Commit 7affd808 authored by Rahul Lakkireddy's avatar Rahul Lakkireddy Committed by David S. Miller

cxgb4: fix throughput drop during Tx backpressure

commit 7c3bebc3 ("cxgb4: request the TX CIDX updates to status page")
reverted back to getting Tx CIDX updates via DMA, instead of interrupts,
introduced by commit d429005f ("cxgb4/cxgb4vf: Add support for SGE
doorbell queue timer")

However, it missed reverting back several code changes where Tx CIDX
updates are not explicitly requested during backpressure when using
interrupt mode. These missed changes cause slow recovery during
backpressure because the corresponding interrupt no longer comes and
hence results in Tx throughput drop.

So, revert back these missed code changes, as well, which will allow
explicitly requesting Tx CIDX updates when backpressure happens.
This enables the corresponding interrupt with Tx CIDX update message
to get generated and hence speed up recovery and restore back
throughput.

Fixes: 7c3bebc3 ("cxgb4: request the TX CIDX updates to status page")
Fixes: d429005f ("cxgb4/cxgb4vf: Add support for SGE doorbell queue timer")
Signed-off-by: default avatarRahul Lakkireddy <rahul.lakkireddy@chelsio.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 22259471
...@@ -1486,16 +1486,7 @@ static netdev_tx_t cxgb4_eth_xmit(struct sk_buff *skb, struct net_device *dev) ...@@ -1486,16 +1486,7 @@ static netdev_tx_t cxgb4_eth_xmit(struct sk_buff *skb, struct net_device *dev)
* has opened up. * has opened up.
*/ */
eth_txq_stop(q); eth_txq_stop(q);
wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
/* If we're using the SGE Doorbell Queue Timer facility, we
* don't need to ask the Firmware to send us Egress Queue CIDX
* Updates: the Hardware will do this automatically. And
* since we send the Ingress Queue CIDX Updates to the
* corresponding Ethernet Response Queue, we'll get them very
* quickly.
*/
if (!q->dbqt)
wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
} }
wr = (void *)&q->q.desc[q->q.pidx]; wr = (void *)&q->q.desc[q->q.pidx];
...@@ -1805,16 +1796,7 @@ static netdev_tx_t cxgb4_vf_eth_xmit(struct sk_buff *skb, ...@@ -1805,16 +1796,7 @@ static netdev_tx_t cxgb4_vf_eth_xmit(struct sk_buff *skb,
* has opened up. * has opened up.
*/ */
eth_txq_stop(txq); eth_txq_stop(txq);
wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
/* If we're using the SGE Doorbell Queue Timer facility, we
* don't need to ask the Firmware to send us Egress Queue CIDX
* Updates: the Hardware will do this automatically. And
* since we send the Ingress Queue CIDX Updates to the
* corresponding Ethernet Response Queue, we'll get them very
* quickly.
*/
if (!txq->dbqt)
wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
} }
/* Start filling in our Work Request. Note that we do _not_ handle /* Start filling in our Work Request. Note that we do _not_ handle
...@@ -3370,26 +3352,6 @@ static void t4_tx_completion_handler(struct sge_rspq *rspq, ...@@ -3370,26 +3352,6 @@ static void t4_tx_completion_handler(struct sge_rspq *rspq,
} }
txq = &s->ethtxq[pi->first_qset + rspq->idx]; txq = &s->ethtxq[pi->first_qset + rspq->idx];
/* We've got the Hardware Consumer Index Update in the Egress Update
* message. If we're using the SGE Doorbell Queue Timer mechanism,
* these Egress Update messages will be our sole CIDX Updates we get
* since we don't want to chew up PCIe bandwidth for both Ingress
* Messages and Status Page writes. However, The code which manages
* reclaiming successfully DMA'ed TX Work Requests uses the CIDX value
* stored in the Status Page at the end of the TX Queue. It's easiest
* to simply copy the CIDX Update value from the Egress Update message
* to the Status Page. Also note that no Endian issues need to be
* considered here since both are Big Endian and we're just copying
* bytes consistently ...
*/
if (txq->dbqt) {
struct cpl_sge_egr_update *egr;
egr = (struct cpl_sge_egr_update *)rsp;
WRITE_ONCE(txq->q.stat->cidx, egr->cidx);
}
t4_sge_eth_txq_egress_update(adapter, txq, -1); t4_sge_eth_txq_egress_update(adapter, txq, -1);
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment