On Thu, Apr 13, 2023 at 10:39?PM Eli Cohen <elic at nvidia.com>
wrote:>
>
> On 10/04/2023 5:00, Jason Wang wrote:
> > On Tue, Apr 4, 2023 at 3:04?PM Eli Cohen <elic at nvidia.com>
wrote:
> >>
> >> On 04/04/2023 5:24, Jason Wang wrote:
> >>> ? 2023/4/4 00:20, Eli Cohen ??:
> >>>> Add support for generation of interrupts from the device
directly to the
> >>>> VM to the VCPU thus avoiding the overhead on the host CPU.
> >>>>
> >>>> When supported, the driver will attempt to allocate
vectors for each
> >>>> data virtqueue. If a vector for a virtqueue cannot be
provided it will
> >>>> use the QP mode where notifications go through the driver.
> >>>>
> >>>> In addition, we add a shutdown callback to make sure
allocated
> >>>> interrupts are released in case of shutdown to allow clean
shutdown.
> >>>>
> >>>> Signed-off-by: Eli Cohen <elic at nvidia.com>
> >>>> Signed-off-by: Saeed Mahameed <saeedm at nvidia.com>
> >>>> ---
> >>>> drivers/vdpa/mlx5/net/mlx5_vnet.c | 139
++++++++++++++++++++++++++++--
> >>>> drivers/vdpa/mlx5/net/mlx5_vnet.h | 14 +++
> >>>> 2 files changed, 144 insertions(+), 9 deletions(-)
> >>>>
> >>>> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> >>>> b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> >>>> index 520646ae7fa0..215a46cf8a98 100644
> >>>> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> >>>> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> >>>> @@ -83,6 +83,7 @@ struct mlx5_vq_restore_info {
> >>>> u64 driver_addr;
> >>>> u16 avail_index;
> >>>> u16 used_index;
> >>>> + struct msi_map map;
> >>>> bool ready;
> >>>> bool restore;
> >>>> };
> >>>> @@ -118,6 +119,7 @@ struct mlx5_vdpa_virtqueue {
> >>>> u16 avail_idx;
> >>>> u16 used_idx;
> >>>> int fw_state;
> >>>> + struct msi_map map;
> >>>
> >>> Just spot that this structure is defined in mlx5_vnet.c but
it's
> >>> nothing net specific. I think it's better to move it with
the
> >>> interrupt bypassing logic here to core/ (we can do it in the
future).
> >> Right.
> >>>
> >>>> /* keep last in the struct */
> >>>> struct mlx5_vq_restore_info ri;
> >>>> @@ -792,6 +794,13 @@ static bool counters_supported(const
struct
> >>>> mlx5_vdpa_dev *mvdev)
> >>>> BIT_ULL(MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
> >>>> }
> >>>> +static bool msix_mode_supported(struct mlx5_vdpa_dev
*mvdev)
> >>>> +{
> >>>> + return (MLX5_CAP_DEV_VDPA_EMULATION(mvdev->mdev,
event_mode) &
> >>>> + (1 << MLX5_VIRTIO_Q_EVENT_MODE_MSIX_MODE)
&&
> >>>> + pci_msix_can_alloc_dyn(mvdev->mdev->pdev));
> >>>> +}
> >>>> +
> >>>> static int create_virtqueue(struct mlx5_vdpa_net *ndev,
struct
> >>>> mlx5_vdpa_virtqueue *mvq)
> >>>> {
> >>>> int inlen =
MLX5_ST_SZ_BYTES(create_virtio_net_q_in);
> >>>> @@ -829,9 +838,15 @@ static int create_virtqueue(struct
mlx5_vdpa_net
> >>>> *ndev, struct mlx5_vdpa_virtque
> >>>> if (vq_is_tx(mvq->index))
> >>>> MLX5_SET(virtio_net_q_object, obj_context,
tisn_or_qpn,
> >>>> ndev->res.tisn);
> >>>> - MLX5_SET(virtio_q, vq_ctx, event_mode,
> >>>> MLX5_VIRTIO_Q_EVENT_MODE_QP_MODE);
> >>>> + if (mvq->map.virq) {
> >>>> + MLX5_SET(virtio_q, vq_ctx, event_mode,
> >>>> MLX5_VIRTIO_Q_EVENT_MODE_MSIX_MODE);
> >>>> + MLX5_SET(virtio_q, vq_ctx, event_qpn_or_msix,
mvq->map.index);
> >>>> + } else {
> >>>> + MLX5_SET(virtio_q, vq_ctx, event_mode,
> >>>> MLX5_VIRTIO_Q_EVENT_MODE_QP_MODE);
> >>>> + MLX5_SET(virtio_q, vq_ctx, event_qpn_or_msix,
> >>>> mvq->fwqp.mqp.qpn);
> >>>> + }
> >>>> +
> >>>> MLX5_SET(virtio_q, vq_ctx, queue_index,
mvq->index);
> >>>> - MLX5_SET(virtio_q, vq_ctx, event_qpn_or_msix,
mvq->fwqp.mqp.qpn);
> >>>> MLX5_SET(virtio_q, vq_ctx, queue_size,
mvq->num_ent);
> >>>> MLX5_SET(virtio_q, vq_ctx, virtio_version_1_0,
> >>>> !!(ndev->mvdev.actual_features &
> >>>> BIT_ULL(VIRTIO_F_VERSION_1)));
> >>>> @@ -1174,6 +1189,32 @@ static void
counter_set_dealloc(struct
> >>>> mlx5_vdpa_net *ndev, struct mlx5_vdpa_vir
> >>>> mlx5_vdpa_warn(&ndev->mvdev,
"dealloc counter set 0x%x\n",
> >>>> mvq->counter_set_id);
> >>>> }
> >>>> +static void alloc_vector(struct mlx5_vdpa_net *ndev,
> >>>> + struct mlx5_vdpa_virtqueue *mvq)
> >>>> +{
> >>>> + struct mlx5_vdpa_irq_pool *irqp = &ndev->irqp;
> >>>> + int i;
> >>>> +
> >>>> + for (i = 0; i < irqp->num_ent; i++) {
> >>>> + if (!irqp->entries[i].usecount) {
> >>>> + irqp->entries[i].usecount++;
> >>>
> >>> If we can't get a usercount which is greater than 1, it
might be
> >>> better to use boolean instead.
> >> Will fix.
> >>>
> >>>> + mvq->map = irqp->entries[i].map;
> >>>> + return;
> >>>> + }
> >>>> + }
> >>>> +}
> >>>> +
> >>>> +static void dealloc_vector(struct mlx5_vdpa_net *ndev,
> >>>> + struct mlx5_vdpa_virtqueue *mvq)
> >>>> +{
> >>>> + struct mlx5_vdpa_irq_pool *irqp = &ndev->irqp;
> >>>> + int i;
> >>>> +
> >>>> + for (i = 0; i < irqp->num_ent; i++)
> >>>> + if (mvq->map.virq ==
irqp->entries[i].map.virq)
> >>>> + irqp->entries[i].usecount--;
> >>>> +}
> >>>> +
> >>>> static int setup_vq(struct mlx5_vdpa_net *ndev, struct
> >>>> mlx5_vdpa_virtqueue *mvq)
> >>>> {
> >>>> u16 idx = mvq->index;
> >>>> @@ -1203,27 +1244,31 @@ static int setup_vq(struct
mlx5_vdpa_net
> >>>> *ndev, struct mlx5_vdpa_virtqueue *mvq)
> >>>> err = counter_set_alloc(ndev, mvq);
> >>>> if (err)
> >>>> - goto err_counter;
> >>>> + goto err_connect;
> >>>> + alloc_vector(ndev, mvq);
> >>>> err = create_virtqueue(ndev, mvq);
> >>>> if (err)
> >>>> - goto err_connect;
> >>>> + goto err_vq;
> >>>> if (mvq->ready) {
> >>>> err = modify_virtqueue(ndev, mvq,
> >>>> MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
> >>>> if (err) {
> >>>> mlx5_vdpa_warn(&ndev->mvdev,
"failed to modify to ready
> >>>> vq idx %d(%d)\n",
> >>>> idx, err);
> >>>> - goto err_connect;
> >>>> + goto err_modify;
> >>>> }
> >>>> }
> >>>> mvq->initialized = true;
> >>>> return 0;
> >>>> -err_connect:
> >>>> +err_modify:
> >>>> + destroy_virtqueue(ndev, mvq);
> >>>> +err_vq:
> >>>> + dealloc_vector(ndev, mvq);
> >>>> counter_set_dealloc(ndev, mvq);
> >>>> -err_counter:
> >>>> +err_connect:
> >>>> qp_destroy(ndev, &mvq->vqqp);
> >>>> err_vqqp:
> >>>> qp_destroy(ndev, &mvq->fwqp);
> >>>> @@ -1267,6 +1312,7 @@ static void teardown_vq(struct
mlx5_vdpa_net
> >>>> *ndev, struct mlx5_vdpa_virtqueue *
> >>>> return;
> >>>> suspend_vq(ndev, mvq);
> >>>> + dealloc_vector(ndev, mvq);
