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/goexperiment" 106 "internal/goos" 107 "internal/runtime/atomic" 108 "internal/runtime/gc" 109 "internal/runtime/math" 110 "internal/runtime/sys" 111 "unsafe" 112 ) 113 114 const ( 115 maxTinySize = _TinySize 116 tinySizeClass = _TinySizeClass 117 maxSmallSize = gc.MaxSmallSize 118 pageSize = 1 << gc.PageShift 119 pageMask = pageSize - 1 120 121 // Unused. Left for viewcore. 122 _PageSize = pageSize 123 minSizeForMallocHeader = gc.MinSizeForMallocHeader 124 mallocHeaderSize = gc.MallocHeaderSize 125 126 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems 127 _64bit = 1 << (^uintptr(0) >> 63) / 2 128 129 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go. 130 _TinySize = gc.TinySize 131 _TinySizeClass = int8(gc.TinySizeClass) 132 133 _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc 134 135 // Per-P, per order stack segment cache size. 136 _StackCacheSize = 32 * 1024 137 138 // Number of orders that get caching. Order 0 is FixedStack 139 // and each successive order is twice as large. 140 // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks 141 // will be allocated directly. 142 // Since FixedStack is different on different systems, we 143 // must vary NumStackOrders to keep the same maximum cached size. 144 // OS | FixedStack | NumStackOrders 145 // -----------------+------------+--------------- 146 // linux/darwin/bsd | 2KB | 4 147 // windows/32 | 4KB | 3 148 // windows/64 | 8KB | 2 149 // plan9 | 4KB | 3 150 _NumStackOrders = 4 - goarch.PtrSize/4*goos.IsWindows - 1*goos.IsPlan9 151 152 // heapAddrBits is the number of bits in a heap address. On 153 // amd64, addresses are sign-extended beyond heapAddrBits. On 154 // other arches, they are zero-extended. 155 // 156 // On most 64-bit platforms, we limit this to 48 bits based on a 157 // combination of hardware and OS limitations. 158 // 159 // amd64 hardware limits addresses to 48 bits, sign-extended 160 // to 64 bits. Addresses where the top 16 bits are not either 161 // all 0 or all 1 are "non-canonical" and invalid. Because of 162 // these "negative" addresses, we offset addresses by 1<<47 163 // (arenaBaseOffset) on amd64 before computing indexes into 164 // the heap arenas index. In 2017, amd64 hardware added 165 // support for 57 bit addresses; however, currently only Linux 166 // supports this extension and the kernel will never choose an 167 // address above 1<<47 unless mmap is called with a hint 168 // address above 1<<47 (which we never do). 169 // 170 // arm64 hardware (as of ARMv8) limits user addresses to 48 171 // bits, in the range [0, 1<<48). 172 // 173 // ppc64, mips64, and s390x support arbitrary 64 bit addresses 174 // in hardware. On Linux, Go leans on stricter OS limits. Based 175 // on Linux's processor.h, the user address space is limited as 176 // follows on 64-bit architectures: 177 // 178 // Architecture Name Maximum Value (exclusive) 179 // --------------------------------------------------------------------- 180 // amd64 TASK_SIZE_MAX 0x007ffffffff000 (47 bit addresses) 181 // arm64 TASK_SIZE_64 0x01000000000000 (48 bit addresses) 182 // ppc64{,le} TASK_SIZE_USER64 0x00400000000000 (46 bit addresses) 183 // mips64{,le} TASK_SIZE64 0x00010000000000 (40 bit addresses) 184 // s390x TASK_SIZE 1<<64 (64 bit addresses) 185 // 186 // These limits may increase over time, but are currently at 187 // most 48 bits except on s390x. On all architectures, Linux 188 // starts placing mmap'd regions at addresses that are 189 // significantly below 48 bits, so even if it's possible to 190 // exceed Go's 48 bit limit, it's extremely unlikely in 191 // practice. 192 // 193 // On 32-bit platforms, we accept the full 32-bit address 194 // space because doing so is cheap. 195 // mips32 only has access to the low 2GB of virtual memory, so 196 // we further limit it to 31 bits. 197 // 198 // On ios/arm64, although 64-bit pointers are presumably 199 // available, pointers are truncated to 33 bits in iOS <14. 200 // Furthermore, only the top 4 GiB of the address space are 201 // actually available to the application. In iOS >=14, more 202 // of the address space is available, and the OS can now 203 // provide addresses outside of those 33 bits. Pick 40 bits 204 // as a reasonable balance between address space usage by the 205 // page allocator, and flexibility for what mmap'd regions 206 // we'll accept for the heap. We can't just move to the full 207 // 48 bits because this uses too much address space for older 208 // iOS versions. 209 // TODO(mknyszek): Once iOS <14 is deprecated, promote ios/arm64 210 // to a 48-bit address space like every other arm64 platform. 211 // 212 // WebAssembly currently has a limit of 4GB linear memory. 213 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 214 215 // maxAlloc is the maximum size of an allocation. On 64-bit, 216 // it's theoretically possible to allocate 1<<heapAddrBits bytes. On 217 // 32-bit, however, this is one less than 1<<32 because the 218 // number of bytes in the address space doesn't actually fit 219 // in a uintptr. 220 maxAlloc = (1 << heapAddrBits) - (1-_64bit)*1 221 222 // The number of bits in a heap address, the size of heap 223 // arenas, and the L1 and L2 arena map sizes are related by 224 // 225 // (1 << addr bits) = arena size * L1 entries * L2 entries 226 // 227 // Currently, we balance these as follows: 228 // 229 // Platform Addr bits Arena size L1 entries L2 entries 230 // -------------- --------- ---------- ---------- ----------- 231 // */64-bit 48 64MB 1 4M (32MB) 232 // windows/64-bit 48 4MB 64 1M (8MB) 233 // ios/arm64 40 4MB 1 256K (2MB) 234 // */32-bit 32 4MB 1 1024 (4KB) 235 // */mips(le) 31 4MB 1 512 (2KB) 236 // wasm 32 512KB 1 8192 (64KB) 237 238 // heapArenaBytes is the size of a heap arena. The heap 239 // consists of mappings of size heapArenaBytes, aligned to 240 // heapArenaBytes. The initial heap mapping is one arena. 241 // 242 // This is currently 64MB on 64-bit non-Windows, 4MB on 243 // 32-bit and on Windows, and 512KB on Wasm. We use smaller 244 // arenas on Windows because all committed memory is charged 245 // to the process, even if it's not touched. Hence, for 246 // processes with small heaps, the mapped arena space needs 247 // to be commensurate. This is particularly important with 248 // the race detector, since it significantly amplifies the 249 // cost of committed memory. We use smaller arenas on Wasm 250 // because some Wasm programs have very small heap, and 251 // everything in the Wasm linear memory is charged. 252 heapArenaBytes = 1 << logHeapArenaBytes 253 254 heapArenaWords = heapArenaBytes / goarch.PtrSize 255 256 // logHeapArenaBytes is log_2 of heapArenaBytes. For clarity, 257 // prefer using heapArenaBytes where possible (we need the 258 // constant to compute some other constants). 259 logHeapArenaBytes = (6+20)*(_64bit*(1-goos.IsWindows)*(1-goarch.IsWasm)*(1-goos.IsIos*goarch.IsArm64)) + (2+20)*(_64bit*goos.IsWindows) + (2+20)*(1-_64bit) + (9+10)*goarch.IsWasm + (2+20)*goos.IsIos*goarch.IsArm64 260 261 // heapArenaBitmapWords is the size of each heap arena's bitmap in uintptrs. 262 heapArenaBitmapWords = heapArenaWords / (8 * goarch.PtrSize) 263 264 pagesPerArena = heapArenaBytes / pageSize 265 266 // arenaL1Bits is the number of bits of the arena number 267 // covered by the first level arena map. 268 // 269 // This number should be small, since the first level arena 270 // map requires PtrSize*(1<<arenaL1Bits) of space in the 271 // binary's BSS. It can be zero, in which case the first level 272 // index is effectively unused. There is a performance benefit 273 // to this, since the generated code can be more efficient, 274 // but comes at the cost of having a large L2 mapping. 275 // 276 // We use the L1 map on 64-bit Windows because the arena size 277 // is small, but the address space is still 48 bits, and 278 // there's a high cost to having a large L2. 279 arenaL1Bits = 6 * (_64bit * goos.IsWindows) 280 281 // arenaL2Bits is the number of bits of the arena number 282 // covered by the second level arena index. 283 // 284 // The size of each arena map allocation is proportional to 285 // 1<<arenaL2Bits, so it's important that this not be too 286 // large. 48 bits leads to 32MB arena index allocations, which 287 // is about the practical threshold. 288 arenaL2Bits = heapAddrBits - logHeapArenaBytes - arenaL1Bits 289 290 // arenaL1Shift is the number of bits to shift an arena frame 291 // number by to compute an index into the first level arena map. 292 arenaL1Shift = arenaL2Bits 293 294 // arenaBits is the total bits in a combined arena map index. 295 // This is split between the index into the L1 arena map and 296 // the L2 arena map. 297 arenaBits = arenaL1Bits + arenaL2Bits 298 299 // arenaBaseOffset is the pointer value that corresponds to 300 // index 0 in the heap arena map. 301 // 302 // On amd64, the address space is 48 bits, sign extended to 64 303 // bits. This offset lets us handle "negative" addresses (or 304 // high addresses if viewed as unsigned). 305 // 306 // On aix/ppc64, this offset allows to keep the heapAddrBits to 307 // 48. Otherwise, it would be 60 in order to handle mmap addresses 308 // (in range 0x0a00000000000000 - 0x0afffffffffffff). But in this 309 // case, the memory reserved in (s *pageAlloc).init for chunks 310 // is causing important slowdowns. 311 // 312 // On other platforms, the user address space is contiguous 313 // and starts at 0, so no offset is necessary. 314 arenaBaseOffset = 0xffff800000000000*goarch.IsAmd64 + 0x0a00000000000000*goos.IsAix 315 // A typed version of this constant that will make it into DWARF (for viewcore). 316 arenaBaseOffsetUintptr = uintptr(arenaBaseOffset) 317 318 // Max number of threads to run garbage collection. 319 // 2, 3, and 4 are all plausible maximums depending 320 // on the hardware details of the machine. The garbage 321 // collector scales well to 32 cpus. 322 _MaxGcproc = 32 323 324 // minLegalPointer is the smallest possible legal pointer. 