On 14.08.19 20:38, Julien Grall wrote:
Hi Oleksandr,
Hi Julien.
On 02/08/2019 17:39, Oleksandr Tyshchenko wrote:
+static int ipmmu_iommu_domain_init(struct domain *d)
+{
+ struct ipmmu_vmsa_xen_domain *xen_domain;
+
+ xen_domain = xzalloc(struct ipmmu_vmsa_xen_domain);
+ if ( !xen_domain )
+ return -ENOMEM;
+
+ spin_lock_init(&xen_domain->lock);
+ INIT_LIST_HEAD(&xen_domain->cache_domains);
+ /*
+ * We don't create Root IPMMU domain here, it will be created on
demand
+ * only, when attaching the first master device to this Xen
domain in
+ * ipmmu_assign_device().
+ * xen_domain->root_domain = NULL;
+ */
+
+ dom_iommu(d)->arch.priv = xen_domain;
While looking at other part of Xen I realized you don't set
IOMMU_FEAT_COHERENT_WALK. Does it mean the IOMMU walker does not
support coherent walk (i.e snooping the cache)?
*AFAIK*, not supported.
Linux driver reports coherent_walk is not supported as well.
Note that when this feature is not set, the p2m code will require to
clean each P2M entry when updated. So if the IPMMU supports coherent
walk, I would strongly suggest to set the flag :).
When playing with non-shared IOMMU in Xen (two years ago), I noticed
that I had forgotten to use clean_dcache after updating a page table
entry. I could face faults when
shattering superpages for example. Once I added it, the faults went away
completely.
So, leave IOMMU_FEAT_COHERENT_WALK in disabled state, but will keep your
suggestion in mind.
Cheers,
--
Regards,
Oleksandr Tyshchenko
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel
