@dfaure does this makes sense to you ?

I don't really understand the relation to ksmserver, surely kioslaves are not part of session management.

I also don't really understand what would happen in the new event loop (select/poll) of the kioslave while the non-blocking I/O is happening. It's not like the slave can handle other commands in parallel. Slaves are in a separate process exactly so that they can block.

Regarding O_DIRECT, I have some doubts. The man page you linked to, says

In summary, O_DIRECT is a potentially powerful tool that should be used with caution. It is recommended that applications treat use of O_DIRECT as a performance option which is disabled by default.
"The thing that has always disturbed me about O_DIRECT is that the whole interface is just stupid, and was probably designed by a deranged monkey on some serious mind-controlling substances."--Linus"`

Are you sure you want to use that? :-)
More precisely, it says it will break with NFS (in some cases).
And it sounds like a bad buffer size with O_DIRECT will slow things down (too many direct writes instead of buffering).
Overall, I'd say -- did you check what the cp source code does? That's what we're competing against :-)

copy https://github.com/coreutils/coreutils/blob/master/src/copy.c#L301 use https://man7.org/linux/man-pages/man2/copy_file_range.2.html or https://github.com/coreutils/gnulib/blob/master/lib/full-write.h
Did not know about copy_file_range, it was introduced in kernel 4.5 and improved/reworked in 5.3 apparently.
It is simply newer than sendfile (but older than io_uring), it seems to be just almost a drop-in replacement for sendfile with reflink aka COW support for supporting fs built-in.
That should be the next area for investigation, it is way simpler than our current reflink support, but it is Linux only.

The io_uring cp example is a good reference to make use of the API.
https://github.com/axboe/liburing/blob/master/examples/io_uring-cp.c

And splice test.
https://github.com/axboe/liburing/blob/a209abe65befe603311b6fcd861e14d28e3e2f2d/test/splice.c

I have a first step : split to two loops use copy_file_range

https://invent.kde.org/frameworks/kio/-/merge_requests/602

Found some hints how to help the kernel manage its buffers and page cache efficiently when writing
http://lkml.iu.edu/hypermail/linux/kernel/1005.2/01845.html
http://lkml.iu.edu/hypermail/linux/kernel/1005.2/01953.html

Also additional fadvise might be worse testing.

(void)posix_fadvise(fd, 0, 0, POSIX_FADV_WILLNEED);
(void)posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL);
(void)posix_fadvise(fd, 0, 0, POSIX_FADV_NOREUSE);

In any case, I desperately need a benchmark tool as kioclient can't really play that role.
Something that would link directly with file IOSLAVE.