Re: [RFC V1 RESEND 5/6] PCI/MSI: Free MSI-X resources by group

[Thomas Gleixner <tglx@linutronix.de>]

[-- Attachment #1: Type: text/plain, Size: 3233 bytes --]

On Tue, 6 Aug 2019, Megha Dey wrote:

> On Sat, 2019-06-29 at 10:08 +0200, Thomas Gleixner wrote:
> > Megha,
> > 
> > On Fri, 21 Jun 2019, Megha Dey wrote:
> > > 
> > > +static int free_msi_irqs_grp(struct pci_dev *dev, int group_id)
> > > +{
> > > 
> > > +
> > > +	for_each_pci_msi_entry(entry, dev) {
> > > +		if (entry->group_id == group_id && entry->irq)
> > > +			for (i = 0; i < entry->nvec_used; i++)
> > > +				BUG_ON(irq_has_action(entry->irq +
> > > i));
> > BUG_ON is wrong here. This can and must be handled gracefully.
> > 
> 
> Hmm, I reused this code from the 'free_msi_irqs' function. I am not
> sure why it is wrong to use BUG_ON here but ok to use it there, please
> let me know.

We are not adding BUG_ON() anymore except for situations where there is
absolutely no way out. Just because there is still older code having
BUG_ON() does not make it any better. Copying it surely is no
justification.

If there is really no way out, then you need to explain it.
 
> > > +static void pci_msix_shutdown_grp(struct pci_dev *dev, int
> > > group_id)
> > > +{
> > > +	struct msi_desc *entry;
> > > +	int grp_present = 0;
> > > +
> > > +	if (pci_dev_is_disconnected(dev)) {
> > > +		dev->msix_enabled = 0;
> > Huch? What's that? I can't figure out why this is needed and of
> > course it
> > completely lacks a comment explaining this. 
> > 
> 
> Again, I have reused this code from the pci_msix_shutdown() function.
> So for the group case, this is not required?

Copy and paste is not an argument, really. Can this happen here? If so,
then please add a comment.

> > > 
> > > +		return;
> > > +	}
> > > +
> > > +	/* Return the device with MSI-X masked as initial states
> > > */
> > > +	for_each_pci_msi_entry(entry, dev) {
> > > +		if (entry->group_id == group_id) {
> > > +			/* Keep cached states to be restored */
> > > +			__pci_msix_desc_mask_irq(entry, 1);
> > > +			grp_present = 1;
> > > +		}
> > > +	}
> > > +
> > > +	if (!grp_present) {
> > > +		pci_err(dev, "Group to be disabled not
> > > present\n");
> > > +		return;
> > So you print an error and silently return
> > 
> 
> This is a void function, hence no error value can be returned. What do
> you think is the right thing to do if someone wants to delete a group
> which is not present?

Well, you made it a void function. 
 
> > > 
> > > +	}
> > > +}
> > > +
> > > +int pci_disable_msix_grp(struct pci_dev *dev, int group_id)
> > > +{
> > > +	int num_vecs;
> > > +
> > > +	if (!pci_msi_enable || !dev)
> > > +		return -EINVAL;
> > > +
> > > +	pci_msix_shutdown_grp(dev, group_id);
> > > +	num_vecs = free_msi_irqs_grp(dev, group_id);
> > just to call in another function which has to do the same group_id
> > lookup
> > muck again.
> 
> Even with the new proposal, we are to have 2 sets of functions: one to
> delete all the msic_desc entries associated with the device, and the
> other to delete those only belonging a 'user specified' group. So we do
> need to pass a group_id to these functions right? Yes, internally the
> deletion would be straightforward with the new approach.

That does not matter. If pci_msix_shutdown_grp() does not find a group, why
proceeding instead of having a proper error return and telling the caller?

Thanks,

	tglx