| blob | 028872259d944195a18b3dc952f5639339148ad8 |
1 // Copyright 2009 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 // See malloc.h for overview.
6 //
7 // TODO(rsc): double-check stats.
9 package runtime
10 #include <stddef.h>
11 #include <errno.h>
12 #include <stdlib.h>
13 #include "go-alloc.h"
14 #include "runtime.h"
15 #include "arch.h"
16 #include "malloc.h"
17 #include "interface.h"
18 #include "go-type.h"
19 #include "race.h"
21 // Map gccgo field names to gc field names.
22 // Eface aka __go_empty_interface.
23 #define type __type_descriptor
24 // Type aka __go_type_descriptor
25 #define kind __code
26 #define string __reflection
27 #define KindPtr GO_PTR
28 #define KindNoPointers GO_NO_POINTERS
30 // GCCGO SPECIFIC CHANGE
31 //
32 // There is a long comment in runtime_mallocinit about where to put the heap
33 // on a 64-bit system. It makes assumptions that are not valid on linux/arm64
34 // -- it assumes user space can choose the lower 47 bits of a pointer, but on
35 // linux/arm64 we can only choose the lower 39 bits. This means the heap is
36 // roughly a quarter of the available address space and we cannot choose a bit
37 // pattern that all pointers will have -- luckily the GC is mostly precise
38 // these days so this doesn't matter all that much. The kernel (as of 3.13)
39 // will allocate address space starting either down from 0x7fffffffff or up
40 // from 0x2000000000, so we put the heap roughly in the middle of these two
41 // addresses to minimize the chance that a non-heap allocation will get in the
42 // way of the heap.
43 //
44 // This all means that there isn't much point in trying 256 different
45 // locations for the heap on such systems.
46 #ifdef __aarch64__
47 #define HeapBase(i) ((void*)(uintptr)(0x40ULL<<32))
48 #define HeapBaseOptions 1
49 #else
50 #define HeapBase(i) ((void*)(uintptr)(i<<40|0x00c0ULL<<32))
51 #define HeapBaseOptions 0x80
52 #endif
53 // END GCCGO SPECIFIC CHANGE
55 // Mark mheap as 'no pointers', it does not contain interesting pointers but occupies ~45K.
56 MHeap runtime_mheap;
57 MStats mstats;
59 int32 runtime_checking;
61 extern MStats mstats; // defined in zruntime_def_$GOOS_$GOARCH.go
63 extern volatile intgo runtime_MemProfileRate
64 __asm__ (GOSYM_PREFIX "runtime.MemProfileRate");
66 static MSpan* largealloc(uint32, uintptr*);
67 static void profilealloc(void *v, uintptr size);
68 static void settype(MSpan *s, void *v, uintptr typ);
70 // Allocate an object of at least size bytes.
71 // Small objects are allocated from the per-thread cache's free lists.
72 // Large objects (> 32 kB) are allocated straight from the heap.
73 // If the block will be freed with runtime_free(), typ must be 0.
74 void*
75 runtime_mallocgc(uintptr size, uintptr typ, uint32 flag)
76 {
77 M *m;
78 G *g;
79 int32 sizeclass;
80 uintptr tinysize, size1;
81 intgo rate;
82 MCache *c;
83 MSpan *s;
84 MLink *v, *next;
85 byte *tiny;
86 bool incallback;
88 if(size == 0) {
89 // All 0-length allocations use this pointer.
90 // The language does not require the allocations to
91 // have distinct values.
92 return &runtime_zerobase;
93 }
95 m = runtime_m();
96 g = runtime_g();
98 incallback = false;
99 if(m->mcache == nil && g->ncgo > 0) {
100 // For gccgo this case can occur when a cgo or SWIG function
101 // has an interface return type and the function
102 // returns a non-pointer, so memory allocation occurs
103 // after syscall.Cgocall but before syscall.CgocallDone.
104 // We treat it as a callback.
105 runtime_exitsyscall();
106 m = runtime_m();
107 incallback = true;
108 flag |= FlagNoInvokeGC;
109 }
111 if(runtime_gcwaiting() && g != m->g0 && m->locks == 0 && !(flag & FlagNoInvokeGC)) {
112 runtime_gosched();
113 m = runtime_m();
114 }
115 if(m->mallocing)
116 runtime_throw("malloc/free - deadlock");
117 // Disable preemption during settype.
118 // We can not use m->mallocing for this, because settype calls mallocgc.
119 m->locks++;
120 m->mallocing = 1;
122 if(DebugTypeAtBlockEnd)
123 size += sizeof(uintptr);
125 c = m->mcache;
126 if(!runtime_debug.efence && size <= MaxSmallSize) {
127 if((flag&(FlagNoScan|FlagNoGC)) == FlagNoScan && size < TinySize) {
128 // Tiny allocator.
