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.
5 // Memory allocator, based on tcmalloc.
6 // http://goog-perftools.sourceforge.net/doc/tcmalloc.html
8 // The main allocator works in runs of pages.
9 // Small allocation sizes (up to and including 32 kB) are
10 // rounded to one of about 100 size classes, each of which
11 // has its own free list of objects of exactly that size.
12 // Any free page of memory can be split into a set of objects
13 // of one size class, which are then managed using free list
16 // The allocator's data structures are:
18 // FixAlloc: a free-list allocator for fixed-size objects,
19 // used to manage storage used by the allocator.
20 // MHeap: the malloc heap, managed at page (4096-byte) granularity.
21 // MSpan: a run of pages managed by the MHeap.
22 // MCentral: a shared free list for a given size class.
23 // MCache: a per-thread (in Go, per-P) cache for small objects.
24 // MStats: allocation statistics.
26 // Allocating a small object proceeds up a hierarchy of caches:
28 // 1. Round the size up to one of the small size classes
29 // and look in the corresponding MCache free list.
30 // If the list is not empty, allocate an object from it.
31 // This can all be done without acquiring a lock.
33 // 2. If the MCache free list is empty, replenish it by
34 // taking a bunch of objects from the MCentral free list.
35 // Moving a bunch amortizes the cost of acquiring the MCentral lock.
37 // 3. If the MCentral free list is empty, replenish it by
38 // allocating a run of pages from the MHeap and then
39 // chopping that memory into objects of the given size.
40 // Allocating many objects amortizes the cost of locking
43 // 4. If the MHeap is empty or has no page runs large enough,
44 // allocate a new group of pages (at least 1MB) from the
45 // operating system. Allocating a large run of pages
46 // amortizes the cost of talking to the operating system.
48 // Freeing a small object proceeds up the same hierarchy:
50 // 1. Look up the size class for the object and add it to
51 // the MCache free list.
53 // 2. If the MCache free list is too long or the MCache has
54 // too much memory, return some to the MCentral free lists.
56 // 3. If all the objects in a given span have returned to
57 // the MCentral list, return that span to the page heap.
59 // 4. If the heap has too much memory, return some to the
62 // TODO(rsc): Step 4 is not implemented.
64 // Allocating and freeing a large object uses the page heap
65 // directly, bypassing the MCache and MCentral free lists.
67 // The small objects on the MCache and MCentral free lists
68 // may or may not be zeroed. They are zeroed if and only if
69 // the second word of the object is zero. A span in the
70 // page heap is zeroed unless s->needzero is set. When a span
71 // is allocated to break into small objects, it is zeroed if needed
72 // and s->needzero is set. There are two main benefits to delaying the
75 // 1. stack frames allocated from the small object lists
76 // or the page heap can avoid zeroing altogether.
77 // 2. the cost of zeroing when reusing a small object is
78 // charged to the mutator, not the garbage collector.
80 // This code was written with an eye toward translating to Go
81 // in the future. Methods have the form Type_Method(Type *t, ...).
90 flagNoScan = _FlagNoScan
91 flagNoZero = _FlagNoZero
93 maxTinySize = _TinySize
94 tinySizeClass = _TinySizeClass
95 maxSmallSize = _MaxSmallSize
97 pageShift = _PageShift
101 mSpanInUse = _MSpanInUse
103 concurrentSweep = _ConcurrentSweep
108 _PageSize = 1 << _PageShift
109 _PageMask = _PageSize - 1
113 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems
114 _64bit = 1 << (^uintptr(0) >> 63) / 2
116 // Computed constant. The definition of MaxSmallSize and the
117 // algorithm in msize.go produces some number of different allocation
118 // size classes. NumSizeClasses is that number. It's needed here
119 // because there are static arrays of this length; when msize runs its
120 // size choosing algorithm it double-checks that NumSizeClasses agrees.
123 // Tunable constants.
124 _MaxSmallSize = 32 << 10
126 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go.
130 _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc
131 _MaxMHeapList = 1 << (20 - _PageShift) // Maximum page length for fixed-size list in MHeap.
132 _HeapAllocChunk = 1 << 20 // Chunk size for heap growth
134 // Per-P, per order stack segment cache size.
135 _StackCacheSize = 32 * 1024
137 // Number of orders that get caching. Order 0 is FixedStack
138 // and each successive order is twice as large.
