Re: [PATCH v3] PCI/PM: Switch D3Hot delay to also use usleep_range

[William McVicker]

On 09/02/2022, Bjorn Helgaas wrote:
> On Wed, Aug 17, 2022 at 11:08:21PM +0000, Will McVicker wrote:
> > From: Sajid Dalvi <sdalvi@xxxxxxxxxx>
> > 
> > Since the PCI spec requires a 10ms D3Hot delay (defined by
> > PCI_PM_D3HOT_WAIT) and a few of the PCI quirks update the D3Hot delay up
> > to 120ms, let's add support for both usleep_range and msleep based on
> > the delay time to improve the delay accuracy.
> > 
> > This patch is based off of a commit from Sajid Dalvi <sdalvi@xxxxxxxxxx>
> > in the Pixel 6 kernel tree [1]. Testing on a Pixel 6, found that the
> > 10ms delay for the Exynos PCIe device was on average delaying for 19ms
> > when the spec requires 10ms. Switching from msleep to uslseep_range
> > therefore decreases the resume time on a Pixel 6 on average by 9ms.
> 
> Add the "PCIe r6.0, sec 5.9" spec reference for the 10ms delay for
> transitions to or from D3hot.
> 
> s/D3Hot/D3hot/ to match other usage (at least in Linux; the spec does
> use "D3Hot")
> 
> s/uslseep_range/usleep_range/
> 
> Add "()" after function names.

Thanks for the suggestions! I'll update these in the next patchset.

> 
> In the subject, "Switch ... to *also* use usleep_range": what does the
> "also" mean?  It sounds like it's referring to some other place where
> we also use usleep_range()?

I was intending to mean usleep_range() will be used when the the delay <=20 ms,
else msleep() will be used. If we drop msleep() altogether as suggested in your
comments below, then I can change this to be: "Switch D3hot delay to use
usleep_range."

> 
> > [1] https://android.googlesource.com/kernel/gs/+/18a8cad68d8e6d50f339a716a18295e6d987cee3
> > 
> > Signed-off-by: Sajid Dalvi <sdalvi@xxxxxxxxxx>
> > Signed-off-by: Will McVicker <willmcvicker@xxxxxxxxxx>
> > ---
> >  drivers/pci/pci.c | 16 +++++++++++-----
> >  1 file changed, 11 insertions(+), 5 deletions(-)
> > 
> > v3:
> >  * Use DIV_ROUND_CLOSEST instead of bit manipulation.
> >  * Minor refactor to use max() where relavant.
> > 
> > v2:
> >  * Update to use 20-25% upper bound
> >  * Update to use usleep_range() for <=20ms, else use msleep()
> > 
> > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> > index 95bc329e74c0..cfa8386314f2 100644
> > --- a/drivers/pci/pci.c
> > +++ b/drivers/pci/pci.c
> > @@ -66,13 +66,19 @@ struct pci_pme_device {
> >  
> >  static void pci_dev_d3_sleep(struct pci_dev *dev)
> >  {
> > -	unsigned int delay = dev->d3hot_delay;
> > +	unsigned int delay_ms = max(dev->d3hot_delay, pci_pm_d3hot_delay);
> >  
> > -	if (delay < pci_pm_d3hot_delay)
> > -		delay = pci_pm_d3hot_delay;
> > +	if (delay_ms) {
> > +		if (delay_ms <= 20) {
> > +			/* Use a 20% upper bound with 1ms minimum */
> > +			unsigned int upper = max(DIV_ROUND_CLOSEST(delay_ms, 5), 1U);
> >  
> > -	if (delay)
> > -		msleep(delay);
> > +			usleep_range(delay_ms * USEC_PER_MSEC,
> > +				     (delay_ms + upper) * USEC_PER_MSEC);
> > +		} else {
> > +			msleep(delay_ms);
> 
> I hate the fact that we have to check for those ancient Intel chips at
> all, but having to read through the usleep_range() vs msleep() thing
> is just painful.  
> 
> fsleep() would be much simpler, but I am sympathetic that the fsleep()
> range of 10-20ms probably wouldn't get the benefit you want.
> 
> I wish Documentation/timers/timers-howto.rst included a reason why we
> should use msleep() instead of usleep_range() for longer sleeps.  Is
> there a reason not to do this:

I'm not familiar with the reasons behind these two beside what the
documentation states. I don't know what happens when you use usleep_range()
with a delay >20ms.

> 
>    static void pci_dev_d3_sleep(struct pci_dev *dev)
>    {
>         unsigned int delay_ms = max(dev->d3hot_delay, pci_pm_d3hot_delay);
>         unsigned int upper;
> 
>         if (delay_ms) {
>                 /* 20% upper bound, 1ms minimum */
>                 upper = max(DIV_ROUND_CLOSEST(delay_ms, 5), 1U)
>                 usleep_range(delay_ms * USEC_PER_MSEC,
>                              (delay_ms + upper) * USEC_PER_MSEC);
>         }
>    }
> 
> Since the Intel quirk is for 120ms, a 20% upper bound would make the
> range 120-144ms.  Would that be a problem?  Those chips are ancient;
> the list is untouched since it was added in 2006.  The point of
> usleep_range() is to allow the scheduler to coalesce the wakeup with
> other events, so it seems unlikely we'd ever wait the whole 144ms.  I
> vote for optimizing the readability over sleep/resume time for
> already-broken chips.

I'm totally fine with this, but I don't really know what the impact would be to
those old Intel chips.

> 
> Bjorn

Thanks,
Will