Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.37 as permitted sender) client-ip=23.128.96.37;
Message-ID: <a1f8b9f8c2f9aecde8ac17831b66f72319bf425a.camel@linux.ibm.com>
Subject: Re: [PATCH net] net/mlx5: fix calling mlx5_cmd_init() before DMA
 mask is set
From:   Niklas Schnelle <schnelle@linux.ibm.com>
To:     Leon Romanovsky <leon@kernel.org>
Cc:     Saeed Mahameed <saeedm@nvidia.com>, Jason Gunthorpe <jgg@ziepe.ca>,
        Matthew Rosato <mjrosato@linux.ibm.com>,
        Joerg Roedel <joro@8bytes.org>,
        Robin Murphy <robin.murphy@arm.com>,
        "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>, Shay Drory <shayd@nvidia.com>,
        Moshe Shemesh <moshe@nvidia.com>,
        Heiko Carstens <hca@linux.ibm.com>,
        Alexander Gordeev <agordeev@linux.ibm.com>,
        linux-s390@vger.kernel.org, netdev@vger.kernel.org,
        linux-rdma@vger.kernel.org, linux-kernel@vger.kernel.org
Date:   Fri, 29 Sep 2023 11:40:49 +0200
In-Reply-To: <20230928175959.GU1642130@unreal>
References: <20230928-mlx5_init_fix-v1-1-79749d45ce60@linux.ibm.com>
         <20230928175959.GU1642130@unreal>
Content-Type: text/plain; charset="UTF-8"
User-Agent: Evolution 3.48.4 (3.48.4-1.fc38) 
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0
Precedence: bulk

On Thu, 2023-09-28 at 20:59 +0300, Leon Romanovsky wrote:
> On Thu, Sep 28, 2023 at 03:55:47PM +0200, Niklas Schnelle wrote:
> > Since commit 06cd555f73ca ("net/mlx5: split mlx5_cmd_init() to probe and
> > reload routines") mlx5_cmd_init() is called in mlx5_mdev_init() which is
> > called in probe_one() before mlx5_pci_init(). This is a problem because
> > mlx5_pci_init() is where the DMA and coherent mask is set but
> > mlx5_cmd_init() already does a dma_alloc_coherent(). Thus a DMA
> > allocation is done during probe before the correct mask is set. This
> > causes probe to fail initialization of the cmdif SW structs on s390x
> > after that is converted to the common dma-iommu code. This is because on
> > s390x DMA addresses below 4 GiB are reserved on current machines and
> > unlike the old s390x specific DMA API implementation common code
> > enforces DMA masks. Fix this by switching the order of the
> > mlx5_mdev_init() and mlx5_pci_init() in probe_one().
> >=20
> > Link: https://lore.kernel.org/linux-iommu/cfc9e9128ed5571d2e36421e34730=
1057662a09e.camel@linux.ibm.com/
> > Fixes: 06cd555f73ca ("net/mlx5: split mlx5_cmd_init() to probe and relo=
ad routines")
> > Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
> > ---
> > Note: I ran into this while testing the linked series for converting
> > s390x to use dma-iommu. The existing s390x specific DMA API
> > implementation doesn't respect DMA masks and is thus not affected
> > despite of course also only supporting DMA addresses above 4 GiB.
> > That said ConnectX VFs are the primary users of native PCI on s390x and
> > we'd really like to get the DMA API conversion into v6.7 so this has
> > high priority for us.
> > ---
> >  drivers/net/ethernet/mellanox/mlx5/core/main.c | 12 ++++++------
> >  1 file changed, 6 insertions(+), 6 deletions(-)
> >=20
> > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/n=
et/ethernet/mellanox/mlx5/core/main.c
> > index 15561965d2af..06744dedd928 100644
> > --- a/drivers/net/ethernet/mellanox/mlx5/core/main.c
> > +++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c
> > @@ -1908,10 +1908,6 @@ static int probe_one(struct pci_dev *pdev, const=
 struct pci_device_id *id)
> >  		goto adev_init_err;
> >  	}
> >=20=20
> > -	err =3D mlx5_mdev_init(dev, prof_sel);
> > -	if (err)
> > -		goto mdev_init_err;
> > -
> >  	err =3D mlx5_pci_init(dev, pdev, id);
> >  	if (err) {
> >  		mlx5_core_err(dev, "mlx5_pci_init failed with error code %d\n",
> > @@ -1919,6 +1915,10 @@ static int probe_one(struct pci_dev *pdev, const=
 struct pci_device_id *id)
> >  		goto pci_init_err;
> >  	}
> >=20=20
> > +	err =3D mlx5_mdev_init(dev, prof_sel);
> > +	if (err)
> > +		goto mdev_init_err;
> > +
>=20
> I had something different in mind as I'm worry that call to pci_enable_de=
vice()
> in mlx5_pci_init() before we finished FW command interface initialization=
 is a bit
> premature.
>=20
> What about the following patch?
>=20
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/net=
/ethernet/mellanox/mlx5/core/main.c
> index 15561965d2af..31f1d701116a 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/main.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c
> @@ -905,12 +905,6 @@ static int mlx5_pci_init(struct mlx5_core_dev *dev, =
struct pci_dev *pdev,
>=20=20
>         pci_set_master(pdev);
>=20=20
> -       err =3D set_dma_caps(pdev);
> -       if (err) {
> -               mlx5_core_err(dev, "Failed setting DMA capabilities mask,=
 aborting\n");
> -               goto err_clr_master;
> -       }
> -
>         if (pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_CO=
MP32) &&
>             pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_CO=
MP64) &&
>             pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_CO=
MP128))
> @@ -1908,9 +1902,15 @@ static int probe_one(struct pci_dev *pdev, const s=
truct pci_device_id *id)
>                 goto adev_init_err;
>         }
>=20=20
> +       err =3D set_dma_caps(pdev);
> +       if (err) {
> +               mlx5_core_err(dev, "Failed setting DMA capabilities mask,=
 aborting\n");
> +               goto dma_cap_err;
> +       }
> +
>         err =3D mlx5_mdev_init(dev, prof_sel);
>         if (err)
> -               goto mdev_init_err;
> +               goto dma_cap_err;
>=20=20
>         err =3D mlx5_pci_init(dev, pdev, id);
>         if (err) {
> @@ -1942,7 +1942,7 @@ static int probe_one(struct pci_dev *pdev, const st=
ruct pci_device_id *id)
>         mlx5_pci_close(dev);
>  pci_init_err:
>         mlx5_mdev_uninit(dev);
> -mdev_init_err:
> +dma_cap_err:
>         mlx5_adev_idx_free(dev->priv.adev_idx);
>  adev_init_err:
>         mlx5_devlink_free(devlink);
>=20
> Thanks

The above works too. Maybe for consistency within probe_one() it would
then make sense to also rename set_dma_caps() to mlx5_dma_init()?

Thanks,
Niklas