139 // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks
140 // will be allocated directly.
141 // Since FixedStack is different on different systems, we
142 // must vary NumStackOrders to keep the same maximum cached size.
143 // OS | FixedStack | NumStackOrders
144 // -----------------+------------+---------------
145 // linux/darwin/bsd | 2KB | 4
146 // windows/32 | 4KB | 3
147 // windows/64 | 8KB | 2
149 _NumStackOrders = 4 - ptrSize/4*goos_windows - 1*goos_plan9
151 // Number of bits in page to span calculations (4k pages).
152 // On Windows 64-bit we limit the arena to 32GB or 35 bits.
153 // Windows counts memory used by page table into committed memory
154 // of the process, so we can't reserve too much memory.
155 // See https://golang.org/issue/5402 and https://golang.org/issue/5236.
156 // On other 64-bit platforms, we limit the arena to 512GB, or 39 bits.
157 // On 32-bit, we don't bother limiting anything, so we use the full 32-bit address.
158 // On Darwin/arm64, we cannot reserve more than ~5GB of virtual memory,
159 // but as most devices have less than 4GB of physical memory anyway, we
160 // try to be conservative here, and only ask for a 2GB heap.
161 _MHeapMap_TotalBits = (_64bit*goos_windows)*35 + (_64bit*(1-goos_windows)*(1-goos_darwin*goarch_arm64))*39 + goos_darwin*goarch_arm64*31 + (1-_64bit)*32
162 _MHeapMap_Bits = _MHeapMap_TotalBits - _PageShift
164 _MaxMem = uintptr(1<<_MHeapMap_TotalBits - 1)
166 // Max number of threads to run garbage collection.
167 // 2, 3, and 4 are all plausible maximums depending
168 // on the hardware details of the machine. The garbage
169 // collector scales well to 32 cpus.
173 // Page number (address>>pageShift)
176 const _MaxArena32 = 2 << 30
178 // OS-defined helpers:
180 // sysAlloc obtains a large chunk of zeroed memory from the
181 // operating system, typically on the order of a hundred kilobytes
183 // NOTE: sysAlloc returns OS-aligned memory, but the heap allocator
184 // may use larger alignment, so the caller must be careful to realign the
185 // memory obtained by sysAlloc.
187 // SysUnused notifies the operating system that the contents
188 // of the memory region are no longer needed and can be reused
189 // for other purposes.
190 // SysUsed notifies the operating system that the contents
191 // of the memory region are needed again.
193 // SysFree returns it unconditionally; this is only used if
194 // an out-of-memory error has been detected midway through
195 // an allocation. It is okay if SysFree is a no-op.
197 // SysReserve reserves address space without allocating memory.
198 // If the pointer passed to it is non-nil, the caller wants the
199 // reservation there, but SysReserve can still choose another
200 // location if that one is unavailable. On some systems and in some
201 // cases SysReserve will simply check that the address space is
202 // available and not actually reserve it. If SysReserve returns
203 // non-nil, it sets *reserved to true if the address space is
204 // reserved, false if it has merely been checked.
205 // NOTE: SysReserve returns OS-aligned memory, but the heap allocator
206 // may use larger alignment, so the caller must be careful to realign the
207 // memory obtained by sysAlloc.
209 // SysMap maps previously reserved address space for use.
210 // The reserved argument is true if the address space was really
211 // reserved, not merely checked.
213 // SysFault marks a (already sysAlloc'd) region to fault
214 // if accessed. Used only for debugging the runtime.
219 if class_to_size[_TinySizeClass] != _TinySize {
220 throw("bad TinySizeClass")
223 var p, bitmapSize, spansSize, pSize, limit uintptr
226 // limit = runtime.memlimit();
227 // See https://golang.org/issue/5049
228 // TODO(rsc): Fix after 1.1.
231 // Set up the allocation arena, a contiguous area of memory where
232 // allocated data will be found. The arena begins with a bitmap large
233 // enough to hold 4 bits per allocated word.
234 if ptrSize == 8 && (limit == 0 || limit > 1<<30) {
235 // On a 64-bit machine, allocate from a single contiguous reservation.
236 // 512 GB (MaxMem) should be big enough for now.
238 // The code will work with the reservation at any address, but ask
239 // SysReserve to use 0x0000XXc000000000 if possible (XX=00...7f).
