Date: Mon, 9 Jun 2014 13:43:39 -0400 From: John Baldwin <jhb@freebsd.org> To: "Alexander V. Chernikov" <melifaro@freebsd.org> Cc: Konstantin Belousov <kostikbel@gmail.com>, freebsd-hackers@freebsd.org, hackers@freebsd.org Subject: Re: Permit init(8) use its own cpuset group. Message-ID: <201406091343.39584.jhb@freebsd.org> In-Reply-To: <53923C09.6020302@FreeBSD.org> References: <538C8F9A.4020301@FreeBSD.org> <5390DB64.6010704@FreeBSD.org> <53923C09.6020302@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Friday, June 06, 2014 6:09:13 pm Alexander V. Chernikov wrote:
> On 06.06.2014 01:04, Alexander V. Chernikov wrote:
> > On 05.06.2014 23:59, John Baldwin wrote:
> >> On Thursday, June 05, 2014 3:46:15 pm Alexander V. Chernikov wrote:
> >>> On 05.06.2014 18:09, John Baldwin wrote:
> >>>> On Wednesday, June 04, 2014 3:16:59 pm Alexander V. Chernikov wrote:
> >>>>> On 04.06.2014 19:06, John Baldwin wrote:
> >>>>>> On Monday, June 02, 2014 12:48:50 pm Konstantin Belousov wrote:
> >>>>>>> On Mon, Jun 02, 2014 at 06:52:10PM +0400, Alexander V. Chernikov wrote:
> >>>>>>>> Hello list!
> >>>>>>>>
> >>>>>>>> Currently init(8) uses group 1 which is root group.
> >>>>>>>> Modifications of this group affects both kernel and userland threads.
> >>>>>>>> Additionally, such modifications are impossible, for example, in
> >>>> presence
> >>>>>>>> of multi-queue NIC drivers (like igb or ixgbe) which binds their
> >> threads
> >>>>>> to
> >>>>>>>> particular cpus.
> >>>>>>>>
> >>>>>>>> Proposed change ("init_cpuset" loader tunable) permits changing cpu
> >>>>>>>> masks for
> >>>>>>>> userland more easily. Restricting user processes to migrate to/from
> >> CPU
> >>>>>>>> cores
> >>>>>>>> used for network traffic processing is one of the cases.
> >>>>>>>>
> >>>>>>>> Phabricator: https://phabric.freebsd.org/D141 (the same version
> >> attached
> >>>>>>>> inline)
> >>>>>>>>
> >>>>>>>> If there are no objections, I'll commit this next week.
> >>>>>>> Why is the tunable needed ?
> >>>>>> Because some people already depend on doing 'cpuset -l 0 -s 1'. It is
> >>>> also
> >>>>>> documented in our manpages that processes start in cpuset 1 by default
> >> so
> >>>>>> that you can use 'cpuset -l 0 -s 1' to move all processes, etc.
> >>>>>>
> >>>>>> For the stated problem (bound ithreads in drivers), I would actually
> >> like
> >>>> to
> >>>>>> fix ithreads that are bound to a specific CPU to create a different
> >> cpuset
> >>>>>> instead so they don't conflict with set 1.
> >>>>> Yes, this seem to be much better approach.
> >>>>> Please take a look on the new patch (here or in the phabricator).
> >>>>> Comments:
> >>>>>
> >>>>> Use different approach for modifyable root set:
> >>>>>
> >>>>> * Make sets in 0..15 internal
> >>>>> * Define CPUSET_SET1 & CPUSET_ITHREAD in that range
> >>>>> * Add cpuset_lookup_builtin() to retrieve such cpu sets by id
> >>>>> * Create additional root set for ithreads
> >>>>> * Use this set in ithread_create()
> >>>>> * Change intr_setaffinity() to use cpuset_iroot (do we really need
> >> this)?
> >>>>>
> >>>>> We can probably do the same for kprocs, but I'm unsure if we really need
> >> it.
> >>>>
> >>>> I imagined something a bit simpler. Just create a new set in
> >> intr_event_bind
> >>>> and bind the ithread to the new set. No need to have more magic set ids,
> >> etc.
> >>> Well, we also have userland which can modify given changes via `cpuset
> >>> -x', so we need to be able to add some more logic on set
> >>> allocation/keeping. Additionally, we can try to do the same via `cpuset
> >>> -t', so introducing something like cpuset_setIthread() and hooking into
> >>> intr_event_bind() won't probably be enough. At least I can't think out a
> >>> quick and easy way to do this.
> >>
> >> cpuset -x calls intr_event_bind(). If you just do it there you fix both
> >> places.
> Yup. I was wrong. This version seems to be simpler while doing the right
> thing.
Hmm, I do think this patch is doing the right thing, but I'm worried you
might have weird effects, e.g. I worry that because the 'intr' proc is in
set 1 still, the thread masks will be (incorrectly) checked when changing
set 1 and you still won't be able to 'cpuset -l 0 -s 1'.
Also, strictly speaking, it would probably be best to return threads to
set 1 when NOCPU is passed to intr_event_bind() to unbind an interrupt.
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201406091343.39584.jhb>