Hi Tony,
On 11/21/24 9:33 AM, Luck, Tony wrote:
> On Tue, Nov 19, 2024 at 05:08:42PM -0800, Reinette Chatre wrote:
>> Hi Tony,
>>
>> On 11/13/24 4:17 PM, Tony Luck wrote:
>>> Resctrl uses local memory bandwidth event as input to the feedback
>>> loop when the mba_MBps mount option is used. This means that this
>>> mount option cannot be used on systems that only support monitoring
>>> of total bandwidth.
>>>
>>> Prepare to allow users to choose the input event independently for
>>> each ctrl_mon group.
>>
> Hi Reinette,
>
>> The lack of detail on design and implementation leaves a lot for the
>> reader to decipher. For example,
>> * the change appears to create a contract that rdtgroup.mba_mbps_event
>> is only valid if mba_sc is enabled, this is "documented" in the
>> structure member comment but not connected to the rest of implementation, not
>> here nor later in series.
>
> I'll add text documenting this to the commit comment here, and also a
> comment in the code that defines mba_mbps_default_event.
>
>> * the patch uses *three* different checks to manage new variables:
>> is_mbm_local_enabled(), is_mba_sc(), and supports_mba_mbps(). Reader is
>> left to decipher that all checks are built on is_mbm_local_enabled()
>> and thus it is ok to use these checks before using the value that is only
>> assigned when is_mbm_local_enabled().
>
> With some refactoring I've got that down to just one new additon of
> "is_mba_sc()" (protecting the assignment of rdtgrp->mba_mbps_event
> in rdtgroup_mkdir_ctrl_mon().
>
Just to clarify, I am not stating that there should not be three different checks. Clearly
if some initialization is done before resctrl mount then the mba_sc checks may not apply.
My goal was to highlight the investigations a reader is forced to do without
guidance from the changelog.
...
>>> @@ -3611,6 +3613,8 @@ static int rdtgroup_mkdir_ctrl_mon(struct kernfs_node *parent_kn,
>>> rdt_last_cmd_puts("kernfs subdir error\n");
>>> goto out_del_list;
>>> }
>>> + if (is_mba_sc(NULL))
>>> + rdtgrp->mba_mbps_event = mba_mbps_default_event;
>>> }
>>>
>>> goto out_unlock;
>>> @@ -3970,6 +3974,8 @@ static void __init rdtgroup_setup_default(void)
>>> rdtgroup_default.closid = RESCTRL_RESERVED_CLOSID;
>>> rdtgroup_default.mon.rmid = RESCTRL_RESERVED_RMID;
>>> rdtgroup_default.type = RDTCTRL_GROUP;
>>> + if (supports_mba_mbps())
>>> + rdtgroup_default.mba_mbps_event = mba_mbps_default_event;
>>> INIT_LIST_HEAD(&rdtgroup_default.mon.crdtgrp_list);
>>>
>>> list_add(&rdtgroup_default.rdtgroup_list, &rdt_all_groups);
>>
>> I do not see the default resource group's mba_mbps_event ever being reset. This means
>> that if the user mounts resctrl, changes mba_mbps_event, umount resctrl, remount
>> resctrl, then the default resource group will not have the default mba_mbps_event
>> but whatever was set on previous mount. Is this intended? No mention of this behavior in
>> changelog.
>
> Good catch. You are correct that a changed event value in the default
> CTRL_MON group is preserved across unmount/remount. This was not
> intentional. Moving the initialization of rdtgroup_default.mba_mbps_event
> into set_mba_sc() fixes this (with the added benefit of removing the
> supports_mba_mbps() check).
I am curious how this will turn out considering the fragmentation of the rdtgroup_default
initialization.
Reinette