129 //
130 // Tiny allocator combines several tiny allocation requests
131 // into a single memory block. The resulting memory block
132 // is freed when all subobjects are unreachable. The subobjects
133 // must be FlagNoScan (don't have pointers), this ensures that
134 // the amount of potentially wasted memory is bounded.
135 //
136 // Size of the memory block used for combining (TinySize) is tunable.
137 // Current setting is 16 bytes, which relates to 2x worst case memory
138 // wastage (when all but one subobjects are unreachable).
139 // 8 bytes would result in no wastage at all, but provides less
140 // opportunities for combining.
141 // 32 bytes provides more opportunities for combining,
142 // but can lead to 4x worst case wastage.
143 // The best case winning is 8x regardless of block size.
144 //
145 // Objects obtained from tiny allocator must not be freed explicitly.
146 // So when an object will be freed explicitly, we ensure that
147 // its size >= TinySize.
148 //
149 // SetFinalizer has a special case for objects potentially coming
150 // from tiny allocator, it such case it allows to set finalizers
151 // for an inner byte of a memory block.
152 //
153 // The main targets of tiny allocator are small strings and
154 // standalone escaping variables. On a json benchmark
155 // the allocator reduces number of allocations by ~12% and
156 // reduces heap size by ~20%.
158 tinysize = c->tinysize;
159 if(size <= tinysize) {
160 tiny = c->tiny;
161 // Align tiny pointer for required (conservative) alignment.
162 if((size&7) == 0)
163 tiny = (byte*)ROUND((uintptr)tiny, 8);
164 else if((size&3) == 0)
165 tiny = (byte*)ROUND((uintptr)tiny, 4);
166 else if((size&1) == 0)
167 tiny = (byte*)ROUND((uintptr)tiny, 2);
168 size1 = size + (tiny - c->tiny);
169 if(size1 <= tinysize) {
170 // The object fits into existing tiny block.
171 v = (MLink*)tiny;
172 c->tiny += size1;
173 c->tinysize -= size1;
174 m->mallocing = 0;
175 m->locks--;
176 if(incallback)
177 runtime_entersyscall();
178 return v;
179 }
180 }
181 // Allocate a new TinySize block.
182 s = c->alloc[TinySizeClass];
183 if(s->freelist == nil)
184 s = runtime_MCache_Refill(c, TinySizeClass);
185 v = s->freelist;
186 next = v->next;
187 s->freelist = next;
188 s->ref++;
189 if(next != nil) // prefetching nil leads to a DTLB miss
190 PREFETCH(next);
191 ((uint64*)v)[0] = 0;
192 ((uint64*)v)[1] = 0;
193 // See if we need to replace the existing tiny block with the new one
194 // based on amount of remaining free space.
195 if(TinySize-size > tinysize) {
196 c->tiny = (byte*)v + size;
197 c->tinysize = TinySize - size;
198 }
199 size = TinySize;
200 goto done;
201 }
202 // Allocate from mcache free lists.
203 // Inlined version of SizeToClass().
204 if(size <= 1024-8)
205 sizeclass = runtime_size_to_class8[(size+7)>>3];
206 else
207 sizeclass = runtime_size_to_class128[(size-1024+127) >> 7];
208 size = runtime_class_to_size[sizeclass];
209 s = c->alloc[sizeclass];
210 if(s->freelist == nil)
211 s = runtime_MCache_Refill(c, sizeclass);
212 v = s->freelist;
213 next = v->next;
214 s->freelist = next;
215 s->ref++;
216 if(next != nil) // prefetching nil leads to a DTLB miss
217 PREFETCH(next);
218 if(!(flag & FlagNoZero)) {
219 v->next = nil;
220 // block is zeroed iff second word is zero ...
221 if(size > 2*sizeof(uintptr) && ((uintptr*)v)[1] != 0)
222 runtime_memclr((byte*)v, size);
223 }
224 done:
225 c->local_cachealloc += size;
226 } else {
227 // Allocate directly from heap.
228 s = largealloc(flag, &size);
229 v = (void*)(s->start << PageShift);
230 }
232 if(flag & FlagNoGC)
233 runtime_marknogc(v);
234 else if(!(flag & FlagNoScan))
235 runtime_markscan(v);
237 if(DebugTypeAtBlockEnd)
238 *(uintptr*)((uintptr)v+size-sizeof(uintptr)) = typ;
240 m->mallocing = 0;
241 // TODO: save type even if FlagNoScan? Potentially expensive but might help
242 // heap profiling/tracing.
243 if(UseSpanType && !(flag & FlagNoScan) && typ != 0)
244 settype(s, v, typ);
246 if(raceenabled)
247 runtime_racemalloc(v, size);
249 if(runtime_debug.allocfreetrace)
250 runtime_tracealloc(v, size, typ);
252 if(!(flag & FlagNoProfiling) && (rate = runtime_MemProfileRate) > 0) {
253 if(size < (uintptr)rate && size < (uintptr)(uint32)c->next_sample)
254 c->next_sample -= size;
255 else
256 profilealloc(v, size);
257 }
259 m->locks--;
261 if(!(flag & FlagNoInvokeGC) && mstats.heap_alloc >= mstats.next_gc)
262 runtime_gc(0);
264 if(incallback)
265 runtime_entersyscall();
267 return v;
268 }
270 static MSpan*
271 largealloc(uint32 flag, uintptr *sizep)
272 {
273 uintptr npages, size;
274 MSpan *s;
275 void *v;
277 // Allocate directly from heap.
278 size = *sizep;
279 if(size + PageSize < size)
280 runtime_throw("out of memory");
281 npages = size >> PageShift;
282 if((size & PageMask) != 0)
283 npages++;
284 s = runtime_MHeap_Alloc(&runtime_mheap, npages, 0, 1, !(flag & FlagNoZero));
285 if(s == nil)
286 runtime_throw("out of memory");
287 s->limit = (byte*)(s->start<<PageShift) + size;
288 *sizep = npages<<PageShift;
289 v = (void*)(s->start << PageShift);
290 // setup for mark sweep
291 runtime_markspan(v, 0, 0, true);
292 return s;
293 }
295 static void
296 profilealloc(void *v, uintptr size)
297 {
298 uintptr rate;
299 int32 next;
300 MCache *c;
302 c = runtime_m()->mcache;
303 rate = runtime_MemProfileRate;
304 if(size < rate) {
305 // pick next profile time
306 // If you change this, also change allocmcache.
307 if(rate > 0x3fffffff) // make 2*rate not overflow
308 rate = 0x3fffffff;
309 next = runtime_fastrand1() % (2*rate);
310 // Subtract the "remainder" of the current allocation.
311 // Otherwise objects that are close in size to sampling rate
312 // will be under-sampled, because we consistently discard this remainder.
313 next -= (size - c->next_sample);
314 if(next < 0)
315 next = 0;
316 c->next_sample = next;
317 }
318 runtime_MProf_Malloc(v, size);
319 }
321 void*
322 __go_alloc(uintptr size)
323 {
324 return runtime_mallocgc(size, 0, FlagNoInvokeGC);
325 }
327 // Free the object whose base pointer is v.
328 void
329 __go_free(void *v)
330 {
331 M *m;
332 int32 sizeclass;
333 MSpan *s;
334 MCache *c;
335 uintptr size;
337 if(v == nil)
338 return;
340 // If you change this also change mgc0.c:/^sweep,
341 // which has a copy of the guts of free.
343 m = runtime_m();
344 if(m->mallocing)
345 runtime_throw("malloc/free - deadlock");
346 m->mallocing = 1;
348 if(!runtime_mlookup(v, nil, nil, &s)) {
349 runtime_printf("free %p: not an allocated block\n", v);
350 runtime_throw("free runtime_mlookup");
351 }
352 size = s->elemsize;
353 sizeclass = s->sizeclass;
354 // Objects that are smaller than TinySize can be allocated using tiny alloc,
355 // if then such object is combined with an object with finalizer, we will crash.
356 if(size < TinySize)
357 runtime_throw("freeing too small block");
359 if(runtime_debug.allocfreetrace)
360 runtime_tracefree(v, size);
362 // Ensure that the span is swept.
363 // If we free into an unswept span, we will corrupt GC bitmaps.
364 runtime_MSpan_EnsureSwept(s);
366 if(s->specials != nil)
367 runtime_freeallspecials(s, v, size);
369 c = m->mcache;
370 if(sizeclass == 0) {
371 // Large object.
372 s->needzero = 1;
373 // Must mark v freed before calling unmarkspan and MHeap_Free:
374 // they might coalesce v into other spans and change the bitmap further.
375 runtime_markfreed(v);
376 runtime_unmarkspan(v, 1<<PageShift);
377 // NOTE(rsc,dvyukov): The original implementation of efence
378 // in CL 22060046 used SysFree instead of SysFault, so that
379 // the operating system would eventually give the memory
380 // back to us again, so that an efence program could run
381 // longer without running out of memory. Unfortunately,
382 // calling SysFree here without any kind of adjustment of the
383 // heap data structures means that when the memory does
384 // come back to us, we have the wrong metadata for it, either in
385 // the MSpan structures or in the garbage collection bitmap.
386 // Using SysFault here means that the program will run out of
387 // memory fairly quickly in efence mode, but at least it won't
388 // have mysterious crashes due to confused memory reuse.
389 // It should be possible to switch back to SysFree if we also
390 // implement and then call some kind of MHeap_DeleteSpan.
