submit_and_wait(1) does not return w/ nop and recvmsg_multishot

[TwoClocks]

I'm having all kinds of issues trying to get recvmsg_multishot to work on some multicast addresses over localhost. Sometimes I get data on some sockets, but not others. Sometimes I don't get anything. Sometimes it works fine. I'm sure I'm doing something wrong, but I'm not sure what.

But the oddest thing, is that submit() is not returning w/ a NOP sqe in the list.

Here is the SQE when I submit_and_wait:

os.linux.io_uring_sqe.io_uring_sqe{ .opcode = os.linux.IORING_OP.RECVMSG, .flags = 37, .ioprio = 2, .fd = 4, .off = 0, .addr = 140100692133480, .len = 1, .rw_flags = 0, .user_data = 1369234387412764160, .buf_index = 43520, .personality = 0, .splice_fd_in = 0, .addr3 = 0, .resv = 0 }
    -- flags:0b100101 ioprio:0b10
os.linux.io_uring_sqe.io_uring_sqe{ .opcode = os.linux.IORING_OP.RECVMSG, .flags = 37, .ioprio = 2, .fd = 5, .off = 0, .addr = 140100692135016, .len = 1, .rw_flags = 0, .user_data = 1369234387412765696, .buf_index = 43521, .personality = 0, .splice_fd_in = 0, .addr3 = 0, .resv = 0 }
    -- flags:0b100101 ioprio:0b10
os.linux.io_uring_sqe.io_uring_sqe{ .opcode = os.linux.IORING_OP.NOP, .flags = 0, .ioprio = 0, .fd = 0, .off = 0, .addr = 0, .len = 0, .rw_flags = 0, .user_data = 864831229147275264, .buf_index = 0, .personality = 0, .splice_fd_in = 0, .addr3 = 0, .resv = 0 }
    -- flags:0b0 ioprio:0b0	

Even if my setup of the recvs is wrong, I would expect this .NOP to return right away. Instead submit_and_wait(1) never returns. The behavior is the same if I setup the ring with IORING_SETUP_SQPOLL or not.

If I re-order the requests, or I pre-pend other requests in the same submission, the behavior changes, but at least submit() returns.

running a 6.12.16 kernel.

What's the best way to debug/trace how the kernel is processing this queue?

[axboe]

Please provide a complete code example that shows it. I can almost certainly guarantee that this is a coding issue, it's not an io_uring issue, as it's a pretty basic usage.

[axboe]

For tracing, you can do:

echo 1 > /sys/kernel/debug/tracing/events/io_uring_enable

run your code, and then

cat /sys/kernel/debug/tracing/trace

[TwoClocks]

Thanks!

Using the trace was clutch. Is that in the man page someplace?
Found my bug. Was setting a bad flag and linking requested that shouldn't be linked. (similar name, different flag field).

I'm still a bit surprised that the submit did not return w/ a valid NOP. The kernel trace said it completed the NOP fine.

[TwoClocks]

oh! I see. My bug was thus:
I was setting the flags with IORING_RECVSEND_FIXED_BUF when I meant to be using IOSQE_FIXED_FILE (auto-complete). Causing the NOP to be linked to a recv requests. It all makes sense now.

[axboe]

Using the trace was clutch. Is that in the man page someplace?

Don't think it is... Should probably be, but not quite sure where to put it.