Re: [PATCH v3 0/3] nvme: Add sysfs interface for APST configuration management

[Yaxiong Tian]

在 2025/4/3 12:25, Christoph Hellwig 写道:
> On Tue, Apr 01, 2025 at 05:22:06PM +0800, Yaxiong Tian wrote:
> > From: Yaxiong Tian <tianyaxiong@xxxxxxxxxx>
> >
> > This series enhances NVMe APST (Autonomous Power State Transition) support by:
> > 1. Adding warnings for PST table allocation failures
>
> That looks fine.
>
> > 2. Exposing APST tables via sysfs for runtime inspection
> > 3. Providing per-controller sysfs interface for APST configuration
>
> Who is going to use this and how?  We'll need proper tools for that,
> and in general I'd prefer to have a common set of policies for APST
> configurations and not random people coming up with lots of random
> policies.

These two patches don't fundamentally change the APST configuration 
policy, but rather enable users to configure various APST parameters in 
real-time across different devices. As mentioned in commit 
<ebd8a93aa4f5> ("nvme: extend and modify the APST configuration algorithm"):

1）This patch only introduces partial functionality from the Windows 
driver - it doesn't enable dynamic regeneration of the APST table when 
switching between AC power and battery power.

2) Additionally, using default configurations on certain brand SSDs has 
been observed to increase power consumption.

Therefore, when a system contains multiple drives and users require 
different configurations for AC vs battery power scenarios, real-time 
APST configuration updates across different devices become necessary.It 
must be said that configuring these parameters is difficult for average 
users, but they can still use the original default settings. However, 
for advanced users looking to optimize their devices, this provides them 
with the necessary interface.

As for providing suitable tools for users, I suppose every advanced user 
has their own preferences and approaches. For example, one could trace 
device idle time statistics through trace events (e.g., nvme_complete_rq 
→ nvme_setup_cmd) and use that data as a reference to configure 
apst_primary_timeout_ms.

For example，The device idle time distribution under specific operating
conditions is shown below (unit: nanoseconds):
We can set apst_primary_timeout_ms to 32ms, which ensures that over 99% 
of NVMe commands remain unaffected by APST.

@intervals:
[512, 1K)            749 |@@@@@@@@ 
   |
[1K, 2K)            4856 
|@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[2K, 4K)            1002 |@@@@@@@@@@ 
     |
[4K, 8K)             756 |@@@@@@@@ 
     |
[8K, 16K)            413 |@@@@ 
     |
[16K, 32K)            36 | 
     |
[32K, 64K)            14 | 
     |
[64K, 128K)            9 | 
     |
[128K, 256K)          13 | 
     |
[256K, 512K)           8 | 
     |
[512K, 1M)             4 | 
     |
[1M, 2M)               4 | 
     |
[2M, 4M)               3 | 
     |
[4M, 8M)               4 | 
     |
[8M, 16M)              3 | 
     |
[16M, 32M)             0 | 
     |
[32M, 64M)             6 | 
     |
[64M, 128M)            5 | 
     |
[128M, 256M)           5 | 
     |
[256M, 512M)          15 | 
     |
[512M, 1G)             7 | 
     |