Re: [PATCH net-next v2 12/14] sfc: unmap VF's MCDI buffer when switching to vDPA mode

[Gautam Dawar <gdawar@amd.com>]

On 3/10/23 10:35, Jason Wang wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> On Tue, Mar 7, 2023 at 7:38 PM Gautam Dawar <gautam.dawar@amd.com> wrote:
>> To avoid clash of IOVA range of VF's MCDI DMA buffer with the guest
>> buffer IOVAs, unmap the MCDI buffer when switching to vDPA mode
>> and use PF's IOMMU domain for running VF's MCDI commands.
>>
>> Signed-off-by: Gautam Dawar <gautam.dawar@amd.com>
>> ---
>>   drivers/net/ethernet/sfc/ef100_nic.c      |  1 -
>>   drivers/net/ethernet/sfc/ef100_vdpa.c     | 25 ++++++++++++++++
>>   drivers/net/ethernet/sfc/ef100_vdpa.h     |  3 ++
>>   drivers/net/ethernet/sfc/ef100_vdpa_ops.c | 35 +++++++++++++++++++++++
>>   drivers/net/ethernet/sfc/mcdi.h           |  3 ++
>>   drivers/net/ethernet/sfc/net_driver.h     | 12 ++++++++
>>   6 files changed, 78 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/ethernet/sfc/ef100_nic.c b/drivers/net/ethernet/sfc/ef100_nic.c
>> index cd9f724a9e64..1bffc1994ed8 100644
>> --- a/drivers/net/ethernet/sfc/ef100_nic.c
>> +++ b/drivers/net/ethernet/sfc/ef100_nic.c
>> @@ -33,7 +33,6 @@
>>
>>   #define EF100_MAX_VIS 4096
>>   #define EF100_NUM_MCDI_BUFFERS 1
>> -#define MCDI_BUF_LEN (8 + MCDI_CTL_SDU_LEN_MAX)
>>
>>   #define EF100_RESET_PORT ((ETH_RESET_MAC | ETH_RESET_PHY) << ETH_RESET_SHARED_SHIFT)
>>
>> diff --git a/drivers/net/ethernet/sfc/ef100_vdpa.c b/drivers/net/ethernet/sfc/ef100_vdpa.c
>> index 5c9f29f881a6..30ca4ab00175 100644
>> --- a/drivers/net/ethernet/sfc/ef100_vdpa.c
>> +++ b/drivers/net/ethernet/sfc/ef100_vdpa.c
>> @@ -223,10 +223,19 @@ static int vdpa_allocate_vis(struct efx_nic *efx, unsigned int *allocated_vis)
>>   static void ef100_vdpa_delete(struct efx_nic *efx)
>>   {
>>          struct vdpa_device *vdpa_dev;
>> +       int rc;
>>
>>          if (efx->vdpa_nic) {
>>                  vdpa_dev = &efx->vdpa_nic->vdpa_dev;
>>                  ef100_vdpa_reset(vdpa_dev);
>> +               if (efx->mcdi_buf_mode == EFX_BUF_MODE_VDPA) {
>> +                       rc = ef100_vdpa_map_mcdi_buffer(efx);
>> +                       if (rc) {
>> +                               pci_err(efx->pci_dev,
>> +                                       "map_mcdi_buffer failed, err: %d\n",
>> +                                       rc);
>> +                       }
>> +               }
>>
>>                  /* replace with _vdpa_unregister_device later */
>>                  put_device(&vdpa_dev->dev);
>> @@ -276,6 +285,21 @@ static int get_net_config(struct ef100_vdpa_nic *vdpa_nic)
>>          return 0;
>>   }
>>
>> +static void unmap_mcdi_buffer(struct efx_nic *efx)
>> +{
>> +       struct ef100_nic_data *nic_data = efx->nic_data;
>> +       struct efx_mcdi_iface *mcdi;
>> +
>> +       mcdi = efx_mcdi(efx);
>> +       spin_lock_bh(&mcdi->iface_lock);
