io_uring_prep_splice using registered files

[manticore-projects]

Greetings!

apologies upfront, I am just a simple Java guy trying to wrap libUring into Java using FFM -- and I have basically no idea what I am doing.

Though I got a basic C sample running copying a file using Pipes and Splice. Everything works fine until I try to register the files and the pipe and provide io_uring_prep_splice with the indices.

This works:

io_uring_prep_splice(sqe1, src_fd, offset, pipefd[1], -1, to_copy, SPLICE_F_MOVE);
io_uring_prep_splice(sqe2, pipefd[0], -1, dst_fd, offset, to_copy, SPLICE_F_MOVE);
submit_count+=2;

This does not work:

int fds[4] = {src_fd, dst_fd, pipefd[0], pipefd[1]};
io_uring_register_files(&ring, fds, 4);

io_uring_prep_splice(sqe1, 0, offset, 3, -1, to_copy,  SPLICE_F_FD_IN_FIXED | SPLICE_F_MOVE);

io_uring_prep_splice(sqe2, 2, -1, 1, offset, to_copy, SPLICE_F_MOVE);
sqe2->flags |= IOSQE_FIXED_FILE;

(Calling io_uring_prep_splice does not fail, but after io_uring_submit I can't access any CQEs.)

Would you like to point out please, what I am doing wrong here?
The documentation is a bit scarce on this topic.

Thanks a lot in advance, best and cheers!

[manticore-projects]

Never mind, I eventually figured it out (brute forcing my way through all possible combinations :-) ).

This works:

int fds[4] = {src_fd, dst_fd, pipefd[0], pipefd[1]};

io_uring_prep_splice(sqe1, 0, offset, 3, -1, to_copy, SPLICE_F_FD_IN_FIXED);
sqe1->flags |= IOSQE_FIXED_FILE;
sqe1->flags |= IOSQE_IO_LINK;  // Link to next

io_uring_prep_splice(sqe2, 2, -1, 1, offset, to_copy, SPLICE_F_FD_IN_FIXED);
sqe2->flags |= IOSQE_FIXED_FILE;

And it seems to be reliably faster than using FileDescriptors and on par with CoreUtils cp:

[manticore-projects]

In case anyone is interested: I got Zero Copy File transfer with Pipes and Fixed FileDescriptors working in Java: https://github.com/manticore-projects/JUring/blob/8ce605afbd45b3a2caf15f012ddba96f89b7e2f8/src/test/java/com/davidvlijmincx/lio/api/JUringTest.java#L289

My ultimate goal was to write a libUring backed Java File Channel that can be plugged into Java H2 Database for faster IO.

[manticore-projects]

We have managed to get on par with Java's Zero-Copy NIO Filechannel copy:

• jUring unoptimized block transfer is already faster than NIO based transfer w/o Zero Copy
• jUring zero copy is faster than NIO based Zero Copy

Benchmark	Mode	Threads	Samples	Score	Score Error (99.9%)	Unit
jUringCopy	ss	1	200	68.406348	1.405607	ms/op
jUringCopyFixedFiles	ss	1	200	66.854152	0.860164	ms/op
jUringTransferTo	ss	1	200	143.726446	3.302417	ms/op
nioTransferTo	ss	1	200	154.855567	37.892569	ms/op
nioTransferToZeroCopy	ss	1	200	67.54023	0.875534	ms/op

Though I wonder, why Fixed File Splice is slower (contrary to what the C reference implementation has shown). Does anyone have an idea or an hint please?