Changkun's Blog欧长坤的博客
idea想法

Proebsting's LawProebsting 定律

Today I read an extra paper. Although it’s not directly related to Go, I think it offers some insightful perspective on the current state of the Go language, so I’d like to share it. The paper is called “On Proebsting’s Law.”

We all know Moore’s Law says the number of transistors on integrated circuits doubles every 18 months, but this paper studies and validates the so-called Proebsting’s Law: the performance improvement brought by compiler optimization techniques doubles every 18 years. Proebsting’s Law was proposed in 1998, and its author Todd Proebsting was probably half-joking, because he suggested that the compiler and programming language research community should reduce their focus on performance optimization and instead pay more attention to improving programmer productivity.

Now, looking back at this suggestion with hindsight, we can see it’s not without merit: although Go’s compiler has gone through several major optimization versions, the techniques it uses aren’t particularly fancy – rather, they are quite traditional and conventional optimization techniques. However, this hasn’t hindered Go’s success, because what it tries to address is exactly programmer productivity:

  1. By avoiding circular dependencies, it greatly reduces the time programmers spend waiting for compilation
  2. Its very concise language design and feature set greatly reduces the time programmers spend thinking about how to use the language
  3. Forward compatibility guarantees almost entirely eliminate the migration and maintenance time caused by version upgrades

今天额外读了一篇论文，虽然跟 Go 没有直接关系，但我觉得对理解目前 Go 语言的现状是有一定启发意义的，所以来分享一下。这篇论文叫做 “On Proebsting’s Law”。

我们都知道 Moore 定律说集成电路上晶体管数量每 18 个月番一番，但这篇论文则研究并验证了所谓的Proebsting 定律: 编译器优化技术带来的性能提升每 18 年番一番。Proebsting 定律是在 1998 年提出的，当时的提出者 Todd Proebsting 可能只是在开玩笑，因为他建议编译器和编程语言研究界应该减少对性能优化的关注，而应该更多的关注程序员工作效率的提升。

现在我们来事后诸葛亮评价这一建议就能发现其实这并不是无道理的: Go 语言的编译器虽然经历过几大版本的优化，但其使用的技术并不够 fancy，相反而是很传统且中规中矩的优化技术。然而这并不影响 Go 语言的成功，因为它尝试解决的正是程序员的工作效率:

  1. 通过避免循环以来而极大的减少了程序员等待编译的时间
  2. 非常简洁的语言设计与特性极大的减少了程序员思考如何使用语言的时间
  3. 向前的兼容性保障几乎彻底消除了因为版本升级给程序员带来的迁移和维护时间
idea想法

Telegram BotTelegram 机器人

Because the COVID situation in Europe is still terrible, even shopping at an Apple Store requires an appointment in advance. Since I urgently needed to visit an Apple Store recently but couldn’t find any available appointment slots, I quickly hacked together a tool to check availability and send a reminder message via Telegram when an appointment becomes available. Tool link: https://changkun.de/s/apreserve

Interacting with Telegram using Go is straightforward:

  1. Create a bot from BotFather
  2. Obtain the bot’s token and the chat ID for your conversation with it
  3. Then you can handle messages
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

package main

import (
	tg "github.com/go-telegram-bot-api/telegram-bot-api"
)

func main() {
	token := …
	chatid := …
	bot, err := tg.NewBotAPI(token)
	if err != nil { … }
	bot.Send(tg.NewMessage(chatid, "text"))
}

因为欧洲疫情依然很糟糕，所以现在甚至于想去苹果店购物都要提前预约。因为最近急需要去苹果店一次，又苦于刷不到可用的预约位置，刚刚顺手就糊一个工具来检查，当预约可用时给telegram发送一条提醒消息。工具地址: https://changkun.de/s/apreserve

用 Go 和 telegram 进行交互没有任何难度:

