled && i > 0x3f) || (raceenabled && i > 0x5f) { 535 hintList = &mheap_.userArena.arenaHints 536 } 537 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 538 hint.addr = p 539 hint.next, *hintList = *hintList, hint 540 } 541 } else { 542 // On a 32-bit machine, we're much more concerned 543 // about keeping the usable heap contiguous. 544 // Hence: 545 // 546 // 1. We reserve space for all heapArenas up front so 547 // they don't get interleaved with the heap. They're 548 // ~258MB, so this isn't too bad. (We could reserve a 549 // smaller amount of space up front if this is a 550 // problem.) 551 // 552 // 2. We hint the heap to start right above the end of 553 // the binary so we have the best chance of keeping it 554 // contiguous. 555 // 556 // 3. We try to stake out a reasonably large initial 557 // heap reservation. 558 559 const arenaMetaSize = (1 << arenaBits) * unsafe.Sizeof(heapArena{}) 560 meta := uintptr(sysReserve(nil, arenaMetaSize)) 561 if meta != 0 { 562 mheap_.heapArenaAlloc.init(meta, arenaMetaSize, true) 563 } 564 565 // We want to start the arena low, but if we're linked 566 // against C code, it's possible global constructors 567 // have called malloc and adjusted the process' brk. 568 // Query the brk so we can avoid trying to map the 569 // region over it (which will cause the kernel to put 570 // the region somewhere else, likely at a high 571 // address). 572 procBrk := sbrk0() 573 574 // If we ask for the end of the data segment but the 575 // operating system requires a little more space 576 // before we can start allocating, it will give out a 577 // slightly higher pointer. Except QEMU, which is 578 // buggy, as usual: it won't adjust the pointer 579 // upward. So adjust it upward a little bit ourselves: 580 // 1/4 MB to get away from the running binary image. 581 p := firstmoduledata.end 582 if p < procBrk { 583 p = procBrk 584 } 585 if mheap_.heapArenaAlloc.next <= p && p < mheap_.heapArenaAlloc.end { 586 p = mheap_.heapArenaAlloc.end 587 } 588 p = alignUp(p+(256<<10), heapArenaBytes) 589 // Because we're worried about fragmentation on 590 // 32-bit, we try to make a large initial reservation. 591 arenaSizes := []uintptr{ 592 512 << 20, 593 256 << 20, 594 128 << 20, 595 } 596 for _, arenaSize := range arenaSizes { 597 a, size := sysReserveAligned(unsafe.Pointer(p), arenaSize, heapArenaBytes) 598 if a != nil { 599 mheap_.arena.init(uintptr(a), size, false) 600 p = mheap_.arena.end // For hint below 601 break 602 } 603 } 604 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 605 hint.addr = p 606 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 607 608 // Place the hint for user arenas just after the large reservation. 609 // 610 // While this potentially competes with the hint above, in practice we probably 611 // aren't going to be getting this far anyway on 32-bit platforms. 612 userArenaHint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 613 userArenaHint.addr = p 614 userArenaHint.next, mheap_.userArena.arenaHints = mheap_.userArena.arenaHints, userArenaHint 615 } 616 // Initialize the memory limit here because the allocator is going to look at it 617 // but we haven't called gcinit yet and we're definitely going to allocate memory before then. 618 gcController.memoryLimit.Store(maxInt64) 619 } 620 621 // sysAlloc allocates heap arena space for at least n bytes. The 622 // returned pointer is always heapArenaBytes-aligned and backed by 623 // h.arenas metadata. The returned size is always a multiple of 624 // heapArenaBytes. sysAlloc returns nil on failure. 625 // There is no corresponding free function. 626 // 627 // hintList is a list of hint addresses for where to allocate new 628 // heap arenas. It must be non-nil. 629 // 630 // register indicates whether the heap arena should be registered 631 // in allArenas. 632 // 633 // sysAlloc returns a memory region in the Reserved state. This region must 634 // be transitioned to Prepared and then Ready before use. 635 // 636 // h must be locked. 637 func (h *mheap) sysAlloc(n uintptr, hintList **arenaHint, register bool) (v unsafe.Pointer, size uintptr) { 638 assertLockHeld(&h.lock) 639 640 n = alignUp(n, heapArenaBytes) 641 642 if hintList == &h.arenaHints { 643 // First, try the arena pre-reservation. 644 // Newly-used mappings are considered released. 645 // 646 // Only do this if we're using the regular heap arena hints. 647 // This behavior is only for the heap. 648 v = h.arena.alloc(n, heapArenaBytes, &gcController.heapReleased) 649 if v != nil { 650 size = n 651 goto mapped 652 } 653 } 654 655 // Try to grow the heap at a hint address. 