* [PATCH] forcedeth: jumbo frame support
@ 2005-07-16 18:14 Manfred Spraul
2005-07-16 21:09 ` Tim Mattox
0 siblings, 1 reply; 3+ messages in thread
From: Manfred Spraul @ 2005-07-16 18:14 UTC (permalink / raw)
To: Jeff Garzik; +Cc: Netdev
[-- Attachment #1: Type: text/plain, Size: 298 bytes --]
Hi,
Attached is the missing 0.36 patch that adds jumbo frame support for the
nForce versions that support it.
The performance difference is quite huge:
- MTU 1500: 57 MB/sec, around 85% cpu load
- MTU 7200: 92 MB/sec, around 37% cpu load
Signed-Off-By: Manfred Spraul <manfred@colorfullife.com>
[-- Attachment #2: patch-forcedeth-036-jumbo --]
[-- Type: text/plain, Size: 6564 bytes --]
--- 2.6/drivers/net/forcedeth.c 2005-06-28 22:51:26.000000000 +0200
+++ build-2.6/drivers/net/forcedeth.c 2005-06-28 22:51:40.000000000 +0200
@@ -85,6 +85,7 @@
* 0.33: 16 May 2005: Support for MCP51 added.
* 0.34: 18 Jun 2005: Add DEV_NEED_LINKTIMER to all nForce nics.
* 0.35: 26 Jun 2005: Support for MCP55 added.
+ * 0.36: 28 Jul 2005: Add jumbo frame support.
*
* Known bugs:
* We suspect that on some hardware no TX done interrupts are generated.
@@ -96,7 +97,7 @@
* DEV_NEED_TIMERIRQ will not harm you on sane hardware, only generating a few
* superfluous timer interrupts from the nic.
*/
-#define FORCEDETH_VERSION "0.35"
+#define FORCEDETH_VERSION "0.36"
#define DRV_NAME "forcedeth"
#include <linux/module.h>
@@ -379,9 +380,13 @@
#define TX_LIMIT_START 62
/* rx/tx mac addr + type + vlan + align + slack*/
-#define RX_NIC_BUFSIZE (ETH_DATA_LEN + 64)
-/* even more slack */
-#define RX_ALLOC_BUFSIZE (ETH_DATA_LEN + 128)
+#define NV_RX_HEADERS (64)
+/* even more slack. */
+#define NV_RX_ALLOC_PAD (64)
+
+/* maximum mtu size */
+#define NV_PKTLIMIT_1 ETH_DATA_LEN /* hard limit not known */
+#define NV_PKTLIMIT_2 9100 /* Actual limit according to NVidia: 9202 */
#define OOM_REFILL (1+HZ/20)
#define POLL_WAIT (1+HZ/100)
@@ -473,6 +478,7 @@
struct sk_buff *rx_skbuff[RX_RING];
dma_addr_t rx_dma[RX_RING];
unsigned int rx_buf_sz;
+ unsigned int pkt_limit;
struct timer_list oom_kick;
struct timer_list nic_poll;
@@ -792,7 +798,7 @@
nr = refill_rx % RX_RING;
if (np->rx_skbuff[nr] == NULL) {
- skb = dev_alloc_skb(RX_ALLOC_BUFSIZE);
+ skb = dev_alloc_skb(np->rx_buf_sz + NV_RX_ALLOC_PAD);
if (!skb)
break;
@@ -805,7 +811,7 @@
PCI_DMA_FROMDEVICE);
np->rx_ring[nr].PacketBuffer = cpu_to_le32(np->rx_dma[nr]);
wmb();
- np->rx_ring[nr].FlagLen = cpu_to_le32(RX_NIC_BUFSIZE | NV_RX_AVAIL);
+ np->rx_ring[nr].FlagLen = cpu_to_le32(np->rx_buf_sz | NV_RX_AVAIL);
dprintk(KERN_DEBUG "%s: nv_alloc_rx: Packet %d marked as Available\n",
dev->name, refill_rx);
refill_rx++;
@@ -831,19 +837,31 @@
enable_irq(dev->irq);
}
-static int nv_init_ring(struct net_device *dev)
+static void nv_init_rx(struct net_device *dev)
{
struct fe_priv *np = get_nvpriv(dev);
int i;
- np->next_tx = np->nic_tx = 0;
- for (i = 0; i < TX_RING; i++)
- np->tx_ring[i].FlagLen = 0;
-
np->cur_rx = RX_RING;
np->refill_rx = 0;
for (i = 0; i < RX_RING; i++)
np->rx_ring[i].FlagLen = 0;
+}
+
+static void nv_init_tx(struct net_device *dev)
+{
+ struct fe_priv *np = get_nvpriv(dev);
+ int i;
+
+ np->next_tx = np->nic_tx = 0;
+ for (i = 0; i < TX_RING; i++)
+ np->tx_ring[i].FlagLen = 0;
+}
+
