Re: [PATCH 2/3] perf/amlogic: Fix large number of counter issue

From: Will Deacon <will@kernel.org>
To: Jiucheng Xu <jiuchengxu@163.com>
Cc: Mark Rutland <mark.rutland@arm.com>,
	Neil Armstrong <neil.armstrong@linaro.org>,
	Kevin Hilman <khilman@baylibre.com>,
	Jerome Brunet <jbrunet@baylibre.com>,
	Martin Blumenstingl <martin.blumenstingl@googlemail.com>,
	Jianxin Pan <jianxin.pan@amlogic.com>,
	Kelvin Zhang <kelvin.zhang@amlogic.com>,
	Chris Healy <cphealy@gmail.com>, Chris Healy <healych@amazon.com>,
	linux-amlogic@lists.infradead.org,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 2/3] perf/amlogic: Fix large number of counter issue
Date: Tue, 28 Mar 2023 12:55:43 +0100	[thread overview]
Message-ID: <20230328115543.GA1159@willie-the-truck> (raw)
In-Reply-To: <79968005-28f3-d890-dfea-faca2e7d7947@163.com>

On Tue, Mar 28, 2023 at 10:29:04AM +0800, Jiucheng Xu wrote:
> 
> My Amlogic email box has some issues. Use my personal email
> <jiucheng.xu@163.com> to reply.
> 
> On 2023/3/27 22:10, Will Deacon wrote:
> > [ EXTERNAL EMAIL ]
> > 
> > On Thu, Feb 09, 2023 at 07:54:02PM +0800, Jiucheng Xu wrote:
> > > When use 1ms interval, very large number of counter happens
> > > once in a while as below:
> > > 
> > > 25.968654513 281474976710655.84 MB meson_ddr_bw/chan_1_rw_bytes,arm=1/
> > > 26.118657346 281474976710655.88 MB meson_ddr_bw/chan_1_rw_bytes,arm=1/
> > > 26.180137180 281474976710655.66 MB meson_ddr_bw/chan_1_rw_bytes,arm=1/
> > > 
> > > Root cause is the race between irq handler
> > > and pmu.read callback. Use spin lock to protect the sw&hw
> > > counters.
> > > 
> > > Signed-off-by: Jiucheng Xu <jiucheng.xu@amlogic.com>
> > > ---
> > >   drivers/perf/amlogic/meson_ddr_pmu_core.c | 10 +++++++++-
> > >   1 file changed, 9 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/perf/amlogic/meson_ddr_pmu_core.c b/drivers/perf/amlogic/meson_ddr_pmu_core.c
> > > index 0b24dee1ed3c..9b2e5d5c0626 100644
> > > --- a/drivers/perf/amlogic/meson_ddr_pmu_core.c
> > > +++ b/drivers/perf/amlogic/meson_ddr_pmu_core.c
> > > @@ -14,6 +14,7 @@
> > >   #include <linux/perf_event.h>
> > >   #include <linux/platform_device.h>
> > >   #include <linux/printk.h>
> > > +#include <linux/spinlock.h>
> > >   #include <linux/sysfs.h>
> > >   #include <linux/types.h>
> > > @@ -23,6 +24,7 @@ struct ddr_pmu {
> > >   	struct pmu pmu;
> > >   	struct dmc_info info;
> > >   	struct dmc_counter counters;	/* save counters from hw */
> > > +	spinlock_t lock;		/* protect hw/sw counter */
> > >   	bool pmu_enabled;
> > >   	struct device *dev;
> > >   	char *name;
> > > @@ -92,10 +94,12 @@ static void meson_ddr_perf_event_update(struct perf_event *event)
> > >   	int idx;
> > >   	int chann_nr = pmu->info.hw_info->chann_nr;
> > > +	spin_lock(&pmu->lock);
> > Why doesn't this need the _irqsave() variant if we're racing with the irq
> > handler?
> > 
> > Will
> I think meson_ddr_perf_event_update function is called with hard irq off.
> So update function couldn't be interrupted by irq handler. Right?

I'm just confused about the race, then. The commit message says you have a
race between an irq handler and a callback, which you fix with a spinlock
that isn't irq safe. So either the race is real and the lock needs to be
irqsafe, or the race is something else entirely, no?

Will

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel