Re: [PATCH mlx5-next v1] net/mlx5: Check device memory pointer before usage

[Leon Romanovsky <leon@kernel.org>]

On Wed, Jul 02, 2025 at 03:07:35PM +0100, Simon Horman wrote:
> On Wed, Jul 02, 2025 at 11:28:47AM +0300, Leon Romanovsky wrote:
> > On Tue, Jul 01, 2025 at 08:38:58PM +0100, Simon Horman wrote:
> > > On Tue, Jul 01, 2025 at 03:08:12PM +0300, Leon Romanovsky wrote:
> > > > From: Stav Aviram <saviram@nvidia.com>
> > > > 
> > > > Add a NULL check before accessing device memory to prevent a crash if
> > > > dev->dm allocation in mlx5_init_once() fails.
> > > > 
> > > > Fixes: c9b9dcb430b3 ("net/mlx5: Move device memory management to mlx5_core")
> > > > Signed-off-by: Stav Aviram <saviram@nvidia.com>
> > > > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > > > ---
> > > > Changelog:
> > > > v1:
> > > >  * Removed extra IS_ERR(dm) check.
> > > > v0:
> > > > https://lore.kernel.org/all/e389fa6ef075af1049cd7026b912d736ebe3ad23.1751279408.git.leonro@nvidia.com
> > > > ---
> > > >  drivers/infiniband/hw/mlx5/dm.c                  | 2 +-
> > > >  drivers/net/ethernet/mellanox/mlx5/core/lib/dm.c | 4 ++--
> > > >  drivers/net/ethernet/mellanox/mlx5/core/main.c   | 2 +-
> > > >  3 files changed, 4 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/drivers/infiniband/hw/mlx5/dm.c b/drivers/infiniband/hw/mlx5/dm.c
> > > > index b4c97fb62abf..9ded2b7c1e31 100644
> > > > --- a/drivers/infiniband/hw/mlx5/dm.c
> > > > +++ b/drivers/infiniband/hw/mlx5/dm.c
> > > > @@ -282,7 +282,7 @@ static struct ib_dm *handle_alloc_dm_memic(struct ib_ucontext *ctx,
> > > >  	int err;
> > > >  	u64 address;
> > > >  
> > > > -	if (!MLX5_CAP_DEV_MEM(dm_db->dev, memic))
> > > > +	if (!dm_db || !MLX5_CAP_DEV_MEM(dm_db->dev, memic))
> > > >  		return ERR_PTR(-EOPNOTSUPP);
> > > 
> > > nit: -EOPNOTSUPP doesn't feel like the right error code
> > >      in the !dm_db case.
> > 
> > Why? This error is returned to the user through mlx5_ib_alloc_dm().
> 
> Because, as I understand things, such a case would be due to a memory
> allocation failure, not by the device not supporting a feature.
> 
> handle_alloc_dm_memic() already returns ERR_PTR(-ENOMEM) if kzalloc() fails.
> I'd suggest doing so for the !dm_db case too.

!dm_db case can be because of missing capabilities and EOPNOTSUPP is a
way to inform users about it.

> 
> But I don't feel particularly strongly about this.