* [PATCH] sh_eth: unmap DMA buffers when freeing rings
@ 2017-04-16 20:01 Sergei Shtylyov
2017-04-17 8:57 ` Sergei Shtylyov
0 siblings, 1 reply; 3+ messages in thread
From: Sergei Shtylyov @ 2017-04-16 20:01 UTC (permalink / raw)
To: netdev; +Cc: linux-renesas-soc, Sergei Shtylyov
[-- Attachment #1: sh_eth-unmap-DMA-buffers-when-freeing-rings.patch --]
[-- Type: text/plain, Size: 6255 bytes --]
The DMA API debugging (when enabled) causes:
WARNING: CPU: 0 PID: 1445 at lib/dma-debug.c:519 add_dma_entry+0xe0/0x12c
DMA-API: exceeded 7 overlapping mappings of cacheline 0x01b2974d
to be printed after repeated initialization of the Ether device, e.g.
suspend/resume or 'ifconfig' up/down. This is because DMA buffers mapped
using dma_map_single() in sh_eth_ring_format() and sh_eth_start_xmit() are
never unmapped. Resolve this problem by unmapping the buffers when freeing
the descriptor rings; in order to do it right, we'd have to add an extra
parameter to sh_eth_txfree() (we rename this function to sh_eth_tx_free(),
while at it).
Based on the commit a47b70ea86bd ("ravb: unmap descriptors when freeing
rings").
Signed-off-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
---
drivers/net/ethernet/renesas/sh_eth.c | 122 ++++++++++++++++++----------------
1 file changed, 67 insertions(+), 55 deletions(-)
Index: net/drivers/net/ethernet/renesas/sh_eth.c
===================================================================
--- net.orig/drivers/net/ethernet/renesas/sh_eth.c
+++ net/drivers/net/ethernet/renesas/sh_eth.c
@@ -1127,12 +1127,70 @@ static struct mdiobb_ops bb_ops = {
.get_mdio_data = sh_get_mdio,
};
+/* free Tx skb function */
+static int sh_eth_tx_free(struct net_device *ndev, bool sent_only)
+{
+ struct sh_eth_private *mdp = netdev_priv(ndev);
+ struct sh_eth_txdesc *txdesc;
+ int free_num = 0;
+ int entry;
+ bool sent;
+
+ for (; mdp->cur_tx - mdp->dirty_tx > 0; mdp->dirty_tx++) {
+ entry = mdp->dirty_tx % mdp->num_tx_ring;
+ txdesc = &mdp->tx_ring[entry];
+ sent = txdesc->status & cpu_to_le32(TD_TACT);
+ if (sent_only && !sent)
+ break;
+ /* TACT bit must be checked before all the following reads */
+ dma_rmb();
+ netif_info(mdp, tx_done, ndev,
+ "tx entry %d status 0x%08x\n",
+ entry, le32_to_cpu(txdesc->status));
+ /* Free the original skb. */
+ if (mdp->tx_skbuff[entry]) {
+ dma_unmap_single(&ndev->dev, le32_to_cpu(txdesc->addr),
+ le32_to_cpu(txdesc->len) >> 16,
+ DMA_TO_DEVICE);
+ dev_kfree_skb_irq(mdp->tx_skbuff[entry]);
+ mdp->tx_skbuff[entry] = NULL;
+ free_num++;
+ }
+ txdesc->status = cpu_to_le32(TD_TFP);
+ if (entry >= mdp->num_tx_ring - 1)
+ txdesc->status |= cpu_to_le32(TD_TDLE);
+
+ if (sent) {
+ ndev->stats.tx_packets++;
+ ndev->stats.tx_bytes += le32_to_cpu(txdesc->len) >> 16;
+ }
+ }
+ return free_num;
+}
+
/* free skb and descriptor buffer */
static void sh_eth_ring_free(struct net_device *ndev)
{
struct sh_eth_private *mdp = netdev_priv(ndev);
int ringsize, i;
+ if (mdp->rx_ring) {
+ for (i = 0; i < mdp->num_rx_ring; i++) {
+ if (mdp->rx_skbuff[i]) {
+ struct sh_eth_rxdesc *rxdesc = &mdp->rx_ring[i];
+
+ dma_unmap_single(&ndev->dev,
+ le32_to_cpu(rxdesc->addr),
+ ALIGN(mdp->rx_buf_sz, 32),
+ DMA_FROM_DEVICE);
+ }
+ }
+ ringsize = sizeof(struct sh_eth_rxdesc) * mdp->num_rx_ring;
+ dma_free_coherent(NULL, ringsize, mdp->rx_ring,
+ mdp->rx_desc_dma);
+ mdp->rx_ring = NULL;
+ }
+
/* Free Rx skb ringbuffer */
if (mdp->rx_skbuff) {
for (i = 0; i < mdp->num_rx_ring; i++)
@@ -1141,27 +1199,18 @@ static void sh_eth_ring_free(struct net_
kfree(mdp->rx_skbuff);
