Nouveau - Sep 2020 - [PATCH 22/28] sgiseeq: convert from dma_cache_sync to dma_sync_single_for_device

On Tue, Sep 01, 2020 at 07:12:41PM +0200, Thomas Bogendoerfer
wrote:
> On Tue, Sep 01, 2020 at 05:22:09PM +0200, Thomas Bogendoerfer wrote:
> > On Wed, Aug 19, 2020 at 08:55:49AM +0200, Christoph Hellwig wrote:
> > > Use the proper modern API to transfer cache ownership for
incoherent DMA.
> > > 
> > > Signed-off-by: Christoph Hellwig <hch at lst.de>
> > > ---
> > >  drivers/net/ethernet/seeq/sgiseeq.c | 12 ++++++++----
> > >  1 file changed, 8 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/drivers/net/ethernet/seeq/sgiseeq.c
b/drivers/net/ethernet/seeq/sgiseeq.c
> > > index 39599bbb5d45b6..f91dae16d69a19 100644
> > > --- a/drivers/net/ethernet/seeq/sgiseeq.c
> > > +++ b/drivers/net/ethernet/seeq/sgiseeq.c
> > > @@ -112,14 +112,18 @@ struct sgiseeq_private {
> > >  
> > >  static inline void dma_sync_desc_cpu(struct net_device *dev,
void *addr)
> > >  {
> > > -	dma_cache_sync(dev->dev.parent, addr, sizeof(struct
sgiseeq_rx_desc),
> > > -		       DMA_FROM_DEVICE);
> > > +	struct sgiseeq_private *sp = netdev_priv(dev);
> > > +
> > > +	dma_sync_single_for_cpu(dev->dev.parent, VIRT_TO_DMA(sp,
addr),
> > > +			sizeof(struct sgiseeq_rx_desc), DMA_BIDIRECTIONAL);
> > >  }
> > >  
> > >  static inline void dma_sync_desc_dev(struct net_device *dev,
void *addr)
> > >  {
> > > -	dma_cache_sync(dev->dev.parent, addr, sizeof(struct
sgiseeq_rx_desc),
> > > -		       DMA_TO_DEVICE);
> > > +	struct sgiseeq_private *sp = netdev_priv(dev);
> > > +
> > > +	dma_sync_single_for_device(dev->dev.parent, VIRT_TO_DMA(sp,
addr),
> > > +			sizeof(struct sgiseeq_rx_desc), DMA_BIDIRECTIONAL);
> > >  }
> > 
> > this breaks ethernet on IP22 completely, but I haven't figured out
why, yet.
> 
> the problem is that dma_sync_single_for_cpu() doesn't flush anything
> for IP22, because it only flushes for CPUs which do speculation. So
> either MIPS arch_sync_dma_for_cpu() should always flush or sgiseeq
> needs to use a different sync funktion, when it wants to re-read
descriptors
> from memory.
Well, if IP22 doesn't speculate (which I'm pretty sure is the case),
dma_sync_single_for_cpu should indeeed be a no-op.  But then there
also shouldn't be anything in the cache, as the previous
dma_sync_single_for_device should have invalidated it.  So it seems like
we are missing one (or more) ownership transfers to the device.  I'll
try to look at the the ownership management in a little more detail
tomorrow.

On Tue, Sep 01, 2020 at 07:16:27PM +0200, Christoph Hellwig
wrote:
> Well, if IP22 doesn't speculate (which I'm pretty sure is the
case),
> dma_sync_single_for_cpu should indeeed be a no-op.  But then there
> also shouldn't be anything in the cache, as the previous
> dma_sync_single_for_device should have invalidated it.  So it seems like
> we are missing one (or more) ownership transfers to the device.  I'll
> try to look at the the ownership management in a little more detail
> tomorrow.
this is the problem:

       /* Always check for received packets. */
        sgiseeq_rx(dev, sp, hregs, sregs);

so the driver will look at the rx descriptor on every interrupt, so
we cache the rx descriptor on the first interrupt and if there was
$no rx packet, we will only see it, if cache line gets flushed for
some other reason. kick_tx() does a busy loop checking tx descriptors,
with just sync_desc_cpu...

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]

On Tue, Sep 01, 2020 at 07:38:10PM +0200, Thomas Bogendoerfer
wrote:
> On Tue, Sep 01, 2020 at 07:16:27PM +0200, Christoph Hellwig wrote:
> > Well, if IP22 doesn't speculate (which I'm pretty sure is the
case),
> > dma_sync_single_for_cpu should indeeed be a no-op.  But then there
> > also shouldn't be anything in the cache, as the previous
> > dma_sync_single_for_device should have invalidated it.  So it seems
like
> > we are missing one (or more) ownership transfers to the device. 
I'll
> > try to look at the the ownership management in a little more detail
> > tomorrow.
> 
> this is the problem:
> 
>        /* Always check for received packets. */
>         sgiseeq_rx(dev, sp, hregs, sregs);
> 
> so the driver will look at the rx descriptor on every interrupt, so
> we cache the rx descriptor on the first interrupt and if there was
> $no rx packet, we will only see it, if cache line gets flushed for
> some other reason. kick_tx() does a busy loop checking tx descriptors,
> with just sync_desc_cpu...
the patch below fixes the problem.

Thomas.


diff --git a/drivers/net/ethernet/seeq/sgiseeq.c
b/drivers/net/ethernet/seeq/sgiseeq.c
index 8507ff242014..876e3700a0e4 100644
--- a/drivers/net/ethernet/seeq/sgiseeq.c
+++ b/drivers/net/ethernet/seeq/sgiseeq.c
@@ -112,14 +112,18 @@ struct sgiseeq_private {
 
 static inline void dma_sync_desc_cpu(struct net_device *dev, void *addr)
 {
-       dma_cache_sync(dev->dev.parent, addr, sizeof(struct sgiseeq_rx_desc),
-                      DMA_FROM_DEVICE);
+       struct sgiseeq_private *sp = netdev_priv(dev);
+
+       dma_sync_single_for_device(dev->dev.parent, VIRT_TO_DMA(sp, addr),
+                       sizeof(struct sgiseeq_rx_desc), DMA_FROM_DEVICE);
 }
 
 static inline void dma_sync_desc_dev(struct net_device *dev, void *addr)
 {
-       dma_cache_sync(dev->dev.parent, addr, sizeof(struct sgiseeq_rx_desc),
-                      DMA_TO_DEVICE);
+       struct sgiseeq_private *sp = netdev_priv(dev);
+
+       dma_sync_single_for_device(dev->dev.parent, VIRT_TO_DMA(sp, addr),
+                       sizeof(struct sgiseeq_rx_desc), DMA_TO_DEVICE);
 }
 
-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]

On Tue, Sep 01, 2020 at 07:38:10PM +0200, Thomas Bogendoerfer
wrote:
> this is the problem:
> 
>        /* Always check for received packets. */
>         sgiseeq_rx(dev, sp, hregs, sregs);
> 
> so the driver will look at the rx descriptor on every interrupt, so
> we cache the rx descriptor on the first interrupt and if there was
> $no rx packet, we will only see it, if cache line gets flushed for
> some other reason.
That means a transfer back to device ownership is missing after a
(negative) check.
> kick_tx() does a busy loop checking tx descriptors,
> with just sync_desc_cpu...
> 
> Thomas.
> 
> -- 
> Crap can work. Given enough thrust pigs will fly, but it's not
necessarily a
> good idea.                                                [ RFC1925, 2.3 ]
---end quoted text---