Gigabit Ethernet driver of Topcliff PCH
Subsystems (2)
PATCH v3 1/1 Gigabit Ethernet driver of Topcliff PCH
Hi Fujita, Jili and Stephen
I have modified for your comment.
Please confirm below.
On: Thu, 26 Aug 2010 08:57:58 -0700
Quoted 44 lines · Stephen Hemminger
Quoted 29 lines
(quoted diff) +29
-0+ /* Performs checksum processing */ + /* + * It is because the hardware accelerator does not support a checksum, + * when the received data size is less than 64 bytes. + */ + if ((skb->len < PCH_GBE_SHORT_PKT) && (adapter->tx_csum == TRUE)) { + frame_ctrl |= + PCH_GBE_TXD_CTRL_APAD | PCH_GBE_TXD_CTRL_TCPIP_ACC_OFF; + if (skb->protocol == htons(ETH_P_IP)) { + struct iphdr *iph = ip_hdr(skb); + unsigned int offset; + iph->check = 0; + iph->check = ip_fast_csum((u8 *) iph, iph->ihl); + offset = skb_transport_offset(skb); + if (iph->protocol == IPPROTO_TCP) { + skb->csum = 0; + tcp_hdr(skb)->check = 0; + skb->csum = + skb_checksum(skb, offset, + skb->len - offset, 0); + tcp_hdr(skb)->check = + csum_tcpudp_magic(iph->saddr, + iph->daddr, + skb->len - offset, + IPPROTO_TCP, skb->csum); + } else if (iph->protocol == IPPROTO_UDP) { + skb->csum = 0; + udp_hdr(skb)->check = 0; + skb->csum =Review
Quoted 10 lines
(quoted diff) +10
-0+ skb_checksum(skb, offset, + skb->len - offset, 0); + udp_hdr(skb)->check = + csum_tcpudp_magic(iph->saddr, + iph->daddr, + skb->len - offset, + IPPROTO_UDP, skb->csum); + } + } + }Why not something simpler using
skb_checksum_helpwhich is safer since it handles case of shared skb correctly.
Quoted 8 lines
if (skb->ip_summed = CHECKSUM_COMPLETE &&
skb->len < PCH_GBE_SHORT_PKT) {
frame_ctrl |=
PCH_GBE_TXD_CTRL_APAD | PCH_GBE_TXD_CTRL_TCPIP_ACC_OFF;
skb->ip_summed = CHECKSUM_NONE;
if (skb_checksum_help(skb))
goto drop; /* could not expand header to find space */
}Since a driver performs a checksam Instead of the checksum of HW, It modified as follows.
However, it did not move normally.
So, modification was given back.
+ if ((skb->len < PCH_GBE_SHORT_PKT) && (adapter->tx_csum)) {
+ frame_ctrl |= PCH_GBE_TXD_CTRL_APAD |
+ PCH_GBE_TXD_CTRL_TCPIP_ACC_OFF;
+ if (skb->protocol == htons(ETH_P_IP)) {
+ if ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
+ (ip_hdr(skb)->protocol == IPPROTO_UDP)) {
+ skb->ip_summed = CHECKSUM_NONE;
+ if (skb_checksum_help(skb)) {
+ dev_kfree_skb_any(skb);
+ adapter->stats.tx_dropped++;
+ return;
+ }
+ }
+ }
+ }
--- drivers/net/Kconfig | 12 + drivers/net/Makefile | 1 + drivers/net/pch_gbe/Makefile | 4 + drivers/net/pch_gbe/pch_gbe.h | 661 +++++++++ drivers/net/pch_gbe/pch_gbe_api.c | 245 ++++ drivers/net/pch_gbe/pch_gbe_api.h | 36 + drivers/net/pch_gbe/pch_gbe_ethtool.c | 587 ++++++++ drivers/net/pch_gbe/pch_gbe_main.c | 2526 +++++++++++++++++++++++++++++++++ drivers/net/pch_gbe/pch_gbe_param.c | 501 +++++++ drivers/net/pch_gbe/pch_gbe_phy.c | 274 ++++ drivers/net/pch_gbe/pch_gbe_phy.h | 37 + 11 files changed, 4884 insertions(+), 0 deletions(-) create mode 100644 drivers/net/pch_gbe/Makefile create mode 100644 drivers/net/pch_gbe/pch_gbe.h create mode 100644 drivers/net/pch_gbe/pch_gbe_api.c create mode 100644 drivers/net/pch_gbe/pch_gbe_api.h create mode 100644 drivers/net/pch_gbe/pch_gbe_ethtool.c create mode 100644 drivers/net/pch_gbe/pch_gbe_main.c create mode 100644 drivers/net/pch_gbe/pch_gbe_param.c create mode 100644 drivers/net/pch_gbe/pch_gbe_phy.c create mode 100644 drivers/net/pch_gbe/pch_gbe_phy.h
drivers/net/Kconfig +12 -0
@@ -2502,6 +2502,18 @@ config S6GMAC source "drivers/net/stmmac/Kconfig" +config PCH_GBE + tristate "PCH Gigabit Ethernet" + depends on PCI + ---help--- + This is a gigabit ethernet driver for Topcliff PCH. + Topcliff PCH is the platform controller hub that is used in Intel's + general embedded platform. + Topcliff PCH has Gigabit Ethernet interface. + Using this interface, it is able to access system devices connected + to Gigabit Ethernet. + This driver enables Gigabit Ethernet function. + endif # NETDEV_1000 #
