We assumes the iommu_ops were attached to the root region of address
space. This may not be true for all kinds of IOMMU implementation and
especially after commit 3716d5902d74 ("pci: introduce a bus master
container"). So fix this by not assuming as->root has iommu_ops,
instead depending on the regions reported by memory listener through:
- register a memory listener to dma_as
- during region_add, if it's a region of IOMMU, register a specific
IOMMU notifier, and store all notifiers in a list.
- during region_del, compare and delete the IOMMU notifier from the list
This is also a must for making vhost device IOTLB works for all types
of IOMMUs. Note, since we register one notifier during each
.region_add, the IOTLB may be flushed more than one times, this is
suboptimal and could be optimized in the future.
Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Fixes: 3716d5902d74 ("pci: introduce a bus master container")
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
hw/virtio/vhost.c | 84 ++++++++++++++++++++++++++++++++++++-----------
include/hw/virtio/vhost.h | 11 +++++++
2 files changed, 75 insertions(+), 20 deletions(-)
diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
index febe519..613494d 100644
--- a/hw/virtio/vhost.c
+++ b/hw/virtio/vhost.c
@@ -425,10 +425,8 @@ static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
static int vhost_dev_has_iommu(struct vhost_dev *dev)
{
VirtIODevice *vdev = dev->vdev;
- AddressSpace *dma_as = vdev->dma_as;
- return memory_region_is_iommu(dma_as->root) &&
- virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
+ return virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
}
static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
@@ -720,6 +718,63 @@ static void vhost_region_del(MemoryListener *listener,
}
}
+static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
+{
+ struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
+ struct vhost_dev *hdev = iommu->hdev;
+ hwaddr iova = iotlb->iova + iommu->iommu_offset;
+
+ if (hdev->vhost_ops->vhost_invalidate_device_iotlb(hdev, iova,
+ iotlb->addr_mask + 1)) {
+ error_report("Fail to invalidate device iotlb");
+ }
+}
+
+static void vhost_iommu_region_add(MemoryListener *listener,
+ MemoryRegionSection *section)
+{
+ struct vhost_dev *dev = container_of(listener, struct vhost_dev,
+ iommu_listener);
+ struct vhost_iommu *iommu;
+
+ if (!memory_region_is_iommu(section->mr)) {
+ return;
+ }
+
+ iommu = g_malloc0(sizeof(*iommu));
+ iommu->n.notify = vhost_iommu_unmap_notify;
+ iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
+ iommu->mr = section->mr;
+ iommu->iommu_offset = section->offset_within_address_space -
+ section->offset_within_region;
+ iommu->hdev = dev;
+ memory_region_register_iommu_notifier(section->mr, &iommu->n);
+ QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
+ /* TODO: can replay help performance here? */
+}
+
+static void vhost_iommu_region_del(MemoryListener *listener,
+ MemoryRegionSection *section)
+{
+ struct vhost_dev *dev = container_of(listener, struct vhost_dev,
+ iommu_listener);
+ struct vhost_iommu *iommu;
+
+ if (!memory_region_is_iommu(section->mr)) {
+ return;
+ }
+
+ QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
+ if (iommu->mr == section->mr) {
+ memory_region_unregister_iommu_notifier(iommu->mr,
+ &iommu->n);
+ QLIST_REMOVE(iommu, iommu_next);
+ g_free(iommu);
+ break;
+ }
+ }
+}
+
static void vhost_region_nop(MemoryListener *listener,
MemoryRegionSection *section)
{
@@ -1161,17 +1216,6 @@ static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
event_notifier_cleanup(&vq->masked_notifier);
}
-static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
-{
- struct vhost_dev *hdev = container_of(n, struct vhost_dev, n);
-
- if (hdev->vhost_ops->vhost_invalidate_device_iotlb(hdev,
- iotlb->iova,
- iotlb->addr_mask + 1)) {
- error_report("Fail to invalidate device iotlb");
- }
-}
-
int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
VhostBackendType backend_type, uint32_t busyloop_timeout)
{
@@ -1244,8 +1288,10 @@ int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
.priority = 10
};
- hdev->n.notify = vhost_iommu_unmap_notify;
- hdev->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
+ hdev->iommu_listener = (MemoryListener) {
+ .region_add = vhost_iommu_region_add,
+ .region_del = vhost_iommu_region_del,
+ };
if (hdev->migration_blocker == NULL) {
if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
@@ -1455,8 +1501,7 @@ int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
}
if (vhost_dev_has_iommu(hdev)) {
- memory_region_register_iommu_notifier(vdev->dma_as->root,
- &hdev->n);
+ memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
}
r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
@@ -1538,8 +1583,7 @@ void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
if (vhost_dev_has_iommu(hdev)) {
hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
- memory_region_unregister_iommu_notifier(vdev->dma_as->root,
- &hdev->n);
+ memory_listener_unregister(&hdev->iommu_listener);
}
vhost_log_put(hdev, true);
hdev->started = false;
diff --git a/include/hw/virtio/vhost.h b/include/hw/virtio/vhost.h
index 52f633e..a450321 100644
--- a/include/hw/virtio/vhost.h
+++ b/include/hw/virtio/vhost.h
@@ -37,10 +37,20 @@ struct vhost_log {
vhost_log_chunk_t *log;
};
+struct vhost_dev;
+struct vhost_iommu {
+ struct vhost_dev *hdev;
+ MemoryRegion *mr;
+ hwaddr iommu_offset;
+ IOMMUNotifier n;
+ QLIST_ENTRY(vhost_iommu) iommu_next;
+};
+
struct vhost_memory;
struct vhost_dev {
VirtIODevice *vdev;
MemoryListener memory_listener;
+ MemoryListener iommu_listener;
struct vhost_memory *mem;
int n_mem_sections;
MemoryRegionSection *mem_sections;
@@ -64,6 +74,7 @@ struct vhost_dev {
void *opaque;
struct vhost_log *log;
QLIST_ENTRY(vhost_dev) entry;
+ QLIST_HEAD(, vhost_iommu) iommu_list;
IOMMUNotifier n;
};
--
2.7.4
On Wed, Mar 29, 2017 at 12:10:04PM +0800, Jason Wang wrote:
> We assumes the iommu_ops were attached to the root region of address
> space. This may not be true for all kinds of IOMMU implementation and
> especially after commit 3716d5902d74 ("pci: introduce a bus master
> container"). So fix this by not assuming as->root has iommu_ops,
> instead depending on the regions reported by memory listener through:
>
> - register a memory listener to dma_as
> - during region_add, if it's a region of IOMMU, register a specific
> IOMMU notifier, and store all notifiers in a list.
> - during region_del, compare and delete the IOMMU notifier from the list
>
> This is also a must for making vhost device IOTLB works for all types
> of IOMMUs. Note, since we register one notifier during each
> .region_add, the IOTLB may be flushed more than one times, this is
> suboptimal and could be optimized in the future.
I guess I'll need to tune vtd-vfio series as well for this - now vhost
would have the same dup-flush issue just like vfio, and in that series
we need to settle both.
I'm thinking maybe I should post a new version for that series as well
even before 2.9 release for some early review comments.
>
> Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> Fixes: 3716d5902d74 ("pci: introduce a bus master container")
> Cc: Peter Xu <peterx@redhat.com>
> Signed-off-by: Jason Wang <jasowang@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
If this patch will make it for 2.9, I'll drop the corresponding one in
my series, but only keep the other one ("intel_iommu: use the correct
memory region for device IOTLB notification"). Thanks,
> ---
> hw/virtio/vhost.c | 84 ++++++++++++++++++++++++++++++++++++-----------
> include/hw/virtio/vhost.h | 11 +++++++
> 2 files changed, 75 insertions(+), 20 deletions(-)
>
> diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> index febe519..613494d 100644
> --- a/hw/virtio/vhost.c
> +++ b/hw/virtio/vhost.c
> @@ -425,10 +425,8 @@ static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
> static int vhost_dev_has_iommu(struct vhost_dev *dev)
> {
> VirtIODevice *vdev = dev->vdev;
> - AddressSpace *dma_as = vdev->dma_as;
>
> - return memory_region_is_iommu(dma_as->root) &&
> - virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
> + return virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
> }
>
> static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
> @@ -720,6 +718,63 @@ static void vhost_region_del(MemoryListener *listener,
> }
> }
>
> +static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
> +{
> + struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
> + struct vhost_dev *hdev = iommu->hdev;
> + hwaddr iova = iotlb->iova + iommu->iommu_offset;
> +
> + if (hdev->vhost_ops->vhost_invalidate_device_iotlb(hdev, iova,
> + iotlb->addr_mask + 1)) {
> + error_report("Fail to invalidate device iotlb");
> + }
> +}
> +
> +static void vhost_iommu_region_add(MemoryListener *listener,
> + MemoryRegionSection *section)
> +{
> + struct vhost_dev *dev = container_of(listener, struct vhost_dev,
> + iommu_listener);
> + struct vhost_iommu *iommu;
> +
> + if (!memory_region_is_iommu(section->mr)) {
> + return;
> + }
> +
> + iommu = g_malloc0(sizeof(*iommu));
> + iommu->n.notify = vhost_iommu_unmap_notify;
> + iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
> + iommu->mr = section->mr;
> + iommu->iommu_offset = section->offset_within_address_space -
> + section->offset_within_region;
> + iommu->hdev = dev;
> + memory_region_register_iommu_notifier(section->mr, &iommu->n);
> + QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
> + /* TODO: can replay help performance here? */
> +}
> +
> +static void vhost_iommu_region_del(MemoryListener *listener,
> + MemoryRegionSection *section)
> +{
> + struct vhost_dev *dev = container_of(listener, struct vhost_dev,
> + iommu_listener);
> + struct vhost_iommu *iommu;
> +
> + if (!memory_region_is_iommu(section->mr)) {
> + return;
> + }
> +
> + QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
> + if (iommu->mr == section->mr) {
> + memory_region_unregister_iommu_notifier(iommu->mr,
> + &iommu->n);
> + QLIST_REMOVE(iommu, iommu_next);
> + g_free(iommu);
> + break;
> + }
> + }
> +}
> +
> static void vhost_region_nop(MemoryListener *listener,
> MemoryRegionSection *section)
> {
> @@ -1161,17 +1216,6 @@ static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
> event_notifier_cleanup(&vq->masked_notifier);
> }
>
> -static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
> -{
> - struct vhost_dev *hdev = container_of(n, struct vhost_dev, n);
> -
> - if (hdev->vhost_ops->vhost_invalidate_device_iotlb(hdev,
> - iotlb->iova,
> - iotlb->addr_mask + 1)) {
> - error_report("Fail to invalidate device iotlb");
> - }
> -}
> -
> int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
> VhostBackendType backend_type, uint32_t busyloop_timeout)
> {
> @@ -1244,8 +1288,10 @@ int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
> .priority = 10
> };
>
> - hdev->n.notify = vhost_iommu_unmap_notify;
> - hdev->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
> + hdev->iommu_listener = (MemoryListener) {
> + .region_add = vhost_iommu_region_add,
> + .region_del = vhost_iommu_region_del,
> + };
>
> if (hdev->migration_blocker == NULL) {
> if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
> @@ -1455,8 +1501,7 @@ int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
> }
>
> if (vhost_dev_has_iommu(hdev)) {
> - memory_region_register_iommu_notifier(vdev->dma_as->root,
> - &hdev->n);
> + memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
> }
>
> r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
> @@ -1538,8 +1583,7 @@ void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
>
> if (vhost_dev_has_iommu(hdev)) {
> hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
> - memory_region_unregister_iommu_notifier(vdev->dma_as->root,
> - &hdev->n);
> + memory_listener_unregister(&hdev->iommu_listener);
> }
> vhost_log_put(hdev, true);
> hdev->started = false;
> diff --git a/include/hw/virtio/vhost.h b/include/hw/virtio/vhost.h
> index 52f633e..a450321 100644
> --- a/include/hw/virtio/vhost.h
> +++ b/include/hw/virtio/vhost.h
> @@ -37,10 +37,20 @@ struct vhost_log {
> vhost_log_chunk_t *log;
> };
>
> +struct vhost_dev;
> +struct vhost_iommu {
> + struct vhost_dev *hdev;
> + MemoryRegion *mr;
> + hwaddr iommu_offset;
> + IOMMUNotifier n;
> + QLIST_ENTRY(vhost_iommu) iommu_next;
> +};
> +
> struct vhost_memory;
> struct vhost_dev {
> VirtIODevice *vdev;
> MemoryListener memory_listener;
> + MemoryListener iommu_listener;
> struct vhost_memory *mem;
> int n_mem_sections;
> MemoryRegionSection *mem_sections;
> @@ -64,6 +74,7 @@ struct vhost_dev {
> void *opaque;
> struct vhost_log *log;
> QLIST_ENTRY(vhost_dev) entry;
> + QLIST_HEAD(, vhost_iommu) iommu_list;
> IOMMUNotifier n;
> };
>
> --
> 2.7.4
>
-- peterx
On 03/29/2017 06:10 AM, Jason Wang wrote:
> We assumes the iommu_ops were attached to the root region of address
> space. This may not be true for all kinds of IOMMU implementation and
> especially after commit 3716d5902d74 ("pci: introduce a bus master
> container"). So fix this by not assuming as->root has iommu_ops,
> instead depending on the regions reported by memory listener through:
>
> - register a memory listener to dma_as
> - during region_add, if it's a region of IOMMU, register a specific
> IOMMU notifier, and store all notifiers in a list.
> - during region_del, compare and delete the IOMMU notifier from the list
>
> This is also a must for making vhost device IOTLB works for all types
> of IOMMUs. Note, since we register one notifier during each
> .region_add, the IOTLB may be flushed more than one times, this is
> suboptimal and could be optimized in the future.
>
> Reported-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> Fixes: 3716d5902d74 ("pci: introduce a bus master container")
> Cc: Peter Xu <peterx@redhat.com>
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
> hw/virtio/vhost.c | 84 ++++++++++++++++++++++++++++++++++++-----------
> include/hw/virtio/vhost.h | 11 +++++++
> 2 files changed, 75 insertions(+), 20 deletions(-)
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Thanks!
Maxime
© 2016 - 2024 Red Hat, Inc.