Re: [PATCH v10 16/24] x86/resctrl: Add interface to the assign counter

["Moger, Babu" <bmoger@amd.com>]

Hi Reinette,

On 12/18/2024 4:01 PM, Reinette Chatre wrote:
> 
> 
> On 12/13/24 8:54 AM, Moger, Babu wrote:
>> On 12/13/2024 10:24 AM, Luck, Tony wrote:
>>>> It is right thing to continue assignment if one of the domain is out of
>>>> counters. In that case how about we save the error(say error_domain) and
>>>> continue. And finally return success if both ret and error_domain are zeros.
>>>>
>>>>      return ret ?  ret : error_domain:
>>>
>>> If there are many domains, then you might have 3 succeed and 5 fail.
>>>
>>> I think the best you can do is return success if everything succeeded
>>> and an error if any failed.
>>
>> Yes. The above check should take care of this case.
>>
> 
> If I understand correctly "error_domain" can capture the ID of
> a single failing domain. If there are multiple failing domains like
> in Tony's example then "error_domain" will not be accurate and thus
> can never be trusted. Instead of a single check of a failure user
> space is then forced to parse the more complex "mbm_assign_control"
> file to learn what succeeded and failed.
> 
> Would it not be simpler to process sequentially and then fail on
> first error encountered with detailed error message? With that
> user space can determine exactly which portion of request
> succeeded and which portion failed.

One more option is to print the error for each failure and continue. And 
finally return error.

"Group mon1, domain:1 Out of MBM counters"

We have the error information as well as the convenience of assignment 
on domains where counters are available when user is working with 
"*"(all domains).

Note: I will be out of office starting next week Until Jan 10.

> 
>>>
>>> You have the same issue if someone tries to update multiple things
>>> with a single write to mbm_assign_control:
>>>
>>> # cat > mbm_assign_control << EOF
>>> c1/m78/0=t;1=l;
>>> c1/m79/0=t;1=l
>>> c1/m80/0=t;1=l;
>>> c1/m81/0=t;1=l;
>>> EOF
>>>
>>> Those get processed in order, some may succeed, but once a domain
>>> is out of counters the rest for that domain will fail.
>>
>> Yes. I see the similar type of processing for schemata.
>> It is processed sequentially. If one fails, it returns immediately.
>>
>> ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
>> if (ret)
>>        goto out;
>>
>> I feel it is ok to keep same level of processing.
>>
> 
> resctrl also does sequential processing when, for example, the user requests
> move of several tasks. resctrl returns with failure right away with error message
> containing failing PID. This gives clear information to user what
> portion of request succeeded without requiring user space to
> do additional queries.
> 
> 
> Reinette
>