Emulate status feature in vhost-vdpa net
PATCH 3/3 vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
In reply to Eugenio PérezNow that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)
include/net/vhost-vdpa.h +1 -0
@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */
hw/net/vhost_net.c +14 -2
@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)
net/vhost-vdpa.c +3 -0
@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Signature · 1 line
2.31.1
PATCH 1/3 virtio_net: Modify virtio_net_get_config to early return
In reply to Eugenio PérezNext patches introduce more code on vhost-vdpa branch, with already have too much indentation.
--- hw/net/virtio-net.c | 28 +++++++++++++++------------- 1 file changed, 15 insertions(+), 13 deletions(-)
hw/net/virtio-net.c +15 -13
@@ -158,20 +158,22 @@ static void virtio_net_get_config(VirtIODevice *vdev, uint8_t *config)if (nc->peer && nc->peer->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { ret =vhost_net_get_config (get_vhost_net (nc->peer), (uint8_t *)&netcfg, n->config_size); -if (ret !=-1 ) { -/* -* Some NIC/kernel combinations present 0 as the mac address. As -* that is not a legal address, try to proceed with the -* address from the QEMU command line in the hope that the -* address has been configured correctly elsewhere - just not -* reported by the device. -*/ -if (memcmp (&netcfg.mac, &zero,sizeof (zero)) ==0 ) { -info_report ("Zero hardware mac address detected. Ignoring." ); -memcpy (netcfg.mac, n->mac,ETH_ALEN ); - } -memcpy (config, &netcfg, n->config_size); +if (ret ==-1 ) { +return ; } + +/* +* Some NIC/kernel combinations present 0 as the mac address. As that +* is not a legal address, try to proceed with the address from the +* QEMU command line in the hope that the address has been configured +* correctly elsewhere - just not reported by the device. +*/ +if (memcmp (&netcfg.mac, &zero,sizeof (zero)) ==0 ) { +info_report ("Zero hardware mac address detected. Ignoring." ); +memcpy (netcfg.mac, n->mac,ETH_ALEN ); + } + +memcpy (config, &netcfg, n->config_size); } }
Signature · 1 line
2.31.1
PATCH 2/3 virtio_net: Handle _F_STATUS emulation in virtio_net_get_config
In reply to Eugenio PérezAt this moment this code path is not reached, but vdpa devices can offer VIRTIO_NET_F_STATUS unconditionally. While the guest must assume that link is always up by the standard, qemu will set the status bit to 1 always in this case.
This makes little use by itself, but VIRTIO_NET_F_STATUS is needed for the guest to read status bit VIRTIO_NET_F_GUEST_ANNOUNCE, used by feature VIRTIO_NET_F_GUEST_ANNOUNCE. So qemu must emulate status feature in case it needs to emulate the guest announce feature.
--- hw/net/virtio-net.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
hw/net/virtio-net.c +9 -2
@@ -156,8 +156,9 @@ static void virtio_net_get_config(VirtIODevice *vdev, uint8_t *config)* disconnect/reconnect a VDPA peer. */ if (nc->peer && nc->peer->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { - ret =vhost_net_get_config (get_vhost_net (nc->peer), (uint8_t *)&netcfg, - n->config_size); +struct vhost_net *net =get_vhost_net (nc->peer); + + ret =vhost_net_get_config (net, (uint8_t *)&netcfg, n->config_size);if (ret ==-1 ) {return ; }
@@ -173,6 +174,12 @@ static void virtio_net_get_config(VirtIODevice *vdev, uint8_t *config)memcpy (netcfg.mac, n->mac,ETH_ALEN ); } +if (vdev->guest_features &BIT_ULL (VIRTIO_NET_F_STATUS ) && + !(net->dev.features &BIT_ULL (VIRTIO_NET_F_STATUS ))) { +/* Emulating link up in qemu */ + netcfg.status |=virtio_tswap16 (vdev,VIRTIO_NET_S_LINK_UP ); + } +memcpy (config, &netcfg, n->config_size); } }
Signature · 1 line
2.31.1
cover Emulate status feature in vhost-vdpa net
The net config space is already copied from the device so it can me modified by qemu. In particular, this is already done to fix cases where the NIC does not expose the right fields.
It's trivial to emulate _F_STATE with qemu if not supported by the device, sice a valid approach is to always show the link as up. If the feature is already supported by the device, no config space modification is needed.
This is a pre requisite to use other features like _F_GUEST_ANNOUNCE, since _F_STATUS is needed for the guest to access the status.
These patches are sent on top of [1] series, so trivial conflicts could arise if it is applied directly on master. Future versions can be not based on it is more convenient.
Eugenio Pérez (3):
virtio_net: Modify virtio_net_get_config to early return
virtio_net: Handle _F_STATUS emulation in virtio_net_get_config
vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- hw/net/virtio-net.c | 39 ++++++++++++++++++++++++--------------- net/vhost-vdpa.c | 3 +++ 4 files changed, 42 insertions(+), 17 deletions(-)
Signature · 1 line
2.31.1
Re: [PATCH 1/3] virtio_net: Modify virtio_net_get_config to early return
In reply to Eugenio PérezQuoted 7 lines · Eugenio Pérez
Next patches introduce more code on vhost-vdpa branch, with already have too much indentation.
Signed-off-by Eugenio Pérez <[email protected]>--- hw/net/virtio-net.c | 28 +++++++++++++++------------- 1 file changed, 15 insertions(+), 13 deletions(-)
Quoted 8 lines · Eugenio Pérez
The net config space is already copied from the device so it can me modified by qemu. In particular, this is already done to fix cases where the NIC does not expose the right fields.
It's trivial to emulate
_F_STATEwith qemu if not supported by the device, sice a valid approach is to always show the link as up. If the feature is already supported by the device, no config space modification is needed.
Quoted 6 lines
This is a pre requisite to use other features like
_F_GUEST_ANNOUNCE, since_F_STATUSis needed for the guest to access the status.These patches are sent on top of [1] series, so trivial conflicts could arise if it is applied directly on master. Future versions can be not based on it is more convenient.
Jason seems to be taking all this work through net tree.
Quoted 14 lines
Eugenio Pérez (3):
virtio_net: Modifyvirtio_net_get_configto early return
virtio_net: Handle_F_STATUSemulation invirtio_net_get_config
vdpa: ExposeVIRTIO_NET_F_STATUSunconditionallyinclude/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- hw/net/virtio-net.c | 39 ++++++++++++++++++++++++--------------- net/vhost-vdpa.c | 3 +++ 4 files changed, 42 insertions(+), 17 deletions(-)Signature · 1 line
2.31.1
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
In reply to Eugenio PérezQuoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>
I may miss something but isn't more easier to simply remove the _F_STATUS from vdpa_feature_bits[]?
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?
How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
The goal with this patch series is to let the guest access the status always, even if the device doesn't support _F_STATUS.
Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 11 lines · Eugenio Perez Martin
Quoted 9 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>
Quoted 6 lines
Quoted 3 lines
I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device. So if we remove _F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
uint64_t features)
{
const int *bit = feature_bits;
while (*bit != VHOST_INVALID_FEATURE_BIT) {
uint64_t bit_mask = (1ULL << *bit);
if (!(hdev->features & bit_mask)) {
features &= ~bit_mask;
}
bit++;
}
return features;
}Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 18 lines · Jason Wang
Quoted 15 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?
Quoted 6 lines
Quoted 2 lines
How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
From what I see these are the only features that will be forwarded to the guest as device_features. If it is not in the list, the feature will be masked out, as if the device does not support it.
So now _F_STATUS it was forwarded only if the device supports it. If we remove it from bit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for _F_GUEST_ANNOUNCE.
Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report link_down. We will only emulate LINK_UP always in the case the device does not support _F_STATUS.
Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)
Quoted 12 lines
{
const int *bit = feature_bits;
while (*bit != VHOST_INVALID_FEATURE_BIT) {
uint64_t bit_mask = (1ULL << *bit);
if (!(hdev->features & bit_mask)) {
features &= ~bit_mask;
}
bit++;
}
return features;
}Now maybe I'm the one missing something, but why is this not done as a masking directly?
Instead of making feature_bits an array of ints, to declare it as a uint64_t with the valid feature bits and simply return features & feature_bits.
Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = {
VIRTIO_F_NOTIFY_ON_EMPTY,
VIRTIO_RING_F_INDIRECT_DESC,
VIRTIO_RING_F_EVENT_IDX,
VIRTIO_NET_F_MRG_RXBUF,
VIRTIO_F_VERSION_1,
VIRTIO_NET_F_MTU,
VIRTIO_F_IOMMU_PLATFORM,
VIRTIO_F_RING_PACKED,
VIRTIO_NET_F_HASH_REPORT,
VHOST_INVALID_FEATURE_BIT
};You can see there's no STATUS bit there since it is emulated by Qemu.
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,
Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }
Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in feature_bits, we need validate the hdev->features and mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUS
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };
Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in vdpa_feature_bits are not acked to the device. More on this later.
Quoted 40 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Quoted 8 lines
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUS
That's useful for _F_GUEST_ANNOUNCE, but we need to ack _F_STATUS for the device if it supports it. QEMU cannot detect by itself when the link is not up. I think that setting unconditionally VIRTIO_NET_S_LINK_UP is actually a regression, since the guest cannot detect the link down that way.
To enable _F_STATUS unconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support _F_GUEST_ANNOUNCE in all cases without device's cooperation.
Having said that, should we go the opposite route and ack _F_STATE as long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 107 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 44 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
Quoted 7 lines
Quoted 4 lines
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUSThat's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.
Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
Quoted 4 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.
I think the idea is to still read status from config if the device supports this.
Quoted 9 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?
So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in vhost_net since we do the features validation there. We probably need another table as input for get/set features there?
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 115 lines · Jason Wang
Quoted 112 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 48 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUS
Quoted 6 lines
Quoted 2 lines
That's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
It is needed for the guest to read the status bit:
"""
status only exists if VIRTIO_NET_F_STATUS is set. Two read-only bits (for the driver) are currently defined for the status field:VIRTIO_NET_S_LINK_UP and VIRTIO_NET_S_ANNOUNCE.
"""
A change on the standard could be possible, like "status only exists if VIRTIO_NET_F_STATUS or VIRTIO_NET_F_GUEST_ANNOUNCE is set". However, Linux drivers already expect _F_STATUS to read _S_ANNOUNCE and to emulate _F_STATUS in case the device doesn't support it should not be a big deal in my opinion.
Quoted 8 lines
Quoted 4 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.I think the idea is to still read status from config if the device supports this.
Yes, that's my point. If I delete it from vdpa_feature_bits, it will not be acked to the device, so we cannot read status from the device.
Quoted 6 lines
Quoted 6 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.
Quoted 9 lines
Quoted 3 lines
Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in
vhost_netsince we do the features validation there. We probably need another table as input for get/set features there?
We can discuss how to do it for sure. But as you pointed out, vhost_net and virtio_net already modify the features received from the devices, so it makes sense to me to modify the features set by the guest.
The problem is that we need to transmit to vhost when ack _F_STATUS and when not. The first proposal was to add a new member of vhost_vdpa but this is not optimal.
If we add a new table it should be a static const one, and vhost_vdpa should have a pointer to it, isn't it? Something like features that are emulated by qemu so they must be offered always to the guest?
Thanks!
