[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <0cd2c8ac-8dee-4280-b726-af0119baa4a1@amd.com>
Date: Tue, 7 Oct 2025 12:36:18 -0500
From: Babu Moger <bmoger@....com>
To: Reinette Chatre <reinette.chatre@...el.com>, babu.moger@....com,
tony.luck@...el.com, Dave.Martin@....com, james.morse@....com,
dave.hansen@...ux.intel.com, bp@...en8.de
Cc: kas@...nel.org, rick.p.edgecombe@...el.com, linux-kernel@...r.kernel.org,
x86@...nel.org, linux-coco@...ts.linux.dev, kvm@...r.kernel.org
Subject: Re: [PATCH] fs/resctrl: Fix MBM events being unconditionally enabled
in mbm_event mode
Hi Reinette,
On 10/6/25 20:23, Reinette Chatre wrote:
> Hi Babu,
>
> On 10/6/25 1:38 PM, Moger, Babu wrote:
>> Hi Reinette,
>>
>> On 10/6/25 12:56, Reinette Chatre wrote:
>>> Hi Babu,
>>>
>>> On 9/30/25 1:26 PM, Babu Moger wrote:
>>>> resctrl features can be enabled or disabled using boot-time kernel
>>>> parameters. To turn off the memory bandwidth events (mbmtotal and
>>>> mbmlocal), users need to pass the following parameter to the kernel:
>>>> "rdt=!mbmtotal,!mbmlocal".
>>>
>>> ah, indeed ... although, the intention behind the mbmtotal and mbmlocal kernel
>>> parameters was to connect them to the actual hardware features identified
>>> by X86_FEATURE_CQM_MBM_TOTAL and X86_FEATURE_CQM_MBM_LOCAL respectively.
>>>
>>>
>>>> Found that memory bandwidth events (mbmtotal and mbmlocal) cannot be
>>>> disabled when mbm_event mode is enabled. resctrl_mon_resource_init()
>>>> unconditionally enables these events without checking if the underlying
>>>> hardware supports them.
>>>
>>> Technically this is correct since if hardware supports ABMC then the
>>> hardware is no longer required to support X86_FEATURE_CQM_MBM_TOTAL and
>>> X86_FEATURE_CQM_MBM_LOCAL in order to provide mbm_total_bytes
>>> and mbm_local_bytes.
>>>
>>> I can see how this may be confusing to user space though ...
>>>
>>>>
>>>> Remove the unconditional enablement of MBM features in
>>>> resctrl_mon_resource_init() to fix the problem. The hardware support
>>>> verification is already done in get_rdt_mon_resources().
>>>
>>> I believe by "hardware support" you mean hardware support for
>>> X86_FEATURE_CQM_MBM_TOTAL and X86_FEATURE_CQM_MBM_LOCAL. Wouldn't a fix like
>>> this then require any system that supports ABMC to also support
>>> X86_FEATURE_CQM_MBM_TOTAL and X86_FEATURE_CQM_MBM_LOCAL to be able to
>>> support mbm_total_bytes and mbm_local_bytes?
>>
>> Yes. That is correct. Right now, ABMC and X86_FEATURE_CQM_MBM_TOTAL/
>> X86_FEATURE_CQM_MBM_LOCAL are kind of tightly coupled. We have not clearly
>> separated the that.
>
> Are you speaking from resctrl side since from what I understand these are
> independent features from the hardware side?
It is independent from hardware side. I meant we still use legacy events
from "default" mode.
>
>>> This problem seems to be similar to the one solved by [1] since
>>> by supporting ABMC there is no "hardware does not support mbmtotal/mbmlocal"
>>> but instead there only needs to be a check if the feature has been disabled
>>> by command line. That is, add a rdt_is_feature_enabled() check to the
>>> existing "!resctrl_is_mon_event_enabled()" check?
>>
>> Enable or disable needs to be done at get_rdt_mon_resources(). It needs to
>> be done early in the initialization before calling domain_add_cpu() where
>> event data structures (mbm_states aarch_mbm_states) are allocated.
>
> Good point. My mistake to suggest the event should be enabled by
> resctrl fs.
How about adding another check in get_rdt_mon_resources()?
if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)
|| rdt_is_feature_enabled(mbmtotal)) {
resctrl_enable_mon_event(QOS_L3_MBM_TOTAL_EVENT_ID);
ret = true;
}
I need to take Tony's patch for this.
>
>>
>>>
>>> But wait ... I think there may be a bigger problem when considering systems
>>> that support ABMC but not X86_FEATURE_CQM_MBM_TOTAL and X86_FEATURE_CQM_MBM_LOCAL.
>>> Shouldn't resctrl prevent such a system from switching to "default"
>>> mbm_assign_mode? Otherwise resctrl will happily let such a system switch
>>> to default mode and when user attempts to read an event file resctrl will
>>> attempt to read it via MSRs that are not supported.
>>> Looks like ABMC may need something similar to CONFIG_RESCTRL_ASSIGN_FIXED
>>> to handle this case in show() while preventing user space from switching to
>>> "default" mode on write()?
>>
>> This may not be an issue right now. When X86_FEATURE_CQM_MBM_TOTAL and
>> X86_FEATURE_CQM_MBM_LOCAL are not supported then mon_data files of these
>> events are not created.
>
> By "right now" I assume you mean the current implementation? I think your statement
> assumes that no CPUs come or go after resctrl_mon_resource_init() enables the MBM events?
> Current implementation will enable MBM events if ABMC is supported. When the
> first CPU of a domain comes online after that then resctrl will create the mon_data
> files. These files will remain if a user then switches to default mode and if
> the user then attempts to read one of these counters then I expect problems.
Yes. It will be a problem in the that case.
I am not clear on using config option you mentioned above.
What about using the check resctrl_is_mon_event_enabled() in
resctrl_mbm_assign_mode_show() and resctrl_mbm_assign_mode_write() ?
Thanks
Babu
Powered by blists - more mailing lists