[PATCHSET,0/4,v2] Support for mapping SQ/CQ rings into huge page

Message

Jens Axboe May 13, 2023, 2:16 p.m. UTC

Hi,

io_uring SQ/CQ rings are allocated by the kernel from contigious, normal
pages, and then the application mmap()'s the rings into userspace. This
works fine, but does require contigious pages to be available for the
given SQ and CQ ring sizes. As uptime increases on a given system, so
does memory fragmentation. Entropy is invevitable.

This patchset adds support for the application passing in a pre-allocated
huge page, and then placing the rings in that. This reduces the need for
contigious pages, and also reduces the TLB pressure for larger rings.

The liburing huge.2 branch has support for using this trivially.
Applications may use the normal ring init helpers and set
IORING_SETUP_NO_MMAP, in which case a huge page will get allocated for
them and used. Or they may use io_uring_queue_init_mem() and pass in
a pre-allocated huge page, getting the amount of it used returned. This
allows placing multiple rings into a single huge page.

Changes since v1:
- Mandate that we're using a single page. May be a normal page if we
  don't need a lot of memory, or a huge page if the ring itself takes
  up more space than a single normal page.