LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
To: Josh Triplett <josh@kernel.org>
Cc: linux-kernel@vger.kernel.org, mingo@elte.hu, tglx@linuxtronix.de,
dipankar@in.ibm.com, tytso@us.ibm.com, dvhltc@us.ibm.com,
oleg@tv-sign.ru, twoerner.k@gmail.com, billh@gnuppy.monkey.org,
nielsen.esben@googlemail.com, corbet@lwn.net
Subject: Re: [RFC PATCH -rt 2/2] RCU priority boosting additions to rcutorture
Date: Thu, 25 Jan 2007 17:52:12 -0800 [thread overview]
Message-ID: <20070126015212.GI1705@linux.vnet.ibm.com> (raw)
In-Reply-To: <45B8FFBB.4090307@kernel.org>
On Thu, Jan 25, 2007 at 11:06:35AM -0800, Josh Triplett wrote:
> Paul E. McKenney wrote:
> > On Thu, Jan 25, 2007 at 12:47:04AM -0800, Josh Triplett wrote:
> >> One major item: this new test feature really needs a new module parameter to
> >> enable or disable it.
> >
> > CONFIG_PREEMPT_RCU_BOOST is the parameter -- if not set, then no test.
> > This parameter is provided by the accompanying RCU-boost patch.
>
> It seems useful for rcutorture to use or not use the preempting thread
> independently of CONFIG_PREEMPT_RCU_BOOST. That would bring you from two
> cases to four, and the two new cases both make sense:
>
> * CONFIG_PREEMPT_RCU_BOOST=n, but run rcutorture with the preempting thread.
> This configuration allows you to demonstrate the need for
> CONFIG_PREEMPT_RCU_BOOST, by showing what happens when you need it and don't
> have it.
>
> * CONFIG_PREEMPT_RCU_BOOST=y, but run rcutorture without the preempting
> thread. This configuration allows you to test with rcutorture while running
> a *real* real-time workload rather than the simple preempting thread, or
> just test basic RCU functionality.
>
> A simple boolean module_param would work here.
OK, sold! I will add this. Perhaps CONFIG_PREEMPT_RCU_TORTURE.
> At some point, we may want to add the ability to run multiple preempting
> threads, but that doesn't need to happen for this patch.
I considered that for this initial round, but you only need to preempt
a single RCU reader to force the RCU booster to do something. ;-)
> >> Paul E. McKenney wrote:
> >>> diff -urpNa -X dontdiff linux-2.6.20-rc4-rt1/kernel/rcutorture.c linux-2.6.20-rc4-rt1-rcubtorture/kernel/rcutorture.c
> >>> --- linux-2.6.20-rc4-rt1/kernel/rcutorture.c 2007-01-09 10:59:54.000000000 -0800
> >>> +++ linux-2.6.20-rc4-rt1-rcubtorture/kernel/rcutorture.c 2007-01-23 11:27:49.000000000 -0800
>
> >>> +static int rcu_torture_preempt(void *arg)
> >>> +{
> >>> + int completedstart;
> >>> + time_t gcstart;
> >>> + struct sched_param sp;
> >>> +
> >>> + sp.sched_priority = MAX_RT_PRIO - 1;
> >>> + sched_setscheduler(current, SCHED_RR, &sp);
> >>> + current->flags |= PF_NOFREEZE;
> >>> +
> >>> + do {
> >>> + completedstart = rcu_torture_completed();
> >>> + gcstart = xtime.tv_sec;
> >>> + while ((xtime.tv_sec - gcstart < 10) &&
> >>> + (rcu_torture_completed() == completedstart))
> >>> + cond_resched();
> >>> + if (rcu_torture_completed() == completedstart)
> >>> + rcu_torture_preempt_errors++;
> >>> + schedule_timeout_interruptible(shuffle_interval * HZ);
> >> Why call schedule_timeout_interruptible here without actually handling
> >> interruptions? So that you can send it a signal to cause the shuffle early?
> >
> > It allows you to kill the process in order to get the module unload to
> > happen more quickly in case someone specified an overly long interval.
>
> I didn't actually know that you could kill a kthread from userspace. :)
>
> That rationale makes sense.
It won't actually die, but if I understand correctly (a big "if") the
signal would cause schedule_timeout_interruptible() to return, allowing
the kthread_should_stop() check to happen.
> > But now that you mention this, a simple one-second sleep is probably
> > appropriate here.
>
> OK.
>
> >>> + } while (!kthread_should_stop());
> >>> + return NULL;
> >>> +}
> >>> +
> >>> +static void rcu_preempt_start(void)
> >>> +{
> >>> + rcu_preeempt_task = kthread_run(rcu_torture_preempt, NULL,
> >>> + "rcu_torture_preempt");
> >>> + if (IS_ERR(rcu_preeempt_task)) {
> >>> + VERBOSE_PRINTK_ERRSTRING("Failed to create preempter");
> >> This ought to include the errno value, PTR_ERR(rcu_preempt_task).
> >
> > Good point -- what I should do is return this value so that
> > rcu_torture_init() can return it, failing the module-load process
> > and unwinding.
