| blob | c5e64c893b83e6fff5f7f936227e5bf11a61617f |
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;
87 void *closure;
89 if(size == 0) {
90 // All 0-length allocations use this pointer.
91 // The language does not require the allocations to
92 // have distinct values.
93 return &runtime_zerobase;
94 }
96 m = runtime_m();
97 g = runtime_g();
99 // We should not be called in between __go_set_closure and the
100 // actual function call, but cope with it if we are.
101 closure = g->closure;
103 incallback = false;
104 if(m->mcache == nil && g->ncgo > 0) {
105 // For gccgo this case can occur when a cgo or SWIG function
106 // has an interface return type and the function
107 // returns a non-pointer, so memory allocation occurs
108 // after syscall.Cgocall but before syscall.CgocallDone.
109 // We treat it as a callback.
110 runtime_exitsyscall();
111 m = runtime_m();
112 incallback = true;
113 flag |= FlagNoInvokeGC;
114 }
116 if(runtime_gcwaiting() && g != m->g0 && m->locks == 0 && !(flag & FlagNoInvokeGC)) {
117 runtime_gosched();
118 m = runtime_m();
119 }
120 if(m->mallocing)
121 runtime_throw("malloc/free - deadlock");
122 // Disable preemption during settype.
123 // We can not use m->mallocing for this, because settype calls mallocgc.
124 m->locks++;
125 m->mallocing = 1;
127 if(DebugTypeAtBlockEnd)
128 size += sizeof(uintptr);
130 c = m->mcache;
131 if(!runtime_debug.efence && size <= MaxSmallSize) {
132 if((flag&(FlagNoScan|FlagNoGC)) == FlagNoScan && size < TinySize) {
133 // Tiny allocator.
134 //
135 // Tiny allocator combines several tiny allocation requests
136 // into a single memory block. The resulting memory block
137 // is freed when all subobjects are unreachable. The subobjects
138 // must be FlagNoScan (don't have pointers), this ensures that
139 // the amount of potentially wasted memory is bounded.
140 //
141 // Size of the memory block used for combining (TinySize) is tunable.
142 // Current setting is 16 bytes, which relates to 2x worst case memory
143 // wastage (when all but one subobjects are unreachable).
144 // 8 bytes would result in no wastage at all, but provides less
145 // opportunities for combining.
146 // 32 bytes provides more opportunities for combining,
147 // but can lead to 4x worst case wastage.
148 // The best case winning is 8x regardless of block size.
149 //
150 // Objects obtained from tiny allocator must not be freed explicitly.
151 // So when an object will be freed explicitly, we ensure that
152 // its size >= TinySize.
153 //
154 // SetFinalizer has a special case for objects potentially coming
155 // from tiny allocator, it such case it allows to set finalizers
156 // for an inner byte of a memory block.
157 //
158 // The main targets of tiny allocator are small strings and
159 // standalone escaping variables. On a json benchmark
160 // the allocator reduces number of allocations by ~12% and
161 // reduces heap size by ~20%.
163 tinysize = c->tinysize;
164 if(size <= tinysize) {
165 tiny = c->tiny;
166 // Align tiny pointer for required (conservative) alignment.
167 if((size&7) == 0)
168 tiny = (byte*)ROUND((uintptr)tiny, 8);
169 else if((size&3) == 0)
170 tiny = (byte*)ROUND((uintptr)tiny, 4);
171 else if((size&1) == 0)
172 tiny = (byte*)ROUND((uintptr)tiny, 2);
173 size1 = size + (tiny - c->tiny);
174 if(size1 <= tinysize) {
175 // The object fits into existing tiny block.
176 v = (MLink*)tiny;
177 c->tiny += size1;
178 c->tinysize -= size1;
179 m->mallocing = 0;
180 m->locks--;
181 if(incallback)
182 runtime_entersyscall();
183 g->closure = closure;
184 return v;
185 }
186 }
187 // Allocate a new TinySize block.
188 s = c->alloc[TinySizeClass];
189 if(s->freelist == nil)
190 s = runtime_MCache_Refill(c, TinySizeClass);
191 v = s->freelist;
192 next = v->next;
193 s->freelist = next;
194 s->ref++;
195 if(next != nil) // prefetching nil leads to a DTLB miss
196 PREFETCH(next);
197 ((uint64*)v)[0] = 0;
198 ((uint64*)v)[1] = 0;
199 // See if we need to replace the existing tiny block with the new one
200 // based on amount of remaining free space.
201 if(TinySize-size > tinysize) {
202 c->tiny = (byte*)v + size;
203 c->tinysize = TinySize - size;
204 }
205 size = TinySize;
206 goto done;
207 }
208 // Allocate from mcache free lists.
209 // Inlined version of SizeToClass().
210 if(size <= 1024-8)
211 sizeclass = runtime_size_to_class8[(size+7)>>3];
212 else
213 sizeclass = runtime_size_to_class128[(size-1024+127) >> 7];
214 size = runtime_class_to_size[sizeclass];
215 s = c->alloc[sizeclass];
216 if(s->freelist == nil)
217 s = runtime_MCache_Refill(c, sizeclass);
218 v = s->freelist;
219 next = v->next;
220 s->freelist = next;
221 s->ref++;
222 if(next != nil) // prefetching nil leads to a DTLB miss
223 PREFETCH(next);
224 if(!(flag & FlagNoZero)) {
225 v->next = nil;
226 // block is zeroed iff second word is zero ...
227 if(size > 2*sizeof(uintptr) && ((uintptr*)v)[1] != 0)
228 runtime_memclr((byte*)v, size);
229 }
230 done:
231 c->local_cachealloc += size;
232 } else {
233 // Allocate directly from heap.
234 s = largealloc(flag, &size);
235 v = (void*)(s->start << PageShift);
236 }
238 if(flag & FlagNoGC)
239 runtime_marknogc(v);
240 else if(!(flag & FlagNoScan))
241 runtime_markscan(v);
243 if(DebugTypeAtBlockEnd)
244 *(uintptr*)((uintptr)v+size-sizeof(uintptr)) = typ;
246 m->mallocing = 0;
247 // TODO: save type even if FlagNoScan? Potentially expensive but might help
248 // heap profiling/tracing.
249 if(UseSpanType && !(flag & FlagNoScan) && typ != 0)
250 settype(s, v, typ);
252 if(raceenabled)
253 runtime_racemalloc(v, size);
255 if(runtime_debug.allocfreetrace)
256 runtime_tracealloc(v, size, typ);
258 if(!(flag & FlagNoProfiling) && (rate = runtime_MemProfileRate) > 0) {
259 if(size < (uintptr)rate && size < (uintptr)(uint32)c->next_sample)
260 c->next_sample -= size;
261 else
262 profilealloc(v, size);
263 }
265 m->locks--;
267 if(!(flag & FlagNoInvokeGC) && mstats.heap_alloc >= mstats.next_gc)
268 runtime_gc(0);
270 if(incallback)
271 runtime_entersyscall();
273 g->closure = closure;
275 return v;
276 }
278 static MSpan*
279 largealloc(uint32 flag, uintptr *sizep)
280 {
281 uintptr npages, size;
282 MSpan *s;
283 void *v;
285 // Allocate directly from heap.
286 size = *sizep;
287 if(size + PageSize < size)
288 runtime_throw("out of memory");
289 npages = size >> PageShift;
290 if((size & PageMask) != 0)
291 npages++;
292 s = runtime_MHeap_Alloc(&runtime_mheap, npages, 0, 1, !(flag & FlagNoZero));
293 if(s == nil)
294 runtime_throw("out of memory");
295 s->limit = (byte*)(s->start<<PageShift) + size;
296 *sizep = npages<<PageShift;
297 v = (void*)(s->start << PageShift);
298 // setup for mark sweep
299 runtime_markspan(v, 0, 0, true);
300 return s;
301 }
303 static void
304 profilealloc(void *v, uintptr size)
305 {
306 uintptr rate;
307 int32 next;
308 MCache *c;
310 c = runtime_m()->mcache;
311 rate = runtime_MemProfileRate;
312 if(size < rate) {
313 // pick next profile time
314 // If you change this, also change allocmcache.
315 if(rate > 0x3fffffff) // make 2*rate not overflow
316 rate = 0x3fffffff;
317 next = runtime_fastrand1() % (2*rate);
318 // Subtract the "remainder" of the current allocation.
319 // Otherwise objects that are close in size to sampling rate
320 // will be under-sampled, because we consistently discard this remainder.
321 next -= (size - c->next_sample);
322 if(next < 0)
323 next = 0;
324 c->next_sample = next;
325 }
326 runtime_MProf_Malloc(v, size);
327 }
329 void*
330 __go_alloc(uintptr size)
331 {
332 return runtime_mallocgc(size, 0, FlagNoInvokeGC);
333 }
335 // Free the object whose base pointer is v.
