Re: [PATCH v2 1/7] vfio/pci: Disable auto-enable of exclusive INTx IRQ

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 11 Mar 2024 08:36:07 +0100
Eric Auger <eric.auger@redhat.com> wrote:

> Hi Alex,
> 
> On 3/9/24 00:05, Alex Williamson wrote:
> > Currently for devices requiring masking at the irqchip for INTx, ie.
> > devices without DisINTx support, the IRQ is enabled in request_irq()
> > and subsequently disabled as necessary to align with the masked status
> > flag.  This presents a window where the interrupt could fire between
> > these events, resulting in the IRQ incrementing the disable depth twice.
> > This would be unrecoverable for a user since the masked flag prevents
> > nested enables through vfio.
> >
> > Instead, invert the logic using IRQF_NO_AUTOEN such that exclusive INTx
> > is never auto-enabled, then unmask as required.
> > Cc: stable@vger.kernel.org
> > Fixes: 89e1f7d4c66d ("vfio: Add PCI device driver")
> > Reviewed-by: Kevin Tian <kevin.tian@intel.com>
> > Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> > ---
> >  drivers/vfio/pci/vfio_pci_intrs.c | 17 ++++++++++-------
> >  1 file changed, 10 insertions(+), 7 deletions(-)
> >
> > diff --git a/drivers/vfio/pci/vfio_pci_intrs.c b/drivers/vfio/pci/vfio_pci_intrs.c
> > index 237beac83809..136101179fcb 100644
> > --- a/drivers/vfio/pci/vfio_pci_intrs.c
> > +++ b/drivers/vfio/pci/vfio_pci_intrs.c
> > @@ -296,8 +296,15 @@ static int vfio_intx_set_signal(struct vfio_pci_core_device *vdev, int fd)
> >  
> >  	ctx->trigger = trigger;
> >  
> > +	/*
> > +	 * Devices without DisINTx support require an exclusive interrupt,
> > +	 * IRQ masking is performed at the IRQ chip.  The masked status is
> > +	 * protected by vdev->irqlock. Setup the IRQ without auto-enable and
> > +	 * unmask as necessary below under lock.  DisINTx is unmodified by
> > +	 * the IRQ configuration and may therefore use auto-enable.  
> If I remember correctly the main reason why the
> 
> vdev->pci_2_3 path is left unchanged is due to the fact the irq may not be exclusive
> and setting IRQF_NO_AUTOEN could be wrong in that case. May be worth to
> precise in the commit msg or here? Besides Reviewed-by: Eric Auger
> <eric.auger@redhat.com> Eric   

IRQF_SHARED and IRQF_NO_AUTOEN are in fact mutually exclusive.  Even if
we could disable auto-enable, the driver sharing the interrupt could
independently enable it.  But really the basis for using IRQF_SHARED is
that we have device level INTx detection and masking.  The comment here
is only to note that request_irq() doesn't gratuitously clear DisINTx,
so the mask state previously applied through config space of the device
is persistent.  Thanks,

Alex
 
> > +	 */
> >  	if (!vdev->pci_2_3)
> > -		irqflags = 0;
> > +		irqflags = IRQF_NO_AUTOEN;
> >  
> >  	ret = request_irq(pdev->irq, vfio_intx_handler,
> >  			  irqflags, ctx->name, vdev);
> > @@ -308,13 +315,9 @@ static int vfio_intx_set_signal(struct vfio_pci_core_device *vdev, int fd)
> >  		return ret;
> >  	}
> >  
> > -	/*
> > -	 * INTx disable will stick across the new irq setup,
> > -	 * disable_irq won't.
> > -	 */
> >  	spin_lock_irqsave(&vdev->irqlock, flags);
> > -	if (!vdev->pci_2_3 && ctx->masked)
> > -		disable_irq_nosync(pdev->irq);
> > +	if (!vdev->pci_2_3 && !ctx->masked)
> > +		enable_irq(pdev->irq);
> >  	spin_unlock_irqrestore(&vdev->irqlock, flags);
> >  
> >  	return 0;  
>