325 // This is the smallest possible architectural page size, 326 // since we assume that the first page is never mapped. 327 // 328 // This should agree with minZeroPage in the compiler. 329 minLegalPointer uintptr = 4096 330 331 // minHeapForMetadataHugePages sets a threshold on when certain kinds of 332 // heap metadata, currently the arenas map L2 entries and page alloc bitmap 333 // mappings, are allowed to be backed by huge pages. If the heap goal ever 334 // exceeds this threshold, then huge pages are enabled. 335 // 336 // These numbers are chosen with the assumption that huge pages are on the 337 // order of a few MiB in size. 338 // 339 // The kind of metadata this applies to has a very low overhead when compared 340 // to address space used, but their constant overheads for small heaps would 341 // be very high if they were to be backed by huge pages (e.g. a few MiB makes 342 // a huge difference for an 8 MiB heap, but barely any difference for a 1 GiB 343 // heap). The benefit of huge pages is also not worth it for small heaps, 344 // because only a very, very small part of the metadata is used for small heaps. 345 // 346 // N.B. If the heap goal exceeds the threshold then shrinks to a very small size 347 // again, then huge pages will still be enabled for this mapping. The reason is that 348 // there's no point unless we're also returning the physical memory for these 349 // metadata mappings back to the OS. That would be quite complex to do in general 350 // as the heap is likely fragmented after a reduction in heap size. 351 minHeapForMetadataHugePages = 1 << 30 352 353 // randomizeHeapBase indicates if the heap base address should be randomized. 354 // See comment in mallocinit for how the randomization is performed. 355 randomizeHeapBase = goexperiment.RandomizedHeapBase64 && goarch.PtrSize == 8 && !isSbrkPlatform && !raceenabled && !msanenabled && !asanenabled 356 357 // randHeapBasePrefixMask is used to extract the top byte of the randomized 358 // heap base address. 359 randHeapBasePrefixMask = ^uintptr(0xff << (heapAddrBits - 8)) 360 ) 361 362 // physPageSize is the size in bytes of the OS's physical pages. 363 // Mapping and unmapping operations must be done at multiples of 364 // physPageSize. 365 // 366 // This must be set by the OS init code (typically in osinit) before 367 // mallocinit. 368 var physPageSize uintptr 369 370 // physHugePageSize is the size in bytes of the OS's default physical huge 371 // page size whose allocation is opaque to the application. It is assumed 372 // and verified to be a power of two. 373 // 374 // If set, this must be set by the OS init code (typically in osinit) before 375 // mallocinit. However, setting it at all is optional, and leaving the default 376 // value is always safe (though potentially less efficient). 377 // 378 // Since physHugePageSize is always assumed to be a power of two, 379 // physHugePageShift is defined as physHugePageSize == 1 << physHugePageShift. 380 // The purpose of physHugePageShift is to avoid doing divisions in 381 // performance critical functions. 382 var ( 383 physHugePageSize uintptr 384 physHugePageShift uint 385 ) 386 387 var ( 388 // heapRandSeed is a random value that is populated in mallocinit if 389 // randomizeHeapBase is set. It is used in mallocinit, and mheap.grow, to 390 // randomize the base heap address. 391 heapRandSeed uintptr 392 heapRandSeedBitsRemaining int 393 ) 394 395 func nextHeapRandBits(bits int) uintptr { 396 if bits > heapRandSeedBitsRemaining { 397 throw("not enough heapRandSeed bits remaining") 398 } 399 r := heapRandSeed >> (64 - bits) 400 heapRandSeed <<= bits 401 heapRandSeedBitsRemaining -= bits 402 return r 403 } 404 405 func mallocinit() { 406 if gc.SizeClassToSize[tinySizeClass] != maxTinySize { 407 throw("bad TinySizeClass") 408 } 409 410 if heapArenaBitmapWords&(heapArenaBitmapWords-1) != 0 { 411 // heapBits expects modular arithmetic on bitmap 412 // addresses to work. 413 throw("heapArenaBitmapWords not a power of 2") 414 } 415 416 // Check physPageSize. 417 if physPageSize == 0 { 418 // The OS init code failed to fetch the physical page size. 419 throw("failed to get system page size") 420 } 421 if physPageSize > maxPhysPageSize { 422 print("system page size (", physPageSize, ") is larger than maximum page size (", maxPhysPageSize, ")\n") 423 throw("bad system page size") 424 } 425 if physPageSize < minPhysPageSize { 426 print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n") 427 throw("bad system page size") 428 } 429 if physPageSize&(physPageSize-1) != 0 { 430 print("system page size (", physPageSize, ") must be a power of 2\n") 431 throw("bad system page size") 432 } 433 if physHugePageSize&(physHugePageSize-1) != 0 { 434 print("system huge page size (", physHugePageSize, ") must be a power of 2\n") 435 throw("bad system huge page size") 436 } 437 if physHugePageSize > maxPhysHugePageSize { 438 // physHugePageSize is greater than the maximum supported huge page size. 439 // Don't throw here, like in the other cases, since a system configured 440 // in this way isn't wrong, we just don't have the code to support them. 441 // Instead, silently set the huge page size to zero. 442 physHugePageSize = 0 443 } 444 if physHugePageSize != 0 { 445 // Since physHugePageSize is a power of 2, it suffices to increase 446 // physHugePageShift until 1<<physHugePageShift == physHugePageSize. 447 for 1<<physHugePageShift != physHugePageSize { 448 physHugePageShift++ 449 } 450 } 451 if pagesPerArena%pagesPerSpanRoot != 0 { 452 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerSpanRoot (", pagesPerSpanRoot, ")\n") 453 throw("bad pagesPerSpanRoot") 454 } 455 if pagesPerArena%pagesPerReclaimerChunk != 0 { 456 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerReclaimerChunk (", pagesPerReclaimerChunk, ")\n") 457 throw("bad pagesPerReclaimerChunk") 458 } 459 // Check that the minimum size (exclusive) for a malloc header is also 460 // a size class boundary. This is important to making sure checks align 461 // across different parts of the runtime. 462 // 463 // While we're here, also check to make sure all these size classes' 464 // span sizes are one page. Some code relies on this. 465 minSizeForMallocHeaderIsSizeClass := false 466 sizeClassesUpToMinSizeForMallocHeaderAreOnePage := true 467 for i := 0; i < len(gc.SizeClassToSize); i++ { 468 if gc.SizeClassToNPages[i] > 1 { 469 sizeClassesUpToMinSizeForMallocHeaderAreOnePage = false 470 } 471 if gc.MinSizeForMallocHeader == uintptr(gc.SizeClassToSize[i]) { 472 minSizeForMallocHeaderIsSizeClass = true 473 break 474 } 475 } 476 if !minSizeForMallocHeaderIsSizeClass { 477 throw("min size of malloc header is not a size class boundary") 478 } 479 if !sizeClassesUpToMinSizeForMallocHeaderAreOnePage { 480 throw("expected all size classes up to min size for malloc header to fit in one-page spans") 481 } 482 // Check that the pointer bitmap for all small sizes without a malloc header 483 // fits in a word. 484 if gc.MinSizeForMallocHeader/goarch.PtrSize > 8*goarch.PtrSize { 485 throw("max pointer/scan bitmap size for headerless objects is too large") 486 } 487 488 if minTagBits > tagBits { 489 throw("tagBits too small") 490 } 491 492 // Initialize the heap. 493 mheap_.init() 494 mcache0 = allocmcache() 495 lockInit(&gcBitsArenas.lock, lockRankGcBitsArenas) 496 lockInit(&profInsertLock, lockRankProfInsert) 497 lockInit(&profBlockLock, lockRankProfBlock) 498 lockInit(&profMemActiveLock, lockRankProfMemActive) 499 for i := range profMemFutureLock { 500 lockInit(&profMemFutureLock[i], lockRankProfMemFuture) 501 } 502 lockInit(&globalAlloc.mutex, lockRankGlobalAlloc) 503 504 // Create initial arena growth hints. 505 if isSbrkPlatform { 506 // Don't generate hints on sbrk platforms. We can 507 // only grow the break sequentially. 508 } else if goarch.PtrSize == 8 { 509 // On a 64-bit machine, we pick the following hints 510 // because: 511 // 512 // 1. Starting from the middle of the address space 513 // makes it easier to grow out a contiguous range 514 // without running in to some other mapping. 515 // 516 // 2. This makes Go heap addresses more easily 517 // recognizable when debugging. 518 // 519 // 3. Stack scanning in gccgo is still conservative, 520 // so it's important that addresses be distinguishable 521 // from other data. 522 // 523 // Starting at 0x00c0 means that the valid memory addresses 524 // will begin 0x00c0, 0x00c1, ... 525 // In little-endian, that's c0 00, c1 00, ... None of those are valid 526 // UTF-8 sequences, and they are otherwise as far away from 527 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0 528 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors 529 // on OS X during thread allocations. 0x00c0 causes conflicts with 530 // AddressSanitizer which reserves all memory up to 0x0100. 531 // These choices reduce the odds of a conservative garbage collector 532 // not collecting memory because some non-pointer block of memory 533 // had a bit pattern that matched a memory address. 534 // 535 // However, on arm64, we ignore all this advice above and slam the 536 // allocation at 0x40 << 32 because when using 4k pages with 3-level 537 // translation buffers, the user address space is limited to 39 bits 538 // On ios/arm64, the address space is even smaller. 539 // 540 // On AIX, mmaps starts at 0x0A00000000000000 for 64-bit. 541 // processes. 542 // 543 // Space mapped for user arenas comes immediately after the range 544 // originally reserved for the regular heap when race mode is not 545 // enabled because user arena chunks can never be used for regular heap 546 // allocations and we want to avoid fragmenting the address space. 547 // 548 // In race mode we have no choice but to just use the same hints because 549 // the race detector requires that the heap be mapped contiguously. 550 // 551 // If randomizeHeapBase is set, we attempt to randomize the base address 552 // as much as possible. We do this by generating a random uint64 via 553 // bootstrapRand and using it's bits to randomize portions of the base 554 // address as follows: 555 // * We first generate a random heapArenaBytes aligned address that we use for 556 // generating the hints. 557 // * On the first call to mheap.grow, we then generate a random PallocChunkBytes 558 // aligned offset into the mmap'd heap region, which we use as the base for 559 // the heap region. 560 // * We then select a page offset in that PallocChunkBytes region to start the 561 // heap at, and mark all the pages up to that offset as allocated. 562 // 563 // Our final randomized "heap base address" becomes the first byte of 564 // the first available page returned by the page allocator. This results 565 // in an address with at least heapAddrBits-gc.PageShift-2-(1*goarch.IsAmd64) 566 // bits of entropy. 567 568 var randHeapBase uintptr 569 var randHeapBasePrefix byte 570 // heapAddrBits is 48 on most platforms, but we only use 47 of those 571 // bits in order to provide a good amount of room for the heap to grow 572 // contiguously. On amd64, there are 48 bits, but the top bit is sign 573 // extended, so we throw away another bit, just to be safe. 574 randHeapAddrBits := heapAddrBits - 1 - (goarch.IsAmd64 * 1) 575 if randomizeHeapBase { 576 // Generate a random value, and take the bottom heapAddrBits-logHeapArenaBytes 577 // bits, using them as the top bits for randHeapBase. 578 heapRandSeed, heapRandSeedBitsRemaining = uintptr(bootstrapRand()), 64 579 580 topBits := (randHeapAddrBits - logHeapArenaBytes) 581 randHeapBase = nextHeapRandBits(topBits) << (randHeapAddrBits - topBits) 582 randHeapBase = alignUp(randHeapBase, heapArenaBytes) 583 randHeapBasePrefix = byte(randHeapBase >> (randHeapAddrBits - 8)) 584 } 585 586 var vmaSize int 587 if GOARCH == "riscv64" { 588 // Identify which memory layout is in use based on the system 589 // stack address, knowing that the bottom half of virtual memory 590 // is user space. This should result in 39, 48 or 57. It may be 591 // possible to use RISCV_HWPROBE_KEY_HIGHEST_VIRT_ADDRESS at some 592 // point in the future - for now use the system stack address. 593 vmaSize = sys.Len64(uint64(getg().m.g0.stack.hi)) + 1 594 if raceenabled && vmaSize != 39 && vmaSize != 48 { 595 println("vma size = ", vmaSize) 596 throw("riscv64 vma size is unknown and race mode is enabled") 597 } 598 } 599 600 for i := 0x7f; i >= 0; i-- { 601 var p uintptr 602 switch { 603 case raceenabled && GOARCH == "riscv64" && vmaSize == 39: 604 p = uintptr(i)<<28 | uintptrMask&(0x0013<<28) 605 if p >= uintptrMask&0x000f00000000 { 606 continue 607 } 608 case raceenabled: 609 // The TSAN runtime requires the heap 610 // to be in the range [0x00c000000000, 611 // 0x00e000000000). 612 p = uintptr(i)<<32 | uintptrMask&(0x00c0<<32) 613 if p >= uintptrMask&0x00e000000000 { 614 continue 615 } 616 case randomizeHeapBase: 617 prefix := uintptr(randHeapBasePrefix+byte(i)) << (randHeapAddrBits - 8) 618 p = prefix | (randHeapBase & randHeapBasePrefixMask) 619 case GOARCH == "arm64" && GOOS == "ios": 620 p = uintptr(i)<<40 | uintptrMask&(0x0013<<28) 621 case GOARCH == "arm64": 622 p = uintptr(i)<<40 | uintptrMask&(0x0040<<32) 623 case GOARCH == "riscv64" && vmaSize == 39: 624 p = uintptr(i)<<32 | uintptrMask&(0x0013<<28) 625 case GOOS == "aix": 626 if i == 0 { 627 // We don't use addresses directly after 0x0A00000000000000 628 // to avoid collisions with others mmaps done by non-go programs. 629 continue 630 } 631 p = uintptr(i)<<40 | uintptrMask&(0xa0<<52) 632 default: 633 p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32) 634 } 635 // Switch to generating hints for user arenas if we've gone 636 // through about half the hints. In race mode, take only about 637 // a quarter; we don't have very much space to work with. 638 hintList := &mheap_.arenaHints 639 if (!raceenabled && i > 0x3f) || (raceenabled && i > 0x5f) { 640 hintList = &mheap_.userArena.arenaHints 641 } 642 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 643 hint.addr = p 644 hint.next, *hintList = *hintList, hint 645 } 646 } else { 647 // On a 32-bit machine, we're much more concerned 648 // about keeping the usable heap contiguous. 649 // Hence: 650 // 651 // 1. We reserve space for all heapArenas up front so 652 // they don't get interleaved with the heap. They're 653 // ~258MB, so this isn't too bad. (We could reserve a 654 // smaller amount of space up front if this is a 655 // problem.) 656 // 657 // 2. We hint the heap to start right above the end of 658 // the binary so we have the best chance of keeping it 659 // contiguous. 660 // 661 // 3. We try to stake out a reasonably large initial 662 // heap reservation. 663 664 const arenaMetaSize = (1 << arenaBits) * unsafe.Sizeof(heapArena{}) 665 meta := uintptr(sysReserve(nil, arenaMetaSize, "heap reservation")) 666 if meta != 0 { 667 mheap_.heapArenaAlloc.init(meta, arenaMetaSize, true) 668 } 669 670 // We want to start the arena low, but if we're linked 671 // against C code, it's possible global constructors 672 // have called malloc and adjusted the process' brk. 673 // Query the brk so we can avoid trying to map the 674 // region over it (which will cause the kernel to put 675 // the region somewhere else, likely at a high 676 // address). 677 procBrk := sbrk0() 678 679 // If we ask for the end of the data segment but the 680 // operating system requires a little more space 681 // before we can start allocating, it will give out a 682 // slightly higher pointer. Except QEMU, which is 683 // buggy, as usual: it won't adjust the pointer 684 // upward. So adjust it upward a little bit ourselves: 685 // 1/4 MB to get away from the running binary image. 686 p := firstmoduledata.end 687 if p < procBrk { 688 p = procBrk 689 } 690 if mheap_.heapArenaAlloc.next <= p && p < mheap_.heapArenaAlloc.end { 691 p = mheap_.heapArenaAlloc.end 692 } 693 p = alignUp(p+(256<<10), heapArenaBytes) 694 // Because we're worried about fragmentation on 695 // 32-bit, we try to make a large initial reservation. 696 arenaSizes := []uintptr{ 697 512 << 20, 698 256 << 20, 699 128 << 20, 700 } 701 for _, arenaSize := range arenaSizes { 702 a, size := sysReserveAligned(unsafe.Pointer(p), arenaSize, heapArenaBytes, "heap reservation") 703 if a != nil { 704 mheap_.arena.init(uintptr(a), size, false) 705 p = mheap_.arena.end // For hint below 706 break 707 } 708 } 709 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 710 hint.addr = p 711 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 712 713 // Place the hint for user arenas just after the large reservation. 714 // 715 // While this potentially competes with the hint above, in practice we probably 716 // aren't going to be getting this far anyway on 32-bit platforms. 717 userArenaHint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) 718 userArenaHint.addr = p 719 userArenaHint.next, mheap_.userArena.arenaHints = mheap_.userArena.arenaHints, userArenaHint 720 } 721 // Initialize the memory limit here because the allocator is going to look at it 722 // but we haven't called gcinit yet and we're definitely going to allocate memory before then. 723 gcController.memoryLimit.Store(math.MaxInt64) 724 } 725 726 // sysAlloc allocates heap arena space for at least n bytes. The 727 // returned pointer is always heapArenaBytes-aligned and backed by 728 // h.arenas metadata. The returned size is always a multiple of 729 // heapArenaBytes. sysAlloc returns nil on failure. 730 // There is no corresponding free function. 731 // 732 // hintList is a list of hint addresses for where to allocate new 733 // heap arenas. It must be non-nil. 734 // 735 // sysAlloc returns a memory region in the Reserved state. This region must 736 // be transitioned to Prepared and then Ready before use. 737 // 738 // arenaList is the list the arena should be added to. 739 // 740 // h must be locked. 741 func (h *mheap) sysAlloc(n uintptr, hintList **arenaHint, arenaList *[]arenaIdx) (v unsafe.Pointer, size uintptr) { 742 assertLockHeld(&h.lock) 743 744 n = alignUp(n, heapArenaBytes) 745 746 if hintList == &h.arenaHints { 747 // First, try the arena pre-reservation. 748 // Newly-used mappings are considered released. 749 // 750 // Only do this if we're using the regular heap arena hints. 751 // This behavior is only for the heap. 752 v = h.arena.alloc(n, heapArenaBytes, &gcController.heapReleased, "heap") 753 if v != nil { 754 size = n 755 goto mapped 756 } 757 } 758 759 // Try to grow the heap at a hint address. 760 for *hintList != nil { 761 hint := *hintList 762 p := hint.addr 763 if hint.down { 764 p -= n 765 } 766 if p+n < p { 767 // We can't use this, so don't ask. 768 v = nil 769 } else if arenaIndex(p+n-1) >= 1<<arenaBits { 770 // Outside addressable heap. Can't use. 771 v = nil 772 } else { 773 v = sysReserve(unsafe.Pointer(p), n, "heap reservation") 774 } 775 if p == uintptr(v) { 776 // Success. Update the hint. 777 if !hint.down { 778 p += n 779 } 780 hint.addr = p 781 size = n 782 break 783 } 784 // Failed. Discard this hint and try the next. 785 // 786 // TODO: This would be cleaner if sysReserve could be 787 // told to only return the requested address. In 788 // particular, this is already how Windows behaves, so 789 // it would simplify things there. 790 if v != nil { 791 sysFreeOS(v, n) 792 } 793 *hintList = hint.next 794 h.arenaHintAlloc.free(unsafe.Pointer(hint)) 795 } 796 797 if size == 0 { 798 if raceenabled { 799 // The race detector assumes the heap lives in 800 // [0x00c000000000, 0x00e000000000), but we 801 // just ran out of hints in this region. Give 802 // a nice failure. 803 throw("too many address space collisions for -race mode") 804 } 805 806 // All of the hints failed, so we'll take any 807 // (sufficiently aligned) address the kernel will give 808 // us. 809 v, size = sysReserveAligned(nil, n, heapArenaBytes, "heap") 810 if v == nil { 811 return nil, 0 812 } 813 814 // Create new hints for extending this region. 