Re: regression: gpiolib: switch the line state notifier to atomic unexpected impact on performance

[David Jander <david@xxxxxxxxxxx>]

On Tue, 11 Mar 2025 12:45:51 +0100
Bartosz Golaszewski <brgl@xxxxxxxx> wrote:

> On Tue, Mar 11, 2025 at 11:01 AM David Jander <david@xxxxxxxxxxx> wrote:
> >
> > On kernel 6.13, after git revert -n fcc8b637c542 time is back to what it was
> > on 6.12.
> >  
> 
> Interestingly: I cannot reproduce it. Obviously gpiofind doesn't exist
> in libgpiod v2 but I'm running gpiodetect with and without reverting
> these changes and am getting roughly the same results: ~0.050s real
> time for 1 up to 4 chips.
> 
> Any idea why that could be? Can you reproduce it with libgpiod v2 (I
> don't know why that wouldn't be the case but worth double checking).

I had libgpiod version 1.6.3 and I have now built libgpiod version 2.1.3.

Here are my findings:

1. time gpiodetect on kernel 6.13 and gpiod 1.6.3:

$ time gpiodetect 
gpiochip0 [GPIOA] (16 lines)
gpiochip1 [GPIOB] (16 lines)
gpiochip10 [GPIOK] (8 lines)
gpiochip11 [GPIOZ] (8 lines)
gpiochip12 [unknown] (22 lines)
gpiochip13 [mcp23s17.0] (16 lines)
gpiochip14 [0-0020] (16 lines)
gpiochip15 [0-0021] (16 lines)
gpiochip2 [GPIOC] (16 lines)
gpiochip3 [GPIOD] (16 lines)
gpiochip4 [GPIOE] (16 lines)
gpiochip5 [GPIOF] (16 lines)
gpiochip6 [GPIOG] (16 lines)
gpiochip7 [GPIOH] (16 lines)
gpiochip8 [GPIOI] (16 lines)
gpiochip9 [GPIOJ] (16 lines)
real    0m 0.19s
user    0m 0.00s
sys     0m 0.01s

2. time gpiodetect on kernel 6.13 and gpiod 2.1.3:

$ time gpiodetect 
gpiochip0 [GPIOA] (16 lines)
gpiochip1 [GPIOB] (16 lines)
gpiochip2 [GPIOC] (16 lines)
gpiochip3 [GPIOD] (16 lines)
gpiochip4 [GPIOE] (16 lines)
gpiochip5 [GPIOF] (16 lines)
gpiochip6 [GPIOG] (16 lines)
gpiochip7 [GPIOH] (16 lines)
gpiochip8 [GPIOI] (16 lines)
gpiochip9 [GPIOJ] (16 lines)
gpiochip10 [GPIOK] (8 lines)
gpiochip11 [GPIOZ] (8 lines)
gpiochip12 [unknown] (22 lines)
gpiochip13 [mcp23s17.0] (16 lines)
gpiochip14 [0-0020] (16 lines)
gpiochip15 [0-0021] (16 lines)
real    0m 0.22s
user    0m 0.00s
sys     0m 0.06s

(note that it became slightly slower from v1 -> v2)

3. time gpiodetect on kernel 6.12 and gpiod 1.6.3:

$ time gpiodetect 
gpiochip0 [GPIOA] (16 lines)
gpiochip1 [GPIOB] (16 lines)
gpiochip10 [GPIOK] (8 lines)
gpiochip11 [GPIOZ] (8 lines)
gpiochip12 [unknown] (22 lines)
gpiochip13 [mcp23s17.0] (16 lines)
gpiochip14 [0-0020] (16 lines)
gpiochip15 [0-0021] (16 lines)
gpiochip2 [GPIOC] (16 lines)
gpiochip3 [GPIOD] (16 lines)
gpiochip4 [GPIOE] (16 lines)
gpiochip5 [GPIOF] (16 lines)
gpiochip6 [GPIOG] (16 lines)
gpiochip7 [GPIOH] (16 lines)
gpiochip8 [GPIOI] (16 lines)
gpiochip9 [GPIOJ] (16 lines)
real    0m 0.03s
user    0m 0.00s
sys     0m 0.01s

4. time gpiodetect on kernel 6.12 and gpiod 2.1.3:

$ time gpiodetect 
gpiochip0 [GPIOA] (16 lines)
gpiochip1 [GPIOB] (16 lines)
gpiochip2 [GPIOC] (16 lines)
gpiochip3 [GPIOD] (16 lines)
gpiochip4 [GPIOE] (16 lines)
gpiochip5 [GPIOF] (16 lines)
gpiochip6 [GPIOG] (16 lines)
gpiochip7 [GPIOH] (16 lines)
gpiochip8 [GPIOI] (16 lines)
gpiochip9 [GPIOJ] (16 lines)
gpiochip10 [GPIOK] (8 lines)
gpiochip11 [GPIOZ] (8 lines)
gpiochip12 [unknown] (22 lines)
gpiochip13 [mcp23s17.0] (16 lines)
gpiochip14 [0-0020] (16 lines)
gpiochip15 [0-0021] (16 lines)
real    0m 0.07s
user    0m 0.00s
sys     0m 0.06s

(roughly same speed difference from v1 -> v2).

Can you describe your platform? Is it a multi-core or single-core CPU? What
RCU implementation does it use? Tree or tiny? If it is multi-core, is there a
difference if you disable all but one core?
Maybe some kernel CONFIG option that makes a difference? I am not an expert in
RCU (in fact I barely know what it does), so maybe I am missing something that
makes this problem go away?

Best regards,

-- 
David Jander