virtio-net: count VIRTIO_NET_F_MAC when calculating config_len
PATCH 1/1 virtio-net: count VIRTIO_NET_F_MAC when calculating config_len
Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.
Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address.
--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)
hw/net/virtio-net.c +2 -1
@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);
Signature · 1 line
1.7.1
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0
hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;
This would be cleaner:
host_features |= (1 << VIRTIO_NET_F_MAC);
no need for a comment then.
Quoted 5 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Quoted 29 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 3 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
Sure, will send V2.
Quoted 5 lines
Quoted 5 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Quoted 32 lines · Jason Wang
Quoted 31 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
host_features |= (1 << VIRTIO_NET_F_MAC);
Quoted 3 lines
Quoted 1 line
no need for a comment then.
Sure, will send V2.
Maybe ass assert(config_size) in core just to make sure other devices do not have this bug.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Quoted 33 lines · Michael S. Tsirkin
Quoted 32 lines · Jason Wang
Quoted 31 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
host_features |= (1 << VIRTIO_NET_F_MAC);
Quoted 4 lines
Quoted 2 lines
Quoted 1 line
no need for a comment then.
Sure, will send V2.
Maybe ass
assert(config_size)in core just to make sure other devices do not have this bug.
Prepare a patch for this, will send soon.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Quoted 30 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why:
static int virtio_device_init(DeviceState *qdev)
{
VirtIODevice *vdev = VIRTIO_DEVICE(qdev);
VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev);
assert(k->init != NULL);
if (k->init(vdev) < 0) {
return -1;
}
virtio_bus_plug_device(vdev);
return 0;
}virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device().
After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so:
static int virtio_device_init(DeviceState *qdev)
{
VirtIODevice *vdev = VIRTIO_DEVICE(qdev);
VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev);
assert(k->init != NULL);
if (k->init(vdev) < 0) {
return -1;
}
virtio_bus_plug_device(vdev);
+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) {
+ return -1;
+ }
return 0;
}This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size.
Quoted 6 lines
Quoted 5 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]
Quoted 73 lines · Jesse Larrew
Quoted 29 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1268
Quoted 6 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.
Good point, I didn't saw that.
but this was not the case with commit 14f9b664 no?
Quoted 13 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?
I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon.
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1
Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.
Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET,
n->config_size);
is called in virtio_net_device_init (k->init).
Is there a way to resize the config after that?
Quoted 12 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]
Quoted 84 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.
Quoted 9 lines
Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
Quoted 13 lines
Quoted 12 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]
Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Quoted 90 lines · Jesse Larrew
Quoted 89 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);
Quoted 7 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
So we need almost all of virtio_bus_plug_device before init, and then after init send the signal:
if (klass->device_plugged != NULL) {
klass->device_plugged(qbus->parent);
}Quoted 22 lines
Quoted 13 lines
Quoted 12 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }
Seems the interesting part is in virtio_pci_device_plugged at the end:
proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY;
proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;
proxy->host_features = virtio_bus_get_vdev_features(bus,proxy->host_features);
This is and was called after set_config_size, isn't that the issue?
Quoted 21 lines
Quoted 21 lines
Quoted 13 lines
Quoted 13 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Quoted 108 lines · KONRAD Frédéric
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }
Quoted 8 lines
Seems the interesting part is in
virtio_pci_device_pluggedat the end:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);This is and was called after
set_config_size, isn't that the issue?
Basically devices expected everything to be setup at the point where init is called.
We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data.
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 13 lines
Quoted 13 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Quoted 116 lines · Michael S. Tsirkin
Quoted 115 lines · KONRAD Frédéric
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }Seems the interesting part is in
virtio_pci_device_pluggedat the end:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);
Quoted 29 lines
Quoted 1 line
This is and was called after
set_config_size, isn't that the issue?Basically devices expected everything to be setup at the point where init is called.
We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data.Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 13 lines
Quoted 13 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value).