  1. 从 botfather 创建一个 bot
  2. 获得这个 bot 的 token 以及跟它对话的 chatid
  3. 于是可以处理消息了
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

package main

import (
	tg "github.com/go-telegram-bot-api/telegram-bot-api"
)

func main() {
	token := …
	chatid := …
	bot, err := tg.NewBotAPI(token)
	if err != nil { … }
	bot.Send(tg.NewMessage(chatid, "text"))
}
idea想法

Apple SiliconApple Silicon

How is Go’s compilation performance on darwin/arm64? I did a rough and non-rigorous comparison of Go compilation performance between an Intel Mac and an M1 Mac. This compilation report was generated with the following commands:

$ go build -gcflags='-bench=bench.out' -a $ cat bench.out

where -a disables the compilation cache.

MacBook Air (M1, 2020), Apple M1, 16 GB:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

commit: devel +7716a2fbb7 Sun Nov 22 11:44:49 2020 -0500
goos: darwin
goarch: arm64
BenchmarkCompile:main:fe:init              1     305167 ns/op     17.86 %
BenchmarkCompile:main:fe:loadsys           1     164916 ns/op      9.65 %
BenchmarkCompile:main:fe:parse             1     199209 ns/op     11.66 %    23 lines    115457 lines/s
BenchmarkCompile:main:fe:typecheck:top1    1       7625 ns/op      0.45 %
BenchmarkCompile:main:fe:typecheck:top2    1       2375 ns/op      0.14 %
BenchmarkCompile:main:fe:typecheck:func    1      34000 ns/op      1.99 %     1 funcs     29412 funcs/s
BenchmarkCompile:main:fe:capturevars       1        250 ns/op      0.01 %
BenchmarkCompile:main:fe:inlining          1       9500 ns/op      0.56 %
BenchmarkCompile:main:fe:escapes           1       5834 ns/op      0.34 %
BenchmarkCompile:main:fe:xclosures         1      49583 ns/op      2.90 %
BenchmarkCompile:main:fe:subtotal          1     778459 ns/op     45.56 %
BenchmarkCompile:main:be:compilefuncs      1     279125 ns/op     16.34 %     1 funcs      3583 funcs/s
BenchmarkCompile:main:be:externaldcls      1        500 ns/op      0.03 %
BenchmarkCompile:main:be:dumpobj           1     639667 ns/op     37.44 %
BenchmarkCompile:main:be:subtotal          1     919292 ns/op     53.81 %
BenchmarkCompile:main:unaccounted          1      10791 ns/op      0.63 %
BenchmarkCompile:main:total                1    1708542 ns/op    100.00 %

Mac mini (2018), 3 GHz 6-Core Intel Core i5, 8 GB 2667 MHz DDR4:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

commit: go1.15.6
goos: darwin
goarch: amd64
BenchmarkCompile:main:fe:init              1     333752 ns/op     15.31 %
BenchmarkCompile:main:fe:loadsys           1     246343 ns/op     11.30 %
BenchmarkCompile:main:fe:parse             1     372343 ns/op     17.08 %    23 lines    61771 lines/s
BenchmarkCompile:main:fe:typecheck:top1    1      19620 ns/op      0.90 %
BenchmarkCompile:main:fe:typecheck:top2    1       7347 ns/op      0.34 %
BenchmarkCompile:main:fe:typecheck:func    1      23073 ns/op      1.06 %     1 funcs    43341 funcs/s
BenchmarkCompile:main:fe:capturevars       1        238 ns/op      0.01 %
BenchmarkCompile:main:fe:inlining          1      13277 ns/op      0.61 %
BenchmarkCompile:main:fe:escapes           1      19283 ns/op      0.88 %
BenchmarkCompile:main:fe:xclosures         1     102213 ns/op      4.69 %
BenchmarkCompile:main:fe:subtotal          1    1137489 ns/op     52.19 %
BenchmarkCompile:main:be:compilefuncs      1     598194 ns/op     27.45 %     1 funcs     1672 funcs/s
BenchmarkCompile:main:be:externaldcls      1        766 ns/op      0.04 %
BenchmarkCompile:main:be:dumpobj           1     415450 ns/op     19.06 %
BenchmarkCompile:main:be:subtotal          1    1014410 ns/op     46.54 %
BenchmarkCompile:main:unaccounted          1      27696 ns/op      1.27 %
BenchmarkCompile:main:total                1    2179595 ns/op    100.00

