Go Practices 101
Theme: dark/light
Three new books, Go Optimizations 101,
Go Details & Tips 101
and Go Generics 101
are published now.
It is most cost-effective to buy all of them through this book bundle
in the Leanpub book store.
Channel is an important built-in feature in Go. It is one of the features that makes Go unique. Along with another unique feature,
goroutine, channel makes concurrent programming convenient, fun and lowers the difficulties of concurrent programming.
Channel mainly acts as a
concurrency synchronization technique. This article will list all the channel related concepts, syntax and rules. To understand channels better, the internal structure of channels and some implementation details by the standard Go compiler/runtime are also simply described.
The information in this article may be slightly challenging for new gophers. Some parts of this article may need to be read several times to be fully understood.
One suggestion (made by Rob Pike) for concurrent programming is don't (let computations) communicate by sharing memory, (let them) share memory by communicating (through channels). (We can view each computation as a goroutine in Go programming.)
Communicating by sharing memory and sharing memory by communicating are two programming manners in concurrent programming. When goroutines communicate by sharing memory, we use traditional concurrency synchronization techniques, such as mutex locks, to protect the shared memory to prevent data races. We can use channels to implement sharing memory by communicating.
Go provides a unique concurrency synchronization technique, channel. Channels make goroutines share memory by communicating. We can view a channel as an internal FIFO (first in, first out) queue within a program. Some goroutines send values to the queue (the channel) and some other goroutines receive values from the queue.
Along with transferring values (through channels), the ownership of some values may also be transferred between goroutines. When a goroutine sends a value to a channel, we can view the goroutine releases the ownership of some values (which could be accessed through the sent value). When a goroutine receives a value from a channel, we can view the goroutine acquires the ownership of some values (which could be accessed through the received value).
Surely, there may also not be any ownership transferred along with channel communications.
The values (whose ownerships are transferred) are often referenced (but are not required to be referenced) by the transferred value. Please note, here, when we talk about ownership, we mean the ownership from the logic view. Go channels can help programmers write data races free code easily, but Go channels can't prevent programmers from writing bad concurrent code from the syntax level.
Although Go also supports traditional concurrency synchronization techniques. only channel is first-class citizen in Go. Channel is one kind of types in Go, so we can use channels without importing any packages. On the other hand, those traditional concurrency synchronization techniques are provided in the sync
and sync/atomic
standard packages.
Honestly, each concurrency synchronization technique has its own best use scenarios. But channel has
a wider application range and has more variety in using. One problem of channels is, the experience of programming with channels is so enjoyable and fun that programmers often even prefer to use channels for the scenarios which channels are not best for.
Like array, slice and map, each channel type has an element type. A channel can only transfer values of the element type of the channel.
Channel types can be bi-directional or single-directional. Assume T
is an arbitrary type,
-
chan T
denotes a bidirectional channel type. Compilers allow both receiving values from and sending values to bidirectional channels.
-
chan<- T
denotes a send-only channel type. Compilers don't allow receiving values from send-only channels.
-
<-chan T
denotes a receive-only channel type. Compilers don't allow sending values to receive-only channels.
T
is called the element type of these channel types.
Values of bidirectional channel type chan T
can be implicitly converted to both send-only type chan<- T
and receive-only type <-chan T
, but not vice versa (even if explicitly). Values of send-only type chan<- T
can't be converted to receive-only type <-chan T
, and vice versa. Note that the <-
signs in channel type literals are modifiers.
Each channel value has a capacity, which will be explained in the section after next. A channel value with a zero capacity is called unbuffered channel and a channel value with a non-zero capacity is called buffered channel.
The zero values of channel types are represented with the predeclared identifier nil
. A non-nil channel value must be created by using the built-in make
function. For example, make(chan int, 10)
will create a channel whose element type is int
. The second argument of the make
function call specifies the capacity of the new created channel. The second parameter is optional and its default value is zero.
All channel types are comparable types.
From the article
value parts, we know that non-nil channel values are multi-part values. If one channel value is assigned to another, the two channels share the same underlying part(s). In other words, those two channels represent the same internal channel object. The result of comparing them is
true
.
