(I think this diff contains the prior change as well?) I would also suggest mstosbt() here as well.

In D34230#775828, @jhb wrote:

(I think this diff contains the prior change as well?) I would also suggest mstosbt() here as well.

Yes, this is a diff against D34222, which would have to be applied first. I believe this is slightly different behavior since the change to git, where arcanist/phabricator will show the full diff here against main, while a patch created would only give the delta...

But only the intended delta is marked red/green.

As a general point, I wonder if it really makes sense to overload 3 separate events into a single callout vs having separate functions for the different callouts. I think while you are waiting for iscsid you don't want to trigger pings anyway, and this would avoid needing is_timeout at all for the non-ping timeouts (which is now a bit more confusing as it is no longer a count of seconds, but a count of timeouts). There's also no reason to even schedule the timeout when waiting for iscsid I think if iscsid_timeout is zero?

I didn't want to modify the iscsi_callout() function completely. I guess the current design was picked to keep the housekeeping overhead low. Splitting iscsi_callout into a iscsid timeout, login timeout and ping hello function is certainly possible.

However, I propose to do

1. per-session timeouts (D34198) and allow these to be
2. configured at milliseconds granularity (D32540 once re-worked) first - leaving all the sysctl global settings unchanged. Anyone wanting sub-second timeouts has to configure this on a per-session basis.
3. This Diff around changing the core behavior (three pings over "timeout" interval instead one ping every second when no other IO was performed) would follow last.
4. I would want to so some socket sendbuffer probing also, between pings - for a faster resumption after a loss of connectivity