391 if(runtime_debug.efence)
392 runtime_SysFault((void*)(s->start<<PageShift), size);
393 else
394 runtime_MHeap_Free(&runtime_mheap, s, 1);
395 c->local_nlargefree++;
396 c->local_largefree += size;
397 } else {
398 // Small object.
399 if(size > 2*sizeof(uintptr))
400 ((uintptr*)v)[1] = (uintptr)0xfeedfeedfeedfeedll; // mark as "needs to be zeroed"
401 else if(size > sizeof(uintptr))
402 ((uintptr*)v)[1] = 0;
403 // Must mark v freed before calling MCache_Free:
404 // it might coalesce v and other blocks into a bigger span
405 // and change the bitmap further.
406 c->local_nsmallfree[sizeclass]++;
407 c->local_cachealloc -= size;
408 if(c->alloc[sizeclass] == s) {
409 // We own the span, so we can just add v to the freelist
410 runtime_markfreed(v);
411 ((MLink*)v)->next = s->freelist;
412 s->freelist = v;
413 s->ref--;
414 } else {
415 // Someone else owns this span. Add to free queue.
416 runtime_MCache_Free(c, v, sizeclass, size);
417 }
418 }
419 m->mallocing = 0;
420 }
422 int32
423 runtime_mlookup(void *v, byte **base, uintptr *size, MSpan **sp)
424 {
425 M *m;
426 uintptr n, i;
427 byte *p;
428 MSpan *s;
430 m = runtime_m();
432 m->mcache->local_nlookup++;
433 if (sizeof(void*) == 4 && m->mcache->local_nlookup >= (1<<30)) {
434 // purge cache stats to prevent overflow
435 runtime_lock(&runtime_mheap);
436 runtime_purgecachedstats(m->mcache);
437 runtime_unlock(&runtime_mheap);
438 }
440 s = runtime_MHeap_LookupMaybe(&runtime_mheap, v);
441 if(sp)
442 *sp = s;
443 if(s == nil) {
444 runtime_checkfreed(v, 1);
445 if(base)
446 *base = nil;
447 if(size)
448 *size = 0;
449 return 0;
450 }
452 p = (byte*)((uintptr)s->start<<PageShift);
453 if(s->sizeclass == 0) {
454 // Large object.
455 if(base)
456 *base = p;
457 if(size)
458 *size = s->npages<<PageShift;
459 return 1;
460 }
462 n = s->elemsize;
463 if(base) {
464 i = ((byte*)v - p)/n;
465 *base = p + i*n;
466 }
467 if(size)
468 *size = n;
470 return 1;
471 }
473 void
474 runtime_purgecachedstats(MCache *c)
475 {
476 MHeap *h;
477 int32 i;
479 // Protected by either heap or GC lock.
480 h = &runtime_mheap;
481 mstats.heap_alloc += c->local_cachealloc;
482 c->local_cachealloc = 0;
483 mstats.nlookup += c->local_nlookup;
484 c->local_nlookup = 0;
485 h->largefree += c->local_largefree;
486 c->local_largefree = 0;
487 h->nlargefree += c->local_nlargefree;
488 c->local_nlargefree = 0;
489 for(i=0; i<(int32)nelem(c->local_nsmallfree); i++) {
490 h->nsmallfree[i] += c->local_nsmallfree[i];
491 c->local_nsmallfree[i] = 0;
492 }
493 }
495 extern uintptr runtime_sizeof_C_MStats
496 __asm__ (GOSYM_PREFIX "runtime.Sizeof_C_MStats");
498 // Size of the trailing by_size array differs between Go and C,
499 // NumSizeClasses was changed, but we can not change Go struct because of backward compatibility.
500 // sizeof_C_MStats is what C thinks about size of Go struct.
502 // Initialized in mallocinit because it's defined in go/runtime/mem.go.
504 #define MaxArena32 (2U<<30)
506 void
507 runtime_mallocinit(void)
508 {
509 byte *p, *p1;
510 uintptr arena_size, bitmap_size, spans_size, p_size;
511 extern byte _end[];
512 uintptr limit;
513 uint64 i;
514 bool reserved;
516 runtime_sizeof_C_MStats = sizeof(MStats) - (NumSizeClasses - 61) * sizeof(mstats.by_size[0]);
518 p = nil;
519 p_size = 0;
520 arena_size = 0;
521 bitmap_size = 0;
522 spans_size = 0;
523 reserved = false;
525 // for 64-bit build
526 USED(p);
527 USED(p_size);
528 USED(arena_size);
529 USED(bitmap_size);
530 USED(spans_size);
532 runtime_InitSizes();
534 if(runtime_class_to_size[TinySizeClass] != TinySize)
535 runtime_throw("bad TinySizeClass");
537 // limit = runtime_memlimit();
538 // See https://code.google.com/p/go/issues/detail?id=5049
539 // TODO(rsc): Fix after 1.1.
540 limit = 0;
542 // Set up the allocation arena, a contiguous area of memory where
543 // allocated data will be found. The arena begins with a bitmap large
544 // enough to hold 4 bits per allocated word.
545 if(sizeof(void*) == 8 && (limit == 0 || limit > (1<<30))) {
546 // On a 64-bit machine, allocate from a single contiguous reservation.
547 // 128 GB (MaxMem) should be big enough for now.
548 //
549 // The code will work with the reservation at any address, but ask
550 // SysReserve to use 0x0000XXc000000000 if possible (XX=00...7f).
551 // Allocating a 128 GB region takes away 37 bits, and the amd64
552 // doesn't let us choose the top 17 bits, so that leaves the 11 bits
553 // in the middle of 0x00c0 for us to choose. Choosing 0x00c0 means
554 // that the valid memory addresses will begin 0x00c0, 0x00c1, ..., 0x00df.
555 // In little-endian, that's c0 00, c1 00, ..., df 00. None of those are valid
556 // UTF-8 sequences, and they are otherwise as far away from
557 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0
558 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors
559 // on OS X during thread allocations. 0x00c0 causes conflicts with
560 // AddressSanitizer which reserves all memory up to 0x0100.
561 // These choices are both for debuggability and to reduce the
562 // odds of the conservative garbage collector not collecting memory
563 // because some non-pointer block of memory had a bit pattern
564 // that matched a memory address.
565 //
566 // Actually we reserve 136 GB (because the bitmap ends up being 8 GB)
567 // but it hardly matters: e0 00 is not valid UTF-8 either.
568 //
569 // If this fails we fall back to the 32 bit memory mechanism
570 arena_size = MaxMem;
571 bitmap_size = arena_size / (sizeof(void*)*8/4);
572 spans_size = arena_size / PageSize * sizeof(runtime_mheap.spans[0]);
573 spans_size = ROUND(spans_size, PageSize);
574 for(i = 0; i < HeapBaseOptions; i++) {
575 p = HeapBase(i);
576 p_size = bitmap_size + spans_size + arena_size + PageSize;
577 p = runtime_SysReserve(p, p_size, &reserved);
578 if(p != nil)
579 break;
580 }
581 }
582 if (p == nil) {
583 // On a 32-bit machine, we can't typically get away
584 // with a giant virtual address space reservation.
585 // Instead we map the memory information bitmap
586 // immediately after the data segment, large enough
587 // to handle another 2GB of mappings (256 MB),
588 // along with a reservation for another 512 MB of memory.
589 // When that gets used up, we'll start asking the kernel
590 // for any memory anywhere and hope it's in the 2GB
591 // following the bitmap (presumably the executable begins
592 // near the bottom of memory, so we'll have to use up
593 // most of memory before the kernel resorts to giving out
594 // memory before the beginning of the text segment).
595 //
596 // Alternatively we could reserve 512 MB bitmap, enough
597 // for 4GB of mappings, and then accept any memory the
598 // kernel threw at us, but normally that's a waste of 512 MB
599 // of address space, which is probably too much in a 32-bit world.
600 bitmap_size = MaxArena32 / (sizeof(void*)*8/4);
601 arena_size = 512<<20;
602 spans_size = MaxArena32 / PageSize * sizeof(runtime_mheap.spans[0]);
603 if(limit > 0 && arena_size+bitmap_size+spans_size > limit) {
604 bitmap_size = (limit / 9) & ~((1<<PageShift) - 1);
605 arena_size = bitmap_size * 8;
606 spans_size = arena_size / PageSize * sizeof(runtime_mheap.spans[0]);
607 }
608 spans_size = ROUND(spans_size, PageSize);
610 // SysReserve treats the address we ask for, end, as a hint,
611 // not as an absolute requirement. If we ask for the end
612 // of the data segment but the operating system requires
613 // a little more space before we can start allocating, it will
614 // give out a slightly higher pointer. Except QEMU, which
615 // is buggy, as usual: it won't adjust the pointer upward.
616 // So adjust it upward a little bit ourselves: 1/4 MB to get
617 // away from the running binary image and then round up
618 // to a MB boundary.
619 p = (byte*)ROUND((uintptr)_end + (1<<18), 1<<20);
620 p_size = bitmap_size + spans_size + arena_size + PageSize;
621 p = runtime_SysReserve(p, p_size, &reserved);
622 if(p == nil)
623 runtime_throw("runtime: cannot reserve arena virtual address space");
624 }
626 // PageSize can be larger than OS definition of page size,
627 // so SysReserve can give us a PageSize-unaligned pointer.
628 // To overcome this we ask for PageSize more and round up the pointer.
629 p1 = (byte*)ROUND((uintptr)p, PageSize);
631 runtime_mheap.spans = (MSpan**)p1;
632 runtime_mheap.bitmap = p1 + spans_size;
633 runtime_mheap.arena_start = p1 + spans_size + bitmap_size;
634 runtime_mheap.arena_used = runtime_mheap.arena_start;
635 runtime_mheap.arena_end = p + p_size;
636 runtime_mheap.arena_reserved = reserved;
638 if(((uintptr)runtime_mheap.arena_start & (PageSize-1)) != 0)
639 runtime_throw("misrounded allocation in mallocinit");
641 // Initialize the rest of the allocator.
642 runtime_MHeap_Init(&runtime_mheap);
643 runtime_m()->mcache = runtime_allocmcache();
645 // See if it works.
646 runtime_free(runtime_malloc(TinySize));
647 }
649 void*
650 runtime_MHeap_SysAlloc(MHeap *h, uintptr n)
651 {
652 byte *p, *p_end;
653 uintptr p_size;
654 bool reserved;
657 if(n > (uintptr)(h->arena_end - h->arena_used)) {
658 // We are in 32-bit mode, maybe we didn't use all possible address space yet.
659 // Reserve some more space.
660 byte *new_end;
662 p_size = ROUND(n + PageSize, 256<<20);
663 new_end = h->arena_end + p_size;
664 if(new_end <= h->arena_start + MaxArena32) {
665 // TODO: It would be bad if part of the arena
666 // is reserved and part is not.
667 p = runtime_SysReserve(h->arena_end, p_size, &reserved);
668 if(p == h->arena_end) {
669 h->arena_end = new_end;
670 h->arena_reserved = reserved;
671 }
672 else if(p+p_size <= h->arena_start + MaxArena32) {
673 // Keep everything page-aligned.
674 // Our pages are bigger than hardware pages.
675 h->arena_end = p+p_size;
676 h->arena_used = p + (-(uintptr)p&(PageSize-1));
677 h->arena_reserved = reserved;
678 } else {
679 uint64 stat;
680 stat = 0;
681 runtime_SysFree(p, p_size, &stat);
682 }
683 }
684 }
685 if(n <= (uintptr)(h->arena_end - h->arena_used)) {
686 // Keep taking from our reservation.
687 p = h->arena_used;
688 runtime_SysMap(p, n, h->arena_reserved, &mstats.heap_sys);
689 h->arena_used += n;
690 runtime_MHeap_MapBits(h);
691 runtime_MHeap_MapSpans(h);
692 if(raceenabled)
693 runtime_racemapshadow(p, n);
695 if(((uintptr)p & (PageSize-1)) != 0)
696 runtime_throw("misrounded allocation in MHeap_SysAlloc");
697 return p;
698 }
700 // If using 64-bit, our reservation is all we have.
701 if((uintptr)(h->arena_end - h->arena_start) >= MaxArena32)
702 return nil;
704 // On 32-bit, once the reservation is gone we can
705 // try to get memory at a location chosen by the OS
706 // and hope that it is in the range we allocated bitmap for.
707 p_size = ROUND(n, PageSize) + PageSize;
708 p = runtime_SysAlloc(p_size, &mstats.heap_sys);
709 if(p == nil)
710 return nil;
712 if(p < h->arena_start || (uintptr)(p+p_size - h->arena_start) >= MaxArena32) {
713 runtime_printf("runtime: memory allocated by OS (%p) not in usable range [%p,%p)\n",
714 p, h->arena_start, h->arena_start+MaxArena32);
715 runtime_SysFree(p, p_size, &mstats.heap_sys);
716 return nil;
717 }
719 p_end = p + p_size;
720 p += -(uintptr)p & (PageSize-1);
721 if(p+n > h->arena_used) {
722 h->arena_used = p+n;
723 if(p_end > h->arena_end)
724 h->arena_end = p_end;
725 runtime_MHeap_MapBits(h);
726 runtime_MHeap_MapSpans(h);
727 if(raceenabled)
728 runtime_racemapshadow(p, n);
729 }
731 if(((uintptr)p & (PageSize-1)) != 0)
732 runtime_throw("misrounded allocation in MHeap_SysAlloc");
733 return p;
734 }
736 static struct
737 {
738 Lock;
739 byte* pos;
740 byte* end;
741 } persistent;
743 enum
744 {
745 PersistentAllocChunk = 256<<10,
746 PersistentAllocMaxBlock = 64<<10, // VM reservation granularity is 64K on windows
747 };
749 // Wrapper around SysAlloc that can allocate small chunks.
750 // There is no associated free operation.
751 // Intended for things like function/type/debug-related persistent data.
752 // If align is 0, uses default align (currently 8).
