On Mon, Aug 25, 2025 at 09:53:27PM -0700, Caleb Sander Mateos wrote:

On Mon, Aug 25, 2025 at 5:49 AM Ming Lei [off-list ref] wrote:

quoted

Add a new command line option --no_ublk_fixed_fd that excludes the ublk
control device (/dev/ublkcN) from io_uring's registered files array.
When this option is used, only backing files are registered starting
from index 1, while the ublk control device is accessed using its raw
file descriptor.

Add ublk_get_registered_fd() helper function that returns the appropriate
file descriptor for use with io_uring operations, taking ublk_queue *
parameter instead of ublk_thread * for better performance.

Key optimizations implemented:
- Cache UBLKS_Q_NO_UBLK_FIXED_FD flag in ublk_queue.flags to avoid
  reading dev->no_ublk_fixed_fd in fast path
- Cache ublk char device fd in ublk_queue.ublk_fd for fast access
- Update ublk_get_registered_fd() to use ublk_queue * parameter
- Update io_uring_prep_buf_register/unregister() to use ublk_queue *
- Replace ublk_device * access with ublk_queue * access in fast paths

This improves performance by:
- Eliminating device structure traversal in hot paths
- Better cache locality with queue-local data access
- Reduced indirection for flag and fd lookups

Are you saying that performance is better when using the raw
/dev/ublkcN fd instead of an io_uring registered file? That would be
really surprising to me, since the whole point of io_uring file
registration is to avoid the file reference counting in the I/O path.

No, here it just describes one implementation detail by caching per-device
flag(no_ublk_fixed_fd) in queue's flag.

The test can trigger hang with patch V1 by passing --no_ublk_fixed_fd
because /dev/ublkcN and io_uring closes can be depended, both are run
from task work context in current task.


Thanks,
Ming