240 // Allocating a 512 GB region takes away 39 bits, and the amd64
241 // doesn't let us choose the top 17 bits, so that leaves the 9 bits
242 // in the middle of 0x00c0 for us to choose. Choosing 0x00c0 means
243 // that the valid memory addresses will begin 0x00c0, 0x00c1, ..., 0x00df.
244 // In little-endian, that's c0 00, c1 00, ..., df 00. None of those are valid
245 // UTF-8 sequences, and they are otherwise as far away from
246 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0
247 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors
248 // on OS X during thread allocations. 0x00c0 causes conflicts with
249 // AddressSanitizer which reserves all memory up to 0x0100.
250 // These choices are both for debuggability and to reduce the
251 // odds of a conservative garbage collector (as is still used in gccgo)
252 // not collecting memory because some non-pointer block of memory
253 // had a bit pattern that matched a memory address.
255 // Actually we reserve 544 GB (because the bitmap ends up being 32 GB)
256 // but it hardly matters: e0 00 is not valid UTF-8 either.
258 // If this fails we fall back to the 32 bit memory mechanism
260 // However, on arm64, we ignore all this advice above and slam the
261 // allocation at 0x40 << 32 because when using 4k pages with 3-level
262 // translation buffers, the user address space is limited to 39 bits
263 // On darwin/arm64, the address space is even smaller.
264 arenaSize := round(_MaxMem, _PageSize)
265 bitmapSize = arenaSize / (ptrSize * 8 / 4)
266 spansSize = arenaSize / _PageSize * ptrSize
267 spansSize = round(spansSize, _PageSize)
268 for i := 0; i <= 0x7f; i++ {
270 case GOARCH == "arm64" && GOOS == "darwin":
271 p = uintptr(i)<<40 | uintptrMask&(0x0013<<28)
272 case GOARCH == "arm64":
273 p = uintptr(i)<<40 | uintptrMask&(0x0040<<32)
275 p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32)
277 pSize = bitmapSize + spansSize + arenaSize + _PageSize
278 p = uintptr(sysReserve(unsafe.Pointer(p), pSize, &reserved))
286 // On a 32-bit machine, we can't typically get away
287 // with a giant virtual address space reservation.
288 // Instead we map the memory information bitmap
289 // immediately after the data segment, large enough
290 // to handle another 2GB of mappings (256 MB),
291 // along with a reservation for an initial arena.
292 // When that gets used up, we'll start asking the kernel
293 // for any memory anywhere and hope it's in the 2GB
294 // following the bitmap (presumably the executable begins
295 // near the bottom of memory, so we'll have to use up
296 // most of memory before the kernel resorts to giving out
297 // memory before the beginning of the text segment).
299 // Alternatively we could reserve 512 MB bitmap, enough
300 // for 4GB of mappings, and then accept any memory the
301 // kernel threw at us, but normally that's a waste of 512 MB
302 // of address space, which is probably too much in a 32-bit world.
304 // If we fail to allocate, try again with a smaller arena.
305 // This is necessary on Android L where we share a process
306 // with ART, which reserves virtual memory aggressively.
307 arenaSizes := []uintptr{
313 for _, arenaSize := range arenaSizes {
314 bitmapSize = _MaxArena32 / (ptrSize * 8 / 4)
315 spansSize = _MaxArena32 / _PageSize * ptrSize
316 if limit > 0 && arenaSize+bitmapSize+spansSize > limit {
317 bitmapSize = (limit / 9) &^ ((1 << _PageShift) - 1)
318 arenaSize = bitmapSize * 8
319 spansSize = arenaSize / _PageSize * ptrSize
321 spansSize = round(spansSize, _PageSize)
323 // SysReserve treats the address we ask for, end, as a hint,
324 // not as an absolute requirement. If we ask for the end
325 // of the data segment but the operating system requires
326 // a little more space before we can start allocating, it will
327 // give out a slightly higher pointer. Except QEMU, which
328 // is buggy, as usual: it won't adjust the pointer upward.
329 // So adjust it upward a little bit ourselves: 1/4 MB to get
330 // away from the running binary image and then round up
332 p = round(firstmoduledata.end+(1<<18), 1<<20)
333 pSize = bitmapSize + spansSize + arenaSize + _PageSize
334 p = uintptr(sysReserve(unsafe.Pointer(p), pSize, &reserved))
340 throw("runtime: cannot reserve arena virtual address space")
344 // PageSize can be larger than OS definition of page size,
345 // so SysReserve can give us a PageSize-unaligned pointer.
