Re: [PATCH 1/2] PCI: Avoid pointless capability searches

[Ilpo Järvinen]

On Thu, 13 Feb 2025, Bjorn Helgaas wrote:

> On Thu, Feb 13, 2025 at 03:52:05PM +0200, Ilpo Järvinen wrote:
> > On Fri, 7 Feb 2025, Bjorn Helgaas wrote:
> > > Many of the save/restore functions in the pci_save_state() and
> > > pci_restore_state() paths depend on both a PCI capability of the device and
> > > a pci_cap_saved_state structure to hold the configuration data, and they
> > > skip the operation if either is missing.
> > > 
> > > Look for the pci_cap_saved_state first so if we don't have one, we can skip
> > > searching for the device capability, which requires several slow config
> > > space accesses.
> 
> > > +++ b/drivers/pci/vc.c
> > > @@ -355,20 +355,17 @@ int pci_save_vc_state(struct pci_dev *dev)
> > >  	int i;
> > >  
> > >  	for (i = 0; i < ARRAY_SIZE(vc_caps); i++) {
> > > -		int pos, ret;
> > >  		struct pci_cap_saved_state *save_state;
> > > +		int pos, ret;
> > > +
> > > +		save_state = pci_find_saved_ext_cap(dev, vc_caps[i].id);
> > > +		if (!save_state)
> > > +			return -ENOMEM;
> > >  
> > >  		pos = pci_find_ext_capability(dev, vc_caps[i].id);
> > >  		if (!pos)
> > >  			continue;
> > >  
> > > -		save_state = pci_find_saved_ext_cap(dev, vc_caps[i].id);
> > > -		if (!save_state) {
> > > -			pci_err(dev, "%s buffer not found in %s\n",
> > > -				vc_caps[i].name, __func__);
> > > -			return -ENOMEM;
> > > -		}
> > 
> > I think this order change will cause a functional change because 
> > pci_allocate_vc_save_buffers() only allocated for those capabilities that 
> > are exist for dev. Thus, the loop will prematurely exit.
> 
> Oof, thank you for catching this!  I'll drop this for now.
> 
> It would be nice to make pci_save_vc_state() parallel with
> pci_restore_vc_state() (and with most other pci_save_*_state()
> functions) and have it return void.  But pci_save_state() returns the
> pci_save_vc_state() return value, and there are ~20 pci_save_state()
> callers that pay attention to that return value.
> 
> I'm not convinced there's real value in pci_save_state() error
> returns, given that so few callers check it, but it definitely
> requires more analysis before removing it.

Indeed, I also though that -ENOMEM even in the original is questionable.
These are not the real sources of the failure but just secondary effect 
from the failure that occurred earlier in _pci_add_cap_save_buffer().

-- 
 i.

> > >  		ret = pci_vc_do_save_buffer(dev, pos, save_state, true);
> > >  		if (ret) {
> > >  			pci_err(dev, "%s save unsuccessful %s\n",
> > > @@ -392,12 +389,15 @@ void pci_restore_vc_state(struct pci_dev *dev)
> > >  	int i;
> > >  
> > >  	for (i = 0; i < ARRAY_SIZE(vc_caps); i++) {
> > > -		int pos;
> > >  		struct pci_cap_saved_state *save_state;
> > > +		int pos;
> > > +
> > > +		save_state = pci_find_saved_ext_cap(dev, vc_caps[i].id);
> > > +		if (!save_state)
> > > +			continue;
> > >  
> > >  		pos = pci_find_ext_capability(dev, vc_caps[i].id);
> > > -		save_state = pci_find_saved_ext_cap(dev, vc_caps[i].id);
> > > -		if (!save_state || !pos)
> > > +		if (!pos)
> > >  			continue;
> > >  
> > >  		pci_vc_do_save_buffer(dev, pos, save_state, false);