815 hint := (*arenaHint)(h.arenaHintAlloc.alloc()) 816 hint.addr, hint.down = uintptr(v), true 817 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 818 hint = (*arenaHint)(h.arenaHintAlloc.alloc()) 819 hint.addr = uintptr(v) + size 820 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint 821 } 822 823 // Check for bad pointers or pointers we can't use. 824 { 825 var bad string 826 p := uintptr(v) 827 if p+size < p { 828 bad = "region exceeds uintptr range" 829 } else if arenaIndex(p) >= 1<<arenaBits { 830 bad = "base outside usable address space" 831 } else if arenaIndex(p+size-1) >= 1<<arenaBits { 832 bad = "end outside usable address space" 833 } 834 if bad != "" { 835 // This should be impossible on most architectures, 836 // but it would be really confusing to debug. 837 print("runtime: memory allocated by OS [", hex(p), ", ", hex(p+size), ") not in usable address space: ", bad, "\n") 838 throw("memory reservation exceeds address space limit") 839 } 840 } 841 842 if uintptr(v)&(heapArenaBytes-1) != 0 { 843 throw("misrounded allocation in sysAlloc") 844 } 845 846 mapped: 847 if valgrindenabled { 848 valgrindCreateMempool(v) 849 valgrindMakeMemNoAccess(v, size) 850 } 851 852 // Create arena metadata. 853 for ri := arenaIndex(uintptr(v)); ri <= arenaIndex(uintptr(v)+size-1); ri++ { 854 l2 := h.arenas[ri.l1()] 855 if l2 == nil { 856 // Allocate an L2 arena map. 857 // 858 // Use sysAllocOS instead of sysAlloc or persistentalloc because there's no 859 // statistic we can comfortably account for this space in. With this structure, 860 // we rely on demand paging to avoid large overheads, but tracking which memory 861 // is paged in is too expensive. Trying to account for the whole region means 862 // that it will appear like an enormous memory overhead in statistics, even though 863 // it is not. 864 l2 = (*[1 << arenaL2Bits]*heapArena)(sysAllocOS(unsafe.Sizeof(*l2), "heap index")) 865 if l2 == nil { 866 throw("out of memory allocating heap arena map") 867 } 868 if h.arenasHugePages { 869 sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 870 } else { 871 sysNoHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 872 } 873 atomic.StorepNoWB(unsafe.Pointer(&h.arenas[ri.l1()]), unsafe.Pointer(l2)) 874 } 875 876 if l2[ri.l2()] != nil { 877 throw("arena already initialized") 878 } 879 var r *heapArena 880 r = (*heapArena)(h.heapArenaAlloc.alloc(unsafe.Sizeof(*r), goarch.PtrSize, &memstats.gcMiscSys, "heap metadata")) 881 if r == nil { 882 r = (*heapArena)(persistentalloc(unsafe.Sizeof(*r), goarch.PtrSize, &memstats.gcMiscSys)) 883 if r == nil { 884 throw("out of memory allocating heap arena metadata") 885 } 886 } 887 888 // Register the arena in allArenas if requested. 889 if len((*arenaList)) == cap((*arenaList)) { 890 size := 2 * uintptr(cap((*arenaList))) * goarch.PtrSize 891 if size == 0 { 892 size = physPageSize 893 } 894 newArray := (*notInHeap)(persistentalloc(size, goarch.PtrSize, &memstats.gcMiscSys)) 895 if newArray == nil { 896 throw("out of memory allocating allArenas") 897 } 898 oldSlice := (*arenaList) 899 *(*notInHeapSlice)(unsafe.Pointer(&(*arenaList))) = notInHeapSlice{newArray, len((*arenaList)), int(size / goarch.PtrSize)} 900 copy((*arenaList), oldSlice) 901 // Do not free the old backing array because 902 // there may be concurrent readers. Since we 903 // double the array each time, this can lead 904 // to at most 2x waste. 905 } 906 (*arenaList) = (*arenaList)[:len((*arenaList))+1] 907 (*arenaList)[len((*arenaList))-1] = ri 908 909 // Store atomically just in case an object from the 910 // new heap arena becomes visible before the heap lock 911 // is released (which shouldn't happen, but there's 912 // little downside to this). 913 atomic.StorepNoWB(unsafe.Pointer(&l2[ri.l2()]), unsafe.Pointer(r)) 914 } 915 916 // Tell the race detector about the new heap memory. 917 if raceenabled { 918 racemapshadow(v, size) 919 } 920 921 return 922 } 923 924 // sysReserveAligned is like sysReserve, but the returned pointer is 925 // aligned to align bytes. It may reserve either n or n+align bytes, 926 // so it returns the size that was reserved. 927 func sysReserveAligned(v unsafe.Pointer, size, align uintptr, vmaName string) (unsafe.Pointer, uintptr) { 928 if isSbrkPlatform { 929 if v != nil { 930 throw("unexpected heap arena hint on sbrk platform") 931 } 932 return sysReserveAlignedSbrk(size, align) 933 } 934 // Since the alignment is rather large in uses of this 935 // function, we're not likely to get it by chance, so we ask 936 // for a larger region and remove the parts we don't need. 937 retries := 0 938 retry: 939 p := uintptr(sysReserve(v, size+align, vmaName)) 940 switch { 941 case p == 0: 942 return nil, 0 943 case p&(align-1) == 0: 944 return unsafe.Pointer(p), size + align 945 case GOOS == "windows": 946 // On Windows we can't release pieces of a 947 // reservation, so we release the whole thing and 948 // re-reserve the aligned sub-region. This may race, 949 // so we may have to try again. 950 sysFreeOS(unsafe.Pointer(p), size+align) 951 p = alignUp(p, align) 952 p2 := sysReserve(unsafe.Pointer(p), size, vmaName) 953 if p != uintptr(p2) { 954 // Must have raced. Try again. 955 sysFreeOS(p2, size) 956 if retries++; retries == 100 { 957 throw("failed to allocate aligned heap memory; too many retries") 958 } 959 goto retry 960 } 961 // Success. 962 return p2, size 963 default: 964 // Trim off the unaligned parts. 965 pAligned := alignUp(p, align) 966 sysFreeOS(unsafe.Pointer(p), pAligned-p) 967 end := pAligned + size 968 endLen := (p + size + align) - end 969 if endLen > 0 { 970 sysFreeOS(unsafe.Pointer(end), endLen) 971 } 972 return unsafe.Pointer(pAligned), size 973 } 974 } 975 976 // enableMetadataHugePages enables huge pages for various sources of heap metadata. 977 // 978 // A note on latency: for sufficiently small heaps (<10s of GiB) this function will take constant 979 // time, but may take time proportional to the size of the mapped heap beyond that. 980 // 981 // This function is idempotent. 982 // 983 // The heap lock must not be held over this operation, since it will briefly acquire 984 // the heap lock. 985 // 986 // Must be called on the system stack because it acquires the heap lock. 987 // 988 //go:systemstack 989 func (h *mheap) enableMetadataHugePages() { 990 // Enable huge pages for page structure. 991 h.pages.enableChunkHugePages() 992 993 // Grab the lock and set arenasHugePages if it's not. 994 // 995 // Once arenasHugePages is set, all new L2 entries will be eligible for 996 // huge pages. We'll set all the old entries after we release the lock. 997 lock(&h.lock) 998 if h.arenasHugePages { 999 unlock(&h.lock) 1000 return 1001 } 1002 h.arenasHugePages = true 1003 unlock(&h.lock) 1004 1005 // N.B. The arenas L1 map is quite small on all platforms, so it's fine to 1006 // just iterate over the whole thing. 1007 for i := range h.arenas { 1008 l2 := (*[1 << arenaL2Bits]*heapArena)(atomic.Loadp(unsafe.Pointer(&h.arenas[i]))) 1009 if l2 == nil { 1010 continue 1011 } 1012 sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) 1013 } 1014 } 1015 1016 // base address for all 0-byte allocations 1017 var zerobase uintptr 1018 1019 // nextFreeFast returns the next free object if one is quickly available. 1020 // Otherwise it returns 0. 1021 func nextFreeFast(s *mspan) gclinkptr { 1022 theBit := sys.TrailingZeros64(s.allocCache) // Is there a free object in the allocCache? 1023 if theBit < 64 { 1024 result := s.freeindex + uint16(theBit) 1025 if result < s.nelems { 1026 freeidx := result + 1 1027 if freeidx%64 == 0 && freeidx != s.nelems { 1028 return 0 1029 } 1030 s.allocCache >>= uint(theBit + 1) 1031 s.freeindex = freeidx 1032 s.allocCount++ 1033 return gclinkptr(uintptr(result)*s.elemsize + s.base()) 1034 } 1035 } 1036 return 0 1037 } 1038 1039 // nextFree returns the next free object from the cached span if one is available. 1040 // Otherwise it refills the cache with a span with an available object and 1041 // returns that object along with a flag indicating that this was a heavy 1042 // weight allocation. If it is a heavy weight allocation the caller must 1043 // determine whether a new GC cycle needs to be started or if the GC is active 1044 // whether this goroutine needs to assist the GC. 1045 // 1046 // Must run in a non-preemptible context since otherwise the owner of 1047 // c could change. 1048 func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, checkGCTrigger bool) { 1049 s = c.alloc[spc] 1050 checkGCTrigger = false 1051 freeIndex := s.nextFreeIndex() 1052 if freeIndex == s.nelems { 1053 // The span is full. 1054 if s.allocCount != s.nelems { 1055 println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 1056 throw("s.allocCount != s.nelems && freeIndex == s.nelems") 1057 } 1058 c.refill(spc) 1059 checkGCTrigger = true 1060 s = c.alloc[spc] 1061 1062 freeIndex = s.nextFreeIndex() 1063 } 1064 1065 if freeIndex >= s.nelems { 1066 throw("freeIndex is not valid") 1067 } 1068 1069 v = gclinkptr(uintptr(freeIndex)*s.elemsize + s.base()) 1070 s.allocCount++ 1071 if s.allocCount > s.nelems { 1072 println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems) 1073 throw("s.allocCount > s.nelems") 1074 } 1075 return 1076 } 1077 1078 // doubleCheckMalloc enables a bunch of extra checks to malloc to double-check 1079 // that various invariants are upheld. 1080 // 1081 // We might consider turning these on by default; many of them previously were. 1082 // They account for a few % of mallocgc's cost though, which does matter somewhat 1083 // at scale. 1084 const doubleCheckMalloc = false 1085 1086 // sizeSpecializedMallocEnabled is the set of conditions where we enable the size-specialized 1087 // mallocgc implementation: the experiment must be enabled, and none of the sanitizers should 1088 // be enabled. The tables used to select the size-specialized malloc function do not compile 1089 // properly on plan9, so size-specialized malloc is also disabled on plan9. 1090 const sizeSpecializedMallocEnabled = goexperiment.SizeSpecializedMalloc && GOOS != "plan9" && !asanenabled && !raceenabled && !msanenabled && !valgrindenabled 1091 1092 // Allocate an object of size bytes. 1093 // Small objects are allocated from the per-P cache's free lists. 