e1000: Fixes for packet split related issues
Mallikarjuna R Chilakala [Tue, 4 Oct 2005 11:05:44 +0000 (07:05 -0400)]
Fixes for packet split related issues
  * On platforms where PAGE_SIZE > 4K, driver will use only required number of
    pages compared to always using 3 pages.
  * Packet split won't be used if the PAGE_SIZE is > 16K
  * Adds a statistics counter to splits.
  * Setting the non Null ptr to zero sized buffers to solve packet split
    receive descriptor error
  * When the no of pages needed is calculated, the header buffer is not
    included for a given MTU.

Signed-off-by: Mallikarjuna R Chilakala <mallikarjuna.chilakala@intel.com>
Signed-off-by: Ganesh Venkatesan <ganesh.venkatesan@intel.com>
Signed-off-by: John Ronciak <john.ronciak@intel.com>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>

drivers/net/e1000/e1000.h
drivers/net/e1000/e1000_ethtool.c
drivers/net/e1000/e1000_main.c

index 9b7274b..3f653a9 100644 (file)
@@ -169,8 +169,8 @@ struct e1000_buffer {
        uint16_t next_to_watch;
 };
 
-struct e1000_ps_page { struct page *ps_page[MAX_PS_BUFFERS]; };
-struct e1000_ps_page_dma { uint64_t ps_page_dma[MAX_PS_BUFFERS]; };
+struct e1000_ps_page { struct page *ps_page[PS_PAGE_BUFFERS]; };
+struct e1000_ps_page_dma { uint64_t ps_page_dma[PS_PAGE_BUFFERS]; };
 
 struct e1000_tx_ring {
        /* pointer to the descriptor ring memory */
@@ -300,10 +300,11 @@ struct e1000_adapter {
 
        uint64_t hw_csum_err;
        uint64_t hw_csum_good;
+       uint64_t rx_hdr_split;
        uint32_t rx_int_delay;
        uint32_t rx_abs_int_delay;
        boolean_t rx_csum;
-       boolean_t rx_ps;
+       unsigned int rx_ps_pages;
        uint32_t gorcl;
        uint64_t gorcl_old;
        uint16_t rx_ps_bsize0;
index 6e7e34f..183b583 100644 (file)
@@ -91,7 +91,8 @@ static const struct e1000_stats e1000_gstrings_stats[] = {
        { "tx_flow_control_xoff", E1000_STAT(stats.xofftxc) },
        { "rx_long_byte_count", E1000_STAT(stats.gorcl) },
        { "rx_csum_offload_good", E1000_STAT(hw_csum_good) },
-       { "rx_csum_offload_errors", E1000_STAT(hw_csum_err) }
+       { "rx_csum_offload_errors", E1000_STAT(hw_csum_err) },
+       { "rx_header_split", E1000_STAT(rx_hdr_split) },
 };
 #define E1000_STATS_LEN        \
        sizeof(e1000_gstrings_stats) / sizeof(struct e1000_stats)
index ad92115..090229f 100644 (file)
@@ -1484,12 +1484,16 @@ e1000_setup_all_rx_resources(struct e1000_adapter *adapter)
  * e1000_setup_rctl - configure the receive control registers
  * @adapter: Board private structure
  **/
-
+#define PAGE_USE_COUNT(S) (((S) >> PAGE_SHIFT) + \
+                       (((S) & (PAGE_SIZE - 1)) ? 1 : 0))
 static void
 e1000_setup_rctl(struct e1000_adapter *adapter)
 {
        uint32_t rctl, rfctl;
        uint32_t psrctl = 0;
+#ifdef CONFIG_E1000_PACKET_SPLIT
+       uint32_t pages = 0;
+#endif
 
        rctl = E1000_READ_REG(&adapter->hw, RCTL);
 
@@ -1543,11 +1547,14 @@ e1000_setup_rctl(struct e1000_adapter *adapter)
         * followed by the page buffers.  Therefore, skb->data is
         * sized to hold the largest protocol header.
         */
-       adapter->rx_ps = (adapter->hw.mac_type > e1000_82547_rev_2) 
-                         && (adapter->netdev->mtu 
-                             < ((3 * PAGE_SIZE) + adapter->rx_ps_bsize0));
+       pages = PAGE_USE_COUNT(adapter->netdev->mtu);
+       if ((adapter->hw.mac_type > e1000_82547_rev_2) && (pages <= 3) &&
+           PAGE_SIZE <= 16384)
+               adapter->rx_ps_pages = pages;
+       else
+               adapter->rx_ps_pages = 0;
 #endif
-       if(adapter->rx_ps) {
+       if (adapter->rx_ps_pages) {
                /* Configure extra packet-split registers */
                rfctl = E1000_READ_REG(&adapter->hw, RFCTL);
                rfctl |= E1000_RFCTL_EXTEN;
@@ -1559,12 +1566,19 @@ e1000_setup_rctl(struct e1000_adapter *adapter)
                
                psrctl |= adapter->rx_ps_bsize0 >>
                        E1000_PSRCTL_BSIZE0_SHIFT;
-               psrctl |= PAGE_SIZE >>
-                       E1000_PSRCTL_BSIZE1_SHIFT;
-               psrctl |= PAGE_SIZE <<
-                       E1000_PSRCTL_BSIZE2_SHIFT;
-               psrctl |= PAGE_SIZE <<
-                       E1000_PSRCTL_BSIZE3_SHIFT;
+
+               switch (adapter->rx_ps_pages) {
+               case 3:
+                       psrctl |= PAGE_SIZE <<
+                               E1000_PSRCTL_BSIZE3_SHIFT;
+               case 2:
+                       psrctl |= PAGE_SIZE <<
+                               E1000_PSRCTL_BSIZE2_SHIFT;
+               case 1:
+                       psrctl |= PAGE_SIZE >>
+                               E1000_PSRCTL_BSIZE1_SHIFT;
+                       break;
+               }
 
