Thinking about the cooperation process of Network Library -- Some Thoughts on reading brpc

bthread One of the functions of ： take Reactor Synchronization in IO Change asynchronous IO.

Theoretically Reactor Not as efficient as Proactor, The concrete manifestation is that although Reactor The model avoids waiting to be readable / The time when the writable event is ready , however read/write The timing of the call itself is inevitable （ That is, the data copy between the kernel buffer and the user buffer occupies the current thread CPU The time of the ）, and Proactor But this time can be avoided , So it's more efficient .

So why do you say Proactor More efficient ？ although Proactor No need to wait read/write In itself , But only the read and write processing time is transferred to the operating system kernel thread , The total processing time does not decrease .

The key is to put read/write The time is transferred to the operating system , The current thread can continue to process the next connection , Although the total throughput has not changed , But the average processing delay per request is reduced . therefore brpc Zhongba IO Operation to new bthread perform （https://github.com/apache/incubator-brpc/blob/master/docs/cn/io.md#%E6%94%B6%E6%B6%88%E6%81%AF） It is equivalent to handing over to support asynchronous IO Operating system to execute , Both of them can reduce the average request delay .

hook system call

With asynchrony IO and bthread The Apocalypse of , We would think of rewriting all system calls that would block in this way , This leads to hook.

Simply speaking hook Library code is a wrapper around system calls , The purpose is to block the system call , Not directly blocking the current thread , It blocks the current bthread, The current thread immediately executes another bthread, This avoids thread blocking .

This is equivalent to calling the system + The thread model is delegated to hook System call after + coroutines .（ Thread switching occurs when the system call is blocked ,hook When the subsequent system call is blocked, a coprocessor switch occurs , The cost of CO process switching is less than that of threads , So the latter has better performance ）.

however brpc did not hook system call , This is because ：

• bthread Blocking may switch system threads , Dependent systems TLS The behavior of the function of is undefined .
• And blocking pthread May deadlock when mixed functions of .
• This kind of hook Functions themselves are generally less efficient , Because there are often additional system calls , Such as epoll.