On 2025/07/11 22:02, Paolo Abeni wrote:
Extend the features configuration space to 128 bits, and allow the
common read/write operation to access all of it.
On migration, save the 128 bit version of the features only if the
upper bits are non zero; after load zero the upper bits if the extended
features were not loaded.
Note that we must clear the proxy-ed features on device reset, otherwise
a guest kernel not supporting extended features booted after an extended
features enabled one could end-up wrongly inheriting extended features.
Signed-off-by: Paolo Abeni <pab...@redhat.com>
---
v1 -> v2:
- use separate VMStateDescription and pre/post load to avoid breaking
migration
- clear proxy features on device reset
---
hw/virtio/virtio-pci.c | 101 +++++++++++++++++++++++++++++----
include/hw/virtio/virtio-pci.h | 6 +-
2 files changed, 96 insertions(+), 11 deletions(-)
diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
index fba2372c93..dc5e7eaf81 100644
--- a/hw/virtio/virtio-pci.c
+++ b/hw/virtio/virtio-pci.c
@@ -108,6 +108,39 @@ static const VMStateDescription
vmstate_virtio_pci_modern_queue_state = {
}
};
+static bool virtio_pci_modern_state_features128_needed(void *opaque)
+{
+ VirtIOPCIProxy *proxy = opaque;
+ uint32_t features = 0;
+ int i;
+
+ for (i = 2; i < ARRAY_SIZE(proxy->guest_features128); ++i) {
+ features |= proxy->guest_features128[i];
+ }
+ return !!features;
"!!" is unnecessary; the implicit cast will clamp the value into true/false.
+}
+
+static int virtio_pci_modern_state_features128_post_load(void *opaque,
+ int version_id)
+{
+ VirtIOPCIProxy *proxy = opaque;
+
+ proxy->extended_features_loaded = true;
+ return 0;
+}
+
+static const VMStateDescription vmstate_virtio_pci_modern_state_features128 = {
+ .name = "virtio_pci/modern_state/features128",
+ .version_id = 1,
+ .minimum_version_id = 1,
+ .post_load = &virtio_pci_modern_state_features128_post_load,
+ .needed = &virtio_pci_modern_state_features128_needed,
+ .fields = (const VMStateField[]) {
+ VMSTATE_UINT32_ARRAY(guest_features128, VirtIOPCIProxy, 4),
+ VMSTATE_END_OF_LIST()
+ }
+};
+
static bool virtio_pci_modern_state_needed(void *opaque)
{
VirtIOPCIProxy *proxy = opaque;
@@ -115,10 +148,40 @@ static bool virtio_pci_modern_state_needed(void *opaque)
return virtio_pci_modern(proxy);
}
+static int virtio_pci_modern_state_pre_load(void *opaque)
+{
+ VirtIOPCIProxy *proxy = opaque;
+
+ proxy->extended_features_loaded = false;
+ return 0;
+}
+
+static int virtio_pci_modern_state_post_load(void *opaque, int version_id)
+{
+ VirtIOPCIProxy *proxy = opaque;
+ int i;
+
+ if (proxy->extended_features_loaded) {
+ return 0;
+ }
+
+ QEMU_BUILD_BUG_ON(offsetof(VirtIOPCIProxy, guest_features[0]) !=
+ offsetof(VirtIOPCIProxy, guest_features128[0]));
+ QEMU_BUILD_BUG_ON(offsetof(VirtIOPCIProxy, guest_features[1]) !=
+ offsetof(VirtIOPCIProxy, guest_features128[1]));
+
+ for (i = 2; i < ARRAY_SIZE(proxy->guest_features128); ++i) {
+ proxy->guest_features128[i] = 0;
+ }
+ return 0;
+}
+
You can expect the device is in the reset state when migrating so expect
guest_features128 is initialized as zero; there are already plenty of
code expecting the reset state.
static const VMStateDescription vmstate_virtio_pci_modern_state_sub = {
.name = "virtio_pci/modern_state",
.version_id = 1,
.minimum_version_id = 1,
+ .pre_load = &virtio_pci_modern_state_pre_load,
+ .post_load = &virtio_pci_modern_state_post_load,
.needed = &virtio_pci_modern_state_needed,
.fields = (const VMStateField[]) {
VMSTATE_UINT32(dfselect, VirtIOPCIProxy),
@@ -128,6 +191,10 @@ static const VMStateDescription
vmstate_virtio_pci_modern_state_sub = {
vmstate_virtio_pci_modern_queue_state,
VirtIOPCIQueue),
VMSTATE_END_OF_LIST()
+ },
+ .subsections = (const VMStateDescription * const []) {
+ &vmstate_virtio_pci_modern_state_features128,
+ NULL
}
};
@@ -1493,19 +1560,22 @@ static uint64_t virtio_pci_common_read(void *opaque, hwaddr addr,
val = proxy->dfselect;
break;
case VIRTIO_PCI_COMMON_DF:
- if (proxy->dfselect <= 1) {
+ if (proxy->dfselect < VIRTIO_FEATURES_WORDS) {
VirtioDeviceClass *vdc = VIRTIO_DEVICE_GET_CLASS(vdev);
- val = (vdev->host_features & ~vdc->legacy_features) >>
- (32 * proxy->dfselect);
+ val = vdev->host_features_array[proxy->dfselect >> 1] >>
+ (32 * (proxy->dfselect & 1));
+ if (proxy->dfselect <= 1) {
+ val &= (~vdc->legacy_features) >> (32 * proxy->dfselect);
+ }
}
break;
case VIRTIO_PCI_COMMON_GFSELECT:
val = proxy->gfselect;
break;
case VIRTIO_PCI_COMMON_GF:
- if (proxy->gfselect < ARRAY_SIZE(proxy->guest_features)) {
- val = proxy->guest_features[proxy->gfselect];
+ if (proxy->gfselect < ARRAY_SIZE(proxy->guest_features128)) {
+ val = proxy->guest_features128[proxy->gfselect];
}
break;
case VIRTIO_PCI_COMMON_MSIX:
@@ -1587,11 +1657,18 @@ static void virtio_pci_common_write(void *opaque,
hwaddr addr,
proxy->gfselect = val;
break;
case VIRTIO_PCI_COMMON_GF:
- if (proxy->gfselect < ARRAY_SIZE(proxy->guest_features)) {
- proxy->guest_features[proxy->gfselect] = val;
- virtio_set_features(vdev,
- (((uint64_t)proxy->guest_features[1]) << 32) |
- proxy->guest_features[0]);
+ if (proxy->gfselect < ARRAY_SIZE(proxy->guest_features128)) {
+ uint64_t features[VIRTIO_FEATURES_DWORDS];
+ int i;
+
+ proxy->guest_features128[proxy->gfselect] = val;
+ virtio_features_clear(features);
+ for (i = 0; i < ARRAY_SIZE(proxy->guest_features128); ++i) {
+ uint64_t cur = proxy->guest_features128[i];
+
+ features[i >> 1] |= cur << ((i & 1) * 32);
+ }
+ virtio_set_features_ex(vdev, features);
}
break;
case VIRTIO_PCI_COMMON_MSIX:
@@ -2310,6 +2387,10 @@ static void virtio_pci_reset(DeviceState *qdev)
virtio_bus_reset(bus);
msix_unuse_all_vectors(&proxy->pci_dev);
+ /* be sure to not carry over any feature across reset */
It's obvious so I don't think the comment makes difference.
+ memset(proxy->guest_features128, 0, sizeof(uint32_t) *
+ ARRAY_SIZE(proxy->guest_features128));
Simpler:
memset(proxy->guest_features128, 0, sizeof(proxy->guest_features128);
+
for (i = 0; i < VIRTIO_QUEUE_MAX; i++) {
proxy->vqs[i].enabled = 0;
proxy->vqs[i].reset = 0;
diff --git a/include/hw/virtio/virtio-pci.h b/include/hw/virtio/virtio-pci.h
index eab5394898..1868e3b106 100644
--- a/include/hw/virtio/virtio-pci.h
+++ b/include/hw/virtio/virtio-pci.h
@@ -151,6 +151,7 @@ struct VirtIOPCIProxy {
uint32_t flags;
bool disable_modern;
bool ignore_backend_features;
+ bool extended_features_loaded;
OnOffAuto disable_legacy;
/* Transitional device id */
uint16_t trans_devid;
@@ -158,7 +159,10 @@ struct VirtIOPCIProxy {
uint32_t nvectors;
uint32_t dfselect;
uint32_t gfselect;
- uint32_t guest_features[2];
+ union {
+ uint32_t guest_features[2];
+ uint32_t guest_features128[4];
+ };
I don't see anything preventing you from directly extending guest_features.
VirtIOPCIQueue vqs[VIRTIO_QUEUE_MAX];
VirtIOIRQFD *vector_irqfd;