656 for *hintList != nil { 657 hint := *hintList 658 p := hint.addr 659 if hint.down { 660 p -= n 661 } 662 if p+n < p { 663 // We can't use this, so don't ask. 664 v = nil 665 } else if arenaIndex(p+n-1) >= 1<<arenaBits { 666 // Outside addressable heap. Can't use. 667 v = nil 668 } else { 669 v = sysReserve(unsafe.Pointer(p), n) 670 } 671 if p == uintptr(v) { 672 // Success. Update the hint. 673 if !hint.down { 674 p += n 675 } 676 hint.addr = p 677 size = n 678 break 679 } 680 // Failed. Discard this hint and try the next. 681 // 682 // TODO: This would be cleaner if sysReserve could be 683 // told to only return the requested address. In 684 // particular, this is already how Windows behaves, so 685 // it would simplify things there. 686 if v != nil { 687 sysFreeOS(v, n) 688 } 689 *hintList = hint.next 690 h.arenaHintAlloc.free(unsafe.Pointer(hint)) 691 } 692 693 if size == 0 { 694 if raceenabled { 695 // The race detector assumes the heap lives in 696 // [0x00c000000000, 0x00e000000000), but we 697 // just ran out of hints in this region. Give 698 // a nice failure. 699 throw("too many address space collisions for -race mode") 700 } 701 702 // All of the hints failed, so we'll take any 703 // (sufficiently aligned) address the kernel will give 704 // us. 705 v, size = sysReserveAligned(nil, n, heapArenaBytes) 706 if v == nil { 707 return nil, 0 708 } 709 710 // Create new hints for extending this region. 711 hint := (*arenaHint)(h.arenaHintAlloc.alloc()) 712 hint.addr, hint.down = uintptr(v), true 713 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 714 hint = (*arenaHint)(h.arenaHintAlloc.alloc()) 715 hint.addr = uintptr(v) + size 716 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 717 } 718 719 // Check for bad pointers or pointers we can't use. 720 { 721 var bad string 722 p := uintptr(v) 723 if p+size < p { 724 bad = "region exceeds uintptr range" 725 } else if arenaIndex(p) >= 1<<arenaBits { 726 bad = "base outside usable address space" 727 } else if arenaIndex(p+size-1) >= 1<<arenaBits { 728 bad = "end outside usable address space" 729 } 730 if bad != "" { 731 // This should be impossible on most architectures, 732 // but it would be really confusing to debug. 733 print("runtime: memory allocated by OS [", hex(p), ", ", hex(p+size), ") not in usable address space: ", bad, "\n") 734 throw("memory reservation exceeds address space limit") 735 } 736 } 737 738 if uintptr(v)&(heapArenaBytes-1) != 0 { 739 throw("misrounded allocation in sysAlloc") 740 } 741 742 mapped: 743 // Create arena metadata. 744 for ri := arenaIndex(uintptr(v)); ri <= arenaIndex(uintptr(v)+size-1); ri++ { 745 l2 := h.arenas[ri.l1()] 746 if l2 == nil { 747 // Allocate an L2 arena map. 748 // 749 // Use sysAllocOS instead of sysAlloc or persistentalloc because there's no 750 // statistic we can comfortably account for this space in. With this structure, 751 // we rely on demand paging to avoid large overheads, but tracking which memory 752 // is paged in is too expensive. Trying to account for the whole region means 753 // that it will appear like an enormous memory overhead in statistics, even though 754 // it is not. 755 l2 = (*[1 << arenaL2Bits]*heapArena)(sysAllocOS(unsafe.Sizeof(*l2))) 756 if l2 == nil { 757 throw("out of memory allocating heap arena map") 758 } 759 if h.arenasHugePages { 760 sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 761 } else { 762 sysNoHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 763 } 764 atomic.StorepNoWB(unsafe.Pointer(&h.arenas[ri.l1()]), unsafe.Pointer(l2)) 765 } 766 767 if l2[ri.l2()] != nil { 768 throw("arena already initialized") 769 } 770 var r *heapArena 771 r = (*heapArena)(h.heapArenaAlloc.alloc(unsafe.Sizeof(*r), goarch.PtrSize, &memstats.gcMiscSys)) 772 if r == nil { 773 r = (*heapArena)(persistentalloc(unsafe.Sizeof(*r), goarch.PtrSize, &memstats.gcMiscSys)) 774 if r == nil { 775 throw("out of memory allocating heap arena metadata") 776 } 777 } 778 779 // Register the arena in allArenas if requested. 780 if register { 781 if len(h.allArenas) == cap(h.allArenas) { 782 size := 2 * uintptr(cap(h.allArenas)) * goarch.PtrSize 783 if size == 0 { 784 size = physPageSize 785 } 786 newArray := (*notInHeap)(persistentalloc(size, goarch.PtrSize, &memstats.gcMiscSys)) 787 if newArray == nil { 788 throw("out of memory allocating allArenas") 789 } 790 oldSlice := h.allArenas 791 *(*notInHeapSlice)(unsafe.Pointer(&h.allArenas)) = notInHeapSlice{newArray, len(h.allArenas), int(size / goarch.PtrSize)} 792 copy(h.allArenas, oldSlice) 793 // Do not free the old backing array because 794 // there may be concurrent readers. Since we 795 // double the array each time, this can lead 796 // to at most 2x waste. 797 } 798 h.allArenas = h.allArenas[:len(h.allArenas)+1] 799 h.allArenas[len(h.allArenas)-1] = ri 800 } 801 802 // Store atomically just in case an object from the 803 // new heap arena becomes visible before the heap lock 804 // is released (which shouldn't happen, but there's 805 // little downside to this). 806 atomic.StorepNoWB(unsafe.Pointer(&l2[ri.l2()]), unsafe.Pointer(r)) 807 } 808 809 // Tell the race detector about the new heap memory. 810 if raceenabled { 811 racemapshadow(v, size) 812 } 813 814 return 815 } 816 817 // sysReserveAligned is like sysReserve, but the returned pointer is 818 // aligned to align bytes. It may reserve either n or n+align bytes, 819 // so it returns the size that was reserved. 