Re: [RFC][PATCH v2 11/11] perf top: demo of how to use the sysfs interface

[Corey Ashford <cjashfor@linux.vnet.ibm.com>]

On 5/20/2010 2:21 AM, Lin Ming wrote:
> On Thu, 2010-05-20 at 16:27 +0800, Corey Ashford wrote:
>> On 05/19/2010 07:08 PM, Lin Ming wrote:
>>> On Thu, 2010-05-20 at 09:33 +0800, Corey Ashford wrote:
>>>> On 5/19/2010 6:17 PM, Lin Ming wrote:
>>>>> On Thu, 2010-05-20 at 02:18 +0800, Corey Ashford wrote:
>>>>>>
>>>>>> On 5/18/2010 6:49 PM, Lin Ming wrote:
>>>>>>> Just a temporary patch to show how to use the pmu sysfs interface...
>>>>>>>
>>>>>>> Signed-off-by: Lin Ming<ming.m.lin@intel.com>
-- snip --
>>>>>>> +		 * This is just an ugly demo of how to use the sysfs interface.
>>>>>>> +		 * You can also parse the<event-name>  and open sys file as,
>>>>>>> +		 * sys_fd = open("/sys/devices/system/cpu/events/<event-name>/event_source/id", O_RDONLY);
>>>>>>> +		 */
>>>>>>
>>>>>> In this above case, does this sys_fd also specify the event I am going to open, in addition to its event source?  I'd assume not since event_source is a symlink to /sys/devices/system/cpu/event_source (right?)
>>>>>
>>>>> Right, this sys_fd only specifies the event source.
>>>>>
>>>>>> How do I specify the exact event id via the sysfs interface?
>>>>>
>>>>> /sys/devices/system/cpu/events/<event-name>/id
>>>>>
>>>>> But in this patch series, the event id sysfs interface is not used yet.
>>>>
>>>> So, I would open that id and then read the id code and place it in attr->config or maybe place
>>>> the fd into attr (somewhere) ?
>>>
>>> Place the id code in attr->config.
>>>
>>>>
>>>> We also need to take into account event "attributes" - other data that is needed to configure a specific event.  For example, think about a memory controller which has a PMU can count events in a particular memory range; we need to be able to supply the memory range somehow, and I don't think that can be accomplished by passing in the fd of a sysfs file that we've opened.
>>>>
>>>
>>> Each event is a directory in the sysfs, so we can put all the event
>>> "attributes" under it.
>>>
>>> For your example,
>>> /sys/devices/system/node/events/<event-name>/id
>>> /sys/devices/system/node/events/<event-name>/memory_range
>>> ....
>>>
>>> Then we can read these attributes and pass the value into the syscall.
>>>
>>
>> I'm not sure I made the example clear.
>>
>> Let's say I have a memory controller event called memory_write and it 
>> has two attributes: low_addr and high_addr... writes to addresses 
>> between the low_addr and high_addr will increment the counter.
>>
>> As a user, I want to be able to specify a particular memory range, let's 
>> say 0x1000000..0x2000000
>>
>> A sysfs structure like this might be constructed:
>> /sys/devices/system/node/events/memory_write
>> /sys/devices/system/node/events/memory_write/attr
>> /sys/devices/system/node/events/memory_write/attr/low_addr
>> /sys/devices/system/node/events/memory_write/attr/low_addr/min
>> /sys/devices/system/node/events/memory_write/attr/low_addr/max
>>
>> In another posting I had also added a bit shift value, but there are so 
>> few bits left in the attr->config, that I'm not sure this is a very 
>> extensible mechanism, but just for the sake of illustration of the basic 
>> idea, I'll add it here:
>> /sys/devices/system/node/events/memory_write/attr/low_addr/shift
>>
>> Then the same thing is repeated for the other attribute:
>> /sys/devices/system/node/events/memory_write/attr/high_addr
>> /sys/devices/system/node/events/memory_write/attr/high_addr/min
>> /sys/devices/system/node/events/memory_write/attr/high_addr/max
>> /sys/devices/system/node/events/memory_write/attr/high_addr/shift
>>
>> In this scenario, a user tool (like perf) would be able to see that 
>> there are attributes associated with the memory_write event, and it 
>> knows the names and range of allowed values for these attributes.  The 
>> shift value tells the tool how much to shift the attribute value before 
>> OR'ing it into the attr->config value.
> 
> So different events may have different shift value for its attributes.

Yes, depending on the hardware, different attributes will be needed with different
[sets of] events.  When more than one event shares a set of attributes (for example memory_read and memory_write events would share low_addr and high_addr), the attributes can be placed at the node level and then symlinked from the event.

So for example:
/sys/devices/system/node/attrs/memory/low_addr
/sys/devices/system/node/attrs/memory/low_addr/min
/sys/devices/system/node/attrs/memory/low_addr/max
/sys/devices/system/node/attrs/memory/low_addr/shift

Then in the specific event:
/sys/devices/system/node/events/memory_write/attr/low_addr --> /sys/devices/system/node/attrs/memory/low_addr
/sys/devices/system/node/events/memory_read/attr/low_addr --> /sys/devices/system/node/attrs/memory/low_addr

So we could save some space this way.

> If an event has 4 attributes, each attribute will have a "shift" in
> sysfs, right?

Exactly.  It allows the user tool to construct a attr->config (+ config_extra[]) value with many multi-bit fields.  

I'm not convinced that this is the best way to go, so any alternative ideas would be welcome.

We are doing something similar with the IBM Wirespeed processor implementation of perf_events, where we are packing the attribute values into the config value, but we are using Stephane Eranian's libpfm4 to handle all the details of doing the packing, instead of exporting the packing info via sysfs.

- Corey