1. Synchronous, asynchronous, blocking, non-blocking
What are synchronous IO and asynchronous IO, blocking IO and non-blocking IO, and what is the difference?
The background discussed in this article is network IO in the Linux environment. The most important reference for this article is "UNIX® Network Programming Volume 1 , Third Edition: The Sockets Networking" by Richard Stevens, Section 6.2 "I/ O Models", where Stevens details the characteristics and differences of various IOs , If English is good enough, it is recommended to read directly. Stevens' style of writing is famous for explaining the simple language, so don't worry about not understanding it. The flow charts in this article are also taken from references. Stevens compared a total of five IO Models in the article: * blocking IO #Blocking IO * nonblocking IO # Non-blocking IO * IO multiplexing # IO multiplexing * signal driven IO * asynchronous IO #Asynchronous IO is not commonly used in practice by signal driven IO (signal driven IO), so the main Introduce the remaining four IO Models.
2. The objects and steps involved when IO occurs.
For a network IO (here we take read as an example), it will involve two system objects, (1) Call the process (or thread) of this IO, (2) System kernel (kernel).
It's important to keep these two points in mind, because the difference between these IO models is that there are different situations in the two phases.
When a read operation occurs, the operation goes through two phases: (1) Waiting for the data to be ready (2) Copying the data from the kernel to the process
3
blocking IO
In Linux, all sockets are blocked by default. A typical read operation flow is like this:
When the user process calls the recvfrom system call, the kernel begins the first stage of IO: preparing data. For network io, many times the data has not arrived at the beginning (for example, a complete UDP packet has not been received), at this time the kernel has to wait for enough data to arrive.
而在用户进程这边，整个进程会被阻塞。当kernel一直等到数据准备好了，它就会将数据从kernel中拷贝到用户内存，
然后kernel返回结果，用户进程才解除block的状态，重新运行起来。
Therefore, the characteristic of blocking IO is that both stages of IO execution (waiting for data and copying data) are blocked.
几乎所有的程序员第一次接触到的网络编程都是从listen\(\)、send\(\)、recv\(\) 等接口开始的，
使用这些接口可以很方便的构建服务器/客户机的模型。然而大部分的socket接口都是阻塞型的。如下图
ps：
所谓阻塞型接口是指系统调用（一般是IO接口）不返回调用结果并让当前线程一直阻塞
只有当该系统调用获得结果或者超时出错时才返回。
In fact, unless otherwise specified, almost all IO interfaces (including socket interfaces) are blocking. This brings a big problem to network programming. For example, when calling recv(1024), the thread will be blocked. During this period, the thread will not be able to perform any operations or respond to any network requests.
A simple solution:
在服务器端使用多线程（或多进程）。多线程（或多进程）的目的是让每个连接都拥有独立的线程（或进程），
这样任何一个连接的阻塞都不会影响其他的连接。
The problem with this program is:
开启多进程或都线程的方式，在遇到要同时响应成百上千路的连接请求，则无论多线程还是多进程都会严重占据系统资源，
降低系统对外界响应效率，而且线程与进程本身也更容易进入假死状态。
improve proposals:
很多程序员可能会考虑使用“线程池”或“连接池”。“线程池”旨在减少创建和销毁线程的频率，
其维持一定合理数量的线程，并让空闲的线程重新承担新的执行任务。“连接池”维持连接的缓存池，尽量重用已有的连接、
减少创建和关闭连接的频率。这两种技术都可以很好的降低系统开销，都被广泛应用很多大型系统，如websphere、tomcat和各种数据库等。
There are actually problems with the improved solution:
“线程池”和“连接池”技术也只是在一定程度上缓解了频繁调用IO接口带来的资源占用。而且，所谓“池”始终有其上限，
当请求大大超过上限时，“池”构成的系统对外界的响应并不比没有池的时候效果好多少。所以使用“池”必须考虑其面临的响应规模，
并根据响应规模调整“池”的大小。
Corresponding to the thousands or even tens of thousands of client requests that may appear at the same time in the above example, "thread pool" or "connection pool" may relieve some of the pressure, but it cannot solve all problems. In short, the multi-threaded model can easily and efficiently solve small-scale service requests, but in the face of large-scale service requests, the multi-threaded model will also encounter bottlenecks. Non-blocking interfaces can be used to try to solve this problem.
