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.
7 // This was originally based on tcmalloc, but has diverged quite a bit.
8 // http://goog-perftools.sourceforge.net/doc/tcmalloc.html
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.
17 // The allocator's data structures are:
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.
27 // Allocating a small object proceeds up a hierarchy of caches:
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.
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
41 // 3. If the mcentral's mspan list is empty, obtain a run
42 // of pages from the mheap to use for the mspan.
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.
49 // Sweeping an mspan and freeing objects on it proceeds up a similar
52 // 1. If the mspan is being swept in response to allocation, it
53 // is returned to the mcache to satisfy the allocation.
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
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.
62 // Allocating and freeing a large object uses the mheap
63 // directly, bypassing the mcache and mcentral.
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
70 // 1. Stack frame allocation can avoid zeroing altogether.
72 // 2. It exhibits better temporal locality, since the program is
73 // probably about to write to the memory.
75 // 3. We don't zero pages that never get reused.
77 // Virtual memory layout
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.
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.
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.
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.
104 "runtime/internal/atomic"
105 "runtime/internal/math"
106 "runtime/internal/sys"
113 maxTinySize = _TinySize
114 tinySizeClass = _TinySizeClass
115 maxSmallSize = _MaxSmallSize
117 pageShift = _PageShift
120 // By construction, single page spans of the smallest object class
121 // have the most objects per span.
122 maxObjsPerSpan = pageSize / 8
124 concurrentSweep = _ConcurrentSweep
126 _PageSize = 1 << _PageShift
127 _PageMask = _PageSize - 1
129 // _64bit = 1 on 64-bit systems, 0 on 32-bit systems
130 _64bit = 1 << (^uintptr(0) >> 63) / 2
132 // Tiny allocator parameters, see "Tiny allocator" comment in malloc.go.
134 _TinySizeClass = int8(2)
136 _FixAllocChunk = 16 << 10 // Chunk size for FixAlloc
138 // Per-P, per order stack segment cache size.
139 _StackCacheSize = 32 * 1024
141 // Number of orders that get caching. Order 0 is FixedStack
142 // and each successive order is twice as large.
143 // We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks
144 // will be allocated directly.
145 // Since FixedStack is different on different systems, we
146 // must vary NumStackOrders to keep the same maximum cached size.
147 // OS | FixedStack | NumStackOrders
148 // -----------------+------------+---------------
149 // linux/darwin/bsd | 2KB | 4
150 // windows/32 | 4KB | 3
151 // windows/64 | 8KB | 2
153 _NumStackOrders = 4 - sys.PtrSize/4*sys.GoosWindows - 1*sys.GoosPlan9
155 // heapAddrBits is the number of bits in a heap address. On
156 // amd64, addresses are sign-extended beyond heapAddrBits. On
157 // other arches, they are zero-extended.
159 // On most 64-bit platforms, we limit this to 48 bits based on a
160 // combination of hardware and OS limitations.
162 // amd64 hardware limits addresses to 48 bits, sign-extended
163 // to 64 bits. Addresses where the top 16 bits are not either
164 // all 0 or all 1 are "non-canonical" and invalid. Because of
165 // these "negative" addresses, we offset addresses by 1<<47
166 // (arenaBaseOffset) on amd64 before computing indexes into
167 // the heap arenas index. In 2017, amd64 hardware added
168 // support for 57 bit addresses; however, currently only Linux
169 // supports this extension and the kernel will never choose an
170 // address above 1<<47 unless mmap is called with a hint
171 // address above 1<<47 (which we never do).
173 // arm64 hardware (as of ARMv8) limits user addresses to 48
174 // bits, in the range [0, 1<<48).
176 // ppc64, mips64, and s390x support arbitrary 64 bit addresses
177 // in hardware. On Linux, Go leans on stricter OS limits. Based
178 // on Linux's processor.h, the user address space is limited as
179 // follows on 64-bit architectures:
181 // Architecture Name Maximum Value (exclusive)
182 // ---------------------------------------------------------------------
183 // amd64 TASK_SIZE_MAX 0x007ffffffff000 (47 bit addresses)
184 // arm64 TASK_SIZE_64 0x01000000000000 (48 bit addresses)
185 // ppc64{,le} TASK_SIZE_USER64 0x00400000000000 (46 bit addresses)
186 // mips64{,le} TASK_SIZE64 0x00010000000000 (40 bit addresses)
187 // s390x TASK_SIZE 1<<64 (64 bit addresses)
189 // These limits may increase over time, but are currently at
190 // most 48 bits except on s390x. On all architectures, Linux
191 // starts placing mmap'd regions at addresses that are
192 // significantly below 48 bits, so even if it's possible to
193 // exceed Go's 48 bit limit, it's extremely unlikely in
196 // On 32-bit platforms, we accept the full 32-bit address
197 // space because doing so is cheap.
198 // mips32 only has access to the low 2GB of virtual memory, so
199 // we further limit it to 31 bits.
201 // On ios/arm64, although 64-bit pointers are presumably
202 // available, pointers are truncated to 33 bits. Furthermore,
203 // only the top 4 GiB of the address space are actually available
204 // to the application, but we allow the whole 33 bits anyway for
206 // TODO(mknyszek): Consider limiting it to 32 bits and using
207 // arenaBaseOffset to offset into the top 4 GiB.
209 // WebAssembly currently has a limit of 4GB linear memory.
210 heapAddrBits = (_64bit*(1-sys.GoarchWasm)*(1-sys.GoosIos*sys.GoarchArm64))*48 + (1-_64bit+sys.GoarchWasm)*(32-(sys.GoarchMips+sys.GoarchMipsle)) + 33*sys.GoosIos*sys.GoarchArm64
212 // maxAlloc is the maximum size of an allocation. On 64-bit,
213 // it's theoretically possible to allocate 1<<heapAddrBits bytes. On
214 // 32-bit, however, this is one less than 1<<32 because the
215 // number of bytes in the address space doesn't actually fit
217 maxAlloc = (1 << heapAddrBits) - (1-_64bit)*1
219 // The number of bits in a heap address, the size of heap
220 // arenas, and the L1 and L2 arena map sizes are related by
222 // (1 << addr bits) = arena size * L1 entries * L2 entries
224 // Currently, we balance these as follows:
226 // Platform Addr bits Arena size L1 entries L2 entries
227 // -------------- --------- ---------- ---------- -----------
228 // */64-bit 48 64MB 1 4M (32MB)
229 // windows/64-bit 48 4MB 64 1M (8MB)
230 // ios/arm64 33 4MB 1 2048 (8KB)
231 // */32-bit 32 4MB 1 1024 (4KB)
232 // */mips(le) 31 4MB 1 512 (2KB)
234 // heapArenaBytes is the size of a heap arena. The heap
235 // consists of mappings of size heapArenaBytes, aligned to
236 // heapArenaBytes. The initial heap mapping is one arena.
238 // This is currently 64MB on 64-bit non-Windows and 4MB on
239 // 32-bit and on Windows. We use smaller arenas on Windows
240 // because all committed memory is charged to the process,
241 // even if it's not touched. Hence, for processes with small
242 // heaps, the mapped arena space needs to be commensurate.
243 // This is particularly important with the race detector,
244 // since it significantly amplifies the cost of committed
246 heapArenaBytes = 1 << logHeapArenaBytes
248 // logHeapArenaBytes is log_2 of heapArenaBytes. For clarity,
249 // prefer using heapArenaBytes where possible (we need the
250 // constant to compute some other constants).
251 logHeapArenaBytes = (6+20)*(_64bit*(1-sys.GoosWindows)*(1-sys.GoarchWasm)*(1-sys.GoosIos*sys.GoarchArm64)) + (2+20)*(_64bit*sys.GoosWindows) + (2+20)*(1-_64bit) + (2+20)*sys.GoarchWasm + (2+20)*sys.GoosIos*sys.GoarchArm64
253 // heapArenaBitmapBytes is the size of each heap arena's bitmap.
254 heapArenaBitmapBytes = heapArenaBytes / (sys.PtrSize * 8 / 2)
256 pagesPerArena = heapArenaBytes / pageSize
258 // arenaL1Bits is the number of bits of the arena number
259 // covered by the first level arena map.
261 // This number should be small, since the first level arena
262 // map requires PtrSize*(1<<arenaL1Bits) of space in the
263 // binary's BSS. It can be zero, in which case the first level
264 // index is effectively unused. There is a performance benefit
265 // to this, since the generated code can be more efficient,
266 // but comes at the cost of having a large L2 mapping.
268 // We use the L1 map on 64-bit Windows because the arena size
269 // is small, but the address space is still 48 bits, and
270 // there's a high cost to having a large L2.
271 arenaL1Bits = 6 * (_64bit * sys.GoosWindows)
273 // arenaL2Bits is the number of bits of the arena number
274 // covered by the second level arena index.
276 // The size of each arena map allocation is proportional to
277 // 1<<arenaL2Bits, so it's important that this not be too
278 // large. 48 bits leads to 32MB arena index allocations, which
279 // is about the practical threshold.
280 arenaL2Bits = heapAddrBits - logHeapArenaBytes - arenaL1Bits
282 // arenaL1Shift is the number of bits to shift an arena frame
283 // number by to compute an index into the first level arena map.
284 arenaL1Shift = arenaL2Bits
286 // arenaBits is the total bits in a combined arena map index.
287 // This is split between the index into the L1 arena map and
289 arenaBits = arenaL1Bits + arenaL2Bits
291 // arenaBaseOffset is the pointer value that corresponds to
292 // index 0 in the heap arena map.
294 // On amd64, the address space is 48 bits, sign extended to 64
295 // bits. This offset lets us handle "negative" addresses (or
296 // high addresses if viewed as unsigned).
298 // On aix/ppc64, this offset allows to keep the heapAddrBits to
299 // 48. Otherwise, it would be 60 in order to handle mmap addresses
300 // (in range 0x0a00000000000000 - 0x0afffffffffffff). But in this
301 // case, the memory reserved in (s *pageAlloc).init for chunks
302 // is causing important slowdowns.
304 // On other platforms, the user address space is contiguous
305 // and starts at 0, so no offset is necessary.
306 arenaBaseOffset = 0xffff800000000000*sys.GoarchAmd64 + 0x0a00000000000000*sys.GoosAix
307 // A typed version of this constant that will make it into DWARF (for viewcore).
308 arenaBaseOffsetUintptr = uintptr(arenaBaseOffset)
310 // Max number of threads to run garbage collection.
311 // 2, 3, and 4 are all plausible maximums depending
312 // on the hardware details of the machine. The garbage
313 // collector scales well to 32 cpus.
316 // minLegalPointer is the smallest possible legal pointer.
317 // This is the smallest possible architectural page size,
318 // since we assume that the first page is never mapped.
320 // This should agree with minZeroPage in the compiler.
321 minLegalPointer uintptr = 4096
324 // physPageSize is the size in bytes of the OS's physical pages.
325 // Mapping and unmapping operations must be done at multiples of
328 // This must be set by the OS init code (typically in osinit) before
330 var physPageSize uintptr
332 // physHugePageSize is the size in bytes of the OS's default physical huge
333 // page size whose allocation is opaque to the application. It is assumed
334 // and verified to be a power of two.
336 // If set, this must be set by the OS init code (typically in osinit) before
337 // mallocinit. However, setting it at all is optional, and leaving the default
338 // value is always safe (though potentially less efficient).
340 // Since physHugePageSize is always assumed to be a power of two,
341 // physHugePageShift is defined as physHugePageSize == 1 << physHugePageShift.
342 // The purpose of physHugePageShift is to avoid doing divisions in
343 // performance critical functions.
345 physHugePageSize uintptr
346 physHugePageShift uint
349 // OS memory management abstraction layer
351 // Regions of the address space managed by the runtime may be in one of four
352 // states at any given time:
353 // 1) None - Unreserved and unmapped, the default state of any region.
354 // 2) Reserved - Owned by the runtime, but accessing it would cause a fault.
355 // Does not count against the process' memory footprint.
