DIOCGDISCARDINFO and DIOCDISCARD

[David Holland <dholland-tech%netbsd.org@localhost>]

I had almost forgotten about this; but a few months back when I came
into contact with the wd TRIM support in current I wanted to change
the interface around before it appears in a release.

The current interface is two ioctls:
   - first you call DIOCGDISCARDINFO; this tells you the maximum
     number of blocks that can be discarded at once on the current
     device.
   - Then you call DIOCDISCARD with a block number and (I think)
     block count to actually do it.

I have the following objections to this interface:

1. The interface doesn't need to and shouldn't expose the hardware
limit of how many blocks can be discarded at once; instead, for large
discards (which are not the common case) the loop over block ranges
should be done in the driver.

2. The interface uses DEV_BSIZE units. Rather than promote the further
proliferation of DEV_BSIZE, it should use either byte offsets or the
device's own addressable block size. (I think byte offsets are best;
making everything byte offsets minimizes the chance of sending in the
wrong units.)

3. The interface is ioctl-based. I would like to see VOP_DISCARD for
vnodes (which on regular files creates holes) and a corresponding
block device op on struct bdevsw. And a syscall for user access to the
operation. This would provide access to functionality that's currently
lacking; also it will allow discard to work on vnd devices.

Thoughts?

(we can argue about what to call the syscall; I think there's prior
art but can't remember what at the moment)

-- 
David A. Holland
dholland%netbsd.org@localhost