336 void
337 __go_free(void *v)
338 {
339 M *m;
340 int32 sizeclass;
341 MSpan *s;
342 MCache *c;
343 uintptr size;
345 if(v == nil)
346 return;
348 // If you change this also change mgc0.c:/^sweep,
349 // which has a copy of the guts of free.
351 m = runtime_m();
352 if(m->mallocing)
353 runtime_throw("malloc/free - deadlock");
354 m->mallocing = 1;
356 if(!runtime_mlookup(v, nil, nil, &s)) {
357 runtime_printf("free %p: not an allocated block\n", v);
358 runtime_throw("free runtime_mlookup");
359 }
360 size = s->elemsize;
361 sizeclass = s->sizeclass;
362 // Objects that are smaller than TinySize can be allocated using tiny alloc,
363 // if then such object is combined with an object with finalizer, we will crash.
364 if(size < TinySize)
365 runtime_throw("freeing too small block");
367 if(runtime_debug.allocfreetrace)
368 runtime_tracefree(v, size);
370 // Ensure that the span is swept.
371 // If we free into an unswept span, we will corrupt GC bitmaps.
372 runtime_MSpan_EnsureSwept(s);
374 if(s->specials != nil)
375 runtime_freeallspecials(s, v, size);
377 c = m->mcache;
378 if(sizeclass == 0) {
379 // Large object.
380 s->needzero = 1;
381 // Must mark v freed before calling unmarkspan and MHeap_Free:
382 // they might coalesce v into other spans and change the bitmap further.
383 runtime_markfreed(v);
384 runtime_unmarkspan(v, 1<<PageShift);
385 // NOTE(rsc,dvyukov): The original implementation of efence
386 // in CL 22060046 used SysFree instead of SysFault, so that
387 // the operating system would eventually give the memory
388 // back to us again, so that an efence program could run
389 // longer without running out of memory. Unfortunately,
390 // calling SysFree here without any kind of adjustment of the
391 // heap data structures means that when the memory does
392 // come back to us, we have the wrong metadata for it, either in
393 // the MSpan structures or in the garbage collection bitmap.
394 // Using SysFault here means that the program will run out of
395 // memory fairly quickly in efence mode, but at least it won't
396 // have mysterious crashes due to confused memory reuse.
397 // It should be possible to switch back to SysFree if we also
398 // implement and then call some kind of MHeap_DeleteSpan.
399 if(runtime_debug.efence)
400 runtime_SysFault((void*)(s->start<<PageShift), size);
401 else
402 runtime_MHeap_Free(&runtime_mheap, s, 1);
403 c->local_nlargefree++;
404 c->local_largefree += size;
405 } else {
406 // Small object.
407 if(size > 2*sizeof(uintptr))
408 ((uintptr*)v)[1] = (uintptr)0xfeedfeedfeedfeedll; // mark as "needs to be zeroed"
409 else if(size > sizeof(uintptr))
410 ((uintptr*)v)[1] = 0;
411 // Must mark v freed before calling MCache_Free:
412 // it might coalesce v and other blocks into a bigger span
413 // and change the bitmap further.
414 c->local_nsmallfree[sizeclass]++;
415 c->local_cachealloc -= size;
416 if(c->alloc[sizeclass] == s) {
417 // We own the span, so we can just add v to the freelist
418 runtime_markfreed(v);
419 ((MLink*)v)->next = s->freelist;
420 s->freelist = v;
421 s->ref--;
422 } else {
423 // Someone else owns this span. Add to free queue.
424 runtime_MCache_Free(c, v, sizeclass, size);
425 }
426 }
427 m->mallocing = 0;
428 }
430 int32
431 runtime_mlookup(void *v, byte **base, uintptr *size, MSpan **sp)
432 {
433 M *m;
434 uintptr n, i;
435 byte *p;
436 MSpan *s;
438 m = runtime_m();
440 m->mcache->local_nlookup++;
441 if (sizeof(void*) == 4 && m->mcache->local_nlookup >= (1<<30)) {
442 // purge cache stats to prevent overflow
443 runtime_lock(&runtime_mheap);
444 runtime_purgecachedstats(m->mcache);
445 runtime_unlock(&runtime_mheap);
446 }
448 s = runtime_MHeap_LookupMaybe(&runtime_mheap, v);
449 if(sp)
450 *sp = s;
451 if(s == nil) {
452 runtime_checkfreed(v, 1);
453 if(base)
454 *base = nil;
455 if(size)
456 *size = 0;
457 return 0;
458 }
460 p = (byte*)((uintptr)s->start<<PageShift);
461 if(s->sizeclass == 0) {
462 // Large object.
463 if(base)
464 *base = p;
465 if(size)
466 *size = s->npages<<PageShift;
467 return 1;
468 }
470 n = s->elemsize;
471 if(base) {
472 i = ((byte*)v - p)/n;
473 *base = p + i*n;
474 }
475 if(size)
476 *size = n;
478 return 1;
479 }
481 void
482 runtime_purgecachedstats(MCache *c)
483 {
484 MHeap *h;
485 int32 i;
487 // Protected by either heap or GC lock.
488 h = &runtime_mheap;
489 mstats.heap_alloc += c->local_cachealloc;
490 c->local_cachealloc = 0;
491 mstats.nlookup += c->local_nlookup;
492 c->local_nlookup = 0;
493 h->largefree += c->local_largefree;
494 c->local_largefree = 0;
495 h->nlargefree += c->local_nlargefree;
496 c->local_nlargefree = 0;
497 for(i=0; i<(int32)nelem(c->local_nsmallfree); i++) {
498 h->nsmallfree[i] += c->local_nsmallfree[i];
499 c->local_nsmallfree[i] = 0;
500 }
501 }
503 extern uintptr runtime_sizeof_C_MStats
504 __asm__ (GOSYM_PREFIX "runtime.Sizeof_C_MStats");
506 // Size of the trailing by_size array differs between Go and C,
507 // NumSizeClasses was changed, but we can not change Go struct because of backward compatibility.
508 // sizeof_C_MStats is what C thinks about size of Go struct.
510 // Initialized in mallocinit because it's defined in go/runtime/mem.go.
512 #define MaxArena32 (2U<<30)
514 void
515 runtime_mallocinit(void)
516 {
517 byte *p, *p1;
518 uintptr arena_size, bitmap_size, spans_size, p_size;
519 extern byte _end[];
520 uintptr limit;
521 uint64 i;
522 bool reserved;
524 runtime_sizeof_C_MStats = sizeof(MStats) - (NumSizeClasses - 61) * sizeof(mstats.by_size[0]);
526 p = nil;
527 p_size = 0;
528 arena_size = 0;
529 bitmap_size = 0;
530 spans_size = 0;
531 reserved = false;
533 // for 64-bit build
534 USED(p);
535 USED(p_size);
536 USED(arena_size);
537 USED(bitmap_size);
538 USED(spans_size);
540 runtime_InitSizes();
542 if(runtime_class_to_size[TinySizeClass] != TinySize)
543 runtime_throw("bad TinySizeClass");
545 // limit = runtime_memlimit();
546 // See https://code.google.com/p/go/issues/detail?id=5049
547 // TODO(rsc): Fix after 1.1.
548 limit = 0;
550 // Set up the allocation arena, a contiguous area of memory where
551 // allocated data will be found. The arena begins with a bitmap large
552 // enough to hold 4 bits per allocated word.
553 if(sizeof(void*) == 8 && (limit == 0 || limit > (1<<30))) {
554 // On a 64-bit machine, allocate from a single contiguous reservation.
555 // 128 GB (MaxMem) should be big enough for now.
556 //
557 // The code will work with the reservation at any address, but ask
558 // SysReserve to use 0x0000XXc000000000 if possible (XX=00...7f).
559 // Allocating a 128 GB region takes away 37 bits, and the amd64
560 // doesn't let us choose the top 17 bits, so that leaves the 11 bits
561 // in the middle of 0x00c0 for us to choose. Choosing 0x00c0 means
562 // that the valid memory addresses will begin 0x00c0, 0x00c1, ..., 0x00df.
563 // In little-endian, that's c0 00, c1 00, ..., df 00. None of those are valid
564 // UTF-8 sequences, and they are otherwise as far away from
565 // ff (likely a common byte) as possible. If that fails, we try other 0xXXc0
566 // addresses. An earlier attempt to use 0x11f8 caused out of memory errors
567 // on OS X during thread allocations. 0x00c0 causes conflicts with
568 // AddressSanitizer which reserves all memory up to 0x0100.
569 // These choices are both for debuggability and to reduce the
570 // odds of the conservative garbage collector not collecting memory
571 // because some non-pointer block of memory had a bit pattern
572 // that matched a memory address.
573 //
574 // Actually we reserve 136 GB (because the bitmap ends up being 8 GB)
575 // but it hardly matters: e0 00 is not valid UTF-8 either.
576 //
577 // If this fails we fall back to the 32 bit memory mechanism
578 arena_size = MaxMem;
579 bitmap_size = arena_size / (sizeof(void*)*8/4);
580 spans_size = arena_size / PageSize * sizeof(runtime_mheap.spans[0]);
581 spans_size = ROUND(spans_size, PageSize);
582 for(i = 0; i < HeapBaseOptions; i++) {
583 p = HeapBase(i);
584 p_size = bitmap_size + spans_size + arena_size + PageSize;
585 p = runtime_SysReserve(p, p_size, &reserved);
586 if(p != nil)
587 break;
588 }
589 }
590 if (p == nil) {
591 // On a 32-bit machine, we can't typically get away
592 // with a giant virtual address space reservation.
593 // Instead we map the memory information bitmap
594 // immediately after the data segment, large enough
595 // to handle another 2GB of mappings (256 MB),
596 // along with a reservation for another 512 MB of memory.
597 // When that gets used up, we'll start asking the kernel
598 // for any memory anywhere and hope it's in the 2GB
599 // following the bitmap (presumably the executable begins
600 // near the bottom of memory, so we'll have to use up
601 // most of memory before the kernel resorts to giving out
602 // memory before the beginning of the text segment).
603 //
604 // Alternatively we could reserve 512 MB bitmap, enough
605 // for 4GB of mappings, and then accept any memory the
606 // kernel threw at us, but normally that's a waste of 512 MB
607 // of address space, which is probably too much in a 32-bit world.
608 bitmap_size = MaxArena32 / (sizeof(void*)*8/4);
609 arena_size = 512<<20;
610 spans_size = MaxArena32 / PageSize * sizeof(runtime_mheap.spans[0]);
611 if(limit > 0 && arena_size+bitmap_size+spans_size > limit) {
612 bitmap_size = (limit / 9) & ~((1<<PageShift) - 1);
613 arena_size = bitmap_size * 8;
614 spans_size = arena_size / PageSize * sizeof(runtime_mheap.spans[0]);
615 }
616 spans_size = ROUND(spans_size, PageSize);
618 // SysReserve treats the address we ask for, end, as a hint,
619 // not as an absolute requirement. If we ask for the end
620 // of the data segment but the operating system requires
621 // a little more space before we can start allocating, it will
622 // give out a slightly higher pointer. Except QEMU, which
623 // is buggy, as usual: it won't adjust the pointer upward.
624 // So adjust it upward a little bit ourselves: 1/4 MB to get
625 // away from the running binary image and then round up
626 // to a MB boundary.
627 p = (byte*)ROUND((uintptr)_end + (1<<18), 1<<20);
628 p_size = bitmap_size + spans_size + arena_size + PageSize;
629 p = runtime_SysReserve(p, p_size, &reserved);
630 if(p == nil)
631 runtime_throw("runtime: cannot reserve arena virtual address space");
632 }
634 // PageSize can be larger than OS definition of page size,
635 // so SysReserve can give us a PageSize-unaligned pointer.
636 // To overcome this we ask for PageSize more and round up the pointer.
637 p1 = (byte*)ROUND((uintptr)p, PageSize);
639 runtime_mheap.spans = (MSpan**)p1;
640 runtime_mheap.bitmap = p1 + spans_size;
641 runtime_mheap.arena_start = p1 + spans_size + bitmap_size;
642 runtime_mheap.arena_used = runtime_mheap.arena_start;
643 runtime_mheap.arena_end = p + p_size;
644 runtime_mheap.arena_reserved = reserved;
646 if(((uintptr)runtime_mheap.arena_start & (PageSize-1)) != 0)
647 runtime_throw("misrounded allocation in mallocinit");
649 // Initialize the rest of the allocator.
650 runtime_MHeap_Init(&runtime_mheap);
651 runtime_m()->mcache = runtime_allocmcache();
653 // See if it works.
654 runtime_free(runtime_malloc(TinySize));
655 }
657 void*
658 runtime_MHeap_SysAlloc(MHeap *h, uintptr n)
659 {
660 byte *p, *p_end;
661 uintptr p_size;
662 bool reserved;
665 if(n > (uintptr)(h->arena_end - h->arena_used)) {
666 // We are in 32-bit mode, maybe we didn't use all possible address space yet.
