Re: [PATCH v2 1/9] perf: Add EVENT_ATTR_ID to simplify event attributes

From: Will Deacon <will@kernel.org>
To: "liuqi (BA)" <liuqi115@huawei.com>
Cc: linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Mark Rutland <mark.rutland@arm.com>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Linuxarm <linuxarm@huawei.com>
Subject: Re: [PATCH v2 1/9] perf: Add EVENT_ATTR_ID to simplify event attributes
Date: Wed, 2 Jun 2021 10:49:22 +0100	[thread overview]
Message-ID: <20210602094922.GA30503@willie-the-truck> (raw)
In-Reply-To: <30abdbec-3174-1f8a-47d4-63a4de3b1e47@huawei.com>

On Wed, Jun 02, 2021 at 04:45:23PM +0800, liuqi (BA) wrote:
> 
> Hi Will,
> 
> Thanks for reviewing this patch.
> 
> On 2021/6/1 21:10, Will Deacon wrote:
> > On Wed, May 19, 2021 at 05:51:51PM +0800, Qi Liu wrote:
> > > Similar EVENT_ATTR macros are defined in many PMU drivers,
> > > like HiSilicon PMU driver, Arm PMU driver, Arm SMMU PMU
> > > driver. So Add a generic macro to simplify code.
> > > 
> > > Cc: Peter Zijlstra <peterz@infradead.org>
> > > Cc: Ingo Molnar <mingo@redhat.com>
> > > Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> > > Cc: Mark Rutland <mark.rutland@arm.com>
> > > Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> > > Signed-off-by: Qi Liu <liuqi115@huawei.com>
> > > ---
> > >   include/linux/perf_event.h | 6 ++++++
> > >   kernel/events/core.c       | 2 ++
> > >   2 files changed, 8 insertions(+)
> > > 
> > > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > > index f5a6a2f..d0aa74e 100644
> > > --- a/include/linux/perf_event.h
> > > +++ b/include/linux/perf_event.h
> > > @@ -1576,6 +1576,12 @@ static struct perf_pmu_events_attr _var = {				    \
> > >   	.event_str	= _str,						    \
> > >   };
> > > +#define PMU_EVENT_ATTR_ID(_name, _id)					     \
> > > +	(&((struct perf_pmu_events_attr[]) {				     \
> > > +		{ .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL),  \
> > > +		  .id = _id, }						     \
> > > +	})[0].attr.attr)
> > > +
> > >   #define PMU_FORMAT_ATTR(_name, _format)					\
> > >   static ssize_t								\
> > >   _name##_show(struct device *dev,					\
> > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > index 0ac818b..330d9cc 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -13295,6 +13295,8 @@ ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
> > >   	if (pmu_attr->event_str)
> > >   		return sprintf(page, "%s\n", pmu_attr->event_str);
> > > +	else
> > > +		return sprintf(page, "config=%#llx\n", pmu_attr->id);
> > 
> > I think it's a really bad idea to hardcode this here. For example, I think
> > this patch series breaks user ABI for the SMMU PMU which used to print:
> > 
> > 	"event=0x%02llx\n"
> > 
> > and by the looks of it many of the other conversions are unsound too.
> > 
> Got it, so I'll use pmu_attr->event_str here, for example,
> SMMU_EVENT_ATTR(cycles, "event=0x00")

You could, but honestly I don't really see this being any better than the
current code. The advantage of using "event=0x%02llx\n" is that things are
consistent by construction and therefore userspace can parse the information
easily. We lose that if we just hardcode the string and it's error-prone to
extend.

Will