acenic.c

linux 内核源代码

/*
 * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
 *           and other Tigon based cards.
 *
 * Copyright 1998-2002 by Jes Sorensen, <[e-mail]>.
 *
 * Thanks to Alteon and 3Com for providing hardware and documentation
 * enabling me to write this driver.
 *
 * A mailing list for discussing the use of this driver has been
 * setup, please subscribe to the lists if you have any questions
 * about the driver. Send mail to [e-mail] to
 * see how to subscribe.
 *
 * 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; either version 2 of the License, or
 * (at your option) any later version.
 *
 * Additional credits:
 *   Pete Wyckoff <[e-mail]>: Initial Linux/Alpha and trace
 *       dump support. The trace dump support has not been
 *       integrated yet however.
 *   Troy Benjegerdes: Big Endian (PPC) patches.
 *   Nate Stahl: Better out of memory handling and stats support.
 *   Aman Singla: Nasty race between interrupt handler and tx code dealing
 *                with 'testing the tx_ret_csm and setting tx_full'
 *   David S. Miller <[e-mail]>: conversion to new PCI dma mapping
 *                                       infrastructure and Sparc support
 *   Pierrick Pinasseau (CERN): For lending me an Ultra 5 to test the
 *                              driver under Linux/Sparc64
 *   Matt Domsch <[e-mail]>: Detect Alteon 1000baseT cards
 *                                       ETHTOOL_GDRVINFO support
 *   Chip Salzenberg <[e-mail]>: Fix race condition between tx
 *                                       handler and close() cleanup.
 *   Ken Aaker <[e-mail]>: Correct check for whether
 *                                       memory mapped IO is enabled to
 *                                       make the driver work on RS/6000.
 *   Takayoshi Kouchi <[e-mail]>: Identifying problem
 *                                       where the driver would disable
 *                                       bus master mode if it had to disable
 *                                       write and invalidate.
 *   Stephen Hack <[e-mail]>: Fixed ace_set_mac_addr for little
 *                                       endian systems.
 *   Val Henson <[e-mail]>:    Reset Jumbo skb producer and
 *                                       rx producer index when
 *                                       flushing the Jumbo ring.
 *   Hans Grobler <[e-mail]>:     Memory leak fixes in the
 *                                       driver init path.
 *   Grant Grundler <[e-mail]>: PCI write posting fixes.
 */

#include <linux/module.h>
#include <linux/moduleparam.h>
#include <linux/version.h>
#include <linux/types.h>
#include <linux/errno.h>
#include <linux/ioport.h>
#include <linux/pci.h>
#include <linux/dma-mapping.h>
#include <linux/kernel.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/skbuff.h>
#include <linux/init.h>
#include <linux/delay.h>
#include <linux/mm.h>
#include <linux/highmem.h>
#include <linux/sockios.h>

#if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
#include <linux/if_vlan.h>
#endif

#ifdef SIOCETHTOOL
#include <linux/ethtool.h>
#endif

#include <net/sock.h>
#include <net/ip.h>

#include <asm/system.h>
#include <asm/io.h>
#include <asm/irq.h>
#include <asm/byteorder.h>
#include <asm/uaccess.h>


#define DRV_NAME "acenic"

#undef INDEX_DEBUG

#ifdef CONFIG_ACENIC_OMIT_TIGON_I
#define ACE_IS_TIGON_I(ap)	0
#define ACE_TX_RING_ENTRIES(ap)	MAX_TX_RING_ENTRIES
#else
#define ACE_IS_TIGON_I(ap)	(ap->version == 1)
#define ACE_TX_RING_ENTRIES(ap)	ap->tx_ring_entries
#endif

#ifndef PCI_VENDOR_ID_ALTEON
#define PCI_VENDOR_ID_ALTEON		0x12ae
#endif
#ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
#define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE  0x0001
#define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
#endif
#ifndef PCI_DEVICE_ID_3COM_3C985
#define PCI_DEVICE_ID_3COM_3C985	0x0001
#endif
#ifndef PCI_VENDOR_ID_NETGEAR
#define PCI_VENDOR_ID_NETGEAR		0x1385
#define PCI_DEVICE_ID_NETGEAR_GA620	0x620a
#endif
#ifndef PCI_DEVICE_ID_NETGEAR_GA620T
#define PCI_DEVICE_ID_NETGEAR_GA620T	0x630a
#endif


/*
 * Farallon used the DEC vendor ID by mistake and they seem not
 * to care - stinky!
 */
#ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
#define PCI_DEVICE_ID_FARALLON_PN9000SX	0x1a
#endif
#ifndef PCI_DEVICE_ID_FARALLON_PN9100T
#define PCI_DEVICE_ID_FARALLON_PN9100T  0xfa
#endif
#ifndef PCI_VENDOR_ID_SGI
#define PCI_VENDOR_ID_SGI		0x10a9
#endif
#ifndef PCI_DEVICE_ID_SGI_ACENIC
#define PCI_DEVICE_ID_SGI_ACENIC	0x0009
#endif

static struct pci_device_id acenic_pci_tbl[] = {
	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_3COM, PCI_DEVICE_ID_3COM_3C985,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620T,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	/*
	 * Farallon used the DEC vendor ID on their cards incorrectly,
	 * then later Alteon's ID.
	 */
	{ PCI_VENDOR_ID_DEC, PCI_DEVICE_ID_FARALLON_PN9000SX,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_FARALLON_PN9100T,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ PCI_VENDOR_ID_SGI, PCI_DEVICE_ID_SGI_ACENIC,
	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
	{ }
};
MODULE_DEVICE_TABLE(pci, acenic_pci_tbl);

#define ace_sync_irq(irq)	synchronize_irq(irq)

#ifndef offset_in_page
#define offset_in_page(ptr)	((unsigned long)(ptr) & ~PAGE_MASK)
#endif

#define ACE_MAX_MOD_PARMS	8
#define BOARD_IDX_STATIC	0
#define BOARD_IDX_OVERFLOW	-1

#if (defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)) && \
	defined(NETIF_F_HW_VLAN_RX)
#define ACENIC_DO_VLAN		1
#define ACE_RCB_VLAN_FLAG	RCB_FLG_VLAN_ASSIST
#else
#define ACENIC_DO_VLAN		0
#define ACE_RCB_VLAN_FLAG	0
#endif

#include "acenic.h"

/*
 * These must be defined before the firmware is included.
 */
#define MAX_TEXT_LEN	96*1024
#define MAX_RODATA_LEN	8*1024
#define MAX_DATA_LEN	2*1024

#include "acenic_firmware.h"

#ifndef tigon2FwReleaseLocal
#define tigon2FwReleaseLocal 0
#endif

/*
 * This driver currently supports Tigon I and Tigon II based cards
 * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
 * GA620. The driver should also work on the SGI, DEC and Farallon
 * versions of the card, however I have not been able to test that
 * myself.
 *
 * This card is really neat, it supports receive hardware checksumming
 * and jumbo frames (up to 9000 bytes) and does a lot of work in the
 * firmware. Also the programming interface is quite neat, except for
 * the parts dealing with the i2c eeprom on the card ;-)
 *
 * Using jumbo frames:
 *
 * To enable jumbo frames, simply specify an mtu between 1500 and 9000
 * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
 * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
 * interface number and <MTU> being the MTU value.
 *
 * Module parameters:
 *
 * When compiled as a loadable module, the driver allows for a number
 * of module parameters to be specified. The driver supports the
 * following module parameters:
 *
 *  trace=<val> - Firmware trace level. This requires special traced
 *                firmware to replace the firmware supplied with
 *                the driver - for debugging purposes only.
 *
 *  link=<val>  - Link state. Normally you want to use the default link
 *                parameters set by the driver. This can be used to
 *                override these in case your switch doesn't negotiate
 *                the link properly. Valid values are:
 *         0x0001 - Force half duplex link.
 *         0x0002 - Do not negotiate line speed with the other end.
 *         0x0010 - 10Mbit/sec link.
 *         0x0020 - 100Mbit/sec link.
 *         0x0040 - 1000Mbit/sec link.
 *         0x0100 - Do not negotiate flow control.
 *         0x0200 - Enable RX flow control Y
 *         0x0400 - Enable TX flow control Y (Tigon II NICs only).
 *                Default value is 0x0270, ie. enable link+flow
 *                control negotiation. Negotiating the highest
 *                possible link speed with RX flow control enabled.
 *
 *                When disabling link speed negotiation, only one link
 *                speed is allowed to be specified!
 *
 *  tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
 *                to wait for more packets to arive before
 *                interrupting the host, from the time the first
 *                packet arrives.
 *
 *  rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
 *                to wait for more packets to arive in the transmit ring,
 *                before interrupting the host, after transmitting the
 *                first packet in the ring.
 *
 *  max_tx_desc=<val> - maximum number of transmit descriptors
 *                (packets) transmitted before interrupting the host.
 *
 *  max_rx_desc=<val> - maximum number of receive descriptors
 *                (packets) received before interrupting the host.
 *
 *  tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
 *                increments of the NIC's on board memory to be used for
 *                transmit and receive buffers. For the 1MB NIC app. 800KB
 *                is available, on the 1/2MB NIC app. 300KB is available.
 *                68KB will always be available as a minimum for both
 *                directions. The default value is a 50/50 split.
 *  dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
 *                operations, default (1) is to always disable this as
 *                that is what Alteon does on NT. I have not been able
 *                to measure any real performance differences with
 *                this on my systems. Set <val>=0 if you want to
 *                enable these operations.
 *
 * If you use more than one NIC, specify the parameters for the
 * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
 * run tracing on NIC #2 but not on NIC #1 and #3.
 *
 * TODO:
 *
 * - Proper multicast support.
 * - NIC dump support.
 * - More tuning parameters.
 *
 * The mini ring is not used under Linux and I am not sure it makes sense
 * to actually use it.
 *
 * New interrupt handler strategy:
 *
 * The old interrupt handler worked using the traditional method of
 * replacing an skbuff with a new one when a packet arrives. However
 * the rx rings do not need to contain a static number of buffer
 * descriptors, thus it makes sense to move the memory allocation out
 * of the main interrupt handler and do it in a bottom half handler
 * and only allocate new buffers when the number of buffers in the
 * ring is below a certain threshold. In order to avoid starving the
 * NIC under heavy load it is however necessary to force allocation
 * when hitting a minimum threshold. The strategy for alloction is as
 * follows:
 *
 *     RX_LOW_BUF_THRES    - allocate buffers in the bottom half
 *     RX_PANIC_LOW_THRES  - we are very low on buffers, allocate
 *                           the buffers in the interrupt handler
 *     RX_RING_THRES       - maximum number of buffers in the rx ring
 *     RX_MINI_THRES       - maximum number of buffers in the mini ring
 *     RX_JUMBO_THRES      - maximum number of buffers in the jumbo ring
 *
 * One advantagous side effect of this allocation approach is that the
 * entire rx processing can be done without holding any spin lock
 * since the rx rings and registers are totally independent of the tx
 * ring and its registers.  This of course includes the kmalloc's of
 * new skb's. Thus start_xmit can run in parallel with rx processing
 * and the memory allocation on SMP systems.
 *
 * Note that running the skb reallocation in a bottom half opens up
 * another can of races which needs to be handled properly. In
 * particular it can happen that the interrupt handler tries to run
 * the reallocation while the bottom half is either running on another
 * CPU or was interrupted on the same CPU. To get around this the
 * driver uses bitops to prevent the reallocation routines from being
 * reentered.
 *
 * TX handling can also be done without holding any spin lock, wheee
 * this is fun! since tx_ret_csm is only written to by the interrupt
 * handler. The case to be aware of is when shutting down the device
 * and cleaning up where it is necessary to make sure that
 * start_xmit() is not running while this is happening. Well DaveM
 * informs me that this case is already protected against ... bye bye
 * Mr. Spin Lock, it was nice to know you.
 *
 * TX interrupts are now partly disabled so the NIC will only generate
 * TX interrupts for the number of coal ticks, not for the number of
 * TX packets in the queue. This should reduce the number of TX only,
 * ie. when no RX processing is done, interrupts seen.
 */

/*
 * Threshold values for RX buffer allocation - the low water marks for
 * when to start refilling the rings are set to 75% of the ring
 * sizes. It seems to make sense to refill the rings entirely from the
 * intrrupt handler once it gets below the panic threshold, that way
 * we don't risk that the refilling is moved to another CPU when the
 * one running the interrupt handler just got the slab code hot in its
 * cache.
 */
#define RX_RING_SIZE		72
#define RX_MINI_SIZE		64
#define RX_JUMBO_SIZE		48

#define RX_PANIC_STD_THRES	16
#define RX_PANIC_STD_REFILL	(3*RX_PANIC_STD_THRES)/2
#define RX_LOW_STD_THRES	(3*RX_RING_SIZE)/4
#define RX_PANIC_MINI_THRES	12
#define RX_PANIC_MINI_REFILL	(3*RX_PANIC_MINI_THRES)/2
#define RX_LOW_MINI_THRES	(3*RX_MINI_SIZE)/4
#define RX_PANIC_JUMBO_THRES	6
#define RX_PANIC_JUMBO_REFILL	(3*RX_PANIC_JUMBO_THRES)/2
#define RX_LOW_JUMBO_THRES	(3*RX_JUMBO_SIZE)/4


/*
 * Size of the mini ring entries, basically these just should be big
 * enough to take TCP ACKs
 */
#define ACE_MINI_SIZE		100

#define ACE_MINI_BUFSIZE	ACE_MINI_SIZE
#define ACE_STD_BUFSIZE		(ACE_STD_MTU + ETH_HLEN + 4)
#define ACE_JUMBO_BUFSIZE	(ACE_JUMBO_MTU + ETH_HLEN + 4)

/*
 * There seems to be a magic difference in the effect between 995 and 996
 * but little difference between 900 and 995 ... no idea why.
 *
 * There is now a default set of tuning parameters which is set, depending
 * on whether or not the user enables Jumbo frames. It's assumed that if
 * Jumbo frames are enabled, the user wants optimal tuning for that case.
 */
#define DEF_TX_COAL		400 /* 996 */
#define DEF_TX_MAX_DESC		60  /* was 40 */
#define DEF_RX_COAL		120 /* 1000 */
#define DEF_RX_MAX_DESC		25
#define DEF_TX_RATIO		21 /* 24 */

#define DEF_JUMBO_TX_COAL	20
#define DEF_JUMBO_TX_MAX_DESC	60
#define DEF_JUMBO_RX_COAL	30
#define DEF_JUMBO_RX_MAX_DESC	6
#define DEF_JUMBO_TX_RATIO	21

#if tigon2FwReleaseLocal < 20001118
/*
 * Standard firmware and early modifications duplicate
 * IRQ load without this flag (coal timer is never reset).
 * Note that with this flag tx_coal should be less than
 * time to xmit full tx ring.
 * 400usec is not so bad for tx ring size of 128.
 */
#define TX_COAL_INTS_ONLY	1	/* worth it */
#else
/*
 * With modified firmware, this is not necessary, but still useful.
 */
#define TX_COAL_INTS_ONLY	1
#endif

#define DEF_TRACE		0
#define DEF_STAT		(2 * TICKS_PER_SEC)


static int link_state[ACE_MAX_MOD_PARMS];
static int trace[ACE_MAX_MOD_PARMS];
static int tx_coal_tick[ACE_MAX_MOD_PARMS];
static int rx_coal_tick[ACE_MAX_MOD_PARMS];
static int max_tx_desc[ACE_MAX_MOD_PARMS];
static int max_rx_desc[ACE_MAX_MOD_PARMS];
static int tx_ratio[ACE_MAX_MOD_PARMS];
static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};

MODULE_AUTHOR("Jes Sorensen <[e-mail]>");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");

module_param_array_named(link, link_state, int, NULL, 0);
module_param_array(trace, int, NULL, 0);
module_param_array(tx_coal_tick, int, NULL, 0);
module_param_array(max_tx_desc, int, NULL, 0);
module_param_array(rx_coal_tick, int, NULL, 0);
module_param_array(max_rx_desc, int, NULL, 0);
module_param_array(tx_ratio, int, NULL, 0);
MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");


static char version[] __devinitdata =
  "acenic.c: v0.92 08/05/2002  Jes Sorensen, [e-mail]\n"
  "                            http://home.cern.ch/~jes/gige/acenic.html\n";

static int ace_get_settings(struct net_device *, struct ethtool_cmd *);
static int ace_set_settings(struct net_device *, struct ethtool_cmd *);
static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);

static const struct ethtool_ops ace_ethtool_ops = {
	.get_settings = ace_get_settings,
	.set_settings = ace_set_settings,
	.get_drvinfo = ace_get_drvinfo,
};

static void ace_watchdog(struct net_device *dev);

static int __devinit acenic_probe_one(struct pci_dev *pdev,
		const struct pci_device_id *id)
{
	struct net_device *dev;
	struct ace_private *ap;
	static int boards_found;

	dev = alloc_etherdev(sizeof(struct ace_private));