4
Non-blocking IO (non-blocking IO)
Under Linux, you can make it non-blocking by setting the socket. When performing a read operation on a non-blocking socket, the flow looks like this:
As can be seen from the figure, when the user process issues a read operation, if the data in the kernel is not ready, it does not block the user process, but returns an error immediately. From the perspective of the user process, after it initiates a read operation, it does not need to wait, but gets a result immediately. When the user process judges that the result is an error, it knows that the data is not ready, so the user can do other things within the time interval from this time to the next time when the read query is issued, or directly send the read operation again. Once the data in the kernel is ready and the system call of the user process is received again, it immediately copies the data to the user memory (it is still blocked at this stage) and returns.
也就是说非阻塞的recvform系统调用调用之后，进程并没有被阻塞，内核马上返回给进程，如果数据还没准备好，
此时会返回一个error。进程在返回之后，可以干点别的事情，然后再发起recvform系统调用。重复上面的过程，
循环往复的进行recvform系统调用。这个过程通常被称之为轮询。轮询检查内核数据，直到数据准备好，再拷贝数据到进程，
进行数据处理。需要注意，拷贝数据整个过程，进程仍然是属于阻塞的状态。
Therefore, in non-blocking IO, the user process actually needs to continuously and actively ask whether the kernel data is ready.
Non-blocking IO example
#服务端
from socket import *
server = socket(AF_INET, SOCK_STREAM)
server.bind(('127.0.0.1',8099))
server.listen(5)
server.setblocking(False)
rlist=[]
wlist=[]
while True:
try:
conn, addr = server.accept()
rlist.append(conn)
print(rlist)
except BlockingIOError:
del_rlist=[]
for sock in rlist:
try:
data=sock.recv(1024)
if not data:
del_rlist.append(sock)
wlist.append((sock,data.upper()))
except BlockingIOError:
continue
except Exception:
sock.close()
del_rlist.append(sock)
del_wlist=[]
for item in wlist:
try:
sock = item[0]
data = item[1]
sock.send(data)
del_wlist.append(item)
except BlockingIOError:
pass
for item in del_wlist:
wlist.remove(item)
for sock in del_rlist:
rlist.remove(sock)
server.close()
#客户端
from socket import *
c=socket(AF_INET,SOCK_STREAM)
c.connect(('127.0.0.1',8080))
while True:
msg=input('>>: ')
if not msg:continue
c.send(msg.encode('utf-8'))
data=c.recv(1024)
print(data.decode('utf-8'))
But the non-blocking IO model is by no means recommended.
We can't otherwise have its advantages: being able to do other work while waiting for the task to complete (including submitting other tasks, that is, "background" can have multiple tasks ""simultaneously"" executing).
But it can't hide its shortcomings:
1. 循环调用recv()将大幅度推高CPU占用率；这也是我们在代码中留一句time.sleep(2)的原因,否则在低配主机下极容易出现卡机情况
2. 任务完成的响应延迟增大了，因为每过一段时间才去轮询一次read操作，而任务可能在两次轮询之间的任意时间完成。
这会导致整体数据吞吐量的降低。
In addition, in this scheme, recv() is more used to detect "whether the operation is completed", and the actual operating system provides a more efficient interface for detecting "whether the operation is completed", such as select() multiplexing With this mode, it is possible to detect if multiple connections are active at once.
5
Multiplexing IO (IO multiplexing)
The term IO multiplexing may be a bit unfamiliar, but if I say select/epoll, I can probably understand. In some places, this IO method is also called event-driven IO
(event driven IO). We all know that the advantage of select/epoll is that a single process can handle the IO of multiple network connections at the same time. Its basic principle is that the select/epoll function will continuously poll all the sockets it is responsible for, and when a certain socket has data arriving, it will notify the user process. Its process is as follows:
当用户进程调用了select，那么整个进程会被block，而同时，kernel会“监视”所有select负责的socket，
当任何一个socket中的数据准备好了，select就会返回。这个时候用户进程再调用read操作，将数据从kernel拷贝到用户进程。
这个图和blocking IO的图其实并没有太大的不同，事实上还更差一些。因为这里需要使用两个系统调用\(select和recvfrom\)，
而blocking IO只调用了一个系统调用\(recvfrom\)。但是，用select的优势在于它可以同时处理多个connection。
emphasize:
1. If the number of connections processed is not very high, the web server using select/epoll is not necessarily better than the web server using multi-threading + blocking IO, and the delay may be greater. The advantage of select/epoll is not that it can handle a single connection faster, but that it can handle more connections.
2. In the multiplexing model, for each socket, it is generally set to non-blocking, but, as shown in the figure above, the entire user process is actually blocked all the time. It's just that the process is blocked by the select function, not by socket IO.
Conclusion: The advantage of select is that it can handle multiple connections, not a single connection
select network IO model example
#服务端
from socket import *
import select
server = socket(AF_INET, SOCK_STREAM)
server.bind(('127.0.0.1',8093))
server.listen(5)
server.setblocking(False)
print('starting...')
rlist=[server,]
wlist=[]
wdata={}
while True:
rl,wl,xl=select.select(rlist,wlist,[],0.5)
print(wl)
for sock in rl:
if sock == server:
conn,addr=sock.accept()
rlist.append(conn)
else:
try:
data=sock.recv(1024)
if not data:
sock.close()
rlist.remove(sock)
continue
wlist.append(sock)
wdata[sock]=data.upper()
except Exception:
sock.close()
rlist.remove(sock)
for sock in wl:
sock.send(wdata[sock])
wlist.remove(sock)
wdata.pop(sock)
#客户端
from socket import *
c=socket(AF_INET,SOCK_STREAM)
c.connect(('127.0.0.1',8081))
while True:
msg=input('>>: ')
if not msg:continue
c.send(msg.encode('utf-8'))
data=c.recv(1024)
print(data.decode('utf-8'))
Process analysis of select monitoring fd changes:
用户进程创建socket对象，拷贝监听的fd到内核空间，每一个fd会对应一张系统文件表，内核空间的fd响应到数据后，
就会发送信号给用户进程数据已到；
用户进程再发送系统调用，比如（accept）将内核空间的数据copy到用户空间，同时作为接受数据端内核空间的数据清除，
这样重新监听时fd再有新的数据又可以响应到了（发送端因为基于TCP协议所以需要收到应答后才会清除）。
Advantages of this model:
相比其他模型，使用select() 的事件驱动模型只用单线程（进程）执行，占用资源少，不消耗太多 CPU，同时能够为多客户端提供服务。
如果试图建立一个简单的事件驱动的服务器程序，这个模型有一定的参考价值。
Disadvantages of this model:
首先select()接口并不是实现“事件驱动”的最好选择。因为当需要探测的句柄值较大时，select()接口本身需要消耗大量时间去轮询各个句柄。
很多操作系统提供了更为高效的接口，如linux提供了epoll，BSD提供了kqueue，Solaris提供了/dev/poll，…。
如果需要实现更高效的服务器程序，类似epoll这样的接口更被推荐。遗憾的是不同的操作系统特供的epoll接口有很大差异，
所以使用类似于epoll的接口实现具有较好跨平台能力的服务器会比较困难。
其次，该模型将事件探测和事件响应夹杂在一起，一旦事件响应的执行体庞大，则对整个模型是灾难性的。3
An understanding of select, poll, epoll
IO复用：为了解释这个名词，首先来理解下复用这个概念，复用也就是共用的意思，这样理解还是有些抽象，
为此，咱们来理解下复用在通信领域的使用，在通信领域中为了充分利用网络连接的物理介质，
往往在同一条网络链路上采用时分复用或频分复用的技术使其在同一链路上传输多路信号，到这里我们就基本上理解了复用的含义，
即公用某个“介质”来尽可能多的做同一类(性质)的事，那IO复用的“介质”是什么呢？为此我们首先来看看服务器编程的模型，
客户端发来的请求服务端会产生一个进程来对其进行服务，每当来一个客户请求就产生一个进程来服务，然而进程不可能无限制的产生，
因此为了解决大量客户端访问的问题，引入了IO复用技术，即：一个进程可以同时对多个客户请求进行服务。
也就是说IO复用的“介质”是进程(准确的说复用的是select和poll，因为进程也是靠调用select和poll来实现的)，
复用一个进程(select和poll)来对多个IO进行服务，虽然客户端发来的IO是并发的但是IO所需的读写数据多数情况下是没有准备好的，
因此就可以利用一个函数(select和poll)来监听IO所需的这些数据的状态，一旦IO有数据可以进行读写了，进程就来对这样的IO进行服务。
理解完IO复用后，我们在来看下实现IO复用中的三个API(select、poll和epoll)的区别和联系
select，poll，epoll都是IO多路复用的机制，I/O多路复用就是通过一种机制，可以监视多个描述符，一旦某个描述符就绪（
一般是读就绪或者写就绪），能够通知应用程序进行相应的读写操作。但select，poll，epoll本质上都是同步I/O，
因为他们都需要在读写事件就绪后自己负责进行读写，也就是说这个读写过程是阻塞的，而异步I/O则无需自己负责进行读写，
异步I/O的实现会负责把数据从内核拷贝到用户空间。三者的原型如下所示：
int select(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds, struct timeval *timeout);
int poll(struct pollfd *fds, nfds_t nfds, int timeout);
int epoll_wait(int epfd, struct epoll_event *events, int maxevents, int timeout);
1.select的第一个参数nfds为fdset集合中最大描述符值加1，fdset是一个位数组，其大小限制为__FD_SETSIZE（1024），
位数组的每一位代表其对应的描述符是否需要被检查。第二三四参数表示需要关注读、写、错误事件的文件描述符位数组，
这些参数既是输入参数也是输出参数，可能会被内核修改用于标示哪些描述符上发生了关注的事件，
所以每次调用select前都需要重新初始化fdset。timeout参数为超时时间，该结构会被内核修改，其值为超时剩余的时间。
select的调用步骤如下：
（1）使用copy_from_user从用户空间拷贝fdset到内核空间
（2）注册回调函数__pollwait
（3）遍历所有fd，调用其对应的poll方法（对于socket，这个poll方法是sock_poll，sock_poll根据情况会调用到tcp_poll,udp_poll
或者datagram_poll）
（4）以tcp_poll为例，其核心实现就是__pollwait，也就是上面注册的回调函数。
（5）__pollwait的主要工作就是把current（当前进程）挂到设备的等待队列中，不同的设备有不同的等待队列，对于tcp_poll 来说，
其等待队列是sk->sk_sleep（注意把进程挂到等待队列中并不代表进程已经睡眠了）。在设备收到一条消息（网络设备）或填写完文件数据
（磁盘设备）后，会唤醒设备等待队列上睡眠的进程，这时current便被唤醒了。
（6）poll方法返回时会返回一个描述读写操作是否就绪的mask掩码，根据这个mask掩码给fd_set赋值。
（7）如果遍历完所有的fd，还没有返回一个可读写的mask掩码，则会调用schedule_timeout是调用select的进程（也就是 current）
进入睡眠。当设备驱动发生自身资源可读写后，会唤醒其等待队列上睡眠的进程。如果超过一定的超时时间（schedule_timeout 指定），
还是没人唤醒，则调用select的进程会重新被唤醒获得CPU，进而重新遍历fd，判断有没有就绪的fd。
（8）把fd_set从内核空间拷贝到用户空间。
总结下select的几大缺点：
（1）每次调用select，都需要把fd集合从用户态拷贝到内核态，这个开销在fd很多时会很大
（2）同时每次调用select都需要在内核遍历传递进来的所有fd，这个开销在fd很多时也很大
（3）select支持的文件描述符数量太小了，默认是1024
2． poll与select不同，通过一个pollfd数组向内核传递需要关注的事件，故没有描述符个数的限制，pollfd中的events字段和revents分别
用于标示关注的事件和发生的事件，故pollfd数组只需要被初始化一次。
poll的实现机制与select类似，其对应内核中的sys_poll，只不过poll向内核传递pollfd数组，然后对pollfd中的每个描述符进行poll，
