mjhjmtu

Let's say I have a file descriptor 4 that is used for reading/writing with the O_APPEND flag for writes. If I have multiple threads writing to this file and let's say periodically I want to switch the file pointed by fd 4, I'd want to do a dup2 call to replace fd 4 with new file. Meanwhile multiple threads will still point at fd 4 for writing. Will I be losing any data or will threads who are writing face errors during the switch?

The Linux man page and the POSIX description rather clearly say that it is, indeed, atomic:

Linux man page on dup2():

The steps of closing and reusing the file descriptor newfd are
performed atomically.

POSIX, under "Rationale":

On the other hand, the dup2() function provides unique services, as no other interface is able to atomically replace an existing file descriptor.

I can't see any reason it should be atomic with regard opening, closing and dup'ing the file descriptor, but not with regard to reads or writes.