I fear this is the last bad choice we have.

Fix some comment wrapping.

mixed up min / max

Improve comments. Wait for reset to finish. stay in the 'resetting'
state until resume completes (though this seems to only affect
queueing of new commands on I/O completions which shouldn't be
possible w/o the hardware queues present....)

Redid with the code move into nvme_ctrlr... tests fine.. like that better (will make ahci easier to do too)...

Move things to export less from nvme_ctrlr

Seems not at all right to me...

It looks good to my eye, modulo my suggestion.

LGTM, but I didn't test it.

From IRC this morning:
[09:40:36] <DamnHippi> bsdimp: I like the idea. I sure wish there was a way to get the relationship to interrupts or intr config hooks into the name (but the name is already pretty long).
[09:41:17] <bsdimp> I struggled with the name as well...
[09:41:34] <bsdimp> bus_delayed_attach(dev, BUS_INTR);
[09:41:43] <bsdimp> was the second choice I had to have a mask of things that you delay it for.
[09:41:46] <DamnHippi> hmmm, interesting.
[09:42:03] <bsdimp> but then that isn't just the children, etc
[09:42:12] <DamnHippi> there was another change pending having to do with delaying attachment.
[09:42:17] <bsdimp> so I went with the simple, but I'm open to something more generic
[09:42:23] <DamnHippi> hang on a sec
[09:42:25] <bsdimp> BUS_NEWDEV
[09:42:39] <bsdimp> or whatever so that nathanw's ideas could be done this way too.
[09:42:54] <bsdimp> The device is partially attach, and finishes attaching when the right device(s) appear.
[09:43:15] <DamnHippi> https://reviews.freebsd.org/D15229
[09:43:41] <bsdimp> yea, that's the one. I love the idea, and hate how he did it.
[09:43:48] <DamnHippi> I haven't even looked at it, just saw it mentioned the other day.
[09:43:51] <bsdimp> and was looking for a better way.
[09:43:56] <bsdimp> and this may be it.
[09:44:42] <DamnHippi> but I can't sidetrack into this in depth right now, I'm kinda slammed at $work.
[09:44:43] <bsdimp> It's more similar to config_intrhook() than this though. config_dependend_device_now_here() but you can't have a device_t to pass to it because that device doesn't exist :)
[09:45:16] <bsdimp> DamnHippi works for me. lemme think about your feedback and we'll chat tomorrow. I have a couple of weekend tests I want to run and I gotta get them setup.

Looks great to me too... Forgot to click Accept yesterday when I looked at it.

Seems a bit wasteful to have multiple files with the same bits, but since some are arrays and others are pointers, I'm unsure of what to suggest instead.

Thanks for doing the compat shim. I think this would have been one area where it would have been hard to recover from a make installworld before a new kernel was in place...

I built this on -current and got an insta-panic when trying to lock a null spin lock in attach...

Looks sane to me. Thanks for digging to the bottom of this rathole...