On 07/16/10 07:49, jeff.liu wrote: > For now, I am inclined to separate efficient read through fiemap > and improve the write and allocation stuff via fallocate() or other ways > later.
I haven't had time to look at it carefully, but here's a very brief review. The code you sent, like what's in the fiemap branch, has a separate version of a chunk of copy.c that does both reading and writing and optimizes both reading and writing by invoking the fiemap ioctls at strategic locations. Instead, it would be better to have a module that separates out the efficient-read stuff by telling copy.c where the next significant input extent is, and then modify copy.c to use that module. On hosts that do not support fiemap, the module would simply report the entire input file as that file's only extent. Surely such an approach would be more modular, and would result in less duplication of code. I can write something along those lines if there's interest and if nobody else wants to take a crack at it.
