iscsi: Use calloutng instead of ticks in iscsi initiator

[rscheff]

Summary

This is the initial patch to improve the timing system for
the iscsi initiator.

Preliminary plan to reduce ping/timeout scheduling overhead, while
also improving the functionality to more closely align with the
expected behavior:

1. move to calloutng - with same 1-sec callouts (and a 10 ms

window, to allow some kernel scheduling improvements)

2. change internal representation of (per session) timeouts

from seconds to milliseconds

3. allow per-session timeouts to be freely configurable, with

bounded minimum of 100ms

4. perform iscsid/login timeout callout only once (if configured)

instead of once every second.

5. perform iscsi ping timeout three times in the

defined interval to transmit SCSI_NOP packets - rather than sending
one ping every second even when very long timeouts are configured.

6. improve loss-of-connectivity detection by tracking socket buffer

levels between pings - if the socket buffer does not drain between
pings, better not wait for TCP RTO firing (which could be up to 512
seconds in between), but rather re-connect session quickly.

sysctls retain 1-sec granularity, only allow per-session settings
to be parsed as floats (e.g. 0.100 for 100 milliseconds)

Retain log output unless fractional-second timeouts are in use.

[rscheff]

• make the call centered around the 1 sec mark

[jhb]

sys/dev/iscsi/icl_soft_proxy.c

188–190

Perhaps define these in terms of mstosbt()? I'm not sure rounding errors matter as much for milliseconds, but I think in general we prefer the APIs in <sys/time.h> over direct math with the SBT_* constants, something like:

sbt = mstosbt(995);
pr = mstosbt(10);

[rscheff]

• make use of mstosbt()

[rscheff]

sys/dev/iscsi/icl_soft_proxy.c

188–190

Thanks for this hint!

[mav]

I'm sorry for late look. I was on vacation. I have no objections, just some comments.

sys/dev/iscsi/icl_soft_proxy.c

188–190

I just hope compiler preprocess those function calls all the way into a single value, as it would do for old SBT_* math. I'd also do it inline without using variables. I am also not sure it makes much sense to shift down the sbt value by half of precision, since I doubt there will be even distribution.