| Message ID | 1341052528-2444-1-git-send-email-jeffrey.t.kirsher@intel.com |
|---|---|
| State | Changes Requested, archived |
| Delegated to: | David Miller |
| Headers | show |
On Sat, 2012-06-30 at 03:35 -0700, Jeff Kirsher wrote: > From: Bruce Allan <bruce.w.allan@intel.com> > > Currently only used when packet split mode is enabled with jumbo frames, > IP payload checksum (for fragmented UDP packets) is mutually exclusive with > receive hashing offload since the hardware uses the same space in the > receive descriptor for the hardware-provided packet checksum and the RSS > hash, respectively. Users currently must disable jumbos when receive > hashing offload is enabled, or vice versa, because of this incompatibility. > Since testing has shown that IP payload checksum does not provide any real > benefit, just remove it so that there is no longer a choice between jumbos > or receive hashing offload but not both as done in other Intel GbE drivers > (e.g. e1000, igb). > > Also, add a missing check for IP checksum error reported by the hardware; > let the stack verify the checksum when this happens. [...] The change to enable RX hashing in 3.4, with this odd restriction seems to have broken most existing systems using jumbo MTU on e1000e. None of the distro scripts or network management daemons will automatically change offload configuration before MTU; how could they know? Therefore this needs to be fixed in 3.5 and 3.4.y, not net-next. Ben.
From: Ben Hutchings <ben@decadent.org.uk> Date: Sat, 30 Jun 2012 22:36:36 +0100 > On Sat, 2012-06-30 at 03:35 -0700, Jeff Kirsher wrote: >> From: Bruce Allan <bruce.w.allan@intel.com> >> >> Currently only used when packet split mode is enabled with jumbo frames, >> IP payload checksum (for fragmented UDP packets) is mutually exclusive with >> receive hashing offload since the hardware uses the same space in the >> receive descriptor for the hardware-provided packet checksum and the RSS >> hash, respectively. Users currently must disable jumbos when receive >> hashing offload is enabled, or vice versa, because of this incompatibility. >> Since testing has shown that IP payload checksum does not provide any real >> benefit, just remove it so that there is no longer a choice between jumbos >> or receive hashing offload but not both as done in other Intel GbE drivers >> (e.g. e1000, igb). >> >> Also, add a missing check for IP checksum error reported by the hardware; >> let the stack verify the checksum when this happens. > [...] > > The change to enable RX hashing in 3.4, with this odd restriction seems > to have broken most existing systems using jumbo MTU on e1000e. None of > the distro scripts or network management daemons will automatically > change offload configuration before MTU; how could they know? > > Therefore this needs to be fixed in 3.5 and 3.4.y, not net-next. Agreed. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sat, 2012-06-30 at 17:37 -0700, David Miller wrote: > From: Ben Hutchings <ben@decadent.org.uk> > Date: Sat, 30 Jun 2012 22:36:36 +0100 > > > On Sat, 2012-06-30 at 03:35 -0700, Jeff Kirsher wrote: > >> From: Bruce Allan <bruce.w.allan@intel.com> > >> > >> Currently only used when packet split mode is enabled with jumbo frames, > >> IP payload checksum (for fragmented UDP packets) is mutually exclusive with > >> receive hashing offload since the hardware uses the same space in the > >> receive descriptor for the hardware-provided packet checksum and the RSS > >> hash, respectively. Users currently must disable jumbos when receive > >> hashing offload is enabled, or vice versa, because of this incompatibility. > >> Since testing has shown that IP payload checksum does not provide any real > >> benefit, just remove it so that there is no longer a choice between jumbos > >> or receive hashing offload but not both as done in other Intel GbE drivers > >> (e.g. e1000, igb). > >> > >> Also, add a missing check for IP checksum error reported by the hardware; > >> let the stack verify the checksum when this happens. > > [...] > > > > The change to enable RX hashing in 3.4, with this odd restriction seems > > to have broken most existing systems using jumbo MTU on e1000e. None of > > the distro scripts or network management daemons will automatically > > change offload configuration before MTU; how could they know? > > > > Therefore this needs to be fixed in 3.5 and 3.4.y, not net-next. > > Agreed. Ok, I will prepare it for net and stable 3.4. I know it will require a backported patch for stable 3.4.y since the current patch only applied to net & net-next. Bruce was wanting to have it applied to net & stable, and I was not sure based on the patch content and description, so I that is why I submitted it for net-next.