346 // To overcome this we ask for PageSize more and round up the pointer.
347 p1 := round(p, _PageSize)
349 mheap_.spans = (**mspan)(unsafe.Pointer(p1))
350 mheap_.bitmap = p1 + spansSize
351 mheap_.arena_start = p1 + (spansSize + bitmapSize)
352 mheap_.arena_used = mheap_.arena_start
353 mheap_.arena_end = p + pSize
354 mheap_.arena_reserved = reserved
356 if mheap_.arena_start&(_PageSize-1) != 0 {
357 println("bad pagesize", hex(p), hex(p1), hex(spansSize), hex(bitmapSize), hex(_PageSize), "start", hex(mheap_.arena_start))
358 throw("misrounded allocation in mallocinit")
361 // Initialize the rest of the allocator.
362 mHeap_Init(&mheap_, spansSize)
364 _g_.m.mcache = allocmcache()
367 // sysReserveHigh reserves space somewhere high in the address space.
368 // sysReserve doesn't actually reserve the full amount requested on
369 // 64-bit systems, because of problems with ulimit. Instead it checks
370 // that it can get the first 64 kB and assumes it can grab the rest as
371 // needed. This doesn't work well with the "let the kernel pick an address"
372 // mode, so don't do that. Pick a high address instead.
373 func sysReserveHigh(n uintptr, reserved *bool) unsafe.Pointer {
375 return sysReserve(nil, n, reserved)
378 for i := 0; i <= 0x7f; i++ {
379 p := uintptr(i)<<40 | uintptrMask&(0x00c0<<32)
381 p = uintptr(sysReserve(unsafe.Pointer(p), n, reserved))
383 return unsafe.Pointer(p)
387 return sysReserve(nil, n, reserved)
390 func mHeap_SysAlloc(h *mheap, n uintptr) unsafe.Pointer {
391 if n > uintptr(h.arena_end)-uintptr(h.arena_used) {
392 // We are in 32-bit mode, maybe we didn't use all possible address space yet.
393 // Reserve some more space.
394 p_size := round(n+_PageSize, 256<<20)
395 new_end := h.arena_end + p_size
396 if new_end <= h.arena_start+_MaxArena32 {
397 // TODO: It would be bad if part of the arena
398 // is reserved and part is not.
400 p := uintptr(sysReserve(unsafe.Pointer(h.arena_end), p_size, &reserved))
404 if p == h.arena_end {
405 h.arena_end = new_end
406 h.arena_reserved = reserved
407 } else if p+p_size <= h.arena_start+_MaxArena32 {
408 // Keep everything page-aligned.
409 // Our pages are bigger than hardware pages.
410 h.arena_end = p + p_size
411 used := p + (-uintptr(p) & (_PageSize - 1))
412 mHeap_MapBits(h, used)
413 mHeap_MapSpans(h, used)
415 h.arena_reserved = reserved
418 sysFree(unsafe.Pointer(p), p_size, &stat)
423 if n <= uintptr(h.arena_end)-uintptr(h.arena_used) {
424 // Keep taking from our reservation.
426 sysMap(unsafe.Pointer(p), n, h.arena_reserved, &memstats.heap_sys)
427 mHeap_MapBits(h, p+n)
428 mHeap_MapSpans(h, p+n)
431 racemapshadow(unsafe.Pointer(p), n)
434 if uintptr(p)&(_PageSize-1) != 0 {
435 throw("misrounded allocation in MHeap_SysAlloc")
437 return unsafe.Pointer(p)
440 // If using 64-bit, our reservation is all we have.
441 if uintptr(h.arena_end)-uintptr(h.arena_start) >= _MaxArena32 {
445 // On 32-bit, once the reservation is gone we can
446 // try to get memory at a location chosen by the OS
447 // and hope that it is in the range we allocated bitmap for.