753 void*
754 runtime_persistentalloc(uintptr size, uintptr align, uint64 *stat)
755 {
756 byte *p;
758 if(align != 0) {
759 if(align&(align-1))
760 runtime_throw("persistentalloc: align is not a power of 2");
761 if(align > PageSize)
762 runtime_throw("persistentalloc: align is too large");
763 } else
764 align = 8;
765 if(size >= PersistentAllocMaxBlock)
766 return runtime_SysAlloc(size, stat);
767 runtime_lock(&persistent);
768 persistent.pos = (byte*)ROUND((uintptr)persistent.pos, align);
769 if(persistent.pos + size > persistent.end) {
770 persistent.pos = runtime_SysAlloc(PersistentAllocChunk, &mstats.other_sys);
771 if(persistent.pos == nil) {
772 runtime_unlock(&persistent);
773 runtime_throw("runtime: cannot allocate memory");
774 }
775 persistent.end = persistent.pos + PersistentAllocChunk;
776 }
777 p = persistent.pos;
778 persistent.pos += size;
779 runtime_unlock(&persistent);
780 if(stat != &mstats.other_sys) {
781 // reaccount the allocation against provided stat
782 runtime_xadd64(stat, size);
783 runtime_xadd64(&mstats.other_sys, -(uint64)size);
784 }
785 return p;
786 }
788 static void
789 settype(MSpan *s, void *v, uintptr typ)
790 {
791 uintptr size, ofs, j, t;
792 uintptr ntypes, nbytes2, nbytes3;
793 uintptr *data2;
794 byte *data3;
796 if(s->sizeclass == 0) {
797 s->types.compression = MTypes_Single;
798 s->types.data = typ;
799 return;
800 }
801 size = s->elemsize;
802 ofs = ((uintptr)v - (s->start<<PageShift)) / size;
804 switch(s->types.compression) {
805 case MTypes_Empty:
806 ntypes = (s->npages << PageShift) / size;
807 nbytes3 = 8*sizeof(uintptr) + 1*ntypes;
808 data3 = runtime_mallocgc(nbytes3, 0, FlagNoProfiling|FlagNoScan|FlagNoInvokeGC);
809 s->types.compression = MTypes_Bytes;
810 s->types.data = (uintptr)data3;
811 ((uintptr*)data3)[1] = typ;
812 data3[8*sizeof(uintptr) + ofs] = 1;
813 break;
815 case MTypes_Words:
816 ((uintptr*)s->types.data)[ofs] = typ;
817 break;
819 case MTypes_Bytes:
820 data3 = (byte*)s->types.data;
821 for(j=1; j<8; j++) {
822 if(((uintptr*)data3)[j] == typ) {
823 break;
824 }
825 if(((uintptr*)data3)[j] == 0) {
826 ((uintptr*)data3)[j] = typ;
827 break;
828 }
829 }
830 if(j < 8) {
831 data3[8*sizeof(uintptr) + ofs] = j;
832 } else {
833 ntypes = (s->npages << PageShift) / size;
834 nbytes2 = ntypes * sizeof(uintptr);
835 data2 = runtime_mallocgc(nbytes2, 0, FlagNoProfiling|FlagNoScan|FlagNoInvokeGC);
836 s->types.compression = MTypes_Words;
837 s->types.data = (uintptr)data2;
839 // Move the contents of data3 to data2. Then deallocate data3.
840 for(j=0; j<ntypes; j++) {
841 t = data3[8*sizeof(uintptr) + j];
842 t = ((uintptr*)data3)[t];
843 data2[j] = t;
844 }
845 data2[ofs] = typ;
846 }
847 break;
848 }
849 }
851 uintptr
852 runtime_gettype(void *v)
853 {
854 MSpan *s;
855 uintptr t, ofs;
856 byte *data;
858 s = runtime_MHeap_LookupMaybe(&runtime_mheap, v);
859 if(s != nil) {
860 t = 0;
861 switch(s->types.compression) {
862 case MTypes_Empty:
863 break;
864 case MTypes_Single:
865 t = s->types.data;
866 break;
867 case MTypes_Words:
868 ofs = (uintptr)v - (s->start<<PageShift);
869 t = ((uintptr*)s->types.data)[ofs/s->elemsize];
870 break;
871 case MTypes_Bytes:
872 ofs = (uintptr)v - (s->start<<PageShift);
873 data = (byte*)s->types.data;
874 t = data[8*sizeof(uintptr) + ofs/s->elemsize];
875 t = ((uintptr*)data)[t];
876 break;
877 default:
878 runtime_throw("runtime_gettype: invalid compression kind");
879 }
880 if(0) {
881 runtime_printf("%p -> %d,%X\n", v, (int32)s->types.compression, (int64)t);
882 }
883 return t;
884 }
885 return 0;
886 }
888 // Runtime stubs.
890 void*
891 runtime_mal(uintptr n)
892 {
893 return runtime_mallocgc(n, 0, 0);
894 }
896 func new(typ *Type) (ret *uint8) {
897 ret = runtime_mallocgc(typ->__size, (uintptr)typ | TypeInfo_SingleObject, typ->kind&KindNoPointers ? FlagNoScan : 0);
898 }
900 static void*
901 cnew(const Type *typ, intgo n, int32 objtyp)
902 {
903 if((objtyp&(PtrSize-1)) != objtyp)
904 runtime_throw("runtime: invalid objtyp");
905 if(n < 0 || (typ->__size > 0 && (uintptr)n > (MaxMem/typ->__size)))
906 runtime_panicstring("runtime: allocation size out of range");
907 return runtime_mallocgc(typ->__size*n, (uintptr)typ | objtyp, typ->kind&KindNoPointers ? FlagNoScan : 0);
908 }
910 // same as runtime_new, but callable from C
911 void*
912 runtime_cnew(const Type *typ)
913 {
914 return cnew(typ, 1, TypeInfo_SingleObject);
915 }
917 void*
918 runtime_cnewarray(const Type *typ, intgo n)
919 {
920 return cnew(typ, n, TypeInfo_Array);
921 }
923 func GC() {
924 runtime_gc(2); // force GC and do eager sweep
925 }
927 func SetFinalizer(obj Eface, finalizer Eface) {
928 byte *base;
929 uintptr size;
930 const FuncType *ft;
931 const Type *fint;
932 const PtrType *ot;
934 if(obj.__type_descriptor == nil) {
935 runtime_printf("runtime.SetFinalizer: first argument is nil interface\n");
936 goto throw;
937 }
938 if(obj.__type_descriptor->__code != GO_PTR) {
939 runtime_printf("runtime.SetFinalizer: first argument is %S, not pointer\n", *obj.__type_descriptor->__reflection);
940 goto throw;
941 }
942 ot = (const PtrType*)obj.type;
943 // As an implementation detail we do not run finalizers for zero-sized objects,
944 // because we use &runtime_zerobase for all such allocations.
945 if(ot->__element_type != nil && ot->__element_type->__size == 0)
946 return;
947 // The following check is required for cases when a user passes a pointer to composite literal,
948 // but compiler makes it a pointer to global. For example:
949 // var Foo = &Object{}
950 // func main() {
951 // runtime.SetFinalizer(Foo, nil)
952 // }
953 // See issue 7656.
954 if((byte*)obj.__object < runtime_mheap.arena_start || runtime_mheap.arena_used <= (byte*)obj.__object)
955 return;
956 if(!runtime_mlookup(obj.__object, &base, &size, nil) || obj.__object != base) {
957 // As an implementation detail we allow to set finalizers for an inner byte
958 // of an object if it could come from tiny alloc (see mallocgc for details).
959 if(ot->__element_type == nil || (ot->__element_type->__code&KindNoPointers) == 0 || ot->__element_type->__size >= TinySize) {
960 runtime_printf("runtime.SetFinalizer: pointer not at beginning of allocated block (%p)\n", obj.__object);
961 goto throw;
962 }
963 }
964 if(finalizer.__type_descriptor != nil) {
965 runtime_createfing();
966 if(finalizer.__type_descriptor->__code != GO_FUNC)
967 goto badfunc;
968 ft = (const FuncType*)finalizer.__type_descriptor;
969 if(ft->__dotdotdot || ft->__in.__count != 1)
970 goto badfunc;
971 fint = *(Type**)ft->__in.__values;
972 if(__go_type_descriptors_equal(fint, obj.__type_descriptor)) {
973 // ok - same type
974 } else if(fint->__code == GO_PTR && (fint->__uncommon == nil || fint->__uncommon->__name == nil || obj.type->__uncommon == nil || obj.type->__uncommon->__name == nil) && __go_type_descriptors_equal(((const PtrType*)fint)->__element_type, ((const PtrType*)obj.type)->__element_type)) {
975 // ok - not same type, but both pointers,
976 // one or the other is unnamed, and same element type, so assignable.
977 } else if(fint->kind == GO_INTERFACE && ((const InterfaceType*)fint)->__methods.__count == 0) {
978 // ok - satisfies empty interface
979 } else if(fint->kind == GO_INTERFACE && __go_convert_interface_2(fint, obj.__type_descriptor, 1) != nil) {
980 // ok - satisfies non-empty interface
981 } else
982 goto badfunc;
984 ot = (const PtrType*)obj.__type_descriptor;
985 if(!runtime_addfinalizer(obj.__object, *(FuncVal**)finalizer.__object, ft, ot)) {
986 runtime_printf("runtime.SetFinalizer: finalizer already set\n");
987 goto throw;
988 }
989 } else {
990 // NOTE: asking to remove a finalizer when there currently isn't one set is OK.
991 runtime_removefinalizer(obj.__object);
992 }
993 return;
995 badfunc:
996 runtime_printf("runtime.SetFinalizer: cannot pass %S to finalizer %S\n", *obj.__type_descriptor->__reflection, *finalizer.__type_descriptor->__reflection);
997 throw:
998 runtime_throw("runtime.SetFinalizer");
999 }