Re: [PATCH v5 05/18] iommu: Pass in reset result to pci_dev_reset_iommu_done()

[Baolu Lu]

On 7/14/26 02:32, Nicolin Chen wrote:
> On Mon, Jul 13, 2026 at 07:48:53PM +0800, Baolu Lu wrote:
> > On 7/3/2026 12:06 PM, Nicolin Chen wrote:
> > > @@ -762,6 +763,9 @@ static void __iommu_group_remove_device(struct device *dev)
> > >    		if (device->dev != dev)
> > >    			continue;
> > > +		/* Must drop the recovery_cnt when removing a blocked device */
> > > +		if (device->blocked && !WARN_ON(group->recovery_cnt == 0))
> > > +			group->recovery_cnt--;
> >
> > I feel that the change above is unrelated to the purpose of this patch.
> > It looks like an independent fix that belongs to a separate patch.
>
> It's actually related; I had this in the commit message:
> "
> This introduces a new situation where a blocked device is being unplugged.
> Decrement the group->recovery_cnt accordingly.
> "

Okay, I see your point now. Previously, reset_prepare and reset_done
were always paired regardless of whether the reset succeeded or failed.

This patch changes that model. The DMA blocking state of a device will
not be released after a failed reset. Thus, it is now possible for a
device to be released by the IOMMU subsystem while its DMA remains
blocked waiting for a successful recovery reset.

> So, it was not needed prior to this patch: it wouldn't fix anything
> if applied before, and it would be a bit odd to apply/fix after.
>
> It's a small three-line change, so I think folding it in is okay?

That makes sense then.

Thanks,
baolu