// Copyright 2014 The Go Authors. All rights reserved. // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. // Memory allocator. // // This was originally based on tcmalloc, but has diverged quite a bit. // http://goog-perftools.sourceforge.net/doc/tcmalloc.html // The main allocator works in runs of pages. // Small allocation sizes (up to and including 32 kB) are // rounded to one of about 70 size classes, each of which // has its own free set of objects of exactly that size. // Any free page of memory can be split into a set of objects // of one size class, which are then managed using a free bitmap. // // The allocator's data structures are: // // fixalloc: a free-list allocator for fixed-size off-heap objects, // used to manage storage used by the allocator. // mheap: the malloc heap, managed at page (8192-byte) granularity. // mspan: a run of in-use pages managed by the mheap. // mcentral: collects all spans of a given size class. // mcache: a per-P cache of mspans with free space. // mstats: allocation statistics. // // Allocating a small object proceeds up a hierarchy of caches: // // 1. Round the size up to one of the small size classes // and look in the corresponding mspan in this P's mcache. // Scan the mspan's free bitmap to find a free slot. // If there is a free slot, allocate it. // This can all be done without acquiring a lock. // // 2. If the mspan has no free slots, obtain a new mspan // from the mcentral's list of mspans of the required size // class that have free space. // Obtaining a whole span amortizes the cost of locking // the mcentral. // // 3. If the mcentral's mspan list is empty, obtain a run // of pages from the mheap to use for the mspan. // // 4. If the mheap is empty or has no page runs large enough, // allocate a new group of pages (at least 1MB) from the // operating system. Allocating a large run of pages // amortizes the cost of talking to the operating system. // // Sweeping an mspan and freeing objects on it proceeds up a similar // hierarchy: // // 1. If the mspan is being swept in response to allocation, it // is returned to the mcache to satisfy the allocation. // // 2. Otherwise, if the mspan still has allocated objects in it, // it is placed on the mcentral free list for the mspan's size // class. // // 3. Otherwise, if all objects in the mspan are free, the mspan's // pages are returned to the mheap and the mspan is now dead. // // Allocating and freeing a large object uses the mheap // directly, bypassing the mcache and mcentral. // // If mspan.needzero is false, then free object slots in the mspan are // already zeroed. Otherwise if needzero is true, objects are zeroed as // they are allocated. There are various benefits to delaying zeroing // this way: // // 1. Stack frame allocation can avoid zeroing altogether. // // 2. It exhibits better temporal locality, since the program is // probably about to write to the memory. // // 3. We don't zero pages that never get reused. // Virtual memory layout // // The heap consists of a set of arenas, which are 64MB on 64-bit and // 4MB on 32-bit (heapArenaBytes). Each arena's start address is also // aligned to the arena size. // // Each arena has an associated heapArena object that stores the // metadata for that arena: the heap bitmap for all words in the arena // and the span map for all pages in the arena. heapArena objects are // themselves allocated off-heap. // // Since arenas are aligned, the address space can be viewed as a // series of arena frames. The arena map (mheap_.arenas) maps from // arena frame number to *heapArena, or nil for parts of the address // space not backed by the Go heap. The arena map is structured as a // two-level array consisting of a "L1" arena map and many "L2" arena // maps; however, since arenas are large, on many architectures, the // arena map consists of a single, large L2 map. // // The arena map covers the entire possible address space, allowing // the Go heap to use any part of the address space. The allocator // attempts to keep arenas contiguous so that large spans (and hence // large objects) can cross arenas. package runtime import ( "internal/goarch" "internal/goexperiment" "internal/goos" "runtime/internal/atomic" "runtime/internal/math" "runtime/internal/sys" "unsafe" ) const ( maxTinySize = _TinySize tinySizeClass = _TinySizeClass maxSmallSize = _MaxSmallSize pageShift = _PageShift pageSize = _PageSize _PageSize = 1 << _PageShift _PageMask = _PageSize - 1 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems _64bit = 1 << (^uintptr(0) >> 63) / 2 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go. _TinySize = 16 _TinySizeClass = int8(2) _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc // Per-P, per order stack segment cache size. _StackCacheSize = 32 * 1024 // Number of orders that get caching. Order 0 is FixedStack // and each successive order is twice as large. // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks // will be allocated directly. // Since FixedStack is different on different systems, we // must vary NumStackOrders to keep the same maximum cached size. // OS | FixedStack | NumStackOrders // -----------------+------------+--------------- // linux/darwin/bsd | 2KB | 4 // windows/32 | 4KB | 