>> +       /* Save current MCDI mode to be restored later */
>> +       efx->vdpa_nic->mcdi_mode = mcdi->mode;
>> +       efx->mcdi_buf_mode = EFX_BUF_MODE_VDPA;
>> +       mcdi->mode = MCDI_MODE_FAIL;
>> +       spin_unlock_bh(&mcdi->iface_lock);
>> +       efx_nic_free_buffer(efx, &nic_data->mcdi_buf);
>> +}
>> +
>>   static struct ef100_vdpa_nic *ef100_vdpa_create(struct efx_nic *efx,
>>                                                  const char *dev_name,
>>                                                  enum ef100_vdpa_class dev_type,
>> @@ -342,6 +366,7 @@ static struct ef100_vdpa_nic *ef100_vdpa_create(struct efx_nic *efx,
>>          for (i = 0; i < EF100_VDPA_MAC_FILTER_NTYPES; i++)
>>                  vdpa_nic->filters[i].filter_id = EFX_INVALID_FILTER_ID;
>>
>> +       unmap_mcdi_buffer(efx);
>>          rc = get_net_config(vdpa_nic);
>>          if (rc)
>>                  goto err_put_device;
>> diff --git a/drivers/net/ethernet/sfc/ef100_vdpa.h b/drivers/net/ethernet/sfc/ef100_vdpa.h
>> index 49fb6be04eb3..0913ac2519cb 100644
>> --- a/drivers/net/ethernet/sfc/ef100_vdpa.h
>> +++ b/drivers/net/ethernet/sfc/ef100_vdpa.h
>> @@ -147,6 +147,7 @@ struct ef100_vdpa_filter {
>>    * @status: device status as per VIRTIO spec
>>    * @features: negotiated feature bits
>>    * @max_queue_pairs: maximum number of queue pairs supported
>> + * @mcdi_mode: MCDI mode at the time of unmapping VF mcdi buffer
>>    * @net_config: virtio_net_config data
>>    * @vring: vring information of the vDPA device.
>>    * @mac_address: mac address of interface associated with this vdpa device
>> @@ -166,6 +167,7 @@ struct ef100_vdpa_nic {
>>          u8 status;
>>          u64 features;
>>          u32 max_queue_pairs;
>> +       enum efx_mcdi_mode mcdi_mode;
>>          struct virtio_net_config net_config;
>>          struct ef100_vdpa_vring_info vring[EF100_VDPA_MAX_QUEUES_PAIRS * 2];
>>          u8 *mac_address;
>> @@ -185,6 +187,7 @@ int ef100_vdpa_add_filter(struct ef100_vdpa_nic *vdpa_nic,
>>   int ef100_vdpa_init_vring(struct ef100_vdpa_nic *vdpa_nic, u16 idx);
>>   void ef100_vdpa_irq_vectors_free(void *data);
>>   int ef100_vdpa_reset(struct vdpa_device *vdev);
>> +int ef100_vdpa_map_mcdi_buffer(struct efx_nic *efx);
>>
>>   static inline bool efx_vdpa_is_little_endian(struct ef100_vdpa_nic *vdpa_nic)
>>   {
>> diff --git a/drivers/net/ethernet/sfc/ef100_vdpa_ops.c b/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> index db86c2693950..c6c9458f0e6f 100644
>> --- a/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> +++ b/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> @@ -711,12 +711,47 @@ static int ef100_vdpa_suspend(struct vdpa_device *vdev)
>>          mutex_unlock(&vdpa_nic->lock);
>>          return rc;
>>   }
>> +
>> +int ef100_vdpa_map_mcdi_buffer(struct efx_nic *efx)
>> +{
> The name of this function is confusing, it's actually map buffer for
> ef100 netdev mode.

Yeah, I get your point. Actually, I am using the prefix ef100_vdpa_ for 
extern functions and this function is just mapping the  MCDI buffer 
resulting in the name ef100_vdpa_map_mcdi_buffer().

I think ef100_vdpa_map_vf_mcdi_buffer() would remove this confusion. 
What do you think?

>
> Actually, I wonder why not moving this to init/fini of bar config ops
> or if we use aux bus, it should be done during aux driver
> probe/remove.
It makes sense, however we store the last mcdi mode (poll or events) in 
vdpa_nic->mcdi_mode to restore later, which requires vdpa_nic to be 
allocated that happens later than switching to vdpa bar_config.
> Thanks
>
>
>> +       struct ef100_nic_data *nic_data = efx->nic_data;
>> +       struct efx_mcdi_iface *mcdi;
>> +       int rc;
>> +
>> +       /* Update VF's MCDI buffer when switching out of vdpa mode */
>> +       rc = efx_nic_alloc_buffer(efx, &nic_data->mcdi_buf,
>> +                                 MCDI_BUF_LEN, GFP_KERNEL);
>> +       if (rc)
>> +               return rc;
>> +
>> +       mcdi = efx_mcdi(efx);
>> +       spin_lock_bh(&mcdi->iface_lock);
>> +       mcdi->mode = efx->vdpa_nic->mcdi_mode;
>> +       efx->mcdi_buf_mode = EFX_BUF_MODE_EF100;
>> +       spin_unlock_bh(&mcdi->iface_lock);
>> +
>> +       return 0;
>> +}
>> +
>>   static void ef100_vdpa_free(struct vdpa_device *vdev)
>>   {
>>          struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +       int rc;
>>          int i;
>>
>>          if (vdpa_nic) {
>> +               if (vdpa_nic->efx->mcdi_buf_mode == EFX_BUF_MODE_VDPA) {
>> +                       /* This will only be called via call to put_device()
>> +                        * on vdpa device creation failure
>> +                        */
>> +                       rc = ef100_vdpa_map_mcdi_buffer(vdpa_nic->efx);
>> +                       if (rc) {
>> +                               dev_err(&vdev->dev,
>> +                                       "map_mcdi_buffer failed, err: %d\n",
>> +                                       rc);
>> +                       }
>> +               }
>> +
>>                  for (i = 0; i < (vdpa_nic->max_queue_pairs * 2); i++) {
>>                          reset_vring(vdpa_nic, i);
>>                          if (vdpa_nic->vring[i].vring_ctx)
>> diff --git a/drivers/net/ethernet/sfc/mcdi.h b/drivers/net/ethernet/sfc/mcdi.h
>> index 2c526d2edeb6..bc4de3b4e6f3 100644
>> --- a/drivers/net/ethernet/sfc/mcdi.h
>> +++ b/drivers/net/ethernet/sfc/mcdi.h
>> @@ -6,6 +6,9 @@
>>
>>   #ifndef EFX_MCDI_H
>>   #define EFX_MCDI_H
>> +#include "mcdi_pcol.h"
>> +
>> +#define MCDI_BUF_LEN (8 + MCDI_CTL_SDU_LEN_MAX)
>>
>>   /**
>>    * enum efx_mcdi_state - MCDI request handling state
>> diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
>> index 948c7a06403a..9cdfeb6ad05a 100644
>> --- a/drivers/net/ethernet/sfc/net_driver.h
>> +++ b/drivers/net/ethernet/sfc/net_driver.h
>> @@ -848,6 +848,16 @@ enum efx_xdp_tx_queues_mode {
>>
>>   struct efx_mae;
>>
>> +/**
>> + * enum efx_buf_alloc_mode - buffer allocation mode
>> + * @EFX_BUF_MODE_EF100: buffer setup in ef100 mode
>> + * @EFX_BUF_MODE_VDPA: buffer setup in vdpa mode
>> + */
>> +enum efx_buf_alloc_mode {
>> +       EFX_BUF_MODE_EF100,
>> +       EFX_BUF_MODE_VDPA
>> +};
>> +
>>   /**
>>    * struct efx_nic - an Efx NIC
>>    * @name: Device name (net device name or bus id before net device registered)
>> @@ -877,6 +887,7 @@ struct efx_mae;
>>    * @irq_rx_mod_step_us: Step size for IRQ moderation for RX event queues
>>    * @irq_rx_moderation_us: IRQ moderation time for RX event queues
>>    * @msg_enable: Log message enable flags
>> + * @mcdi_buf_mode: mcdi buffer allocation mode
>>    * @state: Device state number (%STATE_*). Serialised by the rtnl_lock.
>>    * @reset_pending: Bitmask for pending resets
>>    * @tx_queue: TX DMA queues
>> @@ -1045,6 +1056,7 @@ struct efx_nic {
>>          u32 msg_enable;
>>
>>          enum nic_state state;
>> +       enum efx_buf_alloc_mode mcdi_buf_mode;
>>          unsigned long reset_pending;
>>
>>          struct efx_channel *channel[EFX_MAX_CHANNELS];
>> --
>> 2.30.1
>>