Allow UMA allocated memory to be freed when VNET jails are torn down.

[• rodrigc]

Summary

MFp4: @181627

Allow UMA allocated memory to be freed when VNET jails are torn down.

Originally based on patch by bz.

[adrian]

Interesting; TCP has the same behaviour. Why's that?

[adrian]

.. in fact, is this what UMA_ZONE_NOFREE means? My 30 second check of uma source + manpage says that the slabs are never returned back to the VM - but I haven't yet established whether this is just the case during paging, or whether the path through zone destroy also won't return the keg slabs back to the VM.

[rwatson]

This may be safe, but I'd like to see a clear argument grounded in the revision history as to why that is true. Historically, this would not have been safe, which is why the flags are there.

(I may have made it safe during the network-stack locking work, but my recollection of the final state of things is a bit hazy, hence thinking a good spelunking is the right answer here.)

[gnn]

I agree with Robert. We need to know why this will work, because the diff doesn't tell us that.

[bz]

Ok, I'll leave the comment here as well; in 2010(?) I had convinced myself that it would be ok and ran this for a while.
However the code has changed since (normal churn, UDP-Lite, blah bleu blupp) so we need to re-evaluate this.
For everything else, I think the discussion on net@ is more fruitful at the moment.

[rwatson]

This patch is probably safe if limited to UDP.

[gnn]

No objection from me.

[• rodrigc]

Committed in rS275555