Go is a language which supports automatic memory management, such as automatic memory allocation and automatic garbage collection. So Go programmers can do programming without handling the underlying verbose memory management. This not only brings much convenience and saves Go programmers lots of time, but also helps Go programmers avoid many careless bugs.
Although knowing the underlying memory management implementation details is not necessary for Go programmers to write Go code, understanding some concepts and being aware of some facts in the memory management implementation by the standard Go compiler and runtime is very helpful for Go programmers to write high quality Go code.
This article will explain some concepts and list some facts of the implementation of memory block allocation and garbage collection by the standard Go compiler and runtime. Other aspects, such as memory apply and memory release in memory management, will not be touched in this article.
A memory block is a continuous memory segment to host value parts at run time. Different memory blocks may have different sizes, to host different value parts. One memory block may host multiple value parts at the same time, but each value part can only be hosted within one memory block, no matter how large the size of that value part is. In other words, for any value part, it never crosses memory blocks.
We have known that a value part can reference another value part.
Here, we extend the reference definition by saying
a memory block is referenced by all the value parts it hosts.
So if a value part v
is referenced by another value part,
then the other value will also reference the memory block hosting v
, indirectly.
new
and make
built-in functions.
A new
call will always allocate exact one memory block.
A make
call will allocate more than one memory blocks to host
the direct part and underlying part(s) of the created slice, map or channel value.
append
function (when the capacity of the base slice is not large enough).
For every Go program compiled by the official standard Go compiler, at run time, each goroutine will maintain a stack, which is a memory segment. It acts as a memory pool for some memory blocks to be allocated from/on. Before Go Toolchain 1.19, the initial size of a stack is always 2KiB. Since Go Toolchain 1.19, the initial size is adaptive. The stack of a goroutine will grow and shrink as needed in goroutine running. The minimum stack size is 2KiB.
(Please note, there is a global limit of stack size each goroutine may reach.
If a goroutine exceeds the limit while growing its stack, the program crashes.
As of Go Toolchain 1.20, the default maximum stack size
is 1 GB on 64-bit systems, and 250 MB on 32-bit systems.
We can call the SetMaxStack
function in the
runtime/debug
standard package to change the size.
And please note that, by the current official standard Go compiler implementation,
the actual allowed maximum stack size is the largest power of 2
which is not larger than then MaxStack setting.
So for the default setting, the actual allowed maximum stack size is 512 MiB
on 64-bit systems, and 128 MiB on 32-bit systems.)
Memory blocks can be allocated on stacks. Memory blocks allocated on the stack of a goroutine can only be used (referenced) in the goroutine internally. They are goroutine localized resources. They are not safe to be referenced crossing goroutines. A goroutine can access or modify the value parts hosted on a memory block allocated on the stack of the goroutine without using any data synchronization techniques.
Heap is a singleton in each program. It is a virtual concept. If a memory block is not allocated on any goroutine stack, then we say the memory block is allocated on heap. Value parts hosted on memory blocks allocated on heap can be used by multiple goroutines. In other words, they can be used concurrently. Their uses should be synchronized when needed.
Heap is a conservative place to allocate memory blocks on. If compilers detect a memory block will be referenced crossing goroutines or can't easily confirm whether or not the memory block is safe to be put on the stack of a goroutine, then the memory block will be allocated on heap at run time. This means some values which can be safely allocated on stacks may also be allocated on heap.
If a memory block is allocated somewhere, we can also say the value parts hosted on the memory block are allocated on the same place.
If some value parts of a local variable declared in a function is allocated on heap,
we can say the value parts (and the variable) escape to heap.
By using Go Toolchain, we can run go build -gcflags -m
to check
which local values (value parts) will escape to heap at run time.
As mentioned above, the current escape analyzer in the standard Go compiler
is still not perfect, many local value parts can be allocated on stacks safely
will still escape to heap.
An active value part allocated on heap still in use must be referenced by at least one value part allocated on a stack.
If a value escaping to heap is a declared local variable, and assume its type is T
,
Go runtime will create (a memory block for)
an implicit pointer of type *T
on the stack of the current goroutine.
The value of the pointer stores the address of the memory block allocated for the variable on heap
(a.k.a., the address of the local variable of type T
).
Go compiler will also replace all uses of the variable with
dereferences of the pointer value at compile time.
The *T
pointer value on stack may be marked as dead since a later time,
so the reference relation from it to the T
value on heap will disappear.
The reference relation from the *T
value on stack to the T
value on heap
plays an important role in the garbage collection process which will be described below.
Similarly, we can view each package-level variable is allocated on heap, and the variable is referenced by an implicit pointer which is allocated on a global memory zone. In fact, the implicit pointer references the direct part of the package-level variable, and the direct part of the variable references some other value parts.
A memory block allocated on heap may be referenced by multiple value parts allocated on different stacks at the same time.
v
is referenced by a value (part) which escapes to heap,
then the value (part) v
will also escape to heap.
A memory block created by calling new
function may be allocated on heap or stacks.