>
> Even better, yes.
>
> >>> + rcu_preeempt_task = NULL;
> >>> + }
> >>> +}
> >>> +
> >>> +static void rcu_preempt_end(void)
> >>> +{
> >>> + if (rcu_preeempt_task != NULL) {
> >> if (rcu_preempt_task) would work just as well here.
> >
> > True, but was being consistent with usage elsewhere in this file.
>
> Fair enough; don't worry about it for this patch, then. I'll deal with that
> particular style cleanup later, throughout rcutorture.
Sounds good to me! ;-)
> >>> static struct rcu_torture_ops rcu_ops = {
> >>> .init = NULL,
> >>> .cleanup = NULL,
> >>> @@ -267,7 +334,9 @@ static struct rcu_torture_ops rcu_ops =
> >>> .completed = rcu_torture_completed,
> >>> .deferredfree = rcu_torture_deferred_free,
> >>> .sync = synchronize_rcu,
> >>> - .stats = NULL,
> >>> + .preemptstart = rcu_preempt_start,
> >>> + .preemptend = rcu_preempt_end,
> >>> + .stats = rcu_preempt_stats,
> >>> .name = "rcu"
> >>> };
> >>>
> >>> @@ -306,6 +375,8 @@ static struct rcu_torture_ops rcu_sync_o
> >>> .completed = rcu_torture_completed,
> >>> .deferredfree = rcu_sync_torture_deferred_free,
> >>> .sync = synchronize_rcu,
> >>> + .preemptstart = NULL,
> >>> + .preemptend = NULL,
> >>> .stats = NULL,
> >>> .name = "rcu_sync"
> >>> };
> >> Much like other common structures such as struct file_operations, no need to
> >> explicitly specify members as NULL here; any member you don't specify will get
> >> a NULL value. That avoids the need to update every use of this structure
> >> whenever you add a new member used by only some of them.
> >
> > Untrusting, aren't I? ;-)
>
> Heh. I have that problem as well; I always hestitate to trust the compiler to
> initialize values.
>
> > I removed all the "= NULL" entries.
>
> Thanks.
>
> >>> @@ -856,6 +935,8 @@ rcu_torture_cleanup(void)
> >>> kthread_stop(stats_task);
> >>> }
> >>> stats_task = NULL;
> >>> + if (cur_ops->preemptend != NULL)
> >> if (cur_ops->preemptend) would work as well.
> >
> > True, though again there is a lot of existing "!= NULL" in this file
> > and elsewhere. Many thousands of them through the kernel. ;-)
>
> As before, don't worry about it for this patch then.
>
> > I will run this through the mill and repost.
But with the new kernel parameter this time. ;-)
Thanx, Paul
next prev parent reply other threads:[~2007-01-26 1:52 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2007-01-25 2:11 [RFC PATCH -rt 0/2] RCU priority boosting that survives semi-vicious testing Paul E. McKenney
2007-01-25 2:14 ` [RFC PATCH -rt 1/2] " Paul E. McKenney
2007-01-25 2:23 ` [RFC PATCH -rt 2/2] RCU priority boosting additions to rcutorture Paul E. McKenney
2007-01-25 8:47 ` Josh Triplett
2007-01-25 18:01 ` Paul E. McKenney
2007-01-25 19:06 ` Josh Triplett
2007-01-26 1:52 ` Paul E. McKenney [this message]
2007-01-26 6:29 ` Josh Triplett
2007-01-29 2:11 ` Paul E. McKenney
2007-01-29 6:05 ` Josh Triplett
2007-01-25 9:29 ` [RFC PATCH -rt 1/2] RCU priority boosting that survives semi-vicious testing Josh Triplett
2007-01-25 19:58 ` Paul E. McKenney
2007-01-26 1:47 ` Paul E. McKenney
2007-01-26 6:13 ` Josh Triplett
2007-02-01 1:21 [RFC PATCH -rt 0/2] RCU priority boosting that survives vicious testing Paul E. McKenney
2007-02-01 1:26 ` [RFC PATCH -rt 2/2] RCU priority boosting additions to rcutorture Paul E. McKenney
2007-02-01 2:12 ` Nigel Cunningham
2007-02-01 2:31 ` Paul E. McKenney
2007-02-01 2:42 ` Nigel Cunningham
2007-02-01 5:46 ` Paul E. McKenney
2007-02-01 22:13 ` Nigel Cunningham
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=20070126015212.GI1705@linux.vnet.ibm.com \
--to=paulmck@linux.vnet.ibm.com \
--cc=billh@gnuppy.monkey.org \
--cc=corbet@lwn.net \
--cc=dipankar@in.ibm.com \
--cc=dvhltc@us.ibm.com \
--cc=josh@kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@elte.hu \
--cc=nielsen.esben@googlemail.com \
--cc=oleg@tv-sign.ru \
--cc=tglx@linuxtronix.de \
--cc=twoerner.k@gmail.com \
--cc=tytso@us.ibm.com \
--subject='Re: [RFC PATCH -rt 2/2] RCU priority boosting additions to rcutorture' \
/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
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).