linux-kernel - Re: [PATCH v5 00/21] x86/resctrl: Make resctrl_arch_rmid

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <a05e0397-50be-2f99-e461-60852c8efb3c@linux.alibaba.com>
Date:   Wed, 7 Sep 2022 13:46:17 +0800
From:   haoxin <xhao@...ux.alibaba.com>
To:     Reinette Chatre <reinette.chatre@...el.com>,
        James Morse <james.morse@....com>, x86@...nel.org,
        linux-kernel@...r.kernel.org
Cc:     Fenghua Yu <fenghua.yu@...el.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
        H Peter Anvin <hpa@...or.com>,
        Babu Moger <Babu.Moger@....com>,
        shameerali.kolothum.thodi@...wei.com,
        D Scott Phillips OS <scott@...amperecomputing.com>,
        lcherian@...vell.com, bobo.shaobowang@...wei.com,
        tan.shaopeng@...itsu.com, Jamie Iles <quic_jiles@...cinc.com>,
        Cristian Marussi <cristian.marussi@....com>,
        xingxin.hx@...nanolis.org, baolin.wang@...ux.alibaba.com
Subject: Re: [PATCH v5 00/21] x86/resctrl: Make resctrl_arch_rmid_read()
 return values in bytes


在 2022/8/24 上午1:09, Reinette Chatre 写道:
> Hi,
>
> On 7/3/2022 8:54 AM, Xin Hao wrote:
>> Hi  james,
>>
>> I have a review all of the patches, it looks goot to me, but i also test them once again, i have a little confusion with my test.
>>
>> # mkdir p1
>>
>> # echo "L3:0=001;1=001" > p1/schemata
>>
>> # [root@...p1bu26qv0j3ddyusot3Z p1]# cat schemata
>>      MB:0=100;1=100
>>      L3:0=001;1=001
>>
>> # memhog -r1000000 1000m > /mnt/log &
>>
>> [1] 53023
>> [root@...p1bu26qv0j3ddyusot3Z p1]# echo 53023 > tasks
>> [
>> [root@...p1bu26qv0j3ddyusot3Z p1]# cat mon_data/mon_L3_00/llc_occupancy
>> 3833856
>> [root@...p1bu26qv0j3ddyusot3Z p1]# cat mon_data/mon_L3_00/llc_occupancy
>> 3620864
>> [root@...p1bu26qv0j3ddyusot3Z p1]# cat mon_data/mon_L3_00/llc_occupancy
>> 3727360
>> [root@...p1bu26qv0j3ddyusot3Z p1]# cat size
>>      MB:0=100;1=100
>>      L3:0=3407872;1=3407872
>>
>> Obviously, the value has been overflowed,  Can you explain why？
> Are you seeing different behavior before and after you apply this
> series?
No,they have the same test result。
>
> I do not think the conclusion should immediately be that there is an
> overflow issue. Have you perhaps run into the scenario "Notes on
> cache occupancy monitoring and control" described in
> Documentation/x86/resctrl.rst?
>
> When "memhog" starts it can allocate to the entire L3 for a while
> before it is moved to the constrained resource group. It's cache
> lines are not evicted as part of this move so it is not unusual for
> it to have more lines in L3 than it is allowed to allocate into.

Yes as you said, the mon_data/mon_L3_00/llc_occupancy does not 
immediately become the value small than the set by schemata,  it may 
takes a few minutes to reduce to the set value.

I don't quite understand why it takes so long to see the llc_occupancy 
degrage.

>
> Understanding the occupancy values require understanding of the workload
> as well as the system environment.
>
> Depending on the workload's data usage (for example if it keeps loading
> new data - note that if the workload keeps loading the same data and the
> data is already present in an area of cache that the workload cannot
> allocate into then the data read would still result in a cache hit for the
> workload, the data would not be moved to the area the
> workload can allocate into) and other workloads on the system (there is
> other load present also that evicts the lines owned by the workload) the
> L3 occupancy rate should go down after a while to match the space it
> can allocate into.
>
> Reinette