user: imporve #463, impact on the performance of the tested program

[cfc4n]

设计思路

问题

• 为了读取到TLS握手完成后的client_random等密钥，必需要选择一个合适的HOOK函数。
• SSL_write\SSL_read时，TLS握手是建立完成的，但调用过于频繁，会带来性能问题，参见 TLS 模式下，对被检测程序的性能影响。 #463
• 综合来看，合适的HOOK函数需要满足以下几个条件:
  1. 函数是在TLS握手完成后调用
  2. 函数名在动态链接库的符号表中是导出状态
  3. 函数是低频调用

解决思路

• 在 openssl 类库中，以客户端角色调用 SSL_connect 或者以服务端角色 SSL_accept ，最终都会进入 ssl/statem/statem.c 的 state_machine 函数进行TLS握手。
• 所以，可选范围是在这个函数内以大写SSL开头的函数。
• 当使用openssl的方式为同步调用时，TLS握手成功会返回1，也就是ret = 1，即需要在这个变量赋值后，被调用的函数，才能拿到符合要求的内存数据。 state_machine函数内符合要求的就只有SSL_get_wbio了。
• 当使用openssl的方式为异步 (BIO)调用时，还需要增加SSL_in_before函数。

---

Design concept

Question

• In order to read the client_random and other keys after TLS handshake is completed, it is necessary to select a suitable HOOK function.
• When SSL_write\SSL_read, the TLS handshake is already established, but calling it too frequently will cause performance issues, see TLS 模式下，对被检测程序的性能影响。 #463 for reference.
• Overall, a suitable HOOK function needs to meet the following conditions:
  1. The function is called after TLS handshake completion
  2. The function name is exported in the symbol table of the dynamic link library
  3. The function has low-frequency calls

Solution

• When calling SSL_connect in the OpenSSL library in a client role or SSL_accept in a server role, the execution flow ultimately enters the state_machine function in ssl/statem/statem.c for TLS handshake.
• Therefore, the optional scope is functions within this state_machine function that start with an uppercase SSL.
• When using OpenSSL synchronously, a successful TLS handshake returns 1, i.e., ret = 1. Thus, after this variable is assigned, the called functions can obtain the desired memory data. the only function within the state_machine function that meets the requirements is SSL_get_wbio.
• When using openssl in an asynchronous (BIO) manner, it is also necessary to add the SSL_in_before function.

[cfc4n]

测试结果 Result

[cfc4n]

reviewed.

[yifannir]

最终是选择换了hook点的函数么？可否再测试一种场景，本地大文件wget下载场景，在修改之前，对性能影响会有约20%的下载速度下降，本地nginx，约700m->500m