820 func sysReserveAligned(v unsafe.Pointer, size, align uintptr) (unsafe.Pointer, uintptr) { 821 // Since the alignment is rather large in uses of this 822 // function, we're not likely to get it by chance, so we ask 823 // for a larger region and remove the parts we don't need. 824 retries := 0 825 retry: 826 p := uintptr(sysReserve(v, size+align)) 827 switch { 828 case p == 0: 829 return nil, 0 830 case p&(align-1) == 0: 831 return unsafe.Pointer(p), size + align 832 case GOOS == "windows": 833 // On Windows we can't release pieces of a 834 // reservation, so we release the whole thing and 835 // re-reserve the aligned sub-region. This may race, 836 // so we may have to try again. 837 sysFreeOS(unsafe.Pointer(p), size+align) 838 p = alignUp(p, align) 839 p2 := sysReserve(unsafe.Pointer(p), size) 840 if p != uintptr(p2) { 841 // Must have raced. Try again. 842 sysFreeOS(p2, size) 843 if retries++; retries == 100 { 844 throw("failed to allocate aligned heap memory; too many retries") 845 } 846 goto retry 847 } 848 // Success. 849 return p2, size 850 default: 851 // Trim off the unaligned parts. 852 pAligned := alignUp(p, align) 853 sysFreeOS(unsafe.Pointer(p), pAligned-p) 854 end := pAligned + size 855 endLen := (p + size + align) - end 856 if endLen > 0 { 857 sysFreeOS(unsafe.Pointer(end), endLen) 858 } 859 return unsafe.Pointer(pAligned), size 860 } 861 } 862 863 // enableMetadataHugePages enables huge pages for various sources of heap metadata. 864 // 865 // A note on latency: for sufficiently small heaps (<10s of GiB) this function will take constant 866 // time, but may take time proportional to the size of the mapped heap beyond that. 867 // 868 // This function is idempotent. 869 // 870 // The heap lock must not be held over this operation, since it will briefly acquire 871 // the heap lock. 872 // 873 // Must be called on the system stack because it acquires the heap lock. 874 // 875 //go:systemstack 876 func (h *mheap) enableMetadataHugePages() { 877 // Enable huge pages for page structure. 878 h.pages.enableChunkHugePages() 879 880 // Grab the lock and set arenasHugePages if it's not. 881 // 882 // Once arenasHugePages is set, all new L2 entries will be eligible for 883 // huge pages. We'll set all the old entries after we release the lock. 884 lock(&h.lock) 885 if h.arenasHugePages { 886 unlock(&h.lock) 887 return 888 } 889 h.arenasHugePages = true 890 unlock(&h.lock) 891 892 // N.B. The arenas L1 map is quite small on all platforms, so it's fine to 893 // just iterate over the whole thing. 894 for i := range h.arenas { 895 l2 := (*[1 << arenaL2Bits]*heapArena)(atomic.Loadp(unsafe.Pointer(&h.arenas[i]))) 896 if l2 == nil { 897 continue 898 } 899 sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 900 } 901 } 902 903 // base address for all 0-byte allocations 904 var zerobase uintptr 905 906 // nextFreeFast returns the next free object if one is quickly available. 907 // Otherwise it returns 0. 908 func nextFreeFast(s *mspan) gclinkptr { 909 theBit := sys.TrailingZeros64(s.allocCache) // Is there a free object in the allocCache? 910 if theBit < 64 { 911 result := s.freeindex + uint16(theBit) 912 if result < s.nelems { 913 freeidx := result + 1 914 if freeidx%64 == 0 && freeidx != s.nelems { 915 return 0 916 } 917 s.allocCache >>= uint(theBit + 1) 918 s.freeindex = freeidx 919 s.allocCount++ 920 return gclinkptr(uintptr(result)*s.elemsize + s.base()) 921 } 922 } 923 return 0 924 } 925 926 // nextFree returns the next free object from the cached span if one is available. 927 // Otherwise it refills the cache with a span with an available object and 928 // returns that object along with a flag indicating that this was a heavy 929 // weight allocation. If it is a heavy weight allocation the caller must 930 // determine whether a new GC cycle needs to be started or if the GC is active 931 // whether this goroutine needs to assist the GC. 932 // 933 // Must run in a non-preemptible context since otherwise the owner of 934 // c could change. 935 func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) { 936 s = c.alloc[spc] 937 shouldhelpgc = false 938 freeIndex := s.nextFreeIndex() 939 if freeIndex == s.nelems { 940 // The span is full. 941 if s.allocCount != s.nelems { 942 println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 943 throw("s.allocCount != s.nelems && freeIndex == s.nelems") 944 } 945 c.refill(spc) 946 shouldhelpgc = true 947 s = c.alloc[spc] 948 949 freeIndex = s.nextFreeIndex() 950 } 951 952 if freeIndex >= s.nelems { 953 throw("freeIndex is not valid") 954 } 955 956 v = gclinkptr(uintptr(freeIndex)*s.elemsize + s.base()) 957 s.allocCount++ 958 if s.allocCount > s.nelems { 959 println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 960 throw("s.allocCount > s.nelems") 961 } 962 return 963 } 964 965 // Allocate an object of size bytes. 966 // Small objects are allocated from the per-P cache's free lists. 