resctrl2 - status

[James Morse]

Hi Tony,

On 06/09/2023 19:21, Tony Luck wrote:
> I've just pushed an updated set of patches to:
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux.git resctrl_v65
> 
> Rebased to v6.5. Includes the module auto-load code discussed in
> previous e-mail.

I've taken a look at your resctrl2_v65rc7 branch. I don't think my reply to your first
series got sent, sorry if I'm actually repeating myself.

It goes without saying that I'd prefer we refactor the existing filesystem into the shape
we need, instead of a wholesale replacement.


All this focusses on the structure of the code, but we can change that anytime. I'd prefer
we work out what user-space needs and go from there.

I think the problems are:
 * Discovering what the platform supports, including possible conflicting schemes.
 * Architecture/Platform specific schema. e.g. SMBA. MPAM will have a handful of these too.
* User-space should know which schema are portable, and which are not. e.g. L2/L3/MB.
* Different control types for the same resource. e.g. Intel uses a percentage for
bandwidth control, AMD an absolute number. We should standardise this sort of thing and
make it discoverable.
* Conflicting schemes for the same hardware. e.g. CDP and mba_MBps

~

I'd really like to keep the 'core' schema in the base driver. This is to ensure that 'L3'
doesn't behave differently due to differing software implementation on Intel/AMD/Arm/riscv.

The L{2.3} schema were really well defined as they reference other things that are already
visible in sysfs, this makes them nice and portable. I think this is the standard we
should aim for with any new schema, and only resort to making them arch-specific if the
hardware is just too strange.
I tried to find a way to do this for SMBA as arm platforms will certainly have something
similar, but the 'any other NUMA node' behaviour was just too specific to what was built.

I suspect this could be done with a frontend/backend split, with a 'common' directory
holding the frontend driver for schema that apply to all architectures if the backend has
shown up.

~

Because I think the types a schema is configured with should be portable across
architectures, I'd prefer the string parsing code lives in the core code, and isn't
duplicated by every submodule! String parsing in the kernel is bad enough!

The obvious types we have so far are: bitmap/percentage/number/bandwidth. I think it makes
sense for user-space to be told what the schema is configured with, and let the core code
do that parsing.

~

I don't have a good answer for conflicting drivers for the same hardware resource. I think
we'd need to find a way of making the existing command-line arguments work, causing the
corresponding module to auto-load.

But this sucks for distros, who would need somewhere to describe the policy of which
modules get loaded. The good news is things like libvirt aren't mounting resctrl themselves.

~

I suspect the CDP schemes should be made arch-specific as x86 needs the odd/even pairing,
whereas MPAM does not. This would allow a scheme where the Iside CLOSIDs can be shared to
avoid halving the available CLOSID.

Having somewhere sensible to put the MPAM driver is useful. It's currently dumped in
drivers/platform as it doesn't really fit anywhere!

Allowing each submodule to add extra files to the info directories looks useful. MPAM's
priority control has a property to say whether 0 is the top of the scale or not, this
would allow it to be exposed to user-space, instead of having to shift the range up/down
to hide the difference in behaviour.

MPAM only needs to update the hardware on a CPU that is part of target domain on some
platforms. The vast majority would allow the MMIO writes to come from anywhere. Having the
applychanges behaviour specific to the submodule would reduce the number of IPI.
>From what I've seen, riscv never needs an IPI here.

It looks like all the core code runs at process context without touching the pre-empt
counter or masking interrupts - this is a really nice property. Most of my headaches have
come from this area.

The limbo work isn't arch specific, that is a property of caches, it should really be core
code behaviour to avoid duplication. MPAM needs that, I expect riscv does too.

Making the CLOSID/RMID allocation behaviour arch specific saves some headaches. MPAM is
particularly different in this area. I don't know what riscv's behaviour is here.



> James: 
> 
> I'm now hoping for some feedback from ARM folks on whether this is a
> useful direction. Is it possible to implement MPAM features on top of
> this base architecture independent layer. If not, am I just missing
> some simple extensions that you will need. Or is there some fundamental
> problem that looks hard/impossible to resolve?


You've got an rdt_l2_cat driver which is really a symlink to rdt_l3_cat built with
different pre-processor defines. It's a good trick, but it only works because you know the
full range of hardware that has been built.
MPAM can't do things like this, it really is a bag of bits, and it isn't until the ACPI
tables are parsed to find where the resources are - and then that hardware is probed -
that we have a clue what was built.
For example, the cache controls could be on any cache in the hierarchy, and the number is
derived from the structure of the PPTT table. You've already got a fair amount of
duplication when multiple struct resctrl_resources are defined. MPAM would have to
allocate and populate these dynamically to avoid building a module for 'L5' ... just in
case someone built that...

The mba_MBps logic is really just software, I see no reason to duplicate it on multiple
architectures. This should be in the core of the filesystem. We already have a mount
option to enable it.

I see the arch/submodules can't influence the domain numbers ... this is a problem as it
hardcodes "you can have L2 or L3", which is something you were trying to avoid. MPAM will
want to use NUMA-IDs for memory-side-caches, (I'd hope this to be a core/portable schema),
as well as IOMMU-IDs for the I/O side of this.
I don't think this is really a problem, as I'd like to add the things I need in this area
as core/portable schema.

Arm's IOMMU has support to label traffic with the equivalent to CLOSID/RMID. My prototype
support for this adds IOMMU-groups to the resctrl tasks file so that the devices can be
moved between control/monitor groups as if they were tasks. I think this would work for
other architectures if they had similar support as IOMMU-groups are an existing concept.


Thanks,

James