Question: In `SQPOLL` mode, how does `io_uring` avoid data race in SQ between the user thread and the kernel polling thread?

[ZhenbangYou]

This question is all about SQ polling. Say, I want to submit a read request, so I do the following three steps:

• io_uring_get_sqe,
• io_uring_prep_read,
• setting user_data .

However, since the kernel is polling, is it possible for the following two scenarios to happen:

• Between step 1 and 2, the kernel takes this SQE. Since this SQE is not initialized yet, the kernel treats it as nop, so I will lose this read operation.
• Between step 2 and 3, the kernel takes this SQE. Since user_data is still not set, I won't get a CQE with expected user_data in the CQ.

I tried very hard to find an answer to this, but I couldn't make it. Thanks so much for answering my question!

[mjasny]

No, because the SQE is only visible to the kernel thread after you do io_uring_submit(ring);. This would be step 4.

Internally liburing does increment the tail of the ring which is polled by the kernel in submit using __io_uring_flush_sq()

[ZhenbangYou]

I see! No wonder without io_uring_submit my SQEs weren't executed in SQPOLL mode. Thank you so much!