相比处理fdset来说，poll效率更高。poll返回后，需要对pollfd中的每个元素检查其revents值，来得指事件是否发生。
3．直到Linux2.6才出现了由内核直接支持的实现方法，那就是epoll，被公认为Linux2.6下性能最好的多路I/O就绪通知方法。
epoll可以同时支持水平触发和边缘触发（Edge Triggered，只告诉进程哪些文件描述符刚刚变为就绪状态，它只说一遍，如果我们没有采取行动，
那么它将不会再次告知，这种方式称为边缘触发），理论上边缘触发的性能要更高一些，但是代码实现相当复杂。
epoll同样只告知那些就绪的文件描述符，而且当我们调用epoll_wait()获得就绪文件描述符时，返回的不是实际的描述符，
而是一个代表就绪描述符数量的值，你只需要去epoll指定的一个数组中依次取得相应数量的文件描述符即可，这里也使用了内存映射（mmap）技术，
这样便彻底省掉了这些文件描述符在系统调用时复制的开销。另一个本质的改进在于epoll采用基于事件的就绪通知方式。在select/poll中，
进程只有在调用一定的方法后，内核才对所有监视的文件描述符进行扫描，而epoll事先通过epoll_ctl()来注册一个文件描述符，
一旦基于某个文件描述符就绪时，内核会采用类似callback的回调机制，迅速激活这个文件描述符，当进程调用epoll_wait()时便得到通知。
epoll既然是对select和poll的改进，就应该能避免上述的三个缺点。那epoll都是怎么解决的呢？在此之前，我们先看一下epoll 和select和
poll的调用接口上的不同，select和poll都只提供了一个函数——select或者poll函数。而epoll提供了三个函 数，
epoll_create,epoll_ctl和epoll_wait，epoll_create是创建一个epoll句柄；epoll_ctl是注 册要监听的事件类型；
epoll_wait则是等待事件的产生。
对于第一个缺点，epoll的解决方案在epoll_ctl函数中。每次注册新的事件到epoll句柄中时（在epoll_ctl中指定 EPOLL_CTL_ADD），
会把所有的fd拷贝进内核，而不是在epoll_wait的时候重复拷贝。epoll保证了每个fd在整个过程中只会拷贝 一次。
对于第二个缺点，epoll的解决方案不像select或poll一样每次都把current轮流加入fd对应的设备等待队列中，而只在 epoll_ctl时把
current挂一遍（这一遍必不可少）并为每个fd指定一个回调函数，当设备就绪，唤醒等待队列上的等待者时，就会调用这个回调 函数，
而这个回调函数会把就绪的fd加入一个就绪链表）。epoll_wait的工作实际上就是在这个就绪链表中查看有没有就绪的fd
（利用 schedule_timeout()实现睡一会，判断一会的效果，和select实现中的第7步是类似的）。
对于第三个缺点，epoll没有这个限制，它所支持的FD上限是最大可以打开文件的数目，这个数字一般远大于2048,举个例子,
在1GB内存的机器上大约是10万左右，具体数目可以cat /proc/sys/fs/file-max察看,一般来说这个数目和系统内存关系很大。
总结：
（1）select，poll实现需要自己不断轮询所有fd集合，直到设备就绪，期间可能要睡眠和唤醒多次交替。而epoll其实也需要调用 epoll_wait
不断轮询就绪链表，期间也可能多次睡眠和唤醒交替，但是它是设备就绪时，调用回调函数，把就绪fd放入就绪链表中，并唤醒在 epoll_wait中
进入睡眠的进程。虽然都要睡眠和交替，但是select和poll在“醒着”的时候要遍历整个fd集合，而epoll在“醒着”的 时候只要判断一下就绪链表
是否为空就行了，这节省了大量的CPU时间，这就是回调机制带来的性能提升。
（2）select，poll每次调用都要把fd集合从用户态往内核态拷贝一次，并且要把current往设备等待队列中挂一次，而epoll只要 一次拷贝，
而且把current往等待队列上挂也只挂一次（在epoll_wait的开始，注意这里的等待队列并不是设备等待队列，只是一个epoll内
部定义的等待队列），这也能节省不少的开销。
Two selectors module
These three IO multiplexing models have different support on different platforms, and epoll does not support it under Windows. Fortunately, we have the selectors module, which helps us to choose the most suitable one under the current platform by default.