356 // 3) Prepared - Reserved, intended not to be backed by physical memory (though
357 // an OS may implement this lazily). Can transition efficiently to
358 // Ready. Accessing memory in such a region is undefined (may
359 // fault, may give back unexpected zeroes, etc.).
360 // 4) Ready - may be accessed safely.
362 // This set of states is more than is strictly necessary to support all the
363 // currently supported platforms. One could get by with just None, Reserved, and
364 // Ready. However, the Prepared state gives us flexibility for performance
365 // purposes. For example, on POSIX-y operating systems, Reserved is usually a
366 // private anonymous mmap'd region with PROT_NONE set, and to transition
367 // to Ready would require setting PROT_READ|PROT_WRITE. However the
368 // underspecification of Prepared lets us use just MADV_FREE to transition from
369 // Ready to Prepared. Thus with the Prepared state we can set the permission
370 // bits just once early on, we can efficiently tell the OS that it's free to
371 // take pages away from us when we don't strictly need them.
373 // For each OS there is a common set of helpers defined that transition
374 // memory regions between these states. The helpers are as follows:
376 // sysAlloc transitions an OS-chosen region of memory from None to Ready.
377 // More specifically, it obtains a large chunk of zeroed memory from the
378 // operating system, typically on the order of a hundred kilobytes
379 // or a megabyte. This memory is always immediately available for use.
381 // sysFree transitions a memory region from any state to None. Therefore, it
382 // returns memory unconditionally. It is used if an out-of-memory error has been
383 // detected midway through an allocation or to carve out an aligned section of
384 // the address space. It is okay if sysFree is a no-op only if sysReserve always
385 // returns a memory region aligned to the heap allocator's alignment
388 // sysReserve transitions a memory region from None to Reserved. It reserves
389 // address space in such a way that it would cause a fatal fault upon access
390 // (either via permissions or not committing the memory). Such a reservation is
391 // thus never backed by physical memory.
392 // If the pointer passed to it is non-nil, the caller wants the
393 // reservation there, but sysReserve can still choose another
394 // location if that one is unavailable.
395 // NOTE: sysReserve returns OS-aligned memory, but the heap allocator
396 // may use larger alignment, so the caller must be careful to realign the
397 // memory obtained by sysReserve.
399 // sysMap transitions a memory region from Reserved to Prepared. It ensures the
400 // memory region can be efficiently transitioned to Ready.
402 // sysUsed transitions a memory region from Prepared to Ready. It notifies the
403 // operating system that the memory region is needed and ensures that the region
404 // may be safely accessed. This is typically a no-op on systems that don't have
405 // an explicit commit step and hard over-commit limits, but is critical on
406 // Windows, for example.
408 // sysUnused transitions a memory region from Ready to Prepared. It notifies the
409 // operating system that the physical pages backing this memory region are no
410 // longer needed and can be reused for other purposes. The contents of a
411 // sysUnused memory region are considered forfeit and the region must not be
412 // accessed again until sysUsed is called.
414 // sysFault transitions a memory region from Ready or Prepared to Reserved. It
415 // marks a region such that it will always fault if accessed. Used only for
416 // debugging the runtime.
419 if class_to_size[_TinySizeClass] != _TinySize {
420 throw("bad TinySizeClass")
423 if heapArenaBitmapBytes&(heapArenaBitmapBytes-1) != 0 {
424 // heapBits expects modular arithmetic on bitmap
425 // addresses to work.
426 throw("heapArenaBitmapBytes not a power of 2")
429 // Copy class sizes out for statistics table.
430 for i := range class_to_size {
431 memstats.by_size[i].size = uint32(class_to_size[i])
434 // Check physPageSize.
435 if physPageSize == 0 {
436 // The OS init code failed to fetch the physical page size.
437 throw("failed to get system page size")
439 if physPageSize > maxPhysPageSize {
440 print("system page size (", physPageSize, ") is larger than maximum page size (", maxPhysPageSize, ")\n")
441 throw("bad system page size")
443 if physPageSize < minPhysPageSize {
444 print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n")
445 throw("bad system page size")
447 if physPageSize&(physPageSize-1) != 0 {
448 print("system page size (", physPageSize, ") must be a power of 2\n")
449 throw("bad system page size")
451 if physHugePageSize&(physHugePageSize-1) != 0 {
452 print("system huge page size (", physHugePageSize, ") must be a power of 2\n")
453 throw("bad system huge page size")
455 if physHugePageSize > maxPhysHugePageSize {
456 // physHugePageSize is greater than the maximum supported huge page size.
457 // Don't throw here, like in the other cases, since a system configured
458 // in this way isn't wrong, we just don't have the code to support them.
459 // Instead, silently set the huge page size to zero.
462 if physHugePageSize != 0 {
463 // Since physHugePageSize is a power of 2, it suffices to increase
464 // physHugePageShift until 1<<physHugePageShift == physHugePageSize.
465 for 1<<physHugePageShift != physHugePageSize {
469 if pagesPerArena%pagesPerSpanRoot != 0 {
470 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerSpanRoot (", pagesPerSpanRoot, ")\n")
471 throw("bad pagesPerSpanRoot")
473 if pagesPerArena%pagesPerReclaimerChunk != 0 {
474 print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerReclaimerChunk (", pagesPerReclaimerChunk, ")\n")
475 throw("bad pagesPerReclaimerChunk")
478 // Initialize the heap.
480 mcache0 = allocmcache()
481 lockInit(&gcBitsArenas.lock, lockRankGcBitsArenas)
482 lockInit(&proflock, lockRankProf)
483 lockInit(&globalAlloc.mutex, lockRankGlobalAlloc)
485 // Create initial arena growth hints.
486 if sys.PtrSize == 8 {
487 // On a 64-bit machine, we pick the following hints
490 // 1. Starting from the middle of the address space
491 // makes it easier to grow out a contiguous range
492 // without running in to some other mapping.
494 // 2. This makes Go heap addresses more easily
495 // recognizable when debugging.
497 // 3. Stack scanning in gccgo is still conservative,
498 // so it's important that addresses be distinguishable
501 // Starting at 0x00c0 means that the valid memory addresses
502 // will begin 0x00c0, 0x00c1, ...
