Re: [PATCH] PM: QoS: Add support for CPU affinity mask-based CPUs latency QoS

[Christian Loehle <christian.loehle@arm.com>]

On 4/24/25 10:52, Zhongqiu Han wrote:
> Currently, the PM QoS framework supports global CPU latency QoS and
> per-device CPU latency QoS requests. An example of using global CPU
> latency QoS is a commit 2777e73fc154 ("scsi: ufs: core: Add CPU latency
> QoS support for UFS driver") that improved random io performance by 15%
> for ufs on specific device platform.
> 
> However, this prevents all CPUs in the system from entering C states.
> Typically, some threads or drivers know which specific CPUs they are
> interested in. For example, drivers with IRQ affinity only want interrupts
> to wake up and be handled on specific CPUs. Similarly, kernel thread bound
> to specific CPUs through affinity only care about the latency of those
> particular CPUs.
> 
> This patch introduces support for partial CPUs PM QoS using a CPU affinity
> mask, allowing flexible and more precise latency QoS settings for specific
> CPUs. This can help save power, especially on heterogeneous platforms with
> big and little cores, as well as some power-conscious embedded systems for
> example:
> 
>                   driver A       rt kthread B      module C
> QoS cpu mask:       0-3              2-5              6-7
> target latency:     20               30               50
>                     |                |                |
>                     v                v                v
>                     +---------------------------------+
>                     |        PM  QoS  Framework       |
>                     +---------------------------------+
>                     |                |                |
>                     v                v                v
> cpu mask:          0-3            2-3,4-5            6-7
> actual latency:    20             20, 30             50
> 
> Implement this support based on per-device CPU latency PM QoS.
> 
> Signed-off-by: Zhongqiu Han <quic_zhonhan@quicinc.com>

I like the idea!
The interface does need an in-tree user, why not convert the UFS driver?