Go在darwin/arm64上的编译性能怎么样？我很不严谨的粗略比较了Intel Mac 和 M1 Mac 的 Go 编译性能。这个编译报告由如下指令生成:

$ go build -gcflags='-bench=bench.out' -a $ cat bench.out

其中-a用于禁用编译缓存。

MacBook Air (M1, 2020), Apple M1, 16 GB:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

commit: devel +7716a2fbb7 Sun Nov 22 11:44:49 2020 -0500
goos: darwin
goarch: arm64
BenchmarkCompile:main:fe:init              1     305167 ns/op     17.86 %
BenchmarkCompile:main:fe:loadsys           1     164916 ns/op      9.65 %
BenchmarkCompile:main:fe:parse             1     199209 ns/op     11.66 %    23 lines    115457 lines/s
BenchmarkCompile:main:fe:typecheck:top1    1       7625 ns/op      0.45 %
BenchmarkCompile:main:fe:typecheck:top2    1       2375 ns/op      0.14 %
BenchmarkCompile:main:fe:typecheck:func    1      34000 ns/op      1.99 %     1 funcs     29412 funcs/s
BenchmarkCompile:main:fe:capturevars       1        250 ns/op      0.01 %
BenchmarkCompile:main:fe:inlining          1       9500 ns/op      0.56 %
BenchmarkCompile:main:fe:escapes           1       5834 ns/op      0.34 %
BenchmarkCompile:main:fe:xclosures         1      49583 ns/op      2.90 %
BenchmarkCompile:main:fe:subtotal          1     778459 ns/op     45.56 %
BenchmarkCompile:main:be:compilefuncs      1     279125 ns/op     16.34 %     1 funcs      3583 funcs/s
BenchmarkCompile:main:be:externaldcls      1        500 ns/op      0.03 %
BenchmarkCompile:main:be:dumpobj           1     639667 ns/op     37.44 %
BenchmarkCompile:main:be:subtotal          1     919292 ns/op     53.81 %
BenchmarkCompile:main:unaccounted          1      10791 ns/op      0.63 %
BenchmarkCompile:main:total                1    1708542 ns/op    100.00 %

Mac mini (2018), 3 GHz 6-Core Intel Core i5, 8 GB 2667 MHz DDR4:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

commit: go1.15.6
goos: darwin
goarch: amd64
BenchmarkCompile:main:fe:init              1     333752 ns/op     15.31 %
BenchmarkCompile:main:fe:loadsys           1     246343 ns/op     11.30 %
BenchmarkCompile:main:fe:parse             1     372343 ns/op     17.08 %    23 lines    61771 lines/s
BenchmarkCompile:main:fe:typecheck:top1    1      19620 ns/op      0.90 %
BenchmarkCompile:main:fe:typecheck:top2    1       7347 ns/op      0.34 %
BenchmarkCompile:main:fe:typecheck:func    1      23073 ns/op      1.06 %     1 funcs    43341 funcs/s
BenchmarkCompile:main:fe:capturevars       1        238 ns/op      0.01 %
BenchmarkCompile:main:fe:inlining          1      13277 ns/op      0.61 %
BenchmarkCompile:main:fe:escapes           1      19283 ns/op      0.88 %
BenchmarkCompile:main:fe:xclosures         1     102213 ns/op      4.69 %
BenchmarkCompile:main:fe:subtotal          1    1137489 ns/op     52.19 %
BenchmarkCompile:main:be:compilefuncs      1     598194 ns/op     27.45 %     1 funcs     1672 funcs/s
BenchmarkCompile:main:be:externaldcls      1        766 ns/op      0.04 %
BenchmarkCompile:main:be:dumpobj           1     415450 ns/op     19.06 %
BenchmarkCompile:main:be:subtotal          1    1014410 ns/op     46.54 %
BenchmarkCompile:main:unaccounted          1      27696 ns/op      1.27 %
BenchmarkCompile:main:total                1    2179595 ns/op    100.00
idea想法

Deprecating ioutil弃用 ioutil

ioutil will be fully deprecated in Go 1.16. Although these APIs will continue to exist due to the compatibility guarantee, they are no longer recommended for use. So the question is: what should we use instead? Here are all the APIs in the ioutil package:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

package ioutil

var Discard io.Writer
func NopCloser(r io.Reader) io.ReadCloser
func ReadAll(r io.Reader) ([]byte, error)

func ReadDir(dirname string) ([]os.FileInfo, error)
func ReadFile(filename string) ([]byte, error)
func WriteFile(filename string, data []byte, perm os.FileMode) error

func TempDir(dir, pattern string) (name string, err error)
func TempFile(dir, pattern string) (f *os.File, err error)

The corresponding replacement APIs in 1.16:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

package io

var Discard Writer
func NopCloser(r Reader) ReadCloser
func ReadAll(r Reader) ([]byte, error)


package os

func ReadDir(name string) ([]DirEntry, error)
func ReadFile(name string) ([]byte, error)
func WriteFile(name string, data []byte, perm fs.FileMode) error

func MkdirTemp(dir, pattern string) (string, error)
func CreateTemp(dir, pattern string) (f *File, err error)

In summary, three key changes:

  1. Discard, NopCloser, ReadAll have been moved to the io package
  2. ReadDir, ReadFile, WriteFile have been moved to the os package
  3. TempDir, TempFile have been renamed to MkdirTemp, CreateTemp and moved to the os package

ioutil 将在 Go 1.16 中被彻底弃用，虽然由于兼容性保障这些 API 还会继续存在，但不再被推荐使用了。那么问题来了，我们应该用什么？这是 ioutil 包所有的 API:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

package ioutil

var Discard io.Writer
func NopCloser(r io.Reader) io.ReadCloser
func ReadAll(r io.Reader) ([]byte, error)

func ReadDir(dirname string) ([]os.FileInfo, error)
func ReadFile(filename string) ([]byte, error)
func WriteFile(filename string, data []byte, perm os.FileMode) error

func TempDir(dir, pattern string) (name string, err error)
func TempFile(dir, pattern string) (f *os.File, err error)

1.16 中取而代之的与之对应的 API:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

package io

var Discard Writer
func NopCloser(r Reader) ReadCloser
func ReadAll(r Reader) ([]byte, error)


package os

func ReadDir(name string) ([]DirEntry, error)
func ReadFile(name string) ([]byte, error)
func WriteFile(name string, data []byte, perm fs.FileMode) error

func MkdirTemp(dir, pattern string) (string, error)
func CreateTemp(dir, pattern string) (f *File, err error)

总结起来就是三点:

  1. Discard, NopCloser, ReadAll 挪到了 io 包中
  2. ReadDir, ReadFile, WriteFile 挪到了 os 包中
  3. TempDir, TempFile 更名为了 MkdirTemp, CreateTemp 并挪到了 os 包中
idea想法

Testing io/fs Implementations测试 io/fs 的实现

io/fs is getting closer and closer. The functionality is great, but how do we test it? There is a function in testing/fstest that can do just that.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

package fstest

// TestFS tests a file system implementation.
// It walks the entire tree of files in fsys,
// opening and checking that each file behaves correctly.
// It also checks that the file system contains at least the expected files.
// As a special case, if no expected files are listed, fsys must be empty.
// Otherwise, fsys must only contain at least the listed files: it can also contain others.
//
// If TestFS finds any misbehaviors, it returns an error reporting all of them.
// The error text spans multiple lines, one per detected misbehavior.
//
// Typical usage inside a test is:
//
//	if err := fstest.TestFS(myFS, "file/that/should/be/present"); err != nil {
//		t.Fatal(err)
//	}
//
func TestFS(fsys fs.FS, expected ...string) error

io/fs 越来越近了 功能很好但我们怎么才能测试它呢？testing/fstest 中有一个函数可以做到这件事情。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

package fstest

// TestFS tests a file system implementation.
// It walks the entire tree of files in fsys,
// opening and checking that each file behaves correctly.
// It also checks that the file system contains at least the expected files.
// As a special case, if no expected files are listed, fsys must be empty.
// Otherwise, fsys must only contain at least the listed files: it can also contain others.
//
// If TestFS finds any misbehaviors, it returns an error reporting all of them.
// The error text spans multiple lines, one per detected misbehavior.
//
// Typical usage inside a test is:
//
//	if err := fstest.TestFS(myFS, "file/that/should/be/present"); err != nil {
//		t.Fatal(err)
//	}
//
func TestFS(fsys fs.FS, expected ...string) error
idea想法

Revisiting Asynchronous Preemption回顾异步抢占

Are you sure you understand asynchronous preemption? Today, while discussing with Cao Da (@Xargin) about how the interrupted G in the asynchronous preemption flow restores its previous execution context, I realized my understanding of asynchronous preemption was not comprehensive enough. In “Go Under The Hood,” asynchronous preemption is described as follows: let’s name the two running threads M1 and M2. The overall logic of a preemption call can be summarized as:

  1. M1 sends an interrupt signal (signalM(mp, sigPreempt))
  2. M2 receives the signal; the OS interrupts its executing code and switches to the signal handler (sighandler(signum, info, ctxt, gp))
  3. M2 modifies the execution context and resumes at the modified location (asyncPreempt)
  4. Re-enters the scheduling loop to schedule other Goroutines (preemptPark and gopreempt_m)

This summary is not entirely correct, because it does not clearly explain the difference between preemptPark and gopreempt_m. This week, let’s briefly supplement the overall behavior of asynchronous preemption:

Assuming the system monitor acts as M1, after the system monitor sends the interrupt signal, execution arrives at asyncPreempt2:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

//go:nosplit
func asyncPreempt2() {
	gp := getg()
	gp.asyncSafePoint = true
	if gp.preemptStop {
		mcall(preemptPark)
	} else {
		mcall(gopreempt_m)
	}
	gp.asyncSafePoint = false
}

But will it ultimately choose preemptPark or gopreempt_m? The asynchronous preemption issued by sysmon calling preemptone does not set the preemptStop flag on G, so it enters the gopreempt_m flow. gopreempt_m ultimately calls goschedImpl, which places the preempted G into the global queue to be scheduled later.

So what about the other half (preemptPark)? When we carefully examine the implementation of preemptPark, we find that the preempted G is not added to the scheduling queue at all — instead, it directly calls schedule:

1
2
3
4
5
6
7
8

//go:systemstack
func preemptPark(gp *g) {
	...
	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
	dropg()
	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
	schedule()
}

So how does the preempted G get back into the scheduling loop? It turns out that the branch where gp.preemptStop is true occurs when the GC needs it (markroot): it uses suspendG to mark the running G (gp.preemptStop = true), sends the preemption signal (preemptM), and returns the state of the interrupted G. When the GC’s marking work is complete and preemption ends, it passes this state and calls resumeG, which ultimately calls ready to resume the interrupted G:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