3 // windows/64 | 8KB | 2 // plan9 | 4KB | 3 _NumStackOrders = 4 - goarch.PtrSize/4*goos.IsWindows - 1*goos.IsPlan9 // heapAddrBits is the number of bits in a heap address. On // amd64, addresses are sign-extended beyond heapAddrBits. On // other arches, they are zero-extended. // // On most 64-bit platforms, we limit this to 48 bits based on a // combination of hardware and OS limitations. // // amd64 hardware limits addresses to 48 bits, sign-extended // to 64 bits. Addresses where the top 16 bits are not either // all 0 or all 1 are "non-canonical" and invalid. Because of // these "negative" addresses, we offset addresses by 1<<47 // (arenaBaseOffset) on amd64 before computing indexes into // the heap arenas index. In 2017, amd64 hardware added // support for 57 bit addresses; however, currently only Linux // supports this extension and the kernel will never choose an // address above 1<<47 unless mmap is called with a hint // address above 1<<47 (which we never do). // // arm64 hardware (as of ARMv8) limits user addresses to 48 // bits, in the range [0, 1<<48). // // ppc64, mips64, and s390x support arbitrary 64 bit addresses // in hardware. On Linux, Go leans on stricter OS limits. Based // on Linux's processor.h, the user address space is limited as // follows on 64-bit architectures: // // Architecture Name Maximum Value (exclusive) // --------------------------------------------------------------------- // amd64 TASK_SIZE_MAX 0x007ffffffff000 (47 bit addresses) // arm64 TASK_SIZE_64 0x01000000000000 (48 bit addresses) // ppc64{,le} TASK_SIZE_USER64 0x00400000000000 (46 bit addresses) // mips64{,le} TASK_SIZE64 0x00010000000000 (40 bit addresses) // s390x TASK_SIZE 1<<64 (64 bit addresses) // // These limits may increase over time, but are currently at // most 48 bits except on s390x. On all architectures, Linux // starts placing mmap'd regions at addresses that are // significantly below 48 bits, so even if it's possible to // exceed Go's 48 bit limit, it's extremely unlikely in // practice. // // On 32-bit platforms, we accept the full 32-bit address // space because doing so is cheap. // mips32 only has access to the low 2GB of virtual memory, so // we further limit it to 31 bits. // // On ios/arm64, although 64-bit pointers are presumably // available, pointers are truncated to 33 bits in iOS <14. // Furthermore, only the top 4 GiB of the address space are // actually available to the application. In iOS >=14, more // of the address space is available, and the OS can now // provide addresses outside of those 33 bits. Pick 40 bits // as a reasonable balance between address space usage by the // page allocator, and flexibility for what mmap'd regions // we'll accept for the heap. We can't just move to the full // 48 bits because this uses too much address space for older // iOS versions. // TODO(mknyszek): Once iOS <14 is deprecated, promote ios/arm64 // to a 48-bit address space like every other arm64 platform. // // WebAssembly currently has a limit of 4GB linear memory. 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 // maxAlloc is the maximum size of an allocation. On 64-bit, // it's theoretically possible to allocate 1< maxPhysPageSize { print("system page size (", physPageSize, ") is larger than maximum page size (", maxPhysPageSize, ")\n") throw("bad system page size") } if physPageSize < minPhysPageSize { print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n") throw("bad system page size") } if physPageSize&(physPageSize-1) != 0 { print("system page size (", physPageSize, ") must be a power of 2\n") throw("bad system page size") } if physHugePageSize&(physHugePageSize-1) != 0 { print("system huge page size (", physHugePageSize, ") must be a power of 2\n") throw("bad system huge page size") } if physHugePageSize > maxPhysHugePageSize { // physHugePageSize is greater than the maximum supported huge page size. // Don't throw here, like in the other cases, since a system configured // in this way isn't wrong, we just don't have the code to support them. // Instead, silently set the huge page size to zero. physHugePageSize = 0 } if physHugePageSize != 0 { // Since physHugePageSize is a power of 2, it suffices to increase // physHugePageShift until 1< 8*goarch.PtrSize { throw("max pointer/scan bitmap size for headerless objects is too large") } } if minTagBits > taggedPointerBits { throw("taggedPointerbits too small") } // Initialize the heap. mheap_.init() mcache0 = allocmcache() lockInit(&gcBitsArenas.lock, lockRankGcBitsArenas) lockInit(&profInsertLock, lockRankProfInsert) lockInit(&profBlockLock, lockRankProfBlock) lockInit(&profMemActiveLock, lockRankProfMemActive) for i := range profMemFutureLock { lockInit(&profMemFutureLock[i], lockRankProfMemFuture) } lockInit(&globalAlloc.mutex, lockRankGlobalAlloc) // Create initial arena growth hints. if goarch.PtrSize == 8 { // On a 64-bit machine, we pick the following hints // because: // // 1. Starting from the middle of the address space // makes it easier to grow out a contiguous range // without running in to some other mapping. // // 2. This makes Go heap addresses more easily // recognizable when debugging. // // 3. Stack scanning in gccgo is still conservative, // so it's important that addresses be distinguishable // from other data. // // Starting at 0x00c0 means that the valid memory addresses // will begin 0x00c0, 0x00c1, ... // In little-endian, that's c0 00, c1 00, ... None of those are valid // UTF-8 sequences, and they are otherwise as far away from // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors // on OS X during thread allocations. 0x00c0 causes conflicts with // AddressSanitizer which reserves all memory up to 0x0100. // These choices reduce the odds of a conservative garbage collector // not collecting memory because some non-pointer block of memory // had a bit pattern that matched a memory address. // // However, on arm64, we ignore all this advice above and slam the // allocation at 0x40 << 32 because when using 4k pages with 3-level // translation buffers, the user address space is limited to 39 bits // On ios/arm64, the address space is even smaller. // // On AIX, mmaps starts at 0x0A00000000000000 for 64-bit. // processes. // // Space mapped for user arenas comes immediately after the range // originally reserved for the regular heap when race mode is not // enabled because user arena chunks can never be used for regular heap // allocations and we want to avoid fragmenting the address space. // // In race mode we have no choice but to just use the same hints because // the race detector requires that the heap be mapped contiguously. for i := 0x7f; i >= 0; i-- { var p uintptr switch { case raceenabled: // The TSAN runtime requires the heap // to be in the range [0x00c000000000, // 0x00e000000000). p = uintptr(i)<<32 | uintptrMask&(0x00c0<<32) if p >= uintptrMask&0x00e000000000 { continue } case GOARCH == "arm64" && GOOS == "ios": p = uintptr(i)<<40 | uintptrMask&(0x0013<<28) case GOARCH == "arm64": p = uintptr(i)<<40 | uintptrMask&(0x0040<<32) case GOOS == "aix": if i == 0 { // We don't use addresses directly after 0x0A00000000000000 // to avoid collisions with others mmaps done by non-go programs. continue } p = uintptr(i)<<40 | uintptrMask&(0xa0<<52) default: p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32) } // Switch to generating hints for user arenas if we've gone // through about half the hints. In race mode, take only about // a quarter; we don't have very much space to work with. hintList := &mheap_.arenaHints if (!raceenabled && i > 0x3f) || (raceenabled && i > 0x5f) { hintList = &mheap_.userArena.arenaHints } hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) hint.addr = p hint.next, *hintList = *hintList, hint } } else { // On a 32-bit machine, we're much more concerned // about keeping the usable heap contiguous. // Hence: // // 1. We reserve space for all heapArenas up front so // they don't get interleaved with the heap. They're // ~258MB, so this isn't too bad. (We could reserve a // smaller amount of space up front if this is a // problem.) // // 2. We hint the heap to start right above the end of // the binary so we have the best chance of keeping it // contiguous. // // 3. We try to stake out a reasonably large initial // heap reservation. const arenaMetaSize = (1 << arenaBits) * unsafe.Sizeof(heapArena{}) meta := uintptr(sysReserve(nil, arenaMetaSize)) if meta != 0 { mheap_.heapArenaAlloc.init(meta, arenaMetaSize, true) } // We want to start the arena low, but if we're linked // against C code, it's possible global constructors // have called malloc and adjusted the process' brk. // Query the brk so we can avoid trying to map the // region over it (which will cause the kernel to put // the region somewhere else, likely at a high // address). procBrk := sbrk0() // If we ask for the end of the data segment but the // operating system requires a little more space // before we can start allocating, it will give out a // slightly higher pointer. Except QEMU, which is // buggy, as usual: it won't adjust the pointer // upward. So adjust it upward a little bit ourselves: // 1/4 MB to get away from the running binary image. p := firstmoduledata.end if p < procBrk { p = procBrk } if mheap_.heapArenaAlloc.next <= p && p < mheap_.heapArenaAlloc.end { p = mheap_.heapArenaAlloc.end } p = alignUp(p+(256<<10), heapArenaBytes) // Because we're worried about fragmentation on // 32-bit, we try to make a large initial reservation. arenaSizes := []uintptr{ 512 << 20, 256 << 20, 128 << 20, } for _, arenaSize := range arenaSizes { a, size := sysReserveAligned(unsafe.Pointer(p), arenaSize, heapArenaBytes) if a != nil { mheap_.arena.init(uintptr(a), size, false) p = mheap_.arena.end // For hint below break } } hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) hint.addr = p hint.next, mheap_.arenaHints = mheap_.arenaHints, hint // Place the hint for user arenas just after the large reservation. // // While this potentially competes with the hint above, in practice we probably // aren't going to be getting this far anyway on 32-bit platforms. userArenaHint := (*arenaHint)(mheap_.arenaHintAlloc.alloc()) userArenaHint.addr = p userArenaHint.next, mheap_.userArena.arenaHints = mheap_.userArena.arenaHints, userArenaHint } // Initialize the memory limit here because the allocator is going to look at it // but we haven't called gcinit yet and we're definitely going to allocate memory before then. gcController.memoryLimit.Store(maxInt64) } // sysAlloc allocates heap arena space for at least n bytes. The // returned pointer is always heapArenaBytes-aligned and backed by // h.arenas metadata. The returned size is always a multiple of // heapArenaBytes. sysAlloc returns nil on failure. // There is no corresponding free function. // // hintList is a list of hint addresses for where to allocate new // heap arenas. It must be non-nil. // // register indicates whether the heap arena should be registered // in allArenas. // // sysAlloc returns a memory region in the Reserved state. This region must // be transitioned to Prepared and then Ready before use. // // h must be locked. func (h *mheap) sysAlloc(n uintptr, hintList **arenaHint, register bool) (v unsafe.Pointer, size uintptr) { assertLockHeld(&h.lock) n = alignUp(n, heapArenaBytes) if hintList == &h.arenaHints { // First, try the arena pre-reservation. // Newly-used mappings are considered released. // // Only do this if we're using the regular heap arena hints. // This behavior is only for the heap. v = h.arena.alloc(n, heapArenaBytes, &gcController.heapReleased) if v != nil { size = n goto mapped } } // Try to grow the heap at a hint address. for *hintList != nil { hint := *hintList p := hint.addr if hint.down { p -= n } if p+n < p { // We can't use this, so don't ask. v = nil } else if arenaIndex(p+n-1) >= 1<= 1<= 1< 0 { sysFreeOS(unsafe.Pointer(end), endLen) } return unsafe.Pointer(pAligned), size } } // enableMetadataHugePages enables huge pages for various sources of heap metadata. // // A note on latency: for sufficiently small heaps (<10s of GiB) this function will take constant // time, but may take time proportional to the size of the mapped heap beyond that. // // This function is idempotent. // // The heap lock must not be held over this operation, since it will briefly acquire // the heap lock. // // Must be called on the system stack because it acquires the heap lock. // //go:systemstack func (h *mheap) enableMetadataHugePages() { // Enable huge pages for page structure. h.pages.enableChunkHugePages() // Grab the lock and set arenasHugePages if it's not. // // Once arenasHugePages is set, all new L2 entries will be eligible for // huge pages. We'll set all the old entries after we release the lock. lock(&h.lock) if h.arenasHugePages { unlock(&h.lock) return } h.arenasHugePages = true unlock(&h.lock) // N.B. The arenas L1 map is quite small on all platforms, so it's fine to // just iterate over the whole thing. for i := range h.arenas { l2 := (*[1 << arenaL2Bits]*heapArena)(atomic.Loadp(unsafe.Pointer(&h.arenas[i]))) if l2 == nil { continue } sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2)) } } // base address for all 0-byte allocations var zerobase uintptr // nextFreeFast returns the next free object if one is quickly available. // Otherwise it returns 0. func nextFreeFast(s *mspan) gclinkptr { theBit := sys.TrailingZeros64(s.allocCache) // Is there a free object in the allocCache? if theBit < 64 { result := s.freeindex + uint16(theBit) if result < s.nelems { freeidx := result + 1 if freeidx%64 == 0 && freeidx != s.nelems { return 0 } s.allocCache >>= uint(theBit + 1) s.freeindex = freeidx s.allocCount++ return gclinkptr(uintptr(result)*s.elemsize + s.base()) } } return 0 } // nextFree returns the next free object from the cached span if one is available. // Otherwise it refills the cache with a span with an available object and // returns that object along with a flag indicating that this was a heavy // weight allocation. If it is a heavy weight allocation the caller must // determine whether a new GC cycle needs to be started or if the GC is active // whether this goroutine needs to assist the GC. // // Must run in a non-preemptible context since otherwise the owner of // c could change. func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) { s = c.alloc[spc] shouldhelpgc = false freeIndex := s.nextFreeIndex() if freeIndex == s.nelems { // The span is full. if s.allocCount != s.nelems { println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems) throw("s.allocCount != s.nelems && freeIndex == s.nelems") } c.refill(spc) shouldhelpgc = true s = c.alloc[spc] freeIndex = s.nextFreeIndex() } if freeIndex >= s.nelems { throw("freeIndex is not valid") } v = gclinkptr(uintptr(freeIndex)*s.elemsize + s.base()) s.allocCount++ if s.allocCount > s.nelems { println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems) throw("s.allocCount > s.nelems") } return } // Allocate an object of size bytes. // Small objects are allocated from the per-P cache's free lists. // Large objects (> 32 kB) are allocated straight from the heap. func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer { if gcphase == _GCmarktermination { throw("mallocgc called with gcphase == _GCmarktermination") } if size == 0 { return unsafe.Pointer(&zerobase) } // It's possible for any malloc to trigger sweeping, which may in // turn queue finalizers. Record this dynamic lock edge. lockRankMayQueueFinalizer() userSize := size if asanenabled { // Refer to ASAN runtime library, the malloc() function allocates extra memory, // the redzone, around the user requested memory region. And the redzones are marked // as unaddressable. We perform the same operations in Go to detect the overflows or // underflows. size += computeRZlog(size) } if debug.malloc { if debug.sbrk != 0 { align := uintptr(16) if typ != nil { // TODO(austin): This should be just // align = uintptr(typ.align) // but that's only 4 on 32-bit platforms, // even if there's a uint64 field in typ (see #599). // This causes 64-bit atomic accesses to panic. // Hence, we use stricter alignment that matches // the normal allocator better. if size&7 == 0 { align = 8 } else if size&3 == 0 { align = 4 } else if size&1 == 0 { align = 2 } else { align = 1 } } return persistentalloc(size, align, &memstats.other_sys) } if inittrace.active && inittrace.id == getg().goid { // Init functions are executed sequentially in a single goroutine. inittrace.allocs += 1 } } // assistG is the G to charge for this allocation, or nil if // GC is not currently active. assistG := deductAssistCredit(size) // Set mp.mallocing to keep from being preempted by GC. mp := acquirem() if mp.mallocing != 0 { throw("malloc deadlock") } if mp.gsignal == getg() { throw("malloc during signal") } mp.mallocing = 1 shouldhelpgc := false dataSize := userSize c := getMCache(mp) if c == nil { throw("mallocgc called without a P or outside bootstrapping") } var span *mspan var header **_type var x unsafe.Pointer noscan := typ == nil || typ.PtrBytes == 0 // In some cases block zeroing can profitably (for latency reduction purposes) // be delayed till preemption is possible; delayedZeroing tracks that state. delayedZeroing := false // Determine if it's a 'small' object that goes into a size-classed span. // // Note: This comparison looks a little strange, but it exists to smooth out // the crossover between the largest size class and large objects that have // their own spans. The small window of object sizes between maxSmallSize-mallocHeaderSize // and maxSmallSize will be considered large, even though they might fit in // a size class. In practice this is completely fine, since the largest small // size class has a single object in it already, precisely to make the transition // to large objects smooth. if size <= maxSmallSize-mallocHeaderSize { if noscan && size < maxTinySize { // Tiny allocator. // // Tiny allocator combines several tiny allocation requests // into a single memory block. The resulting memory block // is freed when all subobjects are unreachable. The subobjects // must be noscan (don't have pointers), this ensures that // the amount of potentially wasted memory is bounded. // // Size of the memory block used for combining (maxTinySize) is tunable. // Current setting is 16 bytes, which relates to 2x worst case memory // wastage (when all but one subobjects are unreachable). // 8 bytes would result in no wastage at all, but provides less // opportunities for combining. // 32 bytes provides more opportunities for combining, // but can lead to 4x worst case wastage. // The best case winning is 8x regardless of block size. // // Objects obtained from tiny allocator must not be freed explicitly. // So when an object will be freed explicitly, we ensure that // its size >= maxTinySize. // // SetFinalizer has a special case for objects potentially coming // from tiny allocator, it such case it allows to set finalizers // for an inner byte of a memory block. // // The main targets of tiny allocator are small strings and // standalone escaping variables. On a json benchmark // the allocator reduces number of allocations by ~12% and // reduces heap size by ~20%. off := c.tinyoffset // Align tiny pointer for required (conservative) alignment. if size&7 == 0 { off = alignUp(off, 8) } else if goarch.PtrSize == 4 && size == 12 { // Conservatively align 12-byte objects to 8 bytes on 32-bit // systems so that objects whose first field is a 64-bit // value is aligned to 8 bytes and does not cause a fault on // atomic access. See issue 37262. // TODO(mknyszek): Remove this workaround if/when issue 36606 // is resolved. off = alignUp(off, 8) } else if size&3 == 0 { off = alignUp(off, 4) } else if size&1 == 0 { off = alignUp(off, 2) } if off+size <= maxTinySize && c.tiny != 0 { // The object fits into existing tiny block. x = unsafe.Pointer(c.tiny + off) c.tinyoffset = off + size c.tinyAllocs++ mp.mallocing = 0 releasem(mp) return x } // Allocate a new maxTinySize block. span = c.alloc[tinySpanClass] v := nextFreeFast(span) if v == 0 { v, span, shouldhelpgc = c.nextFree(tinySpanClass) } x = unsafe.Pointer(v) (*[2]uint64)(x)[0] = 0 (*[2]uint64)(x)[1] = 0 // See if we need to replace the existing tiny block with the new one // based on amount of remaining free space. if !raceenabled && (size < c.tinyoffset || c.tiny == 0) { // Note: disabled when race detector is on, see comment near end of this function. c.tiny = uintptr(x) c.tinyoffset = size } size = maxTinySize } else { hasHeader := !noscan && !heapBitsInSpan(size) if goexperiment.AllocHeaders && hasHeader { size += mallocHeaderSize } var sizeclass uint8 if size <= smallSizeMax-8 { sizeclass = size_to_class8[divRoundUp(size, smallSizeDiv)] } else { sizeclass = size_to_class128[divRoundUp(size-smallSizeMax, largeSizeDiv)] } size = uintptr(class_to_size[sizeclass]) spc := makeSpanClass(sizeclass, noscan) span = c.alloc[spc] v := nextFreeFast(span) if v == 0 { v, span, shouldhelpgc = c.nextFree(spc) } x = unsafe.Pointer(v) if needzero && span.needzero != 0 { memclrNoHeapPointers(x, size) } if goexperiment.AllocHeaders && hasHeader { header = (**_type)(x) x = add(x, mallocHeaderSize) size -= mallocHeaderSize } } } else { shouldhelpgc = true // For large allocations, keep track of zeroed state so that // bulk zeroing can be happen later in a preemptible context. span = c.allocLarge(size, noscan) span.freeindex = 1 span.allocCount = 1 size = span.elemsize x = unsafe.Pointer(span.base()) if needzero && span.needzero != 0 { if noscan { delayedZeroing = true } else { memclrNoHeapPointers(x, size) } } if goexperiment.AllocHeaders && !noscan { header = &span.largeType } } if !noscan { if goexperiment.AllocHeaders { c.scanAlloc += heapSetType(uintptr(x), dataSize, typ, header, span) } else { var scanSize uintptr heapBitsSetType(uintptr(x), size, dataSize, typ) if dataSize > typ.Size_ { // Array allocation. If there are any // pointers, GC has to scan to the last // element. if typ.PtrBytes != 0 { scanSize = dataSize - typ.Size_ + typ.PtrBytes } } else { scanSize = typ.PtrBytes } c.scanAlloc += scanSize } } // Ensure that the stores above that initialize x to // type-safe memory and set the heap bits occur before // the caller can make x observable to the garbage // collector. Otherwise, on weakly ordered machines, // the garbage collector could follow a pointer to x, // but see uninitialized memory or stale heap bits. publicationBarrier() // As x and the heap bits are initialized, update // freeIndexForScan now so x is seen by the GC // (including conservative scan) as an allocated object. // While this pointer can't escape into user code as a // _live_ pointer until we return, conservative scanning // may find a dead pointer that happens to point into this // object. Delaying this update until now ensures that // conservative scanning considers this pointer dead until // this point. span.freeIndexForScan = span.freeindex // Allocate black during GC. // All slots hold nil so no scanning is needed. // This may be racing with GC so do it atomically if there can be // a race marking the bit. if gcphase != _GCoff { gcmarknewobject(span, uintptr(x)) } if raceenabled { racemalloc(x, size) } if msanenabled { msanmalloc(x, size) } if asanenabled { // We should only read/write the memory with the size asked by the user. // The rest of the allocated memory should be poisoned, so that we can report // errors when accessing poisoned memory. // The allocated memory is larger than required userSize, it will also include // redzone and some other padding bytes. rzBeg := unsafe.Add(x, userSize) asanpoison(rzBeg, size-userSize) asanunpoison(x, userSize) } // If !goexperiment.AllocHeaders, "size" doesn't include the // allocation header, so use span.elemsize as the "full" size // for various computations below. // // TODO(mknyszek): We should really count the header as part // of gc_sys or something, but it's risky to change the // accounting so much right now. Just pretend its internal // fragmentation and match the GC's accounting by using the // whole allocation slot. fullSize := size if goexperiment.AllocHeaders { fullSize = span.elemsize } if rate := MemProfileRate; rate > 0 { // Note cache c only valid while m acquired; see #47302 // // N.B. Use the full size because that matches how the GC // will update the mem profile on the "free" side. if rate != 1 && fullSize < c.nextSample { c.nextSample -= fullSize } else { profilealloc(mp, x, fullSize) } } mp.mallocing = 0 releasem(mp) // Pointerfree data can be zeroed late in a context where preemption can occur. // x will keep the memory alive. if delayedZeroing { if !noscan { throw("delayed zeroing on data that may contain pointers") } if goexperiment.AllocHeaders && header != nil { throw("unexpected malloc header in delayed zeroing of large object") } // N.B. size == fullSize always in this case. memclrNoHeapPointersChunked(size, x) // This is a possible preemption point: see #47302 } if debug.malloc { if debug.allocfreetrace != 0 { tracealloc(x, size, typ) } if inittrace.active && inittrace.id == getg().goid { // Init functions are executed sequentially in a single goroutine. inittrace.bytes += uint64(fullSize) } } if assistG != nil { // Account for internal fragmentation in the assist // debt now that we know it. // // N.B. Use the full size because that's how the rest // of the GC accounts for bytes marked. assistG.gcAssistBytes -= int64(fullSize - dataSize) } if shouldhelpgc { if t := (gcTrigger{kind: gcTriggerHeap}); t.test() { gcStart(t) } } if raceenabled && noscan && dataSize < maxTinySize { // Pad tinysize allocations so they are aligned with the end // of the tinyalloc region. This ensures that any arithmetic // that goes off the top end of the object will be detectable // by checkptr (issue 38872). // Note that we disable tinyalloc when raceenabled for this to work. // TODO: This padding is only performed when the race detector // is enabled. It would be nice to enable it if any package // was compiled with checkptr, but there's no easy way to // detect that (especially at compile time). // TODO: enable this padding for all allocations, not just // tinyalloc