                E1000_WRITE_REG(&adapter->hw, PSRCTL, psrctl);
        }
@@ -1590,7 +1604,7 @@ e1000_configure_rx(struct e1000_adapter *adapter)
        int i;
 #endif
 
-       if(adapter->rx_ps) {
+       if (adapter->rx_ps_pages) {
                rdlen = adapter->rx_ring[0].count *
                        sizeof(union e1000_rx_desc_packet_split);
                adapter->clean_rx = e1000_clean_rx_irq_ps;
@@ -1700,8 +1714,8 @@ e1000_configure_rx(struct e1000_adapter *adapter)
 
                        /* Enable 82571 IPv4 payload checksum for UDP fragments
                         * Must be used in conjunction with packet-split. */
-                       if((adapter->hw.mac_type > e1000_82547_rev_2) && 
-                          (adapter->rx_ps)) {
+                       if ((hw->mac_type >= e1000_82571) && 
+                          (adapter->rx_ps_pages)) {
                                rxcsum |= E1000_RXCSUM_IPPCSE;
                        }
                } else {
@@ -1906,7 +1920,7 @@ e1000_clean_rx_ring(struct e1000_adapter *adapter,
                        dev_kfree_skb(buffer_info->skb);
                        buffer_info->skb = NULL;
 
-                       for(j = 0; j < PS_PAGE_BUFFERS; j++) {
+                       for(j = 0; j < adapter->rx_ps_pages; j++) {
                                if(!ps_page->ps_page[j]) break;
                                pci_unmap_single(pdev,
                                                 ps_page_dma->ps_page_dma[j],
@@ -3551,7 +3565,7 @@ e1000_clean_rx_irq_ps(struct e1000_adapter *adapter,
                /* Good Receive */
                skb_put(skb, length);
 
-               for(j = 0; j < PS_PAGE_BUFFERS; j++) {
+               for(j = 0; j < adapter->rx_ps_pages; j++) {
                        if(!(length = le16_to_cpu(rx_desc->wb.upper.length[j])))
                                break;
 
@@ -3572,11 +3586,13 @@ e1000_clean_rx_irq_ps(struct e1000_adapter *adapter,
                                  rx_desc->wb.lower.hi_dword.csum_ip.csum, skb);
                skb->protocol = eth_type_trans(skb, netdev);
 
-#ifdef HAVE_RX_ZERO_COPY
                if(likely(rx_desc->wb.upper.header_status &
-                         E1000_RXDPS_HDRSTAT_HDRSP))
+                         E1000_RXDPS_HDRSTAT_HDRSP)) {
+                       adapter->rx_hdr_split++;
+#ifdef HAVE_RX_ZERO_COPY
                        skb_shinfo(skb)->zero_copy = TRUE;
 #endif
+               }
 #ifdef CONFIG_E1000_NAPI
                if(unlikely(adapter->vlgrp && (staterr & E1000_RXD_STAT_VP))) {
                        vlan_hwaccel_receive_skb(skb, adapter->vlgrp,
@@ -3740,22 +3756,26 @@ e1000_alloc_rx_buffers_ps(struct e1000_adapter *adapter,
                rx_desc = E1000_RX_DESC_PS(*rx_ring, i);
 
                for(j = 0; j < PS_PAGE_BUFFERS; j++) {
-                       if(unlikely(!ps_page->ps_page[j])) {
-                               ps_page->ps_page[j] =
-                                       alloc_page(GFP_ATOMIC);
-                               if(unlikely(!ps_page->ps_page[j]))
-                                       goto no_buffers;
-                               ps_page_dma->ps_page_dma[j] =
-                                       pci_map_page(pdev,
-                                                    ps_page->ps_page[j],
-                                                    0, PAGE_SIZE,
-                                                    PCI_DMA_FROMDEVICE);
-                       }
-                       /* Refresh the desc even if buffer_addrs didn't
-                        * change because each write-back erases this info.
-                        */
-                       rx_desc->read.buffer_addr[j+1] =
-                               cpu_to_le64(ps_page_dma->ps_page_dma[j]);
+                       if (j < adapter->rx_ps_pages) {
+                               if (likely(!ps_page->ps_page[j])) {
+                                       ps_page->ps_page[j] =
+                                               alloc_page(GFP_ATOMIC);
+                                       if (unlikely(!ps_page->ps_page[j]))
+                                               goto no_buffers;
+                                       ps_page_dma->ps_page_dma[j] =
+                                               pci_map_page(pdev,
+                                                           ps_page->ps_page[j],
+                                                           0, PAGE_SIZE,
+                                                           PCI_DMA_FROMDEVICE);
+                               }
+                               /* Refresh the desc even if buffer_addrs didn't
+                                * change because each write-back erases 
+                                * this info.
+                                */
+                               rx_desc->read.buffer_addr[j+1] =
+                                    cpu_to_le64(ps_page_dma->ps_page_dma[j]);
+                       } else
+                               rx_desc->read.buffer_addr[j+1] = ~0;
                }
 
                skb = dev_alloc_skb(adapter->rx_ps_bsize0 + NET_IP_ALIGN);