[bug report] RDMA/mlx5: Add support in MEMIC operations

[bug report] RDMA/mlx5: Add support in MEMIC operations

[Dan Carpenter]

Hello Maor Gottlieb,

The patch cea85fa5dbc2: "RDMA/mlx5: Add support in MEMIC operations"
from Apr 11, 2021, leads to the following static checker warning:

	drivers/infiniband/hw/mlx5/dm.c:220 mlx5_ib_handler_MLX5_IB_METHOD_DM_MAP_OP_ADDR()
	error: undefined (user controlled) shift '(((1))) << op'

drivers/infiniband/hw/mlx5/dm.c
   204  static int UVERBS_HANDLER(MLX5_IB_METHOD_DM_MAP_OP_ADDR)(
   205          struct uverbs_attr_bundle *attrs)
   206  {
   207          struct ib_uobject *uobj = uverbs_attr_get_uobject(
   208                  attrs, MLX5_IB_ATTR_DM_MAP_OP_ADDR_REQ_HANDLE);
   209          struct mlx5_ib_dev *dev = to_mdev(uobj->context->device);
   210          struct ib_dm *ibdm = uobj->object;
   211          struct mlx5_ib_dm_memic *dm = to_memic(ibdm);
   212          struct mlx5_ib_dm_op_entry *op_entry;
   213          int err;
   214          u8 op;
   215  
   216          err = uverbs_copy_from(&op, attrs, MLX5_IB_ATTR_DM_MAP_OP_ADDR_REQ_OP);
                                        ^^
op is user controlled and in the 0-255 range.

   217          if (err)
   218                  return err;
   219  
   220          if (!(MLX5_CAP_DEV_MEM(dev->mdev, memic_operations) & BIT(op)))
                                                                      ^^^^^^^
If it's more than 31 then this is undefined (shift wrapping generally).
Plus it might trigger a UBSan warning at run time.

   221                  return -EOPNOTSUPP;
   222  
   223          mutex_lock(&dm->ops_xa_lock);

regards,
dan carpenter

Re: [bug report] RDMA/mlx5: Add support in MEMIC operations

[Leon Romanovsky]

On Tue, May 04, 2021 at 10:50:44AM +0300, Dan Carpenter wrote:
> Hello Maor Gottlieb,
> 
> The patch cea85fa5dbc2: "RDMA/mlx5: Add support in MEMIC operations"
> from Apr 11, 2021, leads to the following static checker warning:
> 
> 	drivers/infiniband/hw/mlx5/dm.c:220 mlx5_ib_handler_MLX5_IB_METHOD_DM_MAP_OP_ADDR()
> 	error: undefined (user controlled) shift '(((1))) << op'
> 
> drivers/infiniband/hw/mlx5/dm.c
>    204  static int UVERBS_HANDLER(MLX5_IB_METHOD_DM_MAP_OP_ADDR)(
>    205          struct uverbs_attr_bundle *attrs)
>    206  {
>    207          struct ib_uobject *uobj = uverbs_attr_get_uobject(
>    208                  attrs, MLX5_IB_ATTR_DM_MAP_OP_ADDR_REQ_HANDLE);
>    209          struct mlx5_ib_dev *dev = to_mdev(uobj->context->device);
>    210          struct ib_dm *ibdm = uobj->object;
>    211          struct mlx5_ib_dm_memic *dm = to_memic(ibdm);
>    212          struct mlx5_ib_dm_op_entry *op_entry;
>    213          int err;
>    214          u8 op;
>    215  
>    216          err = uverbs_copy_from(&op, attrs, MLX5_IB_ATTR_DM_MAP_OP_ADDR_REQ_OP);
>                                         ^^
> op is user controlled and in the 0-255 range.
> 
>    217          if (err)
>    218                  return err;
>    219  
>    220          if (!(MLX5_CAP_DEV_MEM(dev->mdev, memic_operations) & BIT(op)))
>                                                                       ^^^^^^^
> If it's more than 31 then this is undefined (shift wrapping generally).
> Plus it might trigger a UBSan warning at run time.


Thanks Dan, we will prepare the proper patch.
It should be something like this:
diff --git a/drivers/infiniband/hw/mlx5/dm.c b/drivers/infiniband/hw/mlx5/dm.c
index 094bf85589db..dd4480aed1aa 100644
--- a/drivers/infiniband/hw/mlx5/dm.c
+++ b/drivers/infiniband/hw/mlx5/dm.c
@@ -217,6 +217,9 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_DM_MAP_OP_ADDR)(
        if (err)
                return err;

+       if (op > 31)
+               return -EINVAL;
+
        if (!(MLX5_CAP_DEV_MEM(dev->mdev, memic_operations) & BIT(op)))
                return -EOPNOTSUPP;


> 
>    221                  return -EOPNOTSUPP;
>    222  
>    223          mutex_lock(&dm->ops_xa_lock);
> 
> regards,
> dan carpenter