667 // Reserve some more space.
668 byte *new_end;
670 p_size = ROUND(n + PageSize, 256<<20);
671 new_end = h->arena_end + p_size;
672 if(new_end <= h->arena_start + MaxArena32) {
673 // TODO: It would be bad if part of the arena
674 // is reserved and part is not.
675 p = runtime_SysReserve(h->arena_end, p_size, &reserved);
676 if(p == h->arena_end) {
677 h->arena_end = new_end;
678 h->arena_reserved = reserved;
679 }
680 else if(p+p_size <= h->arena_start + MaxArena32) {
681 // Keep everything page-aligned.
682 // Our pages are bigger than hardware pages.
683 h->arena_end = p+p_size;
684 h->arena_used = p + (-(uintptr)p&(PageSize-1));
685 h->arena_reserved = reserved;
686 } else {
687 uint64 stat;
688 stat = 0;
689 runtime_SysFree(p, p_size, &stat);
690 }
691 }
692 }
693 if(n <= (uintptr)(h->arena_end - h->arena_used)) {
694 // Keep taking from our reservation.
695 p = h->arena_used;
696 runtime_SysMap(p, n, h->arena_reserved, &mstats.heap_sys);
697 h->arena_used += n;
698 runtime_MHeap_MapBits(h);
699 runtime_MHeap_MapSpans(h);
700 if(raceenabled)
701 runtime_racemapshadow(p, n);
703 if(((uintptr)p & (PageSize-1)) != 0)
704 runtime_throw("misrounded allocation in MHeap_SysAlloc");
705 return p;
706 }
708 // If using 64-bit, our reservation is all we have.
709 if((uintptr)(h->arena_end - h->arena_start) >= MaxArena32)
710 return nil;
712 // On 32-bit, once the reservation is gone we can
713 // try to get memory at a location chosen by the OS
714 // and hope that it is in the range we allocated bitmap for.
715 p_size = ROUND(n, PageSize) + PageSize;
716 p = runtime_SysAlloc(p_size, &mstats.heap_sys);
717 if(p == nil)
718 return nil;
720 if(p < h->arena_start || (uintptr)(p+p_size - h->arena_start) >= MaxArena32) {
721 runtime_printf("runtime: memory allocated by OS (%p) not in usable range [%p,%p)\n",
722 p, h->arena_start, h->arena_start+MaxArena32);
723 runtime_SysFree(p, p_size, &mstats.heap_sys);
724 return nil;
725 }
727 p_end = p + p_size;
728 p += -(uintptr)p & (PageSize-1);
729 if(p+n > h->arena_used) {
730 h->arena_used = p+n;
731 if(p_end > h->arena_end)
732 h->arena_end = p_end;
733 runtime_MHeap_MapBits(h);
734 runtime_MHeap_MapSpans(h);
735 if(raceenabled)
736 runtime_racemapshadow(p, n);
737 }
739 if(((uintptr)p & (PageSize-1)) != 0)
740 runtime_throw("misrounded allocation in MHeap_SysAlloc");
741 return p;
742 }
744 static struct
745 {
746 Lock;
747 byte* pos;
748 byte* end;
749 } persistent;
751 enum
752 {
753 PersistentAllocChunk = 256<<10,
754 PersistentAllocMaxBlock = 64<<10, // VM reservation granularity is 64K on windows
755 };
757 // Wrapper around SysAlloc that can allocate small chunks.
758 // There is no associated free operation.
759 // Intended for things like function/type/debug-related persistent data.
760 // If align is 0, uses default align (currently 8).
761 void*
762 runtime_persistentalloc(uintptr size, uintptr align, uint64 *stat)
763 {
764 byte *p;
766 if(align != 0) {
767 if(align&(align-1))
768 runtime_throw("persistentalloc: align is not a power of 2");
769 if(align > PageSize)
770 runtime_throw("persistentalloc: align is too large");
771 } else
772 align = 8;
773 if(size >= PersistentAllocMaxBlock)
774 return runtime_SysAlloc(size, stat);
775 runtime_lock(&persistent);
776 persistent.pos = (byte*)ROUND((uintptr)persistent.pos, align);
777 if(persistent.pos + size > persistent.end) {
778 persistent.pos = runtime_SysAlloc(PersistentAllocChunk, &mstats.other_sys);
779 if(persistent.pos == nil) {
780 runtime_unlock(&persistent);
781 runtime_throw("runtime: cannot allocate memory");
782 }
783 persistent.end = persistent.pos + PersistentAllocChunk;
784 }
785 p = persistent.pos;
786 persistent.pos += size;
787 runtime_unlock(&persistent);
788 if(stat != &mstats.other_sys) {
789 // reaccount the allocation against provided stat
790 runtime_xadd64(stat, size);
791 runtime_xadd64(&mstats.other_sys, -(uint64)size);
792 }
793 return p;
794 }
796 static void
797 settype(MSpan *s, void *v, uintptr typ)
798 {
799 uintptr size, ofs, j, t;
800 uintptr ntypes, nbytes2, nbytes3;
801 uintptr *data2;
802 byte *data3;
804 if(s->sizeclass == 0) {
805 s->types.compression = MTypes_Single;
806 s->types.data = typ;
807 return;
808 }
809 size = s->elemsize;
810 ofs = ((uintptr)v - (s->start<<PageShift)) / size;
812 switch(s->types.compression) {
813 case MTypes_Empty:
814 ntypes = (s->npages << PageShift) / size;
815 nbytes3 = 8*sizeof(uintptr) + 1*ntypes;
816 data3 = runtime_mallocgc(nbytes3, 0, FlagNoProfiling|FlagNoScan|FlagNoInvokeGC);
817 s->types.compression = MTypes_Bytes;
818 s->types.data = (uintptr)data3;
819 ((uintptr*)data3)[1] = typ;
820 data3[8*sizeof(uintptr) + ofs] = 1;
821 break;
823 case MTypes_Words:
824 ((uintptr*)s->types.data)[ofs] = typ;
825 break;
827 case MTypes_Bytes:
828 data3 = (byte*)s->types.data;
829 for(j=1; j<8; j++) {
830 if(((uintptr*)data3)[j] == typ) {
831 break;
832 }
833 if(((uintptr*)data3)[j] == 0) {
834 ((uintptr*)data3)[j] = typ;
835 break;
836 }
837 }
838 if(j < 8) {
839 data3[8*sizeof(uintptr) + ofs] = j;
840 } else {
841 ntypes = (s->npages << PageShift) / size;
842 nbytes2 = ntypes * sizeof(uintptr);
843 data2 = runtime_mallocgc(nbytes2, 0, FlagNoProfiling|FlagNoScan|FlagNoInvokeGC);
844 s->types.compression = MTypes_Words;
845 s->types.data = (uintptr)data2;
847 // Move the contents of data3 to data2. Then deallocate data3.
848 for(j=0; j<ntypes; j++) {
849 t = data3[8*sizeof(uintptr) + j];
850 t = ((uintptr*)data3)[t];
851 data2[j] = t;
852 }
853 data2[ofs] = typ;
854 }
855 break;
856 }
857 }
859 uintptr
860 runtime_gettype(void *v)
861 {
862 MSpan *s;
863 uintptr t, ofs;
864 byte *data;
866 s = runtime_MHeap_LookupMaybe(&runtime_mheap, v);
867 if(s != nil) {
868 t = 0;
869 switch(s->types.compression) {
870 case MTypes_Empty:
871 break;
872 case MTypes_Single:
873 t = s->types.data;
874 break;
875 case MTypes_Words:
876 ofs = (uintptr)v - (s->start<<PageShift);
877 t = ((uintptr*)s->types.data)[ofs/s->elemsize];
878 break;
879 case MTypes_Bytes:
880 ofs = (uintptr)v - (s->start<<PageShift);
881 data = (byte*)s->types.data;
882 t = data[8*sizeof(uintptr) + ofs/s->elemsize];
883 t = ((uintptr*)data)[t];
884 break;
885 default:
886 runtime_throw("runtime_gettype: invalid compression kind");
887 }
888 if(0) {
889 runtime_printf("%p -> %d,%X\n", v, (int32)s->types.compression, (int64)t);
890 }
891 return t;
892 }
893 return 0;
894 }
896 // Runtime stubs.