This is different to C++.
package main
// The following directive is to prevent
// calls to the function f being inlined.
//go:noinline
func f(i int) byte {
var a [1<<20]byte // make stack grow
return a[i]
}
func main(){
var x int
println(&x)
f(100)
println(&x)
}
We will find that the two printed addresses are different (as of the standard Go compiler v1.20).
Memory blocks allocated for direct parts of package-level variables will never be collected.
The stack of a goroutine will be collected as a whole when the goroutine exits. So there is no need to collect the memory blocks allocated on stacks, individually, one by one. Stacks are not collected by the garbage collector.
For a memory block allocated on heap, it can be safely collected only if it is no longer referenced (either directly or indirectly) by all the value parts allocated on goroutine stacks and the global memory zone. We call such memory blocks as unused memory blocks. Unused memory blocks on heap will be collected by the garbage collector.
package main
var p *int
func main() {
done := make(chan bool)
// "done" will be used in main and the following
// new goroutine, so it will be allocated on heap.
go func() {
x, y, z := 123, 456, 789
_ = z // z can be allocated on stack safely.
p = &x // For x and y are both ever referenced
p = &y // by the global p, so they will be both
// allocated on heap.
// Now, x is not referenced by anyone, so
// its memory block can be collected now.
p = nil
// Now, y is also not referenced by anyone,
// so its memory block can be collected now.
done <- true
}()
<-done
// Now the above goroutine exits, the done channel
// is not used any more, a smart compiler may
// think it can be collected now.
// ...
}
Sometimes, smart compilers, such as the standard Go compiler, may make some optimizations so that some references are removed earlier than we expect. Here is such an example.
package main
import "fmt"
func main() {
// Assume the length of the slice is so large
// that its elements must be allocated on heap.
bs := make([]byte, 1 << 31)
// A smart compiler can detect that the
// underlying part of the slice bs will never be
// used later, so that the underlying part of the
// slice bs can be garbage collected safely now.
fmt.Println(len(bs))
}
Please read value parts to learn the internal structures of slice values.
By the way, sometimes, we may hope the slice bs
is guaranteed
to not being garbage collected until fmt.Println
is called,
then we can use a runtime.KeepAlive
function call to tell
garbage collectors that the slice bs
and
the value parts referenced by it are still in use.
package main
import "fmt"
import "runtime"
func main() {
bs := make([]int, 1000000)
fmt.Println(len(bs))
// A runtime.KeepAlive(bs) call is also
// okay for this specified example.
runtime.KeepAlive(&bs)
}
The current standard Go compiler (version 1.20) uses a concurrent, tri-color, mark-sweep garbage collector. Here this article will only make a simple explanation for the algorithm.
A garbage collection (GC) process is divided into two phases, the mark phase and the sweep phase. In the mark phase, the collector (a group of goroutines actually) uses the tri-color algorithm to analyze which memory blocks are unused.
The following quote is taken from a Go blog article and is modified a bit to make it clearer.At the start of a GC cycle all heap memory blocks are white. The GC visits all roots, which are objects directly accessible by the application such as globals and things on the stack, and colors these grey. The GC then chooses a grey object, blackens it, and then scans it for pointers to other objects. When this scan finds a pointer to a white memory block, it turns that object grey. This process repeats until there are no more grey objects. At this point, white (heap) memory blocks are known to be unreachable and can be reused.
(About why the algorithm uses three colors instead of two colors, please search "write barrier golang" for details. Here only provides two references: eliminate STW stack re-scanning and mbarrier.go.)
In the sweep phase, the marked unused memory blocks will be collected.
An unused memory block may not be released to OS immediately after it is collected, so that it can be reused for new some value parts. Don't worry, the official Go runtime is much less memory greedy than most Java runtimes.
The GC algorithm is a non-compacting one, so it will not move memory blocks to rearrange them.
Garbage collection processes will consume much CPU resources and some memory resources. So there is not always a garbage collection process in running. A new garbage collection process will be only triggered when some run-time metrics reach certain conditions. How the conditions are defined is a garbage collection pacer problem.
The garbage collection pacer implementation of the official standard Go runtime is still being improved from version to version. So it is hard to describe the implementation precisely and keep the descriptions up-to-date at the same time. Here, I just list some reference articles on this topic:GOGC
and GOMEMLIMIT
environment variables (note that the GOMEMLIMIT
environment variable is only supported since Go 1.19).
The Go 101 프로젝트는 Github 에서 호스팅됩니다. 오타, 문법 오류, 부정확한 표현, 설명 결함, 코드 버그, 끊어진 링크와 같은 모든 종류의 실수에 대한 수정 사항을 제출하여 Go 101을 개선을 돕는 것은 언제나 환영합니다.
주기적으로 Go에 대한 깊이 있는 정보를 얻고 싶다면 Go 101의 공식 트위터 계정인 @go100and1을 팔로우하거나 Go 101 슬랙 채널에j가입해주세요.
reflect
표준 패키지sync
표준 패키지sync/atomic
표준 패키지