Quoted 95 lines
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 131 lines · Eugenio Perez Martin
Quoted 115 lines · Jason Wang
Quoted 112 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 48 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUSQuoted 6 lines
Quoted 2 lines
That's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
It is needed for the guest to read the status bit:
"""
status only exists ifVIRTIO_NET_F_STATUSis set. Two read-only bits (for the driver) are currently defined for the status field:VIRTIO_NET_S_LINK_UPandVIRTIO_NET_S_ANNOUNCE.
"""
Quoted 5 lines
A change on the standard could be possible, like "status only exists if
VIRTIO_NET_F_STATUSorVIRTIO_NET_F_GUEST_ANNOUNCEis set". However, Linux drivers already expect_F_STATUSto read_S_ANNOUNCEand to emulate_F_STATUSin case the device doesn't support it should not be a big deal in my opinion.
RIght, so I think we need a spec patch to clarify the dependency, currently, spec said ANNOUNCE depends on CTRL_VQ.
Quoted 34 lines
Quoted 9 lines
Quoted 5 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.I think the idea is to still read status from config if the device supports this.
Yes, that's my point. If I delete it from
vdpa_feature_bits, it will not be acked to the device, so we cannot read status from the device.Quoted 6 lines
Quoted 6 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.Quoted 9 lines
Quoted 3 lines
Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in
vhost_netsince we do the features validation there. We probably need another table as input for get/set features there?We can discuss how to do it for sure. But as you pointed out,
vhost_netandvirtio_netalready modify the features received from the devices, so it makes sense to me to modify the features set by the guest.
Quoted 6 lines
The problem is that we need to transmit to vhost when ack
_F_STATUSand when not. The first proposal was to add a new member ofvhost_vdpabut this is not optimal.If we add a new table it should be a static const one, and
vhost_vdpashould have a pointer to it, isn't it?
Yes.
Quoted 2 lines
Something like features that
are emulated by qemu so they must be offered always to the guest?
Kind of, actually it should be the features:
1) could be always seen by guest
2) when vhost device have this feature, use that
3) when vhost device doesn't have this feature, emulate one
But a question still, is there a vDPA parent that can't do _F_STATUS now (if not, we probably don't need to bother now).
Thanks
Quoted 99 lines
Thanks!
Quoted 95 lines
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 133 lines · Jason Wang
Quoted 131 lines · Eugenio Perez Martin
Quoted 115 lines · Jason Wang
Quoted 112 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 48 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUSQuoted 6 lines
Quoted 2 lines
That's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
It is needed for the guest to read the status bit:
"""
status only exists ifVIRTIO_NET_F_STATUSis set. Two read-only bits (for the driver) are currently defined for the status field:VIRTIO_NET_S_LINK_UPandVIRTIO_NET_S_ANNOUNCE.
"""
Quoted 9 lines
Quoted 5 lines
A change on the standard could be possible, like "status only exists if
VIRTIO_NET_F_STATUSorVIRTIO_NET_F_GUEST_ANNOUNCEis set". However, Linux drivers already expect_F_STATUSto read_S_ANNOUNCEand to emulate_F_STATUSin case the device doesn't support it should not be a big deal in my opinion.RIght, so I think we need a spec patch to clarify the dependency, currently, spec said ANNOUNCE depends on
CTRL_VQ.
Would it be enough to expand it under the "Feature bit requirements" section?
Quoted 43 lines
Quoted 34 lines
Quoted 9 lines
Quoted 5 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.I think the idea is to still read status from config if the device supports this.
Yes, that's my point. If I delete it from
vdpa_feature_bits, it will not be acked to the device, so we cannot read status from the device.Quoted 6 lines
Quoted 6 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.Quoted 9 lines
Quoted 3 lines
Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in
vhost_netsince we do the features validation there. We probably need another table as input for get/set features there?We can discuss how to do it for sure. But as you pointed out,
vhost_netandvirtio_netalready modify the features received from the devices, so it makes sense to me to modify the features set by the guest.Quoted 6 lines
The problem is that we need to transmit to vhost when ack
_F_STATUSand when not. The first proposal was to add a new member ofvhost_vdpabut this is not optimal.If we add a new table it should be a static const one, and
vhost_vdpashould have a pointer to it, isn't it?Yes.
Quoted 9 lines
Quoted 2 lines
Something like features that
are emulated by qemu so they must be offered always to the guest?Kind of, actually it should be the features:
1) could be always seen by guest
2) when vhost device have this feature, use that
3) when vhost device doesn't have this feature, emulate one
I'm fine with that approach, but restricting the changes to either vhost_net or virtio_net makes more sense to me. The net config space interception goes to virtio_net anyway, not to vhost-vdpa.
I'll try to prepare the patches with a new array.
Quoted 3 lines
But a question still, is there a vDPA parent that can't do
_F_STATUSnow (if not, we probably don't need to bother now).
Only mlx have _F_STATUS at this moment.
If I understand you correctly, you are proposing not to emulate _F_STATUS if the device does not support it? To emulate _F_STATUS is not a big deal emulating _F_GUEST_ANNOUNCE on top anyway.
Thanks!
Quoted 102 lines
Thanks
Quoted 99 lines
Thanks!
Quoted 95 lines
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 141 lines · Eugenio Perez Martin
Quoted 139 lines · Jason Wang
Quoted 136 lines · Eugenio Perez Martin
Quoted 115 lines · Jason Wang
Quoted 112 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 48 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUSQuoted 6 lines
Quoted 2 lines
That's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
It is needed for the guest to read the status bit:
"""
status only exists ifVIRTIO_NET_F_STATUSis set. Two read-only bits (for the driver) are currently defined for the status field:VIRTIO_NET_S_LINK_UPandVIRTIO_NET_S_ANNOUNCE.
"""A change on the standard could be possible, like "status only exists if
VIRTIO_NET_F_STATUSorVIRTIO_NET_F_GUEST_ANNOUNCEis set". However, Linux drivers already expect_F_STATUSto read_S_ANNOUNCEand to emulate_F_STATUSin case the device doesn't support it should not be a big deal in my opinion.
Quoted 6 lines
Quoted 3 lines
RIght, so I think we need a spec patch to clarify the dependency, currently, spec said ANNOUNCE depends on
CTRL_VQ.Would it be enough to expand it under the "Feature bit requirements" section?
Yes.
Quoted 53 lines
Quoted 43 lines
Quoted 34 lines
Quoted 9 lines
Quoted 5 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.I think the idea is to still read status from config if the device supports this.
Yes, that's my point. If I delete it from
vdpa_feature_bits, it will not be acked to the device, so we cannot read status from the device.Quoted 6 lines
Quoted 6 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.Quoted 9 lines
Quoted 3 lines
Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in
vhost_netsince we do the features validation there. We probably need another table as input for get/set features there?We can discuss how to do it for sure. But as you pointed out,
vhost_netandvirtio_netalready modify the features received from the devices, so it makes sense to me to modify the features set by the guest.Quoted 6 lines
The problem is that we need to transmit to vhost when ack
_F_STATUSand when not. The first proposal was to add a new member ofvhost_vdpabut this is not optimal.If we add a new table it should be a static const one, and
vhost_vdpashould have a pointer to it, isn't it?Yes.
Quoted 9 lines
Quoted 2 lines
Something like features that
are emulated by qemu so they must be offered always to the guest?Kind of, actually it should be the features:
1) could be always seen by guest
2) when vhost device have this feature, use that
3) when vhost device doesn't have this feature, emulate one
Quoted 5 lines
I'm fine with that approach, but restricting the changes to either
vhost_netorvirtio_netmakes more sense to me. The net config space interception goes tovirtio_netanyway, not to vhost-vdpa.I'll try to prepare the patches with a new array.
I'm ok if Michael is ok with this. I think it might help if we add a comment in general vhost code like vhost_get_features(), then readers know that each individual vhost implementation can mediate the feature by their own.
Quoted 10 lines
Quoted 4 lines
But a question still, is there a vDPA parent that can't do
_F_STATUSnow (if not, we probably don't need to bother now).Only mlx have
_F_STATUSat this moment.If I understand you correctly, you are proposing not to emulate
_F_STATUSif the device does not support it? To emulate_F_STATUSis not a big deal emulating_F_GUEST_ANNOUNCEon top anyway.
I meant if most of the vDPA parent supports _F_STATUS, there's probably no need to emulate it:
1) simulator, not hard to add status
2) vp_vdpa, should support _F_STATUS
3) IFCVF, I guess it should have this support, Ling Shan, can you clarify this?
4) ENI, not sure but anyhow it's a legacy parent
Thanks
Quoted 106 lines
Thanks!
Quoted 102 lines
Thanks
Quoted 99 lines
Thanks!
Quoted 95 lines
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);
Re: [PATCH 3/3] vdpa: Expose VIRTIO_NET_F_STATUS unconditionally
Quoted 227 lines · Jason Wang
Quoted 141 lines · Eugenio Perez Martin
Quoted 139 lines · Jason Wang
Quoted 136 lines · Eugenio Perez Martin
Quoted 115 lines · Jason Wang
Quoted 112 lines · Eugenio Perez Martin
Quoted 52 lines · Jason Wang
Quoted 23 lines · Eugenio Perez Martin
Quoted 21 lines · Jason Wang
Quoted 17 lines · Eugenio Perez Martin
Quoted 12 lines · Jason Wang
Quoted 4 lines · Eugenio Pérez
Now that qemu can handle and emulate it if the vdpa backend does not support it we can offer it always.
Signed-off-by Eugenio Pérez <[email protected]>I may miss something but isn't more easier to simply remove the
_F_STATUSfrom vdpa_feature_bits[]?How is that? if we remove it, the guest cannot ack it so it cannot access the net status, isn't it?
Quoted 7 lines
Quoted 3 lines
My understanding is that the bits stored in the vdpa_feature_bits[] are the features that must be explicitly supported by the vhost device.
(Non English native here, so maybe I don't get what you mean :) ) The device may not support them. net simulator lacks some of them actually, and it works.
Speaking too fast, I think I meant that, if the bit doesn't belong to vdpa_feature_bits[], it is assumed to be supported by the Qemu without the support of the vhost. So Qemu won't even try to validate if vhost has this support. E.g for vhost-net, we only have:
static const int kernel_feature_bits[] = { VIRTIO_F_NOTIFY_ON_EMPTY, VIRTIO_RING_F_INDIRECT_DESC, VIRTIO_RING_F_EVENT_IDX, VIRTIO_NET_F_MRG_RXBUF, VIRTIO_F_VERSION_1, VIRTIO_NET_F_MTU, VIRTIO_F_IOMMU_PLATFORM, VIRTIO_F_RING_PACKED, VIRTIO_NET_F_HASH_REPORT, VHOST_INVALID_FEATURE_BIT };Quoted 2 lines
You can see there's no STATUS bit there since it is emulated by Qemu.
Ok now I get what you mean, and yes we may modify the patches in that direction.
But if we go then we need to modify how qemu ack the features, because the features that are not in
vdpa_feature_bitsare not acked to the device. More on this later.Quoted 48 lines
Quoted 4 lines
From what I see these are the only features that will be forwarded to the guest as
device_features. If it is not in the list, the feature will be masked out,Only when there's no support for this feature from the vhost.
Quoted 30 lines
as if the device does not support it.
So now
_F_STATUSit was forwarded only if the device supports it. If we remove it frombit_mask, it will never be offered to the guest. But we want to offer it always, since we will need it for_F_GUEST_ANNOUNCE.Things get more complex because we actually need to ack it back if the device offers it, so the vdpa device can report
link_down. We will only emulateLINK_UPalways in the case the device does not support_F_STATUS.Quoted 5 lines
So if we remove
_F_STATUS, Qemu vhost code won't validate if vhost-vdpa device has this support:
uint64_tvhost_get_features(structvhost_dev*hdev, const int*feature_bits,
uint64_tfeatures)Quoted 12 lines
{ const int *bit = feature_bits; while (*bit != VHOST_INVALID_FEATURE_BIT) { uint64_t bit_mask = (1ULL << *bit); if (!(hdev->features & bit_mask)) { features &= ~bit_mask; } bit++; } return features; }Quoted 2 lines
Now maybe I'm the one missing something, but why is this not done as a masking directly?
Not sure, the code has been there since day 0.
But you can see from the code:
1) if STATUS is in
feature_bits, we need validate thehdev->featuresand mask it if the vhost doesn't have the support
2) if STATUS is not, we don't do the check and driver may still see STATUSQuoted 6 lines
Quoted 2 lines
That's useful for
_F_GUEST_ANNOUNCE, but we need to ack_F_STATUSfor the device if it supports it.Rethink about this, I don't see why ANNOUNCE depends on STATUS (spec doesn't say so).
It is needed for the guest to read the status bit:
"""
status only exists ifVIRTIO_NET_F_STATUSis set. Two read-only bits (for the driver) are currently defined for the status field:VIRTIO_NET_S_LINK_UPandVIRTIO_NET_S_ANNOUNCE.
"""A change on the standard could be possible, like "status only exists if
VIRTIO_NET_F_STATUSorVIRTIO_NET_F_GUEST_ANNOUNCEis set". However, Linux drivers already expect_F_STATUSto read_S_ANNOUNCEand to emulate_F_STATUSin case the device doesn't support it should not be a big deal in my opinion.Quoted 6 lines
Quoted 3 lines
RIght, so I think we need a spec patch to clarify the dependency, currently, spec said ANNOUNCE depends on
CTRL_VQ.Would it be enough to expand it under the "Feature bit requirements" section?
Yes.
Quoted 53 lines
Quoted 43 lines
Quoted 34 lines
Quoted 9 lines
Quoted 5 lines
QEMU cannot detect by itself when the
link is not up. I think that setting unconditionallyVIRTIO_NET_S_LINK_UPis actually a regression, since the guest cannot detect the link down that way.I think the idea is to still read status from config if the device supports this.
Yes, that's my point. If I delete it from
vdpa_feature_bits, it will not be acked to the device, so we cannot read status from the device.Quoted 6 lines
Quoted 6 lines
To enable
_F_STATUSunconditionally is only done in the case the device does not support it, because its emulation is very easy. That way we support_F_GUEST_ANNOUNCEin all cases without device's cooperation.Quoted 9 lines
Quoted 3 lines
Having said that, should we go the opposite route and ack
_F_STATEas long as the device supports it? As an advantage, all backends should support that at this moment, isn't it?So I think the method used in this patch is fine, but I wonder if it's better to move it to the vhost layer instead of doing it in
vhost_netsince we do the features validation there. We probably need another table as input for get/set features there?We can discuss how to do it for sure. But as you pointed out,
vhost_netandvirtio_netalready modify the features received from the devices, so it makes sense to me to modify the features set by the guest.Quoted 6 lines
The problem is that we need to transmit to vhost when ack
_F_STATUSand when not. The first proposal was to add a new member ofvhost_vdpabut this is not optimal.If we add a new table it should be a static const one, and
vhost_vdpashould have a pointer to it, isn't it?Yes.
Quoted 9 lines
Quoted 2 lines
Something like features that
are emulated by qemu so they must be offered always to the guest?Kind of, actually it should be the features:
1) could be always seen by guest
2) when vhost device have this feature, use that
3) when vhost device doesn't have this feature, emulate oneQuoted 5 lines
I'm fine with that approach, but restricting the changes to either
vhost_netorvirtio_netmakes more sense to me. The net config space interception goes tovirtio_netanyway, not to vhost-vdpa.I'll try to prepare the patches with a new array.