---
hw/virtio/virtio-bus.c +9 -3
@@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0)#endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device (VirtIODevice *vdev) +void virtio_bus_plug_device (VirtIODevice *vdev) {DeviceState *qdev =DEVICE (vdev);BusState *qbus =BUS (qdev_get_parent_bus (qdev));
@@ -64,12 +64,18 @@ int virtio_bus_plug_device(VirtIODevice *vdev) bus->bindings.set_host_notifier = klass->set_host_notifier; bus->bindings.vmstate_change = klass->vmstate_change;virtio_bind_device (bus->vdev, &bus->bindings, qbus->parent); +} + +void virtio_bus_plugged (VirtIODevice *vdev) +{ +DeviceState *qdev =DEVICE (vdev); +BusState *qbus =BUS (qdev_get_parent_bus (qdev)); +VirtioBusState *bus =VIRTIO_BUS (qbus); +VirtioBusClass *klass =VIRTIO_BUS_GET_CLASS (bus);if (klass->device_plugged !=NULL ) { klass->device_plugged (qbus->parent); } - -return 0 ; }/* Reset the virtio_bus */
hw/virtio/virtio.c +6 -1
@@ -1088,11 +1088,16 @@ static int virtio_device_init(DeviceState *qdev) {VirtIODevice *vdev =VIRTIO_DEVICE (qdev);VirtioDeviceClass *k =VIRTIO_DEVICE_GET_CLASS (qdev); + +virtio_bus_plug_device (vdev); +assert (k->init !=NULL );if (k->init (vdev) <0 ) {return -1 ; } -virtio_bus_plug_device (vdev); + +virtio_bus_plugged (vdev); +return 0 ; }
include/hw/virtio/virtio-bus.h +2 -1
@@ -76,7 +76,8 @@ struct VirtioBusState {VirtIOBindings bindings; }; -int virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plugged (VirtIODevice *vdev);void virtio_bus_reset (VirtioBusState *bus);void virtio_bus_destroy_device (VirtioBusState *bus);/* Get the device id of the plugged device. */
Quoted 212 lines · Michael S. Tsirkin
Quoted 115 lines · Michael S. Tsirkin
Quoted 114 lines · KONRAD Frédéric
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }Seems the interesting part is in
virtio_pci_device_pluggedat the end:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);Quoted 26 lines
Quoted 1 line
This is and was called after
set_config_size, isn't that the issue?Basically devices expected everything to be setup at the point where init is called.
We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data.Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
Quoted 12 lines
Quoted 12 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value).
Signed-off-by Michael S. Tsirkin <[email protected]>---
hw/virtio/virtio-bus.c +9
-3@@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0)#endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device (VirtIODevice *vdev) +void virtio_bus_plug_device (VirtIODevice *vdev) {DeviceState *qdev =DEVICE (vdev);BusState *qbus =BUS (qdev_get_parent_bus (qdev));@@ -64,12 +64,18 @@ int virtio_bus_plug_device(VirtIODevice *vdev) bus->bindings.set_host_notifier = klass->set_host_notifier; bus->bindings.vmstate_change = klass->vmstate_change;virtio_bind_device (bus->vdev, &bus->bindings, qbus->parent); +} + +void virtio_bus_plugged (VirtIODevice *vdev) +{ +DeviceState *qdev =DEVICE (vdev); +BusState *qbus =BUS (qdev_get_parent_bus (qdev)); +VirtioBusState *bus =VIRTIO_BUS (qbus); +VirtioBusClass *klass =VIRTIO_BUS_GET_CLASS (bus);if (klass->device_plugged !=NULL ) { klass->device_plugged (qbus->parent); } - -return 0 ; }/* Reset the virtio_bus */ hw/virtio/virtio.c +6
-1@@ -1088,11 +1088,16 @@ static int virtio_device_init(DeviceState *qdev) {VirtIODevice *vdev =VIRTIO_DEVICE (qdev);VirtioDeviceClass *k =VIRTIO_DEVICE_GET_CLASS (qdev); + +virtio_bus_plug_device (vdev); +assert (k->init !=NULL );if (k->init (vdev) <0 ) {return -1 ; } -virtio_bus_plug_device (vdev); + +virtio_bus_plugged (vdev); +return 0 ; }include/hw/virtio/virtio-bus.h +0
-0
include/hw/virtio/virtio-bus.h struct VirtioBusState { · line 83
Quoted 9 lines
include/hw/virtio/virtio-bus.h +2
-1@@ -76,7 +76,8 @@ struct VirtioBusState {VirtIOBindings bindings; }; -int virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plugged (VirtIODevice *vdev);void virtio_bus_reset (VirtioBusState *bus);void virtio_bus_destroy_device (VirtioBusState *bus);/* Get the device id of the plugged device. */
Which tree are you using? Is it up to date?