//go:nowritebarrier
func markroot(gcw *gcWork, i uint32) {
	...
	var gp *g
	systemstack(func() {
		...
		stopped := suspendG(gp)
		...
		scanstack(gp, gcw) // GC stack scanning
		gp.gcscandone = true
		resumeG(stopped)
		...
	})
}
//go:systemstack
func suspendG(gp *g) suspendGState {
	...
	stopped := false
	for i := 0; ; i++ {
		switch s := readgstatus(gp); s {
		...
		case _Gpreempted:
			...
			stopped = true
			gp.preemptStop = false
			gp.preempt = false
			gp.stackguard0 = gp.stack.lo + _StackGuard
			return suspendGState{g: gp, stopped: stopped}

		case _Grunning:
			...
			gp.preemptStop = true
			gp.preempt = true
			gp.stackguard0 = stackPreempt
			casfrom_Gscanstatus(gp, _Gscanrunning, _Grunning)
			preemptM(gp.m)
		}
		...
	}
}
func resumeG(state suspendGState) {
	...
	gp := state.g
	ready(gp, 0, true)
}

你确定你看懂异步抢占了吗？今天跟曹大@Xargin 交流起异步抢占的流程里被中断的 G 是如何恢复到之前的执行现场时才发现对异步抢占的理解还不够全面。在《Go 语言原本》中是这样描述异步抢占的: 不妨给正在运行的两个线程命名为 M1 和 M2，抢占调用的整体逻辑可以被总结为:

  1. M1 发送中断信号（signalM(mp, sigPreempt)）
  2. M2 收到信号，操作系统中断其执行代码，并切换到信号处理函数（sighandler(signum, info, ctxt, gp)）
  3. M2 修改执行的上下文，并恢复到修改后的位置（asyncPreempt）
  4. 重新进入调度循环进而调度其他 Goroutine（preemptPark 和 gopreempt_m）

这个总结并不完全正确，因为它并没有总结清楚 preemptPark 和 gopreempt_m 这两者之间的区别。这周我们来简单补充一下异步抢占的整体行为:

假设系统监控充当 M1，当系统监控发送中断信号后，会来到 asyncPreempt2:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

//go:nosplit
func asyncPreempt2() {
	gp := getg()
	gp.asyncSafePoint = true
	if gp.preemptStop {
		mcall(preemptPark)
	} else {
		mcall(gopreempt_m)
	}
	gp.asyncSafePoint = false
}

但最终会选择 preemptPark 还是 gopreempt_m 呢？sysmon 调用 preemptone 的代码中发出的异步抢占并不会为 G 设置 preemptStop 标记，从而会进入 gopreempt_m 的流程，而 gopreempt_m 最终会调用 goschedImpl 将被抢占的 G 放入全局队列，等待日后被调度。

那么另一半（preemptPark）呢？当我们仔细查看 preemptPark 的实现则会发现，被抢占的 G 其实并没有被加入到调度队列中，而是直接就调用了 schedule:

1
2
3
4
5
6
7
8

//go:systemstack
func preemptPark(gp *g) {
	...
	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
	dropg()
	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
	schedule()
}

那这时被抢占的 G 怎样才会恢复到调度循环呢？原来 gp.preemptStop 为 true 的分支发生在 GC 需要时（markroot）通过 suspendG 来标记正在运行的 G（gp.preemptStop = true），再发送抢占信号（preemptM），返回被中断 G 的状态。当 GC 的标记工作完成，抢占结束后，在将这个状态传递并调用 resumeG，最终 ready 并恢复这个被中断的 G:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

//go:nowritebarrier
func markroot(gcw *gcWork, i uint32) {
	...
	var gp *g
	systemstack(func() {
		...
		stopped := suspendG(gp)
		...
		scanstack(gp, gcw) // GC 栈扫描
		gp.gcscandone = true
		resumeG(stopped)
		...
	})
}
//go:systemstack
func suspendG(gp *g) suspendGState {
	...
	stopped := false
	for i := 0; ; i++ {
		switch s := readgstatus(gp); s {
		...
		case _Gpreempted:
			...
			stopped = true
			gp.preemptStop = false
			gp.preempt = false
			gp.stackguard0 = gp.stack.lo + _StackGuard
			return suspendGState{g: gp, stopped: stopped}

		case _Grunning:
			...
			gp.preemptStop = true
			gp.preempt = true
			gp.stackguard0 = stackPreempt
			casfrom_Gscanstatus(gp, _Gscanrunning, _Grunning)
			preemptM(gp.m)
		}
		...
	}
}
func resumeG(state suspendGState) {
	...
	gp := state.g
	ready(gp, 0, true)
}
idea想法

Data Races and the Memory Model数据竞争和内存模型

Does this code have a data race?