1094 // Large objects (> 32 kB) are allocated straight from the heap. 1095 // 1096 // mallocgc should be an internal detail, 1097 // but widely used packages access it using linkname. 1098 // Notable members of the hall of shame include: 1099 // - github.com/bytedance/gopkg 1100 // - github.com/bytedance/sonic 1101 // - github.com/cloudwego/frugal 1102 // - github.com/cockroachdb/cockroach 1103 // - github.com/cockroachdb/pebble 1104 // - github.com/ugorji/go/codec 1105 // 1106 // Do not remove or change the type signature. 1107 // See go.dev/issue/67401. 1108 // 1109 //go:linkname mallocgc 1110 func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer { 1111 if doubleCheckMalloc { 1112 if gcphase == _GCmarktermination { 1113 throw("mallocgc called with gcphase == _GCmarktermination") 1114 } 1115 } 1116 1117 // Short-circuit zero-sized allocation requests. 1118 if size == 0 { 1119 return unsafe.Pointer(&zerobase) 1120 } 1121 1122 if sizeSpecializedMallocEnabled && heapBitsInSpan(size) { 1123 if typ == nil || !typ.Pointers() { 1124 return mallocNoScanTable[size](size, typ, needzero) 1125 } else { 1126 if !needzero { 1127 throw("objects with pointers must be zeroed") 1128 } 1129 return mallocScanTable[size](size, typ, needzero) 1130 } 1131 } 1132 1133 // It's possible for any malloc to trigger sweeping, which may in 1134 // turn queue finalizers. Record this dynamic lock edge. 1135 // N.B. Compiled away if lockrank experiment is not enabled. 1136 lockRankMayQueueFinalizer() 1137 1138 // Pre-malloc debug hooks. 1139 if debug.malloc { 1140 if x := preMallocgcDebug(size, typ); x != nil { 1141 return x 1142 } 1143 } 1144 1145 // For ASAN, we allocate extra memory around each allocation called the "redzone." 1146 // These "redzones" are marked as unaddressable. 1147 var asanRZ uintptr 1148 if asanenabled { 1149 asanRZ = redZoneSize(size) 1150 size += asanRZ 1151 } 1152 1153 // Assist the GC if needed. 1154 if gcBlackenEnabled != 0 { 1155 deductAssistCredit(size) 1156 } 1157 1158 // Actually do the allocation. 1159 var x unsafe.Pointer 1160 var elemsize uintptr 1161 if sizeSpecializedMallocEnabled { 1162 // we know that heapBitsInSpan is true. 1163 if size <= maxSmallSize-gc.MallocHeaderSize { 1164 if typ == nil || !typ.Pointers() { 1165 x, elemsize = mallocgcSmallNoscan(size, typ, needzero) 1166 } else { 1167 if !needzero { 1168 throw("objects with pointers must be zeroed") 1169 } 1170 x, elemsize = mallocgcSmallScanHeader(size, typ) 1171 } 1172 } else { 1173 x, elemsize = mallocgcLarge(size, typ, needzero) 1174 } 1175 } else { 1176 if size <= maxSmallSize-gc.MallocHeaderSize { 1177 if typ == nil || !typ.Pointers() { 1178 if size < maxTinySize { 1179 x, elemsize = mallocgcTiny(size, typ) 1180 } else { 1181 x, elemsize = mallocgcSmallNoscan(size, typ, needzero) 1182 } 1183 } else { 1184 if !needzero { 1185 throw("objects with pointers must be zeroed") 1186 } 1187 if heapBitsInSpan(size) { 1188 x, elemsize = mallocgcSmallScanNoHeader(size, typ) 1189 } else { 1190 x, elemsize = mallocgcSmallScanHeader(size, typ) 1191 } 1192 } 1193 } else { 1194 x, elemsize = mallocgcLarge(size, typ, needzero) 1195 } 1196 } 1197 1198 // Notify sanitizers, if enabled. 1199 if raceenabled { 1200 racemalloc(x, size-asanRZ) 1201 } 1202 if msanenabled { 1203 msanmalloc(x, size-asanRZ) 1204 } 1205 if asanenabled { 1206 // Poison the space between the end of the requested size of x 1207 // and the end of the slot. Unpoison the requested allocation. 1208 frag := elemsize - size 1209 if typ != nil && typ.Pointers() && !heapBitsInSpan(elemsize) && size <= maxSmallSize-gc.MallocHeaderSize { 1210 frag -= gc.MallocHeaderSize 1211 } 1212 asanpoison(unsafe.Add(x, size-asanRZ), asanRZ) 1213 asanunpoison(x, size-asanRZ) 1214 } 1215 if valgrindenabled { 1216 valgrindMalloc(x, size-asanRZ) 1217 } 1218 1219 // Adjust our GC assist debt to account for internal fragmentation. 1220 if gcBlackenEnabled != 0 && elemsize != 0 { 1221 if assistG := getg().m.curg; assistG != nil { 1222 assistG.gcAssistBytes -= int64(elemsize - size) 1223 } 1224 } 1225 1226 // Post-malloc debug hooks. 1227 if debug.malloc { 1228 postMallocgcDebug(x, elemsize, typ) 1229 } 1230 return x 1231 } 1232 1233 func mallocgcTiny(size uintptr, typ *_type) (unsafe.Pointer, uintptr) { 1234 // Set mp.mallocing to keep from being preempted by GC. 1235 mp := acquirem() 1236 if doubleCheckMalloc { 1237 if mp.mallocing != 0 { 1238 throw("malloc deadlock") 1239 } 1240 if mp.gsignal == getg() { 1241 throw("malloc during signal") 1242 } 1243 if typ != nil && typ.Pointers() { 1244 throw("expected noscan for tiny alloc") 1245 } 1246 } 1247 mp.mallocing = 1 1248 1249 // Tiny allocator. 1250 // 1251 // Tiny allocator combines several tiny allocation requests 1252 // into a single memory block. The resulting memory block 1253 // is freed when all subobjects are unreachable. The subobjects 1254 // must be noscan (don't have pointers), this ensures that 1255 // the amount of potentially wasted memory is bounded. 1256 // 1257 // Size of the memory block used for combining (maxTinySize) is tunable. 1258 // Current setting is 16 bytes, which relates to 2x worst case memory 1259 // wastage (when all but one subobjects are unreachable). 1260 // 8 bytes would result in no wastage at all, but provides less 1261 // opportunities for combining. 1262 // 32 bytes provides more opportunities for combining, 1263 // but can lead to 4x worst case wastage. 1264 // The best case winning is 8x regardless of block size. 1265 // 1266 // Objects obtained from tiny allocator must not be freed explicitly. 1267 // So when an object will be freed explicitly, we ensure that 1268 // its size >= maxTinySize. 1269 // 1270 // SetFinalizer has a special case for objects potentially coming 1271 // from tiny allocator, it such case it allows to set finalizers 1272 // for an inner byte of a memory block. 1273 // 1274 // The main targets of tiny allocator are small strings and 1275 // standalone escaping variables. On a json benchmark 1276 // the allocator reduces number of allocations by ~12% and 1277 // reduces heap size by ~20%. 1278 c := getMCache(mp) 1279 off := c.tinyoffset 1280 // Align tiny pointer for required (conservative) alignment. 1281 if size&7 == 0 { 1282 off = alignUp(off, 8) 1283 } else if goarch.PtrSize == 4 && size == 12 { 1284 // Conservatively align 12-byte objects to 8 bytes on 32-bit 1285 // systems so that objects whose first field is a 64-bit 1286 // value is aligned to 8 bytes and does not cause a fault on 1287 // atomic access. See issue 37262. 1288 // TODO(mknyszek): Remove this workaround if/when issue 36606 1289 // is resolved. 1290 off = alignUp(off, 8) 1291 } else if size&3 == 0 { 1292 off = alignUp(off, 4) 1293 } else if size&1 == 0 { 1294 off = alignUp(off, 2) 1295 } 1296 if off+size <= maxTinySize && c.tiny != 0 { 1297 // The object fits into existing tiny block. 1298 x := unsafe.Pointer(c.tiny + off) 1299 c.tinyoffset = off + size 1300 c.tinyAllocs++ 1301 mp.mallocing = 0 1302 releasem(mp) 1303 return x, 0 1304 } 1305 // Allocate a new maxTinySize block. 1306 checkGCTrigger := false 1307 span := c.alloc[tinySpanClass] 1308 v := nextFreeFast(span) 1309 if v == 0 { 1310 v, span, checkGCTrigger = c.nextFree(tinySpanClass) 1311 } 1312 x := unsafe.Pointer(v) 1313 (*[2]uint64)(x)[0] = 0 // Always zero 1314 (*[2]uint64)(x)[1] = 0 1315 // See if we need to replace the existing tiny block with the new one 1316 // based on amount of remaining free space. 1317 if !raceenabled && (size < c.tinyoffset || c.tiny == 0) { 1318 // Note: disabled when race detector is on, see comment near end of this function. 1319 c.tiny = uintptr(x) 1320 c.tinyoffset = size 1321 } 1322 1323 // Ensure that the stores above that initialize x to 1324 // type-safe memory and set the heap bits occur before 1325 // the caller can make x observable to the garbage 1326 // collector. Otherwise, on weakly ordered machines, 1327 // the garbage collector could follow a pointer to x, 1328 // but see uninitialized memory or stale heap bits. 1329 publicationBarrier() 1330 1331 if writeBarrier.enabled { 1332 // Allocate black during GC. 1333 // All slots hold nil so no scanning is needed. 1334 // This may be racing with GC so do it atomically if there can be 1335 // a race marking the bit. 1336 gcmarknewobject(span, uintptr(x)) 1337 } else { 1338 // Track the last free index before the mark phase. This field 1339 // is only used by the garbage collector. During the mark phase 1340 // this is used by the conservative scanner to filter out objects 1341 // that are both free and recently-allocated. It's safe to do that 1342 // because we allocate-black if the GC is enabled. The conservative 1343 // scanner produces pointers out of thin air, so without additional 1344 // synchronization it might otherwise observe a partially-initialized 1345 // object, which could crash the program. 1346 span.freeIndexForScan = span.freeindex 1347 } 1348 1349 // Note cache c only valid while m acquired; see #47302 1350 // 1351 // N.B. Use the full size because that matches how the GC 1352 // will update the mem profile on the "free" side. 1353 // 1354 // TODO(mknyszek): We should really count the header as part 1355 // of gc_sys or something. The code below just pretends it is 1356 // internal fragmentation and matches the GC's accounting by 1357 // using the whole allocation slot. 1358 c.nextSample -= int64(span.elemsize) 1359 if c.nextSample < 0 || MemProfileRate != c.memProfRate { 1360 profilealloc(mp, x, span.elemsize) 1361 } 1362 mp.mallocing = 0 1363 releasem(mp) 1364 1365 if checkGCTrigger { 1366 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1367 gcStart(t) 1368 } 1369 } 1370 1371 if raceenabled { 1372 // Pad tinysize allocations so they are aligned with the end 1373 // of the tinyalloc region. This ensures that any arithmetic 1374 // that goes off the top end of the object will be detectable 1375 // by checkptr (issue 38872). 1376 // Note that we disable tinyalloc when raceenabled for this to work. 1377 // TODO: This padding is only performed when the race detector 1378 // is enabled. It would be nice to enable it if any package 1379 // was compiled with checkptr, but there's no easy way to 1380 // detect that (especially at compile time). 1381 // TODO: enable this padding for all allocations, not just 1382 // tinyalloc ones. It's tricky because of pointer maps. 1383 // Maybe just all noscan objects? 1384 x = add(x, span.elemsize-size) 1385 } 1386 return x, span.elemsize 1387 } 1388 1389 func mallocgcSmallNoscan(size uintptr, typ *_type, needzero bool) (unsafe.Pointer, uintptr) { 1390 // Set mp.mallocing to keep from being preempted by GC. 1391 mp := acquirem() 1392 if doubleCheckMalloc { 1393 if mp.mallocing != 0 { 1394 throw("malloc deadlock") 1395 } 1396 if mp.gsignal == getg() { 1397 throw("malloc during signal") 1398 } 1399 if typ != nil && typ.Pointers() { 1400 throw("expected noscan type for noscan alloc") 1401 } 1402 } 1403 mp.mallocing = 1 1404 1405 checkGCTrigger := false 1406 c := getMCache(mp) 1407 var sizeclass uint8 1408 if size <= gc.SmallSizeMax-8 { 1409 sizeclass = gc.SizeToSizeClass8[divRoundUp(size, gc.SmallSizeDiv)] 1410 } else { 1411 sizeclass = gc.SizeToSizeClass128[divRoundUp(size-gc.SmallSizeMax, gc.LargeSizeDiv)] 1412 } 1413 size = uintptr(gc.SizeClassToSize[sizeclass]) 1414 spc := makeSpanClass(sizeclass, true) 1415 span := c.alloc[spc] 1416 v := nextFreeFast(span) 1417 if v == 0 { 1418 v, span, checkGCTrigger = c.nextFree(spc) 1419 } 1420 x := unsafe.Pointer(v) 1421 if needzero && span.needzero != 0 { 1422 memclrNoHeapPointers(x, size) 1423 } 1424 1425 // Ensure that the stores above that initialize x to 1426 // type-safe memory and set the heap bits occur before 1427 // the caller can make x observable to the garbage 1428 // collector. Otherwise, on weakly ordered machines, 1429 // the garbage collector could follow a pointer to x, 1430 // but see uninitialized memory or stale heap bits. 1431 publicationBarrier() 1432 1433 if writeBarrier.enabled { 1434 // Allocate black during GC. 1435 // All slots hold nil so no scanning is needed. 1436 // This may be racing with GC so do it atomically if there can be 1437 // a race marking the bit. 1438 gcmarknewobject(span, uintptr(x)) 1439 } else { 1440 // Track the last free index before the mark phase. This field 1441 // is only used by the garbage collector. During the mark phase 1442 // this is used by the conservative scanner to filter out objects 1443 // that are both free and recently-allocated. It's safe to do that 1444 // because we allocate-black if the GC is enabled. The conservative 1445 // scanner produces pointers out of thin air, so without additional 1446 // synchronization it might otherwise observe a partially-initialized 1447 // object, which could crash the program. 1448 span.freeIndexForScan = span.freeindex 1449 } 1450 1451 // Note cache c only valid while m acquired; see #47302 1452 // 1453 // N.B. Use the full size because that matches how the GC 1454 // will update the mem profile on the "free" side. 1455 // 1456 // TODO(mknyszek): We should really count the header as part 1457 // of gc_sys or something. The code below just pretends it is 1458 // internal fragmentation and matches the GC's accounting by 1459 // using the whole allocation slot. 1460 c.nextSample -= int64(size) 1461 if c.nextSample < 0 || MemProfileRate != c.memProfRate { 1462 profilealloc(mp, x, size) 1463 } 1464 mp.mallocing = 0 1465 releasem(mp) 1466 1467 if checkGCTrigger { 1468 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1469 gcStart(t) 1470 } 1471 } 1472 return x, size 1473 } 1474 1475 func mallocgcSmallScanNoHeader(size uintptr, typ *_type) (unsafe.Pointer, uintptr) { 1476 // Set mp.mallocing to keep from being preempted by GC. 1477 mp := acquirem() 1478 if doubleCheckMalloc { 1479 if mp.mallocing != 0 { 1480 throw("malloc deadlock") 1481 } 1482 if mp.gsignal == getg() { 1483 throw("malloc during signal") 1484 } 1485 if typ == nil || !typ.Pointers() { 1486 throw("noscan allocated in scan-only path") 1487 } 1488 if !heapBitsInSpan(size) { 1489 throw("heap bits in not in span for non-header-only path") 1490 } 1491 } 1492 mp.mallocing = 1 1493 1494 checkGCTrigger := false 1495 c := getMCache(mp) 1496 sizeclass := gc.SizeToSizeClass8[divRoundUp(size, gc.SmallSizeDiv)] 1497 spc := makeSpanClass(sizeclass, false) 1498 span := c.alloc[spc] 1499 v := nextFreeFast(span) 1500 if v == 0 { 1501 v, span, checkGCTrigger = c.nextFree(spc) 1502 } 1503 x := unsafe.Pointer(v) 1504 if span.needzero != 0 { 1505 memclrNoHeapPointers(x, size) 1506 } 1507 if goarch.PtrSize == 8 && sizeclass == 1 { 1508 // initHeapBits already set the pointer bits for the 8-byte sizeclass 1509 // on 64-bit platforms. 1510 c.scanAlloc += 8 1511 } else { 1512 c.scanAlloc += heapSetTypeNoHeader(uintptr(x), size, typ, span) 1513 } 1514 size = uintptr(gc.SizeClassToSize[sizeclass]) 1515 1516 // Ensure that the stores above that initialize x to 1517 // type-safe memory and set the heap bits occur before 1518 // the caller can make x observable to the garbage 1519 // collector. Otherwise, on weakly ordered machines, 1520 // the garbage collector could follow a pointer to x, 1521 // but see uninitialized memory or stale heap bits. 1522 publicationBarrier() 1523 1524 if writeBarrier.enabled { 1525 // Allocate black during GC. 1526 // All slots hold nil so no scanning is needed. 1527 // This may be racing with GC so do it atomically if there can be 1528 // a race marking the bit. 1529 gcmarknewobject(span, uintptr(x)) 1530 } else { 1531 // Track the last free index before the mark phase. This field 1532 // is only used by the garbage collector. During the mark phase 1533 // this is used by the conservative scanner to filter out objects 1534 // that are both free and recently-allocated. It's safe to do that 1535 // because we allocate-black if the GC is enabled. The conservative 1536 // scanner produces pointers out of thin air, so without additional 1537 // synchronization it might otherwise observe a partially-initialized 1538 // object, which could crash the program. 1539 span.freeIndexForScan = span.freeindex 1540 } 1541 1542 // Note cache c only valid while m acquired; see #47302 1543 // 1544 // N.B. Use the full size because that matches how the GC 1545 // will update the mem profile on the "free" side. 1546 // 1547 // TODO(mknyszek): We should really count the header as part 1548 // of gc_sys or something. The code below just pretends it is 1549 // internal fragmentation and matches the GC's accounting by 1550 // using the whole allocation slot. 1551 c.nextSample -= int64(size) 1552 if c.nextSample < 0 || MemProfileRate != c.memProfRate { 1553 profilealloc(mp, x, size) 1554 } 1555 mp.mallocing = 0 1556 releasem(mp) 1557 1558 if checkGCTrigger { 1559 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1560 gcStart(t) 1561 } 1562 } 1563 return x, size 1564 } 1565 1566 func mallocgcSmallScanHeader(size uintptr, typ *_type) (unsafe.Pointer, uintptr) { 1567 // Set mp.mallocing to keep from being preempted by GC. 1568 mp := acquirem() 1569 if doubleCheckMalloc { 1570 if mp.mallocing != 0 { 1571 throw("malloc deadlock") 1572 } 1573 if mp.gsignal == getg() { 1574 throw("malloc during signal") 1575 } 1576 if typ == nil || !typ.Pointers() { 1577 throw("noscan allocated in scan-only path") 1578 } 1579 if heapBitsInSpan(size) { 1580 throw("heap bits in span for header-only path") 1581 } 1582 } 1583 mp.mallocing = 1 1584 1585 checkGCTrigger := false 1586 c := getMCache(mp) 1587 size += gc.MallocHeaderSize 1588 var sizeclass uint8 1589 if size <= gc.SmallSizeMax-8 { 1590 sizeclass = gc.SizeToSizeClass8[divRoundUp(size, gc.SmallSizeDiv)] 1591 } else { 1592 sizeclass = gc.SizeToSizeClass128[divRoundUp(size-gc.SmallSizeMax, gc.LargeSizeDiv)] 1593 } 1594 size = uintptr(gc.SizeClassToSize[sizeclass]) 1595 spc := makeSpanClass(sizeclass, false) 1596 span := c.alloc[spc] 1597 v := nextFreeFast(span) 1598 if v == 0 { 1599 v, span, checkGCTrigger = c.nextFree(spc) 1600 } 1601 x := unsafe.Pointer(v) 1602 if span.needzero != 0 { 1603 memclrNoHeapPointers(x, size) 1604 } 1605 header := (**_type)(x) 1606 x = add(x, gc.MallocHeaderSize) 1607 c.scanAlloc += heapSetTypeSmallHeader(uintptr(x), size-gc.MallocHeaderSize, typ, header, span) 1608 1609 // Ensure that the stores above that initialize x to 1610 // type-safe memory and set the heap bits occur before 1611 // the caller can make x observable to the garbage 1612 // collector. Otherwise, on weakly ordered machines, 1613 // the garbage collector could follow a pointer to x, 1614 // but see uninitialized memory or stale heap bits. 1615 publicationBarrier() 1616 1617 if writeBarrier.enabled { 1618 // Allocate black during GC. 1619 // All slots hold nil so no scanning is needed. 1620 // This may be racing with GC so do it atomically if there can be 1621 // a race marking the bit. 1622 gcmarknewobject(span, uintptr(x)) 1623 } else { 1624 // Track the last free index before the mark phase. This field 1625 // is only used by the garbage collector. During the mark phase 1626 // this is used by the conservative scanner to filter out objects 1627 // that are both free and recently-allocated. It's safe to do that 1628 // because we allocate-black if the GC is enabled. The conservative 1629 // scanner produces pointers out of thin air, so without additional 1630 // synchronization it might otherwise observe a partially-initialized 1631 // object, which could crash the program. 