Quoted 144 lines · KONRAD Frédéric
Quoted 143 lines · Michael S. Tsirkin
Quoted 141 lines · Michael S. Tsirkin
Quoted 107 lines · KONRAD Frédéric
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 21 lines · Jason Wang
Commit
14f9b664(hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets theVIRTIO_NET_F_MACwere always set for qemu later. This will lead a zero config len for virtio-net device when bothVIRTIO_NET_F_STATUSandVIRTIO_NET_F_MQwere disabled form command line. Then qemu will crash when user tries to read the config of virtio-net.Fix this by counting
VIRTIO_NET_F_MACand make sure the config at least contains the mac address.Cc Jesse Larrew <[email protected]>Signed-off-by Jason Wang <[email protected]>--- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)hw/net/virtio-net.c +0
-0hw/net/virtio-net.c static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, · line 1267
Quoted 5 lines
hw/net/virtio-net.c +2
-1@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,void virtio_net_set_config_size (VirtIONet *n,uint32_t host_features) { -int i, config_size =0 ; +/* VIRTIO_NET_F_MAC can't be disabled from qemu side */ +int i, config_size = feature_sizes[0 ].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }Quoted 8 lines
Seems the interesting part is in
virtio_pci_device_pluggedat the end:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);This is and was called after
set_config_size, isn't that the issue?Basically devices expected everything to be setup at the point where init is called.
We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data.Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
Quoted 12 lines
Quoted 12 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
hw/net/virtio-net.c +0
-0for (i =0 ; feature_sizes[i].flags !=0 ; i++) {if (host_features & feature_sizes[i].flags) { config_size =MAX (feature_sizes[i].end, config_size);Signature · 1 line
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]
Quoted 6 lines
Quoted 5 lines
Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value).
Signed-off-by Michael S. Tsirkin <[email protected]>Which tree are you using? Is it up to date?
Heh, more changes came in. So now, it's even simpler:
---
hw/virtio/virtio-bus.c +1 -3
@@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0)#endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device (VirtIODevice *vdev) +void virtio_bus_plug_device (VirtIODevice *vdev) {DeviceState *qdev =DEVICE (vdev);BusState *qbus =BUS (qdev_get_parent_bus (qdev));
@@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev)if (klass->device_plugged !=NULL ) { klass->device_plugged (qbus->parent); } - -return 0 ; }/* Reset the virtio_bus */
hw/virtio/virtio.c +1 -1
@@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) {VirtIODevice *vdev =VIRTIO_DEVICE (qdev);VirtioDeviceClass *k =VIRTIO_DEVICE_GET_CLASS (qdev); +virtio_bus_plug_device (vdev);assert (k->init !=NULL );if (k->init (vdev) <0 ) {return -1 ; } -virtio_bus_plug_device (vdev);return 0 ; }
include/hw/virtio/virtio-bus.h +1 -1
@@ -72,7 +72,7 @@ struct VirtioBusState {VirtIODevice *vdev; }; -int virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plug_device (VirtIODevice *vdev);void virtio_bus_reset (VirtioBusState *bus);void virtio_bus_destroy_device (VirtioBusState *bus);/* Get the device id of the plugged device. */
Quoted 183 lines · Michael S. Tsirkin
Quoted 142 lines · KONRAD Frédéric
Quoted 141 lines · Michael S. Tsirkin
Quoted 140 lines · Michael S. Tsirkin
Quoted 107 lines · KONRAD Frédéric
Quoted 100 lines · Michael S. Tsirkin
Quoted 89 lines · Jesse Larrew
Quoted 88 lines · KONRAD Frédéric
Quoted 74 lines · Jesse Larrew
Quoted 73 lines · KONRAD Frédéric
Quoted 72 lines · Jesse Larrew
Quoted 28 lines · Michael S. Tsirkin
Quoted 26 lines · Jason Wang
Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to
calculate config size based on the host features. But it forgets the
VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config
len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were
disabled form command line. Then qemu will crash when user tries to read the
config of virtio-net.Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains
the mac address.Cc: Jesse Larrew <[email protected]>
Signed-off-by: Jason Wang <[email protected]>
---
hw/net/virtio-net.c | 3 ++-
1 files changed, 2 insertions(+), 1 deletions(-)diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
index 70c8fce..33a70ef 100644
--- a/hw/net/virtio-net.c
+++ b/hw/net/virtio-net.c
@@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx,
void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features)
{
- int i, config_size = 0;
+ /* VIRTIO_NET_F_MAC can't be disabled from qemu side */
+ int i, config_size = feature_sizes[0].end;This would be cleaner:
Quoted 4 lines
host_features |= (1 << VIRTIO_NET_F_MAC);no need for a comment then.
It seems to me that the real problem here is that
host_featuresisn't properly populated beforevirtio_net_set_config_size()is called. Looking atvirtio_device_init(), we can see why:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; }
virtio_net_set_config_size()is currently being called as part of thek->initcall, but thehost_featuresaren't properly setup until the device is plugged into the bus usingvirtio_bus_plug_device().After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a
calculate_config_size()method that can be called at the appropriate time during device initialization like so:static int
virtio_device_init(DeviceState *qdev){ VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev);(quoted diff) +3
-0+ if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0;}
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
This would ensure that
host_featurescontains the proper data before any devices try to make use of it to calculate the config size.Good point, I didn't saw that.
but this was not the case with commit
14f9b664no?I suspect this bug was present in
14f9b664as well. We just hadn't triggered it yet. I'll confirm this afternoon.Ok, I think there is a problem with your patch:
virtio_init(VIRTIO_DEVICE(n), "virtio-net",
VIRTIO_ID_NET,
n->config_size);Quoted 6 lines
Quoted 4 lines
is called in
virtio_net_device_init(k->init).Is there a way to resize the config after that?
Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)?
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
Quoted 6 lines
So we need almost all of
virtio_bus_plug_devicebefore init, and then after init send the signal:if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); }Quoted 8 lines
Seems the interesting part is in
virtio_pci_device_pluggedat the end:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);This is and was called after
set_config_size, isn't that the issue?Basically devices expected everything to be setup at the point where init is called.
We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data.Quoted 18 lines
Quoted 18 lines
Quoted 18 lines
Quoted 11 lines
Quoted 11 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
for (i = 0; feature_sizes[i].flags != 0; i++) {
if (host_features & feature_sizes[i].flags) {
config_size = MAX(feature_sizes[i].end, config_size);
--
1.7.1Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
[email protected]Signature · 5 lines
Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected]Quoted 6 lines
Quoted 5 lines
Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value).
Signed-off-by Michael S. Tsirkin <[email protected]>Which tree are you using? Is it up to date?
Heh, more changes came in. So now, it's even simpler:
Signed-off-by Michael S. Tsirkin <[email protected]>---
hw/virtio/virtio-bus.c +1
-3@@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0)#endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device (VirtIODevice *vdev) +void virtio_bus_plug_device (VirtIODevice *vdev) {DeviceState *qdev =DEVICE (vdev);BusState *qbus =BUS (qdev_get_parent_bus (qdev));@@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev)if (klass->device_plugged !=NULL ) { klass->device_plugged (qbus->parent); } - -return 0 ; }/* Reset the virtio_bus */ hw/virtio/virtio.c +0
-0@@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) {
hw/virtio/virtio.c static int virtio_device_init(DeviceState *qdev) · line 1100
Quoted 10 lines
hw/virtio/virtio.c +1
-1VirtIODevice *vdev =VIRTIO_DEVICE (qdev);VirtioDeviceClass *k =VIRTIO_DEVICE_GET_CLASS (qdev); +virtio_bus_plug_device (vdev);assert (k->init !=NULL );if (k->init (vdev) <0 ) {return -1 ; } -virtio_bus_plug_device (vdev);return 0 ; }
Not sure this is what you want to do.
The device would be plugged before it is inited :/.
You said:
"
This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue.
So we need almost all of virtio_bus_plug_device before init, and then after init send the signal:
if (klass->device_plugged != NULL) {
klass->device_plugged(qbus->parent);
}"
And as virtio_bus_plugged is only this chunk of code:
if (klass->device_plugged != NULL) {
klass->device_plugged(qbus->parent);
}It is the right place now.
Quoted 14 lines
include/hw/virtio/virtio-bus.h +1
-1@@ -72,7 +72,7 @@ struct VirtioBusState {VirtIODevice *vdev; }; -int virtio_bus_plug_device (VirtIODevice *vdev); +void virtio_bus_plug_device (VirtIODevice *vdev);void virtio_bus_reset (VirtioBusState *bus);void virtio_bus_destroy_device (VirtioBusState *bus);/* Get the device id of the plugged device. */
Quoted 224 lines · KONRAD Frédéric
On 29/04/2013 19:01, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; }
Quoted 2 lines
Not sure this is what you want to do.
The device would be plugged before it is inited :/.
I think this is exacly what we want to do.
In fact, this is what other buses do, because
devices simply can't init properly if they
do not know on which bus they reside.
E.g. with pci:
do_pci_register_device (adds device on bus) init
We can add an analog of hotplug bus callback
if bus wants to get notified after device initialization. I don't see a need for this though.
Do you?
Signature · 1 line
MST
Quoted 232 lines · Michael S. Tsirkin
Quoted 224 lines · KONRAD Frédéric
On 29/04/2013 19:01, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; }Quoted 2 lines
Not sure this is what you want to do.
The device would be plugged before it is inited :/.I think this is exacly what we want to do.
In fact, this is what other buses do, because
devices simply can't init properly if they
do not know on which bus they reside.
Quoted 11 lines
E.g. with pci:
do_pci_register_device(adds device on bus) initWe can add an analog of hotplug bus callback
if bus wants to get notified after device initialization. I don't see a need for this though.
Do you?
No, as we don't want to allow virtio-device hotplug.
but look at the plugged callback in virtio-pci:
pci_set_word(config + PCI_SUBSYSTEM_ID, virtio_bus_get_vdev_id(bus));
proxy->host_features = virtio_bus_get_vdev_features(bus,proxy->host_features);
are impossible before the virtio-net init.
Quoted 246 lines · KONRAD Frédéric
Quoted 232 lines · Michael S. Tsirkin
Quoted 224 lines · KONRAD Frédéric
On 29/04/2013 19:01, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; }Quoted 2 lines
Not sure this is what you want to do.
The device would be plugged before it is inited :/.I think this is exacly what we want to do.
In fact, this is what other buses do, because
devices simply can't init properly if they
do not know on which bus they reside.Quoted 11 lines
E.g. with pci:
do_pci_register_device(adds device on bus) initWe can add an analog of hotplug bus callback
if bus wants to get notified after device initialization. I don't see a need for this though.
Do you?No, as we don't want to allow virtio-device hotplug.
Quoted 8 lines
but look at the plugged callback in virtio-pci:
pci_set_word(config + PCI_SUBSYSTEM_ID, virtio_bus_get_vdev_id(bus)); proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);are impossible before the virtio-net init.
At this point I have to admit I don't understand what's going on any more.
virtio_net_instance_init sets config size to
some value, then virtio_net_set_config_size overrides it... Help!
I am guessing it's this hack that backfires somehow.
Signature · 1 line
MST
Quoted 256 lines · Michael S. Tsirkin
Quoted 246 lines · KONRAD Frédéric
Quoted 232 lines · Michael S. Tsirkin
Quoted 224 lines · KONRAD Frédéric
On 29/04/2013 19:01, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; }Quoted 2 lines
Not sure this is what you want to do.
The device would be plugged before it is inited :/.I think this is exacly what we want to do.
In fact, this is what other buses do, because
devices simply can't init properly if they
do not know on which bus they reside.Quoted 11 lines
E.g. with pci:
do_pci_register_device(adds device on bus) initWe can add an analog of hotplug bus callback
if bus wants to get notified after device initialization. I don't see a need for this though.
Do you?No, as we don't want to allow virtio-device hotplug.
Quoted 8 lines
but look at the plugged callback in virtio-pci:
pci_set_word(config + PCI_SUBSYSTEM_ID, virtio_bus_get_vdev_id(bus)); proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);are impossible before the virtio-net init.
Quoted 9 lines
At this point I have to admit I don't understand what's going on any more.
virtio_net_instance_initsets config size to
some value, thenvirtio_net_set_config_sizeoverrides it... Help!I am guessing it's this hack that backfires somehow.
It would be interferring if virtio_net_set_config_size was not called.
The best I think is to set the config size in the virtio_net_init function, then remove the instance init.
The issue is that in the init function, the host_feature is not completely computed:
it lacks the three line in virtio pci plugged function.
Maybe we can move the two firsts line in the virtio_net_pci_init before the init if they are usefull in the config_size computing:
proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY;
proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;Then compute the last one directly in the init function which is the harder:
virtio_net_get_features
Note that there is in virtio_net_get_features:
features |= (1 << VIRTIO_NET_F_MAC);
Which is exacty the issue the initial patch is solving.
Fred
Quoted 285 lines · KONRAD Frédéric
On 29/04/2013 20:15, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 08:01:52PM +0200, KONRAD Frédéric wrote: On 29/04/2013 19:52, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:23:49PM +0200, KONRAD Frédéric wrote: On 29/04/2013 19:01, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; } Not sure this is what you want to do. The device would be plugged before it is inited :/. I think this is exacly what we want to do. In fact, this is what other buses do, because devices simply can't init properly if they do not know on which bus they reside. E.g. with pci: do_pci_register_device (adds device on bus) init We can add an analog of hotplug bus callback if bus wants to get notified after device initialization. I don't see a need for this though. Do you? No, as we don't want to allow virtio-device hotplug. but look at the plugged callback in virtio-pci: pci_set_word(config + PCI_SUBSYSTEM_ID, virtio_bus_get_vdev_id(bus)); proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); are impossible before the virtio-net init. At this point I have to admit I don't understand what's going on any more. virtio_net_instance_init sets config size to some value, then virtio_net_set_config_size overrides it... Help! I am guessing it's this hack that backfires somehow.It would be interferring if
virtio_net_set_config_sizewas not called.The best I think is to set the config size in the
virtio_net_initfunction, then remove the instance init.The issue is that in the init function, the
host_featureis not completely computed:it lacks the three line in virtio pci plugged function.
Quoted 7 lines
Maybe we can move the two firsts line in the
virtio_net_pci_initbefore the init if they are usefull in theconfig_sizecomputing:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;
Bus can set VIRTIO_F_NOTIFY_ON_EMPTY and VIRTIO_F_BAD_FEATURE later, this is not going to affect anything.
Reason is, VIRTIO_F_NOTIFY_ON_EMPTY and VIRTIO_F_BAD_FEATURE are compatibility hacks which is why we dont
have them for s390. It's probably a bug that
we have them for virtio-ccw.
Quoted 4 lines
Then compute the last one directly in the init function which is the harder:
virtio_net_get_features
The real fix is to set features in init.
Can we move host_features to struct VirtIODevice, and init to the device init function?
The reason we didn't do this initially is exactly
because we need to specify them in -device flag,
and there was no way to do this for VirtIODevice,
since it's the proxy that is instanciated.
Does the new bus infrastructure allow this?
Quoted 7 lines
Note that there is in
virtio_net_get_features:features |= (1 << VIRTIO_NET_F_MAC);Which is exacty the issue the initial patch is solving.
Fred
Yes. the lifetime of host features is as follows:
- configured in device by user, mostly set to "on" by default
- depending on device specific logic,
override some features - mostly to "off",
but we also force some required features to "on"
- Two exceptions (notify on empty+bad) are transport specific. they are also not user-controllable.
Signature · 1 line
MST
Quoted 306 lines · Michael S. Tsirkin
Quoted 285 lines · KONRAD Frédéric
On 29/04/2013 20:15, Michael S. Tsirkin wrote:
On Mon, Apr 29, 2013 at 08:01:52PM +0200, KONRAD Frédéric wrote: On 29/04/2013 19:52, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:23:49PM +0200, KONRAD Frédéric wrote: On 29/04/2013 19:01, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:41:08PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:30, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 07:21:06PM +0300, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 06:14:41PM +0200, KONRAD Frédéric wrote: On 29/04/2013 18:02, Michael S. Tsirkin wrote: On Mon, Apr 29, 2013 at 10:55:36AM -0500, Jesse Larrew wrote: On 04/29/2013 10:29 AM, KONRAD Frédéric wrote: On 29/04/2013 17:14, Jesse Larrew wrote: On 04/29/2013 09:55 AM, KONRAD Frédéric wrote: On 29/04/2013 16:42, Jesse Larrew wrote: On 04/25/2013 01:59 AM, Michael S. Tsirkin wrote: On Thu, Apr 25, 2013 at 02:21:29PM +0800, Jason Wang wrote: Commit 14f9b664 (hw/virtio-net.c: set config size using host features) tries to calculate config size based on the host features. But it forgets the VIRTIO_NET_F_MAC were always set for qemu later. This will lead a zero config len for virtio-net device when both VIRTIO_NET_F_STATUS and VIRTIO_NET_F_MQ were disabled form command line. Then qemu will crash when user tries to read the config of virtio-net. Fix this by counting VIRTIO_NET_F_MAC and make sure the config at least contains the mac address. Cc: Jesse Larrew <[email protected]> Signed-off-by: Jason Wang <[email protected]> --- hw/net/virtio-net.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c index 70c8fce..33a70ef 100644 --- a/hw/net/virtio-net.c +++ b/hw/net/virtio-net.c @@ -1264,7 +1264,8 @@ static void virtio_net_guest_notifier_mask(VirtIODevice *vdev, int idx, void virtio_net_set_config_size(VirtIONet *n, uint32_t host_features) { - int i, config_size = 0; + /* VIRTIO_NET_F_MAC can't be disabled from qemu side */ + int i, config_size = feature_sizes[0].end; This would be cleaner: host_features |= (1 << VIRTIO_NET_F_MAC); no need for a comment then. It seems to me that the real problem here is that host_features isn't properly populated before virtio_net_set_config_size() is called. Looking at virtio_device_init(), we can see why: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); return 0; } virtio_net_set_config_size() is currently being called as part of the k->init call, but the host_features aren't properly setup until the device is plugged into the bus using virtio_bus_plug_device(). After talking with mdroth, I think the proper way to fix this would be to extend VirtioDeviceClass to include a calculate_config_size() method that can be called at the appropriate time during device initialization like so: static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } virtio_bus_plug_device(vdev); + if (k->calculate_config_size && k->calculate_config_size(vdev) < 0) { + return -1; + } return 0; } This would ensure that host_features contains the proper data before any devices try to make use of it to calculate the config size. Good point, I didn't saw that. but this was not the case with commit 14f9b664 no? I suspect this bug was present in 14f9b664 as well. We just hadn't triggered it yet. I'll confirm this afternoon. Ok, I think there is a problem with your patch: virtio_init(VIRTIO_DEVICE(n), "virtio-net", VIRTIO_ID_NET, n->config_size); is called in virtio_net_device_init (k->init). Is there a way to resize the config after that? Nope. That's definitely a bug. I can send a patch to fix this today. Any objection to the method suggested above (extending VirtioDeviceClass)? This needs more thought. All devices expected everything is setup during the init call. config size is likely not the only issue. So we need almost all of virtio_bus_plug_device before init, and then after init send the signal: if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } Seems the interesting part is in virtio_pci_device_plugged at the end: proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE; proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); This is and was called after set_config_size, isn't that the issue? Basically devices expected everything to be setup at the point where init is called. We found one bug but let's fix it properly. There's no reason to delay parts of setup until after init, if we do, we'll trip on more uses of uninitialized data. for (i = 0; feature_sizes[i].flags != 0; i++) { if (host_features & feature_sizes[i].flags) { config_size = MAX(feature_sizes[i].end, config_size); -- 1.7.1 Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] -- Jesse Larrew Software Engineer, KVM Team IBM Linux Technology Center Phone: (512) 973-2052 (T/L: 363-2052) [email protected] Basically this is what I propose. Warning! compile-tested only. (I also dropped an unused return value). Signed-off-by: Michael S. Tsirkin <[email protected]> Which tree are you using? Is it up to date? Heh, more changes came in. So now, it's even simpler: Signed-off-by: Michael S. Tsirkin <[email protected]> --- diff --git a/hw/virtio/virtio-bus.c b/hw/virtio/virtio-bus.c index aab72ff..447ba16 100644 --- a/hw/virtio/virtio-bus.c +++ b/hw/virtio/virtio-bus.c @@ -38,7 +38,7 @@ do { printf("virtio_bus: " fmt , ## __VA_ARGS__); } while (0) #endif /* Plug the VirtIODevice */ -int virtio_bus_plug_device(VirtIODevice *vdev) +void virtio_bus_plug_device(VirtIODevice *vdev) { DeviceState *qdev = DEVICE(vdev); BusState *qbus = BUS(qdev_get_parent_bus(qdev)); @@ -51,8 +51,6 @@ int virtio_bus_plug_device(VirtIODevice *vdev) if (klass->device_plugged != NULL) { klass->device_plugged(qbus->parent); } - - return 0; } /* Reset the virtio_bus */ diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c index 0f88c25..c5228e6 100644 --- a/hw/virtio/virtio.c +++ b/hw/virtio/virtio.c @@ -1091,11 +1091,11 @@ static int virtio_device_init(DeviceState *qdev) { VirtIODevice *vdev = VIRTIO_DEVICE(qdev); VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(qdev); + virtio_bus_plug_device(vdev); assert(k->init != NULL); if (k->init(vdev) < 0) { return -1; } - virtio_bus_plug_device(vdev); return 0; } Not sure this is what you want to do. The device would be plugged before it is inited :/. I think this is exacly what we want to do. In fact, this is what other buses do, because devices simply can't init properly if they do not know on which bus they reside. E.g. with pci: do_pci_register_device (adds device on bus) init We can add an analog of hotplug bus callback if bus wants to get notified after device initialization. I don't see a need for this though. Do you? No, as we don't want to allow virtio-device hotplug. but look at the plugged callback in virtio-pci: pci_set_word(config + PCI_SUBSYSTEM_ID, virtio_bus_get_vdev_id(bus)); proxy->host_features = virtio_bus_get_vdev_features(bus, proxy->host_features); are impossible before the virtio-net init. At this point I have to admit I don't understand what's going on any more. virtio_net_instance_init sets config size to some value, then virtio_net_set_config_size overrides it... Help! I am guessing it's this hack that backfires somehow.It would be interferring if
virtio_net_set_config_sizewas not called.The best I think is to set the config size in the
virtio_net_initfunction, then remove the instance init.The issue is that in the init function, the
host_featureis not completely computed:it lacks the three line in virtio pci plugged function.
Quoted 7 lines
Maybe we can move the two firsts line in the
virtio_net_pci_initbefore the init if they are usefull in theconfig_sizecomputing:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;Bus can set
VIRTIO_F_NOTIFY_ON_EMPTYandVIRTIO_F_BAD_FEATURElater, this is not going to affect anything.Reason is,
VIRTIO_F_NOTIFY_ON_EMPTYandVIRTIO_F_BAD_FEATUREare compatibility hacks which is why we dont
have them for s390. It's probably a bug that
we have them for virtio-ccw.Quoted 3 lines
Then compute the last one directly in the init function which is the harder:
virtio_net_get_featuresThe real fix is to set features in init.
