type WaitGroup struct {
noCopy noCopy
state atomic.Uint64
sema uint32
}
type noCopy struct{}
func (*noCopy) Lock() {}
func (*noCopy) Unlock() {}
登入後複製
在 Go 中，只需將結構分配給另一個變數即可輕鬆複製結構。但有些結構，例如 WaitGroup，確實不應該被複製。
Copying a WaitGroup can mess things up because the internal state that tracks the goroutines and their synchronization can get out of sync between the copies. If you've read the mutex post, you'll get the idea, imagine what could go wrong if we copied the internal state of a mutex.
The same kind of issues can happen with WaitGroup.
noCopy
The noCopy struct is included in WaitGroup as a way to help prevent copying mistakes, not by throwing errors, but by serving as a warning. It was contributed by Aliaksandr Valialkin, CTO of VictoriaMetrics, and was introduced in change #22015.
The noCopy struct doesn't actually affect how your program runs. Instead, it acts as a marker that tools like go vet can pick up on to detect when a struct has been copied in a way that it shouldn't be.
It has no fields, so it doesn't take up any meaningful space in memory.
It has two methods, Lock and Unlock, which do nothing (no-op). These methods are there just to work with the -copylocks checker in the go vet tool.
When you run go vet on your code, it checks to see if any structs with a noCopy field, like WaitGroup, have been copied in a way that could cause issues.
It will throw an error to let you know there might be a problem. This gives you a heads-up to fix it before it turns into a bug:
func main() {
var a sync.WaitGroup
b := a
fmt.Println(a, b)
}
// go vet:
// assignment copies lock value to b: sync.WaitGroup contains sync.noCopy
// call of fmt.Println copies lock value: sync.WaitGroup contains sync.noCopy
// call of fmt.Println copies lock value: sync.WaitGroup contains sync.noCopy
登入後複製
In this case, go vet will warn you about 3 different spots where the copying happens. You can try it yourself at: Go Playground.
Note that it's purely a safeguard for when we're writing and testing our code, we can still run it like normal.
Internal State
The state of a WaitGroup is stored in an atomic.Uint64 variable. You might have guessed this if you've read the mutex post, there are several things packed into this single value.
WaitGroup structure
Here's how it breaks down:
Counter (high 32 bits): This part keeps track of the number of goroutines the WaitGroup is waiting for. When you call wg.Add() with a positive value, it bumps up this counter, and when you call wg.Done(), it decreases the counter by one.
Waiter (low 32 bits): This tracks the number of goroutines currently waiting for that counter (the high 32 bits) to hit zero. Every time you call wg.Wait(), it increases this "waiter" count. Once the counter reaches zero, it releases all the goroutines that were waiting.
Then there's the final field, sema uint32, which is an internal semaphore managed by the Go runtime.
when a goroutine calls wg.Wait() and the counter isn't zero, it increases the waiter count and then blocks by calling runtime_Semacquire(&wg.sema). This function call puts the goroutine to sleep until it gets woken up by a corresponding runtime_Semrelease(&wg.sema) call.
We'll dive deeper into this in another article, but for now, I want to focus on the alignment issues.
Alignment Problem
I know, talking about history might seem dull, especially when you just want to get to the point. But trust me, knowing the past is the best way to understand where we are now.
Let's take a quick look at how WaitGroup has evolved over several Go versions:
sync.WaitGroup in different Go versions
I can tell you, the core of WaitGroup (the counter, waiter, and semaphore) hasn't really changed across different Go versions. However, the way these elements are structured has been modified many times.
When we talk about alignment, we're referring to the need for data types to be stored at specific memory addresses to allow for efficient access.
For example, on a 64-bit system, a 64-bit value like uint64 should ideally be stored at a memory address that's a multiple of 8 bytes. The reason is, the CPU can grab aligned data in one go, but if the data isn't aligned, it might take multiple operations to access it.
Alignment issues
Now, here's where things get tricky:
On 32-bit architectures, the compiler doesn't guarantee that 64-bit values will be aligned on an 8-byte boundary. Instead, they might only be aligned on a 4-byte boundary.
This becomes a problem when we use the atomic package to perform operations on the state variable. The atomic package specifically notes:
"On ARM, 386, and 32-bit MIPS, it is the caller's responsibility to arrange for 64-bit alignment of 64-bit words accessed atomically via the primitive atomic functions." - atomic package note
What this means is that if we don't align the state uint64 variable to an 8-byte boundary on these 32-bit architectures, it could cause the program to crash.
So, what's the fix? Let's take a look at how this has been handled across different versions.
Go 1.5: state1 [12]byte
I'd recommend taking a moment to guess the underlying logic of this solution as you read the code below, then we'll walk through it together.
Instead of directly using a uint64 for state, WaitGroup sets aside 12 bytes in an array (state1 [12]byte). This might seem like more than you'd need, but there's a reason behind it.
WaitGroup in Go 1.5
The purpose of using 12 bytes is to ensure there's enough room to find an 8-byte segment that's properly aligned.
The full post is available here: https://victoriametrics.com/blog/go-sync-waitgroup/
