Message ID | 1318264943-10009-1-git-send-email-cascardo@linux.vnet.ibm.com (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
On Mon, Oct 10, 2011 at 01:42:23PM -0300, Thadeu Lima de Souza Cascardo wrote: > The doorbell register was being unconditionally swapped. In x86, that > meant it was being swapped to BE and written to the descriptor and to > memory, depending on the case of blue frame support or writing to > doorbell register. On PPC, this meant it was being swapped to LE and > then swapped back to BE while writing to the register. But in the blue > frame case, it was being written as LE to the descriptor. > > The fix is not to swap doorbell unconditionally, write it to the > register as BE and convert it to BE when writing it to the descriptor. > > Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com> > Reported-by: Richard Hendrickson <richhend@us.ibm.com> > Cc: Eli Cohen <eli@dev.mellanox.co.il> > Cc: Yevgeny Petrilin <yevgenyp@mellanox.co.il> > Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> > --- So I tested this patch and it works for me. Thanks Ben and Eli for finding out the problem with doorbell in the descriptor. Regards, Cascardo. > drivers/net/mlx4/en_tx.c | 6 +++--- > 1 files changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/net/mlx4/en_tx.c b/drivers/net/mlx4/en_tx.c > index 6e03de0..f76ab6b 100644 > --- a/drivers/net/mlx4/en_tx.c > +++ b/drivers/net/mlx4/en_tx.c > @@ -172,7 +172,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, > memset(ring->buf, 0, ring->buf_size); > > ring->qp_state = MLX4_QP_STATE_RST; > - ring->doorbell_qpn = swab32(ring->qp.qpn << 8); > + ring->doorbell_qpn = ring->qp.qpn << 8; > > mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn, > ring->cqn, &ring->context); > @@ -791,7 +791,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev) > skb_orphan(skb); > > if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tag) { > - *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn; > + *(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn); > op_own |= htonl((bf_index & 0xffff) << 8); > /* Ensure new descirptor hits memory > * before setting ownership of this descriptor to HW */ > @@ -812,7 +812,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev) > wmb(); > tx_desc->ctrl.owner_opcode = op_own; > wmb(); > - writel(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL); > + iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL); > } > > /* Poll CQ here */ > -- > 1.7.4.4 >
From: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com> Date: Mon, 10 Oct 2011 13:46:54 -0300 > On Mon, Oct 10, 2011 at 01:42:23PM -0300, Thadeu Lima de Souza Cascardo wrote: >> The doorbell register was being unconditionally swapped. In x86, that >> meant it was being swapped to BE and written to the descriptor and to >> memory, depending on the case of blue frame support or writing to >> doorbell register. On PPC, this meant it was being swapped to LE and >> then swapped back to BE while writing to the register. But in the blue >> frame case, it was being written as LE to the descriptor. >> >> The fix is not to swap doorbell unconditionally, write it to the >> register as BE and convert it to BE when writing it to the descriptor. >> >> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com> >> Reported-by: Richard Hendrickson <richhend@us.ibm.com> >> Cc: Eli Cohen <eli@dev.mellanox.co.il> >> Cc: Yevgeny Petrilin <yevgenyp@mellanox.co.il> >> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> >> --- > > So I tested this patch and it works for me. Thanks Ben and Eli for > finding out the problem with doorbell in the descriptor. Applied, thanks everyone.
diff --git a/drivers/net/mlx4/en_tx.c b/drivers/net/mlx4/en_tx.c index 6e03de0..f76ab6b 100644 --- a/drivers/net/mlx4/en_tx.c +++ b/drivers/net/mlx4/en_tx.c @@ -172,7 +172,7 @@ int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv, memset(ring->buf, 0, ring->buf_size); ring->qp_state = MLX4_QP_STATE_RST; - ring->doorbell_qpn = swab32(ring->qp.qpn << 8); + ring->doorbell_qpn = ring->qp.qpn << 8; mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn, ring->cqn, &ring->context); @@ -791,7 +791,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev) skb_orphan(skb); if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tag) { - *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn; + *(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn); op_own |= htonl((bf_index & 0xffff) << 8); /* Ensure new descirptor hits memory * before setting ownership of this descriptor to HW */ @@ -812,7 +812,7 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev) wmb(); tx_desc->ctrl.owner_opcode = op_own; wmb(); - writel(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL); + iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL); } /* Poll CQ here */
The doorbell register was being unconditionally swapped. In x86, that meant it was being swapped to BE and written to the descriptor and to memory, depending on the case of blue frame support or writing to doorbell register. On PPC, this meant it was being swapped to LE and then swapped back to BE while writing to the register. But in the blue frame case, it was being written as LE to the descriptor. The fix is not to swap doorbell unconditionally, write it to the register as BE and convert it to BE when writing it to the descriptor. Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com> Reported-by: Richard Hendrickson <richhend@us.ibm.com> Cc: Eli Cohen <eli@dev.mellanox.co.il> Cc: Yevgeny Petrilin <yevgenyp@mellanox.co.il> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> --- drivers/net/mlx4/en_tx.c | 6 +++--- 1 files changed, 3 insertions(+), 3 deletions(-)