898 void*
899 runtime_mal(uintptr n)
900 {
901 return runtime_mallocgc(n, 0, 0);
902 }
904 func new(typ *Type) (ret *uint8) {
905 ret = runtime_mallocgc(typ->__size, (uintptr)typ | TypeInfo_SingleObject, typ->kind&KindNoPointers ? FlagNoScan : 0);
906 }
908 static void*
909 cnew(const Type *typ, intgo n, int32 objtyp)
910 {
911 if((objtyp&(PtrSize-1)) != objtyp)
912 runtime_throw("runtime: invalid objtyp");
913 if(n < 0 || (typ->__size > 0 && (uintptr)n > (MaxMem/typ->__size)))
914 runtime_panicstring("runtime: allocation size out of range");
915 return runtime_mallocgc(typ->__size*n, (uintptr)typ | objtyp, typ->kind&KindNoPointers ? FlagNoScan : 0);
916 }
918 // same as runtime_new, but callable from C
919 void*
920 runtime_cnew(const Type *typ)
921 {
922 return cnew(typ, 1, TypeInfo_SingleObject);
923 }
925 void*
926 runtime_cnewarray(const Type *typ, intgo n)
927 {
928 return cnew(typ, n, TypeInfo_Array);
929 }
931 func GC() {
932 runtime_gc(2); // force GC and do eager sweep
933 }
935 func SetFinalizer(obj Eface, finalizer Eface) {
936 byte *base;
937 uintptr size;
938 const FuncType *ft;
939 const Type *fint;
940 const PtrType *ot;
942 if(obj.__type_descriptor == nil) {
943 runtime_printf("runtime.SetFinalizer: first argument is nil interface\n");
944 goto throw;
945 }
946 if(obj.__type_descriptor->__code != GO_PTR) {
947 runtime_printf("runtime.SetFinalizer: first argument is %S, not pointer\n", *obj.__type_descriptor->__reflection);
948 goto throw;
949 }
950 ot = (const PtrType*)obj.type;
951 // As an implementation detail we do not run finalizers for zero-sized objects,
952 // because we use &runtime_zerobase for all such allocations.
953 if(ot->__element_type != nil && ot->__element_type->__size == 0)
954 return;
955 // The following check is required for cases when a user passes a pointer to composite literal,
956 // but compiler makes it a pointer to global. For example:
957 // var Foo = &Object{}
958 // func main() {
959 // runtime.SetFinalizer(Foo, nil)
960 // }
961 // See issue 7656.
962 if((byte*)obj.__object < runtime_mheap.arena_start || runtime_mheap.arena_used <= (byte*)obj.__object)
963 return;
964 if(!runtime_mlookup(obj.__object, &base, &size, nil) || obj.__object != base) {
965 // As an implementation detail we allow to set finalizers for an inner byte
966 // of an object if it could come from tiny alloc (see mallocgc for details).
967 if(ot->__element_type == nil || (ot->__element_type->__code&KindNoPointers) == 0 || ot->__element_type->__size >= TinySize) {
968 runtime_printf("runtime.SetFinalizer: pointer not at beginning of allocated block (%p)\n", obj.__object);
969 goto throw;
970 }
971 }
972 if(finalizer.__type_descriptor != nil) {
973 runtime_createfing();
974 if(finalizer.__type_descriptor->__code != GO_FUNC)
975 goto badfunc;
976 ft = (const FuncType*)finalizer.__type_descriptor;
977 if(ft->__dotdotdot || ft->__in.__count != 1)
978 goto badfunc;
979 fint = *(Type**)ft->__in.__values;
980 if(__go_type_descriptors_equal(fint, obj.__type_descriptor)) {
981 // ok - same type
982 } 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)) {
983 // ok - not same type, but both pointers,
984 // one or the other is unnamed, and same element type, so assignable.
985 } else if(fint->kind == GO_INTERFACE && ((const InterfaceType*)fint)->__methods.__count == 0) {
986 // ok - satisfies empty interface
987 } else if(fint->kind == GO_INTERFACE && __go_convert_interface_2(fint, obj.__type_descriptor, 1) != nil) {
988 // ok - satisfies non-empty interface
989 } else
990 goto badfunc;
992 ot = (const PtrType*)obj.__type_descriptor;
993 if(!runtime_addfinalizer(obj.__object, *(FuncVal**)finalizer.__object, ft, ot)) {
994 runtime_printf("runtime.SetFinalizer: finalizer already set\n");
995 goto throw;
996 }
997 } else {
998 // NOTE: asking to remove a finalizer when there currently isn't one set is OK.
999 runtime_removefinalizer(obj.__object);
1000 }
1001 return;
1003 badfunc:
1004 runtime_printf("runtime.SetFinalizer: cannot pass %S to finalizer %S\n", *obj.__type_descriptor->__reflection, *finalizer.__type_descriptor->__reflection);
1005 throw:
1006 runtime_throw("runtime.SetFinalizer");
1007 }