Received-SPF: pass (google.com: domain of linux-kernel+bounces-51501-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
From: Yicong Yang <yangyicong@huawei.com>
To: <jonathan.cameron@huawei.com>, <will@kernel.org>, <mark.rutland@arm.com>,
	<hejunhao3@huawei.com>, <linux-arm-kernel@lists.infradead.org>,
	<linux-kernel@vger.kernel.org>
CC: <yangyicong@hisilicon.com>, <linuxarm@huawei.com>,
	<prime.zeng@hisilicon.com>, <fanghao11@huawei.com>
Subject: [PATCH 5/7] drivers/perf: hisi_pcie: Relax the check on related events
Date: Sun, 4 Feb 2024 15:45:25 +0800
Message-ID: <20240204074527.47110-6-yangyicong@huawei.com>
In-Reply-To: <20240204074527.47110-1-yangyicong@huawei.com>
References: <20240204074527.47110-1-yangyicong@huawei.com>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Content-Type: text/plain

From: Junhao He <hejunhao3@huawei.com>

If we use two events with the same filter and related event type
(see the following example), the driver check whether they are related
events and are in the same group, otherwise the function
hisi_pcie_pmu_find_related_event() return -EINVAL, then the 2nd event
cannot count but the 1st event is running, although the PCIe PMU has
other idle counters.

In this case, The perf event scheduler will make the two events to
multiplex a counter, if the user use the formula
(1st event_value / 2nd event_value) to calculate the bandwidth, he/she
won't get the correct value, because they are not counting at the
same period.

This patch tries to fix this by making the related events to use
different idle counters if they are not in the same event group.

And finally, I'm going to say. The related events are best used in the
same group [1]. There are two ways to know if they are related events.
a) By event name, such as the latency events "xxx_latency, xxx_cnt" or
bandwidth events "xxx_flux, xxx_time".
b) By event type, such as "event=0xXXXX, event=0x1XXXX".

Use group to count the related events:
  [1] -e "{pmu_name/xxx_latency,port=1/,pmu_name/xxx_cnt,port=1/}"

  example:
    1st event: hisi_pcie0_core1/event=0x804,port=1
    2nd event: hisi_pcie0_core1/event=0x10804,port=1

  test cmd:
    perf stata -e hisi_pcie0_core1/event=0x804,port=1/ \
               -e hisi_pcie0_core1/event=0x10804,port=1/

  before patch:
            25,281      hisi_pcie0_core1/event=0x804,port=1/    (49.91%)
           470,598      hisi_pcie0_core1/event=0x10804,port=1/    (50.09%)

  after patch:
            24,147      hisi_pcie0_core1/event=0x804,port=1/
           474,558      hisi_pcie0_core1/event=0x10804,port=1/

Signed-off-by: Junhao He <hejunhao3@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
---
 drivers/perf/hisilicon/hisi_pcie_pmu.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/drivers/perf/hisilicon/hisi_pcie_pmu.c b/drivers/perf/hisilicon/hisi_pcie_pmu.c
index b91f03c02c57..1b45aeb82859 100644
--- a/drivers/perf/hisilicon/hisi_pcie_pmu.c
+++ b/drivers/perf/hisilicon/hisi_pcie_pmu.c
@@ -408,14 +408,10 @@ static int hisi_pcie_pmu_find_related_event(struct hisi_pcie_pmu *pcie_pmu,
 		if (!sibling)
 			continue;
 
-		if (!hisi_pcie_pmu_cmp_event(sibling, event))
-			continue;
-
 		/* Related events must be used in group */
-		if (sibling->group_leader == event->group_leader)
+		if (hisi_pcie_pmu_cmp_event(sibling, event) &&
+		    sibling->group_leader == event->group_leader)
 			return idx;
-		else
-			return -EINVAL;
 	}
 
 	return idx;
-- 
2.24.0