|
Hello Julian,
> May be you know, only bug fixes to 2.4. I don't know
> if Marcelo will agree for such changes, may be they were forgotten
> long time ago when IPVS was included into 2.4. Now it is too late.
Let me handle this for post 2.4.32 time frame. It's a matter how you
sell something, even if it's almost a bubble :).
So in this case we tell Marcelo that we need this change to happen
because of L2 cache issues. It saves us some TLB invalidation in case we
get preemted or a soft IRQ happens between waiting for ip_vs_bind_dest()
to finish and return the stack to ip_vs_schedule(). Also the amount of
possible time spent spinning on locks is reduced in SMP. Hey, at least
theoretically :), but he won't check.
In case we really need a reason to mark it as bugfix, I'll come up with
one, and I'm quite sure there is a hidden one. Like atomic race between
timer expiration on peristent templates while ip_vs_conn_new schedules a
new conn entry and mucks the refct of the dest struct :).
> As for me, i agree with you, ip_vs_bind_dest() is the right place
> to touch inactconns, that is why Wensong changed it in this way but
> i don't have enough time to play with IPVS.
Neither do I, however last week and this week I'm assigned to a project
where we need LVS since the existing Cisco HW cannot handle what we need
for the architecture and thus I'm back with some development. Feels good
to dive into the LVS code again after so much time.
Thanks for your help, Julian, as always it's much appreciated. I'll send
an updated patch shortly (next few days) regarding this issue.
Have a nice day and best regards,
Roberto Nibali, ratz
--
-------------------------------------------------------------
addr://Kasinostrasse 30, CH-5001 Aarau tel://++41 62 823 9355
http://www.terreactive.com fax://++41 62 823 9356
-------------------------------------------------------------
terreActive AG Wir sichern Ihren Erfolg
-------------------------------------------------------------
|
