[PATCH v7 2/2] drivers/perf: hisi: Add driver for HiSilicon PCIe PMU

[liuqi (BA)]

On 2021/6/26 19:44, John Garry wrote:
> On 26/06/2021 03:13, liuqi (BA) wrote:
>>
>>
>> On 2021/6/25 23:53, John Garry wrote:
>>> On 24/06/2021 11:59, Qi Liu wrote:
>>>> +
>>>> +/*
>>>> + * Events with the "dl" suffix in their names count performance in 
>>>> DL layer,
>>>> + * otherswise, events count performance in TL layer.
>>>> + */
>>>> +static struct attribute *hisi_pcie_pmu_events_attr[] = {
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_rx_mwr, 0x010004),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_rx_mrd, 0x100005),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_tx_mwr, 0x010005),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_tx_mrd, 0x200004),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(lat_rx_mwr, 0x000010),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(lat_rx_mrd, 0x020010),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(lat_tx_mrd, 0x000011),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_rx_dl, 0x010084),
>>>> +    HISI_PCIE_PMU_EVENT_ATTR(bw_tx_dl, 0x030084),
>>>> +    NULL
>>>> +};
>>>> +
>>>> +static struct attribute_group hisi_pcie_pmu_events_group = {
>>>> +    .name = "events",
>>>> +    .attrs = hisi_pcie_pmu_events_attr,
>>>> +};
>>>> +
>>>> +static struct attribute *hisi_pcie_pmu_format_attr[] = {
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(event, "config:0-15"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(subevent, "config:16-23"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(thr_len, "config1:0-3"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(thr_mode, "config1:4"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(trig_len, "config1:5-8"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(trig_mode, "config1:9"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(port, "config2:0-15"),
>>>> +    HISI_PCIE_PMU_FORMAT_ATTR(bdf, "config2:16-31"),
>>>> +    NULL
>>>> +};
>>>
>>> I am just wondering how this now works.
>>>
>>> So if the user programs the following:
>>> ./perf stat -v -e hisi_pcieX/lat_rx_mrd/
>>>
>>> Then the value (incremented) in HISI_PCIE_CNT (I think that's the 
>>> right one) is returned as the event count. But one would expect 
>>> bandwidth from that event, while we only return here the delay cycles 
>>> - how is the count in HISI_PCIE_CNT_EXT exposed, so userspace can do 
>>> the calc for bw?
>>>
>>
> 
> Hi Qi,
> 
>>
>> Hardware counter and ext_counter work together for bandwidth, latency,
>> bus utilization and buffer occupancy events. For example, for latency
>> events(idx = 0x10), counter counts total delay cycles and ext_counter
>> counts PCIe packets number.
>>
>> As we don't want PMU driver to process these two data, "delay cycles"
>> can be treated as an event(id = 0x10), "packets number" as another event
>> (id = 0x10 << 8), and driver could export these data separately.
>>
>> if the user want to calculate latency of rx memory read, they should:
>> ./perf stat -v -e '{hisi_pcieX/event=0x10, 
>> subevent=0x01/,hisi_pcieX/event=0x0400, subevent=0x01/
>>
>> and for bandwidth event:
>> ./perf stat -v -e '{hisi_pcieX/event=0x4, 
>> subevent=0x02/,hisi_pcieX/event=0x1000, subevent=0x02/
> 
Hi John,
> I would suggest supporting a perf metric for this then, which would be 
> like:
> 
> {
>     "BriefDescription": "Latency for inbound traffic...",
>     "MetricName": "hisi_pcie_lat_rx_mrd",
>     "MetricExpr": "hisi_pcieX at event\\=0x4 at subevent\\=0x02 \ 
> hisi_pcieX at event\\=0x1000 at subevent\\=0x02 \",
>     "Unit": "hisi_pci",
>     "Compat": "v1"
> },
> 
> (syntax may be incorrect - illustration only)
> 
yes, we could add these metrics in json file, thanks.
>>
>> Then the value in HISI_PCIE_CNT and HISI_PCIE_EXT_CNT returned 
>> separately, and userspace could do the calculation.
> 
> But I am still curious about lat_rx_mrd and the other events which we 
> continue to advertise. They don't really provide latency or bandwidth on 
> their own, but only half the necessary data. So I doubt their purpose.
> 
So how about changing the event name to show the real purpose of this 
event, like changing "bw_rx_mrd" to "flux_rx_mrd", and changing 
"lat_rx_mrd" to "delay_rx_mrd"?

Thanks, Qi

> Thanks,
> John
> 
> .