path: root/src/pkg/runtime/malloc.goc

// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// See malloc.h for overview.
//
// TODO(rsc): double-check stats.

package runtime
#include "runtime.h"
#include "malloc.h"
#include "defs.h"
#include "type.h"

MHeap runtime·mheap;
extern MStats mstats;	// defined in extern.go

extern volatile int32 runtime·MemProfileRate;

// Same algorithm from chan.c, but a different
// instance of the static uint32 x.
// Not protected by a lock - let the threads use
// the same random number if they like.
static uint32
fastrand1(void)
{
	static uint32 x = 0x49f6428aUL;

	x += x;
	if(x & 0x80000000L)
		x ^= 0x88888eefUL;
	return x;
}

// Allocate an object of at least size bytes.
// Small objects are allocated from the per-thread cache's free lists.
// Large objects (> 32 kB) are allocated straight from the heap.
void*
runtime·mallocgc(uintptr size, uint32 refflag, int32 dogc, int32 zeroed)
{
	int32 sizeclass, rate;
	MCache *c;
	uintptr npages;
	MSpan *s;
	void *v;
	uint32 *ref;

	if(runtime·gcwaiting && g != m->g0 && m->locks == 0)
		runtime·gosched();
	if(m->mallocing)
		runtime·throw("malloc/free - deadlock");
	m->mallocing = 1;
	if(size == 0)
		size = 1;

	mstats.nmalloc++;
	if(size <= MaxSmallSize) {
		// Allocate from mcache free lists.
		sizeclass = runtime·SizeToClass(size);
		size = runtime·class_to_size[sizeclass];
		c = m->mcache;
		v = runtime·MCache_Alloc(c, sizeclass, size, zeroed);
		if(v == nil)
			runtime·throw("out of memory");
		mstats.alloc += size;
		mstats.total_alloc += size;
		mstats.by_size[sizeclass].nmalloc++;

		if(!runtime·mlookup(v, nil, nil, nil, &ref)) {
			runtime·printf("malloc %D; runtime·mlookup failed\n", (uint64)size);
			runtime·throw("malloc runtime·mlookup");
		}
		*ref = RefNone | refflag;
	} else {
		// TODO(rsc): Report tracebacks for very large allocations.

		// Allocate directly from heap.
		npages = size >> PageShift;
		if((size & PageMask) != 0)
			npages++;
		s = runtime·MHeap_Alloc(&runtime·mheap, npages, 0, 1);
		if(s == nil)
			runtime·throw("out of memory");
		size = npages<<PageShift;
		mstats.alloc += size;
		mstats.total_alloc += size;
		v = (void*)(s->start << PageShift);

		// setup for mark sweep
		s->gcref0 = RefNone | refflag;
		ref = &s->gcref0;
	}

	m->mallocing = 0;

	if(!(refflag & RefNoProfiling) && (rate = runtime·MemProfileRate) > 0) {
		if(size >= rate)
			goto profile;
		if(m->mcache->next_sample > size)
			m->mcache->next_sample -= size;
		else {
			// pick next profile time
			if(rate > 0x3fffffff)	// make 2*rate not overflow
				rate = 0x3fffffff;
			m->mcache->next_sample = fastrand1() % (2*rate);
		profile:
			*ref |= RefProfiled;
			runtime·MProf_Malloc(v, size);
		}
	}

	if(dogc && mstats.heap_alloc >= mstats.next_gc)
		runtime·gc(0);
	return v;
}

void*
runtime·malloc(uintptr size)
{
	return runtime·mallocgc(size, 0, 0, 1);
}

// Free the object whose base pointer is v.
void
runtime·free(void *v)
{
	int32 sizeclass, size;
	MSpan *s;
	MCache *c;
	uint32 prof, *ref;

	if(v == nil)
		return;

	if(m->mallocing)
		runtime·throw("malloc/free - deadlock");
	m->mallocing = 1;

	if(!runtime·mlookup(v, nil, nil, &s, &ref)) {
		runtime·printf("free %p: not an allocated block\n", v);
		runtime·throw("free runtime·mlookup");
	}
	prof = *ref & RefProfiled;
	*ref = RefFree;