967 // Large objects (> 32 kB) are allocated straight from the heap. 968 func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer { 969 if gcphase == _GCmarktermination { 970 throw("mallocgc called with gcphase == _GCmarktermination") 971 } 972 973 if size == 0 { 974 return unsafe.Pointer(&zerobase) 975 } 976 977 // It's possible for any malloc to trigger sweeping, which may in 978 // turn queue finalizers. Record this dynamic lock edge. 979 lockRankMayQueueFinalizer() 980 981 userSize := size 982 if asanenabled { 983 // Refer to ASAN runtime library, the malloc() function allocates extra memory, 984 // the redzone, around the user requested memory region. And the redzones are marked 985 // as unaddressable. We perform the same operations in Go to detect the overflows or 986 // underflows. 987 size += computeRZlog(size) 988 } 989 990 if debug.malloc { 991 if debug.sbrk != 0 { 992 align := uintptr(16) 993 if typ != nil { 994 // TODO(austin): This should be just 995 // align = uintptr(typ.align) 996 // but that's only 4 on 32-bit platforms, 997 // even if there's a uint64 field in typ (see #599). 998 // This causes 64-bit atomic accesses to panic. 999 // Hence, we use stricter alignment that matches 1000 // the normal allocator better. 1001 if size&7 == 0 { 1002 align = 8 1003 } else if size&3 == 0 { 1004 align = 4 1005 } else if size&1 == 0 { 1006 align = 2 1007 } else { 1008 align = 1 1009 } 1010 } 1011 return persistentalloc(size, align, &memstats.other_sys) 1012 } 1013 1014 if inittrace.active && inittrace.id == getg().goid { 1015 // Init functions are executed sequentially in a single goroutine. 1016 inittrace.allocs += 1 1017 } 1018 } 1019 1020 // assistG is the G to charge for this allocation, or nil if 1021 // GC is not currently active. 1022 assistG := deductAssistCredit(size) 1023 1024 // Set mp.mallocing to keep from being preempted by GC. 1025 mp := acquirem() 1026 if mp.mallocing != 0 { 1027 throw("malloc deadlock") 1028 } 1029 if mp.gsignal == getg() { 1030 throw("malloc during signal") 1031 } 1032 mp.mallocing = 1 1033 1034 shouldhelpgc := false 1035 dataSize := userSize 1036 c := getMCache(mp) 1037 if c == nil { 1038 throw("mallocgc called without a P or outside bootstrapping") 1039 } 1040 var span *mspan 1041 var header **_type 1042 var x unsafe.Pointer 1043 noscan := typ == nil || !typ.Pointers() 1044 // In some cases block zeroing can profitably (for latency reduction purposes) 1045 // be delayed till preemption is possible; delayedZeroing tracks that state. 1046 delayedZeroing := false 1047 // Determine if it's a 'small' object that goes into a size-classed span. 1048 // 1049 // Note: This comparison looks a little strange, but it exists to smooth out 1050 // the crossover between the largest size class and large objects that have 1051 // their own spans. The small window of object sizes between maxSmallSize-mallocHeaderSize 1052 // and maxSmallSize will be considered large, even though they might fit in 1053 // a size class. In practice this is completely fine, since the largest small 1054 // size class has a single object in it already, precisely to make the transition 1055 // to large objects smooth. 1056 if size <= maxSmallSize-mallocHeaderSize { 1057 if noscan && size < maxTinySize { 1058 // Tiny allocator. 1059 // 1060 // Tiny allocator combines several tiny allocation requests 1061 // into a single memory block. The resulting memory block 1062 // is freed when all subobjects are unreachable. The subobjects 1063 // must be noscan (don't have pointers), this ensures that 1064 // the amount of potentially wasted memory is bounded. 1065 // 1066 // Size of the memory block used for combining (maxTinySize) is tunable. 1067 // Current setting is 16 bytes, which relates to 2x worst case memory 1068 // wastage (when all but one subobjects are unreachable). 1069 // 8 bytes would result in no wastage at all, but provides less 1070 // opportunities for combining. 1071 // 32 bytes provides more opportunities for combining, 1072 // but can lead to 4x worst case wastage. 1073 // The best case winning is 8x regardless of block size. 1074 // 1075 // Objects obtained from tiny allocator must not be freed explicitly. 1076 // So when an object will be freed explicitly, we ensure that 1077 // its size >= maxTinySize. 1078 // 1079 // SetFinalizer has a special case for objects potentially coming 1080 // from tiny allocator, it such case it allows to set finalizers 1081 // for an inner byte of a memory block. 1082 // 1083 // The main targets of tiny allocator are small strings and 1084 // standalone escaping variables. On a json benchmark 1085 // the allocator reduces number of allocations by ~12% and 1086 // reduces heap size by ~20%. 1087 off := c.tinyoffset 1088 // Align tiny pointer for required (conservative) alignment. 