Date: Tue, 19 Jul 2016 16:57:57 +0200 From: Randall Stewart <rrs@netflix.com> To: Glen Barber <gjb@freebsd.org> Cc: Gleb Smirnoff <glebius@freebsd.org>, src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r302998 - head/sys/kern Message-ID: <F3837B91-B5C5-4475-9DCC-411643164C09@netflix.com> In-Reply-To: <20160719135614.GE65494@FreeBSD.org> References: <201607180929.u6I9T9Uw063705@repo.freebsd.org> <64C1543A-3EDE-4852-88EA-5B0B78FCF016@netflix.com> <58868615-8255-4D8B-BD9E-8E19A734CB6C@netflix.com> <5E551FA8-C779-45AE-B038-D8B51B53EEAA@netflix.com> <20160719122545.GC65494@FreeBSD.org> <03E667D5-A4C9-47B9-A8DD-49ECD0E2B7EB@netflix.com> <20160719135614.GE65494@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Well=20 The code itself I had up on machines for probably about 2 months. But = then I switched over to Gleb=E2=80=99s changes here just recently .. which = caused me all kinds of fun :) I had to go back into Mercurial to pull back my changes.. I have had the resurrected changes running on my netflix machines for about 20 or so hours generating about anywhere from 14Gbps to 32Gbps depending on the = machine type. I plan on waiting until tomorrow to sync it down into the NF code base.=20= Note that if you do decide instead to roll back to the 10.x = kern_timeout.c you will need to roll back a bunch of tcp changes as well that use the new async_drain() interface. I am game either way for you to proceed.. I will commit this current = code to head as long as I hear no objections (from Gleb or others)=E2=80=A6. R > On Jul 19, 2016, at 3:56 PM, Glen Barber <gjb@freebsd.org> wrote: >=20 > On Tue, Jul 19, 2016 at 03:46:54PM +0200, Randall Stewart wrote: >> Glen: >>=20 >> My changes work.. I have them running in NF in at least 1/2 dozen = machines. >>=20 >=20 > For how long? What are the uptimes on these machines? >=20 > This is the blocker for 11.0-BETA2, and I don't want to see more > regressions being introduced at this point of the cycle. >=20 > Glen >=20 >> I am more than willing to commit them.. they actually are not much = different than >> whats in stable 10.. though I don=E2=80=99t know if the async-drain = was MFC=E2=80=99d there.. it >> needs to be in for TCP.. or else you will have yet another mess in = that >> respect (TCP depends on ASYNC-drain). >>=20 >> I can commit what I have.. if you like.. or not.. I really don=E2=80=99= t care (I hate kern_timeout.c :-o) >>=20 >> R >>> On Jul 19, 2016, at 2:25 PM, Glen Barber <gjb@freebsd.org> wrote: >>>=20 >>> On Tue, Jul 19, 2016 at 01:43:16PM +0200, Randall Stewart wrote: >>>> Gleb >>>>=20 >>>> Ok >>>>=20 >>>> I have now updated >>>>=20 >>>> https://reviews.freebsd.org/D7135 >>>>=20 >>>> You can take this or not=E2=80=A6 I really don=E2=80=99t care = either way=E2=80=A6 (you are welcome to >>>> own the kern_timeout.c code I hate it) :-) >>>>=20 >>>> Basically when you went off and re-factored kern_timeout.c I had = worked in parallel on fixing >>>> the bugs you were seeing.. There were three distinct problems that = I fixed=E2=80=A6 but then >>>> you had refactored the stop() routine.. and I thought ok.. thats = fine. I had actually thought about >>>> doing something similar to what you did and was too chicken to poke = that much at it.. it has >>>> always had a nasty habit of biting back when you make a lot of = changes :-D >>>>=20 >>>> I know my version has worked for quite some time in my testing so I = brought it back. >>>> Complete with its 3 return codes (I only recently switched to your = version and thus >>>> started having difficulties with leaks and crashes)=E2=80=A6. >>>>=20 >>>> You are welcome not to use this.. I know it works (it ran >>>> on a number of machines at NF last night.. and we will of course = continue testing >>>> it as we finish our dev testing for the upcoming OCA software = release).. For now >>>> this is what will be going out into the OCA=E2=80=99s at least :-) >>>>=20 >>>=20 >>> I'm honestly done with this topic, and at the point now where I'm >>> considering backing out all changes to callout(9) and related = changes to >>> the state they were at in stable/10. >>>=20 >>> This changes the KBI, and if it needs to be done, it needs to happen >>> now. We cannot wait for RC1 phase for this, and the amount of churn = to >>> get things into a working state with the current implementation far >>> outweighs the benefit of the dangers. >>>=20 >>> Glen >>>=20 >>=20 >> -------- >> Randall Stewart >> rrs@netflix.com >> 803-317-4952 >>=20 >>=20 >>=20 >>=20 >>=20 -------- Randall Stewart rrs@netflix.com 803-317-4952
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?F3837B91-B5C5-4475-9DCC-411643164C09>