redo log seems use too many expensive operations for NFS

[hicqu]

Before asking a question, make sure you have

• Searched existing Stack Overflow questions
• Googled your question
• Searched open and closed GitHub issues
• Read the documentation

What is your question?

Some users store redo-log on NFS and they found some operations are too expensive for NFS:

We can see the latency of fsync, rename and trunc is too high.

fsync is called in WriteLog and flushLogMeta; rename is called in flush and flushLogMeta. trunc is only called in flushLogMeta.

And, flushLogMeta holds LogWriter.metaLock so it can block setMaxCommitTs and GetCurrentResolvedTs.

We need to

1. move all I/O operations into background goroutines, so that all redo-log interfaces can be no-blocking;
2. improve the usage of LogWrite.metaLock.