There are five channel specified operations. Assume the channel is ch
, their syntax and function calls of these operations are listed here.
-
Close the channel by using the following function call
close(ch)
where close
is a built-in function. The argument of a close
function call must be a channel value, and the channel ch
must not be a receive-only channel.
-
Send a value, v
, to the channel by using the following syntax
ch <- v
where v
must be a value which is assignable to the element type of channel ch
, and the channel ch
must not be a receive-only channel. Note that here <-
is a channel-send operator.
-
Receive a value from the channel by using the following syntax
<-ch
A channel receive operation always returns at least one result, which is a value of the element type of the channel, and the channel ch
must not be a send-only channel. Note that here <-
is a channel-receive operator. Yes, its representation is the same as a channel-send operator.
For most scenarios, a channel receive operation is viewed as a single-value expression. However, when a channel operation is used as the only source value expression in an assignment, it can result a second optional untyped boolean value and become a multi-value expression. The untyped boolean value indicates whether or not the first result is sent before the channel is closed. (Below we will learn that we can receive unlimited number of values from a closed channel.)
Two channel receive operations which are used as source values in assignments:
v = <-ch
v, sentBeforeClosed = <-ch
-
Query the value buffer capacity of the channel by using the following function call
cap(ch)
where
cap
is a built-in function which has ever been introduced in
containers in Go. The return result of a
cap
function call is an
int
value.
-
Query the current number of values in the value buffer (or the length) of the channel by using the following function call
len(ch)
where len
is a built-in function which also has ever been introduced before. The return value of a len
function call is an int
value. The result length is the number of elements which have already been sent successfully to the queried channel but haven't been received (taken out) yet.
Most basic operations in Go are not synchronized. In other words, they are not concurrency-safe. These operations include value assignments, argument passing and container element manipulations, etc. However, all the just introduced channel operations are already synchronized, so no further synchronizations are needed to safely perform these operations.
Like most other operations in Go, channel value assignments are not synchronized. Similarly, assigning the received value to another value is also not synchronized, though any channel receive operation is synchronized.
If the queried channel is a nil channel, both of the built-in cap
and len
functions return zero. The two query operations are so simple that they will not get further explanations later. In fact, the two operations are seldom used in practice.
Channel send, receive and close operations will be explained in detail in the next section.
To make the explanations for channel operations simple and clear, in the remaining of this article, channels will be classified into three categories:
-
nil channels.
-
non-nil but closed channels.
-
not-closed non-nil channels.
The following table simply summarizes the behaviors for all kinds of operations applying on nil, closed and not-closed non-nil channels.
Operation
|
A Nil Channel
|
A Closed Channel
|
A Not-Closed Non-Nil Channel
|
Close
|
panic
|
panic
|
succeed to close (C)
|
Send Value To
|
block for ever
|
panic
|
block or succeed to send (B)
|
Receive Value From
|
block for ever
|
never block (D)
|
block or succeed to receive (A)
|
For the five cases shown without superscripts, the behaviors are very clear.
-
Closing a nil or an already closed channel produces a panic in the current goroutine.
-
Sending a value to a closed channel also produces a panic in the current goroutine.
-
Sending a value to or receiving a value from a nil channel makes the current goroutine enter and stay in blocking state for ever.
The following will make more explanations for the four cases shown with superscripts (A, B, C and D).
To better understand channel types and values, and to make some explanations easier, looking in the raw internal structures of internal channel objects is very helpful.
We can think of each channel consisting of three queues (all can be viewed as FIFO queues) internally:
-
the receiving goroutine queue (generally FIFO). The queue is a linked list without size limitation. Goroutines in this queue are all in blocking state and waiting to receive values from that channel.
-
the sending goroutine queue (generally FIFO). The queue is also a linked list without size limitation. Goroutines in this queue are all in blocking state and waiting to send values to that channel. The value (or the address of the value, depending on compiler implementation) each goroutine is trying to send is also stored in the queue along with that goroutine.
