I tried using a tasklet for the TX part, but that did not help.
Another way is to prepare the buffer for the next read and read the data 
immediately.
Something like this:

1. Call tty_prepare_flip_string while startup.
2. When the read interrupt arrives, read the data immediately.
3. Call tty_prepare_flip_string for the next read.

Only you then don't know the size of space required/

Again, the problem here is that we need to use global variables to store the 
pre-allocated buffers

You don't. You can store them per port in the existing objects !

The best way is if we can keep the current implementation, one extra copy is 
not hurting us
as we do it after the read_data API.

No problem with that at all.