nvmecontrol devlist: Handle disconnected Fabrics hosts

[jhb]

Summary

If a Fabrics host is disconnected, use the cached controller data
instead of reading the cdata via a pass-through command. In addition,
annotate disconnected hosts including the amount of time since the
connection was lost.

Sponsored by: Chelsio Communications

[imp]

sbin/nvmecontrol/devlist.c

134

Why _FAST? It's called once.

Also, you could base this on the tv_sec field being != 0. That can Never happen for CLOCK_REALTIME.

[jhb]

sbin/nvmecontrol/devlist.c

134

Just that there's no reason for any real precision here given we only care about seconds granularity. I was assuming that we should use _FAST by default unless more precision is required?

[imp]

sbin/nvmecontrol/devlist.c

134

In a loop, definitely optimize for fast. For a one-off call, optimizing for fast is harder to read, imho, so I use the standard form. It saves hitting the hardware once to get the latest TC, which is super cheap on modern time counters, and not that expensive on the exotics.

But where does last_disconnect get populated from? The kernel sets it, I know, but how/where is that set?

[jhb]

Use CLOCK_REALTIME

[jhb]

sbin/nvmecontrol/devlist.c

134

last_disconnect is updated via nanotime() in the softc each time a host disconnects. That field and the call to nanotime() are in the previous review in this stack (D48219) adding the connection status ioctl.

[jhb]

Sample output:

root@sam:~ # nvmecontrol devlist
 nvme0: FreeBSD (disconnected for 3 seconds)