-
the value buffer queue (absolutely FIFO). This is a circular queue. Its size is equal to the capacity of the channel. The types of the values stored in this buffer queue are all the element type of that channel. If the current number of values stored in the value buffer queue of the channel reaches the capacity of the channel, the channel is called in full status. If no values are stored in the value buffer queue of the channel currently, the channel is called in empty status. For a zero-capacity (unbuffered) channel, it is always in both full and empty status.
Each channel internally holds a mutex lock which is used to avoid data races in all kinds of operations.
Channel operation case A: when a goroutine R
tries to receive a value from a not-closed non-nil channel, the goroutine R
will acquire the lock associated with the channel firstly, then do the following steps until one condition is satisfied.
-
If the value buffer queue of the channel is not empty, in which case the receiving goroutine queue of the channel must be empty, the goroutine R
will receive (by shifting) a value from the value buffer queue. If the sending goroutine queue of the channel is also not empty, a sending goroutine will be shifted out of the sending goroutine queue and resumed to running state again. The value that the just shifted sending goroutine is trying to send will be pushed into the value buffer queue of the channel. The receiving goroutine R
continues running. For this scenario, the channel receive operation is called a non-blocking operation.
-
Otherwise (the value buffer queue of the channel is empty), if the sending goroutine queue of the channel is not empty, in which case the channel must be an unbuffered channel, the receiving goroutine R
will shift a sending goroutine from the sending goroutine queue of the channel and receive the value that the just shifted sending goroutine is trying to send. The just shifted sending goroutine will get unblocked and resumed to running state again. The receiving goroutine R
continues running. For this scenario, the channel receive operation is called a non-blocking operation.
-
If value buffer queue and the sending goroutine queue of the channel are both empty, the goroutine R
will be pushed into the receiving goroutine queue of the channel and enter (and stay in) blocking state. It may be resumed to running state when another goroutine sends a value to the channel later. For this scenario, the channel receive operation is called a blocking operation.
Channel rule case B: when a goroutine S
tries to send a value to a not-closed non-nil channel, the goroutine S
will acquire the lock associated with the channel firstly, then do the following steps until one step condition is satisfied.
-
If the receiving goroutine queue of the channel is not empty, in which case the value buffer queue of the channel must be empty, the sending goroutine S
will shift a receiving goroutine from the receiving goroutine queue of the channel and send the value to the just shifted receiving goroutine. The just shifted receiving goroutine will get unblocked and resumed to running state again. The sending goroutine S
continues running. For this scenario, the channel send operation is called a non-blocking operation.
-
Otherwise (the receiving goroutine queue is empty), if the value buffer queue of the channel is not full, in which case the sending goroutine queue must be also empty, the value the sending goroutine S
trying to send will be pushed into the value buffer queue, and the sending goroutine S
continues running. For this scenario, the channel send operation is called a non-blocking operation.
-
If the receiving goroutine queue is empty and the value buffer queue of the channel is already full, the sending goroutine S
will be pushed into the sending goroutine queue of the channel and enter (and stay in) blocking state. It may be resumed to running state when another goroutine receives a value from the channel later. For this scenario, the channel send operation is called a blocking operation.
Above has mentioned, once a non-nil channel is closed, sending a value to the channel will produce a runtime panic in the current goroutine. Note, sending data to a closed channel is viewed as a non-blocking operation.
Channel operation case C: when a goroutine tries to close a not-closed non-nil channel, once the goroutine has acquired the lock of the channel, both of the following two steps will be performed by the following order.
-
If the receiving goroutine queue of the channel is not empty, in which case the value buffer of the channel must be empty, all the goroutines in the receiving goroutine queue of the channel will be shifted one by one, each of them will receive a zero value of the element type of the channel and be resumed to running state.
-
If the sending goroutine queue of the channel is not empty, all the goroutines in the sending goroutine queue of the channel will be shifted one by one and each of them will produce a panic for sending on a closed channel. This is the reason why we should avoid concurrent send and close operations on the same channel. In fact, when
the go command's data race detector option (
-race
) is enabled, concurrent send and close operation cases might be detected at run time and a runtime panic will be thrown.
Note: after a channel is closed, the values which have been already pushed into the value buffer of the channel are still there. Please read the closely following explanations for case D for details.
Channel operation case D: after a non-nil channel is closed, channel receive operations on the channel will never block. The values in the value buffer of the channel can still be received. The accompanying second optional bool return values are still true
. Once all the values in the value buffer are taken out and received, infinite zero values of the element type of the channel will be received by any of the following receive operations on the channel. As mentioned above, the optional second return result of a channel receive operation is an untyped boolean value which indicates whether or not the first result (the received value) is sent before the channel is closed. If the second return result is false
, then the first return result (the received value) must be a zero value of the element type of the channel.
Knowing what are blocking and non-blocking channel send or receive operations is important to understand the mechanism of select
control flow blocks which will be introduced in a later section.
In the above explanations, if a goroutine is shifted out of a queue (either the sending or the receiving goroutine queue) of a channel, and the goroutine was blocked for being pushed into the queue at a
select
control flow code block, then the goroutine will be resumed to running state at step
9 of the
select
control flow code block execution. It may be dequeued from the corresponding goroutine queue of several channels involved in the
select
control flow code block.
According to the explanations listed above, we can get some facts about the internal queues of a channel.
-
If the channel is closed, both its sending goroutine queue and receiving goroutine queue must be empty, but its value buffer queue may not be empty.
-
At any time, if the value buffer is not empty, then its receiving goroutine queue must be empty.
-
At any time, if the value buffer is not full, then its sending goroutine queue must be empty.
-
If the channel is buffered, then at any time, at least one of the channel's goroutine queues must be empty (sending, receiving, or both).
-
If the channel is unbuffered, most of the time one of its sending goroutine queue and the receiving goroutine queue must be empty, with one exception. The exception is that a goroutine may be pushed into both of the two queues when executing a
select
control flow code block.
Now that you've read the above section, let's view some examples which use channels to enhance your understanding.
A simple request/response example. The two goroutines in this example talk to each other through an unbuffered channel.
package main
import (
"fmt"
"time"
)
func main() {
c := make(chan int) // an unbuffered channel
go func(ch chan<- int, x int) {
time.Sleep(time.Second)
// <-ch // fails to compile
// Send the value and block until the result is received.
ch <- x*x // 9 is sent
}(c, 3)
done := make(chan struct{})
go func(ch <-chan int) {
// Block until 9 is received.
n := <-ch
fmt.Println(n) // 9
// ch <- 123 // fails to compile
time.Sleep(time.Second)
done <- struct{}{}
}(c)
// Block here until a value is received by
// the channel "done".
<-done
fmt.Println("bye")
}
The output:
9
bye
A demo of using a buffered channel. This program is not a concurrent one, it just shows how to use buffered channels.
package main
import "fmt"
func main() {
c := make(chan int, 2) // a buffered channel
c <- 3
c <- 5
close(c)
fmt.Println(len(c), cap(c)) // 2 2
x, ok := <-c
fmt.Println(x, ok) // 3 true
fmt.Println(len(c), cap(c)) // 1 2
x, ok = <-c
fmt.Println(x, ok) // 5 true
fmt.Println(len(c), cap(c)) // 0 2
x, ok = <-c
fmt.Println(x, ok) // 0 false
x, ok = <-c
fmt.Println(x, ok) // 0 false
fmt.Println(len(c), cap(c)) // 0 2
close(c) // panic!
// The send will also panic if the above
// close call is removed.
c <- 7
}
A never-ending football game.
package main
import (
"fmt"
"time"
)
func main() {
var ball = make(chan string)
kickBall := func(playerName string) {
for {
fmt.Println(<-ball, "kicked the ball.")
time.Sleep(time.Second)
ball <- playerName
}
}
go kickBall("John")
go kickBall("Alice")
go kickBall("Bob")
go kickBall("Emily")
ball <- "referee" // kick off
var c chan bool // nil
<-c // blocking here for ever
}
When a value is transferred from one goroutine to another goroutine, the value will be copied at least one time. If the transferred value ever stayed in the value buffer of a channel, then two copies will happen in the transfer process. One copy happens when the value is copied from the sender goroutine into the value buffer, the other happens when the value is copied from the value buffer to the receiver goroutine. Like value assignments and function argument passing, when a value is transferred,
only its direct part is copied.