1089 if size&7 == 0 { 1090 off = alignUp(off, 8) 1091 } else if goarch.PtrSize == 4 && size == 12 { 1092 // Conservatively align 12-byte objects to 8 bytes on 32-bit 1093 // systems so that objects whose first field is a 64-bit 1094 // value is aligned to 8 bytes and does not cause a fault on 1095 // atomic access. See issue 37262. 1096 // TODO(mknyszek): Remove this workaround if/when issue 36606 1097 // is resolved. 1098 off = alignUp(off, 8) 1099 } else if size&3 == 0 { 1100 off = alignUp(off, 4) 1101 } else if size&1 == 0 { 1102 off = alignUp(off, 2) 1103 } 1104 if off+size <= maxTinySize && c.tiny != 0 { 1105 // The object fits into existing tiny block. 1106 x = unsafe.Pointer(c.tiny + off) 1107 c.tinyoffset = off + size 1108 c.tinyAllocs++ 1109 mp.mallocing = 0 1110 releasem(mp) 1111 return x 1112 } 1113 // Allocate a new maxTinySize block. 1114 span = c.alloc[tinySpanClass] 1115 v := nextFreeFast(span) 1116 if v == 0 { 1117 v, span, shouldhelpgc = c.nextFree(tinySpanClass) 1118 } 1119 x = unsafe.Pointer(v) 1120 (*[2]uint64)(x)[0] = 0 1121 (*[2]uint64)(x)[1] = 0 1122 // See if we need to replace the existing tiny block with the new one 1123 // based on amount of remaining free space. 1124 if !raceenabled && (size < c.tinyoffset || c.tiny == 0) { 1125 // Note: disabled when race detector is on, see comment near end of this function. 1126 c.tiny = uintptr(x) 1127 c.tinyoffset = size 1128 } 1129 size = maxTinySize 1130 } else { 1131 hasHeader := !noscan && !heapBitsInSpan(size) 1132 if hasHeader { 1133 size += mallocHeaderSize 1134 } 1135 var sizeclass uint8 1136 if size <= smallSizeMax-8 { 1137 sizeclass = size_to_class8[divRoundUp(size, smallSizeDiv)] 1138 } else { 1139 sizeclass = size_to_class128[divRoundUp(size-smallSizeMax, largeSizeDiv)] 1140 } 1141 size = uintptr(class_to_size[sizeclass]) 1142 spc := makeSpanClass(sizeclass, noscan) 1143 span = c.alloc[spc] 1144 v := nextFreeFast(span) 1145 if v == 0 { 1146 v, span, shouldhelpgc = c.nextFree(spc) 1147 } 1148 x = unsafe.Pointer(v) 1149 if needzero && span.needzero != 0 { 1150 memclrNoHeapPointers(x, size) 1151 } 1152 if hasHeader { 1153 header = (**_type)(x) 1154 x = add(x, mallocHeaderSize) 1155 size -= mallocHeaderSize 1156 } 1157 } 1158 } else { 1159 shouldhelpgc = true 1160 // For large allocations, keep track of zeroed state so that 1161 // bulk zeroing can be happen later in a preemptible context. 1162 span = c.allocLarge(size, noscan) 1163 span.freeindex = 1 1164 span.allocCount = 1 1165 size = span.elemsize 1166 x = unsafe.Pointer(span.base()) 1167 if needzero && span.needzero != 0 { 1168 delayedZeroing = true 1169 } 1170 if !noscan { 1171 // Tell the GC not to look at this yet. 1172 span.largeType = nil 1173 header = &span.largeType 1174 } 1175 } 1176 if !noscan && !delayedZeroing { 1177 c.scanAlloc += heapSetType(uintptr(x), dataSize, typ, header, span) 1178 } 1179 1180 // Ensure that the stores above that initialize x to 1181 // type-safe memory and set the heap bits occur before 1182 // the caller can make x observable to the garbage 1183 // collector. Otherwise, on weakly ordered machines, 1184 // the garbage collector could follow a pointer to x, 1185 // but see uninitialized memory or stale heap bits. 1186 publicationBarrier() 1187 // As x and the heap bits are initialized, update 1188 // freeIndexForScan now so x is seen by the GC 1189 // (including conservative scan) as an allocated object. 1190 // While this pointer can't escape into user code as a 1191 // _live_ pointer until we return, conservative scanning 1192 // may find a dead pointer that happens to point into this 1193 // object. Delaying this update until now ensures that 1194 // conservative scanning considers this pointer dead until 1195 // this point. 1196 span.freeIndexForScan = span.freeindex 1197 1198 // Allocate black during GC. 1199 // All slots hold nil so no scanning is needed. 1200 // This may be racing with GC so do it atomically if there can be 1201 // a race marking the bit. 1202 if gcphase != _GCoff { 1203 gcmarknewobject(span, uintptr(x)) 1204 } 1205 1206 if raceenabled { 1207 racemalloc(x, size) 1208 } 1209 1210 if msanenabled { 1211 msanmalloc(x, size) 1212 } 1213 1214 if asanenabled { 1215 // We should only read/write the memory with the size asked by the user. 1216 // The rest of the allocated memory should be poisoned, so that we can report 1217 // errors when accessing poisoned memory. 1218 // The allocated memory is larger than required userSize, it will also include 1219 // redzone and some other padding bytes. 1220 rzBeg := unsafe.Add(x, userSize) 1221 asanpoison(rzBeg, size-userSize) 1222 asanunpoison(x, userSize) 1223 } 1224 1225 // TODO(mknyszek): We should really count the header as part 1226 // of gc_sys or something. The code below just pretends it is 1227 // internal fragmentation and matches the GC's accounting by 1228 // using the whole allocation slot. 1229 fullSize := span.elemsize 1230 if rate := MemProfileRate; rate > 0 { 1231 // Note cache c only valid while m acquired; see #47302 1232 // 1233 // N.B. Use the full size because that matches how the GC 1234 // will update the mem profile on the "free" side. 1235 if rate != 1 && fullSize < c.nextSample { 1236 c.nextSample -= fullSize 1237 } else { 1238 profilealloc(mp, x, fullSize) 1239 } 1240 } 1241 mp.mallocing = 0 1242 releasem(mp) 1243 1244 // Objects can be zeroed late in a context where preemption can occur. 