Date: Mon, 15 Feb 2010 11:33:41 +1100
From: Paul Mackerras <paulus@samba.org>
To: Don Zickus <dzickus@redhat.com>
Cc: linux-kernel@vger.kernel.org, mingo@elte.hu, peterz@infradead.org,
       gorcunov@gmail.com, aris@redhat.com
Subject: Re: [PATCH 3/4] nmi_watchdog: fallback to software events when no
 hardware pmu detected
Message-ID: <20100215003341.GH13769@brick.ozlabs.ibm.com>
References: <1266013161-31197-1-git-send-email-dzickus@redhat.com>
 <1266013161-31197-3-git-send-email-dzickus@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1266013161-31197-3-git-send-email-dzickus@redhat.com>
User-Agent: Mutt/1.5.20 (2009-06-14)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1660
Lines: 34

On Fri, Feb 12, 2010 at 05:19:20PM -0500, Don Zickus wrote:

> diff --git a/kernel/nmi_watchdog.c b/kernel/nmi_watchdog.c
> index 73c1954..4f23505 100644
> --- a/kernel/nmi_watchdog.c
> +++ b/kernel/nmi_watchdog.c
> @@ -166,8 +166,12 @@ cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
>  		wd_attr.sample_period = hw_nmi_get_sample_period();
>  		event = perf_event_create_kernel_counter(&wd_attr, hotcpu, -1, wd_overflow);
>  		if (IS_ERR(event)) {
> -			printk(KERN_ERR "nmi watchdog failed to create perf event on %i: %p\n", hotcpu, event);
> -			return NOTIFY_BAD;
> +			wd_attr.type = PERF_TYPE_SOFTWARE;
> +			event = perf_event_create_kernel_counter(&wd_attr, hotcpu, -1, wd_overflow);

Here you don't explicitly set wd_attr.config or wd_attr.sample_period
for the software event.  So PERF_COUNT_HW_CPU_CYCLES (which is 0)
becomes PERF_COUNT_SW_CPU_CLOCK (also 0).  Which is either a happy
accident or really really subtle. :)  I suggest you either set
wd_attr.config or add a comment pointing out that you are using the
fact that PERF_COUNT_HW_CPU_CYCLES == PERF_COUNT_SW_CPU_CLOCK.

Also you don't explicitly set wd_attr.sample_period, so it will be
whatever hw_nmi_get_sample_period() returned, except now measured in
nanoseconds rather than (presumably) cpu clock cycles.  Are you aiming
for 1 interrupt per second?  If so you should set .sample_period to
NSEC_PER_SEC.

Paul.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/