For the standard Go compiler, the size of channel element types must be smaller than 65536
. However, generally, we shouldn't create channels with large-size element types, to avoid too large copy cost in the process of transferring values between goroutines. So if the passed value size is too large, it is best to use a pointer element type instead, to avoid a large value copy cost.
Note, a channel is referenced by all the goroutines in either the sending or the receiving goroutine queue of the channel, so if neither of the queues of the channel is empty, the channel cannot be garbage collected. On the other hand, if a goroutine is blocked and stays in either the sending or the receiving goroutine queue of a channel, then the goroutine also cannot be garbage collected, even if the channel is referenced only by this goroutine. In fact, a goroutine can only be garbage collected when it has already exited.
Channel send operations and receive operations are
simple statements. A channel receive operation can be always used as a single-value expression. Simple statements and expressions can be used at certain portions of
basic control flow blocks.
An example in which channel send and receive operations appear as simple statements in two for
control flow blocks.
package main
import (
"fmt"
"time"
)
func main() {
fibonacci := func() chan uint64 {
c := make(chan uint64)
go func() {
var x, y uint64 = 0, 1
for ; y < (1 << 63); c <- y { // here
x, y = y, x+y
}
close(c)
}()
return c
}
c := fibonacci()
for x, ok := <-c; ok; x, ok = <-c { // here
time.Sleep(time.Second)
fmt.Println(x)
}
}
The for-range
control flow code block applies to channels. The loop will try to iteratively receive the values sent to a channel, until the channel is closed and its value buffer queue becomes blank. With for-range
syntax on arrays, slices and maps, multiple iteration variables are allowed. However, for for-range
blocks applied to channels, you can use at most one iteration variable, which is used to store the received values.
for v := range aChannel {
// use v
}
is equivalent to
for {
v, ok = <-aChannel
if !ok {
break
}
// use v
}
Surely, here the aChannel
value must not be a send-only channel. If it is a nil channel, the loop will block there for ever.
For example, the second for
loop block in the example shown in the last section can be simplified to
for x := range c {
time.Sleep(time.Second)
fmt.Println(x)
}
There is a select-case
code block syntax which is specially designed for channels. The syntax is much like the switch-case
block syntax. For example, there can be multiple case
branches and at most one default
branch in the select-case
code block. But there are also some obvious differences between them.
-
No expressions and statements are allowed to follow the select
keyword (before {
).
-
No fallthrough
statements are allowed to be used in case
branches.
-
Each statement following a case
keyword in a select-case
code block must be either a channel receive operation or a channel send operation statement. A channel receive operation can appear as the source value of a simple assignment statement. Later, a channel operation following a case
keyword will be called a case
operation.
-
If there are one or more non-blocking case
operations, Go runtime will randomly select one of these non-blocking operations to execute, then continue to execute the corresponding case
branch.
-
If all the case
operations in a select-case
code block are blocking operations, the default
branch will be selected to execute if the default
branch is present. If the default
branch is absent, the current goroutine will be pushed into the corresponding sending goroutine queue or receiving goroutine queue of every channel involved in all case
operations, then enter blocking state.
By the rules, a select-case
code block without any branches, select{}
, will make the current goroutine stay in blocking state forever.
The following program will enter the default
branch for sure.
package main
import "fmt"
func main() {
var c chan struct{} // nil
select {
case <-c: // blocking operation
case c <- struct{}{}: // blocking operation
default:
fmt.Println("Go here.")
}
}
An example showing how to use try-send and try-receive:
package main
import "fmt"
func main() {
c := make(chan string, 2)
trySend := func(v string) {
select {
case c <- v:
default: // go here if c is full.
}
}
tryReceive := func() string {
select {
case v := <-c: return v
default: return "-" // go here if c is empty
}
}
trySend("Hello!") // succeed to send
trySend("Hi!") // succeed to send
// Fail to send, but will not block.
trySend("Bye!")
// The following two lines will
// both succeed to receive.
fmt.Println(tryReceive()) // Hello!
fmt.Println(tryReceive()) // Hi!
// The following line fails to receive.
fmt.Println(tryReceive()) // -
}
The following example has 50% possibility to panic. Both of the two case
operations are non-blocking in this example.
package main
func main() {
c := make(chan struct{})
close(c)
select {
case c <- struct{}{}:
// Panic if the first case is selected.
case <-c:
}
}
There are several steps to execute a select-case
block:
-
evaluate all involved channel expressions and value expressions to be potentially sent in case
operations, from top to bottom and left to right. Destination values for receive operations (as source values) in assignments needn't be evaluated at this time.
-
randomize the branch orders for polling in step 5. The default
branch is always put at the last position in the result order. Channels may be duplicate in the case
operations.
-
sort all involved channels in the case
operations to avoid deadlock (with other goroutines) in the next step. No duplicate channels stay in the first N
channels of the sorted result, where N
is the number of involved channels in the case
operations. Below, the channel lock order is a concept for the first N
channels in the sorted result.
-
lock (a.k.a., acquire the locks of) all involved channels by the channel lock order produced in last step.
-
poll each branch in the select block by the randomized order produced in step 2:
-
if this is a case
branch and the corresponding channel operation is a send-value-to-closed-channel operation, unlock all channels by the inverse channel lock order and make the current goroutine panic. Go to step 12.
-
if this is a case
branch and the corresponding channel operation is non-blocking, perform the channel operation and unlock all channels by the inverse channel lock order, then execute the corresponding case
branch body. The channel operation may wake up another goroutine in blocking state. Go to step 12.
-
if this is the default
branch, then unlock all channels by the inverse channel lock order and execute the default
branch body. Go to step 12.
(Up to here, the default
branch is absent and all case
operations are blocking operations.)
-
push (enqueue) the current goroutine (along with the information of the corresponding case
branch) into the receiving or sending goroutine queue of the involved channel in each case
operation. The current goroutine may be pushed into the queues of a channel for multiple times, for the involved channels in multiple cases may be the same one.
-
make the current goroutine enter blocking state and unlock all channels by the inverse channel lock order.
-
wait in blocking state until other channel operations wake up the current goroutine, ...
-
the current goroutine is waken up by another channel operation in another goroutine. The other operation may be a channel close operation or a channel send/receive operation. If it is a channel send/receive operation, there must be a case
channel receive/send operation (in the current being explained select-case
block) cooperating with it (by transferring a value). In the cooperation, the current goroutine will be dequeued from the receiving/sending goroutine queue of the channel.
-
lock all involved channels by the channel lock order.
-
dequeue the current goroutine from the receiving goroutine queue or sending goroutine queue of the involved channel in each case
operation,
-
if the current goroutine is waken up by a channel close operation, go to step 5.
-
if the current goroutine is waken up by a channel send/receive operation, the corresponding case
branch of the cooperating receive/send operation has already been found in the dequeuing process, so just unlock all channels by the inverse channel lock order and execute the corresponding case
branch.
-
done.
From the implementation, we know that
-
a goroutine may stay in the sending goroutine queues and the receiving goroutine queues of multiple channels at the same time. It can even stay in the sending goroutine queue and the receiving goroutine queue of the same channel at the same time.
-
when a goroutine currently being blocked at a select-case
code block gets resumed later, it will be removed from all the sending goroutine queues and the receiving goroutine queues of all channels involved in the channel operations following case
keywords in the select-case
code block.
Channel may be not always the best solution for all use cases for data synchronizations. Please read
this article and
this article for more synchronization techniques in Go.
The digital versions of this book are available at the following places:
Tapir, the author of Go 101, has been on writing the Go 101 series books
and maintaining the go101.org website since 2016 July.
New contents will be continually added to the book and the website from time to time.
Tapir is also an indie game developer.
You can also support Go 101 by playing
Tapir's games
(made for both Android and iPhone/iPad):
Individual donations via PayPal are also welcome.
Articles in this book:
-
Become Familiar With Go Code