drivers/net/Makefile +1 -0
@@ -295,3 +295,4 @@ obj-$(CONFIG_WIMAX) += wimax/ obj-$(CONFIG_CAIF) += caif/ obj-$(CONFIG_OCTEON_MGMT_ETHERNET) += octeon/ +obj-$(CONFIG_PCH_GBE) += pch_gbe/
drivers/net/pch_gbe/Makefile new file +4 -0
@@ -0,0 +1,4 @@ +obj-$(CONFIG_PCH_GBE) += pch_gbe.o + +pch_gbe-y := pch_gbe_phy.o pch_gbe_ethtool.o pch_gbe_param.o +pch_gbe-y += pch_gbe_api.o pch_gbe_main.o
drivers/net/pch_gbe/pch_gbe.h new file +661 -0
661 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_api.c new file +245 -0
245 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_api.h new file +36 -0
@@ -0,0 +1,36 @@ +/* +* Copyright (C) 1999 - 2010 Intel Corporation. +* Copyright (C) 2010 OKI SEMICONDUCTOR Co., LTD. +* +* This code was derived from the Intel e1000e Linux driver. +* +* This program is free software; you can redistribute it and/or modify +* it under the terms of the GNU General Public License as published by +* the Free Software Foundation; version 2 of the License. +* +* This program is distributed in the hope that it will be useful, +* but WITHOUT ANY WARRANTY; without even the implied warranty of +* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +* GNU General Public License for more details. +* +* You should have received a copy of the GNU General Public License +* along with this program; if not, write to the Free Software +* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307, USA. +*/ +#ifndef _PCH_GBE_API_H_ +#define _PCH_GBE_API_H_ + +#include "pch_gbe_phy.h" + +s32 pch_gbe_hal_setup_init_funcs (struct pch_gbe_hw *hw); +void pch_gbe_hal_get_bus_info (struct pch_gbe_hw *hw); +s32 pch_gbe_hal_init_hw (struct pch_gbe_hw *hw); +s32 pch_gbe_hal_read_phy_reg (struct pch_gbe_hw *hw,u32 offset,u16 *data); +s32 pch_gbe_hal_write_phy_reg (struct pch_gbe_hw *hw,u32 offset,u16 data); +void pch_gbe_hal_phy_hw_reset (struct pch_gbe_hw *hw); +void pch_gbe_hal_phy_sw_reset (struct pch_gbe_hw *hw); +s32 pch_gbe_hal_read_mac_addr (struct pch_gbe_hw *hw); +void pch_gbe_hal_power_up_phy (struct pch_gbe_hw *hw); +void pch_gbe_hal_power_down_phy (struct pch_gbe_hw *hw); + +#endif
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +587 -0
587 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_main.c new file +2526 -0
1,406 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_param.c new file +501 -0
501 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_phy.c new file +274 -0
274 lines not on this page · raw message
drivers/net/pch_gbe/pch_gbe_phy.h new file +37 -0
Lines of this file are not shown. View the raw message.
The rest of this diff is not shown. View the raw message.
Signature · 1 line
1.6.2.5
Part of the diff is not shown here. View the raw message.
drivers/net/pch_gbe/pch_gbe_ethtool.c
Quoted 1 line · Masayuki Ohtake
drivers/net/pch_gbe/pch_gbe_ethtool.c +0
-0
[]
drivers/net/pch_gbe/pch_gbe_ethtool.c line 183
Quoted 11 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +11
-0+static void pch_gbe_get_regs (struct net_device *netdev, +struct ethtool_regs *regs,void *p) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +struct pci_dev *pdev = adapter->pdev; +u32 *regs_buff = p; +u16 i, reg, tmp; + + regs->version = pdev->revision; + regs->version =0x1000000 | (regs->version <<16 ) | pdev->device;
Might be simpler as:
regs->version = 0x1000000 | (__u32)pdev->revision << 16 | pdev->device;
The block below is a bit confusing to me.
drivers/net/pch_gbe/pch_gbe_ethtool.c line 184
Quoted 1 line
drivers/net/pch_gbe/pch_gbe_ethtool.c +1
-0+memset (p,0 ,PCH_GBE_REGS_LEN * (int )sizeof (u32));
It seems the memset is unnecessary as it's completely overwritten below.
