nvme: Switch to using unmapped pages for passthrough data transfer

[imp]

Summary

Use unmapped I/O for the data transfers for passthrough ioctls. Nothing
looks into them, so there's no need to map and unmap VAs we'll never
dereference through.

Sponsored by: Netflix

[kib]

sys/dev/nvme/nvme_sysctl.c

56 ↗

(On Diff #160791)

Why 'try'? Would it be enough to say 'use unmapped ...'?

[jhb]

I must say, I would perhaps find it simpler to avoid vmapbuf() and just use vm_fault_quick_hold_pages() if you didn't care about the sysctl. I suspect no one will ever care about the sysctl or toggle it?

[jhb]

This is fine for now (especially given the code slush). I might try to come up with a thin wrapper around vm_fault_quick_hold_pages that can be used here (and in a few other places) and replace the usage of vmapbuf. It's interesting that prior to this change, all callers vmapbuf pass 1 (mapped bufs) to vmapbuf. It might be nice to simplify that API again to remove the unmapped case.

sys/dev/nvme/nvme_sysctl.c

56 ↗

(On Diff #160791)

I agree, this doesn't really try, it is always unmapped.

[imp]

sys/dev/nvme/nvme_sysctl.c

56 ↗

(On Diff #160791)

if (mapbuf || !unmapped_buf_allowed) {
        pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
        bp->b_data = bp->b_kvabase + bp->b_offset;
} else
        bp->b_data = unmapped_buf;

inside of vmapbuf is why i wrote 'try'.

56 ↗

(On Diff #160791)

But I do agree with 'just do it'.

[imp]

So I'll update this to always do it and remove the sysctl.
I'll work on a wrapper around vm_fault_quick_hold_pages for future use and see about converting current users, maybe after the branch.

[imp]

Just do it (I've updated the commit message locally, but arc doesn't autochange)