Re: [PATCH v9 18/26] x86/resctrl: Add the interface to assign/update counter assignment

[Reinette Chatre <reinette.chatre@intel.com>]

Hi Babu,

On 11/22/24 1:04 PM, Moger, Babu wrote:
> Hi Reinette,
> 
> On 11/21/2024 2:50 PM, Reinette Chatre wrote:
>> Hi Babu,
>>
>> On 11/20/24 10:05 AM, Moger, Babu wrote:
>>> Hi Reinette,
>>>
>>> On 11/15/24 18:57, Reinette Chatre wrote:
>>>> Hi Babu,
>>>>
>>>> On 10/29/24 4:21 PM, Babu Moger wrote:
>>>>> The mbm_cntr_assign mode offers several hardware counters that can be
>>>>> assigned to an RMID, event pair and monitor the bandwidth as long as it
>>>>> is assigned.
>>>>>
>>>>> Counters are managed at two levels. The global assignment is tracked
>>>>> using the mbm_cntr_free_map field in the struct resctrl_mon, while
>>>>> domain-specific assignments are tracked using the mbm_cntr_map field
>>>>> in the struct rdt_mon_domain. Allocation begins at the global level
>>>>> and is then applied individually to each domain.
>>>>>
>>>>> Introduce an interface to allocate these counters and update the
>>>>> corresponding domains accordingly.
>>>>>
>>>>> Signed-off-by: Babu Moger <babu.moger@amd.com>
>>>>> ---
>>>>
>>>> ...
>>>>
>>>>> diff --git a/arch/x86/kernel/cpu/resctrl/internal.h b/arch/x86/kernel/cpu/resctrl/internal.h
>>>>> index 00f7bf60e16a..cb496bd97007 100644
>>>>> --- a/arch/x86/kernel/cpu/resctrl/internal.h
>>>>> +++ b/arch/x86/kernel/cpu/resctrl/internal.h
>>>>> @@ -717,6 +717,8 @@ unsigned int mon_event_config_index_get(u32 evtid);
>>>>>   int resctrl_arch_config_cntr(struct rdt_resource *r, struct rdt_mon_domain *d,
>>>>>                    enum resctrl_event_id evtid, u32 rmid, u32 closid,
>>>>>                    u32 cntr_id, bool assign);
>>>>> +int rdtgroup_assign_cntr_event(struct rdt_resource *r, struct rdtgroup *rdtgrp,
>>>>> +                   struct rdt_mon_domain *d, enum resctrl_event_id evtid);
>>>>>   void rdt_staged_configs_clear(void);
>>>>>   bool closid_allocated(unsigned int closid);
>>>>>   int resctrl_find_cleanest_closid(void);
>>>>> diff --git a/arch/x86/kernel/cpu/resctrl/rdtgroup.c b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>>>> index 1b5529c212f5..bc3752967c44 100644
>>>>> --- a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>>>> +++ b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>>>> @@ -1924,6 +1924,93 @@ int resctrl_arch_config_cntr(struct rdt_resource *r, struct rdt_mon_domain *d,
>>>>>       return 0;
>>>>>   }
>>>>>   +/*
>>>>> + * Configure the counter for the event, RMID pair for the domain.
>>>>> + * Update the bitmap and reset the architectural state.
>>>>> + */
>>>>> +static int resctrl_config_cntr(struct rdt_resource *r, struct rdt_mon_domain *d,
>>>>> +                   enum resctrl_event_id evtid, u32 rmid, u32 closid,
>>>>> +                   u32 cntr_id, bool assign)
>>>>> +{
>>>>> +    int ret;
>>>>> +
>>>>> +    ret = resctrl_arch_config_cntr(r, d, evtid, rmid, closid, cntr_id, assign);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>> +
>>>>> +    if (assign)
>>>>> +        __set_bit(cntr_id, d->mbm_cntr_map);
>>>>> +    else
>>>>> +        __clear_bit(cntr_id, d->mbm_cntr_map);
>>>>> +
>>>>> +    /*
>>>>> +     * Reset the architectural state so that reading of hardware
>>>>> +     * counter is not considered as an overflow in next update.
>>>>> +     */
>>>>> +    resctrl_arch_reset_rmid(r, d, closid, rmid, evtid);
>>>>
>>>> resctrl_arch_reset_rmid() expects to be run on a CPU that is in the domain
>>>> @d ... note that after the architectural state is reset it initializes the
>>>> state by reading the event on the current CPU. By running it here it is
>>>> run on a random CPU that may not be in the right domain.
>>>
>>> Yes. That is correct.  We can move this part to our earlier
>>> implementation. We dont need to read the RMID.  We just have to reset the
>>> counter.
>>>
>>> https://lore.kernel.org/lkml/16d88cc4091cef1999b7ec329364e12dd0dc748d.1728495588.git.babu.moger@amd.com/
>>>
>>> diff --git a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>> b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>> index 9fe419d0c536..bc3654ec3a08 100644
>>> --- a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>> +++ b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
>>> @@ -2371,6 +2371,13 @@ int resctrl_arch_config_cntr(struct rdt_resource
>>> *r, struct rdt_mon_domain *d,
>>>          smp_call_function_any(&d->hdr.cpu_mask, resctrl_abmc_config_one_amd,
>>>                                &abmc_cfg, 1);
>>>
>>> +       /*
>>> +        * Reset the architectural state so that reading of hardware
>>> +        * counter is not considered as an overflow in next update.
>>> +        */
>>> +       if (arch_mbm)
>>> +               memset(arch_mbm, 0, sizeof(struct arch_mbm_state));
>>> +
>>>          return 0;
>>>   }
>>>
>>>
>>
>> I am not sure what you envision here. One motivation for the move out of
>> resctrl_arch_config_cntr() was to avoid architectural state being reset twice. For reference,
>> mbm_config_write_domain()->resctrl_arch_reset_rmid_all(). Will architectural state
>> be reset twice again?
> 
> That is good point. We don't have to do it twice.
> 
> We can move the whole reset(arch_mbm) in  resctrl_arch_config_cntr().

This is not clear to me. The architectural state needs to be reset on MBM config write even
when assignable mode is not supported and/or enabled. Moving it to resctrl_arch_config_cntr()
will break this, no?

I wonder if it may not simplify things to call resctrl_arch_reset_rmid() from
resctrl_abmc_config_one_amd()?

>> One thing that I did not notice before is that the non-architectural MBM state is not
>> reset. Care should be taken to reset this also when considering that there is a plan
>> to use that MBM state to build a generic rate event for all platforms:
>> https://lore.kernel.org/all/CALPaoCgFRFgQqG00Uc0GhMHK47bsbtFw6Bxy5O9A_HeYmGa5sA@mail.gmail.com/
> 
> Did you mean we should add the following code in resctrl_arch_config_cntr()?
> 
> m = get_mbm_state(d, closid, rmid, evtid);
> if (m)
>      memset(m, 0, sizeof(struct mbm_state));

This is not arch code but instead resctrl fs, so resctrl_config_cntr() may be more appropriate?

Reinette