1632 span.freeIndexForScan = span.freeindex 1633 } 1634 1635 // Note cache c only valid while m acquired; see #47302 1636 // 1637 // N.B. Use the full size because that matches how the GC 1638 // will update the mem profile on the "free" side. 1639 // 1640 // TODO(mknyszek): We should really count the header as part 1641 // of gc_sys or something. The code below just pretends it is 1642 // internal fragmentation and matches the GC's accounting by 1643 // using the whole allocation slot. 1644 c.nextSample -= int64(size) 1645 if c.nextSample < 0 || MemProfileRate != c.memProfRate { 1646 profilealloc(mp, x, size) 1647 } 1648 mp.mallocing = 0 1649 releasem(mp) 1650 1651 if checkGCTrigger { 1652 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1653 gcStart(t) 1654 } 1655 } 1656 return x, size 1657 } 1658 1659 func mallocgcLarge(size uintptr, typ *_type, needzero bool) (unsafe.Pointer, uintptr) { 1660 // Set mp.mallocing to keep from being preempted by GC. 1661 mp := acquirem() 1662 if doubleCheckMalloc { 1663 if mp.mallocing != 0 { 1664 throw("malloc deadlock") 1665 } 1666 if mp.gsignal == getg() { 1667 throw("malloc during signal") 1668 } 1669 } 1670 mp.mallocing = 1 1671 1672 c := getMCache(mp) 1673 // For large allocations, keep track of zeroed state so that 1674 // bulk zeroing can be happen later in a preemptible context. 1675 span := c.allocLarge(size, typ == nil || !typ.Pointers()) 1676 span.freeindex = 1 1677 span.allocCount = 1 1678 span.largeType = nil // Tell the GC not to look at this yet. 1679 size = span.elemsize 1680 x := unsafe.Pointer(span.base()) 1681 1682 // Ensure that the store above that sets largeType to 1683 // nil happens before the caller can make x observable 1684 // to the garbage collector. 1685 // 1686 // Otherwise, on weakly ordered machines, the garbage 1687 // collector could follow a pointer to x, but see a stale 1688 // largeType value. 1689 publicationBarrier() 1690 1691 if writeBarrier.enabled { 1692 // Allocate black during GC. 1693 // All slots hold nil so no scanning is needed. 1694 // This may be racing with GC so do it atomically if there can be 1695 // a race marking the bit. 1696 gcmarknewobject(span, uintptr(x)) 1697 } else { 1698 // Track the last free index before the mark phase. This field 1699 // is only used by the garbage collector. During the mark phase 1700 // this is used by the conservative scanner to filter out objects 1701 // that are both free and recently-allocated. It's safe to do that 1702 // because we allocate-black if the GC is enabled. The conservative 1703 // scanner produces pointers out of thin air, so without additional 1704 // synchronization it might otherwise observe a partially-initialized 1705 // object, which could crash the program. 1706 span.freeIndexForScan = span.freeindex 1707 } 1708 1709 // Note cache c only valid while m acquired; see #47302 1710 // 1711 // N.B. Use the full size because that matches how the GC 1712 // will update the mem profile on the "free" side. 1713 // 1714 // TODO(mknyszek): We should really count the header as part 1715 // of gc_sys or something. The code below just pretends it is 1716 // internal fragmentation and matches the GC's accounting by 1717 // using the whole allocation slot. 1718 c.nextSample -= int64(size) 1719 if c.nextSample < 0 || MemProfileRate != c.memProfRate { 1720 profilealloc(mp, x, size) 1721 } 1722 mp.mallocing = 0 1723 releasem(mp) 1724 1725 // Check to see if we need to trigger the GC. 1726 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { 1727 gcStart(t) 1728 } 1729 1730 // Objects can be zeroed late in a context where preemption can occur. 1731 // 1732 // x will keep the memory alive. 1733 if needzero && span.needzero != 0 { 1734 // N.B. size == fullSize always in this case. 1735 memclrNoHeapPointersChunked(size, x) // This is a possible preemption point: see #47302 1736 } 1737 1738 // Set the type and run the publication barrier while non-preemptible. We need to make 1739 // sure that between heapSetTypeLarge and publicationBarrier we cannot get preempted, 1740 // otherwise the GC could potentially observe non-zeroed memory but largeType set on weak 1741 // memory architectures. 1742 // 1743 // The GC can also potentially observe non-zeroed memory if conservative scanning spuriously 1744 // observes a partially-allocated object, see the freeIndexForScan update above. This case is 1745 // handled by synchronization inside heapSetTypeLarge. 1746 mp = acquirem() 1747 if typ != nil && typ.Pointers() { 1748 // Finish storing the type information, now that we're certain the memory is zeroed. 1749 getMCache(mp).scanAlloc += heapSetTypeLarge(uintptr(x), size, typ, span) 1750 } 1751 // Publish the object again, now with zeroed memory and initialized type information. 1752 // 1753 // Even if we didn't update any type information, this is necessary to ensure that, for example, 1754 // x written to a global without any synchronization still results in other goroutines observing 1755 // zeroed memory. 1756 publicationBarrier() 1757 releasem(mp) 1758 return x, size 1759 } 1760 1761 func preMallocgcDebug(size uintptr, typ *_type) unsafe.Pointer { 1762 if debug.sbrk != 0 { 1763 align := uintptr(16) 1764 if typ != nil { 1765 // TODO(austin): This should be just 1766 // align = uintptr(typ.align) 1767 // but that's only 4 on 32-bit platforms, 1768 // even if there's a uint64 field in typ (see #599). 1769 // This causes 64-bit atomic accesses to panic. 1770 // Hence, we use stricter alignment that matches 1771 // the normal allocator better. 1772 if size&7 == 0 { 1773 align = 8 1774 } else if size&3 == 0 { 1775 align = 4 1776 } else if size&1 == 0 { 1777 align = 2 1778 } else { 1779 align = 1 1780 } 1781 } 1782 return persistentalloc(size, align, &memstats.other_sys) 1783 } 1784 if inittrace.active && inittrace.id == getg().goid { 1785 // Init functions are executed sequentially in a single goroutine. 1786 inittrace.allocs += 1 1787 } 1788 return nil 1789 } 1790 1791 func postMallocgcDebug(x unsafe.Pointer, elemsize uintptr, typ *_type) { 1792 if inittrace.active && inittrace.id == getg().goid { 1793 // Init functions are executed sequentially in a single goroutine. 1794 inittrace.bytes += uint64(elemsize) 1795 } 1796 1797 if traceAllocFreeEnabled() { 1798 trace := traceAcquire() 1799 if trace.ok() { 1800 trace.HeapObjectAlloc(uintptr(x), typ) 1801 traceRelease(trace) 1802 } 1803 } 1804 1805 // N.B. elemsize == 0 indicates a tiny allocation, since no new slot was 1806 // allocated to fulfill this call to mallocgc. This means checkfinalizer 1807 // will only flag an error if there is actually any risk. If an allocation 1808 // has the tiny block to itself, it will not get flagged, because we won't 1809 // mark the block as a tiny block. 1810 if debug.checkfinalizers != 0 && elemsize == 0 { 1811 setTinyBlockContext(unsafe.Pointer(alignDown(uintptr(x), maxTinySize))) 1812 } 1813 } 1814 1815 // deductAssistCredit reduces the current G's assist credit 1816 // by size bytes, and assists the GC if necessary. 1817 // 1818 // Caller must be preemptible. 1819 // 1820 // Returns the G for which the assist credit was accounted. 1821 func deductAssistCredit(size uintptr) { 1822 // Charge the current user G for this allocation. 1823 assistG := getg() 1824 if assistG.m.curg != nil { 1825 assistG = assistG.m.curg 1826 } 1827 // Charge the allocation against the G. We'll account 1828 // for internal fragmentation at the end of mallocgc. 1829 assistG.gcAssistBytes -= int64(size) 1830 1831 if assistG.gcAssistBytes < 0 { 1832 // This G is in debt. Assist the GC to correct 1833 // this before allocating. This must happen 1834 // before disabling preemption. 1835 gcAssistAlloc(assistG) 1836 } 1837 } 1838 1839 // memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers 1840 // on chunks of the buffer to be zeroed, with opportunities for preemption 1841 // along the way. memclrNoHeapPointers contains no safepoints and also 1842 // cannot be preemptively scheduled, so this provides a still-efficient 1843 // block copy that can also be preempted on a reasonable granularity. 1844 // 1845 // Use this with care; if the data being cleared is tagged to contain 1846 // pointers, this allows the GC to run before it is all cleared. 1847 func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) { 1848 v := uintptr(x) 1849 // got this from benchmarking. 128k is too small, 512k is too large. 1850 const chunkBytes = 256 * 1024 1851 vsize := v + size 1852 for voff := v; voff < vsize; voff = voff + chunkBytes { 1853 if getg().preempt { 1854 // may hold locks, e.g., profiling 1855 goschedguarded() 1856 } 1857 // clear min(avail, lump) bytes 1858 n := vsize - voff 1859 if n > chunkBytes { 1860 n = chunkBytes 1861 } 1862 memclrNoHeapPointers(unsafe.Pointer(voff), n) 1863 } 1864 } 1865 1866 // implementation of new builtin 1867 // compiler (both frontend and SSA backend) knows the signature 1868 // of this function. 1869 func newobject(typ *_type) unsafe.Pointer { 1870 return mallocgc(typ.Size_, typ, true) 1871 } 1872 1873 //go:linkname maps_newobject internal/runtime/maps.newobject 1874 func maps_newobject(typ *_type) unsafe.Pointer { 1875 return newobject(typ) 1876 } 1877 1878 // reflect_unsafe_New is meant for package reflect, 1879 // but widely used packages access it using linkname. 1880 // Notable members of the hall of shame include: 1881 // - gitee.com/quant1x/gox 1882 // - github.com/goccy/json 1883 // - github.com/modern-go/reflect2 1884 // - github.com/v2pro/plz 1885 // 1886 // Do not remove or change the type signature. 1887 // See go.dev/issue/67401. 1888 // 1889 //go:linkname reflect_unsafe_New reflect.unsafe_New 1890 func reflect_unsafe_New(typ *_type) unsafe.Pointer { 1891 return mallocgc(typ.Size_, typ, true) 1892 } 1893 1894 //go:linkname reflectlite_unsafe_New internal/reflectlite.unsafe_New 1895 func reflectlite_unsafe_New(typ *_type) unsafe.Pointer { 1896 return mallocgc(typ.Size_, typ, true) 1897 } 1898 1899 // newarray allocates an array of n elements of type typ. 1900 // 1901 // newarray should be an internal detail, 1902 // but widely used packages access it using linkname. 1903 // Notable members of the hall of shame include: 1904 // - github.com/RomiChan/protobuf 1905 // - github.com/segmentio/encoding 1906 // - github.com/ugorji/go/codec 1907 // 1908 // Do not remove or change the type signature. 1909 // See go.dev/issue/67401. 