ones. It's tricky because of pointer maps. // Maybe just all noscan objects? x = add(x, size-dataSize) } return x } // deductAssistCredit reduces the current G's assist credit // by size bytes, and assists the GC if necessary. // // Caller must be preemptible. // // Returns the G for which the assist credit was accounted. func deductAssistCredit(size uintptr) *g { var assistG *g if gcBlackenEnabled != 0 { // Charge the current user G for this allocation. assistG = getg() if assistG.m.curg != nil { assistG = assistG.m.curg } // Charge the allocation against the G. We'll account // for internal fragmentation at the end of mallocgc. assistG.gcAssistBytes -= int64(size) if assistG.gcAssistBytes < 0 { // This G is in debt. Assist the GC to correct // this before allocating. This must happen // before disabling preemption. gcAssistAlloc(assistG) } } return assistG } // memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers // on chunks of the buffer to be zeroed, with opportunities for preemption // along the way. memclrNoHeapPointers contains no safepoints and also // cannot be preemptively scheduled, so this provides a still-efficient // block copy that can also be preempted on a reasonable granularity. // // Use this with care; if the data being cleared is tagged to contain // pointers, this allows the GC to run before it is all cleared. func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) { v := uintptr(x) // got this from benchmarking. 128k is too small, 512k is too large. const chunkBytes = 256 * 1024 vsize := v + size for voff := v; voff < vsize; voff = voff + chunkBytes { if getg().preempt { // may hold locks, e.g., profiling goschedguarded() } // clear min(avail, lump) bytes n := vsize - voff if n > chunkBytes { n = chunkBytes } memclrNoHeapPointers(unsafe.Pointer(voff), n) } } // implementation of new builtin // compiler (both frontend and SSA backend) knows the signature // of this function. func newobject(typ *_type) unsafe.Pointer { return mallocgc(typ.Size_, typ, true) } //go:linkname reflect_unsafe_New reflect.unsafe_New func reflect_unsafe_New(typ *_type) unsafe.Pointer { return mallocgc(typ.Size_, typ, true) } //go:linkname reflectlite_unsafe_New internal/reflectlite.unsafe_New func reflectlite_unsafe_New(typ *_type) unsafe.Pointer { return mallocgc(typ.Size_, typ, true) } // newarray allocates an array of n elements of type typ. func newarray(typ *_type, n int) unsafe.Pointer { if n == 1 { return mallocgc(typ.Size_, typ, true) } mem, overflow := math.MulUintptr(typ.Size_, uintptr(n)) if overflow || mem > maxAlloc || n < 0 { panic(plainError("runtime: allocation size out of range")) } return mallocgc(mem, typ, true) } //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer { return newarray(typ, n) } func profilealloc(mp *m, x unsafe.Pointer, size uintptr) { c := getMCache(mp) if c == nil { throw("profilealloc called without a P or outside bootstrapping") } c.nextSample = nextSample() mProf_Malloc(x, size) } // nextSample returns the next sampling point for heap profiling. The goal is // to sample allocations on average every MemProfileRate bytes, but with a // completely random distribution over the allocation timeline; this // corresponds to a Poisson process with parameter MemProfileRate. In Poisson // processes, the distance between two samples follows the exponential // distribution (exp(MemProfileRate)), so the best return value is a random // number taken from an exponential distribution whose mean is MemProfileRate. func nextSample() uintptr { if MemProfileRate == 1 { // Callers assign our return value to // mcache.next_sample, but next_sample is not used // when the rate is 1. So avoid the math below and // just return something. return 0 } if GOOS == "plan9" { // Plan 9 doesn't support floating point in note handler. if gp := getg(); gp == gp.m.gsignal { return nextSampleNoFP() } } return uintptr(fastexprand(MemProfileRate)) } // fastexprand returns a random number from an exponential distribution with // the specified mean. func fastexprand(mean int) int32 { // Avoid overflow. Maximum possible step is // -ln(1/(1< 0x7000000: mean = 0x7000000 case mean == 0: return 0 } // Take a random sample of the exponential distribution exp(-mean*x). // The probability distribution function is mean*exp(-mean*x), so the CDF is // p = 1 - exp(-mean*x), so // q = 1 - p == exp(-mean*x) // log_e(q) = -mean*x // -log_e(q)/mean = x // x = -log_e(q) * mean // x = log_2(q) * (-log_e(2)) * mean ; Using log_2 for efficiency const randomBitCount = 26 q := cheaprandn(1< 0 { qlog = 0 } const minusLog2 = -0.6931471805599453 // -ln(2) return int32(qlog*(minusLog2*float64(mean))) + 1 } // nextSampleNoFP is similar to nextSample, but uses older, // simpler code to avoid floating point. func nextSampleNoFP() uintptr { // Set first allocation sample size. rate := MemProfileRate if rate > 0x3fffffff { // make 2*rate not overflow rate = 0x3fffffff } if rate != 0 { return uintptr(cheaprandn(uint32(2 * rate))) } return 0 } type persistentAlloc struct { base *notInHeap off uintptr } var globalAlloc struct { mutex persistentAlloc } // persistentChunkSize is the number of bytes we allocate when we grow // a persistentAlloc. const persistentChunkSize = 256 << 10 // persistentChunks