| Message ID | 1433250405-2888-4-git-send-email-igvtee@gmail.com |
|---|---|
| State | Accepted |
| Headers | show |
On 2015-06-02 15:06, michael lee wrote: > use pktgen to verify on rt3662. can improve transmit rate. > pkt_size 1500 > burst 1 : 807Mb/sec > burst 8 : 984Mb/sec > > pkt_size 60 > burst 1 : 57Mb/sec > burst 8 : 236Mb/sec > > Signed-off-by: michael lee <igvtee@gmail.com> > --- > .../drivers/net/ethernet/ralink/ralink_soc_eth.c | 51 +++++++++++++--------- > .../drivers/net/ethernet/ralink/ralink_soc_eth.h | 2 + > 2 files changed, 33 insertions(+), 20 deletions(-) > > diff --git a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c > index b2304bb..4b39825 100644 > --- a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c > +++ b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c > @@ -337,6 +337,8 @@ static int fe_alloc_tx(struct fe_priv *priv) > struct fe_tx_ring *ring = &priv->tx_ring; > > ring->tx_free_idx = 0; > + ring->tx_next_idx = 0; > + ring->tx_thresh = max((unsigned long)ring->tx_ring_size >> 2, MAX_SKB_FRAGS); > > ring->tx_buf = kcalloc(ring->tx_ring_size, sizeof(*ring->tx_buf), > GFP_KERNEL); > @@ -525,8 +527,16 @@ static int fe_vlan_rx_kill_vid(struct net_device *dev, > return 0; > } > > +static inline u32 fe_empty_txd(struct fe_tx_ring *ring) > +{ > + barrier(); What is this barrier for? > + return (u32)(ring->tx_ring_size - > + ((ring->tx_next_idx - ring->tx_free_idx) & > + (ring->tx_ring_size - 1))); > +} > + > static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, > - int idx, int tx_num, struct fe_tx_ring *ring) > + int tx_num, struct fe_tx_ring *ring) > { > struct fe_priv *priv = netdev_priv(dev); > struct skb_frag_struct *frag; > @@ -649,14 +659,22 @@ static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, > netdev_sent_queue(dev, skb->len); > skb_tx_timestamp(skb); > > - j = NEXT_TX_DESP_IDX(j); > + ring->tx_next_idx = NEXT_TX_DESP_IDX(j); > wmb(); > - fe_reg_w32(j, FE_REG_TX_CTX_IDX0); > + if (unlikely(fe_empty_txd(ring) <= ring->tx_thresh)) { > + netif_stop_queue(dev); > + smp_mb(); > + if (unlikely(fe_empty_txd(ring) > ring->tx_thresh)) > + netif_wake_queue(dev); Why do that queue wake check here, when it's already being done during tx cleanup. > + } > + > + if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)) || !skb->xmit_more) > + fe_reg_w32(ring->tx_next_idx, FE_REG_TX_CTX_IDX0); > > return 0; > > err_dma: > - j = idx; > + j = ring->tx_next_idx; > for (i = 0; i < tx_num; i++) { > ptxd = &ring->tx_dma[j]; > tx_buf = &ring->tx_buf[j]; - Felix
this feature i reference tg3.c driver at function tg3_start_xmit.
it says therer is a small possibility that start_xmit will miss it
and cause the queue to be stopped forever.
below is the tg3 drivers code
=== function tg3_tx ===
tnapi->tx_cons = sw_idx;
/* Need to make the tx_cons update visible to tg3_start_xmit()
* before checking for netif_queue_stopped(). Without the
* memory barrier, there is a small possibility that
tg3_start_xmit()
* will miss it and cause the queue to be stopped forever.
*/
smp_mb();
if (unlikely(netif_tx_queue_stopped(txq) &&
(tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi)))) {
__netif_tx_lock(txq, smp_processor_id());
if (netif_tx_queue_stopped(txq) &&
(tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi)))
netif_tx_wake_queue(txq);
__netif_tx_unlock(txq);
}
=== tg3_start_xmit ===
budget = tg3_tx_avail(tnapi);
/* We are running in BH disabled context with netif_tx_lock
* and TX reclaim runs via tp->napi.poll inside of a software
* interrupt. Furthermore, IRQ processing runs lockless so we have
* no IRQ context deadlocks to worry about either. Rejoice!
*/
if (unlikely(budget <= (skb_shinfo(skb)->nr_frags + 1))) {
if (!netif_tx_queue_stopped(txq)) {
netif_tx_stop_queue(txq);
/* This is a hard error, log it. */
netdev_err(dev,
"BUG! Tx Ring full when queue awake!\n");
}
return NETDEV_TX_BUSY;
}
2015-06-05 16:40 GMT+08:00 Felix Fietkau <nbd@openwrt.org>:
> On 2015-06-02 15:06, michael lee wrote:
> > use pktgen to verify on rt3662. can improve transmit rate.
> > pkt_size 1500
> > burst 1 : 807Mb/sec
> > burst 8 : 984Mb/sec
> >
> > pkt_size 60
> > burst 1 : 57Mb/sec
> > burst 8 : 236Mb/sec
> >
> > Signed-off-by: michael lee <igvtee@gmail.com>
> > ---
> > .../drivers/net/ethernet/ralink/ralink_soc_eth.c | 51
> +++++++++++++---------
> > .../drivers/net/ethernet/ralink/ralink_soc_eth.h | 2 +
> > 2 files changed, 33 insertions(+), 20 deletions(-)
> >
> > diff --git
> a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c
> b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c
> > index b2304bb..4b39825 100644
> > ---
> a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c
> > +++
> b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c
> > @@ -337,6 +337,8 @@ static int fe_alloc_tx(struct fe_priv *priv)
> > struct fe_tx_ring *ring = &priv->tx_ring;
> >
> > ring->tx_free_idx = 0;
> > + ring->tx_next_idx = 0;
> > + ring->tx_thresh = max((unsigned long)ring->tx_ring_size >> 2,
> MAX_SKB_FRAGS);
> >
> > ring->tx_buf = kcalloc(ring->tx_ring_size, sizeof(*ring->tx_buf),
> > GFP_KERNEL);
> > @@ -525,8 +527,16 @@ static int fe_vlan_rx_kill_vid(struct net_device
> *dev,
> > return 0;
> > }
> >
> > +static inline u32 fe_empty_txd(struct fe_tx_ring *ring)
> > +{
> > + barrier();
> What is this barrier for?
>
> > + return (u32)(ring->tx_ring_size -
> > + ((ring->tx_next_idx - ring->tx_free_idx) &
> > + (ring->tx_ring_size - 1)));
> > +}
> > +
> > static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev,
> > - int idx, int tx_num, struct fe_tx_ring *ring)
> > + int tx_num, struct fe_tx_ring *ring)
> > {
> > struct fe_priv *priv = netdev_priv(dev);
> > struct skb_frag_struct *frag;
> > @@ -649,14 +659,22 @@ static int fe_tx_map_dma(struct sk_buff *skb,
> struct net_device *dev,
> > netdev_sent_queue(dev, skb->len);
> > skb_tx_timestamp(skb);
> >
> > - j = NEXT_TX_DESP_IDX(j);
> > + ring->tx_next_idx = NEXT_TX_DESP_IDX(j);
> > wmb();
> > - fe_reg_w32(j, FE_REG_TX_CTX_IDX0);
> > + if (unlikely(fe_empty_txd(ring) <= ring->tx_thresh)) {
> > + netif_stop_queue(dev);
> > + smp_mb();
> > + if (unlikely(fe_empty_txd(ring) > ring->tx_thresh))
> > + netif_wake_queue(dev);
> Why do that queue wake check here, when it's already being done during
> tx cleanup.
>
> > + }
> > +
> > + if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)) ||
> !skb->xmit_more)
> > + fe_reg_w32(ring->tx_next_idx, FE_REG_TX_CTX_IDX0);
> >
> > return 0;
> >
> > err_dma:
> > - j = idx;
> > + j = ring->tx_next_idx;
> > for (i = 0; i < tx_num; i++) {
> > ptxd = &ring->tx_dma[j];
> > tx_buf = &ring->tx_buf[j];
>
> - Felix
>
Dear Sir. have any suggestion? Best Regards. 2015-06-05 18:09 GMT+08:00 Mingyu Li <igvtee@gmail.com>: > this feature i reference tg3.c driver at function tg3_start_xmit. > it says therer is a small possibility that start_xmit will miss it > and cause the queue to be stopped forever. > > below is the tg3 drivers code > === function tg3_tx === > tnapi->tx_cons = sw_idx; > > /* Need to make the tx_cons update visible to tg3_start_xmit() > * before checking for netif_queue_stopped(). Without the > * memory barrier, there is a small possibility that > tg3_start_xmit() > * will miss it and cause the queue to be stopped forever. > */ > smp_mb(); > > if (unlikely(netif_tx_queue_stopped(txq) && > (tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi)))) > { > __netif_tx_lock(txq, smp_processor_id()); > if (netif_tx_queue_stopped(txq) && > (tg3_tx_avail(tnapi) > TG3_TX_WAKEUP_THRESH(tnapi))) > netif_tx_wake_queue(txq); > __netif_tx_unlock(txq); > } > > === tg3_start_xmit === > budget = tg3_tx_avail(tnapi); > > /* We are running in BH disabled context with netif_tx_lock > * and TX reclaim runs via tp->napi.poll inside of a software > * interrupt. Furthermore, IRQ processing runs lockless so we have > * no IRQ context deadlocks to worry about either. Rejoice! > */ > if (unlikely(budget <= (skb_shinfo(skb)->nr_frags + 1))) { > if (!netif_tx_queue_stopped(txq)) { > netif_tx_stop_queue(txq); > > /* This is a hard error, log it. */ > netdev_err(dev, > "BUG! Tx Ring full when queue > awake!\n"); > } > return NETDEV_TX_BUSY; > } > > > 2015-06-05 16:40 GMT+08:00 Felix Fietkau <nbd@openwrt.org>: > >> On 2015-06-02 15:06, michael lee wrote: >> > use pktgen to verify on rt3662. can improve transmit rate. >> > pkt_size 1500 >> > burst 1 : 807Mb/sec >> > burst 8 : 984Mb/sec >> > >> > pkt_size 60 >> > burst 1 : 57Mb/sec >> > burst 8 : 236Mb/sec >> > >> > Signed-off-by: michael lee <igvtee@gmail.com> >> > --- >> > .../drivers/net/ethernet/ralink/ralink_soc_eth.c | 51 >> +++++++++++++--------- >> > .../drivers/net/ethernet/ralink/ralink_soc_eth.h | 2 + >> > 2 files changed, 33 insertions(+), 20 deletions(-) >> > >> > diff --git >> a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c >> b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c >> > index b2304bb..4b39825 100644 >> > --- >> a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c >> > +++ >> b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c >> > @@ -337,6 +337,8 @@ static int fe_alloc_tx(struct fe_priv *priv) >> > struct fe_tx_ring *ring = &priv->tx_ring; >> > >> > ring->tx_free_idx = 0; >> > + ring->tx_next_idx = 0; >> > + ring->tx_thresh = max((unsigned long)ring->tx_ring_size >> 2, >> MAX_SKB_FRAGS); >> > >> > ring->tx_buf = kcalloc(ring->tx_ring_size, sizeof(*ring->tx_buf), >> > GFP_KERNEL); >> > @@ -525,8 +527,16 @@ static int fe_vlan_rx_kill_vid(struct net_device >> *dev, >> > return 0; >> > } >> > >> > +static inline u32 fe_empty_txd(struct fe_tx_ring *ring) >> > +{ >> > + barrier(); >> What