Re: [PATCH] iommu/s390: Fix race with release_device ops

[Robin Murphy]

On 2022-08-25 12:11, Niklas Schnelle wrote:
> On Thu, 2022-08-25 at 09:22 +0200, Alexander Gordeev wrote:
> > On Wed, Aug 24, 2022 at 04:25:19PM -0400, Matthew Rosato wrote:
> > > > > @@ -90,15 +90,39 @@ static int s390_iommu_attach_device(struct iommu_domain *domain,
> > > > >        struct zpci_dev *zdev = to_zpci_dev(dev);
> > > > >        struct s390_domain_device *domain_device;
> > > > >        unsigned long flags;
> > > > > -    int cc, rc;
> > > > > +    int cc, rc = 0;
> > > > >          if (!zdev)
> > > > >            return -ENODEV;
> > > > >    +    /* First check compatibility */
> > > > > +    spin_lock_irqsave(&s390_domain->list_lock, flags);
> > > > > +    /* First device defines the DMA range limits */
> > > > > +    if (list_empty(&s390_domain->devices)) {
> > > > > +        domain->geometry.aperture_start = zdev->start_dma;
> > > > > +        domain->geometry.aperture_end = zdev->end_dma;
> > > > > +        domain->geometry.force_aperture = true;
> > > > > +    /* Allow only devices with identical DMA range limits */
> > > > > +    } else if (domain->geometry.aperture_start != zdev->start_dma ||
> > > > > +           domain->geometry.aperture_end != zdev->end_dma) {
> > > > > +        rc = -EINVAL;
> > > > > +    }
> > > > > +    spin_unlock_irqrestore(&s390_domain->list_lock, flags);
> > > > > +    if (rc)
> > > > > +        return rc;
> > > > > +
> > > > >        domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL);
> > > > >        if (!domain_device)
> > > > >            return -ENOMEM;
> > > > >    +    /* Leave now if the device has already been released */
> > > > > +    spin_lock_irqsave(&zdev->dma_domain_lock, flags);
> > > > > +    if (!dev_iommu_priv_get(dev)) {
> > > > > +        spin_unlock_irqrestore(&zdev->dma_domain_lock, flags);
> > > > > +        kfree(domain_device);
> > > > > +        return 0;
> > > > > +    }
> > > > > +
> > > > >        if (zdev->dma_table && !zdev->s390_domain) {
> > > > >            cc = zpci_dma_exit_device(zdev);
> > > > >            if (cc) {
> > > >
> > > > Am I wrong? It seems to me that zpci_dma_exit_device here is called with the spin_lock locked but this function zpci_dma_exit_device calls vfree which may sleep.
> > >
> > > Oh, good point, I just enabled lockdep to verify that.
> > >
> > > I think we could just replace this with a mutex instead, it's not a performance path.  I've been running tests successfully today with this patch modified to instead use a mutex for dma_domain_lock.
> >
> > But your original version uses irq-savvy spinlocks.
> > Are there data that need to be protected against interrupts?
> >
> > Thanks!
>
> I think that was a carry over from my original attempt that used the
> zdev->dma_domain_lock in some more places including in interrupt
> context. I think these are gone now so I think Matt is right in his
> version this can be a mutex.

Yes, probe/release/attach/detach should absolutely not be happening from 
atomic/IRQ context. At the very least, the IOMMU core itself needs to 
take the group mutex in those paths.

Cheers,
Robin.