503 // In little-endian, that's c0 00, c1 00, ... None of those are valid
504 // UTF-8 sequences, and they are otherwise as far away from
505 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0
506 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors
507 // on OS X during thread allocations. 0x00c0 causes conflicts with
508 // AddressSanitizer which reserves all memory up to 0x0100.
509 // These choices reduce the odds of a conservative garbage collector
510 // not collecting memory because some non-pointer block of memory
511 // had a bit pattern that matched a memory address.
513 // However, on arm64, we ignore all this advice above and slam the
514 // allocation at 0x40 << 32 because when using 4k pages with 3-level
515 // translation buffers, the user address space is limited to 39 bits
516 // On ios/arm64, the address space is even smaller.
518 // On AIX, mmaps starts at 0x0A00000000000000 for 64-bit.
520 for i := 0x7f; i >= 0; i-- {
524 // The TSAN runtime requires the heap
525 // to be in the range [0x00c000000000,
527 p = uintptr(i)<<32 | uintptrMask&(0x00c0<<32)
528 if p >= uintptrMask&0x00e000000000 {
531 case GOARCH == "arm64" && GOOS == "ios":
532 p = uintptr(i)<<40 | uintptrMask&(0x0013<<28)
533 case GOARCH == "arm64":
534 p = uintptr(i)<<40 | uintptrMask&(0x0040<<32)
537 // We don't use addresses directly after 0x0A00000000000000
538 // to avoid collisions with others mmaps done by non-go programs.
541 p = uintptr(i)<<40 | uintptrMask&(0xa0<<52)
543 p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32)
545 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc())
547 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
550 // On a 32-bit machine, we're much more concerned
551 // about keeping the usable heap contiguous.
554 // 1. We reserve space for all heapArenas up front so
555 // they don't get interleaved with the heap. They're
556 // ~258MB, so this isn't too bad. (We could reserve a
557 // smaller amount of space up front if this is a
560 // 2. We hint the heap to start right above the end of
561 // the binary so we have the best chance of keeping it
564 // 3. We try to stake out a reasonably large initial
567 const arenaMetaSize = (1 << arenaBits) * unsafe.Sizeof(heapArena{})
568 meta := uintptr(sysReserve(nil, arenaMetaSize))
570 mheap_.heapArenaAlloc.init(meta, arenaMetaSize, true)
573 // We want to start the arena low, but if we're linked
574 // against C code, it's possible global constructors
575 // have called malloc and adjusted the process' brk.
576 // Query the brk so we can avoid trying to map the
577 // region over it (which will cause the kernel to put
578 // the region somewhere else, likely at a high
582 // If we ask for the end of the data segment but the
583 // operating system requires a little more space
584 // before we can start allocating, it will give out a
585 // slightly higher pointer. Except QEMU, which is
586 // buggy, as usual: it won't adjust the pointer
587 // upward. So adjust it upward a little bit ourselves:
588 // 1/4 MB to get away from the running binary image.
589 p := firstmoduledata.end
593 if mheap_.heapArenaAlloc.next <= p && p < mheap_.heapArenaAlloc.end {
594 p = mheap_.heapArenaAlloc.end
596 p = alignUp(p+(256<<10), heapArenaBytes)
597 // Because we're worried about fragmentation on
598 // 32-bit, we try to make a large initial reservation.
599 arenaSizes := []uintptr{
604 for _, arenaSize := range arenaSizes {
605 a, size := sysReserveAligned(unsafe.Pointer(p), arenaSize, heapArenaBytes)
607 mheap_.arena.init(uintptr(a), size, false)
608 p = mheap_.arena.end // For hint below
612 hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc())
614 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
618 // sysAlloc allocates heap arena space for at least n bytes. The
619 // returned pointer is always heapArenaBytes-aligned and backed by
620 // h.arenas metadata. The returned size is always a multiple of
621 // heapArenaBytes. sysAlloc returns nil on failure.
622 // There is no corresponding free function.
624 // sysAlloc returns a memory region in the Reserved state. This region must
625 // be transitioned to Prepared and then Ready before use.
628 func (h *mheap) sysAlloc(n uintptr) (v unsafe.Pointer, size uintptr) {
629 assertLockHeld(&h.lock)
631 n = alignUp(n, heapArenaBytes)
633 // First, try the arena pre-reservation.
634 v = h.arena.alloc(n, heapArenaBytes, &memstats.heap_sys)
640 // Try to grow the heap at a hint address.
641 for h.arenaHints != nil {
648 // We can't use this, so don't ask.
650 } else if arenaIndex(p+n-1) >= 1<<arenaBits {
651 // Outside addressable heap. Can't use.
654 v = sysReserve(unsafe.Pointer(p), n)
657 // Success. Update the hint.
665 // Failed. Discard this hint and try the next.
667 // TODO: This would be cleaner if sysReserve could be
668 // told to only return the requested address. In
669 // particular, this is already how Windows behaves, so
670 // it would simplify things there.
674 h.arenaHints = hint.next
675 h.arenaHintAlloc.free(unsafe.Pointer(hint))
680 // The race detector assumes the heap lives in
681 // [0x00c000000000, 0x00e000000000), but we
682 // just ran out of hints in this region. Give
684 throw("too many address space collisions for -race mode")
687 // All of the hints failed, so we'll take any
688 // (sufficiently aligned) address the kernel will give
690 v, size = sysReserveAligned(nil, n, heapArenaBytes)
695 // Create new hints for extending this region.
696 hint := (*arenaHint)(h.arenaHintAlloc.alloc())
697 hint.addr, hint.down = uintptr(v), true
698 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
699 hint = (*arenaHint)(h.arenaHintAlloc.alloc())
700 hint.addr = uintptr(v) + size
701 hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
704 // Check for bad pointers or pointers we can't use.
709 bad = "region exceeds uintptr range"
710 } else if arenaIndex(p) >= 1<<arenaBits {
711 bad = "base outside usable address space"
712 } else if arenaIndex(p+size-1) >= 1<<arenaBits {
713 bad = "end outside usable address space"
716 // This should be impossible on most architectures,
717 // but it would be really confusing to debug.
718 print("runtime: memory allocated by OS [", hex(p), ", ", hex(p+size), ") not in usable address space: ", bad, "\n")
719 throw("memory reservation exceeds address space limit")
723 if uintptr(v)&(heapArenaBytes-1) != 0 {
724 throw("misrounded allocation in sysAlloc")
728 // Create arena metadata.
729 for ri := arenaIndex(uintptr(v)); ri <= arenaIndex(uintptr(v)+size-1); ri++ {
730 l2 := h.arenas[ri.l1()]
732 // Allocate an L2 arena map.
733 l2 = (*[1 << arenaL2Bits]*heapArena)(persistentalloc(unsafe.Sizeof(*l2), sys.PtrSize, nil))
735 throw("out of memory allocating heap arena map")
737 atomic.StorepNoWB(unsafe.Pointer(&h.arenas[ri.l1()]), unsafe.Pointer(l2))
740 if l2[ri.l2()] != nil {
741 throw("arena already initialized")
744 r = (*heapArena)(h.heapArenaAlloc.alloc(unsafe.Sizeof(*r), sys.PtrSize, &memstats.gcMiscSys))
746 r = (*heapArena)(persistentalloc(unsafe.Sizeof(*r), sys.PtrSize, &memstats.gcMiscSys))
748 throw("out of memory allocating heap arena metadata")
752 // Add the arena to the arenas list.
753 if len(h.allArenas) == cap(h.allArenas) {
754 size := 2 * uintptr(cap(h.allArenas)) * sys.PtrSize
758 newArray := (*notInHeap)(persistentalloc(size, sys.PtrSize, &memstats.gcMiscSys))
760 throw("out of memory allocating allArenas")
762 oldSlice := h.allArenas
763 *(*notInHeapSlice)(unsafe.Pointer(&h.allArenas)) = notInHeapSlice{newArray, len(h.allArenas), int(size / sys.PtrSize)}
764 copy(h.allArenas, oldSlice)
765 // Do not free the old backing array because
766 // there may be concurrent readers. Since we
767 // double the array each time, this can lead
768 // to at most 2x waste.
770 h.allArenas = h.allArenas[:len(h.allArenas)+1]
771 h.allArenas[len(h.allArenas)-1] = ri
773 // Store atomically just in case an object from the
774 // new heap arena becomes visible before the heap lock
775 // is released (which shouldn't happen, but there's
776 // little downside to this).
777 atomic.StorepNoWB(unsafe.Pointer(&l2[ri.l2()]), unsafe.Pointer(r))
780 // Tell the race detector about the new heap memory.
782 racemapshadow(v, size)
788 // sysReserveAligned is like sysReserve, but the returned pointer is
789 // aligned to align bytes. It may reserve either n or n+align bytes,
790 // so it returns the size that was reserved.
791 func sysReserveAligned(v unsafe.Pointer, size, align uintptr) (unsafe.Pointer, uintptr) {
792 // Since the alignment is rather large in uses of this
793 // function, we're not likely to get it by chance, so we ask
794 // for a larger region and remove the parts we don't need.
797 p := uintptr(sysReserve(v, size+align))
801 case p&(align-1) == 0:
802 // We got lucky and got an aligned region, so we can
803 // use the whole thing.
804 return unsafe.Pointer(p), size + align
805 case GOOS == "windows":
806 // On Windows we can't release pieces of a
807 // reservation, so we release the whole thing and
808 // re-reserve the aligned sub-region. This may race,
809 // so we may have to try again.
810 sysFree(unsafe.Pointer(p), size+align, nil)
811 p = alignUp(p, align)
812 p2 := sysReserve(unsafe.Pointer(p), size)
813 if p != uintptr(p2) {
814 // Must have raced. Try again.
815 sysFree(p2, size, nil)
816 if retries++; retries == 100 {
817 throw("failed to allocate aligned heap memory; too many retries")
824 // Trim off the unaligned parts.
825 pAligned := alignUp(p, align)
826 sysFree(unsafe.Pointer(p), pAligned-p, nil)
827 end := pAligned + size
828 endLen := (p + size + align) - end
830 sysFree(unsafe.Pointer(end), endLen, nil)
832 return unsafe.Pointer(pAligned), size
836 // base address for all 0-byte allocations
839 // nextFreeFast returns the next free object if one is quickly available.
840 // Otherwise it returns 0.
841 func nextFreeFast(s *mspan) gclinkptr {
842 theBit := sys.Ctz64(s.allocCache) // Is there a free object in the allocCache?
844 result := s.freeindex + uintptr(theBit)
845 if result < s.nelems {
846 freeidx := result + 1
847 if freeidx%64 == 0 && freeidx != s.nelems {
850 s.allocCache >>= uint(theBit + 1)
851 s.freeindex = freeidx
853 return gclinkptr(result*s.elemsize + s.base())
859 // nextFree returns the next free object from the cached span if one is available.
860 // Otherwise it refills the cache with a span with an available object and
861 // returns that object along with a flag indicating that this was a heavy
862 // weight allocation. If it is a heavy weight allocation the caller must
863 // determine whether a new GC cycle needs to be started or if the GC is active
864 // whether this goroutine needs to assist the GC.
866 // Must run in a non-preemptible context since otherwise the owner of
868 func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) {
871 freeIndex := s.nextFreeIndex()
872 if freeIndex == s.nelems {
874 if uintptr(s.allocCount) != s.nelems {
875 println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems)
876 throw("s.allocCount != s.nelems && freeIndex == s.nelems")
882 freeIndex = s.nextFreeIndex()
885 if freeIndex >= s.nelems {
886 throw("freeIndex is not valid")
889 v = gclinkptr(freeIndex*s.elemsize + s.base())
891 if uintptr(s.allocCount) > s.nelems {
892 println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems)
893 throw("s.allocCount > s.nelems")
898 // Allocate an object of size bytes.
899 // Small objects are allocated from the per-P cache's free lists.
