Server-A:
State - BACKUP
priority - 200
nopreempt
Server-B:
State - BACKUP
priority - 100

First of all, VIP is at Server-A.
Then i plug out Server-A's cable, so VIP is at Server-B.
Finally I plug in Server-A's cable, and Server-A preempt the VIP.

Questions:
1. It seems that the flag 'nopreempt' doesn't work, WHY?
2. What can I do if I want to forbid Server-A from preempting the VIP?

You can change the behaviour of the kernel by recompiling this.

I don't think recompling the kernel will fix this problem in keepalived
It seems the switch's prolem, when I re-plugin the cable, the IP address is set to eth0 immediately, but cannot ping it from other hosts in about 30 seconds. I have read the source code of keepalived and the RFC2338, If a host in BACKUP state, it should receive advertisement ARP from MASTER.
So in my case, when I re-plugin the cable of Server-A, Server-A runs into BACKUP state and waits for MASTER's advertisement, but at this time Server-A cannot link with LAN in 30 seconds, so Server-A itself become the MASTER, when it can link with other hosts, it sents advertisment ARP to them and preempt the VIP from the other hosts.

Above is my guess