Date: Mon, 24 Jul 2017 21:30:22 +0800
From: Jonathan Cameron <Jonathan.Cameron@huawei.com>
To: Jan Glauber <jglauber@cavium.com>
CC: Mark Rutland <mark.rutland@arm.com>, Will Deacon <will.deacon@arm.com>,
        <linux-kernel@vger.kernel.org>, <linux-arm-kernel@lists.infradead.org>
Subject: Re: [PATCH v7 3/3] perf: cavium: Add Documentation
Message-ID: <20170724213022.0000196f@huawei.com>
In-Reply-To: <20170719120847.21183-4-jglauber@cavium.com>
References: <20170719120847.21183-1-jglauber@cavium.com>
        <20170719120847.21183-4-jglauber@cavium.com>
Organization: Huawei
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3814
Lines: 94

On Wed, 19 Jul 2017 14:08:47 +0200
Jan Glauber <jglauber@cavium.com> wrote:

> Document Cavium SoC PMUs.
> 
> Signed-off-by: Jan Glauber <jglauber@cavium.com>
And just to finish - the utterly trivial ;)
> ---
>  Documentation/perf/cavium-pmu.txt | 74 +++++++++++++++++++++++++++++++++++++++
>  1 file changed, 74 insertions(+)
>  create mode 100644 Documentation/perf/cavium-pmu.txt
> 
> diff --git a/Documentation/perf/cavium-pmu.txt b/Documentation/perf/cavium-pmu.txt
> new file mode 100644
> index 0000000..37669b9
> --- /dev/null
> +++ b/Documentation/perf/cavium-pmu.txt
> @@ -0,0 +1,74 @@
> +Cavium ThunderX and OcteonTx Performance Monitoring Unit (PMU)
> +==============================================================
> +
> +Cavium SoCs contain various system devices such as L2 caches, processor
> +interconnect and memory controllers. Unfortunately the PMU counters
> +are not following a common design so each device has a slightly different
> +approach how to control and use the PMU counters.
> +
> +Common properties of all devices carrying PMU counters:
> +- The devices are PCI devices and the counters are embedded somewhere
> +  in the PCI register space.
> +- All counters are 64 bit wide.
> +- There are no overflow interrupts (unnecessary because of the 64 bit wide counters).
Too long a line > 80 chars.
> +
> +Properties depending on the device type:
> +- How to start/stop the counters
> +- Programmable vs. fixed purpose counters
> +- Stoppable vs. always running counters
> +- Independent vs. grouped counters
> +- Read-only vs. writable counters
> +- PCI device to PMU group relationship
> +
> +
> +Devices with PMU counters
> +-------------------------
> +
> +Memory controller (LMC):
> +- one PCI device per LMC
> +- fixed-purpose counters
> +- always running counters without start/stop/reset control
> +- read-only counters
> +
> +CCPI interface controller (OCX) Transmit link (TLK) counters:
> +- writable counters
> +- only one PCI device exposes multiple TLK units (3 units on T88)
> +- start/stop control per unit
> +- only present on multi-socket systems
> +
> +PMU (perf) driver
> +-----------------
> +
> +The cavium-pmu driver registers several perf PMU drivers. Each of the perf
> +driver provides description of its available events and configuration options
> +in sysfs, see /sys/devices/<lmcX/ocx_tlkX>/.
> +
> +The "format" directory describes format of the config (event ID),
> +The "events" directory shows the names of the events and provides configuration
> +templates for all supported event types that can be used with perf tool. For
> +example, "lmc0/dclk_cnt/" is an equivalent of "lmc0/config=2/".
> +
> +Each perf driver also provides a "cpumask" sysfs attribute, which contains a
> +single CPU ID of the processor which will be used to handle all the PMU events.
> +
> +Example for perf tool use:
> +
> + / # perf list | grep -e lmc
> +   lmc0/bank_conflict1/                               [Kernel PMU event]
> +   lmc0/bank_conflict2/                               [Kernel PMU event]
> +   lmc0/dclk_cnt/                                     [Kernel PMU event]
> +   lmc0/ifb_cnt/                                      [Kernel PMU event]
> +   lmc0/ops_cnt/                                      [Kernel PMU event]
> +
> + / # perf stat -a -e lmc0/ops_cnt/,lmc0/dclk_cnt/ -- sleep 1
> +
> +   Performance counter stats for 'system wide':
> +
> +             176,133      lmc0/ops_cnt/                                               
> +         670,243,653      lmc0/dclk_cnt/                                              
> +
> +         1.005479295 seconds time elapsed
> +
> +The driver does not support sampling, therefore "perf record" will
> +not work. System wide mode ("-a") must be used as per-task (without "-a")
> +perf sessions are not supported.