Hi,

On 14/07/16 10:34, Marc Zyngier wrote:

On 13/07/16 02:59, Andre Przywara wrote:

quoted

The LPI pending status for a GICv3 redistributor is held in a table
in (guest) memory. To achieve reasonable performance, we cache the
pending bit in our struct vgic_irq. The initial pending state must be
read from guest memory upon enabling LPIs for this redistributor.
As we can't access the guest memory while we hold the lpi_list spinlock,
we create a snapshot of the LPI list and iterate over that.

Signed-off-by: Andre Przywara <andre.przywara@arm.com>
---
 virt/kvm/arm/vgic/vgic-its.c | 91 ++++++++++++++++++++++++++++++++++++++++++++
 virt/kvm/arm/vgic/vgic.h     |  6 +++
 2 files changed, 97 insertions(+)

diff --git a/virt/kvm/arm/vgic/vgic-its.c b/virt/kvm/arm/vgic/vgic-its.c
index 4fc830c..f400ef1 100644
--- a/virt/kvm/arm/vgic/vgic-its.c
+++ b/virt/kvm/arm/vgic/vgic-its.c

@@ -63,6 +63,90 @@ struct its_itte {
 };
 
 #define CBASER_ADDRESS(x)	((x) & GENMASK_ULL(51, 12))
+#define PENDBASER_ADDRESS(x)	((x) & GENMASK_ULL(51, 16))
+
+/*
+ * Create a snapshot of the current LPI list, so that we can enumerate all
+ * LPIs without holding any lock.
+ * Returns the array length and puts the kmalloc'ed array into intid_ptr.
+ */
+static int vgic_copy_lpi_list(struct kvm *kvm, u32 **intid_ptr)
+{
+	struct vgic_dist *dist = &kvm->arch.vgic;
+	struct vgic_irq *irq;
+	u32 *intids;
+	int irq_count = dist->lpi_list_count, i = 0;
+
+	/*
+	 * We use the current value of the list length, which may change
+	 * after the kmalloc. We don't care, because the guest shouldn't
+	 * change anything while the command handling is still running,
+	 * and in the worst case we would miss a new IRQ, which one wouldn't
+	 * expect to be covered by this command anyway.
+	 */
+	intids = kmalloc_array(irq_count, sizeof(intids[0]), GFP_KERNEL);
+	if (!intids)
+		return -ENOMEM;
+
+	spin_lock(&dist->lpi_list_lock);
+	list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
+		/* We don't need to "get" the IRQ, as we hold the list lock. */
+		intids[i] = irq->intid;
+		if (i++ == irq_count)
+			break;

So I can perfectly rewrite this as:

		if (i == irq_count)
			break;
		i++;

Do you now see the bug and how you are performing out of bound accesses?

Ouch!

quoted

+	}
+	spin_unlock(&dist->lpi_list_lock);
+
+	*intid_ptr = intids;
+	return irq_count;
+}
+
+/*
+ * Scan the whole LPI pending table and sync the pending bit in there
+ * with our own data structures. This relies on the LPI being
+ * mapped before.
+ */
+static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
+{
+	gpa_t pendbase = PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
+	struct vgic_irq *irq;
+	u8 pendmask;
+	int ret = 0;
+	u32 *intids;
+	int nr_irqs, i;
+
+	nr_irqs = vgic_copy_lpi_list(vcpu->kvm, &intids);
+	if (nr_irqs < 0)
+		return nr_irqs;
+
+	for (i = 0; i < nr_irqs; i++) {
+		int byte_offset, bit_nr;
+		int last_byte_offset = -1;

Nice try. But by keeping the last_byte_offset inside the loop, you've
made sure that it is reset to -1 on each iteration. Wall <- head.

Ouch again.

Thanks for catching those! Fixed.

Cheers,
Andre.