> >>>> destroy_virtqueue(ndev, mvq);
> >>>> counter_set_dealloc(ndev, mvq);
> >>>> qp_destroy(ndev, &mvq->vqqp);
> >>>> @@ -2374,6 +2420,7 @@ static int save_channel_info(struct
> >>>> mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqu
> >>>> ri->desc_addr = mvq->desc_addr;
> >>>> ri->device_addr = mvq->device_addr;
> >>>> ri->driver_addr = mvq->driver_addr;
> >>>> + ri->map = mvq->map;
> >>>> ri->restore = true;
> >>>> return 0;
> >>>> }
> >>>> @@ -2418,6 +2465,7 @@ static void
restore_channels_info(struct
> >>>> mlx5_vdpa_net *ndev)
> >>>> mvq->desc_addr = ri->desc_addr;
> >>>> mvq->device_addr = ri->device_addr;
> >>>> mvq->driver_addr = ri->driver_addr;
> >>>> + mvq->map = ri->map;
> >>>> }
> >>>> }
> >>>> @@ -2693,6 +2741,22 @@ static struct device
> >>>> *mlx5_get_vq_dma_dev(struct vdpa_device *vdev, u16 idx)
> >>>> return mvdev->vdev.dma_dev;
> >>>> }
> >>>> +static void free_irqs(struct mlx5_vdpa_net *ndev)
> >>>> +{
> >>>> + struct mlx5_vdpa_irq_pool_entry *ent;
> >>>> + int i;
> >>>> +
> >>>> + if (!msix_mode_supported(&ndev->mvdev))
> >>>> + return;
> >>>> +
> >>>> + for (i = ndev->irqp.num_ent - 1; i >= 0; i--) {
> >>>> + ent = ndev->irqp.entries + i;
> >>>> + mlx5_msix_free(ndev->mvdev.mdev, ent->map);
> >>>> + ndev->irqp.num_ent--;
> >>>> + }
> >>>> + kfree(ndev->irqp.entries);
> >>>> +}
> >>>> +
> >>>> static void mlx5_vdpa_free(struct vdpa_device *vdev)
> >>>> {
> >>>> struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >>>> @@ -2708,6 +2772,7 @@ static void mlx5_vdpa_free(struct
vdpa_device
> >>>> *vdev)
> >>>> mlx5_mpfs_del_mac(pfmdev, ndev->config.mac);
> >>>> }
> >>>> mlx5_vdpa_free_resources(&ndev->mvdev);
> >>>> + free_irqs(ndev);
> >>>> kfree(ndev->event_cbs);
> >>>> kfree(ndev->vqs);
> >>>> }
> >>>> @@ -2736,9 +2801,23 @@ static struct
vdpa_notification_area
> >>>> mlx5_get_vq_notification(struct vdpa_device
> >>>> return ret;
> >>>> }
> >>>> -static int mlx5_get_vq_irq(struct vdpa_device *vdv,
u16 idx)
> >>>> +static int mlx5_get_vq_irq(struct vdpa_device *vdev, u16
idx)
> >>>> {
> >>>> - return -EOPNOTSUPP;
> >>>> + struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >>>> + struct mlx5_vdpa_net *ndev =
to_mlx5_vdpa_ndev(mvdev);
> >>>> + struct mlx5_vdpa_virtqueue *mvq;
> >>>> +
> >>>> + if (!is_index_valid(mvdev, idx))
> >>>> + return -EINVAL;
> >>>> +
> >>>> + if (is_ctrl_vq_idx(mvdev, idx))
> >>>> + return -EOPNOTSUPP;
> >>>> +
> >>>> + mvq = &ndev->vqs[idx];
> >>>> + if (!mvq->map.virq)
> >>>> + return -EOPNOTSUPP;
> >>>> +
> >>>> + return mvq->map.virq;
> >>>> }
> >>>> static u64 mlx5_vdpa_get_driver_features(struct
vdpa_device *vdev)
> >>>> @@ -3095,6 +3174,35 @@ static int config_func_mtu(struct
> >>>> mlx5_core_dev *mdev, u16 mtu)
> >>>> return err;
> >>>> }
> >>>> +static irqreturn_t int_handler(int irq, void *nh)
> >>>> +{
> >>>> + return IRQ_HANDLED;
> >>>
> >>> I'm not sure I get the meaning of the irq handler. Note
that even if
> >>> we support get_vq_irq() it does not means the posted interrupt
can
> >>> work for each interrupt. It work only if several conditions
are met
> >>> (for example the VCPU is running).
> >> Not sure I get it. Can you be explicit what needs to be done to
make
> >> sure the interrupts will go through directly to the vCPU?
> > Firstly, the code should work in the setups without PI, e.g for VMX it
> > has a check vmx_can_use_vtd_pi():
> >
> > static bool vmx_can_use_vtd_pi(struct kvm *kvm)
> > {
> > return irqchip_in_kernel(kvm) && enable_apicv
&&
> > kvm_arch_has_assigned_device(kvm) &&
> > irq_remapping_cap(IRQ_POSTING_CAP);
> > }
> >
> > We can see it might only work when
> >
> > - irq chip is emulated in the kernel
> > - apicv is enabled
> > - IRQ remapping is supported
>
>
> OK, so we probably need a public API to tell you that. I am not aware of
> any. Do you think we should introduce one?
We don't need an API to do that, we can simply fallback to the slow
patch which is triggering the callback in the interrupt handler. Then
everything is fine.
>
> BTW, I didn't see intel ifcvf making those checks. Does it mean the
> driver is broken or is there anything I am missing.
>
> >
> > And what's more even if we had all of those, it needs to satisfy
extra
> > requirements like virq affinity but the virq affinity is under the
> > control of the guest, so we can't guarantee that the post
interrupt
> > can work for each time. See the comments in vmx_pi_update_irte():
> >
> > /*
> > * VT-d PI cannot support posting multicast/broadcast
> > * interrupts to a vCPU, we still use interrupt
remapping
> > * for these kind of interrupts.
> > *
> > * For lowest-priority interrupts, we only support
> > * those with single CPU as the destination, e.g.
user
> > * configures the interrupts via /proc/irq or uses
> > * irqbalance to make the interrupts single-CPU.
> > *
> > * We will support full lowest-priority interrupt
later.
> > *
> > * In addition, we can only inject generic interrupts
using
> > * the PI mechanism, refuse to route others through
it.
> > */
> >
> > So I think what I don't understand is how the code works if PI is
not supported?
>
> So it means that on the guest we need to make sure the affinity of each
> VQ interrupts is affiliated to a single CPU (eg. one bit only set).
>
> BTW, on my system the mask is 0xff and I don't have any issue with
> posted interrupts.
Such affinity is determined by the guest driver (e.g for Linux, users
could tweak it via procfs). So we can't have such an assumption.
Thanks
>
> >
> > Thanks
> >
> >>> So if this is the fallback irq handler when we can do post
interrupt,
> >>> should we trigger the virtqueue callback here?
> >> It's not a fallback. I should remove it completely.
