Alternative API for io_uring_register_buffers for multiple rings with shared backend

[yash760]

I have a HugePage memory backed application that can receive IO on 30 threads. All these 30 threads have their own ring which have shared backend worker thread (io_uring_shared_wq). Given these conditions I want to register my complete HugePage memory on initialisation with the rings of all the threads using io_uring_register_buffers .
Is there any other alternative API other than to call io_uring_register_buffers for each ring of 30 threads. Also does calling register buffer only on the shared worker queue ring works ?

[axboe]

Registered buffers can only be used on a ring where they have been registered. So you'd need to register the same buffer on all rings where you'd be doing IO to/from registered buffers with read/write fixed.

[yash760]

Thanks for the answer @axboe !!! Had a follow up question on this.
How much memory will be used by the library/ring for storing the the struct iovec that point to the registered buffers. For example lets say I register 1024 such iovec, then how much memory utilization will increase? I tried checking the size of ring by using io_uring_mlock_size before and after registering buffers and it returned the same value.

[axboe]

The library doesn't store anything, the kernel side overhead is 16 bytes per registered buffer. So 1024 buffers => 16kb of memory.