Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 6 Jul 2018 14:53:51 -0600
From:      Warner Losh <imp@bsdimp.com>
To:        Tijl Coosemans <tijl@freebsd.org>
Cc:        "Rodney W. Grimes" <freebsd@pdx.rh.cn85.dnsmgr.net>,  "Rodney W. Grimes" <rgrimes@freebsd.org>, Hans Petter Selasky <hselasky@freebsd.org>,  src-committers <src-committers@freebsd.org>, svn-src-all@freebsd.org,  svn-src-head@freebsd.org
Subject:   Re: svn commit: r336025 - in head/sys: amd64/include i386/include
Message-ID:  <CANCZdfrv55qtqX-iFJKVj_oH1oOJKkOsU_C4B8%2BSdcco7X-Oag@mail.gmail.com>
In-Reply-To: <20180706225030.2e689882@kalimero.tijl.coosemans.org>
References:  <CANCZdfrzJK47xroYRHO1aG6Qdos-RZFvN7H7ME4zjX9hhYx-0A@mail.gmail.com> <201807061809.w66I9RVR053596@pdx.rh.CN85.dnsmgr.net> <20180706225030.2e689882@kalimero.tijl.coosemans.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Jul 6, 2018, 3:50 PM T=C4=B3l Coosemans <tijl@freebsd.org> wrote:

> On Fri, 6 Jul 2018 11:09:27 -0700 (PDT) "Rodney W. Grimes" <
> freebsd@pdx.rh.CN85.dnsmgr.net> wrote:
> > > On Fri, Jul 6, 2018, 12:27 PM Rodney W. Grimes <
> > > freebsd@pdx.rh.cn85.dnsmgr.net> wrote:
> > >
> > > > > On Fri, Jul 6, 2018 at 9:52 AM, Rodney W. Grimes <
> > > > > freebsd@pdx.rh.cn85.dnsmgr.net> wrote:
> > > > >
> > > > > > > On Fri, Jul 6, 2018 at 9:32 AM, Rodney W. Grimes <
> > > > > > > freebsd@pdx.rh.cn85.dnsmgr.net> wrote:
> > > > > > >
> > > > > > > > > Author: hselasky
> > > > > > > > > Date: Fri Jul  6 10:13:42 2018
> > > > > > > > > New Revision: 336025
> > > > > > > > > URL: https://svnweb.freebsd.org/changeset/base/336025
> > > > > > > > >
> > > > > > > > > Log:
> > > > > > > > >   Make sure kernel modules built by default are portable
> between
> > > > UP
> > > > > > and
> > > > > > > > >   SMP systems by extending defined(SMP) to include
> > > > > > defined(KLD_MODULE).
> > > > > > > > >
> > > > > > > > >   This is a regression issue after r335873 .
> > > > > > > > >
> > > > > > > > >   Discussed with:             mmacy@
> > > > > > > > >   Sponsored by:               Mellanox Technologies
> > > > > > > >
> > > > > > > > Though this fixes the issue, it also means that now when
> > > > > > > > anyone intentionally builds a UP kernel with modules
> > > > > > > > they are getting SMP support in the modules and I am
> > > > > > > > not sure they would want that.  I know I don't.
> > > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > On UP systems, these additional opcodes are harmless. They
> take a few
> > > > > > extra
> > > > > > > cycles (since they lock an uncontested bus) and add a couple
> extra
> > > > memory
> > > > > > > barriers (which will be NOPs). On MP systems, atomics now wor=
k
> by
> > > > > > default.
> > > > > > > Had we not defaulted like this, all modules built outside of =
a
> kernel
> > > > > > build
> > > > > > > env would have broken atomics. Given that (a) the
> overwhelming
> > > > majority
> > > > > > > (99% or more) is SMP and (b) the MP code merely adds a few
> cycles to
> > > > > > what's
> > > > > > > already a not-too-expensive operation, this was the right
> choice.
> > > > > > >
> > > > > > > It simply doesn't matter for systems that are relevant to the
> project
> > > > > > > today. While one could try to optimize this a little (for
> example, by
> > > > > > > having SMP defined to be 0 or 1, say, and changing all the
> ifdef SMP
> > > > to
> > > > > > if
> > > > > > > (defined(SMP) && SMP !=3D 0)), it's likely not going to matte=
r
> enough
> > > > for
> > > > > > > anybody to make the effort. UP on x86 is simply not relevant
> enough
> > > > to
> > > > > > > optimize for it. Even in VMs, people run SMP kernels typicall=
y
> even
> > > > when
> > > > > > > they just allocate one CPU to the VM.
> > > > > > >
> > > > > > > So while we still support the UP config, and we'll let people
> build
> > > > > > > optimized kernels for x86, we've flipped the switch from
> pessimized
> > > > for
> > > > > > SMP
> > > > > > > modules to pessimized for UP modules, which seems like quite
> the
> > > > > > reasonable
> > > > > > > trade-off.
> > > > > > >
> > > > > > > Were it practical to do so, I'd suggest de-orbiting UP on
> x86.
> > > > However,
> > > > > > > it's a lot of work for not much benefit and we'd need to
> invent much
> > > > > > crazy
> > > > > > > to get there.
> > > > > >
> > > > > > Trivial to fix this with
> > > > > > +#if defined(SMP) || !defined(_KERNEL) || defined(KLD_MODULE) |=
|
> > > > > > !defined(KLD_UP_MODULES)
> > > > >
> > > > >
> > > > > Nope. Not so trivial. Who defines KLD_UP_MODULES?
> > > >
> > > > Call it SMP_KLD_MODULES, and it gets defined the same place SMP doe=
s.
> > > >
> > >
> > > Not so simple. SMP is defined in the config file, and winds up in one
> of
> > No problem, that is where I would be defining this anyway, or in the
> > latest case removing it and SMP for my UP kernel build.
> >
> > > the option files. It will be absent for stand alone builds,
> > I am ok with that.  And it would be reasonable to default to SMP.
> >
> > > though. These
> > > change tweak the default yo be inlined and to include the sequence th=
at
> > > works everywhere.
> > >
> > > >
> > > > > And really, it's absolutely not worth it unless someone shows up
> with
> > > > > numbers to show the old 'function call to optimal routine' is
> actually
> > > > > faster than the new 'inline to slightly unoptimal code'. Since I
> think
> > > > the
> > > > > function call overhead is larger than the pessmizations, I'm not
> sure
> > > > what
> > > > > the fuss is about.
> > > >
> > > > I have no issues with the SMP converting from function calls to
> > > > inline locks, I just want to retain the exact same code I had
> > > > before any of these changes, and that was A UP built system
> > > > without any SMP locking.  Is it too much to ask to keep what
> > > > already worked?
> > > >
> > >
> > > This doesn't enable or disable locks in the muted sense. It just
> changes
> > > the atomic ops for the kernel from a function call to an inlined
> function.
> > > The inlining is more efficient than the call, even with the overhead
> added
> > > by always inlining the same stuff. It still is faster than before.
> > >
> > > And userland has done this forever...
> > >
> > > So I honestly think even UP builds are better off, even if it's not
> hyper
> > > optimized for UP. The lock instruction prefix is minimal overhead (a
> cycle
> > > I think).
> >
> > I do not believe, and Bruce seems to have evidence, that LOCK is not
> > a one cycle cost.  And in my head I know that it can not be that
> > simple as it causes lots of very special things to happen in the
> > pipeline to ensure you are locked.
> >
> > > This is different than the mutexes we optimize for the UP cases
> > > (and which aren't affected by this change). It's really not a big
> deal.
> >
> > CPU's are not getting any faster, cycles are cycles, and I think we
> > should at least investigate further before we just start making
> > assumptions about the lock prefix being a 1 cycle cheap thing to
> > do.
>
>
> Just install opt_*.h headers already.  It's not just about the SMP option=
.
> The nvidia-driver ports want to know if PAE is enabled on i386.
>

Sadly, I don't think it will be that simple...

Warner

>



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CANCZdfrv55qtqX-iFJKVj_oH1oOJKkOsU_C4B8%2BSdcco7X-Oag>