900 // Large objects (> 32 kB) are allocated straight from the heap.
901 func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
902 if gcphase == _GCmarktermination {
903 throw("mallocgc called with gcphase == _GCmarktermination")
907 return unsafe.Pointer(&zerobase)
914 // TODO(austin): This should be just
915 // align = uintptr(typ.align)
916 // but that's only 4 on 32-bit platforms,
917 // even if there's a uint64 field in typ (see #599).
918 // This causes 64-bit atomic accesses to panic.
919 // Hence, we use stricter alignment that matches
920 // the normal allocator better.
923 } else if size&3 == 0 {
925 } else if size&1 == 0 {
931 return persistentalloc(size, align, &memstats.other_sys)
934 if inittrace.active && inittrace.id == getg().goid {
935 // Init functions are executed sequentially in a single goroutine.
936 inittrace.allocs += 1
940 // assistG is the G to charge for this allocation, or nil if
941 // GC is not currently active.
943 if gcBlackenEnabled != 0 {
944 // Charge the current user G for this allocation.
946 if assistG.m.curg != nil {
947 assistG = assistG.m.curg
949 // Charge the allocation against the G. We'll account
950 // for internal fragmentation at the end of mallocgc.
951 assistG.gcAssistBytes -= int64(size)
953 if assistG.gcAssistBytes < 0 {
954 // This G is in debt. Assist the GC to correct
955 // this before allocating. This must happen
956 // before disabling preemption.
957 gcAssistAlloc(assistG)
961 // Set mp.mallocing to keep from being preempted by GC.
963 if mp.mallocing != 0 {
964 throw("malloc deadlock")
966 if mp.gsignal == getg() {
967 throw("malloc during signal")
971 shouldhelpgc := false
975 throw("mallocgc called without a P or outside bootstrapping")
979 noscan := typ == nil || typ.ptrdata == 0
980 // In some cases block zeroing can profitably (for latency reduction purposes)
981 // be delayed till preemption is possible; isZeroed tracks that state.
983 if size <= maxSmallSize {
984 if noscan && size < maxTinySize {
987 // Tiny allocator combines several tiny allocation requests
988 // into a single memory block. The resulting memory block
989 // is freed when all subobjects are unreachable. The subobjects
990 // must be noscan (don't have pointers), this ensures that
991 // the amount of potentially wasted memory is bounded.
993 // Size of the memory block used for combining (maxTinySize) is tunable.
994 // Current setting is 16 bytes, which relates to 2x worst case memory
995 // wastage (when all but one subobjects are unreachable).
996 // 8 bytes would result in no wastage at all, but provides less
997 // opportunities for combining.
998 // 32 bytes provides more opportunities for combining,
999 // but can lead to 4x worst case wastage.
1000 // The best case winning is 8x regardless of block size.
1002 // Objects obtained from tiny allocator must not be freed explicitly.
1003 // So when an object will be freed explicitly, we ensure that
1004 // its size >= maxTinySize.
1006 // SetFinalizer has a special case for objects potentially coming
1007 // from tiny allocator, it such case it allows to set finalizers
1008 // for an inner byte of a memory block.
1010 // The main targets of tiny allocator are small strings and
1011 // standalone escaping variables. On a json benchmark
1012 // the allocator reduces number of allocations by ~12% and
1013 // reduces heap size by ~20%.
1015 // Align tiny pointer for required (conservative) alignment.
1017 off = alignUp(off, 8)
1018 } else if sys.PtrSize == 4 && size == 12 {
1019 // Conservatively align 12-byte objects to 8 bytes on 32-bit
1020 // systems so that objects whose first field is a 64-bit
1021 // value is aligned to 8 bytes and does not cause a fault on
1022 // atomic access. See issue 37262.
1023 // TODO(mknyszek): Remove this workaround if/when issue 36606
1025 off = alignUp(off, 8)
1026 } else if size&3 == 0 {
1027 off = alignUp(off, 4)
1028 } else if size&1 == 0 {
1029 off = alignUp(off, 2)
1031 if off+size <= maxTinySize && c.tiny != 0 {
1032 // The object fits into existing tiny block.
1033 x = unsafe.Pointer(c.tiny + off)
1034 c.tinyoffset = off + size
1040 // Allocate a new maxTinySize block.
1041 span = c.alloc[tinySpanClass]
1042 v := nextFreeFast(span)
1044 v, span, shouldhelpgc = c.nextFree(tinySpanClass)
1046 x = unsafe.Pointer(v)
1047 (*[2]uint64)(x)[0] = 0
1048 (*[2]uint64)(x)[1] = 0
1049 // See if we need to replace the existing tiny block with the new one
1050 // based on amount of remaining free space.
1051 if !raceenabled && (size < c.tinyoffset || c.tiny == 0) {
1052 // Note: disabled when race detector is on, see comment near end of this function.
1059 if size <= smallSizeMax-8 {
1060 sizeclass = size_to_class8[divRoundUp(size, smallSizeDiv)]
1062 sizeclass = size_to_class128[divRoundUp(size-smallSizeMax, largeSizeDiv)]
1064 size = uintptr(class_to_size[sizeclass])
1065 spc := makeSpanClass(sizeclass, noscan)
1067 v := nextFreeFast(span)
1069 v, span, shouldhelpgc = c.nextFree(spc)
1071 x = unsafe.Pointer(v)
1072 if needzero && span.needzero != 0 {
1073 memclrNoHeapPointers(unsafe.Pointer(v), size)
1078 // For large allocations, keep track of zeroed state so that
1079 // bulk zeroing can be happen later in a preemptible context.
1080 span, isZeroed = c.allocLarge(size, needzero && !noscan, noscan)
1083 x = unsafe.Pointer(span.base())
1084 size = span.elemsize
1087 var scanSize uintptr
1089 heapBitsSetType(uintptr(x), size, dataSize, typ)
1090 if dataSize > typ.size {
1091 // Array allocation. If there are any
1092 // pointers, GC has to scan to the last
1094 if typ.ptrdata != 0 {
1095 scanSize = dataSize - typ.size + typ.ptrdata
1098 scanSize = typ.ptrdata
1100 c.scanAlloc += scanSize
1103 // Ensure that the stores above that initialize x to
1104 // type-safe memory and set the heap bits occur before
1105 // the caller can make x observable to the garbage
1106 // collector. Otherwise, on weakly ordered machines,
1107 // the garbage collector could follow a pointer to x,
1108 // but see uninitialized memory or stale heap bits.
