Re: [dpdk-dev] [PATCH] eal: move DMA mapping from bus-specific to generic driver

[Thomas Monjalon <thomas@monjalon.net>]

01/04/2021 00:45, Thomas Monjalon:
> The operations of DMA mapping and unmapping are controlled in some
> bus drivers, following rte_bus specification.
> If the device driver don't provide any specific mapping operation,
> the bus driver may have a fallback (VFIO case for PCI).
> 
> The DMA mapping done by the device drivers are called
> from the bus drivers via function pointers in bus-specific structures:
> rte_vdev_driver and rte_pci_driver.
> 
> The device driver DMA mapping is not specific to the bus,
> so it can be generalized to all device drivers, based on rte_device.
> That's why the function pointers dma_map and dma_unmap
> are moved to rte_driver, avoiding useless casts of device object.
> 
> The function prototypes rte_dev_dma_map_t and rte_dev_dma_unmap_t
> are removed from rte_bus.h because the definition in rte_dev.h is enough,
> but they are still used in rte_bus for bus drivers,
> while being added in rte_driver for device drivers,
> and removed from rte_vdev_driver/rte_pci_driver.
> 
> The impacted device drivers are mlx5 (PCI) and virtio_user (vdev).
> 
> Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> ---
> Depends-on: series-16017 ("pci: add rte prefix")
> ---
>  drivers/bus/pci/pci_common.c            |  8 ++--
>  drivers/bus/pci/rte_bus_pci.h           | 40 --------------------
>  drivers/bus/vdev/rte_bus_vdev.h         | 40 --------------------
>  drivers/bus/vdev/vdev.c                 | 32 +++++-----------
>  drivers/common/mlx5/mlx5_common_pci.c   | 30 ++++++++-------
>  drivers/net/mlx5/mlx5.c                 |  4 +-
>  drivers/net/mlx5/mlx5_mr.c              | 50 +++++++++++++------------
>  drivers/net/mlx5/mlx5_rxtx.h            |  4 +-
>  drivers/net/virtio/virtio_user_ethdev.c | 22 +++++------
>  lib/librte_eal/include/rte_bus.h        | 42 ---------------------
>  lib/librte_eal/include/rte_dev.h        | 49 +++++++++++++++++++++++-
>  11 files changed, 117 insertions(+), 204 deletions(-)

The ABI checker reports some issues on the driver interface.
It needs to be carefully analyzed, because driver interface
should not be part of the ABI compatibility contract.