From: "Paul E. McKenney" <paulmck@us.ibm.com>
To: Ingo Molnar <mingo@elte.hu>
Cc: dipankar@in.ibm.com, shemminger@osdl.org, akpm@osdl.org,
torvalds@osdl.org, rusty@au1.ibm.com, tgall@us.ibm.com,
jim.houston@comcast.net, manfred@colorfullife.com, gh@us.ibm.com,
linux-kernel@vger.kernel.org
Subject: Re: Real-Time Preemption and RCU
Date: Fri, 18 Mar 2005 08:43:51 -0800 [thread overview]
Message-ID: <20050318164351.GE1299@us.ibm.com> (raw)
In-Reply-To: <20050318074903.GA27656@elte.hu>
On Fri, Mar 18, 2005 at 08:49:03AM +0100, Ingo Molnar wrote:
>
> * Paul E. McKenney <paulmck@us.ibm.com> wrote:
>
> > Seems to me that it would be good to have an RCU implementation that
> > meet the requirements of the Real-Time Preemption patch, but that is
> > 100% compatible with the "classic RCU" API. Such an implementation
> > must meet a number of requirements, which are listed at the end of
> > this message (search for "REQUIREMENTS").
>
> [ Wow. you must be a secret telepatic mind-reader - yesterday i was
> thinking about mailing you, because my approach to RCU preemptability
> (the API variants) clearly sucked and caused problems all around, both
> in terms of maintainability and in terms of stability and
> scalability. ]
Wish I could claim that, but must credit dumb luck. ;-)
> > 5. The final version, which both scales and meets realtime
> > requirements, as well as exactly matching the "classic RCU"
> > API.
> >
> > I have tested this approach, but in user-level scaffolding. All of
> > these implementations should therefore be regarded with great
> > suspicion: untested, probably don't even compile. Besides which, I
> > certainly can't claim to fully understand the real-time preempt patch,
> > so I am bound to have gotten something wrong somewhere. In any case,
> > none of these implementations are a suitable replacement for "classic
> > RCU" on large servers, since they acquire locks in the RCU read-side
> > critical sections. However, they should scale enough to support small
> > SMP systems, inflicting only a modest performance penalty.
>
> basically for PREEMPT_RT the only constraint is that RCU sections should
> be preemptable. Whatever the performance cost. If PREEMPT_RT is merged
> into the upstream kernel then it will (at least initially) be at a
> status similar to NOMMU: it will be tolerated as long as it causes no
> 'drag' on the main code. The RCU API variants i introduced clearly
> violated this requirement, and were my #1 worry wrt. upstream
> mergability.
Fair enough!
> > I believe that implementation #5 is most appropriate for real-time
> > preempt kernels. [...]
>
> yeah, agreed - it looks perfect - both the read and write side is
> preemptable. Can i just plug the code you sent into rcupudate.c and
> expect it to work, or would you like to send a patch? If you prefer you
> can make it an unconditional patch against an upstream kernel to keep
> things simple for you - i'll then massage it to be properly PREEMPT_RT
> dependent.
I will give it a shot, but feel free to grab anything at any point,
starting either from the patch I will send in a bit or from my earlier
email. Do whatever works best for you.
For the patch, here are my questions:
o What is the best way to select between classic RCU and this
scheme?
1. Massive #ifdef across rcupdate.c
2. Create an rcupdate_rt.c and browbeat the build system
into picking one or the other (no clue if this is
possible...)
3. Create an rcupdate_rt.c and rely on the linker to pick
one or the other, with rcupdate.h generating different
external symbol names to make the choice.
Left to myself, I would grit my teeth and do #1, since I
understand how to do it. Flames welcome, as long as accompanied
by good advice!
o How best to interface to OOM? Left to myself, I leave this
for later. ;-)
I will take the cowardly approach of patching against the upstream kernel.
> > [...] In theory, #3 might be appropriate, but if I understand the
> > real-time preempt implementation of reader-writer lock, it will not
> > perform well if there are long RCU read-side critical sections, even
> > in UP kernels.