is a list of all the persistent chunks we have // allocated. The list is maintained through the first word in the // persistent chunk. This is updated atomically. var persistentChunks *notInHeap // Wrapper around sysAlloc that can allocate small chunks. // There is no associated free operation. // Intended for things like function/type/debug-related persistent data. // If align is 0, uses default align (currently 8). // The returned memory will be zeroed. // sysStat must be non-nil. // // Consider marking persistentalloc'd types not in heap by embedding // runtime/internal/sys.NotInHeap. func persistentalloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer { var p *notInHeap systemstack(func() { p = persistentalloc1(size, align, sysStat) }) return unsafe.Pointer(p) } // Must run on system stack because stack growth can (re)invoke it. // See issue 9174. // //go:systemstack func persistentalloc1(size, align uintptr, sysStat *sysMemStat) *notInHeap { const ( maxBlock = 64 << 10 // VM reservation granularity is 64K on windows ) if size == 0 { throw("persistentalloc: size == 0") } if align != 0 { if align&(align-1) != 0 { throw("persistentalloc: align is not a power of 2") } if align > _PageSize { throw("persistentalloc: align is too large") } } else { align = 8 } if size >= maxBlock { return (*notInHeap)(sysAlloc(size, sysStat)) } mp := acquirem() var persistent *persistentAlloc if mp != nil && mp.p != 0 { persistent = &mp.p.ptr().palloc } else { lock(&globalAlloc.mutex) persistent = &globalAlloc.persistentAlloc } persistent.off = alignUp(persistent.off, align) if persistent.off+size > persistentChunkSize || persistent.base == nil { persistent.base = (*notInHeap)(sysAlloc(persistentChunkSize, &memstats.other_sys)) if persistent.base == nil { if persistent == &globalAlloc.persistentAlloc { unlock(&globalAlloc.mutex) } throw("runtime: cannot allocate memory") } // Add the new chunk to the persistentChunks list. for { chunks := uintptr(unsafe.Pointer(persistentChunks)) *(*uintptr)(unsafe.Pointer(persistent.base)) = chunks if atomic.Casuintptr((*uintptr)(unsafe.Pointer(&persistentChunks)), chunks, uintptr(unsafe.Pointer(persistent.base))) { break } } persistent.off = alignUp(goarch.PtrSize, align) } p := persistent.base.add(persistent.off) persistent.off += size releasem(mp) if persistent == &globalAlloc.persistentAlloc { unlock(&globalAlloc.mutex) } if sysStat != &memstats.other_sys { sysStat.add(int64(size)) memstats.other_sys.add(-int64(size)) } return p } // inPersistentAlloc reports whether p points to memory allocated by // persistentalloc. This must be nosplit because it is called by the // cgo checker code, which is called by the write barrier code. // //go:nosplit func inPersistentAlloc(p uintptr) bool { chunk := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&persistentChunks))) for chunk != 0 { if p >= chunk && p < chunk+persistentChunkSize { return true } chunk = *(*uintptr)(unsafe.Pointer(chunk)) } return false } // linearAlloc is a simple linear allocator that pre-reserves a region // of memory and then optionally maps that region into the Ready state // as needed. // // The caller is responsible for locking. type linearAlloc struct { next uintptr // next free byte mapped uintptr // one byte past end of mapped space end uintptr // end of reserved space mapMemory bool // transition memory from Reserved to Ready if true } func (l *linearAlloc) init(base, size uintptr, mapMemory bool) { if base+size < base { // Chop off the last byte. The runtime isn't prepared // to deal with situations where the bounds could overflow. // Leave that memory reserved, though, so we don't map it // later. size -= 1 } l.next, l.mapped = base, base l.end = base + size l.mapMemory = mapMemory } func (l *linearAlloc) alloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer { p := alignUp(l.next, align) if p+size > l.end { return nil } l.next = p + size if pEnd := alignUp(l.next-1, physPageSize); pEnd > l.mapped { if l.mapMemory { // Transition from Reserved to Prepared to Ready. n := pEnd - l.mapped sysMap(unsafe.Pointer(l.mapped), n, sysStat) sysUsed(unsafe.Pointer(l.mapped), n, n) } l.mapped = pEnd } return unsafe.Pointer(p) } // notInHeap is off-heap memory allocated by a lower-level allocator // like sysAlloc or persistentAlloc. // // In general, it's better to use real types which embed // runtime/internal/sys.NotInHeap, but this serves as a generic type // for situations where that isn't possible (like in the allocators). // // TODO: Use this as the return type of sysAlloc, persistentAlloc, etc? type notInHeap struct{ _ sys.NotInHeap } func (p *notInHeap) add(bytes uintptr) *notInHeap { return (*notInHeap)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + bytes)) } // computeRZlog computes the size of the redzone. // Refer to the implementation of the compiler-rt. func computeRZlog(userSize uintptr) uintptr { switch { case userSize <= (64 - 16): return 16 << 0 case userSize <= (128 - 32): return 16 << 1 case userSize <= (512 - 64): return 16 << 2 case userSize <= (4096 - 128): return 16 << 3 case userSize <= (1<<14)-256: return 16 << 4 case userSize <= (1<<15)-512: return 16 << 5 case userSize <= (1<<16)-1024: return 16 << 6 default: return 16 << 7 } }