1245 // If the object contains pointers, its pointer data must be cleared 1246 // or otherwise indicate that the GC shouldn't scan it. 1247 // x will keep the memory alive. 1248 if delayedZeroing { 1249 // N.B. size == fullSize always in this case. 1250 memclrNoHeapPointersChunked(size, x) // This is a possible preemption point: see #47302 1251 1252 // Finish storing the type information for this case. 1253 if !noscan { 1254 mp := acquirem() 1255 getMCache(mp).scanAlloc += heapSetType(uintptr(x), dataSize, typ, header, span) 1256 1257 // Publish the type information with the zeroed memory. 1258 publicationBarrier() 1259 releasem(mp) 1260 } 1261 } 1262 1263 if debug.malloc { 1264 if debug.allocfreetrace != 0 { 1265 tracealloc(x, size, typ) 1266 } 1267 1268 if inittrace.active && inittrace.id == getg().goid { 1269 // Init functions are executed sequentially in a single goroutine. 1270 inittrace.bytes += uint64(fullSize) 1271 } 1272 } 1273 1274 if assistG != nil { 1275 // Account for internal fragmentation in the assist 1276 // debt now that we know it. 1277 // 1278 // N.B. Use the full size because that's how the rest 1279 // of the GC accounts for bytes marked. 1280 assistG.gcAssistBytes -= int64(fullSize - dataSize) 1281 } 1282 1283 if shouldhelpgc { 1284 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1285 gcStart(t) 1286 } 1287 } 1288 1289 if raceenabled && noscan && dataSize < maxTinySize { 1290 // Pad tinysize allocations so they are aligned with the end 1291 // of the tinyalloc region. This ensures that any arithmetic 1292 // that goes off the top end of the object will be detectable 1293 // by checkptr (issue 38872). 1294 // Note that we disable tinyalloc when raceenabled for this to work. 1295 // TODO: This padding is only performed when the race detector 1296 // is enabled. It would be nice to enable it if any package 1297 // was compiled with checkptr, but there's no easy way to 1298 // detect that (especially at compile time). 1299 // TODO: enable this padding for all allocations, not just 1300 // tinyalloc ones. It's tricky because of pointer maps. 1301 // Maybe just all noscan objects? 1302 x = add(x, size-dataSize) 1303 } 1304 1305 return x 1306 } 1307 1308 // deductAssistCredit reduces the current G's assist credit 1309 // by size bytes, and assists the GC if necessary. 1310 // 1311 // Caller must be preemptible. 1312 // 1313 // Returns the G for which the assist credit was accounted. 1314 func deductAssistCredit(size uintptr) *g { 1315 var assistG *g 1316 if gcBlackenEnabled != 0 { 1317 // Charge the current user G for this allocation. 1318 assistG = getg() 1319 if assistG.m.curg != nil { 1320 assistG = assistG.m.curg 1321 } 1322 // Charge the allocation against the G. We'll account 1323 // for internal fragmentation at the end of mallocgc. 1324 assistG.gcAssistBytes -= int64(size) 1325 1326 if assistG.gcAssistBytes < 0 { 1327 // This G is in debt. Assist the GC to correct 1328 // this before allocating. This must happen 1329 // before disabling preemption. 1330 gcAssistAlloc(assistG) 1331 } 1332 } 1333 return assistG 1334 } 1335 1336 // memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers 1337 // on chunks of the buffer to be zeroed, with opportunities for preemption 1338 // along the way. memclrNoHeapPointers contains no safepoints and also 1339 // cannot be preemptively scheduled, so this provides a still-efficient 1340 // block copy that can also be preempted on a reasonable granularity. 1341 // 1342 // Use this with care; if the data being cleared is tagged to contain 1343 // pointers, this allows the GC to run before it is all cleared. 1344 func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) { 1345 v := uintptr(x) 1346 // got this from benchmarking. 128k is too small, 512k is too large. 1347 const chunkBytes = 256 * 1024 1348 vsize := v + size 1349 for voff := v; voff < vsize; voff = voff + chunkBytes { 1350 if getg().preempt { 1351 // may hold locks, e.g., profiling 1352 goschedguarded() 1353 } 1354 // clear min(avail, lump) bytes 1355 n := vsize - voff 1356 if n > chunkBytes { 1357 n = chunkBytes 1358 } 1359 memclrNoHeapPointers(unsafe.Pointer(voff), n) 1360 } 1361 } 1362 1363 // implementation of new builtin 1364 // compiler (both frontend and SSA backend) knows the signature 1365 // of this function. 