I'm ok if Michael is ok with this. I think it might help if we add a comment in general vhost code like
vhost_get_features(), then readers know that each individual vhost implementation can mediate the feature by their own.Quoted 10 lines
Quoted 4 lines
But a question still, is there a vDPA parent that can't do
_F_STATUSnow (if not, we probably don't need to bother now).Only mlx have
_F_STATUSat this moment.If I understand you correctly, you are proposing not to emulate
_F_STATUSif the device does not support it? To emulate_F_STATUSis not a big deal emulating_F_GUEST_ANNOUNCEon top anyway.
Quoted 8 lines
I meant if most of the vDPA parent supports
_F_STATUS, there's probably no need to emulate it:1) simulator, not hard to add status
2)vp_vdpa, should support_F_STATUS
3) IFCVF, I guess it should have this support, Ling Shan, can you clarify this?
4) ENI, not sure but anyhow it's a legacy parent
If this is the case then yes, it might make more sense to add _F_STATUS support to at least vdpa simulator.
We can always emulate it in qemu later for sure.
Thanks!
Quoted 109 lines
Thanks
Quoted 106 lines
Thanks!
Quoted 102 lines
Thanks
Quoted 99 lines
Thanks!
Quoted 95 lines
Thanks
Quoted 92 lines
Thanks!
Quoted 85 lines
Thanks
Quoted 82 lines
Instead of making
feature_bitsan array of ints, to declare it as auint64_twith the valid feature bits and simply return features &feature_bits.Thanks!
Quoted 74 lines
Thanks
Quoted 69 lines
The goal with this patch series is to let the guest access the status always, even if the device doesn't support
_F_STATUS.Quoted 64 lines
Thanks
Quoted 60 lines
--- include/net/vhost-vdpa.h | 1 + hw/net/vhost_net.c | 16 ++++++++++++++-- net/vhost-vdpa.c | 3 +++ 3 files changed, 18 insertions(+), 2 deletions(-)include/net/vhost-vdpa.h +1
-0@@ -17,5 +17,6 @@struct vhost_net *vhost_vdpa_get_vhost_net (NetClientState *nc);extern const int vdpa_feature_bits[]; +extern const uint64_t vhost_vdpa_net_added_feature_bits;#endif /* VHOST_VDPA_H */ hw/net/vhost_net.c +14
-2@@ -109,10 +109,22 @@ static const int *vhost_net_get_feature_bits(struct vhost_net *net)return feature_bits; } +static uint64_t vhost_net_add_feature_bits (struct vhost_net *net) +{ +if (net->nc->info->type ==NET_CLIENT_DRIVER_VHOST_VDPA ) { +return vhost_vdpa_net_added_feature_bits; + } + +return 0 ; +} +uint64_t vhost_net_get_features (struct vhost_net *net,uint64_t features) { -return vhost_get_features (&net->dev,vhost_net_get_feature_bits (net), - features); +uint64_t ret =vhost_get_features (&net->dev, +vhost_net_get_feature_bits (net), + features); + +return ret |vhost_net_add_feature_bits (net); }int vhost_net_get_config (struct vhost_net *net,uint8_t *config,uint32_t config_len)net/vhost-vdpa.c +3
-0@@ -99,6 +99,9 @@ static const uint64_t vdpa_svq_device_features =BIT_ULL (VIRTIO_NET_F_RSC_EXT ) |BIT_ULL (VIRTIO_NET_F_STANDBY ); +const uint64_t vhost_vdpa_net_added_feature_bits = +BIT_ULL (VIRTIO_NET_F_STATUS ); +VHostNetState *vhost_vdpa_get_vhost_net (NetClientState *nc) {VhostVDPAState *s =DO_UPCAST (VhostVDPAState, nc, nc);