#服务端
from socket import *
import selectors
sel=selectors.DefaultSelector()
def accept(server_fileobj,mask):
conn,addr=server_fileobj.accept()
sel.register(conn,selectors.EVENT_READ,read)
def read(conn,mask):
try:
data=conn.recv(1024)
if not data:
print('closing',conn)
sel.unregister(conn)
conn.close()
return
conn.send(data.upper()+b'_SB')
except Exception:
print('closing', conn)
sel.unregister(conn)
conn.close()
server_fileobj=socket(AF_INET,SOCK_STREAM)
server_fileobj.setsockopt(SOL_SOCKET,SO_REUSEADDR,1)
server_fileobj.bind(('127.0.0.1',8088))
server_fileobj.listen(5)
server_fileobj.setblocking(False) #设置socket的接口为非阻塞
sel.register(server_fileobj,selectors.EVENT_READ,accept) #相当于网select的读列表里append了一个文件句柄
#server_fileobj,并且绑定了一个回调函数accept
while True:
events=sel.select() #检测所有的fileobj，是否有完成wait data的
for sel_obj,mask in events:
callback=sel_obj.data #callback=accpet
callback(sel_obj.fileobj,mask) #accpet(server_fileobj,1)
#客户端
from socket import *
c=socket(AF_INET,SOCK_STREAM)
c.connect(('127.0.0.1',8088))
while True:
msg=input('>>: ')
if not msg:continue
c.send(msg.encode('utf-8'))
data=c.recv(1024)
print(data.decode('utf-8'))
Implementing concurrent FTP based on selectors module
Reference: Link: https://pan.baidu.com/s/1qYPrHCg Password: 9is4
6
Asynchronous IO (Asynchronous I/O)
Asynchronous IO under Linux is not used much, and it was introduced from the kernel version 2.6. Let's take a look at its process:
After the user process initiates the read operation, it can start doing other things immediately. On the other hand, from the perspective of the kernel, when it receives an asynchronous read, it will return immediately, so it will not generate any block for the user process. Then, the kernel will wait for the data preparation to complete, and then copy the data to user memory. When all this is done, the kernel will send a signal to the user process, telling it that the read operation is complete.
4
Comparative analysis of IO models
So far, all four IO Models have been introduced. Now go back and answer the first few questions: what is the difference between blocking and non-blocking, and what is the difference between synchronous IO and asynchronous IO.
Answer the simplest one first: blocking vs non-blocking. In fact, the previous introduction has clearly explained the difference between the two. Calling blocking IO will block the corresponding process until the operation is completed, while non-blocking IO will return immediately when the kernel is still preparing data.
再说明synchronous IO和asynchronous IO的区别之前，需要先给出两者的定义。Stevens给出的定义（其实是POSIX的定义）是这样子的：
A synchronous I/O operation causes the requesting process to be blocked until that I/O operationcompletes;
An asynchronous I/O operation does not cause the requesting process to be blocked;
两者的区别就在于synchronous IO做”IO operation”的时候会将process阻塞。按照这个定义，四个IO模型可以分为两大类，
之前所述的blocking IO，non-blocking IO，IO multiplexing都属于synchronous IO这一类，而 asynchronous I/O后一类 。
有人可能会说，non-blocking IO并没有被block啊。这里有个非常“狡猾”的地方，定义中所指的”IO operation”是指真实的IO操作，
就是例子中的recvfrom这个system call。non-blocking IO在执行recvfrom这个system call的时候，如果kernel的数据没有准备好，
这时候不会block进程。但是，当kernel中数据准备好的时候，recvfrom会将数据从kernel拷贝到用户内存中，这个时候进程是被block了，
在这段时间内，进程是被block的。而asynchronous IO则不一样，当进程发起IO 操作之后，就直接返回再也不理睬了，直到kernel发送一个信号，
告诉进程说IO完成。在这整个过程中，进程完全没有被block。
各个IO Model的比较如图所示：
After the above introduction, you will find that the difference between non-blocking IO and asynchronous IO is still obvious. In non-blocking IO, although the process will not be blocked most of the time, it still requires the process to actively check, and when the data preparation is completed, the process also needs to actively call recvfrom again to copy the data to the user memory . And asynchronous IO is completely different. It is like the user process handing over the entire IO operation to someone else (kernel) to complete, and then signaling when the other person is done. During this period, the user process does not need to check the status of IO operations, nor does it need to actively copy data.
5