| Message ID | 1348218675-3804-1-git-send-email-Joakim.Tjernlund@transmode.se |
|---|---|
| State | Changes Requested, archived |
| Delegated to: | David Miller |
| Headers | show |
Joakim Tjernlund <Joakim.Tjernlund@transmode.se> : > Currently ucc_geth_start_xmit wraps IRQ off for the > whole body just to be safe. By rearranging the code a bit > one can avoid the lock completely. Afaics you went a bit too lockless with the queueing disable / enable logic. The hard_start_xmit handler is run in a locally softirq disabled section but it will happily race with the napi handler on a different CPU. Grep netif_tx_lock in tg3.c for it. The Tx skb free logic probably requires some smp memory barriers as well since the current skb is used by the ucc_geth driver to sync the Tx xmit with the napi completion handler.
Francois Romieu <romieu@fr.zoreil.com> wrote on 2012/09/21 19:35:35: > > Please keep netdev in the Cc:. Sorry, wrong reply button. Now added. > > Joakim Tjernlund <joakim.tjernlund@transmode.se> : > [...] > > This is what I could come up with, what do you think? > > In its current state the driver implicitely relies on the Tx xmit vs Tx > completion exclusion to work. As long as they exclude each other the > ugeth->tx_skbuff[txQ][...] and the "bd_status & T_R" states are consistent. > > This scheme can not work without locking. The skb alone won't provide a > synchronization point. I don't get it. The skb test is there just for one special case, when the BD ring is empty the (bd_status & T_R) == 0 will be true as well so one need something more than the bd_status test. > > An usual solution would be some skb dirtytx vs curtx comparison + smp > barriers (Documentation/memory-barriers.txt). You mean adding an extra test in addition to !skb? Something like: if (!skb && ugeth->skb_dirtytx[txQ] == ugeth->skb_curtx[txQ]) break; > > Sidenote: is there some reason why modulo (%) operations should be > avoided on this platform ? Not that I know. The original driver author did it this way. > > [...] > > @@ -3380,8 +3380,12 @@ static int ucc_geth_tx(struct net_device *dev, u8 txQ) > > 1) & TX_RING_MOD_MASK(ugeth->ug_info->bdRingLenTx[txQ]); > > > > /* We freed a buffer, so now we can restart transmission */ > > - if (netif_queue_stopped(dev)) > > - netif_wake_queue(dev); > > + if (netif_queue_stopped(dev)) { > > + netif_tx_lock(dev); > > + if (netif_queue_stopped(dev)) > > + netif_wake_queue(dev); > > + netif_tx_unlock(dev); > > + } > > Without exclusion we don't know if it was stopped before or after the packet > was freed. There must be some "are there available Tx slots ?" test in the > locked section. This makes sense, but stopping relies on ugeth->confBd[txQ] which is updated last in ucc_geth_tx(). Looks a bit suboptimal though so it should probably be changed. > > Btw the second netif_queue_stopped test can be removed if tx queueing can > only be stopped in a single place (I did not check). There is a netif_stop_queue in xmit and one in ucc_geth_close() so I guess one can say there is only one place. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Joakim Tjernlund <joakim.tjernlund@transmode.se> : [...] > I don't get it. The skb test is there just for one special case, when > the BD ring is empty the (bd_status & T_R) == 0 will be true as well so > one need something more than the bd_status test. Sure but the converse is not true : (bd_status & T_R) == 0 && skb does not mean that the skb has been sent. It happens when said skb is about to be given to the hardware by hard_start_xmit as well.
Francois Romieu <romieu@fr.zoreil.com> wrote on 2012/09/24 23:10:14: > > Joakim Tjernlund <joakim.tjernlund@transmode.se> : > [...] > > I don't get it. The skb test is there just for one special case, when > > the BD ring is empty the (bd_status & T_R) == 0 will be true as well so > > one need something more than the bd_status test. > > Sure but the converse is not true : (bd_status & T_R) == 0 && skb does not > mean that the skb has been sent. It happens when said skb is about to be > given to the hardware by hard_start_xmit as well. duhh, I was too tired when trying to make sense of smp & racing in general, thanks. Will probably be some time before I get to this again due to other stuff though. The other patches are independent of this one, I hope they are good/accepted? Jocke -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/net/ethernet/freescale/ucc_geth.c b/drivers/net/ethernet/freescale/ucc_geth.c index 9ac14f8..040aa70 100644 --- a/drivers/net/ethernet/freescale/ucc_geth.c +++ b/drivers/net/ethernet/freescale/ucc_geth.c @@ -3177,19 +3177,20 @@ static int ucc_geth_start_xmit(struct sk_buff *skb, struct net_device *dev) u8 __iomem *bd; /* BD pointer */ u32 bd_status; u8 txQ = 0; - unsigned long flags; ugeth_vdbg("%s: IN", __func__); - spin_lock_irqsave(&ugeth->lock, flags); - dev->stats.tx_bytes += skb->len; + /* 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! + */ + /* Start from the next BD that should be filled */ bd = ugeth->txBd[txQ]; bd_status = in_be32((u32 __iomem *)bd); - /* Save the skb pointer so we can free it later */ - ugeth->tx_skbuff[txQ][ugeth->skb_curtx[txQ]] = skb; /* Update the current skb pointer (wrapping if this was the last) */ ugeth->skb_curtx[txQ] = @@ -3207,6 +3208,8 @@ static int ucc_geth_start_xmit(struct sk_buff *skb, struct net_device *dev) /* set bd status and length */ out_be32((u32 __iomem *)bd, bd_status); + /* Save the skb pointer so we can free it later */ + ugeth->tx_skbuff[txQ][ugeth->skb_curtx[txQ]] = skb; /* Move to next BD in the ring */ if (!(bd_status & T_W)) @@ -3238,8 +3241,6 @@ static int ucc_geth_start_xmit(struct sk_buff *skb, struct net_device *dev) uccf = ugeth->uccf; out_be16(uccf->p_utodr, UCC_FAST_TOD); #endif - spin_unlock_irqrestore(&ugeth->lock, flags); - return NETDEV_TX_OK; } @@ -3387,10 +3388,8 @@ static int ucc_geth_poll(struct napi_struct *napi, int budget) ug_info = ugeth->ug_info; /* Tx event processing */ - spin_lock(&ugeth->lock); for (i = 0; i < ug_info->numQueuesTx; i++) ucc_geth_tx(ugeth->ndev, i); - spin_unlock(&ugeth->lock); howmany = 0; for (i = 0; i < ug_info->numQueuesRx; i++)
Currently ucc_geth_start_xmit wraps IRQ off for the whole body just to be safe. By rearranging the code a bit one can avoid the lock completely. Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se> --- v2: Move assignment of ugeth->tx_skbuff[txQ][ugeth->skb_curtx[txQ]] inside IRQ off section to prevent racing against ucc_geth_tx(). Spotted by Francois Romieu <romieu@fr.zoreil.com> v3: Lockless xmit Here is my attemept to do lockless xmit. Thanks to Francois Romieu <romieu@fr.zoreil.com> for the idea. drivers/net/ethernet/freescale/ucc_geth.c | 17 ++++++++--------- 1 files changed, 8 insertions(+), 9 deletions(-)