Re: [PATCH for-rc 1/9] RDMA/hns: Check atomic wr length

[Leon Romanovsky <leon@kernel.org>]

On Fri, Jul 05, 2024 at 04:59:29PM +0800, Junxian Huang wrote:
> 8 bytes is the only supported length of atomic. Return an error if
> it is not.
> 
> Fixes: 384f88185112 ("RDMA/hns: Add atomic support")
> Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
> ---
>  drivers/infiniband/hw/hns/hns_roce_device.h |  2 ++
>  drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 19 +++++++++++++++----
>  2 files changed, 17 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
> index ff0b3f68ee3a..05005079258c 100644
> --- a/drivers/infiniband/hw/hns/hns_roce_device.h
> +++ b/drivers/infiniband/hw/hns/hns_roce_device.h
> @@ -91,6 +91,8 @@
>  /* Configure to HW for PAGE_SIZE larger than 4KB */
>  #define PG_SHIFT_OFFSET				(PAGE_SHIFT - 12)
>  
> +#define ATOMIC_WR_LEN				8
> +
>  #define HNS_ROCE_IDX_QUE_ENTRY_SZ		4
>  #define SRQ_DB_REG				0x230
>  
> diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
> index 4287818a737f..a5d746a5cc68 100644
> --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
> +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
> @@ -164,15 +164,23 @@ static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
>  	hr_reg_clear(fseg, FRMR_BLK_MODE);
>  }
>  
> -static void set_atomic_seg(const struct ib_send_wr *wr,
> -			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
> -			   unsigned int valid_num_sge)
> +static int set_atomic_seg(struct hns_roce_dev *hr_dev,
> +			  const struct ib_send_wr *wr,
> +			  struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
> +			  unsigned int valid_num_sge, u32 msg_len)
>  {
>  	struct hns_roce_v2_wqe_data_seg *dseg =
>  		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
>  	struct hns_roce_wqe_atomic_seg *aseg =
>  		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
>  
> +	if (msg_len != ATOMIC_WR_LEN) {
> +		ibdev_err_ratelimited(&hr_dev->ib_dev,
> +				      "invalid atomic wr len, len = %u.\n",
> +				      msg_len);
> +		return -EINVAL;

1. Please don't add prints in data-path.
2. You most likely need to add this check before calling to set_atomic_seg().
3. You shouldn't continue to process the WQE if the length is invalid.
Need to return from set_rc_wqe() and not continue.
4. I wonder if it is right place to put this limitation and can't be
enforced much earlier.

Thanks