IORING_SETUP_NO_MMAP questions

[lano1106]

I have a hard time keeping up with all the new features added to io_uring...

the latest that I have looked into is using the IORING_SETUP_NO_MMAP.

Here are my questions:

1. Do you have any benchmarks or does it make any performance difference in using IORING_SETUP_NO_MMAP is it just a matter to better use the system memory resources?

I have 3 rings in my app.

all 3 are configured to have 256 sq entries.

One of them, the default cqe number is overriden to 8192 with IORING_SETUP_CQSIZE.
That ring do use IORING_RECV_MULTISHOT to manager 25 sockets and without that much cqes, it is constantly overflowing.

I am currently calculating the memory usage to see if I could cram all 3 in a single huge page...

2. Where does the MAP_SHARED requirement comes from?

If it is a bit of mystery to me considering that io_uring is fine with memory coming from a private anonymous mapping for the provided buffers.

3. Where does the 2MB huge page size come from?

There is a fix me comment in the setup.c file above the 2MB value. I will go read Documentation/mm/...

but in the Michael Kerrisk book, they mention that on x86, the size is 4MB... and I am pretty sure that the huge page size is configurable on the kernel command line param... I am pretty sure to have already seen a 3GB huge page sizes...

So please comment on the huge page size setting because this is something that is confusing me...

thx!