As discussed in the previous email, I had an minor & optional comment:

Personally, I think the code would read better if we just get rid of
the push_free state and use 2 separate code paths:

release_and_unlock:
idpf_vc_xn_release_bufs(xn);
idpf_vc_xn_unlock(xn);
idpf_vc_xn_push_free(&adapter->vcxn_mngr, xn);
return retval;

only_unlock:
idpf_vc_xn_unlock(xn);
return retval;

But I don't have strong opinions whether or not the suggestion above
is taken, and the patch LGTM.

Thank you Emil!

Reviewed-by: Li Li <[email protected]>



On Mon, Mar 16, 2026 at 4:28 PM Emil Tantilov <[email protected]> wrote:
>
> Refactor the VC logic dealing with transaction accounting where
> both push and pop_free are using the same spinlock free_xn_bm.
> This resolves potential race when setting and clearing the bits
> in the free_xn_bm bitmask.
>
> Fixes: 34c21fa894a1 ("idpf: implement virtchnl transaction manager")
> Reported-by: Ray Zhang <[email protected]>
> Signed-off-by: Emil Tantilov <[email protected]>
> Reviewed-by: Aleksandr Loktionov <[email protected]>
> ---
>  drivers/net/ethernet/intel/idpf/idpf_virtchnl.c | 17 ++++++++++++++---
>  1 file changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c 
> b/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
> index 113ecfc16dd7..21a6c9d22085 100644
> --- a/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
> +++ b/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
> @@ -401,12 +401,18 @@ struct idpf_vc_xn *idpf_vc_xn_pop_free(struct 
> idpf_vc_xn_manager *vcxn_mngr)
>   * idpf_vc_xn_push_free - Push a free transaction to free list
>   * @vcxn_mngr: transaction manager to push to
>   * @xn: transaction to push
> + *
> + * Callers must ensure idpf_vc_xn_release_bufs() has been called (under
> + * idpf_vc_xn_lock) before invoking this function. This function must
> + * be called without holding idpf_vc_xn_lock to avoid nesting a sleepable
> + * spinlock inside a raw_spinlock on PREEMPT_RT kernels.
>   */
>  static void idpf_vc_xn_push_free(struct idpf_vc_xn_manager *vcxn_mngr,
>                                  struct idpf_vc_xn *xn)
>  {
> -       idpf_vc_xn_release_bufs(xn);
> +       spin_lock_bh(&vcxn_mngr->xn_bm_lock);
>         set_bit(xn->idx, vcxn_mngr->free_xn_bm);
> +       spin_unlock_bh(&vcxn_mngr->xn_bm_lock);
>  }
>
>  /**
> @@ -428,6 +434,7 @@ ssize_t idpf_vc_xn_exec(struct idpf_adapter *adapter,
>                         const struct idpf_vc_xn_params *params)
>  {
>         const struct kvec *send_buf = &params->send_buf;
> +       bool push_free = false;
>         struct idpf_vc_xn *xn;
>         ssize_t retval;
>         u16 cookie;
> @@ -514,10 +521,13 @@ ssize_t idpf_vc_xn_exec(struct idpf_adapter *adapter,
>         }
>
>  release_and_unlock:
> -       idpf_vc_xn_push_free(adapter->vcxn_mngr, xn);
> +       idpf_vc_xn_release_bufs(xn);
> +       push_free = true;
>         /* If we receive a VC reply after here, it will be dropped. */
>  only_unlock:
>         idpf_vc_xn_unlock(xn);
> +       if (push_free)
> +               idpf_vc_xn_push_free(adapter->vcxn_mngr, xn);
>
>         return retval;
>  }
> @@ -559,7 +569,7 @@ idpf_vc_xn_forward_async(struct idpf_adapter *adapter, 
> struct idpf_vc_xn *xn,
>         }
>
>  release_bufs:
> -       idpf_vc_xn_push_free(adapter->vcxn_mngr, xn);
> +       idpf_vc_xn_release_bufs(xn);
>
>         return err;
>  }
> @@ -619,6 +629,7 @@ idpf_vc_xn_forward_reply(struct idpf_adapter *adapter,
>         case IDPF_VC_XN_ASYNC:
>                 err = idpf_vc_xn_forward_async(adapter, xn, ctlq_msg);
>                 idpf_vc_xn_unlock(xn);
> +               idpf_vc_xn_push_free(adapter->vcxn_mngr, xn);
>                 return err;
>         default:
>                 dev_err_ratelimited(&adapter->pdev->dev, "Overwriting VC 
> reply (op %d)\n",
> --
> 2.37.3
>

Reply via email to