On Thu, Mar 15, 2018 at 02:01:53PM +0100, Sebastian Andrzej Siewior wrote:
> On 2018-03-15 13:53:42 [+0100], Joerg Roedel wrote:
> > On Fri, Feb 23, 2018 at 11:27:31PM +0100, Sebastian Andrzej Siewior wrote:
> > > @@ -4103,10 +4093,26 @@ static int irq_remapping_alloc(struct irq_domain
> > > *domain, unsigned int virq,
> > > return ret;
> > >
> > > if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
> > > - if (get_irq_table(devid, true))
> > > + struct irq_remap_table *table;
> > > + struct amd_iommu *iommu;
> > > +
> > > + table = get_irq_table(devid);
> > > + if (table) {
> > > + if (!table->min_index) {
> > > + /*
> > > + * Keep the first 32 indexes free for IOAPIC
> > > + * interrupts.
> > > + */
> > > + table->min_index = 32;
> > > + iommu = amd_iommu_rlookup_table[devid];
> > > + for (i = 0; i < 32; ++i)
> > > + iommu->irte_ops->set_allocated(table,
> > > i);
> > > + }
> >
> > I think this needs to be protected by the table->lock.
>
> Which part any why? The !->min_index part plus extending(setting to 32)?
In particular the set_allocated part, when get_irq_table() returns the
table is visible to other users as well. I have not checked the
irq-layer locking to be sure that can happen, though.
