Message-ID: <aKc7D78owL_op3Ei@google.com>
Date: Thu, 21 Aug 2025 08:28:15 -0700
From: Brian Norris <briannorris@...omium.org>
To: Ethan Zhao <etzhao1900@...il.com>
Cc: Bjorn Helgaas <bhelgaas@...gle.com>, linux-kernel@...r.kernel.org,
	linux-pci@...r.kernel.org, stable@...r.kernel.org
Subject: Re: [PATCH] PCI/sysfs: Ensure devices are powered for config reads

Hi Ethan,

Note: I'm having a hard time reading your emails sometimes, because you
aren't really adding in appropriate newlines that separate your reply
from quoted text. So your own sentences just run together with parts of
my sentences at times. I've tried to resolve this as best I can.

On Thu, Aug 21, 2025 at 08:41:28PM +0800, Ethan Zhao wrote:
> 
> 
> On 8/21/2025 10:56 AM, Brian Norris wrote:
> > On Thu, Aug 21, 2025 at 08:54:52AM +0800, Ethan Zhao wrote:
> > > On 8/21/2025 1:26 AM, Brian Norris wrote:
> > > > From: Brian Norris <briannorris@...gle.com>
> > > > 
> > > > max_link_speed, max_link_width, current_link_speed, current_link_width,
> > > > secondary_bus_number, and subordinate_bus_number all access config
> > > > registers, but they don't check the runtime PM state. If the device is
> > > > in D3cold, we may see -EINVAL or even bogus values.
> > > My understanding, if your device is in D3cold, returning of -EINVAL is
> > > the right behavior.
> > 
> > That's not the guaranteed result though. Some hosts don't properly
> > return PCIBIOS_DEVICE_NOT_FOUND, for one. But also, it's racy -- because
> > we don't even try to hold a pm_runtime reference, the device could
> > possibly enter D3cold while we're in the middle of reading from it. If
> > you're lucky, that'll get you a completion timeout and an all-1's
> > result, and we'll return a garbage result.
> > 
> > So if we want to purposely not resume the device and retain "I can't
> > give you what you asked for" behavior, we'd at least need a
> > pm_runtime_get_noresume() or similar.
> I understand you just want the stable result of these caps,

Yes, I'd like a valid result, not EINVAL. Why would I check this file if
I didn't want the result?

> meanwhile
> you don't want the side effect either.

Personally, I think side effect is completely fine. Or, it's just as
fine as it is for the 'config' attribute or for 'resource_N_size'
attributes that already do the same.

> > > > Wrap these access in pci_config_pm_runtime_{get,put}() like most of the
> > > > rest of the similar sysfs attributes.
> > > > 
> > > > Fixes: 56c1af4606f0 ("PCI: Add sysfs max_link_speed/width, current_link_speed/width, etc")
> > > > Cc: stable@...r.kernel.org
> > > > Signed-off-by: Brian Norris <briannorris@...gle.com>
> > > > Signed-off-by: Brian Norris <briannorris@...omium.org>
> > > > ---
> > > > 
> > > >    drivers/pci/pci-sysfs.c | 32 +++++++++++++++++++++++++++++---
> > > >    1 file changed, 29 insertions(+), 3 deletions(-)
> > > > 
> > > > diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
> > > > index 5eea14c1f7f5..160df897dc5e 100644
> > > > --- a/drivers/pci/pci-sysfs.c
> > > > +++ b/drivers/pci/pci-sysfs.c
> > > > @@ -191,9 +191,16 @@ static ssize_t max_link_speed_show(struct device *dev,
> > > >    				   struct device_attribute *attr, char *buf)
> > > >    {
> > > >    	struct pci_dev *pdev = to_pci_dev(dev);
> > > > +	ssize_t ret;
> > > > +
> > > > +	pci_config_pm_runtime_get(pdev);
> > > This function would potentially change the power state of device,
> > > that would be a complex process, beyond the meaning of
> > > max_link_speed_show(), given the semantics of these functions (
> > > max_link_speed_show()/max_link_width_show()/current_link_speed_show()/
> > > ....),
> > > this cannot be done !
> > 
> > What makes this different than the 'config' attribute (i.e., "read
> > config register")? Why shouldn't that just return -EINVAL? I don't
> > really buy your reasoning -- "it's a complex process" is not a reason
> It is a reason to know there is side effect to be taken into account.

OK, agreed, there's a side effect. I don't think you've convinced me the
side effect is bad though.

> > not
> > to do something. The user asked for the link speed; why not give it?
> > If the user wanted to know if the device was powered, they could check
> > the 'power_state' attribute instead.
> > 
> > (Side note: these attributes don't show up anywhere in Documentation/,
> > so it's also a bit hard to declare "best" semantics for them.)
> > 
> > To flip this question around a bit: if I have a system that aggressively
> > suspends devices when there's no recent activity, how am I supposed to
> > check what the link speed is? Probabilistically hammer the file while
> > hoping some other activity wakes the device, so I can find the small
> > windows of time where it's RPM_ACTIVE? Disable runtime_pm for the device
> > while I check?
> Hold a PM reference by pci_config_pm_runtime_get() and then write some
> data to the PCIe config space, no objection.
> 
> To know about the linkspeed etc capabilities/not status, how about
> creating a cached version of these caps, no need to change their
> power state.

For static values like the "max" attributes, maybe that's fine.

But Linux is not always the one changing the link speed. I've seen PCI
devices that autonomously request link-speed changes, and AFAICT, the
only way we'd know in host software is to go reread the config
registers. So caching just produces cache invalidation problems.

> If there is aggressive power saving requirement, and the polling
> of these caps will make up wakeup/poweron bugs.

If you're worried about wakeup frequency, I think that's a matter of
user space / system administraction to decide -- if it doesn't want to
potentially wake up the link, it shouldn't be poking at config-based
sysfs attributes.

Brian

Powered by blists - more mailing lists