+static int nv_init_ring(struct net_device *dev)
+{
+ nv_init_tx(dev);
+ nv_init_rx(dev);
return nv_alloc_rx(dev);
}
@@ -1207,15 +1225,82 @@
}
}
+static void set_bufsize(struct net_device *dev)
+{
+ struct fe_priv *np = netdev_priv(dev);
+
+ if (dev->mtu <= ETH_DATA_LEN)
+ np->rx_buf_sz = ETH_DATA_LEN + NV_RX_HEADERS;
+ else
+ np->rx_buf_sz = dev->mtu + NV_RX_HEADERS;
+}
+
/*
* nv_change_mtu: dev->change_mtu function
* Called with dev_base_lock held for read.
*/
static int nv_change_mtu(struct net_device *dev, int new_mtu)
{
- if (new_mtu > ETH_DATA_LEN)
+ struct fe_priv *np = get_nvpriv(dev);
+ int old_mtu;
+
+ if (new_mtu < 64 || new_mtu > np->pkt_limit)
return -EINVAL;
+
+ old_mtu = dev->mtu;
dev->mtu = new_mtu;
+
+ /* return early if the buffer sizes will not change */
+ if (old_mtu <= ETH_DATA_LEN && new_mtu <= ETH_DATA_LEN)
+ return 0;
+ if (old_mtu == new_mtu)
+ return 0;
+
+ /* synchronized against open : rtnl_lock() held by caller */
+ if (netif_running(dev)) {
+ u8 *base = get_hwbase(dev);
+ /*
+ * It seems that the nic preloads valid ring entries into an
+ * internal buffer. The procedure for flushing everything is
+ * guessed, there is probably a simpler approach.
+ * Changing the MTU is a rare event, it shouldn't matter.
+ */
+ disable_irq(dev->irq);
+ spin_lock_bh(&dev->xmit_lock);
+ spin_lock(&np->lock);
+ /* stop engines */
+ nv_stop_rx(dev);
+ nv_stop_tx(dev);
+ nv_txrx_reset(dev);
+ /* drain rx queue */
+ nv_drain_rx(dev);
+ nv_drain_tx(dev);
+ /* reinit driver view of the rx queue */
+ nv_init_rx(dev);
+ nv_init_tx(dev);
+ /* alloc new rx buffers */
+ set_bufsize(dev);
+ if (nv_alloc_rx(dev)) {
+ if (!np->in_shutdown)
+ mod_timer(&np->oom_kick, jiffies + OOM_REFILL);
+ }
+ /* reinit nic view of the rx queue */
+ writel(np->rx_buf_sz, base + NvRegOffloadConfig);
+ writel((u32) np->ring_addr, base + NvRegRxRingPhysAddr);
+ writel((u32) (np->ring_addr + RX_RING*sizeof(struct ring_desc)), base + NvRegTxRingPhysAddr);
+ writel( ((RX_RING-1) << NVREG_RINGSZ_RXSHIFT) + ((TX_RING-1) << NVREG_RINGSZ_TXSHIFT),
+ base + NvRegRingSizes);
+ pci_push(base);
+ writel(NVREG_TXRXCTL_KICK|np->desc_ver, get_hwbase(dev) + NvRegTxRxControl);
+ pci_push(base);
+
+ /* restart rx engine */
+ nv_start_rx(dev);
+ nv_start_tx(dev);
+ spin_unlock(&np->lock);
+ spin_unlock_bh(&dev->xmit_lock);
+ enable_irq(dev->irq);
+ }
return 0;
}
@@ -1792,6 +1877,7 @@
writel(0, base + NvRegAdapterControl);
/* 2) initialize descriptor rings */
+ set_bufsize(dev);
oom = nv_init_ring(dev);
writel(0, base + NvRegLinkSpeed);
@@ -1837,7 +1923,7 @@
writel(NVREG_MISC1_FORCE | NVREG_MISC1_HD, base + NvRegMisc1);
writel(readl(base + NvRegTransmitterStatus), base + NvRegTransmitterStatus);
writel(NVREG_PFF_ALWAYS, base + NvRegPacketFilterFlags);
- writel(NVREG_OFFLOAD_NORMAL, base + NvRegOffloadConfig);
+ writel(np->rx_buf_sz, base + NvRegOffloadConfig);
writel(readl(base + NvRegReceiverStatus), base + NvRegReceiverStatus);
get_random_bytes(&i, sizeof(i));
@@ -2007,13 +2093,16 @@
/* handle different descriptor versions */
if (pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_1 ||
- pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_2 ||
- pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_3 ||
- pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_12 ||
- pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_13)
+ pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_2 ||
+ pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_3 ||
+ pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_12 ||
+ pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_13) {
np->desc_ver = DESC_VER_1;
- else
+ np->pkt_limit = NV_PKTLIMIT_1;
+ } else {
np->desc_ver = DESC_VER_2;
+ np->pkt_limit = NV_PKTLIMIT_1;
+ }
err = -ENOMEM;
np->base = ioremap(addr, NV_PCI_REGSZ);
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] forcedeth: jumbo frame support
2005-07-16 18:14 [PATCH] forcedeth: jumbo frame support Manfred Spraul
@ 2005-07-16 21:09 ` Tim Mattox
2005-07-16 21:15 ` Manfred Spraul
0 siblings, 1 reply; 3+ messages in thread
From: Tim Mattox @ 2005-07-16 21:09 UTC (permalink / raw)
To: Manfred Spraul; +Cc: Jeff Garzik, Netdev
This is just from a simple read of Manfred's patch, but shouldn't one side of
the if-else use the NV_PKTLIMIT_2 value? If not, why is
np->pkt_limit = NV_PKTLIMIT_1; inside the if-else at all?
On 7/16/05, Manfred Spraul <manfred@colorfullife.com> wrote:
[snip]
> +/* maximum mtu size */
> +#define NV_PKTLIMIT_1 ETH_DATA_LEN /* hard limit not known */
> +#define NV_PKTLIMIT_2 9100 /* Actual limit according to NVidia: 9202 */
[snip]
> @@ -2007,13 +2093,16 @@
>
> /* handle different descriptor versions */
> if (pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_1 ||
> - pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_2 ||
> - pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_3 ||
> - pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_12 ||
> - pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_13)
> + pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_2 ||
> + pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_3 ||
> + pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_12 ||
> + pci_dev->device == PCI_DEVICE_ID_NVIDIA_NVENET_13) {
> np->desc_ver = DESC_VER_1;
> - else
> + np->pkt_limit = NV_PKTLIMIT_1;
> + } else {
> np->desc_ver = DESC_VER_2;
> + np->pkt_limit = NV_PKTLIMIT_1;
> + }
--
Tim Mattox - tmattox@gmail.com
http://homepage.mac.com/tmattox/
I'm a bright... http://www.the-brights.net/
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] forcedeth: jumbo frame support
2005-07-16 21:09 ` Tim Mattox
@ 2005-07-16 21:15 ` Manfred Spraul
0 siblings, 0 replies; 3+ messages in thread
From: Manfred Spraul @ 2005-07-16 21:15 UTC (permalink / raw)
To: Tim Mattox; +Cc: Jeff Garzik, Netdev
Tim Mattox wrote:
>This is just from a simple read of Manfred's patch, but shouldn't one side of
>the if-else use the NV_PKTLIMIT_2 value?
>
It should - and it's already fixed in the following patch for 0.37,
which is already in Jeff's queue.
I didn't want to break the patch in Jeff's queue, thus I submitted the
patch as it was.
It doesn't hurt much: 0.36 adds the jumbo frame support, 0.37 actually
enables it.
--
Manfred
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2005-07-16 21:15 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2005-07-16 18:14 [PATCH] forcedeth: jumbo frame support Manfred Spraul
2005-07-16 21:09 ` Tim Mattox
2005-07-16 21:15 ` Manfred Spraul
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).