RE: MPAM branch verification (was RE: [RFC PATCH 2/2] ACPI / PPTT: cacheinfo: Label caches based on fw_token)

[Shameerali Kolothum Thodi <shameerali.kolothum.thodi@huawei.com>]

Hi James,

Sorry for the delay. It took a while to get back into this.

> -----Original Message-----
> From: James Morse [mailto:james.morse@arm.com]
> Sent: 19 July 2019 16:30
> To: Shameerali Kolothum Thodi <shameerali.kolothum.thodi@huawei.com>
> Cc: Vijaya Kumar K <vkilari@codeaurora.org>; Lorenzo Pieralisi
> <lorenzo.pieralisi@arm.com>; Tomasz Nowicki
> <Tomasz.Nowicki@cavium.com>; Jeffrey Hugo <jhugo@codeaurora.org>;
> Guohanjun (Hanjun Guo) <guohanjun@huawei.com>; Linuxarm
> <linuxarm@huawei.com>; Jeremy Linton <jeremy.linton@arm.com>;
> linux-acpi@vger.kernel.org; linux-arm-kernel@lists.infradead.org; Sudeep Holla
> <sudeep.holla@arm.com>; wangxiongfeng (C)
> <wangxiongfeng2@huawei.com>; Richard Ruigrok
> <rruigrok@qti.qualcomm.com>
> Subject: Re: MPAM branch verification (was RE: [RFC PATCH 2/2] ACPI / PPTT:
> cacheinfo: Label caches based on fw_token)
> 
> Hi Shameer,
> 
> On 03/07/2019 13:27, Shameerali Kolothum Thodi wrote:
> >> -----Original Message-----
> >> On 21/06/2019 16:57, Shameerali Kolothum Thodi wrote:
> >>>> -----Original Message-----
> >>>> From: James Morse [mailto:james.morse@arm.com]
> 
> >> The domid bitfield not being big enough for the width of the cacheinfo id field
> >> looks like
> >> a bug in the existing resctrl code. Could you spin that as a patch against
> >> mainline?
> >
> > Yes it could be a bug. But I am not sure about the assumption on x86
> platforms with
> > respect to cache id width. Also any need to consider 32 bit systems at all or
> not.
> >
> >> It won't affect any x86 system, but I don't want to 'fix' anything as part of
> the
> >> mpam
> >> support.
> >
> > Does that mean the cache id width on x86 will never be >14 bits?
> 
> I have no idea. Today they're 0,1,2, so its unlikely?, but
> Documentation/x86/resctrl.rst's
> "Cache IDs" section says "it isn't guaranteed to be a contiguous sequence", so
> maybe?
> 
> The problem is 'struct cacheinfo's id field is an int, its exposed via sysfs as an
> int,
> but resctrl packs it into a smaller size. That's going to bite one day, it would be
> good
> to fix it now we know its a problem.
> 
> 
> >> We almost certainly need to compress the cache-id numbers down to {0,1,2}
> if
> >> only so we
> >> haven't filled all the exposed bits on day-1. (so it might not matter for arm64
> >> either...)
> >
> > That will be nice if we can compress it like that> I think we can leave the fix
> for now
> > and come up with a solution when things gets really going.
> >
> > Mean time I am trying to probe memory controller as well on our system and
> it looks
> > like there are still issues.
> 
> Typo in the MBA picking code? Should be:
> | if (!mpam_has_feature(mpam_feat_mbw_part, class->features) &&
> |     !mpam_has_feature(mpam_feat_mbw_max, class->features)) {
> 
> It can do something useful with either of those features, but the (!part || !max)
> previously forced it to have both.
> 
> (This still doesn't work on the model as its describing a 0-bit bitmap
> MBW_PART)

I think what happens on our hardware is, the MBA reports PMG_MAX = 0 and that
upsets mpam_pmg_bits() -->ilog2(). I am not entirely sure whether PMG_MAX= 0 is
allowed as per spec when the resource reports HAS_MSMON =1. But hasn't found
anything in spec that forbids this as the filter is a combination of PRATID:PMG.

I have a temp hack here to keep it going,

https://github.com/hisilicon/kernel-dev/commit/5e0881c4cdded4066dfac7603c53242385417a3a
 
> 
> > I will debug and update if it really is a problem. Please
> > let me know if you have any plans to update the branch so that I can try the
> latest.
> 
> I hope to push a new version by the end of June. (whoosh! There goes June).
> http://www.linux-arm.org/git?p=linux-jm.git;a=shortlog;h=refs/heads/mpam/s
> napshot/jun

Thanks for that. I am using this now. (And I see a more recent one mpam/5.3-tmp
now. Has anything changed other than rebase?)

>
> The changes in there are to avoid the known-issues when the same 'thing' is
> picked as both
> L3 resource and the MBA resource.

Now with the above fix for PMG_MAX=0, I am hitting another issue.
mount -t resctrl resctrl /sys/fs/resctrl fails with "File exists" error.

Debugging points to,
rdt_get_tree() 
  mkdir_mondata_all()
    mkdir_mondata_subdir_alldom()
      mkdir_mondata_subdir() 
        mon_addfile()

It looks like r->evt_list gets corrupted somehow and has duplicate entries. I haven’t
gone into the bottom of this issue, but please let me know if you have any idea.

Cheers,
Shameer

> I think the risk of sleeping-while-atomic if not all mpam:devices are accessible
> from all
> CPUs in the resctrl:domain is my next highest priority issue...
> 
> 
> Thanks,
> 
> James