[PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Simon Horman]

From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>

"swiotlb buffer is full" errors occur after repeated initialisation of a
device - f.e. suspend/resume or ip link set up/down. This is because memory
mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
is not released.  Resolve this problem by unmapping descriptors when
freeing rings.

Note, ravb_tx_free() is moved but not otherwise modified by this patch.

Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
[simon: reworked]
Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
--
v3 [Simon Horman]
* As suggested by Sergei Shtylyov
  - consistently use le32_to_cpu(desc->dptr)
  - Do not clear desc->ds_cc as it is not used
* Paramatise ravb_tx_free() to allow it to free non-transmitted buffers

v2 [Simon Horman]
* As suggested by Sergei Shtylyov
  - Use dma_mapping_error() and rx_desc->ds_cc when unmapping RX descriptors;
    this is consistent with the way that they are mapped
  - Use ravb_tx_free() to clear TX descriptors
* Reduce scope of new local variable

v1 [Kazuya Mizuguchi]
---
 drivers/net/ethernet/renesas/ravb_main.c | 113 ++++++++++++++++++-------------
 1 file changed, 65 insertions(+), 48 deletions(-)

diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c
index 89ac1e3f6175..57fe1411bb9d 100644
--- a/drivers/net/ethernet/renesas/ravb_main.c
+++ b/drivers/net/ethernet/renesas/ravb_main.c
@@ -179,6 +179,51 @@ static struct mdiobb_ops bb_ops = {
 	.get_mdio_data = ravb_get_mdio_data,
 };
 
+enum ravb_tx_free_mode {
+	ravb_tx_free_all,
+	ravb_tx_free_txed_only,
+};
+
+/* Free TX skb function for AVB-IP */
+static int ravb_tx_free(struct net_device *ndev, int q,
+			enum ravb_tx_free_mode free_mode)
+{
+	struct ravb_private *priv = netdev_priv(ndev);
+	struct net_device_stats *stats = &priv->stats[q];
+	struct ravb_tx_desc *desc;
+	int free_num = 0;
+	int entry;
+	u32 size;
+
+	for (; priv->cur_tx[q] - priv->dirty_tx[q] > 0; priv->dirty_tx[q]++) {
+		entry = priv->dirty_tx[q] % (priv->num_tx_ring[q] *
+					     NUM_TX_DESC);
+		desc = &priv->tx_ring[q][entry];
+		if (free_mode == ravb_tx_free_txed_only &&
+		    desc->die_dt != DT_FEMPTY)
+			break;
+		/* Descriptor type must be checked before all other reads */
+		dma_rmb();
+		size = le16_to_cpu(desc->ds_tagl) & TX_DS;
+		/* Free the original skb. */
+		if (priv->tx_skb[q][entry / NUM_TX_DESC]) {
+			dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr),
+					 size, DMA_TO_DEVICE);
+			/* Last packet descriptor? */
+			if (entry % NUM_TX_DESC == NUM_TX_DESC - 1) {
+				entry /= NUM_TX_DESC;
+				dev_kfree_skb_any(priv->tx_skb[q][entry]);
+				priv->tx_skb[q][entry] = NULL;
+				stats->tx_packets++;
+			}
+			free_num++;
+		}
+		stats->tx_bytes += size;
+		desc->die_dt = DT_EEMPTY;
+	}
+	return free_num;
+}
+
 /* Free skb's and DMA buffers for Ethernet AVB */
 static void ravb_ring_free(struct net_device *ndev, int q)
 {
@@ -194,19 +239,21 @@ static void ravb_ring_free(struct net_device *ndev, int q)
 	kfree(priv->rx_skb[q]);
 	priv->rx_skb[q] = NULL;
 
-	/* Free TX skb ringbuffer */
-	if (priv->tx_skb[q]) {
-		for (i = 0; i < priv->num_tx_ring[q]; i++)
-			dev_kfree_skb(priv->tx_skb[q][i]);
-	}
-	kfree(priv->tx_skb[q]);
-	priv->tx_skb[q] = NULL;
-
 	/* Free aligned TX buffers */
 	kfree(priv->tx_align[q]);
 	priv->tx_align[q] = NULL;
 
 	if (priv->rx_ring[q]) {
+		for (i = 0; i < priv->num_rx_ring[q]; i++) {
+			struct ravb_ex_rx_desc *desc = &priv->rx_ring[q][i];
+
+			if (!dma_mapping_error(ndev->dev.parent,
+					       le32_to_cpu(desc->dptr)))
+				dma_unmap_single(ndev->dev.parent,
+						 le32_to_cpu(desc->dptr),
+						 PKT_BUF_SZ,
+						 DMA_FROM_DEVICE);
+		}
 		ring_size = sizeof(struct ravb_ex_rx_desc) *
 			    (priv->num_rx_ring[q] + 1);
 		dma_free_coherent(ndev->dev.parent, ring_size, priv->rx_ring[q],
@@ -215,12 +262,19 @@ static void ravb_ring_free(struct net_device *ndev, int q)
 	}
 
 	if (priv->tx_ring[q]) {
+		ravb_tx_free(ndev, q, ravb_tx_free_all);
+
 		ring_size = sizeof(struct ravb_tx_desc) *
 			    (priv->num_tx_ring[q] * NUM_TX_DESC + 1);
 		dma_free_coherent(ndev->dev.parent, ring_size, priv->tx_ring[q],
 				  priv->tx_desc_dma[q]);
 		priv->tx_ring[q] = NULL;
 	}
+
+	/* Free TX skb ringbuffer.
+	 * SKBs are freed by ravb_tx_free() call above. */
+	kfree(priv->tx_skb[q]);
+	priv->tx_skb[q] = NULL;
 }
 
 /* Format skb and descriptor buffer for Ethernet AVB */
@@ -431,44 +485,6 @@ static int ravb_dmac_init(struct net_device *ndev)
 	return 0;
 }
 
-/* Free TX skb function for AVB-IP */
-static int ravb_tx_free(struct net_device *ndev, int q)
-{
-	struct ravb_private *priv = netdev_priv(ndev);
-	struct net_device_stats *stats = &priv->stats[q];
-	struct ravb_tx_desc *desc;
-	int free_num = 0;
-	int entry;
-	u32 size;
-
-	for (; priv->cur_tx[q] - priv->dirty_tx[q] > 0; priv->dirty_tx[q]++) {
-		entry = priv->dirty_tx[q] % (priv->num_tx_ring[q] *
-					     NUM_TX_DESC);
-		desc = &priv->tx_ring[q][entry];
-		if (desc->die_dt != DT_FEMPTY)
-			break;
-		/* Descriptor type must be checked before all other reads */
-		dma_rmb();
-		size = le16_to_cpu(desc->ds_tagl) & TX_DS;
-		/* Free the original skb. */
-		if (priv->tx_skb[q][entry / NUM_TX_DESC]) {
-			dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr),
-					 size, DMA_TO_DEVICE);
-			/* Last packet descriptor? */
-			if (entry % NUM_TX_DESC == NUM_TX_DESC - 1) {
-				entry /= NUM_TX_DESC;
-				dev_kfree_skb_any(priv->tx_skb[q][entry]);
-				priv->tx_skb[q][entry] = NULL;
-				stats->tx_packets++;
-			}
-			free_num++;
-		}
-		stats->tx_bytes += size;
-		desc->die_dt = DT_EEMPTY;
-	}
-	return free_num;
-}
-
 static void ravb_get_tx_tstamp(struct net_device *ndev)
 {
 	struct ravb_private *priv = netdev_priv(ndev);
@@ -902,7 +918,7 @@ static int ravb_poll(struct napi_struct *napi, int budget)
 			spin_lock_irqsave(&priv->lock, flags);
 			/* Clear TX interrupt */
 			ravb_write(ndev, ~mask, TIS);
-			ravb_tx_free(ndev, q);
+			ravb_tx_free(ndev, q, ravb_tx_free_txed_only);
 			netif_wake_subqueue(ndev, q);
 			mmiowb();
 			spin_unlock_irqrestore(&priv->lock, flags);
@@ -1567,7 +1583,8 @@ static netdev_tx_t ravb_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 
 	priv->cur_tx[q] += NUM_TX_DESC;
 	if (priv->cur_tx[q] - priv->dirty_tx[q] >
-	    (priv->num_tx_ring[q] - 1) * NUM_TX_DESC && !ravb_tx_free(ndev, q))
+	    (priv->num_tx_ring[q] - 1) * NUM_TX_DESC &&
+	    !ravb_tx_free(ndev, q, ravb_tx_free_txed_only))
 		netif_stop_subqueue(ndev, q);
 
 exit:
-- 
2.7.0.rc3.207.g0ac5344

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Sergei Shtylyov]

Hello.

On 01/24/2017 09:21 PM, Simon Horman wrote:

> From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
>
> "swiotlb buffer is full" errors occur after repeated initialisation of a
> device - f.e. suspend/resume or ip link set up/down. This is because memory
> mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
> is not released.  Resolve this problem by unmapping descriptors when
> freeing rings.
>
> Note, ravb_tx_free() is moved but not otherwise modified by this patch.
>
> Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> [simon: reworked]
> Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
> --
> v3 [Simon Horman]
> * As suggested by Sergei Shtylyov
>   - consistently use le32_to_cpu(desc->dptr)
>   - Do not clear desc->ds_cc as it is not used
> * Paramatise ravb_tx_free() to allow it to free non-transmitted buffers
>
> v2 [Simon Horman]
> * As suggested by Sergei Shtylyov
>   - Use dma_mapping_error() and rx_desc->ds_cc when unmapping RX descriptors;
>     this is consistent with the way that they are mapped
>   - Use ravb_tx_free() to clear TX descriptors
> * Reduce scope of new local variable
>
> v1 [Kazuya Mizuguchi]
> ---
>  drivers/net/ethernet/renesas/ravb_main.c | 113 ++++++++++++++++++-------------
>  1 file changed, 65 insertions(+), 48 deletions(-)
>
> diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c
> index 89ac1e3f6175..57fe1411bb9d 100644
> --- a/drivers/net/ethernet/renesas/ravb_main.c
> +++ b/drivers/net/ethernet/renesas/ravb_main.c
> @@ -179,6 +179,51 @@ static struct mdiobb_ops bb_ops = {
>  	.get_mdio_data = ravb_get_mdio_data,
>  };
>
> +enum ravb_tx_free_mode {
> +	ravb_tx_free_all,
> +	ravb_tx_free_txed_only,
> +};
> +
> +/* Free TX skb function for AVB-IP */
> +static int ravb_tx_free(struct net_device *ndev, int q,
> +			enum ravb_tx_free_mode free_mode)

    Hmm... Sorry but this looks over-engineered. A *bool* parameter (named e.g 
'all) would suffice IMHO.

> +{
> +	struct ravb_private *priv = netdev_priv(ndev);
> +	struct net_device_stats *stats = &priv->stats[q];
> +	struct ravb_tx_desc *desc;
> +	int free_num = 0;
> +	int entry;
> +	u32 size;
> +
> +	for (; priv->cur_tx[q] - priv->dirty_tx[q] > 0; priv->dirty_tx[q]++) {
> +		entry = priv->dirty_tx[q] % (priv->num_tx_ring[q] *
> +					     NUM_TX_DESC);
> +		desc = &priv->tx_ring[q][entry];
> +		if (free_mode == ravb_tx_free_txed_only &&
> +		    desc->die_dt != DT_FEMPTY)
> +			break;
> +		/* Descriptor type must be checked before all other reads */
> +		dma_rmb();
> +		size = le16_to_cpu(desc->ds_tagl) & TX_DS;
> +		/* Free the original skb. */
> +		if (priv->tx_skb[q][entry / NUM_TX_DESC]) {
> +			dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr),
> +					 size, DMA_TO_DEVICE);
> +			/* Last packet descriptor? */
> +			if (entry % NUM_TX_DESC == NUM_TX_DESC - 1) {
> +				entry /= NUM_TX_DESC;
> +				dev_kfree_skb_any(priv->tx_skb[q][entry]);
> +				priv->tx_skb[q][entry] = NULL;
> +				stats->tx_packets++;
> +			}
> +			free_num++;
> +		}
> +		stats->tx_bytes += size;

    Hmmm... we shouldn't count the discarded unsent packets/bytes as sent, right?

[...]
> @@ -215,12 +262,19 @@ static void ravb_ring_free(struct net_device *ndev, int q)
>  	}
>
>  	if (priv->tx_ring[q]) {
> +		ravb_tx_free(ndev, q, ravb_tx_free_all);
> +
>  		ring_size = sizeof(struct ravb_tx_desc) *
>  			    (priv->num_tx_ring[q] * NUM_TX_DESC + 1);
>  		dma_free_coherent(ndev->dev.parent, ring_size, priv->tx_ring[q],
>  				  priv->tx_desc_dma[q]);
>  		priv->tx_ring[q] = NULL;
>  	}
> +
> +	/* Free TX skb ringbuffer.
> +	 * SKBs are freed by ravb_tx_free() call above. */

    This is not a recommended comment format:

/* bla
  * bla
  */

[...]

