dect
/
linux-2.6
Archived
13
0
Fork 0
Code Releases Activity

e1000: Buffer optimizations for small MTU 

Remove multi-descriptor support from legacy recieve path

Add memory usage efficiency by using more correct size descriptors for
small MTU sizes and optimize using LPE for <= 1522 byte frame sizes

An extra performance fix that effected our TCP window size growth
as a receiver.  Set our initial buffer to be 128 bytes instead of 256
to prevent over-socket charge when truesize is computed in the stack.
old way: truesize = 256 + l1 = 256 + 1460 = 1716
new way: truesize = 128 + l1 = 128 + 1460 = 1588
The magic value that we can't cross is 1648.


Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Auke Kok <auke-jan.h.kok@intel.com>
Signed-off-by: John Ronciak <john.ronciak@intel.com>
This commit is contained in:
Auke Kok 2006-04-14 19:05:18 -07:00
parent 6fc7a7eca7
commit 9e2feace1a
2 changed files with 51 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
drivers/net/e1000/e1000.h
View File

@ -114,6 +114,8 @@ struct e1000_adapter;
/* Supported Rx Buffer Sizes */
#define E1000_RXBUFFER_128 128 /* Used for packet split */
#define E1000_RXBUFFER_256 256 /* Used for packet split */
#define E1000_RXBUFFER_512 512
#define E1000_RXBUFFER_1024 1024
#define E1000_RXBUFFER_2048 2048
#define E1000_RXBUFFER_4096 4096
#define E1000_RXBUFFER_8192 8192

