Re: [PATCH] PCI: Take __pci_set_master in do_pci_disable_device

[Bjorn Helgaas <helgaas@kernel.org>]

On Thu, Mar 04, 2021 at 01:40:13PM +0900, Minwoo Im wrote:
> On 21-02-24 23:46:00, Krzysztof Wilczyński wrote:
> > Hi Minwoo,
> > 
> > Sorry for a very late reply!
> > 
> > [...]
> > > > You might need to improve the subject a little - it should be brief but
> > > > still informative.
> > > > 
> > > > > __pci_set_mater() has debug log in there so that it would be better to
> > > > > take this function.  So take __pci_set_master() function rather than
> > > > > open coding it.  This patch didn't move __pci_set_master() to above to
> > > > > avoid churns.
> > > > [...]
> > > > 
> > > > It would be __pci_set_master() in the sentence above.  Also, perhaps
> > > > "use" would be better than "take".  Generally, this commit message might
> > > > need a little improvement to be more clear why are you do doing this.
> > > 
> > > Sure, if we consolidate bus master enable clear functions to a single
> > > one, it would be better to debug and tracing the kernel behaviors.
> > > 
> > > Let me describe this 'why' to the description.
> > 
> > Sounds great!  Thank you!
> > 
> > [...]
> > > > You could use pci_clear_master(), which we export and that internally
> > > > calls __pci_set_master(), so there would be no need to add any forward
> > > > declarations or to move anything around in the file.
> > > 
> > > Moving delcaration to above might be churn, and I agree with your point.
> > 
> > I am sure that when it makes sense, then probably folks would not
> > object, especially since "churn" can be subjective.
> > 
> > > > Having said that, there is a difference between do_pci_disable_device()
> > > > and how __pci_set_master() works - the latter sets the is_busmaster flag
> > > > accordingly on the given device whereas the former does not.  This might
> > > > be of some significance - not sure if we should or should not set this,
> > > > since the do_pci_disable_device() does not do that (perhaps it's on
> > > > purpose or due to some hisoric reasons).
> > > 
> > > Thanks for pointing out this.  I think the difference about
> > > `is_busmaster` flag looks like it should not be cleared in case of power
> > > suspend case:
> > > 
> > > 	# Suspend
> > > 	pci_pm_default_suspend()
> > > 		pci_disable_enabled_device()
> > > 
> > > 	# Resume
> > > 	pci_pm_reenable_device()
> > > 		pci_set_master()  <-- This is based on (is_busmaster)
> > > 
> > > 
> > > Please let me know if I'm missing here, and appreciate pointing that
> > > out.  Maybe I can post v2 patch with add an argument of whether
> > > `is_busmaster` shoud be set inside of the function or not to
> > > __pci_set_master()?
> > [...]
> > 
> > Nothing is ever simple, isn't it? :-)
> > 
> > We definitely need to make sure that PM can keep relying on the
> > is_busmaster flag to restore bus mastering to previous state after the
> > device would resume after being suspended.
> 
> Yes,
> 
> > If we add another boolean argument, then we would need to update the
> > __pci_set_master() only in two other places, aside of using it in the
> > do_pci_disable_device() function, as per (as of 5.11.1 kernel):
> 
> I agree with this approach.  I can try it by adding another bool
> argument to decide whether to update the is_busmaster flag or not inside
> of the __pci_set_master.
> 
> > 
> >   File              Line Content
> >   drivers/pci/pci.c 4308 __pci_set_master(dev, true);
> >   drivers/pci/pci.c 4319 __pci_set_master(dev, false);
> > 
> > This is not all that terrible, provided that we _really_ do want to
> > change this function signature and then add another condition inside.
> > 
> > What do you think?  If you still like the idea, then send second version
> > over with all the other proposed changes.
> 
> Let me prepare the next version of this patch. Thanks!

Can you clarify what the purpose of this patch is?  Is it to fix a
defect, improve debug output, make the code cleaner, etc?

The commit log really just describes *what* the patch does, and I'm
looking for the *why*.

Bjorn