MBR, Sergei

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Simon Horman]

On Wed, Jan 25, 2017 at 07:05:08PM +0300, Sergei Shtylyov wrote:
> Hello.
> 
> On 01/24/2017 09:21 PM, Simon Horman wrote:
> 
> >From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> >
> >"swiotlb buffer is full" errors occur after repeated initialisation of a
> >device - f.e. suspend/resume or ip link set up/down. This is because memory
> >mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
> >is not released.  Resolve this problem by unmapping descriptors when
> >freeing rings.
> >
> >Note, ravb_tx_free() is moved but not otherwise modified by this patch.
> >
> >Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> >[simon: reworked]
> >Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
> >--
> >v3 [Simon Horman]
> >* As suggested by Sergei Shtylyov
> >  - consistently use le32_to_cpu(desc->dptr)
> >  - Do not clear desc->ds_cc as it is not used
> >* Paramatise ravb_tx_free() to allow it to free non-transmitted buffers
> >
> >v2 [Simon Horman]
> >* As suggested by Sergei Shtylyov
> >  - Use dma_mapping_error() and rx_desc->ds_cc when unmapping RX descriptors;
> >    this is consistent with the way that they are mapped
> >  - Use ravb_tx_free() to clear TX descriptors
> >* Reduce scope of new local variable
> >
> >v1 [Kazuya Mizuguchi]
> >---
> > drivers/net/ethernet/renesas/ravb_main.c | 113 ++++++++++++++++++-------------
> > 1 file changed, 65 insertions(+), 48 deletions(-)
> >
> >diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c
> >index 89ac1e3f6175..57fe1411bb9d 100644
> >--- a/drivers/net/ethernet/renesas/ravb_main.c
> >+++ b/drivers/net/ethernet/renesas/ravb_main.c
> >@@ -179,6 +179,51 @@ static struct mdiobb_ops bb_ops = {
> > 	.get_mdio_data = ravb_get_mdio_data,
> > };
> >
> >+enum ravb_tx_free_mode {
> >+	ravb_tx_free_all,
> >+	ravb_tx_free_txed_only,
> >+};
> >+
> >+/* Free TX skb function for AVB-IP */
> >+static int ravb_tx_free(struct net_device *ndev, int q,
> >+			enum ravb_tx_free_mode free_mode)
> 
>    Hmm... Sorry but this looks over-engineered. A *bool* parameter (named
> e.g 'all) would suffice IMHO.

Ha! The last time I used a bool for something like this I was encouraged
to use an enum, admittedly that was not kernel code but I was unsure
which way to go this time. I'll change things to bool as you sugget.

> >+{
> >+	struct ravb_private *priv = netdev_priv(ndev);
> >+	struct net_device_stats *stats = &priv->stats[q];
> >+	struct ravb_tx_desc *desc;
> >+	int free_num = 0;
> >+	int entry;
> >+	u32 size;
> >+
> >+	for (; priv->cur_tx[q] - priv->dirty_tx[q] > 0; priv->dirty_tx[q]++) {
> >+		entry = priv->dirty_tx[q] % (priv->num_tx_ring[q] *
> >+					     NUM_TX_DESC);
> >+		desc = &priv->tx_ring[q][entry];
> >+		if (free_mode == ravb_tx_free_txed_only &&
> >+		    desc->die_dt != DT_FEMPTY)
> >+			break;
> >+		/* Descriptor type must be checked before all other reads */
> >+		dma_rmb();
> >+		size = le16_to_cpu(desc->ds_tagl) & TX_DS;
> >+		/* Free the original skb. */
> >+		if (priv->tx_skb[q][entry / NUM_TX_DESC]) {
> >+			dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr),
> >+					 size, DMA_TO_DEVICE);
> >+			/* Last packet descriptor? */
> >+			if (entry % NUM_TX_DESC == NUM_TX_DESC - 1) {
> >+				entry /= NUM_TX_DESC;
> >+				dev_kfree_skb_any(priv->tx_skb[q][entry]);
> >+				priv->tx_skb[q][entry] = NULL;
> >+				stats->tx_packets++;
> >+			}
> >+			free_num++;
> >+		}
> >+		stats->tx_bytes += size;
> 
>    Hmmm... we shouldn't count the discarded unsent packets/bytes as sent, right?

Yes, I think so. Sorry for missing that.