	// Find size class for v.
	sizeclass = s->sizeclass;
	if(sizeclass == 0) {
		// Large object.
		if(prof)
			runtime·MProf_Free(v, s->npages<<PageShift);
		mstats.alloc -= s->npages<<PageShift;
		runtime·memclr(v, s->npages<<PageShift);
		runtime·MHeap_Free(&runtime·mheap, s, 1);
	} else {
		// Small object.
		c = m->mcache;
		size = runtime·class_to_size[sizeclass];
		if(size > sizeof(uintptr))
			((uintptr*)v)[1] = 1;	// mark as "needs to be zeroed"
		if(prof)
			runtime·MProf_Free(v, size);
		mstats.alloc -= size;
		mstats.by_size[sizeclass].nfree++;
		runtime·MCache_Free(c, v, sizeclass, size);
	}
	m->mallocing = 0;
}

int32
runtime·mlookup(void *v, byte **base, uintptr *size, MSpan **sp, uint32 **ref)
{
	uintptr n, nobj, i;
	byte *p;
	MSpan *s;

	mstats.nlookup++;
	s = runtime·MHeap_LookupMaybe(&runtime·mheap, v);
	if(sp)
		*sp = s;
	if(s == nil) {
		if(base)
			*base = nil;
		if(size)
			*size = 0;
		if(ref)
			*ref = 0;
		return 0;
	}

	p = (byte*)((uintptr)s->start<<PageShift);
	if(s->sizeclass == 0) {
		// Large object.
		if(base)
			*base = p;
		if(size)
			*size = s->npages<<PageShift;
		if(ref)
			*ref = &s->gcref0;
		return 1;
	}

	if((byte*)v >= (byte*)s->gcref) {
		// pointers into the gc ref counts
		// do not count as pointers.
		return 0;
	}

	n = runtime·class_to_size[s->sizeclass];
	i = ((byte*)v - p)/n;
	if(base)
		*base = p + i*n;
	if(size)
		*size = n;

	// good for error checking, but expensive
	if(0) {
		nobj = (s->npages << PageShift) / (n + RefcountOverhead);
		if((byte*)s->gcref < p || (byte*)(s->gcref+nobj) > p+(s->npages<<PageShift)) {
			runtime·printf("odd span state=%d span=%p base=%p sizeclass=%d n=%D size=%D npages=%D\n",
				s->state, s, p, s->sizeclass, (uint64)nobj, (uint64)n, (uint64)s->npages);
			runtime·printf("s->base sizeclass %d v=%p base=%p gcref=%p blocksize=%D nobj=%D size=%D end=%p end=%p\n",
				s->sizeclass, v, p, s->gcref, (uint64)s->npages<<PageShift,
				(uint64)nobj, (uint64)n, s->gcref + nobj, p+(s->npages<<PageShift));
			runtime·throw("bad gcref");
		}
	}
	if(ref)
		*ref = &s->gcref[i];

	return 1;
}

MCache*
runtime·allocmcache(void)
{
	MCache *c;

	runtime·lock(&runtime·mheap);
	c = runtime·FixAlloc_Alloc(&runtime·mheap.cachealloc);
	mstats.mcache_inuse = runtime·mheap.cachealloc.inuse;
	mstats.mcache_sys = runtime·mheap.cachealloc.sys;
	runtime·unlock(&runtime·mheap);
	return c;
}

int32 runtime·sizeof_C_MStats = sizeof(MStats);

void
runtime·mallocinit(void)
{
	byte *p;
	uintptr arena_size;

	runtime·InitSizes();

	if(sizeof(void*) == 8) {
		// On a 64-bit machine, allocate from a single contiguous reservation.
		// 16 GB should be big enough for now.
		//
		// The code will work with the reservation at any address, but ask
		// SysReserve to use 0x000000f800000000 if possible.
		// Allocating a 16 GB region takes away 36 bits, and the amd64
		// doesn't let us choose the top 17 bits, so that leaves the 11 bits
		// in the middle of 0x00f8 for us to choose.  Choosing 0x00f8 means
		// that the valid memory addresses will begin 0x00f8, 0x00f9, 0x00fa, 0x00fb.
		// None of the bytes f8 f9 fa fb can appear in valid UTF-8, and
		// they are otherwise as far from ff (likely a common byte) as possible.
		// Choosing 0x00 for the leading 6 bits was more arbitrary, but it
		// is not a common ASCII code point either.  Using 0x11f8 instead
		// caused out of memory errors on OS X during thread allocations.
		// These choices are both for debuggability and to reduce the
		// odds of the conservative garbage collector not collecting memory
		// because some non-pointer block of memory had a bit pattern
		// that matched a memory address.
		arena_size = 16LL<<30;
		p = runtime·SysReserve((void*)(0x00f8ULL<<32), arena_size);
		if(p == nil)
			runtime·throw("runtime: cannot reserve arena virtual address space");
		runtime·mheap.arena_start = p;
		runtime·mheap.arena_used = p;
		runtime·mheap.arena_end = p + arena_size;
	} else {
		// On a 32-bit machine, we'll take what we can get for each allocation
		// and maintain arena_start and arena_end as min, max we've seen.
		runtime·mheap.arena_start = (byte*)0xffffffff;
		runtime·mheap.arena_end = 0;
	}

	// Initialize the rest of the allocator.	
	runtime·MHeap_Init(&runtime·mheap, runtime·SysAlloc);
	m->mcache = runtime·allocmcache();

	// See if it works.
	runtime·free(runtime·malloc(1));
}

void*
runtime·MHeap_SysAlloc(MHeap *h, uintptr n)
{
	byte *p;
	
	if(sizeof(void*) == 8) {
		// Keep taking from our reservation.
		if(h->arena_end - h->arena_used < n)
			return nil;
		p = h->arena_used;
		runtime·SysMap(p, n);
		h->arena_used += n;
		return p;
	} else {
		// Take what we can get from the OS.
		p = runtime·SysAlloc(n);
		if(p == nil)
			return nil;
		if(p+n > h->arena_used)
			h->arena_used = p+n;
		if(p > h->arena_end)
			h->arena_end = p;
		return p;		
	}
}

// Runtime stubs.

void*
runtime·mal(uintptr n)
{
	return runtime·mallocgc(n, 0, 1, 1);
}

func new(n uint32) (ret *uint8) {
	ret = runtime·mal(n);
}

// Stack allocator uses malloc/free most of the time,
// but if we're in the middle of malloc and need stack,
// we have to do something else to avoid deadlock.
// In that case, we fall back on a fixed-size free-list
// allocator, assuming that inside malloc all the stack
// frames are small, so that all the stack allocations
// will be a single size, the minimum (right now, 5k).
static struct {
	Lock;
	FixAlloc;
} stacks;

enum {
	FixedStack = StackBig + StackExtra
};

void*
runtime·stackalloc(uint32 n)
{
	void *v;
	uint32 *ref;

	if(m->mallocing || m->gcing || n == FixedStack) {
		runtime·lock(&stacks);
		if(stacks.size == 0)
			runtime·FixAlloc_Init(&stacks, n, runtime·SysAlloc, nil, nil);
		if(stacks.size != n) {
			runtime·printf("stackalloc: in malloc, size=%D want %d", (uint64)stacks.size, n);
			runtime·throw("stackalloc");
		}
		v = runtime·FixAlloc_Alloc(&stacks);
		mstats.stacks_inuse = stacks.inuse;
		mstats.stacks_sys = stacks.sys;
		runtime·unlock(&stacks);
		return v;
	}
	v = runtime·mallocgc(n, RefNoProfiling, 0, 0);
	if(!runtime·mlookup(v, nil, nil, nil, &ref))
		runtime·throw("stackalloc runtime·mlookup");
	*ref = RefStack;
	return v;
}

void
runtime·stackfree(void *v, uintptr n)
{
	if(m->mallocing || m->gcing || n == FixedStack) {
		runtime·lock(&stacks);
		runtime·FixAlloc_Free(&stacks, v);
		mstats.stacks_inuse = stacks.inuse;
		mstats.stacks_sys = stacks.sys;
		runtime·unlock(&stacks);
		return;
	}
	runtime·free(v);
}

func Alloc(n uintptr) (p *byte) {
	p = runtime·malloc(n);
}

func Free(p *byte) {
	runtime·free(p);
}

func Lookup(p *byte) (base *byte, size uintptr) {
	runtime·mlookup(p, &base, &size, nil, nil);
}

func GC() {
	runtime·gc(1);
}

func SetFinalizer(obj Eface, finalizer Eface) {
	byte *base;
	uintptr size;
	FuncType *ft;
	int32 i, nret;
	Type *t;

	if(obj.type == nil) {
		runtime·printf("runtime.SetFinalizer: first argument is nil interface\n");
	throw:
		runtime·throw("runtime.SetFinalizer");
	}
	if(obj.type->kind != KindPtr) {
		runtime·printf("runtime.SetFinalizer: first argument is %S, not pointer\n", *obj.type->string);
		goto throw;
	}
	if(!runtime·mlookup(obj.data, &base, &size, nil, nil) || obj.data != base) {
		runtime·printf("runtime.SetFinalizer: pointer not at beginning of allocated block\n");
		goto throw;
	}
	nret = 0;
	if(finalizer.type != nil) {
		if(finalizer.type->kind != KindFunc) {
		badfunc:
			runtime·printf("runtime.SetFinalizer: second argument is %S, not func(%S)\n", *finalizer.type->string, *obj.type->string);
			goto throw;
		}
		ft = (FuncType*)finalizer.type;
		if(ft->dotdotdot || ft->in.len != 1 || *(Type**)ft->in.array != obj.type)
			goto badfunc;

		// compute size needed for return parameters
		for(i=0; i<ft->out.len; i++) {
			t = ((Type**)ft->out.array)[i];
			nret = (nret + t->align - 1) & ~(t->align - 1);
			nret += t->size;
		}
		nret = (nret + sizeof(void*)-1) & ~(sizeof(void*)-1);

		if(runtime·getfinalizer(obj.data, 0)) {
			runtime·printf("runtime.SetFinalizer: finalizer already set");
			goto throw;
		}
	}
	runtime·addfinalizer(obj.data, finalizer.data, nret);
}