85
drivers/net/e1000/e1000_main.c
View File

@ -972,8 +972,8 @@ e1000_sw_init(struct e1000_adapter *adapter)
pci_read_config_word(pdev, PCI_COMMAND, &hw->pci_cmd_word);
adapter->rx_buffer_len = E1000_RXBUFFER_2048;
adapter->rx_ps_bsize0 = E1000_RXBUFFER_256;
adapter->rx_buffer_len = MAXIMUM_ETHERNET_FRAME_SIZE;
adapter->rx_ps_bsize0 = E1000_RXBUFFER_128;
hw->max_frame_size = netdev->mtu +
ENET_HEADER_SIZE + ETHERNET_FCS_SIZE;
hw->min_frame_size = MINIMUM_ETHERNET_FRAME_SIZE;
@ -1599,14 +1599,21 @@ e1000_setup_rctl(struct e1000_adapter *adapter)
rctl |= E1000_RCTL_LPE;
/* Setup buffer sizes */
if (adapter->hw.mac_type >= e1000_82571) {
/* We can now specify buffers in 1K increments.
* BSIZE and BSEX are ignored in this case. */
rctl |= adapter->rx_buffer_len << 0x11;
} else {
rctl &= ~E1000_RCTL_SZ_4096;
rctl |= E1000_RCTL_BSEX;
switch (adapter->rx_buffer_len) {
rctl &= ~E1000_RCTL_SZ_4096;
rctl |= E1000_RCTL_BSEX;
switch (adapter->rx_buffer_len) {
case E1000_RXBUFFER_256:
rctl |= E1000_RCTL_SZ_256;
rctl &= ~E1000_RCTL_BSEX;
break;
case E1000_RXBUFFER_512:
rctl |= E1000_RCTL_SZ_512;
rctl &= ~E1000_RCTL_BSEX;
break;
case E1000_RXBUFFER_1024:
rctl |= E1000_RCTL_SZ_1024;
rctl &= ~E1000_RCTL_BSEX;
break;
case E1000_RXBUFFER_2048:
default:
rctl |= E1000_RCTL_SZ_2048;
@ -1621,7 +1628,6 @@ e1000_setup_rctl(struct e1000_adapter *adapter)
case E1000_RXBUFFER_16384:
rctl |= E1000_RCTL_SZ_16384;
break;
}
}
#ifndef CONFIG_E1000_DISABLE_PACKET_SPLIT
@ -2982,8 +2988,7 @@ e1000_change_mtu(struct net_device *netdev, int new_mtu)
/* Adapter-specific max frame size limits. */
switch (adapter->hw.mac_type) {
case e1000_82542_rev2_0:
case e1000_82542_rev2_1:
case e1000_undefined ... e1000_82542_rev2_1:
if (max_frame > MAXIMUM_ETHERNET_FRAME_SIZE) {
DPRINTK(PROBE, ERR, "Jumbo Frames not supported.\n");
return -EINVAL;
@ -3017,27 +3022,32 @@ e1000_change_mtu(struct net_device *netdev, int new_mtu)
break;
}
/* NOTE: dev_alloc_skb reserves 16 bytes, and typically NET_IP_ALIGN
* means we reserve 2 more, this pushes us to allocate from the next
* larger slab size
* i.e. RXBUFFER_2048 --> size-4096 slab */
if (adapter->hw.mac_type > e1000_82547_rev_2) {
adapter->rx_buffer_len = max_frame;
E1000_ROUNDUP(adapter->rx_buffer_len, 1024);
} else {
if(unlikely((adapter->hw.mac_type < e1000_82543) &&
(max_frame > MAXIMUM_ETHERNET_FRAME_SIZE))) {
DPRINTK(PROBE, ERR, "Jumbo Frames not supported "
"on 82542\n");
return -EINVAL;
} else {
if(max_frame <= E1000_RXBUFFER_2048)
adapter->rx_buffer_len = E1000_RXBUFFER_2048;
else if(max_frame <= E1000_RXBUFFER_4096)
adapter->rx_buffer_len = E1000_RXBUFFER_4096;
else if(max_frame <= E1000_RXBUFFER_8192)
adapter->rx_buffer_len = E1000_RXBUFFER_8192;
else if(max_frame <= E1000_RXBUFFER_16384)
adapter->rx_buffer_len = E1000_RXBUFFER_16384;
}
}
if (max_frame <= E1000_RXBUFFER_256)
adapter->rx_buffer_len = E1000_RXBUFFER_256;
else if (max_frame <= E1000_RXBUFFER_512)
adapter->rx_buffer_len = E1000_RXBUFFER_512;
else if (max_frame <= E1000_RXBUFFER_1024)
adapter->rx_buffer_len = E1000_RXBUFFER_1024;
else if (max_frame <= E1000_RXBUFFER_2048)
adapter->rx_buffer_len = E1000_RXBUFFER_2048;
else if (max_frame <= E1000_RXBUFFER_4096)
adapter->rx_buffer_len = E1000_RXBUFFER_4096;
else if (max_frame <= E1000_RXBUFFER_8192)
adapter->rx_buffer_len = E1000_RXBUFFER_8192;
else if (max_frame <= E1000_RXBUFFER_16384)
adapter->rx_buffer_len = E1000_RXBUFFER_16384;
/* adjust allocation if LPE protects us, and we aren't using SBP */
#define MAXIMUM_ETHERNET_VLAN_SIZE 1522
if (!adapter->hw.tbi_compatibility_on &&
((max_frame == MAXIMUM_ETHERNET_FRAME_SIZE) ||
(max_frame == MAXIMUM_ETHERNET_VLAN_SIZE)))
adapter->rx_buffer_len = MAXIMUM_ETHERNET_VLAN_SIZE;
netdev->mtu = new_mtu;
@ -3568,10 +3578,12 @@ e1000_clean_rx_irq(struct e1000_adapter *adapter,
flags);
length--;
} else {
dev_kfree_skb_irq(skb);
/* recycle */
buffer_info->skb = skb;
goto next_desc;
}
}
} else
skb_put(skb, length);
/* code added for copybreak, this should improve
* performance for small packets with large amounts
@ -3676,6 +3688,7 @@ e1000_clean_rx_irq_ps(struct e1000_adapter *adapter,
i = rx_ring->next_to_clean;
rx_desc = E1000_RX_DESC_PS(*rx_ring, i);
staterr = le32_to_cpu(rx_desc->wb.middle.status_error);
buffer_info = &rx_ring->buffer_info[i];
while (staterr & E1000_RXD_STAT_DD) {
buffer_info = &rx_ring->buffer_info[i];
@ -3736,7 +3749,7 @@ e1000_clean_rx_irq_ps(struct e1000_adapter *adapter,
/* page alloc/put takes too long and effects small packet
* throughput, so unsplit small packets and save the alloc/put*/
if (l1 && ((length + l1) < E1000_CB_LENGTH)) {
if (l1 && ((length + l1) <= adapter->rx_ps_bsize0)) {
u8 *vaddr;
/* there is no documentation about how to call
* kmap_atomic, so we can't hold the mapping