1109 publicationBarrier()
1111 // Allocate black during GC.
1112 // All slots hold nil so no scanning is needed.
1113 // This may be racing with GC so do it atomically if there can be
1114 // a race marking the bit.
1115 if gcphase != _GCoff {
1116 gcmarknewobject(span, uintptr(x), size, scanSize)
1130 // Pointerfree data can be zeroed late in a context where preemption can occur.
1131 // x will keep the memory alive.
1132 if !isZeroed && needzero {
1133 memclrNoHeapPointersChunked(size, x)
1137 if debug.allocfreetrace != 0 {
1138 tracealloc(x, size, typ)
1141 if inittrace.active && inittrace.id == getg().goid {
1142 // Init functions are executed sequentially in a single goroutine.
1143 inittrace.bytes += uint64(size)
1147 if rate := MemProfileRate; rate > 0 {
1148 if rate != 1 && size < c.nextSample {
1149 c.nextSample -= size
1152 profilealloc(mp, x, size)
1158 // Account for internal fragmentation in the assist
1159 // debt now that we know it.
1160 assistG.gcAssistBytes -= int64(size - dataSize)
1164 if t := (gcTrigger{kind: gcTriggerHeap}); t.test() {
1169 if raceenabled && noscan && dataSize < maxTinySize {
1170 // Pad tinysize allocations so they are aligned with the end
1171 // of the tinyalloc region. This ensures that any arithmetic
1172 // that goes off the top end of the object will be detectable
1173 // by checkptr (issue 38872).
1174 // Note that we disable tinyalloc when raceenabled for this to work.
1175 // TODO: This padding is only performed when the race detector
1176 // is enabled. It would be nice to enable it if any package
1177 // was compiled with checkptr, but there's no easy way to
1178 // detect that (especially at compile time).
1179 // TODO: enable this padding for all allocations, not just
1180 // tinyalloc ones. It's tricky because of pointer maps.
1181 // Maybe just all noscan objects?
1182 x = add(x, size-dataSize)
1188 // memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers
1189 // on chunks of the buffer to be zeroed, with opportunities for preemption
1190 // along the way. memclrNoHeapPointers contains no safepoints and also
1191 // cannot be preemptively scheduled, so this provides a still-efficient
1192 // block copy that can also be preempted on a reasonable granularity.
1194 // Use this with care; if the data being cleared is tagged to contain
1195 // pointers, this allows the GC to run before it is all cleared.
1196 func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) {
1198 // got this from benchmarking. 128k is too small, 512k is too large.
1199 const chunkBytes = 256 * 1024
1201 for voff := v; voff < vsize; voff = voff + chunkBytes {
1203 // may hold locks, e.g., profiling
1206 // clear min(avail, lump) bytes
1211 memclrNoHeapPointers(unsafe.Pointer(voff), n)
1215 // implementation of new builtin
1216 // compiler (both frontend and SSA backend) knows the signature
1218 func newobject(typ *_type) unsafe.Pointer {
1219 return mallocgc(typ.size, typ, true)
1222 //go:linkname reflect_unsafe_New reflect.unsafe_New
1223 func reflect_unsafe_New(typ *_type) unsafe.Pointer {
1224 return mallocgc(typ.size, typ, true)
1227 //go:linkname reflectlite_unsafe_New internal/reflectlite.unsafe_New
1228 func reflectlite_unsafe_New(typ *_type) unsafe.Pointer {
1229 return mallocgc(typ.size, typ, true)
1232 // newarray allocates an array of n elements of type typ.
1233 func newarray(typ *_type, n int) unsafe.Pointer {
1235 return mallocgc(typ.size, typ, true)
1237 mem, overflow := math.MulUintptr(typ.size, uintptr(n))
1238 if overflow || mem > maxAlloc || n < 0 {
1239 panic(plainError("runtime: allocation size out of range"))
1241 return mallocgc(mem, typ, true)
1244 //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray
1245 func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer {
1246 return newarray(typ, n)
1249 func profilealloc(mp *m, x unsafe.Pointer, size uintptr) {
1252 throw("profilealloc called without a P or outside bootstrapping")
1254 c.nextSample = nextSample()
1255 mProf_Malloc(x, size)
1258 // nextSample returns the next sampling point for heap profiling. The goal is
1259 // to sample allocations on average every MemProfileRate bytes, but with a
1260 // completely random distribution over the allocation timeline; this
1261 // corresponds to a Poisson process with parameter MemProfileRate. In Poisson
1262 // processes, the distance between two samples follows the exponential
1263 // distribution (exp(MemProfileRate)), so the best return value is a random
1264 // number taken from an exponential distribution whose mean is MemProfileRate.
1265 func nextSample() uintptr {
1266 if MemProfileRate == 1 {
1267 // Callers assign our return value to
1268 // mcache.next_sample, but next_sample is not used
1269 // when the rate is 1. So avoid the math below and
1270 // just return something.
1273 if GOOS == "plan9" {
1274 // Plan 9 doesn't support floating point in note handler.
1275 if g := getg(); g == g.m.gsignal {
1276 return nextSampleNoFP()
1280 return uintptr(fastexprand(MemProfileRate))
1283 // fastexprand returns a random number from an exponential distribution with
1284 // the specified mean.
1285 func fastexprand(mean int) int32 {
1286 // Avoid overflow. Maximum possible step is
1287 // -ln(1/(1<<randomBitCount)) * mean, approximately 20 * mean.
1289 case mean > 0x7000000:
1295 // Take a random sample of the exponential distribution exp(-mean*x).