448 p_size := round(n, _PageSize) + _PageSize
449 p := uintptr(sysAlloc(p_size, &memstats.heap_sys))
454 if p < h.arena_start || uintptr(p)+p_size-uintptr(h.arena_start) >= _MaxArena32 {
455 print("runtime: memory allocated by OS (", p, ") not in usable range [", hex(h.arena_start), ",", hex(h.arena_start+_MaxArena32), ")\n")
456 sysFree(unsafe.Pointer(p), p_size, &memstats.heap_sys)
461 p += -p & (_PageSize - 1)
462 if uintptr(p)+n > uintptr(h.arena_used) {
463 mHeap_MapBits(h, p+n)
464 mHeap_MapSpans(h, p+n)
466 if p_end > h.arena_end {
470 racemapshadow(unsafe.Pointer(p), n)
474 if uintptr(p)&(_PageSize-1) != 0 {
475 throw("misrounded allocation in MHeap_SysAlloc")
477 return unsafe.Pointer(p)
480 // base address for all 0-byte allocations
485 _FlagNoScan = 1 << 0 // GC doesn't have to scan object
486 _FlagNoZero = 1 << 1 // don't zero memory
489 // Allocate an object of size bytes.
490 // Small objects are allocated from the per-P cache's free lists.
491 // Large objects (> 32 kB) are allocated straight from the heap.
492 func mallocgc(size uintptr, typ *_type, flags uint32) unsafe.Pointer {
493 if gcphase == _GCmarktermination {
494 throw("mallocgc called with gcphase == _GCmarktermination")
498 return unsafe.Pointer(&zerobase)
501 if flags&flagNoScan == 0 && typ == nil {
502 throw("malloc missing type")
508 align = uintptr(typ.align)
510 return persistentalloc(size, align, &memstats.other_sys)
513 // assistG is the G to charge for this allocation, or nil if
514 // GC is not currently active.
516 if gcBlackenEnabled != 0 {
517 // Charge the current user G for this allocation.
519 if assistG.m.curg != nil {
520 assistG = assistG.m.curg
522 // Charge the allocation against the G. We'll account
523 // for internal fragmentation at the end of mallocgc.
524 assistG.gcAssistBytes -= int64(size)
526 if assistG.gcAssistBytes < 0 {
527 // This G is in debt. Assist the GC to correct
528 // this before allocating. This must happen
529 // before disabling preemption.
530 gcAssistAlloc(assistG)
534 // Set mp.mallocing to keep from being preempted by GC.
536 if mp.mallocing != 0 {
537 throw("malloc deadlock")
539 if mp.gsignal == getg() {
540 throw("malloc during signal")
544 shouldhelpgc := false
549 if size <= maxSmallSize {
550 if flags&flagNoScan != 0 && size < maxTinySize {
553 // Tiny allocator combines several tiny allocation requests
554 // into a single memory block. The resulting memory block
555 // is freed when all subobjects are unreachable. The subobjects
556 // must be FlagNoScan (don't have pointers), this ensures that
557 // the amount of potentially wasted memory is bounded.
559 // Size of the memory block used for combining (maxTinySize) is tunable.
560 // Current setting is 16 bytes, which relates to 2x worst case memory
561 // wastage (when all but one subobjects are unreachable).
562 // 8 bytes would result in no wastage at all, but provides less
563 // opportunities for combining.
564 // 32 bytes provides more opportunities for combining,
565 // but can lead to 4x worst case wastage.
566 // The best case winning is 8x regardless of block size.
568 // Objects obtained from tiny allocator must not be freed explicitly.
569 // So when an object will be freed explicitly, we ensure that
570 // its size >= maxTinySize.
572 // SetFinalizer has a special case for objects potentially coming
573 // from tiny allocator, it such case it allows to set finalizers
574 // for an inner byte of a memory block.
576 // The main targets of tiny allocator are small strings and
577 // standalone escaping variables. On a json benchmark
578 // the allocator reduces number of allocations by ~12% and
579 // reduces heap size by ~20%.
581 // Align tiny pointer for required (conservative) alignment.
584 } else if size&3 == 0 {
586 } else if size&1 == 0 {
589 if off+size <= maxTinySize && c.tiny != nil {
590 // The object fits into existing tiny block.
592 c.tinyoffset = off + size
598 // Allocate a new maxTinySize block.
599 s = c.alloc[tinySizeClass]
603 mCache_Refill(c, tinySizeClass)
606 s = c.alloc[tinySizeClass]
609 s.freelist = v.ptr().next
611 // prefetchnta offers best performance, see change list message.
612 prefetchnta(uintptr(v.ptr().next))
613 x = unsafe.Pointer(v)
614 (*[2]uint64)(x)[0] = 0
615 (*[2]uint64)(x)[1] = 0
616 // See if we need to replace the existing tiny block with the new one
617 // based on amount of remaining free space.