mdp->rx_skbuff = NULL;
- /* Free Tx skb ringbuffer */
- if (mdp->tx_skbuff) {
- for (i = 0; i < mdp->num_tx_ring; i++)
- dev_kfree_skb(mdp->tx_skbuff[i]);
- }
- kfree(mdp->tx_skbuff);
- mdp->tx_skbuff = NULL;
-
- if (mdp->rx_ring) {
- ringsize = sizeof(struct sh_eth_rxdesc) * mdp->num_rx_ring;
- dma_free_coherent(NULL, ringsize, mdp->rx_ring,
- mdp->rx_desc_dma);
- mdp->rx_ring = NULL;
- }
-
if (mdp->tx_ring) {
+ sh_eth_tx_free(ndev, false);
+
ringsize = sizeof(struct sh_eth_txdesc) * mdp->num_tx_ring;
dma_free_coherent(NULL, ringsize, mdp->tx_ring,
mdp->tx_desc_dma);
mdp->tx_ring = NULL;
}
+
+ /* Free Tx skb ringbuffer */
+ kfree(mdp->tx_skbuff);
+ mdp->tx_skbuff = NULL;
}
/* format skb and descriptor buffer */
@@ -1409,43 +1458,6 @@ static void sh_eth_dev_exit(struct net_d
update_mac_address(ndev);
}
-/* free Tx skb function */
-static int sh_eth_txfree(struct net_device *ndev)
-{
- struct sh_eth_private *mdp = netdev_priv(ndev);
- struct sh_eth_txdesc *txdesc;
- int free_num = 0;
- int entry;
-
- for (; mdp->cur_tx - mdp->dirty_tx > 0; mdp->dirty_tx++) {
- entry = mdp->dirty_tx % mdp->num_tx_ring;
- txdesc = &mdp->tx_ring[entry];
- if (txdesc->status & cpu_to_le32(TD_TACT))
- break;
- /* TACT bit must be checked before all the following reads */
- dma_rmb();
- netif_info(mdp, tx_done, ndev,
- "tx entry %d status 0x%08x\n",
- entry, le32_to_cpu(txdesc->status));
- /* Free the original skb. */
- if (mdp->tx_skbuff[entry]) {
- dma_unmap_single(&ndev->dev, le32_to_cpu(txdesc->addr),
- le32_to_cpu(txdesc->len) >> 16,
- DMA_TO_DEVICE);
- dev_kfree_skb_irq(mdp->tx_skbuff[entry]);
- mdp->tx_skbuff[entry] = NULL;
- free_num++;
- }
- txdesc->status = cpu_to_le32(TD_TFP);
- if (entry >= mdp->num_tx_ring - 1)
- txdesc->status |= cpu_to_le32(TD_TDLE);
-
- ndev->stats.tx_packets++;
- ndev->stats.tx_bytes += le32_to_cpu(txdesc->len) >> 16;
- }
- return free_num;
-}
-
/* Packet receive function */
static int sh_eth_rx(struct net_device *ndev, u32 intr_status, int *quota)
{
@@ -1690,7 +1702,7 @@ static void sh_eth_error(struct net_devi
intr_status, mdp->cur_tx, mdp->dirty_tx,
(u32)ndev->state, edtrr);
/* dirty buffer free */
- sh_eth_txfree(ndev);
+ sh_eth_tx_free(ndev, true);
/* SH7712 BUG */
if (edtrr ^ sh_eth_get_edtrr_trns(mdp)) {
@@ -1751,7 +1763,7 @@ static irqreturn_t sh_eth_interrupt(int
/* Clear Tx interrupts */
sh_eth_write(ndev, intr_status & cd->tx_check, EESR);
- sh_eth_txfree(ndev);
+ sh_eth_tx_free(ndev, true);
netif_wake_queue(ndev);
}
@@ -2412,7 +2424,7 @@ static int sh_eth_start_xmit(struct sk_b
spin_lock_irqsave(&mdp->lock, flags);
if ((mdp->cur_tx - mdp->dirty_tx) >= (mdp->num_tx_ring - 4)) {
- if (!sh_eth_txfree(ndev)) {
+ if (!sh_eth_tx_free(ndev, true)) {
netif_warn(mdp, tx_queued, ndev, "TxFD exhausted.\n");
netif_stop_queue(ndev);
spin_unlock_irqrestore(&mdp->lock, flags);
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] sh_eth: unmap DMA buffers when freeing rings
2017-04-16 20:01 [PATCH] sh_eth: unmap DMA buffers when freeing rings Sergei Shtylyov
@ 2017-04-17 8:57 ` Sergei Shtylyov
2017-04-17 12:45 ` Sergei Shtylyov
0 siblings, 1 reply; 3+ messages in thread
From: Sergei Shtylyov @ 2017-04-17 8:57 UTC (permalink / raw)
To: netdev; +Cc: linux-renesas-soc
Hello!
On 4/16/2017 11:01 PM, Sergei Shtylyov wrote:
> The DMA API debugging (when enabled) causes:
>
> WARNING: CPU: 0 PID: 1445 at lib/dma-debug.c:519 add_dma_entry+0xe0/0x12c
> DMA-API: exceeded 7 overlapping mappings of cacheline 0x01b2974d
>
> to be printed after repeated initialization of the Ether device, e.g.