> [...]
> >@@ -215,12 +262,19 @@ static void ravb_ring_free(struct net_device *ndev, int q)
> > 	}
> >
> > 	if (priv->tx_ring[q]) {
> >+		ravb_tx_free(ndev, q, ravb_tx_free_all);
> >+
> > 		ring_size = sizeof(struct ravb_tx_desc) *
> > 			    (priv->num_tx_ring[q] * NUM_TX_DESC + 1);
> > 		dma_free_coherent(ndev->dev.parent, ring_size, priv->tx_ring[q],
> > 				  priv->tx_desc_dma[q]);
> > 		priv->tx_ring[q] = NULL;
> > 	}
> >+
> >+	/* Free TX skb ringbuffer.
> >+	 * SKBs are freed by ravb_tx_free() call above. */
> 
>    This is not a recommended comment format:
> 
> /* bla
>  * bla
>  */

Thanks, I will fix that.

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Sergei Shtylyov]

On 01/24/2017 09:21 PM, Simon Horman wrote:

> From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
>
> "swiotlb buffer is full" errors occur after repeated initialisation of a
> device - f.e. suspend/resume or ip link set up/down. This is because memory
> mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
> is not released.  Resolve this problem by unmapping descriptors when
> freeing rings.

    Could you look into the sh_eth driver which seems to have the same issue?

> Note, ravb_tx_free() is moved but not otherwise modified by this patch.

    This is not true anymore BTW.

> Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> [simon: reworked]
> Signed-off-by: Simon Horman <horms+renesas@verge.net.au>

MBR, Sergei

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Simon Horman]

On Wed, Jan 25, 2017 at 07:18:15PM +0300, Sergei Shtylyov wrote:
> On 01/24/2017 09:21 PM, Simon Horman wrote:
> 
> >From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> >
> >"swiotlb buffer is full" errors occur after repeated initialisation of a
> >device - f.e. suspend/resume or ip link set up/down. This is because memory
> >mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
> >is not released.  Resolve this problem by unmapping descriptors when
> >freeing rings.
> 
>    Could you look into the sh_eth driver which seems to have the same issue?

Sure, I will check.

> >Note, ravb_tx_free() is moved but not otherwise modified by this patch.
> 
>    This is not true anymore BTW.

Thanks for noticing, I'll fix that.

> >Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> >[simon: reworked]
> >Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
> 
> MBR, Sergei
> 

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Geert Uytterhoeven]

On Wed, Jan 25, 2017 at 5:18 PM, Sergei Shtylyov
<sergei.shtylyov@cogentembedded.com> wrote:
> On 01/24/2017 09:21 PM, Simon Horman wrote:
>
>> From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
>>
>> "swiotlb buffer is full" errors occur after repeated initialisation of a
>> device - f.e. suspend/resume or ip link set up/down. This is because
>> memory
>> mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
>> is not released.  Resolve this problem by unmapping descriptors when
>> freeing rings.
>
>
>    Could you look into the sh_eth driver which seems to have the same issue?

Indeed, after a few suspend/resume cycles on r8a7791/koelsch:

WARNING: CPU: 1 PID: 1699 at lib/dma-debug.c:517 add_dma_entry+0xfc/0x148
DMA-API: exceeded 7 overlapping mappings of cacheline 0x0000000001a827e3

Gr{oetje,eeting}s,

                        Geert

Re: [PATCH/RFC v3 net] ravb: unmap descriptors when freeing rings

[Simon Horman]

On Fri, Feb 10, 2017 at 02:52:50PM +0100, Geert Uytterhoeven wrote:
> On Wed, Jan 25, 2017 at 5:18 PM, Sergei Shtylyov
> <sergei.shtylyov@cogentembedded.com> wrote:
> > On 01/24/2017 09:21 PM, Simon Horman wrote:
> >
> >> From: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
> >>
> >> "swiotlb buffer is full" errors occur after repeated initialisation of a
> >> device - f.e. suspend/resume or ip link set up/down. This is because
> >> memory
> >> mapped using dma_map_single() in ravb_ring_format() and ravb_start_xmit()
> >> is not released.  Resolve this problem by unmapping descriptors when
> >> freeing rings.
> >
> >
> >    Could you look into the sh_eth driver which seems to have the same issue?
> 
> Indeed, after a few suspend/resume cycles on r8a7791/koelsch:
> 
> WARNING: CPU: 1 PID: 1699 at lib/dma-debug.c:517 add_dma_entry+0xfc/0x148
> DMA-API: exceeded 7 overlapping mappings of cacheline 0x0000000001a827e3

Thanks for confirming that. It matches my expectation after
reading of the sh_eth code.