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

// from https://golang.org/issue/42598
type module struct {
	v int
}

func foo() {
	mods := []*module{
		&module{v: 0},
		&module{v: 1},
		&module{v: 2},
	}
	type token struct{}
	sem := make(chan token, runtime.GOMAXPROCS(0))
	for _, m := range mods {
		add := func(m *module) {
			sem <- token{}
			go func() {
				*m = module{v: 42} // write
				<-sem
			}()
		}
		add(m) // read
	}
	// Fill semaphore channel to wait for all tasks to finish.
	for n := cap(sem); n > 0; n-- {
		sem <- token{}
	}
}

An issue submitted yesterday appears to point out a bug in the current Go memory model. Further reading:

这段代码有 data race 吗？

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

// from https://golang.org/issue/42598
type module struct {
	v int
}

func foo() {
	mods := []*module{
		&module{v: 0},
		&module{v: 1},
		&module{v: 2},
	}
	type token struct{}
	sem := make(chan token, runtime.GOMAXPROCS(0))
	for _, m := range mods {
		add := func(m *module) {
			sem <- token{}
			go func() {
				*m = module{v: 42} // write
				<-sem
			}()
		}
		add(m) // read
	}
	// Fill semaphore channel to wait for all tasks to finish.
	for n := cap(sem); n > 0; n-- {
		sem <- token{}
	}
}

昨天提交的一个 issue 似乎指出了目前 Go 内存模型中的一个错误。进一步阅读:

idea想法

Further Thoughts on Go Error Handling关于 Go 错误处理的进一步看法

Continued: In my view, the reason many people are dissatisfied with error handling is a lack of patience to understand how Go approaches problem-solving. An important lesson Jonathan drew is that errors are inherently domain-specific — some domains focus on better tracing of error origins, though stack traces themselves are sometimes not that useful; some domains focus on more flexible aggregation of multiple error messages, but many people just want to get the happy path right and throw a unified error at the end, and so on. In his Q&A session, he also mentioned that he does not recommend using xerrors, etc. It should be (not so) obvious that only solutions tailored to the specific problem are the best ones. Developers should take the time to think carefully about how to design error handling for a particular problem. Complaining about the lack of try/catch at the syntax level, or crying about how ugly if err is everywhere, is just as meaningless and life-wasting as debating which brackets to use for generics.

续: 很多人不满错误处理的原因在我看来是没有耐心去理解 Go 里处理问题的方式，Jonathan 总结得到的一个重要教训就是错误本身就是领域特定的，有些领域关注如何更好的追踪错误来源，但堆栈信息本身有时候也不那么有用；有些领域关注如何更加灵活的对多个错误信息进行整合，但很多人可能只想把正常逻辑给写对了然后统一扔一个错误出去等等，后续他的QA中还提到不建议使用xerrors等。（不那么）显然，只有针对问题本身给出的方案才是最好的，开发者应该静下心来思考怎么对某个具体问题设计错误处理，吐槽什么语法层面有没有 try/catch 、 if err 满天飞丑到哭泣就跟讨论泛型用什么括号一样没有意义且浪费生命。

idea想法

The Regret of Go 1.13 Error Values ProposalGo 1.13 错误值提案的遗憾

At today’s GopherCon 2020, the author of the Go 1.13 error values proposal mentioned in hindsight that he regrets the lack of error formatting support, and that there will be no further improvement plans for many years to come. One of the reasons he gave is that error handling is a domain-specific problem, and it is simply beyond his ability to produce a solution that satisfies everyone. Nevertheless, at the end of his talk, he offered some advice on error wrapping — namely, implementing fmt.Formatter. Below is a simple example.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