>
> all RCU-locked sections must be preemptable in -RT. Basically RCU is a
> mainstream API that is used by lots of code and will be introduced in
> many other areas as well. From the -RT kernel's POV sees this as an
> 'uncontrollable latency source', which keeps introducing critical
> sections. One major goal of PREEMPT_RT is to convert all popular
> critical section APIs into preemptible sections, so that the amount of
> code that is non-preemptable is drastically reduced and can be managed
> (and thus can be trusted). This goal has a higher priority than any
> performance consideration, because it doesnt matter what performance you
> have, if you cannot trust the kernel to be deterministic.
Makes sense to me!
Thanx, Paul
next prev parent reply other threads:[~2005-03-18 16:44 UTC|newest]
Thread overview: 58+ messages / expand[flat|nested] mbox.gz Atom feed top
2005-03-18 0:20 Real-Time Preemption and RCU Paul E. McKenney
2005-03-18 7:49 ` Ingo Molnar
2005-03-18 16:43 ` Paul E. McKenney [this message]
2005-03-18 17:11 ` Ingo Molnar
2005-03-18 17:29 ` Paul E. McKenney
2005-03-18 20:35 ` Paul E. McKenney
2005-03-18 22:22 ` Paul E. McKenney
2005-03-19 0:48 ` Paul E. McKenney
2005-03-18 8:44 ` Ingo Molnar
2005-03-18 9:04 ` Ingo Molnar
2005-03-18 9:38 ` Ingo Molnar
2005-03-18 9:13 ` Ingo Molnar
2005-03-18 9:28 ` Ingo Molnar
2005-03-18 9:53 ` Ingo Molnar
2005-03-18 15:33 ` Paul E. McKenney
2005-03-19 5:03 ` Manfred Spraul
2005-03-19 16:26 ` Ingo Molnar
2005-03-20 6:36 ` Manfred Spraul
2005-03-20 9:25 ` Thomas Gleixner
2005-03-20 16:57 ` Manfred Spraul
2005-03-20 21:38 ` Bill Huey
2005-03-20 21:59 ` Bill Huey
2005-03-18 10:03 ` Ingo Molnar
2005-03-18 11:30 ` Ingo Molnar
2005-03-18 16:48 ` Esben Nielsen
2005-03-18 17:19 ` Ingo Molnar
2005-03-20 13:29 ` Esben Nielsen
2005-03-20 22:38 ` Paul E. McKenney
2005-03-20 23:23 ` Esben Nielsen
2005-03-22 5:53 ` Paul E. McKenney
2005-03-22 8:55 ` Esben Nielsen
2005-03-22 9:20 ` Ingo Molnar
2005-03-22 10:19 ` Esben Nielsen
2005-03-23 5:40 ` Paul E. McKenney
2005-03-23 11:44 ` Esben Nielsen
2005-03-24 7:02 ` Paul E. McKenney
2005-03-22 10:56 ` Ingo Molnar
2005-03-22 11:39 ` Esben Nielsen
2005-03-22 13:10 ` Ingo Molnar
2005-03-22 15:08 ` Esben Nielsen
2005-03-18 15:48 ` Paul E. McKenney
2005-03-18 11:38 ` Ingo Molnar
2005-03-18 12:56 ` Bill Huey
2005-03-18 13:17 ` Bill Huey
2005-03-18 15:57 ` Paul E. McKenney
2005-03-18 16:02 ` Ingo Molnar
2005-03-18 16:55 ` Esben Nielsen
2005-03-22 10:04 ` Bill Huey
2005-03-22 10:17 ` Bill Huey
2005-03-22 10:34 ` Bill Huey
2005-03-22 10:38 ` Esben Nielsen
2005-03-18 22:26 ` Herbert Xu
2005-03-19 16:31 ` Ingo Molnar
2005-03-20 8:01 ` Kyle Moffett
2005-03-22 8:08 ` Ingo Molnar
2005-03-18 15:54 ` Paul E. McKenney
2005-03-18 15:58 ` Ingo Molnar
2009-06-11 22:57 real-time preemption " James Huang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20050318164351.GE1299@us.ibm.com \
--to=paulmck@us.ibm.com \
--cc=akpm@osdl.org \
--cc=dipankar@in.ibm.com \
--cc=gh@us.ibm.com \
--cc=jim.houston@comcast.net \
--cc=linux-kernel@vger.kernel.org \
--cc=manfred@colorfullife.com \
--cc=mingo@elte.hu \
--cc=rusty@au1.ibm.com \
--cc=shemminger@osdl.org \
--cc=tgall@us.ibm.com \
--cc=torvalds@osdl.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).