Quoted 8 lines
Can we move
host_featuresto struct VirtIODevice, and init to the device init function?The reason we didn't do this initially is exactly
because we need to specify them in -device flag,
and there was no way to do this for VirtIODevice,
since it's the proxy that is instanciated.
Does the new bus infrastructure allow this?
Yes, I think it's possible for PCI and S390, but it seems more difficult for CCW.
I don't really understand how it's working with CCW devices, there is an array of host_features?
Cornelia any idea?
Quoted 18 lines
Quoted 8 lines
Note that there is in
virtio_net_get_features:features |= (1 << VIRTIO_NET_F_MAC);Which is exacty the issue the initial patch is solving.
Fred
Yes. the lifetime of host features is as follows:
- configured in device by user, mostly set to "on" by default
- depending on device specific logic,
override some features - mostly to "off",
but we also force some required features to "on"
- Two exceptions (notify on empty+bad) are transport specific. they are also not user-controllable.
Quoted 8 lines · KONRAD Frédéric
Quoted 8 lines
Quoted 3 lines
Then compute the last one directly in the init function which is the harder:
virtio_net_get_featuresThe real fix is to set features in init.
Can we move
host_featuresto struct VirtIODevice, and init to the device init function?
Quoted 8 lines
Quoted 5 lines
The reason we didn't do this initially is exactly because we need to specify them in -device flag, and there was no way to do this for VirtIODevice, since it's the proxy that is instanciated. Does the new bus infrastructure allow this?
Yes, I think it's possible for PCI and S390, but it seems more difficult for CCW.
Can you send the patch for pci to let everyone see what you have in mind? The main issue is passing properties from proxy to the device.
Quoted 2 lines
I don't really understand how it's working with CCW devices, there is an array of
host_features?
In practice the array is of size 1. I'd suggest just assuming that for now. When we extend features (which will happen pretty soon)
we'll do this for all transports and change features to uint64_t everywhere.
After that, we'll have a bit of breathing space.
Quoted 20 lines
Cornelia any idea?
Quoted 18 lines
Quoted 8 lines
Note that there is in
virtio_net_get_features:features |= (1 << VIRTIO_NET_F_MAC);Which is exacty the issue the initial patch is solving.
Fred
Yes. the lifetime of host features is as follows:
- configured in device by user, mostly set to "on" by default
- depending on device specific logic,
override some features - mostly to "off",
but we also force some required features to "on"
- Two exceptions (notify on empty+bad) are transport specific. they are also not user-controllable.
Quoted 15 lines · On Tue, 30 Apr 2013 12:02:12 +0300 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 14 lines · KONRAD Frédéric
Quoted 13 lines
Quoted 3 lines
Then compute the last one directly in the init function which is the harder:
virtio_net_get_featuresThe real fix is to set features in init.
Can we move
host_featuresto struct VirtIODevice, and init to the device init function?The reason we didn't do this initially is exactly because we need to specify them in -device flag, and there was no way to do this for VirtIODevice, since it's the proxy that is instanciated. Does the new bus infrastructure allow this?
Quoted 6 lines
Quoted 2 lines
Yes, I think it's possible for PCI and S390, but it seems more difficult for CCW.
Can you send the patch for pci to let everyone see what you have in mind? The main issue is passing properties from proxy to the device.
Should probably not be hard to adapt to ccw.
Quoted 9 lines
Quoted 3 lines
I don't really understand how it's working with CCW devices, there is an array of
host_features?In practice the array is of size 1. I'd suggest just assuming that for now. When we extend features (which will happen pretty soon)
we'll do this for all transports and change features touint64_teverywhere.
After that, we'll have a bit of breathing space.
Yes, the idea was to keep features easily extendable. Going 64 bit will work fine; the transport should be able to handle even larger features.
Quoted 10 lines · On Mon, 29 Apr 2013 23:09:52 +0300 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 7 lines · KONRAD Frédéric
Maybe we can move the two firsts line in the
virtio_net_pci_initbefore the init if they are usefull in theconfig_sizecomputing:proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY; proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;
Quoted 8 lines
Bus can set
VIRTIO_F_NOTIFY_ON_EMPTYandVIRTIO_F_BAD_FEATURElater, this is not going to affect anything.Reason is,
VIRTIO_F_NOTIFY_ON_EMPTYandVIRTIO_F_BAD_FEATUREare compatibility hacks which is why we dont
have them for s390. It's probably a bug that
we have them for virtio-ccw.
Yes, probably (I used virtio-pci as my inspiration a lot...)
I'll check whether killing this works as expected.