From Denys Vlasenko:

From Denys Vlasenko:

This makes more clear what fds are closed when, when you read
caller function code.

This is a preparation to not clobber fdp[] in the child, so that
we can switch to using vfork() in a later patch.

Add ///BUG comments:
(*) we wait() in writechild, thus, if write() in writing child blocks
waiting for decompressor to consume data, and decompressor blocks waiting
ofr _us_ to read its data, we would deadlock.
(*) we can't know that the child we spawned is the only running child
of the process. We should use waitpid() to wait specifically for
the child we need.

Will fix in patch 4.