diff --git a/drivers/net/ethernet/intel/e1000e/defines.h b/drivers/net/ethernet/intel/e1000e/defines.h index 351a409..76edbc1 100644 --- a/drivers/net/ethernet/intel/e1000e/defines.h +++ b/drivers/net/ethernet/intel/e1000e/defines.h @@ -103,6 +103,7 @@ #define E1000_RXD_ERR_SEQ 0x04 /* Sequence Error */ #define E1000_RXD_ERR_CXE 0x10 /* Carrier Extension Error */ #define E1000_RXD_ERR_TCPE 0x20 /* TCP/UDP Checksum Error */ +#define E1000_RXD_ERR_IPE 0x40 /* IP Checksum Error */ #define E1000_RXD_ERR_RXE 0x80 /* Rx Data Error */ #define E1000_RXD_SPC_VLAN_MASK 0x0FFF /* VLAN ID is in lower 12 bits */ diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/intel/e1000e/netdev.c index ba86b3f..a166efc 100644 --- a/drivers/net/ethernet/intel/e1000e/netdev.c +++ b/drivers/net/ethernet/intel/e1000e/netdev.c @@ -496,7 +496,7 @@ static void e1000_receive_skb(struct e1000_adapter *adapter, * @sk_buff: socket buffer with received data **/ static void e1000_rx_checksum(struct e1000_adapter *adapter, u32 status_err, - __le16 csum, struct sk_buff *skb) + struct sk_buff *skb) { u16 status = (u16)status_err; u8 errors = (u8)(status_err >> 24); @@ -511,8 +511,8 @@ static void e1000_rx_checksum(struct e1000_adapter *adapter, u32 status_err, if (status & E1000_RXD_STAT_IXSM) return; - /* TCP/UDP checksum error bit is set */ - if (errors & E1000_RXD_ERR_TCPE) { + /* TCP/UDP checksum error bit or IP checksum error bit is set */ + if (errors & (E1000_RXD_ERR_TCPE | E1000_RXD_ERR_IPE)) { /* let the stack verify checksum errors */ adapter->hw_csum_err++; return; @@ -523,19 +523,7 @@ static void e1000_rx_checksum(struct e1000_adapter *adapter, u32 status_err, return; /* It must be a TCP or UDP packet with a valid checksum */ - if (status & E1000_RXD_STAT_TCPCS) { - /* TCP checksum is good */ - skb->ip_summed = CHECKSUM_UNNECESSARY; - } else { - /* - * IP fragment with UDP payload - * Hardware complements the payload checksum, so we undo it - * and then put the value in host order for further stack use. - */ - __sum16 sum = (__force __sum16)swab16((__force u16)csum); - skb->csum = csum_unfold(~sum); - skb->ip_summed = CHECKSUM_COMPLETE; - } + skb->ip_summed = CHECKSUM_UNNECESSARY; adapter->hw_csum_good++; } @@ -954,8 +942,7 @@ static bool e1000_clean_rx_irq(struct e1000_ring *rx_ring, int *work_done, skb_put(skb, length); /* Receive Checksum Offload */ - e1000_rx_checksum(adapter, staterr, - rx_desc->wb.lower.hi_dword.csum_ip.csum, skb); + e1000_rx_checksum(adapter, staterr, skb); e1000_rx_hash(netdev, rx_desc->wb.lower.hi_dword.rss, skb); @@ -1341,8 +1328,7 @@ copydone: total_rx_bytes += skb->len; total_rx_packets++; - e1000_rx_checksum(adapter, staterr, - rx_desc->wb.lower.hi_dword.csum_ip.csum, skb); + e1000_rx_checksum(adapter, staterr, skb); e1000_rx_hash(netdev, rx_desc->wb.lower.hi_dword.rss, skb); @@ -1512,9 +1498,8 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_ring *rx_ring, int *work_done, } } - /* Receive Checksum Offload XXX recompute due to CRC strip? */ - e1000_rx_checksum(adapter, staterr, - rx_desc->wb.lower.hi_dword.csum_ip.csum, skb); + /* Receive Checksum Offload */ + e1000_rx_checksum(adapter, staterr, skb); e1000_rx_hash(netdev, rx_desc->wb.lower.hi_dword.rss, skb); @@ -3098,19 +3083,10 @@ static void e1000_configure_rx(struct e1000_adapter *adapter) /* Enable Receive Checksum Offload for TCP and UDP */ rxcsum = er32(RXCSUM); - if (adapter->netdev->features & NETIF_F_RXCSUM) { + if (adapter->netdev->features & NETIF_F_RXCSUM) rxcsum |= E1000_RXCSUM_TUOFL; - - /* - * IPv4 payload checksum for UDP fragments must be - * used in conjunction with packet-split. - */ - if (adapter->rx_ps_pages) - rxcsum |= E1000_RXCSUM_IPPCSE; - } else { + else rxcsum &= ~E1000_RXCSUM_TUOFL; - /* no need to clear IPPCSE as it defaults to 0 */ - } ew32(RXCSUM, rxcsum); if (adapter->hw.mac.type == e1000_pch2lan) { @@ -5241,22 +5217,10 @@ static int e1000_change_mtu(struct net_device *netdev, int new_mtu) int max_frame = new_mtu + ETH_HLEN + ETH_FCS_LEN; /* Jumbo frame support */ - if (max_frame > ETH_FRAME_LEN + ETH_FCS_LEN) { - if (!(adapter->flags & FLAG_HAS_JUMBO_FRAMES)) { - e_err("Jumbo Frames not supported.\n"); - return -EINVAL; - } - - /* - * IP payload checksum (enabled with jumbos/packet-split when - * Rx checksum is enabled) and generation of RSS hash is - * mutually exclusive in the hardware. - */ - if ((netdev->features & NETIF_F_RXCSUM) && - (netdev->features & NETIF_F_RXHASH)) { - e_err("Jumbo frames cannot be enabled when both receive checksum offload and receive hashing are enabled. Disable one of the receive offload features before enabling jumbos.\n"); - return -EINVAL; - } + if ((max_frame > ETH_FRAME_LEN + ETH_FCS_LEN) && + !(adapter->flags & FLAG_HAS_JUMBO_FRAMES)) { + e_err("Jumbo Frames not supported.\n"); + return -EINVAL; } /* Supported frame sizes */ @@ -6030,17 +5994,6 @@ static int e1000_set_features(struct net_device *netdev, NETIF_F_RXALL))) return 0; - /* - * IP payload checksum (enabled with jumbos/packet-split when Rx - * checksum is enabled) and generation of RSS hash is mutually - * exclusive in the hardware. - */ - if (adapter->rx_ps_pages && - (features & NETIF_F_RXCSUM) && (features & NETIF_F_RXHASH)) { - e_err("Enabling both receive checksum offload and receive hashing is not possible with jumbo frames. Disable jumbos or enable only one of the receive offload features.\n"); - return -EINVAL; - } - if (changed & NETIF_F_RXFCS) { if (features & NETIF_F_RXFCS) { adapter->flags2 &= ~FLAG2_CRC_STRIPPING;