> suspend/resume or 'ifconfig' up/down. This is because DMA buffers mapped
> using dma_map_single() in sh_eth_ring_format() and sh_eth_start_xmit() are
> never unmapped. Resolve this problem by unmapping the buffers when freeing
> the descriptor rings; in order to do it right, we'd have to add an extra
> parameter to sh_eth_txfree() (we rename this function to sh_eth_tx_free(),
> while at it).
>
> Based on the commit a47b70ea86bd ("ravb: unmap descriptors when freeing
> rings").
>
> Signed-off-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
>
> ---
> drivers/net/ethernet/renesas/sh_eth.c | 122 ++++++++++++++++++----------------
> 1 file changed, 67 insertions(+), 55 deletions(-)
>
> Index: net/drivers/net/ethernet/renesas/sh_eth.c
> ===================================================================
> --- net.orig/drivers/net/ethernet/renesas/sh_eth.c
> +++ net/drivers/net/ethernet/renesas/sh_eth.c
> @@ -1127,12 +1127,70 @@ static struct mdiobb_ops bb_ops = {
> .get_mdio_data = sh_get_mdio,
> };
>
> +/* free Tx skb function */
> +static int sh_eth_tx_free(struct net_device *ndev, bool sent_only)
> +{
> + struct sh_eth_private *mdp = netdev_priv(ndev);
> + struct sh_eth_txdesc *txdesc;
> + int free_num = 0;
> + int entry;
> + bool sent;
> +
> + for (; mdp->cur_tx - mdp->dirty_tx > 0; mdp->dirty_tx++) {
> + entry = mdp->dirty_tx % mdp->num_tx_ring;
> + txdesc = &mdp->tx_ring[entry];
> + sent = txdesc->status & cpu_to_le32(TD_TACT);
Oops, forgot to reverse the condition here. Which means I haven't goved in
enough testing... :-/
> + if (sent_only && !sent)
> + break;
[...]
MBR, Sergei
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] sh_eth: unmap DMA buffers when freeing rings
2017-04-17 8:57 ` Sergei Shtylyov
@ 2017-04-17 12:45 ` Sergei Shtylyov
0 siblings, 0 replies; 3+ messages in thread
From: Sergei Shtylyov @ 2017-04-17 12:45 UTC (permalink / raw)
To: netdev; +Cc: linux-renesas-soc
Hello!
On 04/17/2017 11:57 AM, Sergei Shtylyov wrote:
>> WARNING: CPU: 0 PID: 1445 at lib/dma-debug.c:519 add_dma_entry+0xe0/0x12c
>> DMA-API: exceeded 7 overlapping mappings of cacheline 0x01b2974d
>>
>> to be printed after repeated initialization of the Ether device, e.g.
>> suspend/resume or 'ifconfig' up/down. This is because DMA buffers mapped
>> using dma_map_single() in sh_eth_ring_format() and sh_eth_start_xmit() are
>> never unmapped. Resolve this problem by unmapping the buffers when freeing
>> the descriptor rings; in order to do it right, we'd have to add an extra
>> parameter to sh_eth_txfree() (we rename this function to sh_eth_tx_free(),
>> while at it).
>>
>> Based on the commit a47b70ea86bd ("ravb: unmap descriptors when freeing
>> rings").
>>
>> Signed-off-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
>>
>> ---
>> drivers/net/ethernet/renesas/sh_eth.c | 122
>> ++++++++++++++++++----------------
>> 1 file changed, 67 insertions(+), 55 deletions(-)
>>
>> Index: net/drivers/net/ethernet/renesas/sh_eth.c
>> ===================================================================
>> --- net.orig/drivers/net/ethernet/renesas/sh_eth.c
>> +++ net/drivers/net/ethernet/renesas/sh_eth.c
>> @@ -1127,12 +1127,70 @@ static struct mdiobb_ops bb_ops = {
>> .get_mdio_data = sh_get_mdio,
>> };
>>
>> +/* free Tx skb function */
>> +static int sh_eth_tx_free(struct net_device *ndev, bool sent_only)
>> +{
>> + struct sh_eth_private *mdp = netdev_priv(ndev);
>> + struct sh_eth_txdesc *txdesc;
>> + int free_num = 0;
>> + int entry;
>> + bool sent;
>> +
>> + for (; mdp->cur_tx - mdp->dirty_tx > 0; mdp->dirty_tx++) {
>> + entry = mdp->dirty_tx % mdp->num_tx_ring;
>> + txdesc = &mdp->tx_ring[entry];
>> + sent = txdesc->status & cpu_to_le32(TD_TACT);
>
> Oops, forgot to reverse the condition here. Which means I haven't goved in
s/goved in/given it/. :-)
> enough testing... :-/
TX statistics remained all 0s with this patch...
>> + if (sent_only && !sent)
>> + break;
[...]
MBR, Sergei
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2017-04-17 12:45 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-04-16 20:01 [PATCH] sh_eth: unmap DMA buffers when freeing rings Sergei Shtylyov
2017-04-17 8:57 ` Sergei Shtylyov
2017-04-17 12:45 ` Sergei Shtylyov
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.