1366 func newobject(typ *_type) unsafe.Pointer { 1367 return mallocgc(typ.Size_, typ, true) 1368 } 1369 1370 //go:linkname reflect_unsafe_New reflect.unsafe_New 1371 func reflect_unsafe_New(typ *_type) unsafe.Pointer { 1372 return mallocgc(typ.Size_, typ, true) 1373 } 1374 1375 //go:linkname reflectlite_unsafe_New internal/reflectlite.unsafe_New 1376 func reflectlite_unsafe_New(typ *_type) unsafe.Pointer { 1377 return mallocgc(typ.Size_, typ, true) 1378 } 1379 1380 // newarray allocates an array of n elements of type typ. 1381 func newarray(typ *_type, n int) unsafe.Pointer { 1382 if n == 1 { 1383 return mallocgc(typ.Size_, typ, true) 1384 } 1385 mem, overflow := math.MulUintptr(typ.Size_, uintptr(n)) 1386 if overflow || mem > maxAlloc || n < 0 { 1387 panic(plainError("runtime: allocation size out of range")) 1388 } 1389 return mallocgc(mem, typ, true) 1390 } 1391 1392 //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray 1393 func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer { 1394 return newarray(typ, n) 1395 } 1396 1397 func profilealloc(mp *m, x unsafe.Pointer, size uintptr) { 1398 c := getMCache(mp) 1399 if c == nil { 1400 throw("profilealloc called without a P or outside bootstrapping") 1401 } 1402 c.nextSample = nextSample() 1403 mProf_Malloc(x, size) 1404 } 1405 1406 // nextSample returns the next sampling point for heap profiling. The goal is 1407 // to sample allocations on average every MemProfileRate bytes, but with a 1408 // completely random distribution over the allocation timeline; this 1409 // corresponds to a Poisson process with parameter MemProfileRate. In Poisson 1410 // processes, the distance between two samples follows the exponential 1411 // distribution (exp(MemProfileRate)), so the best return value is a random 1412 // number taken from an exponential distribution whose mean is MemProfileRate. 1413 func nextSample() uintptr { 1414 if MemProfileRate == 1 { 1415 // Callers assign our return value to 1416 // mcache.next_sample, but next_sample is not used 1417 // when the rate is 1. So avoid the math below and 1418 // just return something. 1419 return 0 1420 } 1421 if GOOS == "plan9" { 1422 // Plan 9 doesn't support floating point in note handler. 1423 if gp := getg(); gp == gp.m.gsignal { 1424 return nextSampleNoFP() 1425 } 1426 } 1427 1428 return uintptr(fastexprand(MemProfileRate)) 1429 } 1430 1431 // fastexprand returns a random number from an exponential distribution with 1432 // the specified mean. 1433 func fastexprand(mean int) int32 { 1434 // Avoid overflow. Maximum possible step is 1435 // -ln(1/(1<<randomBitCount)) * mean, approximately 20 * mean. 1436 switch { 1437 case mean > 0x7000000: 1438 mean = 0x7000000 1439 case mean == 0: 1440 return 0 1441 } 1442 1443 // Take a random sample of the exponential distribution exp(-mean*x). 1444 // The probability distribution function is mean*exp(-mean*x), so the CDF is 1445 // p = 1 - exp(-mean*x), so 1446 // q = 1 - p == exp(-mean*x) 1447 // log_e(q) = -mean*x 1448 // -log_e(q)/mean = x 1449 // x = -log_e(q) * mean 1450 // x = log_2(q) * (-log_e(2)) * mean ; Using log_2 for efficiency 1451 const randomBitCount = 26 1452 q := cheaprandn(1<<randomBitCount) + 1 1453 qlog := fastlog2(float64(q)) - randomBitCount 1454 if qlog > 0 { 1455 qlog = 0 1456 } 1457 const minusLog2 = -0.6931471805599453 // -ln(2) 1458 return int32(qlog*(minusLog2*float64(mean))) + 1 1459 } 1460 1461 // nextSampleNoFP is similar to nextSample, but uses older, 1462 // simpler code to avoid floating point. 1463 func nextSampleNoFP() uintptr { 1464 // Set first allocation sample size. 1465 rate := MemProfileRate 1466 if rate > 0x3fffffff { // make 2*rate not overflow 1467 rate = 0x3fffffff 1468 } 1469 if rate != 0 { 1470 return uintptr(cheaprandn(uint32(2 * rate))) 1471 } 1472 return 0 1473 } 1474 1475 type persistentAlloc struct { 1476 base *notInHeap 1477 off uintptr 1478 } 1479 1480 var globalAlloc struct { 1481 mutex 1482 persistentAlloc 1483 } 1484 1485 // persistentChunkSize is the number of bytes we allocate when we grow 1486 // a persistentAlloc. 1487 const persistentChunkSize = 256 << 10 1488 1489 // persistentChunks is a list of all the persistent chunks we have 1490 // allocated. The list is maintained through the first word in the 1491 // persistent chunk. This is updated atomically. 1492 var persistentChunks *notInHeap 1493 1494 // Wrapper around sysAlloc that can allocate small chunks. 1495 // There is no associated free operation. 1496 // Intended for things like function/type/debug-related persistent data. 1497 // If align is 0, uses default align (currently 8). 1498 // The returned memory will be zeroed. 1499 // sysStat must be non-nil. 1500 // 1501 // Consider marking persistentalloc'd types not in heap by embedding 1502 // runtime/internal/sys.NotInHeap. 1503 func persistentalloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer { 1504 var p *notInHeap 1505 systemstack(func() { 1506 p = persistentalloc1(size, align, sysStat) 1507 }) 1508 return unsafe.Pointer(p) 1509 } 1510 1511 // Must run on system stack because stack growth can (re)invoke it. 1512 // See issue 9174. 1513 // 1514 //go:systemstack 1515 func persistentalloc1(size, align uintptr, sysStat *sysMemStat) *notInHeap { 1516 const ( 1517 maxBlock = 64 << 10 // VM reservation granularity is 64K on windows 1518 ) 1519 1520 if size == 0 { 1521 throw("persistentalloc: size == 0") 1522 } 1523 if align != 0 { 1524 if align&(align-1) != 0 { 1525 throw("persistentalloc: align is not a power of 2") 1526 } 1527 if align > _PageSize { 1528 throw("persistentalloc: align is too large") 1529 } 1530 } else { 1531 align = 8 1532 } 1533 1534 if size >= maxBlock { 1535 return (*notInHeap)(sysAlloc(size, sysStat)) 1536 } 1537 1538 mp := acquirem() 1539 var persistent *persistentAlloc 1540 if mp != nil && mp.p != 0 { 1541 persistent = &mp.p.ptr().palloc 1542 } else { 1543 lock(&globalAlloc.mutex) 1544 persistent = &globalAlloc.persistentAlloc 1545 } 1546 persistent.off = alignUp(persistent.off, align) 1547 if persistent.off+size > persistentChunkSize || persistent.base == nil { 1548 persistent.base = (*notInHeap)(sysAlloc(persistentChunkSize, &memstats.other_sys)) 1549 if persistent.base == nil { 1550 if persistent == &globalAlloc.persistentAlloc { 1551 unlock(&globalAlloc.mutex) 1552 } 1553 throw("runtime: cannot allocate memory") 1554 } 1555 1556 // Add the new chunk to the persistentChunks list. 1557 for { 1558 chunks := uintptr(unsafe.Pointer(persistentChunks)) 1559 *(*uintptr)(unsafe.Pointer(persistent.base)) = chunks 1560 if atomic.Casuintptr((*uintptr)(unsafe.Pointer(&persistentChunks)), chunks, uintptr(unsafe.Pointer(persistent.base))) { 1561 break 1562 } 1563 } 1564 persistent.off = alignUp(goarch.PtrSize, align) 1565 } 1566 p := persistent.base.add(persistent.off) 1567 persistent.off += size 1568 releasem(mp) 1569 if persistent == &globalAlloc.persistentAlloc { 1570 unlock(&globalAlloc.mutex) 1571 } 1572 1573 if sysStat != &memstats.other_sys { 1574 sysStat.add(int64(size)) 1575 memstats.other_sys.add(-int64(size)) 1576 } 1577 return p 1578 } 1579 1580 // inPersistentAlloc reports whether p points to memory allocated by 1581 // persistentalloc. This must be nosplit because it is called by the 1582 // cgo checker code, which is called by the write barrier code. 1583 // 1584 //go:nosplit 1585 func inPersistentAlloc(p uintptr) bool { 1586 chunk := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&persistentChunks))) 1587 for chunk != 0 { 1588 if p >= chunk && p < chunk+persistentChunkSize { 1589 return true 1590 } 1591 chunk = *(*uintptr)(unsafe.Pointer(chunk)) 1592 } 1593 return false 1594 } 1595 1596 // linearAlloc is a simple linear allocator that pre-reserves a region 1597 // of memory and then optionally maps that region into the Ready state 1598 // as needed. 1599 // 1600 // The caller is responsible for locking. 1601 type linearAlloc struct { 1602 next uintptr // next free byte 1603 mapped uintptr // one byte past end of mapped space 1604 end uintptr // end of reserved space 1605 1606 mapMemory bool // transition memory from Reserved to Ready if true 1607 } 1608 1609 func (l *linearAlloc) init(base, size uintptr, mapMemory bool) { 1610 if base+size < base { 1611 // Chop off the last byte. The runtime isn't prepared 1612 // to deal with situations where the bounds could overflow. 1613 // Leave that memory reserved, though, so we don't map it 1614 // later. 1615 size -= 1 1616 } 1617 l.next, l.mapped = base, base 1618 l.end = base + size 1619 l.mapMemory = mapMemory 1620 } 1621 1622 func (l *linearAlloc) alloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer { 1623 p := alignUp(l.next, align) 1624 if p+size > l.end { 1625 return nil 1626 } 1627 l.next = p + size 1628 if pEnd := alignUp(l.next-1, physPageSize); pEnd > l.mapped { 1629 if l.mapMemory { 1630 // Transition from Reserved to Prepared to Ready. 1631 n := pEnd - l.mapped 1632 sysMap(unsafe.Pointer(l.mapped), n, sysStat) 1633 sysUsed(unsafe.Pointer(l.mapped), n, n) 1634 } 1635 l.mapped = pEnd 1636 } 1637 return unsafe.Pointer(p) 1638 } 1639 1640 // notInHeap is off-heap memory allocated by a lower-level allocator 1641 // like sysAlloc or persistentAlloc. 1642 // 1643 // In general, it's better to use real types which embed 1644 // runtime/internal/sys.NotInHeap, but this serves as a generic type 1645 // for situations where that isn't possible (like in the allocators). 1646 // 1647 // TODO: Use this as the return type of sysAlloc, persistentAlloc, etc? 1648 type notInHeap struct{ _ sys.NotInHeap } 1649 1650 func (p *notInHeap) add(bytes uintptr) *notInHeap { 1651 return (*notInHeap)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + bytes)) 1652 } 1653 1654 // computeRZlog computes the size of the redzone. 1655 // Refer to the implementation of the compiler-rt. 1656 func computeRZlog(userSize uintptr) uintptr { 1657 switch { 1658 case userSize <= (64 - 16): 1659 return 16 << 0 1660 case userSize <= (128 - 32): 1661 return 16 << 1 1662 case userSize <= (512 - 64): 1663 return 16 << 2 1664 case userSize <= (4096 - 128): 1665 return 16 << 3 1666 case userSize <= (1<<14)-256: 1667 return 16 << 4 1668 case userSize <= (1<<15)-512: 1669 return 16 << 5 1670 case userSize <= (1<<16)-1024: 1671 return 16 << 6 1672 default: 1673 return 16 << 7 1674 } 1675 } 1676