1910 // 1911 //go:linkname newarray 1912 func newarray(typ *_type, n int) unsafe.Pointer { 1913 if n == 1 { 1914 return mallocgc(typ.Size_, typ, true) 1915 } 1916 mem, overflow := math.MulUintptr(typ.Size_, uintptr(n)) 1917 if overflow || mem > maxAlloc || n < 0 { 1918 panic(plainError("runtime: allocation size out of range")) 1919 } 1920 return mallocgc(mem, typ, true) 1921 } 1922 1923 // reflect_unsafe_NewArray is meant for package reflect, 1924 // but widely used packages access it using linkname. 1925 // Notable members of the hall of shame include: 1926 // - gitee.com/quant1x/gox 1927 // - github.com/bytedance/sonic 1928 // - github.com/goccy/json 1929 // - github.com/modern-go/reflect2 1930 // - github.com/segmentio/encoding 1931 // - github.com/segmentio/kafka-go 1932 // - github.com/v2pro/plz 1933 // 1934 // Do not remove or change the type signature. 1935 // See go.dev/issue/67401. 1936 // 1937 //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray 1938 func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer { 1939 return newarray(typ, n) 1940 } 1941 1942 //go:linkname maps_newarray internal/runtime/maps.newarray 1943 func maps_newarray(typ *_type, n int) unsafe.Pointer { 1944 return newarray(typ, n) 1945 } 1946 1947 // profilealloc resets the current mcache's nextSample counter and 1948 // records a memory profile sample. 1949 // 1950 // The caller must be non-preemptible and have a P. 1951 func profilealloc(mp *m, x unsafe.Pointer, size uintptr) { 1952 c := getMCache(mp) 1953 if c == nil { 1954 throw("profilealloc called without a P or outside bootstrapping") 1955 } 1956 c.memProfRate = MemProfileRate 1957 c.nextSample = nextSample() 1958 mProf_Malloc(mp, x, size) 1959 } 1960 1961 // nextSample returns the next sampling point for heap profiling. The goal is 1962 // to sample allocations on average every MemProfileRate bytes, but with a 1963 // completely random distribution over the allocation timeline; this 1964 // corresponds to a Poisson process with parameter MemProfileRate. In Poisson 1965 // processes, the distance between two samples follows the exponential 1966 // distribution (exp(MemProfileRate)), so the best return value is a random 1967 // number taken from an exponential distribution whose mean is MemProfileRate. 1968 func nextSample() int64 { 1969 if MemProfileRate == 0 { 1970 // Basically never sample. 1971 return math.MaxInt64 1972 } 1973 if MemProfileRate == 1 { 1974 // Sample immediately. 1975 return 0 1976 } 1977 return int64(fastexprand(MemProfileRate)) 1978 } 1979 1980 // fastexprand returns a random number from an exponential distribution with 1981 // the specified mean. 1982 func fastexprand(mean int) int32 { 1983 // Avoid overflow. Maximum possible step is 1984 // -ln(1/(1<<randomBitCount)) * mean, approximately 20 * mean. 1985 switch { 1986 case mean > 0x7000000: 1987 mean = 0x7000000 1988 case mean == 0: 1989 return 0 1990 } 1991 1992 // Take a random sample of the exponential distribution exp(-mean*x). 1993 // The probability distribution function is mean*exp(-mean*x), so the CDF is 1994 // p = 1 - exp(-mean*x), so 1995 // q = 1 - p == exp(-mean*x) 1996 // log_e(q) = -mean*x 1997 // -log_e(q)/mean = x 1998 // x = -log_e(q) * mean 1999 // x = log_2(q) * (-log_e(2)) * mean ; Using log_2 for efficiency 2000 const randomBitCount = 26 2001 q := cheaprandn(1<<randomBitCount) + 1 2002 qlog := fastlog2(float64(q)) - randomBitCount 2003 if qlog > 0 { 2004 qlog = 0 2005 } 2006 const minusLog2 = -0.6931471805599453 // -ln(2) 2007 return int32(qlog*(minusLog2*float64(mean))) + 1 2008 } 2009 2010 type persistentAlloc struct { 2011 base *notInHeap 2012 off uintptr 2013 } 2014 2015 var globalAlloc struct { 2016 mutex 2017 persistentAlloc 2018 } 2019 2020 // persistentChunkSize is the number of bytes we allocate when we grow 2021 // a persistentAlloc. 2022 const persistentChunkSize = 256 << 10 2023 2024 // persistentChunks is a list of all the persistent chunks we have 2025 // allocated. The list is maintained through the first word in the 2026 // persistent chunk. This is updated atomically. 2027 var persistentChunks *notInHeap 2028 2029 // Wrapper around sysAlloc that can allocate small chunks. 2030 // There is no associated free operation. 2031 // Intended for things like function/type/debug-related persistent data. 2032 // If align is 0, uses default align (currently 8). 2033 // The returned memory will be zeroed. 2034 // sysStat must be non-nil. 2035 // 2036 // Consider marking persistentalloc'd types not in heap by embedding 2037 // internal/runtime/sys.NotInHeap. 2038 // 2039 // nosplit because it is used during write barriers and must not be preempted. 2040 // 2041 //go:nosplit 2042 func persistentalloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer { 2043 var p *notInHeap 2044 systemstack(func() { 2045 p = persistentalloc1(size, align, sysStat) 2046 }) 2047 return unsafe.Pointer(p) 2048 } 2049 2050 // Must run on system stack because stack growth can (re)invoke it. 2051 // See issue 9174. 2052 // 2053 //go:systemstack 2054 func persistentalloc1(size, align uintptr, sysStat *sysMemStat) *notInHeap { 2055 const ( 2056 maxBlock = 64 << 10 // VM reservation granularity is 64K on windows 2057 ) 2058 2059 if size == 0 { 2060 throw("persistentalloc: size == 0") 2061 } 2062 if align != 0 { 2063 if align&(align-1) != 0 { 2064 throw("persistentalloc: align is not a power of 2") 2065 } 2066 if align > pageSize { 2067 throw("persistentalloc: align is too large") 2068 } 2069 } else { 2070 align = 8 2071 } 2072 2073 if size >= maxBlock { 2074 return (*notInHeap)(sysAlloc(size, sysStat, "immortal metadata")) 2075 } 2076 2077 mp := acquirem() 2078 var persistent *persistentAlloc 2079 if mp != nil && mp.p != 0 { 2080 persistent = &mp.p.ptr().palloc 2081 } else { 2082 lock(&globalAlloc.mutex) 2083 persistent = &globalAlloc.persistentAlloc 2084 } 2085 persistent.off = alignUp(persistent.off, align) 2086 if persistent.off+size > persistentChunkSize || persistent.base == nil { 2087 persistent.base = (*notInHeap)(sysAlloc(persistentChunkSize, &memstats.other_sys, "immortal metadata")) 2088 if persistent.base == nil { 2089 if persistent == &globalAlloc.persistentAlloc { 2090 unlock(&globalAlloc.mutex) 2091 } 2092 throw("runtime: cannot allocate memory") 2093 } 2094 2095 // Add the new chunk to the persistentChunks list. 2096 for { 2097 chunks := uintptr(unsafe.Pointer(persistentChunks)) 2098 *(*uintptr)(unsafe.Pointer(persistent.base)) = chunks 2099 if atomic.Casuintptr((*uintptr)(unsafe.Pointer(&persistentChunks)), chunks, uintptr(unsafe.Pointer(persistent.base))) { 2100 break 2101 } 2102 } 2103 persistent.off = alignUp(goarch.PtrSize, align) 2104 } 2105 p := persistent.base.add(persistent.off) 2106 persistent.off += size 2107 releasem(mp) 2108 if persistent == &globalAlloc.persistentAlloc { 2109 unlock(&globalAlloc.mutex) 2110 } 2111 2112 if sysStat != &memstats.other_sys { 2113 sysStat.add(int64(size)) 2114 memstats.other_sys.add(-int64(size)) 2115 } 2116 return p 2117 } 2118 2119 // inPersistentAlloc reports whether p points to memory allocated by 2120 // persistentalloc. This must be nosplit because it is called by the 2121 // cgo checker code, which is called by the write barrier code. 2122 // 2123 //go:nosplit 2124 func inPersistentAlloc(p uintptr) bool { 2125 chunk := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&persistentChunks))) 2126 for chunk != 0 { 2127 if p >= chunk && p < chunk+persistentChunkSize { 2128 return true 2129 } 2130 chunk = *(*uintptr)(unsafe.Pointer(chunk)) 2131 } 2132 return false 2133 } 2134 2135 // linearAlloc is a simple linear allocator that pre-reserves a region 2136 // of memory and then optionally maps that region into the Ready state 2137 // as needed. 2138 // 2139 // The caller is responsible for locking. 2140 type linearAlloc struct { 2141 next uintptr // next free byte 2142 mapped uintptr // one byte past end of mapped space 2143 end uintptr // end of reserved space 2144 2145 mapMemory bool // transition memory from Reserved to Ready if true 2146 } 2147 2148 func (l *linearAlloc) init(base, size uintptr, mapMemory bool) { 2149 if base+size < base { 2150 // Chop off the last byte. The runtime isn't prepared 2151 // to deal with situations where the bounds could overflow. 2152 // Leave that memory reserved, though, so we don't map it 2153 // later. 2154 size -= 1 2155 } 2156 l.next, l.mapped = base, base 2157 l.end = base + size 2158 l.mapMemory = mapMemory 2159 } 2160 2161 func (l *linearAlloc) alloc(size, align uintptr, sysStat *sysMemStat, vmaName string) unsafe.Pointer { 2162 p := alignUp(l.next, align) 2163 if p+size > l.end { 2164 return nil 2165 } 2166 l.next = p + size 2167 if pEnd := alignUp(l.next-1, physPageSize); pEnd > l.mapped { 2168 if l.mapMemory { 2169 // Transition from Reserved to Prepared to Ready. 2170 n := pEnd - l.mapped 2171 sysMap(unsafe.Pointer(l.mapped), n, sysStat, vmaName) 2172 sysUsed(unsafe.Pointer(l.mapped), n, n) 2173 } 2174 l.mapped = pEnd 2175 } 2176 return unsafe.Pointer(p) 2177 } 2178 2179 // notInHeap is off-heap memory allocated by a lower-level allocator 2180 // like sysAlloc or persistentAlloc. 2181 // 2182 // In general, it's better to use real types which embed 2183 // internal/runtime/sys.NotInHeap, but this serves as a generic type 2184 // for situations where that isn't possible (like in the allocators). 2185 // 2186 // TODO: Use this as the return type of sysAlloc, persistentAlloc, etc? 2187 type notInHeap struct{ _ sys.NotInHeap } 2188 2189 func (p *notInHeap) add(bytes uintptr) *notInHeap { 2190 return (*notInHeap)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + bytes)) 2191 } 2192 2193 // redZoneSize computes the size of the redzone for a given allocation. 2194 // Refer to the implementation of the compiler-rt. 2195 func redZoneSize(userSize uintptr) uintptr { 2196 switch { 2197 case userSize <= (64 - 16): 2198 return 16 << 0 2199 case userSize <= (128 - 32): 2200 return 16 << 1 2201 case userSize <= (512 - 64): 2202 return 16 << 2 2203 case userSize <= (4096 - 128): 2204 return 16 << 3 2205 case userSize <= (1<<14)-256: 2206 return 16 << 4 2207 case userSize <= (1<<15)-512: 2208 return 16 << 5 2209 case userSize <= (1<<16)-1024: 2210 return 16 << 6 2211 default: 2212 return 16 << 7 2213 } 2214 } 2215