618 if size < c.tinyoffset {
626 sizeclass = size_to_class8[(size+7)>>3]
628 sizeclass = size_to_class128[(size-1024+127)>>7]
630 size = uintptr(class_to_size[sizeclass])
631 s = c.alloc[sizeclass]
635 mCache_Refill(c, int32(sizeclass))
638 s = c.alloc[sizeclass]
641 s.freelist = v.ptr().next
643 // prefetchnta offers best performance, see change list message.
644 prefetchnta(uintptr(v.ptr().next))
645 x = unsafe.Pointer(v)
646 if flags&flagNoZero == 0 {
648 if size > 2*ptrSize && ((*[2]uintptr)(x))[1] != 0 {
649 memclr(unsafe.Pointer(v), size)
653 c.local_cachealloc += size
658 s = largeAlloc(size, uint32(flags))
660 x = unsafe.Pointer(uintptr(s.start << pageShift))
661 size = uintptr(s.elemsize)
664 if flags&flagNoScan != 0 {
665 // All objects are pre-marked as noscan. Nothing to do.
667 // If allocating a defer+arg block, now that we've picked a malloc size
668 // large enough to hold everything, cut the "asked for" size down to
669 // just the defer header, so that the GC bitmap will record the arg block
670 // as containing nothing at all (as if it were unused space at the end of
671 // a malloc block caused by size rounding).
672 // The defer arg areas are scanned as part of scanstack.
673 if typ == deferType {
674 dataSize = unsafe.Sizeof(_defer{})
676 heapBitsSetType(uintptr(x), size, dataSize, typ)
677 if dataSize > typ.size {
678 // Array allocation. If there are any
679 // pointers, GC has to scan to the last
681 if typ.ptrdata != 0 {
682 c.local_scan += dataSize - typ.size + typ.ptrdata
685 c.local_scan += typ.ptrdata
688 // Ensure that the stores above that initialize x to
689 // type-safe memory and set the heap bits occur before
690 // the caller can make x observable to the garbage
691 // collector. Otherwise, on weakly ordered machines,
692 // the garbage collector could follow a pointer to x,
693 // but see uninitialized memory or stale heap bits.
697 // GCmarkterminate allocates black
698 // All slots hold nil so no scanning is needed.
699 // This may be racing with GC so do it atomically if there can be
700 // a race marking the bit.
701 if gcphase == _GCmarktermination || gcBlackenPromptly {
703 gcmarknewobject_m(uintptr(x), size)
714 if debug.allocfreetrace != 0 {
715 tracealloc(x, size, typ)
718 if rate := MemProfileRate; rate > 0 {
719 if size < uintptr(rate) && int32(size) < c.next_sample {
720 c.next_sample -= int32(size)
723 profilealloc(mp, x, size)
729 // Account for internal fragmentation in the assist
730 // debt now that we know it.
731 assistG.gcAssistBytes -= int64(size - dataSize)
734 if shouldhelpgc && shouldtriggergc() {
735 startGC(gcBackgroundMode, false)
736 } else if shouldhelpgc && bggc.working != 0 && gcBlackenEnabled == 0 {
737 // The GC is starting up or shutting down, so we can't
738 // assist, but we also can't allocate unabated. Slow
739 // down this G's allocation and help the GC stay
740 // scheduled by yielding.
742 // TODO: This is a workaround. Either help the GC make
743 // the transition or block.
745 if gp != gp.m.g0 && gp.m.locks == 0 && gp.m.preemptoff == "" {
753 func largeAlloc(size uintptr, flag uint32) *mspan {
754 // print("largeAlloc size=", size, "\n")
756 if size+_PageSize < size {
757 throw("out of memory")
759 npages := size >> _PageShift
760 if size&_PageMask != 0 {
764 // Deduct credit for this span allocation and sweep if
765 // necessary. mHeap_Alloc will also sweep npages, so this only
766 // pays the debt down to npage pages.
