lkml.org 
[lkml]   [2020]   [Oct]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [PATCH 1/4] ftgmac100: Fix race issue on TX descriptor[0]
From
Date
On Tue, 2020-10-20 at 04:13 +0000, Joel Stanley wrote:
> On Mon, 19 Oct 2020 at 23:20, Benjamin Herrenschmidt
> <benh@kernel.crashing.org> wrote:
> >
> > On Mon, 2020-10-19 at 16:57 +0800, Dylan Hung wrote:
> > > These rules must be followed when accessing the TX descriptor:
> > >
> > > 1. A TX descriptor is "cleanable" only when its value is non-zero
> > > and the owner bit is set to "software"
> >
> > Can you elaborate ? What is the point of that change ? The owner
> > bit
> > should be sufficient, why do we need to check other fields ?
>
> I would like Dylan to clarify too. The datasheet has a footnote below
> the descriptor layout:
>
> - TXDES#0: Bits 27 ~ 14 are valid only when FTS = 1
> - TXDES#1: Bits 31 ~ 0 are valid only when FTS = 1
>
> So the ownership bit (31) is not valid unless FTS is set. However,
> this isn't what his patch does. It adds checks for EDOTR.

No I think it adds a check for everything except EDOTR which just marks
the end of ring and needs to be ignored in the comparison.

That said, we do need a better explanation.

One potential bug I did find by looking at my code however is:

static bool ftgmac100_tx_complete_packet(struct ftgmac100 *priv)
{
struct net_device *netdev = priv->netdev;
struct ftgmac100_txdes *txdes;
struct sk_buff *skb;
unsigned int pointer;
u32 ctl_stat;

pointer = priv->tx_clean_pointer;
txdes = &priv->txdes[pointer];

ctl_stat = le32_to_cpu(txdes->txdes0);
if (ctl_stat & FTGMAC100_TXDES0_TXDMA_OWN)
return false;

skb = priv->tx_skbs[pointer];
netdev->stats.tx_packets++;
netdev->stats.tx_bytes += skb->len;
ftgmac100_free_tx_packet(priv, pointer, skb, txdes, ctl_stat);
txdes->txdes0 = cpu_to_le32(ctl_stat & priv->txdes0_edotr_mask);

^^^^ There should probably be an smp_wmb() here to ensure that all the above
stores are visible before the tx clean pointer is updated.

priv->tx_clean_pointer = ftgmac100_next_tx_pointer(priv, pointer);

return true;
}

Similarly we probablu should have one before setting tx_pointer in start_xmit().

As for the read side of this, I'm not 100% sure, I'll have to think more about
it, it *think* the existing barriers are sufficient at first sight.

Cheers,
Ben.

> >
> > > 2. A TX descriptor is "writable" only when its value is zero
> > > regardless the edotr mask.
> >
> > Again, why is that ? Can you elaborate ? What race are you trying
> > to
> > address here ?
> >
> > Cheers,
> > Ben.
> >
> > > Fixes: 52c0cae87465 ("ftgmac100: Remove tx descriptor accessors")
> > > Signed-off-by: Dylan Hung <dylan_hung@aspeedtech.com>
> > > Signed-off-by: Joel Stanley <joel@jms.id.au>
> > > ---
> > > drivers/net/ethernet/faraday/ftgmac100.c | 10 ++++++++++
> > > 1 file changed, 10 insertions(+)
> > >
> > > diff --git a/drivers/net/ethernet/faraday/ftgmac100.c
> > > b/drivers/net/ethernet/faraday/ftgmac100.c
> > > index 00024dd41147..7cacbe4aecb7 100644
> > > --- a/drivers/net/ethernet/faraday/ftgmac100.c
> > > +++ b/drivers/net/ethernet/faraday/ftgmac100.c
> > > @@ -647,6 +647,9 @@ static bool
> > > ftgmac100_tx_complete_packet(struct
> > > ftgmac100 *priv)
> > > if (ctl_stat & FTGMAC100_TXDES0_TXDMA_OWN)
> > > return false;
> > >
> > > + if ((ctl_stat & ~(priv->txdes0_edotr_mask)) == 0)
> > > + return false;
> > > +
> > > skb = priv->tx_skbs[pointer];
> > > netdev->stats.tx_packets++;
> > > netdev->stats.tx_bytes += skb->len;
> > > @@ -756,6 +759,9 @@ static netdev_tx_t
> > > ftgmac100_hard_start_xmit(struct sk_buff *skb,
> > > pointer = priv->tx_pointer;
> > > txdes = first = &priv->txdes[pointer];
> > >
> > > + if (le32_to_cpu(txdes->txdes0) & ~priv->txdes0_edotr_mask)
> > > + goto drop;
> > > +
> > > /* Setup it up with the packet head. Don't write the head
> > > to
> > > the
> > > * ring just yet
> > > */
> > > @@ -787,6 +793,10 @@ static netdev_tx_t
> > > ftgmac100_hard_start_xmit(struct sk_buff *skb,
> > > /* Setup descriptor */
> > > priv->tx_skbs[pointer] = skb;
> > > txdes = &priv->txdes[pointer];
> > > +
> > > + if (le32_to_cpu(txdes->txdes0) & ~priv-
> > > > txdes0_edotr_mask)
> > >
> > > + goto dma_err;
> > > +
> > > ctl_stat = ftgmac100_base_tx_ctlstat(priv,
> > > pointer);
> > > ctl_stat |= FTGMAC100_TXDES0_TXDMA_OWN;
> > > ctl_stat |= FTGMAC100_TXDES0_TXBUF_SIZE(len);

\
 
 \ /
  Last update: 2020-10-20 08:24    [W:0.085 / U:0.476 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site