is this barrier for? >> >> > + return (u32)(ring->tx_ring_size - >> > + ((ring->tx_next_idx - ring->tx_free_idx) & >> > + (ring->tx_ring_size - 1))); >> > +} >> > + >> > static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, >> > - int idx, int tx_num, struct fe_tx_ring *ring) >> > + int tx_num, struct fe_tx_ring *ring) >> > { >> > struct fe_priv *priv = netdev_priv(dev); >> > struct skb_frag_struct *frag; >> > @@ -649,14 +659,22 @@ static int fe_tx_map_dma(struct sk_buff *skb, >> struct net_device *dev, >> > netdev_sent_queue(dev, skb->len); >> > skb_tx_timestamp(skb); >> > >> > - j = NEXT_TX_DESP_IDX(j); >> > + ring->tx_next_idx = NEXT_TX_DESP_IDX(j); >> > wmb(); >> > - fe_reg_w32(j, FE_REG_TX_CTX_IDX0); >> > + if (unlikely(fe_empty_txd(ring) <= ring->tx_thresh)) { >> > + netif_stop_queue(dev); >> > + smp_mb(); >> > + if (unlikely(fe_empty_txd(ring) > ring->tx_thresh)) >> > + netif_wake_queue(dev); >> Why do that queue wake check here, when it's already being done during >> tx cleanup. >> >> > + } >> > + >> > + if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)) || >> !skb->xmit_more) >> > + fe_reg_w32(ring->tx_next_idx, FE_REG_TX_CTX_IDX0); >> > >> > return 0; >> > >> > err_dma: >> > - j = idx; >> > + j = ring->tx_next_idx; >> > for (i = 0; i < tx_num; i++) { >> > ptxd = &ring->tx_dma[j]; >> > tx_buf = &ring->tx_buf[j]; >> >> - Felix >> > >
diff --git a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c index b2304bb..4b39825 100644 --- a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c +++ b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.c @@ -337,6 +337,8 @@ static int fe_alloc_tx(struct fe_priv *priv) struct fe_tx_ring *ring = &priv->tx_ring; ring->tx_free_idx = 0; + ring->tx_next_idx = 0; + ring->tx_thresh = max((unsigned long)ring->tx_ring_size >> 2, MAX_SKB_FRAGS); ring->tx_buf = kcalloc(ring->tx_ring_size, sizeof(*ring->tx_buf), GFP_KERNEL); @@ -525,8 +527,16 @@ static int fe_vlan_rx_kill_vid(struct net_device *dev, return 0; } +static inline u32 fe_empty_txd(struct fe_tx_ring *ring) +{ + barrier(); + return (u32)(ring->tx_ring_size - + ((ring->tx_next_idx - ring->tx_free_idx) & + (ring->tx_ring_size - 1))); +} + static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, - int idx, int tx_num, struct fe_tx_ring *ring) + int tx_num, struct fe_tx_ring *ring) { struct fe_priv *priv = netdev_priv(dev); struct skb_frag_struct *frag; @@ -537,7 +547,7 @@ static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, u32 def_txd4; int i, j, k, frag_size, frag_map_size, offset; - tx_buf = &ring->tx_buf[idx]; + tx_buf = &ring->tx_buf[ring->tx_next_idx]; memset(tx_buf, 0, sizeof(*tx_buf)); memset(&txd, 0, sizeof(txd)); nr_frags = skb_shinfo(skb)->nr_frags; @@ -589,7 +599,7 @@ static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, dma_unmap_len_set(tx_buf, dma_len0, skb_headlen(skb)); /* TX SG offload */ - j = idx; + j = ring->tx_next_idx; k = 0; for (i = 0; i < nr_frags; i++) { offset = 0; @@ -649,14 +659,22 @@ static int fe_tx_map_dma(struct sk_buff *skb, struct net_device *dev, netdev_sent_queue(dev, skb->len); skb_tx_timestamp(skb); - j = NEXT_TX_DESP_IDX(j); + ring->tx_next_idx = NEXT_TX_DESP_IDX(j); wmb(); - fe_reg_w32(j, FE_REG_TX_CTX_IDX0); + if (unlikely(fe_empty_txd(ring) <= ring->tx_thresh)) { + netif_stop_queue(dev); + smp_mb(); + if (unlikely(fe_empty_txd(ring) > ring->tx_thresh)) + netif_wake_queue(dev); + } + + if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)) || !skb->xmit_more) + fe_reg_w32(ring->tx_next_idx, FE_REG_TX_CTX_IDX0); return 0; err_dma: - j = idx; + j = ring->tx_next_idx; for (i = 0; i < tx_num; i++) { ptxd = &ring->tx_dma[j]; tx_buf = &ring->tx_buf[j]; @@ -703,12 +721,6 @@ static inline int fe_skb_padto(struct sk_buff *skb, struct fe_priv *priv) { return ret; } -static inline u32 fe_empty_txd(struct fe_tx_ring *ring, u32 tx_fill_idx) -{ - return (u32)(ring->tx_ring_size - ((tx_fill_idx - ring->tx_free_idx) & - (ring->tx_ring_size - 1))); -} - static inline int fe_cal_txd_req(struct sk_buff *skb) { int i, nfrags; @@ -732,7 +744,6 @@ static int fe_start_xmit(struct sk_buff *skb, struct net_device *dev) struct fe_priv *priv = netdev_priv(dev); struct fe_tx_ring *ring = &priv->tx_ring; struct net_device_stats *stats = &dev->stats; - u32 tx; int tx_num; int len = skb->len; @@ -742,8 +753,7 @@ static int fe_start_xmit(struct sk_buff *skb, struct net_device *dev) } tx_num = fe_cal_txd_req(skb); - tx = fe_reg_r32(FE_REG_TX_CTX_IDX0); - if (unlikely(fe_empty_txd(ring, tx) <= tx_num)) + if (unlikely(fe_empty_txd(ring) <= tx_num)) { netif_stop_queue(dev); netif_err(priv, tx_queued,dev, @@ -751,7 +761,7 @@ static int fe_start_xmit(struct sk_buff *skb, struct net_device *dev) return NETDEV_TX_BUSY; } - if (fe_tx_map_dma(skb, dev, tx, tx_num, ring) < 0) { + if (fe_tx_map_dma(skb, dev, tx_num, ring) < 0) { stats->tx_dropped++; } else { stats->tx_packets++; @@ -916,10 +926,10 @@ static int fe_poll_tx(struct fe_priv *priv, int budget, u32 tx_intr, if (done) { netdev_completed_queue(netdev, done, bytes_compl); + smp_mb(); if (unlikely(netif_queue_stopped(netdev) && - netif_carrier_ok(netdev))) { + (fe_empty_txd(ring) > ring->tx_thresh))) netif_wake_queue(netdev); - } } return done; @@ -990,12 +1000,13 @@ static void fe_tx_timeout(struct net_device *dev) netif_info(priv, drv, dev, "dma_cfg:%08x\n", fe_reg_r32(FE_REG_PDMA_GLO_CFG)); netif_info(priv, drv, dev, "tx_ring=%d, " \ - "base=%08x, max=%u, ctx=%u, dtx=%u, fdx=%d\n", 0, + "base=%08x, max=%u, ctx=%u, dtx=%u, fdx=%hu, next=%hu\n", 0, fe_reg_r32(FE_REG_TX_BASE_PTR0), fe_reg_r32(FE_REG_TX_MAX_CNT0), fe_reg_r32(FE_REG_TX_CTX_IDX0), fe_reg_r32(FE_REG_TX_DTX_IDX0), - ring->tx_free_idx + ring->tx_free_idx, + ring->tx_next_idx ); netif_info(priv, drv, dev, "rx_ring=%d, " \ "base=%08x, max=%u, calc=%u, drx=%u\n", 0, diff --git a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.h b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.h index 6614c76..5d4220f 100644 --- a/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.h +++ b/target/linux/ramips/files/drivers/net/ethernet/ralink/ralink_soc_eth.h @@ -461,6 +461,8 @@ struct fe_tx_ring dma_addr_t tx_phys; u16 tx_ring_size; u16 tx_free_idx; + u16 tx_next_idx; + u16 tx_thresh; }; struct fe_priv
use pktgen to verify on rt3662. can improve transmit rate. pkt_size 1500 burst 1 : 807Mb/sec burst 8 : 984Mb/sec pkt_size 60 burst 1 : 57Mb/sec burst 8 : 236Mb/sec Signed-off-by: michael lee <igvtee@gmail.com> --- .../drivers/net/ethernet/ralink/ralink_soc_eth.c | 51 +++++++++++++--------- .../drivers/net/ethernet/ralink/ralink_soc_eth.h | 2 + 2 files changed, 33 insertions(+), 20 deletions(-)