drivers/net/pch_gbe/pch_gbe_ethtool.c line 192
Quoted 8 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +8
-0+for (i =0 ; i <PCH_GBE_MAC_REGS_LEN ; i++) + regs_buff[i] =ioread32 (&hw->reg->INT_ST + i); +/* PHY register */ +for (i =PCH_GBE_MAC_REGS_LEN , reg =0 ; reg <PCH_GBE_PHY_REGS_LEN ; + i++, reg++) { +pch_gbe_hal_read_phy_reg (&adapter->hw, reg, &tmp); + regs_buff[i] = tmp; + }
I think i'd be simpler to do something like:
for (i = 0; i < PCH_GBE_MAC_REGS_LEN; i++)
*regs_buff++ = ioread32(&hw->reg->INT_ST + i);
/* PHY register */
for (i = 0; i < PCH_GBE_PHY_REGS_LEN; i++) {
pch_gbe_hal_read_phy_reg(&adapter->hw, i, &tmp);
*regs_buff++ = tmp;
}[]
+void pch_gbe_mac_mar_set(struct pch_gbe_hw *hw, u8 * addr, u32 index)
Quoted 15 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +15
-0+{ +u32 mar_low, mar_high, adrmask; + +pr_debug ("index : 0x%x\n" , index); + +/* +* HW expects these in little endian so we reverse the byte order +* from network order (big endian) to little endian +*/ + mar_high = ((u32 ) addr[0 ] | ((u32 ) addr[1 ] <<8 ) | + ((u32 ) addr[2 ] <<16 ) | ((u32 ) addr[3 ] <<24 )); + mar_low = ((u32 ) addr[4 ] | ((u32 ) addr[5 ] <<8 )); +/* Stop the MAC Address of index. */ + adrmask =ioread32 (&hw->reg->ADDR_MASK); +iowrite32 ((adrmask | (0x0001 << index)), &hw->reg->ADDR_MASK);
drivers/net/pch_gbe/pch_gbe_ethtool.c
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +10
-0+/* wait busy */ +while ((ioread32 (&hw->reg->ADDR_MASK) &PCH_GBE_BUSY )) { +int tmp =0 ; +udelay (20 ); + tmp++; +if (tmp ==1000 ) { +pr_err ("Address mask bit is not cleared\n" ); +break ; + } + }
You need to move the declaration of tmp out of the while.
Do these really need to be busy-waits?
Quoted 6 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +6
-0+/* Set the MAC address to the MAC address 1A/1B register */ +iowrite32 (mar_high, &hw->reg->mac_adr[index].high); +iowrite32 (mar_low, &hw->reg->mac_adr[index].low); +/* Start the MAC address of index */ +iowrite32 ((adrmask & ~(0x0001 << index)), &hw->reg->ADDR_MASK); +/* wait busy */
drivers/net/pch_gbe/pch_gbe_ethtool.c
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +10
-0+while ((ioread32 (&hw->reg->ADDR_MASK) &PCH_GBE_BUSY )) { +int tmp =0 ; +udelay (20 ); + tmp++; +if (tmp ==1000 ) { +pr_err ("Address mask bit is not cleared\n" ); +break ; + } + } +}
Here too. There are more of these too I'm not listing. Perhaps you could review the logic a bit more.
Hi Joe
Thank you for your comments.
My reply comments are included in the following. I will resubmit after modified.
Thanks Ohtake
-----
Date: Thu, 09 Sep 2010 23:32:07 -0700
From: "Joe Perches" <[email protected]>
Quoted 17 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c
Quoted 1 line
drivers/net/pch_gbe/pch_gbe_ethtool.c +0
-0[]
drivers/net/pch_gbe/pch_gbe_ethtool.c line 183
Quoted 11 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +11
-0+static void pch_gbe_get_regs (struct net_device *netdev, +struct ethtool_regs *regs,void *p) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +struct pci_dev *pdev = adapter->pdev; +u32 *regs_buff = p; +u16 i, reg, tmp; + + regs->version = pdev->revision; + regs->version =0x1000000 | (regs->version <<16 ) | pdev->device;Might be simpler as:
regs->version=0x1000000| (__u32)pdev->revision << 16 |pdev->device;
Quoted 1 line
The block below is a bit confusing to me.
[masa]
This will be modified so.
Quoted 4 lines
Quoted 1 line
+
memset(p, 0, PCH_GBE_REGS_LEN * (int)sizeof(u32));It seems the memset is unnecessary as it's completely overwritten below.
[masa]
This will be deleted.
Quoted 12 lines
Quoted 9 lines
+ for (i = 0; i <
PCH_GBE_MAC_REGS_LEN; i++)
+ regs_buff[i] =ioread32(&hw->reg->INT_ST + i);
+ /* PHY register */
+ for (i =PCH_GBE_MAC_REGS_LEN, reg = 0; reg <PCH_GBE_PHY_REGS_LEN;
+ i++, reg++) {
+pch_gbe_hal_read_phy_reg(&adapter->hw, reg, &tmp);
+ regs_buff[i] = tmp;
+ }I think i'd be simpler to do something like:
Quoted 8 lines
for (i = 0; i <
PCH_GBE_MAC_REGS_LEN; i++)
*regs_buff++ =ioread32(&hw->reg->INT_ST + i);/* PHY register */
for (i = 0; i <PCH_GBE_PHY_REGS_LEN; i++) {pch_gbe_hal_read_phy_reg(&adapter->hw, i, &tmp);
*regs_buff++ = tmp;
}
[masa]
This will be modified so.
Quoted 29 lines
+void
pch_gbe_mac_mar_set(struct pch_gbe_hw *hw, u8 * addr, u32 index)Quoted 15 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +15
-0+{ +u32 mar_low, mar_high, adrmask; + +pr_debug ("index : 0x%x\n" , index); + +/* +* HW expects these in little endian so we reverse the byte order +* from network order (big endian) to little endian +*/ + mar_high = ((u32 ) addr[0 ] | ((u32 ) addr[1 ] <<8 ) | + ((u32 ) addr[2 ] <<16 ) | ((u32 ) addr[3 ] <<24 )); + mar_low = ((u32 ) addr[4 ] | ((u32 ) addr[5 ] <<8 )); +/* Stop the MAC Address of index. */ + adrmask =ioread32 (&hw->reg->ADDR_MASK); +iowrite32 ((adrmask | (0x0001 << index)), &hw->reg->ADDR_MASK);drivers/net/pch_gbe/pch_gbe_ethtool.c
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c +10
-0+/* wait busy */ +while ((ioread32 (&hw->reg->ADDR_MASK) &PCH_GBE_BUSY )) { +int tmp =0 ; +udelay (20 ); + tmp++; +if (tmp ==1000 ) { +pr_err ("Address mask bit is not cleared\n" ); +break ; + } + }You need to move the declaration of tmp out of the while.
Quoted 1 line
Do these really need to be busy-waits?
[masa]
The declaration of tmp will be moved.
It is necessary to wait for BUSY by the specification of hardware.
Quoted 18 lines
Quoted 17 lines
+ /* Set the MAC address to the MAC address 1A/1B register */
+iowrite32(mar_high, &hw->reg->mac_adr[index].high);
+iowrite32(mar_low, &hw->reg->mac_adr[index].low);
+ /* Start the MAC address of index */
+iowrite32((adrmask & ~(0x0001 << index)), &hw->reg->ADDR_MASK);
+ /* wait busy */
+ while ((ioread32(&hw->reg->ADDR_MASK)&PCH_GBE_BUSY)) {
+ int tmp = 0;
+ udelay(20);
+ tmp++;
+ if (tmp == 1000) {
+pr_err("Address mask bit is not cleared\n");
+ break;
+ }
+ }
+}
Quoted 3 lines
Here too. There are more of these too I'm not listing. Perhaps you could review the logic a bit more.
[masa]
This will be modified.
Ccing Rafael (please find pch_gbe_resume here)
Quoted 1 line · Masayuki Ohtake
Hi Fujita, Jili and Stephen
Hi, Jili sounds funny :D.
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +0
-0@@ -0,0 +1,587 @@
...
Quoted 4 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +4
-0+static void pch_gbe_get_regs (struct net_device *netdev, +struct ethtool_regs *regs,void *p) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev);
drivers/net/pch_gbe/pch_gbe_ethtool.c line 186
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +10
-0+struct pch_gbe_hw *hw = &adapter->hw; +struct pci_dev *pdev = adapter->pdev; +u32 *regs_buff = p; +u16 i, reg, tmp; + + regs->version = pdev->revision; + regs->version =0x1000000 | (regs->version <<16 ) | pdev->device; +memset (p,0 ,PCH_GBE_REGS_LEN * (int )sizeof (u32)); +for (i =0 ; i <PCH_GBE_MAC_REGS_LEN ; i++) + regs_buff[i] =ioread32 (&hw->reg->INT_ST + i);
Well, you switched to ioremap, but left ioread/iowrite*. I wrote you either use pci_iomap+ioread/iowrite32 or pci_ioremap_bar+readl/writel which is faster. (But can be used iff you know it is an MMIO space).
drivers/net/pch_gbe/pch_gbe_ethtool.c line 546
Quoted 12 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +12
-0+static void pch_gbe_get_ethtool_stats (struct net_device *netdev, +struct ethtool_stats *stats,u64 *data) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +int i; +const struct pch_gbe_stats *gstats = pch_gbe_gstrings_stats; +char *hw_stats = (char *)&adapter->stats; + +pch_gbe_update_stats (adapter); +for (i =0 ; i <PCH_GBE_GLOBAL_STATS_LEN ; i++) { +char *p = hw_stats + gstats->offset; + data[i] = gstats->size ==sizeof (u64) ? *(u64 *)p:(*(u32 *)p);
Actually, how you protect 64-bit stat numbers on 32-bit to be atomic? You change adapter->stats from the irq context and it may need to update 2 words on 32-bit. So instead of having 0x100000000 you can get here 0x1ffffffff or 0x0 since read/write of u64 on 32bit are not atomic and both words may not reach the memory yet. I don't know if that matters -- what are the stats from ethtool used for, but returning invalid values is not nice. Do you need 64bit stats on 32-bit at all? Wouldn't be netdev_stats (they are ulong) enough?
Example:
Given stat = 0xffffffff; on 32-bit stat is composed of stat_hi|stat_lo, compiler generates for instance:
stat_lo++;
if (overflow) (1)
stat_hi++
Then if you read stat from this ethtool function at point (1), you will get 0x0.
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +3
-0+ gstats++; + } +}
...
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +0
-0@@ -0,0 +1,2526 @@
...
Quoted 16 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +16
-0+void pch_gbe_mac_mar_set (struct pch_gbe_hw *hw,u8 * addr,u32 index) +{ +u32 mar_low, mar_high, adrmask; + +pr_debug ("index : 0x%x\n" , index); + +/* +* HW expects these in little endian so we reverse the byte order +* from network order (big endian) to little endian +*/ + mar_high = ((u32 ) addr[0 ] | ((u32 ) addr[1 ] <<8 ) | + ((u32 ) addr[2 ] <<16 ) | ((u32 ) addr[3 ] <<24 )); + mar_low = ((u32 ) addr[4 ] | ((u32 ) addr[5 ] <<8 )); +/* Stop the MAC Address of index. */ + adrmask =ioread32 (&hw->reg->ADDR_MASK); +iowrite32 ((adrmask | (0x0001 << index)), &hw->reg->ADDR_MASK);
drivers/net/pch_gbe/pch_gbe_main.c line 143
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+/* wait busy */ +while ((ioread32 (&hw->reg->ADDR_MASK) &PCH_GBE_BUSY )) { +int tmp =0 ; +udelay (20 ); + tmp++; +if (tmp ==1000 ) { +pr_err ("Address mask bit is not cleared\n" ); +break ; + } + }
Maybe it would make sense to create a helper for that.
Quoted 4 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +4
-0+static void pch_gbe_tx_queue (struct pch_gbe_adapter *adapter, +struct pch_gbe_tx_ring *tx_ring, +struct sk_buff *skb) +{
...
drivers/net/pch_gbe/pch_gbe_main.c line 1047
Quoted 1 line
drivers/net/pch_gbe/pch_gbe_main.c new file +1
-0+ buffer_info->kernel_skb = skb;
Just curious, why you need ->kernel_skb and ->skb in that structure? Can't you use just ->skb?
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+int pch_gbe_setup_tx_resources (struct pch_gbe_adapter *adapter, +struct pch_gbe_tx_ring *tx_ring) +{ +struct pci_dev *pdev = adapter->pdev; +struct pch_gbe_tx_desc *tx_desc; +int size; +int desNo; + + size = (int )sizeof (struct pch_gbe_buffer ) * tx_ring->count; + tx_ring->buffer_info =vmalloc (size);
drivers/net/pch_gbe/pch_gbe_main.c line 1582
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+if (!tx_ring->buffer_info) { +pr_err ("Unable to allocate memory for the buffer infomation\n" ); +return -ENOMEM ; + } +memset (tx_ring->buffer_info,0 , size); + + tx_ring->size = tx_ring->count * (int )sizeof (struct pch_gbe_tx_desc ); + + tx_ring->desc = +pci_alloc_consistent (pdev, tx_ring->size, &tx_ring->dma);
As you can sleep here now, you may use dma_alloc_coherent(..., GFP_KERNEL) which allows for better memory reclaim and will fail in less cases. For two reasons:
1) pci_alloc_consistent is deprecated AFAIK
2) pci_alloc_consistent calls page allocator with GFP_ATOMIC and may fail for huge allocations early. And as I computed last time, you allocate up to order-4 pages here.
The same for RX dir.
drivers/net/pch_gbe/pch_gbe_main.c line 1899
Quoted 7 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+static int pch_gbe_xmit_frame (struct sk_buff *skb,struct net_device *netdev) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_tx_ring *tx_ring = adapter->tx_ring; +unsigned long flags; + +if (unlikely (skb->len <=0 )) {
Can this happen at all? ndev people should tell you better...
drivers/net/pch_gbe/pch_gbe_main.c line 1908
Quoted 9 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +9
-0+dev_kfree_skb_any (skb); +pr_debug ("Return : OK skb len : %d\n" , skb->len); +return NETDEV_TX_OK ; + } +if (unlikely (skb->len > (adapter->hw.mac.max_frame_size -4 ))) { +pr_err ("Transfer length Error: skb len: %d > max: %d\n" , + skb->len, adapter->hw.mac.max_frame_size); + adapter->stats.tx_length_errors++; +return NETDEV_TX_BUSY ;
Ok, then it's not busy (it will try to send the skb in a loop forever). You should free it and return OK, I think. Anyway, properly set MTU should gurantee that, but it is my guess.
drivers/net/pch_gbe/pch_gbe_main.c line 1939
Quoted 7 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+static struct net_device_stats *pch_gbe_get_stats ( +struct net_device *netdev) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); + +/* only return the current stats */ +return &adapter->net_stats;
Ah, you have your own copy of netstats. You may use netdev->stats all over the code.
Quoted 7 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+#ifdef CONFIG_PM +static int pch_gbe_resume (struct pci_dev *pdev) +{ +struct net_device *netdev =pci_get_drvdata (pdev); +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +u32 err;
drivers/net/pch_gbe/pch_gbe_main.c line 2287
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+ +pci_enable_wake (pdev,PCI_D0 ,0 ); +pci_set_power_state (pdev,PCI_D0 ); +pci_restore_state (pdev); + err =pci_enable_device (pdev); +if (err) { +pr_err ("Cannot enable PCI device from suspend\n" ); +return err; + } +pci_set_master (pdev);
Not sure which of those you need. Rafael may tell you more. You should not set power state and restore state, but I might be wrong.
The same you need not to save state from ->suspend.
Quoted 12 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +12
-0+pch_gbe_hal_power_up_phy (hw); +pch_gbe_reset (adapter); +/* Clear wake on lan control and status */ +pch_gbe_mac_set_wol_event (hw,0 ); + +if (netif_running (netdev)) +pch_gbe_up (adapter); +netif_device_attach (netdev); + + return 0; +} +#endif /* CONFIG_PM */
...
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +0
-0@@ -0,0 +1,501 @@
...
Quoted 1 line
+const struct
pch_gbe_opt_listspeed_list[] = {
static? And all below...
Quoted 5 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +5
-0+ { 0, "" }, + { SPEED_10, "" }, + { SPEED_100, "" }, + { SPEED_1000, "" } +};
...
drivers/net/pch_gbe/pch_gbe_param.c line 286
Quoted 7 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +7
-0+static void pch_gbe_check_copper_options (struct pch_gbe_adapter *adapter) +{ +struct pch_gbe_hw *hw = &adapter->hw; +int speed, dplx; + + {/* Speed */ + struct pch_gbe_option opt = {
So may be this static const? And all below...
drivers/net/pch_gbe/pch_gbe_param.c line 296
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +10
-0+ .type = list_option, + .name = "Speed", + .err = "parameter ignored", + .def = 0, + .arg = { .l = { .nr = (int)ARRAY_SIZE(speed_list), + .p = speed_list } } + }; + speed = Speed; +pch_gbe_validate_option (&speed, &opt, adapter); + }
regards,
Signature · 2 lines
js suse labs
Quoted 2 lines · Rafael J. Wysocki
Quoted 1 line · Jiri Slaby
Ccing Rafael (please find
pch_gbe_resumehere)
Thanks.
...
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +12
-0+#ifdef CONFIG_PM +static int pch_gbe_resume (struct pci_dev *pdev) +{ +struct net_device *netdev =pci_get_drvdata (pdev); +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +u32 err; + +pci_enable_wake (pdev,PCI_D0 ,0 ); +pci_set_power_state (pdev,PCI_D0 ); +pci_restore_state (pdev); + err =pci_enable_device (pdev);
Quoted 8 lines
Quoted 8 lines
drivers/net/pch_gbe/pch_gbe_main.c line 2287
Quoted 5 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +5
-0+if (err) { +pr_err ("Cannot enable PCI device from suspend\n" ); +return err; + } +pci_set_master (pdev);Not sure which of those you need. Rafael may tell you more. You should not set power state and restore state, but I might be wrong.
You're basically right. In fact you're supposed to use struct dev_pm_ops for defining PM callbacks and then you don't need do any PCI-specific operations directly in the driver.
All of this is described in detail in Documentation/power/pci.txt and you can have a look into r8169.c for a reference implementation.
Thanks,
Rafael
Hi Rafael
Thank you for your comment.
I will confirm pci.txt and r8169.c.
Thanks Ohtake.
-----
Date: Mon, 13 Sep 2010 00:19:26 +0200
From: "Rafael J. Wysocki" <[email protected]>
Quoted 31 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +12
-0+#ifdef CONFIG_PM +static int pch_gbe_resume (struct pci_dev *pdev) +{ +struct net_device *netdev =pci_get_drvdata (pdev); +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +u32 err; + +pci_enable_wake (pdev,PCI_D0 ,0 ); +pci_set_power_state (pdev,PCI_D0 ); +pci_restore_state (pdev); + err =pci_enable_device (pdev);Quoted 8 lines
Quoted 8 lines
Quoted 5 lines
+ if (err) {
+pr_err("Cannot enable PCI device from suspend\n");
+ return err;
+ }
+pci_set_master(pdev);Not sure which of those you need. Rafael may tell you more. You should not set power state and restore state, but I might be wrong.
You're basically right. In fact you're supposed to use struct
dev_pm_opsfor defining PM callbacks and then you don't need do any PCI-specific operations directly in the driver.All of this is described in detail in
Documentation/power/pci.txtand you can have a look intor8169.cfor a reference implementation.Thanks,
Rafael
Hi Jiri
Thank you for your comments.
My reply comments are included in the following. I will resubmit after modified.
Thanks Ohtake
-----
Date: Fri, 10 Sep 2010 10:19:30 +0200
From: "Jiri Slaby" <[email protected]>
Quoted 6 lines
Ccing Rafael (please find
pch_gbe_resumehere)Quoted 1 line · Masayuki Ohtake
Hi Fujita, Jili and Stephen
Hi, Jili sounds funny :D.
[masa]
Sorry, :-}
Quoted 15 lines
Quoted 4 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +0
-0@@ -0,0 +1,587 @@...
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +10
-0+static void pch_gbe_get_regs (struct net_device *netdev, +struct ethtool_regs *regs,void *p) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +struct pci_dev *pdev = adapter->pdev; +u32 *regs_buff = p; +u16 i, reg, tmp; + + regs->version = pdev->revision;
Quoted 9 lines
Quoted 4 lines
+
regs->version=0x1000000| (regs->version<< 16) |pdev->device;
+memset(p, 0, PCH_GBE_REGS_LEN * (int)sizeof(u32));
+ for (i = 0; i <PCH_GBE_MAC_REGS_LEN; i++)
+ regs_buff[i] =ioread32(&hw->reg->INT_ST + i);Well, you switched to ioremap, but left ioread/iowrite*. I wrote you either use pci_iomap+ioread/iowrite32 or pci_ioremap_bar+readl/writel which is faster. (But can be used iff you know it is an MMIO space).
[masa]
This will be modified. "pci_iomap+ioread/iowrite32" will be used.
Quoted 29 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c line 546
Quoted 12 lines
drivers/net/pch_gbe/pch_gbe_ethtool.c new file +12
-0+static void pch_gbe_get_ethtool_stats (struct net_device *netdev, +struct ethtool_stats *stats,u64 *data) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +int i; +const struct pch_gbe_stats *gstats = pch_gbe_gstrings_stats; +char *hw_stats = (char *)&adapter->stats; + +pch_gbe_update_stats (adapter); +for (i =0 ; i <PCH_GBE_GLOBAL_STATS_LEN ; i++) { +char *p = hw_stats + gstats->offset; + data[i] = gstats->size ==sizeof (u64) ? *(u64 *)p:(*(u32 *)p);Actually, how you protect 64-bit stat numbers on 32-bit to be atomic? You change
adapter->statsfrom the irq context and it may need to update 2 words on 32-bit. So instead of having0x100000000you can get here0x1ffffffffor0x0since read/write of u64 on 32bit are not atomic and both words may not reach the memory yet. I don't know if that matters -- what are the stats from ethtool used for, but returning invalid values is not nice. Do you need 64bit stats on 32-bit at all? Wouldn't benetdev_stats(they are ulong) enough?Example:
Given stat =0xffffffff; on 32-bit stat is composed of stat_hi|stat_lo, compiler generates for instance:
stat_lo++;
if (overflow) (1)
stat_hi++
Quoted 2 lines
Then if you read stat from this ethtool function at point (1), you will get
0x0.
[masa]
This will be changed to 32bit stats .
Quoted 29 lines
Quoted 4 lines
+ gstats++;
+ }
+}...
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +0
-0@@ -0,0 +1,2526 @@...
Quoted 20 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +20
-0+void pch_gbe_mac_mar_set (struct pch_gbe_hw *hw,u8 * addr,u32 index) +{ +u32 mar_low, mar_high, adrmask; + +pr_debug ("index : 0x%x\n" , index); + +/* +* HW expects these in little endian so we reverse the byte order +* from network order (big endian) to little endian +*/ + mar_high = ((u32 ) addr[0 ] | ((u32 ) addr[1 ] <<8 ) | + ((u32 ) addr[2 ] <<16 ) | ((u32 ) addr[3 ] <<24 )); + mar_low = ((u32 ) addr[4 ] | ((u32 ) addr[5 ] <<8 )); +/* Stop the MAC Address of index. */ + adrmask =ioread32 (&hw->reg->ADDR_MASK); +iowrite32 ((adrmask | (0x0001 << index)), &hw->reg->ADDR_MASK); +/* wait busy */ +while ((ioread32 (&hw->reg->ADDR_MASK) &PCH_GBE_BUSY )) { +int tmp =0 ; +udelay (20 );
Quoted 8 lines
Quoted 6 lines
+ tmp++;
+ if (tmp == 1000) {
+pr_err("Address mask bit is not cleared\n");
+ break;
+ }
+ }Maybe it would make sense to create a helper for that.
[masa]
The function for wait will be added.
Quoted 10 lines
Quoted 5 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +4
-0+static void pch_gbe_tx_queue (struct pch_gbe_adapter *adapter, +struct pch_gbe_tx_ring *tx_ring, +struct sk_buff *skb) +{...
Quoted 1 line
+
buffer_info->kernel_skb= skb;Just curious, why you need ->kernel_skb and ->skb in that structure? Can't you use just ->skb?
[masa]
"->kernel_skb " will be deleted.
Quoted 30 lines
Quoted 11 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+int pch_gbe_setup_tx_resources (struct pch_gbe_adapter *adapter, +struct pch_gbe_tx_ring *tx_ring) +{ +struct pci_dev *pdev = adapter->pdev; +struct pch_gbe_tx_desc *tx_desc; +int size; +int desNo; + + size = (int )sizeof (struct pch_gbe_buffer ) * tx_ring->count; + tx_ring->buffer_info =vmalloc (size);drivers/net/pch_gbe/pch_gbe_main.c line 1582
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+if (!tx_ring->buffer_info) { +pr_err ("Unable to allocate memory for the buffer infomation\n" ); +return -ENOMEM ; + } +memset (tx_ring->buffer_info,0 , size); + + tx_ring->size = tx_ring->count * (int )sizeof (struct pch_gbe_tx_desc ); + + tx_ring->desc = +pci_alloc_consistent (pdev, tx_ring->size, &tx_ring->dma);As you can sleep here now, you may use
dma_alloc_coherent(..., GFP_KERNEL)which allows for better memory reclaim and will fail in less cases. For two reasons:
1)pci_alloc_consistentis deprecated AFAIK
2)pci_alloc_consistentcalls page allocator withGFP_ATOMICand may fail for huge allocations early. And as I computed last time, you allocate up to order-4 pages here.
Quoted 1 line
The same for RX dir.
[masa]
This will be modified so.
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c line 1899
Quoted 8 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+static int pch_gbe_xmit_frame (struct sk_buff *skb,struct net_device *netdev) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_tx_ring *tx_ring = adapter->tx_ring; +unsigned long flags; + +if (unlikely (skb->len <=0 )) {Can this happen at all? ndev people should tell you better...
[masa]
This will be deleted.
Quoted 11 lines
Quoted 10 lines
+
dev_kfree_skb_any(skb);
+pr_debug("Return : OK skb len : %d\n", skb->len);
+ returnNETDEV_TX_OK;
+ }
+ if (unlikely(skb->len > (adapter->hw.mac.max_frame_size - 4))) {
+ pr_err("Transfer length Error: skb len: %d > max: %d\n",
+skb->len,adapter->hw.mac.max_frame_size);
+ adapter->stats.tx_length_errors++;
+ returnNETDEV_TX_BUSY;
Quoted 3 lines
Ok, then it's not busy (it will try to send the skb in a loop forever). You should free it and return OK, I think. Anyway, properly set MTU should gurantee that, but it is my guess.
[masa]dev_kfree_skb_any(skb);
return NETDEV_TX_OK;
Quoted 11 lines
drivers/net/pch_gbe/pch_gbe_main.c line 1939
Quoted 8 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+static struct net_device_stats *pch_gbe_get_stats ( +struct net_device *netdev) +{ +struct pch_gbe_adapter *adapter =netdev_priv (netdev); + +/* only return the current stats */ +return &adapter->net_stats;Ah, you have your own copy of netstats. You may use
netdev->statsall over the code.
[masa]
"netdev->stats" will be used
Quoted 22 lines
Quoted 8 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +7
-0+#ifdef CONFIG_PM +static int pch_gbe_resume (struct pci_dev *pdev) +{ +struct net_device *netdev =pci_get_drvdata (pdev); +struct pch_gbe_adapter *adapter =netdev_priv (netdev); +struct pch_gbe_hw *hw = &adapter->hw; +u32 err;drivers/net/pch_gbe/pch_gbe_main.c line 2287
Quoted 10 lines
drivers/net/pch_gbe/pch_gbe_main.c new file +10
-0+ +pci_enable_wake (pdev,PCI_D0 ,0 ); +pci_set_power_state (pdev,PCI_D0 ); +pci_restore_state (pdev); + err =pci_enable_device (pdev); +if (err) { +pr_err ("Cannot enable PCI device from suspend\n" ); +return err; + } +pci_set_master (pdev);Not sure which of those you need. Rafael may tell you more. You should not set power state and restore state, but I might be wrong.
Quoted 2 lines
The same you need not to save state from ->suspend.
[masa]
I confirm Rafael's comment.
Quoted 13 lines
Quoted 12 lines
+
pch_gbe_hal_power_up_phy(hw);
+pch_gbe_reset(adapter);
+ /* Clear wake on lan control and status */
+pch_gbe_mac_set_wol_event(hw, 0);drivers/net/pch_gbe/pch_gbe_main.c new file +8
-0+ +if (netif_running (netdev)) +pch_gbe_up (adapter); +netif_device_attach (netdev); + + return 0; +} +#endif /* CONFIG_PM */ ...
Quoted 7 lines
Quoted 3 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +0
-0@@ -0,0 +1,501 @@...
Quoted 1 line
+const struct
pch_gbe_opt_listspeed_list[] = {static? And all below...
[masa]
This will be modified.
Quoted 7 lines
Quoted 6 lines
+ { 0, "" },
+ {SPEED_10, "" },
+ {SPEED_100, "" },
+ {SPEED_1000, "" }
+};...
Quoted 9 lines
drivers/net/pch_gbe/pch_gbe_param.c line 286
Quoted 7 lines
drivers/net/pch_gbe/pch_gbe_param.c new file +7
-0+static void pch_gbe_check_copper_options (struct pch_gbe_adapter *adapter) +{ +struct pch_gbe_hw *hw = &adapter->hw; +int speed, dplx; + + {/* Speed */ + struct pch_gbe_option opt = {So may be this static const? And all below...
[masa]
This will be modified.