type DetailError struct {
	msg, detail string
	err         error
}

func (e *DetailError) Unwrap() error { return e.err }

func (e *DetailError) Error() string {
	if e.err == nil {
		return e.msg
	}
	return e.msg + ": " + e.err.Error()
}

func (e *DetailError) Format(s fmt.State, c rune) {
	if s.Flag('#') && c == 'v' {
		type nomethod DetailError
		fmt.Fprintf(s, "%#v", (*nomethod)(e))
		return
	}
	if !s.Flag('+') || c != 'v' {
		fmt.Fprintf(s, spec(s, c), e.Error())
		return
	}
	fmt.Fprintln(s, e.msg)
	if e.detail != "" {
		io.WriteString(s, "\t")
		fmt.Fprintln(s, e.detail)
	}
	if e.err != nil {
		if ferr, ok := e.err.(fmt.Formatter); ok {
			ferr.Format(s, c)
		} else {
			fmt.Fprintf(s, spec(s, c), e.err)
			io.WriteString(s, "\n")
		}
	}
}

func spec(s fmt.State, c rune) string {
	buf := []byte{'%'}
	for _, f := range []int{'+', '-', '#', ' ', '0'} {
		if s.Flag(f) {
			buf = append(buf, byte(f))
		}
	}
	if w, ok := s.Width(); ok {
		buf = strconv.AppendInt(buf, int64(w), 10)
	}
	if p, ok := s.Precision(); ok {
		buf = append(buf, '.')
		buf = strconv.AppendInt(buf, int64(p), 10)
	}
	buf = append(buf, byte(c))
	return string(buf)
}

今天的 GopherCon2020 上，Go 1.13 错误值提案的作者事后提及他对目前错误格式化的缺失表示遗憾，而且在未来很长的好几年内都不会有任何进一步改进计划。对此他本人给出的原因之一是对于错误处理这一领域特定的问题，在他的能力范围内实在是无法给出一个令所有人都满意的方案。尽管如此，在他演讲的最后，还是给出了一些关于错误嵌套的建议，即实现 fmt.Formatter，下面给出了一个简单的例子。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

type DetailError struct {
	msg, detail string
	err         error
}

func (e *DetailError) Unwrap() error { return e.err }

func (e *DetailError) Error() string {
	if e.err == nil {
		return e.msg
	}
	return e.msg + ": " + e.err.Error()
}

func (e *DetailError) Format(s fmt.State, c rune) {
	if s.Flag('#') && c == 'v' {
		type nomethod DetailError
		fmt.Fprintf(s, "%#v", (*nomethod)(e))
		return
	}
	if !s.Flag('+') || c != 'v' {
		fmt.Fprintf(s, spec(s, c), e.Error())
		return
	}
	fmt.Fprintln(s, e.msg)
	if e.detail != "" {
		io.WriteString(s, "\t")
		fmt.Fprintln(s, e.detail)
	}
	if e.err != nil {
		if ferr, ok := e.err.(fmt.Formatter); ok {
			ferr.Format(s, c)
		} else {
			fmt.Fprintf(s, spec(s, c), e.err)
			io.WriteString(s, "\n")
		}
	}
}

func spec(s fmt.State, c rune) string {
	buf := []byte{'%'}
	for _, f := range []int{'+', '-', '#', ' ', '0'} {
		if s.Flag(f) {
			buf = append(buf, byte(f))
		}
	}
	if w, ok := s.Width(); ok {
		buf = strconv.AppendInt(buf, int64(w), 10)
	}
	if p, ok := s.Precision(); ok {
		buf = append(buf, '.')
		buf = strconv.AppendInt(buf, int64(p), 10)
	}
	buf = append(buf, byte(c))
	return string(buf)
}
© 2008 - 2026 Changkun Ou. All rights reserved.保留所有权利。 | PV/UV: /
0%