1296 // The probability distribution function is mean*exp(-mean*x), so the CDF is
1297 // p = 1 - exp(-mean*x), so
1298 // q = 1 - p == exp(-mean*x)
1299 // log_e(q) = -mean*x
1300 // -log_e(q)/mean = x
1301 // x = -log_e(q) * mean
1302 // x = log_2(q) * (-log_e(2)) * mean ; Using log_2 for efficiency
1303 const randomBitCount = 26
1304 q := fastrand()%(1<<randomBitCount) + 1
1305 qlog := fastlog2(float64(q)) - randomBitCount
1309 const minusLog2 = -0.6931471805599453 // -ln(2)
1310 return int32(qlog*(minusLog2*float64(mean))) + 1
1313 // nextSampleNoFP is similar to nextSample, but uses older,
1314 // simpler code to avoid floating point.
1315 func nextSampleNoFP() uintptr {
1316 // Set first allocation sample size.
1317 rate := MemProfileRate
1318 if rate > 0x3fffffff { // make 2*rate not overflow
1322 return uintptr(fastrand() % uint32(2*rate))
1327 type persistentAlloc struct {
1332 var globalAlloc struct {
1337 // persistentChunkSize is the number of bytes we allocate when we grow
1338 // a persistentAlloc.
1339 const persistentChunkSize = 256 << 10
1341 // persistentChunks is a list of all the persistent chunks we have
1342 // allocated. The list is maintained through the first word in the
1343 // persistent chunk. This is updated atomically.
1344 var persistentChunks *notInHeap
1346 // Wrapper around sysAlloc that can allocate small chunks.
1347 // There is no associated free operation.
1348 // Intended for things like function/type/debug-related persistent data.
1349 // If align is 0, uses default align (currently 8).
1350 // The returned memory will be zeroed.
1352 // Consider marking persistentalloc'd types go:notinheap.
1353 func persistentalloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer {
1355 systemstack(func() {
1356 p = persistentalloc1(size, align, sysStat)
1358 return unsafe.Pointer(p)
1361 // Must run on system stack because stack growth can (re)invoke it.
1364 func persistentalloc1(size, align uintptr, sysStat *sysMemStat) *notInHeap {
1366 maxBlock = 64 << 10 // VM reservation granularity is 64K on windows
1370 throw("persistentalloc: size == 0")
1373 if align&(align-1) != 0 {
1374 throw("persistentalloc: align is not a power of 2")
1376 if align > _PageSize {
1377 throw("persistentalloc: align is too large")
1383 if size >= maxBlock {
1384 return (*notInHeap)(sysAlloc(size, sysStat))
1388 var persistent *persistentAlloc
1389 if mp != nil && mp.p != 0 {
1390 persistent = &mp.p.ptr().palloc
1392 lock(&globalAlloc.mutex)
1393 persistent = &globalAlloc.persistentAlloc
1395 persistent.off = alignUp(persistent.off, align)
1396 if persistent.off+size > persistentChunkSize || persistent.base == nil {
1397 persistent.base = (*notInHeap)(sysAlloc(persistentChunkSize, &memstats.other_sys))
1398 if persistent.base == nil {
1399 if persistent == &globalAlloc.persistentAlloc {
1400 unlock(&globalAlloc.mutex)
1402 throw("runtime: cannot allocate memory")
1405 // Add the new chunk to the persistentChunks list.
1407 chunks := uintptr(unsafe.Pointer(persistentChunks))
1408 *(*uintptr)(unsafe.Pointer(persistent.base)) = chunks
1409 if atomic.Casuintptr((*uintptr)(unsafe.Pointer(&persistentChunks)), chunks, uintptr(unsafe.Pointer(persistent.base))) {
1413 persistent.off = alignUp(sys.PtrSize, align)
1415 p := persistent.base.add(persistent.off)
1416 persistent.off += size
1418 if persistent == &globalAlloc.persistentAlloc {
1419 unlock(&globalAlloc.mutex)
1422 if sysStat != &memstats.other_sys {
1423 sysStat.add(int64(size))
1424 memstats.other_sys.add(-int64(size))
1429 // inPersistentAlloc reports whether p points to memory allocated by
1430 // persistentalloc. This must be nosplit because it is called by the
1431 // cgo checker code, which is called by the write barrier code.
1433 func inPersistentAlloc(p uintptr) bool {
1434 chunk := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&persistentChunks)))
1436 if p >= chunk && p < chunk+persistentChunkSize {
1439 chunk = *(*uintptr)(unsafe.Pointer(chunk))
1444 // linearAlloc is a simple linear allocator that pre-reserves a region
1445 // of memory and then optionally maps that region into the Ready state
1448 // The caller is responsible for locking.
1449 type linearAlloc struct {
1450 next uintptr // next free byte
1451 mapped uintptr // one byte past end of mapped space
1452 end uintptr // end of reserved space
1454 mapMemory bool // transition memory from Reserved to Ready if true
1457 func (l *linearAlloc) init(base, size uintptr, mapMemory bool) {
1458 if base+size < base {
1459 // Chop off the last byte. The runtime isn't prepared
1460 // to deal with situations where the bounds could overflow.
1461 // Leave that memory reserved, though, so we don't map it
1465 l.next, l.mapped = base, base
1467 l.mapMemory = mapMemory
1470 func (l *linearAlloc) alloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer {
1471 p := alignUp(l.next, align)
1476 if pEnd := alignUp(l.next-1, physPageSize); pEnd > l.mapped {
1478 // Transition from Reserved to Prepared to Ready.
1479 sysMap(unsafe.Pointer(l.mapped), pEnd-l.mapped, sysStat)
1480 sysUsed(unsafe.Pointer(l.mapped), pEnd-l.mapped)
1484 return unsafe.Pointer(p)
1487 // notInHeap is off-heap memory allocated by a lower-level allocator
1488 // like sysAlloc or persistentAlloc.
1490 // In general, it's better to use real types marked as go:notinheap,
1491 // but this serves as a generic type for situations where that isn't
1492 // possible (like in the allocators).
1494 // TODO: Use this as the return type of sysAlloc, persistentAlloc, etc?
1497 type notInHeap struct{}
1499 func (p *notInHeap) add(bytes uintptr) *notInHeap {
1500 return (*notInHeap)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + bytes))