767 deductSweepCredit(npages*_PageSize, npages)
769 s := mHeap_Alloc(&mheap_, npages, 0, true, flag&_FlagNoZero == 0)
771 throw("out of memory")
773 s.limit = uintptr(s.start)<<_PageShift + size
774 heapBitsForSpan(s.base()).initSpan(s.layout())
778 // implementation of new builtin
779 func newobject(typ *_type) unsafe.Pointer {
781 if typ.kind&kindNoPointers != 0 {
784 return mallocgc(uintptr(typ.size), typ, flags)
787 //go:linkname reflect_unsafe_New reflect.unsafe_New
788 func reflect_unsafe_New(typ *_type) unsafe.Pointer {
789 return newobject(typ)
792 // implementation of make builtin for slices
793 func newarray(typ *_type, n uintptr) unsafe.Pointer {
795 if typ.kind&kindNoPointers != 0 {
798 if int(n) < 0 || (typ.size > 0 && n > _MaxMem/uintptr(typ.size)) {
799 panic("runtime: allocation size out of range")
801 return mallocgc(uintptr(typ.size)*n, typ, flags)
804 //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray
805 func reflect_unsafe_NewArray(typ *_type, n uintptr) unsafe.Pointer {
806 return newarray(typ, n)
809 // rawmem returns a chunk of pointerless memory. It is
811 func rawmem(size uintptr) unsafe.Pointer {
812 return mallocgc(size, nil, flagNoScan|flagNoZero)
815 func profilealloc(mp *m, x unsafe.Pointer, size uintptr) {
816 mp.mcache.next_sample = nextSample()
817 mProf_Malloc(x, size)
820 // nextSample returns the next sampling point for heap profiling.
821 // It produces a random variable with a geometric distribution and
822 // mean MemProfileRate. This is done by generating a uniformly
823 // distributed random number and applying the cumulative distribution
824 // function for an exponential.
825 func nextSample() int32 {
826 period := MemProfileRate
828 // make nextSample not overflow. Maximum possible step is
829 // -ln(1/(1<<kRandomBitCount)) * period, approximately 20 * period.
831 case period > 0x7000000:
837 // Let m be the sample rate,
838 // the probability distribution function is m*exp(-mx), so the CDF is
839 // p = 1 - exp(-mx), so
840 // q = 1 - p == exp(-mx)
843 // x = -log_e(q) * period
844 // x = log_2(q) * (-log_e(2)) * period ; Using log_2 for efficiency
845 const randomBitCount = 26
846 q := uint32(fastrand1())%(1<<randomBitCount) + 1
847 qlog := fastlog2(float64(q)) - randomBitCount
851 const minusLog2 = -0.6931471805599453 // -ln(2)
852 return int32(qlog*(minusLog2*float64(period))) + 1
855 type persistentAlloc struct {
860 var globalAlloc struct {
865 // Wrapper around sysAlloc that can allocate small chunks.
866 // There is no associated free operation.
867 // Intended for things like function/type/debug-related persistent data.
868 // If align is 0, uses default align (currently 8).
869 func persistentalloc(size, align uintptr, sysStat *uint64) unsafe.Pointer {
872 p = persistentalloc1(size, align, sysStat)
877 // Must run on system stack because stack growth can (re)invoke it.
880 func persistentalloc1(size, align uintptr, sysStat *uint64) unsafe.Pointer {
883 maxBlock = 64 << 10 // VM reservation granularity is 64K on windows
887 throw("persistentalloc: size == 0")
890 if align&(align-1) != 0 {
891 throw("persistentalloc: align is not a power of 2")
893 if align > _PageSize {
894 throw("persistentalloc: align is too large")
900 if size >= maxBlock {
901 return sysAlloc(size, sysStat)
905 var persistent *persistentAlloc
906 if mp != nil && mp.p != 0 {
907 persistent = &mp.p.ptr().palloc
909 lock(&globalAlloc.mutex)
910 persistent = &globalAlloc.persistentAlloc
912 persistent.off = round(persistent.off, align)
913 if persistent.off+size > chunk || persistent.base == nil {
914 persistent.base = sysAlloc(chunk, &memstats.other_sys)
915 if persistent.base == nil {
916 if persistent == &globalAlloc.persistentAlloc {
917 unlock(&globalAlloc.mutex)
919 throw("runtime: cannot allocate memory")
923 p := add(persistent.base, persistent.off)
924 persistent.off += size
926 if persistent == &globalAlloc.persistentAlloc {
927 unlock(&globalAlloc.mutex)
930 if sysStat != &memstats.other_sys {
931 mSysStatInc(sysStat, size)
932 mSysStatDec(&memstats.other_sys, size)