> >>
> >>>
> >>>> +}
> >>>> +
> >>>> +static void allocate_irqs(struct mlx5_vdpa_net *ndev)
> >>>> +{
> >>>> + struct mlx5_vdpa_irq_pool_entry *ent;
> >>>> + int i;
> >>>> +
> >>>> + if (!msix_mode_supported(&ndev->mvdev))
> >>>> + return;
> >>>> +
> >>>> + ndev->irqp.entries =
kcalloc(ndev->mvdev.max_vqs,
> >>>> sizeof(*ndev->irqp.entries), GFP_KERNEL);
> >>>> + if (!ndev->irqp.entries)
> >>>> + return;
> >>>> +
> >>>> + for (i = 0; i < ndev->mvdev.max_vqs; i++) {
> >>>> + ent = ndev->irqp.entries + i;
> >>>> + snprintf(ent->name, MLX5_VDPA_IRQ_NAME_LEN,
"%s-vq-%d",
> >>>> + dev_name(&ndev->mvdev.vdev.dev), i);
> >>>> + ent->map =
mlx5_msix_alloc(ndev->mvdev.mdev, int_handler,
> >>>> NULL, ent->name);
> >>>> + if (!ent->map.virq)
> >>>> + return;
> >>>> +
> >>>> + ndev->irqp.num_ent++;
> >>>> + }
> >>>> +}
> >>>> +
> >>>> static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev
*v_mdev, const
> >>>> char *name,
> >>>> const struct vdpa_dev_set_config
*add_config)
> >>>> {
> >>>> @@ -3171,6 +3279,7 @@ static int mlx5_vdpa_dev_add(struct
> >>>> vdpa_mgmt_dev *v_mdev, const char *name,
> >>>> }
> >>>> init_mvqs(ndev);
> >>>> + allocate_irqs(ndev);
> >>>> init_rwsem(&ndev->reslock);
> >>>> config = &ndev->config;
> >>>> @@ -3358,6 +3467,17 @@ static void mlx5v_remove(struct
> >>>> auxiliary_device *adev)
> >>>> kfree(mgtdev);
> >>>> }
> >>>> +static void mlx5v_shutdown(struct auxiliary_device
*auxdev)
> >>>> +{
> >>>> + struct mlx5_vdpa_mgmtdev *mgtdev;
> >>>> + struct mlx5_vdpa_net *ndev;
> >>>> +
> >>>> + mgtdev = auxiliary_get_drvdata(auxdev);
> >>>> + ndev = mgtdev->ndev;
> >>>> +
> >>>> + free_irqs(ndev);
> >>>> +}
> >>>> +
> >>>> static const struct auxiliary_device_id
mlx5v_id_table[] = {
> >>>> { .name = MLX5_ADEV_NAME ".vnet", },
> >>>> {},
> >>>> @@ -3369,6 +3489,7 @@ static struct auxiliary_driver
mlx5v_driver = {
> >>>> .name = "vnet",
> >>>> .probe = mlx5v_probe,
> >>>> .remove = mlx5v_remove,
> >>>> + .shutdown = mlx5v_shutdown,
> >>>
> >>> So we allocate the irq during dev_add, it seems cleaner if we
do that
> >>> in the dev_del.
> >> Not sure what you mean here. You're saying I should not call
free_irqs()
> >> in mlx5v_shutdown()
> >>> Note that mlx5v_remove will call vdpa_mgmtdev_unregister()
which will
> >>> call dev_del.
> >>>
> >>> Thanks
> >>>
> >>>
> >>>> .id_table = mlx5v_id_table,
> >>>> };
> >>>> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.h
> >>>> b/drivers/vdpa/mlx5/net/mlx5_vnet.h
> >>>> index c90a89e1de4d..e5063b310d3c 100644
> >>>> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.h
> >>>> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.h
> >>>> @@ -26,6 +26,19 @@ static inline u16 key2vid(u64 key)
> >>>> return (u16)(key >> 48) & 0xfff;
> >>>> }
> >>>> +#define MLX5_VDPA_IRQ_NAME_LEN 32
> >>>> +
> >>>> +struct mlx5_vdpa_irq_pool_entry {
> >>>> + struct msi_map map;
> >>>> + int usecount;
> >>>> + char name[MLX5_VDPA_IRQ_NAME_LEN];
> >>>> +};
> >>>> +
> >>>> +struct mlx5_vdpa_irq_pool {
> >>>> + int num_ent;
> >>>> + struct mlx5_vdpa_irq_pool_entry *entries;
> >>>> +};
> >>>> +
> >>>> struct mlx5_vdpa_net {
> >>>> struct mlx5_vdpa_dev mvdev;
> >>>> struct mlx5_vdpa_net_resources res;
> >>>> @@ -49,6 +62,7 @@ struct mlx5_vdpa_net {
> >>>> struct vdpa_callback config_cb;
> >>>> struct mlx5_vdpa_wq_ent cvq_ent;
> >>>> struct hlist_head macvlan_hash[MLX5V_MACVLAN_SIZE];
> >>>> + struct mlx5_vdpa_irq_pool irqp;
> >>>> struct dentry *debugfs;
> >>>> };
>